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")];
            bool var_1538 = const()[name = string("op_1538"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1541_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [1]> peak_1_axes_0 = const()[name = string("peak_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_1_cast_fp16 = reduce_max(axes = peak_1_axes_0, keep_dims = var_1538, x = var_1541_cast_fp16)[name = string("peak_1_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = peak_1_cast_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_1537_promoted_to_fp16 = const()[name = string("op_1537_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1545_cast_fp16 = pow(x = x_3_cast_fp16, y = var_1537_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            tensor<int32, [1]> var_1547_axes_0 = const()[name = string("op_1547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1547_keep_dims_0 = const()[name = string("op_1547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1547_cast_fp16 = reduce_mean(axes = var_1547_axes_0, keep_dims = var_1547_keep_dims_0, x = var_1545_cast_fp16)[name = string("op_1547_cast_fp16")];
            fp16 var_1537_promoted_1_to_fp16 = const()[name = string("op_1537_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1548_cast_fp16 = pow(x = peak_1_cast_fp16, y = var_1537_promoted_1_to_fp16)[name = string("op_1548_cast_fp16")];
            fp32 var_1549_epsilon_0 = const()[name = string("op_1549_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1549_cast_fp16 = inverse(epsilon = var_1549_epsilon_0, x = var_1548_cast_fp16)[name = string("op_1549_cast_fp16")];
            fp16 var_1550_to_fp16 = const()[name = string("op_1550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1551_cast_fp16 = mul(x = var_1549_cast_fp16, y = var_1550_to_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1552_cast_fp16 = add(x = var_1547_cast_fp16, y = var_1551_cast_fp16)[name = string("op_1552_cast_fp16")];
            fp32 var_1553_epsilon_0 = const()[name = string("op_1553_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1553_cast_fp16 = rsqrt(epsilon = var_1553_epsilon_0, x = var_1552_cast_fp16)[name = string("op_1553_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1554_cast_fp16 = mul(x = x_3_cast_fp16, y = var_1553_cast_fp16)[name = string("op_1554_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_1554_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_1563 = const()[name = string("op_1563"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1564_cast_fp16 = reshape(shape = var_1563, x = linear_0_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1571 = const()[name = string("op_1571"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_1564_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 16, 16, 128]> var_1574_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<int32, [1]> peak_3_axes_0 = const()[name = string("peak_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_3_cast_fp16 = reduce_max(axes = peak_3_axes_0, keep_dims = var_1571, x = var_1574_cast_fp16)[name = string("peak_3_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = peak_3_cast_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1570_promoted_to_fp16 = const()[name = string("op_1570_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_1578_cast_fp16 = pow(x = x_7_cast_fp16, y = var_1570_promoted_to_fp16)[name = string("op_1578_cast_fp16")];
            tensor<int32, [1]> var_1580_axes_0 = const()[name = string("op_1580_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1580_keep_dims_0 = const()[name = string("op_1580_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1580_cast_fp16 = reduce_mean(axes = var_1580_axes_0, keep_dims = var_1580_keep_dims_0, x = var_1578_cast_fp16)[name = string("op_1580_cast_fp16")];
            fp16 var_1570_promoted_1_to_fp16 = const()[name = string("op_1570_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_1581_cast_fp16 = pow(x = peak_3_cast_fp16, y = var_1570_promoted_1_to_fp16)[name = string("op_1581_cast_fp16")];
            fp32 var_1582_epsilon_0 = const()[name = string("op_1582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_1582_cast_fp16 = inverse(epsilon = var_1582_epsilon_0, x = var_1581_cast_fp16)[name = string("op_1582_cast_fp16")];
            fp16 var_1583_to_fp16 = const()[name = string("op_1583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1584_cast_fp16 = mul(x = var_1582_cast_fp16, y = var_1583_to_fp16)[name = string("op_1584_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_1585_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1584_cast_fp16)[name = string("op_1585_cast_fp16")];
            fp32 var_1586_epsilon_0 = const()[name = string("op_1586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1586_cast_fp16 = rsqrt(epsilon = var_1586_epsilon_0, x = var_1585_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1587_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1586_cast_fp16)[name = string("op_1587_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_13_cast_fp16 = mul(x = var_1587_cast_fp16, y = model_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_13_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_1596 = const()[name = string("op_1596"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1597_cast_fp16 = reshape(shape = var_1596, x = linear_1_cast_fp16)[name = string("op_1597_cast_fp16")];
            tensor<int32, [4]> x_9_perm_0 = const()[name = string("x_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1604 = const()[name = string("op_1604"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = var_1597_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 8, 16, 128]> var_1607_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<int32, [1]> peak_5_axes_0 = const()[name = string("peak_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_5_cast_fp16 = reduce_max(axes = peak_5_axes_0, keep_dims = var_1604, x = var_1607_cast_fp16)[name = string("peak_5_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = peak_5_cast_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1603_promoted_to_fp16 = const()[name = string("op_1603_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_1611_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1603_promoted_to_fp16)[name = string("op_1611_cast_fp16")];
            tensor<int32, [1]> var_1613_axes_0 = const()[name = string("op_1613_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1613_keep_dims_0 = const()[name = string("op_1613_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1613_cast_fp16 = reduce_mean(axes = var_1613_axes_0, keep_dims = var_1613_keep_dims_0, x = var_1611_cast_fp16)[name = string("op_1613_cast_fp16")];
            fp16 var_1603_promoted_1_to_fp16 = const()[name = string("op_1603_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_1614_cast_fp16 = pow(x = peak_5_cast_fp16, y = var_1603_promoted_1_to_fp16)[name = string("op_1614_cast_fp16")];
            fp32 var_1615_epsilon_0 = const()[name = string("op_1615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_1615_cast_fp16 = inverse(epsilon = var_1615_epsilon_0, x = var_1614_cast_fp16)[name = string("op_1615_cast_fp16")];
            fp16 var_1616_to_fp16 = const()[name = string("op_1616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1617_cast_fp16 = mul(x = var_1615_cast_fp16, y = var_1616_to_fp16)[name = string("op_1617_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_1618_cast_fp16 = add(x = var_1613_cast_fp16, y = var_1617_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1619_cast_fp16 = rsqrt(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1620_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1619_cast_fp16)[name = string("op_1620_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_15_cast_fp16 = mul(x = var_1620_cast_fp16, y = model_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_15_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_1629 = const()[name = string("op_1629"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1630_cast_fp16 = reshape(shape = var_1629, x = linear_2_cast_fp16)[name = string("op_1630_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_13_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_13_cast_fp16)[name = string("second_1_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1644_cast_fp16 = mul(x = first_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1644_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1645_cast_fp16 = mul(x = second_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 16, 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, 16, 16, 64]> var_1648_cast_fp16 = mul(x = second_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1649_cast_fp16 = mul(x = first_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1649_cast_fp16")];
            tensor<fp16, [1, 16, 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 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_1653, interleave = q_1_interleave_0, values = (var_1647_cast_fp16, var_1651_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_15_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_15_cast_fp16)[name = string("second_3_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1665_cast_fp16 = mul(x = first_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1665_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1666_cast_fp16 = mul(x = second_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1666_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1668_cast_fp16 = sub(x = var_1665_cast_fp16, y = var_1666_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1669_cast_fp16 = mul(x = second_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1670_cast_fp16 = mul(x = first_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1672_cast_fp16 = add(x = var_1669_cast_fp16, y = var_1670_cast_fp16)[name = string("op_1672_cast_fp16")];
            int32 var_1674 = const()[name = string("op_1674"), 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_1674, interleave = k_1_interleave_0, values = (var_1668_cast_fp16, var_1672_cast_fp16))[name = string("k_1_cast_fp16")];
            tensor<int32, [4]> var_1685_begin_0 = const()[name = string("op_1685_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1685_end_0 = const()[name = string("op_1685_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1685_end_mask_0 = const()[name = string("op_1685_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1685_stride_0 = const()[name = string("op_1685_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1685_cast_fp16 = slice_by_index(begin = var_1685_begin_0, end = var_1685_end_0, end_mask = var_1685_end_mask_0, stride = var_1685_stride_0, x = q_1_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> var_1695_begin_0 = const()[name = string("op_1695_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1695_end_0 = const()[name = string("op_1695_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1695_end_mask_0 = const()[name = string("op_1695_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1695_stride_0 = const()[name = string("op_1695_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1695_cast_fp16 = slice_by_index(begin = var_1695_begin_0, end = var_1695_end_0, end_mask = var_1695_end_mask_0, stride = var_1695_stride_0, x = q_1_cast_fp16)[name = string("op_1695_cast_fp16")];
            int32 var_1697 = const()[name = string("op_1697"), 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_1697, interleave = stacked_1_interleave_0, values = (var_1685_cast_fp16, var_1695_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_1703_begin_0 = const()[name = string("op_1703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1703_end_0 = const()[name = string("op_1703_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_1703_end_mask_0 = const()[name = string("op_1703_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_1703_cast_fp16 = slice_by_index(begin = var_1703_begin_0, end = var_1703_end_0, end_mask = var_1703_end_mask_0, x = read_state_0)[name = string("op_1703_cast_fp16")];
            tensor<int32, [4]> var_1713_begin_0 = const()[name = string("op_1713_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1713_end_0 = const()[name = string("op_1713_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_1713_end_mask_0 = const()[name = string("op_1713_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_1713_cast_fp16 = slice_by_index(begin = var_1713_begin_0, end = var_1713_end_0, end_mask = var_1713_end_mask_0, x = var_1703_cast_fp16)[name = string("op_1713_cast_fp16")];
            bool var_1722_transpose_x_1 = const()[name = string("op_1722_transpose_x_1"), val = bool(false)];
            bool var_1722_transpose_y_1 = const()[name = string("op_1722_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_1722_cast_fp16 = matmul(transpose_x = var_1722_transpose_x_1, transpose_y = var_1722_transpose_y_1, x = var_1713_cast_fp16, y = stacked_1_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<int32, [4]> var_1725_perm_0 = const()[name = string("op_1725_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_1726_to_fp16 = const()[name = string("op_1726_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_1725_cast_fp16 = transpose(perm = var_1725_perm_0, x = var_1722_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 8, 32, 2048]> var_1727_cast_fp16 = mul(x = var_1725_cast_fp16, y = var_1726_to_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_1_cast_fp16 = add(x = var_1727_cast_fp16, y = cacheMask)[name = string("cache_weights_1_cast_fp16")];
            bool var_1733_transpose_x_1 = const()[name = string("op_1733_transpose_x_1"), val = bool(false)];
            bool var_1733_transpose_y_1 = const()[name = string("op_1733_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_1733_cast_fp16 = matmul(transpose_x = var_1733_transpose_x_1, transpose_y = var_1733_transpose_y_1, x = stacked_1_cast_fp16, y = k_1_cast_fp16)[name = string("op_1733_cast_fp16")];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_1735_cast_fp16 = mul(x = var_1733_cast_fp16, y = var_1734_to_fp16)[name = string("op_1735_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_1735_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_1_cast_fp16")];
            int32 var_1739 = const()[name = string("op_1739"), val = int32(-1)];
            bool var_1740_interleave_0 = const()[name = string("op_1740_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_1740_cast_fp16 = concat(axis = var_1739, interleave = var_1740_interleave_0, values = (cache_weights_1_cast_fp16, call_weights_1_cast_fp16))[name = string("op_1740_cast_fp16")];
            int32 var_1741 = const()[name = string("op_1741"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_1_cast_fp16 = softmax(axis = var_1741, x = var_1740_cast_fp16)[name = string("shares_1_cast_fp16")];
            tensor<int32, [4]> var_1748_begin_0 = const()[name = string("op_1748_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1748_end_0 = const()[name = string("op_1748_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_1748_end_mask_0 = const()[name = string("op_1748_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_1748_cast_fp16 = slice_by_index(begin = var_1748_begin_0, end = var_1748_end_0, end_mask = var_1748_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [28, 8, 2064, 128]> read_state_1 = read_state(input = valueCache)[name = string("read_state_1")];
            tensor<int32, [4]> var_1753_begin_0 = const()[name = string("op_1753_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1753_end_0 = const()[name = string("op_1753_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_1753_end_mask_0 = const()[name = string("op_1753_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> 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 = read_state_1)[name = string("op_1753_cast_fp16")];
            tensor<int32, [4]> var_1763_begin_0 = const()[name = string("op_1763_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1763_end_0 = const()[name = string("op_1763_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_1763_end_mask_0 = const()[name = string("op_1763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_1763_cast_fp16 = slice_by_index(begin = var_1763_begin_0, end = var_1763_end_0, end_mask = var_1763_end_mask_0, x = var_1753_cast_fp16)[name = string("op_1763_cast_fp16")];
            bool var_1769_transpose_x_0 = const()[name = string("op_1769_transpose_x_0"), val = bool(false)];
            bool var_1769_transpose_y_0 = const()[name = string("op_1769_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1769_cast_fp16 = matmul(transpose_x = var_1769_transpose_x_0, transpose_y = var_1769_transpose_y_0, x = var_1748_cast_fp16, y = var_1763_cast_fp16)[name = string("op_1769_cast_fp16")];
            tensor<int32, [4]> var_1774_begin_0 = const()[name = string("op_1774_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_1774_end_0 = const()[name = string("op_1774_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_1774_end_mask_0 = const()[name = string("op_1774_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_1774_cast_fp16 = slice_by_index(begin = var_1774_begin_0, end = var_1774_end_0, end_mask = var_1774_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1774_cast_fp16")];
            bool var_1775_transpose_x_0 = const()[name = string("op_1775_transpose_x_0"), val = bool(false)];
            bool var_1775_transpose_y_0 = const()[name = string("op_1775_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_1630_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 8, 32, 128]> var_1775_cast_fp16 = matmul(transpose_x = var_1775_transpose_x_0, transpose_y = var_1775_transpose_y_0, x = var_1774_cast_fp16, y = v_1_cast_fp16)[name = string("op_1775_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_1_cast_fp16 = add(x = var_1769_cast_fp16, y = var_1775_cast_fp16)[name = string("mixed_1_cast_fp16")];
            tensor<int32, [4]> var_1792_begin_0 = const()[name = string("op_1792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1792_end_0 = const()[name = string("op_1792_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_1792_end_mask_0 = const()[name = string("op_1792_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1792_cast_fp16 = slice_by_index(begin = var_1792_begin_0, end = var_1792_end_0, end_mask = var_1792_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1792_cast_fp16")];
            tensor<int32, [4]> var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1807_cast_fp16")];
            int32 var_1809 = const()[name = string("op_1809"), 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_1809, interleave = mixed_3_interleave_0, values = (var_1792_cast_fp16, var_1807_cast_fp16))[name = string("mixed_3_cast_fp16")];
            tensor<int32, [4]> var_1813_perm_0 = const()[name = string("op_1813_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1817 = const()[name = string("op_1817"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_1813_cast_fp16 = transpose(perm = var_1813_perm_0, x = mixed_3_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 16, 2048]> input_3_cast_fp16 = reshape(shape = var_1817, x = var_1813_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_17_cast_fp16 = add(x = x_1_cast_fp16, y = linear_3_cast_fp16)[name = string("x_17_cast_fp16")];
            bool var_1827 = const()[name = string("op_1827"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1830_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_1830_cast_fp16")];
            tensor<int32, [1]> peak_7_axes_0 = const()[name = string("peak_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_7_cast_fp16 = reduce_max(axes = peak_7_axes_0, keep_dims = var_1827, x = var_1830_cast_fp16)[name = string("peak_7_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = peak_7_cast_fp16)[name = string("x_19_cast_fp16")];
            fp16 var_1826_promoted_to_fp16 = const()[name = string("op_1826_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1834_cast_fp16 = pow(x = x_19_cast_fp16, y = var_1826_promoted_to_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [1]> var_1836_axes_0 = const()[name = string("op_1836_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1836_keep_dims_0 = const()[name = string("op_1836_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1836_cast_fp16 = reduce_mean(axes = var_1836_axes_0, keep_dims = var_1836_keep_dims_0, x = var_1834_cast_fp16)[name = string("op_1836_cast_fp16")];
            fp16 var_1826_promoted_1_to_fp16 = const()[name = string("op_1826_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1837_cast_fp16 = pow(x = peak_7_cast_fp16, y = var_1826_promoted_1_to_fp16)[name = string("op_1837_cast_fp16")];
            fp32 var_1838_epsilon_0 = const()[name = string("op_1838_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1838_cast_fp16 = inverse(epsilon = var_1838_epsilon_0, x = var_1837_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 = mul(x = var_1838_cast_fp16, y = var_1839_to_fp16)[name = string("op_1840_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1841_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1840_cast_fp16)[name = string("op_1841_cast_fp16")];
            fp32 var_1842_epsilon_0 = const()[name = string("op_1842_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1842_cast_fp16 = rsqrt(epsilon = var_1842_epsilon_0, x = var_1841_cast_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1843_cast_fp16 = mul(x = x_19_cast_fp16, y = var_1842_cast_fp16)[name = string("op_1843_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_1843_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_1848_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_1848_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_1848_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_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_6_cast_fp16)[name = string("x_21_cast_fp16")];
            bool var_1861 = const()[name = string("op_1861"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1864_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1864_cast_fp16")];
            tensor<int32, [1]> peak_9_axes_0 = const()[name = string("peak_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_9_cast_fp16 = reduce_max(axes = peak_9_axes_0, keep_dims = var_1861, x = var_1864_cast_fp16)[name = string("peak_9_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = peak_9_cast_fp16)[name = string("x_23_cast_fp16")];
            fp16 var_1860_promoted_to_fp16 = const()[name = string("op_1860_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1868_cast_fp16 = pow(x = x_23_cast_fp16, y = var_1860_promoted_to_fp16)[name = string("op_1868_cast_fp16")];
            tensor<int32, [1]> var_1870_axes_0 = const()[name = string("op_1870_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1870_keep_dims_0 = const()[name = string("op_1870_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1870_cast_fp16 = reduce_mean(axes = var_1870_axes_0, keep_dims = var_1870_keep_dims_0, x = var_1868_cast_fp16)[name = string("op_1870_cast_fp16")];
            fp16 var_1860_promoted_1_to_fp16 = const()[name = string("op_1860_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1871_cast_fp16 = pow(x = peak_9_cast_fp16, y = var_1860_promoted_1_to_fp16)[name = string("op_1871_cast_fp16")];
            fp32 var_1872_epsilon_0 = const()[name = string("op_1872_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1872_cast_fp16 = inverse(epsilon = var_1872_epsilon_0, x = var_1871_cast_fp16)[name = string("op_1872_cast_fp16")];
            fp16 var_1873_to_fp16 = const()[name = string("op_1873_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1874_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1873_to_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1875_cast_fp16 = add(x = var_1870_cast_fp16, y = var_1874_cast_fp16)[name = string("op_1875_cast_fp16")];
            fp32 var_1876_epsilon_0 = const()[name = string("op_1876_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1876_cast_fp16 = rsqrt(epsilon = var_1876_epsilon_0, x = var_1875_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1877_cast_fp16 = mul(x = x_23_cast_fp16, y = var_1876_cast_fp16)[name = string("op_1877_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_1877_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_1886 = const()[name = string("op_1886"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1887_cast_fp16 = reshape(shape = var_1886, x = linear_7_cast_fp16)[name = string("op_1887_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1894 = const()[name = string("op_1894"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_1887_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 16, 16, 128]> var_1897_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1897_cast_fp16")];
            tensor<int32, [1]> peak_11_axes_0 = const()[name = string("peak_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_11_cast_fp16 = reduce_max(axes = peak_11_axes_0, keep_dims = var_1894, x = var_1897_cast_fp16)[name = string("peak_11_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = peak_11_cast_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_1893_promoted_to_fp16 = const()[name = string("op_1893_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_1901_cast_fp16 = pow(x = x_27_cast_fp16, y = var_1893_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_1903_axes_0 = const()[name = string("op_1903_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1903_keep_dims_0 = const()[name = string("op_1903_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1903_cast_fp16 = reduce_mean(axes = var_1903_axes_0, keep_dims = var_1903_keep_dims_0, x = var_1901_cast_fp16)[name = string("op_1903_cast_fp16")];
            fp16 var_1893_promoted_1_to_fp16 = const()[name = string("op_1893_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_1904_cast_fp16 = pow(x = peak_11_cast_fp16, y = var_1893_promoted_1_to_fp16)[name = string("op_1904_cast_fp16")];
            fp32 var_1905_epsilon_0 = const()[name = string("op_1905_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_1905_cast_fp16 = inverse(epsilon = var_1905_epsilon_0, x = var_1904_cast_fp16)[name = string("op_1905_cast_fp16")];
            fp16 var_1906_to_fp16 = const()[name = string("op_1906_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1907_cast_fp16 = mul(x = var_1905_cast_fp16, y = var_1906_to_fp16)[name = string("op_1907_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_1908_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1907_cast_fp16)[name = string("op_1908_cast_fp16")];
            fp32 var_1909_epsilon_0 = const()[name = string("op_1909_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1909_cast_fp16 = rsqrt(epsilon = var_1909_epsilon_0, x = var_1908_cast_fp16)[name = string("op_1909_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1910_cast_fp16 = mul(x = x_27_cast_fp16, y = var_1909_cast_fp16)[name = string("op_1910_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_33_cast_fp16 = mul(x = var_1910_cast_fp16, y = model_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_33_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_1919 = const()[name = string("op_1919"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1920_cast_fp16 = reshape(shape = var_1919, x = linear_8_cast_fp16)[name = string("op_1920_cast_fp16")];
            tensor<int32, [4]> x_29_perm_0 = const()[name = string("x_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1927 = const()[name = string("op_1927"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_29_cast_fp16 = transpose(perm = x_29_perm_0, x = var_1920_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 8, 16, 128]> var_1930_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1930_cast_fp16")];
            tensor<int32, [1]> peak_13_axes_0 = const()[name = string("peak_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_13_cast_fp16 = reduce_max(axes = peak_13_axes_0, keep_dims = var_1927, x = var_1930_cast_fp16)[name = string("peak_13_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = peak_13_cast_fp16)[name = string("x_31_cast_fp16")];
            fp16 var_1926_promoted_to_fp16 = const()[name = string("op_1926_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_1934_cast_fp16 = pow(x = x_31_cast_fp16, y = var_1926_promoted_to_fp16)[name = string("op_1934_cast_fp16")];
            tensor<int32, [1]> var_1936_axes_0 = const()[name = string("op_1936_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1936_keep_dims_0 = const()[name = string("op_1936_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1936_cast_fp16 = reduce_mean(axes = var_1936_axes_0, keep_dims = var_1936_keep_dims_0, x = var_1934_cast_fp16)[name = string("op_1936_cast_fp16")];
            fp16 var_1926_promoted_1_to_fp16 = const()[name = string("op_1926_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_1937_cast_fp16 = pow(x = peak_13_cast_fp16, y = var_1926_promoted_1_to_fp16)[name = string("op_1937_cast_fp16")];
            fp32 var_1938_epsilon_0 = const()[name = string("op_1938_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_1938_cast_fp16 = inverse(epsilon = var_1938_epsilon_0, x = var_1937_cast_fp16)[name = string("op_1938_cast_fp16")];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1940_cast_fp16 = mul(x = var_1938_cast_fp16, y = var_1939_to_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_1941_cast_fp16 = add(x = var_1936_cast_fp16, y = var_1940_cast_fp16)[name = string("op_1941_cast_fp16")];
            fp32 var_1942_epsilon_0 = const()[name = string("op_1942_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1942_cast_fp16 = rsqrt(epsilon = var_1942_epsilon_0, x = var_1941_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1943_cast_fp16 = mul(x = x_31_cast_fp16, y = var_1942_cast_fp16)[name = string("op_1943_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_35_cast_fp16 = mul(x = var_1943_cast_fp16, y = model_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_35_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_1952 = const()[name = string("op_1952"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1953_cast_fp16 = reshape(shape = var_1952, x = linear_9_cast_fp16)[name = string("op_1953_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_33_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_33_cast_fp16)[name = string("second_5_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1967_cast_fp16 = mul(x = first_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1968_cast_fp16 = mul(x = second_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1970_cast_fp16 = sub(x = var_1967_cast_fp16, y = var_1968_cast_fp16)[name = string("op_1970_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1971_cast_fp16 = mul(x = second_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1972_cast_fp16 = mul(x = first_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1972_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1974_cast_fp16 = add(x = var_1971_cast_fp16, y = var_1972_cast_fp16)[name = string("op_1974_cast_fp16")];
            int32 var_1976 = const()[name = string("op_1976"), 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_1976, interleave = q_3_interleave_0, values = (var_1970_cast_fp16, var_1974_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_35_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_35_cast_fp16)[name = string("second_7_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1988_cast_fp16 = mul(x = first_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1988_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1989_cast_fp16 = mul(x = second_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1989_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1991_cast_fp16 = sub(x = var_1988_cast_fp16, y = var_1989_cast_fp16)[name = string("op_1991_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1992_cast_fp16 = mul(x = second_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1992_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1993_cast_fp16 = mul(x = first_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1993_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1995_cast_fp16 = add(x = var_1992_cast_fp16, y = var_1993_cast_fp16)[name = string("op_1995_cast_fp16")];
            int32 var_1997 = const()[name = string("op_1997"), 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_1997, interleave = k_3_interleave_0, values = (var_1991_cast_fp16, var_1995_cast_fp16))[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> var_2008_begin_0 = const()[name = string("op_2008_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2008_end_0 = const()[name = string("op_2008_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2008_end_mask_0 = const()[name = string("op_2008_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2008_stride_0 = const()[name = string("op_2008_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2008_cast_fp16 = slice_by_index(begin = var_2008_begin_0, end = var_2008_end_0, end_mask = var_2008_end_mask_0, stride = var_2008_stride_0, x = q_3_cast_fp16)[name = string("op_2008_cast_fp16")];
            tensor<int32, [4]> var_2018_begin_0 = const()[name = string("op_2018_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2018_end_0 = const()[name = string("op_2018_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2018_end_mask_0 = const()[name = string("op_2018_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2018_stride_0 = const()[name = string("op_2018_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2018_cast_fp16 = slice_by_index(begin = var_2018_begin_0, end = var_2018_end_0, end_mask = var_2018_end_mask_0, stride = var_2018_stride_0, x = q_3_cast_fp16)[name = string("op_2018_cast_fp16")];
            int32 var_2020 = const()[name = string("op_2020"), 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_2020, interleave = stacked_3_interleave_0, values = (var_2008_cast_fp16, var_2018_cast_fp16))[name = string("stacked_3_cast_fp16")];
            tensor<int32, [4]> var_2026_begin_0 = const()[name = string("op_2026_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2026_end_0 = const()[name = string("op_2026_end_0"), val = tensor<int32, [4]>([2, 8, 2064, 128])];
            tensor<bool, [4]> var_2026_end_mask_0 = const()[name = string("op_2026_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2026_cast_fp16 = slice_by_index(begin = var_2026_begin_0, end = var_2026_end_0, end_mask = var_2026_end_mask_0, x = read_state_0)[name = string("op_2026_cast_fp16")];
            tensor<int32, [4]> var_2036_begin_0 = const()[name = string("op_2036_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2036_end_0 = const()[name = string("op_2036_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2036_end_mask_0 = const()[name = string("op_2036_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2036_cast_fp16 = slice_by_index(begin = var_2036_begin_0, end = var_2036_end_0, end_mask = var_2036_end_mask_0, x = var_2026_cast_fp16)[name = string("op_2036_cast_fp16")];
            bool var_2045_transpose_x_1 = const()[name = string("op_2045_transpose_x_1"), val = bool(false)];
            bool var_2045_transpose_y_1 = const()[name = string("op_2045_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_2045_cast_fp16 = matmul(transpose_x = var_2045_transpose_x_1, transpose_y = var_2045_transpose_y_1, x = var_2036_cast_fp16, y = stacked_3_cast_fp16)[name = string("op_2045_cast_fp16")];
            tensor<int32, [4]> var_2048_perm_0 = const()[name = string("op_2048_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2049_to_fp16 = const()[name = string("op_2049_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_2048_cast_fp16 = transpose(perm = var_2048_perm_0, x = var_2045_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 32, 2048]> var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = string("op_2050_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_3_cast_fp16 = add(x = var_2050_cast_fp16, y = cacheMask)[name = string("cache_weights_3_cast_fp16")];
            bool var_2056_transpose_x_1 = const()[name = string("op_2056_transpose_x_1"), val = bool(false)];
            bool var_2056_transpose_y_1 = const()[name = string("op_2056_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2056_cast_fp16 = matmul(transpose_x = var_2056_transpose_x_1, transpose_y = var_2056_transpose_y_1, x = stacked_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_2056_cast_fp16")];
            fp16 var_2057_to_fp16 = const()[name = string("op_2057_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2058_cast_fp16 = mul(x = var_2056_cast_fp16, y = var_2057_to_fp16)[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_3_cast_fp16 = add(x = var_2058_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_3_cast_fp16")];
            int32 var_2062 = const()[name = string("op_2062"), val = int32(-1)];
            bool var_2063_interleave_0 = const()[name = string("op_2063_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_2063_cast_fp16 = concat(axis = var_2062, interleave = var_2063_interleave_0, values = (cache_weights_3_cast_fp16, call_weights_3_cast_fp16))[name = string("op_2063_cast_fp16")];
            int32 var_2064 = const()[name = string("op_2064"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_3_cast_fp16 = softmax(axis = var_2064, x = var_2063_cast_fp16)[name = string("shares_3_cast_fp16")];
            tensor<int32, [4]> var_2071_begin_0 = const()[name = string("op_2071_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2071_end_0 = const()[name = string("op_2071_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_2071_end_mask_0 = const()[name = string("op_2071_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_2071_cast_fp16 = slice_by_index(begin = var_2071_begin_0, end = var_2071_end_0, end_mask = var_2071_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2071_cast_fp16")];
            tensor<int32, [4]> var_2076_begin_0 = const()[name = string("op_2076_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2076_end_0 = const()[name = string("op_2076_end_0"), val = tensor<int32, [4]>([2, 8, 2064, 128])];
            tensor<bool, [4]> var_2076_end_mask_0 = const()[name = string("op_2076_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2076_cast_fp16 = slice_by_index(begin = var_2076_begin_0, end = var_2076_end_0, end_mask = var_2076_end_mask_0, x = read_state_1)[name = string("op_2076_cast_fp16")];
            tensor<int32, [4]> var_2086_begin_0 = const()[name = string("op_2086_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2086_end_0 = const()[name = string("op_2086_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2086_end_mask_0 = const()[name = string("op_2086_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2086_cast_fp16 = slice_by_index(begin = var_2086_begin_0, end = var_2086_end_0, end_mask = var_2086_end_mask_0, x = var_2076_cast_fp16)[name = string("op_2086_cast_fp16")];
            bool var_2092_transpose_x_0 = const()[name = string("op_2092_transpose_x_0"), val = bool(false)];
            bool var_2092_transpose_y_0 = const()[name = string("op_2092_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2092_cast_fp16 = matmul(transpose_x = var_2092_transpose_x_0, transpose_y = var_2092_transpose_y_0, x = var_2071_cast_fp16, y = var_2086_cast_fp16)[name = string("op_2092_cast_fp16")];
            tensor<int32, [4]> var_2097_begin_0 = const()[name = string("op_2097_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_2097_end_0 = const()[name = string("op_2097_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_2097_end_mask_0 = const()[name = string("op_2097_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2097_cast_fp16 = slice_by_index(begin = var_2097_begin_0, end = var_2097_end_0, end_mask = var_2097_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2097_cast_fp16")];
            bool var_2098_transpose_x_0 = const()[name = string("op_2098_transpose_x_0"), val = bool(false)];
            bool var_2098_transpose_y_0 = const()[name = string("op_2098_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_1953_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 32, 128]> var_2098_cast_fp16 = matmul(transpose_x = var_2098_transpose_x_0, transpose_y = var_2098_transpose_y_0, x = var_2097_cast_fp16, y = v_3_cast_fp16)[name = string("op_2098_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_5_cast_fp16 = add(x = var_2092_cast_fp16, y = var_2098_cast_fp16)[name = string("mixed_5_cast_fp16")];
            tensor<int32, [4]> var_2115_begin_0 = const()[name = string("op_2115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2115_end_0 = const()[name = string("op_2115_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2115_end_mask_0 = const()[name = string("op_2115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2115_cast_fp16 = slice_by_index(begin = var_2115_begin_0, end = var_2115_end_0, end_mask = var_2115_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<int32, [4]> var_2130_begin_0 = const()[name = string("op_2130_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2130_end_0 = const()[name = string("op_2130_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2130_end_mask_0 = const()[name = string("op_2130_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2130_cast_fp16 = slice_by_index(begin = var_2130_begin_0, end = var_2130_end_0, end_mask = var_2130_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2130_cast_fp16")];
            int32 var_2132 = const()[name = string("op_2132"), 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_2132, interleave = mixed_7_interleave_0, values = (var_2115_cast_fp16, var_2130_cast_fp16))[name = string("mixed_7_cast_fp16")];
            tensor<int32, [4]> var_2136_perm_0 = const()[name = string("op_2136_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2140 = const()[name = string("op_2140"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2136_cast_fp16 = transpose(perm = var_2136_perm_0, x = mixed_7_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 16, 2048]> input_13_cast_fp16 = reshape(shape = var_2140, x = var_2136_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_37_cast_fp16 = add(x = x_21_cast_fp16, y = linear_10_cast_fp16)[name = string("x_37_cast_fp16")];
            bool var_2150 = const()[name = string("op_2150"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2153_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2153_cast_fp16")];
            tensor<int32, [1]> peak_15_axes_0 = const()[name = string("peak_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_15_cast_fp16 = reduce_max(axes = peak_15_axes_0, keep_dims = var_2150, x = var_2153_cast_fp16)[name = string("peak_15_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = peak_15_cast_fp16)[name = string("x_39_cast_fp16")];
            fp16 var_2149_promoted_to_fp16 = const()[name = string("op_2149_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2157_cast_fp16 = pow(x = x_39_cast_fp16, y = var_2149_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, 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_2149_promoted_1_to_fp16 = const()[name = string("op_2149_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2160_cast_fp16 = pow(x = peak_15_cast_fp16, y = var_2149_promoted_1_to_fp16)[name = string("op_2160_cast_fp16")];
            fp32 var_2161_epsilon_0 = const()[name = string("op_2161_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2161_cast_fp16 = inverse(epsilon = var_2161_epsilon_0, x = var_2160_cast_fp16)[name = string("op_2161_cast_fp16")];
            fp16 var_2162_to_fp16 = const()[name = string("op_2162_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2163_cast_fp16 = mul(x = var_2161_cast_fp16, y = var_2162_to_fp16)[name = string("op_2163_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2164_cast_fp16 = add(x = var_2159_cast_fp16, y = var_2163_cast_fp16)[name = string("op_2164_cast_fp16")];
            fp32 var_2165_epsilon_0 = const()[name = string("op_2165_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2165_cast_fp16 = rsqrt(epsilon = var_2165_epsilon_0, x = var_2164_cast_fp16)[name = string("op_2165_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2166_cast_fp16 = mul(x = x_39_cast_fp16, y = var_2165_cast_fp16)[name = string("op_2166_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_2166_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_2171_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_2171_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_2171_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_41_cast_fp16 = add(x = x_37_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            bool var_2184 = const()[name = string("op_2184"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2187_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> peak_17_axes_0 = const()[name = string("peak_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_17_cast_fp16 = reduce_max(axes = peak_17_axes_0, keep_dims = var_2184, x = var_2187_cast_fp16)[name = string("peak_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = peak_17_cast_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_2183_promoted_to_fp16 = const()[name = string("op_2183_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2191_cast_fp16 = pow(x = x_43_cast_fp16, y = var_2183_promoted_to_fp16)[name = string("op_2191_cast_fp16")];
            tensor<int32, [1]> var_2193_axes_0 = const()[name = string("op_2193_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2193_keep_dims_0 = const()[name = string("op_2193_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2193_cast_fp16 = reduce_mean(axes = var_2193_axes_0, keep_dims = var_2193_keep_dims_0, x = var_2191_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp16 var_2183_promoted_1_to_fp16 = const()[name = string("op_2183_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2194_cast_fp16 = pow(x = peak_17_cast_fp16, y = var_2183_promoted_1_to_fp16)[name = string("op_2194_cast_fp16")];
            fp32 var_2195_epsilon_0 = const()[name = string("op_2195_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2195_cast_fp16 = inverse(epsilon = var_2195_epsilon_0, x = var_2194_cast_fp16)[name = string("op_2195_cast_fp16")];
            fp16 var_2196_to_fp16 = const()[name = string("op_2196_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2197_cast_fp16 = mul(x = var_2195_cast_fp16, y = var_2196_to_fp16)[name = string("op_2197_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2198_cast_fp16 = add(x = var_2193_cast_fp16, y = var_2197_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2200_cast_fp16 = mul(x = x_43_cast_fp16, y = var_2199_cast_fp16)[name = string("op_2200_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_2200_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_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_14_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_45_perm_0 = const()[name = string("x_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2217 = const()[name = string("op_2217"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_45_cast_fp16 = transpose(perm = x_45_perm_0, x = var_2210_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 16, 16, 128]> var_2220_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_2220_cast_fp16")];
            tensor<int32, [1]> peak_19_axes_0 = const()[name = string("peak_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_19_cast_fp16 = reduce_max(axes = peak_19_axes_0, keep_dims = var_2217, x = var_2220_cast_fp16)[name = string("peak_19_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = peak_19_cast_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2216_promoted_to_fp16 = const()[name = string("op_2216_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2224_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2216_promoted_to_fp16)[name = string("op_2224_cast_fp16")];
            tensor<int32, [1]> var_2226_axes_0 = const()[name = string("op_2226_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2226_keep_dims_0 = const()[name = string("op_2226_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2226_cast_fp16 = reduce_mean(axes = var_2226_axes_0, keep_dims = var_2226_keep_dims_0, x = var_2224_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp16 var_2216_promoted_1_to_fp16 = const()[name = string("op_2216_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2227_cast_fp16 = pow(x = peak_19_cast_fp16, y = var_2216_promoted_1_to_fp16)[name = string("op_2227_cast_fp16")];
            fp32 var_2228_epsilon_0 = const()[name = string("op_2228_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2228_cast_fp16 = inverse(epsilon = var_2228_epsilon_0, x = var_2227_cast_fp16)[name = string("op_2228_cast_fp16")];
            fp16 var_2229_to_fp16 = const()[name = string("op_2229_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2230_cast_fp16 = mul(x = var_2228_cast_fp16, y = var_2229_to_fp16)[name = string("op_2230_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2231_cast_fp16 = add(x = var_2226_cast_fp16, y = var_2230_cast_fp16)[name = string("op_2231_cast_fp16")];
            fp32 var_2232_epsilon_0 = const()[name = string("op_2232_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2232_cast_fp16 = rsqrt(epsilon = var_2232_epsilon_0, x = var_2231_cast_fp16)[name = string("op_2232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2233_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2232_cast_fp16)[name = string("op_2233_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_53_cast_fp16 = mul(x = var_2233_cast_fp16, y = model_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_53_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_2242 = const()[name = string("op_2242"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2243_cast_fp16 = reshape(shape = var_2242, x = linear_15_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [4]> x_49_perm_0 = const()[name = string("x_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2250 = const()[name = string("op_2250"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_49_cast_fp16 = transpose(perm = x_49_perm_0, x = var_2243_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 8, 16, 128]> var_2253_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2253_cast_fp16")];
            tensor<int32, [1]> peak_21_axes_0 = const()[name = string("peak_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_21_cast_fp16 = reduce_max(axes = peak_21_axes_0, keep_dims = var_2250, x = var_2253_cast_fp16)[name = string("peak_21_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = peak_21_cast_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_2249_promoted_to_fp16 = const()[name = string("op_2249_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2257_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2249_promoted_to_fp16)[name = string("op_2257_cast_fp16")];
            tensor<int32, [1]> var_2259_axes_0 = const()[name = string("op_2259_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2259_keep_dims_0 = const()[name = string("op_2259_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2259_cast_fp16 = reduce_mean(axes = var_2259_axes_0, keep_dims = var_2259_keep_dims_0, x = var_2257_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp16 var_2249_promoted_1_to_fp16 = const()[name = string("op_2249_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2260_cast_fp16 = pow(x = peak_21_cast_fp16, y = var_2249_promoted_1_to_fp16)[name = string("op_2260_cast_fp16")];
            fp32 var_2261_epsilon_0 = const()[name = string("op_2261_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2261_cast_fp16 = inverse(epsilon = var_2261_epsilon_0, x = var_2260_cast_fp16)[name = string("op_2261_cast_fp16")];
            fp16 var_2262_to_fp16 = const()[name = string("op_2262_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2263_cast_fp16 = mul(x = var_2261_cast_fp16, y = var_2262_to_fp16)[name = string("op_2263_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2264_cast_fp16 = add(x = var_2259_cast_fp16, y = var_2263_cast_fp16)[name = string("op_2264_cast_fp16")];
            fp32 var_2265_epsilon_0 = const()[name = string("op_2265_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2265_cast_fp16 = rsqrt(epsilon = var_2265_epsilon_0, x = var_2264_cast_fp16)[name = string("op_2265_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2266_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2265_cast_fp16)[name = string("op_2266_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_55_cast_fp16 = mul(x = var_2266_cast_fp16, y = model_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_55_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_2275 = const()[name = string("op_2275"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2276_cast_fp16 = reshape(shape = var_2275, x = linear_16_cast_fp16)[name = string("op_2276_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_53_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_53_cast_fp16)[name = string("second_9_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2290_cast_fp16 = mul(x = first_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2291_cast_fp16 = mul(x = second_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2293_cast_fp16 = sub(x = var_2290_cast_fp16, y = var_2291_cast_fp16)[name = string("op_2293_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2294_cast_fp16 = mul(x = second_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2295_cast_fp16 = mul(x = first_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2295_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2297_cast_fp16 = add(x = var_2294_cast_fp16, y = var_2295_cast_fp16)[name = string("op_2297_cast_fp16")];
            int32 var_2299 = const()[name = string("op_2299"), 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_2299, interleave = q_5_interleave_0, values = (var_2293_cast_fp16, var_2297_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_55_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_55_cast_fp16)[name = string("second_11_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2311_cast_fp16 = mul(x = first_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2312_cast_fp16 = mul(x = second_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2312_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2314_cast_fp16 = sub(x = var_2311_cast_fp16, y = var_2312_cast_fp16)[name = string("op_2314_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2315_cast_fp16 = mul(x = second_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2315_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2316_cast_fp16 = mul(x = first_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2316_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2318_cast_fp16 = add(x = var_2315_cast_fp16, y = var_2316_cast_fp16)[name = string("op_2318_cast_fp16")];
            int32 var_2320 = const()[name = string("op_2320"), 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_2320, interleave = k_5_interleave_0, values = (var_2314_cast_fp16, var_2318_cast_fp16))[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_2331_begin_0 = const()[name = string("op_2331_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2331_end_0 = const()[name = string("op_2331_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2331_end_mask_0 = const()[name = string("op_2331_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2331_stride_0 = const()[name = string("op_2331_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, stride = var_2331_stride_0, x = q_5_cast_fp16)[name = string("op_2331_cast_fp16")];
            tensor<int32, [4]> var_2341_begin_0 = const()[name = string("op_2341_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2341_end_0 = const()[name = string("op_2341_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2341_end_mask_0 = const()[name = string("op_2341_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2341_stride_0 = const()[name = string("op_2341_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2341_cast_fp16 = slice_by_index(begin = var_2341_begin_0, end = var_2341_end_0, end_mask = var_2341_end_mask_0, stride = var_2341_stride_0, x = q_5_cast_fp16)[name = string("op_2341_cast_fp16")];
            int32 var_2343 = const()[name = string("op_2343"), 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_2343, interleave = stacked_5_interleave_0, values = (var_2331_cast_fp16, var_2341_cast_fp16))[name = string("stacked_5_cast_fp16")];
            tensor<int32, [4]> var_2349_begin_0 = const()[name = string("op_2349_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2349_end_0 = const()[name = string("op_2349_end_0"), val = tensor<int32, [4]>([3, 8, 2064, 128])];
            tensor<bool, [4]> var_2349_end_mask_0 = const()[name = string("op_2349_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2349_cast_fp16 = slice_by_index(begin = var_2349_begin_0, end = var_2349_end_0, end_mask = var_2349_end_mask_0, x = read_state_0)[name = string("op_2349_cast_fp16")];
            tensor<int32, [4]> var_2359_begin_0 = const()[name = string("op_2359_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2359_end_0 = const()[name = string("op_2359_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2359_end_mask_0 = const()[name = string("op_2359_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2359_cast_fp16 = slice_by_index(begin = var_2359_begin_0, end = var_2359_end_0, end_mask = var_2359_end_mask_0, x = var_2349_cast_fp16)[name = string("op_2359_cast_fp16")];
            bool var_2368_transpose_x_1 = const()[name = string("op_2368_transpose_x_1"), val = bool(false)];
            bool var_2368_transpose_y_1 = const()[name = string("op_2368_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_2368_cast_fp16 = matmul(transpose_x = var_2368_transpose_x_1, transpose_y = var_2368_transpose_y_1, x = var_2359_cast_fp16, y = stacked_5_cast_fp16)[name = string("op_2368_cast_fp16")];
            tensor<int32, [4]> var_2371_perm_0 = const()[name = string("op_2371_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_2371_cast_fp16 = transpose(perm = var_2371_perm_0, x = var_2368_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 8, 32, 2048]> var_2373_cast_fp16 = mul(x = var_2371_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_5_cast_fp16 = add(x = var_2373_cast_fp16, y = cacheMask)[name = string("cache_weights_5_cast_fp16")];
            bool var_2379_transpose_x_1 = const()[name = string("op_2379_transpose_x_1"), val = bool(false)];
            bool var_2379_transpose_y_1 = const()[name = string("op_2379_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2379_cast_fp16 = matmul(transpose_x = var_2379_transpose_x_1, transpose_y = var_2379_transpose_y_1, x = stacked_5_cast_fp16, y = k_5_cast_fp16)[name = string("op_2379_cast_fp16")];
            fp16 var_2380_to_fp16 = const()[name = string("op_2380_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2381_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_2380_to_fp16)[name = string("op_2381_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_5_cast_fp16 = add(x = var_2381_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_5_cast_fp16")];
            int32 var_2385 = const()[name = string("op_2385"), val = int32(-1)];
            bool var_2386_interleave_0 = const()[name = string("op_2386_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_2386_cast_fp16 = concat(axis = var_2385, interleave = var_2386_interleave_0, values = (cache_weights_5_cast_fp16, call_weights_5_cast_fp16))[name = string("op_2386_cast_fp16")];
            int32 var_2387 = const()[name = string("op_2387"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_5_cast_fp16 = softmax(axis = var_2387, x = var_2386_cast_fp16)[name = string("shares_5_cast_fp16")];
            tensor<int32, [4]> var_2394_begin_0 = const()[name = string("op_2394_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2394_end_0 = const()[name = string("op_2394_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_2394_end_mask_0 = const()[name = string("op_2394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_2394_cast_fp16 = slice_by_index(begin = var_2394_begin_0, end = var_2394_end_0, end_mask = var_2394_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2394_cast_fp16")];
            tensor<int32, [4]> var_2399_begin_0 = const()[name = string("op_2399_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2399_end_0 = const()[name = string("op_2399_end_0"), val = tensor<int32, [4]>([3, 8, 2064, 128])];
            tensor<bool, [4]> var_2399_end_mask_0 = const()[name = string("op_2399_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2399_cast_fp16 = slice_by_index(begin = var_2399_begin_0, end = var_2399_end_0, end_mask = var_2399_end_mask_0, x = read_state_1)[name = string("op_2399_cast_fp16")];
            tensor<int32, [4]> var_2409_begin_0 = const()[name = string("op_2409_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2409_end_0 = const()[name = string("op_2409_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2409_end_mask_0 = const()[name = string("op_2409_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2409_cast_fp16 = slice_by_index(begin = var_2409_begin_0, end = var_2409_end_0, end_mask = var_2409_end_mask_0, x = var_2399_cast_fp16)[name = string("op_2409_cast_fp16")];
            bool var_2415_transpose_x_0 = const()[name = string("op_2415_transpose_x_0"), val = bool(false)];
            bool var_2415_transpose_y_0 = const()[name = string("op_2415_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2415_cast_fp16 = matmul(transpose_x = var_2415_transpose_x_0, transpose_y = var_2415_transpose_y_0, x = var_2394_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2415_cast_fp16")];
            tensor<int32, [4]> var_2420_begin_0 = const()[name = string("op_2420_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_2420_end_0 = const()[name = string("op_2420_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_2420_end_mask_0 = const()[name = string("op_2420_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2420_cast_fp16 = slice_by_index(begin = var_2420_begin_0, end = var_2420_end_0, end_mask = var_2420_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2420_cast_fp16")];
            bool var_2421_transpose_x_0 = const()[name = string("op_2421_transpose_x_0"), val = bool(false)];
            bool var_2421_transpose_y_0 = const()[name = string("op_2421_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_2276_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 8, 32, 128]> var_2421_cast_fp16 = matmul(transpose_x = var_2421_transpose_x_0, transpose_y = var_2421_transpose_y_0, x = var_2420_cast_fp16, y = v_5_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_9_cast_fp16 = add(x = var_2415_cast_fp16, y = var_2421_cast_fp16)[name = string("mixed_9_cast_fp16")];
            tensor<int32, [4]> var_2438_begin_0 = const()[name = string("op_2438_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2438_end_0 = const()[name = string("op_2438_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2438_end_mask_0 = const()[name = string("op_2438_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2438_cast_fp16 = slice_by_index(begin = var_2438_begin_0, end = var_2438_end_0, end_mask = var_2438_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2438_cast_fp16")];
            tensor<int32, [4]> var_2453_begin_0 = const()[name = string("op_2453_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2453_end_0 = const()[name = string("op_2453_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2453_end_mask_0 = const()[name = string("op_2453_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2453_cast_fp16 = slice_by_index(begin = var_2453_begin_0, end = var_2453_end_0, end_mask = var_2453_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2453_cast_fp16")];
            int32 var_2455 = const()[name = string("op_2455"), 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_2455, interleave = mixed_11_interleave_0, values = (var_2438_cast_fp16, var_2453_cast_fp16))[name = string("mixed_11_cast_fp16")];
            tensor<int32, [4]> var_2459_perm_0 = const()[name = string("op_2459_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2463 = const()[name = string("op_2463"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2459_cast_fp16 = transpose(perm = var_2459_perm_0, x = mixed_11_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 16, 2048]> input_23_cast_fp16 = reshape(shape = var_2463, x = var_2459_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_57_cast_fp16 = add(x = x_41_cast_fp16, y = linear_17_cast_fp16)[name = string("x_57_cast_fp16")];
            bool var_2473 = const()[name = string("op_2473"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2476_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [1]> peak_23_axes_0 = const()[name = string("peak_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_23_cast_fp16 = reduce_max(axes = peak_23_axes_0, keep_dims = var_2473, x = var_2476_cast_fp16)[name = string("peak_23_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = peak_23_cast_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2472_promoted_to_fp16 = const()[name = string("op_2472_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2480_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2472_promoted_to_fp16)[name = string("op_2480_cast_fp16")];
            tensor<int32, [1]> var_2482_axes_0 = const()[name = string("op_2482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2482_keep_dims_0 = const()[name = string("op_2482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2482_cast_fp16 = reduce_mean(axes = var_2482_axes_0, keep_dims = var_2482_keep_dims_0, x = var_2480_cast_fp16)[name = string("op_2482_cast_fp16")];
            fp16 var_2472_promoted_1_to_fp16 = const()[name = string("op_2472_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2483_cast_fp16 = pow(x = peak_23_cast_fp16, y = var_2472_promoted_1_to_fp16)[name = string("op_2483_cast_fp16")];
            fp32 var_2484_epsilon_0 = const()[name = string("op_2484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2484_cast_fp16 = inverse(epsilon = var_2484_epsilon_0, x = var_2483_cast_fp16)[name = string("op_2484_cast_fp16")];
            fp16 var_2485_to_fp16 = const()[name = string("op_2485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2486_cast_fp16 = mul(x = var_2484_cast_fp16, y = var_2485_to_fp16)[name = string("op_2486_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2487_cast_fp16 = add(x = var_2482_cast_fp16, y = var_2486_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp32 var_2488_epsilon_0 = const()[name = string("op_2488_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2488_cast_fp16 = rsqrt(epsilon = var_2488_epsilon_0, x = var_2487_cast_fp16)[name = string("op_2488_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2489_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2488_cast_fp16)[name = string("op_2489_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_2489_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_2494_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_2494_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_2494_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_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_20_cast_fp16)[name = string("x_61_cast_fp16")];
            bool var_2507 = const()[name = string("op_2507"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2510_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<int32, [1]> peak_25_axes_0 = const()[name = string("peak_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_25_cast_fp16 = reduce_max(axes = peak_25_axes_0, keep_dims = var_2507, x = var_2510_cast_fp16)[name = string("peak_25_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = peak_25_cast_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_2506_promoted_to_fp16 = const()[name = string("op_2506_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2514_cast_fp16 = pow(x = x_63_cast_fp16, y = var_2506_promoted_to_fp16)[name = string("op_2514_cast_fp16")];
            tensor<int32, [1]> var_2516_axes_0 = const()[name = string("op_2516_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2516_keep_dims_0 = const()[name = string("op_2516_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2516_cast_fp16 = reduce_mean(axes = var_2516_axes_0, keep_dims = var_2516_keep_dims_0, x = var_2514_cast_fp16)[name = string("op_2516_cast_fp16")];
            fp16 var_2506_promoted_1_to_fp16 = const()[name = string("op_2506_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2517_cast_fp16 = pow(x = peak_25_cast_fp16, y = var_2506_promoted_1_to_fp16)[name = string("op_2517_cast_fp16")];
            fp32 var_2518_epsilon_0 = const()[name = string("op_2518_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2518_cast_fp16 = inverse(epsilon = var_2518_epsilon_0, x = var_2517_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 var_2519_to_fp16 = const()[name = string("op_2519_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2520_cast_fp16 = mul(x = var_2518_cast_fp16, y = var_2519_to_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2521_cast_fp16 = add(x = var_2516_cast_fp16, y = var_2520_cast_fp16)[name = string("op_2521_cast_fp16")];
            fp32 var_2522_epsilon_0 = const()[name = string("op_2522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2522_cast_fp16 = rsqrt(epsilon = var_2522_epsilon_0, x = var_2521_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2523_cast_fp16 = mul(x = x_63_cast_fp16, y = var_2522_cast_fp16)[name = string("op_2523_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_2523_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_2532 = const()[name = string("op_2532"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2533_cast_fp16 = reshape(shape = var_2532, x = linear_21_cast_fp16)[name = string("op_2533_cast_fp16")];
            tensor<int32, [4]> x_65_perm_0 = const()[name = string("x_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2540 = const()[name = string("op_2540"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_65_cast_fp16 = transpose(perm = x_65_perm_0, x = var_2533_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 16, 128]> var_2543_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2543_cast_fp16")];
            tensor<int32, [1]> peak_27_axes_0 = const()[name = string("peak_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_27_cast_fp16 = reduce_max(axes = peak_27_axes_0, keep_dims = var_2540, x = var_2543_cast_fp16)[name = string("peak_27_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = peak_27_cast_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_2539_promoted_to_fp16 = const()[name = string("op_2539_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2547_cast_fp16 = pow(x = x_67_cast_fp16, y = var_2539_promoted_to_fp16)[name = string("op_2547_cast_fp16")];
            tensor<int32, [1]> var_2549_axes_0 = const()[name = string("op_2549_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2549_keep_dims_0 = const()[name = string("op_2549_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2549_cast_fp16 = reduce_mean(axes = var_2549_axes_0, keep_dims = var_2549_keep_dims_0, x = var_2547_cast_fp16)[name = string("op_2549_cast_fp16")];
            fp16 var_2539_promoted_1_to_fp16 = const()[name = string("op_2539_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2550_cast_fp16 = pow(x = peak_27_cast_fp16, y = var_2539_promoted_1_to_fp16)[name = string("op_2550_cast_fp16")];
            fp32 var_2551_epsilon_0 = const()[name = string("op_2551_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2551_cast_fp16 = inverse(epsilon = var_2551_epsilon_0, x = var_2550_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 var_2552_to_fp16 = const()[name = string("op_2552_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2553_cast_fp16 = mul(x = var_2551_cast_fp16, y = var_2552_to_fp16)[name = string("op_2553_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2554_cast_fp16 = add(x = var_2549_cast_fp16, y = var_2553_cast_fp16)[name = string("op_2554_cast_fp16")];
            fp32 var_2555_epsilon_0 = const()[name = string("op_2555_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2555_cast_fp16 = rsqrt(epsilon = var_2555_epsilon_0, x = var_2554_cast_fp16)[name = string("op_2555_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2556_cast_fp16 = mul(x = x_67_cast_fp16, y = var_2555_cast_fp16)[name = string("op_2556_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_73_cast_fp16 = mul(x = var_2556_cast_fp16, y = model_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_73_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_2565 = const()[name = string("op_2565"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2566_cast_fp16 = reshape(shape = var_2565, x = linear_22_cast_fp16)[name = string("op_2566_cast_fp16")];
            tensor<int32, [4]> x_69_perm_0 = const()[name = string("x_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2573 = const()[name = string("op_2573"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_69_cast_fp16 = transpose(perm = x_69_perm_0, x = var_2566_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 16, 128]> var_2576_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> peak_29_axes_0 = const()[name = string("peak_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_29_cast_fp16 = reduce_max(axes = peak_29_axes_0, keep_dims = var_2573, x = var_2576_cast_fp16)[name = string("peak_29_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = peak_29_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_2572_promoted_to_fp16 = const()[name = string("op_2572_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2580_cast_fp16 = pow(x = x_71_cast_fp16, y = var_2572_promoted_to_fp16)[name = string("op_2580_cast_fp16")];
            tensor<int32, [1]> var_2582_axes_0 = const()[name = string("op_2582_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2582_keep_dims_0 = const()[name = string("op_2582_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2582_cast_fp16 = reduce_mean(axes = var_2582_axes_0, keep_dims = var_2582_keep_dims_0, x = var_2580_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp16 var_2572_promoted_1_to_fp16 = const()[name = string("op_2572_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2583_cast_fp16 = pow(x = peak_29_cast_fp16, y = var_2572_promoted_1_to_fp16)[name = string("op_2583_cast_fp16")];
            fp32 var_2584_epsilon_0 = const()[name = string("op_2584_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2584_cast_fp16 = inverse(epsilon = var_2584_epsilon_0, x = var_2583_cast_fp16)[name = string("op_2584_cast_fp16")];
            fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2586_cast_fp16 = mul(x = var_2584_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2587_cast_fp16 = add(x = var_2582_cast_fp16, y = var_2586_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2589_cast_fp16 = mul(x = x_71_cast_fp16, y = var_2588_cast_fp16)[name = string("op_2589_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_75_cast_fp16 = mul(x = var_2589_cast_fp16, y = model_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_75_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_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2599_cast_fp16 = reshape(shape = var_2598, x = linear_23_cast_fp16)[name = string("op_2599_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_73_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_73_cast_fp16)[name = string("second_13_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2613_cast_fp16 = mul(x = first_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2613_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2614_cast_fp16 = mul(x = second_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2614_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2616_cast_fp16 = sub(x = var_2613_cast_fp16, y = var_2614_cast_fp16)[name = string("op_2616_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2617_cast_fp16 = mul(x = second_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2617_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2618_cast_fp16 = mul(x = first_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2618_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2620_cast_fp16 = add(x = var_2617_cast_fp16, y = var_2618_cast_fp16)[name = string("op_2620_cast_fp16")];
            int32 var_2622 = const()[name = string("op_2622"), 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_2622, interleave = q_7_interleave_0, values = (var_2616_cast_fp16, var_2620_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_75_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_75_cast_fp16)[name = string("second_15_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2634_cast_fp16 = mul(x = first_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2635_cast_fp16 = mul(x = second_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2635_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2637_cast_fp16 = sub(x = var_2634_cast_fp16, y = var_2635_cast_fp16)[name = string("op_2637_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2638_cast_fp16 = mul(x = second_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2638_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2639_cast_fp16 = mul(x = first_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2641_cast_fp16 = add(x = var_2638_cast_fp16, y = var_2639_cast_fp16)[name = string("op_2641_cast_fp16")];
            int32 var_2643 = const()[name = string("op_2643"), 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_2643, interleave = k_7_interleave_0, values = (var_2637_cast_fp16, var_2641_cast_fp16))[name = string("k_7_cast_fp16")];
            tensor<int32, [4]> var_2654_begin_0 = const()[name = string("op_2654_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2654_end_0 = const()[name = string("op_2654_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2654_end_mask_0 = const()[name = string("op_2654_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2654_stride_0 = const()[name = string("op_2654_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2654_cast_fp16 = slice_by_index(begin = var_2654_begin_0, end = var_2654_end_0, end_mask = var_2654_end_mask_0, stride = var_2654_stride_0, x = q_7_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2664_begin_0 = const()[name = string("op_2664_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2664_end_0 = const()[name = string("op_2664_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2664_end_mask_0 = const()[name = string("op_2664_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2664_stride_0 = const()[name = string("op_2664_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2664_cast_fp16 = slice_by_index(begin = var_2664_begin_0, end = var_2664_end_0, end_mask = var_2664_end_mask_0, stride = var_2664_stride_0, x = q_7_cast_fp16)[name = string("op_2664_cast_fp16")];
            int32 var_2666 = const()[name = string("op_2666"), 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_2666, interleave = stacked_7_interleave_0, values = (var_2654_cast_fp16, var_2664_cast_fp16))[name = string("stacked_7_cast_fp16")];
            tensor<int32, [4]> var_2672_begin_0 = const()[name = string("op_2672_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2672_end_0 = const()[name = string("op_2672_end_0"), val = tensor<int32, [4]>([4, 8, 2064, 128])];
            tensor<bool, [4]> var_2672_end_mask_0 = const()[name = string("op_2672_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2672_cast_fp16 = slice_by_index(begin = var_2672_begin_0, end = var_2672_end_0, end_mask = var_2672_end_mask_0, x = read_state_0)[name = string("op_2672_cast_fp16")];
            tensor<int32, [4]> var_2682_begin_0 = const()[name = string("op_2682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2682_end_0 = const()[name = string("op_2682_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2682_end_mask_0 = const()[name = string("op_2682_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2682_cast_fp16 = slice_by_index(begin = var_2682_begin_0, end = var_2682_end_0, end_mask = var_2682_end_mask_0, x = var_2672_cast_fp16)[name = string("op_2682_cast_fp16")];
            bool var_2691_transpose_x_1 = const()[name = string("op_2691_transpose_x_1"), val = bool(false)];
            bool var_2691_transpose_y_1 = const()[name = string("op_2691_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_2691_cast_fp16 = matmul(transpose_x = var_2691_transpose_x_1, transpose_y = var_2691_transpose_y_1, x = var_2682_cast_fp16, y = stacked_7_cast_fp16)[name = string("op_2691_cast_fp16")];
            tensor<int32, [4]> var_2694_perm_0 = const()[name = string("op_2694_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2695_to_fp16 = const()[name = string("op_2695_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_2694_cast_fp16 = transpose(perm = var_2694_perm_0, x = var_2691_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 32, 2048]> var_2696_cast_fp16 = mul(x = var_2694_cast_fp16, y = var_2695_to_fp16)[name = string("op_2696_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_7_cast_fp16 = add(x = var_2696_cast_fp16, y = cacheMask)[name = string("cache_weights_7_cast_fp16")];
            bool var_2702_transpose_x_1 = const()[name = string("op_2702_transpose_x_1"), val = bool(false)];
            bool var_2702_transpose_y_1 = const()[name = string("op_2702_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2702_cast_fp16 = matmul(transpose_x = var_2702_transpose_x_1, transpose_y = var_2702_transpose_y_1, x = stacked_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 var_2703_to_fp16 = const()[name = string("op_2703_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2704_cast_fp16 = mul(x = var_2702_cast_fp16, y = var_2703_to_fp16)[name = string("op_2704_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_7_cast_fp16 = add(x = var_2704_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_7_cast_fp16")];
            int32 var_2708 = const()[name = string("op_2708"), val = int32(-1)];
            bool var_2709_interleave_0 = const()[name = string("op_2709_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_2709_cast_fp16 = concat(axis = var_2708, interleave = var_2709_interleave_0, values = (cache_weights_7_cast_fp16, call_weights_7_cast_fp16))[name = string("op_2709_cast_fp16")];
            int32 var_2710 = const()[name = string("op_2710"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_7_cast_fp16 = softmax(axis = var_2710, x = var_2709_cast_fp16)[name = string("shares_7_cast_fp16")];
            tensor<int32, [4]> var_2717_begin_0 = const()[name = string("op_2717_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2717_end_0 = const()[name = string("op_2717_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_2717_end_mask_0 = const()[name = string("op_2717_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_2717_cast_fp16 = slice_by_index(begin = var_2717_begin_0, end = var_2717_end_0, end_mask = var_2717_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2717_cast_fp16")];
            tensor<int32, [4]> var_2722_begin_0 = const()[name = string("op_2722_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2722_end_0 = const()[name = string("op_2722_end_0"), val = tensor<int32, [4]>([4, 8, 2064, 128])];
            tensor<bool, [4]> var_2722_end_mask_0 = const()[name = string("op_2722_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2722_cast_fp16 = slice_by_index(begin = var_2722_begin_0, end = var_2722_end_0, end_mask = var_2722_end_mask_0, x = read_state_1)[name = string("op_2722_cast_fp16")];
            tensor<int32, [4]> var_2732_begin_0 = const()[name = string("op_2732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2732_end_0 = const()[name = string("op_2732_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2732_end_mask_0 = const()[name = string("op_2732_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2732_cast_fp16 = slice_by_index(begin = var_2732_begin_0, end = var_2732_end_0, end_mask = var_2732_end_mask_0, x = var_2722_cast_fp16)[name = string("op_2732_cast_fp16")];
            bool var_2738_transpose_x_0 = const()[name = string("op_2738_transpose_x_0"), val = bool(false)];
            bool var_2738_transpose_y_0 = const()[name = string("op_2738_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2738_cast_fp16 = matmul(transpose_x = var_2738_transpose_x_0, transpose_y = var_2738_transpose_y_0, x = var_2717_cast_fp16, y = var_2732_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> var_2743_begin_0 = const()[name = string("op_2743_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_2743_end_0 = const()[name = string("op_2743_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_2743_end_mask_0 = const()[name = string("op_2743_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2743_cast_fp16 = slice_by_index(begin = var_2743_begin_0, end = var_2743_end_0, end_mask = var_2743_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2743_cast_fp16")];
            bool var_2744_transpose_x_0 = const()[name = string("op_2744_transpose_x_0"), val = bool(false)];
            bool var_2744_transpose_y_0 = const()[name = string("op_2744_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_2599_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 8, 32, 128]> var_2744_cast_fp16 = matmul(transpose_x = var_2744_transpose_x_0, transpose_y = var_2744_transpose_y_0, x = var_2743_cast_fp16, y = v_7_cast_fp16)[name = string("op_2744_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_13_cast_fp16 = add(x = var_2738_cast_fp16, y = var_2744_cast_fp16)[name = string("mixed_13_cast_fp16")];
            tensor<int32, [4]> var_2761_begin_0 = const()[name = string("op_2761_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2761_end_0 = const()[name = string("op_2761_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2761_end_mask_0 = const()[name = string("op_2761_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2761_cast_fp16 = slice_by_index(begin = var_2761_begin_0, end = var_2761_end_0, end_mask = var_2761_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2761_cast_fp16")];
            tensor<int32, [4]> var_2776_begin_0 = const()[name = string("op_2776_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2776_end_0 = const()[name = string("op_2776_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2776_end_mask_0 = const()[name = string("op_2776_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2776_cast_fp16 = slice_by_index(begin = var_2776_begin_0, end = var_2776_end_0, end_mask = var_2776_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2776_cast_fp16")];
            int32 var_2778 = const()[name = string("op_2778"), 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_2778, interleave = mixed_15_interleave_0, values = (var_2761_cast_fp16, var_2776_cast_fp16))[name = string("mixed_15_cast_fp16")];
            tensor<int32, [4]> var_2782_perm_0 = const()[name = string("op_2782_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2786 = const()[name = string("op_2786"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2782_cast_fp16 = transpose(perm = var_2782_perm_0, x = mixed_15_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 16, 2048]> input_33_cast_fp16 = reshape(shape = var_2786, x = var_2782_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_77_cast_fp16 = add(x = x_61_cast_fp16, y = linear_24_cast_fp16)[name = string("x_77_cast_fp16")];
            bool var_2796 = const()[name = string("op_2796"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2799_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_2799_cast_fp16")];
            tensor<int32, [1]> peak_31_axes_0 = const()[name = string("peak_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_31_cast_fp16 = reduce_max(axes = peak_31_axes_0, keep_dims = var_2796, x = var_2799_cast_fp16)[name = string("peak_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = peak_31_cast_fp16)[name = string("x_79_cast_fp16")];
            fp16 var_2795_promoted_to_fp16 = const()[name = string("op_2795_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2803_cast_fp16 = pow(x = x_79_cast_fp16, y = var_2795_promoted_to_fp16)[name = string("op_2803_cast_fp16")];
            tensor<int32, [1]> var_2805_axes_0 = const()[name = string("op_2805_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2805_keep_dims_0 = const()[name = string("op_2805_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2805_cast_fp16 = reduce_mean(axes = var_2805_axes_0, keep_dims = var_2805_keep_dims_0, x = var_2803_cast_fp16)[name = string("op_2805_cast_fp16")];
            fp16 var_2795_promoted_1_to_fp16 = const()[name = string("op_2795_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2806_cast_fp16 = pow(x = peak_31_cast_fp16, y = var_2795_promoted_1_to_fp16)[name = string("op_2806_cast_fp16")];
            fp32 var_2807_epsilon_0 = const()[name = string("op_2807_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2807_cast_fp16 = inverse(epsilon = var_2807_epsilon_0, x = var_2806_cast_fp16)[name = string("op_2807_cast_fp16")];
            fp16 var_2808_to_fp16 = const()[name = string("op_2808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2809_cast_fp16 = mul(x = var_2807_cast_fp16, y = var_2808_to_fp16)[name = string("op_2809_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2810_cast_fp16 = add(x = var_2805_cast_fp16, y = var_2809_cast_fp16)[name = string("op_2810_cast_fp16")];
            fp32 var_2811_epsilon_0 = const()[name = string("op_2811_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2811_cast_fp16 = rsqrt(epsilon = var_2811_epsilon_0, x = var_2810_cast_fp16)[name = string("op_2811_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2812_cast_fp16 = mul(x = x_79_cast_fp16, y = var_2811_cast_fp16)[name = string("op_2812_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_2812_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_2817_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_2817_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_2817_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_81_cast_fp16 = add(x = x_77_cast_fp16, y = linear_27_cast_fp16)[name = string("x_81_cast_fp16")];
            bool var_2830 = const()[name = string("op_2830"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2833_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<int32, [1]> peak_33_axes_0 = const()[name = string("peak_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_33_cast_fp16 = reduce_max(axes = peak_33_axes_0, keep_dims = var_2830, x = var_2833_cast_fp16)[name = string("peak_33_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = peak_33_cast_fp16)[name = string("x_83_cast_fp16")];
            fp16 var_2829_promoted_to_fp16 = const()[name = string("op_2829_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2837_cast_fp16 = pow(x = x_83_cast_fp16, y = var_2829_promoted_to_fp16)[name = string("op_2837_cast_fp16")];
            tensor<int32, [1]> var_2839_axes_0 = const()[name = string("op_2839_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2839_keep_dims_0 = const()[name = string("op_2839_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2839_cast_fp16 = reduce_mean(axes = var_2839_axes_0, keep_dims = var_2839_keep_dims_0, x = var_2837_cast_fp16)[name = string("op_2839_cast_fp16")];
            fp16 var_2829_promoted_1_to_fp16 = const()[name = string("op_2829_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2840_cast_fp16 = pow(x = peak_33_cast_fp16, y = var_2829_promoted_1_to_fp16)[name = string("op_2840_cast_fp16")];
            fp32 var_2841_epsilon_0 = const()[name = string("op_2841_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2841_cast_fp16 = inverse(epsilon = var_2841_epsilon_0, x = var_2840_cast_fp16)[name = string("op_2841_cast_fp16")];
            fp16 var_2842_to_fp16 = const()[name = string("op_2842_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2843_cast_fp16 = mul(x = var_2841_cast_fp16, y = var_2842_to_fp16)[name = string("op_2843_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2844_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2843_cast_fp16)[name = string("op_2844_cast_fp16")];
            fp32 var_2845_epsilon_0 = const()[name = string("op_2845_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2845_cast_fp16 = rsqrt(epsilon = var_2845_epsilon_0, x = var_2844_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2846_cast_fp16 = mul(x = x_83_cast_fp16, y = var_2845_cast_fp16)[name = string("op_2846_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_2846_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_2855 = const()[name = string("op_2855"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2856_cast_fp16 = reshape(shape = var_2855, x = linear_28_cast_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [4]> x_85_perm_0 = const()[name = string("x_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2863 = const()[name = string("op_2863"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_85_cast_fp16 = transpose(perm = x_85_perm_0, x = var_2856_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 16, 16, 128]> var_2866_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_2866_cast_fp16")];
            tensor<int32, [1]> peak_35_axes_0 = const()[name = string("peak_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_35_cast_fp16 = reduce_max(axes = peak_35_axes_0, keep_dims = var_2863, x = var_2866_cast_fp16)[name = string("peak_35_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = peak_35_cast_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_2862_promoted_to_fp16 = const()[name = string("op_2862_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2870_cast_fp16 = pow(x = x_87_cast_fp16, y = var_2862_promoted_to_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [1]> var_2872_axes_0 = const()[name = string("op_2872_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2872_keep_dims_0 = const()[name = string("op_2872_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2872_cast_fp16 = reduce_mean(axes = var_2872_axes_0, keep_dims = var_2872_keep_dims_0, x = var_2870_cast_fp16)[name = string("op_2872_cast_fp16")];
            fp16 var_2862_promoted_1_to_fp16 = const()[name = string("op_2862_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2873_cast_fp16 = pow(x = peak_35_cast_fp16, y = var_2862_promoted_1_to_fp16)[name = string("op_2873_cast_fp16")];
            fp32 var_2874_epsilon_0 = const()[name = string("op_2874_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2874_cast_fp16 = inverse(epsilon = var_2874_epsilon_0, x = var_2873_cast_fp16)[name = string("op_2874_cast_fp16")];
            fp16 var_2875_to_fp16 = const()[name = string("op_2875_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2876_cast_fp16 = mul(x = var_2874_cast_fp16, y = var_2875_to_fp16)[name = string("op_2876_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2877_cast_fp16 = add(x = var_2872_cast_fp16, y = var_2876_cast_fp16)[name = string("op_2877_cast_fp16")];
            fp32 var_2878_epsilon_0 = const()[name = string("op_2878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2878_cast_fp16 = rsqrt(epsilon = var_2878_epsilon_0, x = var_2877_cast_fp16)[name = string("op_2878_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2879_cast_fp16 = mul(x = x_87_cast_fp16, y = var_2878_cast_fp16)[name = string("op_2879_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_93_cast_fp16 = mul(x = var_2879_cast_fp16, y = model_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_93_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_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_29_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> x_89_perm_0 = const()[name = string("x_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2896 = const()[name = string("op_2896"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = var_2889_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 8, 16, 128]> var_2899_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<int32, [1]> peak_37_axes_0 = const()[name = string("peak_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_37_cast_fp16 = reduce_max(axes = peak_37_axes_0, keep_dims = var_2896, x = var_2899_cast_fp16)[name = string("peak_37_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = peak_37_cast_fp16)[name = string("x_91_cast_fp16")];
            fp16 var_2895_promoted_to_fp16 = const()[name = string("op_2895_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2903_cast_fp16 = pow(x = x_91_cast_fp16, y = var_2895_promoted_to_fp16)[name = string("op_2903_cast_fp16")];
            tensor<int32, [1]> var_2905_axes_0 = const()[name = string("op_2905_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2905_keep_dims_0 = const()[name = string("op_2905_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2905_cast_fp16 = reduce_mean(axes = var_2905_axes_0, keep_dims = var_2905_keep_dims_0, x = var_2903_cast_fp16)[name = string("op_2905_cast_fp16")];
            fp16 var_2895_promoted_1_to_fp16 = const()[name = string("op_2895_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2906_cast_fp16 = pow(x = peak_37_cast_fp16, y = var_2895_promoted_1_to_fp16)[name = string("op_2906_cast_fp16")];
            fp32 var_2907_epsilon_0 = const()[name = string("op_2907_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2907_cast_fp16 = inverse(epsilon = var_2907_epsilon_0, x = var_2906_cast_fp16)[name = string("op_2907_cast_fp16")];
            fp16 var_2908_to_fp16 = const()[name = string("op_2908_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2909_cast_fp16 = mul(x = var_2907_cast_fp16, y = var_2908_to_fp16)[name = string("op_2909_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2910_cast_fp16 = add(x = var_2905_cast_fp16, y = var_2909_cast_fp16)[name = string("op_2910_cast_fp16")];
            fp32 var_2911_epsilon_0 = const()[name = string("op_2911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2911_cast_fp16 = rsqrt(epsilon = var_2911_epsilon_0, x = var_2910_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2912_cast_fp16 = mul(x = x_91_cast_fp16, y = var_2911_cast_fp16)[name = string("op_2912_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_95_cast_fp16 = mul(x = var_2912_cast_fp16, y = model_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_95_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_2921 = const()[name = string("op_2921"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2922_cast_fp16 = reshape(shape = var_2921, x = linear_30_cast_fp16)[name = string("op_2922_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_93_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_93_cast_fp16)[name = string("second_17_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2936_cast_fp16 = mul(x = first_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2937_cast_fp16 = mul(x = second_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2937_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2939_cast_fp16 = sub(x = var_2936_cast_fp16, y = var_2937_cast_fp16)[name = string("op_2939_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2940_cast_fp16 = mul(x = second_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2940_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2941_cast_fp16 = mul(x = first_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2943_cast_fp16 = add(x = var_2940_cast_fp16, y = var_2941_cast_fp16)[name = string("op_2943_cast_fp16")];
            int32 var_2945 = const()[name = string("op_2945"), 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_2945, interleave = q_9_interleave_0, values = (var_2939_cast_fp16, var_2943_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_95_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_95_cast_fp16)[name = string("second_19_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2957_cast_fp16 = mul(x = first_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2958_cast_fp16 = mul(x = second_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2960_cast_fp16 = sub(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = string("op_2960_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2961_cast_fp16 = mul(x = second_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2962_cast_fp16 = mul(x = first_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2962_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2964_cast_fp16 = add(x = var_2961_cast_fp16, y = var_2962_cast_fp16)[name = string("op_2964_cast_fp16")];
            int32 var_2966 = const()[name = string("op_2966"), 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_2966, interleave = k_9_interleave_0, values = (var_2960_cast_fp16, var_2964_cast_fp16))[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2977_stride_0 = const()[name = string("op_2977_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, stride = var_2977_stride_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            tensor<int32, [4]> var_2987_begin_0 = const()[name = string("op_2987_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2987_end_0 = const()[name = string("op_2987_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2987_end_mask_0 = const()[name = string("op_2987_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2987_stride_0 = const()[name = string("op_2987_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2987_cast_fp16 = slice_by_index(begin = var_2987_begin_0, end = var_2987_end_0, end_mask = var_2987_end_mask_0, stride = var_2987_stride_0, x = q_9_cast_fp16)[name = string("op_2987_cast_fp16")];
            int32 var_2989 = const()[name = string("op_2989"), 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_2989, interleave = stacked_9_interleave_0, values = (var_2977_cast_fp16, var_2987_cast_fp16))[name = string("stacked_9_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([5, 8, 2064, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = read_state_0)[name = string("op_2995_cast_fp16")];
            tensor<int32, [4]> var_3005_begin_0 = const()[name = string("op_3005_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3005_end_0 = const()[name = string("op_3005_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3005_end_mask_0 = const()[name = string("op_3005_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3005_cast_fp16 = slice_by_index(begin = var_3005_begin_0, end = var_3005_end_0, end_mask = var_3005_end_mask_0, x = var_2995_cast_fp16)[name = string("op_3005_cast_fp16")];
            bool var_3014_transpose_x_1 = const()[name = string("op_3014_transpose_x_1"), val = bool(false)];
            bool var_3014_transpose_y_1 = const()[name = string("op_3014_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_3014_cast_fp16 = matmul(transpose_x = var_3014_transpose_x_1, transpose_y = var_3014_transpose_y_1, x = var_3005_cast_fp16, y = stacked_9_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<int32, [4]> var_3017_perm_0 = const()[name = string("op_3017_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_3017_cast_fp16 = transpose(perm = var_3017_perm_0, x = var_3014_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 8, 32, 2048]> var_3019_cast_fp16 = mul(x = var_3017_cast_fp16, y = var_3018_to_fp16)[name = string("op_3019_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_9_cast_fp16 = add(x = var_3019_cast_fp16, y = cacheMask)[name = string("cache_weights_9_cast_fp16")];
            bool var_3025_transpose_x_1 = const()[name = string("op_3025_transpose_x_1"), val = bool(false)];
            bool var_3025_transpose_y_1 = const()[name = string("op_3025_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3025_cast_fp16 = matmul(transpose_x = var_3025_transpose_x_1, transpose_y = var_3025_transpose_y_1, x = stacked_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_3025_cast_fp16")];
            fp16 var_3026_to_fp16 = const()[name = string("op_3026_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3027_cast_fp16 = mul(x = var_3025_cast_fp16, y = var_3026_to_fp16)[name = string("op_3027_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_9_cast_fp16 = add(x = var_3027_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_9_cast_fp16")];
            int32 var_3031 = const()[name = string("op_3031"), val = int32(-1)];
            bool var_3032_interleave_0 = const()[name = string("op_3032_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_3032_cast_fp16 = concat(axis = var_3031, interleave = var_3032_interleave_0, values = (cache_weights_9_cast_fp16, call_weights_9_cast_fp16))[name = string("op_3032_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_9_cast_fp16 = softmax(axis = var_3033, x = var_3032_cast_fp16)[name = string("shares_9_cast_fp16")];
            tensor<int32, [4]> var_3040_begin_0 = const()[name = string("op_3040_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3040_end_0 = const()[name = string("op_3040_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_3040_end_mask_0 = const()[name = string("op_3040_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_3040_cast_fp16 = slice_by_index(begin = var_3040_begin_0, end = var_3040_end_0, end_mask = var_3040_end_mask_0, x = shares_9_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> var_3045_begin_0 = const()[name = string("op_3045_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_3045_end_0 = const()[name = string("op_3045_end_0"), val = tensor<int32, [4]>([5, 8, 2064, 128])];
            tensor<bool, [4]> var_3045_end_mask_0 = const()[name = string("op_3045_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3045_cast_fp16 = slice_by_index(begin = var_3045_begin_0, end = var_3045_end_0, end_mask = var_3045_end_mask_0, x = read_state_1)[name = string("op_3045_cast_fp16")];
            tensor<int32, [4]> var_3055_begin_0 = const()[name = string("op_3055_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3055_end_0 = const()[name = string("op_3055_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3055_end_mask_0 = const()[name = string("op_3055_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3055_cast_fp16 = slice_by_index(begin = var_3055_begin_0, end = var_3055_end_0, end_mask = var_3055_end_mask_0, x = var_3045_cast_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3061_transpose_x_0 = const()[name = string("op_3061_transpose_x_0"), val = bool(false)];
            bool var_3061_transpose_y_0 = const()[name = string("op_3061_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3061_cast_fp16 = matmul(transpose_x = var_3061_transpose_x_0, transpose_y = var_3061_transpose_y_0, x = var_3040_cast_fp16, y = var_3055_cast_fp16)[name = string("op_3061_cast_fp16")];
            tensor<int32, [4]> var_3066_begin_0 = const()[name = string("op_3066_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_3066_end_0 = const()[name = string("op_3066_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_3066_end_mask_0 = const()[name = string("op_3066_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3066_cast_fp16 = slice_by_index(begin = var_3066_begin_0, end = var_3066_end_0, end_mask = var_3066_end_mask_0, x = shares_9_cast_fp16)[name = string("op_3066_cast_fp16")];
            bool var_3067_transpose_x_0 = const()[name = string("op_3067_transpose_x_0"), val = bool(false)];
            bool var_3067_transpose_y_0 = const()[name = string("op_3067_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_2922_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 8, 32, 128]> var_3067_cast_fp16 = matmul(transpose_x = var_3067_transpose_x_0, transpose_y = var_3067_transpose_y_0, x = var_3066_cast_fp16, y = v_9_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_17_cast_fp16 = add(x = var_3061_cast_fp16, y = var_3067_cast_fp16)[name = string("mixed_17_cast_fp16")];
            tensor<int32, [4]> var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [4]> var_3099_begin_0 = const()[name = string("op_3099_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3099_end_0 = const()[name = string("op_3099_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3099_end_mask_0 = const()[name = string("op_3099_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3099_cast_fp16 = slice_by_index(begin = var_3099_begin_0, end = var_3099_end_0, end_mask = var_3099_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_3099_cast_fp16")];
            int32 var_3101 = const()[name = string("op_3101"), 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_3101, interleave = mixed_19_interleave_0, values = (var_3084_cast_fp16, var_3099_cast_fp16))[name = string("mixed_19_cast_fp16")];
            tensor<int32, [4]> var_3105_perm_0 = const()[name = string("op_3105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3109 = const()[name = string("op_3109"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3105_cast_fp16 = transpose(perm = var_3105_perm_0, x = mixed_19_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 2048]> input_43_cast_fp16 = reshape(shape = var_3109, x = var_3105_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_97_cast_fp16 = add(x = x_81_cast_fp16, y = linear_31_cast_fp16)[name = string("x_97_cast_fp16")];
            bool var_3119 = const()[name = string("op_3119"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3122_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<int32, [1]> peak_39_axes_0 = const()[name = string("peak_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_39_cast_fp16 = reduce_max(axes = peak_39_axes_0, keep_dims = var_3119, x = var_3122_cast_fp16)[name = string("peak_39_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_99_cast_fp16 = real_div(x = x_97_cast_fp16, y = peak_39_cast_fp16)[name = string("x_99_cast_fp16")];
            fp16 var_3118_promoted_to_fp16 = const()[name = string("op_3118_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3126_cast_fp16 = pow(x = x_99_cast_fp16, y = var_3118_promoted_to_fp16)[name = string("op_3126_cast_fp16")];
            tensor<int32, [1]> var_3128_axes_0 = const()[name = string("op_3128_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3128_keep_dims_0 = const()[name = string("op_3128_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3128_cast_fp16 = reduce_mean(axes = var_3128_axes_0, keep_dims = var_3128_keep_dims_0, x = var_3126_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp16 var_3118_promoted_1_to_fp16 = const()[name = string("op_3118_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3129_cast_fp16 = pow(x = peak_39_cast_fp16, y = var_3118_promoted_1_to_fp16)[name = string("op_3129_cast_fp16")];
            fp32 var_3130_epsilon_0 = const()[name = string("op_3130_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3130_cast_fp16 = inverse(epsilon = var_3130_epsilon_0, x = var_3129_cast_fp16)[name = string("op_3130_cast_fp16")];
            fp16 var_3131_to_fp16 = const()[name = string("op_3131_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3132_cast_fp16 = mul(x = var_3130_cast_fp16, y = var_3131_to_fp16)[name = string("op_3132_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3133_cast_fp16 = add(x = var_3128_cast_fp16, y = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            fp32 var_3134_epsilon_0 = const()[name = string("op_3134_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3134_cast_fp16 = rsqrt(epsilon = var_3134_epsilon_0, x = var_3133_cast_fp16)[name = string("op_3134_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3135_cast_fp16 = mul(x = x_99_cast_fp16, y = var_3134_cast_fp16)[name = string("op_3135_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_3135_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_3140_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_3140_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_3140_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_101_cast_fp16 = add(x = x_97_cast_fp16, y = linear_34_cast_fp16)[name = string("x_101_cast_fp16")];
            bool var_3153 = const()[name = string("op_3153"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3156_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<int32, [1]> peak_41_axes_0 = const()[name = string("peak_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_41_cast_fp16 = reduce_max(axes = peak_41_axes_0, keep_dims = var_3153, x = var_3156_cast_fp16)[name = string("peak_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_103_cast_fp16 = real_div(x = x_101_cast_fp16, y = peak_41_cast_fp16)[name = string("x_103_cast_fp16")];
            fp16 var_3152_promoted_to_fp16 = const()[name = string("op_3152_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3160_cast_fp16 = pow(x = x_103_cast_fp16, y = var_3152_promoted_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [1]> var_3162_axes_0 = const()[name = string("op_3162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3162_keep_dims_0 = const()[name = string("op_3162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3162_cast_fp16 = reduce_mean(axes = var_3162_axes_0, keep_dims = var_3162_keep_dims_0, x = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3152_promoted_1_to_fp16 = const()[name = string("op_3152_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3163_cast_fp16 = pow(x = peak_41_cast_fp16, y = var_3152_promoted_1_to_fp16)[name = string("op_3163_cast_fp16")];
            fp32 var_3164_epsilon_0 = const()[name = string("op_3164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3164_cast_fp16 = inverse(epsilon = var_3164_epsilon_0, x = var_3163_cast_fp16)[name = string("op_3164_cast_fp16")];
            fp16 var_3165_to_fp16 = const()[name = string("op_3165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3166_cast_fp16 = mul(x = var_3164_cast_fp16, y = var_3165_to_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3167_cast_fp16 = add(x = var_3162_cast_fp16, y = var_3166_cast_fp16)[name = string("op_3167_cast_fp16")];
            fp32 var_3168_epsilon_0 = const()[name = string("op_3168_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3168_cast_fp16 = rsqrt(epsilon = var_3168_epsilon_0, x = var_3167_cast_fp16)[name = string("op_3168_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3169_cast_fp16 = mul(x = x_103_cast_fp16, y = var_3168_cast_fp16)[name = string("op_3169_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_3169_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_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_35_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> x_105_perm_0 = const()[name = string("x_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3186 = const()[name = string("op_3186"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_105_cast_fp16 = transpose(perm = x_105_perm_0, x = var_3179_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 16, 16, 128]> var_3189_cast_fp16 = abs(x = x_105_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<int32, [1]> peak_43_axes_0 = const()[name = string("peak_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_43_cast_fp16 = reduce_max(axes = peak_43_axes_0, keep_dims = var_3186, x = var_3189_cast_fp16)[name = string("peak_43_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_107_cast_fp16 = real_div(x = x_105_cast_fp16, y = peak_43_cast_fp16)[name = string("x_107_cast_fp16")];
            fp16 var_3185_promoted_to_fp16 = const()[name = string("op_3185_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3193_cast_fp16 = pow(x = x_107_cast_fp16, y = var_3185_promoted_to_fp16)[name = string("op_3193_cast_fp16")];
            tensor<int32, [1]> var_3195_axes_0 = const()[name = string("op_3195_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3195_keep_dims_0 = const()[name = string("op_3195_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3195_cast_fp16 = reduce_mean(axes = var_3195_axes_0, keep_dims = var_3195_keep_dims_0, x = var_3193_cast_fp16)[name = string("op_3195_cast_fp16")];
            fp16 var_3185_promoted_1_to_fp16 = const()[name = string("op_3185_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3196_cast_fp16 = pow(x = peak_43_cast_fp16, y = var_3185_promoted_1_to_fp16)[name = string("op_3196_cast_fp16")];
            fp32 var_3197_epsilon_0 = const()[name = string("op_3197_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3197_cast_fp16 = inverse(epsilon = var_3197_epsilon_0, x = var_3196_cast_fp16)[name = string("op_3197_cast_fp16")];
            fp16 var_3198_to_fp16 = const()[name = string("op_3198_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3199_cast_fp16 = mul(x = var_3197_cast_fp16, y = var_3198_to_fp16)[name = string("op_3199_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3200_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3199_cast_fp16)[name = string("op_3200_cast_fp16")];
            fp32 var_3201_epsilon_0 = const()[name = string("op_3201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3201_cast_fp16 = rsqrt(epsilon = var_3201_epsilon_0, x = var_3200_cast_fp16)[name = string("op_3201_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3202_cast_fp16 = mul(x = x_107_cast_fp16, y = var_3201_cast_fp16)[name = string("op_3202_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_113_cast_fp16 = mul(x = var_3202_cast_fp16, y = model_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_113_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_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_36_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_109_perm_0 = const()[name = string("x_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3219 = const()[name = string("op_3219"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_109_cast_fp16 = transpose(perm = x_109_perm_0, x = var_3212_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 16, 128]> var_3222_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_3222_cast_fp16")];
            tensor<int32, [1]> peak_45_axes_0 = const()[name = string("peak_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_45_cast_fp16 = reduce_max(axes = peak_45_axes_0, keep_dims = var_3219, x = var_3222_cast_fp16)[name = string("peak_45_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_111_cast_fp16 = real_div(x = x_109_cast_fp16, y = peak_45_cast_fp16)[name = string("x_111_cast_fp16")];
            fp16 var_3218_promoted_to_fp16 = const()[name = string("op_3218_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3226_cast_fp16 = pow(x = x_111_cast_fp16, y = var_3218_promoted_to_fp16)[name = string("op_3226_cast_fp16")];
            tensor<int32, [1]> var_3228_axes_0 = const()[name = string("op_3228_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3228_keep_dims_0 = const()[name = string("op_3228_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3228_cast_fp16 = reduce_mean(axes = var_3228_axes_0, keep_dims = var_3228_keep_dims_0, x = var_3226_cast_fp16)[name = string("op_3228_cast_fp16")];
            fp16 var_3218_promoted_1_to_fp16 = const()[name = string("op_3218_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3229_cast_fp16 = pow(x = peak_45_cast_fp16, y = var_3218_promoted_1_to_fp16)[name = string("op_3229_cast_fp16")];
            fp32 var_3230_epsilon_0 = const()[name = string("op_3230_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3230_cast_fp16 = inverse(epsilon = var_3230_epsilon_0, x = var_3229_cast_fp16)[name = string("op_3230_cast_fp16")];
            fp16 var_3231_to_fp16 = const()[name = string("op_3231_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3232_cast_fp16 = mul(x = var_3230_cast_fp16, y = var_3231_to_fp16)[name = string("op_3232_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3233_cast_fp16 = add(x = var_3228_cast_fp16, y = var_3232_cast_fp16)[name = string("op_3233_cast_fp16")];
            fp32 var_3234_epsilon_0 = const()[name = string("op_3234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3234_cast_fp16 = rsqrt(epsilon = var_3234_epsilon_0, x = var_3233_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3235_cast_fp16 = mul(x = x_111_cast_fp16, y = var_3234_cast_fp16)[name = string("op_3235_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_115_cast_fp16 = mul(x = var_3235_cast_fp16, y = model_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_115_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_3244 = const()[name = string("op_3244"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3245_cast_fp16 = reshape(shape = var_3244, x = linear_37_cast_fp16)[name = string("op_3245_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_113_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_113_cast_fp16)[name = string("second_21_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3259_cast_fp16 = mul(x = first_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3260_cast_fp16 = mul(x = second_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3260_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3262_cast_fp16 = sub(x = var_3259_cast_fp16, y = var_3260_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3263_cast_fp16 = mul(x = second_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3264_cast_fp16 = mul(x = first_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3266_cast_fp16 = add(x = var_3263_cast_fp16, y = var_3264_cast_fp16)[name = string("op_3266_cast_fp16")];
            int32 var_3268 = const()[name = string("op_3268"), 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_3268, interleave = q_11_interleave_0, values = (var_3262_cast_fp16, var_3266_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_115_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_115_cast_fp16)[name = string("second_23_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3280_cast_fp16 = mul(x = first_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3280_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3281_cast_fp16 = mul(x = second_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3281_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3283_cast_fp16 = sub(x = var_3280_cast_fp16, y = var_3281_cast_fp16)[name = string("op_3283_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3284_cast_fp16 = mul(x = second_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3285_cast_fp16 = mul(x = first_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3285_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3287_cast_fp16 = add(x = var_3284_cast_fp16, y = var_3285_cast_fp16)[name = string("op_3287_cast_fp16")];
            int32 var_3289 = const()[name = string("op_3289"), 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_3289, interleave = k_11_interleave_0, values = (var_3283_cast_fp16, var_3287_cast_fp16))[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> var_3300_begin_0 = const()[name = string("op_3300_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3300_end_0 = const()[name = string("op_3300_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3300_end_mask_0 = const()[name = string("op_3300_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3300_stride_0 = const()[name = string("op_3300_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3300_cast_fp16 = slice_by_index(begin = var_3300_begin_0, end = var_3300_end_0, end_mask = var_3300_end_mask_0, stride = var_3300_stride_0, x = q_11_cast_fp16)[name = string("op_3300_cast_fp16")];
            tensor<int32, [4]> var_3310_begin_0 = const()[name = string("op_3310_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3310_end_0 = const()[name = string("op_3310_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3310_end_mask_0 = const()[name = string("op_3310_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3310_stride_0 = const()[name = string("op_3310_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3310_cast_fp16 = slice_by_index(begin = var_3310_begin_0, end = var_3310_end_0, end_mask = var_3310_end_mask_0, stride = var_3310_stride_0, x = q_11_cast_fp16)[name = string("op_3310_cast_fp16")];
            int32 var_3312 = const()[name = string("op_3312"), 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_3312, interleave = stacked_11_interleave_0, values = (var_3300_cast_fp16, var_3310_cast_fp16))[name = string("stacked_11_cast_fp16")];
            tensor<int32, [4]> var_3318_begin_0 = const()[name = string("op_3318_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3318_end_0 = const()[name = string("op_3318_end_0"), val = tensor<int32, [4]>([6, 8, 2064, 128])];
            tensor<bool, [4]> var_3318_end_mask_0 = const()[name = string("op_3318_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3318_cast_fp16 = slice_by_index(begin = var_3318_begin_0, end = var_3318_end_0, end_mask = var_3318_end_mask_0, x = read_state_0)[name = string("op_3318_cast_fp16")];
            tensor<int32, [4]> var_3328_begin_0 = const()[name = string("op_3328_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3328_end_0 = const()[name = string("op_3328_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3328_end_mask_0 = const()[name = string("op_3328_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3328_cast_fp16 = slice_by_index(begin = var_3328_begin_0, end = var_3328_end_0, end_mask = var_3328_end_mask_0, x = var_3318_cast_fp16)[name = string("op_3328_cast_fp16")];
            bool var_3337_transpose_x_1 = const()[name = string("op_3337_transpose_x_1"), val = bool(false)];
            bool var_3337_transpose_y_1 = const()[name = string("op_3337_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_3337_cast_fp16 = matmul(transpose_x = var_3337_transpose_x_1, transpose_y = var_3337_transpose_y_1, x = var_3328_cast_fp16, y = stacked_11_cast_fp16)[name = string("op_3337_cast_fp16")];
            tensor<int32, [4]> var_3340_perm_0 = const()[name = string("op_3340_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3341_to_fp16 = const()[name = string("op_3341_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_3340_cast_fp16 = transpose(perm = var_3340_perm_0, x = var_3337_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 8, 32, 2048]> var_3342_cast_fp16 = mul(x = var_3340_cast_fp16, y = var_3341_to_fp16)[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_11_cast_fp16 = add(x = var_3342_cast_fp16, y = cacheMask)[name = string("cache_weights_11_cast_fp16")];
            bool var_3348_transpose_x_1 = const()[name = string("op_3348_transpose_x_1"), val = bool(false)];
            bool var_3348_transpose_y_1 = const()[name = string("op_3348_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3348_cast_fp16 = matmul(transpose_x = var_3348_transpose_x_1, transpose_y = var_3348_transpose_y_1, x = stacked_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_3348_cast_fp16")];
            fp16 var_3349_to_fp16 = const()[name = string("op_3349_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3350_cast_fp16 = mul(x = var_3348_cast_fp16, y = var_3349_to_fp16)[name = string("op_3350_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_11_cast_fp16 = add(x = var_3350_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_11_cast_fp16")];
            int32 var_3354 = const()[name = string("op_3354"), val = int32(-1)];
            bool var_3355_interleave_0 = const()[name = string("op_3355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_3355_cast_fp16 = concat(axis = var_3354, interleave = var_3355_interleave_0, values = (cache_weights_11_cast_fp16, call_weights_11_cast_fp16))[name = string("op_3355_cast_fp16")];
            int32 var_3356 = const()[name = string("op_3356"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_11_cast_fp16 = softmax(axis = var_3356, x = var_3355_cast_fp16)[name = string("shares_11_cast_fp16")];
            tensor<int32, [4]> var_3363_begin_0 = const()[name = string("op_3363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3363_end_0 = const()[name = string("op_3363_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_3363_end_mask_0 = const()[name = string("op_3363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_3363_cast_fp16 = slice_by_index(begin = var_3363_begin_0, end = var_3363_end_0, end_mask = var_3363_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3363_cast_fp16")];
            tensor<int32, [4]> var_3368_begin_0 = const()[name = string("op_3368_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3368_end_0 = const()[name = string("op_3368_end_0"), val = tensor<int32, [4]>([6, 8, 2064, 128])];
            tensor<bool, [4]> var_3368_end_mask_0 = const()[name = string("op_3368_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3368_cast_fp16 = slice_by_index(begin = var_3368_begin_0, end = var_3368_end_0, end_mask = var_3368_end_mask_0, x = read_state_1)[name = string("op_3368_cast_fp16")];
            tensor<int32, [4]> var_3378_begin_0 = const()[name = string("op_3378_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3378_end_0 = const()[name = string("op_3378_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3378_end_mask_0 = const()[name = string("op_3378_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3378_cast_fp16 = slice_by_index(begin = var_3378_begin_0, end = var_3378_end_0, end_mask = var_3378_end_mask_0, x = var_3368_cast_fp16)[name = string("op_3378_cast_fp16")];
            bool var_3384_transpose_x_0 = const()[name = string("op_3384_transpose_x_0"), val = bool(false)];
            bool var_3384_transpose_y_0 = const()[name = string("op_3384_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3384_cast_fp16 = matmul(transpose_x = var_3384_transpose_x_0, transpose_y = var_3384_transpose_y_0, x = var_3363_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<int32, [4]> var_3389_begin_0 = const()[name = string("op_3389_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_3389_end_0 = const()[name = string("op_3389_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_3389_end_mask_0 = const()[name = string("op_3389_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3389_cast_fp16 = slice_by_index(begin = var_3389_begin_0, end = var_3389_end_0, end_mask = var_3389_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3389_cast_fp16")];
            bool var_3390_transpose_x_0 = const()[name = string("op_3390_transpose_x_0"), val = bool(false)];
            bool var_3390_transpose_y_0 = const()[name = string("op_3390_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_3245_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 8, 32, 128]> var_3390_cast_fp16 = matmul(transpose_x = var_3390_transpose_x_0, transpose_y = var_3390_transpose_y_0, x = var_3389_cast_fp16, y = v_11_cast_fp16)[name = string("op_3390_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_21_cast_fp16 = add(x = var_3384_cast_fp16, y = var_3390_cast_fp16)[name = string("mixed_21_cast_fp16")];
            tensor<int32, [4]> var_3407_begin_0 = const()[name = string("op_3407_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3407_end_0 = const()[name = string("op_3407_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3407_end_mask_0 = const()[name = string("op_3407_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3407_cast_fp16")];
            tensor<int32, [4]> var_3422_begin_0 = const()[name = string("op_3422_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3422_end_0 = const()[name = string("op_3422_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3422_end_mask_0 = const()[name = string("op_3422_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3422_cast_fp16 = slice_by_index(begin = var_3422_begin_0, end = var_3422_end_0, end_mask = var_3422_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3422_cast_fp16")];
            int32 var_3424 = const()[name = string("op_3424"), 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_3424, interleave = mixed_23_interleave_0, values = (var_3407_cast_fp16, var_3422_cast_fp16))[name = string("mixed_23_cast_fp16")];
            tensor<int32, [4]> var_3428_perm_0 = const()[name = string("op_3428_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3432 = const()[name = string("op_3432"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3428_cast_fp16 = transpose(perm = var_3428_perm_0, x = mixed_23_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 2048]> input_53_cast_fp16 = reshape(shape = var_3432, x = var_3428_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_117_cast_fp16 = add(x = x_101_cast_fp16, y = linear_38_cast_fp16)[name = string("x_117_cast_fp16")];
            bool var_3442 = const()[name = string("op_3442"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3445_cast_fp16 = abs(x = x_117_cast_fp16)[name = string("op_3445_cast_fp16")];
            tensor<int32, [1]> peak_47_axes_0 = const()[name = string("peak_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_47_cast_fp16 = reduce_max(axes = peak_47_axes_0, keep_dims = var_3442, x = var_3445_cast_fp16)[name = string("peak_47_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_119_cast_fp16 = real_div(x = x_117_cast_fp16, y = peak_47_cast_fp16)[name = string("x_119_cast_fp16")];
            fp16 var_3441_promoted_to_fp16 = const()[name = string("op_3441_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3449_cast_fp16 = pow(x = x_119_cast_fp16, y = var_3441_promoted_to_fp16)[name = string("op_3449_cast_fp16")];
            tensor<int32, [1]> var_3451_axes_0 = const()[name = string("op_3451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3451_keep_dims_0 = const()[name = string("op_3451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3451_cast_fp16 = reduce_mean(axes = var_3451_axes_0, keep_dims = var_3451_keep_dims_0, x = var_3449_cast_fp16)[name = string("op_3451_cast_fp16")];
            fp16 var_3441_promoted_1_to_fp16 = const()[name = string("op_3441_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3452_cast_fp16 = pow(x = peak_47_cast_fp16, y = var_3441_promoted_1_to_fp16)[name = string("op_3452_cast_fp16")];
            fp32 var_3453_epsilon_0 = const()[name = string("op_3453_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3453_cast_fp16 = inverse(epsilon = var_3453_epsilon_0, x = var_3452_cast_fp16)[name = string("op_3453_cast_fp16")];
            fp16 var_3454_to_fp16 = const()[name = string("op_3454_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3455_cast_fp16 = mul(x = var_3453_cast_fp16, y = var_3454_to_fp16)[name = string("op_3455_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3456_cast_fp16 = add(x = var_3451_cast_fp16, y = var_3455_cast_fp16)[name = string("op_3456_cast_fp16")];
            fp32 var_3457_epsilon_0 = const()[name = string("op_3457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3457_cast_fp16 = rsqrt(epsilon = var_3457_epsilon_0, x = var_3456_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3458_cast_fp16 = mul(x = x_119_cast_fp16, y = var_3457_cast_fp16)[name = string("op_3458_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_3458_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_3463_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_3463_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_3463_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_121_cast_fp16 = add(x = x_117_cast_fp16, y = linear_41_cast_fp16)[name = string("x_121_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3479_cast_fp16 = abs(x = x_121_cast_fp16)[name = string("op_3479_cast_fp16")];
            tensor<int32, [1]> peak_49_axes_0 = const()[name = string("peak_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_49_cast_fp16 = reduce_max(axes = peak_49_axes_0, keep_dims = var_3476, x = var_3479_cast_fp16)[name = string("peak_49_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_123_cast_fp16 = real_div(x = x_121_cast_fp16, y = peak_49_cast_fp16)[name = string("x_123_cast_fp16")];
            fp16 var_3475_promoted_to_fp16 = const()[name = string("op_3475_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3483_cast_fp16 = pow(x = x_123_cast_fp16, y = var_3475_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            tensor<int32, [1]> var_3485_axes_0 = const()[name = string("op_3485_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3485_keep_dims_0 = const()[name = string("op_3485_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3485_cast_fp16 = reduce_mean(axes = var_3485_axes_0, keep_dims = var_3485_keep_dims_0, x = var_3483_cast_fp16)[name = string("op_3485_cast_fp16")];
            fp16 var_3475_promoted_1_to_fp16 = const()[name = string("op_3475_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3486_cast_fp16 = pow(x = peak_49_cast_fp16, y = var_3475_promoted_1_to_fp16)[name = string("op_3486_cast_fp16")];
            fp32 var_3487_epsilon_0 = const()[name = string("op_3487_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3487_cast_fp16 = inverse(epsilon = var_3487_epsilon_0, x = var_3486_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp16 var_3488_to_fp16 = const()[name = string("op_3488_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3489_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_3488_to_fp16)[name = string("op_3489_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3490_cast_fp16 = add(x = var_3485_cast_fp16, y = var_3489_cast_fp16)[name = string("op_3490_cast_fp16")];
            fp32 var_3491_epsilon_0 = const()[name = string("op_3491_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3491_cast_fp16 = rsqrt(epsilon = var_3491_epsilon_0, x = var_3490_cast_fp16)[name = string("op_3491_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3492_cast_fp16 = mul(x = x_123_cast_fp16, y = var_3491_cast_fp16)[name = string("op_3492_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_3492_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_3501 = const()[name = string("op_3501"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3502_cast_fp16 = reshape(shape = var_3501, x = linear_42_cast_fp16)[name = string("op_3502_cast_fp16")];
            tensor<int32, [4]> x_125_perm_0 = const()[name = string("x_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3509 = const()[name = string("op_3509"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_125_cast_fp16 = transpose(perm = x_125_perm_0, x = var_3502_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 16, 128]> var_3512_cast_fp16 = abs(x = x_125_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<int32, [1]> peak_51_axes_0 = const()[name = string("peak_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_51_cast_fp16 = reduce_max(axes = peak_51_axes_0, keep_dims = var_3509, x = var_3512_cast_fp16)[name = string("peak_51_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_127_cast_fp16 = real_div(x = x_125_cast_fp16, y = peak_51_cast_fp16)[name = string("x_127_cast_fp16")];
            fp16 var_3508_promoted_to_fp16 = const()[name = string("op_3508_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3516_cast_fp16 = pow(x = x_127_cast_fp16, y = var_3508_promoted_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<int32, [1]> var_3518_axes_0 = const()[name = string("op_3518_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3518_keep_dims_0 = const()[name = string("op_3518_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3518_cast_fp16 = reduce_mean(axes = var_3518_axes_0, keep_dims = var_3518_keep_dims_0, x = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp16 var_3508_promoted_1_to_fp16 = const()[name = string("op_3508_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3519_cast_fp16 = pow(x = peak_51_cast_fp16, y = var_3508_promoted_1_to_fp16)[name = string("op_3519_cast_fp16")];
            fp32 var_3520_epsilon_0 = const()[name = string("op_3520_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3520_cast_fp16 = inverse(epsilon = var_3520_epsilon_0, x = var_3519_cast_fp16)[name = string("op_3520_cast_fp16")];
            fp16 var_3521_to_fp16 = const()[name = string("op_3521_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3522_cast_fp16 = mul(x = var_3520_cast_fp16, y = var_3521_to_fp16)[name = string("op_3522_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3523_cast_fp16 = add(x = var_3518_cast_fp16, y = var_3522_cast_fp16)[name = string("op_3523_cast_fp16")];
            fp32 var_3524_epsilon_0 = const()[name = string("op_3524_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3524_cast_fp16 = rsqrt(epsilon = var_3524_epsilon_0, x = var_3523_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3525_cast_fp16 = mul(x = x_127_cast_fp16, y = var_3524_cast_fp16)[name = string("op_3525_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_133_cast_fp16 = mul(x = var_3525_cast_fp16, y = model_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_133_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_3534 = const()[name = string("op_3534"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3535_cast_fp16 = reshape(shape = var_3534, x = linear_43_cast_fp16)[name = string("op_3535_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3542 = const()[name = string("op_3542"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_3535_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 8, 16, 128]> var_3545_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_3545_cast_fp16")];
            tensor<int32, [1]> peak_53_axes_0 = const()[name = string("peak_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_53_cast_fp16 = reduce_max(axes = peak_53_axes_0, keep_dims = var_3542, x = var_3545_cast_fp16)[name = string("peak_53_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_131_cast_fp16 = real_div(x = x_129_cast_fp16, y = peak_53_cast_fp16)[name = string("x_131_cast_fp16")];
            fp16 var_3541_promoted_to_fp16 = const()[name = string("op_3541_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3549_cast_fp16 = pow(x = x_131_cast_fp16, y = var_3541_promoted_to_fp16)[name = string("op_3549_cast_fp16")];
            tensor<int32, [1]> var_3551_axes_0 = const()[name = string("op_3551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3551_keep_dims_0 = const()[name = string("op_3551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3551_cast_fp16 = reduce_mean(axes = var_3551_axes_0, keep_dims = var_3551_keep_dims_0, x = var_3549_cast_fp16)[name = string("op_3551_cast_fp16")];
            fp16 var_3541_promoted_1_to_fp16 = const()[name = string("op_3541_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3552_cast_fp16 = pow(x = peak_53_cast_fp16, y = var_3541_promoted_1_to_fp16)[name = string("op_3552_cast_fp16")];
            fp32 var_3553_epsilon_0 = const()[name = string("op_3553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3553_cast_fp16 = inverse(epsilon = var_3553_epsilon_0, x = var_3552_cast_fp16)[name = string("op_3553_cast_fp16")];
            fp16 var_3554_to_fp16 = const()[name = string("op_3554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3555_cast_fp16 = mul(x = var_3553_cast_fp16, y = var_3554_to_fp16)[name = string("op_3555_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3556_cast_fp16 = add(x = var_3551_cast_fp16, y = var_3555_cast_fp16)[name = string("op_3556_cast_fp16")];
            fp32 var_3557_epsilon_0 = const()[name = string("op_3557_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3557_cast_fp16 = rsqrt(epsilon = var_3557_epsilon_0, x = var_3556_cast_fp16)[name = string("op_3557_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3558_cast_fp16 = mul(x = x_131_cast_fp16, y = var_3557_cast_fp16)[name = string("op_3558_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_135_cast_fp16 = mul(x = var_3558_cast_fp16, y = model_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_135_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_3567 = const()[name = string("op_3567"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3568_cast_fp16 = reshape(shape = var_3567, x = linear_44_cast_fp16)[name = string("op_3568_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_133_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_133_cast_fp16)[name = string("second_25_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3582_cast_fp16 = mul(x = first_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3582_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3583_cast_fp16 = mul(x = second_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3585_cast_fp16 = sub(x = var_3582_cast_fp16, y = var_3583_cast_fp16)[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3586_cast_fp16 = mul(x = second_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3587_cast_fp16 = mul(x = first_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3587_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3589_cast_fp16 = add(x = var_3586_cast_fp16, y = var_3587_cast_fp16)[name = string("op_3589_cast_fp16")];
            int32 var_3591 = const()[name = string("op_3591"), 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_3591, interleave = q_13_interleave_0, values = (var_3585_cast_fp16, var_3589_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_135_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_135_cast_fp16)[name = string("second_27_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3603_cast_fp16 = mul(x = first_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3603_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3604_cast_fp16 = mul(x = second_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3604_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3606_cast_fp16 = sub(x = var_3603_cast_fp16, y = var_3604_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3607_cast_fp16 = mul(x = second_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3607_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3608_cast_fp16 = mul(x = first_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3608_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3610_cast_fp16 = add(x = var_3607_cast_fp16, y = var_3608_cast_fp16)[name = string("op_3610_cast_fp16")];
            int32 var_3612 = const()[name = string("op_3612"), 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_3612, interleave = k_13_interleave_0, values = (var_3606_cast_fp16, var_3610_cast_fp16))[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_3623_begin_0 = const()[name = string("op_3623_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3623_end_0 = const()[name = string("op_3623_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3623_end_mask_0 = const()[name = string("op_3623_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3623_stride_0 = const()[name = string("op_3623_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3623_cast_fp16 = slice_by_index(begin = var_3623_begin_0, end = var_3623_end_0, end_mask = var_3623_end_mask_0, stride = var_3623_stride_0, x = q_13_cast_fp16)[name = string("op_3623_cast_fp16")];
            tensor<int32, [4]> var_3633_begin_0 = const()[name = string("op_3633_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3633_end_0 = const()[name = string("op_3633_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3633_end_mask_0 = const()[name = string("op_3633_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3633_stride_0 = const()[name = string("op_3633_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3633_cast_fp16 = slice_by_index(begin = var_3633_begin_0, end = var_3633_end_0, end_mask = var_3633_end_mask_0, stride = var_3633_stride_0, x = q_13_cast_fp16)[name = string("op_3633_cast_fp16")];
            int32 var_3635 = const()[name = string("op_3635"), 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_3635, interleave = stacked_13_interleave_0, values = (var_3623_cast_fp16, var_3633_cast_fp16))[name = string("stacked_13_cast_fp16")];
            tensor<int32, [4]> var_3641_begin_0 = const()[name = string("op_3641_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3641_end_0 = const()[name = string("op_3641_end_0"), val = tensor<int32, [4]>([7, 8, 2064, 128])];
            tensor<bool, [4]> var_3641_end_mask_0 = const()[name = string("op_3641_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3641_cast_fp16 = slice_by_index(begin = var_3641_begin_0, end = var_3641_end_0, end_mask = var_3641_end_mask_0, x = read_state_0)[name = string("op_3641_cast_fp16")];
            tensor<int32, [4]> var_3651_begin_0 = const()[name = string("op_3651_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3651_end_0 = const()[name = string("op_3651_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3651_end_mask_0 = const()[name = string("op_3651_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3651_cast_fp16 = slice_by_index(begin = var_3651_begin_0, end = var_3651_end_0, end_mask = var_3651_end_mask_0, x = var_3641_cast_fp16)[name = string("op_3651_cast_fp16")];
            bool var_3660_transpose_x_1 = const()[name = string("op_3660_transpose_x_1"), val = bool(false)];
            bool var_3660_transpose_y_1 = const()[name = string("op_3660_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_3660_cast_fp16 = matmul(transpose_x = var_3660_transpose_x_1, transpose_y = var_3660_transpose_y_1, x = var_3651_cast_fp16, y = stacked_13_cast_fp16)[name = string("op_3660_cast_fp16")];
            tensor<int32, [4]> var_3663_perm_0 = const()[name = string("op_3663_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3664_to_fp16 = const()[name = string("op_3664_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_3663_cast_fp16 = transpose(perm = var_3663_perm_0, x = var_3660_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 8, 32, 2048]> var_3665_cast_fp16 = mul(x = var_3663_cast_fp16, y = var_3664_to_fp16)[name = string("op_3665_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_13_cast_fp16 = add(x = var_3665_cast_fp16, y = cacheMask)[name = string("cache_weights_13_cast_fp16")];
            bool var_3671_transpose_x_1 = const()[name = string("op_3671_transpose_x_1"), val = bool(false)];
            bool var_3671_transpose_y_1 = const()[name = string("op_3671_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3671_cast_fp16 = matmul(transpose_x = var_3671_transpose_x_1, transpose_y = var_3671_transpose_y_1, x = stacked_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_3671_cast_fp16")];
            fp16 var_3672_to_fp16 = const()[name = string("op_3672_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3673_cast_fp16 = mul(x = var_3671_cast_fp16, y = var_3672_to_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_13_cast_fp16 = add(x = var_3673_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_13_cast_fp16")];
            int32 var_3677 = const()[name = string("op_3677"), val = int32(-1)];
            bool var_3678_interleave_0 = const()[name = string("op_3678_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_3678_cast_fp16 = concat(axis = var_3677, interleave = var_3678_interleave_0, values = (cache_weights_13_cast_fp16, call_weights_13_cast_fp16))[name = string("op_3678_cast_fp16")];
            int32 var_3679 = const()[name = string("op_3679"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_13_cast_fp16 = softmax(axis = var_3679, x = var_3678_cast_fp16)[name = string("shares_13_cast_fp16")];
            tensor<int32, [4]> var_3686_begin_0 = const()[name = string("op_3686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3686_end_0 = const()[name = string("op_3686_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_3686_end_mask_0 = const()[name = string("op_3686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_3686_cast_fp16 = slice_by_index(begin = var_3686_begin_0, end = var_3686_end_0, end_mask = var_3686_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3686_cast_fp16")];
            tensor<int32, [4]> var_3691_begin_0 = const()[name = string("op_3691_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3691_end_0 = const()[name = string("op_3691_end_0"), val = tensor<int32, [4]>([7, 8, 2064, 128])];
            tensor<bool, [4]> var_3691_end_mask_0 = const()[name = string("op_3691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3691_cast_fp16 = slice_by_index(begin = var_3691_begin_0, end = var_3691_end_0, end_mask = var_3691_end_mask_0, x = read_state_1)[name = string("op_3691_cast_fp16")];
            tensor<int32, [4]> var_3701_begin_0 = const()[name = string("op_3701_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3701_end_0 = const()[name = string("op_3701_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3701_end_mask_0 = const()[name = string("op_3701_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3701_cast_fp16 = slice_by_index(begin = var_3701_begin_0, end = var_3701_end_0, end_mask = var_3701_end_mask_0, x = var_3691_cast_fp16)[name = string("op_3701_cast_fp16")];
            bool var_3707_transpose_x_0 = const()[name = string("op_3707_transpose_x_0"), val = bool(false)];
            bool var_3707_transpose_y_0 = const()[name = string("op_3707_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3707_cast_fp16 = matmul(transpose_x = var_3707_transpose_x_0, transpose_y = var_3707_transpose_y_0, x = var_3686_cast_fp16, y = var_3701_cast_fp16)[name = string("op_3707_cast_fp16")];
            tensor<int32, [4]> var_3712_begin_0 = const()[name = string("op_3712_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_3712_end_0 = const()[name = string("op_3712_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_3712_end_mask_0 = const()[name = string("op_3712_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3712_cast_fp16 = slice_by_index(begin = var_3712_begin_0, end = var_3712_end_0, end_mask = var_3712_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3712_cast_fp16")];
            bool var_3713_transpose_x_0 = const()[name = string("op_3713_transpose_x_0"), val = bool(false)];
            bool var_3713_transpose_y_0 = const()[name = string("op_3713_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3568_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 8, 32, 128]> var_3713_cast_fp16 = matmul(transpose_x = var_3713_transpose_x_0, transpose_y = var_3713_transpose_y_0, x = var_3712_cast_fp16, y = v_13_cast_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_25_cast_fp16 = add(x = var_3707_cast_fp16, y = var_3713_cast_fp16)[name = string("mixed_25_cast_fp16")];
            tensor<int32, [4]> var_3730_begin_0 = const()[name = string("op_3730_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3730_end_0 = const()[name = string("op_3730_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3730_end_mask_0 = const()[name = string("op_3730_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3730_cast_fp16 = slice_by_index(begin = var_3730_begin_0, end = var_3730_end_0, end_mask = var_3730_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3730_cast_fp16")];
            tensor<int32, [4]> var_3745_begin_0 = const()[name = string("op_3745_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3745_end_0 = const()[name = string("op_3745_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3745_end_mask_0 = const()[name = string("op_3745_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3745_cast_fp16 = slice_by_index(begin = var_3745_begin_0, end = var_3745_end_0, end_mask = var_3745_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3745_cast_fp16")];
            int32 var_3747 = const()[name = string("op_3747"), 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_3747, interleave = mixed_27_interleave_0, values = (var_3730_cast_fp16, var_3745_cast_fp16))[name = string("mixed_27_cast_fp16")];
            tensor<int32, [4]> var_3751_perm_0 = const()[name = string("op_3751_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3755 = const()[name = string("op_3755"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3751_cast_fp16 = transpose(perm = var_3751_perm_0, x = mixed_27_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 2048]> input_63_cast_fp16 = reshape(shape = var_3755, x = var_3751_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_137_cast_fp16 = add(x = x_121_cast_fp16, y = linear_45_cast_fp16)[name = string("x_137_cast_fp16")];
            bool var_3765 = const()[name = string("op_3765"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3768_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_3768_cast_fp16")];
            tensor<int32, [1]> peak_55_axes_0 = const()[name = string("peak_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_55_cast_fp16 = reduce_max(axes = peak_55_axes_0, keep_dims = var_3765, x = var_3768_cast_fp16)[name = string("peak_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_139_cast_fp16 = real_div(x = x_137_cast_fp16, y = peak_55_cast_fp16)[name = string("x_139_cast_fp16")];
            fp16 var_3764_promoted_to_fp16 = const()[name = string("op_3764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3772_cast_fp16 = pow(x = x_139_cast_fp16, y = var_3764_promoted_to_fp16)[name = string("op_3772_cast_fp16")];
            tensor<int32, [1]> var_3774_axes_0 = const()[name = string("op_3774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3774_keep_dims_0 = const()[name = string("op_3774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3774_cast_fp16 = reduce_mean(axes = var_3774_axes_0, keep_dims = var_3774_keep_dims_0, x = var_3772_cast_fp16)[name = string("op_3774_cast_fp16")];
            fp16 var_3764_promoted_1_to_fp16 = const()[name = string("op_3764_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3775_cast_fp16 = pow(x = peak_55_cast_fp16, y = var_3764_promoted_1_to_fp16)[name = string("op_3775_cast_fp16")];
            fp32 var_3776_epsilon_0 = const()[name = string("op_3776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3776_cast_fp16 = inverse(epsilon = var_3776_epsilon_0, x = var_3775_cast_fp16)[name = string("op_3776_cast_fp16")];
            fp16 var_3777_to_fp16 = const()[name = string("op_3777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3778_cast_fp16 = mul(x = var_3776_cast_fp16, y = var_3777_to_fp16)[name = string("op_3778_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3779_cast_fp16 = add(x = var_3774_cast_fp16, y = var_3778_cast_fp16)[name = string("op_3779_cast_fp16")];
            fp32 var_3780_epsilon_0 = const()[name = string("op_3780_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3780_cast_fp16 = rsqrt(epsilon = var_3780_epsilon_0, x = var_3779_cast_fp16)[name = string("op_3780_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3781_cast_fp16 = mul(x = x_139_cast_fp16, y = var_3780_cast_fp16)[name = string("op_3781_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_3781_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_3786_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_3786_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_3786_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_141_cast_fp16 = add(x = x_137_cast_fp16, y = linear_48_cast_fp16)[name = string("x_141_cast_fp16")];
            bool var_3799 = const()[name = string("op_3799"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3802_cast_fp16 = abs(x = x_141_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [1]> peak_57_axes_0 = const()[name = string("peak_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_57_cast_fp16 = reduce_max(axes = peak_57_axes_0, keep_dims = var_3799, x = var_3802_cast_fp16)[name = string("peak_57_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_143_cast_fp16 = real_div(x = x_141_cast_fp16, y = peak_57_cast_fp16)[name = string("x_143_cast_fp16")];
            fp16 var_3798_promoted_to_fp16 = const()[name = string("op_3798_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3806_cast_fp16 = pow(x = x_143_cast_fp16, y = var_3798_promoted_to_fp16)[name = string("op_3806_cast_fp16")];
            tensor<int32, [1]> var_3808_axes_0 = const()[name = string("op_3808_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3808_keep_dims_0 = const()[name = string("op_3808_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3808_cast_fp16 = reduce_mean(axes = var_3808_axes_0, keep_dims = var_3808_keep_dims_0, x = var_3806_cast_fp16)[name = string("op_3808_cast_fp16")];
            fp16 var_3798_promoted_1_to_fp16 = const()[name = string("op_3798_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3809_cast_fp16 = pow(x = peak_57_cast_fp16, y = var_3798_promoted_1_to_fp16)[name = string("op_3809_cast_fp16")];
            fp32 var_3810_epsilon_0 = const()[name = string("op_3810_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3810_cast_fp16 = inverse(epsilon = var_3810_epsilon_0, x = var_3809_cast_fp16)[name = string("op_3810_cast_fp16")];
            fp16 var_3811_to_fp16 = const()[name = string("op_3811_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3812_cast_fp16 = mul(x = var_3810_cast_fp16, y = var_3811_to_fp16)[name = string("op_3812_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3813_cast_fp16 = add(x = var_3808_cast_fp16, y = var_3812_cast_fp16)[name = string("op_3813_cast_fp16")];
            fp32 var_3814_epsilon_0 = const()[name = string("op_3814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3814_cast_fp16 = rsqrt(epsilon = var_3814_epsilon_0, x = var_3813_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3815_cast_fp16 = mul(x = x_143_cast_fp16, y = var_3814_cast_fp16)[name = string("op_3815_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_3815_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_3824 = const()[name = string("op_3824"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3825_cast_fp16 = reshape(shape = var_3824, x = linear_49_cast_fp16)[name = string("op_3825_cast_fp16")];
            tensor<int32, [4]> x_145_perm_0 = const()[name = string("x_145_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_145_cast_fp16 = transpose(perm = x_145_perm_0, x = var_3825_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 16, 16, 128]> var_3835_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_3835_cast_fp16")];
            tensor<int32, [1]> peak_59_axes_0 = const()[name = string("peak_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_59_cast_fp16 = reduce_max(axes = peak_59_axes_0, keep_dims = var_3832, x = var_3835_cast_fp16)[name = string("peak_59_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_147_cast_fp16 = real_div(x = x_145_cast_fp16, y = peak_59_cast_fp16)[name = string("x_147_cast_fp16")];
            fp16 var_3831_promoted_to_fp16 = const()[name = string("op_3831_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3839_cast_fp16 = pow(x = x_147_cast_fp16, y = var_3831_promoted_to_fp16)[name = string("op_3839_cast_fp16")];
            tensor<int32, [1]> var_3841_axes_0 = const()[name = string("op_3841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3841_keep_dims_0 = const()[name = string("op_3841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3841_cast_fp16 = reduce_mean(axes = var_3841_axes_0, keep_dims = var_3841_keep_dims_0, x = var_3839_cast_fp16)[name = string("op_3841_cast_fp16")];
            fp16 var_3831_promoted_1_to_fp16 = const()[name = string("op_3831_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3842_cast_fp16 = pow(x = peak_59_cast_fp16, y = var_3831_promoted_1_to_fp16)[name = string("op_3842_cast_fp16")];
            fp32 var_3843_epsilon_0 = const()[name = string("op_3843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3843_cast_fp16 = inverse(epsilon = var_3843_epsilon_0, x = var_3842_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp16 var_3844_to_fp16 = const()[name = string("op_3844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3845_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3844_to_fp16)[name = string("op_3845_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3846_cast_fp16 = add(x = var_3841_cast_fp16, y = var_3845_cast_fp16)[name = string("op_3846_cast_fp16")];
            fp32 var_3847_epsilon_0 = const()[name = string("op_3847_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3847_cast_fp16 = rsqrt(epsilon = var_3847_epsilon_0, x = var_3846_cast_fp16)[name = string("op_3847_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3848_cast_fp16 = mul(x = x_147_cast_fp16, y = var_3847_cast_fp16)[name = string("op_3848_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_153_cast_fp16 = mul(x = var_3848_cast_fp16, y = model_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_153_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_3857 = const()[name = string("op_3857"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3858_cast_fp16 = reshape(shape = var_3857, x = linear_50_cast_fp16)[name = string("op_3858_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3865 = const()[name = string("op_3865"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_3858_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 8, 16, 128]> var_3868_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_3868_cast_fp16")];
            tensor<int32, [1]> peak_61_axes_0 = const()[name = string("peak_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_61_cast_fp16 = reduce_max(axes = peak_61_axes_0, keep_dims = var_3865, x = var_3868_cast_fp16)[name = string("peak_61_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_151_cast_fp16 = real_div(x = x_149_cast_fp16, y = peak_61_cast_fp16)[name = string("x_151_cast_fp16")];
            fp16 var_3864_promoted_to_fp16 = const()[name = string("op_3864_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3872_cast_fp16 = pow(x = x_151_cast_fp16, y = var_3864_promoted_to_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [1]> var_3874_axes_0 = const()[name = string("op_3874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3874_keep_dims_0 = const()[name = string("op_3874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3874_cast_fp16 = reduce_mean(axes = var_3874_axes_0, keep_dims = var_3874_keep_dims_0, x = var_3872_cast_fp16)[name = string("op_3874_cast_fp16")];
            fp16 var_3864_promoted_1_to_fp16 = const()[name = string("op_3864_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3875_cast_fp16 = pow(x = peak_61_cast_fp16, y = var_3864_promoted_1_to_fp16)[name = string("op_3875_cast_fp16")];
            fp32 var_3876_epsilon_0 = const()[name = string("op_3876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3876_cast_fp16 = inverse(epsilon = var_3876_epsilon_0, x = var_3875_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, 8, 16, 1]> var_3878_cast_fp16 = mul(x = var_3876_cast_fp16, y = var_3877_to_fp16)[name = string("op_3878_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3879_cast_fp16 = add(x = var_3874_cast_fp16, y = var_3878_cast_fp16)[name = string("op_3879_cast_fp16")];
            fp32 var_3880_epsilon_0 = const()[name = string("op_3880_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3880_cast_fp16 = rsqrt(epsilon = var_3880_epsilon_0, x = var_3879_cast_fp16)[name = string("op_3880_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3881_cast_fp16 = mul(x = x_151_cast_fp16, y = var_3880_cast_fp16)[name = string("op_3881_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_155_cast_fp16 = mul(x = var_3881_cast_fp16, y = model_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_155_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_3890 = const()[name = string("op_3890"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3891_cast_fp16 = reshape(shape = var_3890, x = linear_51_cast_fp16)[name = string("op_3891_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_153_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_153_cast_fp16)[name = string("second_29_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3905_cast_fp16 = mul(x = first_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3906_cast_fp16 = mul(x = second_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3908_cast_fp16 = sub(x = var_3905_cast_fp16, y = var_3906_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3909_cast_fp16 = mul(x = second_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3909_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3910_cast_fp16 = mul(x = first_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3910_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3912_cast_fp16 = add(x = var_3909_cast_fp16, y = var_3910_cast_fp16)[name = string("op_3912_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), 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_3914, interleave = q_15_interleave_0, values = (var_3908_cast_fp16, var_3912_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_155_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_155_cast_fp16)[name = string("second_31_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3926_cast_fp16 = mul(x = first_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3927_cast_fp16 = mul(x = second_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3927_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3929_cast_fp16 = sub(x = var_3926_cast_fp16, y = var_3927_cast_fp16)[name = string("op_3929_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3930_cast_fp16 = mul(x = second_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3930_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3931_cast_fp16 = mul(x = first_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3933_cast_fp16 = add(x = var_3930_cast_fp16, y = var_3931_cast_fp16)[name = string("op_3933_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), 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_3935, interleave = k_15_interleave_0, values = (var_3929_cast_fp16, var_3933_cast_fp16))[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> var_3946_begin_0 = const()[name = string("op_3946_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3946_end_0 = const()[name = string("op_3946_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3946_end_mask_0 = const()[name = string("op_3946_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3946_stride_0 = const()[name = string("op_3946_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3946_cast_fp16 = slice_by_index(begin = var_3946_begin_0, end = var_3946_end_0, end_mask = var_3946_end_mask_0, stride = var_3946_stride_0, x = q_15_cast_fp16)[name = string("op_3946_cast_fp16")];
            tensor<int32, [4]> var_3956_begin_0 = const()[name = string("op_3956_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3956_end_0 = const()[name = string("op_3956_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3956_end_mask_0 = const()[name = string("op_3956_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3956_stride_0 = const()[name = string("op_3956_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3956_cast_fp16 = slice_by_index(begin = var_3956_begin_0, end = var_3956_end_0, end_mask = var_3956_end_mask_0, stride = var_3956_stride_0, x = q_15_cast_fp16)[name = string("op_3956_cast_fp16")];
            int32 var_3958 = const()[name = string("op_3958"), 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_3958, interleave = stacked_15_interleave_0, values = (var_3946_cast_fp16, var_3956_cast_fp16))[name = string("stacked_15_cast_fp16")];
            tensor<int32, [4]> var_3964_begin_0 = const()[name = string("op_3964_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3964_end_0 = const()[name = string("op_3964_end_0"), val = tensor<int32, [4]>([8, 8, 2064, 128])];
            tensor<bool, [4]> var_3964_end_mask_0 = const()[name = string("op_3964_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3964_cast_fp16 = slice_by_index(begin = var_3964_begin_0, end = var_3964_end_0, end_mask = var_3964_end_mask_0, x = read_state_0)[name = string("op_3964_cast_fp16")];
            tensor<int32, [4]> var_3974_begin_0 = const()[name = string("op_3974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3974_end_0 = const()[name = string("op_3974_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3974_end_mask_0 = const()[name = string("op_3974_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3974_cast_fp16 = slice_by_index(begin = var_3974_begin_0, end = var_3974_end_0, end_mask = var_3974_end_mask_0, x = var_3964_cast_fp16)[name = string("op_3974_cast_fp16")];
            bool var_3983_transpose_x_1 = const()[name = string("op_3983_transpose_x_1"), val = bool(false)];
            bool var_3983_transpose_y_1 = const()[name = string("op_3983_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_3983_cast_fp16 = matmul(transpose_x = var_3983_transpose_x_1, transpose_y = var_3983_transpose_y_1, x = var_3974_cast_fp16, y = stacked_15_cast_fp16)[name = string("op_3983_cast_fp16")];
            tensor<int32, [4]> var_3986_perm_0 = const()[name = string("op_3986_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3987_to_fp16 = const()[name = string("op_3987_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_3986_cast_fp16 = transpose(perm = var_3986_perm_0, x = var_3983_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 8, 32, 2048]> var_3988_cast_fp16 = mul(x = var_3986_cast_fp16, y = var_3987_to_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_15_cast_fp16 = add(x = var_3988_cast_fp16, y = cacheMask)[name = string("cache_weights_15_cast_fp16")];
            bool var_3994_transpose_x_1 = const()[name = string("op_3994_transpose_x_1"), val = bool(false)];
            bool var_3994_transpose_y_1 = const()[name = string("op_3994_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3994_cast_fp16 = matmul(transpose_x = var_3994_transpose_x_1, transpose_y = var_3994_transpose_y_1, x = stacked_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_3994_cast_fp16")];
            fp16 var_3995_to_fp16 = const()[name = string("op_3995_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3996_cast_fp16 = mul(x = var_3994_cast_fp16, y = var_3995_to_fp16)[name = string("op_3996_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_15_cast_fp16 = add(x = var_3996_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_15_cast_fp16")];
            int32 var_4000 = const()[name = string("op_4000"), val = int32(-1)];
            bool var_4001_interleave_0 = const()[name = string("op_4001_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4001_cast_fp16 = concat(axis = var_4000, interleave = var_4001_interleave_0, values = (cache_weights_15_cast_fp16, call_weights_15_cast_fp16))[name = string("op_4001_cast_fp16")];
            int32 var_4002 = const()[name = string("op_4002"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_15_cast_fp16 = softmax(axis = var_4002, x = var_4001_cast_fp16)[name = string("shares_15_cast_fp16")];
            tensor<int32, [4]> var_4009_begin_0 = const()[name = string("op_4009_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4009_end_0 = const()[name = string("op_4009_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4009_end_mask_0 = const()[name = string("op_4009_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4009_cast_fp16 = slice_by_index(begin = var_4009_begin_0, end = var_4009_end_0, end_mask = var_4009_end_mask_0, x = shares_15_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4014_begin_0 = const()[name = string("op_4014_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_4014_end_0 = const()[name = string("op_4014_end_0"), val = tensor<int32, [4]>([8, 8, 2064, 128])];
            tensor<bool, [4]> var_4014_end_mask_0 = const()[name = string("op_4014_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4014_cast_fp16 = slice_by_index(begin = var_4014_begin_0, end = var_4014_end_0, end_mask = var_4014_end_mask_0, x = read_state_1)[name = string("op_4014_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, 8, 2048, 128])];
            tensor<bool, [4]> var_4024_end_mask_0 = const()[name = string("op_4024_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 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, x = var_4014_cast_fp16)[name = string("op_4024_cast_fp16")];
            bool var_4030_transpose_x_0 = const()[name = string("op_4030_transpose_x_0"), val = bool(false)];
            bool var_4030_transpose_y_0 = const()[name = string("op_4030_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4030_cast_fp16 = matmul(transpose_x = var_4030_transpose_x_0, transpose_y = var_4030_transpose_y_0, x = var_4009_cast_fp16, y = var_4024_cast_fp16)[name = string("op_4030_cast_fp16")];
            tensor<int32, [4]> var_4035_begin_0 = const()[name = string("op_4035_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_4035_end_0 = const()[name = string("op_4035_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_4035_end_mask_0 = const()[name = string("op_4035_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4035_cast_fp16 = slice_by_index(begin = var_4035_begin_0, end = var_4035_end_0, end_mask = var_4035_end_mask_0, x = shares_15_cast_fp16)[name = string("op_4035_cast_fp16")];
            bool var_4036_transpose_x_0 = const()[name = string("op_4036_transpose_x_0"), val = bool(false)];
            bool var_4036_transpose_y_0 = const()[name = string("op_4036_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_3891_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 8, 32, 128]> var_4036_cast_fp16 = matmul(transpose_x = var_4036_transpose_x_0, transpose_y = var_4036_transpose_y_0, x = var_4035_cast_fp16, y = v_15_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_29_cast_fp16 = add(x = var_4030_cast_fp16, y = var_4036_cast_fp16)[name = string("mixed_29_cast_fp16")];
            tensor<int32, [4]> var_4053_begin_0 = const()[name = string("op_4053_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4053_end_0 = const()[name = string("op_4053_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4053_end_mask_0 = const()[name = string("op_4053_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4053_cast_fp16 = slice_by_index(begin = var_4053_begin_0, end = var_4053_end_0, end_mask = var_4053_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_4053_cast_fp16")];
            tensor<int32, [4]> var_4068_begin_0 = const()[name = string("op_4068_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4068_end_0 = const()[name = string("op_4068_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4068_end_mask_0 = const()[name = string("op_4068_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4068_cast_fp16 = slice_by_index(begin = var_4068_begin_0, end = var_4068_end_0, end_mask = var_4068_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_4068_cast_fp16")];
            int32 var_4070 = const()[name = string("op_4070"), 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_4070, interleave = mixed_31_interleave_0, values = (var_4053_cast_fp16, var_4068_cast_fp16))[name = string("mixed_31_cast_fp16")];
            tensor<int32, [4]> var_4074_perm_0 = const()[name = string("op_4074_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4078 = const()[name = string("op_4078"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4074_cast_fp16 = transpose(perm = var_4074_perm_0, x = mixed_31_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 16, 2048]> input_73_cast_fp16 = reshape(shape = var_4078, x = var_4074_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_157_cast_fp16 = add(x = x_141_cast_fp16, y = linear_52_cast_fp16)[name = string("x_157_cast_fp16")];
            bool var_4088 = const()[name = string("op_4088"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4091_cast_fp16 = abs(x = x_157_cast_fp16)[name = string("op_4091_cast_fp16")];
            tensor<int32, [1]> peak_63_axes_0 = const()[name = string("peak_63_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_63_cast_fp16 = reduce_max(axes = peak_63_axes_0, keep_dims = var_4088, x = var_4091_cast_fp16)[name = string("peak_63_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_159_cast_fp16 = real_div(x = x_157_cast_fp16, y = peak_63_cast_fp16)[name = string("x_159_cast_fp16")];
            fp16 var_4087_promoted_to_fp16 = const()[name = string("op_4087_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4095_cast_fp16 = pow(x = x_159_cast_fp16, y = var_4087_promoted_to_fp16)[name = string("op_4095_cast_fp16")];
            tensor<int32, [1]> var_4097_axes_0 = const()[name = string("op_4097_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4097_keep_dims_0 = const()[name = string("op_4097_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4097_cast_fp16 = reduce_mean(axes = var_4097_axes_0, keep_dims = var_4097_keep_dims_0, x = var_4095_cast_fp16)[name = string("op_4097_cast_fp16")];
            fp16 var_4087_promoted_1_to_fp16 = const()[name = string("op_4087_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4098_cast_fp16 = pow(x = peak_63_cast_fp16, y = var_4087_promoted_1_to_fp16)[name = string("op_4098_cast_fp16")];
            fp32 var_4099_epsilon_0 = const()[name = string("op_4099_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4099_cast_fp16 = inverse(epsilon = var_4099_epsilon_0, x = var_4098_cast_fp16)[name = string("op_4099_cast_fp16")];
            fp16 var_4100_to_fp16 = const()[name = string("op_4100_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4101_cast_fp16 = mul(x = var_4099_cast_fp16, y = var_4100_to_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4102_cast_fp16 = add(x = var_4097_cast_fp16, y = var_4101_cast_fp16)[name = string("op_4102_cast_fp16")];
            fp32 var_4103_epsilon_0 = const()[name = string("op_4103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4103_cast_fp16 = rsqrt(epsilon = var_4103_epsilon_0, x = var_4102_cast_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4104_cast_fp16 = mul(x = x_159_cast_fp16, y = var_4103_cast_fp16)[name = string("op_4104_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_4104_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_4109_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_4109_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_4109_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_161_cast_fp16 = add(x = x_157_cast_fp16, y = linear_55_cast_fp16)[name = string("x_161_cast_fp16")];
            bool var_4122 = const()[name = string("op_4122"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4125_cast_fp16 = abs(x = x_161_cast_fp16)[name = string("op_4125_cast_fp16")];
            tensor<int32, [1]> peak_65_axes_0 = const()[name = string("peak_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_65_cast_fp16 = reduce_max(axes = peak_65_axes_0, keep_dims = var_4122, x = var_4125_cast_fp16)[name = string("peak_65_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_163_cast_fp16 = real_div(x = x_161_cast_fp16, y = peak_65_cast_fp16)[name = string("x_163_cast_fp16")];
            fp16 var_4121_promoted_to_fp16 = const()[name = string("op_4121_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4129_cast_fp16 = pow(x = x_163_cast_fp16, y = var_4121_promoted_to_fp16)[name = string("op_4129_cast_fp16")];
            tensor<int32, [1]> var_4131_axes_0 = const()[name = string("op_4131_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4131_keep_dims_0 = const()[name = string("op_4131_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4131_cast_fp16 = reduce_mean(axes = var_4131_axes_0, keep_dims = var_4131_keep_dims_0, x = var_4129_cast_fp16)[name = string("op_4131_cast_fp16")];
            fp16 var_4121_promoted_1_to_fp16 = const()[name = string("op_4121_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4132_cast_fp16 = pow(x = peak_65_cast_fp16, y = var_4121_promoted_1_to_fp16)[name = string("op_4132_cast_fp16")];
            fp32 var_4133_epsilon_0 = const()[name = string("op_4133_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4133_cast_fp16 = inverse(epsilon = var_4133_epsilon_0, x = var_4132_cast_fp16)[name = string("op_4133_cast_fp16")];
            fp16 var_4134_to_fp16 = const()[name = string("op_4134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4135_cast_fp16 = mul(x = var_4133_cast_fp16, y = var_4134_to_fp16)[name = string("op_4135_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4136_cast_fp16 = add(x = var_4131_cast_fp16, y = var_4135_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp32 var_4137_epsilon_0 = const()[name = string("op_4137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4137_cast_fp16 = rsqrt(epsilon = var_4137_epsilon_0, x = var_4136_cast_fp16)[name = string("op_4137_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4138_cast_fp16 = mul(x = x_163_cast_fp16, y = var_4137_cast_fp16)[name = string("op_4138_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_4138_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_4147 = const()[name = string("op_4147"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4148_cast_fp16 = reshape(shape = var_4147, x = linear_56_cast_fp16)[name = string("op_4148_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4155 = const()[name = string("op_4155"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_4148_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 16, 128]> var_4158_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_4158_cast_fp16")];
            tensor<int32, [1]> peak_67_axes_0 = const()[name = string("peak_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_67_cast_fp16 = reduce_max(axes = peak_67_axes_0, keep_dims = var_4155, x = var_4158_cast_fp16)[name = string("peak_67_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_167_cast_fp16 = real_div(x = x_165_cast_fp16, y = peak_67_cast_fp16)[name = string("x_167_cast_fp16")];
            fp16 var_4154_promoted_to_fp16 = const()[name = string("op_4154_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4162_cast_fp16 = pow(x = x_167_cast_fp16, y = var_4154_promoted_to_fp16)[name = string("op_4162_cast_fp16")];
            tensor<int32, [1]> var_4164_axes_0 = const()[name = string("op_4164_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4164_keep_dims_0 = const()[name = string("op_4164_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4164_cast_fp16 = reduce_mean(axes = var_4164_axes_0, keep_dims = var_4164_keep_dims_0, x = var_4162_cast_fp16)[name = string("op_4164_cast_fp16")];
            fp16 var_4154_promoted_1_to_fp16 = const()[name = string("op_4154_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4165_cast_fp16 = pow(x = peak_67_cast_fp16, y = var_4154_promoted_1_to_fp16)[name = string("op_4165_cast_fp16")];
            fp32 var_4166_epsilon_0 = const()[name = string("op_4166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4166_cast_fp16 = inverse(epsilon = var_4166_epsilon_0, x = var_4165_cast_fp16)[name = string("op_4166_cast_fp16")];
            fp16 var_4167_to_fp16 = const()[name = string("op_4167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4168_cast_fp16 = mul(x = var_4166_cast_fp16, y = var_4167_to_fp16)[name = string("op_4168_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4169_cast_fp16 = add(x = var_4164_cast_fp16, y = var_4168_cast_fp16)[name = string("op_4169_cast_fp16")];
            fp32 var_4170_epsilon_0 = const()[name = string("op_4170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4170_cast_fp16 = rsqrt(epsilon = var_4170_epsilon_0, x = var_4169_cast_fp16)[name = string("op_4170_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4171_cast_fp16 = mul(x = x_167_cast_fp16, y = var_4170_cast_fp16)[name = string("op_4171_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_173_cast_fp16 = mul(x = var_4171_cast_fp16, y = model_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_173_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_4180 = const()[name = string("op_4180"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4181_cast_fp16 = reshape(shape = var_4180, x = linear_57_cast_fp16)[name = string("op_4181_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4188 = const()[name = string("op_4188"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_4181_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 8, 16, 128]> var_4191_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_4191_cast_fp16")];
            tensor<int32, [1]> peak_69_axes_0 = const()[name = string("peak_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_69_cast_fp16 = reduce_max(axes = peak_69_axes_0, keep_dims = var_4188, x = var_4191_cast_fp16)[name = string("peak_69_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_171_cast_fp16 = real_div(x = x_169_cast_fp16, y = peak_69_cast_fp16)[name = string("x_171_cast_fp16")];
            fp16 var_4187_promoted_to_fp16 = const()[name = string("op_4187_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4195_cast_fp16 = pow(x = x_171_cast_fp16, y = var_4187_promoted_to_fp16)[name = string("op_4195_cast_fp16")];
            tensor<int32, [1]> var_4197_axes_0 = const()[name = string("op_4197_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4197_keep_dims_0 = const()[name = string("op_4197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4197_cast_fp16 = reduce_mean(axes = var_4197_axes_0, keep_dims = var_4197_keep_dims_0, x = var_4195_cast_fp16)[name = string("op_4197_cast_fp16")];
            fp16 var_4187_promoted_1_to_fp16 = const()[name = string("op_4187_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4198_cast_fp16 = pow(x = peak_69_cast_fp16, y = var_4187_promoted_1_to_fp16)[name = string("op_4198_cast_fp16")];
            fp32 var_4199_epsilon_0 = const()[name = string("op_4199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4199_cast_fp16 = inverse(epsilon = var_4199_epsilon_0, x = var_4198_cast_fp16)[name = string("op_4199_cast_fp16")];
            fp16 var_4200_to_fp16 = const()[name = string("op_4200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4201_cast_fp16 = mul(x = var_4199_cast_fp16, y = var_4200_to_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4202_cast_fp16 = add(x = var_4197_cast_fp16, y = var_4201_cast_fp16)[name = string("op_4202_cast_fp16")];
            fp32 var_4203_epsilon_0 = const()[name = string("op_4203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4203_cast_fp16 = rsqrt(epsilon = var_4203_epsilon_0, x = var_4202_cast_fp16)[name = string("op_4203_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4204_cast_fp16 = mul(x = x_171_cast_fp16, y = var_4203_cast_fp16)[name = string("op_4204_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_175_cast_fp16 = mul(x = var_4204_cast_fp16, y = model_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_175_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_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_58_cast_fp16)[name = string("op_4214_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_173_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_173_cast_fp16)[name = string("second_33_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4228_cast_fp16 = mul(x = first_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4229_cast_fp16 = mul(x = second_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4229_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4231_cast_fp16 = sub(x = var_4228_cast_fp16, y = var_4229_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4232_cast_fp16 = mul(x = second_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4233_cast_fp16 = mul(x = first_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4233_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4235_cast_fp16 = add(x = var_4232_cast_fp16, y = var_4233_cast_fp16)[name = string("op_4235_cast_fp16")];
            int32 var_4237 = const()[name = string("op_4237"), 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_4237, interleave = q_17_interleave_0, values = (var_4231_cast_fp16, var_4235_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_175_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_175_cast_fp16)[name = string("second_35_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4249_cast_fp16 = mul(x = first_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4250_cast_fp16 = mul(x = second_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4252_cast_fp16 = sub(x = var_4249_cast_fp16, y = var_4250_cast_fp16)[name = string("op_4252_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4253_cast_fp16 = mul(x = second_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4254_cast_fp16 = mul(x = first_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4256_cast_fp16 = add(x = var_4253_cast_fp16, y = var_4254_cast_fp16)[name = string("op_4256_cast_fp16")];
            int32 var_4258 = const()[name = string("op_4258"), 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_4258, interleave = k_17_interleave_0, values = (var_4252_cast_fp16, var_4256_cast_fp16))[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_4269_begin_0 = const()[name = string("op_4269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4269_end_0 = const()[name = string("op_4269_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4269_end_mask_0 = const()[name = string("op_4269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4269_stride_0 = const()[name = string("op_4269_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4269_cast_fp16 = slice_by_index(begin = var_4269_begin_0, end = var_4269_end_0, end_mask = var_4269_end_mask_0, stride = var_4269_stride_0, x = q_17_cast_fp16)[name = string("op_4269_cast_fp16")];
            tensor<int32, [4]> var_4279_begin_0 = const()[name = string("op_4279_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4279_end_0 = const()[name = string("op_4279_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4279_end_mask_0 = const()[name = string("op_4279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4279_stride_0 = const()[name = string("op_4279_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4279_cast_fp16 = slice_by_index(begin = var_4279_begin_0, end = var_4279_end_0, end_mask = var_4279_end_mask_0, stride = var_4279_stride_0, x = q_17_cast_fp16)[name = string("op_4279_cast_fp16")];
            int32 var_4281 = const()[name = string("op_4281"), 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_4281, interleave = stacked_17_interleave_0, values = (var_4269_cast_fp16, var_4279_cast_fp16))[name = string("stacked_17_cast_fp16")];
            tensor<int32, [4]> var_4287_begin_0 = const()[name = string("op_4287_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4287_end_0 = const()[name = string("op_4287_end_0"), val = tensor<int32, [4]>([9, 8, 2064, 128])];
            tensor<bool, [4]> var_4287_end_mask_0 = const()[name = string("op_4287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4287_cast_fp16 = slice_by_index(begin = var_4287_begin_0, end = var_4287_end_0, end_mask = var_4287_end_mask_0, x = read_state_0)[name = string("op_4287_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = var_4287_cast_fp16)[name = string("op_4297_cast_fp16")];
            bool var_4306_transpose_x_1 = const()[name = string("op_4306_transpose_x_1"), val = bool(false)];
            bool var_4306_transpose_y_1 = const()[name = string("op_4306_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_4306_cast_fp16 = matmul(transpose_x = var_4306_transpose_x_1, transpose_y = var_4306_transpose_y_1, x = var_4297_cast_fp16, y = stacked_17_cast_fp16)[name = string("op_4306_cast_fp16")];
            tensor<int32, [4]> var_4309_perm_0 = const()[name = string("op_4309_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4310_to_fp16 = const()[name = string("op_4310_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_4309_cast_fp16 = transpose(perm = var_4309_perm_0, x = var_4306_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 32, 2048]> var_4311_cast_fp16 = mul(x = var_4309_cast_fp16, y = var_4310_to_fp16)[name = string("op_4311_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_17_cast_fp16 = add(x = var_4311_cast_fp16, y = cacheMask)[name = string("cache_weights_17_cast_fp16")];
            bool var_4317_transpose_x_1 = const()[name = string("op_4317_transpose_x_1"), val = bool(false)];
            bool var_4317_transpose_y_1 = const()[name = string("op_4317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4317_cast_fp16 = matmul(transpose_x = var_4317_transpose_x_1, transpose_y = var_4317_transpose_y_1, x = stacked_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_4317_cast_fp16")];
            fp16 var_4318_to_fp16 = const()[name = string("op_4318_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4319_cast_fp16 = mul(x = var_4317_cast_fp16, y = var_4318_to_fp16)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_17_cast_fp16 = add(x = var_4319_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_17_cast_fp16")];
            int32 var_4323 = const()[name = string("op_4323"), val = int32(-1)];
            bool var_4324_interleave_0 = const()[name = string("op_4324_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4324_cast_fp16 = concat(axis = var_4323, interleave = var_4324_interleave_0, values = (cache_weights_17_cast_fp16, call_weights_17_cast_fp16))[name = string("op_4324_cast_fp16")];
            int32 var_4325 = const()[name = string("op_4325"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_17_cast_fp16 = softmax(axis = var_4325, x = var_4324_cast_fp16)[name = string("shares_17_cast_fp16")];
            tensor<int32, [4]> var_4332_begin_0 = const()[name = string("op_4332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4332_end_0 = const()[name = string("op_4332_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4332_end_mask_0 = const()[name = string("op_4332_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4332_cast_fp16 = slice_by_index(begin = var_4332_begin_0, end = var_4332_end_0, end_mask = var_4332_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4332_cast_fp16")];
            tensor<int32, [4]> var_4337_begin_0 = const()[name = string("op_4337_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4337_end_0 = const()[name = string("op_4337_end_0"), val = tensor<int32, [4]>([9, 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_1)[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_4353_transpose_x_0 = const()[name = string("op_4353_transpose_x_0"), val = bool(false)];
            bool var_4353_transpose_y_0 = const()[name = string("op_4353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4353_cast_fp16 = matmul(transpose_x = var_4353_transpose_x_0, transpose_y = var_4353_transpose_y_0, x = var_4332_cast_fp16, y = var_4347_cast_fp16)[name = string("op_4353_cast_fp16")];
            tensor<int32, [4]> var_4358_begin_0 = const()[name = string("op_4358_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_4358_end_0 = const()[name = string("op_4358_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_4358_end_mask_0 = const()[name = string("op_4358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4358_cast_fp16 = slice_by_index(begin = var_4358_begin_0, end = var_4358_end_0, end_mask = var_4358_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4358_cast_fp16")];
            bool var_4359_transpose_x_0 = const()[name = string("op_4359_transpose_x_0"), val = bool(false)];
            bool var_4359_transpose_y_0 = const()[name = string("op_4359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_4214_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 32, 128]> var_4359_cast_fp16 = matmul(transpose_x = var_4359_transpose_x_0, transpose_y = var_4359_transpose_y_0, x = var_4358_cast_fp16, y = v_17_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_33_cast_fp16 = add(x = var_4353_cast_fp16, y = var_4359_cast_fp16)[name = string("mixed_33_cast_fp16")];
            tensor<int32, [4]> var_4376_begin_0 = const()[name = string("op_4376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4376_end_0 = const()[name = string("op_4376_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4376_end_mask_0 = const()[name = string("op_4376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4376_cast_fp16 = slice_by_index(begin = var_4376_begin_0, end = var_4376_end_0, end_mask = var_4376_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [4]> var_4391_begin_0 = const()[name = string("op_4391_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4391_end_0 = const()[name = string("op_4391_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4391_end_mask_0 = const()[name = string("op_4391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4391_cast_fp16 = slice_by_index(begin = var_4391_begin_0, end = var_4391_end_0, end_mask = var_4391_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4391_cast_fp16")];
            int32 var_4393 = const()[name = string("op_4393"), 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_4393, interleave = mixed_35_interleave_0, values = (var_4376_cast_fp16, var_4391_cast_fp16))[name = string("mixed_35_cast_fp16")];
            tensor<int32, [4]> var_4397_perm_0 = const()[name = string("op_4397_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4401 = const()[name = string("op_4401"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4397_cast_fp16 = transpose(perm = var_4397_perm_0, x = mixed_35_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 2048]> input_83_cast_fp16 = reshape(shape = var_4401, x = var_4397_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_177_cast_fp16 = add(x = x_161_cast_fp16, y = linear_59_cast_fp16)[name = string("x_177_cast_fp16")];
            bool var_4411 = const()[name = string("op_4411"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4414_cast_fp16 = abs(x = x_177_cast_fp16)[name = string("op_4414_cast_fp16")];
            tensor<int32, [1]> peak_71_axes_0 = const()[name = string("peak_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_71_cast_fp16 = reduce_max(axes = peak_71_axes_0, keep_dims = var_4411, x = var_4414_cast_fp16)[name = string("peak_71_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_179_cast_fp16 = real_div(x = x_177_cast_fp16, y = peak_71_cast_fp16)[name = string("x_179_cast_fp16")];
            fp16 var_4410_promoted_to_fp16 = const()[name = string("op_4410_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4418_cast_fp16 = pow(x = x_179_cast_fp16, y = var_4410_promoted_to_fp16)[name = string("op_4418_cast_fp16")];
            tensor<int32, [1]> var_4420_axes_0 = const()[name = string("op_4420_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4420_keep_dims_0 = const()[name = string("op_4420_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4420_cast_fp16 = reduce_mean(axes = var_4420_axes_0, keep_dims = var_4420_keep_dims_0, x = var_4418_cast_fp16)[name = string("op_4420_cast_fp16")];
            fp16 var_4410_promoted_1_to_fp16 = const()[name = string("op_4410_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4421_cast_fp16 = pow(x = peak_71_cast_fp16, y = var_4410_promoted_1_to_fp16)[name = string("op_4421_cast_fp16")];
            fp32 var_4422_epsilon_0 = const()[name = string("op_4422_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4422_cast_fp16 = inverse(epsilon = var_4422_epsilon_0, x = var_4421_cast_fp16)[name = string("op_4422_cast_fp16")];
            fp16 var_4423_to_fp16 = const()[name = string("op_4423_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4424_cast_fp16 = mul(x = var_4422_cast_fp16, y = var_4423_to_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4425_cast_fp16 = add(x = var_4420_cast_fp16, y = var_4424_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4426_cast_fp16 = rsqrt(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4427_cast_fp16 = mul(x = x_179_cast_fp16, y = var_4426_cast_fp16)[name = string("op_4427_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_4427_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_4432_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_4432_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_4432_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_181_cast_fp16 = add(x = x_177_cast_fp16, y = linear_62_cast_fp16)[name = string("x_181_cast_fp16")];
            bool var_4445 = const()[name = string("op_4445"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4448_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_4448_cast_fp16")];
            tensor<int32, [1]> peak_73_axes_0 = const()[name = string("peak_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_73_cast_fp16 = reduce_max(axes = peak_73_axes_0, keep_dims = var_4445, x = var_4448_cast_fp16)[name = string("peak_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_183_cast_fp16 = real_div(x = x_181_cast_fp16, y = peak_73_cast_fp16)[name = string("x_183_cast_fp16")];
            fp16 var_4444_promoted_to_fp16 = const()[name = string("op_4444_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4452_cast_fp16 = pow(x = x_183_cast_fp16, y = var_4444_promoted_to_fp16)[name = string("op_4452_cast_fp16")];
            tensor<int32, [1]> var_4454_axes_0 = const()[name = string("op_4454_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4454_keep_dims_0 = const()[name = string("op_4454_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4454_cast_fp16 = reduce_mean(axes = var_4454_axes_0, keep_dims = var_4454_keep_dims_0, x = var_4452_cast_fp16)[name = string("op_4454_cast_fp16")];
            fp16 var_4444_promoted_1_to_fp16 = const()[name = string("op_4444_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4455_cast_fp16 = pow(x = peak_73_cast_fp16, y = var_4444_promoted_1_to_fp16)[name = string("op_4455_cast_fp16")];
            fp32 var_4456_epsilon_0 = const()[name = string("op_4456_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4456_cast_fp16 = inverse(epsilon = var_4456_epsilon_0, x = var_4455_cast_fp16)[name = string("op_4456_cast_fp16")];
            fp16 var_4457_to_fp16 = const()[name = string("op_4457_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4458_cast_fp16 = mul(x = var_4456_cast_fp16, y = var_4457_to_fp16)[name = string("op_4458_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4459_cast_fp16 = add(x = var_4454_cast_fp16, y = var_4458_cast_fp16)[name = string("op_4459_cast_fp16")];
            fp32 var_4460_epsilon_0 = const()[name = string("op_4460_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4460_cast_fp16 = rsqrt(epsilon = var_4460_epsilon_0, x = var_4459_cast_fp16)[name = string("op_4460_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4461_cast_fp16 = mul(x = x_183_cast_fp16, y = var_4460_cast_fp16)[name = string("op_4461_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_4461_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_4470 = const()[name = string("op_4470"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4471_cast_fp16 = reshape(shape = var_4470, x = linear_63_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4478 = const()[name = string("op_4478"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_4471_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 16, 128]> var_4481_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_4481_cast_fp16")];
            tensor<int32, [1]> peak_75_axes_0 = const()[name = string("peak_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_75_cast_fp16 = reduce_max(axes = peak_75_axes_0, keep_dims = var_4478, x = var_4481_cast_fp16)[name = string("peak_75_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_187_cast_fp16 = real_div(x = x_185_cast_fp16, y = peak_75_cast_fp16)[name = string("x_187_cast_fp16")];
            fp16 var_4477_promoted_to_fp16 = const()[name = string("op_4477_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4485_cast_fp16 = pow(x = x_187_cast_fp16, y = var_4477_promoted_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<int32, [1]> var_4487_axes_0 = const()[name = string("op_4487_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4487_keep_dims_0 = const()[name = string("op_4487_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4487_cast_fp16 = reduce_mean(axes = var_4487_axes_0, keep_dims = var_4487_keep_dims_0, x = var_4485_cast_fp16)[name = string("op_4487_cast_fp16")];
            fp16 var_4477_promoted_1_to_fp16 = const()[name = string("op_4477_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4488_cast_fp16 = pow(x = peak_75_cast_fp16, y = var_4477_promoted_1_to_fp16)[name = string("op_4488_cast_fp16")];
            fp32 var_4489_epsilon_0 = const()[name = string("op_4489_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4489_cast_fp16 = inverse(epsilon = var_4489_epsilon_0, x = var_4488_cast_fp16)[name = string("op_4489_cast_fp16")];
            fp16 var_4490_to_fp16 = const()[name = string("op_4490_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4491_cast_fp16 = mul(x = var_4489_cast_fp16, y = var_4490_to_fp16)[name = string("op_4491_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4492_cast_fp16 = add(x = var_4487_cast_fp16, y = var_4491_cast_fp16)[name = string("op_4492_cast_fp16")];
            fp32 var_4493_epsilon_0 = const()[name = string("op_4493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4493_cast_fp16 = rsqrt(epsilon = var_4493_epsilon_0, x = var_4492_cast_fp16)[name = string("op_4493_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4494_cast_fp16 = mul(x = x_187_cast_fp16, y = var_4493_cast_fp16)[name = string("op_4494_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_193_cast_fp16 = mul(x = var_4494_cast_fp16, y = model_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_193_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_4503 = const()[name = string("op_4503"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4504_cast_fp16 = reshape(shape = var_4503, x = linear_64_cast_fp16)[name = string("op_4504_cast_fp16")];
            tensor<int32, [4]> x_189_perm_0 = const()[name = string("x_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4511 = const()[name = string("op_4511"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_189_cast_fp16 = transpose(perm = x_189_perm_0, x = var_4504_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 16, 128]> var_4514_cast_fp16 = abs(x = x_189_cast_fp16)[name = string("op_4514_cast_fp16")];
            tensor<int32, [1]> peak_77_axes_0 = const()[name = string("peak_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_77_cast_fp16 = reduce_max(axes = peak_77_axes_0, keep_dims = var_4511, x = var_4514_cast_fp16)[name = string("peak_77_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_191_cast_fp16 = real_div(x = x_189_cast_fp16, y = peak_77_cast_fp16)[name = string("x_191_cast_fp16")];
            fp16 var_4510_promoted_to_fp16 = const()[name = string("op_4510_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4518_cast_fp16 = pow(x = x_191_cast_fp16, y = var_4510_promoted_to_fp16)[name = string("op_4518_cast_fp16")];
            tensor<int32, [1]> var_4520_axes_0 = const()[name = string("op_4520_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4520_keep_dims_0 = const()[name = string("op_4520_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4520_cast_fp16 = reduce_mean(axes = var_4520_axes_0, keep_dims = var_4520_keep_dims_0, x = var_4518_cast_fp16)[name = string("op_4520_cast_fp16")];
            fp16 var_4510_promoted_1_to_fp16 = const()[name = string("op_4510_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4521_cast_fp16 = pow(x = peak_77_cast_fp16, y = var_4510_promoted_1_to_fp16)[name = string("op_4521_cast_fp16")];
            fp32 var_4522_epsilon_0 = const()[name = string("op_4522_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4522_cast_fp16 = inverse(epsilon = var_4522_epsilon_0, x = var_4521_cast_fp16)[name = string("op_4522_cast_fp16")];
            fp16 var_4523_to_fp16 = const()[name = string("op_4523_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4524_cast_fp16 = mul(x = var_4522_cast_fp16, y = var_4523_to_fp16)[name = string("op_4524_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4525_cast_fp16 = add(x = var_4520_cast_fp16, y = var_4524_cast_fp16)[name = string("op_4525_cast_fp16")];
            fp32 var_4526_epsilon_0 = const()[name = string("op_4526_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4526_cast_fp16 = rsqrt(epsilon = var_4526_epsilon_0, x = var_4525_cast_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4527_cast_fp16 = mul(x = x_191_cast_fp16, y = var_4526_cast_fp16)[name = string("op_4527_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_195_cast_fp16 = mul(x = var_4527_cast_fp16, y = model_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_195_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_4536 = const()[name = string("op_4536"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4537_cast_fp16 = reshape(shape = var_4536, x = linear_65_cast_fp16)[name = string("op_4537_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_193_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_193_cast_fp16)[name = string("second_37_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4551_cast_fp16 = mul(x = first_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4551_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4552_cast_fp16 = mul(x = second_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4552_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4554_cast_fp16 = sub(x = var_4551_cast_fp16, y = var_4552_cast_fp16)[name = string("op_4554_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4555_cast_fp16 = mul(x = second_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4555_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4556_cast_fp16 = mul(x = first_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4558_cast_fp16 = add(x = var_4555_cast_fp16, y = var_4556_cast_fp16)[name = string("op_4558_cast_fp16")];
            int32 var_4560 = const()[name = string("op_4560"), 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_4560, interleave = q_19_interleave_0, values = (var_4554_cast_fp16, var_4558_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_195_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_195_cast_fp16)[name = string("second_39_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4572_cast_fp16 = mul(x = first_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4572_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4573_cast_fp16 = mul(x = second_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4573_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4575_cast_fp16 = sub(x = var_4572_cast_fp16, y = var_4573_cast_fp16)[name = string("op_4575_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4576_cast_fp16 = mul(x = second_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4576_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4577_cast_fp16 = mul(x = first_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4579_cast_fp16 = add(x = var_4576_cast_fp16, y = var_4577_cast_fp16)[name = string("op_4579_cast_fp16")];
            int32 var_4581 = const()[name = string("op_4581"), 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_4581, interleave = k_19_interleave_0, values = (var_4575_cast_fp16, var_4579_cast_fp16))[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_4592_begin_0 = const()[name = string("op_4592_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4592_end_0 = const()[name = string("op_4592_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4592_end_mask_0 = const()[name = string("op_4592_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4592_stride_0 = const()[name = string("op_4592_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4592_cast_fp16 = slice_by_index(begin = var_4592_begin_0, end = var_4592_end_0, end_mask = var_4592_end_mask_0, stride = var_4592_stride_0, x = q_19_cast_fp16)[name = string("op_4592_cast_fp16")];
            tensor<int32, [4]> var_4602_begin_0 = const()[name = string("op_4602_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4602_end_0 = const()[name = string("op_4602_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4602_end_mask_0 = const()[name = string("op_4602_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4602_stride_0 = const()[name = string("op_4602_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4602_cast_fp16 = slice_by_index(begin = var_4602_begin_0, end = var_4602_end_0, end_mask = var_4602_end_mask_0, stride = var_4602_stride_0, x = q_19_cast_fp16)[name = string("op_4602_cast_fp16")];
            int32 var_4604 = const()[name = string("op_4604"), 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_4604, interleave = stacked_19_interleave_0, values = (var_4592_cast_fp16, var_4602_cast_fp16))[name = string("stacked_19_cast_fp16")];
            tensor<int32, [4]> var_4610_begin_0 = const()[name = string("op_4610_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4610_end_0 = const()[name = string("op_4610_end_0"), val = tensor<int32, [4]>([10, 8, 2064, 128])];
            tensor<bool, [4]> var_4610_end_mask_0 = const()[name = string("op_4610_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4610_cast_fp16 = slice_by_index(begin = var_4610_begin_0, end = var_4610_end_0, end_mask = var_4610_end_mask_0, x = read_state_0)[name = string("op_4610_cast_fp16")];
            tensor<int32, [4]> var_4620_begin_0 = const()[name = string("op_4620_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4620_end_0 = const()[name = string("op_4620_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4620_end_mask_0 = const()[name = string("op_4620_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4620_cast_fp16 = slice_by_index(begin = var_4620_begin_0, end = var_4620_end_0, end_mask = var_4620_end_mask_0, x = var_4610_cast_fp16)[name = string("op_4620_cast_fp16")];
            bool var_4629_transpose_x_1 = const()[name = string("op_4629_transpose_x_1"), val = bool(false)];
            bool var_4629_transpose_y_1 = const()[name = string("op_4629_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_4629_cast_fp16 = matmul(transpose_x = var_4629_transpose_x_1, transpose_y = var_4629_transpose_y_1, x = var_4620_cast_fp16, y = stacked_19_cast_fp16)[name = string("op_4629_cast_fp16")];
            tensor<int32, [4]> var_4632_perm_0 = const()[name = string("op_4632_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4633_to_fp16 = const()[name = string("op_4633_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_4632_cast_fp16 = transpose(perm = var_4632_perm_0, x = var_4629_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 8, 32, 2048]> var_4634_cast_fp16 = mul(x = var_4632_cast_fp16, y = var_4633_to_fp16)[name = string("op_4634_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_19_cast_fp16 = add(x = var_4634_cast_fp16, y = cacheMask)[name = string("cache_weights_19_cast_fp16")];
            bool var_4640_transpose_x_1 = const()[name = string("op_4640_transpose_x_1"), val = bool(false)];
            bool var_4640_transpose_y_1 = const()[name = string("op_4640_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4640_cast_fp16 = matmul(transpose_x = var_4640_transpose_x_1, transpose_y = var_4640_transpose_y_1, x = stacked_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_4640_cast_fp16")];
            fp16 var_4641_to_fp16 = const()[name = string("op_4641_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4642_cast_fp16 = mul(x = var_4640_cast_fp16, y = var_4641_to_fp16)[name = string("op_4642_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_19_cast_fp16 = add(x = var_4642_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_19_cast_fp16")];
            int32 var_4646 = const()[name = string("op_4646"), val = int32(-1)];
            bool var_4647_interleave_0 = const()[name = string("op_4647_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4647_cast_fp16 = concat(axis = var_4646, interleave = var_4647_interleave_0, values = (cache_weights_19_cast_fp16, call_weights_19_cast_fp16))[name = string("op_4647_cast_fp16")];
            int32 var_4648 = const()[name = string("op_4648"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_19_cast_fp16 = softmax(axis = var_4648, x = var_4647_cast_fp16)[name = string("shares_19_cast_fp16")];
            tensor<int32, [4]> var_4655_begin_0 = const()[name = string("op_4655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4655_end_0 = const()[name = string("op_4655_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4655_end_mask_0 = const()[name = string("op_4655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4655_cast_fp16 = slice_by_index(begin = var_4655_begin_0, end = var_4655_end_0, end_mask = var_4655_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4655_cast_fp16")];
            tensor<int32, [4]> var_4660_begin_0 = const()[name = string("op_4660_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4660_end_0 = const()[name = string("op_4660_end_0"), val = tensor<int32, [4]>([10, 8, 2064, 128])];
            tensor<bool, [4]> var_4660_end_mask_0 = const()[name = string("op_4660_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4660_cast_fp16 = slice_by_index(begin = var_4660_begin_0, end = var_4660_end_0, end_mask = var_4660_end_mask_0, x = read_state_1)[name = string("op_4660_cast_fp16")];
            tensor<int32, [4]> var_4670_begin_0 = const()[name = string("op_4670_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4670_end_0 = const()[name = string("op_4670_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4670_end_mask_0 = const()[name = string("op_4670_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4670_cast_fp16 = slice_by_index(begin = var_4670_begin_0, end = var_4670_end_0, end_mask = var_4670_end_mask_0, x = var_4660_cast_fp16)[name = string("op_4670_cast_fp16")];
            bool var_4676_transpose_x_0 = const()[name = string("op_4676_transpose_x_0"), val = bool(false)];
            bool var_4676_transpose_y_0 = const()[name = string("op_4676_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4676_cast_fp16 = matmul(transpose_x = var_4676_transpose_x_0, transpose_y = var_4676_transpose_y_0, x = var_4655_cast_fp16, y = var_4670_cast_fp16)[name = string("op_4676_cast_fp16")];
            tensor<int32, [4]> var_4681_begin_0 = const()[name = string("op_4681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_4681_end_0 = const()[name = string("op_4681_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_4681_end_mask_0 = const()[name = string("op_4681_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4681_cast_fp16 = slice_by_index(begin = var_4681_begin_0, end = var_4681_end_0, end_mask = var_4681_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4682_transpose_x_0 = const()[name = string("op_4682_transpose_x_0"), val = bool(false)];
            bool var_4682_transpose_y_0 = const()[name = string("op_4682_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_4537_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 8, 32, 128]> var_4682_cast_fp16 = matmul(transpose_x = var_4682_transpose_x_0, transpose_y = var_4682_transpose_y_0, x = var_4681_cast_fp16, y = v_19_cast_fp16)[name = string("op_4682_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_37_cast_fp16 = add(x = var_4676_cast_fp16, y = var_4682_cast_fp16)[name = string("mixed_37_cast_fp16")];
            tensor<int32, [4]> var_4699_begin_0 = const()[name = string("op_4699_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4699_end_0 = const()[name = string("op_4699_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4699_end_mask_0 = const()[name = string("op_4699_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4699_cast_fp16 = slice_by_index(begin = var_4699_begin_0, end = var_4699_end_0, end_mask = var_4699_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<int32, [4]> var_4714_begin_0 = const()[name = string("op_4714_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4714_end_0 = const()[name = string("op_4714_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4714_end_mask_0 = const()[name = string("op_4714_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4714_cast_fp16 = slice_by_index(begin = var_4714_begin_0, end = var_4714_end_0, end_mask = var_4714_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4714_cast_fp16")];
            int32 var_4716 = const()[name = string("op_4716"), 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_4716, interleave = mixed_39_interleave_0, values = (var_4699_cast_fp16, var_4714_cast_fp16))[name = string("mixed_39_cast_fp16")];
            tensor<int32, [4]> var_4720_perm_0 = const()[name = string("op_4720_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4724 = const()[name = string("op_4724"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4720_cast_fp16 = transpose(perm = var_4720_perm_0, x = mixed_39_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 2048]> input_93_cast_fp16 = reshape(shape = var_4724, x = var_4720_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_197_cast_fp16 = add(x = x_181_cast_fp16, y = linear_66_cast_fp16)[name = string("x_197_cast_fp16")];
            bool var_4734 = const()[name = string("op_4734"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4737_cast_fp16 = abs(x = x_197_cast_fp16)[name = string("op_4737_cast_fp16")];
            tensor<int32, [1]> peak_79_axes_0 = const()[name = string("peak_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_79_cast_fp16 = reduce_max(axes = peak_79_axes_0, keep_dims = var_4734, x = var_4737_cast_fp16)[name = string("peak_79_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_199_cast_fp16 = real_div(x = x_197_cast_fp16, y = peak_79_cast_fp16)[name = string("x_199_cast_fp16")];
            fp16 var_4733_promoted_to_fp16 = const()[name = string("op_4733_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4741_cast_fp16 = pow(x = x_199_cast_fp16, y = var_4733_promoted_to_fp16)[name = string("op_4741_cast_fp16")];
            tensor<int32, [1]> var_4743_axes_0 = const()[name = string("op_4743_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4743_keep_dims_0 = const()[name = string("op_4743_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4743_cast_fp16 = reduce_mean(axes = var_4743_axes_0, keep_dims = var_4743_keep_dims_0, x = var_4741_cast_fp16)[name = string("op_4743_cast_fp16")];
            fp16 var_4733_promoted_1_to_fp16 = const()[name = string("op_4733_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4744_cast_fp16 = pow(x = peak_79_cast_fp16, y = var_4733_promoted_1_to_fp16)[name = string("op_4744_cast_fp16")];
            fp32 var_4745_epsilon_0 = const()[name = string("op_4745_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4745_cast_fp16 = inverse(epsilon = var_4745_epsilon_0, x = var_4744_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4747_cast_fp16 = mul(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("op_4747_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4748_cast_fp16 = add(x = var_4743_cast_fp16, y = var_4747_cast_fp16)[name = string("op_4748_cast_fp16")];
            fp32 var_4749_epsilon_0 = const()[name = string("op_4749_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4749_cast_fp16 = rsqrt(epsilon = var_4749_epsilon_0, x = var_4748_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4750_cast_fp16 = mul(x = x_199_cast_fp16, y = var_4749_cast_fp16)[name = string("op_4750_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_4750_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_4755_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_4755_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_4755_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_201_cast_fp16 = add(x = x_197_cast_fp16, y = linear_69_cast_fp16)[name = string("x_201_cast_fp16")];
            bool var_4768 = const()[name = string("op_4768"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4771_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<int32, [1]> peak_81_axes_0 = const()[name = string("peak_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_81_cast_fp16 = reduce_max(axes = peak_81_axes_0, keep_dims = var_4768, x = var_4771_cast_fp16)[name = string("peak_81_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_203_cast_fp16 = real_div(x = x_201_cast_fp16, y = peak_81_cast_fp16)[name = string("x_203_cast_fp16")];
            fp16 var_4767_promoted_to_fp16 = const()[name = string("op_4767_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4775_cast_fp16 = pow(x = x_203_cast_fp16, y = var_4767_promoted_to_fp16)[name = string("op_4775_cast_fp16")];
            tensor<int32, [1]> var_4777_axes_0 = const()[name = string("op_4777_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4777_keep_dims_0 = const()[name = string("op_4777_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4777_cast_fp16 = reduce_mean(axes = var_4777_axes_0, keep_dims = var_4777_keep_dims_0, x = var_4775_cast_fp16)[name = string("op_4777_cast_fp16")];
            fp16 var_4767_promoted_1_to_fp16 = const()[name = string("op_4767_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4778_cast_fp16 = pow(x = peak_81_cast_fp16, y = var_4767_promoted_1_to_fp16)[name = string("op_4778_cast_fp16")];
            fp32 var_4779_epsilon_0 = const()[name = string("op_4779_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4779_cast_fp16 = inverse(epsilon = var_4779_epsilon_0, x = var_4778_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 var_4780_to_fp16 = const()[name = string("op_4780_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16, y = var_4780_to_fp16)[name = string("op_4781_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4782_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4781_cast_fp16)[name = string("op_4782_cast_fp16")];
            fp32 var_4783_epsilon_0 = const()[name = string("op_4783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4783_cast_fp16 = rsqrt(epsilon = var_4783_epsilon_0, x = var_4782_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4784_cast_fp16 = mul(x = x_203_cast_fp16, y = var_4783_cast_fp16)[name = string("op_4784_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_4784_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_4793 = const()[name = string("op_4793"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4794_cast_fp16 = reshape(shape = var_4793, x = linear_70_cast_fp16)[name = string("op_4794_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4801 = const()[name = string("op_4801"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_4794_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 16, 128]> var_4804_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_4804_cast_fp16")];
            tensor<int32, [1]> peak_83_axes_0 = const()[name = string("peak_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_83_cast_fp16 = reduce_max(axes = peak_83_axes_0, keep_dims = var_4801, x = var_4804_cast_fp16)[name = string("peak_83_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_207_cast_fp16 = real_div(x = x_205_cast_fp16, y = peak_83_cast_fp16)[name = string("x_207_cast_fp16")];
            fp16 var_4800_promoted_to_fp16 = const()[name = string("op_4800_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4808_cast_fp16 = pow(x = x_207_cast_fp16, y = var_4800_promoted_to_fp16)[name = string("op_4808_cast_fp16")];
            tensor<int32, [1]> var_4810_axes_0 = const()[name = string("op_4810_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4810_keep_dims_0 = const()[name = string("op_4810_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4810_cast_fp16 = reduce_mean(axes = var_4810_axes_0, keep_dims = var_4810_keep_dims_0, x = var_4808_cast_fp16)[name = string("op_4810_cast_fp16")];
            fp16 var_4800_promoted_1_to_fp16 = const()[name = string("op_4800_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4811_cast_fp16 = pow(x = peak_83_cast_fp16, y = var_4800_promoted_1_to_fp16)[name = string("op_4811_cast_fp16")];
            fp32 var_4812_epsilon_0 = const()[name = string("op_4812_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4812_cast_fp16 = inverse(epsilon = var_4812_epsilon_0, x = var_4811_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp16 var_4813_to_fp16 = const()[name = string("op_4813_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4814_cast_fp16 = mul(x = var_4812_cast_fp16, y = var_4813_to_fp16)[name = string("op_4814_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4815_cast_fp16 = add(x = var_4810_cast_fp16, y = var_4814_cast_fp16)[name = string("op_4815_cast_fp16")];
            fp32 var_4816_epsilon_0 = const()[name = string("op_4816_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4816_cast_fp16 = rsqrt(epsilon = var_4816_epsilon_0, x = var_4815_cast_fp16)[name = string("op_4816_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4817_cast_fp16 = mul(x = x_207_cast_fp16, y = var_4816_cast_fp16)[name = string("op_4817_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_213_cast_fp16 = mul(x = var_4817_cast_fp16, y = model_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_213_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_4826 = const()[name = string("op_4826"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4827_cast_fp16 = reshape(shape = var_4826, x = linear_71_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [4]> x_209_perm_0 = const()[name = string("x_209_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4834 = const()[name = string("op_4834"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_209_cast_fp16 = transpose(perm = x_209_perm_0, x = var_4827_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 8, 16, 128]> var_4837_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_4837_cast_fp16")];
            tensor<int32, [1]> peak_85_axes_0 = const()[name = string("peak_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_85_cast_fp16 = reduce_max(axes = peak_85_axes_0, keep_dims = var_4834, x = var_4837_cast_fp16)[name = string("peak_85_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_211_cast_fp16 = real_div(x = x_209_cast_fp16, y = peak_85_cast_fp16)[name = string("x_211_cast_fp16")];
            fp16 var_4833_promoted_to_fp16 = const()[name = string("op_4833_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4841_cast_fp16 = pow(x = x_211_cast_fp16, y = var_4833_promoted_to_fp16)[name = string("op_4841_cast_fp16")];
            tensor<int32, [1]> var_4843_axes_0 = const()[name = string("op_4843_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4843_keep_dims_0 = const()[name = string("op_4843_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4843_cast_fp16 = reduce_mean(axes = var_4843_axes_0, keep_dims = var_4843_keep_dims_0, x = var_4841_cast_fp16)[name = string("op_4843_cast_fp16")];
            fp16 var_4833_promoted_1_to_fp16 = const()[name = string("op_4833_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4844_cast_fp16 = pow(x = peak_85_cast_fp16, y = var_4833_promoted_1_to_fp16)[name = string("op_4844_cast_fp16")];
            fp32 var_4845_epsilon_0 = const()[name = string("op_4845_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4845_cast_fp16 = inverse(epsilon = var_4845_epsilon_0, x = var_4844_cast_fp16)[name = string("op_4845_cast_fp16")];
            fp16 var_4846_to_fp16 = const()[name = string("op_4846_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4847_cast_fp16 = mul(x = var_4845_cast_fp16, y = var_4846_to_fp16)[name = string("op_4847_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4848_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4847_cast_fp16)[name = string("op_4848_cast_fp16")];
            fp32 var_4849_epsilon_0 = const()[name = string("op_4849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4849_cast_fp16 = rsqrt(epsilon = var_4849_epsilon_0, x = var_4848_cast_fp16)[name = string("op_4849_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4850_cast_fp16 = mul(x = x_211_cast_fp16, y = var_4849_cast_fp16)[name = string("op_4850_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_215_cast_fp16 = mul(x = var_4850_cast_fp16, y = model_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_215_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_4859 = const()[name = string("op_4859"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4860_cast_fp16 = reshape(shape = var_4859, x = linear_72_cast_fp16)[name = string("op_4860_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_213_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_213_cast_fp16)[name = string("second_41_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4874_cast_fp16 = mul(x = first_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4875_cast_fp16 = mul(x = second_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4875_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4877_cast_fp16 = sub(x = var_4874_cast_fp16, y = var_4875_cast_fp16)[name = string("op_4877_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4878_cast_fp16 = mul(x = second_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4878_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4879_cast_fp16 = mul(x = first_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4879_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4881_cast_fp16 = add(x = var_4878_cast_fp16, y = var_4879_cast_fp16)[name = string("op_4881_cast_fp16")];
            int32 var_4883 = const()[name = string("op_4883"), 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_4883, interleave = q_21_interleave_0, values = (var_4877_cast_fp16, var_4881_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_215_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_215_cast_fp16)[name = string("second_43_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4895_cast_fp16 = mul(x = first_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4896_cast_fp16 = mul(x = second_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4896_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4898_cast_fp16 = sub(x = var_4895_cast_fp16, y = var_4896_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4899_cast_fp16 = mul(x = second_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4899_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4900_cast_fp16 = mul(x = first_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4900_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4902_cast_fp16 = add(x = var_4899_cast_fp16, y = var_4900_cast_fp16)[name = string("op_4902_cast_fp16")];
            int32 var_4904 = const()[name = string("op_4904"), 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_4904, interleave = k_21_interleave_0, values = (var_4898_cast_fp16, var_4902_cast_fp16))[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_4915_begin_0 = const()[name = string("op_4915_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4915_end_0 = const()[name = string("op_4915_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4915_end_mask_0 = const()[name = string("op_4915_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4915_stride_0 = const()[name = string("op_4915_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4915_cast_fp16 = slice_by_index(begin = var_4915_begin_0, end = var_4915_end_0, end_mask = var_4915_end_mask_0, stride = var_4915_stride_0, x = q_21_cast_fp16)[name = string("op_4915_cast_fp16")];
            tensor<int32, [4]> var_4925_begin_0 = const()[name = string("op_4925_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4925_end_0 = const()[name = string("op_4925_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4925_end_mask_0 = const()[name = string("op_4925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4925_stride_0 = const()[name = string("op_4925_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4925_cast_fp16 = slice_by_index(begin = var_4925_begin_0, end = var_4925_end_0, end_mask = var_4925_end_mask_0, stride = var_4925_stride_0, x = q_21_cast_fp16)[name = string("op_4925_cast_fp16")];
            int32 var_4927 = const()[name = string("op_4927"), 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_4927, interleave = stacked_21_interleave_0, values = (var_4915_cast_fp16, var_4925_cast_fp16))[name = string("stacked_21_cast_fp16")];
            tensor<int32, [4]> var_4933_begin_0 = const()[name = string("op_4933_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4933_end_0 = const()[name = string("op_4933_end_0"), val = tensor<int32, [4]>([11, 8, 2064, 128])];
            tensor<bool, [4]> var_4933_end_mask_0 = const()[name = string("op_4933_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4933_cast_fp16 = slice_by_index(begin = var_4933_begin_0, end = var_4933_end_0, end_mask = var_4933_end_mask_0, x = read_state_0)[name = string("op_4933_cast_fp16")];
            tensor<int32, [4]> var_4943_begin_0 = const()[name = string("op_4943_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4943_end_0 = const()[name = string("op_4943_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4943_end_mask_0 = const()[name = string("op_4943_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4943_cast_fp16 = slice_by_index(begin = var_4943_begin_0, end = var_4943_end_0, end_mask = var_4943_end_mask_0, x = var_4933_cast_fp16)[name = string("op_4943_cast_fp16")];
            bool var_4952_transpose_x_1 = const()[name = string("op_4952_transpose_x_1"), val = bool(false)];
            bool var_4952_transpose_y_1 = const()[name = string("op_4952_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_4952_cast_fp16 = matmul(transpose_x = var_4952_transpose_x_1, transpose_y = var_4952_transpose_y_1, x = var_4943_cast_fp16, y = stacked_21_cast_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4955_perm_0 = const()[name = string("op_4955_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4956_to_fp16 = const()[name = string("op_4956_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_4955_cast_fp16 = transpose(perm = var_4955_perm_0, x = var_4952_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 32, 2048]> var_4957_cast_fp16 = mul(x = var_4955_cast_fp16, y = var_4956_to_fp16)[name = string("op_4957_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_21_cast_fp16 = add(x = var_4957_cast_fp16, y = cacheMask)[name = string("cache_weights_21_cast_fp16")];
            bool var_4963_transpose_x_1 = const()[name = string("op_4963_transpose_x_1"), val = bool(false)];
            bool var_4963_transpose_y_1 = const()[name = string("op_4963_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4963_cast_fp16 = matmul(transpose_x = var_4963_transpose_x_1, transpose_y = var_4963_transpose_y_1, x = stacked_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_4963_cast_fp16")];
            fp16 var_4964_to_fp16 = const()[name = string("op_4964_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4965_cast_fp16 = mul(x = var_4963_cast_fp16, y = var_4964_to_fp16)[name = string("op_4965_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_21_cast_fp16 = add(x = var_4965_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_21_cast_fp16")];
            int32 var_4969 = const()[name = string("op_4969"), val = int32(-1)];
            bool var_4970_interleave_0 = const()[name = string("op_4970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4970_cast_fp16 = concat(axis = var_4969, interleave = var_4970_interleave_0, values = (cache_weights_21_cast_fp16, call_weights_21_cast_fp16))[name = string("op_4970_cast_fp16")];
            int32 var_4971 = const()[name = string("op_4971"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_21_cast_fp16 = softmax(axis = var_4971, x = var_4970_cast_fp16)[name = string("shares_21_cast_fp16")];
            tensor<int32, [4]> var_4978_begin_0 = const()[name = string("op_4978_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4978_end_0 = const()[name = string("op_4978_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4978_end_mask_0 = const()[name = string("op_4978_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4978_cast_fp16 = slice_by_index(begin = var_4978_begin_0, end = var_4978_end_0, end_mask = var_4978_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4978_cast_fp16")];
            tensor<int32, [4]> var_4983_begin_0 = const()[name = string("op_4983_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4983_end_0 = const()[name = string("op_4983_end_0"), val = tensor<int32, [4]>([11, 8, 2064, 128])];
            tensor<bool, [4]> var_4983_end_mask_0 = const()[name = string("op_4983_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4983_cast_fp16 = slice_by_index(begin = var_4983_begin_0, end = var_4983_end_0, end_mask = var_4983_end_mask_0, x = read_state_1)[name = string("op_4983_cast_fp16")];
            tensor<int32, [4]> var_4993_begin_0 = const()[name = string("op_4993_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4993_end_0 = const()[name = string("op_4993_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4993_end_mask_0 = const()[name = string("op_4993_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4993_cast_fp16 = slice_by_index(begin = var_4993_begin_0, end = var_4993_end_0, end_mask = var_4993_end_mask_0, x = var_4983_cast_fp16)[name = string("op_4993_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, 32, 128]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = var_4978_cast_fp16, y = var_4993_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5004_begin_0 = const()[name = string("op_5004_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_5004_end_0 = const()[name = string("op_5004_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_5004_end_mask_0 = const()[name = string("op_5004_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5004_cast_fp16 = slice_by_index(begin = var_5004_begin_0, end = var_5004_end_0, end_mask = var_5004_end_mask_0, x = shares_21_cast_fp16)[name = string("op_5004_cast_fp16")];
            bool var_5005_transpose_x_0 = const()[name = string("op_5005_transpose_x_0"), val = bool(false)];
            bool var_5005_transpose_y_0 = const()[name = string("op_5005_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_4860_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 32, 128]> var_5005_cast_fp16 = matmul(transpose_x = var_5005_transpose_x_0, transpose_y = var_5005_transpose_y_0, x = var_5004_cast_fp16, y = v_21_cast_fp16)[name = string("op_5005_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_41_cast_fp16 = add(x = var_4999_cast_fp16, y = var_5005_cast_fp16)[name = string("mixed_41_cast_fp16")];
            tensor<int32, [4]> var_5022_begin_0 = const()[name = string("op_5022_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5022_end_0 = const()[name = string("op_5022_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5022_end_mask_0 = const()[name = string("op_5022_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5022_cast_fp16 = slice_by_index(begin = var_5022_begin_0, end = var_5022_end_0, end_mask = var_5022_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_5022_cast_fp16")];
            tensor<int32, [4]> var_5037_begin_0 = const()[name = string("op_5037_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5037_end_0 = const()[name = string("op_5037_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5037_end_mask_0 = const()[name = string("op_5037_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5037_cast_fp16 = slice_by_index(begin = var_5037_begin_0, end = var_5037_end_0, end_mask = var_5037_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_5037_cast_fp16")];
            int32 var_5039 = const()[name = string("op_5039"), 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_5039, interleave = mixed_43_interleave_0, values = (var_5022_cast_fp16, var_5037_cast_fp16))[name = string("mixed_43_cast_fp16")];
            tensor<int32, [4]> var_5043_perm_0 = const()[name = string("op_5043_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5047 = const()[name = string("op_5047"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5043_cast_fp16 = transpose(perm = var_5043_perm_0, x = mixed_43_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 2048]> input_103_cast_fp16 = reshape(shape = var_5047, x = var_5043_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_217_cast_fp16 = add(x = x_201_cast_fp16, y = linear_73_cast_fp16)[name = string("x_217_cast_fp16")];
            bool var_5057 = const()[name = string("op_5057"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5060_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<int32, [1]> peak_87_axes_0 = const()[name = string("peak_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_87_cast_fp16 = reduce_max(axes = peak_87_axes_0, keep_dims = var_5057, x = var_5060_cast_fp16)[name = string("peak_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_219_cast_fp16 = real_div(x = x_217_cast_fp16, y = peak_87_cast_fp16)[name = string("x_219_cast_fp16")];
            fp16 var_5056_promoted_to_fp16 = const()[name = string("op_5056_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5064_cast_fp16 = pow(x = x_219_cast_fp16, y = var_5056_promoted_to_fp16)[name = string("op_5064_cast_fp16")];
            tensor<int32, [1]> var_5066_axes_0 = const()[name = string("op_5066_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5066_keep_dims_0 = const()[name = string("op_5066_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5066_cast_fp16 = reduce_mean(axes = var_5066_axes_0, keep_dims = var_5066_keep_dims_0, x = var_5064_cast_fp16)[name = string("op_5066_cast_fp16")];
            fp16 var_5056_promoted_1_to_fp16 = const()[name = string("op_5056_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5067_cast_fp16 = pow(x = peak_87_cast_fp16, y = var_5056_promoted_1_to_fp16)[name = string("op_5067_cast_fp16")];
            fp32 var_5068_epsilon_0 = const()[name = string("op_5068_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5068_cast_fp16 = inverse(epsilon = var_5068_epsilon_0, x = var_5067_cast_fp16)[name = string("op_5068_cast_fp16")];
            fp16 var_5069_to_fp16 = const()[name = string("op_5069_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5070_cast_fp16 = mul(x = var_5068_cast_fp16, y = var_5069_to_fp16)[name = string("op_5070_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5071_cast_fp16 = add(x = var_5066_cast_fp16, y = var_5070_cast_fp16)[name = string("op_5071_cast_fp16")];
            fp32 var_5072_epsilon_0 = const()[name = string("op_5072_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5072_cast_fp16 = rsqrt(epsilon = var_5072_epsilon_0, x = var_5071_cast_fp16)[name = string("op_5072_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5073_cast_fp16 = mul(x = x_219_cast_fp16, y = var_5072_cast_fp16)[name = string("op_5073_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_5073_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_5078_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_5078_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_5078_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_221_cast_fp16 = add(x = x_217_cast_fp16, y = linear_76_cast_fp16)[name = string("x_221_cast_fp16")];
            bool var_5091 = const()[name = string("op_5091"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5094_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [1]> peak_89_axes_0 = const()[name = string("peak_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_89_cast_fp16 = reduce_max(axes = peak_89_axes_0, keep_dims = var_5091, x = var_5094_cast_fp16)[name = string("peak_89_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_223_cast_fp16 = real_div(x = x_221_cast_fp16, y = peak_89_cast_fp16)[name = string("x_223_cast_fp16")];
            fp16 var_5090_promoted_to_fp16 = const()[name = string("op_5090_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5098_cast_fp16 = pow(x = x_223_cast_fp16, y = var_5090_promoted_to_fp16)[name = string("op_5098_cast_fp16")];
            tensor<int32, [1]> var_5100_axes_0 = const()[name = string("op_5100_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5100_keep_dims_0 = const()[name = string("op_5100_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5100_cast_fp16 = reduce_mean(axes = var_5100_axes_0, keep_dims = var_5100_keep_dims_0, x = var_5098_cast_fp16)[name = string("op_5100_cast_fp16")];
            fp16 var_5090_promoted_1_to_fp16 = const()[name = string("op_5090_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5101_cast_fp16 = pow(x = peak_89_cast_fp16, y = var_5090_promoted_1_to_fp16)[name = string("op_5101_cast_fp16")];
            fp32 var_5102_epsilon_0 = const()[name = string("op_5102_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5102_cast_fp16 = inverse(epsilon = var_5102_epsilon_0, x = var_5101_cast_fp16)[name = string("op_5102_cast_fp16")];
            fp16 var_5103_to_fp16 = const()[name = string("op_5103_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5104_cast_fp16 = mul(x = var_5102_cast_fp16, y = var_5103_to_fp16)[name = string("op_5104_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5105_cast_fp16 = add(x = var_5100_cast_fp16, y = var_5104_cast_fp16)[name = string("op_5105_cast_fp16")];
            fp32 var_5106_epsilon_0 = const()[name = string("op_5106_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5106_cast_fp16 = rsqrt(epsilon = var_5106_epsilon_0, x = var_5105_cast_fp16)[name = string("op_5106_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5107_cast_fp16 = mul(x = x_223_cast_fp16, y = var_5106_cast_fp16)[name = string("op_5107_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_5107_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_5116 = const()[name = string("op_5116"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5117_cast_fp16 = reshape(shape = var_5116, x = linear_77_cast_fp16)[name = string("op_5117_cast_fp16")];
            tensor<int32, [4]> x_225_perm_0 = const()[name = string("x_225_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5124 = const()[name = string("op_5124"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_225_cast_fp16 = transpose(perm = x_225_perm_0, x = var_5117_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 16, 16, 128]> var_5127_cast_fp16 = abs(x = x_225_cast_fp16)[name = string("op_5127_cast_fp16")];
            tensor<int32, [1]> peak_91_axes_0 = const()[name = string("peak_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_91_cast_fp16 = reduce_max(axes = peak_91_axes_0, keep_dims = var_5124, x = var_5127_cast_fp16)[name = string("peak_91_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_227_cast_fp16 = real_div(x = x_225_cast_fp16, y = peak_91_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_5123_promoted_to_fp16 = const()[name = string("op_5123_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5131_cast_fp16 = pow(x = x_227_cast_fp16, y = var_5123_promoted_to_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5133_axes_0 = const()[name = string("op_5133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5133_keep_dims_0 = const()[name = string("op_5133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5133_cast_fp16 = reduce_mean(axes = var_5133_axes_0, keep_dims = var_5133_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5133_cast_fp16")];
            fp16 var_5123_promoted_1_to_fp16 = const()[name = string("op_5123_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5134_cast_fp16 = pow(x = peak_91_cast_fp16, y = var_5123_promoted_1_to_fp16)[name = string("op_5134_cast_fp16")];
            fp32 var_5135_epsilon_0 = const()[name = string("op_5135_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5135_cast_fp16 = inverse(epsilon = var_5135_epsilon_0, x = var_5134_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, 16, 16, 1]> var_5137_cast_fp16 = mul(x = var_5135_cast_fp16, y = var_5136_to_fp16)[name = string("op_5137_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5138_cast_fp16 = add(x = var_5133_cast_fp16, y = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp32 var_5139_epsilon_0 = const()[name = string("op_5139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5139_cast_fp16 = rsqrt(epsilon = var_5139_epsilon_0, x = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5140_cast_fp16 = mul(x = x_227_cast_fp16, y = var_5139_cast_fp16)[name = string("op_5140_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_233_cast_fp16 = mul(x = var_5140_cast_fp16, y = model_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_233_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_5149 = const()[name = string("op_5149"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5150_cast_fp16 = reshape(shape = var_5149, x = linear_78_cast_fp16)[name = string("op_5150_cast_fp16")];
            tensor<int32, [4]> x_229_perm_0 = const()[name = string("x_229_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5157 = const()[name = string("op_5157"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_229_cast_fp16 = transpose(perm = x_229_perm_0, x = var_5150_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 8, 16, 128]> var_5160_cast_fp16 = abs(x = x_229_cast_fp16)[name = string("op_5160_cast_fp16")];
            tensor<int32, [1]> peak_93_axes_0 = const()[name = string("peak_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_93_cast_fp16 = reduce_max(axes = peak_93_axes_0, keep_dims = var_5157, x = var_5160_cast_fp16)[name = string("peak_93_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_231_cast_fp16 = real_div(x = x_229_cast_fp16, y = peak_93_cast_fp16)[name = string("x_231_cast_fp16")];
            fp16 var_5156_promoted_to_fp16 = const()[name = string("op_5156_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5164_cast_fp16 = pow(x = x_231_cast_fp16, y = var_5156_promoted_to_fp16)[name = string("op_5164_cast_fp16")];
            tensor<int32, [1]> var_5166_axes_0 = const()[name = string("op_5166_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5166_keep_dims_0 = const()[name = string("op_5166_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5166_cast_fp16 = reduce_mean(axes = var_5166_axes_0, keep_dims = var_5166_keep_dims_0, x = var_5164_cast_fp16)[name = string("op_5166_cast_fp16")];
            fp16 var_5156_promoted_1_to_fp16 = const()[name = string("op_5156_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5167_cast_fp16 = pow(x = peak_93_cast_fp16, y = var_5156_promoted_1_to_fp16)[name = string("op_5167_cast_fp16")];
            fp32 var_5168_epsilon_0 = const()[name = string("op_5168_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5168_cast_fp16 = inverse(epsilon = var_5168_epsilon_0, x = var_5167_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp16 var_5169_to_fp16 = const()[name = string("op_5169_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5170_cast_fp16 = mul(x = var_5168_cast_fp16, y = var_5169_to_fp16)[name = string("op_5170_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5171_cast_fp16 = add(x = var_5166_cast_fp16, y = var_5170_cast_fp16)[name = string("op_5171_cast_fp16")];
            fp32 var_5172_epsilon_0 = const()[name = string("op_5172_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5172_cast_fp16 = rsqrt(epsilon = var_5172_epsilon_0, x = var_5171_cast_fp16)[name = string("op_5172_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5173_cast_fp16 = mul(x = x_231_cast_fp16, y = var_5172_cast_fp16)[name = string("op_5173_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_235_cast_fp16 = mul(x = var_5173_cast_fp16, y = model_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_235_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_5182 = const()[name = string("op_5182"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5183_cast_fp16 = reshape(shape = var_5182, x = linear_79_cast_fp16)[name = string("op_5183_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_233_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_233_cast_fp16)[name = string("second_45_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5197_cast_fp16 = mul(x = first_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5197_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5198_cast_fp16 = mul(x = second_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5198_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5200_cast_fp16 = sub(x = var_5197_cast_fp16, y = var_5198_cast_fp16)[name = string("op_5200_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5201_cast_fp16 = mul(x = second_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5201_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5202_cast_fp16 = mul(x = first_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5202_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5204_cast_fp16 = add(x = var_5201_cast_fp16, y = var_5202_cast_fp16)[name = string("op_5204_cast_fp16")];
            int32 var_5206 = const()[name = string("op_5206"), 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_5206, interleave = q_23_interleave_0, values = (var_5200_cast_fp16, var_5204_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_235_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_235_cast_fp16)[name = string("second_47_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5218_cast_fp16 = mul(x = first_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5218_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5219_cast_fp16 = mul(x = second_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5219_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5221_cast_fp16 = sub(x = var_5218_cast_fp16, y = var_5219_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5222_cast_fp16 = mul(x = second_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5222_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5223_cast_fp16 = mul(x = first_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5225_cast_fp16 = add(x = var_5222_cast_fp16, y = var_5223_cast_fp16)[name = string("op_5225_cast_fp16")];
            int32 var_5227 = const()[name = string("op_5227"), 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_5227, interleave = k_23_interleave_0, values = (var_5221_cast_fp16, var_5225_cast_fp16))[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> var_5238_begin_0 = const()[name = string("op_5238_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5238_end_0 = const()[name = string("op_5238_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5238_end_mask_0 = const()[name = string("op_5238_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5238_stride_0 = const()[name = string("op_5238_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, stride = var_5238_stride_0, x = q_23_cast_fp16)[name = string("op_5238_cast_fp16")];
            tensor<int32, [4]> var_5248_begin_0 = const()[name = string("op_5248_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5248_end_0 = const()[name = string("op_5248_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5248_end_mask_0 = const()[name = string("op_5248_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5248_stride_0 = const()[name = string("op_5248_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5248_cast_fp16 = slice_by_index(begin = var_5248_begin_0, end = var_5248_end_0, end_mask = var_5248_end_mask_0, stride = var_5248_stride_0, x = q_23_cast_fp16)[name = string("op_5248_cast_fp16")];
            int32 var_5250 = const()[name = string("op_5250"), 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_5250, interleave = stacked_23_interleave_0, values = (var_5238_cast_fp16, var_5248_cast_fp16))[name = string("stacked_23_cast_fp16")];
            tensor<int32, [4]> var_5256_begin_0 = const()[name = string("op_5256_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_5256_end_0 = const()[name = string("op_5256_end_0"), val = tensor<int32, [4]>([12, 8, 2064, 128])];
            tensor<bool, [4]> var_5256_end_mask_0 = const()[name = string("op_5256_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5256_cast_fp16 = slice_by_index(begin = var_5256_begin_0, end = var_5256_end_0, end_mask = var_5256_end_mask_0, x = read_state_0)[name = string("op_5256_cast_fp16")];
            tensor<int32, [4]> var_5266_begin_0 = const()[name = string("op_5266_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5266_end_0 = const()[name = string("op_5266_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5266_end_mask_0 = const()[name = string("op_5266_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5266_cast_fp16 = slice_by_index(begin = var_5266_begin_0, end = var_5266_end_0, end_mask = var_5266_end_mask_0, x = var_5256_cast_fp16)[name = string("op_5266_cast_fp16")];
            bool var_5275_transpose_x_1 = const()[name = string("op_5275_transpose_x_1"), val = bool(false)];
            bool var_5275_transpose_y_1 = const()[name = string("op_5275_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_5275_cast_fp16 = matmul(transpose_x = var_5275_transpose_x_1, transpose_y = var_5275_transpose_y_1, x = var_5266_cast_fp16, y = stacked_23_cast_fp16)[name = string("op_5275_cast_fp16")];
            tensor<int32, [4]> var_5278_perm_0 = const()[name = string("op_5278_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5279_to_fp16 = const()[name = string("op_5279_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_5278_cast_fp16 = transpose(perm = var_5278_perm_0, x = var_5275_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, 32, 2048]> var_5280_cast_fp16 = mul(x = var_5278_cast_fp16, y = var_5279_to_fp16)[name = string("op_5280_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_23_cast_fp16 = add(x = var_5280_cast_fp16, y = cacheMask)[name = string("cache_weights_23_cast_fp16")];
            bool var_5286_transpose_x_1 = const()[name = string("op_5286_transpose_x_1"), val = bool(false)];
            bool var_5286_transpose_y_1 = const()[name = string("op_5286_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5286_cast_fp16 = matmul(transpose_x = var_5286_transpose_x_1, transpose_y = var_5286_transpose_y_1, x = stacked_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_5286_cast_fp16")];
            fp16 var_5287_to_fp16 = const()[name = string("op_5287_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5288_cast_fp16 = mul(x = var_5286_cast_fp16, y = var_5287_to_fp16)[name = string("op_5288_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_23_cast_fp16 = add(x = var_5288_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_23_cast_fp16")];
            int32 var_5292 = const()[name = string("op_5292"), val = int32(-1)];
            bool var_5293_interleave_0 = const()[name = string("op_5293_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_5293_cast_fp16 = concat(axis = var_5292, interleave = var_5293_interleave_0, values = (cache_weights_23_cast_fp16, call_weights_23_cast_fp16))[name = string("op_5293_cast_fp16")];
            int32 var_5294 = const()[name = string("op_5294"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_23_cast_fp16 = softmax(axis = var_5294, x = var_5293_cast_fp16)[name = string("shares_23_cast_fp16")];
            tensor<int32, [4]> var_5301_begin_0 = const()[name = string("op_5301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5301_end_0 = const()[name = string("op_5301_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_5301_end_mask_0 = const()[name = string("op_5301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_5301_cast_fp16 = slice_by_index(begin = var_5301_begin_0, end = var_5301_end_0, end_mask = var_5301_end_mask_0, x = shares_23_cast_fp16)[name = string("op_5301_cast_fp16")];
            tensor<int32, [4]> var_5306_begin_0 = const()[name = string("op_5306_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_5306_end_0 = const()[name = string("op_5306_end_0"), val = tensor<int32, [4]>([12, 8, 2064, 128])];
            tensor<bool, [4]> var_5306_end_mask_0 = const()[name = string("op_5306_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5306_cast_fp16 = slice_by_index(begin = var_5306_begin_0, end = var_5306_end_0, end_mask = var_5306_end_mask_0, x = read_state_1)[name = string("op_5306_cast_fp16")];
            tensor<int32, [4]> var_5316_begin_0 = const()[name = string("op_5316_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5316_end_0 = const()[name = string("op_5316_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5316_end_mask_0 = const()[name = string("op_5316_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5316_cast_fp16 = slice_by_index(begin = var_5316_begin_0, end = var_5316_end_0, end_mask = var_5316_end_mask_0, x = var_5306_cast_fp16)[name = string("op_5316_cast_fp16")];
            bool var_5322_transpose_x_0 = const()[name = string("op_5322_transpose_x_0"), val = bool(false)];
            bool var_5322_transpose_y_0 = const()[name = string("op_5322_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5322_cast_fp16 = matmul(transpose_x = var_5322_transpose_x_0, transpose_y = var_5322_transpose_y_0, x = var_5301_cast_fp16, y = var_5316_cast_fp16)[name = string("op_5322_cast_fp16")];
            tensor<int32, [4]> var_5327_begin_0 = const()[name = string("op_5327_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_5327_end_0 = const()[name = string("op_5327_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_5327_end_mask_0 = const()[name = string("op_5327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5327_cast_fp16 = slice_by_index(begin = var_5327_begin_0, end = var_5327_end_0, end_mask = var_5327_end_mask_0, x = shares_23_cast_fp16)[name = string("op_5327_cast_fp16")];
            bool var_5328_transpose_x_0 = const()[name = string("op_5328_transpose_x_0"), val = bool(false)];
            bool var_5328_transpose_y_0 = const()[name = string("op_5328_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_5183_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 8, 32, 128]> var_5328_cast_fp16 = matmul(transpose_x = var_5328_transpose_x_0, transpose_y = var_5328_transpose_y_0, x = var_5327_cast_fp16, y = v_23_cast_fp16)[name = string("op_5328_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_45_cast_fp16 = add(x = var_5322_cast_fp16, y = var_5328_cast_fp16)[name = string("mixed_45_cast_fp16")];
            tensor<int32, [4]> var_5345_begin_0 = const()[name = string("op_5345_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5345_end_0 = const()[name = string("op_5345_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5345_end_mask_0 = const()[name = string("op_5345_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5345_cast_fp16 = slice_by_index(begin = var_5345_begin_0, end = var_5345_end_0, end_mask = var_5345_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5345_cast_fp16")];
            tensor<int32, [4]> var_5360_begin_0 = const()[name = string("op_5360_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5360_end_0 = const()[name = string("op_5360_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5360_end_mask_0 = const()[name = string("op_5360_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5360_cast_fp16 = slice_by_index(begin = var_5360_begin_0, end = var_5360_end_0, end_mask = var_5360_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5360_cast_fp16")];
            int32 var_5362 = const()[name = string("op_5362"), 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_5362, interleave = mixed_47_interleave_0, values = (var_5345_cast_fp16, var_5360_cast_fp16))[name = string("mixed_47_cast_fp16")];
            tensor<int32, [4]> var_5366_perm_0 = const()[name = string("op_5366_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5370 = const()[name = string("op_5370"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5366_cast_fp16 = transpose(perm = var_5366_perm_0, x = mixed_47_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 16, 2048]> input_113_cast_fp16 = reshape(shape = var_5370, x = var_5366_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_237_cast_fp16 = add(x = x_221_cast_fp16, y = linear_80_cast_fp16)[name = string("x_237_cast_fp16")];
            bool var_5380 = const()[name = string("op_5380"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5383_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_5383_cast_fp16")];
            tensor<int32, [1]> peak_95_axes_0 = const()[name = string("peak_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_95_cast_fp16 = reduce_max(axes = peak_95_axes_0, keep_dims = var_5380, x = var_5383_cast_fp16)[name = string("peak_95_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_239_cast_fp16 = real_div(x = x_237_cast_fp16, y = peak_95_cast_fp16)[name = string("x_239_cast_fp16")];
            fp16 var_5379_promoted_to_fp16 = const()[name = string("op_5379_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5387_cast_fp16 = pow(x = x_239_cast_fp16, y = var_5379_promoted_to_fp16)[name = string("op_5387_cast_fp16")];
            tensor<int32, [1]> var_5389_axes_0 = const()[name = string("op_5389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5389_keep_dims_0 = const()[name = string("op_5389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5389_cast_fp16 = reduce_mean(axes = var_5389_axes_0, keep_dims = var_5389_keep_dims_0, x = var_5387_cast_fp16)[name = string("op_5389_cast_fp16")];
            fp16 var_5379_promoted_1_to_fp16 = const()[name = string("op_5379_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5390_cast_fp16 = pow(x = peak_95_cast_fp16, y = var_5379_promoted_1_to_fp16)[name = string("op_5390_cast_fp16")];
            fp32 var_5391_epsilon_0 = const()[name = string("op_5391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5391_cast_fp16 = inverse(epsilon = var_5391_epsilon_0, x = var_5390_cast_fp16)[name = string("op_5391_cast_fp16")];
            fp16 var_5392_to_fp16 = const()[name = string("op_5392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5393_cast_fp16 = mul(x = var_5391_cast_fp16, y = var_5392_to_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5394_cast_fp16 = add(x = var_5389_cast_fp16, y = var_5393_cast_fp16)[name = string("op_5394_cast_fp16")];
            fp32 var_5395_epsilon_0 = const()[name = string("op_5395_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5395_cast_fp16 = rsqrt(epsilon = var_5395_epsilon_0, x = var_5394_cast_fp16)[name = string("op_5395_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5396_cast_fp16 = mul(x = x_239_cast_fp16, y = var_5395_cast_fp16)[name = string("op_5396_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_5396_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_5401_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_5401_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_5401_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_241_cast_fp16 = add(x = x_237_cast_fp16, y = linear_83_cast_fp16)[name = string("x_241_cast_fp16")];
            bool var_5414 = const()[name = string("op_5414"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5417_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_5417_cast_fp16")];
            tensor<int32, [1]> peak_97_axes_0 = const()[name = string("peak_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_97_cast_fp16 = reduce_max(axes = peak_97_axes_0, keep_dims = var_5414, x = var_5417_cast_fp16)[name = string("peak_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_243_cast_fp16 = real_div(x = x_241_cast_fp16, y = peak_97_cast_fp16)[name = string("x_243_cast_fp16")];
            fp16 var_5413_promoted_to_fp16 = const()[name = string("op_5413_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5421_cast_fp16 = pow(x = x_243_cast_fp16, y = var_5413_promoted_to_fp16)[name = string("op_5421_cast_fp16")];
            tensor<int32, [1]> var_5423_axes_0 = const()[name = string("op_5423_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5423_keep_dims_0 = const()[name = string("op_5423_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5423_cast_fp16 = reduce_mean(axes = var_5423_axes_0, keep_dims = var_5423_keep_dims_0, x = var_5421_cast_fp16)[name = string("op_5423_cast_fp16")];
            fp16 var_5413_promoted_1_to_fp16 = const()[name = string("op_5413_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5424_cast_fp16 = pow(x = peak_97_cast_fp16, y = var_5413_promoted_1_to_fp16)[name = string("op_5424_cast_fp16")];
            fp32 var_5425_epsilon_0 = const()[name = string("op_5425_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5425_cast_fp16 = inverse(epsilon = var_5425_epsilon_0, x = var_5424_cast_fp16)[name = string("op_5425_cast_fp16")];
            fp16 var_5426_to_fp16 = const()[name = string("op_5426_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5427_cast_fp16 = mul(x = var_5425_cast_fp16, y = var_5426_to_fp16)[name = string("op_5427_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5428_cast_fp16 = add(x = var_5423_cast_fp16, y = var_5427_cast_fp16)[name = string("op_5428_cast_fp16")];
            fp32 var_5429_epsilon_0 = const()[name = string("op_5429_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5429_cast_fp16 = rsqrt(epsilon = var_5429_epsilon_0, x = var_5428_cast_fp16)[name = string("op_5429_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5430_cast_fp16 = mul(x = x_243_cast_fp16, y = var_5429_cast_fp16)[name = string("op_5430_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_5430_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_5439 = const()[name = string("op_5439"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5440_cast_fp16 = reshape(shape = var_5439, x = linear_84_cast_fp16)[name = string("op_5440_cast_fp16")];
            tensor<int32, [4]> x_245_perm_0 = const()[name = string("x_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5447 = const()[name = string("op_5447"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_245_cast_fp16 = transpose(perm = x_245_perm_0, x = var_5440_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 16, 128]> var_5450_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_5450_cast_fp16")];
            tensor<int32, [1]> peak_99_axes_0 = const()[name = string("peak_99_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_99_cast_fp16 = reduce_max(axes = peak_99_axes_0, keep_dims = var_5447, x = var_5450_cast_fp16)[name = string("peak_99_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_247_cast_fp16 = real_div(x = x_245_cast_fp16, y = peak_99_cast_fp16)[name = string("x_247_cast_fp16")];
            fp16 var_5446_promoted_to_fp16 = const()[name = string("op_5446_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5454_cast_fp16 = pow(x = x_247_cast_fp16, y = var_5446_promoted_to_fp16)[name = string("op_5454_cast_fp16")];
            tensor<int32, [1]> var_5456_axes_0 = const()[name = string("op_5456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5456_keep_dims_0 = const()[name = string("op_5456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5456_cast_fp16 = reduce_mean(axes = var_5456_axes_0, keep_dims = var_5456_keep_dims_0, x = var_5454_cast_fp16)[name = string("op_5456_cast_fp16")];
            fp16 var_5446_promoted_1_to_fp16 = const()[name = string("op_5446_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5457_cast_fp16 = pow(x = peak_99_cast_fp16, y = var_5446_promoted_1_to_fp16)[name = string("op_5457_cast_fp16")];
            fp32 var_5458_epsilon_0 = const()[name = string("op_5458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5458_cast_fp16 = inverse(epsilon = var_5458_epsilon_0, x = var_5457_cast_fp16)[name = string("op_5458_cast_fp16")];
            fp16 var_5459_to_fp16 = const()[name = string("op_5459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5460_cast_fp16 = mul(x = var_5458_cast_fp16, y = var_5459_to_fp16)[name = string("op_5460_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5461_cast_fp16 = add(x = var_5456_cast_fp16, y = var_5460_cast_fp16)[name = string("op_5461_cast_fp16")];
            fp32 var_5462_epsilon_0 = const()[name = string("op_5462_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5462_cast_fp16 = rsqrt(epsilon = var_5462_epsilon_0, x = var_5461_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5463_cast_fp16 = mul(x = x_247_cast_fp16, y = var_5462_cast_fp16)[name = string("op_5463_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_253_cast_fp16 = mul(x = var_5463_cast_fp16, y = model_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_253_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_5472 = const()[name = string("op_5472"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5473_cast_fp16 = reshape(shape = var_5472, x = linear_85_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<int32, [4]> x_249_perm_0 = const()[name = string("x_249_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5480 = const()[name = string("op_5480"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_249_cast_fp16 = transpose(perm = x_249_perm_0, x = var_5473_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 16, 128]> var_5483_cast_fp16 = abs(x = x_249_cast_fp16)[name = string("op_5483_cast_fp16")];
            tensor<int32, [1]> peak_101_axes_0 = const()[name = string("peak_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_101_cast_fp16 = reduce_max(axes = peak_101_axes_0, keep_dims = var_5480, x = var_5483_cast_fp16)[name = string("peak_101_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_251_cast_fp16 = real_div(x = x_249_cast_fp16, y = peak_101_cast_fp16)[name = string("x_251_cast_fp16")];
            fp16 var_5479_promoted_to_fp16 = const()[name = string("op_5479_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5487_cast_fp16 = pow(x = x_251_cast_fp16, y = var_5479_promoted_to_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5489_axes_0 = const()[name = string("op_5489_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5489_keep_dims_0 = const()[name = string("op_5489_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5489_cast_fp16 = reduce_mean(axes = var_5489_axes_0, keep_dims = var_5489_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5489_cast_fp16")];
            fp16 var_5479_promoted_1_to_fp16 = const()[name = string("op_5479_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5490_cast_fp16 = pow(x = peak_101_cast_fp16, y = var_5479_promoted_1_to_fp16)[name = string("op_5490_cast_fp16")];
            fp32 var_5491_epsilon_0 = const()[name = string("op_5491_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5491_cast_fp16 = inverse(epsilon = var_5491_epsilon_0, x = var_5490_cast_fp16)[name = string("op_5491_cast_fp16")];
            fp16 var_5492_to_fp16 = const()[name = string("op_5492_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5493_cast_fp16 = mul(x = var_5491_cast_fp16, y = var_5492_to_fp16)[name = string("op_5493_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5494_cast_fp16 = add(x = var_5489_cast_fp16, y = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp32 var_5495_epsilon_0 = const()[name = string("op_5495_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5495_cast_fp16 = rsqrt(epsilon = var_5495_epsilon_0, x = var_5494_cast_fp16)[name = string("op_5495_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5496_cast_fp16 = mul(x = x_251_cast_fp16, y = var_5495_cast_fp16)[name = string("op_5496_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_255_cast_fp16 = mul(x = var_5496_cast_fp16, y = model_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_255_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_5505 = const()[name = string("op_5505"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5506_cast_fp16 = reshape(shape = var_5505, x = linear_86_cast_fp16)[name = string("op_5506_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_253_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_253_cast_fp16)[name = string("second_49_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5520_cast_fp16 = mul(x = first_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5520_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5521_cast_fp16 = mul(x = second_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5521_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5523_cast_fp16 = sub(x = var_5520_cast_fp16, y = var_5521_cast_fp16)[name = string("op_5523_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5524_cast_fp16 = mul(x = second_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5525_cast_fp16 = mul(x = first_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5525_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5527_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5525_cast_fp16)[name = string("op_5527_cast_fp16")];
            int32 var_5529 = const()[name = string("op_5529"), 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_5529, interleave = q_25_interleave_0, values = (var_5523_cast_fp16, var_5527_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_255_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_255_cast_fp16)[name = string("second_51_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5541_cast_fp16 = mul(x = first_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5541_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5542_cast_fp16 = mul(x = second_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5542_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5544_cast_fp16 = sub(x = var_5541_cast_fp16, y = var_5542_cast_fp16)[name = string("op_5544_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5545_cast_fp16 = mul(x = second_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5545_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5546_cast_fp16 = mul(x = first_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5548_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5546_cast_fp16)[name = string("op_5548_cast_fp16")];
            int32 var_5550 = const()[name = string("op_5550"), 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_5550, interleave = k_25_interleave_0, values = (var_5544_cast_fp16, var_5548_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_5561_begin_0 = const()[name = string("op_5561_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5561_end_0 = const()[name = string("op_5561_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5561_end_mask_0 = const()[name = string("op_5561_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5561_stride_0 = const()[name = string("op_5561_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5561_cast_fp16 = slice_by_index(begin = var_5561_begin_0, end = var_5561_end_0, end_mask = var_5561_end_mask_0, stride = var_5561_stride_0, x = q_25_cast_fp16)[name = string("op_5561_cast_fp16")];
            tensor<int32, [4]> var_5571_begin_0 = const()[name = string("op_5571_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5571_end_0 = const()[name = string("op_5571_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5571_end_mask_0 = const()[name = string("op_5571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5571_stride_0 = const()[name = string("op_5571_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5571_cast_fp16 = slice_by_index(begin = var_5571_begin_0, end = var_5571_end_0, end_mask = var_5571_end_mask_0, stride = var_5571_stride_0, x = q_25_cast_fp16)[name = string("op_5571_cast_fp16")];
            int32 var_5573 = const()[name = string("op_5573"), 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_5573, interleave = stacked_25_interleave_0, values = (var_5561_cast_fp16, var_5571_cast_fp16))[name = string("stacked_25_cast_fp16")];
            tensor<int32, [4]> var_5579_begin_0 = const()[name = string("op_5579_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5579_end_0 = const()[name = string("op_5579_end_0"), val = tensor<int32, [4]>([13, 8, 2064, 128])];
            tensor<bool, [4]> var_5579_end_mask_0 = const()[name = string("op_5579_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5579_cast_fp16 = slice_by_index(begin = var_5579_begin_0, end = var_5579_end_0, end_mask = var_5579_end_mask_0, x = read_state_0)[name = string("op_5579_cast_fp16")];
            tensor<int32, [4]> var_5589_begin_0 = const()[name = string("op_5589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5589_end_0 = const()[name = string("op_5589_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5589_end_mask_0 = const()[name = string("op_5589_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5589_cast_fp16 = slice_by_index(begin = var_5589_begin_0, end = var_5589_end_0, end_mask = var_5589_end_mask_0, x = var_5579_cast_fp16)[name = string("op_5589_cast_fp16")];
            bool var_5598_transpose_x_1 = const()[name = string("op_5598_transpose_x_1"), val = bool(false)];
            bool var_5598_transpose_y_1 = const()[name = string("op_5598_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_5598_cast_fp16 = matmul(transpose_x = var_5598_transpose_x_1, transpose_y = var_5598_transpose_y_1, x = var_5589_cast_fp16, y = stacked_25_cast_fp16)[name = string("op_5598_cast_fp16")];
            tensor<int32, [4]> var_5601_perm_0 = const()[name = string("op_5601_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5602_to_fp16 = const()[name = string("op_5602_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_5601_cast_fp16 = transpose(perm = var_5601_perm_0, x = var_5598_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 32, 2048]> var_5603_cast_fp16 = mul(x = var_5601_cast_fp16, y = var_5602_to_fp16)[name = string("op_5603_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_25_cast_fp16 = add(x = var_5603_cast_fp16, y = cacheMask)[name = string("cache_weights_25_cast_fp16")];
            bool var_5609_transpose_x_1 = const()[name = string("op_5609_transpose_x_1"), val = bool(false)];
            bool var_5609_transpose_y_1 = const()[name = string("op_5609_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5609_cast_fp16 = matmul(transpose_x = var_5609_transpose_x_1, transpose_y = var_5609_transpose_y_1, x = stacked_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_5609_cast_fp16")];
            fp16 var_5610_to_fp16 = const()[name = string("op_5610_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5611_cast_fp16 = mul(x = var_5609_cast_fp16, y = var_5610_to_fp16)[name = string("op_5611_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_25_cast_fp16 = add(x = var_5611_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_25_cast_fp16")];
            int32 var_5615 = const()[name = string("op_5615"), val = int32(-1)];
            bool var_5616_interleave_0 = const()[name = string("op_5616_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_5616_cast_fp16 = concat(axis = var_5615, interleave = var_5616_interleave_0, values = (cache_weights_25_cast_fp16, call_weights_25_cast_fp16))[name = string("op_5616_cast_fp16")];
            int32 var_5617 = const()[name = string("op_5617"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_25_cast_fp16 = softmax(axis = var_5617, x = var_5616_cast_fp16)[name = string("shares_25_cast_fp16")];
            tensor<int32, [4]> var_5624_begin_0 = const()[name = string("op_5624_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5624_end_0 = const()[name = string("op_5624_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_5624_end_mask_0 = const()[name = string("op_5624_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_5624_cast_fp16 = slice_by_index(begin = var_5624_begin_0, end = var_5624_end_0, end_mask = var_5624_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([13, 8, 2064, 128])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = read_state_1)[name = string("op_5629_cast_fp16")];
            tensor<int32, [4]> var_5639_begin_0 = const()[name = string("op_5639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5639_end_0 = const()[name = string("op_5639_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5639_end_mask_0 = const()[name = string("op_5639_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5639_cast_fp16 = slice_by_index(begin = var_5639_begin_0, end = var_5639_end_0, end_mask = var_5639_end_mask_0, x = var_5629_cast_fp16)[name = string("op_5639_cast_fp16")];
            bool var_5645_transpose_x_0 = const()[name = string("op_5645_transpose_x_0"), val = bool(false)];
            bool var_5645_transpose_y_0 = const()[name = string("op_5645_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5645_cast_fp16 = matmul(transpose_x = var_5645_transpose_x_0, transpose_y = var_5645_transpose_y_0, x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("op_5645_cast_fp16")];
            tensor<int32, [4]> var_5650_begin_0 = const()[name = string("op_5650_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_5650_end_0 = const()[name = string("op_5650_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_5650_end_mask_0 = const()[name = string("op_5650_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5650_cast_fp16 = slice_by_index(begin = var_5650_begin_0, end = var_5650_end_0, end_mask = var_5650_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5650_cast_fp16")];
            bool var_5651_transpose_x_0 = const()[name = string("op_5651_transpose_x_0"), val = bool(false)];
            bool var_5651_transpose_y_0 = const()[name = string("op_5651_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_5506_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 8, 32, 128]> var_5651_cast_fp16 = matmul(transpose_x = var_5651_transpose_x_0, transpose_y = var_5651_transpose_y_0, x = var_5650_cast_fp16, y = v_25_cast_fp16)[name = string("op_5651_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_49_cast_fp16 = add(x = var_5645_cast_fp16, y = var_5651_cast_fp16)[name = string("mixed_49_cast_fp16")];
            tensor<int32, [4]> var_5668_begin_0 = const()[name = string("op_5668_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5668_end_0 = const()[name = string("op_5668_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5668_end_mask_0 = const()[name = string("op_5668_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5668_cast_fp16 = slice_by_index(begin = var_5668_begin_0, end = var_5668_end_0, end_mask = var_5668_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5668_cast_fp16")];
            tensor<int32, [4]> var_5683_begin_0 = const()[name = string("op_5683_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5683_end_0 = const()[name = string("op_5683_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5683_end_mask_0 = const()[name = string("op_5683_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5683_cast_fp16 = slice_by_index(begin = var_5683_begin_0, end = var_5683_end_0, end_mask = var_5683_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5683_cast_fp16")];
            int32 var_5685 = const()[name = string("op_5685"), 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_5685, interleave = mixed_51_interleave_0, values = (var_5668_cast_fp16, var_5683_cast_fp16))[name = string("mixed_51_cast_fp16")];
            tensor<int32, [4]> var_5689_perm_0 = const()[name = string("op_5689_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5693 = const()[name = string("op_5693"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5689_cast_fp16 = transpose(perm = var_5689_perm_0, x = mixed_51_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 2048]> input_123_cast_fp16 = reshape(shape = var_5693, x = var_5689_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_257_cast_fp16 = add(x = x_241_cast_fp16, y = linear_87_cast_fp16)[name = string("x_257_cast_fp16")];
            bool var_5703 = const()[name = string("op_5703"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5706_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_5706_cast_fp16")];
            tensor<int32, [1]> peak_103_axes_0 = const()[name = string("peak_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_103_cast_fp16 = reduce_max(axes = peak_103_axes_0, keep_dims = var_5703, x = var_5706_cast_fp16)[name = string("peak_103_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_259_cast_fp16 = real_div(x = x_257_cast_fp16, y = peak_103_cast_fp16)[name = string("x_259_cast_fp16")];
            fp16 var_5702_promoted_to_fp16 = const()[name = string("op_5702_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5710_cast_fp16 = pow(x = x_259_cast_fp16, y = var_5702_promoted_to_fp16)[name = string("op_5710_cast_fp16")];
            tensor<int32, [1]> var_5712_axes_0 = const()[name = string("op_5712_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5712_keep_dims_0 = const()[name = string("op_5712_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5712_cast_fp16 = reduce_mean(axes = var_5712_axes_0, keep_dims = var_5712_keep_dims_0, x = var_5710_cast_fp16)[name = string("op_5712_cast_fp16")];
            fp16 var_5702_promoted_1_to_fp16 = const()[name = string("op_5702_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5713_cast_fp16 = pow(x = peak_103_cast_fp16, y = var_5702_promoted_1_to_fp16)[name = string("op_5713_cast_fp16")];
            fp32 var_5714_epsilon_0 = const()[name = string("op_5714_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5714_cast_fp16 = inverse(epsilon = var_5714_epsilon_0, x = var_5713_cast_fp16)[name = string("op_5714_cast_fp16")];
            fp16 var_5715_to_fp16 = const()[name = string("op_5715_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5716_cast_fp16 = mul(x = var_5714_cast_fp16, y = var_5715_to_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5717_cast_fp16 = add(x = var_5712_cast_fp16, y = var_5716_cast_fp16)[name = string("op_5717_cast_fp16")];
            fp32 var_5718_epsilon_0 = const()[name = string("op_5718_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5718_cast_fp16 = rsqrt(epsilon = var_5718_epsilon_0, x = var_5717_cast_fp16)[name = string("op_5718_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5719_cast_fp16 = mul(x = x_259_cast_fp16, y = var_5718_cast_fp16)[name = string("op_5719_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_5719_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_5724_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_5724_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_5724_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_261_cast_fp16 = add(x = x_257_cast_fp16, y = linear_90_cast_fp16)[name = string("x_261_cast_fp16")];
            bool var_5737 = const()[name = string("op_5737"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5740_cast_fp16 = abs(x = x_261_cast_fp16)[name = string("op_5740_cast_fp16")];
            tensor<int32, [1]> peak_105_axes_0 = const()[name = string("peak_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_105_cast_fp16 = reduce_max(axes = peak_105_axes_0, keep_dims = var_5737, x = var_5740_cast_fp16)[name = string("peak_105_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_263_cast_fp16 = real_div(x = x_261_cast_fp16, y = peak_105_cast_fp16)[name = string("x_263_cast_fp16")];
            fp16 var_5736_promoted_to_fp16 = const()[name = string("op_5736_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5744_cast_fp16 = pow(x = x_263_cast_fp16, y = var_5736_promoted_to_fp16)[name = string("op_5744_cast_fp16")];
            tensor<int32, [1]> var_5746_axes_0 = const()[name = string("op_5746_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5746_keep_dims_0 = const()[name = string("op_5746_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5746_cast_fp16 = reduce_mean(axes = var_5746_axes_0, keep_dims = var_5746_keep_dims_0, x = var_5744_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5736_promoted_1_to_fp16 = const()[name = string("op_5736_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5747_cast_fp16 = pow(x = peak_105_cast_fp16, y = var_5736_promoted_1_to_fp16)[name = string("op_5747_cast_fp16")];
            fp32 var_5748_epsilon_0 = const()[name = string("op_5748_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5748_cast_fp16 = inverse(epsilon = var_5748_epsilon_0, x = var_5747_cast_fp16)[name = string("op_5748_cast_fp16")];
            fp16 var_5749_to_fp16 = const()[name = string("op_5749_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5750_cast_fp16 = mul(x = var_5748_cast_fp16, y = var_5749_to_fp16)[name = string("op_5750_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5751_cast_fp16 = add(x = var_5746_cast_fp16, y = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            fp32 var_5752_epsilon_0 = const()[name = string("op_5752_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5752_cast_fp16 = rsqrt(epsilon = var_5752_epsilon_0, x = var_5751_cast_fp16)[name = string("op_5752_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5753_cast_fp16 = mul(x = x_263_cast_fp16, y = var_5752_cast_fp16)[name = string("op_5753_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_5753_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_5762 = const()[name = string("op_5762"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5763_cast_fp16 = reshape(shape = var_5762, x = linear_91_cast_fp16)[name = string("op_5763_cast_fp16")];
            tensor<int32, [4]> x_265_perm_0 = const()[name = string("x_265_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5770 = const()[name = string("op_5770"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_265_cast_fp16 = transpose(perm = x_265_perm_0, x = var_5763_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 16, 128]> var_5773_cast_fp16 = abs(x = x_265_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<int32, [1]> peak_107_axes_0 = const()[name = string("peak_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_107_cast_fp16 = reduce_max(axes = peak_107_axes_0, keep_dims = var_5770, x = var_5773_cast_fp16)[name = string("peak_107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_267_cast_fp16 = real_div(x = x_265_cast_fp16, y = peak_107_cast_fp16)[name = string("x_267_cast_fp16")];
            fp16 var_5769_promoted_to_fp16 = const()[name = string("op_5769_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5777_cast_fp16 = pow(x = x_267_cast_fp16, y = var_5769_promoted_to_fp16)[name = string("op_5777_cast_fp16")];
            tensor<int32, [1]> var_5779_axes_0 = const()[name = string("op_5779_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5779_keep_dims_0 = const()[name = string("op_5779_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5779_cast_fp16 = reduce_mean(axes = var_5779_axes_0, keep_dims = var_5779_keep_dims_0, x = var_5777_cast_fp16)[name = string("op_5779_cast_fp16")];
            fp16 var_5769_promoted_1_to_fp16 = const()[name = string("op_5769_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5780_cast_fp16 = pow(x = peak_107_cast_fp16, y = var_5769_promoted_1_to_fp16)[name = string("op_5780_cast_fp16")];
            fp32 var_5781_epsilon_0 = const()[name = string("op_5781_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5781_cast_fp16 = inverse(epsilon = var_5781_epsilon_0, x = var_5780_cast_fp16)[name = string("op_5781_cast_fp16")];
            fp16 var_5782_to_fp16 = const()[name = string("op_5782_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5783_cast_fp16 = mul(x = var_5781_cast_fp16, y = var_5782_to_fp16)[name = string("op_5783_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5784_cast_fp16 = add(x = var_5779_cast_fp16, y = var_5783_cast_fp16)[name = string("op_5784_cast_fp16")];
            fp32 var_5785_epsilon_0 = const()[name = string("op_5785_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5785_cast_fp16 = rsqrt(epsilon = var_5785_epsilon_0, x = var_5784_cast_fp16)[name = string("op_5785_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5786_cast_fp16 = mul(x = x_267_cast_fp16, y = var_5785_cast_fp16)[name = string("op_5786_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_273_cast_fp16 = mul(x = var_5786_cast_fp16, y = model_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_273_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_5795 = const()[name = string("op_5795"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5796_cast_fp16 = reshape(shape = var_5795, x = linear_92_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5803 = const()[name = string("op_5803"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_5796_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 8, 16, 128]> var_5806_cast_fp16 = abs(x = x_269_cast_fp16)[name = string("op_5806_cast_fp16")];
            tensor<int32, [1]> peak_109_axes_0 = const()[name = string("peak_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_109_cast_fp16 = reduce_max(axes = peak_109_axes_0, keep_dims = var_5803, x = var_5806_cast_fp16)[name = string("peak_109_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_271_cast_fp16 = real_div(x = x_269_cast_fp16, y = peak_109_cast_fp16)[name = string("x_271_cast_fp16")];
            fp16 var_5802_promoted_to_fp16 = const()[name = string("op_5802_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5810_cast_fp16 = pow(x = x_271_cast_fp16, y = var_5802_promoted_to_fp16)[name = string("op_5810_cast_fp16")];
            tensor<int32, [1]> var_5812_axes_0 = const()[name = string("op_5812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5812_keep_dims_0 = const()[name = string("op_5812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5812_cast_fp16 = reduce_mean(axes = var_5812_axes_0, keep_dims = var_5812_keep_dims_0, x = var_5810_cast_fp16)[name = string("op_5812_cast_fp16")];
            fp16 var_5802_promoted_1_to_fp16 = const()[name = string("op_5802_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5813_cast_fp16 = pow(x = peak_109_cast_fp16, y = var_5802_promoted_1_to_fp16)[name = string("op_5813_cast_fp16")];
            fp32 var_5814_epsilon_0 = const()[name = string("op_5814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5814_cast_fp16 = inverse(epsilon = var_5814_epsilon_0, x = var_5813_cast_fp16)[name = string("op_5814_cast_fp16")];
            fp16 var_5815_to_fp16 = const()[name = string("op_5815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5816_cast_fp16 = mul(x = var_5814_cast_fp16, y = var_5815_to_fp16)[name = string("op_5816_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5817_cast_fp16 = add(x = var_5812_cast_fp16, y = var_5816_cast_fp16)[name = string("op_5817_cast_fp16")];
            fp32 var_5818_epsilon_0 = const()[name = string("op_5818_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5818_cast_fp16 = rsqrt(epsilon = var_5818_epsilon_0, x = var_5817_cast_fp16)[name = string("op_5818_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5819_cast_fp16 = mul(x = x_271_cast_fp16, y = var_5818_cast_fp16)[name = string("op_5819_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_275_cast_fp16 = mul(x = var_5819_cast_fp16, y = model_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_275_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_5828 = const()[name = string("op_5828"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5829_cast_fp16 = reshape(shape = var_5828, x = linear_93_cast_fp16)[name = string("op_5829_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_273_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_273_cast_fp16)[name = string("second_53_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5843_cast_fp16 = mul(x = first_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5843_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5844_cast_fp16 = mul(x = second_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5844_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5846_cast_fp16 = sub(x = var_5843_cast_fp16, y = var_5844_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5847_cast_fp16 = mul(x = second_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5847_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5848_cast_fp16 = mul(x = first_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5848_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5850_cast_fp16 = add(x = var_5847_cast_fp16, y = var_5848_cast_fp16)[name = string("op_5850_cast_fp16")];
            int32 var_5852 = const()[name = string("op_5852"), 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_5852, interleave = q_27_interleave_0, values = (var_5846_cast_fp16, var_5850_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_275_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_275_cast_fp16)[name = string("second_55_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5864_cast_fp16 = mul(x = first_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5865_cast_fp16 = mul(x = second_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5865_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5867_cast_fp16 = sub(x = var_5864_cast_fp16, y = var_5865_cast_fp16)[name = string("op_5867_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5868_cast_fp16 = mul(x = second_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5868_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5869_cast_fp16 = mul(x = first_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5869_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5871_cast_fp16 = add(x = var_5868_cast_fp16, y = var_5869_cast_fp16)[name = string("op_5871_cast_fp16")];
            int32 var_5873 = const()[name = string("op_5873"), 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_5873, interleave = k_27_interleave_0, values = (var_5867_cast_fp16, var_5871_cast_fp16))[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> var_5884_begin_0 = const()[name = string("op_5884_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5884_end_0 = const()[name = string("op_5884_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5884_end_mask_0 = const()[name = string("op_5884_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5884_stride_0 = const()[name = string("op_5884_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5884_cast_fp16 = slice_by_index(begin = var_5884_begin_0, end = var_5884_end_0, end_mask = var_5884_end_mask_0, stride = var_5884_stride_0, x = q_27_cast_fp16)[name = string("op_5884_cast_fp16")];
            tensor<int32, [4]> var_5894_begin_0 = const()[name = string("op_5894_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5894_end_0 = const()[name = string("op_5894_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5894_end_mask_0 = const()[name = string("op_5894_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5894_stride_0 = const()[name = string("op_5894_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5894_cast_fp16 = slice_by_index(begin = var_5894_begin_0, end = var_5894_end_0, end_mask = var_5894_end_mask_0, stride = var_5894_stride_0, x = q_27_cast_fp16)[name = string("op_5894_cast_fp16")];
            int32 var_5896 = const()[name = string("op_5896"), 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_5896, interleave = stacked_27_interleave_0, values = (var_5884_cast_fp16, var_5894_cast_fp16))[name = string("stacked_27_cast_fp16")];
            tensor<int32, [4]> var_5902_begin_0 = const()[name = string("op_5902_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5902_end_0 = const()[name = string("op_5902_end_0"), val = tensor<int32, [4]>([14, 8, 2064, 128])];
            tensor<bool, [4]> var_5902_end_mask_0 = const()[name = string("op_5902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5902_cast_fp16 = slice_by_index(begin = var_5902_begin_0, end = var_5902_end_0, end_mask = var_5902_end_mask_0, x = read_state_0)[name = string("op_5902_cast_fp16")];
            tensor<int32, [4]> var_5912_begin_0 = const()[name = string("op_5912_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5912_end_0 = const()[name = string("op_5912_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5912_end_mask_0 = const()[name = string("op_5912_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5912_cast_fp16 = slice_by_index(begin = var_5912_begin_0, end = var_5912_end_0, end_mask = var_5912_end_mask_0, x = var_5902_cast_fp16)[name = string("op_5912_cast_fp16")];
            bool var_5921_transpose_x_1 = const()[name = string("op_5921_transpose_x_1"), val = bool(false)];
            bool var_5921_transpose_y_1 = const()[name = string("op_5921_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_5921_cast_fp16 = matmul(transpose_x = var_5921_transpose_x_1, transpose_y = var_5921_transpose_y_1, x = var_5912_cast_fp16, y = stacked_27_cast_fp16)[name = string("op_5921_cast_fp16")];
            tensor<int32, [4]> var_5924_perm_0 = const()[name = string("op_5924_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5925_to_fp16 = const()[name = string("op_5925_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_5924_cast_fp16 = transpose(perm = var_5924_perm_0, x = var_5921_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 8, 32, 2048]> var_5926_cast_fp16 = mul(x = var_5924_cast_fp16, y = var_5925_to_fp16)[name = string("op_5926_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_27_cast_fp16 = add(x = var_5926_cast_fp16, y = cacheMask)[name = string("cache_weights_27_cast_fp16")];
            bool var_5932_transpose_x_1 = const()[name = string("op_5932_transpose_x_1"), val = bool(false)];
            bool var_5932_transpose_y_1 = const()[name = string("op_5932_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5932_cast_fp16 = matmul(transpose_x = var_5932_transpose_x_1, transpose_y = var_5932_transpose_y_1, x = stacked_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_5932_cast_fp16")];
            fp16 var_5933_to_fp16 = const()[name = string("op_5933_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5934_cast_fp16 = mul(x = var_5932_cast_fp16, y = var_5933_to_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_27_cast_fp16 = add(x = var_5934_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_27_cast_fp16")];
            int32 var_5938 = const()[name = string("op_5938"), val = int32(-1)];
            bool var_5939_interleave_0 = const()[name = string("op_5939_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_5939_cast_fp16 = concat(axis = var_5938, interleave = var_5939_interleave_0, values = (cache_weights_27_cast_fp16, call_weights_27_cast_fp16))[name = string("op_5939_cast_fp16")];
            int32 var_5940 = const()[name = string("op_5940"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_27_cast_fp16 = softmax(axis = var_5940, x = var_5939_cast_fp16)[name = string("shares_27_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5947_cast_fp16")];
            tensor<int32, [4]> var_5952_begin_0 = const()[name = string("op_5952_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5952_end_0 = const()[name = string("op_5952_end_0"), val = tensor<int32, [4]>([14, 8, 2064, 128])];
            tensor<bool, [4]> var_5952_end_mask_0 = const()[name = string("op_5952_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5952_cast_fp16 = slice_by_index(begin = var_5952_begin_0, end = var_5952_end_0, end_mask = var_5952_end_mask_0, x = read_state_1)[name = string("op_5952_cast_fp16")];
            tensor<int32, [4]> var_5962_begin_0 = const()[name = string("op_5962_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5962_end_0 = const()[name = string("op_5962_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5962_end_mask_0 = const()[name = string("op_5962_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5962_cast_fp16 = slice_by_index(begin = var_5962_begin_0, end = var_5962_end_0, end_mask = var_5962_end_mask_0, x = var_5952_cast_fp16)[name = string("op_5962_cast_fp16")];
            bool var_5968_transpose_x_0 = const()[name = string("op_5968_transpose_x_0"), val = bool(false)];
            bool var_5968_transpose_y_0 = const()[name = string("op_5968_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5968_cast_fp16 = matmul(transpose_x = var_5968_transpose_x_0, transpose_y = var_5968_transpose_y_0, x = var_5947_cast_fp16, y = var_5962_cast_fp16)[name = string("op_5968_cast_fp16")];
            tensor<int32, [4]> var_5973_begin_0 = const()[name = string("op_5973_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_5973_end_0 = const()[name = string("op_5973_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_5973_end_mask_0 = const()[name = string("op_5973_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5973_cast_fp16 = slice_by_index(begin = var_5973_begin_0, end = var_5973_end_0, end_mask = var_5973_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5973_cast_fp16")];
            bool var_5974_transpose_x_0 = const()[name = string("op_5974_transpose_x_0"), val = bool(false)];
            bool var_5974_transpose_y_0 = const()[name = string("op_5974_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_5829_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 8, 32, 128]> var_5974_cast_fp16 = matmul(transpose_x = var_5974_transpose_x_0, transpose_y = var_5974_transpose_y_0, x = var_5973_cast_fp16, y = v_27_cast_fp16)[name = string("op_5974_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_53_cast_fp16 = add(x = var_5968_cast_fp16, y = var_5974_cast_fp16)[name = string("mixed_53_cast_fp16")];
            tensor<int32, [4]> var_5991_begin_0 = const()[name = string("op_5991_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5991_end_0 = const()[name = string("op_5991_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5991_end_mask_0 = const()[name = string("op_5991_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5991_cast_fp16 = slice_by_index(begin = var_5991_begin_0, end = var_5991_end_0, end_mask = var_5991_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5991_cast_fp16")];
            tensor<int32, [4]> var_6006_begin_0 = const()[name = string("op_6006_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6006_end_0 = const()[name = string("op_6006_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6006_end_mask_0 = const()[name = string("op_6006_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6006_cast_fp16 = slice_by_index(begin = var_6006_begin_0, end = var_6006_end_0, end_mask = var_6006_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_6006_cast_fp16")];
            int32 var_6008 = const()[name = string("op_6008"), 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_6008, interleave = mixed_55_interleave_0, values = (var_5991_cast_fp16, var_6006_cast_fp16))[name = string("mixed_55_cast_fp16")];
            tensor<int32, [4]> var_6012_perm_0 = const()[name = string("op_6012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6016 = const()[name = string("op_6016"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6012_cast_fp16 = transpose(perm = var_6012_perm_0, x = mixed_55_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 2048]> input_133_cast_fp16 = reshape(shape = var_6016, x = var_6012_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_277_cast_fp16 = add(x = x_261_cast_fp16, y = linear_94_cast_fp16)[name = string("x_277_cast_fp16")];
            bool var_6026 = const()[name = string("op_6026"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6029_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_6029_cast_fp16")];
            tensor<int32, [1]> peak_111_axes_0 = const()[name = string("peak_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_111_cast_fp16 = reduce_max(axes = peak_111_axes_0, keep_dims = var_6026, x = var_6029_cast_fp16)[name = string("peak_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_279_cast_fp16 = real_div(x = x_277_cast_fp16, y = peak_111_cast_fp16)[name = string("x_279_cast_fp16")];
            fp16 var_6025_promoted_to_fp16 = const()[name = string("op_6025_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6033_cast_fp16 = pow(x = x_279_cast_fp16, y = var_6025_promoted_to_fp16)[name = string("op_6033_cast_fp16")];
            tensor<int32, [1]> var_6035_axes_0 = const()[name = string("op_6035_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6035_keep_dims_0 = const()[name = string("op_6035_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6035_cast_fp16 = reduce_mean(axes = var_6035_axes_0, keep_dims = var_6035_keep_dims_0, x = var_6033_cast_fp16)[name = string("op_6035_cast_fp16")];
            fp16 var_6025_promoted_1_to_fp16 = const()[name = string("op_6025_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6036_cast_fp16 = pow(x = peak_111_cast_fp16, y = var_6025_promoted_1_to_fp16)[name = string("op_6036_cast_fp16")];
            fp32 var_6037_epsilon_0 = const()[name = string("op_6037_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6037_cast_fp16 = inverse(epsilon = var_6037_epsilon_0, x = var_6036_cast_fp16)[name = string("op_6037_cast_fp16")];
            fp16 var_6038_to_fp16 = const()[name = string("op_6038_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6040_cast_fp16 = add(x = var_6035_cast_fp16, y = var_6039_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6041_cast_fp16 = rsqrt(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6042_cast_fp16 = mul(x = x_279_cast_fp16, y = var_6041_cast_fp16)[name = string("op_6042_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_6042_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_6047_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_6047_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_6047_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_281_cast_fp16 = add(x = x_277_cast_fp16, y = linear_97_cast_fp16)[name = string("x_281_cast_fp16")];
            bool var_6060 = const()[name = string("op_6060"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6063_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_6063_cast_fp16")];
            tensor<int32, [1]> peak_113_axes_0 = const()[name = string("peak_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_113_cast_fp16 = reduce_max(axes = peak_113_axes_0, keep_dims = var_6060, x = var_6063_cast_fp16)[name = string("peak_113_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_283_cast_fp16 = real_div(x = x_281_cast_fp16, y = peak_113_cast_fp16)[name = string("x_283_cast_fp16")];
            fp16 var_6059_promoted_to_fp16 = const()[name = string("op_6059_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6067_cast_fp16 = pow(x = x_283_cast_fp16, y = var_6059_promoted_to_fp16)[name = string("op_6067_cast_fp16")];
            tensor<int32, [1]> var_6069_axes_0 = const()[name = string("op_6069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6069_keep_dims_0 = const()[name = string("op_6069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6069_cast_fp16 = reduce_mean(axes = var_6069_axes_0, keep_dims = var_6069_keep_dims_0, x = var_6067_cast_fp16)[name = string("op_6069_cast_fp16")];
            fp16 var_6059_promoted_1_to_fp16 = const()[name = string("op_6059_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6070_cast_fp16 = pow(x = peak_113_cast_fp16, y = var_6059_promoted_1_to_fp16)[name = string("op_6070_cast_fp16")];
            fp32 var_6071_epsilon_0 = const()[name = string("op_6071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6071_cast_fp16 = inverse(epsilon = var_6071_epsilon_0, x = var_6070_cast_fp16)[name = string("op_6071_cast_fp16")];
            fp16 var_6072_to_fp16 = const()[name = string("op_6072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6073_cast_fp16 = mul(x = var_6071_cast_fp16, y = var_6072_to_fp16)[name = string("op_6073_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6074_cast_fp16 = add(x = var_6069_cast_fp16, y = var_6073_cast_fp16)[name = string("op_6074_cast_fp16")];
            fp32 var_6075_epsilon_0 = const()[name = string("op_6075_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6075_cast_fp16 = rsqrt(epsilon = var_6075_epsilon_0, x = var_6074_cast_fp16)[name = string("op_6075_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6076_cast_fp16 = mul(x = x_283_cast_fp16, y = var_6075_cast_fp16)[name = string("op_6076_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_6076_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_6085 = const()[name = string("op_6085"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6086_cast_fp16 = reshape(shape = var_6085, x = linear_98_cast_fp16)[name = string("op_6086_cast_fp16")];
            tensor<int32, [4]> x_285_perm_0 = const()[name = string("x_285_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6093 = const()[name = string("op_6093"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_285_cast_fp16 = transpose(perm = x_285_perm_0, x = var_6086_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 16, 128]> var_6096_cast_fp16 = abs(x = x_285_cast_fp16)[name = string("op_6096_cast_fp16")];
            tensor<int32, [1]> peak_115_axes_0 = const()[name = string("peak_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_115_cast_fp16 = reduce_max(axes = peak_115_axes_0, keep_dims = var_6093, x = var_6096_cast_fp16)[name = string("peak_115_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_287_cast_fp16 = real_div(x = x_285_cast_fp16, y = peak_115_cast_fp16)[name = string("x_287_cast_fp16")];
            fp16 var_6092_promoted_to_fp16 = const()[name = string("op_6092_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6100_cast_fp16 = pow(x = x_287_cast_fp16, y = var_6092_promoted_to_fp16)[name = string("op_6100_cast_fp16")];
            tensor<int32, [1]> var_6102_axes_0 = const()[name = string("op_6102_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6102_keep_dims_0 = const()[name = string("op_6102_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6102_cast_fp16 = reduce_mean(axes = var_6102_axes_0, keep_dims = var_6102_keep_dims_0, x = var_6100_cast_fp16)[name = string("op_6102_cast_fp16")];
            fp16 var_6092_promoted_1_to_fp16 = const()[name = string("op_6092_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6103_cast_fp16 = pow(x = peak_115_cast_fp16, y = var_6092_promoted_1_to_fp16)[name = string("op_6103_cast_fp16")];
            fp32 var_6104_epsilon_0 = const()[name = string("op_6104_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6104_cast_fp16 = inverse(epsilon = var_6104_epsilon_0, x = var_6103_cast_fp16)[name = string("op_6104_cast_fp16")];
            fp16 var_6105_to_fp16 = const()[name = string("op_6105_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6106_cast_fp16 = mul(x = var_6104_cast_fp16, y = var_6105_to_fp16)[name = string("op_6106_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6107_cast_fp16 = add(x = var_6102_cast_fp16, y = var_6106_cast_fp16)[name = string("op_6107_cast_fp16")];
            fp32 var_6108_epsilon_0 = const()[name = string("op_6108_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6108_cast_fp16 = rsqrt(epsilon = var_6108_epsilon_0, x = var_6107_cast_fp16)[name = string("op_6108_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6109_cast_fp16 = mul(x = x_287_cast_fp16, y = var_6108_cast_fp16)[name = string("op_6109_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_293_cast_fp16 = mul(x = var_6109_cast_fp16, y = model_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_293_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_6118 = const()[name = string("op_6118"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6119_cast_fp16 = reshape(shape = var_6118, x = linear_99_cast_fp16)[name = string("op_6119_cast_fp16")];
            tensor<int32, [4]> x_289_perm_0 = const()[name = string("x_289_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6126 = const()[name = string("op_6126"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_289_cast_fp16 = transpose(perm = x_289_perm_0, x = var_6119_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 16, 128]> var_6129_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_6129_cast_fp16")];
            tensor<int32, [1]> peak_117_axes_0 = const()[name = string("peak_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_117_cast_fp16 = reduce_max(axes = peak_117_axes_0, keep_dims = var_6126, x = var_6129_cast_fp16)[name = string("peak_117_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_291_cast_fp16 = real_div(x = x_289_cast_fp16, y = peak_117_cast_fp16)[name = string("x_291_cast_fp16")];
            fp16 var_6125_promoted_to_fp16 = const()[name = string("op_6125_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6133_cast_fp16 = pow(x = x_291_cast_fp16, y = var_6125_promoted_to_fp16)[name = string("op_6133_cast_fp16")];
            tensor<int32, [1]> var_6135_axes_0 = const()[name = string("op_6135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6135_keep_dims_0 = const()[name = string("op_6135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6135_cast_fp16 = reduce_mean(axes = var_6135_axes_0, keep_dims = var_6135_keep_dims_0, x = var_6133_cast_fp16)[name = string("op_6135_cast_fp16")];
            fp16 var_6125_promoted_1_to_fp16 = const()[name = string("op_6125_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6136_cast_fp16 = pow(x = peak_117_cast_fp16, y = var_6125_promoted_1_to_fp16)[name = string("op_6136_cast_fp16")];
            fp32 var_6137_epsilon_0 = const()[name = string("op_6137_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6137_cast_fp16 = inverse(epsilon = var_6137_epsilon_0, x = var_6136_cast_fp16)[name = string("op_6137_cast_fp16")];
            fp16 var_6138_to_fp16 = const()[name = string("op_6138_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> 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, 16, 1]> var_6140_cast_fp16 = add(x = var_6135_cast_fp16, y = var_6139_cast_fp16)[name = string("op_6140_cast_fp16")];
            fp32 var_6141_epsilon_0 = const()[name = string("op_6141_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6141_cast_fp16 = rsqrt(epsilon = var_6141_epsilon_0, x = var_6140_cast_fp16)[name = string("op_6141_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6142_cast_fp16 = mul(x = x_291_cast_fp16, y = var_6141_cast_fp16)[name = string("op_6142_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_295_cast_fp16 = mul(x = var_6142_cast_fp16, y = model_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_295_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_6151 = const()[name = string("op_6151"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6152_cast_fp16 = reshape(shape = var_6151, x = linear_100_cast_fp16)[name = string("op_6152_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_293_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_293_cast_fp16)[name = string("second_57_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6166_cast_fp16 = mul(x = first_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6166_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6167_cast_fp16 = mul(x = second_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6167_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6169_cast_fp16 = sub(x = var_6166_cast_fp16, y = var_6167_cast_fp16)[name = string("op_6169_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6170_cast_fp16 = mul(x = second_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6171_cast_fp16 = mul(x = first_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6171_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6173_cast_fp16 = add(x = var_6170_cast_fp16, y = var_6171_cast_fp16)[name = string("op_6173_cast_fp16")];
            int32 var_6175 = const()[name = string("op_6175"), 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_6175, interleave = q_29_interleave_0, values = (var_6169_cast_fp16, var_6173_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_295_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_295_cast_fp16)[name = string("second_59_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6187_cast_fp16 = mul(x = first_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6187_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6188_cast_fp16 = mul(x = second_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6190_cast_fp16 = sub(x = var_6187_cast_fp16, y = var_6188_cast_fp16)[name = string("op_6190_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6191_cast_fp16 = mul(x = second_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6191_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6192_cast_fp16 = mul(x = first_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6192_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6194_cast_fp16 = add(x = var_6191_cast_fp16, y = var_6192_cast_fp16)[name = string("op_6194_cast_fp16")];
            int32 var_6196 = const()[name = string("op_6196"), 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_6196, interleave = k_29_interleave_0, values = (var_6190_cast_fp16, var_6194_cast_fp16))[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_6207_begin_0 = const()[name = string("op_6207_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6207_end_0 = const()[name = string("op_6207_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6207_end_mask_0 = const()[name = string("op_6207_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6207_stride_0 = const()[name = string("op_6207_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6207_cast_fp16 = slice_by_index(begin = var_6207_begin_0, end = var_6207_end_0, end_mask = var_6207_end_mask_0, stride = var_6207_stride_0, x = q_29_cast_fp16)[name = string("op_6207_cast_fp16")];
            tensor<int32, [4]> var_6217_begin_0 = const()[name = string("op_6217_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6217_end_0 = const()[name = string("op_6217_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6217_end_mask_0 = const()[name = string("op_6217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6217_stride_0 = const()[name = string("op_6217_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6217_cast_fp16 = slice_by_index(begin = var_6217_begin_0, end = var_6217_end_0, end_mask = var_6217_end_mask_0, stride = var_6217_stride_0, x = q_29_cast_fp16)[name = string("op_6217_cast_fp16")];
            int32 var_6219 = const()[name = string("op_6219"), 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_6219, interleave = stacked_29_interleave_0, values = (var_6207_cast_fp16, var_6217_cast_fp16))[name = string("stacked_29_cast_fp16")];
            tensor<int32, [4]> var_6225_begin_0 = const()[name = string("op_6225_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_6225_end_0 = const()[name = string("op_6225_end_0"), val = tensor<int32, [4]>([15, 8, 2064, 128])];
            tensor<bool, [4]> var_6225_end_mask_0 = const()[name = string("op_6225_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6225_cast_fp16 = slice_by_index(begin = var_6225_begin_0, end = var_6225_end_0, end_mask = var_6225_end_mask_0, x = read_state_0)[name = string("op_6225_cast_fp16")];
            tensor<int32, [4]> var_6235_begin_0 = const()[name = string("op_6235_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6235_end_0 = const()[name = string("op_6235_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6235_end_mask_0 = const()[name = string("op_6235_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6235_cast_fp16 = slice_by_index(begin = var_6235_begin_0, end = var_6235_end_0, end_mask = var_6235_end_mask_0, x = var_6225_cast_fp16)[name = string("op_6235_cast_fp16")];
            bool var_6244_transpose_x_1 = const()[name = string("op_6244_transpose_x_1"), val = bool(false)];
            bool var_6244_transpose_y_1 = const()[name = string("op_6244_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_6244_cast_fp16 = matmul(transpose_x = var_6244_transpose_x_1, transpose_y = var_6244_transpose_y_1, x = var_6235_cast_fp16, y = stacked_29_cast_fp16)[name = string("op_6244_cast_fp16")];
            tensor<int32, [4]> var_6247_perm_0 = const()[name = string("op_6247_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6248_to_fp16 = const()[name = string("op_6248_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_6247_cast_fp16 = transpose(perm = var_6247_perm_0, x = var_6244_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 8, 32, 2048]> var_6249_cast_fp16 = mul(x = var_6247_cast_fp16, y = var_6248_to_fp16)[name = string("op_6249_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_29_cast_fp16 = add(x = var_6249_cast_fp16, y = cacheMask)[name = string("cache_weights_29_cast_fp16")];
            bool var_6255_transpose_x_1 = const()[name = string("op_6255_transpose_x_1"), val = bool(false)];
            bool var_6255_transpose_y_1 = const()[name = string("op_6255_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6255_cast_fp16 = matmul(transpose_x = var_6255_transpose_x_1, transpose_y = var_6255_transpose_y_1, x = stacked_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_6255_cast_fp16")];
            fp16 var_6256_to_fp16 = const()[name = string("op_6256_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6257_cast_fp16 = mul(x = var_6255_cast_fp16, y = var_6256_to_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_29_cast_fp16 = add(x = var_6257_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_29_cast_fp16")];
            int32 var_6261 = const()[name = string("op_6261"), val = int32(-1)];
            bool var_6262_interleave_0 = const()[name = string("op_6262_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_6262_cast_fp16 = concat(axis = var_6261, interleave = var_6262_interleave_0, values = (cache_weights_29_cast_fp16, call_weights_29_cast_fp16))[name = string("op_6262_cast_fp16")];
            int32 var_6263 = const()[name = string("op_6263"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_29_cast_fp16 = softmax(axis = var_6263, x = var_6262_cast_fp16)[name = string("shares_29_cast_fp16")];
            tensor<int32, [4]> var_6270_begin_0 = const()[name = string("op_6270_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6270_end_0 = const()[name = string("op_6270_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_6270_end_mask_0 = const()[name = string("op_6270_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_6270_cast_fp16 = slice_by_index(begin = var_6270_begin_0, end = var_6270_end_0, end_mask = var_6270_end_mask_0, x = shares_29_cast_fp16)[name = string("op_6270_cast_fp16")];
            tensor<int32, [4]> var_6275_begin_0 = const()[name = string("op_6275_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_6275_end_0 = const()[name = string("op_6275_end_0"), val = tensor<int32, [4]>([15, 8, 2064, 128])];
            tensor<bool, [4]> var_6275_end_mask_0 = const()[name = string("op_6275_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6275_cast_fp16 = slice_by_index(begin = var_6275_begin_0, end = var_6275_end_0, end_mask = var_6275_end_mask_0, x = read_state_1)[name = string("op_6275_cast_fp16")];
            tensor<int32, [4]> var_6285_begin_0 = const()[name = string("op_6285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6285_end_0 = const()[name = string("op_6285_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6285_end_mask_0 = const()[name = string("op_6285_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6285_cast_fp16 = slice_by_index(begin = var_6285_begin_0, end = var_6285_end_0, end_mask = var_6285_end_mask_0, x = var_6275_cast_fp16)[name = string("op_6285_cast_fp16")];
            bool var_6291_transpose_x_0 = const()[name = string("op_6291_transpose_x_0"), val = bool(false)];
            bool var_6291_transpose_y_0 = const()[name = string("op_6291_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6291_cast_fp16 = matmul(transpose_x = var_6291_transpose_x_0, transpose_y = var_6291_transpose_y_0, x = var_6270_cast_fp16, y = var_6285_cast_fp16)[name = string("op_6291_cast_fp16")];
            tensor<int32, [4]> var_6296_begin_0 = const()[name = string("op_6296_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_6296_end_0 = const()[name = string("op_6296_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_6296_end_mask_0 = const()[name = string("op_6296_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6296_cast_fp16 = slice_by_index(begin = var_6296_begin_0, end = var_6296_end_0, end_mask = var_6296_end_mask_0, x = shares_29_cast_fp16)[name = string("op_6296_cast_fp16")];
            bool var_6297_transpose_x_0 = const()[name = string("op_6297_transpose_x_0"), val = bool(false)];
            bool var_6297_transpose_y_0 = const()[name = string("op_6297_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_6152_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 8, 32, 128]> var_6297_cast_fp16 = matmul(transpose_x = var_6297_transpose_x_0, transpose_y = var_6297_transpose_y_0, x = var_6296_cast_fp16, y = v_29_cast_fp16)[name = string("op_6297_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_57_cast_fp16 = add(x = var_6291_cast_fp16, y = var_6297_cast_fp16)[name = string("mixed_57_cast_fp16")];
            tensor<int32, [4]> var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6314_cast_fp16 = slice_by_index(begin = var_6314_begin_0, end = var_6314_end_0, end_mask = var_6314_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_6314_cast_fp16")];
            tensor<int32, [4]> var_6329_begin_0 = const()[name = string("op_6329_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6329_end_0 = const()[name = string("op_6329_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6329_end_mask_0 = const()[name = string("op_6329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6329_cast_fp16 = slice_by_index(begin = var_6329_begin_0, end = var_6329_end_0, end_mask = var_6329_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_6329_cast_fp16")];
            int32 var_6331 = const()[name = string("op_6331"), 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_6331, interleave = mixed_59_interleave_0, values = (var_6314_cast_fp16, var_6329_cast_fp16))[name = string("mixed_59_cast_fp16")];
            tensor<int32, [4]> var_6335_perm_0 = const()[name = string("op_6335_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6339 = const()[name = string("op_6339"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6335_cast_fp16 = transpose(perm = var_6335_perm_0, x = mixed_59_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 2048]> input_143_cast_fp16 = reshape(shape = var_6339, x = var_6335_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_297_cast_fp16 = add(x = x_281_cast_fp16, y = linear_101_cast_fp16)[name = string("x_297_cast_fp16")];
            bool var_6349 = const()[name = string("op_6349"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6352_cast_fp16 = abs(x = x_297_cast_fp16)[name = string("op_6352_cast_fp16")];
            tensor<int32, [1]> peak_119_axes_0 = const()[name = string("peak_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_119_cast_fp16 = reduce_max(axes = peak_119_axes_0, keep_dims = var_6349, x = var_6352_cast_fp16)[name = string("peak_119_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_299_cast_fp16 = real_div(x = x_297_cast_fp16, y = peak_119_cast_fp16)[name = string("x_299_cast_fp16")];
            fp16 var_6348_promoted_to_fp16 = const()[name = string("op_6348_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6356_cast_fp16 = pow(x = x_299_cast_fp16, y = var_6348_promoted_to_fp16)[name = string("op_6356_cast_fp16")];
            tensor<int32, [1]> var_6358_axes_0 = const()[name = string("op_6358_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6358_keep_dims_0 = const()[name = string("op_6358_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6358_cast_fp16 = reduce_mean(axes = var_6358_axes_0, keep_dims = var_6358_keep_dims_0, x = var_6356_cast_fp16)[name = string("op_6358_cast_fp16")];
            fp16 var_6348_promoted_1_to_fp16 = const()[name = string("op_6348_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6359_cast_fp16 = pow(x = peak_119_cast_fp16, y = var_6348_promoted_1_to_fp16)[name = string("op_6359_cast_fp16")];
            fp32 var_6360_epsilon_0 = const()[name = string("op_6360_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6360_cast_fp16 = inverse(epsilon = var_6360_epsilon_0, x = var_6359_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6362_cast_fp16 = mul(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("op_6362_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6363_cast_fp16 = add(x = var_6358_cast_fp16, y = var_6362_cast_fp16)[name = string("op_6363_cast_fp16")];
            fp32 var_6364_epsilon_0 = const()[name = string("op_6364_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6364_cast_fp16 = rsqrt(epsilon = var_6364_epsilon_0, x = var_6363_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6365_cast_fp16 = mul(x = x_299_cast_fp16, y = var_6364_cast_fp16)[name = string("op_6365_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_6365_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_6370_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_6370_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_6370_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_301_cast_fp16 = add(x = x_297_cast_fp16, y = linear_104_cast_fp16)[name = string("x_301_cast_fp16")];
            bool var_6383 = const()[name = string("op_6383"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6386_cast_fp16 = abs(x = x_301_cast_fp16)[name = string("op_6386_cast_fp16")];
            tensor<int32, [1]> peak_121_axes_0 = const()[name = string("peak_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_121_cast_fp16 = reduce_max(axes = peak_121_axes_0, keep_dims = var_6383, x = var_6386_cast_fp16)[name = string("peak_121_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_303_cast_fp16 = real_div(x = x_301_cast_fp16, y = peak_121_cast_fp16)[name = string("x_303_cast_fp16")];
            fp16 var_6382_promoted_to_fp16 = const()[name = string("op_6382_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6390_cast_fp16 = pow(x = x_303_cast_fp16, y = var_6382_promoted_to_fp16)[name = string("op_6390_cast_fp16")];
            tensor<int32, [1]> var_6392_axes_0 = const()[name = string("op_6392_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6392_keep_dims_0 = const()[name = string("op_6392_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6392_cast_fp16 = reduce_mean(axes = var_6392_axes_0, keep_dims = var_6392_keep_dims_0, x = var_6390_cast_fp16)[name = string("op_6392_cast_fp16")];
            fp16 var_6382_promoted_1_to_fp16 = const()[name = string("op_6382_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6393_cast_fp16 = pow(x = peak_121_cast_fp16, y = var_6382_promoted_1_to_fp16)[name = string("op_6393_cast_fp16")];
            fp32 var_6394_epsilon_0 = const()[name = string("op_6394_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6394_cast_fp16 = inverse(epsilon = var_6394_epsilon_0, x = var_6393_cast_fp16)[name = string("op_6394_cast_fp16")];
            fp16 var_6395_to_fp16 = const()[name = string("op_6395_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6396_cast_fp16 = mul(x = var_6394_cast_fp16, y = var_6395_to_fp16)[name = string("op_6396_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6397_cast_fp16 = add(x = var_6392_cast_fp16, y = var_6396_cast_fp16)[name = string("op_6397_cast_fp16")];
            fp32 var_6398_epsilon_0 = const()[name = string("op_6398_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6398_cast_fp16 = rsqrt(epsilon = var_6398_epsilon_0, x = var_6397_cast_fp16)[name = string("op_6398_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6399_cast_fp16 = mul(x = x_303_cast_fp16, y = var_6398_cast_fp16)[name = string("op_6399_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_6399_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_6408 = const()[name = string("op_6408"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6409_cast_fp16 = reshape(shape = var_6408, x = linear_105_cast_fp16)[name = string("op_6409_cast_fp16")];
            tensor<int32, [4]> x_305_perm_0 = const()[name = string("x_305_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6416 = const()[name = string("op_6416"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_305_cast_fp16 = transpose(perm = x_305_perm_0, x = var_6409_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 16, 16, 128]> var_6419_cast_fp16 = abs(x = x_305_cast_fp16)[name = string("op_6419_cast_fp16")];
            tensor<int32, [1]> peak_123_axes_0 = const()[name = string("peak_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_123_cast_fp16 = reduce_max(axes = peak_123_axes_0, keep_dims = var_6416, x = var_6419_cast_fp16)[name = string("peak_123_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_307_cast_fp16 = real_div(x = x_305_cast_fp16, y = peak_123_cast_fp16)[name = string("x_307_cast_fp16")];
            fp16 var_6415_promoted_to_fp16 = const()[name = string("op_6415_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6423_cast_fp16 = pow(x = x_307_cast_fp16, y = var_6415_promoted_to_fp16)[name = string("op_6423_cast_fp16")];
            tensor<int32, [1]> var_6425_axes_0 = const()[name = string("op_6425_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6425_keep_dims_0 = const()[name = string("op_6425_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6425_cast_fp16 = reduce_mean(axes = var_6425_axes_0, keep_dims = var_6425_keep_dims_0, x = var_6423_cast_fp16)[name = string("op_6425_cast_fp16")];
            fp16 var_6415_promoted_1_to_fp16 = const()[name = string("op_6415_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6426_cast_fp16 = pow(x = peak_123_cast_fp16, y = var_6415_promoted_1_to_fp16)[name = string("op_6426_cast_fp16")];
            fp32 var_6427_epsilon_0 = const()[name = string("op_6427_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6427_cast_fp16 = inverse(epsilon = var_6427_epsilon_0, x = var_6426_cast_fp16)[name = string("op_6427_cast_fp16")];
            fp16 var_6428_to_fp16 = const()[name = string("op_6428_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6429_cast_fp16 = mul(x = var_6427_cast_fp16, y = var_6428_to_fp16)[name = string("op_6429_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6430_cast_fp16 = add(x = var_6425_cast_fp16, y = var_6429_cast_fp16)[name = string("op_6430_cast_fp16")];
            fp32 var_6431_epsilon_0 = const()[name = string("op_6431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6431_cast_fp16 = rsqrt(epsilon = var_6431_epsilon_0, x = var_6430_cast_fp16)[name = string("op_6431_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6432_cast_fp16 = mul(x = x_307_cast_fp16, y = var_6431_cast_fp16)[name = string("op_6432_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_313_cast_fp16 = mul(x = var_6432_cast_fp16, y = model_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_313_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_6441 = const()[name = string("op_6441"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6442_cast_fp16 = reshape(shape = var_6441, x = linear_106_cast_fp16)[name = string("op_6442_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6449 = const()[name = string("op_6449"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_6442_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 8, 16, 128]> var_6452_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_6452_cast_fp16")];
            tensor<int32, [1]> peak_125_axes_0 = const()[name = string("peak_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_125_cast_fp16 = reduce_max(axes = peak_125_axes_0, keep_dims = var_6449, x = var_6452_cast_fp16)[name = string("peak_125_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_311_cast_fp16 = real_div(x = x_309_cast_fp16, y = peak_125_cast_fp16)[name = string("x_311_cast_fp16")];
            fp16 var_6448_promoted_to_fp16 = const()[name = string("op_6448_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6456_cast_fp16 = pow(x = x_311_cast_fp16, y = var_6448_promoted_to_fp16)[name = string("op_6456_cast_fp16")];
            tensor<int32, [1]> var_6458_axes_0 = const()[name = string("op_6458_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6458_keep_dims_0 = const()[name = string("op_6458_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6458_cast_fp16 = reduce_mean(axes = var_6458_axes_0, keep_dims = var_6458_keep_dims_0, x = var_6456_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6448_promoted_1_to_fp16 = const()[name = string("op_6448_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6459_cast_fp16 = pow(x = peak_125_cast_fp16, y = var_6448_promoted_1_to_fp16)[name = string("op_6459_cast_fp16")];
            fp32 var_6460_epsilon_0 = const()[name = string("op_6460_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6460_cast_fp16 = inverse(epsilon = var_6460_epsilon_0, x = var_6459_cast_fp16)[name = string("op_6460_cast_fp16")];
            fp16 var_6461_to_fp16 = const()[name = string("op_6461_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6462_cast_fp16 = mul(x = var_6460_cast_fp16, y = var_6461_to_fp16)[name = string("op_6462_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_6463_cast_fp16 = add(x = var_6458_cast_fp16, y = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            fp32 var_6464_epsilon_0 = const()[name = string("op_6464_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6464_cast_fp16 = rsqrt(epsilon = var_6464_epsilon_0, x = var_6463_cast_fp16)[name = string("op_6464_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6465_cast_fp16 = mul(x = x_311_cast_fp16, y = var_6464_cast_fp16)[name = string("op_6465_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_315_cast_fp16 = mul(x = var_6465_cast_fp16, y = model_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_315_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_6474 = const()[name = string("op_6474"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6475_cast_fp16 = reshape(shape = var_6474, x = linear_107_cast_fp16)[name = string("op_6475_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_313_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_313_cast_fp16)[name = string("second_61_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6489_cast_fp16 = mul(x = first_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6490_cast_fp16 = mul(x = second_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6490_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6492_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = string("op_6492_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6493_cast_fp16 = mul(x = second_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6493_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6494_cast_fp16 = mul(x = first_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6494_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6496_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = string("op_6496_cast_fp16")];
            int32 var_6498 = const()[name = string("op_6498"), 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_6498, interleave = q_31_interleave_0, values = (var_6492_cast_fp16, var_6496_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_315_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_315_cast_fp16)[name = string("second_63_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6510_cast_fp16 = mul(x = first_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6510_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6511_cast_fp16 = mul(x = second_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6511_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6513_cast_fp16 = sub(x = var_6510_cast_fp16, y = var_6511_cast_fp16)[name = string("op_6513_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6514_cast_fp16 = mul(x = second_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6514_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6515_cast_fp16 = mul(x = first_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6515_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6517_cast_fp16 = add(x = var_6514_cast_fp16, y = var_6515_cast_fp16)[name = string("op_6517_cast_fp16")];
            int32 var_6519 = const()[name = string("op_6519"), 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_6519, interleave = k_31_interleave_0, values = (var_6513_cast_fp16, var_6517_cast_fp16))[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> var_6530_begin_0 = const()[name = string("op_6530_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6530_end_0 = const()[name = string("op_6530_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6530_end_mask_0 = const()[name = string("op_6530_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6530_stride_0 = const()[name = string("op_6530_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6530_cast_fp16 = slice_by_index(begin = var_6530_begin_0, end = var_6530_end_0, end_mask = var_6530_end_mask_0, stride = var_6530_stride_0, x = q_31_cast_fp16)[name = string("op_6530_cast_fp16")];
            tensor<int32, [4]> var_6540_begin_0 = const()[name = string("op_6540_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6540_end_0 = const()[name = string("op_6540_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6540_end_mask_0 = const()[name = string("op_6540_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6540_stride_0 = const()[name = string("op_6540_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6540_cast_fp16 = slice_by_index(begin = var_6540_begin_0, end = var_6540_end_0, end_mask = var_6540_end_mask_0, stride = var_6540_stride_0, x = q_31_cast_fp16)[name = string("op_6540_cast_fp16")];
            int32 var_6542 = const()[name = string("op_6542"), 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_6542, interleave = stacked_31_interleave_0, values = (var_6530_cast_fp16, var_6540_cast_fp16))[name = string("stacked_31_cast_fp16")];
            tensor<int32, [4]> var_6548_begin_0 = const()[name = string("op_6548_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6548_end_0 = const()[name = string("op_6548_end_0"), val = tensor<int32, [4]>([16, 8, 2064, 128])];
            tensor<bool, [4]> var_6548_end_mask_0 = const()[name = string("op_6548_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6548_cast_fp16 = slice_by_index(begin = var_6548_begin_0, end = var_6548_end_0, end_mask = var_6548_end_mask_0, x = read_state_0)[name = string("op_6548_cast_fp16")];
            tensor<int32, [4]> var_6558_begin_0 = const()[name = string("op_6558_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6558_end_0 = const()[name = string("op_6558_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6558_end_mask_0 = const()[name = string("op_6558_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6558_cast_fp16 = slice_by_index(begin = var_6558_begin_0, end = var_6558_end_0, end_mask = var_6558_end_mask_0, x = var_6548_cast_fp16)[name = string("op_6558_cast_fp16")];
            bool var_6567_transpose_x_1 = const()[name = string("op_6567_transpose_x_1"), val = bool(false)];
            bool var_6567_transpose_y_1 = const()[name = string("op_6567_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_6567_cast_fp16 = matmul(transpose_x = var_6567_transpose_x_1, transpose_y = var_6567_transpose_y_1, x = var_6558_cast_fp16, y = stacked_31_cast_fp16)[name = string("op_6567_cast_fp16")];
            tensor<int32, [4]> var_6570_perm_0 = const()[name = string("op_6570_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6571_to_fp16 = const()[name = string("op_6571_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_6570_cast_fp16 = transpose(perm = var_6570_perm_0, x = var_6567_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, 32, 2048]> var_6572_cast_fp16 = mul(x = var_6570_cast_fp16, y = var_6571_to_fp16)[name = string("op_6572_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_31_cast_fp16 = add(x = var_6572_cast_fp16, y = cacheMask)[name = string("cache_weights_31_cast_fp16")];
            bool var_6578_transpose_x_1 = const()[name = string("op_6578_transpose_x_1"), val = bool(false)];
            bool var_6578_transpose_y_1 = const()[name = string("op_6578_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6578_cast_fp16 = matmul(transpose_x = var_6578_transpose_x_1, transpose_y = var_6578_transpose_y_1, x = stacked_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6580_cast_fp16 = mul(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("op_6580_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_31_cast_fp16 = add(x = var_6580_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_31_cast_fp16")];
            int32 var_6584 = const()[name = string("op_6584"), val = int32(-1)];
            bool var_6585_interleave_0 = const()[name = string("op_6585_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_6585_cast_fp16 = concat(axis = var_6584, interleave = var_6585_interleave_0, values = (cache_weights_31_cast_fp16, call_weights_31_cast_fp16))[name = string("op_6585_cast_fp16")];
            int32 var_6586 = const()[name = string("op_6586"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_31_cast_fp16 = softmax(axis = var_6586, x = var_6585_cast_fp16)[name = string("shares_31_cast_fp16")];
            tensor<int32, [4]> var_6593_begin_0 = const()[name = string("op_6593_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6593_end_0 = const()[name = string("op_6593_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_6593_end_mask_0 = const()[name = string("op_6593_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_6593_cast_fp16 = slice_by_index(begin = var_6593_begin_0, end = var_6593_end_0, end_mask = var_6593_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6593_cast_fp16")];
            tensor<int32, [4]> var_6598_begin_0 = const()[name = string("op_6598_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6598_end_0 = const()[name = string("op_6598_end_0"), val = tensor<int32, [4]>([16, 8, 2064, 128])];
            tensor<bool, [4]> var_6598_end_mask_0 = const()[name = string("op_6598_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6598_cast_fp16 = slice_by_index(begin = var_6598_begin_0, end = var_6598_end_0, end_mask = var_6598_end_mask_0, x = read_state_1)[name = string("op_6598_cast_fp16")];
            tensor<int32, [4]> var_6608_begin_0 = const()[name = string("op_6608_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6608_end_0 = const()[name = string("op_6608_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6608_end_mask_0 = const()[name = string("op_6608_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6608_cast_fp16 = slice_by_index(begin = var_6608_begin_0, end = var_6608_end_0, end_mask = var_6608_end_mask_0, x = var_6598_cast_fp16)[name = string("op_6608_cast_fp16")];
            bool var_6614_transpose_x_0 = const()[name = string("op_6614_transpose_x_0"), val = bool(false)];
            bool var_6614_transpose_y_0 = const()[name = string("op_6614_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6614_cast_fp16 = matmul(transpose_x = var_6614_transpose_x_0, transpose_y = var_6614_transpose_y_0, x = var_6593_cast_fp16, y = var_6608_cast_fp16)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6619_cast_fp16")];
            bool var_6620_transpose_x_0 = const()[name = string("op_6620_transpose_x_0"), val = bool(false)];
            bool var_6620_transpose_y_0 = const()[name = string("op_6620_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_6475_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 32, 128]> var_6620_cast_fp16 = matmul(transpose_x = var_6620_transpose_x_0, transpose_y = var_6620_transpose_y_0, x = var_6619_cast_fp16, y = v_31_cast_fp16)[name = string("op_6620_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_61_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6620_cast_fp16)[name = string("mixed_61_cast_fp16")];
            tensor<int32, [4]> var_6637_begin_0 = const()[name = string("op_6637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6637_end_0 = const()[name = string("op_6637_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6637_end_mask_0 = const()[name = string("op_6637_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6637_cast_fp16 = slice_by_index(begin = var_6637_begin_0, end = var_6637_end_0, end_mask = var_6637_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6637_cast_fp16")];
            tensor<int32, [4]> var_6652_begin_0 = const()[name = string("op_6652_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6652_end_0 = const()[name = string("op_6652_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6652_end_mask_0 = const()[name = string("op_6652_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6652_cast_fp16 = slice_by_index(begin = var_6652_begin_0, end = var_6652_end_0, end_mask = var_6652_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6652_cast_fp16")];
            int32 var_6654 = const()[name = string("op_6654"), 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_6654, interleave = mixed_63_interleave_0, values = (var_6637_cast_fp16, var_6652_cast_fp16))[name = string("mixed_63_cast_fp16")];
            tensor<int32, [4]> var_6658_perm_0 = const()[name = string("op_6658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6662 = const()[name = string("op_6662"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6658_cast_fp16 = transpose(perm = var_6658_perm_0, x = mixed_63_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 16, 2048]> input_153_cast_fp16 = reshape(shape = var_6662, x = var_6658_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_317_cast_fp16 = add(x = x_301_cast_fp16, y = linear_108_cast_fp16)[name = string("x_317_cast_fp16")];
            bool var_6672 = const()[name = string("op_6672"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6675_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_6675_cast_fp16")];
            tensor<int32, [1]> peak_127_axes_0 = const()[name = string("peak_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_127_cast_fp16 = reduce_max(axes = peak_127_axes_0, keep_dims = var_6672, x = var_6675_cast_fp16)[name = string("peak_127_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_319_cast_fp16 = real_div(x = x_317_cast_fp16, y = peak_127_cast_fp16)[name = string("x_319_cast_fp16")];
            fp16 var_6671_promoted_to_fp16 = const()[name = string("op_6671_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6679_cast_fp16 = pow(x = x_319_cast_fp16, y = var_6671_promoted_to_fp16)[name = string("op_6679_cast_fp16")];
            tensor<int32, [1]> var_6681_axes_0 = const()[name = string("op_6681_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6681_keep_dims_0 = const()[name = string("op_6681_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6681_cast_fp16 = reduce_mean(axes = var_6681_axes_0, keep_dims = var_6681_keep_dims_0, x = var_6679_cast_fp16)[name = string("op_6681_cast_fp16")];
            fp16 var_6671_promoted_1_to_fp16 = const()[name = string("op_6671_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6682_cast_fp16 = pow(x = peak_127_cast_fp16, y = var_6671_promoted_1_to_fp16)[name = string("op_6682_cast_fp16")];
            fp32 var_6683_epsilon_0 = const()[name = string("op_6683_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6683_cast_fp16 = inverse(epsilon = var_6683_epsilon_0, x = var_6682_cast_fp16)[name = string("op_6683_cast_fp16")];
            fp16 var_6684_to_fp16 = const()[name = string("op_6684_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6685_cast_fp16 = mul(x = var_6683_cast_fp16, y = var_6684_to_fp16)[name = string("op_6685_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6686_cast_fp16 = add(x = var_6681_cast_fp16, y = var_6685_cast_fp16)[name = string("op_6686_cast_fp16")];
            fp32 var_6687_epsilon_0 = const()[name = string("op_6687_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6687_cast_fp16 = rsqrt(epsilon = var_6687_epsilon_0, x = var_6686_cast_fp16)[name = string("op_6687_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6688_cast_fp16 = mul(x = x_319_cast_fp16, y = var_6687_cast_fp16)[name = string("op_6688_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_6688_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_6693_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_6693_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_6693_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_321_cast_fp16 = add(x = x_317_cast_fp16, y = linear_111_cast_fp16)[name = string("x_321_cast_fp16")];
            bool var_6706 = const()[name = string("op_6706"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6709_cast_fp16 = abs(x = x_321_cast_fp16)[name = string("op_6709_cast_fp16")];
            tensor<int32, [1]> peak_129_axes_0 = const()[name = string("peak_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_129_cast_fp16 = reduce_max(axes = peak_129_axes_0, keep_dims = var_6706, x = var_6709_cast_fp16)[name = string("peak_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_323_cast_fp16 = real_div(x = x_321_cast_fp16, y = peak_129_cast_fp16)[name = string("x_323_cast_fp16")];
            fp16 var_6705_promoted_to_fp16 = const()[name = string("op_6705_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6713_cast_fp16 = pow(x = x_323_cast_fp16, y = var_6705_promoted_to_fp16)[name = string("op_6713_cast_fp16")];
            tensor<int32, [1]> var_6715_axes_0 = const()[name = string("op_6715_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6715_keep_dims_0 = const()[name = string("op_6715_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6715_cast_fp16 = reduce_mean(axes = var_6715_axes_0, keep_dims = var_6715_keep_dims_0, x = var_6713_cast_fp16)[name = string("op_6715_cast_fp16")];
            fp16 var_6705_promoted_1_to_fp16 = const()[name = string("op_6705_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6716_cast_fp16 = pow(x = peak_129_cast_fp16, y = var_6705_promoted_1_to_fp16)[name = string("op_6716_cast_fp16")];
            fp32 var_6717_epsilon_0 = const()[name = string("op_6717_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6717_cast_fp16 = inverse(epsilon = var_6717_epsilon_0, x = var_6716_cast_fp16)[name = string("op_6717_cast_fp16")];
            fp16 var_6718_to_fp16 = const()[name = string("op_6718_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6719_cast_fp16 = mul(x = var_6717_cast_fp16, y = var_6718_to_fp16)[name = string("op_6719_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6720_cast_fp16 = add(x = var_6715_cast_fp16, y = var_6719_cast_fp16)[name = string("op_6720_cast_fp16")];
            fp32 var_6721_epsilon_0 = const()[name = string("op_6721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6721_cast_fp16 = rsqrt(epsilon = var_6721_epsilon_0, x = var_6720_cast_fp16)[name = string("op_6721_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6722_cast_fp16 = mul(x = x_323_cast_fp16, y = var_6721_cast_fp16)[name = string("op_6722_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_6722_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_6731 = const()[name = string("op_6731"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6732_cast_fp16 = reshape(shape = var_6731, x = linear_112_cast_fp16)[name = string("op_6732_cast_fp16")];
            tensor<int32, [4]> x_325_perm_0 = const()[name = string("x_325_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6739 = const()[name = string("op_6739"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_325_cast_fp16 = transpose(perm = x_325_perm_0, x = var_6732_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 16, 128]> var_6742_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_6742_cast_fp16")];
            tensor<int32, [1]> peak_131_axes_0 = const()[name = string("peak_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_131_cast_fp16 = reduce_max(axes = peak_131_axes_0, keep_dims = var_6739, x = var_6742_cast_fp16)[name = string("peak_131_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_327_cast_fp16 = real_div(x = x_325_cast_fp16, y = peak_131_cast_fp16)[name = string("x_327_cast_fp16")];
            fp16 var_6738_promoted_to_fp16 = const()[name = string("op_6738_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6746_cast_fp16 = pow(x = x_327_cast_fp16, y = var_6738_promoted_to_fp16)[name = string("op_6746_cast_fp16")];
            tensor<int32, [1]> var_6748_axes_0 = const()[name = string("op_6748_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6748_keep_dims_0 = const()[name = string("op_6748_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6748_cast_fp16 = reduce_mean(axes = var_6748_axes_0, keep_dims = var_6748_keep_dims_0, x = var_6746_cast_fp16)[name = string("op_6748_cast_fp16")];
            fp16 var_6738_promoted_1_to_fp16 = const()[name = string("op_6738_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6749_cast_fp16 = pow(x = peak_131_cast_fp16, y = var_6738_promoted_1_to_fp16)[name = string("op_6749_cast_fp16")];
            fp32 var_6750_epsilon_0 = const()[name = string("op_6750_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6750_cast_fp16 = inverse(epsilon = var_6750_epsilon_0, x = var_6749_cast_fp16)[name = string("op_6750_cast_fp16")];
            fp16 var_6751_to_fp16 = const()[name = string("op_6751_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6752_cast_fp16 = mul(x = var_6750_cast_fp16, y = var_6751_to_fp16)[name = string("op_6752_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6753_cast_fp16 = add(x = var_6748_cast_fp16, y = var_6752_cast_fp16)[name = string("op_6753_cast_fp16")];
            fp32 var_6754_epsilon_0 = const()[name = string("op_6754_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6754_cast_fp16 = rsqrt(epsilon = var_6754_epsilon_0, x = var_6753_cast_fp16)[name = string("op_6754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6755_cast_fp16 = mul(x = x_327_cast_fp16, y = var_6754_cast_fp16)[name = string("op_6755_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_333_cast_fp16 = mul(x = var_6755_cast_fp16, y = model_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("x_333_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_6764 = const()[name = string("op_6764"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6765_cast_fp16 = reshape(shape = var_6764, x = linear_113_cast_fp16)[name = string("op_6765_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6772 = const()[name = string("op_6772"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_6765_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 16, 128]> var_6775_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_6775_cast_fp16")];
            tensor<int32, [1]> peak_133_axes_0 = const()[name = string("peak_133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_133_cast_fp16 = reduce_max(axes = peak_133_axes_0, keep_dims = var_6772, x = var_6775_cast_fp16)[name = string("peak_133_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_331_cast_fp16 = real_div(x = x_329_cast_fp16, y = peak_133_cast_fp16)[name = string("x_331_cast_fp16")];
            fp16 var_6771_promoted_to_fp16 = const()[name = string("op_6771_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6779_cast_fp16 = pow(x = x_331_cast_fp16, y = var_6771_promoted_to_fp16)[name = string("op_6779_cast_fp16")];
            tensor<int32, [1]> var_6781_axes_0 = const()[name = string("op_6781_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6781_keep_dims_0 = const()[name = string("op_6781_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6781_cast_fp16 = reduce_mean(axes = var_6781_axes_0, keep_dims = var_6781_keep_dims_0, x = var_6779_cast_fp16)[name = string("op_6781_cast_fp16")];
            fp16 var_6771_promoted_1_to_fp16 = const()[name = string("op_6771_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6782_cast_fp16 = pow(x = peak_133_cast_fp16, y = var_6771_promoted_1_to_fp16)[name = string("op_6782_cast_fp16")];
            fp32 var_6783_epsilon_0 = const()[name = string("op_6783_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6783_cast_fp16 = inverse(epsilon = var_6783_epsilon_0, x = var_6782_cast_fp16)[name = string("op_6783_cast_fp16")];
            fp16 var_6784_to_fp16 = const()[name = string("op_6784_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6785_cast_fp16 = mul(x = var_6783_cast_fp16, y = var_6784_to_fp16)[name = string("op_6785_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_6786_cast_fp16 = add(x = var_6781_cast_fp16, y = var_6785_cast_fp16)[name = string("op_6786_cast_fp16")];
            fp32 var_6787_epsilon_0 = const()[name = string("op_6787_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6787_cast_fp16 = rsqrt(epsilon = var_6787_epsilon_0, x = var_6786_cast_fp16)[name = string("op_6787_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6788_cast_fp16 = mul(x = x_331_cast_fp16, y = var_6787_cast_fp16)[name = string("op_6788_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_335_cast_fp16 = mul(x = var_6788_cast_fp16, y = model_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("x_335_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_6797 = const()[name = string("op_6797"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6798_cast_fp16 = reshape(shape = var_6797, x = linear_114_cast_fp16)[name = string("op_6798_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_333_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_333_cast_fp16)[name = string("second_65_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6812_cast_fp16 = mul(x = first_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6813_cast_fp16 = mul(x = second_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6813_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6815_cast_fp16 = sub(x = var_6812_cast_fp16, y = var_6813_cast_fp16)[name = string("op_6815_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6816_cast_fp16 = mul(x = second_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6817_cast_fp16 = mul(x = first_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6817_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6819_cast_fp16 = add(x = var_6816_cast_fp16, y = var_6817_cast_fp16)[name = string("op_6819_cast_fp16")];
            int32 var_6821 = const()[name = string("op_6821"), 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_6821, interleave = q_33_interleave_0, values = (var_6815_cast_fp16, var_6819_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_335_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_335_cast_fp16)[name = string("second_67_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6833_cast_fp16 = mul(x = first_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6833_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6834_cast_fp16 = mul(x = second_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6834_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6836_cast_fp16 = sub(x = var_6833_cast_fp16, y = var_6834_cast_fp16)[name = string("op_6836_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6837_cast_fp16 = mul(x = second_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6837_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6838_cast_fp16 = mul(x = first_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6838_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6840_cast_fp16 = add(x = var_6837_cast_fp16, y = var_6838_cast_fp16)[name = string("op_6840_cast_fp16")];
            int32 var_6842 = const()[name = string("op_6842"), 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_6842, interleave = k_33_interleave_0, values = (var_6836_cast_fp16, var_6840_cast_fp16))[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_6853_begin_0 = const()[name = string("op_6853_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6853_end_0 = const()[name = string("op_6853_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6853_end_mask_0 = const()[name = string("op_6853_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6853_stride_0 = const()[name = string("op_6853_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6853_cast_fp16 = slice_by_index(begin = var_6853_begin_0, end = var_6853_end_0, end_mask = var_6853_end_mask_0, stride = var_6853_stride_0, x = q_33_cast_fp16)[name = string("op_6853_cast_fp16")];
            tensor<int32, [4]> var_6863_begin_0 = const()[name = string("op_6863_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6863_end_0 = const()[name = string("op_6863_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6863_end_mask_0 = const()[name = string("op_6863_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6863_stride_0 = const()[name = string("op_6863_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6863_cast_fp16 = slice_by_index(begin = var_6863_begin_0, end = var_6863_end_0, end_mask = var_6863_end_mask_0, stride = var_6863_stride_0, x = q_33_cast_fp16)[name = string("op_6863_cast_fp16")];
            int32 var_6865 = const()[name = string("op_6865"), 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_6865, interleave = stacked_33_interleave_0, values = (var_6853_cast_fp16, var_6863_cast_fp16))[name = string("stacked_33_cast_fp16")];
            tensor<int32, [4]> var_6871_begin_0 = const()[name = string("op_6871_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6871_end_0 = const()[name = string("op_6871_end_0"), val = tensor<int32, [4]>([17, 8, 2064, 128])];
            tensor<bool, [4]> var_6871_end_mask_0 = const()[name = string("op_6871_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6871_cast_fp16 = slice_by_index(begin = var_6871_begin_0, end = var_6871_end_0, end_mask = var_6871_end_mask_0, x = read_state_0)[name = string("op_6871_cast_fp16")];
            tensor<int32, [4]> var_6881_begin_0 = const()[name = string("op_6881_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6881_end_0 = const()[name = string("op_6881_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6881_end_mask_0 = const()[name = string("op_6881_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6881_cast_fp16 = slice_by_index(begin = var_6881_begin_0, end = var_6881_end_0, end_mask = var_6881_end_mask_0, x = var_6871_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6890_transpose_x_1 = const()[name = string("op_6890_transpose_x_1"), val = bool(false)];
            bool var_6890_transpose_y_1 = const()[name = string("op_6890_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_6890_cast_fp16 = matmul(transpose_x = var_6890_transpose_x_1, transpose_y = var_6890_transpose_y_1, x = var_6881_cast_fp16, y = stacked_33_cast_fp16)[name = string("op_6890_cast_fp16")];
            tensor<int32, [4]> var_6893_perm_0 = const()[name = string("op_6893_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6894_to_fp16 = const()[name = string("op_6894_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_6893_cast_fp16 = transpose(perm = var_6893_perm_0, x = var_6890_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 32, 2048]> var_6895_cast_fp16 = mul(x = var_6893_cast_fp16, y = var_6894_to_fp16)[name = string("op_6895_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_33_cast_fp16 = add(x = var_6895_cast_fp16, y = cacheMask)[name = string("cache_weights_33_cast_fp16")];
            bool var_6901_transpose_x_1 = const()[name = string("op_6901_transpose_x_1"), val = bool(false)];
            bool var_6901_transpose_y_1 = const()[name = string("op_6901_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6901_cast_fp16 = matmul(transpose_x = var_6901_transpose_x_1, transpose_y = var_6901_transpose_y_1, x = stacked_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_6901_cast_fp16")];
            fp16 var_6902_to_fp16 = const()[name = string("op_6902_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6903_cast_fp16 = mul(x = var_6901_cast_fp16, y = var_6902_to_fp16)[name = string("op_6903_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_33_cast_fp16 = add(x = var_6903_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_33_cast_fp16")];
            int32 var_6907 = const()[name = string("op_6907"), val = int32(-1)];
            bool var_6908_interleave_0 = const()[name = string("op_6908_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_6908_cast_fp16 = concat(axis = var_6907, interleave = var_6908_interleave_0, values = (cache_weights_33_cast_fp16, call_weights_33_cast_fp16))[name = string("op_6908_cast_fp16")];
            int32 var_6909 = const()[name = string("op_6909"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_33_cast_fp16 = softmax(axis = var_6909, x = var_6908_cast_fp16)[name = string("shares_33_cast_fp16")];
            tensor<int32, [4]> var_6916_begin_0 = const()[name = string("op_6916_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6916_end_0 = const()[name = string("op_6916_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_6916_end_mask_0 = const()[name = string("op_6916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_6916_cast_fp16 = slice_by_index(begin = var_6916_begin_0, end = var_6916_end_0, end_mask = var_6916_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6916_cast_fp16")];
            tensor<int32, [4]> var_6921_begin_0 = const()[name = string("op_6921_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6921_end_0 = const()[name = string("op_6921_end_0"), val = tensor<int32, [4]>([17, 8, 2064, 128])];
            tensor<bool, [4]> var_6921_end_mask_0 = const()[name = string("op_6921_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6921_cast_fp16 = slice_by_index(begin = var_6921_begin_0, end = var_6921_end_0, end_mask = var_6921_end_mask_0, x = read_state_1)[name = string("op_6921_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = var_6921_cast_fp16)[name = string("op_6931_cast_fp16")];
            bool var_6937_transpose_x_0 = const()[name = string("op_6937_transpose_x_0"), val = bool(false)];
            bool var_6937_transpose_y_0 = const()[name = string("op_6937_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6937_cast_fp16 = matmul(transpose_x = var_6937_transpose_x_0, transpose_y = var_6937_transpose_y_0, x = var_6916_cast_fp16, y = var_6931_cast_fp16)[name = string("op_6937_cast_fp16")];
            tensor<int32, [4]> var_6942_begin_0 = const()[name = string("op_6942_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_6942_end_0 = const()[name = string("op_6942_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_6942_end_mask_0 = const()[name = string("op_6942_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6942_cast_fp16 = slice_by_index(begin = var_6942_begin_0, end = var_6942_end_0, end_mask = var_6942_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6942_cast_fp16")];
            bool var_6943_transpose_x_0 = const()[name = string("op_6943_transpose_x_0"), val = bool(false)];
            bool var_6943_transpose_y_0 = const()[name = string("op_6943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_6798_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 8, 32, 128]> var_6943_cast_fp16 = matmul(transpose_x = var_6943_transpose_x_0, transpose_y = var_6943_transpose_y_0, x = var_6942_cast_fp16, y = v_33_cast_fp16)[name = string("op_6943_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_65_cast_fp16 = add(x = var_6937_cast_fp16, y = var_6943_cast_fp16)[name = string("mixed_65_cast_fp16")];
            tensor<int32, [4]> var_6960_begin_0 = const()[name = string("op_6960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6960_end_0 = const()[name = string("op_6960_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6960_end_mask_0 = const()[name = string("op_6960_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6960_cast_fp16 = slice_by_index(begin = var_6960_begin_0, end = var_6960_end_0, end_mask = var_6960_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6960_cast_fp16")];
            tensor<int32, [4]> var_6975_begin_0 = const()[name = string("op_6975_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6975_end_0 = const()[name = string("op_6975_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6975_end_mask_0 = const()[name = string("op_6975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6975_cast_fp16 = slice_by_index(begin = var_6975_begin_0, end = var_6975_end_0, end_mask = var_6975_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6977 = const()[name = string("op_6977"), 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_6977, interleave = mixed_67_interleave_0, values = (var_6960_cast_fp16, var_6975_cast_fp16))[name = string("mixed_67_cast_fp16")];
            tensor<int32, [4]> var_6981_perm_0 = const()[name = string("op_6981_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6985 = const()[name = string("op_6985"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6981_cast_fp16 = transpose(perm = var_6981_perm_0, x = mixed_67_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 2048]> input_163_cast_fp16 = reshape(shape = var_6985, x = var_6981_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_337_cast_fp16 = add(x = x_321_cast_fp16, y = linear_115_cast_fp16)[name = string("x_337_cast_fp16")];
            bool var_6995 = const()[name = string("op_6995"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6998_cast_fp16 = abs(x = x_337_cast_fp16)[name = string("op_6998_cast_fp16")];
            tensor<int32, [1]> peak_135_axes_0 = const()[name = string("peak_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_135_cast_fp16 = reduce_max(axes = peak_135_axes_0, keep_dims = var_6995, x = var_6998_cast_fp16)[name = string("peak_135_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_339_cast_fp16 = real_div(x = x_337_cast_fp16, y = peak_135_cast_fp16)[name = string("x_339_cast_fp16")];
            fp16 var_6994_promoted_to_fp16 = const()[name = string("op_6994_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7002_cast_fp16 = pow(x = x_339_cast_fp16, y = var_6994_promoted_to_fp16)[name = string("op_7002_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_7002_cast_fp16)[name = string("op_7004_cast_fp16")];
            fp16 var_6994_promoted_1_to_fp16 = const()[name = string("op_6994_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7005_cast_fp16 = pow(x = peak_135_cast_fp16, y = var_6994_promoted_1_to_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7009_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7009_cast_fp16")];
            fp32 var_7010_epsilon_0 = const()[name = string("op_7010_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7010_cast_fp16 = rsqrt(epsilon = var_7010_epsilon_0, x = var_7009_cast_fp16)[name = string("op_7010_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7011_cast_fp16 = mul(x = x_339_cast_fp16, y = var_7010_cast_fp16)[name = string("op_7011_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_7011_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_7016_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_7016_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_7016_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_341_cast_fp16 = add(x = x_337_cast_fp16, y = linear_118_cast_fp16)[name = string("x_341_cast_fp16")];
            bool var_7029 = const()[name = string("op_7029"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7032_cast_fp16 = abs(x = x_341_cast_fp16)[name = string("op_7032_cast_fp16")];
            tensor<int32, [1]> peak_137_axes_0 = const()[name = string("peak_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_137_cast_fp16 = reduce_max(axes = peak_137_axes_0, keep_dims = var_7029, x = var_7032_cast_fp16)[name = string("peak_137_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_343_cast_fp16 = real_div(x = x_341_cast_fp16, y = peak_137_cast_fp16)[name = string("x_343_cast_fp16")];
            fp16 var_7028_promoted_to_fp16 = const()[name = string("op_7028_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7036_cast_fp16 = pow(x = x_343_cast_fp16, y = var_7028_promoted_to_fp16)[name = string("op_7036_cast_fp16")];
            tensor<int32, [1]> var_7038_axes_0 = const()[name = string("op_7038_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7038_keep_dims_0 = const()[name = string("op_7038_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7038_cast_fp16 = reduce_mean(axes = var_7038_axes_0, keep_dims = var_7038_keep_dims_0, x = var_7036_cast_fp16)[name = string("op_7038_cast_fp16")];
            fp16 var_7028_promoted_1_to_fp16 = const()[name = string("op_7028_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7039_cast_fp16 = pow(x = peak_137_cast_fp16, y = var_7028_promoted_1_to_fp16)[name = string("op_7039_cast_fp16")];
            fp32 var_7040_epsilon_0 = const()[name = string("op_7040_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7040_cast_fp16 = inverse(epsilon = var_7040_epsilon_0, x = var_7039_cast_fp16)[name = string("op_7040_cast_fp16")];
            fp16 var_7041_to_fp16 = const()[name = string("op_7041_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7042_cast_fp16 = mul(x = var_7040_cast_fp16, y = var_7041_to_fp16)[name = string("op_7042_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7043_cast_fp16 = add(x = var_7038_cast_fp16, y = var_7042_cast_fp16)[name = string("op_7043_cast_fp16")];
            fp32 var_7044_epsilon_0 = const()[name = string("op_7044_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7044_cast_fp16 = rsqrt(epsilon = var_7044_epsilon_0, x = var_7043_cast_fp16)[name = string("op_7044_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7045_cast_fp16 = mul(x = x_343_cast_fp16, y = var_7044_cast_fp16)[name = string("op_7045_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_7045_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_7054 = const()[name = string("op_7054"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7055_cast_fp16 = reshape(shape = var_7054, x = linear_119_cast_fp16)[name = string("op_7055_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7062 = const()[name = string("op_7062"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_7055_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 16, 128]> var_7065_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_7065_cast_fp16")];
            tensor<int32, [1]> peak_139_axes_0 = const()[name = string("peak_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_139_cast_fp16 = reduce_max(axes = peak_139_axes_0, keep_dims = var_7062, x = var_7065_cast_fp16)[name = string("peak_139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_347_cast_fp16 = real_div(x = x_345_cast_fp16, y = peak_139_cast_fp16)[name = string("x_347_cast_fp16")];
            fp16 var_7061_promoted_to_fp16 = const()[name = string("op_7061_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7069_cast_fp16 = pow(x = x_347_cast_fp16, y = var_7061_promoted_to_fp16)[name = string("op_7069_cast_fp16")];
            tensor<int32, [1]> var_7071_axes_0 = const()[name = string("op_7071_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7071_keep_dims_0 = const()[name = string("op_7071_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7071_cast_fp16 = reduce_mean(axes = var_7071_axes_0, keep_dims = var_7071_keep_dims_0, x = var_7069_cast_fp16)[name = string("op_7071_cast_fp16")];
            fp16 var_7061_promoted_1_to_fp16 = const()[name = string("op_7061_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7072_cast_fp16 = pow(x = peak_139_cast_fp16, y = var_7061_promoted_1_to_fp16)[name = string("op_7072_cast_fp16")];
            fp32 var_7073_epsilon_0 = const()[name = string("op_7073_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7073_cast_fp16 = inverse(epsilon = var_7073_epsilon_0, x = var_7072_cast_fp16)[name = string("op_7073_cast_fp16")];
            fp16 var_7074_to_fp16 = const()[name = string("op_7074_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7075_cast_fp16 = mul(x = var_7073_cast_fp16, y = var_7074_to_fp16)[name = string("op_7075_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7076_cast_fp16 = add(x = var_7071_cast_fp16, y = var_7075_cast_fp16)[name = string("op_7076_cast_fp16")];
            fp32 var_7077_epsilon_0 = const()[name = string("op_7077_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7077_cast_fp16 = rsqrt(epsilon = var_7077_epsilon_0, x = var_7076_cast_fp16)[name = string("op_7077_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7078_cast_fp16 = mul(x = x_347_cast_fp16, y = var_7077_cast_fp16)[name = string("op_7078_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_353_cast_fp16 = mul(x = var_7078_cast_fp16, y = model_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("x_353_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_7087 = const()[name = string("op_7087"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7088_cast_fp16 = reshape(shape = var_7087, x = linear_120_cast_fp16)[name = string("op_7088_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7095 = const()[name = string("op_7095"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_7088_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 8, 16, 128]> var_7098_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_7098_cast_fp16")];
            tensor<int32, [1]> peak_141_axes_0 = const()[name = string("peak_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_141_cast_fp16 = reduce_max(axes = peak_141_axes_0, keep_dims = var_7095, x = var_7098_cast_fp16)[name = string("peak_141_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_351_cast_fp16 = real_div(x = x_349_cast_fp16, y = peak_141_cast_fp16)[name = string("x_351_cast_fp16")];
            fp16 var_7094_promoted_to_fp16 = const()[name = string("op_7094_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7102_cast_fp16 = pow(x = x_351_cast_fp16, y = var_7094_promoted_to_fp16)[name = string("op_7102_cast_fp16")];
            tensor<int32, [1]> var_7104_axes_0 = const()[name = string("op_7104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7104_keep_dims_0 = const()[name = string("op_7104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7104_cast_fp16 = reduce_mean(axes = var_7104_axes_0, keep_dims = var_7104_keep_dims_0, x = var_7102_cast_fp16)[name = string("op_7104_cast_fp16")];
            fp16 var_7094_promoted_1_to_fp16 = const()[name = string("op_7094_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7105_cast_fp16 = pow(x = peak_141_cast_fp16, y = var_7094_promoted_1_to_fp16)[name = string("op_7105_cast_fp16")];
            fp32 var_7106_epsilon_0 = const()[name = string("op_7106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7106_cast_fp16 = inverse(epsilon = var_7106_epsilon_0, x = var_7105_cast_fp16)[name = string("op_7106_cast_fp16")];
            fp16 var_7107_to_fp16 = const()[name = string("op_7107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7108_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7107_to_fp16)[name = string("op_7108_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7109_cast_fp16 = add(x = var_7104_cast_fp16, y = var_7108_cast_fp16)[name = string("op_7109_cast_fp16")];
            fp32 var_7110_epsilon_0 = const()[name = string("op_7110_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7110_cast_fp16 = rsqrt(epsilon = var_7110_epsilon_0, x = var_7109_cast_fp16)[name = string("op_7110_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7111_cast_fp16 = mul(x = x_351_cast_fp16, y = var_7110_cast_fp16)[name = string("op_7111_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_355_cast_fp16 = mul(x = var_7111_cast_fp16, y = model_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("x_355_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_7120 = const()[name = string("op_7120"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7121_cast_fp16 = reshape(shape = var_7120, x = linear_121_cast_fp16)[name = string("op_7121_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_353_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_353_cast_fp16)[name = string("second_69_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7135_cast_fp16 = mul(x = first_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7135_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7136_cast_fp16 = mul(x = second_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7136_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7138_cast_fp16 = sub(x = var_7135_cast_fp16, y = var_7136_cast_fp16)[name = string("op_7138_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7139_cast_fp16 = mul(x = second_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7140_cast_fp16 = mul(x = first_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7140_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7142_cast_fp16 = add(x = var_7139_cast_fp16, y = var_7140_cast_fp16)[name = string("op_7142_cast_fp16")];
            int32 var_7144 = const()[name = string("op_7144"), 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_7144, interleave = q_35_interleave_0, values = (var_7138_cast_fp16, var_7142_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_355_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_355_cast_fp16)[name = string("second_71_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7156_cast_fp16 = mul(x = first_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7156_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7157_cast_fp16 = mul(x = second_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7157_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7159_cast_fp16 = sub(x = var_7156_cast_fp16, y = var_7157_cast_fp16)[name = string("op_7159_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7160_cast_fp16 = mul(x = second_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7161_cast_fp16 = mul(x = first_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7161_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7163_cast_fp16 = add(x = var_7160_cast_fp16, y = var_7161_cast_fp16)[name = string("op_7163_cast_fp16")];
            int32 var_7165 = const()[name = string("op_7165"), 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_7165, interleave = k_35_interleave_0, values = (var_7159_cast_fp16, var_7163_cast_fp16))[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> var_7176_begin_0 = const()[name = string("op_7176_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7176_end_0 = const()[name = string("op_7176_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7176_end_mask_0 = const()[name = string("op_7176_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7176_stride_0 = const()[name = string("op_7176_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7176_cast_fp16 = slice_by_index(begin = var_7176_begin_0, end = var_7176_end_0, end_mask = var_7176_end_mask_0, stride = var_7176_stride_0, x = q_35_cast_fp16)[name = string("op_7176_cast_fp16")];
            tensor<int32, [4]> var_7186_begin_0 = const()[name = string("op_7186_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7186_end_0 = const()[name = string("op_7186_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7186_end_mask_0 = const()[name = string("op_7186_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7186_stride_0 = const()[name = string("op_7186_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7186_cast_fp16 = slice_by_index(begin = var_7186_begin_0, end = var_7186_end_0, end_mask = var_7186_end_mask_0, stride = var_7186_stride_0, x = q_35_cast_fp16)[name = string("op_7186_cast_fp16")];
            int32 var_7188 = const()[name = string("op_7188"), 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_7188, interleave = stacked_35_interleave_0, values = (var_7176_cast_fp16, var_7186_cast_fp16))[name = string("stacked_35_cast_fp16")];
            tensor<int32, [4]> var_7194_begin_0 = const()[name = string("op_7194_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_7194_end_0 = const()[name = string("op_7194_end_0"), val = tensor<int32, [4]>([18, 8, 2064, 128])];
            tensor<bool, [4]> var_7194_end_mask_0 = const()[name = string("op_7194_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7194_cast_fp16 = slice_by_index(begin = var_7194_begin_0, end = var_7194_end_0, end_mask = var_7194_end_mask_0, x = read_state_0)[name = string("op_7194_cast_fp16")];
            tensor<int32, [4]> var_7204_begin_0 = const()[name = string("op_7204_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7204_end_0 = const()[name = string("op_7204_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7204_end_mask_0 = const()[name = string("op_7204_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7204_cast_fp16 = slice_by_index(begin = var_7204_begin_0, end = var_7204_end_0, end_mask = var_7204_end_mask_0, x = var_7194_cast_fp16)[name = string("op_7204_cast_fp16")];
            bool var_7213_transpose_x_1 = const()[name = string("op_7213_transpose_x_1"), val = bool(false)];
            bool var_7213_transpose_y_1 = const()[name = string("op_7213_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_7213_cast_fp16 = matmul(transpose_x = var_7213_transpose_x_1, transpose_y = var_7213_transpose_y_1, x = var_7204_cast_fp16, y = stacked_35_cast_fp16)[name = string("op_7213_cast_fp16")];
            tensor<int32, [4]> var_7216_perm_0 = const()[name = string("op_7216_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7217_to_fp16 = const()[name = string("op_7217_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_7216_cast_fp16 = transpose(perm = var_7216_perm_0, x = var_7213_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 8, 32, 2048]> var_7218_cast_fp16 = mul(x = var_7216_cast_fp16, y = var_7217_to_fp16)[name = string("op_7218_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_35_cast_fp16 = add(x = var_7218_cast_fp16, y = cacheMask)[name = string("cache_weights_35_cast_fp16")];
            bool var_7224_transpose_x_1 = const()[name = string("op_7224_transpose_x_1"), val = bool(false)];
            bool var_7224_transpose_y_1 = const()[name = string("op_7224_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7224_cast_fp16 = matmul(transpose_x = var_7224_transpose_x_1, transpose_y = var_7224_transpose_y_1, x = stacked_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_7224_cast_fp16")];
            fp16 var_7225_to_fp16 = const()[name = string("op_7225_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7226_cast_fp16 = mul(x = var_7224_cast_fp16, y = var_7225_to_fp16)[name = string("op_7226_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_35_cast_fp16 = add(x = var_7226_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_35_cast_fp16")];
            int32 var_7230 = const()[name = string("op_7230"), val = int32(-1)];
            bool var_7231_interleave_0 = const()[name = string("op_7231_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_7231_cast_fp16 = concat(axis = var_7230, interleave = var_7231_interleave_0, values = (cache_weights_35_cast_fp16, call_weights_35_cast_fp16))[name = string("op_7231_cast_fp16")];
            int32 var_7232 = const()[name = string("op_7232"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_35_cast_fp16 = softmax(axis = var_7232, x = var_7231_cast_fp16)[name = string("shares_35_cast_fp16")];
            tensor<int32, [4]> var_7239_begin_0 = const()[name = string("op_7239_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7239_end_0 = const()[name = string("op_7239_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_7239_end_mask_0 = const()[name = string("op_7239_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_7239_cast_fp16 = slice_by_index(begin = var_7239_begin_0, end = var_7239_end_0, end_mask = var_7239_end_mask_0, x = shares_35_cast_fp16)[name = string("op_7239_cast_fp16")];
            tensor<int32, [4]> var_7244_begin_0 = const()[name = string("op_7244_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_7244_end_0 = const()[name = string("op_7244_end_0"), val = tensor<int32, [4]>([18, 8, 2064, 128])];
            tensor<bool, [4]> var_7244_end_mask_0 = const()[name = string("op_7244_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7244_cast_fp16 = slice_by_index(begin = var_7244_begin_0, end = var_7244_end_0, end_mask = var_7244_end_mask_0, x = read_state_1)[name = string("op_7244_cast_fp16")];
            tensor<int32, [4]> var_7254_begin_0 = const()[name = string("op_7254_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7254_end_0 = const()[name = string("op_7254_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7254_end_mask_0 = const()[name = string("op_7254_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7254_cast_fp16 = slice_by_index(begin = var_7254_begin_0, end = var_7254_end_0, end_mask = var_7254_end_mask_0, x = var_7244_cast_fp16)[name = string("op_7254_cast_fp16")];
            bool var_7260_transpose_x_0 = const()[name = string("op_7260_transpose_x_0"), val = bool(false)];
            bool var_7260_transpose_y_0 = const()[name = string("op_7260_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7260_cast_fp16 = matmul(transpose_x = var_7260_transpose_x_0, transpose_y = var_7260_transpose_y_0, x = var_7239_cast_fp16, y = var_7254_cast_fp16)[name = string("op_7260_cast_fp16")];
            tensor<int32, [4]> var_7265_begin_0 = const()[name = string("op_7265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_7265_end_0 = const()[name = string("op_7265_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_7265_end_mask_0 = const()[name = string("op_7265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7265_cast_fp16 = slice_by_index(begin = var_7265_begin_0, end = var_7265_end_0, end_mask = var_7265_end_mask_0, x = shares_35_cast_fp16)[name = string("op_7265_cast_fp16")];
            bool var_7266_transpose_x_0 = const()[name = string("op_7266_transpose_x_0"), val = bool(false)];
            bool var_7266_transpose_y_0 = const()[name = string("op_7266_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_7121_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 32, 128]> var_7266_cast_fp16 = matmul(transpose_x = var_7266_transpose_x_0, transpose_y = var_7266_transpose_y_0, x = var_7265_cast_fp16, y = v_35_cast_fp16)[name = string("op_7266_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_69_cast_fp16 = add(x = var_7260_cast_fp16, y = var_7266_cast_fp16)[name = string("mixed_69_cast_fp16")];
            tensor<int32, [4]> var_7283_begin_0 = const()[name = string("op_7283_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7283_end_0 = const()[name = string("op_7283_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7283_end_mask_0 = const()[name = string("op_7283_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7283_cast_fp16 = slice_by_index(begin = var_7283_begin_0, end = var_7283_end_0, end_mask = var_7283_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_7283_cast_fp16")];
            tensor<int32, [4]> var_7298_begin_0 = const()[name = string("op_7298_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7298_end_0 = const()[name = string("op_7298_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7298_end_mask_0 = const()[name = string("op_7298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7298_cast_fp16 = slice_by_index(begin = var_7298_begin_0, end = var_7298_end_0, end_mask = var_7298_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_7298_cast_fp16")];
            int32 var_7300 = const()[name = string("op_7300"), 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_7300, interleave = mixed_71_interleave_0, values = (var_7283_cast_fp16, var_7298_cast_fp16))[name = string("mixed_71_cast_fp16")];
            tensor<int32, [4]> var_7304_perm_0 = const()[name = string("op_7304_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7308 = const()[name = string("op_7308"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7304_cast_fp16 = transpose(perm = var_7304_perm_0, x = mixed_71_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 2048]> input_173_cast_fp16 = reshape(shape = var_7308, x = var_7304_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_357_cast_fp16 = add(x = x_341_cast_fp16, y = linear_122_cast_fp16)[name = string("x_357_cast_fp16")];
            bool var_7318 = const()[name = string("op_7318"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7321_cast_fp16 = abs(x = x_357_cast_fp16)[name = string("op_7321_cast_fp16")];
            tensor<int32, [1]> peak_143_axes_0 = const()[name = string("peak_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_143_cast_fp16 = reduce_max(axes = peak_143_axes_0, keep_dims = var_7318, x = var_7321_cast_fp16)[name = string("peak_143_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_359_cast_fp16 = real_div(x = x_357_cast_fp16, y = peak_143_cast_fp16)[name = string("x_359_cast_fp16")];
            fp16 var_7317_promoted_to_fp16 = const()[name = string("op_7317_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7325_cast_fp16 = pow(x = x_359_cast_fp16, y = var_7317_promoted_to_fp16)[name = string("op_7325_cast_fp16")];
            tensor<int32, [1]> var_7327_axes_0 = const()[name = string("op_7327_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7327_keep_dims_0 = const()[name = string("op_7327_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7327_cast_fp16 = reduce_mean(axes = var_7327_axes_0, keep_dims = var_7327_keep_dims_0, x = var_7325_cast_fp16)[name = string("op_7327_cast_fp16")];
            fp16 var_7317_promoted_1_to_fp16 = const()[name = string("op_7317_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7328_cast_fp16 = pow(x = peak_143_cast_fp16, y = var_7317_promoted_1_to_fp16)[name = string("op_7328_cast_fp16")];
            fp32 var_7329_epsilon_0 = const()[name = string("op_7329_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7329_cast_fp16 = inverse(epsilon = var_7329_epsilon_0, x = var_7328_cast_fp16)[name = string("op_7329_cast_fp16")];
            fp16 var_7330_to_fp16 = const()[name = string("op_7330_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7331_cast_fp16 = mul(x = var_7329_cast_fp16, y = var_7330_to_fp16)[name = string("op_7331_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7332_cast_fp16 = add(x = var_7327_cast_fp16, y = var_7331_cast_fp16)[name = string("op_7332_cast_fp16")];
            fp32 var_7333_epsilon_0 = const()[name = string("op_7333_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7333_cast_fp16 = rsqrt(epsilon = var_7333_epsilon_0, x = var_7332_cast_fp16)[name = string("op_7333_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7334_cast_fp16 = mul(x = x_359_cast_fp16, y = var_7333_cast_fp16)[name = string("op_7334_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_7334_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_7339_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_7339_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_7339_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_361_cast_fp16 = add(x = x_357_cast_fp16, y = linear_125_cast_fp16)[name = string("x_361_cast_fp16")];
            bool var_7352 = const()[name = string("op_7352"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7355_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_7355_cast_fp16")];
            tensor<int32, [1]> peak_145_axes_0 = const()[name = string("peak_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_145_cast_fp16 = reduce_max(axes = peak_145_axes_0, keep_dims = var_7352, x = var_7355_cast_fp16)[name = string("peak_145_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_363_cast_fp16 = real_div(x = x_361_cast_fp16, y = peak_145_cast_fp16)[name = string("x_363_cast_fp16")];
            fp16 var_7351_promoted_to_fp16 = const()[name = string("op_7351_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7359_cast_fp16 = pow(x = x_363_cast_fp16, y = var_7351_promoted_to_fp16)[name = string("op_7359_cast_fp16")];
            tensor<int32, [1]> var_7361_axes_0 = const()[name = string("op_7361_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7361_keep_dims_0 = const()[name = string("op_7361_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7361_cast_fp16 = reduce_mean(axes = var_7361_axes_0, keep_dims = var_7361_keep_dims_0, x = var_7359_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7351_promoted_1_to_fp16 = const()[name = string("op_7351_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7362_cast_fp16 = pow(x = peak_145_cast_fp16, y = var_7351_promoted_1_to_fp16)[name = string("op_7362_cast_fp16")];
            fp32 var_7363_epsilon_0 = const()[name = string("op_7363_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7363_cast_fp16 = inverse(epsilon = var_7363_epsilon_0, x = var_7362_cast_fp16)[name = string("op_7363_cast_fp16")];
            fp16 var_7364_to_fp16 = const()[name = string("op_7364_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7365_cast_fp16 = mul(x = var_7363_cast_fp16, y = var_7364_to_fp16)[name = string("op_7365_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7366_cast_fp16 = add(x = var_7361_cast_fp16, y = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            fp32 var_7367_epsilon_0 = const()[name = string("op_7367_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7367_cast_fp16 = rsqrt(epsilon = var_7367_epsilon_0, x = var_7366_cast_fp16)[name = string("op_7367_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7368_cast_fp16 = mul(x = x_363_cast_fp16, y = var_7367_cast_fp16)[name = string("op_7368_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_7368_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_7377 = const()[name = string("op_7377"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7378_cast_fp16 = reshape(shape = var_7377, x = linear_126_cast_fp16)[name = string("op_7378_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7385 = const()[name = string("op_7385"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_7378_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 16, 128]> var_7388_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_7388_cast_fp16")];
            tensor<int32, [1]> peak_147_axes_0 = const()[name = string("peak_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_147_cast_fp16 = reduce_max(axes = peak_147_axes_0, keep_dims = var_7385, x = var_7388_cast_fp16)[name = string("peak_147_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_367_cast_fp16 = real_div(x = x_365_cast_fp16, y = peak_147_cast_fp16)[name = string("x_367_cast_fp16")];
            fp16 var_7384_promoted_to_fp16 = const()[name = string("op_7384_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7392_cast_fp16 = pow(x = x_367_cast_fp16, y = var_7384_promoted_to_fp16)[name = string("op_7392_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7392_cast_fp16)[name = string("op_7394_cast_fp16")];
            fp16 var_7384_promoted_1_to_fp16 = const()[name = string("op_7384_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7395_cast_fp16 = pow(x = peak_147_cast_fp16, y = var_7384_promoted_1_to_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7399_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7399_cast_fp16")];
            fp32 var_7400_epsilon_0 = const()[name = string("op_7400_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7400_cast_fp16 = rsqrt(epsilon = var_7400_epsilon_0, x = var_7399_cast_fp16)[name = string("op_7400_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7401_cast_fp16 = mul(x = x_367_cast_fp16, y = var_7400_cast_fp16)[name = string("op_7401_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_373_cast_fp16 = mul(x = var_7401_cast_fp16, y = model_layers_18_self_attn_q_norm_weight_to_fp16)[name = string("x_373_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_7410 = const()[name = string("op_7410"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7411_cast_fp16 = reshape(shape = var_7410, x = linear_127_cast_fp16)[name = string("op_7411_cast_fp16")];
            tensor<int32, [4]> x_369_perm_0 = const()[name = string("x_369_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7418 = const()[name = string("op_7418"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_369_cast_fp16 = transpose(perm = x_369_perm_0, x = var_7411_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 8, 16, 128]> var_7421_cast_fp16 = abs(x = x_369_cast_fp16)[name = string("op_7421_cast_fp16")];
            tensor<int32, [1]> peak_149_axes_0 = const()[name = string("peak_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_149_cast_fp16 = reduce_max(axes = peak_149_axes_0, keep_dims = var_7418, x = var_7421_cast_fp16)[name = string("peak_149_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_371_cast_fp16 = real_div(x = x_369_cast_fp16, y = peak_149_cast_fp16)[name = string("x_371_cast_fp16")];
            fp16 var_7417_promoted_to_fp16 = const()[name = string("op_7417_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7425_cast_fp16 = pow(x = x_371_cast_fp16, y = var_7417_promoted_to_fp16)[name = string("op_7425_cast_fp16")];
            tensor<int32, [1]> var_7427_axes_0 = const()[name = string("op_7427_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7427_keep_dims_0 = const()[name = string("op_7427_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7427_cast_fp16 = reduce_mean(axes = var_7427_axes_0, keep_dims = var_7427_keep_dims_0, x = var_7425_cast_fp16)[name = string("op_7427_cast_fp16")];
            fp16 var_7417_promoted_1_to_fp16 = const()[name = string("op_7417_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7428_cast_fp16 = pow(x = peak_149_cast_fp16, y = var_7417_promoted_1_to_fp16)[name = string("op_7428_cast_fp16")];
            fp32 var_7429_epsilon_0 = const()[name = string("op_7429_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7429_cast_fp16 = inverse(epsilon = var_7429_epsilon_0, x = var_7428_cast_fp16)[name = string("op_7429_cast_fp16")];
            fp16 var_7430_to_fp16 = const()[name = string("op_7430_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7431_cast_fp16 = mul(x = var_7429_cast_fp16, y = var_7430_to_fp16)[name = string("op_7431_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7432_cast_fp16 = add(x = var_7427_cast_fp16, y = var_7431_cast_fp16)[name = string("op_7432_cast_fp16")];
            fp32 var_7433_epsilon_0 = const()[name = string("op_7433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7433_cast_fp16 = rsqrt(epsilon = var_7433_epsilon_0, x = var_7432_cast_fp16)[name = string("op_7433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7434_cast_fp16 = mul(x = x_371_cast_fp16, y = var_7433_cast_fp16)[name = string("op_7434_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_375_cast_fp16 = mul(x = var_7434_cast_fp16, y = model_layers_18_self_attn_k_norm_weight_to_fp16)[name = string("x_375_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_7443 = const()[name = string("op_7443"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7444_cast_fp16 = reshape(shape = var_7443, x = linear_128_cast_fp16)[name = string("op_7444_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_373_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_373_cast_fp16)[name = string("second_73_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7458_cast_fp16 = mul(x = first_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7458_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7459_cast_fp16 = mul(x = second_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7459_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7461_cast_fp16 = sub(x = var_7458_cast_fp16, y = var_7459_cast_fp16)[name = string("op_7461_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7462_cast_fp16 = mul(x = second_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7463_cast_fp16 = mul(x = first_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7463_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7465_cast_fp16 = add(x = var_7462_cast_fp16, y = var_7463_cast_fp16)[name = string("op_7465_cast_fp16")];
            int32 var_7467 = const()[name = string("op_7467"), 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_7467, interleave = q_37_interleave_0, values = (var_7461_cast_fp16, var_7465_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_375_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_375_cast_fp16)[name = string("second_75_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7479_cast_fp16 = mul(x = first_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7480_cast_fp16 = mul(x = second_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7480_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7482_cast_fp16 = sub(x = var_7479_cast_fp16, y = var_7480_cast_fp16)[name = string("op_7482_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7483_cast_fp16 = mul(x = second_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7484_cast_fp16 = mul(x = first_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7484_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7486_cast_fp16 = add(x = var_7483_cast_fp16, y = var_7484_cast_fp16)[name = string("op_7486_cast_fp16")];
            int32 var_7488 = const()[name = string("op_7488"), 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_7488, interleave = k_37_interleave_0, values = (var_7482_cast_fp16, var_7486_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_7499_begin_0 = const()[name = string("op_7499_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7499_end_0 = const()[name = string("op_7499_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7499_end_mask_0 = const()[name = string("op_7499_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7499_stride_0 = const()[name = string("op_7499_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7499_cast_fp16 = slice_by_index(begin = var_7499_begin_0, end = var_7499_end_0, end_mask = var_7499_end_mask_0, stride = var_7499_stride_0, x = q_37_cast_fp16)[name = string("op_7499_cast_fp16")];
            tensor<int32, [4]> var_7509_begin_0 = const()[name = string("op_7509_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7509_end_0 = const()[name = string("op_7509_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7509_end_mask_0 = const()[name = string("op_7509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7509_stride_0 = const()[name = string("op_7509_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7509_cast_fp16 = slice_by_index(begin = var_7509_begin_0, end = var_7509_end_0, end_mask = var_7509_end_mask_0, stride = var_7509_stride_0, x = q_37_cast_fp16)[name = string("op_7509_cast_fp16")];
            int32 var_7511 = const()[name = string("op_7511"), 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_7511, interleave = stacked_37_interleave_0, values = (var_7499_cast_fp16, var_7509_cast_fp16))[name = string("stacked_37_cast_fp16")];
            tensor<int32, [4]> var_7517_begin_0 = const()[name = string("op_7517_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7517_end_0 = const()[name = string("op_7517_end_0"), val = tensor<int32, [4]>([19, 8, 2064, 128])];
            tensor<bool, [4]> var_7517_end_mask_0 = const()[name = string("op_7517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7517_cast_fp16 = slice_by_index(begin = var_7517_begin_0, end = var_7517_end_0, end_mask = var_7517_end_mask_0, x = read_state_0)[name = string("op_7517_cast_fp16")];
            tensor<int32, [4]> var_7527_begin_0 = const()[name = string("op_7527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7527_end_0 = const()[name = string("op_7527_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7527_end_mask_0 = const()[name = string("op_7527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7527_cast_fp16 = slice_by_index(begin = var_7527_begin_0, end = var_7527_end_0, end_mask = var_7527_end_mask_0, x = var_7517_cast_fp16)[name = string("op_7527_cast_fp16")];
            bool var_7536_transpose_x_1 = const()[name = string("op_7536_transpose_x_1"), val = bool(false)];
            bool var_7536_transpose_y_1 = const()[name = string("op_7536_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_7536_cast_fp16 = matmul(transpose_x = var_7536_transpose_x_1, transpose_y = var_7536_transpose_y_1, x = var_7527_cast_fp16, y = stacked_37_cast_fp16)[name = string("op_7536_cast_fp16")];
            tensor<int32, [4]> var_7539_perm_0 = const()[name = string("op_7539_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7540_to_fp16 = const()[name = string("op_7540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_7539_cast_fp16 = transpose(perm = var_7539_perm_0, x = var_7536_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 8, 32, 2048]> var_7541_cast_fp16 = mul(x = var_7539_cast_fp16, y = var_7540_to_fp16)[name = string("op_7541_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_37_cast_fp16 = add(x = var_7541_cast_fp16, y = cacheMask)[name = string("cache_weights_37_cast_fp16")];
            bool var_7547_transpose_x_1 = const()[name = string("op_7547_transpose_x_1"), val = bool(false)];
            bool var_7547_transpose_y_1 = const()[name = string("op_7547_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7547_cast_fp16 = matmul(transpose_x = var_7547_transpose_x_1, transpose_y = var_7547_transpose_y_1, x = stacked_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_7547_cast_fp16")];
            fp16 var_7548_to_fp16 = const()[name = string("op_7548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7549_cast_fp16 = mul(x = var_7547_cast_fp16, y = var_7548_to_fp16)[name = string("op_7549_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_37_cast_fp16 = add(x = var_7549_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_37_cast_fp16")];
            int32 var_7553 = const()[name = string("op_7553"), val = int32(-1)];
            bool var_7554_interleave_0 = const()[name = string("op_7554_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_7554_cast_fp16 = concat(axis = var_7553, interleave = var_7554_interleave_0, values = (cache_weights_37_cast_fp16, call_weights_37_cast_fp16))[name = string("op_7554_cast_fp16")];
            int32 var_7555 = const()[name = string("op_7555"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_37_cast_fp16 = softmax(axis = var_7555, x = var_7554_cast_fp16)[name = string("shares_37_cast_fp16")];
            tensor<int32, [4]> var_7562_begin_0 = const()[name = string("op_7562_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7562_end_0 = const()[name = string("op_7562_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_7562_end_mask_0 = const()[name = string("op_7562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_7562_cast_fp16 = slice_by_index(begin = var_7562_begin_0, end = var_7562_end_0, end_mask = var_7562_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7562_cast_fp16")];
            tensor<int32, [4]> var_7567_begin_0 = const()[name = string("op_7567_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7567_end_0 = const()[name = string("op_7567_end_0"), val = tensor<int32, [4]>([19, 8, 2064, 128])];
            tensor<bool, [4]> var_7567_end_mask_0 = const()[name = string("op_7567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7567_cast_fp16 = slice_by_index(begin = var_7567_begin_0, end = var_7567_end_0, end_mask = var_7567_end_mask_0, x = read_state_1)[name = string("op_7567_cast_fp16")];
            tensor<int32, [4]> var_7577_begin_0 = const()[name = string("op_7577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7577_end_0 = const()[name = string("op_7577_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7577_end_mask_0 = const()[name = string("op_7577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7577_cast_fp16 = slice_by_index(begin = var_7577_begin_0, end = var_7577_end_0, end_mask = var_7577_end_mask_0, x = var_7567_cast_fp16)[name = string("op_7577_cast_fp16")];
            bool var_7583_transpose_x_0 = const()[name = string("op_7583_transpose_x_0"), val = bool(false)];
            bool var_7583_transpose_y_0 = const()[name = string("op_7583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7583_cast_fp16 = matmul(transpose_x = var_7583_transpose_x_0, transpose_y = var_7583_transpose_y_0, x = var_7562_cast_fp16, y = var_7577_cast_fp16)[name = string("op_7583_cast_fp16")];
            tensor<int32, [4]> var_7588_begin_0 = const()[name = string("op_7588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_7588_end_0 = const()[name = string("op_7588_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_7588_end_mask_0 = const()[name = string("op_7588_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7588_cast_fp16 = slice_by_index(begin = var_7588_begin_0, end = var_7588_end_0, end_mask = var_7588_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7588_cast_fp16")];
            bool var_7589_transpose_x_0 = const()[name = string("op_7589_transpose_x_0"), val = bool(false)];
            bool var_7589_transpose_y_0 = const()[name = string("op_7589_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_7444_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, 32, 128]> var_7589_cast_fp16 = matmul(transpose_x = var_7589_transpose_x_0, transpose_y = var_7589_transpose_y_0, x = var_7588_cast_fp16, y = v_37_cast_fp16)[name = string("op_7589_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_73_cast_fp16 = add(x = var_7583_cast_fp16, y = var_7589_cast_fp16)[name = string("mixed_73_cast_fp16")];
            tensor<int32, [4]> var_7606_begin_0 = const()[name = string("op_7606_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7606_end_0 = const()[name = string("op_7606_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7606_end_mask_0 = const()[name = string("op_7606_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7606_cast_fp16 = slice_by_index(begin = var_7606_begin_0, end = var_7606_end_0, end_mask = var_7606_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7606_cast_fp16")];
            tensor<int32, [4]> var_7621_begin_0 = const()[name = string("op_7621_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7621_end_0 = const()[name = string("op_7621_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7621_end_mask_0 = const()[name = string("op_7621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7621_cast_fp16 = slice_by_index(begin = var_7621_begin_0, end = var_7621_end_0, end_mask = var_7621_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7621_cast_fp16")];
            int32 var_7623 = const()[name = string("op_7623"), 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_7623, interleave = mixed_75_interleave_0, values = (var_7606_cast_fp16, var_7621_cast_fp16))[name = string("mixed_75_cast_fp16")];
            tensor<int32, [4]> var_7627_perm_0 = const()[name = string("op_7627_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7631 = const()[name = string("op_7631"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7627_cast_fp16 = transpose(perm = var_7627_perm_0, x = mixed_75_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 2048]> input_183_cast_fp16 = reshape(shape = var_7631, x = var_7627_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_377_cast_fp16 = add(x = x_361_cast_fp16, y = linear_129_cast_fp16)[name = string("x_377_cast_fp16")];
            bool var_7641 = const()[name = string("op_7641"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7644_cast_fp16 = abs(x = x_377_cast_fp16)[name = string("op_7644_cast_fp16")];
            tensor<int32, [1]> peak_151_axes_0 = const()[name = string("peak_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_151_cast_fp16 = reduce_max(axes = peak_151_axes_0, keep_dims = var_7641, x = var_7644_cast_fp16)[name = string("peak_151_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_379_cast_fp16 = real_div(x = x_377_cast_fp16, y = peak_151_cast_fp16)[name = string("x_379_cast_fp16")];
            fp16 var_7640_promoted_to_fp16 = const()[name = string("op_7640_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7648_cast_fp16 = pow(x = x_379_cast_fp16, y = var_7640_promoted_to_fp16)[name = string("op_7648_cast_fp16")];
            tensor<int32, [1]> var_7650_axes_0 = const()[name = string("op_7650_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7650_keep_dims_0 = const()[name = string("op_7650_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7650_cast_fp16 = reduce_mean(axes = var_7650_axes_0, keep_dims = var_7650_keep_dims_0, x = var_7648_cast_fp16)[name = string("op_7650_cast_fp16")];
            fp16 var_7640_promoted_1_to_fp16 = const()[name = string("op_7640_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7651_cast_fp16 = pow(x = peak_151_cast_fp16, y = var_7640_promoted_1_to_fp16)[name = string("op_7651_cast_fp16")];
            fp32 var_7652_epsilon_0 = const()[name = string("op_7652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7652_cast_fp16 = inverse(epsilon = var_7652_epsilon_0, x = var_7651_cast_fp16)[name = string("op_7652_cast_fp16")];
            fp16 var_7653_to_fp16 = const()[name = string("op_7653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7654_cast_fp16 = mul(x = var_7652_cast_fp16, y = var_7653_to_fp16)[name = string("op_7654_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7655_cast_fp16 = add(x = var_7650_cast_fp16, y = var_7654_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp32 var_7656_epsilon_0 = const()[name = string("op_7656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7656_cast_fp16 = rsqrt(epsilon = var_7656_epsilon_0, x = var_7655_cast_fp16)[name = string("op_7656_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7657_cast_fp16 = mul(x = x_379_cast_fp16, y = var_7656_cast_fp16)[name = string("op_7657_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_7657_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_7662_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_7662_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_7662_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_381_cast_fp16 = add(x = x_377_cast_fp16, y = linear_132_cast_fp16)[name = string("x_381_cast_fp16")];
            bool var_7675 = const()[name = string("op_7675"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7678_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_7678_cast_fp16")];
            tensor<int32, [1]> peak_153_axes_0 = const()[name = string("peak_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_153_cast_fp16 = reduce_max(axes = peak_153_axes_0, keep_dims = var_7675, x = var_7678_cast_fp16)[name = string("peak_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_383_cast_fp16 = real_div(x = x_381_cast_fp16, y = peak_153_cast_fp16)[name = string("x_383_cast_fp16")];
            fp16 var_7674_promoted_to_fp16 = const()[name = string("op_7674_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7682_cast_fp16 = pow(x = x_383_cast_fp16, y = var_7674_promoted_to_fp16)[name = string("op_7682_cast_fp16")];
            tensor<int32, [1]> var_7684_axes_0 = const()[name = string("op_7684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7684_keep_dims_0 = const()[name = string("op_7684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7684_cast_fp16 = reduce_mean(axes = var_7684_axes_0, keep_dims = var_7684_keep_dims_0, x = var_7682_cast_fp16)[name = string("op_7684_cast_fp16")];
            fp16 var_7674_promoted_1_to_fp16 = const()[name = string("op_7674_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7685_cast_fp16 = pow(x = peak_153_cast_fp16, y = var_7674_promoted_1_to_fp16)[name = string("op_7685_cast_fp16")];
            fp32 var_7686_epsilon_0 = const()[name = string("op_7686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7686_cast_fp16 = inverse(epsilon = var_7686_epsilon_0, x = var_7685_cast_fp16)[name = string("op_7686_cast_fp16")];
            fp16 var_7687_to_fp16 = const()[name = string("op_7687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7688_cast_fp16 = mul(x = var_7686_cast_fp16, y = var_7687_to_fp16)[name = string("op_7688_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7689_cast_fp16 = add(x = var_7684_cast_fp16, y = var_7688_cast_fp16)[name = string("op_7689_cast_fp16")];
            fp32 var_7690_epsilon_0 = const()[name = string("op_7690_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7690_cast_fp16 = rsqrt(epsilon = var_7690_epsilon_0, x = var_7689_cast_fp16)[name = string("op_7690_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7691_cast_fp16 = mul(x = x_383_cast_fp16, y = var_7690_cast_fp16)[name = string("op_7691_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_7691_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_7700 = const()[name = string("op_7700"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7701_cast_fp16 = reshape(shape = var_7700, x = linear_133_cast_fp16)[name = string("op_7701_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7708 = const()[name = string("op_7708"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_7701_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 16, 16, 128]> var_7711_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_7711_cast_fp16")];
            tensor<int32, [1]> peak_155_axes_0 = const()[name = string("peak_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_155_cast_fp16 = reduce_max(axes = peak_155_axes_0, keep_dims = var_7708, x = var_7711_cast_fp16)[name = string("peak_155_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_387_cast_fp16 = real_div(x = x_385_cast_fp16, y = peak_155_cast_fp16)[name = string("x_387_cast_fp16")];
            fp16 var_7707_promoted_to_fp16 = const()[name = string("op_7707_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7715_cast_fp16 = pow(x = x_387_cast_fp16, y = var_7707_promoted_to_fp16)[name = string("op_7715_cast_fp16")];
            tensor<int32, [1]> var_7717_axes_0 = const()[name = string("op_7717_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7717_keep_dims_0 = const()[name = string("op_7717_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7717_cast_fp16 = reduce_mean(axes = var_7717_axes_0, keep_dims = var_7717_keep_dims_0, x = var_7715_cast_fp16)[name = string("op_7717_cast_fp16")];
            fp16 var_7707_promoted_1_to_fp16 = const()[name = string("op_7707_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7718_cast_fp16 = pow(x = peak_155_cast_fp16, y = var_7707_promoted_1_to_fp16)[name = string("op_7718_cast_fp16")];
            fp32 var_7719_epsilon_0 = const()[name = string("op_7719_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7719_cast_fp16 = inverse(epsilon = var_7719_epsilon_0, x = var_7718_cast_fp16)[name = string("op_7719_cast_fp16")];
            fp16 var_7720_to_fp16 = const()[name = string("op_7720_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7721_cast_fp16 = mul(x = var_7719_cast_fp16, y = var_7720_to_fp16)[name = string("op_7721_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7722_cast_fp16 = add(x = var_7717_cast_fp16, y = var_7721_cast_fp16)[name = string("op_7722_cast_fp16")];
            fp32 var_7723_epsilon_0 = const()[name = string("op_7723_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7723_cast_fp16 = rsqrt(epsilon = var_7723_epsilon_0, x = var_7722_cast_fp16)[name = string("op_7723_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7724_cast_fp16 = mul(x = x_387_cast_fp16, y = var_7723_cast_fp16)[name = string("op_7724_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_393_cast_fp16 = mul(x = var_7724_cast_fp16, y = model_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("x_393_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_7733 = const()[name = string("op_7733"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7734_cast_fp16 = reshape(shape = var_7733, x = linear_134_cast_fp16)[name = string("op_7734_cast_fp16")];
            tensor<int32, [4]> x_389_perm_0 = const()[name = string("x_389_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7741 = const()[name = string("op_7741"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_389_cast_fp16 = transpose(perm = x_389_perm_0, x = var_7734_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 8, 16, 128]> var_7744_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_7744_cast_fp16")];
            tensor<int32, [1]> peak_157_axes_0 = const()[name = string("peak_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_157_cast_fp16 = reduce_max(axes = peak_157_axes_0, keep_dims = var_7741, x = var_7744_cast_fp16)[name = string("peak_157_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_391_cast_fp16 = real_div(x = x_389_cast_fp16, y = peak_157_cast_fp16)[name = string("x_391_cast_fp16")];
            fp16 var_7740_promoted_to_fp16 = const()[name = string("op_7740_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7748_cast_fp16 = pow(x = x_391_cast_fp16, y = var_7740_promoted_to_fp16)[name = string("op_7748_cast_fp16")];
            tensor<int32, [1]> var_7750_axes_0 = const()[name = string("op_7750_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7750_keep_dims_0 = const()[name = string("op_7750_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7750_cast_fp16 = reduce_mean(axes = var_7750_axes_0, keep_dims = var_7750_keep_dims_0, x = var_7748_cast_fp16)[name = string("op_7750_cast_fp16")];
            fp16 var_7740_promoted_1_to_fp16 = const()[name = string("op_7740_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7751_cast_fp16 = pow(x = peak_157_cast_fp16, y = var_7740_promoted_1_to_fp16)[name = string("op_7751_cast_fp16")];
            fp32 var_7752_epsilon_0 = const()[name = string("op_7752_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7752_cast_fp16 = inverse(epsilon = var_7752_epsilon_0, x = var_7751_cast_fp16)[name = string("op_7752_cast_fp16")];
            fp16 var_7753_to_fp16 = const()[name = string("op_7753_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7754_cast_fp16 = mul(x = var_7752_cast_fp16, y = var_7753_to_fp16)[name = string("op_7754_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7755_cast_fp16 = add(x = var_7750_cast_fp16, y = var_7754_cast_fp16)[name = string("op_7755_cast_fp16")];
            fp32 var_7756_epsilon_0 = const()[name = string("op_7756_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7756_cast_fp16 = rsqrt(epsilon = var_7756_epsilon_0, x = var_7755_cast_fp16)[name = string("op_7756_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7757_cast_fp16 = mul(x = x_391_cast_fp16, y = var_7756_cast_fp16)[name = string("op_7757_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_395_cast_fp16 = mul(x = var_7757_cast_fp16, y = model_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("x_395_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_7766 = const()[name = string("op_7766"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7767_cast_fp16 = reshape(shape = var_7766, x = linear_135_cast_fp16)[name = string("op_7767_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_393_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_393_cast_fp16)[name = string("second_77_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7781_cast_fp16 = mul(x = first_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7781_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7782_cast_fp16 = mul(x = second_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7782_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7784_cast_fp16 = sub(x = var_7781_cast_fp16, y = var_7782_cast_fp16)[name = string("op_7784_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7785_cast_fp16 = mul(x = second_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7786_cast_fp16 = mul(x = first_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7786_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7788_cast_fp16 = add(x = var_7785_cast_fp16, y = var_7786_cast_fp16)[name = string("op_7788_cast_fp16")];
            int32 var_7790 = const()[name = string("op_7790"), 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_7790, interleave = q_39_interleave_0, values = (var_7784_cast_fp16, var_7788_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_395_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_395_cast_fp16)[name = string("second_79_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7802_cast_fp16 = mul(x = first_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7803_cast_fp16 = mul(x = second_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7803_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7805_cast_fp16 = sub(x = var_7802_cast_fp16, y = var_7803_cast_fp16)[name = string("op_7805_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7806_cast_fp16 = mul(x = second_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7807_cast_fp16 = mul(x = first_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7807_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7809_cast_fp16 = add(x = var_7806_cast_fp16, y = var_7807_cast_fp16)[name = string("op_7809_cast_fp16")];
            int32 var_7811 = const()[name = string("op_7811"), 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_7811, interleave = k_39_interleave_0, values = (var_7805_cast_fp16, var_7809_cast_fp16))[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> var_7822_begin_0 = const()[name = string("op_7822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7822_end_0 = const()[name = string("op_7822_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7822_end_mask_0 = const()[name = string("op_7822_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7822_stride_0 = const()[name = string("op_7822_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7822_cast_fp16 = slice_by_index(begin = var_7822_begin_0, end = var_7822_end_0, end_mask = var_7822_end_mask_0, stride = var_7822_stride_0, x = q_39_cast_fp16)[name = string("op_7822_cast_fp16")];
            tensor<int32, [4]> var_7832_begin_0 = const()[name = string("op_7832_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7832_end_0 = const()[name = string("op_7832_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7832_end_mask_0 = const()[name = string("op_7832_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7832_stride_0 = const()[name = string("op_7832_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7832_cast_fp16 = slice_by_index(begin = var_7832_begin_0, end = var_7832_end_0, end_mask = var_7832_end_mask_0, stride = var_7832_stride_0, x = q_39_cast_fp16)[name = string("op_7832_cast_fp16")];
            int32 var_7834 = const()[name = string("op_7834"), 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_7834, interleave = stacked_39_interleave_0, values = (var_7822_cast_fp16, var_7832_cast_fp16))[name = string("stacked_39_cast_fp16")];
            tensor<int32, [4]> var_7840_begin_0 = const()[name = string("op_7840_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7840_end_0 = const()[name = string("op_7840_end_0"), val = tensor<int32, [4]>([20, 8, 2064, 128])];
            tensor<bool, [4]> var_7840_end_mask_0 = const()[name = string("op_7840_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7840_cast_fp16 = slice_by_index(begin = var_7840_begin_0, end = var_7840_end_0, end_mask = var_7840_end_mask_0, x = read_state_0)[name = string("op_7840_cast_fp16")];
            tensor<int32, [4]> var_7850_begin_0 = const()[name = string("op_7850_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7850_end_0 = const()[name = string("op_7850_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7850_end_mask_0 = const()[name = string("op_7850_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7850_cast_fp16 = slice_by_index(begin = var_7850_begin_0, end = var_7850_end_0, end_mask = var_7850_end_mask_0, x = var_7840_cast_fp16)[name = string("op_7850_cast_fp16")];
            bool var_7859_transpose_x_1 = const()[name = string("op_7859_transpose_x_1"), val = bool(false)];
            bool var_7859_transpose_y_1 = const()[name = string("op_7859_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_7859_cast_fp16 = matmul(transpose_x = var_7859_transpose_x_1, transpose_y = var_7859_transpose_y_1, x = var_7850_cast_fp16, y = stacked_39_cast_fp16)[name = string("op_7859_cast_fp16")];
            tensor<int32, [4]> var_7862_perm_0 = const()[name = string("op_7862_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7863_to_fp16 = const()[name = string("op_7863_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_7862_cast_fp16 = transpose(perm = var_7862_perm_0, x = var_7859_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 32, 2048]> var_7864_cast_fp16 = mul(x = var_7862_cast_fp16, y = var_7863_to_fp16)[name = string("op_7864_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_39_cast_fp16 = add(x = var_7864_cast_fp16, y = cacheMask)[name = string("cache_weights_39_cast_fp16")];
            bool var_7870_transpose_x_1 = const()[name = string("op_7870_transpose_x_1"), val = bool(false)];
            bool var_7870_transpose_y_1 = const()[name = string("op_7870_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7870_cast_fp16 = matmul(transpose_x = var_7870_transpose_x_1, transpose_y = var_7870_transpose_y_1, x = stacked_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_7870_cast_fp16")];
            fp16 var_7871_to_fp16 = const()[name = string("op_7871_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7872_cast_fp16 = mul(x = var_7870_cast_fp16, y = var_7871_to_fp16)[name = string("op_7872_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_39_cast_fp16 = add(x = var_7872_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_39_cast_fp16")];
            int32 var_7876 = const()[name = string("op_7876"), val = int32(-1)];
            bool var_7877_interleave_0 = const()[name = string("op_7877_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_7877_cast_fp16 = concat(axis = var_7876, interleave = var_7877_interleave_0, values = (cache_weights_39_cast_fp16, call_weights_39_cast_fp16))[name = string("op_7877_cast_fp16")];
            int32 var_7878 = const()[name = string("op_7878"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_39_cast_fp16 = softmax(axis = var_7878, x = var_7877_cast_fp16)[name = string("shares_39_cast_fp16")];
            tensor<int32, [4]> var_7885_begin_0 = const()[name = string("op_7885_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7885_end_0 = const()[name = string("op_7885_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_7885_end_mask_0 = const()[name = string("op_7885_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_7885_cast_fp16 = slice_by_index(begin = var_7885_begin_0, end = var_7885_end_0, end_mask = var_7885_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7885_cast_fp16")];
            tensor<int32, [4]> var_7890_begin_0 = const()[name = string("op_7890_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7890_end_0 = const()[name = string("op_7890_end_0"), val = tensor<int32, [4]>([20, 8, 2064, 128])];
            tensor<bool, [4]> var_7890_end_mask_0 = const()[name = string("op_7890_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7890_cast_fp16 = slice_by_index(begin = var_7890_begin_0, end = var_7890_end_0, end_mask = var_7890_end_mask_0, x = read_state_1)[name = string("op_7890_cast_fp16")];
            tensor<int32, [4]> var_7900_begin_0 = const()[name = string("op_7900_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7900_end_0 = const()[name = string("op_7900_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7900_end_mask_0 = const()[name = string("op_7900_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7900_cast_fp16 = slice_by_index(begin = var_7900_begin_0, end = var_7900_end_0, end_mask = var_7900_end_mask_0, x = var_7890_cast_fp16)[name = string("op_7900_cast_fp16")];
            bool var_7906_transpose_x_0 = const()[name = string("op_7906_transpose_x_0"), val = bool(false)];
            bool var_7906_transpose_y_0 = const()[name = string("op_7906_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7906_cast_fp16 = matmul(transpose_x = var_7906_transpose_x_0, transpose_y = var_7906_transpose_y_0, x = var_7885_cast_fp16, y = var_7900_cast_fp16)[name = string("op_7906_cast_fp16")];
            tensor<int32, [4]> var_7911_begin_0 = const()[name = string("op_7911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_7911_end_0 = const()[name = string("op_7911_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_7911_end_mask_0 = const()[name = string("op_7911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7911_cast_fp16 = slice_by_index(begin = var_7911_begin_0, end = var_7911_end_0, end_mask = var_7911_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7911_cast_fp16")];
            bool var_7912_transpose_x_0 = const()[name = string("op_7912_transpose_x_0"), val = bool(false)];
            bool var_7912_transpose_y_0 = const()[name = string("op_7912_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_7767_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 32, 128]> var_7912_cast_fp16 = matmul(transpose_x = var_7912_transpose_x_0, transpose_y = var_7912_transpose_y_0, x = var_7911_cast_fp16, y = v_39_cast_fp16)[name = string("op_7912_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_77_cast_fp16 = add(x = var_7906_cast_fp16, y = var_7912_cast_fp16)[name = string("mixed_77_cast_fp16")];
            tensor<int32, [4]> var_7929_begin_0 = const()[name = string("op_7929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7929_end_0 = const()[name = string("op_7929_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7929_end_mask_0 = const()[name = string("op_7929_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7929_cast_fp16 = slice_by_index(begin = var_7929_begin_0, end = var_7929_end_0, end_mask = var_7929_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7929_cast_fp16")];
            tensor<int32, [4]> var_7944_begin_0 = const()[name = string("op_7944_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7944_end_0 = const()[name = string("op_7944_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7944_end_mask_0 = const()[name = string("op_7944_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7944_cast_fp16 = slice_by_index(begin = var_7944_begin_0, end = var_7944_end_0, end_mask = var_7944_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7944_cast_fp16")];
            int32 var_7946 = const()[name = string("op_7946"), 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_7946, interleave = mixed_79_interleave_0, values = (var_7929_cast_fp16, var_7944_cast_fp16))[name = string("mixed_79_cast_fp16")];
            tensor<int32, [4]> var_7950_perm_0 = const()[name = string("op_7950_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7954 = const()[name = string("op_7954"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7950_cast_fp16 = transpose(perm = var_7950_perm_0, x = mixed_79_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 16, 2048]> input_193_cast_fp16 = reshape(shape = var_7954, x = var_7950_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_397_cast_fp16 = add(x = x_381_cast_fp16, y = linear_136_cast_fp16)[name = string("x_397_cast_fp16")];
            bool var_7964 = const()[name = string("op_7964"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7967_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_7967_cast_fp16")];
            tensor<int32, [1]> peak_159_axes_0 = const()[name = string("peak_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_159_cast_fp16 = reduce_max(axes = peak_159_axes_0, keep_dims = var_7964, x = var_7967_cast_fp16)[name = string("peak_159_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_399_cast_fp16 = real_div(x = x_397_cast_fp16, y = peak_159_cast_fp16)[name = string("x_399_cast_fp16")];
            fp16 var_7963_promoted_to_fp16 = const()[name = string("op_7963_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7971_cast_fp16 = pow(x = x_399_cast_fp16, y = var_7963_promoted_to_fp16)[name = string("op_7971_cast_fp16")];
            tensor<int32, [1]> var_7973_axes_0 = const()[name = string("op_7973_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7973_keep_dims_0 = const()[name = string("op_7973_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7973_cast_fp16 = reduce_mean(axes = var_7973_axes_0, keep_dims = var_7973_keep_dims_0, x = var_7971_cast_fp16)[name = string("op_7973_cast_fp16")];
            fp16 var_7963_promoted_1_to_fp16 = const()[name = string("op_7963_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7974_cast_fp16 = pow(x = peak_159_cast_fp16, y = var_7963_promoted_1_to_fp16)[name = string("op_7974_cast_fp16")];
            fp32 var_7975_epsilon_0 = const()[name = string("op_7975_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7975_cast_fp16 = inverse(epsilon = var_7975_epsilon_0, x = var_7974_cast_fp16)[name = string("op_7975_cast_fp16")];
            fp16 var_7976_to_fp16 = const()[name = string("op_7976_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7977_cast_fp16 = mul(x = var_7975_cast_fp16, y = var_7976_to_fp16)[name = string("op_7977_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7978_cast_fp16 = add(x = var_7973_cast_fp16, y = var_7977_cast_fp16)[name = string("op_7978_cast_fp16")];
            fp32 var_7979_epsilon_0 = const()[name = string("op_7979_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7979_cast_fp16 = rsqrt(epsilon = var_7979_epsilon_0, x = var_7978_cast_fp16)[name = string("op_7979_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7980_cast_fp16 = mul(x = x_399_cast_fp16, y = var_7979_cast_fp16)[name = string("op_7980_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_7980_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_7985_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_7985_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_7985_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_401_cast_fp16 = add(x = x_397_cast_fp16, y = linear_139_cast_fp16)[name = string("x_401_cast_fp16")];
            bool var_7998 = const()[name = string("op_7998"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8001_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<int32, [1]> peak_161_axes_0 = const()[name = string("peak_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_161_cast_fp16 = reduce_max(axes = peak_161_axes_0, keep_dims = var_7998, x = var_8001_cast_fp16)[name = string("peak_161_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_403_cast_fp16 = real_div(x = x_401_cast_fp16, y = peak_161_cast_fp16)[name = string("x_403_cast_fp16")];
            fp16 var_7997_promoted_to_fp16 = const()[name = string("op_7997_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8005_cast_fp16 = pow(x = x_403_cast_fp16, y = var_7997_promoted_to_fp16)[name = string("op_8005_cast_fp16")];
            tensor<int32, [1]> var_8007_axes_0 = const()[name = string("op_8007_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8007_keep_dims_0 = const()[name = string("op_8007_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8007_cast_fp16 = reduce_mean(axes = var_8007_axes_0, keep_dims = var_8007_keep_dims_0, x = var_8005_cast_fp16)[name = string("op_8007_cast_fp16")];
            fp16 var_7997_promoted_1_to_fp16 = const()[name = string("op_7997_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8008_cast_fp16 = pow(x = peak_161_cast_fp16, y = var_7997_promoted_1_to_fp16)[name = string("op_8008_cast_fp16")];
            fp32 var_8009_epsilon_0 = const()[name = string("op_8009_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8009_cast_fp16 = inverse(epsilon = var_8009_epsilon_0, x = var_8008_cast_fp16)[name = string("op_8009_cast_fp16")];
            fp16 var_8010_to_fp16 = const()[name = string("op_8010_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8011_cast_fp16 = mul(x = var_8009_cast_fp16, y = var_8010_to_fp16)[name = string("op_8011_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8012_cast_fp16 = add(x = var_8007_cast_fp16, y = var_8011_cast_fp16)[name = string("op_8012_cast_fp16")];
            fp32 var_8013_epsilon_0 = const()[name = string("op_8013_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8013_cast_fp16 = rsqrt(epsilon = var_8013_epsilon_0, x = var_8012_cast_fp16)[name = string("op_8013_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8014_cast_fp16 = mul(x = x_403_cast_fp16, y = var_8013_cast_fp16)[name = string("op_8014_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_8014_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_8023 = const()[name = string("op_8023"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8024_cast_fp16 = reshape(shape = var_8023, x = linear_140_cast_fp16)[name = string("op_8024_cast_fp16")];
            tensor<int32, [4]> x_405_perm_0 = const()[name = string("x_405_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8031 = const()[name = string("op_8031"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_405_cast_fp16 = transpose(perm = x_405_perm_0, x = var_8024_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 16, 128]> var_8034_cast_fp16 = abs(x = x_405_cast_fp16)[name = string("op_8034_cast_fp16")];
            tensor<int32, [1]> peak_163_axes_0 = const()[name = string("peak_163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_163_cast_fp16 = reduce_max(axes = peak_163_axes_0, keep_dims = var_8031, x = var_8034_cast_fp16)[name = string("peak_163_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_407_cast_fp16 = real_div(x = x_405_cast_fp16, y = peak_163_cast_fp16)[name = string("x_407_cast_fp16")];
            fp16 var_8030_promoted_to_fp16 = const()[name = string("op_8030_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8038_cast_fp16 = pow(x = x_407_cast_fp16, y = var_8030_promoted_to_fp16)[name = string("op_8038_cast_fp16")];
            tensor<int32, [1]> var_8040_axes_0 = const()[name = string("op_8040_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8040_keep_dims_0 = const()[name = string("op_8040_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8040_cast_fp16 = reduce_mean(axes = var_8040_axes_0, keep_dims = var_8040_keep_dims_0, x = var_8038_cast_fp16)[name = string("op_8040_cast_fp16")];
            fp16 var_8030_promoted_1_to_fp16 = const()[name = string("op_8030_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8041_cast_fp16 = pow(x = peak_163_cast_fp16, y = var_8030_promoted_1_to_fp16)[name = string("op_8041_cast_fp16")];
            fp32 var_8042_epsilon_0 = const()[name = string("op_8042_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8042_cast_fp16 = inverse(epsilon = var_8042_epsilon_0, x = var_8041_cast_fp16)[name = string("op_8042_cast_fp16")];
            fp16 var_8043_to_fp16 = const()[name = string("op_8043_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8044_cast_fp16 = mul(x = var_8042_cast_fp16, y = var_8043_to_fp16)[name = string("op_8044_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8045_cast_fp16 = add(x = var_8040_cast_fp16, y = var_8044_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp32 var_8046_epsilon_0 = const()[name = string("op_8046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8046_cast_fp16 = rsqrt(epsilon = var_8046_epsilon_0, x = var_8045_cast_fp16)[name = string("op_8046_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8047_cast_fp16 = mul(x = x_407_cast_fp16, y = var_8046_cast_fp16)[name = string("op_8047_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_413_cast_fp16 = mul(x = var_8047_cast_fp16, y = model_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("x_413_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_8056 = const()[name = string("op_8056"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8057_cast_fp16 = reshape(shape = var_8056, x = linear_141_cast_fp16)[name = string("op_8057_cast_fp16")];
            tensor<int32, [4]> x_409_perm_0 = const()[name = string("x_409_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8064 = const()[name = string("op_8064"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_409_cast_fp16 = transpose(perm = x_409_perm_0, x = var_8057_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, 16, 128]> var_8067_cast_fp16 = abs(x = x_409_cast_fp16)[name = string("op_8067_cast_fp16")];
            tensor<int32, [1]> peak_165_axes_0 = const()[name = string("peak_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_165_cast_fp16 = reduce_max(axes = peak_165_axes_0, keep_dims = var_8064, x = var_8067_cast_fp16)[name = string("peak_165_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_411_cast_fp16 = real_div(x = x_409_cast_fp16, y = peak_165_cast_fp16)[name = string("x_411_cast_fp16")];
            fp16 var_8063_promoted_to_fp16 = const()[name = string("op_8063_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8071_cast_fp16 = pow(x = x_411_cast_fp16, y = var_8063_promoted_to_fp16)[name = string("op_8071_cast_fp16")];
            tensor<int32, [1]> var_8073_axes_0 = const()[name = string("op_8073_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8073_keep_dims_0 = const()[name = string("op_8073_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8073_cast_fp16 = reduce_mean(axes = var_8073_axes_0, keep_dims = var_8073_keep_dims_0, x = var_8071_cast_fp16)[name = string("op_8073_cast_fp16")];
            fp16 var_8063_promoted_1_to_fp16 = const()[name = string("op_8063_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8074_cast_fp16 = pow(x = peak_165_cast_fp16, y = var_8063_promoted_1_to_fp16)[name = string("op_8074_cast_fp16")];
            fp32 var_8075_epsilon_0 = const()[name = string("op_8075_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8075_cast_fp16 = inverse(epsilon = var_8075_epsilon_0, x = var_8074_cast_fp16)[name = string("op_8075_cast_fp16")];
            fp16 var_8076_to_fp16 = const()[name = string("op_8076_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8077_cast_fp16 = mul(x = var_8075_cast_fp16, y = var_8076_to_fp16)[name = string("op_8077_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8078_cast_fp16 = add(x = var_8073_cast_fp16, y = var_8077_cast_fp16)[name = string("op_8078_cast_fp16")];
            fp32 var_8079_epsilon_0 = const()[name = string("op_8079_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8079_cast_fp16 = rsqrt(epsilon = var_8079_epsilon_0, x = var_8078_cast_fp16)[name = string("op_8079_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8080_cast_fp16 = mul(x = x_411_cast_fp16, y = var_8079_cast_fp16)[name = string("op_8080_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_415_cast_fp16 = mul(x = var_8080_cast_fp16, y = model_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("x_415_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_8089 = const()[name = string("op_8089"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8090_cast_fp16 = reshape(shape = var_8089, x = linear_142_cast_fp16)[name = string("op_8090_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_413_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_413_cast_fp16)[name = string("second_81_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8104_cast_fp16 = mul(x = first_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8104_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8105_cast_fp16 = mul(x = second_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8107_cast_fp16 = sub(x = var_8104_cast_fp16, y = var_8105_cast_fp16)[name = string("op_8107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8108_cast_fp16 = mul(x = second_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8108_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8109_cast_fp16 = mul(x = first_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8111_cast_fp16 = add(x = var_8108_cast_fp16, y = var_8109_cast_fp16)[name = string("op_8111_cast_fp16")];
            int32 var_8113 = const()[name = string("op_8113"), 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_8113, interleave = q_41_interleave_0, values = (var_8107_cast_fp16, var_8111_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_415_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_415_cast_fp16)[name = string("second_83_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8125_cast_fp16 = mul(x = first_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8125_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8126_cast_fp16 = mul(x = second_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8126_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8128_cast_fp16 = sub(x = var_8125_cast_fp16, y = var_8126_cast_fp16)[name = string("op_8128_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8129_cast_fp16 = mul(x = second_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8129_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8130_cast_fp16 = mul(x = first_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8130_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8132_cast_fp16 = add(x = var_8129_cast_fp16, y = var_8130_cast_fp16)[name = string("op_8132_cast_fp16")];
            int32 var_8134 = const()[name = string("op_8134"), 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_8134, interleave = k_41_interleave_0, values = (var_8128_cast_fp16, var_8132_cast_fp16))[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_8145_begin_0 = const()[name = string("op_8145_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8145_end_0 = const()[name = string("op_8145_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8145_end_mask_0 = const()[name = string("op_8145_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8145_stride_0 = const()[name = string("op_8145_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8145_cast_fp16 = slice_by_index(begin = var_8145_begin_0, end = var_8145_end_0, end_mask = var_8145_end_mask_0, stride = var_8145_stride_0, x = q_41_cast_fp16)[name = string("op_8145_cast_fp16")];
            tensor<int32, [4]> var_8155_begin_0 = const()[name = string("op_8155_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8155_end_0 = const()[name = string("op_8155_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8155_end_mask_0 = const()[name = string("op_8155_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8155_stride_0 = const()[name = string("op_8155_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8155_cast_fp16 = slice_by_index(begin = var_8155_begin_0, end = var_8155_end_0, end_mask = var_8155_end_mask_0, stride = var_8155_stride_0, x = q_41_cast_fp16)[name = string("op_8155_cast_fp16")];
            int32 var_8157 = const()[name = string("op_8157"), 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_8157, interleave = stacked_41_interleave_0, values = (var_8145_cast_fp16, var_8155_cast_fp16))[name = string("stacked_41_cast_fp16")];
            tensor<int32, [4]> var_8163_begin_0 = const()[name = string("op_8163_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_8163_end_0 = const()[name = string("op_8163_end_0"), val = tensor<int32, [4]>([21, 8, 2064, 128])];
            tensor<bool, [4]> var_8163_end_mask_0 = const()[name = string("op_8163_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, x = read_state_0)[name = string("op_8163_cast_fp16")];
            tensor<int32, [4]> var_8173_begin_0 = const()[name = string("op_8173_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8173_end_0 = const()[name = string("op_8173_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8173_end_mask_0 = const()[name = string("op_8173_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8173_cast_fp16 = slice_by_index(begin = var_8173_begin_0, end = var_8173_end_0, end_mask = var_8173_end_mask_0, x = var_8163_cast_fp16)[name = string("op_8173_cast_fp16")];
            bool var_8182_transpose_x_1 = const()[name = string("op_8182_transpose_x_1"), val = bool(false)];
            bool var_8182_transpose_y_1 = const()[name = string("op_8182_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_8182_cast_fp16 = matmul(transpose_x = var_8182_transpose_x_1, transpose_y = var_8182_transpose_y_1, x = var_8173_cast_fp16, y = stacked_41_cast_fp16)[name = string("op_8182_cast_fp16")];
            tensor<int32, [4]> var_8185_perm_0 = const()[name = string("op_8185_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8186_to_fp16 = const()[name = string("op_8186_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_8185_cast_fp16 = transpose(perm = var_8185_perm_0, x = var_8182_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 32, 2048]> var_8187_cast_fp16 = mul(x = var_8185_cast_fp16, y = var_8186_to_fp16)[name = string("op_8187_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_41_cast_fp16 = add(x = var_8187_cast_fp16, y = cacheMask)[name = string("cache_weights_41_cast_fp16")];
            bool var_8193_transpose_x_1 = const()[name = string("op_8193_transpose_x_1"), val = bool(false)];
            bool var_8193_transpose_y_1 = const()[name = string("op_8193_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8193_cast_fp16 = matmul(transpose_x = var_8193_transpose_x_1, transpose_y = var_8193_transpose_y_1, x = stacked_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_41_cast_fp16 = add(x = var_8195_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_41_cast_fp16")];
            int32 var_8199 = const()[name = string("op_8199"), val = int32(-1)];
            bool var_8200_interleave_0 = const()[name = string("op_8200_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_8200_cast_fp16 = concat(axis = var_8199, interleave = var_8200_interleave_0, values = (cache_weights_41_cast_fp16, call_weights_41_cast_fp16))[name = string("op_8200_cast_fp16")];
            int32 var_8201 = const()[name = string("op_8201"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_41_cast_fp16 = softmax(axis = var_8201, x = var_8200_cast_fp16)[name = string("shares_41_cast_fp16")];
            tensor<int32, [4]> var_8208_begin_0 = const()[name = string("op_8208_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8208_end_0 = const()[name = string("op_8208_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_8208_end_mask_0 = const()[name = string("op_8208_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_8208_cast_fp16 = slice_by_index(begin = var_8208_begin_0, end = var_8208_end_0, end_mask = var_8208_end_mask_0, x = shares_41_cast_fp16)[name = string("op_8208_cast_fp16")];
            tensor<int32, [4]> var_8213_begin_0 = const()[name = string("op_8213_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_8213_end_0 = const()[name = string("op_8213_end_0"), val = tensor<int32, [4]>([21, 8, 2064, 128])];
            tensor<bool, [4]> var_8213_end_mask_0 = const()[name = string("op_8213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8213_cast_fp16 = slice_by_index(begin = var_8213_begin_0, end = var_8213_end_0, end_mask = var_8213_end_mask_0, x = read_state_1)[name = string("op_8213_cast_fp16")];
            tensor<int32, [4]> var_8223_begin_0 = const()[name = string("op_8223_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8223_end_0 = const()[name = string("op_8223_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8223_end_mask_0 = const()[name = string("op_8223_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8223_cast_fp16 = slice_by_index(begin = var_8223_begin_0, end = var_8223_end_0, end_mask = var_8223_end_mask_0, x = var_8213_cast_fp16)[name = string("op_8223_cast_fp16")];
            bool var_8229_transpose_x_0 = const()[name = string("op_8229_transpose_x_0"), val = bool(false)];
            bool var_8229_transpose_y_0 = const()[name = string("op_8229_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8229_cast_fp16 = matmul(transpose_x = var_8229_transpose_x_0, transpose_y = var_8229_transpose_y_0, x = var_8208_cast_fp16, y = var_8223_cast_fp16)[name = string("op_8229_cast_fp16")];
            tensor<int32, [4]> var_8234_begin_0 = const()[name = string("op_8234_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_8234_end_0 = const()[name = string("op_8234_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_8234_end_mask_0 = const()[name = string("op_8234_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8234_cast_fp16 = slice_by_index(begin = var_8234_begin_0, end = var_8234_end_0, end_mask = var_8234_end_mask_0, x = shares_41_cast_fp16)[name = string("op_8234_cast_fp16")];
            bool var_8235_transpose_x_0 = const()[name = string("op_8235_transpose_x_0"), val = bool(false)];
            bool var_8235_transpose_y_0 = const()[name = string("op_8235_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_8090_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 8, 32, 128]> var_8235_cast_fp16 = matmul(transpose_x = var_8235_transpose_x_0, transpose_y = var_8235_transpose_y_0, x = var_8234_cast_fp16, y = v_41_cast_fp16)[name = string("op_8235_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_81_cast_fp16 = add(x = var_8229_cast_fp16, y = var_8235_cast_fp16)[name = string("mixed_81_cast_fp16")];
            tensor<int32, [4]> var_8252_begin_0 = const()[name = string("op_8252_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8252_end_0 = const()[name = string("op_8252_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8252_end_mask_0 = const()[name = string("op_8252_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8252_cast_fp16 = slice_by_index(begin = var_8252_begin_0, end = var_8252_end_0, end_mask = var_8252_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8267_begin_0 = const()[name = string("op_8267_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8267_end_0 = const()[name = string("op_8267_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8267_end_mask_0 = const()[name = string("op_8267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8267_cast_fp16 = slice_by_index(begin = var_8267_begin_0, end = var_8267_end_0, end_mask = var_8267_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_8267_cast_fp16")];
            int32 var_8269 = const()[name = string("op_8269"), 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_8269, interleave = mixed_83_interleave_0, values = (var_8252_cast_fp16, var_8267_cast_fp16))[name = string("mixed_83_cast_fp16")];
            tensor<int32, [4]> var_8273_perm_0 = const()[name = string("op_8273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8277 = const()[name = string("op_8277"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8273_cast_fp16 = transpose(perm = var_8273_perm_0, x = mixed_83_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 2048]> input_203_cast_fp16 = reshape(shape = var_8277, x = var_8273_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_417_cast_fp16 = add(x = x_401_cast_fp16, y = linear_143_cast_fp16)[name = string("x_417_cast_fp16")];
            bool var_8287 = const()[name = string("op_8287"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8290_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_8290_cast_fp16")];
            tensor<int32, [1]> peak_167_axes_0 = const()[name = string("peak_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_167_cast_fp16 = reduce_max(axes = peak_167_axes_0, keep_dims = var_8287, x = var_8290_cast_fp16)[name = string("peak_167_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_419_cast_fp16 = real_div(x = x_417_cast_fp16, y = peak_167_cast_fp16)[name = string("x_419_cast_fp16")];
            fp16 var_8286_promoted_to_fp16 = const()[name = string("op_8286_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8294_cast_fp16 = pow(x = x_419_cast_fp16, y = var_8286_promoted_to_fp16)[name = string("op_8294_cast_fp16")];
            tensor<int32, [1]> var_8296_axes_0 = const()[name = string("op_8296_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8296_keep_dims_0 = const()[name = string("op_8296_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8296_cast_fp16 = reduce_mean(axes = var_8296_axes_0, keep_dims = var_8296_keep_dims_0, x = var_8294_cast_fp16)[name = string("op_8296_cast_fp16")];
            fp16 var_8286_promoted_1_to_fp16 = const()[name = string("op_8286_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8297_cast_fp16 = pow(x = peak_167_cast_fp16, y = var_8286_promoted_1_to_fp16)[name = string("op_8297_cast_fp16")];
            fp32 var_8298_epsilon_0 = const()[name = string("op_8298_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8298_cast_fp16 = inverse(epsilon = var_8298_epsilon_0, x = var_8297_cast_fp16)[name = string("op_8298_cast_fp16")];
            fp16 var_8299_to_fp16 = const()[name = string("op_8299_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8300_cast_fp16 = mul(x = var_8298_cast_fp16, y = var_8299_to_fp16)[name = string("op_8300_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8301_cast_fp16 = add(x = var_8296_cast_fp16, y = var_8300_cast_fp16)[name = string("op_8301_cast_fp16")];
            fp32 var_8302_epsilon_0 = const()[name = string("op_8302_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8302_cast_fp16 = rsqrt(epsilon = var_8302_epsilon_0, x = var_8301_cast_fp16)[name = string("op_8302_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8303_cast_fp16 = mul(x = x_419_cast_fp16, y = var_8302_cast_fp16)[name = string("op_8303_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_8303_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_8308_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_8308_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_8308_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_421_cast_fp16 = add(x = x_417_cast_fp16, y = linear_146_cast_fp16)[name = string("x_421_cast_fp16")];
            bool var_8321 = const()[name = string("op_8321"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8324_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<int32, [1]> peak_169_axes_0 = const()[name = string("peak_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_169_cast_fp16 = reduce_max(axes = peak_169_axes_0, keep_dims = var_8321, x = var_8324_cast_fp16)[name = string("peak_169_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_423_cast_fp16 = real_div(x = x_421_cast_fp16, y = peak_169_cast_fp16)[name = string("x_423_cast_fp16")];
            fp16 var_8320_promoted_to_fp16 = const()[name = string("op_8320_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8328_cast_fp16 = pow(x = x_423_cast_fp16, y = var_8320_promoted_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<int32, [1]> var_8330_axes_0 = const()[name = string("op_8330_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8330_keep_dims_0 = const()[name = string("op_8330_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8330_cast_fp16 = reduce_mean(axes = var_8330_axes_0, keep_dims = var_8330_keep_dims_0, x = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp16 var_8320_promoted_1_to_fp16 = const()[name = string("op_8320_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8331_cast_fp16 = pow(x = peak_169_cast_fp16, y = var_8320_promoted_1_to_fp16)[name = string("op_8331_cast_fp16")];
            fp32 var_8332_epsilon_0 = const()[name = string("op_8332_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8332_cast_fp16 = inverse(epsilon = var_8332_epsilon_0, x = var_8331_cast_fp16)[name = string("op_8332_cast_fp16")];
            fp16 var_8333_to_fp16 = const()[name = string("op_8333_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8334_cast_fp16 = mul(x = var_8332_cast_fp16, y = var_8333_to_fp16)[name = string("op_8334_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8335_cast_fp16 = add(x = var_8330_cast_fp16, y = var_8334_cast_fp16)[name = string("op_8335_cast_fp16")];
            fp32 var_8336_epsilon_0 = const()[name = string("op_8336_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8336_cast_fp16 = rsqrt(epsilon = var_8336_epsilon_0, x = var_8335_cast_fp16)[name = string("op_8336_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8337_cast_fp16 = mul(x = x_423_cast_fp16, y = var_8336_cast_fp16)[name = string("op_8337_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_8337_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_8346 = const()[name = string("op_8346"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8347_cast_fp16 = reshape(shape = var_8346, x = linear_147_cast_fp16)[name = string("op_8347_cast_fp16")];
            tensor<int32, [4]> x_425_perm_0 = const()[name = string("x_425_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8354 = const()[name = string("op_8354"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_425_cast_fp16 = transpose(perm = x_425_perm_0, x = var_8347_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 16, 128]> var_8357_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_8357_cast_fp16")];
            tensor<int32, [1]> peak_171_axes_0 = const()[name = string("peak_171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_171_cast_fp16 = reduce_max(axes = peak_171_axes_0, keep_dims = var_8354, x = var_8357_cast_fp16)[name = string("peak_171_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_427_cast_fp16 = real_div(x = x_425_cast_fp16, y = peak_171_cast_fp16)[name = string("x_427_cast_fp16")];
            fp16 var_8353_promoted_to_fp16 = const()[name = string("op_8353_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8361_cast_fp16 = pow(x = x_427_cast_fp16, y = var_8353_promoted_to_fp16)[name = string("op_8361_cast_fp16")];
            tensor<int32, [1]> var_8363_axes_0 = const()[name = string("op_8363_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8363_keep_dims_0 = const()[name = string("op_8363_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8363_cast_fp16 = reduce_mean(axes = var_8363_axes_0, keep_dims = var_8363_keep_dims_0, x = var_8361_cast_fp16)[name = string("op_8363_cast_fp16")];
            fp16 var_8353_promoted_1_to_fp16 = const()[name = string("op_8353_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8364_cast_fp16 = pow(x = peak_171_cast_fp16, y = var_8353_promoted_1_to_fp16)[name = string("op_8364_cast_fp16")];
            fp32 var_8365_epsilon_0 = const()[name = string("op_8365_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8365_cast_fp16 = inverse(epsilon = var_8365_epsilon_0, x = var_8364_cast_fp16)[name = string("op_8365_cast_fp16")];
            fp16 var_8366_to_fp16 = const()[name = string("op_8366_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8367_cast_fp16 = mul(x = var_8365_cast_fp16, y = var_8366_to_fp16)[name = string("op_8367_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8368_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8367_cast_fp16)[name = string("op_8368_cast_fp16")];
            fp32 var_8369_epsilon_0 = const()[name = string("op_8369_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8369_cast_fp16 = rsqrt(epsilon = var_8369_epsilon_0, x = var_8368_cast_fp16)[name = string("op_8369_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8370_cast_fp16 = mul(x = x_427_cast_fp16, y = var_8369_cast_fp16)[name = string("op_8370_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_433_cast_fp16 = mul(x = var_8370_cast_fp16, y = model_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("x_433_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_8379 = const()[name = string("op_8379"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8380_cast_fp16 = reshape(shape = var_8379, x = linear_148_cast_fp16)[name = string("op_8380_cast_fp16")];
            tensor<int32, [4]> x_429_perm_0 = const()[name = string("x_429_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8387 = const()[name = string("op_8387"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_429_cast_fp16 = transpose(perm = x_429_perm_0, x = var_8380_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 16, 128]> var_8390_cast_fp16 = abs(x = x_429_cast_fp16)[name = string("op_8390_cast_fp16")];
            tensor<int32, [1]> peak_173_axes_0 = const()[name = string("peak_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_173_cast_fp16 = reduce_max(axes = peak_173_axes_0, keep_dims = var_8387, x = var_8390_cast_fp16)[name = string("peak_173_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_431_cast_fp16 = real_div(x = x_429_cast_fp16, y = peak_173_cast_fp16)[name = string("x_431_cast_fp16")];
            fp16 var_8386_promoted_to_fp16 = const()[name = string("op_8386_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8394_cast_fp16 = pow(x = x_431_cast_fp16, y = var_8386_promoted_to_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [1]> var_8396_axes_0 = const()[name = string("op_8396_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8396_keep_dims_0 = const()[name = string("op_8396_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8396_cast_fp16 = reduce_mean(axes = var_8396_axes_0, keep_dims = var_8396_keep_dims_0, x = var_8394_cast_fp16)[name = string("op_8396_cast_fp16")];
            fp16 var_8386_promoted_1_to_fp16 = const()[name = string("op_8386_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8397_cast_fp16 = pow(x = peak_173_cast_fp16, y = var_8386_promoted_1_to_fp16)[name = string("op_8397_cast_fp16")];
            fp32 var_8398_epsilon_0 = const()[name = string("op_8398_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8398_cast_fp16 = inverse(epsilon = var_8398_epsilon_0, x = var_8397_cast_fp16)[name = string("op_8398_cast_fp16")];
            fp16 var_8399_to_fp16 = const()[name = string("op_8399_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8400_cast_fp16 = mul(x = var_8398_cast_fp16, y = var_8399_to_fp16)[name = string("op_8400_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8401_cast_fp16 = add(x = var_8396_cast_fp16, y = var_8400_cast_fp16)[name = string("op_8401_cast_fp16")];
            fp32 var_8402_epsilon_0 = const()[name = string("op_8402_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8402_cast_fp16 = rsqrt(epsilon = var_8402_epsilon_0, x = var_8401_cast_fp16)[name = string("op_8402_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8403_cast_fp16 = mul(x = x_431_cast_fp16, y = var_8402_cast_fp16)[name = string("op_8403_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_435_cast_fp16 = mul(x = var_8403_cast_fp16, y = model_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("x_435_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_8412 = const()[name = string("op_8412"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8413_cast_fp16 = reshape(shape = var_8412, x = linear_149_cast_fp16)[name = string("op_8413_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_433_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_433_cast_fp16)[name = string("second_85_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8427_cast_fp16 = mul(x = first_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8427_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8428_cast_fp16 = mul(x = second_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8428_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8430_cast_fp16 = sub(x = var_8427_cast_fp16, y = var_8428_cast_fp16)[name = string("op_8430_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8431_cast_fp16 = mul(x = second_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8432_cast_fp16 = mul(x = first_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8432_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8434_cast_fp16 = add(x = var_8431_cast_fp16, y = var_8432_cast_fp16)[name = string("op_8434_cast_fp16")];
            int32 var_8436 = const()[name = string("op_8436"), 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_8436, interleave = q_43_interleave_0, values = (var_8430_cast_fp16, var_8434_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_435_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_435_cast_fp16)[name = string("second_87_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8448_cast_fp16 = mul(x = first_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8448_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8449_cast_fp16 = mul(x = second_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8449_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8451_cast_fp16 = sub(x = var_8448_cast_fp16, y = var_8449_cast_fp16)[name = string("op_8451_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8452_cast_fp16 = mul(x = second_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8452_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8453_cast_fp16 = mul(x = first_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8453_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8455_cast_fp16 = add(x = var_8452_cast_fp16, y = var_8453_cast_fp16)[name = string("op_8455_cast_fp16")];
            int32 var_8457 = const()[name = string("op_8457"), 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_8457, interleave = k_43_interleave_0, values = (var_8451_cast_fp16, var_8455_cast_fp16))[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> var_8468_begin_0 = const()[name = string("op_8468_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8468_end_0 = const()[name = string("op_8468_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8468_end_mask_0 = const()[name = string("op_8468_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8468_stride_0 = const()[name = string("op_8468_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8468_cast_fp16 = slice_by_index(begin = var_8468_begin_0, end = var_8468_end_0, end_mask = var_8468_end_mask_0, stride = var_8468_stride_0, x = q_43_cast_fp16)[name = string("op_8468_cast_fp16")];
            tensor<int32, [4]> var_8478_begin_0 = const()[name = string("op_8478_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8478_end_0 = const()[name = string("op_8478_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8478_end_mask_0 = const()[name = string("op_8478_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8478_stride_0 = const()[name = string("op_8478_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8478_cast_fp16 = slice_by_index(begin = var_8478_begin_0, end = var_8478_end_0, end_mask = var_8478_end_mask_0, stride = var_8478_stride_0, x = q_43_cast_fp16)[name = string("op_8478_cast_fp16")];
            int32 var_8480 = const()[name = string("op_8480"), 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_8480, interleave = stacked_43_interleave_0, values = (var_8468_cast_fp16, var_8478_cast_fp16))[name = string("stacked_43_cast_fp16")];
            tensor<int32, [4]> var_8486_begin_0 = const()[name = string("op_8486_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_8486_end_0 = const()[name = string("op_8486_end_0"), val = tensor<int32, [4]>([22, 8, 2064, 128])];
            tensor<bool, [4]> var_8486_end_mask_0 = const()[name = string("op_8486_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8486_cast_fp16 = slice_by_index(begin = var_8486_begin_0, end = var_8486_end_0, end_mask = var_8486_end_mask_0, x = read_state_0)[name = string("op_8486_cast_fp16")];
            tensor<int32, [4]> var_8496_begin_0 = const()[name = string("op_8496_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8496_end_0 = const()[name = string("op_8496_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8496_end_mask_0 = const()[name = string("op_8496_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8496_cast_fp16 = slice_by_index(begin = var_8496_begin_0, end = var_8496_end_0, end_mask = var_8496_end_mask_0, x = var_8486_cast_fp16)[name = string("op_8496_cast_fp16")];
            bool var_8505_transpose_x_1 = const()[name = string("op_8505_transpose_x_1"), val = bool(false)];
            bool var_8505_transpose_y_1 = const()[name = string("op_8505_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_8505_cast_fp16 = matmul(transpose_x = var_8505_transpose_x_1, transpose_y = var_8505_transpose_y_1, x = var_8496_cast_fp16, y = stacked_43_cast_fp16)[name = string("op_8505_cast_fp16")];
            tensor<int32, [4]> var_8508_perm_0 = const()[name = string("op_8508_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8509_to_fp16 = const()[name = string("op_8509_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_8508_cast_fp16 = transpose(perm = var_8508_perm_0, x = var_8505_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 32, 2048]> var_8510_cast_fp16 = mul(x = var_8508_cast_fp16, y = var_8509_to_fp16)[name = string("op_8510_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_43_cast_fp16 = add(x = var_8510_cast_fp16, y = cacheMask)[name = string("cache_weights_43_cast_fp16")];
            bool var_8516_transpose_x_1 = const()[name = string("op_8516_transpose_x_1"), val = bool(false)];
            bool var_8516_transpose_y_1 = const()[name = string("op_8516_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8516_cast_fp16 = matmul(transpose_x = var_8516_transpose_x_1, transpose_y = var_8516_transpose_y_1, x = stacked_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_8516_cast_fp16")];
            fp16 var_8517_to_fp16 = const()[name = string("op_8517_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8518_cast_fp16 = mul(x = var_8516_cast_fp16, y = var_8517_to_fp16)[name = string("op_8518_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_43_cast_fp16 = add(x = var_8518_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_43_cast_fp16")];
            int32 var_8522 = const()[name = string("op_8522"), val = int32(-1)];
            bool var_8523_interleave_0 = const()[name = string("op_8523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_8523_cast_fp16 = concat(axis = var_8522, interleave = var_8523_interleave_0, values = (cache_weights_43_cast_fp16, call_weights_43_cast_fp16))[name = string("op_8523_cast_fp16")];
            int32 var_8524 = const()[name = string("op_8524"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_43_cast_fp16 = softmax(axis = var_8524, x = var_8523_cast_fp16)[name = string("shares_43_cast_fp16")];
            tensor<int32, [4]> var_8531_begin_0 = const()[name = string("op_8531_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8531_end_0 = const()[name = string("op_8531_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_8531_end_mask_0 = const()[name = string("op_8531_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_8531_cast_fp16 = slice_by_index(begin = var_8531_begin_0, end = var_8531_end_0, end_mask = var_8531_end_mask_0, x = shares_43_cast_fp16)[name = string("op_8531_cast_fp16")];
            tensor<int32, [4]> var_8536_begin_0 = const()[name = string("op_8536_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_8536_end_0 = const()[name = string("op_8536_end_0"), val = tensor<int32, [4]>([22, 8, 2064, 128])];
            tensor<bool, [4]> var_8536_end_mask_0 = const()[name = string("op_8536_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8536_cast_fp16 = slice_by_index(begin = var_8536_begin_0, end = var_8536_end_0, end_mask = var_8536_end_mask_0, x = read_state_1)[name = string("op_8536_cast_fp16")];
            tensor<int32, [4]> var_8546_begin_0 = const()[name = string("op_8546_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8546_end_0 = const()[name = string("op_8546_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8546_end_mask_0 = const()[name = string("op_8546_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8546_cast_fp16 = slice_by_index(begin = var_8546_begin_0, end = var_8546_end_0, end_mask = var_8546_end_mask_0, x = var_8536_cast_fp16)[name = string("op_8546_cast_fp16")];
            bool var_8552_transpose_x_0 = const()[name = string("op_8552_transpose_x_0"), val = bool(false)];
            bool var_8552_transpose_y_0 = const()[name = string("op_8552_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8552_cast_fp16 = matmul(transpose_x = var_8552_transpose_x_0, transpose_y = var_8552_transpose_y_0, x = var_8531_cast_fp16, y = var_8546_cast_fp16)[name = string("op_8552_cast_fp16")];
            tensor<int32, [4]> var_8557_begin_0 = const()[name = string("op_8557_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_8557_end_0 = const()[name = string("op_8557_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_8557_end_mask_0 = const()[name = string("op_8557_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8557_cast_fp16 = slice_by_index(begin = var_8557_begin_0, end = var_8557_end_0, end_mask = var_8557_end_mask_0, x = shares_43_cast_fp16)[name = string("op_8557_cast_fp16")];
            bool var_8558_transpose_x_0 = const()[name = string("op_8558_transpose_x_0"), val = bool(false)];
            bool var_8558_transpose_y_0 = const()[name = string("op_8558_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_8413_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 8, 32, 128]> var_8558_cast_fp16 = matmul(transpose_x = var_8558_transpose_x_0, transpose_y = var_8558_transpose_y_0, x = var_8557_cast_fp16, y = v_43_cast_fp16)[name = string("op_8558_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_85_cast_fp16 = add(x = var_8552_cast_fp16, y = var_8558_cast_fp16)[name = string("mixed_85_cast_fp16")];
            tensor<int32, [4]> var_8575_begin_0 = const()[name = string("op_8575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8575_end_0 = const()[name = string("op_8575_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8575_end_mask_0 = const()[name = string("op_8575_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8575_cast_fp16 = slice_by_index(begin = var_8575_begin_0, end = var_8575_end_0, end_mask = var_8575_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_8575_cast_fp16")];
            tensor<int32, [4]> var_8590_begin_0 = const()[name = string("op_8590_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8590_end_0 = const()[name = string("op_8590_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8590_end_mask_0 = const()[name = string("op_8590_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8590_cast_fp16 = slice_by_index(begin = var_8590_begin_0, end = var_8590_end_0, end_mask = var_8590_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_8590_cast_fp16")];
            int32 var_8592 = const()[name = string("op_8592"), 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_8592, interleave = mixed_87_interleave_0, values = (var_8575_cast_fp16, var_8590_cast_fp16))[name = string("mixed_87_cast_fp16")];
            tensor<int32, [4]> var_8596_perm_0 = const()[name = string("op_8596_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8600 = const()[name = string("op_8600"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8596_cast_fp16 = transpose(perm = var_8596_perm_0, x = mixed_87_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 2048]> input_213_cast_fp16 = reshape(shape = var_8600, x = var_8596_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_437_cast_fp16 = add(x = x_421_cast_fp16, y = linear_150_cast_fp16)[name = string("x_437_cast_fp16")];
            bool var_8610 = const()[name = string("op_8610"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8613_cast_fp16 = abs(x = x_437_cast_fp16)[name = string("op_8613_cast_fp16")];
            tensor<int32, [1]> peak_175_axes_0 = const()[name = string("peak_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_175_cast_fp16 = reduce_max(axes = peak_175_axes_0, keep_dims = var_8610, x = var_8613_cast_fp16)[name = string("peak_175_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_439_cast_fp16 = real_div(x = x_437_cast_fp16, y = peak_175_cast_fp16)[name = string("x_439_cast_fp16")];
            fp16 var_8609_promoted_to_fp16 = const()[name = string("op_8609_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8617_cast_fp16 = pow(x = x_439_cast_fp16, y = var_8609_promoted_to_fp16)[name = string("op_8617_cast_fp16")];
            tensor<int32, [1]> var_8619_axes_0 = const()[name = string("op_8619_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8619_keep_dims_0 = const()[name = string("op_8619_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8619_cast_fp16 = reduce_mean(axes = var_8619_axes_0, keep_dims = var_8619_keep_dims_0, x = var_8617_cast_fp16)[name = string("op_8619_cast_fp16")];
            fp16 var_8609_promoted_1_to_fp16 = const()[name = string("op_8609_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8620_cast_fp16 = pow(x = peak_175_cast_fp16, y = var_8609_promoted_1_to_fp16)[name = string("op_8620_cast_fp16")];
            fp32 var_8621_epsilon_0 = const()[name = string("op_8621_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8621_cast_fp16 = inverse(epsilon = var_8621_epsilon_0, x = var_8620_cast_fp16)[name = string("op_8621_cast_fp16")];
            fp16 var_8622_to_fp16 = const()[name = string("op_8622_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8623_cast_fp16 = mul(x = var_8621_cast_fp16, y = var_8622_to_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8624_cast_fp16 = add(x = var_8619_cast_fp16, y = var_8623_cast_fp16)[name = string("op_8624_cast_fp16")];
            fp32 var_8625_epsilon_0 = const()[name = string("op_8625_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8625_cast_fp16 = rsqrt(epsilon = var_8625_epsilon_0, x = var_8624_cast_fp16)[name = string("op_8625_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8626_cast_fp16 = mul(x = x_439_cast_fp16, y = var_8625_cast_fp16)[name = string("op_8626_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_8626_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_8631_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_8631_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_8631_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_441_cast_fp16 = add(x = x_437_cast_fp16, y = linear_153_cast_fp16)[name = string("x_441_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8647_cast_fp16 = abs(x = x_441_cast_fp16)[name = string("op_8647_cast_fp16")];
            tensor<int32, [1]> peak_177_axes_0 = const()[name = string("peak_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_177_cast_fp16 = reduce_max(axes = peak_177_axes_0, keep_dims = var_8644, x = var_8647_cast_fp16)[name = string("peak_177_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_443_cast_fp16 = real_div(x = x_441_cast_fp16, y = peak_177_cast_fp16)[name = string("x_443_cast_fp16")];
            fp16 var_8643_promoted_to_fp16 = const()[name = string("op_8643_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8651_cast_fp16 = pow(x = x_443_cast_fp16, y = var_8643_promoted_to_fp16)[name = string("op_8651_cast_fp16")];
            tensor<int32, [1]> var_8653_axes_0 = const()[name = string("op_8653_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8653_keep_dims_0 = const()[name = string("op_8653_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8653_cast_fp16 = reduce_mean(axes = var_8653_axes_0, keep_dims = var_8653_keep_dims_0, x = var_8651_cast_fp16)[name = string("op_8653_cast_fp16")];
            fp16 var_8643_promoted_1_to_fp16 = const()[name = string("op_8643_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8654_cast_fp16 = pow(x = peak_177_cast_fp16, y = var_8643_promoted_1_to_fp16)[name = string("op_8654_cast_fp16")];
            fp32 var_8655_epsilon_0 = const()[name = string("op_8655_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8655_cast_fp16 = inverse(epsilon = var_8655_epsilon_0, x = var_8654_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp16 var_8656_to_fp16 = const()[name = string("op_8656_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8657_cast_fp16 = mul(x = var_8655_cast_fp16, y = var_8656_to_fp16)[name = string("op_8657_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8658_cast_fp16 = add(x = var_8653_cast_fp16, y = var_8657_cast_fp16)[name = string("op_8658_cast_fp16")];
            fp32 var_8659_epsilon_0 = const()[name = string("op_8659_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8659_cast_fp16 = rsqrt(epsilon = var_8659_epsilon_0, x = var_8658_cast_fp16)[name = string("op_8659_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8660_cast_fp16 = mul(x = x_443_cast_fp16, y = var_8659_cast_fp16)[name = string("op_8660_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_8660_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_8669 = const()[name = string("op_8669"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8670_cast_fp16 = reshape(shape = var_8669, x = linear_154_cast_fp16)[name = string("op_8670_cast_fp16")];
            tensor<int32, [4]> x_445_perm_0 = const()[name = string("x_445_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8677 = const()[name = string("op_8677"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_445_cast_fp16 = transpose(perm = x_445_perm_0, x = var_8670_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 16, 128]> var_8680_cast_fp16 = abs(x = x_445_cast_fp16)[name = string("op_8680_cast_fp16")];
            tensor<int32, [1]> peak_179_axes_0 = const()[name = string("peak_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_179_cast_fp16 = reduce_max(axes = peak_179_axes_0, keep_dims = var_8677, x = var_8680_cast_fp16)[name = string("peak_179_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_447_cast_fp16 = real_div(x = x_445_cast_fp16, y = peak_179_cast_fp16)[name = string("x_447_cast_fp16")];
            fp16 var_8676_promoted_to_fp16 = const()[name = string("op_8676_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8684_cast_fp16 = pow(x = x_447_cast_fp16, y = var_8676_promoted_to_fp16)[name = string("op_8684_cast_fp16")];
            tensor<int32, [1]> var_8686_axes_0 = const()[name = string("op_8686_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8686_keep_dims_0 = const()[name = string("op_8686_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8686_cast_fp16 = reduce_mean(axes = var_8686_axes_0, keep_dims = var_8686_keep_dims_0, x = var_8684_cast_fp16)[name = string("op_8686_cast_fp16")];
            fp16 var_8676_promoted_1_to_fp16 = const()[name = string("op_8676_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8687_cast_fp16 = pow(x = peak_179_cast_fp16, y = var_8676_promoted_1_to_fp16)[name = string("op_8687_cast_fp16")];
            fp32 var_8688_epsilon_0 = const()[name = string("op_8688_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8688_cast_fp16 = inverse(epsilon = var_8688_epsilon_0, x = var_8687_cast_fp16)[name = string("op_8688_cast_fp16")];
            fp16 var_8689_to_fp16 = const()[name = string("op_8689_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8690_cast_fp16 = mul(x = var_8688_cast_fp16, y = var_8689_to_fp16)[name = string("op_8690_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8691_cast_fp16 = add(x = var_8686_cast_fp16, y = var_8690_cast_fp16)[name = string("op_8691_cast_fp16")];
            fp32 var_8692_epsilon_0 = const()[name = string("op_8692_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8692_cast_fp16 = rsqrt(epsilon = var_8692_epsilon_0, x = var_8691_cast_fp16)[name = string("op_8692_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8693_cast_fp16 = mul(x = x_447_cast_fp16, y = var_8692_cast_fp16)[name = string("op_8693_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_453_cast_fp16 = mul(x = var_8693_cast_fp16, y = model_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("x_453_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_8702 = const()[name = string("op_8702"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8703_cast_fp16 = reshape(shape = var_8702, x = linear_155_cast_fp16)[name = string("op_8703_cast_fp16")];
            tensor<int32, [4]> x_449_perm_0 = const()[name = string("x_449_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8710 = const()[name = string("op_8710"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_449_cast_fp16 = transpose(perm = x_449_perm_0, x = var_8703_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 8, 16, 128]> var_8713_cast_fp16 = abs(x = x_449_cast_fp16)[name = string("op_8713_cast_fp16")];
            tensor<int32, [1]> peak_181_axes_0 = const()[name = string("peak_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_181_cast_fp16 = reduce_max(axes = peak_181_axes_0, keep_dims = var_8710, x = var_8713_cast_fp16)[name = string("peak_181_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_451_cast_fp16 = real_div(x = x_449_cast_fp16, y = peak_181_cast_fp16)[name = string("x_451_cast_fp16")];
            fp16 var_8709_promoted_to_fp16 = const()[name = string("op_8709_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8717_cast_fp16 = pow(x = x_451_cast_fp16, y = var_8709_promoted_to_fp16)[name = string("op_8717_cast_fp16")];
            tensor<int32, [1]> var_8719_axes_0 = const()[name = string("op_8719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8719_keep_dims_0 = const()[name = string("op_8719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8719_cast_fp16 = reduce_mean(axes = var_8719_axes_0, keep_dims = var_8719_keep_dims_0, x = var_8717_cast_fp16)[name = string("op_8719_cast_fp16")];
            fp16 var_8709_promoted_1_to_fp16 = const()[name = string("op_8709_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8720_cast_fp16 = pow(x = peak_181_cast_fp16, y = var_8709_promoted_1_to_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8721_cast_fp16 = inverse(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            fp16 var_8722_to_fp16 = const()[name = string("op_8722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8723_cast_fp16 = mul(x = var_8721_cast_fp16, y = var_8722_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8724_cast_fp16 = add(x = var_8719_cast_fp16, y = var_8723_cast_fp16)[name = string("op_8724_cast_fp16")];
            fp32 var_8725_epsilon_0 = const()[name = string("op_8725_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8725_cast_fp16 = rsqrt(epsilon = var_8725_epsilon_0, x = var_8724_cast_fp16)[name = string("op_8725_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8726_cast_fp16 = mul(x = x_451_cast_fp16, y = var_8725_cast_fp16)[name = string("op_8726_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_455_cast_fp16 = mul(x = var_8726_cast_fp16, y = model_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("x_455_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_8735 = const()[name = string("op_8735"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8736_cast_fp16 = reshape(shape = var_8735, x = linear_156_cast_fp16)[name = string("op_8736_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_453_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_453_cast_fp16)[name = string("second_89_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8750_cast_fp16 = mul(x = first_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8750_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8751_cast_fp16 = mul(x = second_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8751_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8753_cast_fp16 = sub(x = var_8750_cast_fp16, y = var_8751_cast_fp16)[name = string("op_8753_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8754_cast_fp16 = mul(x = second_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8755_cast_fp16 = mul(x = first_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8755_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8757_cast_fp16 = add(x = var_8754_cast_fp16, y = var_8755_cast_fp16)[name = string("op_8757_cast_fp16")];
            int32 var_8759 = const()[name = string("op_8759"), 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_8759, interleave = q_45_interleave_0, values = (var_8753_cast_fp16, var_8757_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_455_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_455_cast_fp16)[name = string("second_91_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8771_cast_fp16 = mul(x = first_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8771_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8772_cast_fp16 = mul(x = second_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8772_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8774_cast_fp16 = sub(x = var_8771_cast_fp16, y = var_8772_cast_fp16)[name = string("op_8774_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8775_cast_fp16 = mul(x = second_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8776_cast_fp16 = mul(x = first_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8776_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8778_cast_fp16 = add(x = var_8775_cast_fp16, y = var_8776_cast_fp16)[name = string("op_8778_cast_fp16")];
            int32 var_8780 = const()[name = string("op_8780"), 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_8780, interleave = k_45_interleave_0, values = (var_8774_cast_fp16, var_8778_cast_fp16))[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_8791_begin_0 = const()[name = string("op_8791_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8791_end_0 = const()[name = string("op_8791_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8791_end_mask_0 = const()[name = string("op_8791_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8791_stride_0 = const()[name = string("op_8791_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8791_cast_fp16 = slice_by_index(begin = var_8791_begin_0, end = var_8791_end_0, end_mask = var_8791_end_mask_0, stride = var_8791_stride_0, x = q_45_cast_fp16)[name = string("op_8791_cast_fp16")];
            tensor<int32, [4]> var_8801_begin_0 = const()[name = string("op_8801_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8801_end_0 = const()[name = string("op_8801_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8801_end_mask_0 = const()[name = string("op_8801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8801_stride_0 = const()[name = string("op_8801_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8801_cast_fp16 = slice_by_index(begin = var_8801_begin_0, end = var_8801_end_0, end_mask = var_8801_end_mask_0, stride = var_8801_stride_0, x = q_45_cast_fp16)[name = string("op_8801_cast_fp16")];
            int32 var_8803 = const()[name = string("op_8803"), 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_8803, interleave = stacked_45_interleave_0, values = (var_8791_cast_fp16, var_8801_cast_fp16))[name = string("stacked_45_cast_fp16")];
            tensor<int32, [4]> var_8809_begin_0 = const()[name = string("op_8809_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8809_end_0 = const()[name = string("op_8809_end_0"), val = tensor<int32, [4]>([23, 8, 2064, 128])];
            tensor<bool, [4]> var_8809_end_mask_0 = const()[name = string("op_8809_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8809_cast_fp16 = slice_by_index(begin = var_8809_begin_0, end = var_8809_end_0, end_mask = var_8809_end_mask_0, x = read_state_0)[name = string("op_8809_cast_fp16")];
            tensor<int32, [4]> var_8819_begin_0 = const()[name = string("op_8819_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8819_end_0 = const()[name = string("op_8819_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8819_end_mask_0 = const()[name = string("op_8819_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8819_cast_fp16 = slice_by_index(begin = var_8819_begin_0, end = var_8819_end_0, end_mask = var_8819_end_mask_0, x = var_8809_cast_fp16)[name = string("op_8819_cast_fp16")];
            bool var_8828_transpose_x_1 = const()[name = string("op_8828_transpose_x_1"), val = bool(false)];
            bool var_8828_transpose_y_1 = const()[name = string("op_8828_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_8828_cast_fp16 = matmul(transpose_x = var_8828_transpose_x_1, transpose_y = var_8828_transpose_y_1, x = var_8819_cast_fp16, y = stacked_45_cast_fp16)[name = string("op_8828_cast_fp16")];
            tensor<int32, [4]> var_8831_perm_0 = const()[name = string("op_8831_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8832_to_fp16 = const()[name = string("op_8832_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_8831_cast_fp16 = transpose(perm = var_8831_perm_0, x = var_8828_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 8, 32, 2048]> var_8833_cast_fp16 = mul(x = var_8831_cast_fp16, y = var_8832_to_fp16)[name = string("op_8833_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_45_cast_fp16 = add(x = var_8833_cast_fp16, y = cacheMask)[name = string("cache_weights_45_cast_fp16")];
            bool var_8839_transpose_x_1 = const()[name = string("op_8839_transpose_x_1"), val = bool(false)];
            bool var_8839_transpose_y_1 = const()[name = string("op_8839_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8839_cast_fp16 = matmul(transpose_x = var_8839_transpose_x_1, transpose_y = var_8839_transpose_y_1, x = stacked_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_8839_cast_fp16")];
            fp16 var_8840_to_fp16 = const()[name = string("op_8840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8841_cast_fp16 = mul(x = var_8839_cast_fp16, y = var_8840_to_fp16)[name = string("op_8841_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_45_cast_fp16 = add(x = var_8841_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_45_cast_fp16")];
            int32 var_8845 = const()[name = string("op_8845"), val = int32(-1)];
            bool var_8846_interleave_0 = const()[name = string("op_8846_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_8846_cast_fp16 = concat(axis = var_8845, interleave = var_8846_interleave_0, values = (cache_weights_45_cast_fp16, call_weights_45_cast_fp16))[name = string("op_8846_cast_fp16")];
            int32 var_8847 = const()[name = string("op_8847"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_45_cast_fp16 = softmax(axis = var_8847, x = var_8846_cast_fp16)[name = string("shares_45_cast_fp16")];
            tensor<int32, [4]> var_8854_begin_0 = const()[name = string("op_8854_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8854_end_0 = const()[name = string("op_8854_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_8854_end_mask_0 = const()[name = string("op_8854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8854_cast_fp16")];
            tensor<int32, [4]> var_8859_begin_0 = const()[name = string("op_8859_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8859_end_0 = const()[name = string("op_8859_end_0"), val = tensor<int32, [4]>([23, 8, 2064, 128])];
            tensor<bool, [4]> var_8859_end_mask_0 = const()[name = string("op_8859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8859_cast_fp16 = slice_by_index(begin = var_8859_begin_0, end = var_8859_end_0, end_mask = var_8859_end_mask_0, x = read_state_1)[name = string("op_8859_cast_fp16")];
            tensor<int32, [4]> var_8869_begin_0 = const()[name = string("op_8869_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8869_end_0 = const()[name = string("op_8869_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8869_end_mask_0 = const()[name = string("op_8869_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8869_cast_fp16 = slice_by_index(begin = var_8869_begin_0, end = var_8869_end_0, end_mask = var_8869_end_mask_0, x = var_8859_cast_fp16)[name = string("op_8869_cast_fp16")];
            bool var_8875_transpose_x_0 = const()[name = string("op_8875_transpose_x_0"), val = bool(false)];
            bool var_8875_transpose_y_0 = const()[name = string("op_8875_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8875_cast_fp16 = matmul(transpose_x = var_8875_transpose_x_0, transpose_y = var_8875_transpose_y_0, x = var_8854_cast_fp16, y = var_8869_cast_fp16)[name = string("op_8875_cast_fp16")];
            tensor<int32, [4]> var_8880_begin_0 = const()[name = string("op_8880_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_8880_end_0 = const()[name = string("op_8880_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_8880_end_mask_0 = const()[name = string("op_8880_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8880_cast_fp16 = slice_by_index(begin = var_8880_begin_0, end = var_8880_end_0, end_mask = var_8880_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8880_cast_fp16")];
            bool var_8881_transpose_x_0 = const()[name = string("op_8881_transpose_x_0"), val = bool(false)];
            bool var_8881_transpose_y_0 = const()[name = string("op_8881_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_8736_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, 32, 128]> var_8881_cast_fp16 = matmul(transpose_x = var_8881_transpose_x_0, transpose_y = var_8881_transpose_y_0, x = var_8880_cast_fp16, y = v_45_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_89_cast_fp16 = add(x = var_8875_cast_fp16, y = var_8881_cast_fp16)[name = string("mixed_89_cast_fp16")];
            tensor<int32, [4]> var_8898_begin_0 = const()[name = string("op_8898_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8898_end_0 = const()[name = string("op_8898_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8898_end_mask_0 = const()[name = string("op_8898_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8898_cast_fp16 = slice_by_index(begin = var_8898_begin_0, end = var_8898_end_0, end_mask = var_8898_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8898_cast_fp16")];
            tensor<int32, [4]> var_8913_begin_0 = const()[name = string("op_8913_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8913_end_0 = const()[name = string("op_8913_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8913_end_mask_0 = const()[name = string("op_8913_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8913_cast_fp16 = slice_by_index(begin = var_8913_begin_0, end = var_8913_end_0, end_mask = var_8913_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8913_cast_fp16")];
            int32 var_8915 = const()[name = string("op_8915"), 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_8915, interleave = mixed_91_interleave_0, values = (var_8898_cast_fp16, var_8913_cast_fp16))[name = string("mixed_91_cast_fp16")];
            tensor<int32, [4]> var_8919_perm_0 = const()[name = string("op_8919_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8923 = const()[name = string("op_8923"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8919_cast_fp16 = transpose(perm = var_8919_perm_0, x = mixed_91_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 2048]> input_223_cast_fp16 = reshape(shape = var_8923, x = var_8919_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_457_cast_fp16 = add(x = x_441_cast_fp16, y = linear_157_cast_fp16)[name = string("x_457_cast_fp16")];
            bool var_8933 = const()[name = string("op_8933"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8936_cast_fp16 = abs(x = x_457_cast_fp16)[name = string("op_8936_cast_fp16")];
            tensor<int32, [1]> peak_183_axes_0 = const()[name = string("peak_183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_183_cast_fp16 = reduce_max(axes = peak_183_axes_0, keep_dims = var_8933, x = var_8936_cast_fp16)[name = string("peak_183_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_459_cast_fp16 = real_div(x = x_457_cast_fp16, y = peak_183_cast_fp16)[name = string("x_459_cast_fp16")];
            fp16 var_8932_promoted_to_fp16 = const()[name = string("op_8932_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8940_cast_fp16 = pow(x = x_459_cast_fp16, y = var_8932_promoted_to_fp16)[name = string("op_8940_cast_fp16")];
            tensor<int32, [1]> var_8942_axes_0 = const()[name = string("op_8942_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8942_keep_dims_0 = const()[name = string("op_8942_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8942_cast_fp16 = reduce_mean(axes = var_8942_axes_0, keep_dims = var_8942_keep_dims_0, x = var_8940_cast_fp16)[name = string("op_8942_cast_fp16")];
            fp16 var_8932_promoted_1_to_fp16 = const()[name = string("op_8932_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8943_cast_fp16 = pow(x = peak_183_cast_fp16, y = var_8932_promoted_1_to_fp16)[name = string("op_8943_cast_fp16")];
            fp32 var_8944_epsilon_0 = const()[name = string("op_8944_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8944_cast_fp16 = inverse(epsilon = var_8944_epsilon_0, x = var_8943_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, 1]> var_8946_cast_fp16 = mul(x = var_8944_cast_fp16, y = var_8945_to_fp16)[name = string("op_8946_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8947_cast_fp16 = add(x = var_8942_cast_fp16, y = var_8946_cast_fp16)[name = string("op_8947_cast_fp16")];
            fp32 var_8948_epsilon_0 = const()[name = string("op_8948_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8948_cast_fp16 = rsqrt(epsilon = var_8948_epsilon_0, x = var_8947_cast_fp16)[name = string("op_8948_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8949_cast_fp16 = mul(x = x_459_cast_fp16, y = var_8948_cast_fp16)[name = string("op_8949_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_8949_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_8954_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_8954_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_8954_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_461_cast_fp16 = add(x = x_457_cast_fp16, y = linear_160_cast_fp16)[name = string("x_461_cast_fp16")];
            bool var_8967 = const()[name = string("op_8967"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8970_cast_fp16 = abs(x = x_461_cast_fp16)[name = string("op_8970_cast_fp16")];
            tensor<int32, [1]> peak_185_axes_0 = const()[name = string("peak_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_185_cast_fp16 = reduce_max(axes = peak_185_axes_0, keep_dims = var_8967, x = var_8970_cast_fp16)[name = string("peak_185_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_463_cast_fp16 = real_div(x = x_461_cast_fp16, y = peak_185_cast_fp16)[name = string("x_463_cast_fp16")];
            fp16 var_8966_promoted_to_fp16 = const()[name = string("op_8966_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8974_cast_fp16 = pow(x = x_463_cast_fp16, y = var_8966_promoted_to_fp16)[name = string("op_8974_cast_fp16")];
            tensor<int32, [1]> var_8976_axes_0 = const()[name = string("op_8976_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8976_keep_dims_0 = const()[name = string("op_8976_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8976_cast_fp16 = reduce_mean(axes = var_8976_axes_0, keep_dims = var_8976_keep_dims_0, x = var_8974_cast_fp16)[name = string("op_8976_cast_fp16")];
            fp16 var_8966_promoted_1_to_fp16 = const()[name = string("op_8966_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8977_cast_fp16 = pow(x = peak_185_cast_fp16, y = var_8966_promoted_1_to_fp16)[name = string("op_8977_cast_fp16")];
            fp32 var_8978_epsilon_0 = const()[name = string("op_8978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8978_cast_fp16 = inverse(epsilon = var_8978_epsilon_0, x = var_8977_cast_fp16)[name = string("op_8978_cast_fp16")];
            fp16 var_8979_to_fp16 = const()[name = string("op_8979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8980_cast_fp16 = mul(x = var_8978_cast_fp16, y = var_8979_to_fp16)[name = string("op_8980_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8981_cast_fp16 = add(x = var_8976_cast_fp16, y = var_8980_cast_fp16)[name = string("op_8981_cast_fp16")];
            fp32 var_8982_epsilon_0 = const()[name = string("op_8982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8982_cast_fp16 = rsqrt(epsilon = var_8982_epsilon_0, x = var_8981_cast_fp16)[name = string("op_8982_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8983_cast_fp16 = mul(x = x_463_cast_fp16, y = var_8982_cast_fp16)[name = string("op_8983_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_8983_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_8992 = const()[name = string("op_8992"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8993_cast_fp16 = reshape(shape = var_8992, x = linear_161_cast_fp16)[name = string("op_8993_cast_fp16")];
            tensor<int32, [4]> x_465_perm_0 = const()[name = string("x_465_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9000 = const()[name = string("op_9000"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_465_cast_fp16 = transpose(perm = x_465_perm_0, x = var_8993_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 16, 16, 128]> var_9003_cast_fp16 = abs(x = x_465_cast_fp16)[name = string("op_9003_cast_fp16")];
            tensor<int32, [1]> peak_187_axes_0 = const()[name = string("peak_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_187_cast_fp16 = reduce_max(axes = peak_187_axes_0, keep_dims = var_9000, x = var_9003_cast_fp16)[name = string("peak_187_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_467_cast_fp16 = real_div(x = x_465_cast_fp16, y = peak_187_cast_fp16)[name = string("x_467_cast_fp16")];
            fp16 var_8999_promoted_to_fp16 = const()[name = string("op_8999_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9007_cast_fp16 = pow(x = x_467_cast_fp16, y = var_8999_promoted_to_fp16)[name = string("op_9007_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9007_cast_fp16)[name = string("op_9009_cast_fp16")];
            fp16 var_8999_promoted_1_to_fp16 = const()[name = string("op_8999_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9010_cast_fp16 = pow(x = peak_187_cast_fp16, y = var_8999_promoted_1_to_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9014_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9014_cast_fp16")];
            fp32 var_9015_epsilon_0 = const()[name = string("op_9015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9015_cast_fp16 = rsqrt(epsilon = var_9015_epsilon_0, x = var_9014_cast_fp16)[name = string("op_9015_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9016_cast_fp16 = mul(x = x_467_cast_fp16, y = var_9015_cast_fp16)[name = string("op_9016_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_473_cast_fp16 = mul(x = var_9016_cast_fp16, y = model_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("x_473_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_9025 = const()[name = string("op_9025"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9026_cast_fp16 = reshape(shape = var_9025, x = linear_162_cast_fp16)[name = string("op_9026_cast_fp16")];
            tensor<int32, [4]> x_469_perm_0 = const()[name = string("x_469_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9033 = const()[name = string("op_9033"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_469_cast_fp16 = transpose(perm = x_469_perm_0, x = var_9026_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 16, 128]> var_9036_cast_fp16 = abs(x = x_469_cast_fp16)[name = string("op_9036_cast_fp16")];
            tensor<int32, [1]> peak_189_axes_0 = const()[name = string("peak_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_189_cast_fp16 = reduce_max(axes = peak_189_axes_0, keep_dims = var_9033, x = var_9036_cast_fp16)[name = string("peak_189_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_471_cast_fp16 = real_div(x = x_469_cast_fp16, y = peak_189_cast_fp16)[name = string("x_471_cast_fp16")];
            fp16 var_9032_promoted_to_fp16 = const()[name = string("op_9032_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9040_cast_fp16 = pow(x = x_471_cast_fp16, y = var_9032_promoted_to_fp16)[name = string("op_9040_cast_fp16")];
            tensor<int32, [1]> var_9042_axes_0 = const()[name = string("op_9042_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9042_keep_dims_0 = const()[name = string("op_9042_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9042_cast_fp16 = reduce_mean(axes = var_9042_axes_0, keep_dims = var_9042_keep_dims_0, x = var_9040_cast_fp16)[name = string("op_9042_cast_fp16")];
            fp16 var_9032_promoted_1_to_fp16 = const()[name = string("op_9032_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9043_cast_fp16 = pow(x = peak_189_cast_fp16, y = var_9032_promoted_1_to_fp16)[name = string("op_9043_cast_fp16")];
            fp32 var_9044_epsilon_0 = const()[name = string("op_9044_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9044_cast_fp16 = inverse(epsilon = var_9044_epsilon_0, x = var_9043_cast_fp16)[name = string("op_9044_cast_fp16")];
            fp16 var_9045_to_fp16 = const()[name = string("op_9045_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9046_cast_fp16 = mul(x = var_9044_cast_fp16, y = var_9045_to_fp16)[name = string("op_9046_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9047_cast_fp16 = add(x = var_9042_cast_fp16, y = var_9046_cast_fp16)[name = string("op_9047_cast_fp16")];
            fp32 var_9048_epsilon_0 = const()[name = string("op_9048_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9048_cast_fp16 = rsqrt(epsilon = var_9048_epsilon_0, x = var_9047_cast_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9049_cast_fp16 = mul(x = x_471_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9049_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_475_cast_fp16 = mul(x = var_9049_cast_fp16, y = model_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("x_475_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_9058 = const()[name = string("op_9058"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9059_cast_fp16 = reshape(shape = var_9058, x = linear_163_cast_fp16)[name = string("op_9059_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_473_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_473_cast_fp16)[name = string("second_93_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9073_cast_fp16 = mul(x = first_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9074_cast_fp16 = mul(x = second_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9074_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9076_cast_fp16 = sub(x = var_9073_cast_fp16, y = var_9074_cast_fp16)[name = string("op_9076_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9077_cast_fp16 = mul(x = second_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9077_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9078_cast_fp16 = mul(x = first_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9078_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9080_cast_fp16 = add(x = var_9077_cast_fp16, y = var_9078_cast_fp16)[name = string("op_9080_cast_fp16")];
            int32 var_9082 = const()[name = string("op_9082"), 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_9082, interleave = q_47_interleave_0, values = (var_9076_cast_fp16, var_9080_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_475_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_475_cast_fp16)[name = string("second_95_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9094_cast_fp16 = mul(x = first_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9094_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9095_cast_fp16 = mul(x = second_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9095_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9097_cast_fp16 = sub(x = var_9094_cast_fp16, y = var_9095_cast_fp16)[name = string("op_9097_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9098_cast_fp16 = mul(x = second_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9098_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9099_cast_fp16 = mul(x = first_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9099_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9101_cast_fp16 = add(x = var_9098_cast_fp16, y = var_9099_cast_fp16)[name = string("op_9101_cast_fp16")];
            int32 var_9103 = const()[name = string("op_9103"), 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_9103, interleave = k_47_interleave_0, values = (var_9097_cast_fp16, var_9101_cast_fp16))[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> var_9114_begin_0 = const()[name = string("op_9114_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9114_end_0 = const()[name = string("op_9114_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9114_end_mask_0 = const()[name = string("op_9114_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9114_stride_0 = const()[name = string("op_9114_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9114_cast_fp16 = slice_by_index(begin = var_9114_begin_0, end = var_9114_end_0, end_mask = var_9114_end_mask_0, stride = var_9114_stride_0, x = q_47_cast_fp16)[name = string("op_9114_cast_fp16")];
            tensor<int32, [4]> var_9124_begin_0 = const()[name = string("op_9124_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9124_end_0 = const()[name = string("op_9124_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9124_end_mask_0 = const()[name = string("op_9124_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9124_stride_0 = const()[name = string("op_9124_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9124_cast_fp16 = slice_by_index(begin = var_9124_begin_0, end = var_9124_end_0, end_mask = var_9124_end_mask_0, stride = var_9124_stride_0, x = q_47_cast_fp16)[name = string("op_9124_cast_fp16")];
            int32 var_9126 = const()[name = string("op_9126"), 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_9126, interleave = stacked_47_interleave_0, values = (var_9114_cast_fp16, var_9124_cast_fp16))[name = string("stacked_47_cast_fp16")];
            tensor<int32, [4]> var_9132_begin_0 = const()[name = string("op_9132_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_9132_end_0 = const()[name = string("op_9132_end_0"), val = tensor<int32, [4]>([24, 8, 2064, 128])];
            tensor<bool, [4]> var_9132_end_mask_0 = const()[name = string("op_9132_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9132_cast_fp16 = slice_by_index(begin = var_9132_begin_0, end = var_9132_end_0, end_mask = var_9132_end_mask_0, x = read_state_0)[name = string("op_9132_cast_fp16")];
            tensor<int32, [4]> var_9142_begin_0 = const()[name = string("op_9142_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9142_end_0 = const()[name = string("op_9142_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9142_end_mask_0 = const()[name = string("op_9142_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9142_cast_fp16 = slice_by_index(begin = var_9142_begin_0, end = var_9142_end_0, end_mask = var_9142_end_mask_0, x = var_9132_cast_fp16)[name = string("op_9142_cast_fp16")];
            bool var_9151_transpose_x_1 = const()[name = string("op_9151_transpose_x_1"), val = bool(false)];
            bool var_9151_transpose_y_1 = const()[name = string("op_9151_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_9151_cast_fp16 = matmul(transpose_x = var_9151_transpose_x_1, transpose_y = var_9151_transpose_y_1, x = var_9142_cast_fp16, y = stacked_47_cast_fp16)[name = string("op_9151_cast_fp16")];
            tensor<int32, [4]> var_9154_perm_0 = const()[name = string("op_9154_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9155_to_fp16 = const()[name = string("op_9155_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_9154_cast_fp16 = transpose(perm = var_9154_perm_0, x = var_9151_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, 32, 2048]> var_9156_cast_fp16 = mul(x = var_9154_cast_fp16, y = var_9155_to_fp16)[name = string("op_9156_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_47_cast_fp16 = add(x = var_9156_cast_fp16, y = cacheMask)[name = string("cache_weights_47_cast_fp16")];
            bool var_9162_transpose_x_1 = const()[name = string("op_9162_transpose_x_1"), val = bool(false)];
            bool var_9162_transpose_y_1 = const()[name = string("op_9162_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9162_cast_fp16 = matmul(transpose_x = var_9162_transpose_x_1, transpose_y = var_9162_transpose_y_1, x = stacked_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_9162_cast_fp16")];
            fp16 var_9163_to_fp16 = const()[name = string("op_9163_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9164_cast_fp16 = mul(x = var_9162_cast_fp16, y = var_9163_to_fp16)[name = string("op_9164_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_47_cast_fp16 = add(x = var_9164_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_47_cast_fp16")];
            int32 var_9168 = const()[name = string("op_9168"), val = int32(-1)];
            bool var_9169_interleave_0 = const()[name = string("op_9169_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_9169_cast_fp16 = concat(axis = var_9168, interleave = var_9169_interleave_0, values = (cache_weights_47_cast_fp16, call_weights_47_cast_fp16))[name = string("op_9169_cast_fp16")];
            int32 var_9170 = const()[name = string("op_9170"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_47_cast_fp16 = softmax(axis = var_9170, x = var_9169_cast_fp16)[name = string("shares_47_cast_fp16")];
            tensor<int32, [4]> var_9177_begin_0 = const()[name = string("op_9177_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9177_end_0 = const()[name = string("op_9177_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_9177_end_mask_0 = const()[name = string("op_9177_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_9177_cast_fp16 = slice_by_index(begin = var_9177_begin_0, end = var_9177_end_0, end_mask = var_9177_end_mask_0, x = shares_47_cast_fp16)[name = string("op_9177_cast_fp16")];
            tensor<int32, [4]> var_9182_begin_0 = const()[name = string("op_9182_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_9182_end_0 = const()[name = string("op_9182_end_0"), val = tensor<int32, [4]>([24, 8, 2064, 128])];
            tensor<bool, [4]> var_9182_end_mask_0 = const()[name = string("op_9182_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9182_cast_fp16 = slice_by_index(begin = var_9182_begin_0, end = var_9182_end_0, end_mask = var_9182_end_mask_0, x = read_state_1)[name = string("op_9182_cast_fp16")];
            tensor<int32, [4]> var_9192_begin_0 = const()[name = string("op_9192_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9192_end_0 = const()[name = string("op_9192_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9192_end_mask_0 = const()[name = string("op_9192_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9192_cast_fp16 = slice_by_index(begin = var_9192_begin_0, end = var_9192_end_0, end_mask = var_9192_end_mask_0, x = var_9182_cast_fp16)[name = string("op_9192_cast_fp16")];
            bool var_9198_transpose_x_0 = const()[name = string("op_9198_transpose_x_0"), val = bool(false)];
            bool var_9198_transpose_y_0 = const()[name = string("op_9198_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9198_cast_fp16 = matmul(transpose_x = var_9198_transpose_x_0, transpose_y = var_9198_transpose_y_0, x = var_9177_cast_fp16, y = var_9192_cast_fp16)[name = string("op_9198_cast_fp16")];
            tensor<int32, [4]> var_9203_begin_0 = const()[name = string("op_9203_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_9203_end_0 = const()[name = string("op_9203_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_9203_end_mask_0 = const()[name = string("op_9203_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9203_cast_fp16 = slice_by_index(begin = var_9203_begin_0, end = var_9203_end_0, end_mask = var_9203_end_mask_0, x = shares_47_cast_fp16)[name = string("op_9203_cast_fp16")];
            bool var_9204_transpose_x_0 = const()[name = string("op_9204_transpose_x_0"), val = bool(false)];
            bool var_9204_transpose_y_0 = const()[name = string("op_9204_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_9059_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 8, 32, 128]> var_9204_cast_fp16 = matmul(transpose_x = var_9204_transpose_x_0, transpose_y = var_9204_transpose_y_0, x = var_9203_cast_fp16, y = v_47_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_93_cast_fp16 = add(x = var_9198_cast_fp16, y = var_9204_cast_fp16)[name = string("mixed_93_cast_fp16")];
            tensor<int32, [4]> var_9221_begin_0 = const()[name = string("op_9221_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9221_end_0 = const()[name = string("op_9221_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9221_end_mask_0 = const()[name = string("op_9221_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9221_cast_fp16 = slice_by_index(begin = var_9221_begin_0, end = var_9221_end_0, end_mask = var_9221_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_9221_cast_fp16")];
            tensor<int32, [4]> var_9236_begin_0 = const()[name = string("op_9236_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9236_end_0 = const()[name = string("op_9236_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9236_end_mask_0 = const()[name = string("op_9236_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9236_cast_fp16 = slice_by_index(begin = var_9236_begin_0, end = var_9236_end_0, end_mask = var_9236_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_9236_cast_fp16")];
            int32 var_9238 = const()[name = string("op_9238"), 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_9238, interleave = mixed_95_interleave_0, values = (var_9221_cast_fp16, var_9236_cast_fp16))[name = string("mixed_95_cast_fp16")];
            tensor<int32, [4]> var_9242_perm_0 = const()[name = string("op_9242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9246 = const()[name = string("op_9246"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9242_cast_fp16 = transpose(perm = var_9242_perm_0, x = mixed_95_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 16, 2048]> input_233_cast_fp16 = reshape(shape = var_9246, x = var_9242_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_477_cast_fp16 = add(x = x_461_cast_fp16, y = linear_164_cast_fp16)[name = string("x_477_cast_fp16")];
            bool var_9256 = const()[name = string("op_9256"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9259_cast_fp16 = abs(x = x_477_cast_fp16)[name = string("op_9259_cast_fp16")];
            tensor<int32, [1]> peak_191_axes_0 = const()[name = string("peak_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_191_cast_fp16 = reduce_max(axes = peak_191_axes_0, keep_dims = var_9256, x = var_9259_cast_fp16)[name = string("peak_191_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_479_cast_fp16 = real_div(x = x_477_cast_fp16, y = peak_191_cast_fp16)[name = string("x_479_cast_fp16")];
            fp16 var_9255_promoted_to_fp16 = const()[name = string("op_9255_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9263_cast_fp16 = pow(x = x_479_cast_fp16, y = var_9255_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, 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_9255_promoted_1_to_fp16 = const()[name = string("op_9255_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9266_cast_fp16 = pow(x = peak_191_cast_fp16, y = var_9255_promoted_1_to_fp16)[name = string("op_9266_cast_fp16")];
            fp32 var_9267_epsilon_0 = const()[name = string("op_9267_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9267_cast_fp16 = inverse(epsilon = var_9267_epsilon_0, x = var_9266_cast_fp16)[name = string("op_9267_cast_fp16")];
            fp16 var_9268_to_fp16 = const()[name = string("op_9268_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9269_cast_fp16 = mul(x = var_9267_cast_fp16, y = var_9268_to_fp16)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9270_cast_fp16 = add(x = var_9265_cast_fp16, y = var_9269_cast_fp16)[name = string("op_9270_cast_fp16")];
            fp32 var_9271_epsilon_0 = const()[name = string("op_9271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9271_cast_fp16 = rsqrt(epsilon = var_9271_epsilon_0, x = var_9270_cast_fp16)[name = string("op_9271_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9272_cast_fp16 = mul(x = x_479_cast_fp16, y = var_9271_cast_fp16)[name = string("op_9272_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_9272_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_9277_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_9277_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_9277_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_481_cast_fp16 = add(x = x_477_cast_fp16, y = linear_167_cast_fp16)[name = string("x_481_cast_fp16")];
            bool var_9290 = const()[name = string("op_9290"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9293_cast_fp16 = abs(x = x_481_cast_fp16)[name = string("op_9293_cast_fp16")];
            tensor<int32, [1]> peak_193_axes_0 = const()[name = string("peak_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_193_cast_fp16 = reduce_max(axes = peak_193_axes_0, keep_dims = var_9290, x = var_9293_cast_fp16)[name = string("peak_193_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_483_cast_fp16 = real_div(x = x_481_cast_fp16, y = peak_193_cast_fp16)[name = string("x_483_cast_fp16")];
            fp16 var_9289_promoted_to_fp16 = const()[name = string("op_9289_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9297_cast_fp16 = pow(x = x_483_cast_fp16, y = var_9289_promoted_to_fp16)[name = string("op_9297_cast_fp16")];
            tensor<int32, [1]> var_9299_axes_0 = const()[name = string("op_9299_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9299_keep_dims_0 = const()[name = string("op_9299_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9299_cast_fp16 = reduce_mean(axes = var_9299_axes_0, keep_dims = var_9299_keep_dims_0, x = var_9297_cast_fp16)[name = string("op_9299_cast_fp16")];
            fp16 var_9289_promoted_1_to_fp16 = const()[name = string("op_9289_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9300_cast_fp16 = pow(x = peak_193_cast_fp16, y = var_9289_promoted_1_to_fp16)[name = string("op_9300_cast_fp16")];
            fp32 var_9301_epsilon_0 = const()[name = string("op_9301_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9301_cast_fp16 = inverse(epsilon = var_9301_epsilon_0, x = var_9300_cast_fp16)[name = string("op_9301_cast_fp16")];
            fp16 var_9302_to_fp16 = const()[name = string("op_9302_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9303_cast_fp16 = mul(x = var_9301_cast_fp16, y = var_9302_to_fp16)[name = string("op_9303_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9304_cast_fp16 = add(x = var_9299_cast_fp16, y = var_9303_cast_fp16)[name = string("op_9304_cast_fp16")];
            fp32 var_9305_epsilon_0 = const()[name = string("op_9305_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9305_cast_fp16 = rsqrt(epsilon = var_9305_epsilon_0, x = var_9304_cast_fp16)[name = string("op_9305_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9306_cast_fp16 = mul(x = x_483_cast_fp16, y = var_9305_cast_fp16)[name = string("op_9306_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_9306_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_9315 = const()[name = string("op_9315"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9316_cast_fp16 = reshape(shape = var_9315, x = linear_168_cast_fp16)[name = string("op_9316_cast_fp16")];
            tensor<int32, [4]> x_485_perm_0 = const()[name = string("x_485_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9323 = const()[name = string("op_9323"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_485_cast_fp16 = transpose(perm = x_485_perm_0, x = var_9316_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 16, 128]> var_9326_cast_fp16 = abs(x = x_485_cast_fp16)[name = string("op_9326_cast_fp16")];
            tensor<int32, [1]> peak_195_axes_0 = const()[name = string("peak_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_195_cast_fp16 = reduce_max(axes = peak_195_axes_0, keep_dims = var_9323, x = var_9326_cast_fp16)[name = string("peak_195_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_487_cast_fp16 = real_div(x = x_485_cast_fp16, y = peak_195_cast_fp16)[name = string("x_487_cast_fp16")];
            fp16 var_9322_promoted_to_fp16 = const()[name = string("op_9322_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9330_cast_fp16 = pow(x = x_487_cast_fp16, y = var_9322_promoted_to_fp16)[name = string("op_9330_cast_fp16")];
            tensor<int32, [1]> var_9332_axes_0 = const()[name = string("op_9332_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9332_keep_dims_0 = const()[name = string("op_9332_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9332_cast_fp16 = reduce_mean(axes = var_9332_axes_0, keep_dims = var_9332_keep_dims_0, x = var_9330_cast_fp16)[name = string("op_9332_cast_fp16")];
            fp16 var_9322_promoted_1_to_fp16 = const()[name = string("op_9322_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9333_cast_fp16 = pow(x = peak_195_cast_fp16, y = var_9322_promoted_1_to_fp16)[name = string("op_9333_cast_fp16")];
            fp32 var_9334_epsilon_0 = const()[name = string("op_9334_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9334_cast_fp16 = inverse(epsilon = var_9334_epsilon_0, x = var_9333_cast_fp16)[name = string("op_9334_cast_fp16")];
            fp16 var_9335_to_fp16 = const()[name = string("op_9335_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9336_cast_fp16 = mul(x = var_9334_cast_fp16, y = var_9335_to_fp16)[name = string("op_9336_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9337_cast_fp16 = add(x = var_9332_cast_fp16, y = var_9336_cast_fp16)[name = string("op_9337_cast_fp16")];
            fp32 var_9338_epsilon_0 = const()[name = string("op_9338_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9338_cast_fp16 = rsqrt(epsilon = var_9338_epsilon_0, x = var_9337_cast_fp16)[name = string("op_9338_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9339_cast_fp16 = mul(x = x_487_cast_fp16, y = var_9338_cast_fp16)[name = string("op_9339_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_493_cast_fp16 = mul(x = var_9339_cast_fp16, y = model_layers_24_self_attn_q_norm_weight_to_fp16)[name = string("x_493_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_9348 = const()[name = string("op_9348"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9349_cast_fp16 = reshape(shape = var_9348, x = linear_169_cast_fp16)[name = string("op_9349_cast_fp16")];
            tensor<int32, [4]> x_489_perm_0 = const()[name = string("x_489_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9356 = const()[name = string("op_9356"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_489_cast_fp16 = transpose(perm = x_489_perm_0, x = var_9349_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, 16, 128]> var_9359_cast_fp16 = abs(x = x_489_cast_fp16)[name = string("op_9359_cast_fp16")];
            tensor<int32, [1]> peak_197_axes_0 = const()[name = string("peak_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_197_cast_fp16 = reduce_max(axes = peak_197_axes_0, keep_dims = var_9356, x = var_9359_cast_fp16)[name = string("peak_197_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_491_cast_fp16 = real_div(x = x_489_cast_fp16, y = peak_197_cast_fp16)[name = string("x_491_cast_fp16")];
            fp16 var_9355_promoted_to_fp16 = const()[name = string("op_9355_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9363_cast_fp16 = pow(x = x_491_cast_fp16, y = var_9355_promoted_to_fp16)[name = string("op_9363_cast_fp16")];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9365_keep_dims_0 = const()[name = string("op_9365_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9365_cast_fp16 = reduce_mean(axes = var_9365_axes_0, keep_dims = var_9365_keep_dims_0, x = var_9363_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9355_promoted_1_to_fp16 = const()[name = string("op_9355_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9366_cast_fp16 = pow(x = peak_197_cast_fp16, y = var_9355_promoted_1_to_fp16)[name = string("op_9366_cast_fp16")];
            fp32 var_9367_epsilon_0 = const()[name = string("op_9367_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9367_cast_fp16 = inverse(epsilon = var_9367_epsilon_0, x = var_9366_cast_fp16)[name = string("op_9367_cast_fp16")];
            fp16 var_9368_to_fp16 = const()[name = string("op_9368_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9369_cast_fp16 = mul(x = var_9367_cast_fp16, y = var_9368_to_fp16)[name = string("op_9369_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9370_cast_fp16 = add(x = var_9365_cast_fp16, y = var_9369_cast_fp16)[name = string("op_9370_cast_fp16")];
            fp32 var_9371_epsilon_0 = const()[name = string("op_9371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9371_cast_fp16 = rsqrt(epsilon = var_9371_epsilon_0, x = var_9370_cast_fp16)[name = string("op_9371_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9372_cast_fp16 = mul(x = x_491_cast_fp16, y = var_9371_cast_fp16)[name = string("op_9372_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_495_cast_fp16 = mul(x = var_9372_cast_fp16, y = model_layers_24_self_attn_k_norm_weight_to_fp16)[name = string("x_495_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_9381 = const()[name = string("op_9381"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9382_cast_fp16 = reshape(shape = var_9381, x = linear_170_cast_fp16)[name = string("op_9382_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_493_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_493_cast_fp16)[name = string("second_97_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9396_cast_fp16 = mul(x = first_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9396_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9397_cast_fp16 = mul(x = second_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9397_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9399_cast_fp16 = sub(x = var_9396_cast_fp16, y = var_9397_cast_fp16)[name = string("op_9399_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9400_cast_fp16 = mul(x = second_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9400_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9401_cast_fp16 = mul(x = first_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9401_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9403_cast_fp16 = add(x = var_9400_cast_fp16, y = var_9401_cast_fp16)[name = string("op_9403_cast_fp16")];
            int32 var_9405 = const()[name = string("op_9405"), 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_9405, interleave = q_49_interleave_0, values = (var_9399_cast_fp16, var_9403_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_495_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_495_cast_fp16)[name = string("second_99_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9417_cast_fp16 = mul(x = first_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9417_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9418_cast_fp16 = mul(x = second_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9418_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9420_cast_fp16 = sub(x = var_9417_cast_fp16, y = var_9418_cast_fp16)[name = string("op_9420_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9421_cast_fp16 = mul(x = second_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9422_cast_fp16 = mul(x = first_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9422_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9424_cast_fp16 = add(x = var_9421_cast_fp16, y = var_9422_cast_fp16)[name = string("op_9424_cast_fp16")];
            int32 var_9426 = const()[name = string("op_9426"), 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_9426, interleave = k_49_interleave_0, values = (var_9420_cast_fp16, var_9424_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_9437_begin_0 = const()[name = string("op_9437_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9437_end_0 = const()[name = string("op_9437_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9437_end_mask_0 = const()[name = string("op_9437_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9437_stride_0 = const()[name = string("op_9437_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9437_cast_fp16 = slice_by_index(begin = var_9437_begin_0, end = var_9437_end_0, end_mask = var_9437_end_mask_0, stride = var_9437_stride_0, x = q_49_cast_fp16)[name = string("op_9437_cast_fp16")];
            tensor<int32, [4]> var_9447_begin_0 = const()[name = string("op_9447_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9447_end_0 = const()[name = string("op_9447_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9447_end_mask_0 = const()[name = string("op_9447_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9447_stride_0 = const()[name = string("op_9447_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9447_cast_fp16 = slice_by_index(begin = var_9447_begin_0, end = var_9447_end_0, end_mask = var_9447_end_mask_0, stride = var_9447_stride_0, x = q_49_cast_fp16)[name = string("op_9447_cast_fp16")];
            int32 var_9449 = const()[name = string("op_9449"), 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_9449, interleave = stacked_49_interleave_0, values = (var_9437_cast_fp16, var_9447_cast_fp16))[name = string("stacked_49_cast_fp16")];
            tensor<int32, [4]> var_9455_begin_0 = const()[name = string("op_9455_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_9455_end_0 = const()[name = string("op_9455_end_0"), val = tensor<int32, [4]>([25, 8, 2064, 128])];
            tensor<bool, [4]> var_9455_end_mask_0 = const()[name = string("op_9455_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9455_cast_fp16 = slice_by_index(begin = var_9455_begin_0, end = var_9455_end_0, end_mask = var_9455_end_mask_0, x = read_state_0)[name = string("op_9455_cast_fp16")];
            tensor<int32, [4]> var_9465_begin_0 = const()[name = string("op_9465_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9465_end_0 = const()[name = string("op_9465_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9465_end_mask_0 = const()[name = string("op_9465_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9465_cast_fp16 = slice_by_index(begin = var_9465_begin_0, end = var_9465_end_0, end_mask = var_9465_end_mask_0, x = var_9455_cast_fp16)[name = string("op_9465_cast_fp16")];
            bool var_9474_transpose_x_1 = const()[name = string("op_9474_transpose_x_1"), val = bool(false)];
            bool var_9474_transpose_y_1 = const()[name = string("op_9474_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_9474_cast_fp16 = matmul(transpose_x = var_9474_transpose_x_1, transpose_y = var_9474_transpose_y_1, x = var_9465_cast_fp16, y = stacked_49_cast_fp16)[name = string("op_9474_cast_fp16")];
            tensor<int32, [4]> var_9477_perm_0 = const()[name = string("op_9477_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9478_to_fp16 = const()[name = string("op_9478_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_9477_cast_fp16 = transpose(perm = var_9477_perm_0, x = var_9474_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 32, 2048]> var_9479_cast_fp16 = mul(x = var_9477_cast_fp16, y = var_9478_to_fp16)[name = string("op_9479_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_49_cast_fp16 = add(x = var_9479_cast_fp16, y = cacheMask)[name = string("cache_weights_49_cast_fp16")];
            bool var_9485_transpose_x_1 = const()[name = string("op_9485_transpose_x_1"), val = bool(false)];
            bool var_9485_transpose_y_1 = const()[name = string("op_9485_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9485_cast_fp16 = matmul(transpose_x = var_9485_transpose_x_1, transpose_y = var_9485_transpose_y_1, x = stacked_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_9485_cast_fp16")];
            fp16 var_9486_to_fp16 = const()[name = string("op_9486_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9487_cast_fp16 = mul(x = var_9485_cast_fp16, y = var_9486_to_fp16)[name = string("op_9487_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_49_cast_fp16 = add(x = var_9487_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_49_cast_fp16")];
            int32 var_9491 = const()[name = string("op_9491"), val = int32(-1)];
            bool var_9492_interleave_0 = const()[name = string("op_9492_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_9492_cast_fp16 = concat(axis = var_9491, interleave = var_9492_interleave_0, values = (cache_weights_49_cast_fp16, call_weights_49_cast_fp16))[name = string("op_9492_cast_fp16")];
            int32 var_9493 = const()[name = string("op_9493"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_49_cast_fp16 = softmax(axis = var_9493, x = var_9492_cast_fp16)[name = string("shares_49_cast_fp16")];
            tensor<int32, [4]> var_9500_begin_0 = const()[name = string("op_9500_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9500_end_0 = const()[name = string("op_9500_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_9500_end_mask_0 = const()[name = string("op_9500_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_9500_cast_fp16 = slice_by_index(begin = var_9500_begin_0, end = var_9500_end_0, end_mask = var_9500_end_mask_0, x = shares_49_cast_fp16)[name = string("op_9500_cast_fp16")];
            tensor<int32, [4]> var_9505_begin_0 = const()[name = string("op_9505_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_9505_end_0 = const()[name = string("op_9505_end_0"), val = tensor<int32, [4]>([25, 8, 2064, 128])];
            tensor<bool, [4]> var_9505_end_mask_0 = const()[name = string("op_9505_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9505_cast_fp16 = slice_by_index(begin = var_9505_begin_0, end = var_9505_end_0, end_mask = var_9505_end_mask_0, x = read_state_1)[name = string("op_9505_cast_fp16")];
            tensor<int32, [4]> var_9515_begin_0 = const()[name = string("op_9515_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9515_end_0 = const()[name = string("op_9515_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9515_end_mask_0 = const()[name = string("op_9515_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9515_cast_fp16 = slice_by_index(begin = var_9515_begin_0, end = var_9515_end_0, end_mask = var_9515_end_mask_0, x = var_9505_cast_fp16)[name = string("op_9515_cast_fp16")];
            bool var_9521_transpose_x_0 = const()[name = string("op_9521_transpose_x_0"), val = bool(false)];
            bool var_9521_transpose_y_0 = const()[name = string("op_9521_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9521_cast_fp16 = matmul(transpose_x = var_9521_transpose_x_0, transpose_y = var_9521_transpose_y_0, x = var_9500_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9521_cast_fp16")];
            tensor<int32, [4]> var_9526_begin_0 = const()[name = string("op_9526_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_9526_end_0 = const()[name = string("op_9526_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_9526_end_mask_0 = const()[name = string("op_9526_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9526_cast_fp16 = slice_by_index(begin = var_9526_begin_0, end = var_9526_end_0, end_mask = var_9526_end_mask_0, x = shares_49_cast_fp16)[name = string("op_9526_cast_fp16")];
            bool var_9527_transpose_x_0 = const()[name = string("op_9527_transpose_x_0"), val = bool(false)];
            bool var_9527_transpose_y_0 = const()[name = string("op_9527_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_9382_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 32, 128]> var_9527_cast_fp16 = matmul(transpose_x = var_9527_transpose_x_0, transpose_y = var_9527_transpose_y_0, x = var_9526_cast_fp16, y = v_49_cast_fp16)[name = string("op_9527_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_97_cast_fp16 = add(x = var_9521_cast_fp16, y = var_9527_cast_fp16)[name = string("mixed_97_cast_fp16")];
            tensor<int32, [4]> var_9544_begin_0 = const()[name = string("op_9544_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9544_end_0 = const()[name = string("op_9544_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9544_end_mask_0 = const()[name = string("op_9544_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9544_cast_fp16 = slice_by_index(begin = var_9544_begin_0, end = var_9544_end_0, end_mask = var_9544_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_9544_cast_fp16")];
            tensor<int32, [4]> var_9559_begin_0 = const()[name = string("op_9559_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9559_end_0 = const()[name = string("op_9559_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9559_end_mask_0 = const()[name = string("op_9559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9559_cast_fp16 = slice_by_index(begin = var_9559_begin_0, end = var_9559_end_0, end_mask = var_9559_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_9559_cast_fp16")];
            int32 var_9561 = const()[name = string("op_9561"), 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_9561, interleave = mixed_99_interleave_0, values = (var_9544_cast_fp16, var_9559_cast_fp16))[name = string("mixed_99_cast_fp16")];
            tensor<int32, [4]> var_9565_perm_0 = const()[name = string("op_9565_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9569 = const()[name = string("op_9569"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9565_cast_fp16 = transpose(perm = var_9565_perm_0, x = mixed_99_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 2048]> input_243_cast_fp16 = reshape(shape = var_9569, x = var_9565_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_497_cast_fp16 = add(x = x_481_cast_fp16, y = linear_171_cast_fp16)[name = string("x_497_cast_fp16")];
            bool var_9579 = const()[name = string("op_9579"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9582_cast_fp16 = abs(x = x_497_cast_fp16)[name = string("op_9582_cast_fp16")];
            tensor<int32, [1]> peak_199_axes_0 = const()[name = string("peak_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_199_cast_fp16 = reduce_max(axes = peak_199_axes_0, keep_dims = var_9579, x = var_9582_cast_fp16)[name = string("peak_199_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_499_cast_fp16 = real_div(x = x_497_cast_fp16, y = peak_199_cast_fp16)[name = string("x_499_cast_fp16")];
            fp16 var_9578_promoted_to_fp16 = const()[name = string("op_9578_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9586_cast_fp16 = pow(x = x_499_cast_fp16, y = var_9578_promoted_to_fp16)[name = string("op_9586_cast_fp16")];
            tensor<int32, [1]> var_9588_axes_0 = const()[name = string("op_9588_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9588_keep_dims_0 = const()[name = string("op_9588_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9588_cast_fp16 = reduce_mean(axes = var_9588_axes_0, keep_dims = var_9588_keep_dims_0, x = var_9586_cast_fp16)[name = string("op_9588_cast_fp16")];
            fp16 var_9578_promoted_1_to_fp16 = const()[name = string("op_9578_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9589_cast_fp16 = pow(x = peak_199_cast_fp16, y = var_9578_promoted_1_to_fp16)[name = string("op_9589_cast_fp16")];
            fp32 var_9590_epsilon_0 = const()[name = string("op_9590_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9590_cast_fp16 = inverse(epsilon = var_9590_epsilon_0, x = var_9589_cast_fp16)[name = string("op_9590_cast_fp16")];
            fp16 var_9591_to_fp16 = const()[name = string("op_9591_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9592_cast_fp16 = mul(x = var_9590_cast_fp16, y = var_9591_to_fp16)[name = string("op_9592_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9593_cast_fp16 = add(x = var_9588_cast_fp16, y = var_9592_cast_fp16)[name = string("op_9593_cast_fp16")];
            fp32 var_9594_epsilon_0 = const()[name = string("op_9594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9594_cast_fp16 = rsqrt(epsilon = var_9594_epsilon_0, x = var_9593_cast_fp16)[name = string("op_9594_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9595_cast_fp16 = mul(x = x_499_cast_fp16, y = var_9594_cast_fp16)[name = string("op_9595_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_9595_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_9600_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_9600_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_9600_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_501_cast_fp16 = add(x = x_497_cast_fp16, y = linear_174_cast_fp16)[name = string("x_501_cast_fp16")];
            bool var_9613 = const()[name = string("op_9613"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9616_cast_fp16 = abs(x = x_501_cast_fp16)[name = string("op_9616_cast_fp16")];
            tensor<int32, [1]> peak_201_axes_0 = const()[name = string("peak_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_201_cast_fp16 = reduce_max(axes = peak_201_axes_0, keep_dims = var_9613, x = var_9616_cast_fp16)[name = string("peak_201_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_503_cast_fp16 = real_div(x = x_501_cast_fp16, y = peak_201_cast_fp16)[name = string("x_503_cast_fp16")];
            fp16 var_9612_promoted_to_fp16 = const()[name = string("op_9612_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9620_cast_fp16 = pow(x = x_503_cast_fp16, y = var_9612_promoted_to_fp16)[name = string("op_9620_cast_fp16")];
            tensor<int32, [1]> var_9622_axes_0 = const()[name = string("op_9622_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9622_keep_dims_0 = const()[name = string("op_9622_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9622_cast_fp16 = reduce_mean(axes = var_9622_axes_0, keep_dims = var_9622_keep_dims_0, x = var_9620_cast_fp16)[name = string("op_9622_cast_fp16")];
            fp16 var_9612_promoted_1_to_fp16 = const()[name = string("op_9612_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9623_cast_fp16 = pow(x = peak_201_cast_fp16, y = var_9612_promoted_1_to_fp16)[name = string("op_9623_cast_fp16")];
            fp32 var_9624_epsilon_0 = const()[name = string("op_9624_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9624_cast_fp16 = inverse(epsilon = var_9624_epsilon_0, x = var_9623_cast_fp16)[name = string("op_9624_cast_fp16")];
            fp16 var_9625_to_fp16 = const()[name = string("op_9625_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9626_cast_fp16 = mul(x = var_9624_cast_fp16, y = var_9625_to_fp16)[name = string("op_9626_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9627_cast_fp16 = add(x = var_9622_cast_fp16, y = var_9626_cast_fp16)[name = string("op_9627_cast_fp16")];
            fp32 var_9628_epsilon_0 = const()[name = string("op_9628_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9628_cast_fp16 = rsqrt(epsilon = var_9628_epsilon_0, x = var_9627_cast_fp16)[name = string("op_9628_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9629_cast_fp16 = mul(x = x_503_cast_fp16, y = var_9628_cast_fp16)[name = string("op_9629_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_9629_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_9638 = const()[name = string("op_9638"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9639_cast_fp16 = reshape(shape = var_9638, x = linear_175_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [4]> x_505_perm_0 = const()[name = string("x_505_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9646 = const()[name = string("op_9646"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_505_cast_fp16 = transpose(perm = x_505_perm_0, x = var_9639_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 16, 128]> var_9649_cast_fp16 = abs(x = x_505_cast_fp16)[name = string("op_9649_cast_fp16")];
            tensor<int32, [1]> peak_203_axes_0 = const()[name = string("peak_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_203_cast_fp16 = reduce_max(axes = peak_203_axes_0, keep_dims = var_9646, x = var_9649_cast_fp16)[name = string("peak_203_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_507_cast_fp16 = real_div(x = x_505_cast_fp16, y = peak_203_cast_fp16)[name = string("x_507_cast_fp16")];
            fp16 var_9645_promoted_to_fp16 = const()[name = string("op_9645_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9653_cast_fp16 = pow(x = x_507_cast_fp16, y = var_9645_promoted_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<int32, [1]> var_9655_axes_0 = const()[name = string("op_9655_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9655_keep_dims_0 = const()[name = string("op_9655_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9655_cast_fp16 = reduce_mean(axes = var_9655_axes_0, keep_dims = var_9655_keep_dims_0, x = var_9653_cast_fp16)[name = string("op_9655_cast_fp16")];
            fp16 var_9645_promoted_1_to_fp16 = const()[name = string("op_9645_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9656_cast_fp16 = pow(x = peak_203_cast_fp16, y = var_9645_promoted_1_to_fp16)[name = string("op_9656_cast_fp16")];
            fp32 var_9657_epsilon_0 = const()[name = string("op_9657_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9657_cast_fp16 = inverse(epsilon = var_9657_epsilon_0, x = var_9656_cast_fp16)[name = string("op_9657_cast_fp16")];
            fp16 var_9658_to_fp16 = const()[name = string("op_9658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9659_cast_fp16 = mul(x = var_9657_cast_fp16, y = var_9658_to_fp16)[name = string("op_9659_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9660_cast_fp16 = add(x = var_9655_cast_fp16, y = var_9659_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp32 var_9661_epsilon_0 = const()[name = string("op_9661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9661_cast_fp16 = rsqrt(epsilon = var_9661_epsilon_0, x = var_9660_cast_fp16)[name = string("op_9661_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9662_cast_fp16 = mul(x = x_507_cast_fp16, y = var_9661_cast_fp16)[name = string("op_9662_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_513_cast_fp16 = mul(x = var_9662_cast_fp16, y = model_layers_25_self_attn_q_norm_weight_to_fp16)[name = string("x_513_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_9671 = const()[name = string("op_9671"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9672_cast_fp16 = reshape(shape = var_9671, x = linear_176_cast_fp16)[name = string("op_9672_cast_fp16")];
            tensor<int32, [4]> x_509_perm_0 = const()[name = string("x_509_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9679 = const()[name = string("op_9679"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_509_cast_fp16 = transpose(perm = x_509_perm_0, x = var_9672_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 16, 128]> var_9682_cast_fp16 = abs(x = x_509_cast_fp16)[name = string("op_9682_cast_fp16")];
            tensor<int32, [1]> peak_205_axes_0 = const()[name = string("peak_205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_205_cast_fp16 = reduce_max(axes = peak_205_axes_0, keep_dims = var_9679, x = var_9682_cast_fp16)[name = string("peak_205_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_511_cast_fp16 = real_div(x = x_509_cast_fp16, y = peak_205_cast_fp16)[name = string("x_511_cast_fp16")];
            fp16 var_9678_promoted_to_fp16 = const()[name = string("op_9678_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9686_cast_fp16 = pow(x = x_511_cast_fp16, y = var_9678_promoted_to_fp16)[name = string("op_9686_cast_fp16")];
            tensor<int32, [1]> var_9688_axes_0 = const()[name = string("op_9688_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9688_keep_dims_0 = const()[name = string("op_9688_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9688_cast_fp16 = reduce_mean(axes = var_9688_axes_0, keep_dims = var_9688_keep_dims_0, x = var_9686_cast_fp16)[name = string("op_9688_cast_fp16")];
            fp16 var_9678_promoted_1_to_fp16 = const()[name = string("op_9678_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9689_cast_fp16 = pow(x = peak_205_cast_fp16, y = var_9678_promoted_1_to_fp16)[name = string("op_9689_cast_fp16")];
            fp32 var_9690_epsilon_0 = const()[name = string("op_9690_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9690_cast_fp16 = inverse(epsilon = var_9690_epsilon_0, x = var_9689_cast_fp16)[name = string("op_9690_cast_fp16")];
            fp16 var_9691_to_fp16 = const()[name = string("op_9691_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9692_cast_fp16 = mul(x = var_9690_cast_fp16, y = var_9691_to_fp16)[name = string("op_9692_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9693_cast_fp16 = add(x = var_9688_cast_fp16, y = var_9692_cast_fp16)[name = string("op_9693_cast_fp16")];
            fp32 var_9694_epsilon_0 = const()[name = string("op_9694_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9694_cast_fp16 = rsqrt(epsilon = var_9694_epsilon_0, x = var_9693_cast_fp16)[name = string("op_9694_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9695_cast_fp16 = mul(x = x_511_cast_fp16, y = var_9694_cast_fp16)[name = string("op_9695_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_515_cast_fp16 = mul(x = var_9695_cast_fp16, y = model_layers_25_self_attn_k_norm_weight_to_fp16)[name = string("x_515_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_9704 = const()[name = string("op_9704"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9705_cast_fp16 = reshape(shape = var_9704, x = linear_177_cast_fp16)[name = string("op_9705_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_513_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_513_cast_fp16)[name = string("second_101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9719_cast_fp16 = mul(x = first_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9719_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9720_cast_fp16 = mul(x = second_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9720_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9722_cast_fp16 = sub(x = var_9719_cast_fp16, y = var_9720_cast_fp16)[name = string("op_9722_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9723_cast_fp16 = mul(x = second_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9723_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9724_cast_fp16 = mul(x = first_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9724_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9726_cast_fp16 = add(x = var_9723_cast_fp16, y = var_9724_cast_fp16)[name = string("op_9726_cast_fp16")];
            int32 var_9728 = const()[name = string("op_9728"), 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_9728, interleave = q_51_interleave_0, values = (var_9722_cast_fp16, var_9726_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_515_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_515_cast_fp16)[name = string("second_103_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9740_cast_fp16 = mul(x = first_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9740_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9741_cast_fp16 = mul(x = second_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9741_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9743_cast_fp16 = sub(x = var_9740_cast_fp16, y = var_9741_cast_fp16)[name = string("op_9743_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9744_cast_fp16 = mul(x = second_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9744_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9745_cast_fp16 = mul(x = first_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9745_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9747_cast_fp16 = add(x = var_9744_cast_fp16, y = var_9745_cast_fp16)[name = string("op_9747_cast_fp16")];
            int32 var_9749 = const()[name = string("op_9749"), 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_9749, interleave = k_51_interleave_0, values = (var_9743_cast_fp16, var_9747_cast_fp16))[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> var_9760_begin_0 = const()[name = string("op_9760_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9760_end_0 = const()[name = string("op_9760_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9760_end_mask_0 = const()[name = string("op_9760_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9760_stride_0 = const()[name = string("op_9760_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9760_cast_fp16 = slice_by_index(begin = var_9760_begin_0, end = var_9760_end_0, end_mask = var_9760_end_mask_0, stride = var_9760_stride_0, x = q_51_cast_fp16)[name = string("op_9760_cast_fp16")];
            tensor<int32, [4]> var_9770_begin_0 = const()[name = string("op_9770_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9770_end_0 = const()[name = string("op_9770_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9770_end_mask_0 = const()[name = string("op_9770_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9770_stride_0 = const()[name = string("op_9770_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9770_cast_fp16 = slice_by_index(begin = var_9770_begin_0, end = var_9770_end_0, end_mask = var_9770_end_mask_0, stride = var_9770_stride_0, x = q_51_cast_fp16)[name = string("op_9770_cast_fp16")];
            int32 var_9772 = const()[name = string("op_9772"), 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_9772, interleave = stacked_51_interleave_0, values = (var_9760_cast_fp16, var_9770_cast_fp16))[name = string("stacked_51_cast_fp16")];
            tensor<int32, [4]> var_9778_begin_0 = const()[name = string("op_9778_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9778_end_0 = const()[name = string("op_9778_end_0"), val = tensor<int32, [4]>([26, 8, 2064, 128])];
            tensor<bool, [4]> var_9778_end_mask_0 = const()[name = string("op_9778_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9778_cast_fp16 = slice_by_index(begin = var_9778_begin_0, end = var_9778_end_0, end_mask = var_9778_end_mask_0, x = read_state_0)[name = string("op_9778_cast_fp16")];
            tensor<int32, [4]> var_9788_begin_0 = const()[name = string("op_9788_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9788_end_0 = const()[name = string("op_9788_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9788_end_mask_0 = const()[name = string("op_9788_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9788_cast_fp16 = slice_by_index(begin = var_9788_begin_0, end = var_9788_end_0, end_mask = var_9788_end_mask_0, x = var_9778_cast_fp16)[name = string("op_9788_cast_fp16")];
            bool var_9797_transpose_x_1 = const()[name = string("op_9797_transpose_x_1"), val = bool(false)];
            bool var_9797_transpose_y_1 = const()[name = string("op_9797_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_1, transpose_y = var_9797_transpose_y_1, x = var_9788_cast_fp16, y = stacked_51_cast_fp16)[name = string("op_9797_cast_fp16")];
            tensor<int32, [4]> var_9800_perm_0 = const()[name = string("op_9800_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9801_to_fp16 = const()[name = string("op_9801_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_9800_cast_fp16 = transpose(perm = var_9800_perm_0, x = var_9797_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 32, 2048]> var_9802_cast_fp16 = mul(x = var_9800_cast_fp16, y = var_9801_to_fp16)[name = string("op_9802_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_51_cast_fp16 = add(x = var_9802_cast_fp16, y = cacheMask)[name = string("cache_weights_51_cast_fp16")];
            bool var_9808_transpose_x_1 = const()[name = string("op_9808_transpose_x_1"), val = bool(false)];
            bool var_9808_transpose_y_1 = const()[name = string("op_9808_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9808_cast_fp16 = matmul(transpose_x = var_9808_transpose_x_1, transpose_y = var_9808_transpose_y_1, x = stacked_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_9808_cast_fp16")];
            fp16 var_9809_to_fp16 = const()[name = string("op_9809_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9810_cast_fp16 = mul(x = var_9808_cast_fp16, y = var_9809_to_fp16)[name = string("op_9810_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_51_cast_fp16 = add(x = var_9810_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_51_cast_fp16")];
            int32 var_9814 = const()[name = string("op_9814"), val = int32(-1)];
            bool var_9815_interleave_0 = const()[name = string("op_9815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_9815_cast_fp16 = concat(axis = var_9814, interleave = var_9815_interleave_0, values = (cache_weights_51_cast_fp16, call_weights_51_cast_fp16))[name = string("op_9815_cast_fp16")];
            int32 var_9816 = const()[name = string("op_9816"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_51_cast_fp16 = softmax(axis = var_9816, x = var_9815_cast_fp16)[name = string("shares_51_cast_fp16")];
            tensor<int32, [4]> var_9823_begin_0 = const()[name = string("op_9823_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9823_end_0 = const()[name = string("op_9823_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_9823_end_mask_0 = const()[name = string("op_9823_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_9823_cast_fp16 = slice_by_index(begin = var_9823_begin_0, end = var_9823_end_0, end_mask = var_9823_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9823_cast_fp16")];
            tensor<int32, [4]> var_9828_begin_0 = const()[name = string("op_9828_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9828_end_0 = const()[name = string("op_9828_end_0"), val = tensor<int32, [4]>([26, 8, 2064, 128])];
            tensor<bool, [4]> var_9828_end_mask_0 = const()[name = string("op_9828_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9828_cast_fp16 = slice_by_index(begin = var_9828_begin_0, end = var_9828_end_0, end_mask = var_9828_end_mask_0, x = read_state_1)[name = string("op_9828_cast_fp16")];
            tensor<int32, [4]> var_9838_begin_0 = const()[name = string("op_9838_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9838_end_0 = const()[name = string("op_9838_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9838_end_mask_0 = const()[name = string("op_9838_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9838_cast_fp16 = slice_by_index(begin = var_9838_begin_0, end = var_9838_end_0, end_mask = var_9838_end_mask_0, x = var_9828_cast_fp16)[name = string("op_9838_cast_fp16")];
            bool var_9844_transpose_x_0 = const()[name = string("op_9844_transpose_x_0"), val = bool(false)];
            bool var_9844_transpose_y_0 = const()[name = string("op_9844_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9844_cast_fp16 = matmul(transpose_x = var_9844_transpose_x_0, transpose_y = var_9844_transpose_y_0, x = var_9823_cast_fp16, y = var_9838_cast_fp16)[name = string("op_9844_cast_fp16")];
            tensor<int32, [4]> var_9849_begin_0 = const()[name = string("op_9849_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_9849_end_0 = const()[name = string("op_9849_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_9849_end_mask_0 = const()[name = string("op_9849_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9849_cast_fp16 = slice_by_index(begin = var_9849_begin_0, end = var_9849_end_0, end_mask = var_9849_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9849_cast_fp16")];
            bool var_9850_transpose_x_0 = const()[name = string("op_9850_transpose_x_0"), val = bool(false)];
            bool var_9850_transpose_y_0 = const()[name = string("op_9850_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_9705_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 8, 32, 128]> var_9850_cast_fp16 = matmul(transpose_x = var_9850_transpose_x_0, transpose_y = var_9850_transpose_y_0, x = var_9849_cast_fp16, y = v_51_cast_fp16)[name = string("op_9850_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_101_cast_fp16 = add(x = var_9844_cast_fp16, y = var_9850_cast_fp16)[name = string("mixed_101_cast_fp16")];
            tensor<int32, [4]> var_9867_begin_0 = const()[name = string("op_9867_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9867_end_0 = const()[name = string("op_9867_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9867_end_mask_0 = const()[name = string("op_9867_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9867_cast_fp16 = slice_by_index(begin = var_9867_begin_0, end = var_9867_end_0, end_mask = var_9867_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9867_cast_fp16")];
            tensor<int32, [4]> var_9882_begin_0 = const()[name = string("op_9882_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9882_end_0 = const()[name = string("op_9882_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9882_end_mask_0 = const()[name = string("op_9882_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9882_cast_fp16 = slice_by_index(begin = var_9882_begin_0, end = var_9882_end_0, end_mask = var_9882_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9882_cast_fp16")];
            int32 var_9884 = const()[name = string("op_9884"), 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_9884, interleave = mixed_103_interleave_0, values = (var_9867_cast_fp16, var_9882_cast_fp16))[name = string("mixed_103_cast_fp16")];
            tensor<int32, [4]> var_9888_perm_0 = const()[name = string("op_9888_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9892 = const()[name = string("op_9892"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9888_cast_fp16 = transpose(perm = var_9888_perm_0, x = mixed_103_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 2048]> input_253_cast_fp16 = reshape(shape = var_9892, x = var_9888_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_517_cast_fp16 = add(x = x_501_cast_fp16, y = linear_178_cast_fp16)[name = string("x_517_cast_fp16")];
            bool var_9902 = const()[name = string("op_9902"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9905_cast_fp16 = abs(x = x_517_cast_fp16)[name = string("op_9905_cast_fp16")];
            tensor<int32, [1]> peak_207_axes_0 = const()[name = string("peak_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_207_cast_fp16 = reduce_max(axes = peak_207_axes_0, keep_dims = var_9902, x = var_9905_cast_fp16)[name = string("peak_207_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_519_cast_fp16 = real_div(x = x_517_cast_fp16, y = peak_207_cast_fp16)[name = string("x_519_cast_fp16")];
            fp16 var_9901_promoted_to_fp16 = const()[name = string("op_9901_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9909_cast_fp16 = pow(x = x_519_cast_fp16, y = var_9901_promoted_to_fp16)[name = string("op_9909_cast_fp16")];
            tensor<int32, [1]> var_9911_axes_0 = const()[name = string("op_9911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9911_keep_dims_0 = const()[name = string("op_9911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9911_cast_fp16 = reduce_mean(axes = var_9911_axes_0, keep_dims = var_9911_keep_dims_0, x = var_9909_cast_fp16)[name = string("op_9911_cast_fp16")];
            fp16 var_9901_promoted_1_to_fp16 = const()[name = string("op_9901_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9912_cast_fp16 = pow(x = peak_207_cast_fp16, y = var_9901_promoted_1_to_fp16)[name = string("op_9912_cast_fp16")];
            fp32 var_9913_epsilon_0 = const()[name = string("op_9913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9913_cast_fp16 = inverse(epsilon = var_9913_epsilon_0, x = var_9912_cast_fp16)[name = string("op_9913_cast_fp16")];
            fp16 var_9914_to_fp16 = const()[name = string("op_9914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9915_cast_fp16 = mul(x = var_9913_cast_fp16, y = var_9914_to_fp16)[name = string("op_9915_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9916_cast_fp16 = add(x = var_9911_cast_fp16, y = var_9915_cast_fp16)[name = string("op_9916_cast_fp16")];
            fp32 var_9917_epsilon_0 = const()[name = string("op_9917_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9917_cast_fp16 = rsqrt(epsilon = var_9917_epsilon_0, x = var_9916_cast_fp16)[name = string("op_9917_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9918_cast_fp16 = mul(x = x_519_cast_fp16, y = var_9917_cast_fp16)[name = string("op_9918_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_9918_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_9923_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_9923_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_9923_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_521_cast_fp16 = add(x = x_517_cast_fp16, y = linear_181_cast_fp16)[name = string("x_521_cast_fp16")];
            bool var_9936 = const()[name = string("op_9936"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9939_cast_fp16 = abs(x = x_521_cast_fp16)[name = string("op_9939_cast_fp16")];
            tensor<int32, [1]> peak_209_axes_0 = const()[name = string("peak_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_209_cast_fp16 = reduce_max(axes = peak_209_axes_0, keep_dims = var_9936, x = var_9939_cast_fp16)[name = string("peak_209_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_523_cast_fp16 = real_div(x = x_521_cast_fp16, y = peak_209_cast_fp16)[name = string("x_523_cast_fp16")];
            fp16 var_9935_promoted_to_fp16 = const()[name = string("op_9935_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9943_cast_fp16 = pow(x = x_523_cast_fp16, y = var_9935_promoted_to_fp16)[name = string("op_9943_cast_fp16")];
            tensor<int32, [1]> var_9945_axes_0 = const()[name = string("op_9945_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9945_keep_dims_0 = const()[name = string("op_9945_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9945_cast_fp16 = reduce_mean(axes = var_9945_axes_0, keep_dims = var_9945_keep_dims_0, x = var_9943_cast_fp16)[name = string("op_9945_cast_fp16")];
            fp16 var_9935_promoted_1_to_fp16 = const()[name = string("op_9935_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9946_cast_fp16 = pow(x = peak_209_cast_fp16, y = var_9935_promoted_1_to_fp16)[name = string("op_9946_cast_fp16")];
            fp32 var_9947_epsilon_0 = const()[name = string("op_9947_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9947_cast_fp16 = inverse(epsilon = var_9947_epsilon_0, x = var_9946_cast_fp16)[name = string("op_9947_cast_fp16")];
            fp16 var_9948_to_fp16 = const()[name = string("op_9948_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9949_cast_fp16 = mul(x = var_9947_cast_fp16, y = var_9948_to_fp16)[name = string("op_9949_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9950_cast_fp16 = add(x = var_9945_cast_fp16, y = var_9949_cast_fp16)[name = string("op_9950_cast_fp16")];
            fp32 var_9951_epsilon_0 = const()[name = string("op_9951_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9951_cast_fp16 = rsqrt(epsilon = var_9951_epsilon_0, x = var_9950_cast_fp16)[name = string("op_9951_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9952_cast_fp16 = mul(x = x_523_cast_fp16, y = var_9951_cast_fp16)[name = string("op_9952_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_9952_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_9961 = const()[name = string("op_9961"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9962_cast_fp16 = reshape(shape = var_9961, x = linear_182_cast_fp16)[name = string("op_9962_cast_fp16")];
            tensor<int32, [4]> x_525_perm_0 = const()[name = string("x_525_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9969 = const()[name = string("op_9969"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_525_cast_fp16 = transpose(perm = x_525_perm_0, x = var_9962_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 16, 128]> var_9972_cast_fp16 = abs(x = x_525_cast_fp16)[name = string("op_9972_cast_fp16")];
            tensor<int32, [1]> peak_211_axes_0 = const()[name = string("peak_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_211_cast_fp16 = reduce_max(axes = peak_211_axes_0, keep_dims = var_9969, x = var_9972_cast_fp16)[name = string("peak_211_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_527_cast_fp16 = real_div(x = x_525_cast_fp16, y = peak_211_cast_fp16)[name = string("x_527_cast_fp16")];
            fp16 var_9968_promoted_to_fp16 = const()[name = string("op_9968_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9976_cast_fp16 = pow(x = x_527_cast_fp16, y = var_9968_promoted_to_fp16)[name = string("op_9976_cast_fp16")];
            tensor<int32, [1]> var_9978_axes_0 = const()[name = string("op_9978_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9978_keep_dims_0 = const()[name = string("op_9978_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9978_cast_fp16 = reduce_mean(axes = var_9978_axes_0, keep_dims = var_9978_keep_dims_0, x = var_9976_cast_fp16)[name = string("op_9978_cast_fp16")];
            fp16 var_9968_promoted_1_to_fp16 = const()[name = string("op_9968_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9979_cast_fp16 = pow(x = peak_211_cast_fp16, y = var_9968_promoted_1_to_fp16)[name = string("op_9979_cast_fp16")];
            fp32 var_9980_epsilon_0 = const()[name = string("op_9980_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9980_cast_fp16 = inverse(epsilon = var_9980_epsilon_0, x = var_9979_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp16 var_9981_to_fp16 = const()[name = string("op_9981_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9982_cast_fp16 = mul(x = var_9980_cast_fp16, y = var_9981_to_fp16)[name = string("op_9982_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9983_cast_fp16 = add(x = var_9978_cast_fp16, y = var_9982_cast_fp16)[name = string("op_9983_cast_fp16")];
            fp32 var_9984_epsilon_0 = const()[name = string("op_9984_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9984_cast_fp16 = rsqrt(epsilon = var_9984_epsilon_0, x = var_9983_cast_fp16)[name = string("op_9984_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9985_cast_fp16 = mul(x = x_527_cast_fp16, y = var_9984_cast_fp16)[name = string("op_9985_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_533_cast_fp16 = mul(x = var_9985_cast_fp16, y = model_layers_26_self_attn_q_norm_weight_to_fp16)[name = string("x_533_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_9994 = const()[name = string("op_9994"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9995_cast_fp16 = reshape(shape = var_9994, x = linear_183_cast_fp16)[name = string("op_9995_cast_fp16")];
            tensor<int32, [4]> x_529_perm_0 = const()[name = string("x_529_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10002 = const()[name = string("op_10002"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_529_cast_fp16 = transpose(perm = x_529_perm_0, x = var_9995_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 16, 128]> var_10005_cast_fp16 = abs(x = x_529_cast_fp16)[name = string("op_10005_cast_fp16")];
            tensor<int32, [1]> peak_213_axes_0 = const()[name = string("peak_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_213_cast_fp16 = reduce_max(axes = peak_213_axes_0, keep_dims = var_10002, x = var_10005_cast_fp16)[name = string("peak_213_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_531_cast_fp16 = real_div(x = x_529_cast_fp16, y = peak_213_cast_fp16)[name = string("x_531_cast_fp16")];
            fp16 var_10001_promoted_to_fp16 = const()[name = string("op_10001_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_10009_cast_fp16 = pow(x = x_531_cast_fp16, y = var_10001_promoted_to_fp16)[name = string("op_10009_cast_fp16")];
            tensor<int32, [1]> var_10011_axes_0 = const()[name = string("op_10011_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10011_keep_dims_0 = const()[name = string("op_10011_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_10011_cast_fp16 = reduce_mean(axes = var_10011_axes_0, keep_dims = var_10011_keep_dims_0, x = var_10009_cast_fp16)[name = string("op_10011_cast_fp16")];
            fp16 var_10001_promoted_1_to_fp16 = const()[name = string("op_10001_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_10012_cast_fp16 = pow(x = peak_213_cast_fp16, y = var_10001_promoted_1_to_fp16)[name = string("op_10012_cast_fp16")];
            fp32 var_10013_epsilon_0 = const()[name = string("op_10013_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_10013_cast_fp16 = inverse(epsilon = var_10013_epsilon_0, x = var_10012_cast_fp16)[name = string("op_10013_cast_fp16")];
            fp16 var_10014_to_fp16 = const()[name = string("op_10014_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_10015_cast_fp16 = mul(x = var_10013_cast_fp16, y = var_10014_to_fp16)[name = string("op_10015_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_10016_cast_fp16 = add(x = var_10011_cast_fp16, y = var_10015_cast_fp16)[name = string("op_10016_cast_fp16")];
            fp32 var_10017_epsilon_0 = const()[name = string("op_10017_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_10017_cast_fp16 = rsqrt(epsilon = var_10017_epsilon_0, x = var_10016_cast_fp16)[name = string("op_10017_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_10018_cast_fp16 = mul(x = x_531_cast_fp16, y = var_10017_cast_fp16)[name = string("op_10018_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_535_cast_fp16 = mul(x = var_10018_cast_fp16, y = model_layers_26_self_attn_k_norm_weight_to_fp16)[name = string("x_535_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_10027 = const()[name = string("op_10027"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10028_cast_fp16 = reshape(shape = var_10027, x = linear_184_cast_fp16)[name = string("op_10028_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_533_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_533_cast_fp16)[name = string("second_105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10042_cast_fp16 = mul(x = first_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10042_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10043_cast_fp16 = mul(x = second_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10045_cast_fp16 = sub(x = var_10042_cast_fp16, y = var_10043_cast_fp16)[name = string("op_10045_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10046_cast_fp16 = mul(x = second_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10046_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10047_cast_fp16 = mul(x = first_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10047_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10049_cast_fp16 = add(x = var_10046_cast_fp16, y = var_10047_cast_fp16)[name = string("op_10049_cast_fp16")];
            int32 var_10051 = const()[name = string("op_10051"), 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_10051, interleave = q_53_interleave_0, values = (var_10045_cast_fp16, var_10049_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_535_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_535_cast_fp16)[name = string("second_107_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10063_cast_fp16 = mul(x = first_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10064_cast_fp16 = mul(x = second_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10064_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10066_cast_fp16 = sub(x = var_10063_cast_fp16, y = var_10064_cast_fp16)[name = string("op_10066_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10067_cast_fp16 = mul(x = second_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10067_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10068_cast_fp16 = mul(x = first_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10068_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10070_cast_fp16 = add(x = var_10067_cast_fp16, y = var_10068_cast_fp16)[name = string("op_10070_cast_fp16")];
            int32 var_10072 = const()[name = string("op_10072"), 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_10072, interleave = k_53_interleave_0, values = (var_10066_cast_fp16, var_10070_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> var_10083_begin_0 = const()[name = string("op_10083_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10083_end_0 = const()[name = string("op_10083_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10083_end_mask_0 = const()[name = string("op_10083_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10083_stride_0 = const()[name = string("op_10083_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10083_cast_fp16 = slice_by_index(begin = var_10083_begin_0, end = var_10083_end_0, end_mask = var_10083_end_mask_0, stride = var_10083_stride_0, x = q_53_cast_fp16)[name = string("op_10083_cast_fp16")];
            tensor<int32, [4]> var_10093_begin_0 = const()[name = string("op_10093_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_10093_end_0 = const()[name = string("op_10093_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10093_end_mask_0 = const()[name = string("op_10093_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10093_stride_0 = const()[name = string("op_10093_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10093_cast_fp16 = slice_by_index(begin = var_10093_begin_0, end = var_10093_end_0, end_mask = var_10093_end_mask_0, stride = var_10093_stride_0, x = q_53_cast_fp16)[name = string("op_10093_cast_fp16")];
            int32 var_10095 = const()[name = string("op_10095"), 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_10095, interleave = stacked_53_interleave_0, values = (var_10083_cast_fp16, var_10093_cast_fp16))[name = string("stacked_53_cast_fp16")];
            tensor<int32, [4]> var_10101_begin_0 = const()[name = string("op_10101_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_10101_end_0 = const()[name = string("op_10101_end_0"), val = tensor<int32, [4]>([27, 8, 2064, 128])];
            tensor<bool, [4]> var_10101_end_mask_0 = const()[name = string("op_10101_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_10101_cast_fp16 = slice_by_index(begin = var_10101_begin_0, end = var_10101_end_0, end_mask = var_10101_end_mask_0, x = read_state_0)[name = string("op_10101_cast_fp16")];
            tensor<int32, [4]> var_10111_begin_0 = const()[name = string("op_10111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10111_end_0 = const()[name = string("op_10111_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_10111_end_mask_0 = const()[name = string("op_10111_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, x = var_10101_cast_fp16)[name = string("op_10111_cast_fp16")];
            bool var_10120_transpose_x_1 = const()[name = string("op_10120_transpose_x_1"), val = bool(false)];
            bool var_10120_transpose_y_1 = const()[name = string("op_10120_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_10120_cast_fp16 = matmul(transpose_x = var_10120_transpose_x_1, transpose_y = var_10120_transpose_y_1, x = var_10111_cast_fp16, y = stacked_53_cast_fp16)[name = string("op_10120_cast_fp16")];
            tensor<int32, [4]> var_10123_perm_0 = const()[name = string("op_10123_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_10124_to_fp16 = const()[name = string("op_10124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_10123_cast_fp16 = transpose(perm = var_10123_perm_0, x = var_10120_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 8, 32, 2048]> var_10125_cast_fp16 = mul(x = var_10123_cast_fp16, y = var_10124_to_fp16)[name = string("op_10125_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_53_cast_fp16 = add(x = var_10125_cast_fp16, y = cacheMask)[name = string("cache_weights_53_cast_fp16")];
            bool var_10131_transpose_x_1 = const()[name = string("op_10131_transpose_x_1"), val = bool(false)];
            bool var_10131_transpose_y_1 = const()[name = string("op_10131_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_10131_cast_fp16 = matmul(transpose_x = var_10131_transpose_x_1, transpose_y = var_10131_transpose_y_1, x = stacked_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_10131_cast_fp16")];
            fp16 var_10132_to_fp16 = const()[name = string("op_10132_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_10133_cast_fp16 = mul(x = var_10131_cast_fp16, y = var_10132_to_fp16)[name = string("op_10133_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_53_cast_fp16 = add(x = var_10133_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_53_cast_fp16")];
            int32 var_10137 = const()[name = string("op_10137"), val = int32(-1)];
            bool var_10138_interleave_0 = const()[name = string("op_10138_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_10138_cast_fp16 = concat(axis = var_10137, interleave = var_10138_interleave_0, values = (cache_weights_53_cast_fp16, call_weights_53_cast_fp16))[name = string("op_10138_cast_fp16")];
            int32 var_10139 = const()[name = string("op_10139"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_53_cast_fp16 = softmax(axis = var_10139, x = var_10138_cast_fp16)[name = string("shares_53_cast_fp16")];
            tensor<int32, [4]> var_10146_begin_0 = const()[name = string("op_10146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10146_end_0 = const()[name = string("op_10146_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_10146_end_mask_0 = const()[name = string("op_10146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_10146_cast_fp16 = slice_by_index(begin = var_10146_begin_0, end = var_10146_end_0, end_mask = var_10146_end_mask_0, x = shares_53_cast_fp16)[name = string("op_10146_cast_fp16")];
            tensor<int32, [4]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [4]>([27, 8, 2064, 128])];
            tensor<bool, [4]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, x = read_state_1)[name = string("op_10151_cast_fp16")];
            tensor<int32, [4]> var_10161_begin_0 = const()[name = string("op_10161_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10161_end_0 = const()[name = string("op_10161_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_10161_end_mask_0 = const()[name = string("op_10161_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_10161_cast_fp16 = slice_by_index(begin = var_10161_begin_0, end = var_10161_end_0, end_mask = var_10161_end_mask_0, x = var_10151_cast_fp16)[name = string("op_10161_cast_fp16")];
            bool var_10167_transpose_x_0 = const()[name = string("op_10167_transpose_x_0"), val = bool(false)];
            bool var_10167_transpose_y_0 = const()[name = string("op_10167_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_10167_cast_fp16 = matmul(transpose_x = var_10167_transpose_x_0, transpose_y = var_10167_transpose_y_0, x = var_10146_cast_fp16, y = var_10161_cast_fp16)[name = string("op_10167_cast_fp16")];
            tensor<int32, [4]> var_10172_begin_0 = const()[name = string("op_10172_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_10172_end_0 = const()[name = string("op_10172_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_10172_end_mask_0 = const()[name = string("op_10172_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_10172_cast_fp16 = slice_by_index(begin = var_10172_begin_0, end = var_10172_end_0, end_mask = var_10172_end_mask_0, x = shares_53_cast_fp16)[name = string("op_10172_cast_fp16")];
            bool var_10173_transpose_x_0 = const()[name = string("op_10173_transpose_x_0"), val = bool(false)];
            bool var_10173_transpose_y_0 = const()[name = string("op_10173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_10028_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 32, 128]> var_10173_cast_fp16 = matmul(transpose_x = var_10173_transpose_x_0, transpose_y = var_10173_transpose_y_0, x = var_10172_cast_fp16, y = v_53_cast_fp16)[name = string("op_10173_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_105_cast_fp16 = add(x = var_10167_cast_fp16, y = var_10173_cast_fp16)[name = string("mixed_105_cast_fp16")];
            tensor<int32, [4]> var_10190_begin_0 = const()[name = string("op_10190_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10190_end_0 = const()[name = string("op_10190_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_10190_end_mask_0 = const()[name = string("op_10190_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10190_cast_fp16 = slice_by_index(begin = var_10190_begin_0, end = var_10190_end_0, end_mask = var_10190_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_10190_cast_fp16")];
            tensor<int32, [4]> var_10205_begin_0 = const()[name = string("op_10205_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_10205_end_0 = const()[name = string("op_10205_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_10205_end_mask_0 = const()[name = string("op_10205_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10205_cast_fp16 = slice_by_index(begin = var_10205_begin_0, end = var_10205_end_0, end_mask = var_10205_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_10205_cast_fp16")];
            int32 var_10207 = const()[name = string("op_10207"), 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_10207, interleave = mixed_107_interleave_0, values = (var_10190_cast_fp16, var_10205_cast_fp16))[name = string("mixed_107_cast_fp16")];
            tensor<int32, [4]> var_10211_perm_0 = const()[name = string("op_10211_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_10215 = const()[name = string("op_10215"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_10211_cast_fp16 = transpose(perm = var_10211_perm_0, x = mixed_107_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 2048]> input_263_cast_fp16 = reshape(shape = var_10215, x = var_10211_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_537_cast_fp16 = add(x = x_521_cast_fp16, y = linear_185_cast_fp16)[name = string("x_537_cast_fp16")];
            bool var_10225 = const()[name = string("op_10225"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10228_cast_fp16 = abs(x = x_537_cast_fp16)[name = string("op_10228_cast_fp16")];
            tensor<int32, [1]> peak_215_axes_0 = const()[name = string("peak_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_215_cast_fp16 = reduce_max(axes = peak_215_axes_0, keep_dims = var_10225, x = var_10228_cast_fp16)[name = string("peak_215_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_539_cast_fp16 = real_div(x = x_537_cast_fp16, y = peak_215_cast_fp16)[name = string("x_539_cast_fp16")];
            fp16 var_10224_promoted_to_fp16 = const()[name = string("op_10224_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10232_cast_fp16 = pow(x = x_539_cast_fp16, y = var_10224_promoted_to_fp16)[name = string("op_10232_cast_fp16")];
            tensor<int32, [1]> var_10234_axes_0 = const()[name = string("op_10234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10234_keep_dims_0 = const()[name = string("op_10234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10234_cast_fp16 = reduce_mean(axes = var_10234_axes_0, keep_dims = var_10234_keep_dims_0, x = var_10232_cast_fp16)[name = string("op_10234_cast_fp16")];
            fp16 var_10224_promoted_1_to_fp16 = const()[name = string("op_10224_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10235_cast_fp16 = pow(x = peak_215_cast_fp16, y = var_10224_promoted_1_to_fp16)[name = string("op_10235_cast_fp16")];
            fp32 var_10236_epsilon_0 = const()[name = string("op_10236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10236_cast_fp16 = inverse(epsilon = var_10236_epsilon_0, x = var_10235_cast_fp16)[name = string("op_10236_cast_fp16")];
            fp16 var_10237_to_fp16 = const()[name = string("op_10237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10238_cast_fp16 = mul(x = var_10236_cast_fp16, y = var_10237_to_fp16)[name = string("op_10238_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10239_cast_fp16 = add(x = var_10234_cast_fp16, y = var_10238_cast_fp16)[name = string("op_10239_cast_fp16")];
            fp32 var_10240_epsilon_0 = const()[name = string("op_10240_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10240_cast_fp16 = rsqrt(epsilon = var_10240_epsilon_0, x = var_10239_cast_fp16)[name = string("op_10240_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10241_cast_fp16 = mul(x = x_539_cast_fp16, y = var_10240_cast_fp16)[name = string("op_10241_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_10241_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_10246_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_10246_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_10246_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_541_cast_fp16 = add(x = x_537_cast_fp16, y = linear_188_cast_fp16)[name = string("x_541_cast_fp16")];
            bool var_10259 = const()[name = string("op_10259"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10262_cast_fp16 = abs(x = x_541_cast_fp16)[name = string("op_10262_cast_fp16")];
            tensor<int32, [1]> peak_217_axes_0 = const()[name = string("peak_217_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_217_cast_fp16 = reduce_max(axes = peak_217_axes_0, keep_dims = var_10259, x = var_10262_cast_fp16)[name = string("peak_217_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_543_cast_fp16 = real_div(x = x_541_cast_fp16, y = peak_217_cast_fp16)[name = string("x_543_cast_fp16")];
            fp16 var_10258_promoted_to_fp16 = const()[name = string("op_10258_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10266_cast_fp16 = pow(x = x_543_cast_fp16, y = var_10258_promoted_to_fp16)[name = string("op_10266_cast_fp16")];
            tensor<int32, [1]> var_10268_axes_0 = const()[name = string("op_10268_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10268_keep_dims_0 = const()[name = string("op_10268_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10268_cast_fp16 = reduce_mean(axes = var_10268_axes_0, keep_dims = var_10268_keep_dims_0, x = var_10266_cast_fp16)[name = string("op_10268_cast_fp16")];
            fp16 var_10258_promoted_1_to_fp16 = const()[name = string("op_10258_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10269_cast_fp16 = pow(x = peak_217_cast_fp16, y = var_10258_promoted_1_to_fp16)[name = string("op_10269_cast_fp16")];
            fp32 var_10270_epsilon_0 = const()[name = string("op_10270_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10270_cast_fp16 = inverse(epsilon = var_10270_epsilon_0, x = var_10269_cast_fp16)[name = string("op_10270_cast_fp16")];
            fp16 var_10271_to_fp16 = const()[name = string("op_10271_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10272_cast_fp16 = mul(x = var_10270_cast_fp16, y = var_10271_to_fp16)[name = string("op_10272_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10273_cast_fp16 = add(x = var_10268_cast_fp16, y = var_10272_cast_fp16)[name = string("op_10273_cast_fp16")];
            fp32 var_10274_epsilon_0 = const()[name = string("op_10274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10274_cast_fp16 = rsqrt(epsilon = var_10274_epsilon_0, x = var_10273_cast_fp16)[name = string("op_10274_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10275_cast_fp16 = mul(x = x_543_cast_fp16, y = var_10274_cast_fp16)[name = string("op_10275_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_10275_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_10284 = const()[name = string("op_10284"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_10285_cast_fp16 = reshape(shape = var_10284, x = linear_189_cast_fp16)[name = string("op_10285_cast_fp16")];
            tensor<int32, [4]> x_545_perm_0 = const()[name = string("x_545_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10292 = const()[name = string("op_10292"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_545_cast_fp16 = transpose(perm = x_545_perm_0, x = var_10285_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 16, 16, 128]> var_10295_cast_fp16 = abs(x = x_545_cast_fp16)[name = string("op_10295_cast_fp16")];
            tensor<int32, [1]> peak_219_axes_0 = const()[name = string("peak_219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_219_cast_fp16 = reduce_max(axes = peak_219_axes_0, keep_dims = var_10292, x = var_10295_cast_fp16)[name = string("peak_219_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_547_cast_fp16 = real_div(x = x_545_cast_fp16, y = peak_219_cast_fp16)[name = string("x_547_cast_fp16")];
            fp16 var_10291_promoted_to_fp16 = const()[name = string("op_10291_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_10299_cast_fp16 = pow(x = x_547_cast_fp16, y = var_10291_promoted_to_fp16)[name = string("op_10299_cast_fp16")];
            tensor<int32, [1]> var_10301_axes_0 = const()[name = string("op_10301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10301_keep_dims_0 = const()[name = string("op_10301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_10301_cast_fp16 = reduce_mean(axes = var_10301_axes_0, keep_dims = var_10301_keep_dims_0, x = var_10299_cast_fp16)[name = string("op_10301_cast_fp16")];
            fp16 var_10291_promoted_1_to_fp16 = const()[name = string("op_10291_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_10302_cast_fp16 = pow(x = peak_219_cast_fp16, y = var_10291_promoted_1_to_fp16)[name = string("op_10302_cast_fp16")];
            fp32 var_10303_epsilon_0 = const()[name = string("op_10303_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_10303_cast_fp16 = inverse(epsilon = var_10303_epsilon_0, x = var_10302_cast_fp16)[name = string("op_10303_cast_fp16")];
            fp16 var_10304_to_fp16 = const()[name = string("op_10304_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_10305_cast_fp16 = mul(x = var_10303_cast_fp16, y = var_10304_to_fp16)[name = string("op_10305_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_10306_cast_fp16 = add(x = var_10301_cast_fp16, y = var_10305_cast_fp16)[name = string("op_10306_cast_fp16")];
            fp32 var_10307_epsilon_0 = const()[name = string("op_10307_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_10307_cast_fp16 = rsqrt(epsilon = var_10307_epsilon_0, x = var_10306_cast_fp16)[name = string("op_10307_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_10308_cast_fp16 = mul(x = x_547_cast_fp16, y = var_10307_cast_fp16)[name = string("op_10308_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_553_cast_fp16 = mul(x = var_10308_cast_fp16, y = model_layers_27_self_attn_q_norm_weight_to_fp16)[name = string("x_553_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_10317 = const()[name = string("op_10317"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10318_cast_fp16 = reshape(shape = var_10317, x = linear_190_cast_fp16)[name = string("op_10318_cast_fp16")];
            tensor<int32, [4]> x_549_perm_0 = const()[name = string("x_549_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10325 = const()[name = string("op_10325"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_549_cast_fp16 = transpose(perm = x_549_perm_0, x = var_10318_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 8, 16, 128]> var_10328_cast_fp16 = abs(x = x_549_cast_fp16)[name = string("op_10328_cast_fp16")];
            tensor<int32, [1]> peak_221_axes_0 = const()[name = string("peak_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_221_cast_fp16 = reduce_max(axes = peak_221_axes_0, keep_dims = var_10325, x = var_10328_cast_fp16)[name = string("peak_221_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_551_cast_fp16 = real_div(x = x_549_cast_fp16, y = peak_221_cast_fp16)[name = string("x_551_cast_fp16")];
            fp16 var_10324_promoted_to_fp16 = const()[name = string("op_10324_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_10332_cast_fp16 = pow(x = x_551_cast_fp16, y = var_10324_promoted_to_fp16)[name = string("op_10332_cast_fp16")];
            tensor<int32, [1]> var_10334_axes_0 = const()[name = string("op_10334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10334_keep_dims_0 = const()[name = string("op_10334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_10334_cast_fp16 = reduce_mean(axes = var_10334_axes_0, keep_dims = var_10334_keep_dims_0, x = var_10332_cast_fp16)[name = string("op_10334_cast_fp16")];
            fp16 var_10324_promoted_1_to_fp16 = const()[name = string("op_10324_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_10335_cast_fp16 = pow(x = peak_221_cast_fp16, y = var_10324_promoted_1_to_fp16)[name = string("op_10335_cast_fp16")];
            fp32 var_10336_epsilon_0 = const()[name = string("op_10336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_10336_cast_fp16 = inverse(epsilon = var_10336_epsilon_0, x = var_10335_cast_fp16)[name = string("op_10336_cast_fp16")];
            fp16 var_10337_to_fp16 = const()[name = string("op_10337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_10338_cast_fp16 = mul(x = var_10336_cast_fp16, y = var_10337_to_fp16)[name = string("op_10338_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_10339_cast_fp16 = add(x = var_10334_cast_fp16, y = var_10338_cast_fp16)[name = string("op_10339_cast_fp16")];
            fp32 var_10340_epsilon_0 = const()[name = string("op_10340_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_10340_cast_fp16 = rsqrt(epsilon = var_10340_epsilon_0, x = var_10339_cast_fp16)[name = string("op_10340_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_10341_cast_fp16 = mul(x = x_551_cast_fp16, y = var_10340_cast_fp16)[name = string("op_10341_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_555_cast_fp16 = mul(x = var_10341_cast_fp16, y = model_layers_27_self_attn_k_norm_weight_to_fp16)[name = string("x_555_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_10350 = const()[name = string("op_10350"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10351_cast_fp16 = reshape(shape = var_10350, x = linear_191_cast_fp16)[name = string("op_10351_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_553_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_553_cast_fp16)[name = string("second_109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10365_cast_fp16 = mul(x = first_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10365_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10366_cast_fp16 = mul(x = second_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10366_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10368_cast_fp16 = sub(x = var_10365_cast_fp16, y = var_10366_cast_fp16)[name = string("op_10368_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10369_cast_fp16 = mul(x = second_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10369_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10370_cast_fp16 = mul(x = first_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10370_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10372_cast_fp16 = add(x = var_10369_cast_fp16, y = var_10370_cast_fp16)[name = string("op_10372_cast_fp16")];
            int32 var_10374 = const()[name = string("op_10374"), 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_10374, interleave = q_interleave_0, values = (var_10368_cast_fp16, var_10372_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_555_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_555_cast_fp16)[name = string("second_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10386_cast_fp16 = mul(x = first_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10386_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10387_cast_fp16 = mul(x = second_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10387_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10389_cast_fp16 = sub(x = var_10386_cast_fp16, y = var_10387_cast_fp16)[name = string("op_10389_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10390_cast_fp16 = mul(x = second_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10390_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10391_cast_fp16 = mul(x = first_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10391_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10393_cast_fp16 = add(x = var_10390_cast_fp16, y = var_10391_cast_fp16)[name = string("op_10393_cast_fp16")];
            int32 var_10395 = const()[name = string("op_10395"), 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_10395, interleave = k_interleave_0, values = (var_10389_cast_fp16, var_10393_cast_fp16))[name = string("k_cast_fp16")];
            tensor<int32, [4]> var_10406_begin_0 = const()[name = string("op_10406_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10406_end_0 = const()[name = string("op_10406_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10406_end_mask_0 = const()[name = string("op_10406_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10406_stride_0 = const()[name = string("op_10406_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10406_cast_fp16 = slice_by_index(begin = var_10406_begin_0, end = var_10406_end_0, end_mask = var_10406_end_mask_0, stride = var_10406_stride_0, x = q_cast_fp16)[name = string("op_10406_cast_fp16")];
            tensor<int32, [4]> var_10416_begin_0 = const()[name = string("op_10416_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_10416_end_0 = const()[name = string("op_10416_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10416_end_mask_0 = const()[name = string("op_10416_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10416_stride_0 = const()[name = string("op_10416_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10416_cast_fp16 = slice_by_index(begin = var_10416_begin_0, end = var_10416_end_0, end_mask = var_10416_end_mask_0, stride = var_10416_stride_0, x = q_cast_fp16)[name = string("op_10416_cast_fp16")];
            int32 var_10418 = const()[name = string("op_10418"), 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_10418, interleave = stacked_interleave_0, values = (var_10406_cast_fp16, var_10416_cast_fp16))[name = string("stacked_cast_fp16")];
            tensor<int32, [4]> var_10424_begin_0 = const()[name = string("op_10424_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_10424_end_0 = const()[name = string("op_10424_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_10424_end_mask_0 = const()[name = string("op_10424_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_10424_cast_fp16 = slice_by_index(begin = var_10424_begin_0, end = var_10424_end_0, end_mask = var_10424_end_mask_0, x = read_state_0)[name = string("op_10424_cast_fp16")];
            tensor<int32, [4]> var_10434_begin_0 = const()[name = string("op_10434_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10434_end_0 = const()[name = string("op_10434_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_10434_end_mask_0 = const()[name = string("op_10434_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_10434_cast_fp16 = slice_by_index(begin = var_10434_begin_0, end = var_10434_end_0, end_mask = var_10434_end_mask_0, x = var_10424_cast_fp16)[name = string("op_10434_cast_fp16")];
            bool var_10443_transpose_x_1 = const()[name = string("op_10443_transpose_x_1"), val = bool(false)];
            bool var_10443_transpose_y_1 = const()[name = string("op_10443_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_10443_cast_fp16 = matmul(transpose_x = var_10443_transpose_x_1, transpose_y = var_10443_transpose_y_1, x = var_10434_cast_fp16, y = stacked_cast_fp16)[name = string("op_10443_cast_fp16")];
            tensor<int32, [4]> var_10446_perm_0 = const()[name = string("op_10446_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_10447_to_fp16 = const()[name = string("op_10447_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_10446_cast_fp16 = transpose(perm = var_10446_perm_0, x = var_10443_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, 32, 2048]> var_10448_cast_fp16 = mul(x = var_10446_cast_fp16, y = var_10447_to_fp16)[name = string("op_10448_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_cast_fp16 = add(x = var_10448_cast_fp16, y = cacheMask)[name = string("cache_weights_cast_fp16")];
            bool var_10454_transpose_x_1 = const()[name = string("op_10454_transpose_x_1"), val = bool(false)];
            bool var_10454_transpose_y_1 = const()[name = string("op_10454_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_10454_cast_fp16 = matmul(transpose_x = var_10454_transpose_x_1, transpose_y = var_10454_transpose_y_1, x = stacked_cast_fp16, y = k_cast_fp16)[name = string("op_10454_cast_fp16")];
            fp16 var_10455_to_fp16 = const()[name = string("op_10455_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_10456_cast_fp16 = mul(x = var_10454_cast_fp16, y = var_10455_to_fp16)[name = string("op_10456_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_cast_fp16 = add(x = var_10456_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_cast_fp16")];
            int32 var_10460 = const()[name = string("op_10460"), val = int32(-1)];
            bool var_10461_interleave_0 = const()[name = string("op_10461_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_10461_cast_fp16 = concat(axis = var_10460, interleave = var_10461_interleave_0, values = (cache_weights_cast_fp16, call_weights_cast_fp16))[name = string("op_10461_cast_fp16")];
            int32 var_10462 = const()[name = string("op_10462"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_cast_fp16 = softmax(axis = var_10462, x = var_10461_cast_fp16)[name = string("shares_cast_fp16")];
            tensor<int32, [4]> var_10469_begin_0 = const()[name = string("op_10469_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10469_end_0 = const()[name = string("op_10469_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_10469_end_mask_0 = const()[name = string("op_10469_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_10469_cast_fp16 = slice_by_index(begin = var_10469_begin_0, end = var_10469_end_0, end_mask = var_10469_end_mask_0, x = shares_cast_fp16)[name = string("op_10469_cast_fp16")];
            tensor<int32, [4]> var_10474_begin_0 = const()[name = string("op_10474_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_10474_end_0 = const()[name = string("op_10474_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_10474_end_mask_0 = const()[name = string("op_10474_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_10474_cast_fp16 = slice_by_index(begin = var_10474_begin_0, end = var_10474_end_0, end_mask = var_10474_end_mask_0, x = read_state_1)[name = string("op_10474_cast_fp16")];
            tensor<int32, [4]> var_10484_begin_0 = const()[name = string("op_10484_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10484_end_0 = const()[name = string("op_10484_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_10484_end_mask_0 = const()[name = string("op_10484_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_10484_cast_fp16 = slice_by_index(begin = var_10484_begin_0, end = var_10484_end_0, end_mask = var_10484_end_mask_0, x = var_10474_cast_fp16)[name = string("op_10484_cast_fp16")];
            bool var_10490_transpose_x_0 = const()[name = string("op_10490_transpose_x_0"), val = bool(false)];
            bool var_10490_transpose_y_0 = const()[name = string("op_10490_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_10490_cast_fp16 = matmul(transpose_x = var_10490_transpose_x_0, transpose_y = var_10490_transpose_y_0, x = var_10469_cast_fp16, y = var_10484_cast_fp16)[name = string("op_10490_cast_fp16")];
            tensor<int32, [4]> var_10495_begin_0 = const()[name = string("op_10495_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_10495_end_0 = const()[name = string("op_10495_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_10495_end_mask_0 = const()[name = string("op_10495_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_10495_cast_fp16 = slice_by_index(begin = var_10495_begin_0, end = var_10495_end_0, end_mask = var_10495_end_mask_0, x = shares_cast_fp16)[name = string("op_10495_cast_fp16")];
            bool var_10496_transpose_x_0 = const()[name = string("op_10496_transpose_x_0"), val = bool(false)];
            bool var_10496_transpose_y_0 = const()[name = string("op_10496_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_10351_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 32, 128]> var_10496_cast_fp16 = matmul(transpose_x = var_10496_transpose_x_0, transpose_y = var_10496_transpose_y_0, x = var_10495_cast_fp16, y = v_cast_fp16)[name = string("op_10496_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_109_cast_fp16 = add(x = var_10490_cast_fp16, y = var_10496_cast_fp16)[name = string("mixed_109_cast_fp16")];
            tensor<int32, [4]> var_10513_begin_0 = const()[name = string("op_10513_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10513_end_0 = const()[name = string("op_10513_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_10513_end_mask_0 = const()[name = string("op_10513_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10513_cast_fp16 = slice_by_index(begin = var_10513_begin_0, end = var_10513_end_0, end_mask = var_10513_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_10513_cast_fp16")];
            tensor<int32, [4]> var_10528_begin_0 = const()[name = string("op_10528_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_10528_end_0 = const()[name = string("op_10528_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_10528_end_mask_0 = const()[name = string("op_10528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10528_cast_fp16 = slice_by_index(begin = var_10528_begin_0, end = var_10528_end_0, end_mask = var_10528_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_10528_cast_fp16")];
            int32 var_10530 = const()[name = string("op_10530"), 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_10530, interleave = mixed_interleave_0, values = (var_10513_cast_fp16, var_10528_cast_fp16))[name = string("mixed_cast_fp16")];
            tensor<int32, [4]> var_10534_perm_0 = const()[name = string("op_10534_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_10538 = const()[name = string("op_10538"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_10534_cast_fp16 = transpose(perm = var_10534_perm_0, x = mixed_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 16, 2048]> input_273_cast_fp16 = reshape(shape = var_10538, x = var_10534_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_557_cast_fp16 = add(x = x_541_cast_fp16, y = linear_192_cast_fp16)[name = string("x_557_cast_fp16")];
            bool var_10548 = const()[name = string("op_10548"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10551_cast_fp16 = abs(x = x_557_cast_fp16)[name = string("op_10551_cast_fp16")];
            tensor<int32, [1]> peak_223_axes_0 = const()[name = string("peak_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_223_cast_fp16 = reduce_max(axes = peak_223_axes_0, keep_dims = var_10548, x = var_10551_cast_fp16)[name = string("peak_223_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_559_cast_fp16 = real_div(x = x_557_cast_fp16, y = peak_223_cast_fp16)[name = string("x_559_cast_fp16")];
            fp16 var_10547_promoted_to_fp16 = const()[name = string("op_10547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10555_cast_fp16 = pow(x = x_559_cast_fp16, y = var_10547_promoted_to_fp16)[name = string("op_10555_cast_fp16")];
            tensor<int32, [1]> var_10557_axes_0 = const()[name = string("op_10557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10557_keep_dims_0 = const()[name = string("op_10557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10557_cast_fp16 = reduce_mean(axes = var_10557_axes_0, keep_dims = var_10557_keep_dims_0, x = var_10555_cast_fp16)[name = string("op_10557_cast_fp16")];
            fp16 var_10547_promoted_1_to_fp16 = const()[name = string("op_10547_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10558_cast_fp16 = pow(x = peak_223_cast_fp16, y = var_10547_promoted_1_to_fp16)[name = string("op_10558_cast_fp16")];
            fp32 var_10559_epsilon_0 = const()[name = string("op_10559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10559_cast_fp16 = inverse(epsilon = var_10559_epsilon_0, x = var_10558_cast_fp16)[name = string("op_10559_cast_fp16")];
            fp16 var_10560_to_fp16 = const()[name = string("op_10560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10561_cast_fp16 = mul(x = var_10559_cast_fp16, y = var_10560_to_fp16)[name = string("op_10561_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10562_cast_fp16 = add(x = var_10557_cast_fp16, y = var_10561_cast_fp16)[name = string("op_10562_cast_fp16")];
            fp32 var_10563_epsilon_0 = const()[name = string("op_10563_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10563_cast_fp16 = rsqrt(epsilon = var_10563_epsilon_0, x = var_10562_cast_fp16)[name = string("op_10563_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10564_cast_fp16 = mul(x = x_559_cast_fp16, y = var_10563_cast_fp16)[name = string("op_10564_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_10564_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_10569_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_10569_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_10569_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_561_cast_fp16 = add(x = x_557_cast_fp16, y = linear_195_cast_fp16)[name = string("x_561_cast_fp16")];
            int32 var_10579 = const()[name = string("op_10579"), val = int32(1)];
            int32 x_563_batch_dims_0 = const()[name = string("x_563_batch_dims_0"), val = int32(0)];
            bool x_563_validate_indices_0 = const()[name = string("x_563_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_563_cast_fp16_cast_uint16 = gather(axis = var_10579, batch_dims = x_563_batch_dims_0, indices = answerRow_to_uint16, validate_indices = x_563_validate_indices_0, x = x_561_cast_fp16)[name = string("x_563_cast_fp16_cast_uint16")];
            bool var_10584 = const()[name = string("op_10584"), val = bool(true)];
            tensor<fp16, [1, 1, 1024]> var_10587_cast_fp16 = abs(x = x_563_cast_fp16_cast_uint16)[name = string("op_10587_cast_fp16")];
            tensor<int32, [1]> peak_axes_0 = const()[name = string("peak_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 1]> peak_cast_fp16 = reduce_max(axes = peak_axes_0, keep_dims = var_10584, x = var_10587_cast_fp16)[name = string("peak_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> x_cast_fp16 = real_div(x = x_563_cast_fp16_cast_uint16, y = peak_cast_fp16)[name = string("x_cast_fp16")];
            fp16 var_10583_promoted_to_fp16 = const()[name = string("op_10583_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1024]> var_10591_cast_fp16 = pow(x = x_cast_fp16, y = var_10583_promoted_to_fp16)[name = string("op_10591_cast_fp16")];
            tensor<int32, [1]> var_10593_axes_0 = const()[name = string("op_10593_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10593_keep_dims_0 = const()[name = string("op_10593_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1]> var_10593_cast_fp16 = reduce_mean(axes = var_10593_axes_0, keep_dims = var_10593_keep_dims_0, x = var_10591_cast_fp16)[name = string("op_10593_cast_fp16")];
            fp16 var_10583_promoted_1_to_fp16 = const()[name = string("op_10583_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1]> var_10594_cast_fp16 = pow(x = peak_cast_fp16, y = var_10583_promoted_1_to_fp16)[name = string("op_10594_cast_fp16")];
            fp32 var_10595_epsilon_0 = const()[name = string("op_10595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 1]> var_10595_cast_fp16 = inverse(epsilon = var_10595_epsilon_0, x = var_10594_cast_fp16)[name = string("op_10595_cast_fp16")];
            fp16 var_10596_to_fp16 = const()[name = string("op_10596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_10597_cast_fp16 = mul(x = var_10595_cast_fp16, y = var_10596_to_fp16)[name = string("op_10597_cast_fp16")];
            tensor<fp16, [1, 1, 1]> var_10598_cast_fp16 = add(x = var_10593_cast_fp16, y = var_10597_cast_fp16)[name = string("op_10598_cast_fp16")];
            fp32 var_10599_epsilon_0 = const()[name = string("op_10599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1]> var_10599_cast_fp16 = rsqrt(epsilon = var_10599_epsilon_0, x = var_10598_cast_fp16)[name = string("op_10599_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> var_10600_cast_fp16 = mul(x = x_cast_fp16, y = var_10599_cast_fp16)[name = string("op_10600_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_10600_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_10606_axis_0 = const()[name = string("op_10606_axis_0"), val = int32(-1)];
            bool var_10606_keep_dims_0 = const()[name = string("op_10606_keep_dims_0"), val = bool(false)];
            string var_10606_output_dtype_0 = const()[name = string("op_10606_output_dtype_0"), val = string("int32")];
            tensor<int32, [1, 1]> nextToken = reduce_argmax(axis = var_10606_axis_0, keep_dims = var_10606_keep_dims_0, output_dtype = var_10606_output_dtype_0, x = linear_196_cast_fp16)[name = string("op_10606_cast_fp16")];
            int32 var_10613 = const()[name = string("op_10613"), val = int32(0)];
            bool var_10614_interleave_0 = const()[name = string("op_10614_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newKeys = concat(axis = var_10613, interleave = var_10614_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_10614_cast_fp16")];
            int32 var_10616 = const()[name = string("op_10616"), val = int32(0)];
            bool var_10617_interleave_0 = const()[name = string("op_10617_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newValues = concat(axis = var_10616, interleave = var_10617_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_10617_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")];
            bool var_1538 = const()[name = string("op_1538"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1541_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [1]> peak_1_axes_0 = const()[name = string("peak_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_1_cast_fp16 = reduce_max(axes = peak_1_axes_0, keep_dims = var_1538, x = var_1541_cast_fp16)[name = string("peak_1_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = peak_1_cast_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_1537_promoted_to_fp16 = const()[name = string("op_1537_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1545_cast_fp16 = pow(x = x_3_cast_fp16, y = var_1537_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            tensor<int32, [1]> var_1547_axes_0 = const()[name = string("op_1547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1547_keep_dims_0 = const()[name = string("op_1547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1547_cast_fp16 = reduce_mean(axes = var_1547_axes_0, keep_dims = var_1547_keep_dims_0, x = var_1545_cast_fp16)[name = string("op_1547_cast_fp16")];
            fp16 var_1537_promoted_1_to_fp16 = const()[name = string("op_1537_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1548_cast_fp16 = pow(x = peak_1_cast_fp16, y = var_1537_promoted_1_to_fp16)[name = string("op_1548_cast_fp16")];
            fp32 var_1549_epsilon_0 = const()[name = string("op_1549_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1549_cast_fp16 = inverse(epsilon = var_1549_epsilon_0, x = var_1548_cast_fp16)[name = string("op_1549_cast_fp16")];
            fp16 var_1550_to_fp16 = const()[name = string("op_1550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1551_cast_fp16 = mul(x = var_1549_cast_fp16, y = var_1550_to_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1552_cast_fp16 = add(x = var_1547_cast_fp16, y = var_1551_cast_fp16)[name = string("op_1552_cast_fp16")];
            fp32 var_1553_epsilon_0 = const()[name = string("op_1553_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1553_cast_fp16 = rsqrt(epsilon = var_1553_epsilon_0, x = var_1552_cast_fp16)[name = string("op_1553_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1554_cast_fp16 = mul(x = x_3_cast_fp16, y = var_1553_cast_fp16)[name = string("op_1554_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_1554_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_1563 = const()[name = string("op_1563"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1564_cast_fp16 = reshape(shape = var_1563, x = linear_0_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1571 = const()[name = string("op_1571"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_1564_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 16, 16, 128]> var_1574_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<int32, [1]> peak_3_axes_0 = const()[name = string("peak_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_3_cast_fp16 = reduce_max(axes = peak_3_axes_0, keep_dims = var_1571, x = var_1574_cast_fp16)[name = string("peak_3_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = peak_3_cast_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1570_promoted_to_fp16 = const()[name = string("op_1570_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_1578_cast_fp16 = pow(x = x_7_cast_fp16, y = var_1570_promoted_to_fp16)[name = string("op_1578_cast_fp16")];
            tensor<int32, [1]> var_1580_axes_0 = const()[name = string("op_1580_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1580_keep_dims_0 = const()[name = string("op_1580_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1580_cast_fp16 = reduce_mean(axes = var_1580_axes_0, keep_dims = var_1580_keep_dims_0, x = var_1578_cast_fp16)[name = string("op_1580_cast_fp16")];
            fp16 var_1570_promoted_1_to_fp16 = const()[name = string("op_1570_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_1581_cast_fp16 = pow(x = peak_3_cast_fp16, y = var_1570_promoted_1_to_fp16)[name = string("op_1581_cast_fp16")];
            fp32 var_1582_epsilon_0 = const()[name = string("op_1582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_1582_cast_fp16 = inverse(epsilon = var_1582_epsilon_0, x = var_1581_cast_fp16)[name = string("op_1582_cast_fp16")];
            fp16 var_1583_to_fp16 = const()[name = string("op_1583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1584_cast_fp16 = mul(x = var_1582_cast_fp16, y = var_1583_to_fp16)[name = string("op_1584_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_1585_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1584_cast_fp16)[name = string("op_1585_cast_fp16")];
            fp32 var_1586_epsilon_0 = const()[name = string("op_1586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1586_cast_fp16 = rsqrt(epsilon = var_1586_epsilon_0, x = var_1585_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1587_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1586_cast_fp16)[name = string("op_1587_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_13_cast_fp16 = mul(x = var_1587_cast_fp16, y = model_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_13_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_1596 = const()[name = string("op_1596"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1597_cast_fp16 = reshape(shape = var_1596, x = linear_1_cast_fp16)[name = string("op_1597_cast_fp16")];
            tensor<int32, [4]> x_9_perm_0 = const()[name = string("x_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1604 = const()[name = string("op_1604"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = var_1597_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 8, 16, 128]> var_1607_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<int32, [1]> peak_5_axes_0 = const()[name = string("peak_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_5_cast_fp16 = reduce_max(axes = peak_5_axes_0, keep_dims = var_1604, x = var_1607_cast_fp16)[name = string("peak_5_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = peak_5_cast_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1603_promoted_to_fp16 = const()[name = string("op_1603_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_1611_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1603_promoted_to_fp16)[name = string("op_1611_cast_fp16")];
            tensor<int32, [1]> var_1613_axes_0 = const()[name = string("op_1613_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1613_keep_dims_0 = const()[name = string("op_1613_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1613_cast_fp16 = reduce_mean(axes = var_1613_axes_0, keep_dims = var_1613_keep_dims_0, x = var_1611_cast_fp16)[name = string("op_1613_cast_fp16")];
            fp16 var_1603_promoted_1_to_fp16 = const()[name = string("op_1603_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_1614_cast_fp16 = pow(x = peak_5_cast_fp16, y = var_1603_promoted_1_to_fp16)[name = string("op_1614_cast_fp16")];
            fp32 var_1615_epsilon_0 = const()[name = string("op_1615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_1615_cast_fp16 = inverse(epsilon = var_1615_epsilon_0, x = var_1614_cast_fp16)[name = string("op_1615_cast_fp16")];
            fp16 var_1616_to_fp16 = const()[name = string("op_1616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1617_cast_fp16 = mul(x = var_1615_cast_fp16, y = var_1616_to_fp16)[name = string("op_1617_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_1618_cast_fp16 = add(x = var_1613_cast_fp16, y = var_1617_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1619_cast_fp16 = rsqrt(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1620_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1619_cast_fp16)[name = string("op_1620_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_15_cast_fp16 = mul(x = var_1620_cast_fp16, y = model_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_15_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_1629 = const()[name = string("op_1629"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1630_cast_fp16 = reshape(shape = var_1629, x = linear_2_cast_fp16)[name = string("op_1630_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_13_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_13_cast_fp16)[name = string("second_1_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1644_cast_fp16 = mul(x = first_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1644_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1645_cast_fp16 = mul(x = second_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 16, 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, 16, 16, 64]> var_1648_cast_fp16 = mul(x = second_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1649_cast_fp16 = mul(x = first_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1649_cast_fp16")];
            tensor<fp16, [1, 16, 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 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_1653, interleave = q_1_interleave_0, values = (var_1647_cast_fp16, var_1651_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_15_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_15_cast_fp16)[name = string("second_3_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1665_cast_fp16 = mul(x = first_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1665_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1666_cast_fp16 = mul(x = second_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1666_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1668_cast_fp16 = sub(x = var_1665_cast_fp16, y = var_1666_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1669_cast_fp16 = mul(x = second_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1670_cast_fp16 = mul(x = first_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1672_cast_fp16 = add(x = var_1669_cast_fp16, y = var_1670_cast_fp16)[name = string("op_1672_cast_fp16")];
            int32 var_1674 = const()[name = string("op_1674"), 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_1674, interleave = k_1_interleave_0, values = (var_1668_cast_fp16, var_1672_cast_fp16))[name = string("k_1_cast_fp16")];
            tensor<int32, [4]> var_1685_begin_0 = const()[name = string("op_1685_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1685_end_0 = const()[name = string("op_1685_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1685_end_mask_0 = const()[name = string("op_1685_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1685_stride_0 = const()[name = string("op_1685_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1685_cast_fp16 = slice_by_index(begin = var_1685_begin_0, end = var_1685_end_0, end_mask = var_1685_end_mask_0, stride = var_1685_stride_0, x = q_1_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> var_1695_begin_0 = const()[name = string("op_1695_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1695_end_0 = const()[name = string("op_1695_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1695_end_mask_0 = const()[name = string("op_1695_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1695_stride_0 = const()[name = string("op_1695_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1695_cast_fp16 = slice_by_index(begin = var_1695_begin_0, end = var_1695_end_0, end_mask = var_1695_end_mask_0, stride = var_1695_stride_0, x = q_1_cast_fp16)[name = string("op_1695_cast_fp16")];
            int32 var_1697 = const()[name = string("op_1697"), 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_1697, interleave = stacked_1_interleave_0, values = (var_1685_cast_fp16, var_1695_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_1703_begin_0 = const()[name = string("op_1703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1703_end_0 = const()[name = string("op_1703_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_1703_end_mask_0 = const()[name = string("op_1703_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_1703_cast_fp16 = slice_by_index(begin = var_1703_begin_0, end = var_1703_end_0, end_mask = var_1703_end_mask_0, x = read_state_0)[name = string("op_1703_cast_fp16")];
            tensor<int32, [4]> var_1713_begin_0 = const()[name = string("op_1713_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1713_end_0 = const()[name = string("op_1713_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_1713_end_mask_0 = const()[name = string("op_1713_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_1713_cast_fp16 = slice_by_index(begin = var_1713_begin_0, end = var_1713_end_0, end_mask = var_1713_end_mask_0, x = var_1703_cast_fp16)[name = string("op_1713_cast_fp16")];
            bool var_1722_transpose_x_1 = const()[name = string("op_1722_transpose_x_1"), val = bool(false)];
            bool var_1722_transpose_y_1 = const()[name = string("op_1722_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_1722_cast_fp16 = matmul(transpose_x = var_1722_transpose_x_1, transpose_y = var_1722_transpose_y_1, x = var_1713_cast_fp16, y = stacked_1_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<int32, [4]> var_1725_perm_0 = const()[name = string("op_1725_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_1726_to_fp16 = const()[name = string("op_1726_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_1725_cast_fp16 = transpose(perm = var_1725_perm_0, x = var_1722_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 8, 32, 4096]> var_1727_cast_fp16 = mul(x = var_1725_cast_fp16, y = var_1726_to_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_1_cast_fp16 = add(x = var_1727_cast_fp16, y = cacheMask)[name = string("cache_weights_1_cast_fp16")];
            bool var_1733_transpose_x_1 = const()[name = string("op_1733_transpose_x_1"), val = bool(false)];
            bool var_1733_transpose_y_1 = const()[name = string("op_1733_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_1733_cast_fp16 = matmul(transpose_x = var_1733_transpose_x_1, transpose_y = var_1733_transpose_y_1, x = stacked_1_cast_fp16, y = k_1_cast_fp16)[name = string("op_1733_cast_fp16")];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_1735_cast_fp16 = mul(x = var_1733_cast_fp16, y = var_1734_to_fp16)[name = string("op_1735_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_1735_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_1_cast_fp16")];
            int32 var_1739 = const()[name = string("op_1739"), val = int32(-1)];
            bool var_1740_interleave_0 = const()[name = string("op_1740_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_1740_cast_fp16 = concat(axis = var_1739, interleave = var_1740_interleave_0, values = (cache_weights_1_cast_fp16, call_weights_1_cast_fp16))[name = string("op_1740_cast_fp16")];
            int32 var_1741 = const()[name = string("op_1741"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_1_cast_fp16 = softmax(axis = var_1741, x = var_1740_cast_fp16)[name = string("shares_1_cast_fp16")];
            tensor<int32, [4]> var_1748_begin_0 = const()[name = string("op_1748_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1748_end_0 = const()[name = string("op_1748_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_1748_end_mask_0 = const()[name = string("op_1748_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_1748_cast_fp16 = slice_by_index(begin = var_1748_begin_0, end = var_1748_end_0, end_mask = var_1748_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [28, 8, 4112, 128]> read_state_1 = read_state(input = valueCache)[name = string("read_state_1")];
            tensor<int32, [4]> var_1753_begin_0 = const()[name = string("op_1753_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1753_end_0 = const()[name = string("op_1753_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_1753_end_mask_0 = const()[name = string("op_1753_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> 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 = read_state_1)[name = string("op_1753_cast_fp16")];
            tensor<int32, [4]> var_1763_begin_0 = const()[name = string("op_1763_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1763_end_0 = const()[name = string("op_1763_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_1763_end_mask_0 = const()[name = string("op_1763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_1763_cast_fp16 = slice_by_index(begin = var_1763_begin_0, end = var_1763_end_0, end_mask = var_1763_end_mask_0, x = var_1753_cast_fp16)[name = string("op_1763_cast_fp16")];
            bool var_1769_transpose_x_0 = const()[name = string("op_1769_transpose_x_0"), val = bool(false)];
            bool var_1769_transpose_y_0 = const()[name = string("op_1769_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1769_cast_fp16 = matmul(transpose_x = var_1769_transpose_x_0, transpose_y = var_1769_transpose_y_0, x = var_1748_cast_fp16, y = var_1763_cast_fp16)[name = string("op_1769_cast_fp16")];
            tensor<int32, [4]> var_1774_begin_0 = const()[name = string("op_1774_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_1774_end_0 = const()[name = string("op_1774_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_1774_end_mask_0 = const()[name = string("op_1774_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_1774_cast_fp16 = slice_by_index(begin = var_1774_begin_0, end = var_1774_end_0, end_mask = var_1774_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1774_cast_fp16")];
            bool var_1775_transpose_x_0 = const()[name = string("op_1775_transpose_x_0"), val = bool(false)];
            bool var_1775_transpose_y_0 = const()[name = string("op_1775_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_1630_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 8, 32, 128]> var_1775_cast_fp16 = matmul(transpose_x = var_1775_transpose_x_0, transpose_y = var_1775_transpose_y_0, x = var_1774_cast_fp16, y = v_1_cast_fp16)[name = string("op_1775_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_1_cast_fp16 = add(x = var_1769_cast_fp16, y = var_1775_cast_fp16)[name = string("mixed_1_cast_fp16")];
            tensor<int32, [4]> var_1792_begin_0 = const()[name = string("op_1792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1792_end_0 = const()[name = string("op_1792_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_1792_end_mask_0 = const()[name = string("op_1792_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1792_cast_fp16 = slice_by_index(begin = var_1792_begin_0, end = var_1792_end_0, end_mask = var_1792_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1792_cast_fp16")];
            tensor<int32, [4]> var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1807_cast_fp16")];
            int32 var_1809 = const()[name = string("op_1809"), 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_1809, interleave = mixed_3_interleave_0, values = (var_1792_cast_fp16, var_1807_cast_fp16))[name = string("mixed_3_cast_fp16")];
            tensor<int32, [4]> var_1813_perm_0 = const()[name = string("op_1813_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1817 = const()[name = string("op_1817"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_1813_cast_fp16 = transpose(perm = var_1813_perm_0, x = mixed_3_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 16, 2048]> input_3_cast_fp16 = reshape(shape = var_1817, x = var_1813_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_17_cast_fp16 = add(x = x_1_cast_fp16, y = linear_3_cast_fp16)[name = string("x_17_cast_fp16")];
            bool var_1827 = const()[name = string("op_1827"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1830_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_1830_cast_fp16")];
            tensor<int32, [1]> peak_7_axes_0 = const()[name = string("peak_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_7_cast_fp16 = reduce_max(axes = peak_7_axes_0, keep_dims = var_1827, x = var_1830_cast_fp16)[name = string("peak_7_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = peak_7_cast_fp16)[name = string("x_19_cast_fp16")];
            fp16 var_1826_promoted_to_fp16 = const()[name = string("op_1826_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1834_cast_fp16 = pow(x = x_19_cast_fp16, y = var_1826_promoted_to_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [1]> var_1836_axes_0 = const()[name = string("op_1836_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1836_keep_dims_0 = const()[name = string("op_1836_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1836_cast_fp16 = reduce_mean(axes = var_1836_axes_0, keep_dims = var_1836_keep_dims_0, x = var_1834_cast_fp16)[name = string("op_1836_cast_fp16")];
            fp16 var_1826_promoted_1_to_fp16 = const()[name = string("op_1826_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1837_cast_fp16 = pow(x = peak_7_cast_fp16, y = var_1826_promoted_1_to_fp16)[name = string("op_1837_cast_fp16")];
            fp32 var_1838_epsilon_0 = const()[name = string("op_1838_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1838_cast_fp16 = inverse(epsilon = var_1838_epsilon_0, x = var_1837_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 = mul(x = var_1838_cast_fp16, y = var_1839_to_fp16)[name = string("op_1840_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1841_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1840_cast_fp16)[name = string("op_1841_cast_fp16")];
            fp32 var_1842_epsilon_0 = const()[name = string("op_1842_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1842_cast_fp16 = rsqrt(epsilon = var_1842_epsilon_0, x = var_1841_cast_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1843_cast_fp16 = mul(x = x_19_cast_fp16, y = var_1842_cast_fp16)[name = string("op_1843_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_1843_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_1848_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_1848_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_1848_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_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_6_cast_fp16)[name = string("x_21_cast_fp16")];
            bool var_1861 = const()[name = string("op_1861"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1864_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1864_cast_fp16")];
            tensor<int32, [1]> peak_9_axes_0 = const()[name = string("peak_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_9_cast_fp16 = reduce_max(axes = peak_9_axes_0, keep_dims = var_1861, x = var_1864_cast_fp16)[name = string("peak_9_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = peak_9_cast_fp16)[name = string("x_23_cast_fp16")];
            fp16 var_1860_promoted_to_fp16 = const()[name = string("op_1860_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1868_cast_fp16 = pow(x = x_23_cast_fp16, y = var_1860_promoted_to_fp16)[name = string("op_1868_cast_fp16")];
            tensor<int32, [1]> var_1870_axes_0 = const()[name = string("op_1870_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1870_keep_dims_0 = const()[name = string("op_1870_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1870_cast_fp16 = reduce_mean(axes = var_1870_axes_0, keep_dims = var_1870_keep_dims_0, x = var_1868_cast_fp16)[name = string("op_1870_cast_fp16")];
            fp16 var_1860_promoted_1_to_fp16 = const()[name = string("op_1860_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1871_cast_fp16 = pow(x = peak_9_cast_fp16, y = var_1860_promoted_1_to_fp16)[name = string("op_1871_cast_fp16")];
            fp32 var_1872_epsilon_0 = const()[name = string("op_1872_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1872_cast_fp16 = inverse(epsilon = var_1872_epsilon_0, x = var_1871_cast_fp16)[name = string("op_1872_cast_fp16")];
            fp16 var_1873_to_fp16 = const()[name = string("op_1873_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1874_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1873_to_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1875_cast_fp16 = add(x = var_1870_cast_fp16, y = var_1874_cast_fp16)[name = string("op_1875_cast_fp16")];
            fp32 var_1876_epsilon_0 = const()[name = string("op_1876_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1876_cast_fp16 = rsqrt(epsilon = var_1876_epsilon_0, x = var_1875_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1877_cast_fp16 = mul(x = x_23_cast_fp16, y = var_1876_cast_fp16)[name = string("op_1877_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_1877_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_1886 = const()[name = string("op_1886"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1887_cast_fp16 = reshape(shape = var_1886, x = linear_7_cast_fp16)[name = string("op_1887_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1894 = const()[name = string("op_1894"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_1887_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 16, 16, 128]> var_1897_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1897_cast_fp16")];
            tensor<int32, [1]> peak_11_axes_0 = const()[name = string("peak_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_11_cast_fp16 = reduce_max(axes = peak_11_axes_0, keep_dims = var_1894, x = var_1897_cast_fp16)[name = string("peak_11_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = peak_11_cast_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_1893_promoted_to_fp16 = const()[name = string("op_1893_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_1901_cast_fp16 = pow(x = x_27_cast_fp16, y = var_1893_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_1903_axes_0 = const()[name = string("op_1903_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1903_keep_dims_0 = const()[name = string("op_1903_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1903_cast_fp16 = reduce_mean(axes = var_1903_axes_0, keep_dims = var_1903_keep_dims_0, x = var_1901_cast_fp16)[name = string("op_1903_cast_fp16")];
            fp16 var_1893_promoted_1_to_fp16 = const()[name = string("op_1893_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_1904_cast_fp16 = pow(x = peak_11_cast_fp16, y = var_1893_promoted_1_to_fp16)[name = string("op_1904_cast_fp16")];
            fp32 var_1905_epsilon_0 = const()[name = string("op_1905_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_1905_cast_fp16 = inverse(epsilon = var_1905_epsilon_0, x = var_1904_cast_fp16)[name = string("op_1905_cast_fp16")];
            fp16 var_1906_to_fp16 = const()[name = string("op_1906_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1907_cast_fp16 = mul(x = var_1905_cast_fp16, y = var_1906_to_fp16)[name = string("op_1907_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_1908_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1907_cast_fp16)[name = string("op_1908_cast_fp16")];
            fp32 var_1909_epsilon_0 = const()[name = string("op_1909_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1909_cast_fp16 = rsqrt(epsilon = var_1909_epsilon_0, x = var_1908_cast_fp16)[name = string("op_1909_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1910_cast_fp16 = mul(x = x_27_cast_fp16, y = var_1909_cast_fp16)[name = string("op_1910_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_33_cast_fp16 = mul(x = var_1910_cast_fp16, y = model_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_33_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_1919 = const()[name = string("op_1919"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1920_cast_fp16 = reshape(shape = var_1919, x = linear_8_cast_fp16)[name = string("op_1920_cast_fp16")];
            tensor<int32, [4]> x_29_perm_0 = const()[name = string("x_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1927 = const()[name = string("op_1927"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_29_cast_fp16 = transpose(perm = x_29_perm_0, x = var_1920_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 8, 16, 128]> var_1930_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1930_cast_fp16")];
            tensor<int32, [1]> peak_13_axes_0 = const()[name = string("peak_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_13_cast_fp16 = reduce_max(axes = peak_13_axes_0, keep_dims = var_1927, x = var_1930_cast_fp16)[name = string("peak_13_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = peak_13_cast_fp16)[name = string("x_31_cast_fp16")];
            fp16 var_1926_promoted_to_fp16 = const()[name = string("op_1926_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_1934_cast_fp16 = pow(x = x_31_cast_fp16, y = var_1926_promoted_to_fp16)[name = string("op_1934_cast_fp16")];
            tensor<int32, [1]> var_1936_axes_0 = const()[name = string("op_1936_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1936_keep_dims_0 = const()[name = string("op_1936_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1936_cast_fp16 = reduce_mean(axes = var_1936_axes_0, keep_dims = var_1936_keep_dims_0, x = var_1934_cast_fp16)[name = string("op_1936_cast_fp16")];
            fp16 var_1926_promoted_1_to_fp16 = const()[name = string("op_1926_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_1937_cast_fp16 = pow(x = peak_13_cast_fp16, y = var_1926_promoted_1_to_fp16)[name = string("op_1937_cast_fp16")];
            fp32 var_1938_epsilon_0 = const()[name = string("op_1938_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_1938_cast_fp16 = inverse(epsilon = var_1938_epsilon_0, x = var_1937_cast_fp16)[name = string("op_1938_cast_fp16")];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1940_cast_fp16 = mul(x = var_1938_cast_fp16, y = var_1939_to_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_1941_cast_fp16 = add(x = var_1936_cast_fp16, y = var_1940_cast_fp16)[name = string("op_1941_cast_fp16")];
            fp32 var_1942_epsilon_0 = const()[name = string("op_1942_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1942_cast_fp16 = rsqrt(epsilon = var_1942_epsilon_0, x = var_1941_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1943_cast_fp16 = mul(x = x_31_cast_fp16, y = var_1942_cast_fp16)[name = string("op_1943_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_35_cast_fp16 = mul(x = var_1943_cast_fp16, y = model_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_35_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_1952 = const()[name = string("op_1952"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1953_cast_fp16 = reshape(shape = var_1952, x = linear_9_cast_fp16)[name = string("op_1953_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_33_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_33_cast_fp16)[name = string("second_5_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1967_cast_fp16 = mul(x = first_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1968_cast_fp16 = mul(x = second_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1970_cast_fp16 = sub(x = var_1967_cast_fp16, y = var_1968_cast_fp16)[name = string("op_1970_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1971_cast_fp16 = mul(x = second_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1972_cast_fp16 = mul(x = first_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1972_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1974_cast_fp16 = add(x = var_1971_cast_fp16, y = var_1972_cast_fp16)[name = string("op_1974_cast_fp16")];
            int32 var_1976 = const()[name = string("op_1976"), 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_1976, interleave = q_3_interleave_0, values = (var_1970_cast_fp16, var_1974_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_35_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_35_cast_fp16)[name = string("second_7_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1988_cast_fp16 = mul(x = first_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1988_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1989_cast_fp16 = mul(x = second_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1989_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1991_cast_fp16 = sub(x = var_1988_cast_fp16, y = var_1989_cast_fp16)[name = string("op_1991_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1992_cast_fp16 = mul(x = second_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1992_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1993_cast_fp16 = mul(x = first_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1993_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1995_cast_fp16 = add(x = var_1992_cast_fp16, y = var_1993_cast_fp16)[name = string("op_1995_cast_fp16")];
            int32 var_1997 = const()[name = string("op_1997"), 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_1997, interleave = k_3_interleave_0, values = (var_1991_cast_fp16, var_1995_cast_fp16))[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> var_2008_begin_0 = const()[name = string("op_2008_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2008_end_0 = const()[name = string("op_2008_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2008_end_mask_0 = const()[name = string("op_2008_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2008_stride_0 = const()[name = string("op_2008_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2008_cast_fp16 = slice_by_index(begin = var_2008_begin_0, end = var_2008_end_0, end_mask = var_2008_end_mask_0, stride = var_2008_stride_0, x = q_3_cast_fp16)[name = string("op_2008_cast_fp16")];
            tensor<int32, [4]> var_2018_begin_0 = const()[name = string("op_2018_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2018_end_0 = const()[name = string("op_2018_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2018_end_mask_0 = const()[name = string("op_2018_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2018_stride_0 = const()[name = string("op_2018_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2018_cast_fp16 = slice_by_index(begin = var_2018_begin_0, end = var_2018_end_0, end_mask = var_2018_end_mask_0, stride = var_2018_stride_0, x = q_3_cast_fp16)[name = string("op_2018_cast_fp16")];
            int32 var_2020 = const()[name = string("op_2020"), 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_2020, interleave = stacked_3_interleave_0, values = (var_2008_cast_fp16, var_2018_cast_fp16))[name = string("stacked_3_cast_fp16")];
            tensor<int32, [4]> var_2026_begin_0 = const()[name = string("op_2026_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2026_end_0 = const()[name = string("op_2026_end_0"), val = tensor<int32, [4]>([2, 8, 4112, 128])];
            tensor<bool, [4]> var_2026_end_mask_0 = const()[name = string("op_2026_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2026_cast_fp16 = slice_by_index(begin = var_2026_begin_0, end = var_2026_end_0, end_mask = var_2026_end_mask_0, x = read_state_0)[name = string("op_2026_cast_fp16")];
            tensor<int32, [4]> var_2036_begin_0 = const()[name = string("op_2036_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2036_end_0 = const()[name = string("op_2036_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2036_end_mask_0 = const()[name = string("op_2036_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2036_cast_fp16 = slice_by_index(begin = var_2036_begin_0, end = var_2036_end_0, end_mask = var_2036_end_mask_0, x = var_2026_cast_fp16)[name = string("op_2036_cast_fp16")];
            bool var_2045_transpose_x_1 = const()[name = string("op_2045_transpose_x_1"), val = bool(false)];
            bool var_2045_transpose_y_1 = const()[name = string("op_2045_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_2045_cast_fp16 = matmul(transpose_x = var_2045_transpose_x_1, transpose_y = var_2045_transpose_y_1, x = var_2036_cast_fp16, y = stacked_3_cast_fp16)[name = string("op_2045_cast_fp16")];
            tensor<int32, [4]> var_2048_perm_0 = const()[name = string("op_2048_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2049_to_fp16 = const()[name = string("op_2049_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_2048_cast_fp16 = transpose(perm = var_2048_perm_0, x = var_2045_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 32, 4096]> var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = string("op_2050_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_3_cast_fp16 = add(x = var_2050_cast_fp16, y = cacheMask)[name = string("cache_weights_3_cast_fp16")];
            bool var_2056_transpose_x_1 = const()[name = string("op_2056_transpose_x_1"), val = bool(false)];
            bool var_2056_transpose_y_1 = const()[name = string("op_2056_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2056_cast_fp16 = matmul(transpose_x = var_2056_transpose_x_1, transpose_y = var_2056_transpose_y_1, x = stacked_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_2056_cast_fp16")];
            fp16 var_2057_to_fp16 = const()[name = string("op_2057_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2058_cast_fp16 = mul(x = var_2056_cast_fp16, y = var_2057_to_fp16)[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_3_cast_fp16 = add(x = var_2058_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_3_cast_fp16")];
            int32 var_2062 = const()[name = string("op_2062"), val = int32(-1)];
            bool var_2063_interleave_0 = const()[name = string("op_2063_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_2063_cast_fp16 = concat(axis = var_2062, interleave = var_2063_interleave_0, values = (cache_weights_3_cast_fp16, call_weights_3_cast_fp16))[name = string("op_2063_cast_fp16")];
            int32 var_2064 = const()[name = string("op_2064"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_3_cast_fp16 = softmax(axis = var_2064, x = var_2063_cast_fp16)[name = string("shares_3_cast_fp16")];
            tensor<int32, [4]> var_2071_begin_0 = const()[name = string("op_2071_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2071_end_0 = const()[name = string("op_2071_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_2071_end_mask_0 = const()[name = string("op_2071_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_2071_cast_fp16 = slice_by_index(begin = var_2071_begin_0, end = var_2071_end_0, end_mask = var_2071_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2071_cast_fp16")];
            tensor<int32, [4]> var_2076_begin_0 = const()[name = string("op_2076_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2076_end_0 = const()[name = string("op_2076_end_0"), val = tensor<int32, [4]>([2, 8, 4112, 128])];
            tensor<bool, [4]> var_2076_end_mask_0 = const()[name = string("op_2076_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2076_cast_fp16 = slice_by_index(begin = var_2076_begin_0, end = var_2076_end_0, end_mask = var_2076_end_mask_0, x = read_state_1)[name = string("op_2076_cast_fp16")];
            tensor<int32, [4]> var_2086_begin_0 = const()[name = string("op_2086_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2086_end_0 = const()[name = string("op_2086_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2086_end_mask_0 = const()[name = string("op_2086_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2086_cast_fp16 = slice_by_index(begin = var_2086_begin_0, end = var_2086_end_0, end_mask = var_2086_end_mask_0, x = var_2076_cast_fp16)[name = string("op_2086_cast_fp16")];
            bool var_2092_transpose_x_0 = const()[name = string("op_2092_transpose_x_0"), val = bool(false)];
            bool var_2092_transpose_y_0 = const()[name = string("op_2092_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2092_cast_fp16 = matmul(transpose_x = var_2092_transpose_x_0, transpose_y = var_2092_transpose_y_0, x = var_2071_cast_fp16, y = var_2086_cast_fp16)[name = string("op_2092_cast_fp16")];
            tensor<int32, [4]> var_2097_begin_0 = const()[name = string("op_2097_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_2097_end_0 = const()[name = string("op_2097_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_2097_end_mask_0 = const()[name = string("op_2097_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2097_cast_fp16 = slice_by_index(begin = var_2097_begin_0, end = var_2097_end_0, end_mask = var_2097_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2097_cast_fp16")];
            bool var_2098_transpose_x_0 = const()[name = string("op_2098_transpose_x_0"), val = bool(false)];
            bool var_2098_transpose_y_0 = const()[name = string("op_2098_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_1953_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 32, 128]> var_2098_cast_fp16 = matmul(transpose_x = var_2098_transpose_x_0, transpose_y = var_2098_transpose_y_0, x = var_2097_cast_fp16, y = v_3_cast_fp16)[name = string("op_2098_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_5_cast_fp16 = add(x = var_2092_cast_fp16, y = var_2098_cast_fp16)[name = string("mixed_5_cast_fp16")];
            tensor<int32, [4]> var_2115_begin_0 = const()[name = string("op_2115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2115_end_0 = const()[name = string("op_2115_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2115_end_mask_0 = const()[name = string("op_2115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2115_cast_fp16 = slice_by_index(begin = var_2115_begin_0, end = var_2115_end_0, end_mask = var_2115_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<int32, [4]> var_2130_begin_0 = const()[name = string("op_2130_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2130_end_0 = const()[name = string("op_2130_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2130_end_mask_0 = const()[name = string("op_2130_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2130_cast_fp16 = slice_by_index(begin = var_2130_begin_0, end = var_2130_end_0, end_mask = var_2130_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2130_cast_fp16")];
            int32 var_2132 = const()[name = string("op_2132"), 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_2132, interleave = mixed_7_interleave_0, values = (var_2115_cast_fp16, var_2130_cast_fp16))[name = string("mixed_7_cast_fp16")];
            tensor<int32, [4]> var_2136_perm_0 = const()[name = string("op_2136_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2140 = const()[name = string("op_2140"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2136_cast_fp16 = transpose(perm = var_2136_perm_0, x = mixed_7_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 16, 2048]> input_13_cast_fp16 = reshape(shape = var_2140, x = var_2136_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_37_cast_fp16 = add(x = x_21_cast_fp16, y = linear_10_cast_fp16)[name = string("x_37_cast_fp16")];
            bool var_2150 = const()[name = string("op_2150"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2153_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2153_cast_fp16")];
            tensor<int32, [1]> peak_15_axes_0 = const()[name = string("peak_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_15_cast_fp16 = reduce_max(axes = peak_15_axes_0, keep_dims = var_2150, x = var_2153_cast_fp16)[name = string("peak_15_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = peak_15_cast_fp16)[name = string("x_39_cast_fp16")];
            fp16 var_2149_promoted_to_fp16 = const()[name = string("op_2149_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2157_cast_fp16 = pow(x = x_39_cast_fp16, y = var_2149_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, 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_2149_promoted_1_to_fp16 = const()[name = string("op_2149_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2160_cast_fp16 = pow(x = peak_15_cast_fp16, y = var_2149_promoted_1_to_fp16)[name = string("op_2160_cast_fp16")];
            fp32 var_2161_epsilon_0 = const()[name = string("op_2161_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2161_cast_fp16 = inverse(epsilon = var_2161_epsilon_0, x = var_2160_cast_fp16)[name = string("op_2161_cast_fp16")];
            fp16 var_2162_to_fp16 = const()[name = string("op_2162_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2163_cast_fp16 = mul(x = var_2161_cast_fp16, y = var_2162_to_fp16)[name = string("op_2163_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2164_cast_fp16 = add(x = var_2159_cast_fp16, y = var_2163_cast_fp16)[name = string("op_2164_cast_fp16")];
            fp32 var_2165_epsilon_0 = const()[name = string("op_2165_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2165_cast_fp16 = rsqrt(epsilon = var_2165_epsilon_0, x = var_2164_cast_fp16)[name = string("op_2165_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2166_cast_fp16 = mul(x = x_39_cast_fp16, y = var_2165_cast_fp16)[name = string("op_2166_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_2166_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_2171_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_2171_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_2171_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_41_cast_fp16 = add(x = x_37_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            bool var_2184 = const()[name = string("op_2184"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2187_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> peak_17_axes_0 = const()[name = string("peak_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_17_cast_fp16 = reduce_max(axes = peak_17_axes_0, keep_dims = var_2184, x = var_2187_cast_fp16)[name = string("peak_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = peak_17_cast_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_2183_promoted_to_fp16 = const()[name = string("op_2183_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2191_cast_fp16 = pow(x = x_43_cast_fp16, y = var_2183_promoted_to_fp16)[name = string("op_2191_cast_fp16")];
            tensor<int32, [1]> var_2193_axes_0 = const()[name = string("op_2193_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2193_keep_dims_0 = const()[name = string("op_2193_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2193_cast_fp16 = reduce_mean(axes = var_2193_axes_0, keep_dims = var_2193_keep_dims_0, x = var_2191_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp16 var_2183_promoted_1_to_fp16 = const()[name = string("op_2183_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2194_cast_fp16 = pow(x = peak_17_cast_fp16, y = var_2183_promoted_1_to_fp16)[name = string("op_2194_cast_fp16")];
            fp32 var_2195_epsilon_0 = const()[name = string("op_2195_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2195_cast_fp16 = inverse(epsilon = var_2195_epsilon_0, x = var_2194_cast_fp16)[name = string("op_2195_cast_fp16")];
            fp16 var_2196_to_fp16 = const()[name = string("op_2196_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2197_cast_fp16 = mul(x = var_2195_cast_fp16, y = var_2196_to_fp16)[name = string("op_2197_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2198_cast_fp16 = add(x = var_2193_cast_fp16, y = var_2197_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2200_cast_fp16 = mul(x = x_43_cast_fp16, y = var_2199_cast_fp16)[name = string("op_2200_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_2200_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_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_14_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_45_perm_0 = const()[name = string("x_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2217 = const()[name = string("op_2217"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_45_cast_fp16 = transpose(perm = x_45_perm_0, x = var_2210_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 16, 16, 128]> var_2220_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_2220_cast_fp16")];
            tensor<int32, [1]> peak_19_axes_0 = const()[name = string("peak_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_19_cast_fp16 = reduce_max(axes = peak_19_axes_0, keep_dims = var_2217, x = var_2220_cast_fp16)[name = string("peak_19_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = peak_19_cast_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2216_promoted_to_fp16 = const()[name = string("op_2216_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2224_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2216_promoted_to_fp16)[name = string("op_2224_cast_fp16")];
            tensor<int32, [1]> var_2226_axes_0 = const()[name = string("op_2226_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2226_keep_dims_0 = const()[name = string("op_2226_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2226_cast_fp16 = reduce_mean(axes = var_2226_axes_0, keep_dims = var_2226_keep_dims_0, x = var_2224_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp16 var_2216_promoted_1_to_fp16 = const()[name = string("op_2216_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2227_cast_fp16 = pow(x = peak_19_cast_fp16, y = var_2216_promoted_1_to_fp16)[name = string("op_2227_cast_fp16")];
            fp32 var_2228_epsilon_0 = const()[name = string("op_2228_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2228_cast_fp16 = inverse(epsilon = var_2228_epsilon_0, x = var_2227_cast_fp16)[name = string("op_2228_cast_fp16")];
            fp16 var_2229_to_fp16 = const()[name = string("op_2229_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2230_cast_fp16 = mul(x = var_2228_cast_fp16, y = var_2229_to_fp16)[name = string("op_2230_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2231_cast_fp16 = add(x = var_2226_cast_fp16, y = var_2230_cast_fp16)[name = string("op_2231_cast_fp16")];
            fp32 var_2232_epsilon_0 = const()[name = string("op_2232_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2232_cast_fp16 = rsqrt(epsilon = var_2232_epsilon_0, x = var_2231_cast_fp16)[name = string("op_2232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2233_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2232_cast_fp16)[name = string("op_2233_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_53_cast_fp16 = mul(x = var_2233_cast_fp16, y = model_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_53_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_2242 = const()[name = string("op_2242"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2243_cast_fp16 = reshape(shape = var_2242, x = linear_15_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [4]> x_49_perm_0 = const()[name = string("x_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2250 = const()[name = string("op_2250"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_49_cast_fp16 = transpose(perm = x_49_perm_0, x = var_2243_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 8, 16, 128]> var_2253_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2253_cast_fp16")];
            tensor<int32, [1]> peak_21_axes_0 = const()[name = string("peak_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_21_cast_fp16 = reduce_max(axes = peak_21_axes_0, keep_dims = var_2250, x = var_2253_cast_fp16)[name = string("peak_21_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = peak_21_cast_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_2249_promoted_to_fp16 = const()[name = string("op_2249_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2257_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2249_promoted_to_fp16)[name = string("op_2257_cast_fp16")];
            tensor<int32, [1]> var_2259_axes_0 = const()[name = string("op_2259_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2259_keep_dims_0 = const()[name = string("op_2259_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2259_cast_fp16 = reduce_mean(axes = var_2259_axes_0, keep_dims = var_2259_keep_dims_0, x = var_2257_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp16 var_2249_promoted_1_to_fp16 = const()[name = string("op_2249_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2260_cast_fp16 = pow(x = peak_21_cast_fp16, y = var_2249_promoted_1_to_fp16)[name = string("op_2260_cast_fp16")];
            fp32 var_2261_epsilon_0 = const()[name = string("op_2261_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2261_cast_fp16 = inverse(epsilon = var_2261_epsilon_0, x = var_2260_cast_fp16)[name = string("op_2261_cast_fp16")];
            fp16 var_2262_to_fp16 = const()[name = string("op_2262_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2263_cast_fp16 = mul(x = var_2261_cast_fp16, y = var_2262_to_fp16)[name = string("op_2263_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2264_cast_fp16 = add(x = var_2259_cast_fp16, y = var_2263_cast_fp16)[name = string("op_2264_cast_fp16")];
            fp32 var_2265_epsilon_0 = const()[name = string("op_2265_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2265_cast_fp16 = rsqrt(epsilon = var_2265_epsilon_0, x = var_2264_cast_fp16)[name = string("op_2265_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2266_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2265_cast_fp16)[name = string("op_2266_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_55_cast_fp16 = mul(x = var_2266_cast_fp16, y = model_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_55_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_2275 = const()[name = string("op_2275"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2276_cast_fp16 = reshape(shape = var_2275, x = linear_16_cast_fp16)[name = string("op_2276_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_53_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_53_cast_fp16)[name = string("second_9_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2290_cast_fp16 = mul(x = first_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2291_cast_fp16 = mul(x = second_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2293_cast_fp16 = sub(x = var_2290_cast_fp16, y = var_2291_cast_fp16)[name = string("op_2293_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2294_cast_fp16 = mul(x = second_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2295_cast_fp16 = mul(x = first_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2295_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2297_cast_fp16 = add(x = var_2294_cast_fp16, y = var_2295_cast_fp16)[name = string("op_2297_cast_fp16")];
            int32 var_2299 = const()[name = string("op_2299"), 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_2299, interleave = q_5_interleave_0, values = (var_2293_cast_fp16, var_2297_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_55_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_55_cast_fp16)[name = string("second_11_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2311_cast_fp16 = mul(x = first_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2312_cast_fp16 = mul(x = second_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2312_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2314_cast_fp16 = sub(x = var_2311_cast_fp16, y = var_2312_cast_fp16)[name = string("op_2314_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2315_cast_fp16 = mul(x = second_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2315_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2316_cast_fp16 = mul(x = first_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2316_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2318_cast_fp16 = add(x = var_2315_cast_fp16, y = var_2316_cast_fp16)[name = string("op_2318_cast_fp16")];
            int32 var_2320 = const()[name = string("op_2320"), 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_2320, interleave = k_5_interleave_0, values = (var_2314_cast_fp16, var_2318_cast_fp16))[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_2331_begin_0 = const()[name = string("op_2331_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2331_end_0 = const()[name = string("op_2331_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2331_end_mask_0 = const()[name = string("op_2331_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2331_stride_0 = const()[name = string("op_2331_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, stride = var_2331_stride_0, x = q_5_cast_fp16)[name = string("op_2331_cast_fp16")];
            tensor<int32, [4]> var_2341_begin_0 = const()[name = string("op_2341_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2341_end_0 = const()[name = string("op_2341_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2341_end_mask_0 = const()[name = string("op_2341_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2341_stride_0 = const()[name = string("op_2341_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2341_cast_fp16 = slice_by_index(begin = var_2341_begin_0, end = var_2341_end_0, end_mask = var_2341_end_mask_0, stride = var_2341_stride_0, x = q_5_cast_fp16)[name = string("op_2341_cast_fp16")];
            int32 var_2343 = const()[name = string("op_2343"), 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_2343, interleave = stacked_5_interleave_0, values = (var_2331_cast_fp16, var_2341_cast_fp16))[name = string("stacked_5_cast_fp16")];
            tensor<int32, [4]> var_2349_begin_0 = const()[name = string("op_2349_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2349_end_0 = const()[name = string("op_2349_end_0"), val = tensor<int32, [4]>([3, 8, 4112, 128])];
            tensor<bool, [4]> var_2349_end_mask_0 = const()[name = string("op_2349_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2349_cast_fp16 = slice_by_index(begin = var_2349_begin_0, end = var_2349_end_0, end_mask = var_2349_end_mask_0, x = read_state_0)[name = string("op_2349_cast_fp16")];
            tensor<int32, [4]> var_2359_begin_0 = const()[name = string("op_2359_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2359_end_0 = const()[name = string("op_2359_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2359_end_mask_0 = const()[name = string("op_2359_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2359_cast_fp16 = slice_by_index(begin = var_2359_begin_0, end = var_2359_end_0, end_mask = var_2359_end_mask_0, x = var_2349_cast_fp16)[name = string("op_2359_cast_fp16")];
            bool var_2368_transpose_x_1 = const()[name = string("op_2368_transpose_x_1"), val = bool(false)];
            bool var_2368_transpose_y_1 = const()[name = string("op_2368_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_2368_cast_fp16 = matmul(transpose_x = var_2368_transpose_x_1, transpose_y = var_2368_transpose_y_1, x = var_2359_cast_fp16, y = stacked_5_cast_fp16)[name = string("op_2368_cast_fp16")];
            tensor<int32, [4]> var_2371_perm_0 = const()[name = string("op_2371_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_2371_cast_fp16 = transpose(perm = var_2371_perm_0, x = var_2368_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 8, 32, 4096]> var_2373_cast_fp16 = mul(x = var_2371_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_5_cast_fp16 = add(x = var_2373_cast_fp16, y = cacheMask)[name = string("cache_weights_5_cast_fp16")];
            bool var_2379_transpose_x_1 = const()[name = string("op_2379_transpose_x_1"), val = bool(false)];
            bool var_2379_transpose_y_1 = const()[name = string("op_2379_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2379_cast_fp16 = matmul(transpose_x = var_2379_transpose_x_1, transpose_y = var_2379_transpose_y_1, x = stacked_5_cast_fp16, y = k_5_cast_fp16)[name = string("op_2379_cast_fp16")];
            fp16 var_2380_to_fp16 = const()[name = string("op_2380_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2381_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_2380_to_fp16)[name = string("op_2381_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_5_cast_fp16 = add(x = var_2381_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_5_cast_fp16")];
            int32 var_2385 = const()[name = string("op_2385"), val = int32(-1)];
            bool var_2386_interleave_0 = const()[name = string("op_2386_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_2386_cast_fp16 = concat(axis = var_2385, interleave = var_2386_interleave_0, values = (cache_weights_5_cast_fp16, call_weights_5_cast_fp16))[name = string("op_2386_cast_fp16")];
            int32 var_2387 = const()[name = string("op_2387"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_5_cast_fp16 = softmax(axis = var_2387, x = var_2386_cast_fp16)[name = string("shares_5_cast_fp16")];
            tensor<int32, [4]> var_2394_begin_0 = const()[name = string("op_2394_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2394_end_0 = const()[name = string("op_2394_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_2394_end_mask_0 = const()[name = string("op_2394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_2394_cast_fp16 = slice_by_index(begin = var_2394_begin_0, end = var_2394_end_0, end_mask = var_2394_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2394_cast_fp16")];
            tensor<int32, [4]> var_2399_begin_0 = const()[name = string("op_2399_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2399_end_0 = const()[name = string("op_2399_end_0"), val = tensor<int32, [4]>([3, 8, 4112, 128])];
            tensor<bool, [4]> var_2399_end_mask_0 = const()[name = string("op_2399_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2399_cast_fp16 = slice_by_index(begin = var_2399_begin_0, end = var_2399_end_0, end_mask = var_2399_end_mask_0, x = read_state_1)[name = string("op_2399_cast_fp16")];
            tensor<int32, [4]> var_2409_begin_0 = const()[name = string("op_2409_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2409_end_0 = const()[name = string("op_2409_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2409_end_mask_0 = const()[name = string("op_2409_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2409_cast_fp16 = slice_by_index(begin = var_2409_begin_0, end = var_2409_end_0, end_mask = var_2409_end_mask_0, x = var_2399_cast_fp16)[name = string("op_2409_cast_fp16")];
            bool var_2415_transpose_x_0 = const()[name = string("op_2415_transpose_x_0"), val = bool(false)];
            bool var_2415_transpose_y_0 = const()[name = string("op_2415_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2415_cast_fp16 = matmul(transpose_x = var_2415_transpose_x_0, transpose_y = var_2415_transpose_y_0, x = var_2394_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2415_cast_fp16")];
            tensor<int32, [4]> var_2420_begin_0 = const()[name = string("op_2420_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_2420_end_0 = const()[name = string("op_2420_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_2420_end_mask_0 = const()[name = string("op_2420_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2420_cast_fp16 = slice_by_index(begin = var_2420_begin_0, end = var_2420_end_0, end_mask = var_2420_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2420_cast_fp16")];
            bool var_2421_transpose_x_0 = const()[name = string("op_2421_transpose_x_0"), val = bool(false)];
            bool var_2421_transpose_y_0 = const()[name = string("op_2421_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_2276_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 8, 32, 128]> var_2421_cast_fp16 = matmul(transpose_x = var_2421_transpose_x_0, transpose_y = var_2421_transpose_y_0, x = var_2420_cast_fp16, y = v_5_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_9_cast_fp16 = add(x = var_2415_cast_fp16, y = var_2421_cast_fp16)[name = string("mixed_9_cast_fp16")];
            tensor<int32, [4]> var_2438_begin_0 = const()[name = string("op_2438_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2438_end_0 = const()[name = string("op_2438_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2438_end_mask_0 = const()[name = string("op_2438_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2438_cast_fp16 = slice_by_index(begin = var_2438_begin_0, end = var_2438_end_0, end_mask = var_2438_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2438_cast_fp16")];
            tensor<int32, [4]> var_2453_begin_0 = const()[name = string("op_2453_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2453_end_0 = const()[name = string("op_2453_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2453_end_mask_0 = const()[name = string("op_2453_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2453_cast_fp16 = slice_by_index(begin = var_2453_begin_0, end = var_2453_end_0, end_mask = var_2453_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2453_cast_fp16")];
            int32 var_2455 = const()[name = string("op_2455"), 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_2455, interleave = mixed_11_interleave_0, values = (var_2438_cast_fp16, var_2453_cast_fp16))[name = string("mixed_11_cast_fp16")];
            tensor<int32, [4]> var_2459_perm_0 = const()[name = string("op_2459_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2463 = const()[name = string("op_2463"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2459_cast_fp16 = transpose(perm = var_2459_perm_0, x = mixed_11_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 16, 2048]> input_23_cast_fp16 = reshape(shape = var_2463, x = var_2459_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_57_cast_fp16 = add(x = x_41_cast_fp16, y = linear_17_cast_fp16)[name = string("x_57_cast_fp16")];
            bool var_2473 = const()[name = string("op_2473"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2476_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [1]> peak_23_axes_0 = const()[name = string("peak_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_23_cast_fp16 = reduce_max(axes = peak_23_axes_0, keep_dims = var_2473, x = var_2476_cast_fp16)[name = string("peak_23_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = peak_23_cast_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2472_promoted_to_fp16 = const()[name = string("op_2472_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2480_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2472_promoted_to_fp16)[name = string("op_2480_cast_fp16")];
            tensor<int32, [1]> var_2482_axes_0 = const()[name = string("op_2482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2482_keep_dims_0 = const()[name = string("op_2482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2482_cast_fp16 = reduce_mean(axes = var_2482_axes_0, keep_dims = var_2482_keep_dims_0, x = var_2480_cast_fp16)[name = string("op_2482_cast_fp16")];
            fp16 var_2472_promoted_1_to_fp16 = const()[name = string("op_2472_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2483_cast_fp16 = pow(x = peak_23_cast_fp16, y = var_2472_promoted_1_to_fp16)[name = string("op_2483_cast_fp16")];
            fp32 var_2484_epsilon_0 = const()[name = string("op_2484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2484_cast_fp16 = inverse(epsilon = var_2484_epsilon_0, x = var_2483_cast_fp16)[name = string("op_2484_cast_fp16")];
            fp16 var_2485_to_fp16 = const()[name = string("op_2485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2486_cast_fp16 = mul(x = var_2484_cast_fp16, y = var_2485_to_fp16)[name = string("op_2486_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2487_cast_fp16 = add(x = var_2482_cast_fp16, y = var_2486_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp32 var_2488_epsilon_0 = const()[name = string("op_2488_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2488_cast_fp16 = rsqrt(epsilon = var_2488_epsilon_0, x = var_2487_cast_fp16)[name = string("op_2488_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2489_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2488_cast_fp16)[name = string("op_2489_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_2489_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_2494_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_2494_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_2494_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_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_20_cast_fp16)[name = string("x_61_cast_fp16")];
            bool var_2507 = const()[name = string("op_2507"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2510_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<int32, [1]> peak_25_axes_0 = const()[name = string("peak_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_25_cast_fp16 = reduce_max(axes = peak_25_axes_0, keep_dims = var_2507, x = var_2510_cast_fp16)[name = string("peak_25_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = peak_25_cast_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_2506_promoted_to_fp16 = const()[name = string("op_2506_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2514_cast_fp16 = pow(x = x_63_cast_fp16, y = var_2506_promoted_to_fp16)[name = string("op_2514_cast_fp16")];
            tensor<int32, [1]> var_2516_axes_0 = const()[name = string("op_2516_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2516_keep_dims_0 = const()[name = string("op_2516_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2516_cast_fp16 = reduce_mean(axes = var_2516_axes_0, keep_dims = var_2516_keep_dims_0, x = var_2514_cast_fp16)[name = string("op_2516_cast_fp16")];
            fp16 var_2506_promoted_1_to_fp16 = const()[name = string("op_2506_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2517_cast_fp16 = pow(x = peak_25_cast_fp16, y = var_2506_promoted_1_to_fp16)[name = string("op_2517_cast_fp16")];
            fp32 var_2518_epsilon_0 = const()[name = string("op_2518_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2518_cast_fp16 = inverse(epsilon = var_2518_epsilon_0, x = var_2517_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 var_2519_to_fp16 = const()[name = string("op_2519_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2520_cast_fp16 = mul(x = var_2518_cast_fp16, y = var_2519_to_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2521_cast_fp16 = add(x = var_2516_cast_fp16, y = var_2520_cast_fp16)[name = string("op_2521_cast_fp16")];
            fp32 var_2522_epsilon_0 = const()[name = string("op_2522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2522_cast_fp16 = rsqrt(epsilon = var_2522_epsilon_0, x = var_2521_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2523_cast_fp16 = mul(x = x_63_cast_fp16, y = var_2522_cast_fp16)[name = string("op_2523_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_2523_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_2532 = const()[name = string("op_2532"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2533_cast_fp16 = reshape(shape = var_2532, x = linear_21_cast_fp16)[name = string("op_2533_cast_fp16")];
            tensor<int32, [4]> x_65_perm_0 = const()[name = string("x_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2540 = const()[name = string("op_2540"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_65_cast_fp16 = transpose(perm = x_65_perm_0, x = var_2533_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 16, 128]> var_2543_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2543_cast_fp16")];
            tensor<int32, [1]> peak_27_axes_0 = const()[name = string("peak_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_27_cast_fp16 = reduce_max(axes = peak_27_axes_0, keep_dims = var_2540, x = var_2543_cast_fp16)[name = string("peak_27_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = peak_27_cast_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_2539_promoted_to_fp16 = const()[name = string("op_2539_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2547_cast_fp16 = pow(x = x_67_cast_fp16, y = var_2539_promoted_to_fp16)[name = string("op_2547_cast_fp16")];
            tensor<int32, [1]> var_2549_axes_0 = const()[name = string("op_2549_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2549_keep_dims_0 = const()[name = string("op_2549_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2549_cast_fp16 = reduce_mean(axes = var_2549_axes_0, keep_dims = var_2549_keep_dims_0, x = var_2547_cast_fp16)[name = string("op_2549_cast_fp16")];
            fp16 var_2539_promoted_1_to_fp16 = const()[name = string("op_2539_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2550_cast_fp16 = pow(x = peak_27_cast_fp16, y = var_2539_promoted_1_to_fp16)[name = string("op_2550_cast_fp16")];
            fp32 var_2551_epsilon_0 = const()[name = string("op_2551_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2551_cast_fp16 = inverse(epsilon = var_2551_epsilon_0, x = var_2550_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 var_2552_to_fp16 = const()[name = string("op_2552_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2553_cast_fp16 = mul(x = var_2551_cast_fp16, y = var_2552_to_fp16)[name = string("op_2553_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2554_cast_fp16 = add(x = var_2549_cast_fp16, y = var_2553_cast_fp16)[name = string("op_2554_cast_fp16")];
            fp32 var_2555_epsilon_0 = const()[name = string("op_2555_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2555_cast_fp16 = rsqrt(epsilon = var_2555_epsilon_0, x = var_2554_cast_fp16)[name = string("op_2555_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2556_cast_fp16 = mul(x = x_67_cast_fp16, y = var_2555_cast_fp16)[name = string("op_2556_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_73_cast_fp16 = mul(x = var_2556_cast_fp16, y = model_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_73_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_2565 = const()[name = string("op_2565"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2566_cast_fp16 = reshape(shape = var_2565, x = linear_22_cast_fp16)[name = string("op_2566_cast_fp16")];
            tensor<int32, [4]> x_69_perm_0 = const()[name = string("x_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2573 = const()[name = string("op_2573"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_69_cast_fp16 = transpose(perm = x_69_perm_0, x = var_2566_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 16, 128]> var_2576_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> peak_29_axes_0 = const()[name = string("peak_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_29_cast_fp16 = reduce_max(axes = peak_29_axes_0, keep_dims = var_2573, x = var_2576_cast_fp16)[name = string("peak_29_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = peak_29_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_2572_promoted_to_fp16 = const()[name = string("op_2572_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2580_cast_fp16 = pow(x = x_71_cast_fp16, y = var_2572_promoted_to_fp16)[name = string("op_2580_cast_fp16")];
            tensor<int32, [1]> var_2582_axes_0 = const()[name = string("op_2582_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2582_keep_dims_0 = const()[name = string("op_2582_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2582_cast_fp16 = reduce_mean(axes = var_2582_axes_0, keep_dims = var_2582_keep_dims_0, x = var_2580_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp16 var_2572_promoted_1_to_fp16 = const()[name = string("op_2572_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2583_cast_fp16 = pow(x = peak_29_cast_fp16, y = var_2572_promoted_1_to_fp16)[name = string("op_2583_cast_fp16")];
            fp32 var_2584_epsilon_0 = const()[name = string("op_2584_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2584_cast_fp16 = inverse(epsilon = var_2584_epsilon_0, x = var_2583_cast_fp16)[name = string("op_2584_cast_fp16")];
            fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2586_cast_fp16 = mul(x = var_2584_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2587_cast_fp16 = add(x = var_2582_cast_fp16, y = var_2586_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2589_cast_fp16 = mul(x = x_71_cast_fp16, y = var_2588_cast_fp16)[name = string("op_2589_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_75_cast_fp16 = mul(x = var_2589_cast_fp16, y = model_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_75_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_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2599_cast_fp16 = reshape(shape = var_2598, x = linear_23_cast_fp16)[name = string("op_2599_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_73_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_73_cast_fp16)[name = string("second_13_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2613_cast_fp16 = mul(x = first_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2613_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2614_cast_fp16 = mul(x = second_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2614_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2616_cast_fp16 = sub(x = var_2613_cast_fp16, y = var_2614_cast_fp16)[name = string("op_2616_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2617_cast_fp16 = mul(x = second_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2617_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2618_cast_fp16 = mul(x = first_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2618_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2620_cast_fp16 = add(x = var_2617_cast_fp16, y = var_2618_cast_fp16)[name = string("op_2620_cast_fp16")];
            int32 var_2622 = const()[name = string("op_2622"), 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_2622, interleave = q_7_interleave_0, values = (var_2616_cast_fp16, var_2620_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_75_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_75_cast_fp16)[name = string("second_15_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2634_cast_fp16 = mul(x = first_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2635_cast_fp16 = mul(x = second_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2635_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2637_cast_fp16 = sub(x = var_2634_cast_fp16, y = var_2635_cast_fp16)[name = string("op_2637_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2638_cast_fp16 = mul(x = second_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2638_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2639_cast_fp16 = mul(x = first_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2641_cast_fp16 = add(x = var_2638_cast_fp16, y = var_2639_cast_fp16)[name = string("op_2641_cast_fp16")];
            int32 var_2643 = const()[name = string("op_2643"), 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_2643, interleave = k_7_interleave_0, values = (var_2637_cast_fp16, var_2641_cast_fp16))[name = string("k_7_cast_fp16")];
            tensor<int32, [4]> var_2654_begin_0 = const()[name = string("op_2654_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2654_end_0 = const()[name = string("op_2654_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2654_end_mask_0 = const()[name = string("op_2654_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2654_stride_0 = const()[name = string("op_2654_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2654_cast_fp16 = slice_by_index(begin = var_2654_begin_0, end = var_2654_end_0, end_mask = var_2654_end_mask_0, stride = var_2654_stride_0, x = q_7_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2664_begin_0 = const()[name = string("op_2664_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2664_end_0 = const()[name = string("op_2664_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2664_end_mask_0 = const()[name = string("op_2664_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2664_stride_0 = const()[name = string("op_2664_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2664_cast_fp16 = slice_by_index(begin = var_2664_begin_0, end = var_2664_end_0, end_mask = var_2664_end_mask_0, stride = var_2664_stride_0, x = q_7_cast_fp16)[name = string("op_2664_cast_fp16")];
            int32 var_2666 = const()[name = string("op_2666"), 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_2666, interleave = stacked_7_interleave_0, values = (var_2654_cast_fp16, var_2664_cast_fp16))[name = string("stacked_7_cast_fp16")];
            tensor<int32, [4]> var_2672_begin_0 = const()[name = string("op_2672_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2672_end_0 = const()[name = string("op_2672_end_0"), val = tensor<int32, [4]>([4, 8, 4112, 128])];
            tensor<bool, [4]> var_2672_end_mask_0 = const()[name = string("op_2672_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2672_cast_fp16 = slice_by_index(begin = var_2672_begin_0, end = var_2672_end_0, end_mask = var_2672_end_mask_0, x = read_state_0)[name = string("op_2672_cast_fp16")];
            tensor<int32, [4]> var_2682_begin_0 = const()[name = string("op_2682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2682_end_0 = const()[name = string("op_2682_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2682_end_mask_0 = const()[name = string("op_2682_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2682_cast_fp16 = slice_by_index(begin = var_2682_begin_0, end = var_2682_end_0, end_mask = var_2682_end_mask_0, x = var_2672_cast_fp16)[name = string("op_2682_cast_fp16")];
            bool var_2691_transpose_x_1 = const()[name = string("op_2691_transpose_x_1"), val = bool(false)];
            bool var_2691_transpose_y_1 = const()[name = string("op_2691_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_2691_cast_fp16 = matmul(transpose_x = var_2691_transpose_x_1, transpose_y = var_2691_transpose_y_1, x = var_2682_cast_fp16, y = stacked_7_cast_fp16)[name = string("op_2691_cast_fp16")];
            tensor<int32, [4]> var_2694_perm_0 = const()[name = string("op_2694_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2695_to_fp16 = const()[name = string("op_2695_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_2694_cast_fp16 = transpose(perm = var_2694_perm_0, x = var_2691_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 32, 4096]> var_2696_cast_fp16 = mul(x = var_2694_cast_fp16, y = var_2695_to_fp16)[name = string("op_2696_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_7_cast_fp16 = add(x = var_2696_cast_fp16, y = cacheMask)[name = string("cache_weights_7_cast_fp16")];
            bool var_2702_transpose_x_1 = const()[name = string("op_2702_transpose_x_1"), val = bool(false)];
            bool var_2702_transpose_y_1 = const()[name = string("op_2702_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2702_cast_fp16 = matmul(transpose_x = var_2702_transpose_x_1, transpose_y = var_2702_transpose_y_1, x = stacked_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 var_2703_to_fp16 = const()[name = string("op_2703_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2704_cast_fp16 = mul(x = var_2702_cast_fp16, y = var_2703_to_fp16)[name = string("op_2704_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_7_cast_fp16 = add(x = var_2704_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_7_cast_fp16")];
            int32 var_2708 = const()[name = string("op_2708"), val = int32(-1)];
            bool var_2709_interleave_0 = const()[name = string("op_2709_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_2709_cast_fp16 = concat(axis = var_2708, interleave = var_2709_interleave_0, values = (cache_weights_7_cast_fp16, call_weights_7_cast_fp16))[name = string("op_2709_cast_fp16")];
            int32 var_2710 = const()[name = string("op_2710"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_7_cast_fp16 = softmax(axis = var_2710, x = var_2709_cast_fp16)[name = string("shares_7_cast_fp16")];
            tensor<int32, [4]> var_2717_begin_0 = const()[name = string("op_2717_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2717_end_0 = const()[name = string("op_2717_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_2717_end_mask_0 = const()[name = string("op_2717_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_2717_cast_fp16 = slice_by_index(begin = var_2717_begin_0, end = var_2717_end_0, end_mask = var_2717_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2717_cast_fp16")];
            tensor<int32, [4]> var_2722_begin_0 = const()[name = string("op_2722_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2722_end_0 = const()[name = string("op_2722_end_0"), val = tensor<int32, [4]>([4, 8, 4112, 128])];
            tensor<bool, [4]> var_2722_end_mask_0 = const()[name = string("op_2722_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2722_cast_fp16 = slice_by_index(begin = var_2722_begin_0, end = var_2722_end_0, end_mask = var_2722_end_mask_0, x = read_state_1)[name = string("op_2722_cast_fp16")];
            tensor<int32, [4]> var_2732_begin_0 = const()[name = string("op_2732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2732_end_0 = const()[name = string("op_2732_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2732_end_mask_0 = const()[name = string("op_2732_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2732_cast_fp16 = slice_by_index(begin = var_2732_begin_0, end = var_2732_end_0, end_mask = var_2732_end_mask_0, x = var_2722_cast_fp16)[name = string("op_2732_cast_fp16")];
            bool var_2738_transpose_x_0 = const()[name = string("op_2738_transpose_x_0"), val = bool(false)];
            bool var_2738_transpose_y_0 = const()[name = string("op_2738_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2738_cast_fp16 = matmul(transpose_x = var_2738_transpose_x_0, transpose_y = var_2738_transpose_y_0, x = var_2717_cast_fp16, y = var_2732_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> var_2743_begin_0 = const()[name = string("op_2743_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_2743_end_0 = const()[name = string("op_2743_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_2743_end_mask_0 = const()[name = string("op_2743_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2743_cast_fp16 = slice_by_index(begin = var_2743_begin_0, end = var_2743_end_0, end_mask = var_2743_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2743_cast_fp16")];
            bool var_2744_transpose_x_0 = const()[name = string("op_2744_transpose_x_0"), val = bool(false)];
            bool var_2744_transpose_y_0 = const()[name = string("op_2744_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_2599_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 8, 32, 128]> var_2744_cast_fp16 = matmul(transpose_x = var_2744_transpose_x_0, transpose_y = var_2744_transpose_y_0, x = var_2743_cast_fp16, y = v_7_cast_fp16)[name = string("op_2744_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_13_cast_fp16 = add(x = var_2738_cast_fp16, y = var_2744_cast_fp16)[name = string("mixed_13_cast_fp16")];
            tensor<int32, [4]> var_2761_begin_0 = const()[name = string("op_2761_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2761_end_0 = const()[name = string("op_2761_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2761_end_mask_0 = const()[name = string("op_2761_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2761_cast_fp16 = slice_by_index(begin = var_2761_begin_0, end = var_2761_end_0, end_mask = var_2761_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2761_cast_fp16")];
            tensor<int32, [4]> var_2776_begin_0 = const()[name = string("op_2776_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2776_end_0 = const()[name = string("op_2776_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2776_end_mask_0 = const()[name = string("op_2776_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2776_cast_fp16 = slice_by_index(begin = var_2776_begin_0, end = var_2776_end_0, end_mask = var_2776_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2776_cast_fp16")];
            int32 var_2778 = const()[name = string("op_2778"), 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_2778, interleave = mixed_15_interleave_0, values = (var_2761_cast_fp16, var_2776_cast_fp16))[name = string("mixed_15_cast_fp16")];
            tensor<int32, [4]> var_2782_perm_0 = const()[name = string("op_2782_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2786 = const()[name = string("op_2786"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2782_cast_fp16 = transpose(perm = var_2782_perm_0, x = mixed_15_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 16, 2048]> input_33_cast_fp16 = reshape(shape = var_2786, x = var_2782_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_77_cast_fp16 = add(x = x_61_cast_fp16, y = linear_24_cast_fp16)[name = string("x_77_cast_fp16")];
            bool var_2796 = const()[name = string("op_2796"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2799_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_2799_cast_fp16")];
            tensor<int32, [1]> peak_31_axes_0 = const()[name = string("peak_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_31_cast_fp16 = reduce_max(axes = peak_31_axes_0, keep_dims = var_2796, x = var_2799_cast_fp16)[name = string("peak_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = peak_31_cast_fp16)[name = string("x_79_cast_fp16")];
            fp16 var_2795_promoted_to_fp16 = const()[name = string("op_2795_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2803_cast_fp16 = pow(x = x_79_cast_fp16, y = var_2795_promoted_to_fp16)[name = string("op_2803_cast_fp16")];
            tensor<int32, [1]> var_2805_axes_0 = const()[name = string("op_2805_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2805_keep_dims_0 = const()[name = string("op_2805_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2805_cast_fp16 = reduce_mean(axes = var_2805_axes_0, keep_dims = var_2805_keep_dims_0, x = var_2803_cast_fp16)[name = string("op_2805_cast_fp16")];
            fp16 var_2795_promoted_1_to_fp16 = const()[name = string("op_2795_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2806_cast_fp16 = pow(x = peak_31_cast_fp16, y = var_2795_promoted_1_to_fp16)[name = string("op_2806_cast_fp16")];
            fp32 var_2807_epsilon_0 = const()[name = string("op_2807_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2807_cast_fp16 = inverse(epsilon = var_2807_epsilon_0, x = var_2806_cast_fp16)[name = string("op_2807_cast_fp16")];
            fp16 var_2808_to_fp16 = const()[name = string("op_2808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2809_cast_fp16 = mul(x = var_2807_cast_fp16, y = var_2808_to_fp16)[name = string("op_2809_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2810_cast_fp16 = add(x = var_2805_cast_fp16, y = var_2809_cast_fp16)[name = string("op_2810_cast_fp16")];
            fp32 var_2811_epsilon_0 = const()[name = string("op_2811_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2811_cast_fp16 = rsqrt(epsilon = var_2811_epsilon_0, x = var_2810_cast_fp16)[name = string("op_2811_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2812_cast_fp16 = mul(x = x_79_cast_fp16, y = var_2811_cast_fp16)[name = string("op_2812_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_2812_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_2817_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_2817_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_2817_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_81_cast_fp16 = add(x = x_77_cast_fp16, y = linear_27_cast_fp16)[name = string("x_81_cast_fp16")];
            bool var_2830 = const()[name = string("op_2830"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2833_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<int32, [1]> peak_33_axes_0 = const()[name = string("peak_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_33_cast_fp16 = reduce_max(axes = peak_33_axes_0, keep_dims = var_2830, x = var_2833_cast_fp16)[name = string("peak_33_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = peak_33_cast_fp16)[name = string("x_83_cast_fp16")];
            fp16 var_2829_promoted_to_fp16 = const()[name = string("op_2829_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2837_cast_fp16 = pow(x = x_83_cast_fp16, y = var_2829_promoted_to_fp16)[name = string("op_2837_cast_fp16")];
            tensor<int32, [1]> var_2839_axes_0 = const()[name = string("op_2839_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2839_keep_dims_0 = const()[name = string("op_2839_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2839_cast_fp16 = reduce_mean(axes = var_2839_axes_0, keep_dims = var_2839_keep_dims_0, x = var_2837_cast_fp16)[name = string("op_2839_cast_fp16")];
            fp16 var_2829_promoted_1_to_fp16 = const()[name = string("op_2829_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2840_cast_fp16 = pow(x = peak_33_cast_fp16, y = var_2829_promoted_1_to_fp16)[name = string("op_2840_cast_fp16")];
            fp32 var_2841_epsilon_0 = const()[name = string("op_2841_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2841_cast_fp16 = inverse(epsilon = var_2841_epsilon_0, x = var_2840_cast_fp16)[name = string("op_2841_cast_fp16")];
            fp16 var_2842_to_fp16 = const()[name = string("op_2842_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2843_cast_fp16 = mul(x = var_2841_cast_fp16, y = var_2842_to_fp16)[name = string("op_2843_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2844_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2843_cast_fp16)[name = string("op_2844_cast_fp16")];
            fp32 var_2845_epsilon_0 = const()[name = string("op_2845_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2845_cast_fp16 = rsqrt(epsilon = var_2845_epsilon_0, x = var_2844_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2846_cast_fp16 = mul(x = x_83_cast_fp16, y = var_2845_cast_fp16)[name = string("op_2846_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_2846_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_2855 = const()[name = string("op_2855"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2856_cast_fp16 = reshape(shape = var_2855, x = linear_28_cast_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [4]> x_85_perm_0 = const()[name = string("x_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2863 = const()[name = string("op_2863"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_85_cast_fp16 = transpose(perm = x_85_perm_0, x = var_2856_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 16, 16, 128]> var_2866_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_2866_cast_fp16")];
            tensor<int32, [1]> peak_35_axes_0 = const()[name = string("peak_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_35_cast_fp16 = reduce_max(axes = peak_35_axes_0, keep_dims = var_2863, x = var_2866_cast_fp16)[name = string("peak_35_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = peak_35_cast_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_2862_promoted_to_fp16 = const()[name = string("op_2862_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2870_cast_fp16 = pow(x = x_87_cast_fp16, y = var_2862_promoted_to_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [1]> var_2872_axes_0 = const()[name = string("op_2872_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2872_keep_dims_0 = const()[name = string("op_2872_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2872_cast_fp16 = reduce_mean(axes = var_2872_axes_0, keep_dims = var_2872_keep_dims_0, x = var_2870_cast_fp16)[name = string("op_2872_cast_fp16")];
            fp16 var_2862_promoted_1_to_fp16 = const()[name = string("op_2862_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2873_cast_fp16 = pow(x = peak_35_cast_fp16, y = var_2862_promoted_1_to_fp16)[name = string("op_2873_cast_fp16")];
            fp32 var_2874_epsilon_0 = const()[name = string("op_2874_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2874_cast_fp16 = inverse(epsilon = var_2874_epsilon_0, x = var_2873_cast_fp16)[name = string("op_2874_cast_fp16")];
            fp16 var_2875_to_fp16 = const()[name = string("op_2875_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2876_cast_fp16 = mul(x = var_2874_cast_fp16, y = var_2875_to_fp16)[name = string("op_2876_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2877_cast_fp16 = add(x = var_2872_cast_fp16, y = var_2876_cast_fp16)[name = string("op_2877_cast_fp16")];
            fp32 var_2878_epsilon_0 = const()[name = string("op_2878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2878_cast_fp16 = rsqrt(epsilon = var_2878_epsilon_0, x = var_2877_cast_fp16)[name = string("op_2878_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2879_cast_fp16 = mul(x = x_87_cast_fp16, y = var_2878_cast_fp16)[name = string("op_2879_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_93_cast_fp16 = mul(x = var_2879_cast_fp16, y = model_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_93_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_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_29_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> x_89_perm_0 = const()[name = string("x_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2896 = const()[name = string("op_2896"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = var_2889_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 8, 16, 128]> var_2899_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<int32, [1]> peak_37_axes_0 = const()[name = string("peak_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_37_cast_fp16 = reduce_max(axes = peak_37_axes_0, keep_dims = var_2896, x = var_2899_cast_fp16)[name = string("peak_37_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = peak_37_cast_fp16)[name = string("x_91_cast_fp16")];
            fp16 var_2895_promoted_to_fp16 = const()[name = string("op_2895_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2903_cast_fp16 = pow(x = x_91_cast_fp16, y = var_2895_promoted_to_fp16)[name = string("op_2903_cast_fp16")];
            tensor<int32, [1]> var_2905_axes_0 = const()[name = string("op_2905_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2905_keep_dims_0 = const()[name = string("op_2905_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2905_cast_fp16 = reduce_mean(axes = var_2905_axes_0, keep_dims = var_2905_keep_dims_0, x = var_2903_cast_fp16)[name = string("op_2905_cast_fp16")];
            fp16 var_2895_promoted_1_to_fp16 = const()[name = string("op_2895_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2906_cast_fp16 = pow(x = peak_37_cast_fp16, y = var_2895_promoted_1_to_fp16)[name = string("op_2906_cast_fp16")];
            fp32 var_2907_epsilon_0 = const()[name = string("op_2907_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2907_cast_fp16 = inverse(epsilon = var_2907_epsilon_0, x = var_2906_cast_fp16)[name = string("op_2907_cast_fp16")];
            fp16 var_2908_to_fp16 = const()[name = string("op_2908_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2909_cast_fp16 = mul(x = var_2907_cast_fp16, y = var_2908_to_fp16)[name = string("op_2909_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2910_cast_fp16 = add(x = var_2905_cast_fp16, y = var_2909_cast_fp16)[name = string("op_2910_cast_fp16")];
            fp32 var_2911_epsilon_0 = const()[name = string("op_2911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2911_cast_fp16 = rsqrt(epsilon = var_2911_epsilon_0, x = var_2910_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2912_cast_fp16 = mul(x = x_91_cast_fp16, y = var_2911_cast_fp16)[name = string("op_2912_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_95_cast_fp16 = mul(x = var_2912_cast_fp16, y = model_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_95_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_2921 = const()[name = string("op_2921"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2922_cast_fp16 = reshape(shape = var_2921, x = linear_30_cast_fp16)[name = string("op_2922_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_93_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_93_cast_fp16)[name = string("second_17_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2936_cast_fp16 = mul(x = first_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2937_cast_fp16 = mul(x = second_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2937_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2939_cast_fp16 = sub(x = var_2936_cast_fp16, y = var_2937_cast_fp16)[name = string("op_2939_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2940_cast_fp16 = mul(x = second_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2940_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2941_cast_fp16 = mul(x = first_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2943_cast_fp16 = add(x = var_2940_cast_fp16, y = var_2941_cast_fp16)[name = string("op_2943_cast_fp16")];
            int32 var_2945 = const()[name = string("op_2945"), 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_2945, interleave = q_9_interleave_0, values = (var_2939_cast_fp16, var_2943_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_95_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_95_cast_fp16)[name = string("second_19_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2957_cast_fp16 = mul(x = first_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2958_cast_fp16 = mul(x = second_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2960_cast_fp16 = sub(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = string("op_2960_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2961_cast_fp16 = mul(x = second_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2962_cast_fp16 = mul(x = first_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2962_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2964_cast_fp16 = add(x = var_2961_cast_fp16, y = var_2962_cast_fp16)[name = string("op_2964_cast_fp16")];
            int32 var_2966 = const()[name = string("op_2966"), 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_2966, interleave = k_9_interleave_0, values = (var_2960_cast_fp16, var_2964_cast_fp16))[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2977_stride_0 = const()[name = string("op_2977_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, stride = var_2977_stride_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            tensor<int32, [4]> var_2987_begin_0 = const()[name = string("op_2987_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2987_end_0 = const()[name = string("op_2987_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2987_end_mask_0 = const()[name = string("op_2987_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2987_stride_0 = const()[name = string("op_2987_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2987_cast_fp16 = slice_by_index(begin = var_2987_begin_0, end = var_2987_end_0, end_mask = var_2987_end_mask_0, stride = var_2987_stride_0, x = q_9_cast_fp16)[name = string("op_2987_cast_fp16")];
            int32 var_2989 = const()[name = string("op_2989"), 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_2989, interleave = stacked_9_interleave_0, values = (var_2977_cast_fp16, var_2987_cast_fp16))[name = string("stacked_9_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([5, 8, 4112, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = read_state_0)[name = string("op_2995_cast_fp16")];
            tensor<int32, [4]> var_3005_begin_0 = const()[name = string("op_3005_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3005_end_0 = const()[name = string("op_3005_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3005_end_mask_0 = const()[name = string("op_3005_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3005_cast_fp16 = slice_by_index(begin = var_3005_begin_0, end = var_3005_end_0, end_mask = var_3005_end_mask_0, x = var_2995_cast_fp16)[name = string("op_3005_cast_fp16")];
            bool var_3014_transpose_x_1 = const()[name = string("op_3014_transpose_x_1"), val = bool(false)];
            bool var_3014_transpose_y_1 = const()[name = string("op_3014_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_3014_cast_fp16 = matmul(transpose_x = var_3014_transpose_x_1, transpose_y = var_3014_transpose_y_1, x = var_3005_cast_fp16, y = stacked_9_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<int32, [4]> var_3017_perm_0 = const()[name = string("op_3017_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_3017_cast_fp16 = transpose(perm = var_3017_perm_0, x = var_3014_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 8, 32, 4096]> var_3019_cast_fp16 = mul(x = var_3017_cast_fp16, y = var_3018_to_fp16)[name = string("op_3019_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_9_cast_fp16 = add(x = var_3019_cast_fp16, y = cacheMask)[name = string("cache_weights_9_cast_fp16")];
            bool var_3025_transpose_x_1 = const()[name = string("op_3025_transpose_x_1"), val = bool(false)];
            bool var_3025_transpose_y_1 = const()[name = string("op_3025_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3025_cast_fp16 = matmul(transpose_x = var_3025_transpose_x_1, transpose_y = var_3025_transpose_y_1, x = stacked_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_3025_cast_fp16")];
            fp16 var_3026_to_fp16 = const()[name = string("op_3026_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3027_cast_fp16 = mul(x = var_3025_cast_fp16, y = var_3026_to_fp16)[name = string("op_3027_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_9_cast_fp16 = add(x = var_3027_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_9_cast_fp16")];
            int32 var_3031 = const()[name = string("op_3031"), val = int32(-1)];
            bool var_3032_interleave_0 = const()[name = string("op_3032_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_3032_cast_fp16 = concat(axis = var_3031, interleave = var_3032_interleave_0, values = (cache_weights_9_cast_fp16, call_weights_9_cast_fp16))[name = string("op_3032_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_9_cast_fp16 = softmax(axis = var_3033, x = var_3032_cast_fp16)[name = string("shares_9_cast_fp16")];
            tensor<int32, [4]> var_3040_begin_0 = const()[name = string("op_3040_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3040_end_0 = const()[name = string("op_3040_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_3040_end_mask_0 = const()[name = string("op_3040_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_3040_cast_fp16 = slice_by_index(begin = var_3040_begin_0, end = var_3040_end_0, end_mask = var_3040_end_mask_0, x = shares_9_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> var_3045_begin_0 = const()[name = string("op_3045_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_3045_end_0 = const()[name = string("op_3045_end_0"), val = tensor<int32, [4]>([5, 8, 4112, 128])];
            tensor<bool, [4]> var_3045_end_mask_0 = const()[name = string("op_3045_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3045_cast_fp16 = slice_by_index(begin = var_3045_begin_0, end = var_3045_end_0, end_mask = var_3045_end_mask_0, x = read_state_1)[name = string("op_3045_cast_fp16")];
            tensor<int32, [4]> var_3055_begin_0 = const()[name = string("op_3055_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3055_end_0 = const()[name = string("op_3055_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3055_end_mask_0 = const()[name = string("op_3055_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3055_cast_fp16 = slice_by_index(begin = var_3055_begin_0, end = var_3055_end_0, end_mask = var_3055_end_mask_0, x = var_3045_cast_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3061_transpose_x_0 = const()[name = string("op_3061_transpose_x_0"), val = bool(false)];
            bool var_3061_transpose_y_0 = const()[name = string("op_3061_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3061_cast_fp16 = matmul(transpose_x = var_3061_transpose_x_0, transpose_y = var_3061_transpose_y_0, x = var_3040_cast_fp16, y = var_3055_cast_fp16)[name = string("op_3061_cast_fp16")];
            tensor<int32, [4]> var_3066_begin_0 = const()[name = string("op_3066_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_3066_end_0 = const()[name = string("op_3066_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_3066_end_mask_0 = const()[name = string("op_3066_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3066_cast_fp16 = slice_by_index(begin = var_3066_begin_0, end = var_3066_end_0, end_mask = var_3066_end_mask_0, x = shares_9_cast_fp16)[name = string("op_3066_cast_fp16")];
            bool var_3067_transpose_x_0 = const()[name = string("op_3067_transpose_x_0"), val = bool(false)];
            bool var_3067_transpose_y_0 = const()[name = string("op_3067_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_2922_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 8, 32, 128]> var_3067_cast_fp16 = matmul(transpose_x = var_3067_transpose_x_0, transpose_y = var_3067_transpose_y_0, x = var_3066_cast_fp16, y = v_9_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_17_cast_fp16 = add(x = var_3061_cast_fp16, y = var_3067_cast_fp16)[name = string("mixed_17_cast_fp16")];
            tensor<int32, [4]> var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [4]> var_3099_begin_0 = const()[name = string("op_3099_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3099_end_0 = const()[name = string("op_3099_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3099_end_mask_0 = const()[name = string("op_3099_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3099_cast_fp16 = slice_by_index(begin = var_3099_begin_0, end = var_3099_end_0, end_mask = var_3099_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_3099_cast_fp16")];
            int32 var_3101 = const()[name = string("op_3101"), 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_3101, interleave = mixed_19_interleave_0, values = (var_3084_cast_fp16, var_3099_cast_fp16))[name = string("mixed_19_cast_fp16")];
            tensor<int32, [4]> var_3105_perm_0 = const()[name = string("op_3105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3109 = const()[name = string("op_3109"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3105_cast_fp16 = transpose(perm = var_3105_perm_0, x = mixed_19_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 2048]> input_43_cast_fp16 = reshape(shape = var_3109, x = var_3105_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_97_cast_fp16 = add(x = x_81_cast_fp16, y = linear_31_cast_fp16)[name = string("x_97_cast_fp16")];
            bool var_3119 = const()[name = string("op_3119"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3122_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<int32, [1]> peak_39_axes_0 = const()[name = string("peak_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_39_cast_fp16 = reduce_max(axes = peak_39_axes_0, keep_dims = var_3119, x = var_3122_cast_fp16)[name = string("peak_39_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_99_cast_fp16 = real_div(x = x_97_cast_fp16, y = peak_39_cast_fp16)[name = string("x_99_cast_fp16")];
            fp16 var_3118_promoted_to_fp16 = const()[name = string("op_3118_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3126_cast_fp16 = pow(x = x_99_cast_fp16, y = var_3118_promoted_to_fp16)[name = string("op_3126_cast_fp16")];
            tensor<int32, [1]> var_3128_axes_0 = const()[name = string("op_3128_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3128_keep_dims_0 = const()[name = string("op_3128_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3128_cast_fp16 = reduce_mean(axes = var_3128_axes_0, keep_dims = var_3128_keep_dims_0, x = var_3126_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp16 var_3118_promoted_1_to_fp16 = const()[name = string("op_3118_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3129_cast_fp16 = pow(x = peak_39_cast_fp16, y = var_3118_promoted_1_to_fp16)[name = string("op_3129_cast_fp16")];
            fp32 var_3130_epsilon_0 = const()[name = string("op_3130_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3130_cast_fp16 = inverse(epsilon = var_3130_epsilon_0, x = var_3129_cast_fp16)[name = string("op_3130_cast_fp16")];
            fp16 var_3131_to_fp16 = const()[name = string("op_3131_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3132_cast_fp16 = mul(x = var_3130_cast_fp16, y = var_3131_to_fp16)[name = string("op_3132_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3133_cast_fp16 = add(x = var_3128_cast_fp16, y = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            fp32 var_3134_epsilon_0 = const()[name = string("op_3134_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3134_cast_fp16 = rsqrt(epsilon = var_3134_epsilon_0, x = var_3133_cast_fp16)[name = string("op_3134_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3135_cast_fp16 = mul(x = x_99_cast_fp16, y = var_3134_cast_fp16)[name = string("op_3135_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_3135_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_3140_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_3140_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_3140_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_101_cast_fp16 = add(x = x_97_cast_fp16, y = linear_34_cast_fp16)[name = string("x_101_cast_fp16")];
            bool var_3153 = const()[name = string("op_3153"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3156_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<int32, [1]> peak_41_axes_0 = const()[name = string("peak_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_41_cast_fp16 = reduce_max(axes = peak_41_axes_0, keep_dims = var_3153, x = var_3156_cast_fp16)[name = string("peak_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_103_cast_fp16 = real_div(x = x_101_cast_fp16, y = peak_41_cast_fp16)[name = string("x_103_cast_fp16")];
            fp16 var_3152_promoted_to_fp16 = const()[name = string("op_3152_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3160_cast_fp16 = pow(x = x_103_cast_fp16, y = var_3152_promoted_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [1]> var_3162_axes_0 = const()[name = string("op_3162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3162_keep_dims_0 = const()[name = string("op_3162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3162_cast_fp16 = reduce_mean(axes = var_3162_axes_0, keep_dims = var_3162_keep_dims_0, x = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3152_promoted_1_to_fp16 = const()[name = string("op_3152_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3163_cast_fp16 = pow(x = peak_41_cast_fp16, y = var_3152_promoted_1_to_fp16)[name = string("op_3163_cast_fp16")];
            fp32 var_3164_epsilon_0 = const()[name = string("op_3164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3164_cast_fp16 = inverse(epsilon = var_3164_epsilon_0, x = var_3163_cast_fp16)[name = string("op_3164_cast_fp16")];
            fp16 var_3165_to_fp16 = const()[name = string("op_3165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3166_cast_fp16 = mul(x = var_3164_cast_fp16, y = var_3165_to_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3167_cast_fp16 = add(x = var_3162_cast_fp16, y = var_3166_cast_fp16)[name = string("op_3167_cast_fp16")];
            fp32 var_3168_epsilon_0 = const()[name = string("op_3168_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3168_cast_fp16 = rsqrt(epsilon = var_3168_epsilon_0, x = var_3167_cast_fp16)[name = string("op_3168_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3169_cast_fp16 = mul(x = x_103_cast_fp16, y = var_3168_cast_fp16)[name = string("op_3169_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_3169_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_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_35_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> x_105_perm_0 = const()[name = string("x_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3186 = const()[name = string("op_3186"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_105_cast_fp16 = transpose(perm = x_105_perm_0, x = var_3179_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 16, 16, 128]> var_3189_cast_fp16 = abs(x = x_105_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<int32, [1]> peak_43_axes_0 = const()[name = string("peak_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_43_cast_fp16 = reduce_max(axes = peak_43_axes_0, keep_dims = var_3186, x = var_3189_cast_fp16)[name = string("peak_43_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_107_cast_fp16 = real_div(x = x_105_cast_fp16, y = peak_43_cast_fp16)[name = string("x_107_cast_fp16")];
            fp16 var_3185_promoted_to_fp16 = const()[name = string("op_3185_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3193_cast_fp16 = pow(x = x_107_cast_fp16, y = var_3185_promoted_to_fp16)[name = string("op_3193_cast_fp16")];
            tensor<int32, [1]> var_3195_axes_0 = const()[name = string("op_3195_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3195_keep_dims_0 = const()[name = string("op_3195_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3195_cast_fp16 = reduce_mean(axes = var_3195_axes_0, keep_dims = var_3195_keep_dims_0, x = var_3193_cast_fp16)[name = string("op_3195_cast_fp16")];
            fp16 var_3185_promoted_1_to_fp16 = const()[name = string("op_3185_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3196_cast_fp16 = pow(x = peak_43_cast_fp16, y = var_3185_promoted_1_to_fp16)[name = string("op_3196_cast_fp16")];
            fp32 var_3197_epsilon_0 = const()[name = string("op_3197_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3197_cast_fp16 = inverse(epsilon = var_3197_epsilon_0, x = var_3196_cast_fp16)[name = string("op_3197_cast_fp16")];
            fp16 var_3198_to_fp16 = const()[name = string("op_3198_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3199_cast_fp16 = mul(x = var_3197_cast_fp16, y = var_3198_to_fp16)[name = string("op_3199_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3200_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3199_cast_fp16)[name = string("op_3200_cast_fp16")];
            fp32 var_3201_epsilon_0 = const()[name = string("op_3201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3201_cast_fp16 = rsqrt(epsilon = var_3201_epsilon_0, x = var_3200_cast_fp16)[name = string("op_3201_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3202_cast_fp16 = mul(x = x_107_cast_fp16, y = var_3201_cast_fp16)[name = string("op_3202_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_113_cast_fp16 = mul(x = var_3202_cast_fp16, y = model_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_113_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_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_36_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_109_perm_0 = const()[name = string("x_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3219 = const()[name = string("op_3219"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_109_cast_fp16 = transpose(perm = x_109_perm_0, x = var_3212_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 16, 128]> var_3222_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_3222_cast_fp16")];
            tensor<int32, [1]> peak_45_axes_0 = const()[name = string("peak_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_45_cast_fp16 = reduce_max(axes = peak_45_axes_0, keep_dims = var_3219, x = var_3222_cast_fp16)[name = string("peak_45_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_111_cast_fp16 = real_div(x = x_109_cast_fp16, y = peak_45_cast_fp16)[name = string("x_111_cast_fp16")];
            fp16 var_3218_promoted_to_fp16 = const()[name = string("op_3218_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3226_cast_fp16 = pow(x = x_111_cast_fp16, y = var_3218_promoted_to_fp16)[name = string("op_3226_cast_fp16")];
            tensor<int32, [1]> var_3228_axes_0 = const()[name = string("op_3228_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3228_keep_dims_0 = const()[name = string("op_3228_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3228_cast_fp16 = reduce_mean(axes = var_3228_axes_0, keep_dims = var_3228_keep_dims_0, x = var_3226_cast_fp16)[name = string("op_3228_cast_fp16")];
            fp16 var_3218_promoted_1_to_fp16 = const()[name = string("op_3218_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3229_cast_fp16 = pow(x = peak_45_cast_fp16, y = var_3218_promoted_1_to_fp16)[name = string("op_3229_cast_fp16")];
            fp32 var_3230_epsilon_0 = const()[name = string("op_3230_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3230_cast_fp16 = inverse(epsilon = var_3230_epsilon_0, x = var_3229_cast_fp16)[name = string("op_3230_cast_fp16")];
            fp16 var_3231_to_fp16 = const()[name = string("op_3231_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3232_cast_fp16 = mul(x = var_3230_cast_fp16, y = var_3231_to_fp16)[name = string("op_3232_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3233_cast_fp16 = add(x = var_3228_cast_fp16, y = var_3232_cast_fp16)[name = string("op_3233_cast_fp16")];
            fp32 var_3234_epsilon_0 = const()[name = string("op_3234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3234_cast_fp16 = rsqrt(epsilon = var_3234_epsilon_0, x = var_3233_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3235_cast_fp16 = mul(x = x_111_cast_fp16, y = var_3234_cast_fp16)[name = string("op_3235_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_115_cast_fp16 = mul(x = var_3235_cast_fp16, y = model_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_115_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_3244 = const()[name = string("op_3244"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3245_cast_fp16 = reshape(shape = var_3244, x = linear_37_cast_fp16)[name = string("op_3245_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_113_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_113_cast_fp16)[name = string("second_21_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3259_cast_fp16 = mul(x = first_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3260_cast_fp16 = mul(x = second_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3260_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3262_cast_fp16 = sub(x = var_3259_cast_fp16, y = var_3260_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3263_cast_fp16 = mul(x = second_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3264_cast_fp16 = mul(x = first_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3266_cast_fp16 = add(x = var_3263_cast_fp16, y = var_3264_cast_fp16)[name = string("op_3266_cast_fp16")];
            int32 var_3268 = const()[name = string("op_3268"), 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_3268, interleave = q_11_interleave_0, values = (var_3262_cast_fp16, var_3266_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_115_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_115_cast_fp16)[name = string("second_23_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3280_cast_fp16 = mul(x = first_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3280_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3281_cast_fp16 = mul(x = second_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3281_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3283_cast_fp16 = sub(x = var_3280_cast_fp16, y = var_3281_cast_fp16)[name = string("op_3283_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3284_cast_fp16 = mul(x = second_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3285_cast_fp16 = mul(x = first_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3285_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3287_cast_fp16 = add(x = var_3284_cast_fp16, y = var_3285_cast_fp16)[name = string("op_3287_cast_fp16")];
            int32 var_3289 = const()[name = string("op_3289"), 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_3289, interleave = k_11_interleave_0, values = (var_3283_cast_fp16, var_3287_cast_fp16))[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> var_3300_begin_0 = const()[name = string("op_3300_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3300_end_0 = const()[name = string("op_3300_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3300_end_mask_0 = const()[name = string("op_3300_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3300_stride_0 = const()[name = string("op_3300_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3300_cast_fp16 = slice_by_index(begin = var_3300_begin_0, end = var_3300_end_0, end_mask = var_3300_end_mask_0, stride = var_3300_stride_0, x = q_11_cast_fp16)[name = string("op_3300_cast_fp16")];
            tensor<int32, [4]> var_3310_begin_0 = const()[name = string("op_3310_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3310_end_0 = const()[name = string("op_3310_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3310_end_mask_0 = const()[name = string("op_3310_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3310_stride_0 = const()[name = string("op_3310_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3310_cast_fp16 = slice_by_index(begin = var_3310_begin_0, end = var_3310_end_0, end_mask = var_3310_end_mask_0, stride = var_3310_stride_0, x = q_11_cast_fp16)[name = string("op_3310_cast_fp16")];
            int32 var_3312 = const()[name = string("op_3312"), 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_3312, interleave = stacked_11_interleave_0, values = (var_3300_cast_fp16, var_3310_cast_fp16))[name = string("stacked_11_cast_fp16")];
            tensor<int32, [4]> var_3318_begin_0 = const()[name = string("op_3318_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3318_end_0 = const()[name = string("op_3318_end_0"), val = tensor<int32, [4]>([6, 8, 4112, 128])];
            tensor<bool, [4]> var_3318_end_mask_0 = const()[name = string("op_3318_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3318_cast_fp16 = slice_by_index(begin = var_3318_begin_0, end = var_3318_end_0, end_mask = var_3318_end_mask_0, x = read_state_0)[name = string("op_3318_cast_fp16")];
            tensor<int32, [4]> var_3328_begin_0 = const()[name = string("op_3328_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3328_end_0 = const()[name = string("op_3328_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3328_end_mask_0 = const()[name = string("op_3328_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3328_cast_fp16 = slice_by_index(begin = var_3328_begin_0, end = var_3328_end_0, end_mask = var_3328_end_mask_0, x = var_3318_cast_fp16)[name = string("op_3328_cast_fp16")];
            bool var_3337_transpose_x_1 = const()[name = string("op_3337_transpose_x_1"), val = bool(false)];
            bool var_3337_transpose_y_1 = const()[name = string("op_3337_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_3337_cast_fp16 = matmul(transpose_x = var_3337_transpose_x_1, transpose_y = var_3337_transpose_y_1, x = var_3328_cast_fp16, y = stacked_11_cast_fp16)[name = string("op_3337_cast_fp16")];
            tensor<int32, [4]> var_3340_perm_0 = const()[name = string("op_3340_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3341_to_fp16 = const()[name = string("op_3341_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_3340_cast_fp16 = transpose(perm = var_3340_perm_0, x = var_3337_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 8, 32, 4096]> var_3342_cast_fp16 = mul(x = var_3340_cast_fp16, y = var_3341_to_fp16)[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_11_cast_fp16 = add(x = var_3342_cast_fp16, y = cacheMask)[name = string("cache_weights_11_cast_fp16")];
            bool var_3348_transpose_x_1 = const()[name = string("op_3348_transpose_x_1"), val = bool(false)];
            bool var_3348_transpose_y_1 = const()[name = string("op_3348_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3348_cast_fp16 = matmul(transpose_x = var_3348_transpose_x_1, transpose_y = var_3348_transpose_y_1, x = stacked_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_3348_cast_fp16")];
            fp16 var_3349_to_fp16 = const()[name = string("op_3349_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3350_cast_fp16 = mul(x = var_3348_cast_fp16, y = var_3349_to_fp16)[name = string("op_3350_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_11_cast_fp16 = add(x = var_3350_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_11_cast_fp16")];
            int32 var_3354 = const()[name = string("op_3354"), val = int32(-1)];
            bool var_3355_interleave_0 = const()[name = string("op_3355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_3355_cast_fp16 = concat(axis = var_3354, interleave = var_3355_interleave_0, values = (cache_weights_11_cast_fp16, call_weights_11_cast_fp16))[name = string("op_3355_cast_fp16")];
            int32 var_3356 = const()[name = string("op_3356"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_11_cast_fp16 = softmax(axis = var_3356, x = var_3355_cast_fp16)[name = string("shares_11_cast_fp16")];
            tensor<int32, [4]> var_3363_begin_0 = const()[name = string("op_3363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3363_end_0 = const()[name = string("op_3363_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_3363_end_mask_0 = const()[name = string("op_3363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_3363_cast_fp16 = slice_by_index(begin = var_3363_begin_0, end = var_3363_end_0, end_mask = var_3363_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3363_cast_fp16")];
            tensor<int32, [4]> var_3368_begin_0 = const()[name = string("op_3368_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3368_end_0 = const()[name = string("op_3368_end_0"), val = tensor<int32, [4]>([6, 8, 4112, 128])];
            tensor<bool, [4]> var_3368_end_mask_0 = const()[name = string("op_3368_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3368_cast_fp16 = slice_by_index(begin = var_3368_begin_0, end = var_3368_end_0, end_mask = var_3368_end_mask_0, x = read_state_1)[name = string("op_3368_cast_fp16")];
            tensor<int32, [4]> var_3378_begin_0 = const()[name = string("op_3378_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3378_end_0 = const()[name = string("op_3378_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3378_end_mask_0 = const()[name = string("op_3378_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3378_cast_fp16 = slice_by_index(begin = var_3378_begin_0, end = var_3378_end_0, end_mask = var_3378_end_mask_0, x = var_3368_cast_fp16)[name = string("op_3378_cast_fp16")];
            bool var_3384_transpose_x_0 = const()[name = string("op_3384_transpose_x_0"), val = bool(false)];
            bool var_3384_transpose_y_0 = const()[name = string("op_3384_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3384_cast_fp16 = matmul(transpose_x = var_3384_transpose_x_0, transpose_y = var_3384_transpose_y_0, x = var_3363_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<int32, [4]> var_3389_begin_0 = const()[name = string("op_3389_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_3389_end_0 = const()[name = string("op_3389_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_3389_end_mask_0 = const()[name = string("op_3389_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3389_cast_fp16 = slice_by_index(begin = var_3389_begin_0, end = var_3389_end_0, end_mask = var_3389_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3389_cast_fp16")];
            bool var_3390_transpose_x_0 = const()[name = string("op_3390_transpose_x_0"), val = bool(false)];
            bool var_3390_transpose_y_0 = const()[name = string("op_3390_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_3245_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 8, 32, 128]> var_3390_cast_fp16 = matmul(transpose_x = var_3390_transpose_x_0, transpose_y = var_3390_transpose_y_0, x = var_3389_cast_fp16, y = v_11_cast_fp16)[name = string("op_3390_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_21_cast_fp16 = add(x = var_3384_cast_fp16, y = var_3390_cast_fp16)[name = string("mixed_21_cast_fp16")];
            tensor<int32, [4]> var_3407_begin_0 = const()[name = string("op_3407_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3407_end_0 = const()[name = string("op_3407_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3407_end_mask_0 = const()[name = string("op_3407_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3407_cast_fp16")];
            tensor<int32, [4]> var_3422_begin_0 = const()[name = string("op_3422_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3422_end_0 = const()[name = string("op_3422_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3422_end_mask_0 = const()[name = string("op_3422_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3422_cast_fp16 = slice_by_index(begin = var_3422_begin_0, end = var_3422_end_0, end_mask = var_3422_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3422_cast_fp16")];
            int32 var_3424 = const()[name = string("op_3424"), 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_3424, interleave = mixed_23_interleave_0, values = (var_3407_cast_fp16, var_3422_cast_fp16))[name = string("mixed_23_cast_fp16")];
            tensor<int32, [4]> var_3428_perm_0 = const()[name = string("op_3428_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3432 = const()[name = string("op_3432"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3428_cast_fp16 = transpose(perm = var_3428_perm_0, x = mixed_23_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 2048]> input_53_cast_fp16 = reshape(shape = var_3432, x = var_3428_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_117_cast_fp16 = add(x = x_101_cast_fp16, y = linear_38_cast_fp16)[name = string("x_117_cast_fp16")];
            bool var_3442 = const()[name = string("op_3442"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3445_cast_fp16 = abs(x = x_117_cast_fp16)[name = string("op_3445_cast_fp16")];
            tensor<int32, [1]> peak_47_axes_0 = const()[name = string("peak_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_47_cast_fp16 = reduce_max(axes = peak_47_axes_0, keep_dims = var_3442, x = var_3445_cast_fp16)[name = string("peak_47_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_119_cast_fp16 = real_div(x = x_117_cast_fp16, y = peak_47_cast_fp16)[name = string("x_119_cast_fp16")];
            fp16 var_3441_promoted_to_fp16 = const()[name = string("op_3441_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3449_cast_fp16 = pow(x = x_119_cast_fp16, y = var_3441_promoted_to_fp16)[name = string("op_3449_cast_fp16")];
            tensor<int32, [1]> var_3451_axes_0 = const()[name = string("op_3451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3451_keep_dims_0 = const()[name = string("op_3451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3451_cast_fp16 = reduce_mean(axes = var_3451_axes_0, keep_dims = var_3451_keep_dims_0, x = var_3449_cast_fp16)[name = string("op_3451_cast_fp16")];
            fp16 var_3441_promoted_1_to_fp16 = const()[name = string("op_3441_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3452_cast_fp16 = pow(x = peak_47_cast_fp16, y = var_3441_promoted_1_to_fp16)[name = string("op_3452_cast_fp16")];
            fp32 var_3453_epsilon_0 = const()[name = string("op_3453_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3453_cast_fp16 = inverse(epsilon = var_3453_epsilon_0, x = var_3452_cast_fp16)[name = string("op_3453_cast_fp16")];
            fp16 var_3454_to_fp16 = const()[name = string("op_3454_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3455_cast_fp16 = mul(x = var_3453_cast_fp16, y = var_3454_to_fp16)[name = string("op_3455_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3456_cast_fp16 = add(x = var_3451_cast_fp16, y = var_3455_cast_fp16)[name = string("op_3456_cast_fp16")];
            fp32 var_3457_epsilon_0 = const()[name = string("op_3457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3457_cast_fp16 = rsqrt(epsilon = var_3457_epsilon_0, x = var_3456_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3458_cast_fp16 = mul(x = x_119_cast_fp16, y = var_3457_cast_fp16)[name = string("op_3458_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_3458_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_3463_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_3463_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_3463_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_121_cast_fp16 = add(x = x_117_cast_fp16, y = linear_41_cast_fp16)[name = string("x_121_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3479_cast_fp16 = abs(x = x_121_cast_fp16)[name = string("op_3479_cast_fp16")];
            tensor<int32, [1]> peak_49_axes_0 = const()[name = string("peak_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_49_cast_fp16 = reduce_max(axes = peak_49_axes_0, keep_dims = var_3476, x = var_3479_cast_fp16)[name = string("peak_49_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_123_cast_fp16 = real_div(x = x_121_cast_fp16, y = peak_49_cast_fp16)[name = string("x_123_cast_fp16")];
            fp16 var_3475_promoted_to_fp16 = const()[name = string("op_3475_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3483_cast_fp16 = pow(x = x_123_cast_fp16, y = var_3475_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            tensor<int32, [1]> var_3485_axes_0 = const()[name = string("op_3485_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3485_keep_dims_0 = const()[name = string("op_3485_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3485_cast_fp16 = reduce_mean(axes = var_3485_axes_0, keep_dims = var_3485_keep_dims_0, x = var_3483_cast_fp16)[name = string("op_3485_cast_fp16")];
            fp16 var_3475_promoted_1_to_fp16 = const()[name = string("op_3475_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3486_cast_fp16 = pow(x = peak_49_cast_fp16, y = var_3475_promoted_1_to_fp16)[name = string("op_3486_cast_fp16")];
            fp32 var_3487_epsilon_0 = const()[name = string("op_3487_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3487_cast_fp16 = inverse(epsilon = var_3487_epsilon_0, x = var_3486_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp16 var_3488_to_fp16 = const()[name = string("op_3488_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3489_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_3488_to_fp16)[name = string("op_3489_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3490_cast_fp16 = add(x = var_3485_cast_fp16, y = var_3489_cast_fp16)[name = string("op_3490_cast_fp16")];
            fp32 var_3491_epsilon_0 = const()[name = string("op_3491_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3491_cast_fp16 = rsqrt(epsilon = var_3491_epsilon_0, x = var_3490_cast_fp16)[name = string("op_3491_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3492_cast_fp16 = mul(x = x_123_cast_fp16, y = var_3491_cast_fp16)[name = string("op_3492_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_3492_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_3501 = const()[name = string("op_3501"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3502_cast_fp16 = reshape(shape = var_3501, x = linear_42_cast_fp16)[name = string("op_3502_cast_fp16")];
            tensor<int32, [4]> x_125_perm_0 = const()[name = string("x_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3509 = const()[name = string("op_3509"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_125_cast_fp16 = transpose(perm = x_125_perm_0, x = var_3502_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 16, 128]> var_3512_cast_fp16 = abs(x = x_125_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<int32, [1]> peak_51_axes_0 = const()[name = string("peak_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_51_cast_fp16 = reduce_max(axes = peak_51_axes_0, keep_dims = var_3509, x = var_3512_cast_fp16)[name = string("peak_51_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_127_cast_fp16 = real_div(x = x_125_cast_fp16, y = peak_51_cast_fp16)[name = string("x_127_cast_fp16")];
            fp16 var_3508_promoted_to_fp16 = const()[name = string("op_3508_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3516_cast_fp16 = pow(x = x_127_cast_fp16, y = var_3508_promoted_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<int32, [1]> var_3518_axes_0 = const()[name = string("op_3518_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3518_keep_dims_0 = const()[name = string("op_3518_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3518_cast_fp16 = reduce_mean(axes = var_3518_axes_0, keep_dims = var_3518_keep_dims_0, x = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp16 var_3508_promoted_1_to_fp16 = const()[name = string("op_3508_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3519_cast_fp16 = pow(x = peak_51_cast_fp16, y = var_3508_promoted_1_to_fp16)[name = string("op_3519_cast_fp16")];
            fp32 var_3520_epsilon_0 = const()[name = string("op_3520_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3520_cast_fp16 = inverse(epsilon = var_3520_epsilon_0, x = var_3519_cast_fp16)[name = string("op_3520_cast_fp16")];
            fp16 var_3521_to_fp16 = const()[name = string("op_3521_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3522_cast_fp16 = mul(x = var_3520_cast_fp16, y = var_3521_to_fp16)[name = string("op_3522_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3523_cast_fp16 = add(x = var_3518_cast_fp16, y = var_3522_cast_fp16)[name = string("op_3523_cast_fp16")];
            fp32 var_3524_epsilon_0 = const()[name = string("op_3524_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3524_cast_fp16 = rsqrt(epsilon = var_3524_epsilon_0, x = var_3523_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3525_cast_fp16 = mul(x = x_127_cast_fp16, y = var_3524_cast_fp16)[name = string("op_3525_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_133_cast_fp16 = mul(x = var_3525_cast_fp16, y = model_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_133_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_3534 = const()[name = string("op_3534"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3535_cast_fp16 = reshape(shape = var_3534, x = linear_43_cast_fp16)[name = string("op_3535_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3542 = const()[name = string("op_3542"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_3535_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 8, 16, 128]> var_3545_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_3545_cast_fp16")];
            tensor<int32, [1]> peak_53_axes_0 = const()[name = string("peak_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_53_cast_fp16 = reduce_max(axes = peak_53_axes_0, keep_dims = var_3542, x = var_3545_cast_fp16)[name = string("peak_53_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_131_cast_fp16 = real_div(x = x_129_cast_fp16, y = peak_53_cast_fp16)[name = string("x_131_cast_fp16")];
            fp16 var_3541_promoted_to_fp16 = const()[name = string("op_3541_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3549_cast_fp16 = pow(x = x_131_cast_fp16, y = var_3541_promoted_to_fp16)[name = string("op_3549_cast_fp16")];
            tensor<int32, [1]> var_3551_axes_0 = const()[name = string("op_3551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3551_keep_dims_0 = const()[name = string("op_3551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3551_cast_fp16 = reduce_mean(axes = var_3551_axes_0, keep_dims = var_3551_keep_dims_0, x = var_3549_cast_fp16)[name = string("op_3551_cast_fp16")];
            fp16 var_3541_promoted_1_to_fp16 = const()[name = string("op_3541_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3552_cast_fp16 = pow(x = peak_53_cast_fp16, y = var_3541_promoted_1_to_fp16)[name = string("op_3552_cast_fp16")];
            fp32 var_3553_epsilon_0 = const()[name = string("op_3553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3553_cast_fp16 = inverse(epsilon = var_3553_epsilon_0, x = var_3552_cast_fp16)[name = string("op_3553_cast_fp16")];
            fp16 var_3554_to_fp16 = const()[name = string("op_3554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3555_cast_fp16 = mul(x = var_3553_cast_fp16, y = var_3554_to_fp16)[name = string("op_3555_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3556_cast_fp16 = add(x = var_3551_cast_fp16, y = var_3555_cast_fp16)[name = string("op_3556_cast_fp16")];
            fp32 var_3557_epsilon_0 = const()[name = string("op_3557_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3557_cast_fp16 = rsqrt(epsilon = var_3557_epsilon_0, x = var_3556_cast_fp16)[name = string("op_3557_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3558_cast_fp16 = mul(x = x_131_cast_fp16, y = var_3557_cast_fp16)[name = string("op_3558_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_135_cast_fp16 = mul(x = var_3558_cast_fp16, y = model_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_135_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_3567 = const()[name = string("op_3567"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3568_cast_fp16 = reshape(shape = var_3567, x = linear_44_cast_fp16)[name = string("op_3568_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_133_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_133_cast_fp16)[name = string("second_25_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3582_cast_fp16 = mul(x = first_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3582_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3583_cast_fp16 = mul(x = second_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3585_cast_fp16 = sub(x = var_3582_cast_fp16, y = var_3583_cast_fp16)[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3586_cast_fp16 = mul(x = second_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3587_cast_fp16 = mul(x = first_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3587_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3589_cast_fp16 = add(x = var_3586_cast_fp16, y = var_3587_cast_fp16)[name = string("op_3589_cast_fp16")];
            int32 var_3591 = const()[name = string("op_3591"), 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_3591, interleave = q_13_interleave_0, values = (var_3585_cast_fp16, var_3589_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_135_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_135_cast_fp16)[name = string("second_27_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3603_cast_fp16 = mul(x = first_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3603_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3604_cast_fp16 = mul(x = second_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3604_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3606_cast_fp16 = sub(x = var_3603_cast_fp16, y = var_3604_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3607_cast_fp16 = mul(x = second_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3607_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3608_cast_fp16 = mul(x = first_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3608_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3610_cast_fp16 = add(x = var_3607_cast_fp16, y = var_3608_cast_fp16)[name = string("op_3610_cast_fp16")];
            int32 var_3612 = const()[name = string("op_3612"), 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_3612, interleave = k_13_interleave_0, values = (var_3606_cast_fp16, var_3610_cast_fp16))[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_3623_begin_0 = const()[name = string("op_3623_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3623_end_0 = const()[name = string("op_3623_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3623_end_mask_0 = const()[name = string("op_3623_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3623_stride_0 = const()[name = string("op_3623_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3623_cast_fp16 = slice_by_index(begin = var_3623_begin_0, end = var_3623_end_0, end_mask = var_3623_end_mask_0, stride = var_3623_stride_0, x = q_13_cast_fp16)[name = string("op_3623_cast_fp16")];
            tensor<int32, [4]> var_3633_begin_0 = const()[name = string("op_3633_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3633_end_0 = const()[name = string("op_3633_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3633_end_mask_0 = const()[name = string("op_3633_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3633_stride_0 = const()[name = string("op_3633_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3633_cast_fp16 = slice_by_index(begin = var_3633_begin_0, end = var_3633_end_0, end_mask = var_3633_end_mask_0, stride = var_3633_stride_0, x = q_13_cast_fp16)[name = string("op_3633_cast_fp16")];
            int32 var_3635 = const()[name = string("op_3635"), 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_3635, interleave = stacked_13_interleave_0, values = (var_3623_cast_fp16, var_3633_cast_fp16))[name = string("stacked_13_cast_fp16")];
            tensor<int32, [4]> var_3641_begin_0 = const()[name = string("op_3641_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3641_end_0 = const()[name = string("op_3641_end_0"), val = tensor<int32, [4]>([7, 8, 4112, 128])];
            tensor<bool, [4]> var_3641_end_mask_0 = const()[name = string("op_3641_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3641_cast_fp16 = slice_by_index(begin = var_3641_begin_0, end = var_3641_end_0, end_mask = var_3641_end_mask_0, x = read_state_0)[name = string("op_3641_cast_fp16")];
            tensor<int32, [4]> var_3651_begin_0 = const()[name = string("op_3651_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3651_end_0 = const()[name = string("op_3651_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3651_end_mask_0 = const()[name = string("op_3651_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3651_cast_fp16 = slice_by_index(begin = var_3651_begin_0, end = var_3651_end_0, end_mask = var_3651_end_mask_0, x = var_3641_cast_fp16)[name = string("op_3651_cast_fp16")];
            bool var_3660_transpose_x_1 = const()[name = string("op_3660_transpose_x_1"), val = bool(false)];
            bool var_3660_transpose_y_1 = const()[name = string("op_3660_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_3660_cast_fp16 = matmul(transpose_x = var_3660_transpose_x_1, transpose_y = var_3660_transpose_y_1, x = var_3651_cast_fp16, y = stacked_13_cast_fp16)[name = string("op_3660_cast_fp16")];
            tensor<int32, [4]> var_3663_perm_0 = const()[name = string("op_3663_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3664_to_fp16 = const()[name = string("op_3664_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_3663_cast_fp16 = transpose(perm = var_3663_perm_0, x = var_3660_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 8, 32, 4096]> var_3665_cast_fp16 = mul(x = var_3663_cast_fp16, y = var_3664_to_fp16)[name = string("op_3665_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_13_cast_fp16 = add(x = var_3665_cast_fp16, y = cacheMask)[name = string("cache_weights_13_cast_fp16")];
            bool var_3671_transpose_x_1 = const()[name = string("op_3671_transpose_x_1"), val = bool(false)];
            bool var_3671_transpose_y_1 = const()[name = string("op_3671_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3671_cast_fp16 = matmul(transpose_x = var_3671_transpose_x_1, transpose_y = var_3671_transpose_y_1, x = stacked_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_3671_cast_fp16")];
            fp16 var_3672_to_fp16 = const()[name = string("op_3672_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3673_cast_fp16 = mul(x = var_3671_cast_fp16, y = var_3672_to_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_13_cast_fp16 = add(x = var_3673_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_13_cast_fp16")];
            int32 var_3677 = const()[name = string("op_3677"), val = int32(-1)];
            bool var_3678_interleave_0 = const()[name = string("op_3678_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_3678_cast_fp16 = concat(axis = var_3677, interleave = var_3678_interleave_0, values = (cache_weights_13_cast_fp16, call_weights_13_cast_fp16))[name = string("op_3678_cast_fp16")];
            int32 var_3679 = const()[name = string("op_3679"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_13_cast_fp16 = softmax(axis = var_3679, x = var_3678_cast_fp16)[name = string("shares_13_cast_fp16")];
            tensor<int32, [4]> var_3686_begin_0 = const()[name = string("op_3686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3686_end_0 = const()[name = string("op_3686_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_3686_end_mask_0 = const()[name = string("op_3686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_3686_cast_fp16 = slice_by_index(begin = var_3686_begin_0, end = var_3686_end_0, end_mask = var_3686_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3686_cast_fp16")];
            tensor<int32, [4]> var_3691_begin_0 = const()[name = string("op_3691_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3691_end_0 = const()[name = string("op_3691_end_0"), val = tensor<int32, [4]>([7, 8, 4112, 128])];
            tensor<bool, [4]> var_3691_end_mask_0 = const()[name = string("op_3691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3691_cast_fp16 = slice_by_index(begin = var_3691_begin_0, end = var_3691_end_0, end_mask = var_3691_end_mask_0, x = read_state_1)[name = string("op_3691_cast_fp16")];
            tensor<int32, [4]> var_3701_begin_0 = const()[name = string("op_3701_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3701_end_0 = const()[name = string("op_3701_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3701_end_mask_0 = const()[name = string("op_3701_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3701_cast_fp16 = slice_by_index(begin = var_3701_begin_0, end = var_3701_end_0, end_mask = var_3701_end_mask_0, x = var_3691_cast_fp16)[name = string("op_3701_cast_fp16")];
            bool var_3707_transpose_x_0 = const()[name = string("op_3707_transpose_x_0"), val = bool(false)];
            bool var_3707_transpose_y_0 = const()[name = string("op_3707_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3707_cast_fp16 = matmul(transpose_x = var_3707_transpose_x_0, transpose_y = var_3707_transpose_y_0, x = var_3686_cast_fp16, y = var_3701_cast_fp16)[name = string("op_3707_cast_fp16")];
            tensor<int32, [4]> var_3712_begin_0 = const()[name = string("op_3712_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_3712_end_0 = const()[name = string("op_3712_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_3712_end_mask_0 = const()[name = string("op_3712_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3712_cast_fp16 = slice_by_index(begin = var_3712_begin_0, end = var_3712_end_0, end_mask = var_3712_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3712_cast_fp16")];
            bool var_3713_transpose_x_0 = const()[name = string("op_3713_transpose_x_0"), val = bool(false)];
            bool var_3713_transpose_y_0 = const()[name = string("op_3713_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3568_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 8, 32, 128]> var_3713_cast_fp16 = matmul(transpose_x = var_3713_transpose_x_0, transpose_y = var_3713_transpose_y_0, x = var_3712_cast_fp16, y = v_13_cast_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_25_cast_fp16 = add(x = var_3707_cast_fp16, y = var_3713_cast_fp16)[name = string("mixed_25_cast_fp16")];
            tensor<int32, [4]> var_3730_begin_0 = const()[name = string("op_3730_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3730_end_0 = const()[name = string("op_3730_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3730_end_mask_0 = const()[name = string("op_3730_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3730_cast_fp16 = slice_by_index(begin = var_3730_begin_0, end = var_3730_end_0, end_mask = var_3730_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3730_cast_fp16")];
            tensor<int32, [4]> var_3745_begin_0 = const()[name = string("op_3745_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3745_end_0 = const()[name = string("op_3745_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3745_end_mask_0 = const()[name = string("op_3745_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3745_cast_fp16 = slice_by_index(begin = var_3745_begin_0, end = var_3745_end_0, end_mask = var_3745_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3745_cast_fp16")];
            int32 var_3747 = const()[name = string("op_3747"), 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_3747, interleave = mixed_27_interleave_0, values = (var_3730_cast_fp16, var_3745_cast_fp16))[name = string("mixed_27_cast_fp16")];
            tensor<int32, [4]> var_3751_perm_0 = const()[name = string("op_3751_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3755 = const()[name = string("op_3755"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3751_cast_fp16 = transpose(perm = var_3751_perm_0, x = mixed_27_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 2048]> input_63_cast_fp16 = reshape(shape = var_3755, x = var_3751_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_137_cast_fp16 = add(x = x_121_cast_fp16, y = linear_45_cast_fp16)[name = string("x_137_cast_fp16")];
            bool var_3765 = const()[name = string("op_3765"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3768_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_3768_cast_fp16")];
            tensor<int32, [1]> peak_55_axes_0 = const()[name = string("peak_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_55_cast_fp16 = reduce_max(axes = peak_55_axes_0, keep_dims = var_3765, x = var_3768_cast_fp16)[name = string("peak_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_139_cast_fp16 = real_div(x = x_137_cast_fp16, y = peak_55_cast_fp16)[name = string("x_139_cast_fp16")];
            fp16 var_3764_promoted_to_fp16 = const()[name = string("op_3764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3772_cast_fp16 = pow(x = x_139_cast_fp16, y = var_3764_promoted_to_fp16)[name = string("op_3772_cast_fp16")];
            tensor<int32, [1]> var_3774_axes_0 = const()[name = string("op_3774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3774_keep_dims_0 = const()[name = string("op_3774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3774_cast_fp16 = reduce_mean(axes = var_3774_axes_0, keep_dims = var_3774_keep_dims_0, x = var_3772_cast_fp16)[name = string("op_3774_cast_fp16")];
            fp16 var_3764_promoted_1_to_fp16 = const()[name = string("op_3764_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3775_cast_fp16 = pow(x = peak_55_cast_fp16, y = var_3764_promoted_1_to_fp16)[name = string("op_3775_cast_fp16")];
            fp32 var_3776_epsilon_0 = const()[name = string("op_3776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3776_cast_fp16 = inverse(epsilon = var_3776_epsilon_0, x = var_3775_cast_fp16)[name = string("op_3776_cast_fp16")];
            fp16 var_3777_to_fp16 = const()[name = string("op_3777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3778_cast_fp16 = mul(x = var_3776_cast_fp16, y = var_3777_to_fp16)[name = string("op_3778_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3779_cast_fp16 = add(x = var_3774_cast_fp16, y = var_3778_cast_fp16)[name = string("op_3779_cast_fp16")];
            fp32 var_3780_epsilon_0 = const()[name = string("op_3780_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3780_cast_fp16 = rsqrt(epsilon = var_3780_epsilon_0, x = var_3779_cast_fp16)[name = string("op_3780_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3781_cast_fp16 = mul(x = x_139_cast_fp16, y = var_3780_cast_fp16)[name = string("op_3781_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_3781_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_3786_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_3786_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_3786_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_141_cast_fp16 = add(x = x_137_cast_fp16, y = linear_48_cast_fp16)[name = string("x_141_cast_fp16")];
            bool var_3799 = const()[name = string("op_3799"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3802_cast_fp16 = abs(x = x_141_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [1]> peak_57_axes_0 = const()[name = string("peak_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_57_cast_fp16 = reduce_max(axes = peak_57_axes_0, keep_dims = var_3799, x = var_3802_cast_fp16)[name = string("peak_57_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_143_cast_fp16 = real_div(x = x_141_cast_fp16, y = peak_57_cast_fp16)[name = string("x_143_cast_fp16")];
            fp16 var_3798_promoted_to_fp16 = const()[name = string("op_3798_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3806_cast_fp16 = pow(x = x_143_cast_fp16, y = var_3798_promoted_to_fp16)[name = string("op_3806_cast_fp16")];
            tensor<int32, [1]> var_3808_axes_0 = const()[name = string("op_3808_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3808_keep_dims_0 = const()[name = string("op_3808_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3808_cast_fp16 = reduce_mean(axes = var_3808_axes_0, keep_dims = var_3808_keep_dims_0, x = var_3806_cast_fp16)[name = string("op_3808_cast_fp16")];
            fp16 var_3798_promoted_1_to_fp16 = const()[name = string("op_3798_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3809_cast_fp16 = pow(x = peak_57_cast_fp16, y = var_3798_promoted_1_to_fp16)[name = string("op_3809_cast_fp16")];
            fp32 var_3810_epsilon_0 = const()[name = string("op_3810_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3810_cast_fp16 = inverse(epsilon = var_3810_epsilon_0, x = var_3809_cast_fp16)[name = string("op_3810_cast_fp16")];
            fp16 var_3811_to_fp16 = const()[name = string("op_3811_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3812_cast_fp16 = mul(x = var_3810_cast_fp16, y = var_3811_to_fp16)[name = string("op_3812_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3813_cast_fp16 = add(x = var_3808_cast_fp16, y = var_3812_cast_fp16)[name = string("op_3813_cast_fp16")];
            fp32 var_3814_epsilon_0 = const()[name = string("op_3814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3814_cast_fp16 = rsqrt(epsilon = var_3814_epsilon_0, x = var_3813_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3815_cast_fp16 = mul(x = x_143_cast_fp16, y = var_3814_cast_fp16)[name = string("op_3815_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_3815_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_3824 = const()[name = string("op_3824"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3825_cast_fp16 = reshape(shape = var_3824, x = linear_49_cast_fp16)[name = string("op_3825_cast_fp16")];
            tensor<int32, [4]> x_145_perm_0 = const()[name = string("x_145_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_145_cast_fp16 = transpose(perm = x_145_perm_0, x = var_3825_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 16, 16, 128]> var_3835_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_3835_cast_fp16")];
            tensor<int32, [1]> peak_59_axes_0 = const()[name = string("peak_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_59_cast_fp16 = reduce_max(axes = peak_59_axes_0, keep_dims = var_3832, x = var_3835_cast_fp16)[name = string("peak_59_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_147_cast_fp16 = real_div(x = x_145_cast_fp16, y = peak_59_cast_fp16)[name = string("x_147_cast_fp16")];
            fp16 var_3831_promoted_to_fp16 = const()[name = string("op_3831_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3839_cast_fp16 = pow(x = x_147_cast_fp16, y = var_3831_promoted_to_fp16)[name = string("op_3839_cast_fp16")];
            tensor<int32, [1]> var_3841_axes_0 = const()[name = string("op_3841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3841_keep_dims_0 = const()[name = string("op_3841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3841_cast_fp16 = reduce_mean(axes = var_3841_axes_0, keep_dims = var_3841_keep_dims_0, x = var_3839_cast_fp16)[name = string("op_3841_cast_fp16")];
            fp16 var_3831_promoted_1_to_fp16 = const()[name = string("op_3831_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3842_cast_fp16 = pow(x = peak_59_cast_fp16, y = var_3831_promoted_1_to_fp16)[name = string("op_3842_cast_fp16")];
            fp32 var_3843_epsilon_0 = const()[name = string("op_3843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3843_cast_fp16 = inverse(epsilon = var_3843_epsilon_0, x = var_3842_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp16 var_3844_to_fp16 = const()[name = string("op_3844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3845_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3844_to_fp16)[name = string("op_3845_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3846_cast_fp16 = add(x = var_3841_cast_fp16, y = var_3845_cast_fp16)[name = string("op_3846_cast_fp16")];
            fp32 var_3847_epsilon_0 = const()[name = string("op_3847_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3847_cast_fp16 = rsqrt(epsilon = var_3847_epsilon_0, x = var_3846_cast_fp16)[name = string("op_3847_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3848_cast_fp16 = mul(x = x_147_cast_fp16, y = var_3847_cast_fp16)[name = string("op_3848_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_153_cast_fp16 = mul(x = var_3848_cast_fp16, y = model_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_153_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_3857 = const()[name = string("op_3857"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3858_cast_fp16 = reshape(shape = var_3857, x = linear_50_cast_fp16)[name = string("op_3858_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3865 = const()[name = string("op_3865"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_3858_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 8, 16, 128]> var_3868_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_3868_cast_fp16")];
            tensor<int32, [1]> peak_61_axes_0 = const()[name = string("peak_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_61_cast_fp16 = reduce_max(axes = peak_61_axes_0, keep_dims = var_3865, x = var_3868_cast_fp16)[name = string("peak_61_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_151_cast_fp16 = real_div(x = x_149_cast_fp16, y = peak_61_cast_fp16)[name = string("x_151_cast_fp16")];
            fp16 var_3864_promoted_to_fp16 = const()[name = string("op_3864_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3872_cast_fp16 = pow(x = x_151_cast_fp16, y = var_3864_promoted_to_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [1]> var_3874_axes_0 = const()[name = string("op_3874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3874_keep_dims_0 = const()[name = string("op_3874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3874_cast_fp16 = reduce_mean(axes = var_3874_axes_0, keep_dims = var_3874_keep_dims_0, x = var_3872_cast_fp16)[name = string("op_3874_cast_fp16")];
            fp16 var_3864_promoted_1_to_fp16 = const()[name = string("op_3864_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3875_cast_fp16 = pow(x = peak_61_cast_fp16, y = var_3864_promoted_1_to_fp16)[name = string("op_3875_cast_fp16")];
            fp32 var_3876_epsilon_0 = const()[name = string("op_3876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3876_cast_fp16 = inverse(epsilon = var_3876_epsilon_0, x = var_3875_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, 8, 16, 1]> var_3878_cast_fp16 = mul(x = var_3876_cast_fp16, y = var_3877_to_fp16)[name = string("op_3878_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3879_cast_fp16 = add(x = var_3874_cast_fp16, y = var_3878_cast_fp16)[name = string("op_3879_cast_fp16")];
            fp32 var_3880_epsilon_0 = const()[name = string("op_3880_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3880_cast_fp16 = rsqrt(epsilon = var_3880_epsilon_0, x = var_3879_cast_fp16)[name = string("op_3880_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3881_cast_fp16 = mul(x = x_151_cast_fp16, y = var_3880_cast_fp16)[name = string("op_3881_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_155_cast_fp16 = mul(x = var_3881_cast_fp16, y = model_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_155_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_3890 = const()[name = string("op_3890"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3891_cast_fp16 = reshape(shape = var_3890, x = linear_51_cast_fp16)[name = string("op_3891_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_153_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_153_cast_fp16)[name = string("second_29_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3905_cast_fp16 = mul(x = first_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3906_cast_fp16 = mul(x = second_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3908_cast_fp16 = sub(x = var_3905_cast_fp16, y = var_3906_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3909_cast_fp16 = mul(x = second_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3909_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3910_cast_fp16 = mul(x = first_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3910_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3912_cast_fp16 = add(x = var_3909_cast_fp16, y = var_3910_cast_fp16)[name = string("op_3912_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), 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_3914, interleave = q_15_interleave_0, values = (var_3908_cast_fp16, var_3912_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_155_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_155_cast_fp16)[name = string("second_31_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3926_cast_fp16 = mul(x = first_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3927_cast_fp16 = mul(x = second_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3927_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3929_cast_fp16 = sub(x = var_3926_cast_fp16, y = var_3927_cast_fp16)[name = string("op_3929_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3930_cast_fp16 = mul(x = second_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3930_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3931_cast_fp16 = mul(x = first_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3933_cast_fp16 = add(x = var_3930_cast_fp16, y = var_3931_cast_fp16)[name = string("op_3933_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), 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_3935, interleave = k_15_interleave_0, values = (var_3929_cast_fp16, var_3933_cast_fp16))[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> var_3946_begin_0 = const()[name = string("op_3946_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3946_end_0 = const()[name = string("op_3946_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3946_end_mask_0 = const()[name = string("op_3946_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3946_stride_0 = const()[name = string("op_3946_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3946_cast_fp16 = slice_by_index(begin = var_3946_begin_0, end = var_3946_end_0, end_mask = var_3946_end_mask_0, stride = var_3946_stride_0, x = q_15_cast_fp16)[name = string("op_3946_cast_fp16")];
            tensor<int32, [4]> var_3956_begin_0 = const()[name = string("op_3956_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3956_end_0 = const()[name = string("op_3956_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3956_end_mask_0 = const()[name = string("op_3956_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3956_stride_0 = const()[name = string("op_3956_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3956_cast_fp16 = slice_by_index(begin = var_3956_begin_0, end = var_3956_end_0, end_mask = var_3956_end_mask_0, stride = var_3956_stride_0, x = q_15_cast_fp16)[name = string("op_3956_cast_fp16")];
            int32 var_3958 = const()[name = string("op_3958"), 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_3958, interleave = stacked_15_interleave_0, values = (var_3946_cast_fp16, var_3956_cast_fp16))[name = string("stacked_15_cast_fp16")];
            tensor<int32, [4]> var_3964_begin_0 = const()[name = string("op_3964_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3964_end_0 = const()[name = string("op_3964_end_0"), val = tensor<int32, [4]>([8, 8, 4112, 128])];
            tensor<bool, [4]> var_3964_end_mask_0 = const()[name = string("op_3964_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3964_cast_fp16 = slice_by_index(begin = var_3964_begin_0, end = var_3964_end_0, end_mask = var_3964_end_mask_0, x = read_state_0)[name = string("op_3964_cast_fp16")];
            tensor<int32, [4]> var_3974_begin_0 = const()[name = string("op_3974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3974_end_0 = const()[name = string("op_3974_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3974_end_mask_0 = const()[name = string("op_3974_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3974_cast_fp16 = slice_by_index(begin = var_3974_begin_0, end = var_3974_end_0, end_mask = var_3974_end_mask_0, x = var_3964_cast_fp16)[name = string("op_3974_cast_fp16")];
            bool var_3983_transpose_x_1 = const()[name = string("op_3983_transpose_x_1"), val = bool(false)];
            bool var_3983_transpose_y_1 = const()[name = string("op_3983_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_3983_cast_fp16 = matmul(transpose_x = var_3983_transpose_x_1, transpose_y = var_3983_transpose_y_1, x = var_3974_cast_fp16, y = stacked_15_cast_fp16)[name = string("op_3983_cast_fp16")];
            tensor<int32, [4]> var_3986_perm_0 = const()[name = string("op_3986_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3987_to_fp16 = const()[name = string("op_3987_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_3986_cast_fp16 = transpose(perm = var_3986_perm_0, x = var_3983_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 8, 32, 4096]> var_3988_cast_fp16 = mul(x = var_3986_cast_fp16, y = var_3987_to_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_15_cast_fp16 = add(x = var_3988_cast_fp16, y = cacheMask)[name = string("cache_weights_15_cast_fp16")];
            bool var_3994_transpose_x_1 = const()[name = string("op_3994_transpose_x_1"), val = bool(false)];
            bool var_3994_transpose_y_1 = const()[name = string("op_3994_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3994_cast_fp16 = matmul(transpose_x = var_3994_transpose_x_1, transpose_y = var_3994_transpose_y_1, x = stacked_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_3994_cast_fp16")];
            fp16 var_3995_to_fp16 = const()[name = string("op_3995_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3996_cast_fp16 = mul(x = var_3994_cast_fp16, y = var_3995_to_fp16)[name = string("op_3996_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_15_cast_fp16 = add(x = var_3996_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_15_cast_fp16")];
            int32 var_4000 = const()[name = string("op_4000"), val = int32(-1)];
            bool var_4001_interleave_0 = const()[name = string("op_4001_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4001_cast_fp16 = concat(axis = var_4000, interleave = var_4001_interleave_0, values = (cache_weights_15_cast_fp16, call_weights_15_cast_fp16))[name = string("op_4001_cast_fp16")];
            int32 var_4002 = const()[name = string("op_4002"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_15_cast_fp16 = softmax(axis = var_4002, x = var_4001_cast_fp16)[name = string("shares_15_cast_fp16")];
            tensor<int32, [4]> var_4009_begin_0 = const()[name = string("op_4009_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4009_end_0 = const()[name = string("op_4009_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4009_end_mask_0 = const()[name = string("op_4009_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4009_cast_fp16 = slice_by_index(begin = var_4009_begin_0, end = var_4009_end_0, end_mask = var_4009_end_mask_0, x = shares_15_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4014_begin_0 = const()[name = string("op_4014_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_4014_end_0 = const()[name = string("op_4014_end_0"), val = tensor<int32, [4]>([8, 8, 4112, 128])];
            tensor<bool, [4]> var_4014_end_mask_0 = const()[name = string("op_4014_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4014_cast_fp16 = slice_by_index(begin = var_4014_begin_0, end = var_4014_end_0, end_mask = var_4014_end_mask_0, x = read_state_1)[name = string("op_4014_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, 8, 4096, 128])];
            tensor<bool, [4]> var_4024_end_mask_0 = const()[name = string("op_4024_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 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, x = var_4014_cast_fp16)[name = string("op_4024_cast_fp16")];
            bool var_4030_transpose_x_0 = const()[name = string("op_4030_transpose_x_0"), val = bool(false)];
            bool var_4030_transpose_y_0 = const()[name = string("op_4030_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4030_cast_fp16 = matmul(transpose_x = var_4030_transpose_x_0, transpose_y = var_4030_transpose_y_0, x = var_4009_cast_fp16, y = var_4024_cast_fp16)[name = string("op_4030_cast_fp16")];
            tensor<int32, [4]> var_4035_begin_0 = const()[name = string("op_4035_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_4035_end_0 = const()[name = string("op_4035_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_4035_end_mask_0 = const()[name = string("op_4035_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4035_cast_fp16 = slice_by_index(begin = var_4035_begin_0, end = var_4035_end_0, end_mask = var_4035_end_mask_0, x = shares_15_cast_fp16)[name = string("op_4035_cast_fp16")];
            bool var_4036_transpose_x_0 = const()[name = string("op_4036_transpose_x_0"), val = bool(false)];
            bool var_4036_transpose_y_0 = const()[name = string("op_4036_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_3891_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 8, 32, 128]> var_4036_cast_fp16 = matmul(transpose_x = var_4036_transpose_x_0, transpose_y = var_4036_transpose_y_0, x = var_4035_cast_fp16, y = v_15_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_29_cast_fp16 = add(x = var_4030_cast_fp16, y = var_4036_cast_fp16)[name = string("mixed_29_cast_fp16")];
            tensor<int32, [4]> var_4053_begin_0 = const()[name = string("op_4053_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4053_end_0 = const()[name = string("op_4053_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4053_end_mask_0 = const()[name = string("op_4053_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4053_cast_fp16 = slice_by_index(begin = var_4053_begin_0, end = var_4053_end_0, end_mask = var_4053_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_4053_cast_fp16")];
            tensor<int32, [4]> var_4068_begin_0 = const()[name = string("op_4068_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4068_end_0 = const()[name = string("op_4068_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4068_end_mask_0 = const()[name = string("op_4068_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4068_cast_fp16 = slice_by_index(begin = var_4068_begin_0, end = var_4068_end_0, end_mask = var_4068_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_4068_cast_fp16")];
            int32 var_4070 = const()[name = string("op_4070"), 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_4070, interleave = mixed_31_interleave_0, values = (var_4053_cast_fp16, var_4068_cast_fp16))[name = string("mixed_31_cast_fp16")];
            tensor<int32, [4]> var_4074_perm_0 = const()[name = string("op_4074_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4078 = const()[name = string("op_4078"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4074_cast_fp16 = transpose(perm = var_4074_perm_0, x = mixed_31_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 16, 2048]> input_73_cast_fp16 = reshape(shape = var_4078, x = var_4074_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_157_cast_fp16 = add(x = x_141_cast_fp16, y = linear_52_cast_fp16)[name = string("x_157_cast_fp16")];
            bool var_4088 = const()[name = string("op_4088"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4091_cast_fp16 = abs(x = x_157_cast_fp16)[name = string("op_4091_cast_fp16")];
            tensor<int32, [1]> peak_63_axes_0 = const()[name = string("peak_63_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_63_cast_fp16 = reduce_max(axes = peak_63_axes_0, keep_dims = var_4088, x = var_4091_cast_fp16)[name = string("peak_63_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_159_cast_fp16 = real_div(x = x_157_cast_fp16, y = peak_63_cast_fp16)[name = string("x_159_cast_fp16")];
            fp16 var_4087_promoted_to_fp16 = const()[name = string("op_4087_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4095_cast_fp16 = pow(x = x_159_cast_fp16, y = var_4087_promoted_to_fp16)[name = string("op_4095_cast_fp16")];
            tensor<int32, [1]> var_4097_axes_0 = const()[name = string("op_4097_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4097_keep_dims_0 = const()[name = string("op_4097_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4097_cast_fp16 = reduce_mean(axes = var_4097_axes_0, keep_dims = var_4097_keep_dims_0, x = var_4095_cast_fp16)[name = string("op_4097_cast_fp16")];
            fp16 var_4087_promoted_1_to_fp16 = const()[name = string("op_4087_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4098_cast_fp16 = pow(x = peak_63_cast_fp16, y = var_4087_promoted_1_to_fp16)[name = string("op_4098_cast_fp16")];
            fp32 var_4099_epsilon_0 = const()[name = string("op_4099_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4099_cast_fp16 = inverse(epsilon = var_4099_epsilon_0, x = var_4098_cast_fp16)[name = string("op_4099_cast_fp16")];
            fp16 var_4100_to_fp16 = const()[name = string("op_4100_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4101_cast_fp16 = mul(x = var_4099_cast_fp16, y = var_4100_to_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4102_cast_fp16 = add(x = var_4097_cast_fp16, y = var_4101_cast_fp16)[name = string("op_4102_cast_fp16")];
            fp32 var_4103_epsilon_0 = const()[name = string("op_4103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4103_cast_fp16 = rsqrt(epsilon = var_4103_epsilon_0, x = var_4102_cast_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4104_cast_fp16 = mul(x = x_159_cast_fp16, y = var_4103_cast_fp16)[name = string("op_4104_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_4104_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_4109_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_4109_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_4109_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_161_cast_fp16 = add(x = x_157_cast_fp16, y = linear_55_cast_fp16)[name = string("x_161_cast_fp16")];
            bool var_4122 = const()[name = string("op_4122"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4125_cast_fp16 = abs(x = x_161_cast_fp16)[name = string("op_4125_cast_fp16")];
            tensor<int32, [1]> peak_65_axes_0 = const()[name = string("peak_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_65_cast_fp16 = reduce_max(axes = peak_65_axes_0, keep_dims = var_4122, x = var_4125_cast_fp16)[name = string("peak_65_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_163_cast_fp16 = real_div(x = x_161_cast_fp16, y = peak_65_cast_fp16)[name = string("x_163_cast_fp16")];
            fp16 var_4121_promoted_to_fp16 = const()[name = string("op_4121_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4129_cast_fp16 = pow(x = x_163_cast_fp16, y = var_4121_promoted_to_fp16)[name = string("op_4129_cast_fp16")];
            tensor<int32, [1]> var_4131_axes_0 = const()[name = string("op_4131_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4131_keep_dims_0 = const()[name = string("op_4131_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4131_cast_fp16 = reduce_mean(axes = var_4131_axes_0, keep_dims = var_4131_keep_dims_0, x = var_4129_cast_fp16)[name = string("op_4131_cast_fp16")];
            fp16 var_4121_promoted_1_to_fp16 = const()[name = string("op_4121_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4132_cast_fp16 = pow(x = peak_65_cast_fp16, y = var_4121_promoted_1_to_fp16)[name = string("op_4132_cast_fp16")];
            fp32 var_4133_epsilon_0 = const()[name = string("op_4133_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4133_cast_fp16 = inverse(epsilon = var_4133_epsilon_0, x = var_4132_cast_fp16)[name = string("op_4133_cast_fp16")];
            fp16 var_4134_to_fp16 = const()[name = string("op_4134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4135_cast_fp16 = mul(x = var_4133_cast_fp16, y = var_4134_to_fp16)[name = string("op_4135_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4136_cast_fp16 = add(x = var_4131_cast_fp16, y = var_4135_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp32 var_4137_epsilon_0 = const()[name = string("op_4137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4137_cast_fp16 = rsqrt(epsilon = var_4137_epsilon_0, x = var_4136_cast_fp16)[name = string("op_4137_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4138_cast_fp16 = mul(x = x_163_cast_fp16, y = var_4137_cast_fp16)[name = string("op_4138_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_4138_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_4147 = const()[name = string("op_4147"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4148_cast_fp16 = reshape(shape = var_4147, x = linear_56_cast_fp16)[name = string("op_4148_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4155 = const()[name = string("op_4155"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_4148_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 16, 128]> var_4158_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_4158_cast_fp16")];
            tensor<int32, [1]> peak_67_axes_0 = const()[name = string("peak_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_67_cast_fp16 = reduce_max(axes = peak_67_axes_0, keep_dims = var_4155, x = var_4158_cast_fp16)[name = string("peak_67_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_167_cast_fp16 = real_div(x = x_165_cast_fp16, y = peak_67_cast_fp16)[name = string("x_167_cast_fp16")];
            fp16 var_4154_promoted_to_fp16 = const()[name = string("op_4154_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4162_cast_fp16 = pow(x = x_167_cast_fp16, y = var_4154_promoted_to_fp16)[name = string("op_4162_cast_fp16")];
            tensor<int32, [1]> var_4164_axes_0 = const()[name = string("op_4164_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4164_keep_dims_0 = const()[name = string("op_4164_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4164_cast_fp16 = reduce_mean(axes = var_4164_axes_0, keep_dims = var_4164_keep_dims_0, x = var_4162_cast_fp16)[name = string("op_4164_cast_fp16")];
            fp16 var_4154_promoted_1_to_fp16 = const()[name = string("op_4154_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4165_cast_fp16 = pow(x = peak_67_cast_fp16, y = var_4154_promoted_1_to_fp16)[name = string("op_4165_cast_fp16")];
            fp32 var_4166_epsilon_0 = const()[name = string("op_4166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4166_cast_fp16 = inverse(epsilon = var_4166_epsilon_0, x = var_4165_cast_fp16)[name = string("op_4166_cast_fp16")];
            fp16 var_4167_to_fp16 = const()[name = string("op_4167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4168_cast_fp16 = mul(x = var_4166_cast_fp16, y = var_4167_to_fp16)[name = string("op_4168_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4169_cast_fp16 = add(x = var_4164_cast_fp16, y = var_4168_cast_fp16)[name = string("op_4169_cast_fp16")];
            fp32 var_4170_epsilon_0 = const()[name = string("op_4170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4170_cast_fp16 = rsqrt(epsilon = var_4170_epsilon_0, x = var_4169_cast_fp16)[name = string("op_4170_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4171_cast_fp16 = mul(x = x_167_cast_fp16, y = var_4170_cast_fp16)[name = string("op_4171_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_173_cast_fp16 = mul(x = var_4171_cast_fp16, y = model_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_173_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_4180 = const()[name = string("op_4180"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4181_cast_fp16 = reshape(shape = var_4180, x = linear_57_cast_fp16)[name = string("op_4181_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4188 = const()[name = string("op_4188"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_4181_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 8, 16, 128]> var_4191_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_4191_cast_fp16")];
            tensor<int32, [1]> peak_69_axes_0 = const()[name = string("peak_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_69_cast_fp16 = reduce_max(axes = peak_69_axes_0, keep_dims = var_4188, x = var_4191_cast_fp16)[name = string("peak_69_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_171_cast_fp16 = real_div(x = x_169_cast_fp16, y = peak_69_cast_fp16)[name = string("x_171_cast_fp16")];
            fp16 var_4187_promoted_to_fp16 = const()[name = string("op_4187_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4195_cast_fp16 = pow(x = x_171_cast_fp16, y = var_4187_promoted_to_fp16)[name = string("op_4195_cast_fp16")];
            tensor<int32, [1]> var_4197_axes_0 = const()[name = string("op_4197_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4197_keep_dims_0 = const()[name = string("op_4197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4197_cast_fp16 = reduce_mean(axes = var_4197_axes_0, keep_dims = var_4197_keep_dims_0, x = var_4195_cast_fp16)[name = string("op_4197_cast_fp16")];
            fp16 var_4187_promoted_1_to_fp16 = const()[name = string("op_4187_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4198_cast_fp16 = pow(x = peak_69_cast_fp16, y = var_4187_promoted_1_to_fp16)[name = string("op_4198_cast_fp16")];
            fp32 var_4199_epsilon_0 = const()[name = string("op_4199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4199_cast_fp16 = inverse(epsilon = var_4199_epsilon_0, x = var_4198_cast_fp16)[name = string("op_4199_cast_fp16")];
            fp16 var_4200_to_fp16 = const()[name = string("op_4200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4201_cast_fp16 = mul(x = var_4199_cast_fp16, y = var_4200_to_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4202_cast_fp16 = add(x = var_4197_cast_fp16, y = var_4201_cast_fp16)[name = string("op_4202_cast_fp16")];
            fp32 var_4203_epsilon_0 = const()[name = string("op_4203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4203_cast_fp16 = rsqrt(epsilon = var_4203_epsilon_0, x = var_4202_cast_fp16)[name = string("op_4203_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4204_cast_fp16 = mul(x = x_171_cast_fp16, y = var_4203_cast_fp16)[name = string("op_4204_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_175_cast_fp16 = mul(x = var_4204_cast_fp16, y = model_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_175_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_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_58_cast_fp16)[name = string("op_4214_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_173_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_173_cast_fp16)[name = string("second_33_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4228_cast_fp16 = mul(x = first_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4229_cast_fp16 = mul(x = second_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4229_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4231_cast_fp16 = sub(x = var_4228_cast_fp16, y = var_4229_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4232_cast_fp16 = mul(x = second_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4233_cast_fp16 = mul(x = first_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4233_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4235_cast_fp16 = add(x = var_4232_cast_fp16, y = var_4233_cast_fp16)[name = string("op_4235_cast_fp16")];
            int32 var_4237 = const()[name = string("op_4237"), 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_4237, interleave = q_17_interleave_0, values = (var_4231_cast_fp16, var_4235_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_175_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_175_cast_fp16)[name = string("second_35_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4249_cast_fp16 = mul(x = first_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4250_cast_fp16 = mul(x = second_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4252_cast_fp16 = sub(x = var_4249_cast_fp16, y = var_4250_cast_fp16)[name = string("op_4252_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4253_cast_fp16 = mul(x = second_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4254_cast_fp16 = mul(x = first_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4256_cast_fp16 = add(x = var_4253_cast_fp16, y = var_4254_cast_fp16)[name = string("op_4256_cast_fp16")];
            int32 var_4258 = const()[name = string("op_4258"), 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_4258, interleave = k_17_interleave_0, values = (var_4252_cast_fp16, var_4256_cast_fp16))[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_4269_begin_0 = const()[name = string("op_4269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4269_end_0 = const()[name = string("op_4269_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4269_end_mask_0 = const()[name = string("op_4269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4269_stride_0 = const()[name = string("op_4269_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4269_cast_fp16 = slice_by_index(begin = var_4269_begin_0, end = var_4269_end_0, end_mask = var_4269_end_mask_0, stride = var_4269_stride_0, x = q_17_cast_fp16)[name = string("op_4269_cast_fp16")];
            tensor<int32, [4]> var_4279_begin_0 = const()[name = string("op_4279_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4279_end_0 = const()[name = string("op_4279_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4279_end_mask_0 = const()[name = string("op_4279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4279_stride_0 = const()[name = string("op_4279_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4279_cast_fp16 = slice_by_index(begin = var_4279_begin_0, end = var_4279_end_0, end_mask = var_4279_end_mask_0, stride = var_4279_stride_0, x = q_17_cast_fp16)[name = string("op_4279_cast_fp16")];
            int32 var_4281 = const()[name = string("op_4281"), 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_4281, interleave = stacked_17_interleave_0, values = (var_4269_cast_fp16, var_4279_cast_fp16))[name = string("stacked_17_cast_fp16")];
            tensor<int32, [4]> var_4287_begin_0 = const()[name = string("op_4287_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4287_end_0 = const()[name = string("op_4287_end_0"), val = tensor<int32, [4]>([9, 8, 4112, 128])];
            tensor<bool, [4]> var_4287_end_mask_0 = const()[name = string("op_4287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4287_cast_fp16 = slice_by_index(begin = var_4287_begin_0, end = var_4287_end_0, end_mask = var_4287_end_mask_0, x = read_state_0)[name = string("op_4287_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = var_4287_cast_fp16)[name = string("op_4297_cast_fp16")];
            bool var_4306_transpose_x_1 = const()[name = string("op_4306_transpose_x_1"), val = bool(false)];
            bool var_4306_transpose_y_1 = const()[name = string("op_4306_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_4306_cast_fp16 = matmul(transpose_x = var_4306_transpose_x_1, transpose_y = var_4306_transpose_y_1, x = var_4297_cast_fp16, y = stacked_17_cast_fp16)[name = string("op_4306_cast_fp16")];
            tensor<int32, [4]> var_4309_perm_0 = const()[name = string("op_4309_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4310_to_fp16 = const()[name = string("op_4310_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_4309_cast_fp16 = transpose(perm = var_4309_perm_0, x = var_4306_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 32, 4096]> var_4311_cast_fp16 = mul(x = var_4309_cast_fp16, y = var_4310_to_fp16)[name = string("op_4311_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_17_cast_fp16 = add(x = var_4311_cast_fp16, y = cacheMask)[name = string("cache_weights_17_cast_fp16")];
            bool var_4317_transpose_x_1 = const()[name = string("op_4317_transpose_x_1"), val = bool(false)];
            bool var_4317_transpose_y_1 = const()[name = string("op_4317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4317_cast_fp16 = matmul(transpose_x = var_4317_transpose_x_1, transpose_y = var_4317_transpose_y_1, x = stacked_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_4317_cast_fp16")];
            fp16 var_4318_to_fp16 = const()[name = string("op_4318_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4319_cast_fp16 = mul(x = var_4317_cast_fp16, y = var_4318_to_fp16)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_17_cast_fp16 = add(x = var_4319_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_17_cast_fp16")];
            int32 var_4323 = const()[name = string("op_4323"), val = int32(-1)];
            bool var_4324_interleave_0 = const()[name = string("op_4324_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4324_cast_fp16 = concat(axis = var_4323, interleave = var_4324_interleave_0, values = (cache_weights_17_cast_fp16, call_weights_17_cast_fp16))[name = string("op_4324_cast_fp16")];
            int32 var_4325 = const()[name = string("op_4325"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_17_cast_fp16 = softmax(axis = var_4325, x = var_4324_cast_fp16)[name = string("shares_17_cast_fp16")];
            tensor<int32, [4]> var_4332_begin_0 = const()[name = string("op_4332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4332_end_0 = const()[name = string("op_4332_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4332_end_mask_0 = const()[name = string("op_4332_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4332_cast_fp16 = slice_by_index(begin = var_4332_begin_0, end = var_4332_end_0, end_mask = var_4332_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4332_cast_fp16")];
            tensor<int32, [4]> var_4337_begin_0 = const()[name = string("op_4337_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4337_end_0 = const()[name = string("op_4337_end_0"), val = tensor<int32, [4]>([9, 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_1)[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_4353_transpose_x_0 = const()[name = string("op_4353_transpose_x_0"), val = bool(false)];
            bool var_4353_transpose_y_0 = const()[name = string("op_4353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4353_cast_fp16 = matmul(transpose_x = var_4353_transpose_x_0, transpose_y = var_4353_transpose_y_0, x = var_4332_cast_fp16, y = var_4347_cast_fp16)[name = string("op_4353_cast_fp16")];
            tensor<int32, [4]> var_4358_begin_0 = const()[name = string("op_4358_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_4358_end_0 = const()[name = string("op_4358_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_4358_end_mask_0 = const()[name = string("op_4358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4358_cast_fp16 = slice_by_index(begin = var_4358_begin_0, end = var_4358_end_0, end_mask = var_4358_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4358_cast_fp16")];
            bool var_4359_transpose_x_0 = const()[name = string("op_4359_transpose_x_0"), val = bool(false)];
            bool var_4359_transpose_y_0 = const()[name = string("op_4359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_4214_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 32, 128]> var_4359_cast_fp16 = matmul(transpose_x = var_4359_transpose_x_0, transpose_y = var_4359_transpose_y_0, x = var_4358_cast_fp16, y = v_17_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_33_cast_fp16 = add(x = var_4353_cast_fp16, y = var_4359_cast_fp16)[name = string("mixed_33_cast_fp16")];
            tensor<int32, [4]> var_4376_begin_0 = const()[name = string("op_4376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4376_end_0 = const()[name = string("op_4376_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4376_end_mask_0 = const()[name = string("op_4376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4376_cast_fp16 = slice_by_index(begin = var_4376_begin_0, end = var_4376_end_0, end_mask = var_4376_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [4]> var_4391_begin_0 = const()[name = string("op_4391_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4391_end_0 = const()[name = string("op_4391_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4391_end_mask_0 = const()[name = string("op_4391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4391_cast_fp16 = slice_by_index(begin = var_4391_begin_0, end = var_4391_end_0, end_mask = var_4391_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4391_cast_fp16")];
            int32 var_4393 = const()[name = string("op_4393"), 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_4393, interleave = mixed_35_interleave_0, values = (var_4376_cast_fp16, var_4391_cast_fp16))[name = string("mixed_35_cast_fp16")];
            tensor<int32, [4]> var_4397_perm_0 = const()[name = string("op_4397_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4401 = const()[name = string("op_4401"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4397_cast_fp16 = transpose(perm = var_4397_perm_0, x = mixed_35_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 2048]> input_83_cast_fp16 = reshape(shape = var_4401, x = var_4397_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_177_cast_fp16 = add(x = x_161_cast_fp16, y = linear_59_cast_fp16)[name = string("x_177_cast_fp16")];
            bool var_4411 = const()[name = string("op_4411"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4414_cast_fp16 = abs(x = x_177_cast_fp16)[name = string("op_4414_cast_fp16")];
            tensor<int32, [1]> peak_71_axes_0 = const()[name = string("peak_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_71_cast_fp16 = reduce_max(axes = peak_71_axes_0, keep_dims = var_4411, x = var_4414_cast_fp16)[name = string("peak_71_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_179_cast_fp16 = real_div(x = x_177_cast_fp16, y = peak_71_cast_fp16)[name = string("x_179_cast_fp16")];
            fp16 var_4410_promoted_to_fp16 = const()[name = string("op_4410_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4418_cast_fp16 = pow(x = x_179_cast_fp16, y = var_4410_promoted_to_fp16)[name = string("op_4418_cast_fp16")];
            tensor<int32, [1]> var_4420_axes_0 = const()[name = string("op_4420_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4420_keep_dims_0 = const()[name = string("op_4420_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4420_cast_fp16 = reduce_mean(axes = var_4420_axes_0, keep_dims = var_4420_keep_dims_0, x = var_4418_cast_fp16)[name = string("op_4420_cast_fp16")];
            fp16 var_4410_promoted_1_to_fp16 = const()[name = string("op_4410_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4421_cast_fp16 = pow(x = peak_71_cast_fp16, y = var_4410_promoted_1_to_fp16)[name = string("op_4421_cast_fp16")];
            fp32 var_4422_epsilon_0 = const()[name = string("op_4422_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4422_cast_fp16 = inverse(epsilon = var_4422_epsilon_0, x = var_4421_cast_fp16)[name = string("op_4422_cast_fp16")];
            fp16 var_4423_to_fp16 = const()[name = string("op_4423_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4424_cast_fp16 = mul(x = var_4422_cast_fp16, y = var_4423_to_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4425_cast_fp16 = add(x = var_4420_cast_fp16, y = var_4424_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4426_cast_fp16 = rsqrt(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4427_cast_fp16 = mul(x = x_179_cast_fp16, y = var_4426_cast_fp16)[name = string("op_4427_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_4427_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_4432_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_4432_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_4432_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_181_cast_fp16 = add(x = x_177_cast_fp16, y = linear_62_cast_fp16)[name = string("x_181_cast_fp16")];
            bool var_4445 = const()[name = string("op_4445"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4448_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_4448_cast_fp16")];
            tensor<int32, [1]> peak_73_axes_0 = const()[name = string("peak_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_73_cast_fp16 = reduce_max(axes = peak_73_axes_0, keep_dims = var_4445, x = var_4448_cast_fp16)[name = string("peak_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_183_cast_fp16 = real_div(x = x_181_cast_fp16, y = peak_73_cast_fp16)[name = string("x_183_cast_fp16")];
            fp16 var_4444_promoted_to_fp16 = const()[name = string("op_4444_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4452_cast_fp16 = pow(x = x_183_cast_fp16, y = var_4444_promoted_to_fp16)[name = string("op_4452_cast_fp16")];
            tensor<int32, [1]> var_4454_axes_0 = const()[name = string("op_4454_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4454_keep_dims_0 = const()[name = string("op_4454_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4454_cast_fp16 = reduce_mean(axes = var_4454_axes_0, keep_dims = var_4454_keep_dims_0, x = var_4452_cast_fp16)[name = string("op_4454_cast_fp16")];
            fp16 var_4444_promoted_1_to_fp16 = const()[name = string("op_4444_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4455_cast_fp16 = pow(x = peak_73_cast_fp16, y = var_4444_promoted_1_to_fp16)[name = string("op_4455_cast_fp16")];
            fp32 var_4456_epsilon_0 = const()[name = string("op_4456_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4456_cast_fp16 = inverse(epsilon = var_4456_epsilon_0, x = var_4455_cast_fp16)[name = string("op_4456_cast_fp16")];
            fp16 var_4457_to_fp16 = const()[name = string("op_4457_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4458_cast_fp16 = mul(x = var_4456_cast_fp16, y = var_4457_to_fp16)[name = string("op_4458_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4459_cast_fp16 = add(x = var_4454_cast_fp16, y = var_4458_cast_fp16)[name = string("op_4459_cast_fp16")];
            fp32 var_4460_epsilon_0 = const()[name = string("op_4460_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4460_cast_fp16 = rsqrt(epsilon = var_4460_epsilon_0, x = var_4459_cast_fp16)[name = string("op_4460_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4461_cast_fp16 = mul(x = x_183_cast_fp16, y = var_4460_cast_fp16)[name = string("op_4461_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_4461_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_4470 = const()[name = string("op_4470"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4471_cast_fp16 = reshape(shape = var_4470, x = linear_63_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4478 = const()[name = string("op_4478"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_4471_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 16, 128]> var_4481_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_4481_cast_fp16")];
            tensor<int32, [1]> peak_75_axes_0 = const()[name = string("peak_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_75_cast_fp16 = reduce_max(axes = peak_75_axes_0, keep_dims = var_4478, x = var_4481_cast_fp16)[name = string("peak_75_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_187_cast_fp16 = real_div(x = x_185_cast_fp16, y = peak_75_cast_fp16)[name = string("x_187_cast_fp16")];
            fp16 var_4477_promoted_to_fp16 = const()[name = string("op_4477_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4485_cast_fp16 = pow(x = x_187_cast_fp16, y = var_4477_promoted_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<int32, [1]> var_4487_axes_0 = const()[name = string("op_4487_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4487_keep_dims_0 = const()[name = string("op_4487_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4487_cast_fp16 = reduce_mean(axes = var_4487_axes_0, keep_dims = var_4487_keep_dims_0, x = var_4485_cast_fp16)[name = string("op_4487_cast_fp16")];
            fp16 var_4477_promoted_1_to_fp16 = const()[name = string("op_4477_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4488_cast_fp16 = pow(x = peak_75_cast_fp16, y = var_4477_promoted_1_to_fp16)[name = string("op_4488_cast_fp16")];
            fp32 var_4489_epsilon_0 = const()[name = string("op_4489_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4489_cast_fp16 = inverse(epsilon = var_4489_epsilon_0, x = var_4488_cast_fp16)[name = string("op_4489_cast_fp16")];
            fp16 var_4490_to_fp16 = const()[name = string("op_4490_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4491_cast_fp16 = mul(x = var_4489_cast_fp16, y = var_4490_to_fp16)[name = string("op_4491_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4492_cast_fp16 = add(x = var_4487_cast_fp16, y = var_4491_cast_fp16)[name = string("op_4492_cast_fp16")];
            fp32 var_4493_epsilon_0 = const()[name = string("op_4493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4493_cast_fp16 = rsqrt(epsilon = var_4493_epsilon_0, x = var_4492_cast_fp16)[name = string("op_4493_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4494_cast_fp16 = mul(x = x_187_cast_fp16, y = var_4493_cast_fp16)[name = string("op_4494_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_193_cast_fp16 = mul(x = var_4494_cast_fp16, y = model_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_193_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_4503 = const()[name = string("op_4503"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4504_cast_fp16 = reshape(shape = var_4503, x = linear_64_cast_fp16)[name = string("op_4504_cast_fp16")];
            tensor<int32, [4]> x_189_perm_0 = const()[name = string("x_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4511 = const()[name = string("op_4511"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_189_cast_fp16 = transpose(perm = x_189_perm_0, x = var_4504_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 16, 128]> var_4514_cast_fp16 = abs(x = x_189_cast_fp16)[name = string("op_4514_cast_fp16")];
            tensor<int32, [1]> peak_77_axes_0 = const()[name = string("peak_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_77_cast_fp16 = reduce_max(axes = peak_77_axes_0, keep_dims = var_4511, x = var_4514_cast_fp16)[name = string("peak_77_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_191_cast_fp16 = real_div(x = x_189_cast_fp16, y = peak_77_cast_fp16)[name = string("x_191_cast_fp16")];
            fp16 var_4510_promoted_to_fp16 = const()[name = string("op_4510_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4518_cast_fp16 = pow(x = x_191_cast_fp16, y = var_4510_promoted_to_fp16)[name = string("op_4518_cast_fp16")];
            tensor<int32, [1]> var_4520_axes_0 = const()[name = string("op_4520_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4520_keep_dims_0 = const()[name = string("op_4520_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4520_cast_fp16 = reduce_mean(axes = var_4520_axes_0, keep_dims = var_4520_keep_dims_0, x = var_4518_cast_fp16)[name = string("op_4520_cast_fp16")];
            fp16 var_4510_promoted_1_to_fp16 = const()[name = string("op_4510_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4521_cast_fp16 = pow(x = peak_77_cast_fp16, y = var_4510_promoted_1_to_fp16)[name = string("op_4521_cast_fp16")];
            fp32 var_4522_epsilon_0 = const()[name = string("op_4522_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4522_cast_fp16 = inverse(epsilon = var_4522_epsilon_0, x = var_4521_cast_fp16)[name = string("op_4522_cast_fp16")];
            fp16 var_4523_to_fp16 = const()[name = string("op_4523_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4524_cast_fp16 = mul(x = var_4522_cast_fp16, y = var_4523_to_fp16)[name = string("op_4524_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4525_cast_fp16 = add(x = var_4520_cast_fp16, y = var_4524_cast_fp16)[name = string("op_4525_cast_fp16")];
            fp32 var_4526_epsilon_0 = const()[name = string("op_4526_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4526_cast_fp16 = rsqrt(epsilon = var_4526_epsilon_0, x = var_4525_cast_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4527_cast_fp16 = mul(x = x_191_cast_fp16, y = var_4526_cast_fp16)[name = string("op_4527_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_195_cast_fp16 = mul(x = var_4527_cast_fp16, y = model_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_195_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_4536 = const()[name = string("op_4536"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4537_cast_fp16 = reshape(shape = var_4536, x = linear_65_cast_fp16)[name = string("op_4537_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_193_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_193_cast_fp16)[name = string("second_37_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4551_cast_fp16 = mul(x = first_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4551_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4552_cast_fp16 = mul(x = second_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4552_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4554_cast_fp16 = sub(x = var_4551_cast_fp16, y = var_4552_cast_fp16)[name = string("op_4554_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4555_cast_fp16 = mul(x = second_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4555_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4556_cast_fp16 = mul(x = first_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4558_cast_fp16 = add(x = var_4555_cast_fp16, y = var_4556_cast_fp16)[name = string("op_4558_cast_fp16")];
            int32 var_4560 = const()[name = string("op_4560"), 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_4560, interleave = q_19_interleave_0, values = (var_4554_cast_fp16, var_4558_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_195_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_195_cast_fp16)[name = string("second_39_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4572_cast_fp16 = mul(x = first_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4572_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4573_cast_fp16 = mul(x = second_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4573_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4575_cast_fp16 = sub(x = var_4572_cast_fp16, y = var_4573_cast_fp16)[name = string("op_4575_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4576_cast_fp16 = mul(x = second_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4576_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4577_cast_fp16 = mul(x = first_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4579_cast_fp16 = add(x = var_4576_cast_fp16, y = var_4577_cast_fp16)[name = string("op_4579_cast_fp16")];
            int32 var_4581 = const()[name = string("op_4581"), 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_4581, interleave = k_19_interleave_0, values = (var_4575_cast_fp16, var_4579_cast_fp16))[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_4592_begin_0 = const()[name = string("op_4592_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4592_end_0 = const()[name = string("op_4592_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4592_end_mask_0 = const()[name = string("op_4592_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4592_stride_0 = const()[name = string("op_4592_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4592_cast_fp16 = slice_by_index(begin = var_4592_begin_0, end = var_4592_end_0, end_mask = var_4592_end_mask_0, stride = var_4592_stride_0, x = q_19_cast_fp16)[name = string("op_4592_cast_fp16")];
            tensor<int32, [4]> var_4602_begin_0 = const()[name = string("op_4602_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4602_end_0 = const()[name = string("op_4602_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4602_end_mask_0 = const()[name = string("op_4602_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4602_stride_0 = const()[name = string("op_4602_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4602_cast_fp16 = slice_by_index(begin = var_4602_begin_0, end = var_4602_end_0, end_mask = var_4602_end_mask_0, stride = var_4602_stride_0, x = q_19_cast_fp16)[name = string("op_4602_cast_fp16")];
            int32 var_4604 = const()[name = string("op_4604"), 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_4604, interleave = stacked_19_interleave_0, values = (var_4592_cast_fp16, var_4602_cast_fp16))[name = string("stacked_19_cast_fp16")];
            tensor<int32, [4]> var_4610_begin_0 = const()[name = string("op_4610_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4610_end_0 = const()[name = string("op_4610_end_0"), val = tensor<int32, [4]>([10, 8, 4112, 128])];
            tensor<bool, [4]> var_4610_end_mask_0 = const()[name = string("op_4610_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4610_cast_fp16 = slice_by_index(begin = var_4610_begin_0, end = var_4610_end_0, end_mask = var_4610_end_mask_0, x = read_state_0)[name = string("op_4610_cast_fp16")];
            tensor<int32, [4]> var_4620_begin_0 = const()[name = string("op_4620_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4620_end_0 = const()[name = string("op_4620_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4620_end_mask_0 = const()[name = string("op_4620_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4620_cast_fp16 = slice_by_index(begin = var_4620_begin_0, end = var_4620_end_0, end_mask = var_4620_end_mask_0, x = var_4610_cast_fp16)[name = string("op_4620_cast_fp16")];
            bool var_4629_transpose_x_1 = const()[name = string("op_4629_transpose_x_1"), val = bool(false)];
            bool var_4629_transpose_y_1 = const()[name = string("op_4629_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_4629_cast_fp16 = matmul(transpose_x = var_4629_transpose_x_1, transpose_y = var_4629_transpose_y_1, x = var_4620_cast_fp16, y = stacked_19_cast_fp16)[name = string("op_4629_cast_fp16")];
            tensor<int32, [4]> var_4632_perm_0 = const()[name = string("op_4632_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4633_to_fp16 = const()[name = string("op_4633_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_4632_cast_fp16 = transpose(perm = var_4632_perm_0, x = var_4629_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 8, 32, 4096]> var_4634_cast_fp16 = mul(x = var_4632_cast_fp16, y = var_4633_to_fp16)[name = string("op_4634_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_19_cast_fp16 = add(x = var_4634_cast_fp16, y = cacheMask)[name = string("cache_weights_19_cast_fp16")];
            bool var_4640_transpose_x_1 = const()[name = string("op_4640_transpose_x_1"), val = bool(false)];
            bool var_4640_transpose_y_1 = const()[name = string("op_4640_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4640_cast_fp16 = matmul(transpose_x = var_4640_transpose_x_1, transpose_y = var_4640_transpose_y_1, x = stacked_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_4640_cast_fp16")];
            fp16 var_4641_to_fp16 = const()[name = string("op_4641_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4642_cast_fp16 = mul(x = var_4640_cast_fp16, y = var_4641_to_fp16)[name = string("op_4642_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_19_cast_fp16 = add(x = var_4642_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_19_cast_fp16")];
            int32 var_4646 = const()[name = string("op_4646"), val = int32(-1)];
            bool var_4647_interleave_0 = const()[name = string("op_4647_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4647_cast_fp16 = concat(axis = var_4646, interleave = var_4647_interleave_0, values = (cache_weights_19_cast_fp16, call_weights_19_cast_fp16))[name = string("op_4647_cast_fp16")];
            int32 var_4648 = const()[name = string("op_4648"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_19_cast_fp16 = softmax(axis = var_4648, x = var_4647_cast_fp16)[name = string("shares_19_cast_fp16")];
            tensor<int32, [4]> var_4655_begin_0 = const()[name = string("op_4655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4655_end_0 = const()[name = string("op_4655_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4655_end_mask_0 = const()[name = string("op_4655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4655_cast_fp16 = slice_by_index(begin = var_4655_begin_0, end = var_4655_end_0, end_mask = var_4655_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4655_cast_fp16")];
            tensor<int32, [4]> var_4660_begin_0 = const()[name = string("op_4660_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4660_end_0 = const()[name = string("op_4660_end_0"), val = tensor<int32, [4]>([10, 8, 4112, 128])];
            tensor<bool, [4]> var_4660_end_mask_0 = const()[name = string("op_4660_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4660_cast_fp16 = slice_by_index(begin = var_4660_begin_0, end = var_4660_end_0, end_mask = var_4660_end_mask_0, x = read_state_1)[name = string("op_4660_cast_fp16")];
            tensor<int32, [4]> var_4670_begin_0 = const()[name = string("op_4670_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4670_end_0 = const()[name = string("op_4670_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4670_end_mask_0 = const()[name = string("op_4670_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4670_cast_fp16 = slice_by_index(begin = var_4670_begin_0, end = var_4670_end_0, end_mask = var_4670_end_mask_0, x = var_4660_cast_fp16)[name = string("op_4670_cast_fp16")];
            bool var_4676_transpose_x_0 = const()[name = string("op_4676_transpose_x_0"), val = bool(false)];
            bool var_4676_transpose_y_0 = const()[name = string("op_4676_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4676_cast_fp16 = matmul(transpose_x = var_4676_transpose_x_0, transpose_y = var_4676_transpose_y_0, x = var_4655_cast_fp16, y = var_4670_cast_fp16)[name = string("op_4676_cast_fp16")];
            tensor<int32, [4]> var_4681_begin_0 = const()[name = string("op_4681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_4681_end_0 = const()[name = string("op_4681_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_4681_end_mask_0 = const()[name = string("op_4681_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4681_cast_fp16 = slice_by_index(begin = var_4681_begin_0, end = var_4681_end_0, end_mask = var_4681_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4682_transpose_x_0 = const()[name = string("op_4682_transpose_x_0"), val = bool(false)];
            bool var_4682_transpose_y_0 = const()[name = string("op_4682_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_4537_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 8, 32, 128]> var_4682_cast_fp16 = matmul(transpose_x = var_4682_transpose_x_0, transpose_y = var_4682_transpose_y_0, x = var_4681_cast_fp16, y = v_19_cast_fp16)[name = string("op_4682_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_37_cast_fp16 = add(x = var_4676_cast_fp16, y = var_4682_cast_fp16)[name = string("mixed_37_cast_fp16")];
            tensor<int32, [4]> var_4699_begin_0 = const()[name = string("op_4699_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4699_end_0 = const()[name = string("op_4699_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4699_end_mask_0 = const()[name = string("op_4699_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4699_cast_fp16 = slice_by_index(begin = var_4699_begin_0, end = var_4699_end_0, end_mask = var_4699_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<int32, [4]> var_4714_begin_0 = const()[name = string("op_4714_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4714_end_0 = const()[name = string("op_4714_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4714_end_mask_0 = const()[name = string("op_4714_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4714_cast_fp16 = slice_by_index(begin = var_4714_begin_0, end = var_4714_end_0, end_mask = var_4714_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4714_cast_fp16")];
            int32 var_4716 = const()[name = string("op_4716"), 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_4716, interleave = mixed_39_interleave_0, values = (var_4699_cast_fp16, var_4714_cast_fp16))[name = string("mixed_39_cast_fp16")];
            tensor<int32, [4]> var_4720_perm_0 = const()[name = string("op_4720_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4724 = const()[name = string("op_4724"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4720_cast_fp16 = transpose(perm = var_4720_perm_0, x = mixed_39_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 2048]> input_93_cast_fp16 = reshape(shape = var_4724, x = var_4720_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_197_cast_fp16 = add(x = x_181_cast_fp16, y = linear_66_cast_fp16)[name = string("x_197_cast_fp16")];
            bool var_4734 = const()[name = string("op_4734"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4737_cast_fp16 = abs(x = x_197_cast_fp16)[name = string("op_4737_cast_fp16")];
            tensor<int32, [1]> peak_79_axes_0 = const()[name = string("peak_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_79_cast_fp16 = reduce_max(axes = peak_79_axes_0, keep_dims = var_4734, x = var_4737_cast_fp16)[name = string("peak_79_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_199_cast_fp16 = real_div(x = x_197_cast_fp16, y = peak_79_cast_fp16)[name = string("x_199_cast_fp16")];
            fp16 var_4733_promoted_to_fp16 = const()[name = string("op_4733_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4741_cast_fp16 = pow(x = x_199_cast_fp16, y = var_4733_promoted_to_fp16)[name = string("op_4741_cast_fp16")];
            tensor<int32, [1]> var_4743_axes_0 = const()[name = string("op_4743_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4743_keep_dims_0 = const()[name = string("op_4743_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4743_cast_fp16 = reduce_mean(axes = var_4743_axes_0, keep_dims = var_4743_keep_dims_0, x = var_4741_cast_fp16)[name = string("op_4743_cast_fp16")];
            fp16 var_4733_promoted_1_to_fp16 = const()[name = string("op_4733_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4744_cast_fp16 = pow(x = peak_79_cast_fp16, y = var_4733_promoted_1_to_fp16)[name = string("op_4744_cast_fp16")];
            fp32 var_4745_epsilon_0 = const()[name = string("op_4745_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4745_cast_fp16 = inverse(epsilon = var_4745_epsilon_0, x = var_4744_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4747_cast_fp16 = mul(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("op_4747_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4748_cast_fp16 = add(x = var_4743_cast_fp16, y = var_4747_cast_fp16)[name = string("op_4748_cast_fp16")];
            fp32 var_4749_epsilon_0 = const()[name = string("op_4749_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4749_cast_fp16 = rsqrt(epsilon = var_4749_epsilon_0, x = var_4748_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4750_cast_fp16 = mul(x = x_199_cast_fp16, y = var_4749_cast_fp16)[name = string("op_4750_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_4750_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_4755_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_4755_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_4755_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_201_cast_fp16 = add(x = x_197_cast_fp16, y = linear_69_cast_fp16)[name = string("x_201_cast_fp16")];
            bool var_4768 = const()[name = string("op_4768"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4771_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<int32, [1]> peak_81_axes_0 = const()[name = string("peak_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_81_cast_fp16 = reduce_max(axes = peak_81_axes_0, keep_dims = var_4768, x = var_4771_cast_fp16)[name = string("peak_81_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_203_cast_fp16 = real_div(x = x_201_cast_fp16, y = peak_81_cast_fp16)[name = string("x_203_cast_fp16")];
            fp16 var_4767_promoted_to_fp16 = const()[name = string("op_4767_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4775_cast_fp16 = pow(x = x_203_cast_fp16, y = var_4767_promoted_to_fp16)[name = string("op_4775_cast_fp16")];
            tensor<int32, [1]> var_4777_axes_0 = const()[name = string("op_4777_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4777_keep_dims_0 = const()[name = string("op_4777_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4777_cast_fp16 = reduce_mean(axes = var_4777_axes_0, keep_dims = var_4777_keep_dims_0, x = var_4775_cast_fp16)[name = string("op_4777_cast_fp16")];
            fp16 var_4767_promoted_1_to_fp16 = const()[name = string("op_4767_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4778_cast_fp16 = pow(x = peak_81_cast_fp16, y = var_4767_promoted_1_to_fp16)[name = string("op_4778_cast_fp16")];
            fp32 var_4779_epsilon_0 = const()[name = string("op_4779_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4779_cast_fp16 = inverse(epsilon = var_4779_epsilon_0, x = var_4778_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 var_4780_to_fp16 = const()[name = string("op_4780_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16, y = var_4780_to_fp16)[name = string("op_4781_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4782_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4781_cast_fp16)[name = string("op_4782_cast_fp16")];
            fp32 var_4783_epsilon_0 = const()[name = string("op_4783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4783_cast_fp16 = rsqrt(epsilon = var_4783_epsilon_0, x = var_4782_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4784_cast_fp16 = mul(x = x_203_cast_fp16, y = var_4783_cast_fp16)[name = string("op_4784_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_4784_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_4793 = const()[name = string("op_4793"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4794_cast_fp16 = reshape(shape = var_4793, x = linear_70_cast_fp16)[name = string("op_4794_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4801 = const()[name = string("op_4801"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_4794_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 16, 128]> var_4804_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_4804_cast_fp16")];
            tensor<int32, [1]> peak_83_axes_0 = const()[name = string("peak_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_83_cast_fp16 = reduce_max(axes = peak_83_axes_0, keep_dims = var_4801, x = var_4804_cast_fp16)[name = string("peak_83_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_207_cast_fp16 = real_div(x = x_205_cast_fp16, y = peak_83_cast_fp16)[name = string("x_207_cast_fp16")];
            fp16 var_4800_promoted_to_fp16 = const()[name = string("op_4800_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4808_cast_fp16 = pow(x = x_207_cast_fp16, y = var_4800_promoted_to_fp16)[name = string("op_4808_cast_fp16")];
            tensor<int32, [1]> var_4810_axes_0 = const()[name = string("op_4810_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4810_keep_dims_0 = const()[name = string("op_4810_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4810_cast_fp16 = reduce_mean(axes = var_4810_axes_0, keep_dims = var_4810_keep_dims_0, x = var_4808_cast_fp16)[name = string("op_4810_cast_fp16")];
            fp16 var_4800_promoted_1_to_fp16 = const()[name = string("op_4800_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4811_cast_fp16 = pow(x = peak_83_cast_fp16, y = var_4800_promoted_1_to_fp16)[name = string("op_4811_cast_fp16")];
            fp32 var_4812_epsilon_0 = const()[name = string("op_4812_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4812_cast_fp16 = inverse(epsilon = var_4812_epsilon_0, x = var_4811_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp16 var_4813_to_fp16 = const()[name = string("op_4813_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4814_cast_fp16 = mul(x = var_4812_cast_fp16, y = var_4813_to_fp16)[name = string("op_4814_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4815_cast_fp16 = add(x = var_4810_cast_fp16, y = var_4814_cast_fp16)[name = string("op_4815_cast_fp16")];
            fp32 var_4816_epsilon_0 = const()[name = string("op_4816_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4816_cast_fp16 = rsqrt(epsilon = var_4816_epsilon_0, x = var_4815_cast_fp16)[name = string("op_4816_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4817_cast_fp16 = mul(x = x_207_cast_fp16, y = var_4816_cast_fp16)[name = string("op_4817_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_213_cast_fp16 = mul(x = var_4817_cast_fp16, y = model_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_213_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_4826 = const()[name = string("op_4826"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4827_cast_fp16 = reshape(shape = var_4826, x = linear_71_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [4]> x_209_perm_0 = const()[name = string("x_209_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4834 = const()[name = string("op_4834"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_209_cast_fp16 = transpose(perm = x_209_perm_0, x = var_4827_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 8, 16, 128]> var_4837_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_4837_cast_fp16")];
            tensor<int32, [1]> peak_85_axes_0 = const()[name = string("peak_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_85_cast_fp16 = reduce_max(axes = peak_85_axes_0, keep_dims = var_4834, x = var_4837_cast_fp16)[name = string("peak_85_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_211_cast_fp16 = real_div(x = x_209_cast_fp16, y = peak_85_cast_fp16)[name = string("x_211_cast_fp16")];
            fp16 var_4833_promoted_to_fp16 = const()[name = string("op_4833_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4841_cast_fp16 = pow(x = x_211_cast_fp16, y = var_4833_promoted_to_fp16)[name = string("op_4841_cast_fp16")];
            tensor<int32, [1]> var_4843_axes_0 = const()[name = string("op_4843_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4843_keep_dims_0 = const()[name = string("op_4843_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4843_cast_fp16 = reduce_mean(axes = var_4843_axes_0, keep_dims = var_4843_keep_dims_0, x = var_4841_cast_fp16)[name = string("op_4843_cast_fp16")];
            fp16 var_4833_promoted_1_to_fp16 = const()[name = string("op_4833_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4844_cast_fp16 = pow(x = peak_85_cast_fp16, y = var_4833_promoted_1_to_fp16)[name = string("op_4844_cast_fp16")];
            fp32 var_4845_epsilon_0 = const()[name = string("op_4845_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4845_cast_fp16 = inverse(epsilon = var_4845_epsilon_0, x = var_4844_cast_fp16)[name = string("op_4845_cast_fp16")];
            fp16 var_4846_to_fp16 = const()[name = string("op_4846_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4847_cast_fp16 = mul(x = var_4845_cast_fp16, y = var_4846_to_fp16)[name = string("op_4847_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4848_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4847_cast_fp16)[name = string("op_4848_cast_fp16")];
            fp32 var_4849_epsilon_0 = const()[name = string("op_4849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4849_cast_fp16 = rsqrt(epsilon = var_4849_epsilon_0, x = var_4848_cast_fp16)[name = string("op_4849_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4850_cast_fp16 = mul(x = x_211_cast_fp16, y = var_4849_cast_fp16)[name = string("op_4850_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_215_cast_fp16 = mul(x = var_4850_cast_fp16, y = model_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_215_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_4859 = const()[name = string("op_4859"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4860_cast_fp16 = reshape(shape = var_4859, x = linear_72_cast_fp16)[name = string("op_4860_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_213_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_213_cast_fp16)[name = string("second_41_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4874_cast_fp16 = mul(x = first_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4875_cast_fp16 = mul(x = second_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4875_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4877_cast_fp16 = sub(x = var_4874_cast_fp16, y = var_4875_cast_fp16)[name = string("op_4877_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4878_cast_fp16 = mul(x = second_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4878_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4879_cast_fp16 = mul(x = first_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4879_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4881_cast_fp16 = add(x = var_4878_cast_fp16, y = var_4879_cast_fp16)[name = string("op_4881_cast_fp16")];
            int32 var_4883 = const()[name = string("op_4883"), 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_4883, interleave = q_21_interleave_0, values = (var_4877_cast_fp16, var_4881_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_215_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_215_cast_fp16)[name = string("second_43_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4895_cast_fp16 = mul(x = first_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4896_cast_fp16 = mul(x = second_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4896_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4898_cast_fp16 = sub(x = var_4895_cast_fp16, y = var_4896_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4899_cast_fp16 = mul(x = second_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4899_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4900_cast_fp16 = mul(x = first_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4900_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4902_cast_fp16 = add(x = var_4899_cast_fp16, y = var_4900_cast_fp16)[name = string("op_4902_cast_fp16")];
            int32 var_4904 = const()[name = string("op_4904"), 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_4904, interleave = k_21_interleave_0, values = (var_4898_cast_fp16, var_4902_cast_fp16))[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_4915_begin_0 = const()[name = string("op_4915_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4915_end_0 = const()[name = string("op_4915_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4915_end_mask_0 = const()[name = string("op_4915_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4915_stride_0 = const()[name = string("op_4915_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4915_cast_fp16 = slice_by_index(begin = var_4915_begin_0, end = var_4915_end_0, end_mask = var_4915_end_mask_0, stride = var_4915_stride_0, x = q_21_cast_fp16)[name = string("op_4915_cast_fp16")];
            tensor<int32, [4]> var_4925_begin_0 = const()[name = string("op_4925_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4925_end_0 = const()[name = string("op_4925_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4925_end_mask_0 = const()[name = string("op_4925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4925_stride_0 = const()[name = string("op_4925_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4925_cast_fp16 = slice_by_index(begin = var_4925_begin_0, end = var_4925_end_0, end_mask = var_4925_end_mask_0, stride = var_4925_stride_0, x = q_21_cast_fp16)[name = string("op_4925_cast_fp16")];
            int32 var_4927 = const()[name = string("op_4927"), 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_4927, interleave = stacked_21_interleave_0, values = (var_4915_cast_fp16, var_4925_cast_fp16))[name = string("stacked_21_cast_fp16")];
            tensor<int32, [4]> var_4933_begin_0 = const()[name = string("op_4933_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4933_end_0 = const()[name = string("op_4933_end_0"), val = tensor<int32, [4]>([11, 8, 4112, 128])];
            tensor<bool, [4]> var_4933_end_mask_0 = const()[name = string("op_4933_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4933_cast_fp16 = slice_by_index(begin = var_4933_begin_0, end = var_4933_end_0, end_mask = var_4933_end_mask_0, x = read_state_0)[name = string("op_4933_cast_fp16")];
            tensor<int32, [4]> var_4943_begin_0 = const()[name = string("op_4943_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4943_end_0 = const()[name = string("op_4943_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4943_end_mask_0 = const()[name = string("op_4943_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4943_cast_fp16 = slice_by_index(begin = var_4943_begin_0, end = var_4943_end_0, end_mask = var_4943_end_mask_0, x = var_4933_cast_fp16)[name = string("op_4943_cast_fp16")];
            bool var_4952_transpose_x_1 = const()[name = string("op_4952_transpose_x_1"), val = bool(false)];
            bool var_4952_transpose_y_1 = const()[name = string("op_4952_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_4952_cast_fp16 = matmul(transpose_x = var_4952_transpose_x_1, transpose_y = var_4952_transpose_y_1, x = var_4943_cast_fp16, y = stacked_21_cast_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4955_perm_0 = const()[name = string("op_4955_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4956_to_fp16 = const()[name = string("op_4956_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_4955_cast_fp16 = transpose(perm = var_4955_perm_0, x = var_4952_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 32, 4096]> var_4957_cast_fp16 = mul(x = var_4955_cast_fp16, y = var_4956_to_fp16)[name = string("op_4957_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_21_cast_fp16 = add(x = var_4957_cast_fp16, y = cacheMask)[name = string("cache_weights_21_cast_fp16")];
            bool var_4963_transpose_x_1 = const()[name = string("op_4963_transpose_x_1"), val = bool(false)];
            bool var_4963_transpose_y_1 = const()[name = string("op_4963_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4963_cast_fp16 = matmul(transpose_x = var_4963_transpose_x_1, transpose_y = var_4963_transpose_y_1, x = stacked_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_4963_cast_fp16")];
            fp16 var_4964_to_fp16 = const()[name = string("op_4964_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4965_cast_fp16 = mul(x = var_4963_cast_fp16, y = var_4964_to_fp16)[name = string("op_4965_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_21_cast_fp16 = add(x = var_4965_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_21_cast_fp16")];
            int32 var_4969 = const()[name = string("op_4969"), val = int32(-1)];
            bool var_4970_interleave_0 = const()[name = string("op_4970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4970_cast_fp16 = concat(axis = var_4969, interleave = var_4970_interleave_0, values = (cache_weights_21_cast_fp16, call_weights_21_cast_fp16))[name = string("op_4970_cast_fp16")];
            int32 var_4971 = const()[name = string("op_4971"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_21_cast_fp16 = softmax(axis = var_4971, x = var_4970_cast_fp16)[name = string("shares_21_cast_fp16")];
            tensor<int32, [4]> var_4978_begin_0 = const()[name = string("op_4978_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4978_end_0 = const()[name = string("op_4978_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4978_end_mask_0 = const()[name = string("op_4978_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4978_cast_fp16 = slice_by_index(begin = var_4978_begin_0, end = var_4978_end_0, end_mask = var_4978_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4978_cast_fp16")];
            tensor<int32, [4]> var_4983_begin_0 = const()[name = string("op_4983_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4983_end_0 = const()[name = string("op_4983_end_0"), val = tensor<int32, [4]>([11, 8, 4112, 128])];
            tensor<bool, [4]> var_4983_end_mask_0 = const()[name = string("op_4983_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4983_cast_fp16 = slice_by_index(begin = var_4983_begin_0, end = var_4983_end_0, end_mask = var_4983_end_mask_0, x = read_state_1)[name = string("op_4983_cast_fp16")];
            tensor<int32, [4]> var_4993_begin_0 = const()[name = string("op_4993_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4993_end_0 = const()[name = string("op_4993_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4993_end_mask_0 = const()[name = string("op_4993_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4993_cast_fp16 = slice_by_index(begin = var_4993_begin_0, end = var_4993_end_0, end_mask = var_4993_end_mask_0, x = var_4983_cast_fp16)[name = string("op_4993_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, 32, 128]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = var_4978_cast_fp16, y = var_4993_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5004_begin_0 = const()[name = string("op_5004_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_5004_end_0 = const()[name = string("op_5004_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_5004_end_mask_0 = const()[name = string("op_5004_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5004_cast_fp16 = slice_by_index(begin = var_5004_begin_0, end = var_5004_end_0, end_mask = var_5004_end_mask_0, x = shares_21_cast_fp16)[name = string("op_5004_cast_fp16")];
            bool var_5005_transpose_x_0 = const()[name = string("op_5005_transpose_x_0"), val = bool(false)];
            bool var_5005_transpose_y_0 = const()[name = string("op_5005_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_4860_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 32, 128]> var_5005_cast_fp16 = matmul(transpose_x = var_5005_transpose_x_0, transpose_y = var_5005_transpose_y_0, x = var_5004_cast_fp16, y = v_21_cast_fp16)[name = string("op_5005_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_41_cast_fp16 = add(x = var_4999_cast_fp16, y = var_5005_cast_fp16)[name = string("mixed_41_cast_fp16")];
            tensor<int32, [4]> var_5022_begin_0 = const()[name = string("op_5022_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5022_end_0 = const()[name = string("op_5022_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5022_end_mask_0 = const()[name = string("op_5022_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5022_cast_fp16 = slice_by_index(begin = var_5022_begin_0, end = var_5022_end_0, end_mask = var_5022_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_5022_cast_fp16")];
            tensor<int32, [4]> var_5037_begin_0 = const()[name = string("op_5037_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5037_end_0 = const()[name = string("op_5037_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5037_end_mask_0 = const()[name = string("op_5037_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5037_cast_fp16 = slice_by_index(begin = var_5037_begin_0, end = var_5037_end_0, end_mask = var_5037_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_5037_cast_fp16")];
            int32 var_5039 = const()[name = string("op_5039"), 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_5039, interleave = mixed_43_interleave_0, values = (var_5022_cast_fp16, var_5037_cast_fp16))[name = string("mixed_43_cast_fp16")];
            tensor<int32, [4]> var_5043_perm_0 = const()[name = string("op_5043_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5047 = const()[name = string("op_5047"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5043_cast_fp16 = transpose(perm = var_5043_perm_0, x = mixed_43_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 2048]> input_103_cast_fp16 = reshape(shape = var_5047, x = var_5043_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_217_cast_fp16 = add(x = x_201_cast_fp16, y = linear_73_cast_fp16)[name = string("x_217_cast_fp16")];
            bool var_5057 = const()[name = string("op_5057"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5060_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<int32, [1]> peak_87_axes_0 = const()[name = string("peak_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_87_cast_fp16 = reduce_max(axes = peak_87_axes_0, keep_dims = var_5057, x = var_5060_cast_fp16)[name = string("peak_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_219_cast_fp16 = real_div(x = x_217_cast_fp16, y = peak_87_cast_fp16)[name = string("x_219_cast_fp16")];
            fp16 var_5056_promoted_to_fp16 = const()[name = string("op_5056_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5064_cast_fp16 = pow(x = x_219_cast_fp16, y = var_5056_promoted_to_fp16)[name = string("op_5064_cast_fp16")];
            tensor<int32, [1]> var_5066_axes_0 = const()[name = string("op_5066_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5066_keep_dims_0 = const()[name = string("op_5066_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5066_cast_fp16 = reduce_mean(axes = var_5066_axes_0, keep_dims = var_5066_keep_dims_0, x = var_5064_cast_fp16)[name = string("op_5066_cast_fp16")];
            fp16 var_5056_promoted_1_to_fp16 = const()[name = string("op_5056_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5067_cast_fp16 = pow(x = peak_87_cast_fp16, y = var_5056_promoted_1_to_fp16)[name = string("op_5067_cast_fp16")];
            fp32 var_5068_epsilon_0 = const()[name = string("op_5068_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5068_cast_fp16 = inverse(epsilon = var_5068_epsilon_0, x = var_5067_cast_fp16)[name = string("op_5068_cast_fp16")];
            fp16 var_5069_to_fp16 = const()[name = string("op_5069_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5070_cast_fp16 = mul(x = var_5068_cast_fp16, y = var_5069_to_fp16)[name = string("op_5070_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5071_cast_fp16 = add(x = var_5066_cast_fp16, y = var_5070_cast_fp16)[name = string("op_5071_cast_fp16")];
            fp32 var_5072_epsilon_0 = const()[name = string("op_5072_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5072_cast_fp16 = rsqrt(epsilon = var_5072_epsilon_0, x = var_5071_cast_fp16)[name = string("op_5072_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5073_cast_fp16 = mul(x = x_219_cast_fp16, y = var_5072_cast_fp16)[name = string("op_5073_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_5073_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_5078_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_5078_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_5078_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_221_cast_fp16 = add(x = x_217_cast_fp16, y = linear_76_cast_fp16)[name = string("x_221_cast_fp16")];
            bool var_5091 = const()[name = string("op_5091"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5094_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [1]> peak_89_axes_0 = const()[name = string("peak_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_89_cast_fp16 = reduce_max(axes = peak_89_axes_0, keep_dims = var_5091, x = var_5094_cast_fp16)[name = string("peak_89_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_223_cast_fp16 = real_div(x = x_221_cast_fp16, y = peak_89_cast_fp16)[name = string("x_223_cast_fp16")];
            fp16 var_5090_promoted_to_fp16 = const()[name = string("op_5090_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5098_cast_fp16 = pow(x = x_223_cast_fp16, y = var_5090_promoted_to_fp16)[name = string("op_5098_cast_fp16")];
            tensor<int32, [1]> var_5100_axes_0 = const()[name = string("op_5100_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5100_keep_dims_0 = const()[name = string("op_5100_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5100_cast_fp16 = reduce_mean(axes = var_5100_axes_0, keep_dims = var_5100_keep_dims_0, x = var_5098_cast_fp16)[name = string("op_5100_cast_fp16")];
            fp16 var_5090_promoted_1_to_fp16 = const()[name = string("op_5090_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5101_cast_fp16 = pow(x = peak_89_cast_fp16, y = var_5090_promoted_1_to_fp16)[name = string("op_5101_cast_fp16")];
            fp32 var_5102_epsilon_0 = const()[name = string("op_5102_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5102_cast_fp16 = inverse(epsilon = var_5102_epsilon_0, x = var_5101_cast_fp16)[name = string("op_5102_cast_fp16")];
            fp16 var_5103_to_fp16 = const()[name = string("op_5103_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5104_cast_fp16 = mul(x = var_5102_cast_fp16, y = var_5103_to_fp16)[name = string("op_5104_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5105_cast_fp16 = add(x = var_5100_cast_fp16, y = var_5104_cast_fp16)[name = string("op_5105_cast_fp16")];
            fp32 var_5106_epsilon_0 = const()[name = string("op_5106_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5106_cast_fp16 = rsqrt(epsilon = var_5106_epsilon_0, x = var_5105_cast_fp16)[name = string("op_5106_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5107_cast_fp16 = mul(x = x_223_cast_fp16, y = var_5106_cast_fp16)[name = string("op_5107_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_5107_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_5116 = const()[name = string("op_5116"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5117_cast_fp16 = reshape(shape = var_5116, x = linear_77_cast_fp16)[name = string("op_5117_cast_fp16")];
            tensor<int32, [4]> x_225_perm_0 = const()[name = string("x_225_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5124 = const()[name = string("op_5124"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_225_cast_fp16 = transpose(perm = x_225_perm_0, x = var_5117_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 16, 16, 128]> var_5127_cast_fp16 = abs(x = x_225_cast_fp16)[name = string("op_5127_cast_fp16")];
            tensor<int32, [1]> peak_91_axes_0 = const()[name = string("peak_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_91_cast_fp16 = reduce_max(axes = peak_91_axes_0, keep_dims = var_5124, x = var_5127_cast_fp16)[name = string("peak_91_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_227_cast_fp16 = real_div(x = x_225_cast_fp16, y = peak_91_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_5123_promoted_to_fp16 = const()[name = string("op_5123_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5131_cast_fp16 = pow(x = x_227_cast_fp16, y = var_5123_promoted_to_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5133_axes_0 = const()[name = string("op_5133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5133_keep_dims_0 = const()[name = string("op_5133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5133_cast_fp16 = reduce_mean(axes = var_5133_axes_0, keep_dims = var_5133_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5133_cast_fp16")];
            fp16 var_5123_promoted_1_to_fp16 = const()[name = string("op_5123_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5134_cast_fp16 = pow(x = peak_91_cast_fp16, y = var_5123_promoted_1_to_fp16)[name = string("op_5134_cast_fp16")];
            fp32 var_5135_epsilon_0 = const()[name = string("op_5135_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5135_cast_fp16 = inverse(epsilon = var_5135_epsilon_0, x = var_5134_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, 16, 16, 1]> var_5137_cast_fp16 = mul(x = var_5135_cast_fp16, y = var_5136_to_fp16)[name = string("op_5137_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5138_cast_fp16 = add(x = var_5133_cast_fp16, y = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp32 var_5139_epsilon_0 = const()[name = string("op_5139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5139_cast_fp16 = rsqrt(epsilon = var_5139_epsilon_0, x = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5140_cast_fp16 = mul(x = x_227_cast_fp16, y = var_5139_cast_fp16)[name = string("op_5140_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_233_cast_fp16 = mul(x = var_5140_cast_fp16, y = model_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_233_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_5149 = const()[name = string("op_5149"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5150_cast_fp16 = reshape(shape = var_5149, x = linear_78_cast_fp16)[name = string("op_5150_cast_fp16")];
            tensor<int32, [4]> x_229_perm_0 = const()[name = string("x_229_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5157 = const()[name = string("op_5157"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_229_cast_fp16 = transpose(perm = x_229_perm_0, x = var_5150_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 8, 16, 128]> var_5160_cast_fp16 = abs(x = x_229_cast_fp16)[name = string("op_5160_cast_fp16")];
            tensor<int32, [1]> peak_93_axes_0 = const()[name = string("peak_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_93_cast_fp16 = reduce_max(axes = peak_93_axes_0, keep_dims = var_5157, x = var_5160_cast_fp16)[name = string("peak_93_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_231_cast_fp16 = real_div(x = x_229_cast_fp16, y = peak_93_cast_fp16)[name = string("x_231_cast_fp16")];
            fp16 var_5156_promoted_to_fp16 = const()[name = string("op_5156_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5164_cast_fp16 = pow(x = x_231_cast_fp16, y = var_5156_promoted_to_fp16)[name = string("op_5164_cast_fp16")];
            tensor<int32, [1]> var_5166_axes_0 = const()[name = string("op_5166_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5166_keep_dims_0 = const()[name = string("op_5166_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5166_cast_fp16 = reduce_mean(axes = var_5166_axes_0, keep_dims = var_5166_keep_dims_0, x = var_5164_cast_fp16)[name = string("op_5166_cast_fp16")];
            fp16 var_5156_promoted_1_to_fp16 = const()[name = string("op_5156_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5167_cast_fp16 = pow(x = peak_93_cast_fp16, y = var_5156_promoted_1_to_fp16)[name = string("op_5167_cast_fp16")];
            fp32 var_5168_epsilon_0 = const()[name = string("op_5168_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5168_cast_fp16 = inverse(epsilon = var_5168_epsilon_0, x = var_5167_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp16 var_5169_to_fp16 = const()[name = string("op_5169_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5170_cast_fp16 = mul(x = var_5168_cast_fp16, y = var_5169_to_fp16)[name = string("op_5170_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5171_cast_fp16 = add(x = var_5166_cast_fp16, y = var_5170_cast_fp16)[name = string("op_5171_cast_fp16")];
            fp32 var_5172_epsilon_0 = const()[name = string("op_5172_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5172_cast_fp16 = rsqrt(epsilon = var_5172_epsilon_0, x = var_5171_cast_fp16)[name = string("op_5172_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5173_cast_fp16 = mul(x = x_231_cast_fp16, y = var_5172_cast_fp16)[name = string("op_5173_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_235_cast_fp16 = mul(x = var_5173_cast_fp16, y = model_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_235_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_5182 = const()[name = string("op_5182"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5183_cast_fp16 = reshape(shape = var_5182, x = linear_79_cast_fp16)[name = string("op_5183_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_233_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_233_cast_fp16)[name = string("second_45_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5197_cast_fp16 = mul(x = first_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5197_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5198_cast_fp16 = mul(x = second_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5198_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5200_cast_fp16 = sub(x = var_5197_cast_fp16, y = var_5198_cast_fp16)[name = string("op_5200_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5201_cast_fp16 = mul(x = second_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5201_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5202_cast_fp16 = mul(x = first_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5202_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5204_cast_fp16 = add(x = var_5201_cast_fp16, y = var_5202_cast_fp16)[name = string("op_5204_cast_fp16")];
            int32 var_5206 = const()[name = string("op_5206"), 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_5206, interleave = q_23_interleave_0, values = (var_5200_cast_fp16, var_5204_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_235_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_235_cast_fp16)[name = string("second_47_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5218_cast_fp16 = mul(x = first_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5218_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5219_cast_fp16 = mul(x = second_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5219_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5221_cast_fp16 = sub(x = var_5218_cast_fp16, y = var_5219_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5222_cast_fp16 = mul(x = second_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5222_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5223_cast_fp16 = mul(x = first_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5225_cast_fp16 = add(x = var_5222_cast_fp16, y = var_5223_cast_fp16)[name = string("op_5225_cast_fp16")];
            int32 var_5227 = const()[name = string("op_5227"), 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_5227, interleave = k_23_interleave_0, values = (var_5221_cast_fp16, var_5225_cast_fp16))[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> var_5238_begin_0 = const()[name = string("op_5238_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5238_end_0 = const()[name = string("op_5238_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5238_end_mask_0 = const()[name = string("op_5238_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5238_stride_0 = const()[name = string("op_5238_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, stride = var_5238_stride_0, x = q_23_cast_fp16)[name = string("op_5238_cast_fp16")];
            tensor<int32, [4]> var_5248_begin_0 = const()[name = string("op_5248_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5248_end_0 = const()[name = string("op_5248_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5248_end_mask_0 = const()[name = string("op_5248_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5248_stride_0 = const()[name = string("op_5248_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5248_cast_fp16 = slice_by_index(begin = var_5248_begin_0, end = var_5248_end_0, end_mask = var_5248_end_mask_0, stride = var_5248_stride_0, x = q_23_cast_fp16)[name = string("op_5248_cast_fp16")];
            int32 var_5250 = const()[name = string("op_5250"), 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_5250, interleave = stacked_23_interleave_0, values = (var_5238_cast_fp16, var_5248_cast_fp16))[name = string("stacked_23_cast_fp16")];
            tensor<int32, [4]> var_5256_begin_0 = const()[name = string("op_5256_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_5256_end_0 = const()[name = string("op_5256_end_0"), val = tensor<int32, [4]>([12, 8, 4112, 128])];
            tensor<bool, [4]> var_5256_end_mask_0 = const()[name = string("op_5256_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5256_cast_fp16 = slice_by_index(begin = var_5256_begin_0, end = var_5256_end_0, end_mask = var_5256_end_mask_0, x = read_state_0)[name = string("op_5256_cast_fp16")];
            tensor<int32, [4]> var_5266_begin_0 = const()[name = string("op_5266_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5266_end_0 = const()[name = string("op_5266_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5266_end_mask_0 = const()[name = string("op_5266_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5266_cast_fp16 = slice_by_index(begin = var_5266_begin_0, end = var_5266_end_0, end_mask = var_5266_end_mask_0, x = var_5256_cast_fp16)[name = string("op_5266_cast_fp16")];
            bool var_5275_transpose_x_1 = const()[name = string("op_5275_transpose_x_1"), val = bool(false)];
            bool var_5275_transpose_y_1 = const()[name = string("op_5275_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_5275_cast_fp16 = matmul(transpose_x = var_5275_transpose_x_1, transpose_y = var_5275_transpose_y_1, x = var_5266_cast_fp16, y = stacked_23_cast_fp16)[name = string("op_5275_cast_fp16")];
            tensor<int32, [4]> var_5278_perm_0 = const()[name = string("op_5278_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5279_to_fp16 = const()[name = string("op_5279_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_5278_cast_fp16 = transpose(perm = var_5278_perm_0, x = var_5275_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, 32, 4096]> var_5280_cast_fp16 = mul(x = var_5278_cast_fp16, y = var_5279_to_fp16)[name = string("op_5280_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_23_cast_fp16 = add(x = var_5280_cast_fp16, y = cacheMask)[name = string("cache_weights_23_cast_fp16")];
            bool var_5286_transpose_x_1 = const()[name = string("op_5286_transpose_x_1"), val = bool(false)];
            bool var_5286_transpose_y_1 = const()[name = string("op_5286_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5286_cast_fp16 = matmul(transpose_x = var_5286_transpose_x_1, transpose_y = var_5286_transpose_y_1, x = stacked_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_5286_cast_fp16")];
            fp16 var_5287_to_fp16 = const()[name = string("op_5287_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5288_cast_fp16 = mul(x = var_5286_cast_fp16, y = var_5287_to_fp16)[name = string("op_5288_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_23_cast_fp16 = add(x = var_5288_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_23_cast_fp16")];
            int32 var_5292 = const()[name = string("op_5292"), val = int32(-1)];
            bool var_5293_interleave_0 = const()[name = string("op_5293_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_5293_cast_fp16 = concat(axis = var_5292, interleave = var_5293_interleave_0, values = (cache_weights_23_cast_fp16, call_weights_23_cast_fp16))[name = string("op_5293_cast_fp16")];
            int32 var_5294 = const()[name = string("op_5294"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_23_cast_fp16 = softmax(axis = var_5294, x = var_5293_cast_fp16)[name = string("shares_23_cast_fp16")];
            tensor<int32, [4]> var_5301_begin_0 = const()[name = string("op_5301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5301_end_0 = const()[name = string("op_5301_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_5301_end_mask_0 = const()[name = string("op_5301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_5301_cast_fp16 = slice_by_index(begin = var_5301_begin_0, end = var_5301_end_0, end_mask = var_5301_end_mask_0, x = shares_23_cast_fp16)[name = string("op_5301_cast_fp16")];
            tensor<int32, [4]> var_5306_begin_0 = const()[name = string("op_5306_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_5306_end_0 = const()[name = string("op_5306_end_0"), val = tensor<int32, [4]>([12, 8, 4112, 128])];
            tensor<bool, [4]> var_5306_end_mask_0 = const()[name = string("op_5306_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5306_cast_fp16 = slice_by_index(begin = var_5306_begin_0, end = var_5306_end_0, end_mask = var_5306_end_mask_0, x = read_state_1)[name = string("op_5306_cast_fp16")];
            tensor<int32, [4]> var_5316_begin_0 = const()[name = string("op_5316_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5316_end_0 = const()[name = string("op_5316_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5316_end_mask_0 = const()[name = string("op_5316_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5316_cast_fp16 = slice_by_index(begin = var_5316_begin_0, end = var_5316_end_0, end_mask = var_5316_end_mask_0, x = var_5306_cast_fp16)[name = string("op_5316_cast_fp16")];
            bool var_5322_transpose_x_0 = const()[name = string("op_5322_transpose_x_0"), val = bool(false)];
            bool var_5322_transpose_y_0 = const()[name = string("op_5322_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5322_cast_fp16 = matmul(transpose_x = var_5322_transpose_x_0, transpose_y = var_5322_transpose_y_0, x = var_5301_cast_fp16, y = var_5316_cast_fp16)[name = string("op_5322_cast_fp16")];
            tensor<int32, [4]> var_5327_begin_0 = const()[name = string("op_5327_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_5327_end_0 = const()[name = string("op_5327_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_5327_end_mask_0 = const()[name = string("op_5327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5327_cast_fp16 = slice_by_index(begin = var_5327_begin_0, end = var_5327_end_0, end_mask = var_5327_end_mask_0, x = shares_23_cast_fp16)[name = string("op_5327_cast_fp16")];
            bool var_5328_transpose_x_0 = const()[name = string("op_5328_transpose_x_0"), val = bool(false)];
            bool var_5328_transpose_y_0 = const()[name = string("op_5328_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_5183_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 8, 32, 128]> var_5328_cast_fp16 = matmul(transpose_x = var_5328_transpose_x_0, transpose_y = var_5328_transpose_y_0, x = var_5327_cast_fp16, y = v_23_cast_fp16)[name = string("op_5328_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_45_cast_fp16 = add(x = var_5322_cast_fp16, y = var_5328_cast_fp16)[name = string("mixed_45_cast_fp16")];
            tensor<int32, [4]> var_5345_begin_0 = const()[name = string("op_5345_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5345_end_0 = const()[name = string("op_5345_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5345_end_mask_0 = const()[name = string("op_5345_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5345_cast_fp16 = slice_by_index(begin = var_5345_begin_0, end = var_5345_end_0, end_mask = var_5345_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5345_cast_fp16")];
            tensor<int32, [4]> var_5360_begin_0 = const()[name = string("op_5360_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5360_end_0 = const()[name = string("op_5360_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5360_end_mask_0 = const()[name = string("op_5360_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5360_cast_fp16 = slice_by_index(begin = var_5360_begin_0, end = var_5360_end_0, end_mask = var_5360_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5360_cast_fp16")];
            int32 var_5362 = const()[name = string("op_5362"), 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_5362, interleave = mixed_47_interleave_0, values = (var_5345_cast_fp16, var_5360_cast_fp16))[name = string("mixed_47_cast_fp16")];
            tensor<int32, [4]> var_5366_perm_0 = const()[name = string("op_5366_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5370 = const()[name = string("op_5370"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5366_cast_fp16 = transpose(perm = var_5366_perm_0, x = mixed_47_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 16, 2048]> input_113_cast_fp16 = reshape(shape = var_5370, x = var_5366_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_237_cast_fp16 = add(x = x_221_cast_fp16, y = linear_80_cast_fp16)[name = string("x_237_cast_fp16")];
            bool var_5380 = const()[name = string("op_5380"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5383_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_5383_cast_fp16")];
            tensor<int32, [1]> peak_95_axes_0 = const()[name = string("peak_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_95_cast_fp16 = reduce_max(axes = peak_95_axes_0, keep_dims = var_5380, x = var_5383_cast_fp16)[name = string("peak_95_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_239_cast_fp16 = real_div(x = x_237_cast_fp16, y = peak_95_cast_fp16)[name = string("x_239_cast_fp16")];
            fp16 var_5379_promoted_to_fp16 = const()[name = string("op_5379_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5387_cast_fp16 = pow(x = x_239_cast_fp16, y = var_5379_promoted_to_fp16)[name = string("op_5387_cast_fp16")];
            tensor<int32, [1]> var_5389_axes_0 = const()[name = string("op_5389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5389_keep_dims_0 = const()[name = string("op_5389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5389_cast_fp16 = reduce_mean(axes = var_5389_axes_0, keep_dims = var_5389_keep_dims_0, x = var_5387_cast_fp16)[name = string("op_5389_cast_fp16")];
            fp16 var_5379_promoted_1_to_fp16 = const()[name = string("op_5379_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5390_cast_fp16 = pow(x = peak_95_cast_fp16, y = var_5379_promoted_1_to_fp16)[name = string("op_5390_cast_fp16")];
            fp32 var_5391_epsilon_0 = const()[name = string("op_5391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5391_cast_fp16 = inverse(epsilon = var_5391_epsilon_0, x = var_5390_cast_fp16)[name = string("op_5391_cast_fp16")];
            fp16 var_5392_to_fp16 = const()[name = string("op_5392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5393_cast_fp16 = mul(x = var_5391_cast_fp16, y = var_5392_to_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5394_cast_fp16 = add(x = var_5389_cast_fp16, y = var_5393_cast_fp16)[name = string("op_5394_cast_fp16")];
            fp32 var_5395_epsilon_0 = const()[name = string("op_5395_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5395_cast_fp16 = rsqrt(epsilon = var_5395_epsilon_0, x = var_5394_cast_fp16)[name = string("op_5395_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5396_cast_fp16 = mul(x = x_239_cast_fp16, y = var_5395_cast_fp16)[name = string("op_5396_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_5396_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_5401_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_5401_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_5401_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_241_cast_fp16 = add(x = x_237_cast_fp16, y = linear_83_cast_fp16)[name = string("x_241_cast_fp16")];
            bool var_5414 = const()[name = string("op_5414"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5417_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_5417_cast_fp16")];
            tensor<int32, [1]> peak_97_axes_0 = const()[name = string("peak_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_97_cast_fp16 = reduce_max(axes = peak_97_axes_0, keep_dims = var_5414, x = var_5417_cast_fp16)[name = string("peak_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_243_cast_fp16 = real_div(x = x_241_cast_fp16, y = peak_97_cast_fp16)[name = string("x_243_cast_fp16")];
            fp16 var_5413_promoted_to_fp16 = const()[name = string("op_5413_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5421_cast_fp16 = pow(x = x_243_cast_fp16, y = var_5413_promoted_to_fp16)[name = string("op_5421_cast_fp16")];
            tensor<int32, [1]> var_5423_axes_0 = const()[name = string("op_5423_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5423_keep_dims_0 = const()[name = string("op_5423_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5423_cast_fp16 = reduce_mean(axes = var_5423_axes_0, keep_dims = var_5423_keep_dims_0, x = var_5421_cast_fp16)[name = string("op_5423_cast_fp16")];
            fp16 var_5413_promoted_1_to_fp16 = const()[name = string("op_5413_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5424_cast_fp16 = pow(x = peak_97_cast_fp16, y = var_5413_promoted_1_to_fp16)[name = string("op_5424_cast_fp16")];
            fp32 var_5425_epsilon_0 = const()[name = string("op_5425_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5425_cast_fp16 = inverse(epsilon = var_5425_epsilon_0, x = var_5424_cast_fp16)[name = string("op_5425_cast_fp16")];
            fp16 var_5426_to_fp16 = const()[name = string("op_5426_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5427_cast_fp16 = mul(x = var_5425_cast_fp16, y = var_5426_to_fp16)[name = string("op_5427_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5428_cast_fp16 = add(x = var_5423_cast_fp16, y = var_5427_cast_fp16)[name = string("op_5428_cast_fp16")];
            fp32 var_5429_epsilon_0 = const()[name = string("op_5429_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5429_cast_fp16 = rsqrt(epsilon = var_5429_epsilon_0, x = var_5428_cast_fp16)[name = string("op_5429_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5430_cast_fp16 = mul(x = x_243_cast_fp16, y = var_5429_cast_fp16)[name = string("op_5430_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_5430_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_5439 = const()[name = string("op_5439"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5440_cast_fp16 = reshape(shape = var_5439, x = linear_84_cast_fp16)[name = string("op_5440_cast_fp16")];
            tensor<int32, [4]> x_245_perm_0 = const()[name = string("x_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5447 = const()[name = string("op_5447"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_245_cast_fp16 = transpose(perm = x_245_perm_0, x = var_5440_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 16, 128]> var_5450_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_5450_cast_fp16")];
            tensor<int32, [1]> peak_99_axes_0 = const()[name = string("peak_99_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_99_cast_fp16 = reduce_max(axes = peak_99_axes_0, keep_dims = var_5447, x = var_5450_cast_fp16)[name = string("peak_99_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_247_cast_fp16 = real_div(x = x_245_cast_fp16, y = peak_99_cast_fp16)[name = string("x_247_cast_fp16")];
            fp16 var_5446_promoted_to_fp16 = const()[name = string("op_5446_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5454_cast_fp16 = pow(x = x_247_cast_fp16, y = var_5446_promoted_to_fp16)[name = string("op_5454_cast_fp16")];
            tensor<int32, [1]> var_5456_axes_0 = const()[name = string("op_5456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5456_keep_dims_0 = const()[name = string("op_5456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5456_cast_fp16 = reduce_mean(axes = var_5456_axes_0, keep_dims = var_5456_keep_dims_0, x = var_5454_cast_fp16)[name = string("op_5456_cast_fp16")];
            fp16 var_5446_promoted_1_to_fp16 = const()[name = string("op_5446_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5457_cast_fp16 = pow(x = peak_99_cast_fp16, y = var_5446_promoted_1_to_fp16)[name = string("op_5457_cast_fp16")];
            fp32 var_5458_epsilon_0 = const()[name = string("op_5458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5458_cast_fp16 = inverse(epsilon = var_5458_epsilon_0, x = var_5457_cast_fp16)[name = string("op_5458_cast_fp16")];
            fp16 var_5459_to_fp16 = const()[name = string("op_5459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5460_cast_fp16 = mul(x = var_5458_cast_fp16, y = var_5459_to_fp16)[name = string("op_5460_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5461_cast_fp16 = add(x = var_5456_cast_fp16, y = var_5460_cast_fp16)[name = string("op_5461_cast_fp16")];
            fp32 var_5462_epsilon_0 = const()[name = string("op_5462_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5462_cast_fp16 = rsqrt(epsilon = var_5462_epsilon_0, x = var_5461_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5463_cast_fp16 = mul(x = x_247_cast_fp16, y = var_5462_cast_fp16)[name = string("op_5463_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_253_cast_fp16 = mul(x = var_5463_cast_fp16, y = model_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_253_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_5472 = const()[name = string("op_5472"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5473_cast_fp16 = reshape(shape = var_5472, x = linear_85_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<int32, [4]> x_249_perm_0 = const()[name = string("x_249_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5480 = const()[name = string("op_5480"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_249_cast_fp16 = transpose(perm = x_249_perm_0, x = var_5473_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 16, 128]> var_5483_cast_fp16 = abs(x = x_249_cast_fp16)[name = string("op_5483_cast_fp16")];
            tensor<int32, [1]> peak_101_axes_0 = const()[name = string("peak_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_101_cast_fp16 = reduce_max(axes = peak_101_axes_0, keep_dims = var_5480, x = var_5483_cast_fp16)[name = string("peak_101_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_251_cast_fp16 = real_div(x = x_249_cast_fp16, y = peak_101_cast_fp16)[name = string("x_251_cast_fp16")];
            fp16 var_5479_promoted_to_fp16 = const()[name = string("op_5479_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5487_cast_fp16 = pow(x = x_251_cast_fp16, y = var_5479_promoted_to_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5489_axes_0 = const()[name = string("op_5489_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5489_keep_dims_0 = const()[name = string("op_5489_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5489_cast_fp16 = reduce_mean(axes = var_5489_axes_0, keep_dims = var_5489_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5489_cast_fp16")];
            fp16 var_5479_promoted_1_to_fp16 = const()[name = string("op_5479_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5490_cast_fp16 = pow(x = peak_101_cast_fp16, y = var_5479_promoted_1_to_fp16)[name = string("op_5490_cast_fp16")];
            fp32 var_5491_epsilon_0 = const()[name = string("op_5491_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5491_cast_fp16 = inverse(epsilon = var_5491_epsilon_0, x = var_5490_cast_fp16)[name = string("op_5491_cast_fp16")];
            fp16 var_5492_to_fp16 = const()[name = string("op_5492_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5493_cast_fp16 = mul(x = var_5491_cast_fp16, y = var_5492_to_fp16)[name = string("op_5493_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5494_cast_fp16 = add(x = var_5489_cast_fp16, y = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp32 var_5495_epsilon_0 = const()[name = string("op_5495_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5495_cast_fp16 = rsqrt(epsilon = var_5495_epsilon_0, x = var_5494_cast_fp16)[name = string("op_5495_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5496_cast_fp16 = mul(x = x_251_cast_fp16, y = var_5495_cast_fp16)[name = string("op_5496_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_255_cast_fp16 = mul(x = var_5496_cast_fp16, y = model_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_255_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_5505 = const()[name = string("op_5505"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5506_cast_fp16 = reshape(shape = var_5505, x = linear_86_cast_fp16)[name = string("op_5506_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_253_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_253_cast_fp16)[name = string("second_49_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5520_cast_fp16 = mul(x = first_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5520_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5521_cast_fp16 = mul(x = second_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5521_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5523_cast_fp16 = sub(x = var_5520_cast_fp16, y = var_5521_cast_fp16)[name = string("op_5523_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5524_cast_fp16 = mul(x = second_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5525_cast_fp16 = mul(x = first_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5525_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5527_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5525_cast_fp16)[name = string("op_5527_cast_fp16")];
            int32 var_5529 = const()[name = string("op_5529"), 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_5529, interleave = q_25_interleave_0, values = (var_5523_cast_fp16, var_5527_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_255_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_255_cast_fp16)[name = string("second_51_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5541_cast_fp16 = mul(x = first_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5541_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5542_cast_fp16 = mul(x = second_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5542_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5544_cast_fp16 = sub(x = var_5541_cast_fp16, y = var_5542_cast_fp16)[name = string("op_5544_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5545_cast_fp16 = mul(x = second_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5545_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5546_cast_fp16 = mul(x = first_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5548_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5546_cast_fp16)[name = string("op_5548_cast_fp16")];
            int32 var_5550 = const()[name = string("op_5550"), 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_5550, interleave = k_25_interleave_0, values = (var_5544_cast_fp16, var_5548_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_5561_begin_0 = const()[name = string("op_5561_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5561_end_0 = const()[name = string("op_5561_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5561_end_mask_0 = const()[name = string("op_5561_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5561_stride_0 = const()[name = string("op_5561_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5561_cast_fp16 = slice_by_index(begin = var_5561_begin_0, end = var_5561_end_0, end_mask = var_5561_end_mask_0, stride = var_5561_stride_0, x = q_25_cast_fp16)[name = string("op_5561_cast_fp16")];
            tensor<int32, [4]> var_5571_begin_0 = const()[name = string("op_5571_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5571_end_0 = const()[name = string("op_5571_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5571_end_mask_0 = const()[name = string("op_5571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5571_stride_0 = const()[name = string("op_5571_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5571_cast_fp16 = slice_by_index(begin = var_5571_begin_0, end = var_5571_end_0, end_mask = var_5571_end_mask_0, stride = var_5571_stride_0, x = q_25_cast_fp16)[name = string("op_5571_cast_fp16")];
            int32 var_5573 = const()[name = string("op_5573"), 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_5573, interleave = stacked_25_interleave_0, values = (var_5561_cast_fp16, var_5571_cast_fp16))[name = string("stacked_25_cast_fp16")];
            tensor<int32, [4]> var_5579_begin_0 = const()[name = string("op_5579_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5579_end_0 = const()[name = string("op_5579_end_0"), val = tensor<int32, [4]>([13, 8, 4112, 128])];
            tensor<bool, [4]> var_5579_end_mask_0 = const()[name = string("op_5579_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5579_cast_fp16 = slice_by_index(begin = var_5579_begin_0, end = var_5579_end_0, end_mask = var_5579_end_mask_0, x = read_state_0)[name = string("op_5579_cast_fp16")];
            tensor<int32, [4]> var_5589_begin_0 = const()[name = string("op_5589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5589_end_0 = const()[name = string("op_5589_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5589_end_mask_0 = const()[name = string("op_5589_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5589_cast_fp16 = slice_by_index(begin = var_5589_begin_0, end = var_5589_end_0, end_mask = var_5589_end_mask_0, x = var_5579_cast_fp16)[name = string("op_5589_cast_fp16")];
            bool var_5598_transpose_x_1 = const()[name = string("op_5598_transpose_x_1"), val = bool(false)];
            bool var_5598_transpose_y_1 = const()[name = string("op_5598_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_5598_cast_fp16 = matmul(transpose_x = var_5598_transpose_x_1, transpose_y = var_5598_transpose_y_1, x = var_5589_cast_fp16, y = stacked_25_cast_fp16)[name = string("op_5598_cast_fp16")];
            tensor<int32, [4]> var_5601_perm_0 = const()[name = string("op_5601_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5602_to_fp16 = const()[name = string("op_5602_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_5601_cast_fp16 = transpose(perm = var_5601_perm_0, x = var_5598_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 32, 4096]> var_5603_cast_fp16 = mul(x = var_5601_cast_fp16, y = var_5602_to_fp16)[name = string("op_5603_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_25_cast_fp16 = add(x = var_5603_cast_fp16, y = cacheMask)[name = string("cache_weights_25_cast_fp16")];
            bool var_5609_transpose_x_1 = const()[name = string("op_5609_transpose_x_1"), val = bool(false)];
            bool var_5609_transpose_y_1 = const()[name = string("op_5609_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5609_cast_fp16 = matmul(transpose_x = var_5609_transpose_x_1, transpose_y = var_5609_transpose_y_1, x = stacked_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_5609_cast_fp16")];
            fp16 var_5610_to_fp16 = const()[name = string("op_5610_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5611_cast_fp16 = mul(x = var_5609_cast_fp16, y = var_5610_to_fp16)[name = string("op_5611_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_25_cast_fp16 = add(x = var_5611_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_25_cast_fp16")];
            int32 var_5615 = const()[name = string("op_5615"), val = int32(-1)];
            bool var_5616_interleave_0 = const()[name = string("op_5616_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_5616_cast_fp16 = concat(axis = var_5615, interleave = var_5616_interleave_0, values = (cache_weights_25_cast_fp16, call_weights_25_cast_fp16))[name = string("op_5616_cast_fp16")];
            int32 var_5617 = const()[name = string("op_5617"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_25_cast_fp16 = softmax(axis = var_5617, x = var_5616_cast_fp16)[name = string("shares_25_cast_fp16")];
            tensor<int32, [4]> var_5624_begin_0 = const()[name = string("op_5624_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5624_end_0 = const()[name = string("op_5624_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_5624_end_mask_0 = const()[name = string("op_5624_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_5624_cast_fp16 = slice_by_index(begin = var_5624_begin_0, end = var_5624_end_0, end_mask = var_5624_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([13, 8, 4112, 128])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = read_state_1)[name = string("op_5629_cast_fp16")];
            tensor<int32, [4]> var_5639_begin_0 = const()[name = string("op_5639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5639_end_0 = const()[name = string("op_5639_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5639_end_mask_0 = const()[name = string("op_5639_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5639_cast_fp16 = slice_by_index(begin = var_5639_begin_0, end = var_5639_end_0, end_mask = var_5639_end_mask_0, x = var_5629_cast_fp16)[name = string("op_5639_cast_fp16")];
            bool var_5645_transpose_x_0 = const()[name = string("op_5645_transpose_x_0"), val = bool(false)];
            bool var_5645_transpose_y_0 = const()[name = string("op_5645_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5645_cast_fp16 = matmul(transpose_x = var_5645_transpose_x_0, transpose_y = var_5645_transpose_y_0, x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("op_5645_cast_fp16")];
            tensor<int32, [4]> var_5650_begin_0 = const()[name = string("op_5650_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_5650_end_0 = const()[name = string("op_5650_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_5650_end_mask_0 = const()[name = string("op_5650_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5650_cast_fp16 = slice_by_index(begin = var_5650_begin_0, end = var_5650_end_0, end_mask = var_5650_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5650_cast_fp16")];
            bool var_5651_transpose_x_0 = const()[name = string("op_5651_transpose_x_0"), val = bool(false)];
            bool var_5651_transpose_y_0 = const()[name = string("op_5651_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_5506_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 8, 32, 128]> var_5651_cast_fp16 = matmul(transpose_x = var_5651_transpose_x_0, transpose_y = var_5651_transpose_y_0, x = var_5650_cast_fp16, y = v_25_cast_fp16)[name = string("op_5651_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_49_cast_fp16 = add(x = var_5645_cast_fp16, y = var_5651_cast_fp16)[name = string("mixed_49_cast_fp16")];
            tensor<int32, [4]> var_5668_begin_0 = const()[name = string("op_5668_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5668_end_0 = const()[name = string("op_5668_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5668_end_mask_0 = const()[name = string("op_5668_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5668_cast_fp16 = slice_by_index(begin = var_5668_begin_0, end = var_5668_end_0, end_mask = var_5668_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5668_cast_fp16")];
            tensor<int32, [4]> var_5683_begin_0 = const()[name = string("op_5683_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5683_end_0 = const()[name = string("op_5683_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5683_end_mask_0 = const()[name = string("op_5683_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5683_cast_fp16 = slice_by_index(begin = var_5683_begin_0, end = var_5683_end_0, end_mask = var_5683_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5683_cast_fp16")];
            int32 var_5685 = const()[name = string("op_5685"), 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_5685, interleave = mixed_51_interleave_0, values = (var_5668_cast_fp16, var_5683_cast_fp16))[name = string("mixed_51_cast_fp16")];
            tensor<int32, [4]> var_5689_perm_0 = const()[name = string("op_5689_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5693 = const()[name = string("op_5693"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5689_cast_fp16 = transpose(perm = var_5689_perm_0, x = mixed_51_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 2048]> input_123_cast_fp16 = reshape(shape = var_5693, x = var_5689_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_257_cast_fp16 = add(x = x_241_cast_fp16, y = linear_87_cast_fp16)[name = string("x_257_cast_fp16")];
            bool var_5703 = const()[name = string("op_5703"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5706_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_5706_cast_fp16")];
            tensor<int32, [1]> peak_103_axes_0 = const()[name = string("peak_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_103_cast_fp16 = reduce_max(axes = peak_103_axes_0, keep_dims = var_5703, x = var_5706_cast_fp16)[name = string("peak_103_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_259_cast_fp16 = real_div(x = x_257_cast_fp16, y = peak_103_cast_fp16)[name = string("x_259_cast_fp16")];
            fp16 var_5702_promoted_to_fp16 = const()[name = string("op_5702_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5710_cast_fp16 = pow(x = x_259_cast_fp16, y = var_5702_promoted_to_fp16)[name = string("op_5710_cast_fp16")];
            tensor<int32, [1]> var_5712_axes_0 = const()[name = string("op_5712_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5712_keep_dims_0 = const()[name = string("op_5712_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5712_cast_fp16 = reduce_mean(axes = var_5712_axes_0, keep_dims = var_5712_keep_dims_0, x = var_5710_cast_fp16)[name = string("op_5712_cast_fp16")];
            fp16 var_5702_promoted_1_to_fp16 = const()[name = string("op_5702_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5713_cast_fp16 = pow(x = peak_103_cast_fp16, y = var_5702_promoted_1_to_fp16)[name = string("op_5713_cast_fp16")];
            fp32 var_5714_epsilon_0 = const()[name = string("op_5714_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5714_cast_fp16 = inverse(epsilon = var_5714_epsilon_0, x = var_5713_cast_fp16)[name = string("op_5714_cast_fp16")];
            fp16 var_5715_to_fp16 = const()[name = string("op_5715_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5716_cast_fp16 = mul(x = var_5714_cast_fp16, y = var_5715_to_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5717_cast_fp16 = add(x = var_5712_cast_fp16, y = var_5716_cast_fp16)[name = string("op_5717_cast_fp16")];
            fp32 var_5718_epsilon_0 = const()[name = string("op_5718_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5718_cast_fp16 = rsqrt(epsilon = var_5718_epsilon_0, x = var_5717_cast_fp16)[name = string("op_5718_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5719_cast_fp16 = mul(x = x_259_cast_fp16, y = var_5718_cast_fp16)[name = string("op_5719_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_5719_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_5724_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_5724_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_5724_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_261_cast_fp16 = add(x = x_257_cast_fp16, y = linear_90_cast_fp16)[name = string("x_261_cast_fp16")];
            bool var_5737 = const()[name = string("op_5737"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5740_cast_fp16 = abs(x = x_261_cast_fp16)[name = string("op_5740_cast_fp16")];
            tensor<int32, [1]> peak_105_axes_0 = const()[name = string("peak_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_105_cast_fp16 = reduce_max(axes = peak_105_axes_0, keep_dims = var_5737, x = var_5740_cast_fp16)[name = string("peak_105_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_263_cast_fp16 = real_div(x = x_261_cast_fp16, y = peak_105_cast_fp16)[name = string("x_263_cast_fp16")];
            fp16 var_5736_promoted_to_fp16 = const()[name = string("op_5736_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5744_cast_fp16 = pow(x = x_263_cast_fp16, y = var_5736_promoted_to_fp16)[name = string("op_5744_cast_fp16")];
            tensor<int32, [1]> var_5746_axes_0 = const()[name = string("op_5746_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5746_keep_dims_0 = const()[name = string("op_5746_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5746_cast_fp16 = reduce_mean(axes = var_5746_axes_0, keep_dims = var_5746_keep_dims_0, x = var_5744_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5736_promoted_1_to_fp16 = const()[name = string("op_5736_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5747_cast_fp16 = pow(x = peak_105_cast_fp16, y = var_5736_promoted_1_to_fp16)[name = string("op_5747_cast_fp16")];
            fp32 var_5748_epsilon_0 = const()[name = string("op_5748_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5748_cast_fp16 = inverse(epsilon = var_5748_epsilon_0, x = var_5747_cast_fp16)[name = string("op_5748_cast_fp16")];
            fp16 var_5749_to_fp16 = const()[name = string("op_5749_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5750_cast_fp16 = mul(x = var_5748_cast_fp16, y = var_5749_to_fp16)[name = string("op_5750_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5751_cast_fp16 = add(x = var_5746_cast_fp16, y = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            fp32 var_5752_epsilon_0 = const()[name = string("op_5752_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5752_cast_fp16 = rsqrt(epsilon = var_5752_epsilon_0, x = var_5751_cast_fp16)[name = string("op_5752_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5753_cast_fp16 = mul(x = x_263_cast_fp16, y = var_5752_cast_fp16)[name = string("op_5753_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_5753_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_5762 = const()[name = string("op_5762"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5763_cast_fp16 = reshape(shape = var_5762, x = linear_91_cast_fp16)[name = string("op_5763_cast_fp16")];
            tensor<int32, [4]> x_265_perm_0 = const()[name = string("x_265_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5770 = const()[name = string("op_5770"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_265_cast_fp16 = transpose(perm = x_265_perm_0, x = var_5763_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 16, 128]> var_5773_cast_fp16 = abs(x = x_265_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<int32, [1]> peak_107_axes_0 = const()[name = string("peak_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_107_cast_fp16 = reduce_max(axes = peak_107_axes_0, keep_dims = var_5770, x = var_5773_cast_fp16)[name = string("peak_107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_267_cast_fp16 = real_div(x = x_265_cast_fp16, y = peak_107_cast_fp16)[name = string("x_267_cast_fp16")];
            fp16 var_5769_promoted_to_fp16 = const()[name = string("op_5769_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5777_cast_fp16 = pow(x = x_267_cast_fp16, y = var_5769_promoted_to_fp16)[name = string("op_5777_cast_fp16")];
            tensor<int32, [1]> var_5779_axes_0 = const()[name = string("op_5779_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5779_keep_dims_0 = const()[name = string("op_5779_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5779_cast_fp16 = reduce_mean(axes = var_5779_axes_0, keep_dims = var_5779_keep_dims_0, x = var_5777_cast_fp16)[name = string("op_5779_cast_fp16")];
            fp16 var_5769_promoted_1_to_fp16 = const()[name = string("op_5769_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5780_cast_fp16 = pow(x = peak_107_cast_fp16, y = var_5769_promoted_1_to_fp16)[name = string("op_5780_cast_fp16")];
            fp32 var_5781_epsilon_0 = const()[name = string("op_5781_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5781_cast_fp16 = inverse(epsilon = var_5781_epsilon_0, x = var_5780_cast_fp16)[name = string("op_5781_cast_fp16")];
            fp16 var_5782_to_fp16 = const()[name = string("op_5782_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5783_cast_fp16 = mul(x = var_5781_cast_fp16, y = var_5782_to_fp16)[name = string("op_5783_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5784_cast_fp16 = add(x = var_5779_cast_fp16, y = var_5783_cast_fp16)[name = string("op_5784_cast_fp16")];
            fp32 var_5785_epsilon_0 = const()[name = string("op_5785_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5785_cast_fp16 = rsqrt(epsilon = var_5785_epsilon_0, x = var_5784_cast_fp16)[name = string("op_5785_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5786_cast_fp16 = mul(x = x_267_cast_fp16, y = var_5785_cast_fp16)[name = string("op_5786_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_273_cast_fp16 = mul(x = var_5786_cast_fp16, y = model_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_273_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_5795 = const()[name = string("op_5795"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5796_cast_fp16 = reshape(shape = var_5795, x = linear_92_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5803 = const()[name = string("op_5803"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_5796_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 8, 16, 128]> var_5806_cast_fp16 = abs(x = x_269_cast_fp16)[name = string("op_5806_cast_fp16")];
            tensor<int32, [1]> peak_109_axes_0 = const()[name = string("peak_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_109_cast_fp16 = reduce_max(axes = peak_109_axes_0, keep_dims = var_5803, x = var_5806_cast_fp16)[name = string("peak_109_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_271_cast_fp16 = real_div(x = x_269_cast_fp16, y = peak_109_cast_fp16)[name = string("x_271_cast_fp16")];
            fp16 var_5802_promoted_to_fp16 = const()[name = string("op_5802_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5810_cast_fp16 = pow(x = x_271_cast_fp16, y = var_5802_promoted_to_fp16)[name = string("op_5810_cast_fp16")];
            tensor<int32, [1]> var_5812_axes_0 = const()[name = string("op_5812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5812_keep_dims_0 = const()[name = string("op_5812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5812_cast_fp16 = reduce_mean(axes = var_5812_axes_0, keep_dims = var_5812_keep_dims_0, x = var_5810_cast_fp16)[name = string("op_5812_cast_fp16")];
            fp16 var_5802_promoted_1_to_fp16 = const()[name = string("op_5802_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5813_cast_fp16 = pow(x = peak_109_cast_fp16, y = var_5802_promoted_1_to_fp16)[name = string("op_5813_cast_fp16")];
            fp32 var_5814_epsilon_0 = const()[name = string("op_5814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5814_cast_fp16 = inverse(epsilon = var_5814_epsilon_0, x = var_5813_cast_fp16)[name = string("op_5814_cast_fp16")];
            fp16 var_5815_to_fp16 = const()[name = string("op_5815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5816_cast_fp16 = mul(x = var_5814_cast_fp16, y = var_5815_to_fp16)[name = string("op_5816_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5817_cast_fp16 = add(x = var_5812_cast_fp16, y = var_5816_cast_fp16)[name = string("op_5817_cast_fp16")];
            fp32 var_5818_epsilon_0 = const()[name = string("op_5818_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5818_cast_fp16 = rsqrt(epsilon = var_5818_epsilon_0, x = var_5817_cast_fp16)[name = string("op_5818_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5819_cast_fp16 = mul(x = x_271_cast_fp16, y = var_5818_cast_fp16)[name = string("op_5819_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_275_cast_fp16 = mul(x = var_5819_cast_fp16, y = model_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_275_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_5828 = const()[name = string("op_5828"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5829_cast_fp16 = reshape(shape = var_5828, x = linear_93_cast_fp16)[name = string("op_5829_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_273_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_273_cast_fp16)[name = string("second_53_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5843_cast_fp16 = mul(x = first_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5843_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5844_cast_fp16 = mul(x = second_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5844_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5846_cast_fp16 = sub(x = var_5843_cast_fp16, y = var_5844_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5847_cast_fp16 = mul(x = second_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5847_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5848_cast_fp16 = mul(x = first_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5848_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5850_cast_fp16 = add(x = var_5847_cast_fp16, y = var_5848_cast_fp16)[name = string("op_5850_cast_fp16")];
            int32 var_5852 = const()[name = string("op_5852"), 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_5852, interleave = q_27_interleave_0, values = (var_5846_cast_fp16, var_5850_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_275_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_275_cast_fp16)[name = string("second_55_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5864_cast_fp16 = mul(x = first_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5865_cast_fp16 = mul(x = second_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5865_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5867_cast_fp16 = sub(x = var_5864_cast_fp16, y = var_5865_cast_fp16)[name = string("op_5867_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5868_cast_fp16 = mul(x = second_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5868_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5869_cast_fp16 = mul(x = first_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5869_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5871_cast_fp16 = add(x = var_5868_cast_fp16, y = var_5869_cast_fp16)[name = string("op_5871_cast_fp16")];
            int32 var_5873 = const()[name = string("op_5873"), 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_5873, interleave = k_27_interleave_0, values = (var_5867_cast_fp16, var_5871_cast_fp16))[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> var_5884_begin_0 = const()[name = string("op_5884_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5884_end_0 = const()[name = string("op_5884_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5884_end_mask_0 = const()[name = string("op_5884_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5884_stride_0 = const()[name = string("op_5884_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5884_cast_fp16 = slice_by_index(begin = var_5884_begin_0, end = var_5884_end_0, end_mask = var_5884_end_mask_0, stride = var_5884_stride_0, x = q_27_cast_fp16)[name = string("op_5884_cast_fp16")];
            tensor<int32, [4]> var_5894_begin_0 = const()[name = string("op_5894_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5894_end_0 = const()[name = string("op_5894_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5894_end_mask_0 = const()[name = string("op_5894_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5894_stride_0 = const()[name = string("op_5894_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5894_cast_fp16 = slice_by_index(begin = var_5894_begin_0, end = var_5894_end_0, end_mask = var_5894_end_mask_0, stride = var_5894_stride_0, x = q_27_cast_fp16)[name = string("op_5894_cast_fp16")];
            int32 var_5896 = const()[name = string("op_5896"), 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_5896, interleave = stacked_27_interleave_0, values = (var_5884_cast_fp16, var_5894_cast_fp16))[name = string("stacked_27_cast_fp16")];
            tensor<int32, [4]> var_5902_begin_0 = const()[name = string("op_5902_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5902_end_0 = const()[name = string("op_5902_end_0"), val = tensor<int32, [4]>([14, 8, 4112, 128])];
            tensor<bool, [4]> var_5902_end_mask_0 = const()[name = string("op_5902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5902_cast_fp16 = slice_by_index(begin = var_5902_begin_0, end = var_5902_end_0, end_mask = var_5902_end_mask_0, x = read_state_0)[name = string("op_5902_cast_fp16")];
            tensor<int32, [4]> var_5912_begin_0 = const()[name = string("op_5912_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5912_end_0 = const()[name = string("op_5912_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5912_end_mask_0 = const()[name = string("op_5912_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5912_cast_fp16 = slice_by_index(begin = var_5912_begin_0, end = var_5912_end_0, end_mask = var_5912_end_mask_0, x = var_5902_cast_fp16)[name = string("op_5912_cast_fp16")];
            bool var_5921_transpose_x_1 = const()[name = string("op_5921_transpose_x_1"), val = bool(false)];
            bool var_5921_transpose_y_1 = const()[name = string("op_5921_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_5921_cast_fp16 = matmul(transpose_x = var_5921_transpose_x_1, transpose_y = var_5921_transpose_y_1, x = var_5912_cast_fp16, y = stacked_27_cast_fp16)[name = string("op_5921_cast_fp16")];
            tensor<int32, [4]> var_5924_perm_0 = const()[name = string("op_5924_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5925_to_fp16 = const()[name = string("op_5925_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_5924_cast_fp16 = transpose(perm = var_5924_perm_0, x = var_5921_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 8, 32, 4096]> var_5926_cast_fp16 = mul(x = var_5924_cast_fp16, y = var_5925_to_fp16)[name = string("op_5926_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_27_cast_fp16 = add(x = var_5926_cast_fp16, y = cacheMask)[name = string("cache_weights_27_cast_fp16")];
            bool var_5932_transpose_x_1 = const()[name = string("op_5932_transpose_x_1"), val = bool(false)];
            bool var_5932_transpose_y_1 = const()[name = string("op_5932_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5932_cast_fp16 = matmul(transpose_x = var_5932_transpose_x_1, transpose_y = var_5932_transpose_y_1, x = stacked_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_5932_cast_fp16")];
            fp16 var_5933_to_fp16 = const()[name = string("op_5933_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5934_cast_fp16 = mul(x = var_5932_cast_fp16, y = var_5933_to_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_27_cast_fp16 = add(x = var_5934_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_27_cast_fp16")];
            int32 var_5938 = const()[name = string("op_5938"), val = int32(-1)];
            bool var_5939_interleave_0 = const()[name = string("op_5939_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_5939_cast_fp16 = concat(axis = var_5938, interleave = var_5939_interleave_0, values = (cache_weights_27_cast_fp16, call_weights_27_cast_fp16))[name = string("op_5939_cast_fp16")];
            int32 var_5940 = const()[name = string("op_5940"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_27_cast_fp16 = softmax(axis = var_5940, x = var_5939_cast_fp16)[name = string("shares_27_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5947_cast_fp16")];
            tensor<int32, [4]> var_5952_begin_0 = const()[name = string("op_5952_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5952_end_0 = const()[name = string("op_5952_end_0"), val = tensor<int32, [4]>([14, 8, 4112, 128])];
            tensor<bool, [4]> var_5952_end_mask_0 = const()[name = string("op_5952_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5952_cast_fp16 = slice_by_index(begin = var_5952_begin_0, end = var_5952_end_0, end_mask = var_5952_end_mask_0, x = read_state_1)[name = string("op_5952_cast_fp16")];
            tensor<int32, [4]> var_5962_begin_0 = const()[name = string("op_5962_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5962_end_0 = const()[name = string("op_5962_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5962_end_mask_0 = const()[name = string("op_5962_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5962_cast_fp16 = slice_by_index(begin = var_5962_begin_0, end = var_5962_end_0, end_mask = var_5962_end_mask_0, x = var_5952_cast_fp16)[name = string("op_5962_cast_fp16")];
            bool var_5968_transpose_x_0 = const()[name = string("op_5968_transpose_x_0"), val = bool(false)];
            bool var_5968_transpose_y_0 = const()[name = string("op_5968_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5968_cast_fp16 = matmul(transpose_x = var_5968_transpose_x_0, transpose_y = var_5968_transpose_y_0, x = var_5947_cast_fp16, y = var_5962_cast_fp16)[name = string("op_5968_cast_fp16")];
            tensor<int32, [4]> var_5973_begin_0 = const()[name = string("op_5973_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_5973_end_0 = const()[name = string("op_5973_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_5973_end_mask_0 = const()[name = string("op_5973_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5973_cast_fp16 = slice_by_index(begin = var_5973_begin_0, end = var_5973_end_0, end_mask = var_5973_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5973_cast_fp16")];
            bool var_5974_transpose_x_0 = const()[name = string("op_5974_transpose_x_0"), val = bool(false)];
            bool var_5974_transpose_y_0 = const()[name = string("op_5974_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_5829_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 8, 32, 128]> var_5974_cast_fp16 = matmul(transpose_x = var_5974_transpose_x_0, transpose_y = var_5974_transpose_y_0, x = var_5973_cast_fp16, y = v_27_cast_fp16)[name = string("op_5974_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_53_cast_fp16 = add(x = var_5968_cast_fp16, y = var_5974_cast_fp16)[name = string("mixed_53_cast_fp16")];
            tensor<int32, [4]> var_5991_begin_0 = const()[name = string("op_5991_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5991_end_0 = const()[name = string("op_5991_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5991_end_mask_0 = const()[name = string("op_5991_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5991_cast_fp16 = slice_by_index(begin = var_5991_begin_0, end = var_5991_end_0, end_mask = var_5991_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5991_cast_fp16")];
            tensor<int32, [4]> var_6006_begin_0 = const()[name = string("op_6006_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6006_end_0 = const()[name = string("op_6006_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6006_end_mask_0 = const()[name = string("op_6006_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6006_cast_fp16 = slice_by_index(begin = var_6006_begin_0, end = var_6006_end_0, end_mask = var_6006_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_6006_cast_fp16")];
            int32 var_6008 = const()[name = string("op_6008"), 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_6008, interleave = mixed_55_interleave_0, values = (var_5991_cast_fp16, var_6006_cast_fp16))[name = string("mixed_55_cast_fp16")];
            tensor<int32, [4]> var_6012_perm_0 = const()[name = string("op_6012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6016 = const()[name = string("op_6016"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6012_cast_fp16 = transpose(perm = var_6012_perm_0, x = mixed_55_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 2048]> input_133_cast_fp16 = reshape(shape = var_6016, x = var_6012_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_277_cast_fp16 = add(x = x_261_cast_fp16, y = linear_94_cast_fp16)[name = string("x_277_cast_fp16")];
            bool var_6026 = const()[name = string("op_6026"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6029_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_6029_cast_fp16")];
            tensor<int32, [1]> peak_111_axes_0 = const()[name = string("peak_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_111_cast_fp16 = reduce_max(axes = peak_111_axes_0, keep_dims = var_6026, x = var_6029_cast_fp16)[name = string("peak_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_279_cast_fp16 = real_div(x = x_277_cast_fp16, y = peak_111_cast_fp16)[name = string("x_279_cast_fp16")];
            fp16 var_6025_promoted_to_fp16 = const()[name = string("op_6025_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6033_cast_fp16 = pow(x = x_279_cast_fp16, y = var_6025_promoted_to_fp16)[name = string("op_6033_cast_fp16")];
            tensor<int32, [1]> var_6035_axes_0 = const()[name = string("op_6035_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6035_keep_dims_0 = const()[name = string("op_6035_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6035_cast_fp16 = reduce_mean(axes = var_6035_axes_0, keep_dims = var_6035_keep_dims_0, x = var_6033_cast_fp16)[name = string("op_6035_cast_fp16")];
            fp16 var_6025_promoted_1_to_fp16 = const()[name = string("op_6025_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6036_cast_fp16 = pow(x = peak_111_cast_fp16, y = var_6025_promoted_1_to_fp16)[name = string("op_6036_cast_fp16")];
            fp32 var_6037_epsilon_0 = const()[name = string("op_6037_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6037_cast_fp16 = inverse(epsilon = var_6037_epsilon_0, x = var_6036_cast_fp16)[name = string("op_6037_cast_fp16")];
            fp16 var_6038_to_fp16 = const()[name = string("op_6038_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6040_cast_fp16 = add(x = var_6035_cast_fp16, y = var_6039_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6041_cast_fp16 = rsqrt(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6042_cast_fp16 = mul(x = x_279_cast_fp16, y = var_6041_cast_fp16)[name = string("op_6042_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_6042_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_6047_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_6047_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_6047_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_281_cast_fp16 = add(x = x_277_cast_fp16, y = linear_97_cast_fp16)[name = string("x_281_cast_fp16")];
            bool var_6060 = const()[name = string("op_6060"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6063_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_6063_cast_fp16")];
            tensor<int32, [1]> peak_113_axes_0 = const()[name = string("peak_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_113_cast_fp16 = reduce_max(axes = peak_113_axes_0, keep_dims = var_6060, x = var_6063_cast_fp16)[name = string("peak_113_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_283_cast_fp16 = real_div(x = x_281_cast_fp16, y = peak_113_cast_fp16)[name = string("x_283_cast_fp16")];
            fp16 var_6059_promoted_to_fp16 = const()[name = string("op_6059_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6067_cast_fp16 = pow(x = x_283_cast_fp16, y = var_6059_promoted_to_fp16)[name = string("op_6067_cast_fp16")];
            tensor<int32, [1]> var_6069_axes_0 = const()[name = string("op_6069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6069_keep_dims_0 = const()[name = string("op_6069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6069_cast_fp16 = reduce_mean(axes = var_6069_axes_0, keep_dims = var_6069_keep_dims_0, x = var_6067_cast_fp16)[name = string("op_6069_cast_fp16")];
            fp16 var_6059_promoted_1_to_fp16 = const()[name = string("op_6059_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6070_cast_fp16 = pow(x = peak_113_cast_fp16, y = var_6059_promoted_1_to_fp16)[name = string("op_6070_cast_fp16")];
            fp32 var_6071_epsilon_0 = const()[name = string("op_6071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6071_cast_fp16 = inverse(epsilon = var_6071_epsilon_0, x = var_6070_cast_fp16)[name = string("op_6071_cast_fp16")];
            fp16 var_6072_to_fp16 = const()[name = string("op_6072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6073_cast_fp16 = mul(x = var_6071_cast_fp16, y = var_6072_to_fp16)[name = string("op_6073_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6074_cast_fp16 = add(x = var_6069_cast_fp16, y = var_6073_cast_fp16)[name = string("op_6074_cast_fp16")];
            fp32 var_6075_epsilon_0 = const()[name = string("op_6075_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6075_cast_fp16 = rsqrt(epsilon = var_6075_epsilon_0, x = var_6074_cast_fp16)[name = string("op_6075_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6076_cast_fp16 = mul(x = x_283_cast_fp16, y = var_6075_cast_fp16)[name = string("op_6076_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_6076_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_6085 = const()[name = string("op_6085"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6086_cast_fp16 = reshape(shape = var_6085, x = linear_98_cast_fp16)[name = string("op_6086_cast_fp16")];
            tensor<int32, [4]> x_285_perm_0 = const()[name = string("x_285_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6093 = const()[name = string("op_6093"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_285_cast_fp16 = transpose(perm = x_285_perm_0, x = var_6086_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 16, 128]> var_6096_cast_fp16 = abs(x = x_285_cast_fp16)[name = string("op_6096_cast_fp16")];
            tensor<int32, [1]> peak_115_axes_0 = const()[name = string("peak_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_115_cast_fp16 = reduce_max(axes = peak_115_axes_0, keep_dims = var_6093, x = var_6096_cast_fp16)[name = string("peak_115_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_287_cast_fp16 = real_div(x = x_285_cast_fp16, y = peak_115_cast_fp16)[name = string("x_287_cast_fp16")];
            fp16 var_6092_promoted_to_fp16 = const()[name = string("op_6092_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6100_cast_fp16 = pow(x = x_287_cast_fp16, y = var_6092_promoted_to_fp16)[name = string("op_6100_cast_fp16")];
            tensor<int32, [1]> var_6102_axes_0 = const()[name = string("op_6102_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6102_keep_dims_0 = const()[name = string("op_6102_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6102_cast_fp16 = reduce_mean(axes = var_6102_axes_0, keep_dims = var_6102_keep_dims_0, x = var_6100_cast_fp16)[name = string("op_6102_cast_fp16")];
            fp16 var_6092_promoted_1_to_fp16 = const()[name = string("op_6092_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6103_cast_fp16 = pow(x = peak_115_cast_fp16, y = var_6092_promoted_1_to_fp16)[name = string("op_6103_cast_fp16")];
            fp32 var_6104_epsilon_0 = const()[name = string("op_6104_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6104_cast_fp16 = inverse(epsilon = var_6104_epsilon_0, x = var_6103_cast_fp16)[name = string("op_6104_cast_fp16")];
            fp16 var_6105_to_fp16 = const()[name = string("op_6105_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6106_cast_fp16 = mul(x = var_6104_cast_fp16, y = var_6105_to_fp16)[name = string("op_6106_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6107_cast_fp16 = add(x = var_6102_cast_fp16, y = var_6106_cast_fp16)[name = string("op_6107_cast_fp16")];
            fp32 var_6108_epsilon_0 = const()[name = string("op_6108_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6108_cast_fp16 = rsqrt(epsilon = var_6108_epsilon_0, x = var_6107_cast_fp16)[name = string("op_6108_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6109_cast_fp16 = mul(x = x_287_cast_fp16, y = var_6108_cast_fp16)[name = string("op_6109_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_293_cast_fp16 = mul(x = var_6109_cast_fp16, y = model_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_293_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_6118 = const()[name = string("op_6118"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6119_cast_fp16 = reshape(shape = var_6118, x = linear_99_cast_fp16)[name = string("op_6119_cast_fp16")];
            tensor<int32, [4]> x_289_perm_0 = const()[name = string("x_289_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6126 = const()[name = string("op_6126"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_289_cast_fp16 = transpose(perm = x_289_perm_0, x = var_6119_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 16, 128]> var_6129_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_6129_cast_fp16")];
            tensor<int32, [1]> peak_117_axes_0 = const()[name = string("peak_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_117_cast_fp16 = reduce_max(axes = peak_117_axes_0, keep_dims = var_6126, x = var_6129_cast_fp16)[name = string("peak_117_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_291_cast_fp16 = real_div(x = x_289_cast_fp16, y = peak_117_cast_fp16)[name = string("x_291_cast_fp16")];
            fp16 var_6125_promoted_to_fp16 = const()[name = string("op_6125_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6133_cast_fp16 = pow(x = x_291_cast_fp16, y = var_6125_promoted_to_fp16)[name = string("op_6133_cast_fp16")];
            tensor<int32, [1]> var_6135_axes_0 = const()[name = string("op_6135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6135_keep_dims_0 = const()[name = string("op_6135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6135_cast_fp16 = reduce_mean(axes = var_6135_axes_0, keep_dims = var_6135_keep_dims_0, x = var_6133_cast_fp16)[name = string("op_6135_cast_fp16")];
            fp16 var_6125_promoted_1_to_fp16 = const()[name = string("op_6125_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6136_cast_fp16 = pow(x = peak_117_cast_fp16, y = var_6125_promoted_1_to_fp16)[name = string("op_6136_cast_fp16")];
            fp32 var_6137_epsilon_0 = const()[name = string("op_6137_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6137_cast_fp16 = inverse(epsilon = var_6137_epsilon_0, x = var_6136_cast_fp16)[name = string("op_6137_cast_fp16")];
            fp16 var_6138_to_fp16 = const()[name = string("op_6138_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> 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, 16, 1]> var_6140_cast_fp16 = add(x = var_6135_cast_fp16, y = var_6139_cast_fp16)[name = string("op_6140_cast_fp16")];
            fp32 var_6141_epsilon_0 = const()[name = string("op_6141_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6141_cast_fp16 = rsqrt(epsilon = var_6141_epsilon_0, x = var_6140_cast_fp16)[name = string("op_6141_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6142_cast_fp16 = mul(x = x_291_cast_fp16, y = var_6141_cast_fp16)[name = string("op_6142_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_295_cast_fp16 = mul(x = var_6142_cast_fp16, y = model_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_295_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_6151 = const()[name = string("op_6151"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6152_cast_fp16 = reshape(shape = var_6151, x = linear_100_cast_fp16)[name = string("op_6152_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_293_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_293_cast_fp16)[name = string("second_57_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6166_cast_fp16 = mul(x = first_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6166_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6167_cast_fp16 = mul(x = second_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6167_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6169_cast_fp16 = sub(x = var_6166_cast_fp16, y = var_6167_cast_fp16)[name = string("op_6169_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6170_cast_fp16 = mul(x = second_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6171_cast_fp16 = mul(x = first_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6171_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6173_cast_fp16 = add(x = var_6170_cast_fp16, y = var_6171_cast_fp16)[name = string("op_6173_cast_fp16")];
            int32 var_6175 = const()[name = string("op_6175"), 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_6175, interleave = q_29_interleave_0, values = (var_6169_cast_fp16, var_6173_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_295_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_295_cast_fp16)[name = string("second_59_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6187_cast_fp16 = mul(x = first_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6187_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6188_cast_fp16 = mul(x = second_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6190_cast_fp16 = sub(x = var_6187_cast_fp16, y = var_6188_cast_fp16)[name = string("op_6190_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6191_cast_fp16 = mul(x = second_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6191_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6192_cast_fp16 = mul(x = first_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6192_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6194_cast_fp16 = add(x = var_6191_cast_fp16, y = var_6192_cast_fp16)[name = string("op_6194_cast_fp16")];
            int32 var_6196 = const()[name = string("op_6196"), 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_6196, interleave = k_29_interleave_0, values = (var_6190_cast_fp16, var_6194_cast_fp16))[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_6207_begin_0 = const()[name = string("op_6207_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6207_end_0 = const()[name = string("op_6207_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6207_end_mask_0 = const()[name = string("op_6207_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6207_stride_0 = const()[name = string("op_6207_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6207_cast_fp16 = slice_by_index(begin = var_6207_begin_0, end = var_6207_end_0, end_mask = var_6207_end_mask_0, stride = var_6207_stride_0, x = q_29_cast_fp16)[name = string("op_6207_cast_fp16")];
            tensor<int32, [4]> var_6217_begin_0 = const()[name = string("op_6217_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6217_end_0 = const()[name = string("op_6217_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6217_end_mask_0 = const()[name = string("op_6217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6217_stride_0 = const()[name = string("op_6217_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6217_cast_fp16 = slice_by_index(begin = var_6217_begin_0, end = var_6217_end_0, end_mask = var_6217_end_mask_0, stride = var_6217_stride_0, x = q_29_cast_fp16)[name = string("op_6217_cast_fp16")];
            int32 var_6219 = const()[name = string("op_6219"), 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_6219, interleave = stacked_29_interleave_0, values = (var_6207_cast_fp16, var_6217_cast_fp16))[name = string("stacked_29_cast_fp16")];
            tensor<int32, [4]> var_6225_begin_0 = const()[name = string("op_6225_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_6225_end_0 = const()[name = string("op_6225_end_0"), val = tensor<int32, [4]>([15, 8, 4112, 128])];
            tensor<bool, [4]> var_6225_end_mask_0 = const()[name = string("op_6225_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6225_cast_fp16 = slice_by_index(begin = var_6225_begin_0, end = var_6225_end_0, end_mask = var_6225_end_mask_0, x = read_state_0)[name = string("op_6225_cast_fp16")];
            tensor<int32, [4]> var_6235_begin_0 = const()[name = string("op_6235_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6235_end_0 = const()[name = string("op_6235_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6235_end_mask_0 = const()[name = string("op_6235_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6235_cast_fp16 = slice_by_index(begin = var_6235_begin_0, end = var_6235_end_0, end_mask = var_6235_end_mask_0, x = var_6225_cast_fp16)[name = string("op_6235_cast_fp16")];
            bool var_6244_transpose_x_1 = const()[name = string("op_6244_transpose_x_1"), val = bool(false)];
            bool var_6244_transpose_y_1 = const()[name = string("op_6244_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_6244_cast_fp16 = matmul(transpose_x = var_6244_transpose_x_1, transpose_y = var_6244_transpose_y_1, x = var_6235_cast_fp16, y = stacked_29_cast_fp16)[name = string("op_6244_cast_fp16")];
            tensor<int32, [4]> var_6247_perm_0 = const()[name = string("op_6247_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6248_to_fp16 = const()[name = string("op_6248_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_6247_cast_fp16 = transpose(perm = var_6247_perm_0, x = var_6244_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 8, 32, 4096]> var_6249_cast_fp16 = mul(x = var_6247_cast_fp16, y = var_6248_to_fp16)[name = string("op_6249_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_29_cast_fp16 = add(x = var_6249_cast_fp16, y = cacheMask)[name = string("cache_weights_29_cast_fp16")];
            bool var_6255_transpose_x_1 = const()[name = string("op_6255_transpose_x_1"), val = bool(false)];
            bool var_6255_transpose_y_1 = const()[name = string("op_6255_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6255_cast_fp16 = matmul(transpose_x = var_6255_transpose_x_1, transpose_y = var_6255_transpose_y_1, x = stacked_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_6255_cast_fp16")];
            fp16 var_6256_to_fp16 = const()[name = string("op_6256_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6257_cast_fp16 = mul(x = var_6255_cast_fp16, y = var_6256_to_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_29_cast_fp16 = add(x = var_6257_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_29_cast_fp16")];
            int32 var_6261 = const()[name = string("op_6261"), val = int32(-1)];
            bool var_6262_interleave_0 = const()[name = string("op_6262_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_6262_cast_fp16 = concat(axis = var_6261, interleave = var_6262_interleave_0, values = (cache_weights_29_cast_fp16, call_weights_29_cast_fp16))[name = string("op_6262_cast_fp16")];
            int32 var_6263 = const()[name = string("op_6263"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_29_cast_fp16 = softmax(axis = var_6263, x = var_6262_cast_fp16)[name = string("shares_29_cast_fp16")];
            tensor<int32, [4]> var_6270_begin_0 = const()[name = string("op_6270_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6270_end_0 = const()[name = string("op_6270_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_6270_end_mask_0 = const()[name = string("op_6270_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_6270_cast_fp16 = slice_by_index(begin = var_6270_begin_0, end = var_6270_end_0, end_mask = var_6270_end_mask_0, x = shares_29_cast_fp16)[name = string("op_6270_cast_fp16")];
            tensor<int32, [4]> var_6275_begin_0 = const()[name = string("op_6275_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_6275_end_0 = const()[name = string("op_6275_end_0"), val = tensor<int32, [4]>([15, 8, 4112, 128])];
            tensor<bool, [4]> var_6275_end_mask_0 = const()[name = string("op_6275_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6275_cast_fp16 = slice_by_index(begin = var_6275_begin_0, end = var_6275_end_0, end_mask = var_6275_end_mask_0, x = read_state_1)[name = string("op_6275_cast_fp16")];
            tensor<int32, [4]> var_6285_begin_0 = const()[name = string("op_6285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6285_end_0 = const()[name = string("op_6285_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6285_end_mask_0 = const()[name = string("op_6285_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6285_cast_fp16 = slice_by_index(begin = var_6285_begin_0, end = var_6285_end_0, end_mask = var_6285_end_mask_0, x = var_6275_cast_fp16)[name = string("op_6285_cast_fp16")];
            bool var_6291_transpose_x_0 = const()[name = string("op_6291_transpose_x_0"), val = bool(false)];
            bool var_6291_transpose_y_0 = const()[name = string("op_6291_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6291_cast_fp16 = matmul(transpose_x = var_6291_transpose_x_0, transpose_y = var_6291_transpose_y_0, x = var_6270_cast_fp16, y = var_6285_cast_fp16)[name = string("op_6291_cast_fp16")];
            tensor<int32, [4]> var_6296_begin_0 = const()[name = string("op_6296_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_6296_end_0 = const()[name = string("op_6296_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_6296_end_mask_0 = const()[name = string("op_6296_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6296_cast_fp16 = slice_by_index(begin = var_6296_begin_0, end = var_6296_end_0, end_mask = var_6296_end_mask_0, x = shares_29_cast_fp16)[name = string("op_6296_cast_fp16")];
            bool var_6297_transpose_x_0 = const()[name = string("op_6297_transpose_x_0"), val = bool(false)];
            bool var_6297_transpose_y_0 = const()[name = string("op_6297_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_6152_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 8, 32, 128]> var_6297_cast_fp16 = matmul(transpose_x = var_6297_transpose_x_0, transpose_y = var_6297_transpose_y_0, x = var_6296_cast_fp16, y = v_29_cast_fp16)[name = string("op_6297_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_57_cast_fp16 = add(x = var_6291_cast_fp16, y = var_6297_cast_fp16)[name = string("mixed_57_cast_fp16")];
            tensor<int32, [4]> var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6314_cast_fp16 = slice_by_index(begin = var_6314_begin_0, end = var_6314_end_0, end_mask = var_6314_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_6314_cast_fp16")];
            tensor<int32, [4]> var_6329_begin_0 = const()[name = string("op_6329_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6329_end_0 = const()[name = string("op_6329_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6329_end_mask_0 = const()[name = string("op_6329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6329_cast_fp16 = slice_by_index(begin = var_6329_begin_0, end = var_6329_end_0, end_mask = var_6329_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_6329_cast_fp16")];
            int32 var_6331 = const()[name = string("op_6331"), 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_6331, interleave = mixed_59_interleave_0, values = (var_6314_cast_fp16, var_6329_cast_fp16))[name = string("mixed_59_cast_fp16")];
            tensor<int32, [4]> var_6335_perm_0 = const()[name = string("op_6335_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6339 = const()[name = string("op_6339"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6335_cast_fp16 = transpose(perm = var_6335_perm_0, x = mixed_59_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 2048]> input_143_cast_fp16 = reshape(shape = var_6339, x = var_6335_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_297_cast_fp16 = add(x = x_281_cast_fp16, y = linear_101_cast_fp16)[name = string("x_297_cast_fp16")];
            bool var_6349 = const()[name = string("op_6349"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6352_cast_fp16 = abs(x = x_297_cast_fp16)[name = string("op_6352_cast_fp16")];
            tensor<int32, [1]> peak_119_axes_0 = const()[name = string("peak_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_119_cast_fp16 = reduce_max(axes = peak_119_axes_0, keep_dims = var_6349, x = var_6352_cast_fp16)[name = string("peak_119_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_299_cast_fp16 = real_div(x = x_297_cast_fp16, y = peak_119_cast_fp16)[name = string("x_299_cast_fp16")];
            fp16 var_6348_promoted_to_fp16 = const()[name = string("op_6348_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6356_cast_fp16 = pow(x = x_299_cast_fp16, y = var_6348_promoted_to_fp16)[name = string("op_6356_cast_fp16")];
            tensor<int32, [1]> var_6358_axes_0 = const()[name = string("op_6358_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6358_keep_dims_0 = const()[name = string("op_6358_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6358_cast_fp16 = reduce_mean(axes = var_6358_axes_0, keep_dims = var_6358_keep_dims_0, x = var_6356_cast_fp16)[name = string("op_6358_cast_fp16")];
            fp16 var_6348_promoted_1_to_fp16 = const()[name = string("op_6348_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6359_cast_fp16 = pow(x = peak_119_cast_fp16, y = var_6348_promoted_1_to_fp16)[name = string("op_6359_cast_fp16")];
            fp32 var_6360_epsilon_0 = const()[name = string("op_6360_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6360_cast_fp16 = inverse(epsilon = var_6360_epsilon_0, x = var_6359_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6362_cast_fp16 = mul(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("op_6362_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6363_cast_fp16 = add(x = var_6358_cast_fp16, y = var_6362_cast_fp16)[name = string("op_6363_cast_fp16")];
            fp32 var_6364_epsilon_0 = const()[name = string("op_6364_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6364_cast_fp16 = rsqrt(epsilon = var_6364_epsilon_0, x = var_6363_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6365_cast_fp16 = mul(x = x_299_cast_fp16, y = var_6364_cast_fp16)[name = string("op_6365_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_6365_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_6370_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_6370_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_6370_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_301_cast_fp16 = add(x = x_297_cast_fp16, y = linear_104_cast_fp16)[name = string("x_301_cast_fp16")];
            bool var_6383 = const()[name = string("op_6383"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6386_cast_fp16 = abs(x = x_301_cast_fp16)[name = string("op_6386_cast_fp16")];
            tensor<int32, [1]> peak_121_axes_0 = const()[name = string("peak_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_121_cast_fp16 = reduce_max(axes = peak_121_axes_0, keep_dims = var_6383, x = var_6386_cast_fp16)[name = string("peak_121_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_303_cast_fp16 = real_div(x = x_301_cast_fp16, y = peak_121_cast_fp16)[name = string("x_303_cast_fp16")];
            fp16 var_6382_promoted_to_fp16 = const()[name = string("op_6382_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6390_cast_fp16 = pow(x = x_303_cast_fp16, y = var_6382_promoted_to_fp16)[name = string("op_6390_cast_fp16")];
            tensor<int32, [1]> var_6392_axes_0 = const()[name = string("op_6392_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6392_keep_dims_0 = const()[name = string("op_6392_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6392_cast_fp16 = reduce_mean(axes = var_6392_axes_0, keep_dims = var_6392_keep_dims_0, x = var_6390_cast_fp16)[name = string("op_6392_cast_fp16")];
            fp16 var_6382_promoted_1_to_fp16 = const()[name = string("op_6382_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6393_cast_fp16 = pow(x = peak_121_cast_fp16, y = var_6382_promoted_1_to_fp16)[name = string("op_6393_cast_fp16")];
            fp32 var_6394_epsilon_0 = const()[name = string("op_6394_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6394_cast_fp16 = inverse(epsilon = var_6394_epsilon_0, x = var_6393_cast_fp16)[name = string("op_6394_cast_fp16")];
            fp16 var_6395_to_fp16 = const()[name = string("op_6395_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6396_cast_fp16 = mul(x = var_6394_cast_fp16, y = var_6395_to_fp16)[name = string("op_6396_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6397_cast_fp16 = add(x = var_6392_cast_fp16, y = var_6396_cast_fp16)[name = string("op_6397_cast_fp16")];
            fp32 var_6398_epsilon_0 = const()[name = string("op_6398_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6398_cast_fp16 = rsqrt(epsilon = var_6398_epsilon_0, x = var_6397_cast_fp16)[name = string("op_6398_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6399_cast_fp16 = mul(x = x_303_cast_fp16, y = var_6398_cast_fp16)[name = string("op_6399_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_6399_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_6408 = const()[name = string("op_6408"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6409_cast_fp16 = reshape(shape = var_6408, x = linear_105_cast_fp16)[name = string("op_6409_cast_fp16")];
            tensor<int32, [4]> x_305_perm_0 = const()[name = string("x_305_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6416 = const()[name = string("op_6416"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_305_cast_fp16 = transpose(perm = x_305_perm_0, x = var_6409_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 16, 16, 128]> var_6419_cast_fp16 = abs(x = x_305_cast_fp16)[name = string("op_6419_cast_fp16")];
            tensor<int32, [1]> peak_123_axes_0 = const()[name = string("peak_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_123_cast_fp16 = reduce_max(axes = peak_123_axes_0, keep_dims = var_6416, x = var_6419_cast_fp16)[name = string("peak_123_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_307_cast_fp16 = real_div(x = x_305_cast_fp16, y = peak_123_cast_fp16)[name = string("x_307_cast_fp16")];
            fp16 var_6415_promoted_to_fp16 = const()[name = string("op_6415_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6423_cast_fp16 = pow(x = x_307_cast_fp16, y = var_6415_promoted_to_fp16)[name = string("op_6423_cast_fp16")];
            tensor<int32, [1]> var_6425_axes_0 = const()[name = string("op_6425_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6425_keep_dims_0 = const()[name = string("op_6425_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6425_cast_fp16 = reduce_mean(axes = var_6425_axes_0, keep_dims = var_6425_keep_dims_0, x = var_6423_cast_fp16)[name = string("op_6425_cast_fp16")];
            fp16 var_6415_promoted_1_to_fp16 = const()[name = string("op_6415_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6426_cast_fp16 = pow(x = peak_123_cast_fp16, y = var_6415_promoted_1_to_fp16)[name = string("op_6426_cast_fp16")];
            fp32 var_6427_epsilon_0 = const()[name = string("op_6427_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6427_cast_fp16 = inverse(epsilon = var_6427_epsilon_0, x = var_6426_cast_fp16)[name = string("op_6427_cast_fp16")];
            fp16 var_6428_to_fp16 = const()[name = string("op_6428_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6429_cast_fp16 = mul(x = var_6427_cast_fp16, y = var_6428_to_fp16)[name = string("op_6429_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6430_cast_fp16 = add(x = var_6425_cast_fp16, y = var_6429_cast_fp16)[name = string("op_6430_cast_fp16")];
            fp32 var_6431_epsilon_0 = const()[name = string("op_6431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6431_cast_fp16 = rsqrt(epsilon = var_6431_epsilon_0, x = var_6430_cast_fp16)[name = string("op_6431_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6432_cast_fp16 = mul(x = x_307_cast_fp16, y = var_6431_cast_fp16)[name = string("op_6432_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_313_cast_fp16 = mul(x = var_6432_cast_fp16, y = model_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_313_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_6441 = const()[name = string("op_6441"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6442_cast_fp16 = reshape(shape = var_6441, x = linear_106_cast_fp16)[name = string("op_6442_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6449 = const()[name = string("op_6449"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_6442_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 8, 16, 128]> var_6452_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_6452_cast_fp16")];
            tensor<int32, [1]> peak_125_axes_0 = const()[name = string("peak_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_125_cast_fp16 = reduce_max(axes = peak_125_axes_0, keep_dims = var_6449, x = var_6452_cast_fp16)[name = string("peak_125_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_311_cast_fp16 = real_div(x = x_309_cast_fp16, y = peak_125_cast_fp16)[name = string("x_311_cast_fp16")];
            fp16 var_6448_promoted_to_fp16 = const()[name = string("op_6448_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6456_cast_fp16 = pow(x = x_311_cast_fp16, y = var_6448_promoted_to_fp16)[name = string("op_6456_cast_fp16")];
            tensor<int32, [1]> var_6458_axes_0 = const()[name = string("op_6458_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6458_keep_dims_0 = const()[name = string("op_6458_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6458_cast_fp16 = reduce_mean(axes = var_6458_axes_0, keep_dims = var_6458_keep_dims_0, x = var_6456_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6448_promoted_1_to_fp16 = const()[name = string("op_6448_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6459_cast_fp16 = pow(x = peak_125_cast_fp16, y = var_6448_promoted_1_to_fp16)[name = string("op_6459_cast_fp16")];
            fp32 var_6460_epsilon_0 = const()[name = string("op_6460_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6460_cast_fp16 = inverse(epsilon = var_6460_epsilon_0, x = var_6459_cast_fp16)[name = string("op_6460_cast_fp16")];
            fp16 var_6461_to_fp16 = const()[name = string("op_6461_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6462_cast_fp16 = mul(x = var_6460_cast_fp16, y = var_6461_to_fp16)[name = string("op_6462_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_6463_cast_fp16 = add(x = var_6458_cast_fp16, y = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            fp32 var_6464_epsilon_0 = const()[name = string("op_6464_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6464_cast_fp16 = rsqrt(epsilon = var_6464_epsilon_0, x = var_6463_cast_fp16)[name = string("op_6464_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6465_cast_fp16 = mul(x = x_311_cast_fp16, y = var_6464_cast_fp16)[name = string("op_6465_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_315_cast_fp16 = mul(x = var_6465_cast_fp16, y = model_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_315_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_6474 = const()[name = string("op_6474"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6475_cast_fp16 = reshape(shape = var_6474, x = linear_107_cast_fp16)[name = string("op_6475_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_313_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_313_cast_fp16)[name = string("second_61_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6489_cast_fp16 = mul(x = first_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6490_cast_fp16 = mul(x = second_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6490_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6492_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = string("op_6492_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6493_cast_fp16 = mul(x = second_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6493_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6494_cast_fp16 = mul(x = first_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6494_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6496_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = string("op_6496_cast_fp16")];
            int32 var_6498 = const()[name = string("op_6498"), 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_6498, interleave = q_31_interleave_0, values = (var_6492_cast_fp16, var_6496_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_315_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_315_cast_fp16)[name = string("second_63_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6510_cast_fp16 = mul(x = first_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6510_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6511_cast_fp16 = mul(x = second_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6511_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6513_cast_fp16 = sub(x = var_6510_cast_fp16, y = var_6511_cast_fp16)[name = string("op_6513_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6514_cast_fp16 = mul(x = second_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6514_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6515_cast_fp16 = mul(x = first_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6515_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6517_cast_fp16 = add(x = var_6514_cast_fp16, y = var_6515_cast_fp16)[name = string("op_6517_cast_fp16")];
            int32 var_6519 = const()[name = string("op_6519"), 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_6519, interleave = k_31_interleave_0, values = (var_6513_cast_fp16, var_6517_cast_fp16))[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> var_6530_begin_0 = const()[name = string("op_6530_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6530_end_0 = const()[name = string("op_6530_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6530_end_mask_0 = const()[name = string("op_6530_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6530_stride_0 = const()[name = string("op_6530_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6530_cast_fp16 = slice_by_index(begin = var_6530_begin_0, end = var_6530_end_0, end_mask = var_6530_end_mask_0, stride = var_6530_stride_0, x = q_31_cast_fp16)[name = string("op_6530_cast_fp16")];
            tensor<int32, [4]> var_6540_begin_0 = const()[name = string("op_6540_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6540_end_0 = const()[name = string("op_6540_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6540_end_mask_0 = const()[name = string("op_6540_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6540_stride_0 = const()[name = string("op_6540_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6540_cast_fp16 = slice_by_index(begin = var_6540_begin_0, end = var_6540_end_0, end_mask = var_6540_end_mask_0, stride = var_6540_stride_0, x = q_31_cast_fp16)[name = string("op_6540_cast_fp16")];
            int32 var_6542 = const()[name = string("op_6542"), 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_6542, interleave = stacked_31_interleave_0, values = (var_6530_cast_fp16, var_6540_cast_fp16))[name = string("stacked_31_cast_fp16")];
            tensor<int32, [4]> var_6548_begin_0 = const()[name = string("op_6548_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6548_end_0 = const()[name = string("op_6548_end_0"), val = tensor<int32, [4]>([16, 8, 4112, 128])];
            tensor<bool, [4]> var_6548_end_mask_0 = const()[name = string("op_6548_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6548_cast_fp16 = slice_by_index(begin = var_6548_begin_0, end = var_6548_end_0, end_mask = var_6548_end_mask_0, x = read_state_0)[name = string("op_6548_cast_fp16")];
            tensor<int32, [4]> var_6558_begin_0 = const()[name = string("op_6558_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6558_end_0 = const()[name = string("op_6558_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6558_end_mask_0 = const()[name = string("op_6558_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6558_cast_fp16 = slice_by_index(begin = var_6558_begin_0, end = var_6558_end_0, end_mask = var_6558_end_mask_0, x = var_6548_cast_fp16)[name = string("op_6558_cast_fp16")];
            bool var_6567_transpose_x_1 = const()[name = string("op_6567_transpose_x_1"), val = bool(false)];
            bool var_6567_transpose_y_1 = const()[name = string("op_6567_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_6567_cast_fp16 = matmul(transpose_x = var_6567_transpose_x_1, transpose_y = var_6567_transpose_y_1, x = var_6558_cast_fp16, y = stacked_31_cast_fp16)[name = string("op_6567_cast_fp16")];
            tensor<int32, [4]> var_6570_perm_0 = const()[name = string("op_6570_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6571_to_fp16 = const()[name = string("op_6571_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_6570_cast_fp16 = transpose(perm = var_6570_perm_0, x = var_6567_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, 32, 4096]> var_6572_cast_fp16 = mul(x = var_6570_cast_fp16, y = var_6571_to_fp16)[name = string("op_6572_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_31_cast_fp16 = add(x = var_6572_cast_fp16, y = cacheMask)[name = string("cache_weights_31_cast_fp16")];
            bool var_6578_transpose_x_1 = const()[name = string("op_6578_transpose_x_1"), val = bool(false)];
            bool var_6578_transpose_y_1 = const()[name = string("op_6578_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6578_cast_fp16 = matmul(transpose_x = var_6578_transpose_x_1, transpose_y = var_6578_transpose_y_1, x = stacked_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6580_cast_fp16 = mul(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("op_6580_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_31_cast_fp16 = add(x = var_6580_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_31_cast_fp16")];
            int32 var_6584 = const()[name = string("op_6584"), val = int32(-1)];
            bool var_6585_interleave_0 = const()[name = string("op_6585_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_6585_cast_fp16 = concat(axis = var_6584, interleave = var_6585_interleave_0, values = (cache_weights_31_cast_fp16, call_weights_31_cast_fp16))[name = string("op_6585_cast_fp16")];
            int32 var_6586 = const()[name = string("op_6586"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_31_cast_fp16 = softmax(axis = var_6586, x = var_6585_cast_fp16)[name = string("shares_31_cast_fp16")];
            tensor<int32, [4]> var_6593_begin_0 = const()[name = string("op_6593_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6593_end_0 = const()[name = string("op_6593_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_6593_end_mask_0 = const()[name = string("op_6593_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_6593_cast_fp16 = slice_by_index(begin = var_6593_begin_0, end = var_6593_end_0, end_mask = var_6593_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6593_cast_fp16")];
            tensor<int32, [4]> var_6598_begin_0 = const()[name = string("op_6598_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6598_end_0 = const()[name = string("op_6598_end_0"), val = tensor<int32, [4]>([16, 8, 4112, 128])];
            tensor<bool, [4]> var_6598_end_mask_0 = const()[name = string("op_6598_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6598_cast_fp16 = slice_by_index(begin = var_6598_begin_0, end = var_6598_end_0, end_mask = var_6598_end_mask_0, x = read_state_1)[name = string("op_6598_cast_fp16")];
            tensor<int32, [4]> var_6608_begin_0 = const()[name = string("op_6608_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6608_end_0 = const()[name = string("op_6608_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6608_end_mask_0 = const()[name = string("op_6608_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6608_cast_fp16 = slice_by_index(begin = var_6608_begin_0, end = var_6608_end_0, end_mask = var_6608_end_mask_0, x = var_6598_cast_fp16)[name = string("op_6608_cast_fp16")];
            bool var_6614_transpose_x_0 = const()[name = string("op_6614_transpose_x_0"), val = bool(false)];
            bool var_6614_transpose_y_0 = const()[name = string("op_6614_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6614_cast_fp16 = matmul(transpose_x = var_6614_transpose_x_0, transpose_y = var_6614_transpose_y_0, x = var_6593_cast_fp16, y = var_6608_cast_fp16)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6619_cast_fp16")];
            bool var_6620_transpose_x_0 = const()[name = string("op_6620_transpose_x_0"), val = bool(false)];
            bool var_6620_transpose_y_0 = const()[name = string("op_6620_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_6475_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 32, 128]> var_6620_cast_fp16 = matmul(transpose_x = var_6620_transpose_x_0, transpose_y = var_6620_transpose_y_0, x = var_6619_cast_fp16, y = v_31_cast_fp16)[name = string("op_6620_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_61_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6620_cast_fp16)[name = string("mixed_61_cast_fp16")];
            tensor<int32, [4]> var_6637_begin_0 = const()[name = string("op_6637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6637_end_0 = const()[name = string("op_6637_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6637_end_mask_0 = const()[name = string("op_6637_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6637_cast_fp16 = slice_by_index(begin = var_6637_begin_0, end = var_6637_end_0, end_mask = var_6637_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6637_cast_fp16")];
            tensor<int32, [4]> var_6652_begin_0 = const()[name = string("op_6652_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6652_end_0 = const()[name = string("op_6652_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6652_end_mask_0 = const()[name = string("op_6652_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6652_cast_fp16 = slice_by_index(begin = var_6652_begin_0, end = var_6652_end_0, end_mask = var_6652_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6652_cast_fp16")];
            int32 var_6654 = const()[name = string("op_6654"), 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_6654, interleave = mixed_63_interleave_0, values = (var_6637_cast_fp16, var_6652_cast_fp16))[name = string("mixed_63_cast_fp16")];
            tensor<int32, [4]> var_6658_perm_0 = const()[name = string("op_6658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6662 = const()[name = string("op_6662"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6658_cast_fp16 = transpose(perm = var_6658_perm_0, x = mixed_63_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 16, 2048]> input_153_cast_fp16 = reshape(shape = var_6662, x = var_6658_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_317_cast_fp16 = add(x = x_301_cast_fp16, y = linear_108_cast_fp16)[name = string("x_317_cast_fp16")];
            bool var_6672 = const()[name = string("op_6672"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6675_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_6675_cast_fp16")];
            tensor<int32, [1]> peak_127_axes_0 = const()[name = string("peak_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_127_cast_fp16 = reduce_max(axes = peak_127_axes_0, keep_dims = var_6672, x = var_6675_cast_fp16)[name = string("peak_127_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_319_cast_fp16 = real_div(x = x_317_cast_fp16, y = peak_127_cast_fp16)[name = string("x_319_cast_fp16")];
            fp16 var_6671_promoted_to_fp16 = const()[name = string("op_6671_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6679_cast_fp16 = pow(x = x_319_cast_fp16, y = var_6671_promoted_to_fp16)[name = string("op_6679_cast_fp16")];
            tensor<int32, [1]> var_6681_axes_0 = const()[name = string("op_6681_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6681_keep_dims_0 = const()[name = string("op_6681_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6681_cast_fp16 = reduce_mean(axes = var_6681_axes_0, keep_dims = var_6681_keep_dims_0, x = var_6679_cast_fp16)[name = string("op_6681_cast_fp16")];
            fp16 var_6671_promoted_1_to_fp16 = const()[name = string("op_6671_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6682_cast_fp16 = pow(x = peak_127_cast_fp16, y = var_6671_promoted_1_to_fp16)[name = string("op_6682_cast_fp16")];
            fp32 var_6683_epsilon_0 = const()[name = string("op_6683_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6683_cast_fp16 = inverse(epsilon = var_6683_epsilon_0, x = var_6682_cast_fp16)[name = string("op_6683_cast_fp16")];
            fp16 var_6684_to_fp16 = const()[name = string("op_6684_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6685_cast_fp16 = mul(x = var_6683_cast_fp16, y = var_6684_to_fp16)[name = string("op_6685_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6686_cast_fp16 = add(x = var_6681_cast_fp16, y = var_6685_cast_fp16)[name = string("op_6686_cast_fp16")];
            fp32 var_6687_epsilon_0 = const()[name = string("op_6687_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6687_cast_fp16 = rsqrt(epsilon = var_6687_epsilon_0, x = var_6686_cast_fp16)[name = string("op_6687_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6688_cast_fp16 = mul(x = x_319_cast_fp16, y = var_6687_cast_fp16)[name = string("op_6688_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_6688_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_6693_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_6693_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_6693_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_321_cast_fp16 = add(x = x_317_cast_fp16, y = linear_111_cast_fp16)[name = string("x_321_cast_fp16")];
            bool var_6706 = const()[name = string("op_6706"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6709_cast_fp16 = abs(x = x_321_cast_fp16)[name = string("op_6709_cast_fp16")];
            tensor<int32, [1]> peak_129_axes_0 = const()[name = string("peak_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_129_cast_fp16 = reduce_max(axes = peak_129_axes_0, keep_dims = var_6706, x = var_6709_cast_fp16)[name = string("peak_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_323_cast_fp16 = real_div(x = x_321_cast_fp16, y = peak_129_cast_fp16)[name = string("x_323_cast_fp16")];
            fp16 var_6705_promoted_to_fp16 = const()[name = string("op_6705_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6713_cast_fp16 = pow(x = x_323_cast_fp16, y = var_6705_promoted_to_fp16)[name = string("op_6713_cast_fp16")];
            tensor<int32, [1]> var_6715_axes_0 = const()[name = string("op_6715_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6715_keep_dims_0 = const()[name = string("op_6715_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6715_cast_fp16 = reduce_mean(axes = var_6715_axes_0, keep_dims = var_6715_keep_dims_0, x = var_6713_cast_fp16)[name = string("op_6715_cast_fp16")];
            fp16 var_6705_promoted_1_to_fp16 = const()[name = string("op_6705_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6716_cast_fp16 = pow(x = peak_129_cast_fp16, y = var_6705_promoted_1_to_fp16)[name = string("op_6716_cast_fp16")];
            fp32 var_6717_epsilon_0 = const()[name = string("op_6717_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6717_cast_fp16 = inverse(epsilon = var_6717_epsilon_0, x = var_6716_cast_fp16)[name = string("op_6717_cast_fp16")];
            fp16 var_6718_to_fp16 = const()[name = string("op_6718_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6719_cast_fp16 = mul(x = var_6717_cast_fp16, y = var_6718_to_fp16)[name = string("op_6719_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6720_cast_fp16 = add(x = var_6715_cast_fp16, y = var_6719_cast_fp16)[name = string("op_6720_cast_fp16")];
            fp32 var_6721_epsilon_0 = const()[name = string("op_6721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6721_cast_fp16 = rsqrt(epsilon = var_6721_epsilon_0, x = var_6720_cast_fp16)[name = string("op_6721_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6722_cast_fp16 = mul(x = x_323_cast_fp16, y = var_6721_cast_fp16)[name = string("op_6722_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_6722_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_6731 = const()[name = string("op_6731"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6732_cast_fp16 = reshape(shape = var_6731, x = linear_112_cast_fp16)[name = string("op_6732_cast_fp16")];
            tensor<int32, [4]> x_325_perm_0 = const()[name = string("x_325_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6739 = const()[name = string("op_6739"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_325_cast_fp16 = transpose(perm = x_325_perm_0, x = var_6732_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 16, 128]> var_6742_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_6742_cast_fp16")];
            tensor<int32, [1]> peak_131_axes_0 = const()[name = string("peak_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_131_cast_fp16 = reduce_max(axes = peak_131_axes_0, keep_dims = var_6739, x = var_6742_cast_fp16)[name = string("peak_131_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_327_cast_fp16 = real_div(x = x_325_cast_fp16, y = peak_131_cast_fp16)[name = string("x_327_cast_fp16")];
            fp16 var_6738_promoted_to_fp16 = const()[name = string("op_6738_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6746_cast_fp16 = pow(x = x_327_cast_fp16, y = var_6738_promoted_to_fp16)[name = string("op_6746_cast_fp16")];
            tensor<int32, [1]> var_6748_axes_0 = const()[name = string("op_6748_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6748_keep_dims_0 = const()[name = string("op_6748_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6748_cast_fp16 = reduce_mean(axes = var_6748_axes_0, keep_dims = var_6748_keep_dims_0, x = var_6746_cast_fp16)[name = string("op_6748_cast_fp16")];
            fp16 var_6738_promoted_1_to_fp16 = const()[name = string("op_6738_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6749_cast_fp16 = pow(x = peak_131_cast_fp16, y = var_6738_promoted_1_to_fp16)[name = string("op_6749_cast_fp16")];
            fp32 var_6750_epsilon_0 = const()[name = string("op_6750_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6750_cast_fp16 = inverse(epsilon = var_6750_epsilon_0, x = var_6749_cast_fp16)[name = string("op_6750_cast_fp16")];
            fp16 var_6751_to_fp16 = const()[name = string("op_6751_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6752_cast_fp16 = mul(x = var_6750_cast_fp16, y = var_6751_to_fp16)[name = string("op_6752_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6753_cast_fp16 = add(x = var_6748_cast_fp16, y = var_6752_cast_fp16)[name = string("op_6753_cast_fp16")];
            fp32 var_6754_epsilon_0 = const()[name = string("op_6754_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6754_cast_fp16 = rsqrt(epsilon = var_6754_epsilon_0, x = var_6753_cast_fp16)[name = string("op_6754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6755_cast_fp16 = mul(x = x_327_cast_fp16, y = var_6754_cast_fp16)[name = string("op_6755_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_333_cast_fp16 = mul(x = var_6755_cast_fp16, y = model_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("x_333_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_6764 = const()[name = string("op_6764"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6765_cast_fp16 = reshape(shape = var_6764, x = linear_113_cast_fp16)[name = string("op_6765_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6772 = const()[name = string("op_6772"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_6765_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 16, 128]> var_6775_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_6775_cast_fp16")];
            tensor<int32, [1]> peak_133_axes_0 = const()[name = string("peak_133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_133_cast_fp16 = reduce_max(axes = peak_133_axes_0, keep_dims = var_6772, x = var_6775_cast_fp16)[name = string("peak_133_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_331_cast_fp16 = real_div(x = x_329_cast_fp16, y = peak_133_cast_fp16)[name = string("x_331_cast_fp16")];
            fp16 var_6771_promoted_to_fp16 = const()[name = string("op_6771_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6779_cast_fp16 = pow(x = x_331_cast_fp16, y = var_6771_promoted_to_fp16)[name = string("op_6779_cast_fp16")];
            tensor<int32, [1]> var_6781_axes_0 = const()[name = string("op_6781_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6781_keep_dims_0 = const()[name = string("op_6781_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6781_cast_fp16 = reduce_mean(axes = var_6781_axes_0, keep_dims = var_6781_keep_dims_0, x = var_6779_cast_fp16)[name = string("op_6781_cast_fp16")];
            fp16 var_6771_promoted_1_to_fp16 = const()[name = string("op_6771_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6782_cast_fp16 = pow(x = peak_133_cast_fp16, y = var_6771_promoted_1_to_fp16)[name = string("op_6782_cast_fp16")];
            fp32 var_6783_epsilon_0 = const()[name = string("op_6783_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6783_cast_fp16 = inverse(epsilon = var_6783_epsilon_0, x = var_6782_cast_fp16)[name = string("op_6783_cast_fp16")];
            fp16 var_6784_to_fp16 = const()[name = string("op_6784_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6785_cast_fp16 = mul(x = var_6783_cast_fp16, y = var_6784_to_fp16)[name = string("op_6785_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_6786_cast_fp16 = add(x = var_6781_cast_fp16, y = var_6785_cast_fp16)[name = string("op_6786_cast_fp16")];
            fp32 var_6787_epsilon_0 = const()[name = string("op_6787_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6787_cast_fp16 = rsqrt(epsilon = var_6787_epsilon_0, x = var_6786_cast_fp16)[name = string("op_6787_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6788_cast_fp16 = mul(x = x_331_cast_fp16, y = var_6787_cast_fp16)[name = string("op_6788_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_335_cast_fp16 = mul(x = var_6788_cast_fp16, y = model_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("x_335_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_6797 = const()[name = string("op_6797"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6798_cast_fp16 = reshape(shape = var_6797, x = linear_114_cast_fp16)[name = string("op_6798_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_333_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_333_cast_fp16)[name = string("second_65_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6812_cast_fp16 = mul(x = first_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6813_cast_fp16 = mul(x = second_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6813_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6815_cast_fp16 = sub(x = var_6812_cast_fp16, y = var_6813_cast_fp16)[name = string("op_6815_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6816_cast_fp16 = mul(x = second_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6817_cast_fp16 = mul(x = first_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6817_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6819_cast_fp16 = add(x = var_6816_cast_fp16, y = var_6817_cast_fp16)[name = string("op_6819_cast_fp16")];
            int32 var_6821 = const()[name = string("op_6821"), 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_6821, interleave = q_33_interleave_0, values = (var_6815_cast_fp16, var_6819_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_335_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_335_cast_fp16)[name = string("second_67_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6833_cast_fp16 = mul(x = first_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6833_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6834_cast_fp16 = mul(x = second_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6834_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6836_cast_fp16 = sub(x = var_6833_cast_fp16, y = var_6834_cast_fp16)[name = string("op_6836_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6837_cast_fp16 = mul(x = second_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6837_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6838_cast_fp16 = mul(x = first_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6838_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6840_cast_fp16 = add(x = var_6837_cast_fp16, y = var_6838_cast_fp16)[name = string("op_6840_cast_fp16")];
            int32 var_6842 = const()[name = string("op_6842"), 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_6842, interleave = k_33_interleave_0, values = (var_6836_cast_fp16, var_6840_cast_fp16))[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_6853_begin_0 = const()[name = string("op_6853_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6853_end_0 = const()[name = string("op_6853_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6853_end_mask_0 = const()[name = string("op_6853_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6853_stride_0 = const()[name = string("op_6853_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6853_cast_fp16 = slice_by_index(begin = var_6853_begin_0, end = var_6853_end_0, end_mask = var_6853_end_mask_0, stride = var_6853_stride_0, x = q_33_cast_fp16)[name = string("op_6853_cast_fp16")];
            tensor<int32, [4]> var_6863_begin_0 = const()[name = string("op_6863_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6863_end_0 = const()[name = string("op_6863_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6863_end_mask_0 = const()[name = string("op_6863_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6863_stride_0 = const()[name = string("op_6863_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6863_cast_fp16 = slice_by_index(begin = var_6863_begin_0, end = var_6863_end_0, end_mask = var_6863_end_mask_0, stride = var_6863_stride_0, x = q_33_cast_fp16)[name = string("op_6863_cast_fp16")];
            int32 var_6865 = const()[name = string("op_6865"), 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_6865, interleave = stacked_33_interleave_0, values = (var_6853_cast_fp16, var_6863_cast_fp16))[name = string("stacked_33_cast_fp16")];
            tensor<int32, [4]> var_6871_begin_0 = const()[name = string("op_6871_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6871_end_0 = const()[name = string("op_6871_end_0"), val = tensor<int32, [4]>([17, 8, 4112, 128])];
            tensor<bool, [4]> var_6871_end_mask_0 = const()[name = string("op_6871_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6871_cast_fp16 = slice_by_index(begin = var_6871_begin_0, end = var_6871_end_0, end_mask = var_6871_end_mask_0, x = read_state_0)[name = string("op_6871_cast_fp16")];
            tensor<int32, [4]> var_6881_begin_0 = const()[name = string("op_6881_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6881_end_0 = const()[name = string("op_6881_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6881_end_mask_0 = const()[name = string("op_6881_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6881_cast_fp16 = slice_by_index(begin = var_6881_begin_0, end = var_6881_end_0, end_mask = var_6881_end_mask_0, x = var_6871_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6890_transpose_x_1 = const()[name = string("op_6890_transpose_x_1"), val = bool(false)];
            bool var_6890_transpose_y_1 = const()[name = string("op_6890_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_6890_cast_fp16 = matmul(transpose_x = var_6890_transpose_x_1, transpose_y = var_6890_transpose_y_1, x = var_6881_cast_fp16, y = stacked_33_cast_fp16)[name = string("op_6890_cast_fp16")];
            tensor<int32, [4]> var_6893_perm_0 = const()[name = string("op_6893_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6894_to_fp16 = const()[name = string("op_6894_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_6893_cast_fp16 = transpose(perm = var_6893_perm_0, x = var_6890_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 32, 4096]> var_6895_cast_fp16 = mul(x = var_6893_cast_fp16, y = var_6894_to_fp16)[name = string("op_6895_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_33_cast_fp16 = add(x = var_6895_cast_fp16, y = cacheMask)[name = string("cache_weights_33_cast_fp16")];
            bool var_6901_transpose_x_1 = const()[name = string("op_6901_transpose_x_1"), val = bool(false)];
            bool var_6901_transpose_y_1 = const()[name = string("op_6901_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6901_cast_fp16 = matmul(transpose_x = var_6901_transpose_x_1, transpose_y = var_6901_transpose_y_1, x = stacked_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_6901_cast_fp16")];
            fp16 var_6902_to_fp16 = const()[name = string("op_6902_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6903_cast_fp16 = mul(x = var_6901_cast_fp16, y = var_6902_to_fp16)[name = string("op_6903_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_33_cast_fp16 = add(x = var_6903_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_33_cast_fp16")];
            int32 var_6907 = const()[name = string("op_6907"), val = int32(-1)];
            bool var_6908_interleave_0 = const()[name = string("op_6908_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_6908_cast_fp16 = concat(axis = var_6907, interleave = var_6908_interleave_0, values = (cache_weights_33_cast_fp16, call_weights_33_cast_fp16))[name = string("op_6908_cast_fp16")];
            int32 var_6909 = const()[name = string("op_6909"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_33_cast_fp16 = softmax(axis = var_6909, x = var_6908_cast_fp16)[name = string("shares_33_cast_fp16")];
            tensor<int32, [4]> var_6916_begin_0 = const()[name = string("op_6916_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6916_end_0 = const()[name = string("op_6916_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_6916_end_mask_0 = const()[name = string("op_6916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_6916_cast_fp16 = slice_by_index(begin = var_6916_begin_0, end = var_6916_end_0, end_mask = var_6916_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6916_cast_fp16")];
            tensor<int32, [4]> var_6921_begin_0 = const()[name = string("op_6921_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6921_end_0 = const()[name = string("op_6921_end_0"), val = tensor<int32, [4]>([17, 8, 4112, 128])];
            tensor<bool, [4]> var_6921_end_mask_0 = const()[name = string("op_6921_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6921_cast_fp16 = slice_by_index(begin = var_6921_begin_0, end = var_6921_end_0, end_mask = var_6921_end_mask_0, x = read_state_1)[name = string("op_6921_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = var_6921_cast_fp16)[name = string("op_6931_cast_fp16")];
            bool var_6937_transpose_x_0 = const()[name = string("op_6937_transpose_x_0"), val = bool(false)];
            bool var_6937_transpose_y_0 = const()[name = string("op_6937_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6937_cast_fp16 = matmul(transpose_x = var_6937_transpose_x_0, transpose_y = var_6937_transpose_y_0, x = var_6916_cast_fp16, y = var_6931_cast_fp16)[name = string("op_6937_cast_fp16")];
            tensor<int32, [4]> var_6942_begin_0 = const()[name = string("op_6942_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_6942_end_0 = const()[name = string("op_6942_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_6942_end_mask_0 = const()[name = string("op_6942_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6942_cast_fp16 = slice_by_index(begin = var_6942_begin_0, end = var_6942_end_0, end_mask = var_6942_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6942_cast_fp16")];
            bool var_6943_transpose_x_0 = const()[name = string("op_6943_transpose_x_0"), val = bool(false)];
            bool var_6943_transpose_y_0 = const()[name = string("op_6943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_6798_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 8, 32, 128]> var_6943_cast_fp16 = matmul(transpose_x = var_6943_transpose_x_0, transpose_y = var_6943_transpose_y_0, x = var_6942_cast_fp16, y = v_33_cast_fp16)[name = string("op_6943_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_65_cast_fp16 = add(x = var_6937_cast_fp16, y = var_6943_cast_fp16)[name = string("mixed_65_cast_fp16")];
            tensor<int32, [4]> var_6960_begin_0 = const()[name = string("op_6960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6960_end_0 = const()[name = string("op_6960_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6960_end_mask_0 = const()[name = string("op_6960_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6960_cast_fp16 = slice_by_index(begin = var_6960_begin_0, end = var_6960_end_0, end_mask = var_6960_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6960_cast_fp16")];
            tensor<int32, [4]> var_6975_begin_0 = const()[name = string("op_6975_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6975_end_0 = const()[name = string("op_6975_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6975_end_mask_0 = const()[name = string("op_6975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6975_cast_fp16 = slice_by_index(begin = var_6975_begin_0, end = var_6975_end_0, end_mask = var_6975_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6977 = const()[name = string("op_6977"), 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_6977, interleave = mixed_67_interleave_0, values = (var_6960_cast_fp16, var_6975_cast_fp16))[name = string("mixed_67_cast_fp16")];
            tensor<int32, [4]> var_6981_perm_0 = const()[name = string("op_6981_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6985 = const()[name = string("op_6985"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6981_cast_fp16 = transpose(perm = var_6981_perm_0, x = mixed_67_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 2048]> input_163_cast_fp16 = reshape(shape = var_6985, x = var_6981_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_337_cast_fp16 = add(x = x_321_cast_fp16, y = linear_115_cast_fp16)[name = string("x_337_cast_fp16")];
            bool var_6995 = const()[name = string("op_6995"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6998_cast_fp16 = abs(x = x_337_cast_fp16)[name = string("op_6998_cast_fp16")];
            tensor<int32, [1]> peak_135_axes_0 = const()[name = string("peak_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_135_cast_fp16 = reduce_max(axes = peak_135_axes_0, keep_dims = var_6995, x = var_6998_cast_fp16)[name = string("peak_135_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_339_cast_fp16 = real_div(x = x_337_cast_fp16, y = peak_135_cast_fp16)[name = string("x_339_cast_fp16")];
            fp16 var_6994_promoted_to_fp16 = const()[name = string("op_6994_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7002_cast_fp16 = pow(x = x_339_cast_fp16, y = var_6994_promoted_to_fp16)[name = string("op_7002_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_7002_cast_fp16)[name = string("op_7004_cast_fp16")];
            fp16 var_6994_promoted_1_to_fp16 = const()[name = string("op_6994_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7005_cast_fp16 = pow(x = peak_135_cast_fp16, y = var_6994_promoted_1_to_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7009_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7009_cast_fp16")];
            fp32 var_7010_epsilon_0 = const()[name = string("op_7010_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7010_cast_fp16 = rsqrt(epsilon = var_7010_epsilon_0, x = var_7009_cast_fp16)[name = string("op_7010_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7011_cast_fp16 = mul(x = x_339_cast_fp16, y = var_7010_cast_fp16)[name = string("op_7011_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_7011_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_7016_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_7016_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_7016_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_341_cast_fp16 = add(x = x_337_cast_fp16, y = linear_118_cast_fp16)[name = string("x_341_cast_fp16")];
            bool var_7029 = const()[name = string("op_7029"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7032_cast_fp16 = abs(x = x_341_cast_fp16)[name = string("op_7032_cast_fp16")];
            tensor<int32, [1]> peak_137_axes_0 = const()[name = string("peak_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_137_cast_fp16 = reduce_max(axes = peak_137_axes_0, keep_dims = var_7029, x = var_7032_cast_fp16)[name = string("peak_137_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_343_cast_fp16 = real_div(x = x_341_cast_fp16, y = peak_137_cast_fp16)[name = string("x_343_cast_fp16")];
            fp16 var_7028_promoted_to_fp16 = const()[name = string("op_7028_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7036_cast_fp16 = pow(x = x_343_cast_fp16, y = var_7028_promoted_to_fp16)[name = string("op_7036_cast_fp16")];
            tensor<int32, [1]> var_7038_axes_0 = const()[name = string("op_7038_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7038_keep_dims_0 = const()[name = string("op_7038_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7038_cast_fp16 = reduce_mean(axes = var_7038_axes_0, keep_dims = var_7038_keep_dims_0, x = var_7036_cast_fp16)[name = string("op_7038_cast_fp16")];
            fp16 var_7028_promoted_1_to_fp16 = const()[name = string("op_7028_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7039_cast_fp16 = pow(x = peak_137_cast_fp16, y = var_7028_promoted_1_to_fp16)[name = string("op_7039_cast_fp16")];
            fp32 var_7040_epsilon_0 = const()[name = string("op_7040_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7040_cast_fp16 = inverse(epsilon = var_7040_epsilon_0, x = var_7039_cast_fp16)[name = string("op_7040_cast_fp16")];
            fp16 var_7041_to_fp16 = const()[name = string("op_7041_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7042_cast_fp16 = mul(x = var_7040_cast_fp16, y = var_7041_to_fp16)[name = string("op_7042_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7043_cast_fp16 = add(x = var_7038_cast_fp16, y = var_7042_cast_fp16)[name = string("op_7043_cast_fp16")];
            fp32 var_7044_epsilon_0 = const()[name = string("op_7044_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7044_cast_fp16 = rsqrt(epsilon = var_7044_epsilon_0, x = var_7043_cast_fp16)[name = string("op_7044_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7045_cast_fp16 = mul(x = x_343_cast_fp16, y = var_7044_cast_fp16)[name = string("op_7045_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_7045_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_7054 = const()[name = string("op_7054"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7055_cast_fp16 = reshape(shape = var_7054, x = linear_119_cast_fp16)[name = string("op_7055_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7062 = const()[name = string("op_7062"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_7055_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 16, 128]> var_7065_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_7065_cast_fp16")];
            tensor<int32, [1]> peak_139_axes_0 = const()[name = string("peak_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_139_cast_fp16 = reduce_max(axes = peak_139_axes_0, keep_dims = var_7062, x = var_7065_cast_fp16)[name = string("peak_139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_347_cast_fp16 = real_div(x = x_345_cast_fp16, y = peak_139_cast_fp16)[name = string("x_347_cast_fp16")];
            fp16 var_7061_promoted_to_fp16 = const()[name = string("op_7061_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7069_cast_fp16 = pow(x = x_347_cast_fp16, y = var_7061_promoted_to_fp16)[name = string("op_7069_cast_fp16")];
            tensor<int32, [1]> var_7071_axes_0 = const()[name = string("op_7071_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7071_keep_dims_0 = const()[name = string("op_7071_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7071_cast_fp16 = reduce_mean(axes = var_7071_axes_0, keep_dims = var_7071_keep_dims_0, x = var_7069_cast_fp16)[name = string("op_7071_cast_fp16")];
            fp16 var_7061_promoted_1_to_fp16 = const()[name = string("op_7061_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7072_cast_fp16 = pow(x = peak_139_cast_fp16, y = var_7061_promoted_1_to_fp16)[name = string("op_7072_cast_fp16")];
            fp32 var_7073_epsilon_0 = const()[name = string("op_7073_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7073_cast_fp16 = inverse(epsilon = var_7073_epsilon_0, x = var_7072_cast_fp16)[name = string("op_7073_cast_fp16")];
            fp16 var_7074_to_fp16 = const()[name = string("op_7074_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7075_cast_fp16 = mul(x = var_7073_cast_fp16, y = var_7074_to_fp16)[name = string("op_7075_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7076_cast_fp16 = add(x = var_7071_cast_fp16, y = var_7075_cast_fp16)[name = string("op_7076_cast_fp16")];
            fp32 var_7077_epsilon_0 = const()[name = string("op_7077_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7077_cast_fp16 = rsqrt(epsilon = var_7077_epsilon_0, x = var_7076_cast_fp16)[name = string("op_7077_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7078_cast_fp16 = mul(x = x_347_cast_fp16, y = var_7077_cast_fp16)[name = string("op_7078_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_353_cast_fp16 = mul(x = var_7078_cast_fp16, y = model_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("x_353_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_7087 = const()[name = string("op_7087"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7088_cast_fp16 = reshape(shape = var_7087, x = linear_120_cast_fp16)[name = string("op_7088_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7095 = const()[name = string("op_7095"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_7088_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 8, 16, 128]> var_7098_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_7098_cast_fp16")];
            tensor<int32, [1]> peak_141_axes_0 = const()[name = string("peak_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_141_cast_fp16 = reduce_max(axes = peak_141_axes_0, keep_dims = var_7095, x = var_7098_cast_fp16)[name = string("peak_141_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_351_cast_fp16 = real_div(x = x_349_cast_fp16, y = peak_141_cast_fp16)[name = string("x_351_cast_fp16")];
            fp16 var_7094_promoted_to_fp16 = const()[name = string("op_7094_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7102_cast_fp16 = pow(x = x_351_cast_fp16, y = var_7094_promoted_to_fp16)[name = string("op_7102_cast_fp16")];
            tensor<int32, [1]> var_7104_axes_0 = const()[name = string("op_7104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7104_keep_dims_0 = const()[name = string("op_7104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7104_cast_fp16 = reduce_mean(axes = var_7104_axes_0, keep_dims = var_7104_keep_dims_0, x = var_7102_cast_fp16)[name = string("op_7104_cast_fp16")];
            fp16 var_7094_promoted_1_to_fp16 = const()[name = string("op_7094_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7105_cast_fp16 = pow(x = peak_141_cast_fp16, y = var_7094_promoted_1_to_fp16)[name = string("op_7105_cast_fp16")];
            fp32 var_7106_epsilon_0 = const()[name = string("op_7106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7106_cast_fp16 = inverse(epsilon = var_7106_epsilon_0, x = var_7105_cast_fp16)[name = string("op_7106_cast_fp16")];
            fp16 var_7107_to_fp16 = const()[name = string("op_7107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7108_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7107_to_fp16)[name = string("op_7108_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7109_cast_fp16 = add(x = var_7104_cast_fp16, y = var_7108_cast_fp16)[name = string("op_7109_cast_fp16")];
            fp32 var_7110_epsilon_0 = const()[name = string("op_7110_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7110_cast_fp16 = rsqrt(epsilon = var_7110_epsilon_0, x = var_7109_cast_fp16)[name = string("op_7110_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7111_cast_fp16 = mul(x = x_351_cast_fp16, y = var_7110_cast_fp16)[name = string("op_7111_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_355_cast_fp16 = mul(x = var_7111_cast_fp16, y = model_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("x_355_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_7120 = const()[name = string("op_7120"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7121_cast_fp16 = reshape(shape = var_7120, x = linear_121_cast_fp16)[name = string("op_7121_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_353_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_353_cast_fp16)[name = string("second_69_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7135_cast_fp16 = mul(x = first_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7135_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7136_cast_fp16 = mul(x = second_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7136_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7138_cast_fp16 = sub(x = var_7135_cast_fp16, y = var_7136_cast_fp16)[name = string("op_7138_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7139_cast_fp16 = mul(x = second_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7140_cast_fp16 = mul(x = first_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7140_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7142_cast_fp16 = add(x = var_7139_cast_fp16, y = var_7140_cast_fp16)[name = string("op_7142_cast_fp16")];
            int32 var_7144 = const()[name = string("op_7144"), 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_7144, interleave = q_35_interleave_0, values = (var_7138_cast_fp16, var_7142_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_355_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_355_cast_fp16)[name = string("second_71_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7156_cast_fp16 = mul(x = first_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7156_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7157_cast_fp16 = mul(x = second_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7157_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7159_cast_fp16 = sub(x = var_7156_cast_fp16, y = var_7157_cast_fp16)[name = string("op_7159_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7160_cast_fp16 = mul(x = second_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7161_cast_fp16 = mul(x = first_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7161_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7163_cast_fp16 = add(x = var_7160_cast_fp16, y = var_7161_cast_fp16)[name = string("op_7163_cast_fp16")];
            int32 var_7165 = const()[name = string("op_7165"), 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_7165, interleave = k_35_interleave_0, values = (var_7159_cast_fp16, var_7163_cast_fp16))[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> var_7176_begin_0 = const()[name = string("op_7176_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7176_end_0 = const()[name = string("op_7176_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7176_end_mask_0 = const()[name = string("op_7176_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7176_stride_0 = const()[name = string("op_7176_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7176_cast_fp16 = slice_by_index(begin = var_7176_begin_0, end = var_7176_end_0, end_mask = var_7176_end_mask_0, stride = var_7176_stride_0, x = q_35_cast_fp16)[name = string("op_7176_cast_fp16")];
            tensor<int32, [4]> var_7186_begin_0 = const()[name = string("op_7186_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7186_end_0 = const()[name = string("op_7186_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7186_end_mask_0 = const()[name = string("op_7186_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7186_stride_0 = const()[name = string("op_7186_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7186_cast_fp16 = slice_by_index(begin = var_7186_begin_0, end = var_7186_end_0, end_mask = var_7186_end_mask_0, stride = var_7186_stride_0, x = q_35_cast_fp16)[name = string("op_7186_cast_fp16")];
            int32 var_7188 = const()[name = string("op_7188"), 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_7188, interleave = stacked_35_interleave_0, values = (var_7176_cast_fp16, var_7186_cast_fp16))[name = string("stacked_35_cast_fp16")];
            tensor<int32, [4]> var_7194_begin_0 = const()[name = string("op_7194_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_7194_end_0 = const()[name = string("op_7194_end_0"), val = tensor<int32, [4]>([18, 8, 4112, 128])];
            tensor<bool, [4]> var_7194_end_mask_0 = const()[name = string("op_7194_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7194_cast_fp16 = slice_by_index(begin = var_7194_begin_0, end = var_7194_end_0, end_mask = var_7194_end_mask_0, x = read_state_0)[name = string("op_7194_cast_fp16")];
            tensor<int32, [4]> var_7204_begin_0 = const()[name = string("op_7204_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7204_end_0 = const()[name = string("op_7204_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7204_end_mask_0 = const()[name = string("op_7204_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7204_cast_fp16 = slice_by_index(begin = var_7204_begin_0, end = var_7204_end_0, end_mask = var_7204_end_mask_0, x = var_7194_cast_fp16)[name = string("op_7204_cast_fp16")];
            bool var_7213_transpose_x_1 = const()[name = string("op_7213_transpose_x_1"), val = bool(false)];
            bool var_7213_transpose_y_1 = const()[name = string("op_7213_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_7213_cast_fp16 = matmul(transpose_x = var_7213_transpose_x_1, transpose_y = var_7213_transpose_y_1, x = var_7204_cast_fp16, y = stacked_35_cast_fp16)[name = string("op_7213_cast_fp16")];
            tensor<int32, [4]> var_7216_perm_0 = const()[name = string("op_7216_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7217_to_fp16 = const()[name = string("op_7217_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_7216_cast_fp16 = transpose(perm = var_7216_perm_0, x = var_7213_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 8, 32, 4096]> var_7218_cast_fp16 = mul(x = var_7216_cast_fp16, y = var_7217_to_fp16)[name = string("op_7218_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_35_cast_fp16 = add(x = var_7218_cast_fp16, y = cacheMask)[name = string("cache_weights_35_cast_fp16")];
            bool var_7224_transpose_x_1 = const()[name = string("op_7224_transpose_x_1"), val = bool(false)];
            bool var_7224_transpose_y_1 = const()[name = string("op_7224_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7224_cast_fp16 = matmul(transpose_x = var_7224_transpose_x_1, transpose_y = var_7224_transpose_y_1, x = stacked_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_7224_cast_fp16")];
            fp16 var_7225_to_fp16 = const()[name = string("op_7225_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7226_cast_fp16 = mul(x = var_7224_cast_fp16, y = var_7225_to_fp16)[name = string("op_7226_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_35_cast_fp16 = add(x = var_7226_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_35_cast_fp16")];
            int32 var_7230 = const()[name = string("op_7230"), val = int32(-1)];
            bool var_7231_interleave_0 = const()[name = string("op_7231_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_7231_cast_fp16 = concat(axis = var_7230, interleave = var_7231_interleave_0, values = (cache_weights_35_cast_fp16, call_weights_35_cast_fp16))[name = string("op_7231_cast_fp16")];
            int32 var_7232 = const()[name = string("op_7232"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_35_cast_fp16 = softmax(axis = var_7232, x = var_7231_cast_fp16)[name = string("shares_35_cast_fp16")];
            tensor<int32, [4]> var_7239_begin_0 = const()[name = string("op_7239_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7239_end_0 = const()[name = string("op_7239_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_7239_end_mask_0 = const()[name = string("op_7239_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_7239_cast_fp16 = slice_by_index(begin = var_7239_begin_0, end = var_7239_end_0, end_mask = var_7239_end_mask_0, x = shares_35_cast_fp16)[name = string("op_7239_cast_fp16")];
            tensor<int32, [4]> var_7244_begin_0 = const()[name = string("op_7244_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_7244_end_0 = const()[name = string("op_7244_end_0"), val = tensor<int32, [4]>([18, 8, 4112, 128])];
            tensor<bool, [4]> var_7244_end_mask_0 = const()[name = string("op_7244_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7244_cast_fp16 = slice_by_index(begin = var_7244_begin_0, end = var_7244_end_0, end_mask = var_7244_end_mask_0, x = read_state_1)[name = string("op_7244_cast_fp16")];
            tensor<int32, [4]> var_7254_begin_0 = const()[name = string("op_7254_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7254_end_0 = const()[name = string("op_7254_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7254_end_mask_0 = const()[name = string("op_7254_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7254_cast_fp16 = slice_by_index(begin = var_7254_begin_0, end = var_7254_end_0, end_mask = var_7254_end_mask_0, x = var_7244_cast_fp16)[name = string("op_7254_cast_fp16")];
            bool var_7260_transpose_x_0 = const()[name = string("op_7260_transpose_x_0"), val = bool(false)];
            bool var_7260_transpose_y_0 = const()[name = string("op_7260_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7260_cast_fp16 = matmul(transpose_x = var_7260_transpose_x_0, transpose_y = var_7260_transpose_y_0, x = var_7239_cast_fp16, y = var_7254_cast_fp16)[name = string("op_7260_cast_fp16")];
            tensor<int32, [4]> var_7265_begin_0 = const()[name = string("op_7265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_7265_end_0 = const()[name = string("op_7265_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_7265_end_mask_0 = const()[name = string("op_7265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7265_cast_fp16 = slice_by_index(begin = var_7265_begin_0, end = var_7265_end_0, end_mask = var_7265_end_mask_0, x = shares_35_cast_fp16)[name = string("op_7265_cast_fp16")];
            bool var_7266_transpose_x_0 = const()[name = string("op_7266_transpose_x_0"), val = bool(false)];
            bool var_7266_transpose_y_0 = const()[name = string("op_7266_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_7121_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 32, 128]> var_7266_cast_fp16 = matmul(transpose_x = var_7266_transpose_x_0, transpose_y = var_7266_transpose_y_0, x = var_7265_cast_fp16, y = v_35_cast_fp16)[name = string("op_7266_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_69_cast_fp16 = add(x = var_7260_cast_fp16, y = var_7266_cast_fp16)[name = string("mixed_69_cast_fp16")];
            tensor<int32, [4]> var_7283_begin_0 = const()[name = string("op_7283_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7283_end_0 = const()[name = string("op_7283_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7283_end_mask_0 = const()[name = string("op_7283_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7283_cast_fp16 = slice_by_index(begin = var_7283_begin_0, end = var_7283_end_0, end_mask = var_7283_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_7283_cast_fp16")];
            tensor<int32, [4]> var_7298_begin_0 = const()[name = string("op_7298_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7298_end_0 = const()[name = string("op_7298_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7298_end_mask_0 = const()[name = string("op_7298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7298_cast_fp16 = slice_by_index(begin = var_7298_begin_0, end = var_7298_end_0, end_mask = var_7298_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_7298_cast_fp16")];
            int32 var_7300 = const()[name = string("op_7300"), 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_7300, interleave = mixed_71_interleave_0, values = (var_7283_cast_fp16, var_7298_cast_fp16))[name = string("mixed_71_cast_fp16")];
            tensor<int32, [4]> var_7304_perm_0 = const()[name = string("op_7304_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7308 = const()[name = string("op_7308"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7304_cast_fp16 = transpose(perm = var_7304_perm_0, x = mixed_71_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 2048]> input_173_cast_fp16 = reshape(shape = var_7308, x = var_7304_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_357_cast_fp16 = add(x = x_341_cast_fp16, y = linear_122_cast_fp16)[name = string("x_357_cast_fp16")];
            bool var_7318 = const()[name = string("op_7318"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7321_cast_fp16 = abs(x = x_357_cast_fp16)[name = string("op_7321_cast_fp16")];
            tensor<int32, [1]> peak_143_axes_0 = const()[name = string("peak_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_143_cast_fp16 = reduce_max(axes = peak_143_axes_0, keep_dims = var_7318, x = var_7321_cast_fp16)[name = string("peak_143_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_359_cast_fp16 = real_div(x = x_357_cast_fp16, y = peak_143_cast_fp16)[name = string("x_359_cast_fp16")];
            fp16 var_7317_promoted_to_fp16 = const()[name = string("op_7317_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7325_cast_fp16 = pow(x = x_359_cast_fp16, y = var_7317_promoted_to_fp16)[name = string("op_7325_cast_fp16")];
            tensor<int32, [1]> var_7327_axes_0 = const()[name = string("op_7327_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7327_keep_dims_0 = const()[name = string("op_7327_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7327_cast_fp16 = reduce_mean(axes = var_7327_axes_0, keep_dims = var_7327_keep_dims_0, x = var_7325_cast_fp16)[name = string("op_7327_cast_fp16")];
            fp16 var_7317_promoted_1_to_fp16 = const()[name = string("op_7317_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7328_cast_fp16 = pow(x = peak_143_cast_fp16, y = var_7317_promoted_1_to_fp16)[name = string("op_7328_cast_fp16")];
            fp32 var_7329_epsilon_0 = const()[name = string("op_7329_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7329_cast_fp16 = inverse(epsilon = var_7329_epsilon_0, x = var_7328_cast_fp16)[name = string("op_7329_cast_fp16")];
            fp16 var_7330_to_fp16 = const()[name = string("op_7330_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7331_cast_fp16 = mul(x = var_7329_cast_fp16, y = var_7330_to_fp16)[name = string("op_7331_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7332_cast_fp16 = add(x = var_7327_cast_fp16, y = var_7331_cast_fp16)[name = string("op_7332_cast_fp16")];
            fp32 var_7333_epsilon_0 = const()[name = string("op_7333_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7333_cast_fp16 = rsqrt(epsilon = var_7333_epsilon_0, x = var_7332_cast_fp16)[name = string("op_7333_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7334_cast_fp16 = mul(x = x_359_cast_fp16, y = var_7333_cast_fp16)[name = string("op_7334_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_7334_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_7339_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_7339_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_7339_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_361_cast_fp16 = add(x = x_357_cast_fp16, y = linear_125_cast_fp16)[name = string("x_361_cast_fp16")];
            bool var_7352 = const()[name = string("op_7352"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7355_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_7355_cast_fp16")];
            tensor<int32, [1]> peak_145_axes_0 = const()[name = string("peak_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_145_cast_fp16 = reduce_max(axes = peak_145_axes_0, keep_dims = var_7352, x = var_7355_cast_fp16)[name = string("peak_145_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_363_cast_fp16 = real_div(x = x_361_cast_fp16, y = peak_145_cast_fp16)[name = string("x_363_cast_fp16")];
            fp16 var_7351_promoted_to_fp16 = const()[name = string("op_7351_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7359_cast_fp16 = pow(x = x_363_cast_fp16, y = var_7351_promoted_to_fp16)[name = string("op_7359_cast_fp16")];
            tensor<int32, [1]> var_7361_axes_0 = const()[name = string("op_7361_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7361_keep_dims_0 = const()[name = string("op_7361_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7361_cast_fp16 = reduce_mean(axes = var_7361_axes_0, keep_dims = var_7361_keep_dims_0, x = var_7359_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7351_promoted_1_to_fp16 = const()[name = string("op_7351_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7362_cast_fp16 = pow(x = peak_145_cast_fp16, y = var_7351_promoted_1_to_fp16)[name = string("op_7362_cast_fp16")];
            fp32 var_7363_epsilon_0 = const()[name = string("op_7363_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7363_cast_fp16 = inverse(epsilon = var_7363_epsilon_0, x = var_7362_cast_fp16)[name = string("op_7363_cast_fp16")];
            fp16 var_7364_to_fp16 = const()[name = string("op_7364_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7365_cast_fp16 = mul(x = var_7363_cast_fp16, y = var_7364_to_fp16)[name = string("op_7365_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7366_cast_fp16 = add(x = var_7361_cast_fp16, y = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            fp32 var_7367_epsilon_0 = const()[name = string("op_7367_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7367_cast_fp16 = rsqrt(epsilon = var_7367_epsilon_0, x = var_7366_cast_fp16)[name = string("op_7367_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7368_cast_fp16 = mul(x = x_363_cast_fp16, y = var_7367_cast_fp16)[name = string("op_7368_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_7368_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_7377 = const()[name = string("op_7377"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7378_cast_fp16 = reshape(shape = var_7377, x = linear_126_cast_fp16)[name = string("op_7378_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7385 = const()[name = string("op_7385"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_7378_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 16, 128]> var_7388_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_7388_cast_fp16")];
            tensor<int32, [1]> peak_147_axes_0 = const()[name = string("peak_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_147_cast_fp16 = reduce_max(axes = peak_147_axes_0, keep_dims = var_7385, x = var_7388_cast_fp16)[name = string("peak_147_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_367_cast_fp16 = real_div(x = x_365_cast_fp16, y = peak_147_cast_fp16)[name = string("x_367_cast_fp16")];
            fp16 var_7384_promoted_to_fp16 = const()[name = string("op_7384_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7392_cast_fp16 = pow(x = x_367_cast_fp16, y = var_7384_promoted_to_fp16)[name = string("op_7392_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7392_cast_fp16)[name = string("op_7394_cast_fp16")];
            fp16 var_7384_promoted_1_to_fp16 = const()[name = string("op_7384_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7395_cast_fp16 = pow(x = peak_147_cast_fp16, y = var_7384_promoted_1_to_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7399_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7399_cast_fp16")];
            fp32 var_7400_epsilon_0 = const()[name = string("op_7400_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7400_cast_fp16 = rsqrt(epsilon = var_7400_epsilon_0, x = var_7399_cast_fp16)[name = string("op_7400_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7401_cast_fp16 = mul(x = x_367_cast_fp16, y = var_7400_cast_fp16)[name = string("op_7401_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_373_cast_fp16 = mul(x = var_7401_cast_fp16, y = model_layers_18_self_attn_q_norm_weight_to_fp16)[name = string("x_373_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_7410 = const()[name = string("op_7410"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7411_cast_fp16 = reshape(shape = var_7410, x = linear_127_cast_fp16)[name = string("op_7411_cast_fp16")];
            tensor<int32, [4]> x_369_perm_0 = const()[name = string("x_369_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7418 = const()[name = string("op_7418"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_369_cast_fp16 = transpose(perm = x_369_perm_0, x = var_7411_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 8, 16, 128]> var_7421_cast_fp16 = abs(x = x_369_cast_fp16)[name = string("op_7421_cast_fp16")];
            tensor<int32, [1]> peak_149_axes_0 = const()[name = string("peak_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_149_cast_fp16 = reduce_max(axes = peak_149_axes_0, keep_dims = var_7418, x = var_7421_cast_fp16)[name = string("peak_149_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_371_cast_fp16 = real_div(x = x_369_cast_fp16, y = peak_149_cast_fp16)[name = string("x_371_cast_fp16")];
            fp16 var_7417_promoted_to_fp16 = const()[name = string("op_7417_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7425_cast_fp16 = pow(x = x_371_cast_fp16, y = var_7417_promoted_to_fp16)[name = string("op_7425_cast_fp16")];
            tensor<int32, [1]> var_7427_axes_0 = const()[name = string("op_7427_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7427_keep_dims_0 = const()[name = string("op_7427_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7427_cast_fp16 = reduce_mean(axes = var_7427_axes_0, keep_dims = var_7427_keep_dims_0, x = var_7425_cast_fp16)[name = string("op_7427_cast_fp16")];
            fp16 var_7417_promoted_1_to_fp16 = const()[name = string("op_7417_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7428_cast_fp16 = pow(x = peak_149_cast_fp16, y = var_7417_promoted_1_to_fp16)[name = string("op_7428_cast_fp16")];
            fp32 var_7429_epsilon_0 = const()[name = string("op_7429_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7429_cast_fp16 = inverse(epsilon = var_7429_epsilon_0, x = var_7428_cast_fp16)[name = string("op_7429_cast_fp16")];
            fp16 var_7430_to_fp16 = const()[name = string("op_7430_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7431_cast_fp16 = mul(x = var_7429_cast_fp16, y = var_7430_to_fp16)[name = string("op_7431_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7432_cast_fp16 = add(x = var_7427_cast_fp16, y = var_7431_cast_fp16)[name = string("op_7432_cast_fp16")];
            fp32 var_7433_epsilon_0 = const()[name = string("op_7433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7433_cast_fp16 = rsqrt(epsilon = var_7433_epsilon_0, x = var_7432_cast_fp16)[name = string("op_7433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7434_cast_fp16 = mul(x = x_371_cast_fp16, y = var_7433_cast_fp16)[name = string("op_7434_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_375_cast_fp16 = mul(x = var_7434_cast_fp16, y = model_layers_18_self_attn_k_norm_weight_to_fp16)[name = string("x_375_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_7443 = const()[name = string("op_7443"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7444_cast_fp16 = reshape(shape = var_7443, x = linear_128_cast_fp16)[name = string("op_7444_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_373_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_373_cast_fp16)[name = string("second_73_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7458_cast_fp16 = mul(x = first_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7458_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7459_cast_fp16 = mul(x = second_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7459_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7461_cast_fp16 = sub(x = var_7458_cast_fp16, y = var_7459_cast_fp16)[name = string("op_7461_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7462_cast_fp16 = mul(x = second_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7463_cast_fp16 = mul(x = first_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7463_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7465_cast_fp16 = add(x = var_7462_cast_fp16, y = var_7463_cast_fp16)[name = string("op_7465_cast_fp16")];
            int32 var_7467 = const()[name = string("op_7467"), 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_7467, interleave = q_37_interleave_0, values = (var_7461_cast_fp16, var_7465_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_375_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_375_cast_fp16)[name = string("second_75_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7479_cast_fp16 = mul(x = first_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7480_cast_fp16 = mul(x = second_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7480_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7482_cast_fp16 = sub(x = var_7479_cast_fp16, y = var_7480_cast_fp16)[name = string("op_7482_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7483_cast_fp16 = mul(x = second_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7484_cast_fp16 = mul(x = first_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7484_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7486_cast_fp16 = add(x = var_7483_cast_fp16, y = var_7484_cast_fp16)[name = string("op_7486_cast_fp16")];
            int32 var_7488 = const()[name = string("op_7488"), 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_7488, interleave = k_37_interleave_0, values = (var_7482_cast_fp16, var_7486_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_7499_begin_0 = const()[name = string("op_7499_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7499_end_0 = const()[name = string("op_7499_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7499_end_mask_0 = const()[name = string("op_7499_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7499_stride_0 = const()[name = string("op_7499_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7499_cast_fp16 = slice_by_index(begin = var_7499_begin_0, end = var_7499_end_0, end_mask = var_7499_end_mask_0, stride = var_7499_stride_0, x = q_37_cast_fp16)[name = string("op_7499_cast_fp16")];
            tensor<int32, [4]> var_7509_begin_0 = const()[name = string("op_7509_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7509_end_0 = const()[name = string("op_7509_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7509_end_mask_0 = const()[name = string("op_7509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7509_stride_0 = const()[name = string("op_7509_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7509_cast_fp16 = slice_by_index(begin = var_7509_begin_0, end = var_7509_end_0, end_mask = var_7509_end_mask_0, stride = var_7509_stride_0, x = q_37_cast_fp16)[name = string("op_7509_cast_fp16")];
            int32 var_7511 = const()[name = string("op_7511"), 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_7511, interleave = stacked_37_interleave_0, values = (var_7499_cast_fp16, var_7509_cast_fp16))[name = string("stacked_37_cast_fp16")];
            tensor<int32, [4]> var_7517_begin_0 = const()[name = string("op_7517_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7517_end_0 = const()[name = string("op_7517_end_0"), val = tensor<int32, [4]>([19, 8, 4112, 128])];
            tensor<bool, [4]> var_7517_end_mask_0 = const()[name = string("op_7517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7517_cast_fp16 = slice_by_index(begin = var_7517_begin_0, end = var_7517_end_0, end_mask = var_7517_end_mask_0, x = read_state_0)[name = string("op_7517_cast_fp16")];
            tensor<int32, [4]> var_7527_begin_0 = const()[name = string("op_7527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7527_end_0 = const()[name = string("op_7527_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7527_end_mask_0 = const()[name = string("op_7527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7527_cast_fp16 = slice_by_index(begin = var_7527_begin_0, end = var_7527_end_0, end_mask = var_7527_end_mask_0, x = var_7517_cast_fp16)[name = string("op_7527_cast_fp16")];
            bool var_7536_transpose_x_1 = const()[name = string("op_7536_transpose_x_1"), val = bool(false)];
            bool var_7536_transpose_y_1 = const()[name = string("op_7536_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_7536_cast_fp16 = matmul(transpose_x = var_7536_transpose_x_1, transpose_y = var_7536_transpose_y_1, x = var_7527_cast_fp16, y = stacked_37_cast_fp16)[name = string("op_7536_cast_fp16")];
            tensor<int32, [4]> var_7539_perm_0 = const()[name = string("op_7539_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7540_to_fp16 = const()[name = string("op_7540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_7539_cast_fp16 = transpose(perm = var_7539_perm_0, x = var_7536_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 8, 32, 4096]> var_7541_cast_fp16 = mul(x = var_7539_cast_fp16, y = var_7540_to_fp16)[name = string("op_7541_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_37_cast_fp16 = add(x = var_7541_cast_fp16, y = cacheMask)[name = string("cache_weights_37_cast_fp16")];
            bool var_7547_transpose_x_1 = const()[name = string("op_7547_transpose_x_1"), val = bool(false)];
            bool var_7547_transpose_y_1 = const()[name = string("op_7547_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7547_cast_fp16 = matmul(transpose_x = var_7547_transpose_x_1, transpose_y = var_7547_transpose_y_1, x = stacked_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_7547_cast_fp16")];
            fp16 var_7548_to_fp16 = const()[name = string("op_7548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7549_cast_fp16 = mul(x = var_7547_cast_fp16, y = var_7548_to_fp16)[name = string("op_7549_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_37_cast_fp16 = add(x = var_7549_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_37_cast_fp16")];
            int32 var_7553 = const()[name = string("op_7553"), val = int32(-1)];
            bool var_7554_interleave_0 = const()[name = string("op_7554_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_7554_cast_fp16 = concat(axis = var_7553, interleave = var_7554_interleave_0, values = (cache_weights_37_cast_fp16, call_weights_37_cast_fp16))[name = string("op_7554_cast_fp16")];
            int32 var_7555 = const()[name = string("op_7555"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_37_cast_fp16 = softmax(axis = var_7555, x = var_7554_cast_fp16)[name = string("shares_37_cast_fp16")];
            tensor<int32, [4]> var_7562_begin_0 = const()[name = string("op_7562_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7562_end_0 = const()[name = string("op_7562_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_7562_end_mask_0 = const()[name = string("op_7562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_7562_cast_fp16 = slice_by_index(begin = var_7562_begin_0, end = var_7562_end_0, end_mask = var_7562_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7562_cast_fp16")];
            tensor<int32, [4]> var_7567_begin_0 = const()[name = string("op_7567_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7567_end_0 = const()[name = string("op_7567_end_0"), val = tensor<int32, [4]>([19, 8, 4112, 128])];
            tensor<bool, [4]> var_7567_end_mask_0 = const()[name = string("op_7567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7567_cast_fp16 = slice_by_index(begin = var_7567_begin_0, end = var_7567_end_0, end_mask = var_7567_end_mask_0, x = read_state_1)[name = string("op_7567_cast_fp16")];
            tensor<int32, [4]> var_7577_begin_0 = const()[name = string("op_7577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7577_end_0 = const()[name = string("op_7577_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7577_end_mask_0 = const()[name = string("op_7577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7577_cast_fp16 = slice_by_index(begin = var_7577_begin_0, end = var_7577_end_0, end_mask = var_7577_end_mask_0, x = var_7567_cast_fp16)[name = string("op_7577_cast_fp16")];
            bool var_7583_transpose_x_0 = const()[name = string("op_7583_transpose_x_0"), val = bool(false)];
            bool var_7583_transpose_y_0 = const()[name = string("op_7583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7583_cast_fp16 = matmul(transpose_x = var_7583_transpose_x_0, transpose_y = var_7583_transpose_y_0, x = var_7562_cast_fp16, y = var_7577_cast_fp16)[name = string("op_7583_cast_fp16")];
            tensor<int32, [4]> var_7588_begin_0 = const()[name = string("op_7588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_7588_end_0 = const()[name = string("op_7588_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_7588_end_mask_0 = const()[name = string("op_7588_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7588_cast_fp16 = slice_by_index(begin = var_7588_begin_0, end = var_7588_end_0, end_mask = var_7588_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7588_cast_fp16")];
            bool var_7589_transpose_x_0 = const()[name = string("op_7589_transpose_x_0"), val = bool(false)];
            bool var_7589_transpose_y_0 = const()[name = string("op_7589_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_7444_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, 32, 128]> var_7589_cast_fp16 = matmul(transpose_x = var_7589_transpose_x_0, transpose_y = var_7589_transpose_y_0, x = var_7588_cast_fp16, y = v_37_cast_fp16)[name = string("op_7589_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_73_cast_fp16 = add(x = var_7583_cast_fp16, y = var_7589_cast_fp16)[name = string("mixed_73_cast_fp16")];
            tensor<int32, [4]> var_7606_begin_0 = const()[name = string("op_7606_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7606_end_0 = const()[name = string("op_7606_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7606_end_mask_0 = const()[name = string("op_7606_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7606_cast_fp16 = slice_by_index(begin = var_7606_begin_0, end = var_7606_end_0, end_mask = var_7606_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7606_cast_fp16")];
            tensor<int32, [4]> var_7621_begin_0 = const()[name = string("op_7621_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7621_end_0 = const()[name = string("op_7621_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7621_end_mask_0 = const()[name = string("op_7621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7621_cast_fp16 = slice_by_index(begin = var_7621_begin_0, end = var_7621_end_0, end_mask = var_7621_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7621_cast_fp16")];
            int32 var_7623 = const()[name = string("op_7623"), 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_7623, interleave = mixed_75_interleave_0, values = (var_7606_cast_fp16, var_7621_cast_fp16))[name = string("mixed_75_cast_fp16")];
            tensor<int32, [4]> var_7627_perm_0 = const()[name = string("op_7627_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7631 = const()[name = string("op_7631"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7627_cast_fp16 = transpose(perm = var_7627_perm_0, x = mixed_75_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 2048]> input_183_cast_fp16 = reshape(shape = var_7631, x = var_7627_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_377_cast_fp16 = add(x = x_361_cast_fp16, y = linear_129_cast_fp16)[name = string("x_377_cast_fp16")];
            bool var_7641 = const()[name = string("op_7641"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7644_cast_fp16 = abs(x = x_377_cast_fp16)[name = string("op_7644_cast_fp16")];
            tensor<int32, [1]> peak_151_axes_0 = const()[name = string("peak_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_151_cast_fp16 = reduce_max(axes = peak_151_axes_0, keep_dims = var_7641, x = var_7644_cast_fp16)[name = string("peak_151_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_379_cast_fp16 = real_div(x = x_377_cast_fp16, y = peak_151_cast_fp16)[name = string("x_379_cast_fp16")];
            fp16 var_7640_promoted_to_fp16 = const()[name = string("op_7640_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7648_cast_fp16 = pow(x = x_379_cast_fp16, y = var_7640_promoted_to_fp16)[name = string("op_7648_cast_fp16")];
            tensor<int32, [1]> var_7650_axes_0 = const()[name = string("op_7650_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7650_keep_dims_0 = const()[name = string("op_7650_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7650_cast_fp16 = reduce_mean(axes = var_7650_axes_0, keep_dims = var_7650_keep_dims_0, x = var_7648_cast_fp16)[name = string("op_7650_cast_fp16")];
            fp16 var_7640_promoted_1_to_fp16 = const()[name = string("op_7640_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7651_cast_fp16 = pow(x = peak_151_cast_fp16, y = var_7640_promoted_1_to_fp16)[name = string("op_7651_cast_fp16")];
            fp32 var_7652_epsilon_0 = const()[name = string("op_7652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7652_cast_fp16 = inverse(epsilon = var_7652_epsilon_0, x = var_7651_cast_fp16)[name = string("op_7652_cast_fp16")];
            fp16 var_7653_to_fp16 = const()[name = string("op_7653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7654_cast_fp16 = mul(x = var_7652_cast_fp16, y = var_7653_to_fp16)[name = string("op_7654_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7655_cast_fp16 = add(x = var_7650_cast_fp16, y = var_7654_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp32 var_7656_epsilon_0 = const()[name = string("op_7656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7656_cast_fp16 = rsqrt(epsilon = var_7656_epsilon_0, x = var_7655_cast_fp16)[name = string("op_7656_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7657_cast_fp16 = mul(x = x_379_cast_fp16, y = var_7656_cast_fp16)[name = string("op_7657_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_7657_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_7662_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_7662_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_7662_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_381_cast_fp16 = add(x = x_377_cast_fp16, y = linear_132_cast_fp16)[name = string("x_381_cast_fp16")];
            bool var_7675 = const()[name = string("op_7675"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7678_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_7678_cast_fp16")];
            tensor<int32, [1]> peak_153_axes_0 = const()[name = string("peak_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_153_cast_fp16 = reduce_max(axes = peak_153_axes_0, keep_dims = var_7675, x = var_7678_cast_fp16)[name = string("peak_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_383_cast_fp16 = real_div(x = x_381_cast_fp16, y = peak_153_cast_fp16)[name = string("x_383_cast_fp16")];
            fp16 var_7674_promoted_to_fp16 = const()[name = string("op_7674_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7682_cast_fp16 = pow(x = x_383_cast_fp16, y = var_7674_promoted_to_fp16)[name = string("op_7682_cast_fp16")];
            tensor<int32, [1]> var_7684_axes_0 = const()[name = string("op_7684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7684_keep_dims_0 = const()[name = string("op_7684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7684_cast_fp16 = reduce_mean(axes = var_7684_axes_0, keep_dims = var_7684_keep_dims_0, x = var_7682_cast_fp16)[name = string("op_7684_cast_fp16")];
            fp16 var_7674_promoted_1_to_fp16 = const()[name = string("op_7674_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7685_cast_fp16 = pow(x = peak_153_cast_fp16, y = var_7674_promoted_1_to_fp16)[name = string("op_7685_cast_fp16")];
            fp32 var_7686_epsilon_0 = const()[name = string("op_7686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7686_cast_fp16 = inverse(epsilon = var_7686_epsilon_0, x = var_7685_cast_fp16)[name = string("op_7686_cast_fp16")];
            fp16 var_7687_to_fp16 = const()[name = string("op_7687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7688_cast_fp16 = mul(x = var_7686_cast_fp16, y = var_7687_to_fp16)[name = string("op_7688_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7689_cast_fp16 = add(x = var_7684_cast_fp16, y = var_7688_cast_fp16)[name = string("op_7689_cast_fp16")];
            fp32 var_7690_epsilon_0 = const()[name = string("op_7690_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7690_cast_fp16 = rsqrt(epsilon = var_7690_epsilon_0, x = var_7689_cast_fp16)[name = string("op_7690_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7691_cast_fp16 = mul(x = x_383_cast_fp16, y = var_7690_cast_fp16)[name = string("op_7691_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_7691_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_7700 = const()[name = string("op_7700"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7701_cast_fp16 = reshape(shape = var_7700, x = linear_133_cast_fp16)[name = string("op_7701_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7708 = const()[name = string("op_7708"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_7701_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 16, 16, 128]> var_7711_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_7711_cast_fp16")];
            tensor<int32, [1]> peak_155_axes_0 = const()[name = string("peak_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_155_cast_fp16 = reduce_max(axes = peak_155_axes_0, keep_dims = var_7708, x = var_7711_cast_fp16)[name = string("peak_155_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_387_cast_fp16 = real_div(x = x_385_cast_fp16, y = peak_155_cast_fp16)[name = string("x_387_cast_fp16")];
            fp16 var_7707_promoted_to_fp16 = const()[name = string("op_7707_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7715_cast_fp16 = pow(x = x_387_cast_fp16, y = var_7707_promoted_to_fp16)[name = string("op_7715_cast_fp16")];
            tensor<int32, [1]> var_7717_axes_0 = const()[name = string("op_7717_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7717_keep_dims_0 = const()[name = string("op_7717_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7717_cast_fp16 = reduce_mean(axes = var_7717_axes_0, keep_dims = var_7717_keep_dims_0, x = var_7715_cast_fp16)[name = string("op_7717_cast_fp16")];
            fp16 var_7707_promoted_1_to_fp16 = const()[name = string("op_7707_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7718_cast_fp16 = pow(x = peak_155_cast_fp16, y = var_7707_promoted_1_to_fp16)[name = string("op_7718_cast_fp16")];
            fp32 var_7719_epsilon_0 = const()[name = string("op_7719_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7719_cast_fp16 = inverse(epsilon = var_7719_epsilon_0, x = var_7718_cast_fp16)[name = string("op_7719_cast_fp16")];
            fp16 var_7720_to_fp16 = const()[name = string("op_7720_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7721_cast_fp16 = mul(x = var_7719_cast_fp16, y = var_7720_to_fp16)[name = string("op_7721_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7722_cast_fp16 = add(x = var_7717_cast_fp16, y = var_7721_cast_fp16)[name = string("op_7722_cast_fp16")];
            fp32 var_7723_epsilon_0 = const()[name = string("op_7723_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7723_cast_fp16 = rsqrt(epsilon = var_7723_epsilon_0, x = var_7722_cast_fp16)[name = string("op_7723_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7724_cast_fp16 = mul(x = x_387_cast_fp16, y = var_7723_cast_fp16)[name = string("op_7724_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_393_cast_fp16 = mul(x = var_7724_cast_fp16, y = model_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("x_393_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_7733 = const()[name = string("op_7733"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7734_cast_fp16 = reshape(shape = var_7733, x = linear_134_cast_fp16)[name = string("op_7734_cast_fp16")];
            tensor<int32, [4]> x_389_perm_0 = const()[name = string("x_389_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7741 = const()[name = string("op_7741"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_389_cast_fp16 = transpose(perm = x_389_perm_0, x = var_7734_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 8, 16, 128]> var_7744_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_7744_cast_fp16")];
            tensor<int32, [1]> peak_157_axes_0 = const()[name = string("peak_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_157_cast_fp16 = reduce_max(axes = peak_157_axes_0, keep_dims = var_7741, x = var_7744_cast_fp16)[name = string("peak_157_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_391_cast_fp16 = real_div(x = x_389_cast_fp16, y = peak_157_cast_fp16)[name = string("x_391_cast_fp16")];
            fp16 var_7740_promoted_to_fp16 = const()[name = string("op_7740_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7748_cast_fp16 = pow(x = x_391_cast_fp16, y = var_7740_promoted_to_fp16)[name = string("op_7748_cast_fp16")];
            tensor<int32, [1]> var_7750_axes_0 = const()[name = string("op_7750_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7750_keep_dims_0 = const()[name = string("op_7750_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7750_cast_fp16 = reduce_mean(axes = var_7750_axes_0, keep_dims = var_7750_keep_dims_0, x = var_7748_cast_fp16)[name = string("op_7750_cast_fp16")];
            fp16 var_7740_promoted_1_to_fp16 = const()[name = string("op_7740_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7751_cast_fp16 = pow(x = peak_157_cast_fp16, y = var_7740_promoted_1_to_fp16)[name = string("op_7751_cast_fp16")];
            fp32 var_7752_epsilon_0 = const()[name = string("op_7752_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7752_cast_fp16 = inverse(epsilon = var_7752_epsilon_0, x = var_7751_cast_fp16)[name = string("op_7752_cast_fp16")];
            fp16 var_7753_to_fp16 = const()[name = string("op_7753_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7754_cast_fp16 = mul(x = var_7752_cast_fp16, y = var_7753_to_fp16)[name = string("op_7754_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7755_cast_fp16 = add(x = var_7750_cast_fp16, y = var_7754_cast_fp16)[name = string("op_7755_cast_fp16")];
            fp32 var_7756_epsilon_0 = const()[name = string("op_7756_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7756_cast_fp16 = rsqrt(epsilon = var_7756_epsilon_0, x = var_7755_cast_fp16)[name = string("op_7756_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7757_cast_fp16 = mul(x = x_391_cast_fp16, y = var_7756_cast_fp16)[name = string("op_7757_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_395_cast_fp16 = mul(x = var_7757_cast_fp16, y = model_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("x_395_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_7766 = const()[name = string("op_7766"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7767_cast_fp16 = reshape(shape = var_7766, x = linear_135_cast_fp16)[name = string("op_7767_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_393_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_393_cast_fp16)[name = string("second_77_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7781_cast_fp16 = mul(x = first_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7781_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7782_cast_fp16 = mul(x = second_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7782_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7784_cast_fp16 = sub(x = var_7781_cast_fp16, y = var_7782_cast_fp16)[name = string("op_7784_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7785_cast_fp16 = mul(x = second_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7786_cast_fp16 = mul(x = first_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7786_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7788_cast_fp16 = add(x = var_7785_cast_fp16, y = var_7786_cast_fp16)[name = string("op_7788_cast_fp16")];
            int32 var_7790 = const()[name = string("op_7790"), 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_7790, interleave = q_39_interleave_0, values = (var_7784_cast_fp16, var_7788_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_395_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_395_cast_fp16)[name = string("second_79_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7802_cast_fp16 = mul(x = first_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7803_cast_fp16 = mul(x = second_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7803_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7805_cast_fp16 = sub(x = var_7802_cast_fp16, y = var_7803_cast_fp16)[name = string("op_7805_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7806_cast_fp16 = mul(x = second_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7807_cast_fp16 = mul(x = first_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7807_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7809_cast_fp16 = add(x = var_7806_cast_fp16, y = var_7807_cast_fp16)[name = string("op_7809_cast_fp16")];
            int32 var_7811 = const()[name = string("op_7811"), 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_7811, interleave = k_39_interleave_0, values = (var_7805_cast_fp16, var_7809_cast_fp16))[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> var_7822_begin_0 = const()[name = string("op_7822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7822_end_0 = const()[name = string("op_7822_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7822_end_mask_0 = const()[name = string("op_7822_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7822_stride_0 = const()[name = string("op_7822_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7822_cast_fp16 = slice_by_index(begin = var_7822_begin_0, end = var_7822_end_0, end_mask = var_7822_end_mask_0, stride = var_7822_stride_0, x = q_39_cast_fp16)[name = string("op_7822_cast_fp16")];
            tensor<int32, [4]> var_7832_begin_0 = const()[name = string("op_7832_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7832_end_0 = const()[name = string("op_7832_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7832_end_mask_0 = const()[name = string("op_7832_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7832_stride_0 = const()[name = string("op_7832_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7832_cast_fp16 = slice_by_index(begin = var_7832_begin_0, end = var_7832_end_0, end_mask = var_7832_end_mask_0, stride = var_7832_stride_0, x = q_39_cast_fp16)[name = string("op_7832_cast_fp16")];
            int32 var_7834 = const()[name = string("op_7834"), 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_7834, interleave = stacked_39_interleave_0, values = (var_7822_cast_fp16, var_7832_cast_fp16))[name = string("stacked_39_cast_fp16")];
            tensor<int32, [4]> var_7840_begin_0 = const()[name = string("op_7840_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7840_end_0 = const()[name = string("op_7840_end_0"), val = tensor<int32, [4]>([20, 8, 4112, 128])];
            tensor<bool, [4]> var_7840_end_mask_0 = const()[name = string("op_7840_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7840_cast_fp16 = slice_by_index(begin = var_7840_begin_0, end = var_7840_end_0, end_mask = var_7840_end_mask_0, x = read_state_0)[name = string("op_7840_cast_fp16")];
            tensor<int32, [4]> var_7850_begin_0 = const()[name = string("op_7850_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7850_end_0 = const()[name = string("op_7850_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7850_end_mask_0 = const()[name = string("op_7850_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7850_cast_fp16 = slice_by_index(begin = var_7850_begin_0, end = var_7850_end_0, end_mask = var_7850_end_mask_0, x = var_7840_cast_fp16)[name = string("op_7850_cast_fp16")];
            bool var_7859_transpose_x_1 = const()[name = string("op_7859_transpose_x_1"), val = bool(false)];
            bool var_7859_transpose_y_1 = const()[name = string("op_7859_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_7859_cast_fp16 = matmul(transpose_x = var_7859_transpose_x_1, transpose_y = var_7859_transpose_y_1, x = var_7850_cast_fp16, y = stacked_39_cast_fp16)[name = string("op_7859_cast_fp16")];
            tensor<int32, [4]> var_7862_perm_0 = const()[name = string("op_7862_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7863_to_fp16 = const()[name = string("op_7863_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_7862_cast_fp16 = transpose(perm = var_7862_perm_0, x = var_7859_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 32, 4096]> var_7864_cast_fp16 = mul(x = var_7862_cast_fp16, y = var_7863_to_fp16)[name = string("op_7864_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_39_cast_fp16 = add(x = var_7864_cast_fp16, y = cacheMask)[name = string("cache_weights_39_cast_fp16")];
            bool var_7870_transpose_x_1 = const()[name = string("op_7870_transpose_x_1"), val = bool(false)];
            bool var_7870_transpose_y_1 = const()[name = string("op_7870_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7870_cast_fp16 = matmul(transpose_x = var_7870_transpose_x_1, transpose_y = var_7870_transpose_y_1, x = stacked_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_7870_cast_fp16")];
            fp16 var_7871_to_fp16 = const()[name = string("op_7871_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7872_cast_fp16 = mul(x = var_7870_cast_fp16, y = var_7871_to_fp16)[name = string("op_7872_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_39_cast_fp16 = add(x = var_7872_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_39_cast_fp16")];
            int32 var_7876 = const()[name = string("op_7876"), val = int32(-1)];
            bool var_7877_interleave_0 = const()[name = string("op_7877_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_7877_cast_fp16 = concat(axis = var_7876, interleave = var_7877_interleave_0, values = (cache_weights_39_cast_fp16, call_weights_39_cast_fp16))[name = string("op_7877_cast_fp16")];
            int32 var_7878 = const()[name = string("op_7878"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_39_cast_fp16 = softmax(axis = var_7878, x = var_7877_cast_fp16)[name = string("shares_39_cast_fp16")];
            tensor<int32, [4]> var_7885_begin_0 = const()[name = string("op_7885_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7885_end_0 = const()[name = string("op_7885_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_7885_end_mask_0 = const()[name = string("op_7885_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_7885_cast_fp16 = slice_by_index(begin = var_7885_begin_0, end = var_7885_end_0, end_mask = var_7885_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7885_cast_fp16")];
            tensor<int32, [4]> var_7890_begin_0 = const()[name = string("op_7890_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7890_end_0 = const()[name = string("op_7890_end_0"), val = tensor<int32, [4]>([20, 8, 4112, 128])];
            tensor<bool, [4]> var_7890_end_mask_0 = const()[name = string("op_7890_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7890_cast_fp16 = slice_by_index(begin = var_7890_begin_0, end = var_7890_end_0, end_mask = var_7890_end_mask_0, x = read_state_1)[name = string("op_7890_cast_fp16")];
            tensor<int32, [4]> var_7900_begin_0 = const()[name = string("op_7900_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7900_end_0 = const()[name = string("op_7900_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7900_end_mask_0 = const()[name = string("op_7900_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7900_cast_fp16 = slice_by_index(begin = var_7900_begin_0, end = var_7900_end_0, end_mask = var_7900_end_mask_0, x = var_7890_cast_fp16)[name = string("op_7900_cast_fp16")];
            bool var_7906_transpose_x_0 = const()[name = string("op_7906_transpose_x_0"), val = bool(false)];
            bool var_7906_transpose_y_0 = const()[name = string("op_7906_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7906_cast_fp16 = matmul(transpose_x = var_7906_transpose_x_0, transpose_y = var_7906_transpose_y_0, x = var_7885_cast_fp16, y = var_7900_cast_fp16)[name = string("op_7906_cast_fp16")];
            tensor<int32, [4]> var_7911_begin_0 = const()[name = string("op_7911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_7911_end_0 = const()[name = string("op_7911_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_7911_end_mask_0 = const()[name = string("op_7911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7911_cast_fp16 = slice_by_index(begin = var_7911_begin_0, end = var_7911_end_0, end_mask = var_7911_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7911_cast_fp16")];
            bool var_7912_transpose_x_0 = const()[name = string("op_7912_transpose_x_0"), val = bool(false)];
            bool var_7912_transpose_y_0 = const()[name = string("op_7912_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_7767_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 32, 128]> var_7912_cast_fp16 = matmul(transpose_x = var_7912_transpose_x_0, transpose_y = var_7912_transpose_y_0, x = var_7911_cast_fp16, y = v_39_cast_fp16)[name = string("op_7912_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_77_cast_fp16 = add(x = var_7906_cast_fp16, y = var_7912_cast_fp16)[name = string("mixed_77_cast_fp16")];
            tensor<int32, [4]> var_7929_begin_0 = const()[name = string("op_7929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7929_end_0 = const()[name = string("op_7929_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7929_end_mask_0 = const()[name = string("op_7929_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7929_cast_fp16 = slice_by_index(begin = var_7929_begin_0, end = var_7929_end_0, end_mask = var_7929_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7929_cast_fp16")];
            tensor<int32, [4]> var_7944_begin_0 = const()[name = string("op_7944_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7944_end_0 = const()[name = string("op_7944_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7944_end_mask_0 = const()[name = string("op_7944_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7944_cast_fp16 = slice_by_index(begin = var_7944_begin_0, end = var_7944_end_0, end_mask = var_7944_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7944_cast_fp16")];
            int32 var_7946 = const()[name = string("op_7946"), 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_7946, interleave = mixed_79_interleave_0, values = (var_7929_cast_fp16, var_7944_cast_fp16))[name = string("mixed_79_cast_fp16")];
            tensor<int32, [4]> var_7950_perm_0 = const()[name = string("op_7950_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7954 = const()[name = string("op_7954"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7950_cast_fp16 = transpose(perm = var_7950_perm_0, x = mixed_79_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 16, 2048]> input_193_cast_fp16 = reshape(shape = var_7954, x = var_7950_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_397_cast_fp16 = add(x = x_381_cast_fp16, y = linear_136_cast_fp16)[name = string("x_397_cast_fp16")];
            bool var_7964 = const()[name = string("op_7964"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7967_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_7967_cast_fp16")];
            tensor<int32, [1]> peak_159_axes_0 = const()[name = string("peak_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_159_cast_fp16 = reduce_max(axes = peak_159_axes_0, keep_dims = var_7964, x = var_7967_cast_fp16)[name = string("peak_159_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_399_cast_fp16 = real_div(x = x_397_cast_fp16, y = peak_159_cast_fp16)[name = string("x_399_cast_fp16")];
            fp16 var_7963_promoted_to_fp16 = const()[name = string("op_7963_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7971_cast_fp16 = pow(x = x_399_cast_fp16, y = var_7963_promoted_to_fp16)[name = string("op_7971_cast_fp16")];
            tensor<int32, [1]> var_7973_axes_0 = const()[name = string("op_7973_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7973_keep_dims_0 = const()[name = string("op_7973_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7973_cast_fp16 = reduce_mean(axes = var_7973_axes_0, keep_dims = var_7973_keep_dims_0, x = var_7971_cast_fp16)[name = string("op_7973_cast_fp16")];
            fp16 var_7963_promoted_1_to_fp16 = const()[name = string("op_7963_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7974_cast_fp16 = pow(x = peak_159_cast_fp16, y = var_7963_promoted_1_to_fp16)[name = string("op_7974_cast_fp16")];
            fp32 var_7975_epsilon_0 = const()[name = string("op_7975_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7975_cast_fp16 = inverse(epsilon = var_7975_epsilon_0, x = var_7974_cast_fp16)[name = string("op_7975_cast_fp16")];
            fp16 var_7976_to_fp16 = const()[name = string("op_7976_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7977_cast_fp16 = mul(x = var_7975_cast_fp16, y = var_7976_to_fp16)[name = string("op_7977_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7978_cast_fp16 = add(x = var_7973_cast_fp16, y = var_7977_cast_fp16)[name = string("op_7978_cast_fp16")];
            fp32 var_7979_epsilon_0 = const()[name = string("op_7979_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7979_cast_fp16 = rsqrt(epsilon = var_7979_epsilon_0, x = var_7978_cast_fp16)[name = string("op_7979_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7980_cast_fp16 = mul(x = x_399_cast_fp16, y = var_7979_cast_fp16)[name = string("op_7980_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_7980_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_7985_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_7985_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_7985_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_401_cast_fp16 = add(x = x_397_cast_fp16, y = linear_139_cast_fp16)[name = string("x_401_cast_fp16")];
            bool var_7998 = const()[name = string("op_7998"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8001_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<int32, [1]> peak_161_axes_0 = const()[name = string("peak_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_161_cast_fp16 = reduce_max(axes = peak_161_axes_0, keep_dims = var_7998, x = var_8001_cast_fp16)[name = string("peak_161_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_403_cast_fp16 = real_div(x = x_401_cast_fp16, y = peak_161_cast_fp16)[name = string("x_403_cast_fp16")];
            fp16 var_7997_promoted_to_fp16 = const()[name = string("op_7997_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8005_cast_fp16 = pow(x = x_403_cast_fp16, y = var_7997_promoted_to_fp16)[name = string("op_8005_cast_fp16")];
            tensor<int32, [1]> var_8007_axes_0 = const()[name = string("op_8007_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8007_keep_dims_0 = const()[name = string("op_8007_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8007_cast_fp16 = reduce_mean(axes = var_8007_axes_0, keep_dims = var_8007_keep_dims_0, x = var_8005_cast_fp16)[name = string("op_8007_cast_fp16")];
            fp16 var_7997_promoted_1_to_fp16 = const()[name = string("op_7997_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8008_cast_fp16 = pow(x = peak_161_cast_fp16, y = var_7997_promoted_1_to_fp16)[name = string("op_8008_cast_fp16")];
            fp32 var_8009_epsilon_0 = const()[name = string("op_8009_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8009_cast_fp16 = inverse(epsilon = var_8009_epsilon_0, x = var_8008_cast_fp16)[name = string("op_8009_cast_fp16")];
            fp16 var_8010_to_fp16 = const()[name = string("op_8010_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8011_cast_fp16 = mul(x = var_8009_cast_fp16, y = var_8010_to_fp16)[name = string("op_8011_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8012_cast_fp16 = add(x = var_8007_cast_fp16, y = var_8011_cast_fp16)[name = string("op_8012_cast_fp16")];
            fp32 var_8013_epsilon_0 = const()[name = string("op_8013_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8013_cast_fp16 = rsqrt(epsilon = var_8013_epsilon_0, x = var_8012_cast_fp16)[name = string("op_8013_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8014_cast_fp16 = mul(x = x_403_cast_fp16, y = var_8013_cast_fp16)[name = string("op_8014_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_8014_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_8023 = const()[name = string("op_8023"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8024_cast_fp16 = reshape(shape = var_8023, x = linear_140_cast_fp16)[name = string("op_8024_cast_fp16")];
            tensor<int32, [4]> x_405_perm_0 = const()[name = string("x_405_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8031 = const()[name = string("op_8031"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_405_cast_fp16 = transpose(perm = x_405_perm_0, x = var_8024_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 16, 128]> var_8034_cast_fp16 = abs(x = x_405_cast_fp16)[name = string("op_8034_cast_fp16")];
            tensor<int32, [1]> peak_163_axes_0 = const()[name = string("peak_163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_163_cast_fp16 = reduce_max(axes = peak_163_axes_0, keep_dims = var_8031, x = var_8034_cast_fp16)[name = string("peak_163_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_407_cast_fp16 = real_div(x = x_405_cast_fp16, y = peak_163_cast_fp16)[name = string("x_407_cast_fp16")];
            fp16 var_8030_promoted_to_fp16 = const()[name = string("op_8030_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8038_cast_fp16 = pow(x = x_407_cast_fp16, y = var_8030_promoted_to_fp16)[name = string("op_8038_cast_fp16")];
            tensor<int32, [1]> var_8040_axes_0 = const()[name = string("op_8040_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8040_keep_dims_0 = const()[name = string("op_8040_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8040_cast_fp16 = reduce_mean(axes = var_8040_axes_0, keep_dims = var_8040_keep_dims_0, x = var_8038_cast_fp16)[name = string("op_8040_cast_fp16")];
            fp16 var_8030_promoted_1_to_fp16 = const()[name = string("op_8030_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8041_cast_fp16 = pow(x = peak_163_cast_fp16, y = var_8030_promoted_1_to_fp16)[name = string("op_8041_cast_fp16")];
            fp32 var_8042_epsilon_0 = const()[name = string("op_8042_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8042_cast_fp16 = inverse(epsilon = var_8042_epsilon_0, x = var_8041_cast_fp16)[name = string("op_8042_cast_fp16")];
            fp16 var_8043_to_fp16 = const()[name = string("op_8043_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8044_cast_fp16 = mul(x = var_8042_cast_fp16, y = var_8043_to_fp16)[name = string("op_8044_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8045_cast_fp16 = add(x = var_8040_cast_fp16, y = var_8044_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp32 var_8046_epsilon_0 = const()[name = string("op_8046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8046_cast_fp16 = rsqrt(epsilon = var_8046_epsilon_0, x = var_8045_cast_fp16)[name = string("op_8046_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8047_cast_fp16 = mul(x = x_407_cast_fp16, y = var_8046_cast_fp16)[name = string("op_8047_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_413_cast_fp16 = mul(x = var_8047_cast_fp16, y = model_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("x_413_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_8056 = const()[name = string("op_8056"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8057_cast_fp16 = reshape(shape = var_8056, x = linear_141_cast_fp16)[name = string("op_8057_cast_fp16")];
            tensor<int32, [4]> x_409_perm_0 = const()[name = string("x_409_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8064 = const()[name = string("op_8064"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_409_cast_fp16 = transpose(perm = x_409_perm_0, x = var_8057_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, 16, 128]> var_8067_cast_fp16 = abs(x = x_409_cast_fp16)[name = string("op_8067_cast_fp16")];
            tensor<int32, [1]> peak_165_axes_0 = const()[name = string("peak_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_165_cast_fp16 = reduce_max(axes = peak_165_axes_0, keep_dims = var_8064, x = var_8067_cast_fp16)[name = string("peak_165_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_411_cast_fp16 = real_div(x = x_409_cast_fp16, y = peak_165_cast_fp16)[name = string("x_411_cast_fp16")];
            fp16 var_8063_promoted_to_fp16 = const()[name = string("op_8063_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8071_cast_fp16 = pow(x = x_411_cast_fp16, y = var_8063_promoted_to_fp16)[name = string("op_8071_cast_fp16")];
            tensor<int32, [1]> var_8073_axes_0 = const()[name = string("op_8073_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8073_keep_dims_0 = const()[name = string("op_8073_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8073_cast_fp16 = reduce_mean(axes = var_8073_axes_0, keep_dims = var_8073_keep_dims_0, x = var_8071_cast_fp16)[name = string("op_8073_cast_fp16")];
            fp16 var_8063_promoted_1_to_fp16 = const()[name = string("op_8063_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8074_cast_fp16 = pow(x = peak_165_cast_fp16, y = var_8063_promoted_1_to_fp16)[name = string("op_8074_cast_fp16")];
            fp32 var_8075_epsilon_0 = const()[name = string("op_8075_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8075_cast_fp16 = inverse(epsilon = var_8075_epsilon_0, x = var_8074_cast_fp16)[name = string("op_8075_cast_fp16")];
            fp16 var_8076_to_fp16 = const()[name = string("op_8076_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8077_cast_fp16 = mul(x = var_8075_cast_fp16, y = var_8076_to_fp16)[name = string("op_8077_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8078_cast_fp16 = add(x = var_8073_cast_fp16, y = var_8077_cast_fp16)[name = string("op_8078_cast_fp16")];
            fp32 var_8079_epsilon_0 = const()[name = string("op_8079_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8079_cast_fp16 = rsqrt(epsilon = var_8079_epsilon_0, x = var_8078_cast_fp16)[name = string("op_8079_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8080_cast_fp16 = mul(x = x_411_cast_fp16, y = var_8079_cast_fp16)[name = string("op_8080_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_415_cast_fp16 = mul(x = var_8080_cast_fp16, y = model_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("x_415_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_8089 = const()[name = string("op_8089"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8090_cast_fp16 = reshape(shape = var_8089, x = linear_142_cast_fp16)[name = string("op_8090_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_413_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_413_cast_fp16)[name = string("second_81_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8104_cast_fp16 = mul(x = first_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8104_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8105_cast_fp16 = mul(x = second_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8107_cast_fp16 = sub(x = var_8104_cast_fp16, y = var_8105_cast_fp16)[name = string("op_8107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8108_cast_fp16 = mul(x = second_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8108_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8109_cast_fp16 = mul(x = first_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8111_cast_fp16 = add(x = var_8108_cast_fp16, y = var_8109_cast_fp16)[name = string("op_8111_cast_fp16")];
            int32 var_8113 = const()[name = string("op_8113"), 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_8113, interleave = q_41_interleave_0, values = (var_8107_cast_fp16, var_8111_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_415_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_415_cast_fp16)[name = string("second_83_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8125_cast_fp16 = mul(x = first_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8125_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8126_cast_fp16 = mul(x = second_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8126_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8128_cast_fp16 = sub(x = var_8125_cast_fp16, y = var_8126_cast_fp16)[name = string("op_8128_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8129_cast_fp16 = mul(x = second_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8129_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8130_cast_fp16 = mul(x = first_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8130_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8132_cast_fp16 = add(x = var_8129_cast_fp16, y = var_8130_cast_fp16)[name = string("op_8132_cast_fp16")];
            int32 var_8134 = const()[name = string("op_8134"), 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_8134, interleave = k_41_interleave_0, values = (var_8128_cast_fp16, var_8132_cast_fp16))[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_8145_begin_0 = const()[name = string("op_8145_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8145_end_0 = const()[name = string("op_8145_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8145_end_mask_0 = const()[name = string("op_8145_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8145_stride_0 = const()[name = string("op_8145_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8145_cast_fp16 = slice_by_index(begin = var_8145_begin_0, end = var_8145_end_0, end_mask = var_8145_end_mask_0, stride = var_8145_stride_0, x = q_41_cast_fp16)[name = string("op_8145_cast_fp16")];
            tensor<int32, [4]> var_8155_begin_0 = const()[name = string("op_8155_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8155_end_0 = const()[name = string("op_8155_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8155_end_mask_0 = const()[name = string("op_8155_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8155_stride_0 = const()[name = string("op_8155_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8155_cast_fp16 = slice_by_index(begin = var_8155_begin_0, end = var_8155_end_0, end_mask = var_8155_end_mask_0, stride = var_8155_stride_0, x = q_41_cast_fp16)[name = string("op_8155_cast_fp16")];
            int32 var_8157 = const()[name = string("op_8157"), 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_8157, interleave = stacked_41_interleave_0, values = (var_8145_cast_fp16, var_8155_cast_fp16))[name = string("stacked_41_cast_fp16")];
            tensor<int32, [4]> var_8163_begin_0 = const()[name = string("op_8163_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_8163_end_0 = const()[name = string("op_8163_end_0"), val = tensor<int32, [4]>([21, 8, 4112, 128])];
            tensor<bool, [4]> var_8163_end_mask_0 = const()[name = string("op_8163_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, x = read_state_0)[name = string("op_8163_cast_fp16")];
            tensor<int32, [4]> var_8173_begin_0 = const()[name = string("op_8173_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8173_end_0 = const()[name = string("op_8173_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8173_end_mask_0 = const()[name = string("op_8173_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8173_cast_fp16 = slice_by_index(begin = var_8173_begin_0, end = var_8173_end_0, end_mask = var_8173_end_mask_0, x = var_8163_cast_fp16)[name = string("op_8173_cast_fp16")];
            bool var_8182_transpose_x_1 = const()[name = string("op_8182_transpose_x_1"), val = bool(false)];
            bool var_8182_transpose_y_1 = const()[name = string("op_8182_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_8182_cast_fp16 = matmul(transpose_x = var_8182_transpose_x_1, transpose_y = var_8182_transpose_y_1, x = var_8173_cast_fp16, y = stacked_41_cast_fp16)[name = string("op_8182_cast_fp16")];
            tensor<int32, [4]> var_8185_perm_0 = const()[name = string("op_8185_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8186_to_fp16 = const()[name = string("op_8186_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_8185_cast_fp16 = transpose(perm = var_8185_perm_0, x = var_8182_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 32, 4096]> var_8187_cast_fp16 = mul(x = var_8185_cast_fp16, y = var_8186_to_fp16)[name = string("op_8187_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_41_cast_fp16 = add(x = var_8187_cast_fp16, y = cacheMask)[name = string("cache_weights_41_cast_fp16")];
            bool var_8193_transpose_x_1 = const()[name = string("op_8193_transpose_x_1"), val = bool(false)];
            bool var_8193_transpose_y_1 = const()[name = string("op_8193_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8193_cast_fp16 = matmul(transpose_x = var_8193_transpose_x_1, transpose_y = var_8193_transpose_y_1, x = stacked_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_41_cast_fp16 = add(x = var_8195_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_41_cast_fp16")];
            int32 var_8199 = const()[name = string("op_8199"), val = int32(-1)];
            bool var_8200_interleave_0 = const()[name = string("op_8200_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_8200_cast_fp16 = concat(axis = var_8199, interleave = var_8200_interleave_0, values = (cache_weights_41_cast_fp16, call_weights_41_cast_fp16))[name = string("op_8200_cast_fp16")];
            int32 var_8201 = const()[name = string("op_8201"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_41_cast_fp16 = softmax(axis = var_8201, x = var_8200_cast_fp16)[name = string("shares_41_cast_fp16")];
            tensor<int32, [4]> var_8208_begin_0 = const()[name = string("op_8208_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8208_end_0 = const()[name = string("op_8208_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_8208_end_mask_0 = const()[name = string("op_8208_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_8208_cast_fp16 = slice_by_index(begin = var_8208_begin_0, end = var_8208_end_0, end_mask = var_8208_end_mask_0, x = shares_41_cast_fp16)[name = string("op_8208_cast_fp16")];
            tensor<int32, [4]> var_8213_begin_0 = const()[name = string("op_8213_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_8213_end_0 = const()[name = string("op_8213_end_0"), val = tensor<int32, [4]>([21, 8, 4112, 128])];
            tensor<bool, [4]> var_8213_end_mask_0 = const()[name = string("op_8213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8213_cast_fp16 = slice_by_index(begin = var_8213_begin_0, end = var_8213_end_0, end_mask = var_8213_end_mask_0, x = read_state_1)[name = string("op_8213_cast_fp16")];
            tensor<int32, [4]> var_8223_begin_0 = const()[name = string("op_8223_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8223_end_0 = const()[name = string("op_8223_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8223_end_mask_0 = const()[name = string("op_8223_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8223_cast_fp16 = slice_by_index(begin = var_8223_begin_0, end = var_8223_end_0, end_mask = var_8223_end_mask_0, x = var_8213_cast_fp16)[name = string("op_8223_cast_fp16")];
            bool var_8229_transpose_x_0 = const()[name = string("op_8229_transpose_x_0"), val = bool(false)];
            bool var_8229_transpose_y_0 = const()[name = string("op_8229_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8229_cast_fp16 = matmul(transpose_x = var_8229_transpose_x_0, transpose_y = var_8229_transpose_y_0, x = var_8208_cast_fp16, y = var_8223_cast_fp16)[name = string("op_8229_cast_fp16")];
            tensor<int32, [4]> var_8234_begin_0 = const()[name = string("op_8234_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_8234_end_0 = const()[name = string("op_8234_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_8234_end_mask_0 = const()[name = string("op_8234_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8234_cast_fp16 = slice_by_index(begin = var_8234_begin_0, end = var_8234_end_0, end_mask = var_8234_end_mask_0, x = shares_41_cast_fp16)[name = string("op_8234_cast_fp16")];
            bool var_8235_transpose_x_0 = const()[name = string("op_8235_transpose_x_0"), val = bool(false)];
            bool var_8235_transpose_y_0 = const()[name = string("op_8235_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_8090_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 8, 32, 128]> var_8235_cast_fp16 = matmul(transpose_x = var_8235_transpose_x_0, transpose_y = var_8235_transpose_y_0, x = var_8234_cast_fp16, y = v_41_cast_fp16)[name = string("op_8235_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_81_cast_fp16 = add(x = var_8229_cast_fp16, y = var_8235_cast_fp16)[name = string("mixed_81_cast_fp16")];
            tensor<int32, [4]> var_8252_begin_0 = const()[name = string("op_8252_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8252_end_0 = const()[name = string("op_8252_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8252_end_mask_0 = const()[name = string("op_8252_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8252_cast_fp16 = slice_by_index(begin = var_8252_begin_0, end = var_8252_end_0, end_mask = var_8252_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8267_begin_0 = const()[name = string("op_8267_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8267_end_0 = const()[name = string("op_8267_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8267_end_mask_0 = const()[name = string("op_8267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8267_cast_fp16 = slice_by_index(begin = var_8267_begin_0, end = var_8267_end_0, end_mask = var_8267_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_8267_cast_fp16")];
            int32 var_8269 = const()[name = string("op_8269"), 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_8269, interleave = mixed_83_interleave_0, values = (var_8252_cast_fp16, var_8267_cast_fp16))[name = string("mixed_83_cast_fp16")];
            tensor<int32, [4]> var_8273_perm_0 = const()[name = string("op_8273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8277 = const()[name = string("op_8277"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8273_cast_fp16 = transpose(perm = var_8273_perm_0, x = mixed_83_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 2048]> input_203_cast_fp16 = reshape(shape = var_8277, x = var_8273_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_417_cast_fp16 = add(x = x_401_cast_fp16, y = linear_143_cast_fp16)[name = string("x_417_cast_fp16")];
            bool var_8287 = const()[name = string("op_8287"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8290_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_8290_cast_fp16")];
            tensor<int32, [1]> peak_167_axes_0 = const()[name = string("peak_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_167_cast_fp16 = reduce_max(axes = peak_167_axes_0, keep_dims = var_8287, x = var_8290_cast_fp16)[name = string("peak_167_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_419_cast_fp16 = real_div(x = x_417_cast_fp16, y = peak_167_cast_fp16)[name = string("x_419_cast_fp16")];
            fp16 var_8286_promoted_to_fp16 = const()[name = string("op_8286_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8294_cast_fp16 = pow(x = x_419_cast_fp16, y = var_8286_promoted_to_fp16)[name = string("op_8294_cast_fp16")];
            tensor<int32, [1]> var_8296_axes_0 = const()[name = string("op_8296_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8296_keep_dims_0 = const()[name = string("op_8296_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8296_cast_fp16 = reduce_mean(axes = var_8296_axes_0, keep_dims = var_8296_keep_dims_0, x = var_8294_cast_fp16)[name = string("op_8296_cast_fp16")];
            fp16 var_8286_promoted_1_to_fp16 = const()[name = string("op_8286_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8297_cast_fp16 = pow(x = peak_167_cast_fp16, y = var_8286_promoted_1_to_fp16)[name = string("op_8297_cast_fp16")];
            fp32 var_8298_epsilon_0 = const()[name = string("op_8298_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8298_cast_fp16 = inverse(epsilon = var_8298_epsilon_0, x = var_8297_cast_fp16)[name = string("op_8298_cast_fp16")];
            fp16 var_8299_to_fp16 = const()[name = string("op_8299_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8300_cast_fp16 = mul(x = var_8298_cast_fp16, y = var_8299_to_fp16)[name = string("op_8300_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8301_cast_fp16 = add(x = var_8296_cast_fp16, y = var_8300_cast_fp16)[name = string("op_8301_cast_fp16")];
            fp32 var_8302_epsilon_0 = const()[name = string("op_8302_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8302_cast_fp16 = rsqrt(epsilon = var_8302_epsilon_0, x = var_8301_cast_fp16)[name = string("op_8302_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8303_cast_fp16 = mul(x = x_419_cast_fp16, y = var_8302_cast_fp16)[name = string("op_8303_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_8303_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_8308_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_8308_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_8308_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_421_cast_fp16 = add(x = x_417_cast_fp16, y = linear_146_cast_fp16)[name = string("x_421_cast_fp16")];
            bool var_8321 = const()[name = string("op_8321"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8324_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<int32, [1]> peak_169_axes_0 = const()[name = string("peak_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_169_cast_fp16 = reduce_max(axes = peak_169_axes_0, keep_dims = var_8321, x = var_8324_cast_fp16)[name = string("peak_169_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_423_cast_fp16 = real_div(x = x_421_cast_fp16, y = peak_169_cast_fp16)[name = string("x_423_cast_fp16")];
            fp16 var_8320_promoted_to_fp16 = const()[name = string("op_8320_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8328_cast_fp16 = pow(x = x_423_cast_fp16, y = var_8320_promoted_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<int32, [1]> var_8330_axes_0 = const()[name = string("op_8330_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8330_keep_dims_0 = const()[name = string("op_8330_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8330_cast_fp16 = reduce_mean(axes = var_8330_axes_0, keep_dims = var_8330_keep_dims_0, x = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp16 var_8320_promoted_1_to_fp16 = const()[name = string("op_8320_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8331_cast_fp16 = pow(x = peak_169_cast_fp16, y = var_8320_promoted_1_to_fp16)[name = string("op_8331_cast_fp16")];
            fp32 var_8332_epsilon_0 = const()[name = string("op_8332_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8332_cast_fp16 = inverse(epsilon = var_8332_epsilon_0, x = var_8331_cast_fp16)[name = string("op_8332_cast_fp16")];
            fp16 var_8333_to_fp16 = const()[name = string("op_8333_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8334_cast_fp16 = mul(x = var_8332_cast_fp16, y = var_8333_to_fp16)[name = string("op_8334_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8335_cast_fp16 = add(x = var_8330_cast_fp16, y = var_8334_cast_fp16)[name = string("op_8335_cast_fp16")];
            fp32 var_8336_epsilon_0 = const()[name = string("op_8336_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8336_cast_fp16 = rsqrt(epsilon = var_8336_epsilon_0, x = var_8335_cast_fp16)[name = string("op_8336_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8337_cast_fp16 = mul(x = x_423_cast_fp16, y = var_8336_cast_fp16)[name = string("op_8337_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_8337_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_8346 = const()[name = string("op_8346"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8347_cast_fp16 = reshape(shape = var_8346, x = linear_147_cast_fp16)[name = string("op_8347_cast_fp16")];
            tensor<int32, [4]> x_425_perm_0 = const()[name = string("x_425_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8354 = const()[name = string("op_8354"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_425_cast_fp16 = transpose(perm = x_425_perm_0, x = var_8347_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 16, 128]> var_8357_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_8357_cast_fp16")];
            tensor<int32, [1]> peak_171_axes_0 = const()[name = string("peak_171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_171_cast_fp16 = reduce_max(axes = peak_171_axes_0, keep_dims = var_8354, x = var_8357_cast_fp16)[name = string("peak_171_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_427_cast_fp16 = real_div(x = x_425_cast_fp16, y = peak_171_cast_fp16)[name = string("x_427_cast_fp16")];
            fp16 var_8353_promoted_to_fp16 = const()[name = string("op_8353_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8361_cast_fp16 = pow(x = x_427_cast_fp16, y = var_8353_promoted_to_fp16)[name = string("op_8361_cast_fp16")];
            tensor<int32, [1]> var_8363_axes_0 = const()[name = string("op_8363_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8363_keep_dims_0 = const()[name = string("op_8363_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8363_cast_fp16 = reduce_mean(axes = var_8363_axes_0, keep_dims = var_8363_keep_dims_0, x = var_8361_cast_fp16)[name = string("op_8363_cast_fp16")];
            fp16 var_8353_promoted_1_to_fp16 = const()[name = string("op_8353_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8364_cast_fp16 = pow(x = peak_171_cast_fp16, y = var_8353_promoted_1_to_fp16)[name = string("op_8364_cast_fp16")];
            fp32 var_8365_epsilon_0 = const()[name = string("op_8365_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8365_cast_fp16 = inverse(epsilon = var_8365_epsilon_0, x = var_8364_cast_fp16)[name = string("op_8365_cast_fp16")];
            fp16 var_8366_to_fp16 = const()[name = string("op_8366_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8367_cast_fp16 = mul(x = var_8365_cast_fp16, y = var_8366_to_fp16)[name = string("op_8367_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8368_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8367_cast_fp16)[name = string("op_8368_cast_fp16")];
            fp32 var_8369_epsilon_0 = const()[name = string("op_8369_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8369_cast_fp16 = rsqrt(epsilon = var_8369_epsilon_0, x = var_8368_cast_fp16)[name = string("op_8369_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8370_cast_fp16 = mul(x = x_427_cast_fp16, y = var_8369_cast_fp16)[name = string("op_8370_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_433_cast_fp16 = mul(x = var_8370_cast_fp16, y = model_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("x_433_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_8379 = const()[name = string("op_8379"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8380_cast_fp16 = reshape(shape = var_8379, x = linear_148_cast_fp16)[name = string("op_8380_cast_fp16")];
            tensor<int32, [4]> x_429_perm_0 = const()[name = string("x_429_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8387 = const()[name = string("op_8387"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_429_cast_fp16 = transpose(perm = x_429_perm_0, x = var_8380_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 16, 128]> var_8390_cast_fp16 = abs(x = x_429_cast_fp16)[name = string("op_8390_cast_fp16")];
            tensor<int32, [1]> peak_173_axes_0 = const()[name = string("peak_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_173_cast_fp16 = reduce_max(axes = peak_173_axes_0, keep_dims = var_8387, x = var_8390_cast_fp16)[name = string("peak_173_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_431_cast_fp16 = real_div(x = x_429_cast_fp16, y = peak_173_cast_fp16)[name = string("x_431_cast_fp16")];
            fp16 var_8386_promoted_to_fp16 = const()[name = string("op_8386_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8394_cast_fp16 = pow(x = x_431_cast_fp16, y = var_8386_promoted_to_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [1]> var_8396_axes_0 = const()[name = string("op_8396_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8396_keep_dims_0 = const()[name = string("op_8396_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8396_cast_fp16 = reduce_mean(axes = var_8396_axes_0, keep_dims = var_8396_keep_dims_0, x = var_8394_cast_fp16)[name = string("op_8396_cast_fp16")];
            fp16 var_8386_promoted_1_to_fp16 = const()[name = string("op_8386_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8397_cast_fp16 = pow(x = peak_173_cast_fp16, y = var_8386_promoted_1_to_fp16)[name = string("op_8397_cast_fp16")];
            fp32 var_8398_epsilon_0 = const()[name = string("op_8398_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8398_cast_fp16 = inverse(epsilon = var_8398_epsilon_0, x = var_8397_cast_fp16)[name = string("op_8398_cast_fp16")];
            fp16 var_8399_to_fp16 = const()[name = string("op_8399_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8400_cast_fp16 = mul(x = var_8398_cast_fp16, y = var_8399_to_fp16)[name = string("op_8400_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8401_cast_fp16 = add(x = var_8396_cast_fp16, y = var_8400_cast_fp16)[name = string("op_8401_cast_fp16")];
            fp32 var_8402_epsilon_0 = const()[name = string("op_8402_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8402_cast_fp16 = rsqrt(epsilon = var_8402_epsilon_0, x = var_8401_cast_fp16)[name = string("op_8402_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8403_cast_fp16 = mul(x = x_431_cast_fp16, y = var_8402_cast_fp16)[name = string("op_8403_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_435_cast_fp16 = mul(x = var_8403_cast_fp16, y = model_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("x_435_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_8412 = const()[name = string("op_8412"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8413_cast_fp16 = reshape(shape = var_8412, x = linear_149_cast_fp16)[name = string("op_8413_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_433_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_433_cast_fp16)[name = string("second_85_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8427_cast_fp16 = mul(x = first_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8427_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8428_cast_fp16 = mul(x = second_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8428_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8430_cast_fp16 = sub(x = var_8427_cast_fp16, y = var_8428_cast_fp16)[name = string("op_8430_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8431_cast_fp16 = mul(x = second_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8432_cast_fp16 = mul(x = first_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8432_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8434_cast_fp16 = add(x = var_8431_cast_fp16, y = var_8432_cast_fp16)[name = string("op_8434_cast_fp16")];
            int32 var_8436 = const()[name = string("op_8436"), 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_8436, interleave = q_43_interleave_0, values = (var_8430_cast_fp16, var_8434_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_435_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_435_cast_fp16)[name = string("second_87_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8448_cast_fp16 = mul(x = first_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8448_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8449_cast_fp16 = mul(x = second_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8449_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8451_cast_fp16 = sub(x = var_8448_cast_fp16, y = var_8449_cast_fp16)[name = string("op_8451_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8452_cast_fp16 = mul(x = second_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8452_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8453_cast_fp16 = mul(x = first_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8453_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8455_cast_fp16 = add(x = var_8452_cast_fp16, y = var_8453_cast_fp16)[name = string("op_8455_cast_fp16")];
            int32 var_8457 = const()[name = string("op_8457"), 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_8457, interleave = k_43_interleave_0, values = (var_8451_cast_fp16, var_8455_cast_fp16))[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> var_8468_begin_0 = const()[name = string("op_8468_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8468_end_0 = const()[name = string("op_8468_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8468_end_mask_0 = const()[name = string("op_8468_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8468_stride_0 = const()[name = string("op_8468_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8468_cast_fp16 = slice_by_index(begin = var_8468_begin_0, end = var_8468_end_0, end_mask = var_8468_end_mask_0, stride = var_8468_stride_0, x = q_43_cast_fp16)[name = string("op_8468_cast_fp16")];
            tensor<int32, [4]> var_8478_begin_0 = const()[name = string("op_8478_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8478_end_0 = const()[name = string("op_8478_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8478_end_mask_0 = const()[name = string("op_8478_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8478_stride_0 = const()[name = string("op_8478_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8478_cast_fp16 = slice_by_index(begin = var_8478_begin_0, end = var_8478_end_0, end_mask = var_8478_end_mask_0, stride = var_8478_stride_0, x = q_43_cast_fp16)[name = string("op_8478_cast_fp16")];
            int32 var_8480 = const()[name = string("op_8480"), 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_8480, interleave = stacked_43_interleave_0, values = (var_8468_cast_fp16, var_8478_cast_fp16))[name = string("stacked_43_cast_fp16")];
            tensor<int32, [4]> var_8486_begin_0 = const()[name = string("op_8486_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_8486_end_0 = const()[name = string("op_8486_end_0"), val = tensor<int32, [4]>([22, 8, 4112, 128])];
            tensor<bool, [4]> var_8486_end_mask_0 = const()[name = string("op_8486_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8486_cast_fp16 = slice_by_index(begin = var_8486_begin_0, end = var_8486_end_0, end_mask = var_8486_end_mask_0, x = read_state_0)[name = string("op_8486_cast_fp16")];
            tensor<int32, [4]> var_8496_begin_0 = const()[name = string("op_8496_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8496_end_0 = const()[name = string("op_8496_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8496_end_mask_0 = const()[name = string("op_8496_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8496_cast_fp16 = slice_by_index(begin = var_8496_begin_0, end = var_8496_end_0, end_mask = var_8496_end_mask_0, x = var_8486_cast_fp16)[name = string("op_8496_cast_fp16")];
            bool var_8505_transpose_x_1 = const()[name = string("op_8505_transpose_x_1"), val = bool(false)];
            bool var_8505_transpose_y_1 = const()[name = string("op_8505_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_8505_cast_fp16 = matmul(transpose_x = var_8505_transpose_x_1, transpose_y = var_8505_transpose_y_1, x = var_8496_cast_fp16, y = stacked_43_cast_fp16)[name = string("op_8505_cast_fp16")];
            tensor<int32, [4]> var_8508_perm_0 = const()[name = string("op_8508_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8509_to_fp16 = const()[name = string("op_8509_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_8508_cast_fp16 = transpose(perm = var_8508_perm_0, x = var_8505_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 32, 4096]> var_8510_cast_fp16 = mul(x = var_8508_cast_fp16, y = var_8509_to_fp16)[name = string("op_8510_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_43_cast_fp16 = add(x = var_8510_cast_fp16, y = cacheMask)[name = string("cache_weights_43_cast_fp16")];
            bool var_8516_transpose_x_1 = const()[name = string("op_8516_transpose_x_1"), val = bool(false)];
            bool var_8516_transpose_y_1 = const()[name = string("op_8516_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8516_cast_fp16 = matmul(transpose_x = var_8516_transpose_x_1, transpose_y = var_8516_transpose_y_1, x = stacked_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_8516_cast_fp16")];
            fp16 var_8517_to_fp16 = const()[name = string("op_8517_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8518_cast_fp16 = mul(x = var_8516_cast_fp16, y = var_8517_to_fp16)[name = string("op_8518_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_43_cast_fp16 = add(x = var_8518_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_43_cast_fp16")];
            int32 var_8522 = const()[name = string("op_8522"), val = int32(-1)];
            bool var_8523_interleave_0 = const()[name = string("op_8523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_8523_cast_fp16 = concat(axis = var_8522, interleave = var_8523_interleave_0, values = (cache_weights_43_cast_fp16, call_weights_43_cast_fp16))[name = string("op_8523_cast_fp16")];
            int32 var_8524 = const()[name = string("op_8524"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_43_cast_fp16 = softmax(axis = var_8524, x = var_8523_cast_fp16)[name = string("shares_43_cast_fp16")];
            tensor<int32, [4]> var_8531_begin_0 = const()[name = string("op_8531_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8531_end_0 = const()[name = string("op_8531_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_8531_end_mask_0 = const()[name = string("op_8531_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_8531_cast_fp16 = slice_by_index(begin = var_8531_begin_0, end = var_8531_end_0, end_mask = var_8531_end_mask_0, x = shares_43_cast_fp16)[name = string("op_8531_cast_fp16")];
            tensor<int32, [4]> var_8536_begin_0 = const()[name = string("op_8536_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_8536_end_0 = const()[name = string("op_8536_end_0"), val = tensor<int32, [4]>([22, 8, 4112, 128])];
            tensor<bool, [4]> var_8536_end_mask_0 = const()[name = string("op_8536_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8536_cast_fp16 = slice_by_index(begin = var_8536_begin_0, end = var_8536_end_0, end_mask = var_8536_end_mask_0, x = read_state_1)[name = string("op_8536_cast_fp16")];
            tensor<int32, [4]> var_8546_begin_0 = const()[name = string("op_8546_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8546_end_0 = const()[name = string("op_8546_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8546_end_mask_0 = const()[name = string("op_8546_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8546_cast_fp16 = slice_by_index(begin = var_8546_begin_0, end = var_8546_end_0, end_mask = var_8546_end_mask_0, x = var_8536_cast_fp16)[name = string("op_8546_cast_fp16")];
            bool var_8552_transpose_x_0 = const()[name = string("op_8552_transpose_x_0"), val = bool(false)];
            bool var_8552_transpose_y_0 = const()[name = string("op_8552_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8552_cast_fp16 = matmul(transpose_x = var_8552_transpose_x_0, transpose_y = var_8552_transpose_y_0, x = var_8531_cast_fp16, y = var_8546_cast_fp16)[name = string("op_8552_cast_fp16")];
            tensor<int32, [4]> var_8557_begin_0 = const()[name = string("op_8557_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_8557_end_0 = const()[name = string("op_8557_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_8557_end_mask_0 = const()[name = string("op_8557_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8557_cast_fp16 = slice_by_index(begin = var_8557_begin_0, end = var_8557_end_0, end_mask = var_8557_end_mask_0, x = shares_43_cast_fp16)[name = string("op_8557_cast_fp16")];
            bool var_8558_transpose_x_0 = const()[name = string("op_8558_transpose_x_0"), val = bool(false)];
            bool var_8558_transpose_y_0 = const()[name = string("op_8558_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_8413_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 8, 32, 128]> var_8558_cast_fp16 = matmul(transpose_x = var_8558_transpose_x_0, transpose_y = var_8558_transpose_y_0, x = var_8557_cast_fp16, y = v_43_cast_fp16)[name = string("op_8558_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_85_cast_fp16 = add(x = var_8552_cast_fp16, y = var_8558_cast_fp16)[name = string("mixed_85_cast_fp16")];
            tensor<int32, [4]> var_8575_begin_0 = const()[name = string("op_8575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8575_end_0 = const()[name = string("op_8575_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8575_end_mask_0 = const()[name = string("op_8575_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8575_cast_fp16 = slice_by_index(begin = var_8575_begin_0, end = var_8575_end_0, end_mask = var_8575_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_8575_cast_fp16")];
            tensor<int32, [4]> var_8590_begin_0 = const()[name = string("op_8590_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8590_end_0 = const()[name = string("op_8590_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8590_end_mask_0 = const()[name = string("op_8590_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8590_cast_fp16 = slice_by_index(begin = var_8590_begin_0, end = var_8590_end_0, end_mask = var_8590_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_8590_cast_fp16")];
            int32 var_8592 = const()[name = string("op_8592"), 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_8592, interleave = mixed_87_interleave_0, values = (var_8575_cast_fp16, var_8590_cast_fp16))[name = string("mixed_87_cast_fp16")];
            tensor<int32, [4]> var_8596_perm_0 = const()[name = string("op_8596_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8600 = const()[name = string("op_8600"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8596_cast_fp16 = transpose(perm = var_8596_perm_0, x = mixed_87_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 2048]> input_213_cast_fp16 = reshape(shape = var_8600, x = var_8596_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_437_cast_fp16 = add(x = x_421_cast_fp16, y = linear_150_cast_fp16)[name = string("x_437_cast_fp16")];
            bool var_8610 = const()[name = string("op_8610"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8613_cast_fp16 = abs(x = x_437_cast_fp16)[name = string("op_8613_cast_fp16")];
            tensor<int32, [1]> peak_175_axes_0 = const()[name = string("peak_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_175_cast_fp16 = reduce_max(axes = peak_175_axes_0, keep_dims = var_8610, x = var_8613_cast_fp16)[name = string("peak_175_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_439_cast_fp16 = real_div(x = x_437_cast_fp16, y = peak_175_cast_fp16)[name = string("x_439_cast_fp16")];
            fp16 var_8609_promoted_to_fp16 = const()[name = string("op_8609_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8617_cast_fp16 = pow(x = x_439_cast_fp16, y = var_8609_promoted_to_fp16)[name = string("op_8617_cast_fp16")];
            tensor<int32, [1]> var_8619_axes_0 = const()[name = string("op_8619_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8619_keep_dims_0 = const()[name = string("op_8619_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8619_cast_fp16 = reduce_mean(axes = var_8619_axes_0, keep_dims = var_8619_keep_dims_0, x = var_8617_cast_fp16)[name = string("op_8619_cast_fp16")];
            fp16 var_8609_promoted_1_to_fp16 = const()[name = string("op_8609_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8620_cast_fp16 = pow(x = peak_175_cast_fp16, y = var_8609_promoted_1_to_fp16)[name = string("op_8620_cast_fp16")];
            fp32 var_8621_epsilon_0 = const()[name = string("op_8621_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8621_cast_fp16 = inverse(epsilon = var_8621_epsilon_0, x = var_8620_cast_fp16)[name = string("op_8621_cast_fp16")];
            fp16 var_8622_to_fp16 = const()[name = string("op_8622_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8623_cast_fp16 = mul(x = var_8621_cast_fp16, y = var_8622_to_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8624_cast_fp16 = add(x = var_8619_cast_fp16, y = var_8623_cast_fp16)[name = string("op_8624_cast_fp16")];
            fp32 var_8625_epsilon_0 = const()[name = string("op_8625_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8625_cast_fp16 = rsqrt(epsilon = var_8625_epsilon_0, x = var_8624_cast_fp16)[name = string("op_8625_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8626_cast_fp16 = mul(x = x_439_cast_fp16, y = var_8625_cast_fp16)[name = string("op_8626_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_8626_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_8631_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_8631_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_8631_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_441_cast_fp16 = add(x = x_437_cast_fp16, y = linear_153_cast_fp16)[name = string("x_441_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8647_cast_fp16 = abs(x = x_441_cast_fp16)[name = string("op_8647_cast_fp16")];
            tensor<int32, [1]> peak_177_axes_0 = const()[name = string("peak_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_177_cast_fp16 = reduce_max(axes = peak_177_axes_0, keep_dims = var_8644, x = var_8647_cast_fp16)[name = string("peak_177_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_443_cast_fp16 = real_div(x = x_441_cast_fp16, y = peak_177_cast_fp16)[name = string("x_443_cast_fp16")];
            fp16 var_8643_promoted_to_fp16 = const()[name = string("op_8643_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8651_cast_fp16 = pow(x = x_443_cast_fp16, y = var_8643_promoted_to_fp16)[name = string("op_8651_cast_fp16")];
            tensor<int32, [1]> var_8653_axes_0 = const()[name = string("op_8653_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8653_keep_dims_0 = const()[name = string("op_8653_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8653_cast_fp16 = reduce_mean(axes = var_8653_axes_0, keep_dims = var_8653_keep_dims_0, x = var_8651_cast_fp16)[name = string("op_8653_cast_fp16")];
            fp16 var_8643_promoted_1_to_fp16 = const()[name = string("op_8643_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8654_cast_fp16 = pow(x = peak_177_cast_fp16, y = var_8643_promoted_1_to_fp16)[name = string("op_8654_cast_fp16")];
            fp32 var_8655_epsilon_0 = const()[name = string("op_8655_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8655_cast_fp16 = inverse(epsilon = var_8655_epsilon_0, x = var_8654_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp16 var_8656_to_fp16 = const()[name = string("op_8656_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8657_cast_fp16 = mul(x = var_8655_cast_fp16, y = var_8656_to_fp16)[name = string("op_8657_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8658_cast_fp16 = add(x = var_8653_cast_fp16, y = var_8657_cast_fp16)[name = string("op_8658_cast_fp16")];
            fp32 var_8659_epsilon_0 = const()[name = string("op_8659_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8659_cast_fp16 = rsqrt(epsilon = var_8659_epsilon_0, x = var_8658_cast_fp16)[name = string("op_8659_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8660_cast_fp16 = mul(x = x_443_cast_fp16, y = var_8659_cast_fp16)[name = string("op_8660_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_8660_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_8669 = const()[name = string("op_8669"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8670_cast_fp16 = reshape(shape = var_8669, x = linear_154_cast_fp16)[name = string("op_8670_cast_fp16")];
            tensor<int32, [4]> x_445_perm_0 = const()[name = string("x_445_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8677 = const()[name = string("op_8677"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_445_cast_fp16 = transpose(perm = x_445_perm_0, x = var_8670_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 16, 128]> var_8680_cast_fp16 = abs(x = x_445_cast_fp16)[name = string("op_8680_cast_fp16")];
            tensor<int32, [1]> peak_179_axes_0 = const()[name = string("peak_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_179_cast_fp16 = reduce_max(axes = peak_179_axes_0, keep_dims = var_8677, x = var_8680_cast_fp16)[name = string("peak_179_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_447_cast_fp16 = real_div(x = x_445_cast_fp16, y = peak_179_cast_fp16)[name = string("x_447_cast_fp16")];
            fp16 var_8676_promoted_to_fp16 = const()[name = string("op_8676_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8684_cast_fp16 = pow(x = x_447_cast_fp16, y = var_8676_promoted_to_fp16)[name = string("op_8684_cast_fp16")];
            tensor<int32, [1]> var_8686_axes_0 = const()[name = string("op_8686_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8686_keep_dims_0 = const()[name = string("op_8686_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8686_cast_fp16 = reduce_mean(axes = var_8686_axes_0, keep_dims = var_8686_keep_dims_0, x = var_8684_cast_fp16)[name = string("op_8686_cast_fp16")];
            fp16 var_8676_promoted_1_to_fp16 = const()[name = string("op_8676_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8687_cast_fp16 = pow(x = peak_179_cast_fp16, y = var_8676_promoted_1_to_fp16)[name = string("op_8687_cast_fp16")];
            fp32 var_8688_epsilon_0 = const()[name = string("op_8688_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8688_cast_fp16 = inverse(epsilon = var_8688_epsilon_0, x = var_8687_cast_fp16)[name = string("op_8688_cast_fp16")];
            fp16 var_8689_to_fp16 = const()[name = string("op_8689_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8690_cast_fp16 = mul(x = var_8688_cast_fp16, y = var_8689_to_fp16)[name = string("op_8690_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8691_cast_fp16 = add(x = var_8686_cast_fp16, y = var_8690_cast_fp16)[name = string("op_8691_cast_fp16")];
            fp32 var_8692_epsilon_0 = const()[name = string("op_8692_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8692_cast_fp16 = rsqrt(epsilon = var_8692_epsilon_0, x = var_8691_cast_fp16)[name = string("op_8692_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8693_cast_fp16 = mul(x = x_447_cast_fp16, y = var_8692_cast_fp16)[name = string("op_8693_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_453_cast_fp16 = mul(x = var_8693_cast_fp16, y = model_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("x_453_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_8702 = const()[name = string("op_8702"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8703_cast_fp16 = reshape(shape = var_8702, x = linear_155_cast_fp16)[name = string("op_8703_cast_fp16")];
            tensor<int32, [4]> x_449_perm_0 = const()[name = string("x_449_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8710 = const()[name = string("op_8710"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_449_cast_fp16 = transpose(perm = x_449_perm_0, x = var_8703_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 8, 16, 128]> var_8713_cast_fp16 = abs(x = x_449_cast_fp16)[name = string("op_8713_cast_fp16")];
            tensor<int32, [1]> peak_181_axes_0 = const()[name = string("peak_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_181_cast_fp16 = reduce_max(axes = peak_181_axes_0, keep_dims = var_8710, x = var_8713_cast_fp16)[name = string("peak_181_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_451_cast_fp16 = real_div(x = x_449_cast_fp16, y = peak_181_cast_fp16)[name = string("x_451_cast_fp16")];
            fp16 var_8709_promoted_to_fp16 = const()[name = string("op_8709_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8717_cast_fp16 = pow(x = x_451_cast_fp16, y = var_8709_promoted_to_fp16)[name = string("op_8717_cast_fp16")];
            tensor<int32, [1]> var_8719_axes_0 = const()[name = string("op_8719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8719_keep_dims_0 = const()[name = string("op_8719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8719_cast_fp16 = reduce_mean(axes = var_8719_axes_0, keep_dims = var_8719_keep_dims_0, x = var_8717_cast_fp16)[name = string("op_8719_cast_fp16")];
            fp16 var_8709_promoted_1_to_fp16 = const()[name = string("op_8709_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8720_cast_fp16 = pow(x = peak_181_cast_fp16, y = var_8709_promoted_1_to_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8721_cast_fp16 = inverse(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            fp16 var_8722_to_fp16 = const()[name = string("op_8722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8723_cast_fp16 = mul(x = var_8721_cast_fp16, y = var_8722_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8724_cast_fp16 = add(x = var_8719_cast_fp16, y = var_8723_cast_fp16)[name = string("op_8724_cast_fp16")];
            fp32 var_8725_epsilon_0 = const()[name = string("op_8725_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8725_cast_fp16 = rsqrt(epsilon = var_8725_epsilon_0, x = var_8724_cast_fp16)[name = string("op_8725_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8726_cast_fp16 = mul(x = x_451_cast_fp16, y = var_8725_cast_fp16)[name = string("op_8726_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_455_cast_fp16 = mul(x = var_8726_cast_fp16, y = model_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("x_455_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_8735 = const()[name = string("op_8735"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8736_cast_fp16 = reshape(shape = var_8735, x = linear_156_cast_fp16)[name = string("op_8736_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_453_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_453_cast_fp16)[name = string("second_89_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8750_cast_fp16 = mul(x = first_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8750_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8751_cast_fp16 = mul(x = second_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8751_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8753_cast_fp16 = sub(x = var_8750_cast_fp16, y = var_8751_cast_fp16)[name = string("op_8753_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8754_cast_fp16 = mul(x = second_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8755_cast_fp16 = mul(x = first_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8755_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8757_cast_fp16 = add(x = var_8754_cast_fp16, y = var_8755_cast_fp16)[name = string("op_8757_cast_fp16")];
            int32 var_8759 = const()[name = string("op_8759"), 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_8759, interleave = q_45_interleave_0, values = (var_8753_cast_fp16, var_8757_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_455_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_455_cast_fp16)[name = string("second_91_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8771_cast_fp16 = mul(x = first_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8771_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8772_cast_fp16 = mul(x = second_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8772_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8774_cast_fp16 = sub(x = var_8771_cast_fp16, y = var_8772_cast_fp16)[name = string("op_8774_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8775_cast_fp16 = mul(x = second_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8776_cast_fp16 = mul(x = first_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8776_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8778_cast_fp16 = add(x = var_8775_cast_fp16, y = var_8776_cast_fp16)[name = string("op_8778_cast_fp16")];
            int32 var_8780 = const()[name = string("op_8780"), 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_8780, interleave = k_45_interleave_0, values = (var_8774_cast_fp16, var_8778_cast_fp16))[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_8791_begin_0 = const()[name = string("op_8791_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8791_end_0 = const()[name = string("op_8791_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8791_end_mask_0 = const()[name = string("op_8791_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8791_stride_0 = const()[name = string("op_8791_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8791_cast_fp16 = slice_by_index(begin = var_8791_begin_0, end = var_8791_end_0, end_mask = var_8791_end_mask_0, stride = var_8791_stride_0, x = q_45_cast_fp16)[name = string("op_8791_cast_fp16")];
            tensor<int32, [4]> var_8801_begin_0 = const()[name = string("op_8801_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8801_end_0 = const()[name = string("op_8801_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8801_end_mask_0 = const()[name = string("op_8801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8801_stride_0 = const()[name = string("op_8801_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8801_cast_fp16 = slice_by_index(begin = var_8801_begin_0, end = var_8801_end_0, end_mask = var_8801_end_mask_0, stride = var_8801_stride_0, x = q_45_cast_fp16)[name = string("op_8801_cast_fp16")];
            int32 var_8803 = const()[name = string("op_8803"), 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_8803, interleave = stacked_45_interleave_0, values = (var_8791_cast_fp16, var_8801_cast_fp16))[name = string("stacked_45_cast_fp16")];
            tensor<int32, [4]> var_8809_begin_0 = const()[name = string("op_8809_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8809_end_0 = const()[name = string("op_8809_end_0"), val = tensor<int32, [4]>([23, 8, 4112, 128])];
            tensor<bool, [4]> var_8809_end_mask_0 = const()[name = string("op_8809_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8809_cast_fp16 = slice_by_index(begin = var_8809_begin_0, end = var_8809_end_0, end_mask = var_8809_end_mask_0, x = read_state_0)[name = string("op_8809_cast_fp16")];
            tensor<int32, [4]> var_8819_begin_0 = const()[name = string("op_8819_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8819_end_0 = const()[name = string("op_8819_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8819_end_mask_0 = const()[name = string("op_8819_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8819_cast_fp16 = slice_by_index(begin = var_8819_begin_0, end = var_8819_end_0, end_mask = var_8819_end_mask_0, x = var_8809_cast_fp16)[name = string("op_8819_cast_fp16")];
            bool var_8828_transpose_x_1 = const()[name = string("op_8828_transpose_x_1"), val = bool(false)];
            bool var_8828_transpose_y_1 = const()[name = string("op_8828_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_8828_cast_fp16 = matmul(transpose_x = var_8828_transpose_x_1, transpose_y = var_8828_transpose_y_1, x = var_8819_cast_fp16, y = stacked_45_cast_fp16)[name = string("op_8828_cast_fp16")];
            tensor<int32, [4]> var_8831_perm_0 = const()[name = string("op_8831_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8832_to_fp16 = const()[name = string("op_8832_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_8831_cast_fp16 = transpose(perm = var_8831_perm_0, x = var_8828_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 8, 32, 4096]> var_8833_cast_fp16 = mul(x = var_8831_cast_fp16, y = var_8832_to_fp16)[name = string("op_8833_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_45_cast_fp16 = add(x = var_8833_cast_fp16, y = cacheMask)[name = string("cache_weights_45_cast_fp16")];
            bool var_8839_transpose_x_1 = const()[name = string("op_8839_transpose_x_1"), val = bool(false)];
            bool var_8839_transpose_y_1 = const()[name = string("op_8839_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8839_cast_fp16 = matmul(transpose_x = var_8839_transpose_x_1, transpose_y = var_8839_transpose_y_1, x = stacked_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_8839_cast_fp16")];
            fp16 var_8840_to_fp16 = const()[name = string("op_8840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8841_cast_fp16 = mul(x = var_8839_cast_fp16, y = var_8840_to_fp16)[name = string("op_8841_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_45_cast_fp16 = add(x = var_8841_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_45_cast_fp16")];
            int32 var_8845 = const()[name = string("op_8845"), val = int32(-1)];
            bool var_8846_interleave_0 = const()[name = string("op_8846_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_8846_cast_fp16 = concat(axis = var_8845, interleave = var_8846_interleave_0, values = (cache_weights_45_cast_fp16, call_weights_45_cast_fp16))[name = string("op_8846_cast_fp16")];
            int32 var_8847 = const()[name = string("op_8847"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_45_cast_fp16 = softmax(axis = var_8847, x = var_8846_cast_fp16)[name = string("shares_45_cast_fp16")];
            tensor<int32, [4]> var_8854_begin_0 = const()[name = string("op_8854_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8854_end_0 = const()[name = string("op_8854_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_8854_end_mask_0 = const()[name = string("op_8854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8854_cast_fp16")];
            tensor<int32, [4]> var_8859_begin_0 = const()[name = string("op_8859_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8859_end_0 = const()[name = string("op_8859_end_0"), val = tensor<int32, [4]>([23, 8, 4112, 128])];
            tensor<bool, [4]> var_8859_end_mask_0 = const()[name = string("op_8859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8859_cast_fp16 = slice_by_index(begin = var_8859_begin_0, end = var_8859_end_0, end_mask = var_8859_end_mask_0, x = read_state_1)[name = string("op_8859_cast_fp16")];
            tensor<int32, [4]> var_8869_begin_0 = const()[name = string("op_8869_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8869_end_0 = const()[name = string("op_8869_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8869_end_mask_0 = const()[name = string("op_8869_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8869_cast_fp16 = slice_by_index(begin = var_8869_begin_0, end = var_8869_end_0, end_mask = var_8869_end_mask_0, x = var_8859_cast_fp16)[name = string("op_8869_cast_fp16")];
            bool var_8875_transpose_x_0 = const()[name = string("op_8875_transpose_x_0"), val = bool(false)];
            bool var_8875_transpose_y_0 = const()[name = string("op_8875_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8875_cast_fp16 = matmul(transpose_x = var_8875_transpose_x_0, transpose_y = var_8875_transpose_y_0, x = var_8854_cast_fp16, y = var_8869_cast_fp16)[name = string("op_8875_cast_fp16")];
            tensor<int32, [4]> var_8880_begin_0 = const()[name = string("op_8880_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_8880_end_0 = const()[name = string("op_8880_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_8880_end_mask_0 = const()[name = string("op_8880_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8880_cast_fp16 = slice_by_index(begin = var_8880_begin_0, end = var_8880_end_0, end_mask = var_8880_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8880_cast_fp16")];
            bool var_8881_transpose_x_0 = const()[name = string("op_8881_transpose_x_0"), val = bool(false)];
            bool var_8881_transpose_y_0 = const()[name = string("op_8881_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_8736_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, 32, 128]> var_8881_cast_fp16 = matmul(transpose_x = var_8881_transpose_x_0, transpose_y = var_8881_transpose_y_0, x = var_8880_cast_fp16, y = v_45_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_89_cast_fp16 = add(x = var_8875_cast_fp16, y = var_8881_cast_fp16)[name = string("mixed_89_cast_fp16")];
            tensor<int32, [4]> var_8898_begin_0 = const()[name = string("op_8898_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8898_end_0 = const()[name = string("op_8898_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8898_end_mask_0 = const()[name = string("op_8898_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8898_cast_fp16 = slice_by_index(begin = var_8898_begin_0, end = var_8898_end_0, end_mask = var_8898_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8898_cast_fp16")];
            tensor<int32, [4]> var_8913_begin_0 = const()[name = string("op_8913_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8913_end_0 = const()[name = string("op_8913_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8913_end_mask_0 = const()[name = string("op_8913_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8913_cast_fp16 = slice_by_index(begin = var_8913_begin_0, end = var_8913_end_0, end_mask = var_8913_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8913_cast_fp16")];
            int32 var_8915 = const()[name = string("op_8915"), 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_8915, interleave = mixed_91_interleave_0, values = (var_8898_cast_fp16, var_8913_cast_fp16))[name = string("mixed_91_cast_fp16")];
            tensor<int32, [4]> var_8919_perm_0 = const()[name = string("op_8919_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8923 = const()[name = string("op_8923"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8919_cast_fp16 = transpose(perm = var_8919_perm_0, x = mixed_91_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 2048]> input_223_cast_fp16 = reshape(shape = var_8923, x = var_8919_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_457_cast_fp16 = add(x = x_441_cast_fp16, y = linear_157_cast_fp16)[name = string("x_457_cast_fp16")];
            bool var_8933 = const()[name = string("op_8933"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8936_cast_fp16 = abs(x = x_457_cast_fp16)[name = string("op_8936_cast_fp16")];
            tensor<int32, [1]> peak_183_axes_0 = const()[name = string("peak_183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_183_cast_fp16 = reduce_max(axes = peak_183_axes_0, keep_dims = var_8933, x = var_8936_cast_fp16)[name = string("peak_183_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_459_cast_fp16 = real_div(x = x_457_cast_fp16, y = peak_183_cast_fp16)[name = string("x_459_cast_fp16")];
            fp16 var_8932_promoted_to_fp16 = const()[name = string("op_8932_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8940_cast_fp16 = pow(x = x_459_cast_fp16, y = var_8932_promoted_to_fp16)[name = string("op_8940_cast_fp16")];
            tensor<int32, [1]> var_8942_axes_0 = const()[name = string("op_8942_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8942_keep_dims_0 = const()[name = string("op_8942_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8942_cast_fp16 = reduce_mean(axes = var_8942_axes_0, keep_dims = var_8942_keep_dims_0, x = var_8940_cast_fp16)[name = string("op_8942_cast_fp16")];
            fp16 var_8932_promoted_1_to_fp16 = const()[name = string("op_8932_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8943_cast_fp16 = pow(x = peak_183_cast_fp16, y = var_8932_promoted_1_to_fp16)[name = string("op_8943_cast_fp16")];
            fp32 var_8944_epsilon_0 = const()[name = string("op_8944_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8944_cast_fp16 = inverse(epsilon = var_8944_epsilon_0, x = var_8943_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, 1]> var_8946_cast_fp16 = mul(x = var_8944_cast_fp16, y = var_8945_to_fp16)[name = string("op_8946_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8947_cast_fp16 = add(x = var_8942_cast_fp16, y = var_8946_cast_fp16)[name = string("op_8947_cast_fp16")];
            fp32 var_8948_epsilon_0 = const()[name = string("op_8948_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8948_cast_fp16 = rsqrt(epsilon = var_8948_epsilon_0, x = var_8947_cast_fp16)[name = string("op_8948_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8949_cast_fp16 = mul(x = x_459_cast_fp16, y = var_8948_cast_fp16)[name = string("op_8949_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_8949_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_8954_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_8954_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_8954_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_461_cast_fp16 = add(x = x_457_cast_fp16, y = linear_160_cast_fp16)[name = string("x_461_cast_fp16")];
            bool var_8967 = const()[name = string("op_8967"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8970_cast_fp16 = abs(x = x_461_cast_fp16)[name = string("op_8970_cast_fp16")];
            tensor<int32, [1]> peak_185_axes_0 = const()[name = string("peak_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_185_cast_fp16 = reduce_max(axes = peak_185_axes_0, keep_dims = var_8967, x = var_8970_cast_fp16)[name = string("peak_185_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_463_cast_fp16 = real_div(x = x_461_cast_fp16, y = peak_185_cast_fp16)[name = string("x_463_cast_fp16")];
            fp16 var_8966_promoted_to_fp16 = const()[name = string("op_8966_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8974_cast_fp16 = pow(x = x_463_cast_fp16, y = var_8966_promoted_to_fp16)[name = string("op_8974_cast_fp16")];
            tensor<int32, [1]> var_8976_axes_0 = const()[name = string("op_8976_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8976_keep_dims_0 = const()[name = string("op_8976_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8976_cast_fp16 = reduce_mean(axes = var_8976_axes_0, keep_dims = var_8976_keep_dims_0, x = var_8974_cast_fp16)[name = string("op_8976_cast_fp16")];
            fp16 var_8966_promoted_1_to_fp16 = const()[name = string("op_8966_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8977_cast_fp16 = pow(x = peak_185_cast_fp16, y = var_8966_promoted_1_to_fp16)[name = string("op_8977_cast_fp16")];
            fp32 var_8978_epsilon_0 = const()[name = string("op_8978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8978_cast_fp16 = inverse(epsilon = var_8978_epsilon_0, x = var_8977_cast_fp16)[name = string("op_8978_cast_fp16")];
            fp16 var_8979_to_fp16 = const()[name = string("op_8979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8980_cast_fp16 = mul(x = var_8978_cast_fp16, y = var_8979_to_fp16)[name = string("op_8980_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8981_cast_fp16 = add(x = var_8976_cast_fp16, y = var_8980_cast_fp16)[name = string("op_8981_cast_fp16")];
            fp32 var_8982_epsilon_0 = const()[name = string("op_8982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8982_cast_fp16 = rsqrt(epsilon = var_8982_epsilon_0, x = var_8981_cast_fp16)[name = string("op_8982_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8983_cast_fp16 = mul(x = x_463_cast_fp16, y = var_8982_cast_fp16)[name = string("op_8983_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_8983_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_8992 = const()[name = string("op_8992"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8993_cast_fp16 = reshape(shape = var_8992, x = linear_161_cast_fp16)[name = string("op_8993_cast_fp16")];
            tensor<int32, [4]> x_465_perm_0 = const()[name = string("x_465_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9000 = const()[name = string("op_9000"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_465_cast_fp16 = transpose(perm = x_465_perm_0, x = var_8993_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 16, 16, 128]> var_9003_cast_fp16 = abs(x = x_465_cast_fp16)[name = string("op_9003_cast_fp16")];
            tensor<int32, [1]> peak_187_axes_0 = const()[name = string("peak_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_187_cast_fp16 = reduce_max(axes = peak_187_axes_0, keep_dims = var_9000, x = var_9003_cast_fp16)[name = string("peak_187_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_467_cast_fp16 = real_div(x = x_465_cast_fp16, y = peak_187_cast_fp16)[name = string("x_467_cast_fp16")];
            fp16 var_8999_promoted_to_fp16 = const()[name = string("op_8999_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9007_cast_fp16 = pow(x = x_467_cast_fp16, y = var_8999_promoted_to_fp16)[name = string("op_9007_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9007_cast_fp16)[name = string("op_9009_cast_fp16")];
            fp16 var_8999_promoted_1_to_fp16 = const()[name = string("op_8999_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9010_cast_fp16 = pow(x = peak_187_cast_fp16, y = var_8999_promoted_1_to_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9014_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9014_cast_fp16")];
            fp32 var_9015_epsilon_0 = const()[name = string("op_9015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9015_cast_fp16 = rsqrt(epsilon = var_9015_epsilon_0, x = var_9014_cast_fp16)[name = string("op_9015_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9016_cast_fp16 = mul(x = x_467_cast_fp16, y = var_9015_cast_fp16)[name = string("op_9016_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_473_cast_fp16 = mul(x = var_9016_cast_fp16, y = model_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("x_473_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_9025 = const()[name = string("op_9025"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9026_cast_fp16 = reshape(shape = var_9025, x = linear_162_cast_fp16)[name = string("op_9026_cast_fp16")];
            tensor<int32, [4]> x_469_perm_0 = const()[name = string("x_469_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9033 = const()[name = string("op_9033"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_469_cast_fp16 = transpose(perm = x_469_perm_0, x = var_9026_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 16, 128]> var_9036_cast_fp16 = abs(x = x_469_cast_fp16)[name = string("op_9036_cast_fp16")];
            tensor<int32, [1]> peak_189_axes_0 = const()[name = string("peak_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_189_cast_fp16 = reduce_max(axes = peak_189_axes_0, keep_dims = var_9033, x = var_9036_cast_fp16)[name = string("peak_189_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_471_cast_fp16 = real_div(x = x_469_cast_fp16, y = peak_189_cast_fp16)[name = string("x_471_cast_fp16")];
            fp16 var_9032_promoted_to_fp16 = const()[name = string("op_9032_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9040_cast_fp16 = pow(x = x_471_cast_fp16, y = var_9032_promoted_to_fp16)[name = string("op_9040_cast_fp16")];
            tensor<int32, [1]> var_9042_axes_0 = const()[name = string("op_9042_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9042_keep_dims_0 = const()[name = string("op_9042_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9042_cast_fp16 = reduce_mean(axes = var_9042_axes_0, keep_dims = var_9042_keep_dims_0, x = var_9040_cast_fp16)[name = string("op_9042_cast_fp16")];
            fp16 var_9032_promoted_1_to_fp16 = const()[name = string("op_9032_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9043_cast_fp16 = pow(x = peak_189_cast_fp16, y = var_9032_promoted_1_to_fp16)[name = string("op_9043_cast_fp16")];
            fp32 var_9044_epsilon_0 = const()[name = string("op_9044_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9044_cast_fp16 = inverse(epsilon = var_9044_epsilon_0, x = var_9043_cast_fp16)[name = string("op_9044_cast_fp16")];
            fp16 var_9045_to_fp16 = const()[name = string("op_9045_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9046_cast_fp16 = mul(x = var_9044_cast_fp16, y = var_9045_to_fp16)[name = string("op_9046_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9047_cast_fp16 = add(x = var_9042_cast_fp16, y = var_9046_cast_fp16)[name = string("op_9047_cast_fp16")];
            fp32 var_9048_epsilon_0 = const()[name = string("op_9048_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9048_cast_fp16 = rsqrt(epsilon = var_9048_epsilon_0, x = var_9047_cast_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9049_cast_fp16 = mul(x = x_471_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9049_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_475_cast_fp16 = mul(x = var_9049_cast_fp16, y = model_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("x_475_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_9058 = const()[name = string("op_9058"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9059_cast_fp16 = reshape(shape = var_9058, x = linear_163_cast_fp16)[name = string("op_9059_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_473_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_473_cast_fp16)[name = string("second_93_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9073_cast_fp16 = mul(x = first_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9074_cast_fp16 = mul(x = second_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9074_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9076_cast_fp16 = sub(x = var_9073_cast_fp16, y = var_9074_cast_fp16)[name = string("op_9076_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9077_cast_fp16 = mul(x = second_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9077_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9078_cast_fp16 = mul(x = first_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9078_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9080_cast_fp16 = add(x = var_9077_cast_fp16, y = var_9078_cast_fp16)[name = string("op_9080_cast_fp16")];
            int32 var_9082 = const()[name = string("op_9082"), 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_9082, interleave = q_47_interleave_0, values = (var_9076_cast_fp16, var_9080_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_475_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_475_cast_fp16)[name = string("second_95_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9094_cast_fp16 = mul(x = first_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9094_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9095_cast_fp16 = mul(x = second_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9095_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9097_cast_fp16 = sub(x = var_9094_cast_fp16, y = var_9095_cast_fp16)[name = string("op_9097_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9098_cast_fp16 = mul(x = second_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9098_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9099_cast_fp16 = mul(x = first_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9099_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9101_cast_fp16 = add(x = var_9098_cast_fp16, y = var_9099_cast_fp16)[name = string("op_9101_cast_fp16")];
            int32 var_9103 = const()[name = string("op_9103"), 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_9103, interleave = k_47_interleave_0, values = (var_9097_cast_fp16, var_9101_cast_fp16))[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> var_9114_begin_0 = const()[name = string("op_9114_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9114_end_0 = const()[name = string("op_9114_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9114_end_mask_0 = const()[name = string("op_9114_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9114_stride_0 = const()[name = string("op_9114_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9114_cast_fp16 = slice_by_index(begin = var_9114_begin_0, end = var_9114_end_0, end_mask = var_9114_end_mask_0, stride = var_9114_stride_0, x = q_47_cast_fp16)[name = string("op_9114_cast_fp16")];
            tensor<int32, [4]> var_9124_begin_0 = const()[name = string("op_9124_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9124_end_0 = const()[name = string("op_9124_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9124_end_mask_0 = const()[name = string("op_9124_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9124_stride_0 = const()[name = string("op_9124_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9124_cast_fp16 = slice_by_index(begin = var_9124_begin_0, end = var_9124_end_0, end_mask = var_9124_end_mask_0, stride = var_9124_stride_0, x = q_47_cast_fp16)[name = string("op_9124_cast_fp16")];
            int32 var_9126 = const()[name = string("op_9126"), 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_9126, interleave = stacked_47_interleave_0, values = (var_9114_cast_fp16, var_9124_cast_fp16))[name = string("stacked_47_cast_fp16")];
            tensor<int32, [4]> var_9132_begin_0 = const()[name = string("op_9132_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_9132_end_0 = const()[name = string("op_9132_end_0"), val = tensor<int32, [4]>([24, 8, 4112, 128])];
            tensor<bool, [4]> var_9132_end_mask_0 = const()[name = string("op_9132_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9132_cast_fp16 = slice_by_index(begin = var_9132_begin_0, end = var_9132_end_0, end_mask = var_9132_end_mask_0, x = read_state_0)[name = string("op_9132_cast_fp16")];
            tensor<int32, [4]> var_9142_begin_0 = const()[name = string("op_9142_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9142_end_0 = const()[name = string("op_9142_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9142_end_mask_0 = const()[name = string("op_9142_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9142_cast_fp16 = slice_by_index(begin = var_9142_begin_0, end = var_9142_end_0, end_mask = var_9142_end_mask_0, x = var_9132_cast_fp16)[name = string("op_9142_cast_fp16")];
            bool var_9151_transpose_x_1 = const()[name = string("op_9151_transpose_x_1"), val = bool(false)];
            bool var_9151_transpose_y_1 = const()[name = string("op_9151_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_9151_cast_fp16 = matmul(transpose_x = var_9151_transpose_x_1, transpose_y = var_9151_transpose_y_1, x = var_9142_cast_fp16, y = stacked_47_cast_fp16)[name = string("op_9151_cast_fp16")];
            tensor<int32, [4]> var_9154_perm_0 = const()[name = string("op_9154_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9155_to_fp16 = const()[name = string("op_9155_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_9154_cast_fp16 = transpose(perm = var_9154_perm_0, x = var_9151_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, 32, 4096]> var_9156_cast_fp16 = mul(x = var_9154_cast_fp16, y = var_9155_to_fp16)[name = string("op_9156_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_47_cast_fp16 = add(x = var_9156_cast_fp16, y = cacheMask)[name = string("cache_weights_47_cast_fp16")];
            bool var_9162_transpose_x_1 = const()[name = string("op_9162_transpose_x_1"), val = bool(false)];
            bool var_9162_transpose_y_1 = const()[name = string("op_9162_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9162_cast_fp16 = matmul(transpose_x = var_9162_transpose_x_1, transpose_y = var_9162_transpose_y_1, x = stacked_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_9162_cast_fp16")];
            fp16 var_9163_to_fp16 = const()[name = string("op_9163_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9164_cast_fp16 = mul(x = var_9162_cast_fp16, y = var_9163_to_fp16)[name = string("op_9164_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_47_cast_fp16 = add(x = var_9164_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_47_cast_fp16")];
            int32 var_9168 = const()[name = string("op_9168"), val = int32(-1)];
            bool var_9169_interleave_0 = const()[name = string("op_9169_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_9169_cast_fp16 = concat(axis = var_9168, interleave = var_9169_interleave_0, values = (cache_weights_47_cast_fp16, call_weights_47_cast_fp16))[name = string("op_9169_cast_fp16")];
            int32 var_9170 = const()[name = string("op_9170"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_47_cast_fp16 = softmax(axis = var_9170, x = var_9169_cast_fp16)[name = string("shares_47_cast_fp16")];
            tensor<int32, [4]> var_9177_begin_0 = const()[name = string("op_9177_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9177_end_0 = const()[name = string("op_9177_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_9177_end_mask_0 = const()[name = string("op_9177_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_9177_cast_fp16 = slice_by_index(begin = var_9177_begin_0, end = var_9177_end_0, end_mask = var_9177_end_mask_0, x = shares_47_cast_fp16)[name = string("op_9177_cast_fp16")];
            tensor<int32, [4]> var_9182_begin_0 = const()[name = string("op_9182_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_9182_end_0 = const()[name = string("op_9182_end_0"), val = tensor<int32, [4]>([24, 8, 4112, 128])];
            tensor<bool, [4]> var_9182_end_mask_0 = const()[name = string("op_9182_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9182_cast_fp16 = slice_by_index(begin = var_9182_begin_0, end = var_9182_end_0, end_mask = var_9182_end_mask_0, x = read_state_1)[name = string("op_9182_cast_fp16")];
            tensor<int32, [4]> var_9192_begin_0 = const()[name = string("op_9192_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9192_end_0 = const()[name = string("op_9192_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9192_end_mask_0 = const()[name = string("op_9192_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9192_cast_fp16 = slice_by_index(begin = var_9192_begin_0, end = var_9192_end_0, end_mask = var_9192_end_mask_0, x = var_9182_cast_fp16)[name = string("op_9192_cast_fp16")];
            bool var_9198_transpose_x_0 = const()[name = string("op_9198_transpose_x_0"), val = bool(false)];
            bool var_9198_transpose_y_0 = const()[name = string("op_9198_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9198_cast_fp16 = matmul(transpose_x = var_9198_transpose_x_0, transpose_y = var_9198_transpose_y_0, x = var_9177_cast_fp16, y = var_9192_cast_fp16)[name = string("op_9198_cast_fp16")];
            tensor<int32, [4]> var_9203_begin_0 = const()[name = string("op_9203_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_9203_end_0 = const()[name = string("op_9203_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_9203_end_mask_0 = const()[name = string("op_9203_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9203_cast_fp16 = slice_by_index(begin = var_9203_begin_0, end = var_9203_end_0, end_mask = var_9203_end_mask_0, x = shares_47_cast_fp16)[name = string("op_9203_cast_fp16")];
            bool var_9204_transpose_x_0 = const()[name = string("op_9204_transpose_x_0"), val = bool(false)];
            bool var_9204_transpose_y_0 = const()[name = string("op_9204_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_9059_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 8, 32, 128]> var_9204_cast_fp16 = matmul(transpose_x = var_9204_transpose_x_0, transpose_y = var_9204_transpose_y_0, x = var_9203_cast_fp16, y = v_47_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_93_cast_fp16 = add(x = var_9198_cast_fp16, y = var_9204_cast_fp16)[name = string("mixed_93_cast_fp16")];
            tensor<int32, [4]> var_9221_begin_0 = const()[name = string("op_9221_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9221_end_0 = const()[name = string("op_9221_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9221_end_mask_0 = const()[name = string("op_9221_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9221_cast_fp16 = slice_by_index(begin = var_9221_begin_0, end = var_9221_end_0, end_mask = var_9221_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_9221_cast_fp16")];
            tensor<int32, [4]> var_9236_begin_0 = const()[name = string("op_9236_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9236_end_0 = const()[name = string("op_9236_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9236_end_mask_0 = const()[name = string("op_9236_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9236_cast_fp16 = slice_by_index(begin = var_9236_begin_0, end = var_9236_end_0, end_mask = var_9236_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_9236_cast_fp16")];
            int32 var_9238 = const()[name = string("op_9238"), 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_9238, interleave = mixed_95_interleave_0, values = (var_9221_cast_fp16, var_9236_cast_fp16))[name = string("mixed_95_cast_fp16")];
            tensor<int32, [4]> var_9242_perm_0 = const()[name = string("op_9242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9246 = const()[name = string("op_9246"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9242_cast_fp16 = transpose(perm = var_9242_perm_0, x = mixed_95_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 16, 2048]> input_233_cast_fp16 = reshape(shape = var_9246, x = var_9242_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_477_cast_fp16 = add(x = x_461_cast_fp16, y = linear_164_cast_fp16)[name = string("x_477_cast_fp16")];
            bool var_9256 = const()[name = string("op_9256"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9259_cast_fp16 = abs(x = x_477_cast_fp16)[name = string("op_9259_cast_fp16")];
            tensor<int32, [1]> peak_191_axes_0 = const()[name = string("peak_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_191_cast_fp16 = reduce_max(axes = peak_191_axes_0, keep_dims = var_9256, x = var_9259_cast_fp16)[name = string("peak_191_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_479_cast_fp16 = real_div(x = x_477_cast_fp16, y = peak_191_cast_fp16)[name = string("x_479_cast_fp16")];
            fp16 var_9255_promoted_to_fp16 = const()[name = string("op_9255_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9263_cast_fp16 = pow(x = x_479_cast_fp16, y = var_9255_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, 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_9255_promoted_1_to_fp16 = const()[name = string("op_9255_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9266_cast_fp16 = pow(x = peak_191_cast_fp16, y = var_9255_promoted_1_to_fp16)[name = string("op_9266_cast_fp16")];
            fp32 var_9267_epsilon_0 = const()[name = string("op_9267_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9267_cast_fp16 = inverse(epsilon = var_9267_epsilon_0, x = var_9266_cast_fp16)[name = string("op_9267_cast_fp16")];
            fp16 var_9268_to_fp16 = const()[name = string("op_9268_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9269_cast_fp16 = mul(x = var_9267_cast_fp16, y = var_9268_to_fp16)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9270_cast_fp16 = add(x = var_9265_cast_fp16, y = var_9269_cast_fp16)[name = string("op_9270_cast_fp16")];
            fp32 var_9271_epsilon_0 = const()[name = string("op_9271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9271_cast_fp16 = rsqrt(epsilon = var_9271_epsilon_0, x = var_9270_cast_fp16)[name = string("op_9271_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9272_cast_fp16 = mul(x = x_479_cast_fp16, y = var_9271_cast_fp16)[name = string("op_9272_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_9272_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_9277_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_9277_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_9277_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_481_cast_fp16 = add(x = x_477_cast_fp16, y = linear_167_cast_fp16)[name = string("x_481_cast_fp16")];
            bool var_9290 = const()[name = string("op_9290"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9293_cast_fp16 = abs(x = x_481_cast_fp16)[name = string("op_9293_cast_fp16")];
            tensor<int32, [1]> peak_193_axes_0 = const()[name = string("peak_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_193_cast_fp16 = reduce_max(axes = peak_193_axes_0, keep_dims = var_9290, x = var_9293_cast_fp16)[name = string("peak_193_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_483_cast_fp16 = real_div(x = x_481_cast_fp16, y = peak_193_cast_fp16)[name = string("x_483_cast_fp16")];
            fp16 var_9289_promoted_to_fp16 = const()[name = string("op_9289_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9297_cast_fp16 = pow(x = x_483_cast_fp16, y = var_9289_promoted_to_fp16)[name = string("op_9297_cast_fp16")];
            tensor<int32, [1]> var_9299_axes_0 = const()[name = string("op_9299_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9299_keep_dims_0 = const()[name = string("op_9299_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9299_cast_fp16 = reduce_mean(axes = var_9299_axes_0, keep_dims = var_9299_keep_dims_0, x = var_9297_cast_fp16)[name = string("op_9299_cast_fp16")];
            fp16 var_9289_promoted_1_to_fp16 = const()[name = string("op_9289_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9300_cast_fp16 = pow(x = peak_193_cast_fp16, y = var_9289_promoted_1_to_fp16)[name = string("op_9300_cast_fp16")];
            fp32 var_9301_epsilon_0 = const()[name = string("op_9301_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9301_cast_fp16 = inverse(epsilon = var_9301_epsilon_0, x = var_9300_cast_fp16)[name = string("op_9301_cast_fp16")];
            fp16 var_9302_to_fp16 = const()[name = string("op_9302_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9303_cast_fp16 = mul(x = var_9301_cast_fp16, y = var_9302_to_fp16)[name = string("op_9303_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9304_cast_fp16 = add(x = var_9299_cast_fp16, y = var_9303_cast_fp16)[name = string("op_9304_cast_fp16")];
            fp32 var_9305_epsilon_0 = const()[name = string("op_9305_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9305_cast_fp16 = rsqrt(epsilon = var_9305_epsilon_0, x = var_9304_cast_fp16)[name = string("op_9305_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9306_cast_fp16 = mul(x = x_483_cast_fp16, y = var_9305_cast_fp16)[name = string("op_9306_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_9306_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_9315 = const()[name = string("op_9315"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9316_cast_fp16 = reshape(shape = var_9315, x = linear_168_cast_fp16)[name = string("op_9316_cast_fp16")];
            tensor<int32, [4]> x_485_perm_0 = const()[name = string("x_485_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9323 = const()[name = string("op_9323"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_485_cast_fp16 = transpose(perm = x_485_perm_0, x = var_9316_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 16, 128]> var_9326_cast_fp16 = abs(x = x_485_cast_fp16)[name = string("op_9326_cast_fp16")];
            tensor<int32, [1]> peak_195_axes_0 = const()[name = string("peak_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_195_cast_fp16 = reduce_max(axes = peak_195_axes_0, keep_dims = var_9323, x = var_9326_cast_fp16)[name = string("peak_195_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_487_cast_fp16 = real_div(x = x_485_cast_fp16, y = peak_195_cast_fp16)[name = string("x_487_cast_fp16")];
            fp16 var_9322_promoted_to_fp16 = const()[name = string("op_9322_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9330_cast_fp16 = pow(x = x_487_cast_fp16, y = var_9322_promoted_to_fp16)[name = string("op_9330_cast_fp16")];
            tensor<int32, [1]> var_9332_axes_0 = const()[name = string("op_9332_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9332_keep_dims_0 = const()[name = string("op_9332_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9332_cast_fp16 = reduce_mean(axes = var_9332_axes_0, keep_dims = var_9332_keep_dims_0, x = var_9330_cast_fp16)[name = string("op_9332_cast_fp16")];
            fp16 var_9322_promoted_1_to_fp16 = const()[name = string("op_9322_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9333_cast_fp16 = pow(x = peak_195_cast_fp16, y = var_9322_promoted_1_to_fp16)[name = string("op_9333_cast_fp16")];
            fp32 var_9334_epsilon_0 = const()[name = string("op_9334_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9334_cast_fp16 = inverse(epsilon = var_9334_epsilon_0, x = var_9333_cast_fp16)[name = string("op_9334_cast_fp16")];
            fp16 var_9335_to_fp16 = const()[name = string("op_9335_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9336_cast_fp16 = mul(x = var_9334_cast_fp16, y = var_9335_to_fp16)[name = string("op_9336_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9337_cast_fp16 = add(x = var_9332_cast_fp16, y = var_9336_cast_fp16)[name = string("op_9337_cast_fp16")];
            fp32 var_9338_epsilon_0 = const()[name = string("op_9338_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9338_cast_fp16 = rsqrt(epsilon = var_9338_epsilon_0, x = var_9337_cast_fp16)[name = string("op_9338_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9339_cast_fp16 = mul(x = x_487_cast_fp16, y = var_9338_cast_fp16)[name = string("op_9339_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_493_cast_fp16 = mul(x = var_9339_cast_fp16, y = model_layers_24_self_attn_q_norm_weight_to_fp16)[name = string("x_493_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_9348 = const()[name = string("op_9348"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9349_cast_fp16 = reshape(shape = var_9348, x = linear_169_cast_fp16)[name = string("op_9349_cast_fp16")];
            tensor<int32, [4]> x_489_perm_0 = const()[name = string("x_489_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9356 = const()[name = string("op_9356"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_489_cast_fp16 = transpose(perm = x_489_perm_0, x = var_9349_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, 16, 128]> var_9359_cast_fp16 = abs(x = x_489_cast_fp16)[name = string("op_9359_cast_fp16")];
            tensor<int32, [1]> peak_197_axes_0 = const()[name = string("peak_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_197_cast_fp16 = reduce_max(axes = peak_197_axes_0, keep_dims = var_9356, x = var_9359_cast_fp16)[name = string("peak_197_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_491_cast_fp16 = real_div(x = x_489_cast_fp16, y = peak_197_cast_fp16)[name = string("x_491_cast_fp16")];
            fp16 var_9355_promoted_to_fp16 = const()[name = string("op_9355_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9363_cast_fp16 = pow(x = x_491_cast_fp16, y = var_9355_promoted_to_fp16)[name = string("op_9363_cast_fp16")];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9365_keep_dims_0 = const()[name = string("op_9365_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9365_cast_fp16 = reduce_mean(axes = var_9365_axes_0, keep_dims = var_9365_keep_dims_0, x = var_9363_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9355_promoted_1_to_fp16 = const()[name = string("op_9355_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9366_cast_fp16 = pow(x = peak_197_cast_fp16, y = var_9355_promoted_1_to_fp16)[name = string("op_9366_cast_fp16")];
            fp32 var_9367_epsilon_0 = const()[name = string("op_9367_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9367_cast_fp16 = inverse(epsilon = var_9367_epsilon_0, x = var_9366_cast_fp16)[name = string("op_9367_cast_fp16")];
            fp16 var_9368_to_fp16 = const()[name = string("op_9368_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9369_cast_fp16 = mul(x = var_9367_cast_fp16, y = var_9368_to_fp16)[name = string("op_9369_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9370_cast_fp16 = add(x = var_9365_cast_fp16, y = var_9369_cast_fp16)[name = string("op_9370_cast_fp16")];
            fp32 var_9371_epsilon_0 = const()[name = string("op_9371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9371_cast_fp16 = rsqrt(epsilon = var_9371_epsilon_0, x = var_9370_cast_fp16)[name = string("op_9371_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9372_cast_fp16 = mul(x = x_491_cast_fp16, y = var_9371_cast_fp16)[name = string("op_9372_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_495_cast_fp16 = mul(x = var_9372_cast_fp16, y = model_layers_24_self_attn_k_norm_weight_to_fp16)[name = string("x_495_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_9381 = const()[name = string("op_9381"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9382_cast_fp16 = reshape(shape = var_9381, x = linear_170_cast_fp16)[name = string("op_9382_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_493_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_493_cast_fp16)[name = string("second_97_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9396_cast_fp16 = mul(x = first_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9396_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9397_cast_fp16 = mul(x = second_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9397_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9399_cast_fp16 = sub(x = var_9396_cast_fp16, y = var_9397_cast_fp16)[name = string("op_9399_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9400_cast_fp16 = mul(x = second_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9400_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9401_cast_fp16 = mul(x = first_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9401_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9403_cast_fp16 = add(x = var_9400_cast_fp16, y = var_9401_cast_fp16)[name = string("op_9403_cast_fp16")];
            int32 var_9405 = const()[name = string("op_9405"), 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_9405, interleave = q_49_interleave_0, values = (var_9399_cast_fp16, var_9403_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_495_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_495_cast_fp16)[name = string("second_99_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9417_cast_fp16 = mul(x = first_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9417_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9418_cast_fp16 = mul(x = second_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9418_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9420_cast_fp16 = sub(x = var_9417_cast_fp16, y = var_9418_cast_fp16)[name = string("op_9420_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9421_cast_fp16 = mul(x = second_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9422_cast_fp16 = mul(x = first_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9422_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9424_cast_fp16 = add(x = var_9421_cast_fp16, y = var_9422_cast_fp16)[name = string("op_9424_cast_fp16")];
            int32 var_9426 = const()[name = string("op_9426"), 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_9426, interleave = k_49_interleave_0, values = (var_9420_cast_fp16, var_9424_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_9437_begin_0 = const()[name = string("op_9437_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9437_end_0 = const()[name = string("op_9437_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9437_end_mask_0 = const()[name = string("op_9437_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9437_stride_0 = const()[name = string("op_9437_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9437_cast_fp16 = slice_by_index(begin = var_9437_begin_0, end = var_9437_end_0, end_mask = var_9437_end_mask_0, stride = var_9437_stride_0, x = q_49_cast_fp16)[name = string("op_9437_cast_fp16")];
            tensor<int32, [4]> var_9447_begin_0 = const()[name = string("op_9447_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9447_end_0 = const()[name = string("op_9447_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9447_end_mask_0 = const()[name = string("op_9447_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9447_stride_0 = const()[name = string("op_9447_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9447_cast_fp16 = slice_by_index(begin = var_9447_begin_0, end = var_9447_end_0, end_mask = var_9447_end_mask_0, stride = var_9447_stride_0, x = q_49_cast_fp16)[name = string("op_9447_cast_fp16")];
            int32 var_9449 = const()[name = string("op_9449"), 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_9449, interleave = stacked_49_interleave_0, values = (var_9437_cast_fp16, var_9447_cast_fp16))[name = string("stacked_49_cast_fp16")];
            tensor<int32, [4]> var_9455_begin_0 = const()[name = string("op_9455_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_9455_end_0 = const()[name = string("op_9455_end_0"), val = tensor<int32, [4]>([25, 8, 4112, 128])];
            tensor<bool, [4]> var_9455_end_mask_0 = const()[name = string("op_9455_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9455_cast_fp16 = slice_by_index(begin = var_9455_begin_0, end = var_9455_end_0, end_mask = var_9455_end_mask_0, x = read_state_0)[name = string("op_9455_cast_fp16")];
            tensor<int32, [4]> var_9465_begin_0 = const()[name = string("op_9465_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9465_end_0 = const()[name = string("op_9465_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9465_end_mask_0 = const()[name = string("op_9465_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9465_cast_fp16 = slice_by_index(begin = var_9465_begin_0, end = var_9465_end_0, end_mask = var_9465_end_mask_0, x = var_9455_cast_fp16)[name = string("op_9465_cast_fp16")];
            bool var_9474_transpose_x_1 = const()[name = string("op_9474_transpose_x_1"), val = bool(false)];
            bool var_9474_transpose_y_1 = const()[name = string("op_9474_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_9474_cast_fp16 = matmul(transpose_x = var_9474_transpose_x_1, transpose_y = var_9474_transpose_y_1, x = var_9465_cast_fp16, y = stacked_49_cast_fp16)[name = string("op_9474_cast_fp16")];
            tensor<int32, [4]> var_9477_perm_0 = const()[name = string("op_9477_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9478_to_fp16 = const()[name = string("op_9478_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_9477_cast_fp16 = transpose(perm = var_9477_perm_0, x = var_9474_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 32, 4096]> var_9479_cast_fp16 = mul(x = var_9477_cast_fp16, y = var_9478_to_fp16)[name = string("op_9479_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_49_cast_fp16 = add(x = var_9479_cast_fp16, y = cacheMask)[name = string("cache_weights_49_cast_fp16")];
            bool var_9485_transpose_x_1 = const()[name = string("op_9485_transpose_x_1"), val = bool(false)];
            bool var_9485_transpose_y_1 = const()[name = string("op_9485_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9485_cast_fp16 = matmul(transpose_x = var_9485_transpose_x_1, transpose_y = var_9485_transpose_y_1, x = stacked_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_9485_cast_fp16")];
            fp16 var_9486_to_fp16 = const()[name = string("op_9486_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9487_cast_fp16 = mul(x = var_9485_cast_fp16, y = var_9486_to_fp16)[name = string("op_9487_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_49_cast_fp16 = add(x = var_9487_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_49_cast_fp16")];
            int32 var_9491 = const()[name = string("op_9491"), val = int32(-1)];
            bool var_9492_interleave_0 = const()[name = string("op_9492_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_9492_cast_fp16 = concat(axis = var_9491, interleave = var_9492_interleave_0, values = (cache_weights_49_cast_fp16, call_weights_49_cast_fp16))[name = string("op_9492_cast_fp16")];
            int32 var_9493 = const()[name = string("op_9493"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_49_cast_fp16 = softmax(axis = var_9493, x = var_9492_cast_fp16)[name = string("shares_49_cast_fp16")];
            tensor<int32, [4]> var_9500_begin_0 = const()[name = string("op_9500_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9500_end_0 = const()[name = string("op_9500_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_9500_end_mask_0 = const()[name = string("op_9500_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_9500_cast_fp16 = slice_by_index(begin = var_9500_begin_0, end = var_9500_end_0, end_mask = var_9500_end_mask_0, x = shares_49_cast_fp16)[name = string("op_9500_cast_fp16")];
            tensor<int32, [4]> var_9505_begin_0 = const()[name = string("op_9505_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_9505_end_0 = const()[name = string("op_9505_end_0"), val = tensor<int32, [4]>([25, 8, 4112, 128])];
            tensor<bool, [4]> var_9505_end_mask_0 = const()[name = string("op_9505_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9505_cast_fp16 = slice_by_index(begin = var_9505_begin_0, end = var_9505_end_0, end_mask = var_9505_end_mask_0, x = read_state_1)[name = string("op_9505_cast_fp16")];
            tensor<int32, [4]> var_9515_begin_0 = const()[name = string("op_9515_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9515_end_0 = const()[name = string("op_9515_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9515_end_mask_0 = const()[name = string("op_9515_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9515_cast_fp16 = slice_by_index(begin = var_9515_begin_0, end = var_9515_end_0, end_mask = var_9515_end_mask_0, x = var_9505_cast_fp16)[name = string("op_9515_cast_fp16")];
            bool var_9521_transpose_x_0 = const()[name = string("op_9521_transpose_x_0"), val = bool(false)];
            bool var_9521_transpose_y_0 = const()[name = string("op_9521_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9521_cast_fp16 = matmul(transpose_x = var_9521_transpose_x_0, transpose_y = var_9521_transpose_y_0, x = var_9500_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9521_cast_fp16")];
            tensor<int32, [4]> var_9526_begin_0 = const()[name = string("op_9526_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_9526_end_0 = const()[name = string("op_9526_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_9526_end_mask_0 = const()[name = string("op_9526_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9526_cast_fp16 = slice_by_index(begin = var_9526_begin_0, end = var_9526_end_0, end_mask = var_9526_end_mask_0, x = shares_49_cast_fp16)[name = string("op_9526_cast_fp16")];
            bool var_9527_transpose_x_0 = const()[name = string("op_9527_transpose_x_0"), val = bool(false)];
            bool var_9527_transpose_y_0 = const()[name = string("op_9527_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_9382_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 32, 128]> var_9527_cast_fp16 = matmul(transpose_x = var_9527_transpose_x_0, transpose_y = var_9527_transpose_y_0, x = var_9526_cast_fp16, y = v_49_cast_fp16)[name = string("op_9527_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_97_cast_fp16 = add(x = var_9521_cast_fp16, y = var_9527_cast_fp16)[name = string("mixed_97_cast_fp16")];
            tensor<int32, [4]> var_9544_begin_0 = const()[name = string("op_9544_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9544_end_0 = const()[name = string("op_9544_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9544_end_mask_0 = const()[name = string("op_9544_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9544_cast_fp16 = slice_by_index(begin = var_9544_begin_0, end = var_9544_end_0, end_mask = var_9544_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_9544_cast_fp16")];
            tensor<int32, [4]> var_9559_begin_0 = const()[name = string("op_9559_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9559_end_0 = const()[name = string("op_9559_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9559_end_mask_0 = const()[name = string("op_9559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9559_cast_fp16 = slice_by_index(begin = var_9559_begin_0, end = var_9559_end_0, end_mask = var_9559_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_9559_cast_fp16")];
            int32 var_9561 = const()[name = string("op_9561"), 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_9561, interleave = mixed_99_interleave_0, values = (var_9544_cast_fp16, var_9559_cast_fp16))[name = string("mixed_99_cast_fp16")];
            tensor<int32, [4]> var_9565_perm_0 = const()[name = string("op_9565_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9569 = const()[name = string("op_9569"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9565_cast_fp16 = transpose(perm = var_9565_perm_0, x = mixed_99_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 2048]> input_243_cast_fp16 = reshape(shape = var_9569, x = var_9565_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_497_cast_fp16 = add(x = x_481_cast_fp16, y = linear_171_cast_fp16)[name = string("x_497_cast_fp16")];
            bool var_9579 = const()[name = string("op_9579"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9582_cast_fp16 = abs(x = x_497_cast_fp16)[name = string("op_9582_cast_fp16")];
            tensor<int32, [1]> peak_199_axes_0 = const()[name = string("peak_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_199_cast_fp16 = reduce_max(axes = peak_199_axes_0, keep_dims = var_9579, x = var_9582_cast_fp16)[name = string("peak_199_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_499_cast_fp16 = real_div(x = x_497_cast_fp16, y = peak_199_cast_fp16)[name = string("x_499_cast_fp16")];
            fp16 var_9578_promoted_to_fp16 = const()[name = string("op_9578_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9586_cast_fp16 = pow(x = x_499_cast_fp16, y = var_9578_promoted_to_fp16)[name = string("op_9586_cast_fp16")];
            tensor<int32, [1]> var_9588_axes_0 = const()[name = string("op_9588_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9588_keep_dims_0 = const()[name = string("op_9588_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9588_cast_fp16 = reduce_mean(axes = var_9588_axes_0, keep_dims = var_9588_keep_dims_0, x = var_9586_cast_fp16)[name = string("op_9588_cast_fp16")];
            fp16 var_9578_promoted_1_to_fp16 = const()[name = string("op_9578_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9589_cast_fp16 = pow(x = peak_199_cast_fp16, y = var_9578_promoted_1_to_fp16)[name = string("op_9589_cast_fp16")];
            fp32 var_9590_epsilon_0 = const()[name = string("op_9590_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9590_cast_fp16 = inverse(epsilon = var_9590_epsilon_0, x = var_9589_cast_fp16)[name = string("op_9590_cast_fp16")];
            fp16 var_9591_to_fp16 = const()[name = string("op_9591_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9592_cast_fp16 = mul(x = var_9590_cast_fp16, y = var_9591_to_fp16)[name = string("op_9592_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9593_cast_fp16 = add(x = var_9588_cast_fp16, y = var_9592_cast_fp16)[name = string("op_9593_cast_fp16")];
            fp32 var_9594_epsilon_0 = const()[name = string("op_9594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9594_cast_fp16 = rsqrt(epsilon = var_9594_epsilon_0, x = var_9593_cast_fp16)[name = string("op_9594_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9595_cast_fp16 = mul(x = x_499_cast_fp16, y = var_9594_cast_fp16)[name = string("op_9595_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_9595_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_9600_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_9600_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_9600_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_501_cast_fp16 = add(x = x_497_cast_fp16, y = linear_174_cast_fp16)[name = string("x_501_cast_fp16")];
            bool var_9613 = const()[name = string("op_9613"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9616_cast_fp16 = abs(x = x_501_cast_fp16)[name = string("op_9616_cast_fp16")];
            tensor<int32, [1]> peak_201_axes_0 = const()[name = string("peak_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_201_cast_fp16 = reduce_max(axes = peak_201_axes_0, keep_dims = var_9613, x = var_9616_cast_fp16)[name = string("peak_201_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_503_cast_fp16 = real_div(x = x_501_cast_fp16, y = peak_201_cast_fp16)[name = string("x_503_cast_fp16")];
            fp16 var_9612_promoted_to_fp16 = const()[name = string("op_9612_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9620_cast_fp16 = pow(x = x_503_cast_fp16, y = var_9612_promoted_to_fp16)[name = string("op_9620_cast_fp16")];
            tensor<int32, [1]> var_9622_axes_0 = const()[name = string("op_9622_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9622_keep_dims_0 = const()[name = string("op_9622_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9622_cast_fp16 = reduce_mean(axes = var_9622_axes_0, keep_dims = var_9622_keep_dims_0, x = var_9620_cast_fp16)[name = string("op_9622_cast_fp16")];
            fp16 var_9612_promoted_1_to_fp16 = const()[name = string("op_9612_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9623_cast_fp16 = pow(x = peak_201_cast_fp16, y = var_9612_promoted_1_to_fp16)[name = string("op_9623_cast_fp16")];
            fp32 var_9624_epsilon_0 = const()[name = string("op_9624_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9624_cast_fp16 = inverse(epsilon = var_9624_epsilon_0, x = var_9623_cast_fp16)[name = string("op_9624_cast_fp16")];
            fp16 var_9625_to_fp16 = const()[name = string("op_9625_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9626_cast_fp16 = mul(x = var_9624_cast_fp16, y = var_9625_to_fp16)[name = string("op_9626_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9627_cast_fp16 = add(x = var_9622_cast_fp16, y = var_9626_cast_fp16)[name = string("op_9627_cast_fp16")];
            fp32 var_9628_epsilon_0 = const()[name = string("op_9628_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9628_cast_fp16 = rsqrt(epsilon = var_9628_epsilon_0, x = var_9627_cast_fp16)[name = string("op_9628_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9629_cast_fp16 = mul(x = x_503_cast_fp16, y = var_9628_cast_fp16)[name = string("op_9629_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_9629_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_9638 = const()[name = string("op_9638"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9639_cast_fp16 = reshape(shape = var_9638, x = linear_175_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [4]> x_505_perm_0 = const()[name = string("x_505_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9646 = const()[name = string("op_9646"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_505_cast_fp16 = transpose(perm = x_505_perm_0, x = var_9639_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 16, 128]> var_9649_cast_fp16 = abs(x = x_505_cast_fp16)[name = string("op_9649_cast_fp16")];
            tensor<int32, [1]> peak_203_axes_0 = const()[name = string("peak_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_203_cast_fp16 = reduce_max(axes = peak_203_axes_0, keep_dims = var_9646, x = var_9649_cast_fp16)[name = string("peak_203_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_507_cast_fp16 = real_div(x = x_505_cast_fp16, y = peak_203_cast_fp16)[name = string("x_507_cast_fp16")];
            fp16 var_9645_promoted_to_fp16 = const()[name = string("op_9645_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9653_cast_fp16 = pow(x = x_507_cast_fp16, y = var_9645_promoted_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<int32, [1]> var_9655_axes_0 = const()[name = string("op_9655_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9655_keep_dims_0 = const()[name = string("op_9655_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9655_cast_fp16 = reduce_mean(axes = var_9655_axes_0, keep_dims = var_9655_keep_dims_0, x = var_9653_cast_fp16)[name = string("op_9655_cast_fp16")];
            fp16 var_9645_promoted_1_to_fp16 = const()[name = string("op_9645_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9656_cast_fp16 = pow(x = peak_203_cast_fp16, y = var_9645_promoted_1_to_fp16)[name = string("op_9656_cast_fp16")];
            fp32 var_9657_epsilon_0 = const()[name = string("op_9657_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9657_cast_fp16 = inverse(epsilon = var_9657_epsilon_0, x = var_9656_cast_fp16)[name = string("op_9657_cast_fp16")];
            fp16 var_9658_to_fp16 = const()[name = string("op_9658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9659_cast_fp16 = mul(x = var_9657_cast_fp16, y = var_9658_to_fp16)[name = string("op_9659_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9660_cast_fp16 = add(x = var_9655_cast_fp16, y = var_9659_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp32 var_9661_epsilon_0 = const()[name = string("op_9661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9661_cast_fp16 = rsqrt(epsilon = var_9661_epsilon_0, x = var_9660_cast_fp16)[name = string("op_9661_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9662_cast_fp16 = mul(x = x_507_cast_fp16, y = var_9661_cast_fp16)[name = string("op_9662_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_513_cast_fp16 = mul(x = var_9662_cast_fp16, y = model_layers_25_self_attn_q_norm_weight_to_fp16)[name = string("x_513_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_9671 = const()[name = string("op_9671"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9672_cast_fp16 = reshape(shape = var_9671, x = linear_176_cast_fp16)[name = string("op_9672_cast_fp16")];
            tensor<int32, [4]> x_509_perm_0 = const()[name = string("x_509_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9679 = const()[name = string("op_9679"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_509_cast_fp16 = transpose(perm = x_509_perm_0, x = var_9672_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 16, 128]> var_9682_cast_fp16 = abs(x = x_509_cast_fp16)[name = string("op_9682_cast_fp16")];
            tensor<int32, [1]> peak_205_axes_0 = const()[name = string("peak_205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_205_cast_fp16 = reduce_max(axes = peak_205_axes_0, keep_dims = var_9679, x = var_9682_cast_fp16)[name = string("peak_205_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_511_cast_fp16 = real_div(x = x_509_cast_fp16, y = peak_205_cast_fp16)[name = string("x_511_cast_fp16")];
            fp16 var_9678_promoted_to_fp16 = const()[name = string("op_9678_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9686_cast_fp16 = pow(x = x_511_cast_fp16, y = var_9678_promoted_to_fp16)[name = string("op_9686_cast_fp16")];
            tensor<int32, [1]> var_9688_axes_0 = const()[name = string("op_9688_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9688_keep_dims_0 = const()[name = string("op_9688_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9688_cast_fp16 = reduce_mean(axes = var_9688_axes_0, keep_dims = var_9688_keep_dims_0, x = var_9686_cast_fp16)[name = string("op_9688_cast_fp16")];
            fp16 var_9678_promoted_1_to_fp16 = const()[name = string("op_9678_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9689_cast_fp16 = pow(x = peak_205_cast_fp16, y = var_9678_promoted_1_to_fp16)[name = string("op_9689_cast_fp16")];
            fp32 var_9690_epsilon_0 = const()[name = string("op_9690_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9690_cast_fp16 = inverse(epsilon = var_9690_epsilon_0, x = var_9689_cast_fp16)[name = string("op_9690_cast_fp16")];
            fp16 var_9691_to_fp16 = const()[name = string("op_9691_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9692_cast_fp16 = mul(x = var_9690_cast_fp16, y = var_9691_to_fp16)[name = string("op_9692_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9693_cast_fp16 = add(x = var_9688_cast_fp16, y = var_9692_cast_fp16)[name = string("op_9693_cast_fp16")];
            fp32 var_9694_epsilon_0 = const()[name = string("op_9694_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9694_cast_fp16 = rsqrt(epsilon = var_9694_epsilon_0, x = var_9693_cast_fp16)[name = string("op_9694_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9695_cast_fp16 = mul(x = x_511_cast_fp16, y = var_9694_cast_fp16)[name = string("op_9695_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_515_cast_fp16 = mul(x = var_9695_cast_fp16, y = model_layers_25_self_attn_k_norm_weight_to_fp16)[name = string("x_515_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_9704 = const()[name = string("op_9704"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9705_cast_fp16 = reshape(shape = var_9704, x = linear_177_cast_fp16)[name = string("op_9705_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_513_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_513_cast_fp16)[name = string("second_101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9719_cast_fp16 = mul(x = first_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9719_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9720_cast_fp16 = mul(x = second_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9720_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9722_cast_fp16 = sub(x = var_9719_cast_fp16, y = var_9720_cast_fp16)[name = string("op_9722_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9723_cast_fp16 = mul(x = second_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9723_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9724_cast_fp16 = mul(x = first_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9724_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9726_cast_fp16 = add(x = var_9723_cast_fp16, y = var_9724_cast_fp16)[name = string("op_9726_cast_fp16")];
            int32 var_9728 = const()[name = string("op_9728"), 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_9728, interleave = q_51_interleave_0, values = (var_9722_cast_fp16, var_9726_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_515_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_515_cast_fp16)[name = string("second_103_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9740_cast_fp16 = mul(x = first_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9740_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9741_cast_fp16 = mul(x = second_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9741_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9743_cast_fp16 = sub(x = var_9740_cast_fp16, y = var_9741_cast_fp16)[name = string("op_9743_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9744_cast_fp16 = mul(x = second_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9744_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9745_cast_fp16 = mul(x = first_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9745_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9747_cast_fp16 = add(x = var_9744_cast_fp16, y = var_9745_cast_fp16)[name = string("op_9747_cast_fp16")];
            int32 var_9749 = const()[name = string("op_9749"), 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_9749, interleave = k_51_interleave_0, values = (var_9743_cast_fp16, var_9747_cast_fp16))[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> var_9760_begin_0 = const()[name = string("op_9760_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9760_end_0 = const()[name = string("op_9760_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9760_end_mask_0 = const()[name = string("op_9760_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9760_stride_0 = const()[name = string("op_9760_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9760_cast_fp16 = slice_by_index(begin = var_9760_begin_0, end = var_9760_end_0, end_mask = var_9760_end_mask_0, stride = var_9760_stride_0, x = q_51_cast_fp16)[name = string("op_9760_cast_fp16")];
            tensor<int32, [4]> var_9770_begin_0 = const()[name = string("op_9770_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9770_end_0 = const()[name = string("op_9770_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9770_end_mask_0 = const()[name = string("op_9770_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9770_stride_0 = const()[name = string("op_9770_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9770_cast_fp16 = slice_by_index(begin = var_9770_begin_0, end = var_9770_end_0, end_mask = var_9770_end_mask_0, stride = var_9770_stride_0, x = q_51_cast_fp16)[name = string("op_9770_cast_fp16")];
            int32 var_9772 = const()[name = string("op_9772"), 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_9772, interleave = stacked_51_interleave_0, values = (var_9760_cast_fp16, var_9770_cast_fp16))[name = string("stacked_51_cast_fp16")];
            tensor<int32, [4]> var_9778_begin_0 = const()[name = string("op_9778_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9778_end_0 = const()[name = string("op_9778_end_0"), val = tensor<int32, [4]>([26, 8, 4112, 128])];
            tensor<bool, [4]> var_9778_end_mask_0 = const()[name = string("op_9778_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9778_cast_fp16 = slice_by_index(begin = var_9778_begin_0, end = var_9778_end_0, end_mask = var_9778_end_mask_0, x = read_state_0)[name = string("op_9778_cast_fp16")];
            tensor<int32, [4]> var_9788_begin_0 = const()[name = string("op_9788_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9788_end_0 = const()[name = string("op_9788_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9788_end_mask_0 = const()[name = string("op_9788_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9788_cast_fp16 = slice_by_index(begin = var_9788_begin_0, end = var_9788_end_0, end_mask = var_9788_end_mask_0, x = var_9778_cast_fp16)[name = string("op_9788_cast_fp16")];
            bool var_9797_transpose_x_1 = const()[name = string("op_9797_transpose_x_1"), val = bool(false)];
            bool var_9797_transpose_y_1 = const()[name = string("op_9797_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_1, transpose_y = var_9797_transpose_y_1, x = var_9788_cast_fp16, y = stacked_51_cast_fp16)[name = string("op_9797_cast_fp16")];
            tensor<int32, [4]> var_9800_perm_0 = const()[name = string("op_9800_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9801_to_fp16 = const()[name = string("op_9801_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_9800_cast_fp16 = transpose(perm = var_9800_perm_0, x = var_9797_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 32, 4096]> var_9802_cast_fp16 = mul(x = var_9800_cast_fp16, y = var_9801_to_fp16)[name = string("op_9802_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_51_cast_fp16 = add(x = var_9802_cast_fp16, y = cacheMask)[name = string("cache_weights_51_cast_fp16")];
            bool var_9808_transpose_x_1 = const()[name = string("op_9808_transpose_x_1"), val = bool(false)];
            bool var_9808_transpose_y_1 = const()[name = string("op_9808_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9808_cast_fp16 = matmul(transpose_x = var_9808_transpose_x_1, transpose_y = var_9808_transpose_y_1, x = stacked_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_9808_cast_fp16")];
            fp16 var_9809_to_fp16 = const()[name = string("op_9809_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9810_cast_fp16 = mul(x = var_9808_cast_fp16, y = var_9809_to_fp16)[name = string("op_9810_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_51_cast_fp16 = add(x = var_9810_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_51_cast_fp16")];
            int32 var_9814 = const()[name = string("op_9814"), val = int32(-1)];
            bool var_9815_interleave_0 = const()[name = string("op_9815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_9815_cast_fp16 = concat(axis = var_9814, interleave = var_9815_interleave_0, values = (cache_weights_51_cast_fp16, call_weights_51_cast_fp16))[name = string("op_9815_cast_fp16")];
            int32 var_9816 = const()[name = string("op_9816"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_51_cast_fp16 = softmax(axis = var_9816, x = var_9815_cast_fp16)[name = string("shares_51_cast_fp16")];
            tensor<int32, [4]> var_9823_begin_0 = const()[name = string("op_9823_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9823_end_0 = const()[name = string("op_9823_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_9823_end_mask_0 = const()[name = string("op_9823_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_9823_cast_fp16 = slice_by_index(begin = var_9823_begin_0, end = var_9823_end_0, end_mask = var_9823_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9823_cast_fp16")];
            tensor<int32, [4]> var_9828_begin_0 = const()[name = string("op_9828_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9828_end_0 = const()[name = string("op_9828_end_0"), val = tensor<int32, [4]>([26, 8, 4112, 128])];
            tensor<bool, [4]> var_9828_end_mask_0 = const()[name = string("op_9828_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9828_cast_fp16 = slice_by_index(begin = var_9828_begin_0, end = var_9828_end_0, end_mask = var_9828_end_mask_0, x = read_state_1)[name = string("op_9828_cast_fp16")];
            tensor<int32, [4]> var_9838_begin_0 = const()[name = string("op_9838_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9838_end_0 = const()[name = string("op_9838_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9838_end_mask_0 = const()[name = string("op_9838_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9838_cast_fp16 = slice_by_index(begin = var_9838_begin_0, end = var_9838_end_0, end_mask = var_9838_end_mask_0, x = var_9828_cast_fp16)[name = string("op_9838_cast_fp16")];
            bool var_9844_transpose_x_0 = const()[name = string("op_9844_transpose_x_0"), val = bool(false)];
            bool var_9844_transpose_y_0 = const()[name = string("op_9844_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9844_cast_fp16 = matmul(transpose_x = var_9844_transpose_x_0, transpose_y = var_9844_transpose_y_0, x = var_9823_cast_fp16, y = var_9838_cast_fp16)[name = string("op_9844_cast_fp16")];
            tensor<int32, [4]> var_9849_begin_0 = const()[name = string("op_9849_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_9849_end_0 = const()[name = string("op_9849_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_9849_end_mask_0 = const()[name = string("op_9849_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9849_cast_fp16 = slice_by_index(begin = var_9849_begin_0, end = var_9849_end_0, end_mask = var_9849_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9849_cast_fp16")];
            bool var_9850_transpose_x_0 = const()[name = string("op_9850_transpose_x_0"), val = bool(false)];
            bool var_9850_transpose_y_0 = const()[name = string("op_9850_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_9705_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 8, 32, 128]> var_9850_cast_fp16 = matmul(transpose_x = var_9850_transpose_x_0, transpose_y = var_9850_transpose_y_0, x = var_9849_cast_fp16, y = v_51_cast_fp16)[name = string("op_9850_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_101_cast_fp16 = add(x = var_9844_cast_fp16, y = var_9850_cast_fp16)[name = string("mixed_101_cast_fp16")];
            tensor<int32, [4]> var_9867_begin_0 = const()[name = string("op_9867_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9867_end_0 = const()[name = string("op_9867_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9867_end_mask_0 = const()[name = string("op_9867_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9867_cast_fp16 = slice_by_index(begin = var_9867_begin_0, end = var_9867_end_0, end_mask = var_9867_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9867_cast_fp16")];
            tensor<int32, [4]> var_9882_begin_0 = const()[name = string("op_9882_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9882_end_0 = const()[name = string("op_9882_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9882_end_mask_0 = const()[name = string("op_9882_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9882_cast_fp16 = slice_by_index(begin = var_9882_begin_0, end = var_9882_end_0, end_mask = var_9882_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9882_cast_fp16")];
            int32 var_9884 = const()[name = string("op_9884"), 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_9884, interleave = mixed_103_interleave_0, values = (var_9867_cast_fp16, var_9882_cast_fp16))[name = string("mixed_103_cast_fp16")];
            tensor<int32, [4]> var_9888_perm_0 = const()[name = string("op_9888_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9892 = const()[name = string("op_9892"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9888_cast_fp16 = transpose(perm = var_9888_perm_0, x = mixed_103_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 2048]> input_253_cast_fp16 = reshape(shape = var_9892, x = var_9888_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_517_cast_fp16 = add(x = x_501_cast_fp16, y = linear_178_cast_fp16)[name = string("x_517_cast_fp16")];
            bool var_9902 = const()[name = string("op_9902"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9905_cast_fp16 = abs(x = x_517_cast_fp16)[name = string("op_9905_cast_fp16")];
            tensor<int32, [1]> peak_207_axes_0 = const()[name = string("peak_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_207_cast_fp16 = reduce_max(axes = peak_207_axes_0, keep_dims = var_9902, x = var_9905_cast_fp16)[name = string("peak_207_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_519_cast_fp16 = real_div(x = x_517_cast_fp16, y = peak_207_cast_fp16)[name = string("x_519_cast_fp16")];
            fp16 var_9901_promoted_to_fp16 = const()[name = string("op_9901_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9909_cast_fp16 = pow(x = x_519_cast_fp16, y = var_9901_promoted_to_fp16)[name = string("op_9909_cast_fp16")];
            tensor<int32, [1]> var_9911_axes_0 = const()[name = string("op_9911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9911_keep_dims_0 = const()[name = string("op_9911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9911_cast_fp16 = reduce_mean(axes = var_9911_axes_0, keep_dims = var_9911_keep_dims_0, x = var_9909_cast_fp16)[name = string("op_9911_cast_fp16")];
            fp16 var_9901_promoted_1_to_fp16 = const()[name = string("op_9901_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9912_cast_fp16 = pow(x = peak_207_cast_fp16, y = var_9901_promoted_1_to_fp16)[name = string("op_9912_cast_fp16")];
            fp32 var_9913_epsilon_0 = const()[name = string("op_9913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9913_cast_fp16 = inverse(epsilon = var_9913_epsilon_0, x = var_9912_cast_fp16)[name = string("op_9913_cast_fp16")];
            fp16 var_9914_to_fp16 = const()[name = string("op_9914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9915_cast_fp16 = mul(x = var_9913_cast_fp16, y = var_9914_to_fp16)[name = string("op_9915_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9916_cast_fp16 = add(x = var_9911_cast_fp16, y = var_9915_cast_fp16)[name = string("op_9916_cast_fp16")];
            fp32 var_9917_epsilon_0 = const()[name = string("op_9917_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9917_cast_fp16 = rsqrt(epsilon = var_9917_epsilon_0, x = var_9916_cast_fp16)[name = string("op_9917_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9918_cast_fp16 = mul(x = x_519_cast_fp16, y = var_9917_cast_fp16)[name = string("op_9918_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_9918_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_9923_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_9923_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_9923_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_521_cast_fp16 = add(x = x_517_cast_fp16, y = linear_181_cast_fp16)[name = string("x_521_cast_fp16")];
            bool var_9936 = const()[name = string("op_9936"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9939_cast_fp16 = abs(x = x_521_cast_fp16)[name = string("op_9939_cast_fp16")];
            tensor<int32, [1]> peak_209_axes_0 = const()[name = string("peak_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_209_cast_fp16 = reduce_max(axes = peak_209_axes_0, keep_dims = var_9936, x = var_9939_cast_fp16)[name = string("peak_209_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_523_cast_fp16 = real_div(x = x_521_cast_fp16, y = peak_209_cast_fp16)[name = string("x_523_cast_fp16")];
            fp16 var_9935_promoted_to_fp16 = const()[name = string("op_9935_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9943_cast_fp16 = pow(x = x_523_cast_fp16, y = var_9935_promoted_to_fp16)[name = string("op_9943_cast_fp16")];
            tensor<int32, [1]> var_9945_axes_0 = const()[name = string("op_9945_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9945_keep_dims_0 = const()[name = string("op_9945_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9945_cast_fp16 = reduce_mean(axes = var_9945_axes_0, keep_dims = var_9945_keep_dims_0, x = var_9943_cast_fp16)[name = string("op_9945_cast_fp16")];
            fp16 var_9935_promoted_1_to_fp16 = const()[name = string("op_9935_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9946_cast_fp16 = pow(x = peak_209_cast_fp16, y = var_9935_promoted_1_to_fp16)[name = string("op_9946_cast_fp16")];
            fp32 var_9947_epsilon_0 = const()[name = string("op_9947_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9947_cast_fp16 = inverse(epsilon = var_9947_epsilon_0, x = var_9946_cast_fp16)[name = string("op_9947_cast_fp16")];
            fp16 var_9948_to_fp16 = const()[name = string("op_9948_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9949_cast_fp16 = mul(x = var_9947_cast_fp16, y = var_9948_to_fp16)[name = string("op_9949_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9950_cast_fp16 = add(x = var_9945_cast_fp16, y = var_9949_cast_fp16)[name = string("op_9950_cast_fp16")];
            fp32 var_9951_epsilon_0 = const()[name = string("op_9951_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9951_cast_fp16 = rsqrt(epsilon = var_9951_epsilon_0, x = var_9950_cast_fp16)[name = string("op_9951_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9952_cast_fp16 = mul(x = x_523_cast_fp16, y = var_9951_cast_fp16)[name = string("op_9952_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_9952_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_9961 = const()[name = string("op_9961"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9962_cast_fp16 = reshape(shape = var_9961, x = linear_182_cast_fp16)[name = string("op_9962_cast_fp16")];
            tensor<int32, [4]> x_525_perm_0 = const()[name = string("x_525_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9969 = const()[name = string("op_9969"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_525_cast_fp16 = transpose(perm = x_525_perm_0, x = var_9962_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 16, 128]> var_9972_cast_fp16 = abs(x = x_525_cast_fp16)[name = string("op_9972_cast_fp16")];
            tensor<int32, [1]> peak_211_axes_0 = const()[name = string("peak_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_211_cast_fp16 = reduce_max(axes = peak_211_axes_0, keep_dims = var_9969, x = var_9972_cast_fp16)[name = string("peak_211_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_527_cast_fp16 = real_div(x = x_525_cast_fp16, y = peak_211_cast_fp16)[name = string("x_527_cast_fp16")];
            fp16 var_9968_promoted_to_fp16 = const()[name = string("op_9968_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9976_cast_fp16 = pow(x = x_527_cast_fp16, y = var_9968_promoted_to_fp16)[name = string("op_9976_cast_fp16")];
            tensor<int32, [1]> var_9978_axes_0 = const()[name = string("op_9978_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9978_keep_dims_0 = const()[name = string("op_9978_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9978_cast_fp16 = reduce_mean(axes = var_9978_axes_0, keep_dims = var_9978_keep_dims_0, x = var_9976_cast_fp16)[name = string("op_9978_cast_fp16")];
            fp16 var_9968_promoted_1_to_fp16 = const()[name = string("op_9968_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9979_cast_fp16 = pow(x = peak_211_cast_fp16, y = var_9968_promoted_1_to_fp16)[name = string("op_9979_cast_fp16")];
            fp32 var_9980_epsilon_0 = const()[name = string("op_9980_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9980_cast_fp16 = inverse(epsilon = var_9980_epsilon_0, x = var_9979_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp16 var_9981_to_fp16 = const()[name = string("op_9981_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9982_cast_fp16 = mul(x = var_9980_cast_fp16, y = var_9981_to_fp16)[name = string("op_9982_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9983_cast_fp16 = add(x = var_9978_cast_fp16, y = var_9982_cast_fp16)[name = string("op_9983_cast_fp16")];
            fp32 var_9984_epsilon_0 = const()[name = string("op_9984_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9984_cast_fp16 = rsqrt(epsilon = var_9984_epsilon_0, x = var_9983_cast_fp16)[name = string("op_9984_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9985_cast_fp16 = mul(x = x_527_cast_fp16, y = var_9984_cast_fp16)[name = string("op_9985_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_533_cast_fp16 = mul(x = var_9985_cast_fp16, y = model_layers_26_self_attn_q_norm_weight_to_fp16)[name = string("x_533_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_9994 = const()[name = string("op_9994"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9995_cast_fp16 = reshape(shape = var_9994, x = linear_183_cast_fp16)[name = string("op_9995_cast_fp16")];
            tensor<int32, [4]> x_529_perm_0 = const()[name = string("x_529_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10002 = const()[name = string("op_10002"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_529_cast_fp16 = transpose(perm = x_529_perm_0, x = var_9995_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 16, 128]> var_10005_cast_fp16 = abs(x = x_529_cast_fp16)[name = string("op_10005_cast_fp16")];
            tensor<int32, [1]> peak_213_axes_0 = const()[name = string("peak_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_213_cast_fp16 = reduce_max(axes = peak_213_axes_0, keep_dims = var_10002, x = var_10005_cast_fp16)[name = string("peak_213_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_531_cast_fp16 = real_div(x = x_529_cast_fp16, y = peak_213_cast_fp16)[name = string("x_531_cast_fp16")];
            fp16 var_10001_promoted_to_fp16 = const()[name = string("op_10001_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_10009_cast_fp16 = pow(x = x_531_cast_fp16, y = var_10001_promoted_to_fp16)[name = string("op_10009_cast_fp16")];
            tensor<int32, [1]> var_10011_axes_0 = const()[name = string("op_10011_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10011_keep_dims_0 = const()[name = string("op_10011_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_10011_cast_fp16 = reduce_mean(axes = var_10011_axes_0, keep_dims = var_10011_keep_dims_0, x = var_10009_cast_fp16)[name = string("op_10011_cast_fp16")];
            fp16 var_10001_promoted_1_to_fp16 = const()[name = string("op_10001_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_10012_cast_fp16 = pow(x = peak_213_cast_fp16, y = var_10001_promoted_1_to_fp16)[name = string("op_10012_cast_fp16")];
            fp32 var_10013_epsilon_0 = const()[name = string("op_10013_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_10013_cast_fp16 = inverse(epsilon = var_10013_epsilon_0, x = var_10012_cast_fp16)[name = string("op_10013_cast_fp16")];
            fp16 var_10014_to_fp16 = const()[name = string("op_10014_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_10015_cast_fp16 = mul(x = var_10013_cast_fp16, y = var_10014_to_fp16)[name = string("op_10015_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_10016_cast_fp16 = add(x = var_10011_cast_fp16, y = var_10015_cast_fp16)[name = string("op_10016_cast_fp16")];
            fp32 var_10017_epsilon_0 = const()[name = string("op_10017_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_10017_cast_fp16 = rsqrt(epsilon = var_10017_epsilon_0, x = var_10016_cast_fp16)[name = string("op_10017_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_10018_cast_fp16 = mul(x = x_531_cast_fp16, y = var_10017_cast_fp16)[name = string("op_10018_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_535_cast_fp16 = mul(x = var_10018_cast_fp16, y = model_layers_26_self_attn_k_norm_weight_to_fp16)[name = string("x_535_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_10027 = const()[name = string("op_10027"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10028_cast_fp16 = reshape(shape = var_10027, x = linear_184_cast_fp16)[name = string("op_10028_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_533_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_533_cast_fp16)[name = string("second_105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10042_cast_fp16 = mul(x = first_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10042_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10043_cast_fp16 = mul(x = second_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10045_cast_fp16 = sub(x = var_10042_cast_fp16, y = var_10043_cast_fp16)[name = string("op_10045_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10046_cast_fp16 = mul(x = second_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10046_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10047_cast_fp16 = mul(x = first_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10047_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10049_cast_fp16 = add(x = var_10046_cast_fp16, y = var_10047_cast_fp16)[name = string("op_10049_cast_fp16")];
            int32 var_10051 = const()[name = string("op_10051"), 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_10051, interleave = q_53_interleave_0, values = (var_10045_cast_fp16, var_10049_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_535_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_535_cast_fp16)[name = string("second_107_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10063_cast_fp16 = mul(x = first_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10064_cast_fp16 = mul(x = second_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10064_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10066_cast_fp16 = sub(x = var_10063_cast_fp16, y = var_10064_cast_fp16)[name = string("op_10066_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10067_cast_fp16 = mul(x = second_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10067_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10068_cast_fp16 = mul(x = first_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10068_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10070_cast_fp16 = add(x = var_10067_cast_fp16, y = var_10068_cast_fp16)[name = string("op_10070_cast_fp16")];
            int32 var_10072 = const()[name = string("op_10072"), 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_10072, interleave = k_53_interleave_0, values = (var_10066_cast_fp16, var_10070_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> var_10083_begin_0 = const()[name = string("op_10083_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10083_end_0 = const()[name = string("op_10083_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10083_end_mask_0 = const()[name = string("op_10083_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10083_stride_0 = const()[name = string("op_10083_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10083_cast_fp16 = slice_by_index(begin = var_10083_begin_0, end = var_10083_end_0, end_mask = var_10083_end_mask_0, stride = var_10083_stride_0, x = q_53_cast_fp16)[name = string("op_10083_cast_fp16")];
            tensor<int32, [4]> var_10093_begin_0 = const()[name = string("op_10093_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_10093_end_0 = const()[name = string("op_10093_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10093_end_mask_0 = const()[name = string("op_10093_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10093_stride_0 = const()[name = string("op_10093_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10093_cast_fp16 = slice_by_index(begin = var_10093_begin_0, end = var_10093_end_0, end_mask = var_10093_end_mask_0, stride = var_10093_stride_0, x = q_53_cast_fp16)[name = string("op_10093_cast_fp16")];
            int32 var_10095 = const()[name = string("op_10095"), 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_10095, interleave = stacked_53_interleave_0, values = (var_10083_cast_fp16, var_10093_cast_fp16))[name = string("stacked_53_cast_fp16")];
            tensor<int32, [4]> var_10101_begin_0 = const()[name = string("op_10101_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_10101_end_0 = const()[name = string("op_10101_end_0"), val = tensor<int32, [4]>([27, 8, 4112, 128])];
            tensor<bool, [4]> var_10101_end_mask_0 = const()[name = string("op_10101_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_10101_cast_fp16 = slice_by_index(begin = var_10101_begin_0, end = var_10101_end_0, end_mask = var_10101_end_mask_0, x = read_state_0)[name = string("op_10101_cast_fp16")];
            tensor<int32, [4]> var_10111_begin_0 = const()[name = string("op_10111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10111_end_0 = const()[name = string("op_10111_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_10111_end_mask_0 = const()[name = string("op_10111_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, x = var_10101_cast_fp16)[name = string("op_10111_cast_fp16")];
            bool var_10120_transpose_x_1 = const()[name = string("op_10120_transpose_x_1"), val = bool(false)];
            bool var_10120_transpose_y_1 = const()[name = string("op_10120_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_10120_cast_fp16 = matmul(transpose_x = var_10120_transpose_x_1, transpose_y = var_10120_transpose_y_1, x = var_10111_cast_fp16, y = stacked_53_cast_fp16)[name = string("op_10120_cast_fp16")];
            tensor<int32, [4]> var_10123_perm_0 = const()[name = string("op_10123_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_10124_to_fp16 = const()[name = string("op_10124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_10123_cast_fp16 = transpose(perm = var_10123_perm_0, x = var_10120_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 8, 32, 4096]> var_10125_cast_fp16 = mul(x = var_10123_cast_fp16, y = var_10124_to_fp16)[name = string("op_10125_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_53_cast_fp16 = add(x = var_10125_cast_fp16, y = cacheMask)[name = string("cache_weights_53_cast_fp16")];
            bool var_10131_transpose_x_1 = const()[name = string("op_10131_transpose_x_1"), val = bool(false)];
            bool var_10131_transpose_y_1 = const()[name = string("op_10131_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_10131_cast_fp16 = matmul(transpose_x = var_10131_transpose_x_1, transpose_y = var_10131_transpose_y_1, x = stacked_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_10131_cast_fp16")];
            fp16 var_10132_to_fp16 = const()[name = string("op_10132_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_10133_cast_fp16 = mul(x = var_10131_cast_fp16, y = var_10132_to_fp16)[name = string("op_10133_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_53_cast_fp16 = add(x = var_10133_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_53_cast_fp16")];
            int32 var_10137 = const()[name = string("op_10137"), val = int32(-1)];
            bool var_10138_interleave_0 = const()[name = string("op_10138_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_10138_cast_fp16 = concat(axis = var_10137, interleave = var_10138_interleave_0, values = (cache_weights_53_cast_fp16, call_weights_53_cast_fp16))[name = string("op_10138_cast_fp16")];
            int32 var_10139 = const()[name = string("op_10139"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_53_cast_fp16 = softmax(axis = var_10139, x = var_10138_cast_fp16)[name = string("shares_53_cast_fp16")];
            tensor<int32, [4]> var_10146_begin_0 = const()[name = string("op_10146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10146_end_0 = const()[name = string("op_10146_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_10146_end_mask_0 = const()[name = string("op_10146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_10146_cast_fp16 = slice_by_index(begin = var_10146_begin_0, end = var_10146_end_0, end_mask = var_10146_end_mask_0, x = shares_53_cast_fp16)[name = string("op_10146_cast_fp16")];
            tensor<int32, [4]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [4]>([27, 8, 4112, 128])];
            tensor<bool, [4]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, x = read_state_1)[name = string("op_10151_cast_fp16")];
            tensor<int32, [4]> var_10161_begin_0 = const()[name = string("op_10161_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10161_end_0 = const()[name = string("op_10161_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_10161_end_mask_0 = const()[name = string("op_10161_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_10161_cast_fp16 = slice_by_index(begin = var_10161_begin_0, end = var_10161_end_0, end_mask = var_10161_end_mask_0, x = var_10151_cast_fp16)[name = string("op_10161_cast_fp16")];
            bool var_10167_transpose_x_0 = const()[name = string("op_10167_transpose_x_0"), val = bool(false)];
            bool var_10167_transpose_y_0 = const()[name = string("op_10167_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_10167_cast_fp16 = matmul(transpose_x = var_10167_transpose_x_0, transpose_y = var_10167_transpose_y_0, x = var_10146_cast_fp16, y = var_10161_cast_fp16)[name = string("op_10167_cast_fp16")];
            tensor<int32, [4]> var_10172_begin_0 = const()[name = string("op_10172_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_10172_end_0 = const()[name = string("op_10172_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_10172_end_mask_0 = const()[name = string("op_10172_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_10172_cast_fp16 = slice_by_index(begin = var_10172_begin_0, end = var_10172_end_0, end_mask = var_10172_end_mask_0, x = shares_53_cast_fp16)[name = string("op_10172_cast_fp16")];
            bool var_10173_transpose_x_0 = const()[name = string("op_10173_transpose_x_0"), val = bool(false)];
            bool var_10173_transpose_y_0 = const()[name = string("op_10173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_10028_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 32, 128]> var_10173_cast_fp16 = matmul(transpose_x = var_10173_transpose_x_0, transpose_y = var_10173_transpose_y_0, x = var_10172_cast_fp16, y = v_53_cast_fp16)[name = string("op_10173_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_105_cast_fp16 = add(x = var_10167_cast_fp16, y = var_10173_cast_fp16)[name = string("mixed_105_cast_fp16")];
            tensor<int32, [4]> var_10190_begin_0 = const()[name = string("op_10190_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10190_end_0 = const()[name = string("op_10190_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_10190_end_mask_0 = const()[name = string("op_10190_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10190_cast_fp16 = slice_by_index(begin = var_10190_begin_0, end = var_10190_end_0, end_mask = var_10190_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_10190_cast_fp16")];
            tensor<int32, [4]> var_10205_begin_0 = const()[name = string("op_10205_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_10205_end_0 = const()[name = string("op_10205_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_10205_end_mask_0 = const()[name = string("op_10205_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10205_cast_fp16 = slice_by_index(begin = var_10205_begin_0, end = var_10205_end_0, end_mask = var_10205_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_10205_cast_fp16")];
            int32 var_10207 = const()[name = string("op_10207"), 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_10207, interleave = mixed_107_interleave_0, values = (var_10190_cast_fp16, var_10205_cast_fp16))[name = string("mixed_107_cast_fp16")];
            tensor<int32, [4]> var_10211_perm_0 = const()[name = string("op_10211_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_10215 = const()[name = string("op_10215"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_10211_cast_fp16 = transpose(perm = var_10211_perm_0, x = mixed_107_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 2048]> input_263_cast_fp16 = reshape(shape = var_10215, x = var_10211_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_537_cast_fp16 = add(x = x_521_cast_fp16, y = linear_185_cast_fp16)[name = string("x_537_cast_fp16")];
            bool var_10225 = const()[name = string("op_10225"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10228_cast_fp16 = abs(x = x_537_cast_fp16)[name = string("op_10228_cast_fp16")];
            tensor<int32, [1]> peak_215_axes_0 = const()[name = string("peak_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_215_cast_fp16 = reduce_max(axes = peak_215_axes_0, keep_dims = var_10225, x = var_10228_cast_fp16)[name = string("peak_215_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_539_cast_fp16 = real_div(x = x_537_cast_fp16, y = peak_215_cast_fp16)[name = string("x_539_cast_fp16")];
            fp16 var_10224_promoted_to_fp16 = const()[name = string("op_10224_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10232_cast_fp16 = pow(x = x_539_cast_fp16, y = var_10224_promoted_to_fp16)[name = string("op_10232_cast_fp16")];
            tensor<int32, [1]> var_10234_axes_0 = const()[name = string("op_10234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10234_keep_dims_0 = const()[name = string("op_10234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10234_cast_fp16 = reduce_mean(axes = var_10234_axes_0, keep_dims = var_10234_keep_dims_0, x = var_10232_cast_fp16)[name = string("op_10234_cast_fp16")];
            fp16 var_10224_promoted_1_to_fp16 = const()[name = string("op_10224_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10235_cast_fp16 = pow(x = peak_215_cast_fp16, y = var_10224_promoted_1_to_fp16)[name = string("op_10235_cast_fp16")];
            fp32 var_10236_epsilon_0 = const()[name = string("op_10236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10236_cast_fp16 = inverse(epsilon = var_10236_epsilon_0, x = var_10235_cast_fp16)[name = string("op_10236_cast_fp16")];
            fp16 var_10237_to_fp16 = const()[name = string("op_10237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10238_cast_fp16 = mul(x = var_10236_cast_fp16, y = var_10237_to_fp16)[name = string("op_10238_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10239_cast_fp16 = add(x = var_10234_cast_fp16, y = var_10238_cast_fp16)[name = string("op_10239_cast_fp16")];
            fp32 var_10240_epsilon_0 = const()[name = string("op_10240_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10240_cast_fp16 = rsqrt(epsilon = var_10240_epsilon_0, x = var_10239_cast_fp16)[name = string("op_10240_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10241_cast_fp16 = mul(x = x_539_cast_fp16, y = var_10240_cast_fp16)[name = string("op_10241_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_10241_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_10246_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_10246_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_10246_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_541_cast_fp16 = add(x = x_537_cast_fp16, y = linear_188_cast_fp16)[name = string("x_541_cast_fp16")];
            bool var_10259 = const()[name = string("op_10259"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10262_cast_fp16 = abs(x = x_541_cast_fp16)[name = string("op_10262_cast_fp16")];
            tensor<int32, [1]> peak_217_axes_0 = const()[name = string("peak_217_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_217_cast_fp16 = reduce_max(axes = peak_217_axes_0, keep_dims = var_10259, x = var_10262_cast_fp16)[name = string("peak_217_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_543_cast_fp16 = real_div(x = x_541_cast_fp16, y = peak_217_cast_fp16)[name = string("x_543_cast_fp16")];
            fp16 var_10258_promoted_to_fp16 = const()[name = string("op_10258_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10266_cast_fp16 = pow(x = x_543_cast_fp16, y = var_10258_promoted_to_fp16)[name = string("op_10266_cast_fp16")];
            tensor<int32, [1]> var_10268_axes_0 = const()[name = string("op_10268_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10268_keep_dims_0 = const()[name = string("op_10268_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10268_cast_fp16 = reduce_mean(axes = var_10268_axes_0, keep_dims = var_10268_keep_dims_0, x = var_10266_cast_fp16)[name = string("op_10268_cast_fp16")];
            fp16 var_10258_promoted_1_to_fp16 = const()[name = string("op_10258_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10269_cast_fp16 = pow(x = peak_217_cast_fp16, y = var_10258_promoted_1_to_fp16)[name = string("op_10269_cast_fp16")];
            fp32 var_10270_epsilon_0 = const()[name = string("op_10270_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10270_cast_fp16 = inverse(epsilon = var_10270_epsilon_0, x = var_10269_cast_fp16)[name = string("op_10270_cast_fp16")];
            fp16 var_10271_to_fp16 = const()[name = string("op_10271_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10272_cast_fp16 = mul(x = var_10270_cast_fp16, y = var_10271_to_fp16)[name = string("op_10272_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10273_cast_fp16 = add(x = var_10268_cast_fp16, y = var_10272_cast_fp16)[name = string("op_10273_cast_fp16")];
            fp32 var_10274_epsilon_0 = const()[name = string("op_10274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10274_cast_fp16 = rsqrt(epsilon = var_10274_epsilon_0, x = var_10273_cast_fp16)[name = string("op_10274_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10275_cast_fp16 = mul(x = x_543_cast_fp16, y = var_10274_cast_fp16)[name = string("op_10275_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_10275_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_10284 = const()[name = string("op_10284"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_10285_cast_fp16 = reshape(shape = var_10284, x = linear_189_cast_fp16)[name = string("op_10285_cast_fp16")];
            tensor<int32, [4]> x_545_perm_0 = const()[name = string("x_545_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10292 = const()[name = string("op_10292"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_545_cast_fp16 = transpose(perm = x_545_perm_0, x = var_10285_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 16, 16, 128]> var_10295_cast_fp16 = abs(x = x_545_cast_fp16)[name = string("op_10295_cast_fp16")];
            tensor<int32, [1]> peak_219_axes_0 = const()[name = string("peak_219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_219_cast_fp16 = reduce_max(axes = peak_219_axes_0, keep_dims = var_10292, x = var_10295_cast_fp16)[name = string("peak_219_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_547_cast_fp16 = real_div(x = x_545_cast_fp16, y = peak_219_cast_fp16)[name = string("x_547_cast_fp16")];
            fp16 var_10291_promoted_to_fp16 = const()[name = string("op_10291_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_10299_cast_fp16 = pow(x = x_547_cast_fp16, y = var_10291_promoted_to_fp16)[name = string("op_10299_cast_fp16")];
            tensor<int32, [1]> var_10301_axes_0 = const()[name = string("op_10301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10301_keep_dims_0 = const()[name = string("op_10301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_10301_cast_fp16 = reduce_mean(axes = var_10301_axes_0, keep_dims = var_10301_keep_dims_0, x = var_10299_cast_fp16)[name = string("op_10301_cast_fp16")];
            fp16 var_10291_promoted_1_to_fp16 = const()[name = string("op_10291_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_10302_cast_fp16 = pow(x = peak_219_cast_fp16, y = var_10291_promoted_1_to_fp16)[name = string("op_10302_cast_fp16")];
            fp32 var_10303_epsilon_0 = const()[name = string("op_10303_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_10303_cast_fp16 = inverse(epsilon = var_10303_epsilon_0, x = var_10302_cast_fp16)[name = string("op_10303_cast_fp16")];
            fp16 var_10304_to_fp16 = const()[name = string("op_10304_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_10305_cast_fp16 = mul(x = var_10303_cast_fp16, y = var_10304_to_fp16)[name = string("op_10305_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_10306_cast_fp16 = add(x = var_10301_cast_fp16, y = var_10305_cast_fp16)[name = string("op_10306_cast_fp16")];
            fp32 var_10307_epsilon_0 = const()[name = string("op_10307_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_10307_cast_fp16 = rsqrt(epsilon = var_10307_epsilon_0, x = var_10306_cast_fp16)[name = string("op_10307_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_10308_cast_fp16 = mul(x = x_547_cast_fp16, y = var_10307_cast_fp16)[name = string("op_10308_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_553_cast_fp16 = mul(x = var_10308_cast_fp16, y = model_layers_27_self_attn_q_norm_weight_to_fp16)[name = string("x_553_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_10317 = const()[name = string("op_10317"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10318_cast_fp16 = reshape(shape = var_10317, x = linear_190_cast_fp16)[name = string("op_10318_cast_fp16")];
            tensor<int32, [4]> x_549_perm_0 = const()[name = string("x_549_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10325 = const()[name = string("op_10325"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_549_cast_fp16 = transpose(perm = x_549_perm_0, x = var_10318_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 8, 16, 128]> var_10328_cast_fp16 = abs(x = x_549_cast_fp16)[name = string("op_10328_cast_fp16")];
            tensor<int32, [1]> peak_221_axes_0 = const()[name = string("peak_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_221_cast_fp16 = reduce_max(axes = peak_221_axes_0, keep_dims = var_10325, x = var_10328_cast_fp16)[name = string("peak_221_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_551_cast_fp16 = real_div(x = x_549_cast_fp16, y = peak_221_cast_fp16)[name = string("x_551_cast_fp16")];
            fp16 var_10324_promoted_to_fp16 = const()[name = string("op_10324_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_10332_cast_fp16 = pow(x = x_551_cast_fp16, y = var_10324_promoted_to_fp16)[name = string("op_10332_cast_fp16")];
            tensor<int32, [1]> var_10334_axes_0 = const()[name = string("op_10334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10334_keep_dims_0 = const()[name = string("op_10334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_10334_cast_fp16 = reduce_mean(axes = var_10334_axes_0, keep_dims = var_10334_keep_dims_0, x = var_10332_cast_fp16)[name = string("op_10334_cast_fp16")];
            fp16 var_10324_promoted_1_to_fp16 = const()[name = string("op_10324_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_10335_cast_fp16 = pow(x = peak_221_cast_fp16, y = var_10324_promoted_1_to_fp16)[name = string("op_10335_cast_fp16")];
            fp32 var_10336_epsilon_0 = const()[name = string("op_10336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_10336_cast_fp16 = inverse(epsilon = var_10336_epsilon_0, x = var_10335_cast_fp16)[name = string("op_10336_cast_fp16")];
            fp16 var_10337_to_fp16 = const()[name = string("op_10337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_10338_cast_fp16 = mul(x = var_10336_cast_fp16, y = var_10337_to_fp16)[name = string("op_10338_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_10339_cast_fp16 = add(x = var_10334_cast_fp16, y = var_10338_cast_fp16)[name = string("op_10339_cast_fp16")];
            fp32 var_10340_epsilon_0 = const()[name = string("op_10340_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_10340_cast_fp16 = rsqrt(epsilon = var_10340_epsilon_0, x = var_10339_cast_fp16)[name = string("op_10340_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_10341_cast_fp16 = mul(x = x_551_cast_fp16, y = var_10340_cast_fp16)[name = string("op_10341_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_555_cast_fp16 = mul(x = var_10341_cast_fp16, y = model_layers_27_self_attn_k_norm_weight_to_fp16)[name = string("x_555_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_10350 = const()[name = string("op_10350"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10351_cast_fp16 = reshape(shape = var_10350, x = linear_191_cast_fp16)[name = string("op_10351_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_553_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_553_cast_fp16)[name = string("second_109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10365_cast_fp16 = mul(x = first_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10365_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10366_cast_fp16 = mul(x = second_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10366_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10368_cast_fp16 = sub(x = var_10365_cast_fp16, y = var_10366_cast_fp16)[name = string("op_10368_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10369_cast_fp16 = mul(x = second_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10369_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10370_cast_fp16 = mul(x = first_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10370_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10372_cast_fp16 = add(x = var_10369_cast_fp16, y = var_10370_cast_fp16)[name = string("op_10372_cast_fp16")];
            int32 var_10374 = const()[name = string("op_10374"), 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_10374, interleave = q_interleave_0, values = (var_10368_cast_fp16, var_10372_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_555_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_555_cast_fp16)[name = string("second_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10386_cast_fp16 = mul(x = first_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10386_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10387_cast_fp16 = mul(x = second_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10387_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10389_cast_fp16 = sub(x = var_10386_cast_fp16, y = var_10387_cast_fp16)[name = string("op_10389_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10390_cast_fp16 = mul(x = second_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10390_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10391_cast_fp16 = mul(x = first_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10391_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10393_cast_fp16 = add(x = var_10390_cast_fp16, y = var_10391_cast_fp16)[name = string("op_10393_cast_fp16")];
            int32 var_10395 = const()[name = string("op_10395"), 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_10395, interleave = k_interleave_0, values = (var_10389_cast_fp16, var_10393_cast_fp16))[name = string("k_cast_fp16")];
            tensor<int32, [4]> var_10406_begin_0 = const()[name = string("op_10406_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10406_end_0 = const()[name = string("op_10406_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10406_end_mask_0 = const()[name = string("op_10406_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10406_stride_0 = const()[name = string("op_10406_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10406_cast_fp16 = slice_by_index(begin = var_10406_begin_0, end = var_10406_end_0, end_mask = var_10406_end_mask_0, stride = var_10406_stride_0, x = q_cast_fp16)[name = string("op_10406_cast_fp16")];
            tensor<int32, [4]> var_10416_begin_0 = const()[name = string("op_10416_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_10416_end_0 = const()[name = string("op_10416_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10416_end_mask_0 = const()[name = string("op_10416_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10416_stride_0 = const()[name = string("op_10416_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10416_cast_fp16 = slice_by_index(begin = var_10416_begin_0, end = var_10416_end_0, end_mask = var_10416_end_mask_0, stride = var_10416_stride_0, x = q_cast_fp16)[name = string("op_10416_cast_fp16")];
            int32 var_10418 = const()[name = string("op_10418"), 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_10418, interleave = stacked_interleave_0, values = (var_10406_cast_fp16, var_10416_cast_fp16))[name = string("stacked_cast_fp16")];
            tensor<int32, [4]> var_10424_begin_0 = const()[name = string("op_10424_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_10424_end_0 = const()[name = string("op_10424_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_10424_end_mask_0 = const()[name = string("op_10424_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_10424_cast_fp16 = slice_by_index(begin = var_10424_begin_0, end = var_10424_end_0, end_mask = var_10424_end_mask_0, x = read_state_0)[name = string("op_10424_cast_fp16")];
            tensor<int32, [4]> var_10434_begin_0 = const()[name = string("op_10434_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10434_end_0 = const()[name = string("op_10434_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_10434_end_mask_0 = const()[name = string("op_10434_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_10434_cast_fp16 = slice_by_index(begin = var_10434_begin_0, end = var_10434_end_0, end_mask = var_10434_end_mask_0, x = var_10424_cast_fp16)[name = string("op_10434_cast_fp16")];
            bool var_10443_transpose_x_1 = const()[name = string("op_10443_transpose_x_1"), val = bool(false)];
            bool var_10443_transpose_y_1 = const()[name = string("op_10443_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_10443_cast_fp16 = matmul(transpose_x = var_10443_transpose_x_1, transpose_y = var_10443_transpose_y_1, x = var_10434_cast_fp16, y = stacked_cast_fp16)[name = string("op_10443_cast_fp16")];
            tensor<int32, [4]> var_10446_perm_0 = const()[name = string("op_10446_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_10447_to_fp16 = const()[name = string("op_10447_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_10446_cast_fp16 = transpose(perm = var_10446_perm_0, x = var_10443_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, 32, 4096]> var_10448_cast_fp16 = mul(x = var_10446_cast_fp16, y = var_10447_to_fp16)[name = string("op_10448_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_cast_fp16 = add(x = var_10448_cast_fp16, y = cacheMask)[name = string("cache_weights_cast_fp16")];
            bool var_10454_transpose_x_1 = const()[name = string("op_10454_transpose_x_1"), val = bool(false)];
            bool var_10454_transpose_y_1 = const()[name = string("op_10454_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_10454_cast_fp16 = matmul(transpose_x = var_10454_transpose_x_1, transpose_y = var_10454_transpose_y_1, x = stacked_cast_fp16, y = k_cast_fp16)[name = string("op_10454_cast_fp16")];
            fp16 var_10455_to_fp16 = const()[name = string("op_10455_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_10456_cast_fp16 = mul(x = var_10454_cast_fp16, y = var_10455_to_fp16)[name = string("op_10456_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_cast_fp16 = add(x = var_10456_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_cast_fp16")];
            int32 var_10460 = const()[name = string("op_10460"), val = int32(-1)];
            bool var_10461_interleave_0 = const()[name = string("op_10461_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_10461_cast_fp16 = concat(axis = var_10460, interleave = var_10461_interleave_0, values = (cache_weights_cast_fp16, call_weights_cast_fp16))[name = string("op_10461_cast_fp16")];
            int32 var_10462 = const()[name = string("op_10462"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_cast_fp16 = softmax(axis = var_10462, x = var_10461_cast_fp16)[name = string("shares_cast_fp16")];
            tensor<int32, [4]> var_10469_begin_0 = const()[name = string("op_10469_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10469_end_0 = const()[name = string("op_10469_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_10469_end_mask_0 = const()[name = string("op_10469_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_10469_cast_fp16 = slice_by_index(begin = var_10469_begin_0, end = var_10469_end_0, end_mask = var_10469_end_mask_0, x = shares_cast_fp16)[name = string("op_10469_cast_fp16")];
            tensor<int32, [4]> var_10474_begin_0 = const()[name = string("op_10474_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_10474_end_0 = const()[name = string("op_10474_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_10474_end_mask_0 = const()[name = string("op_10474_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_10474_cast_fp16 = slice_by_index(begin = var_10474_begin_0, end = var_10474_end_0, end_mask = var_10474_end_mask_0, x = read_state_1)[name = string("op_10474_cast_fp16")];
            tensor<int32, [4]> var_10484_begin_0 = const()[name = string("op_10484_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10484_end_0 = const()[name = string("op_10484_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_10484_end_mask_0 = const()[name = string("op_10484_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_10484_cast_fp16 = slice_by_index(begin = var_10484_begin_0, end = var_10484_end_0, end_mask = var_10484_end_mask_0, x = var_10474_cast_fp16)[name = string("op_10484_cast_fp16")];
            bool var_10490_transpose_x_0 = const()[name = string("op_10490_transpose_x_0"), val = bool(false)];
            bool var_10490_transpose_y_0 = const()[name = string("op_10490_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_10490_cast_fp16 = matmul(transpose_x = var_10490_transpose_x_0, transpose_y = var_10490_transpose_y_0, x = var_10469_cast_fp16, y = var_10484_cast_fp16)[name = string("op_10490_cast_fp16")];
            tensor<int32, [4]> var_10495_begin_0 = const()[name = string("op_10495_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_10495_end_0 = const()[name = string("op_10495_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_10495_end_mask_0 = const()[name = string("op_10495_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_10495_cast_fp16 = slice_by_index(begin = var_10495_begin_0, end = var_10495_end_0, end_mask = var_10495_end_mask_0, x = shares_cast_fp16)[name = string("op_10495_cast_fp16")];
            bool var_10496_transpose_x_0 = const()[name = string("op_10496_transpose_x_0"), val = bool(false)];
            bool var_10496_transpose_y_0 = const()[name = string("op_10496_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_10351_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 32, 128]> var_10496_cast_fp16 = matmul(transpose_x = var_10496_transpose_x_0, transpose_y = var_10496_transpose_y_0, x = var_10495_cast_fp16, y = v_cast_fp16)[name = string("op_10496_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_109_cast_fp16 = add(x = var_10490_cast_fp16, y = var_10496_cast_fp16)[name = string("mixed_109_cast_fp16")];
            tensor<int32, [4]> var_10513_begin_0 = const()[name = string("op_10513_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10513_end_0 = const()[name = string("op_10513_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_10513_end_mask_0 = const()[name = string("op_10513_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10513_cast_fp16 = slice_by_index(begin = var_10513_begin_0, end = var_10513_end_0, end_mask = var_10513_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_10513_cast_fp16")];
            tensor<int32, [4]> var_10528_begin_0 = const()[name = string("op_10528_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_10528_end_0 = const()[name = string("op_10528_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_10528_end_mask_0 = const()[name = string("op_10528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10528_cast_fp16 = slice_by_index(begin = var_10528_begin_0, end = var_10528_end_0, end_mask = var_10528_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_10528_cast_fp16")];
            int32 var_10530 = const()[name = string("op_10530"), 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_10530, interleave = mixed_interleave_0, values = (var_10513_cast_fp16, var_10528_cast_fp16))[name = string("mixed_cast_fp16")];
            tensor<int32, [4]> var_10534_perm_0 = const()[name = string("op_10534_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_10538 = const()[name = string("op_10538"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_10534_cast_fp16 = transpose(perm = var_10534_perm_0, x = mixed_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 16, 2048]> input_273_cast_fp16 = reshape(shape = var_10538, x = var_10534_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_557_cast_fp16 = add(x = x_541_cast_fp16, y = linear_192_cast_fp16)[name = string("x_557_cast_fp16")];
            bool var_10548 = const()[name = string("op_10548"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10551_cast_fp16 = abs(x = x_557_cast_fp16)[name = string("op_10551_cast_fp16")];
            tensor<int32, [1]> peak_223_axes_0 = const()[name = string("peak_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_223_cast_fp16 = reduce_max(axes = peak_223_axes_0, keep_dims = var_10548, x = var_10551_cast_fp16)[name = string("peak_223_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_559_cast_fp16 = real_div(x = x_557_cast_fp16, y = peak_223_cast_fp16)[name = string("x_559_cast_fp16")];
            fp16 var_10547_promoted_to_fp16 = const()[name = string("op_10547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10555_cast_fp16 = pow(x = x_559_cast_fp16, y = var_10547_promoted_to_fp16)[name = string("op_10555_cast_fp16")];
            tensor<int32, [1]> var_10557_axes_0 = const()[name = string("op_10557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10557_keep_dims_0 = const()[name = string("op_10557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10557_cast_fp16 = reduce_mean(axes = var_10557_axes_0, keep_dims = var_10557_keep_dims_0, x = var_10555_cast_fp16)[name = string("op_10557_cast_fp16")];
            fp16 var_10547_promoted_1_to_fp16 = const()[name = string("op_10547_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10558_cast_fp16 = pow(x = peak_223_cast_fp16, y = var_10547_promoted_1_to_fp16)[name = string("op_10558_cast_fp16")];
            fp32 var_10559_epsilon_0 = const()[name = string("op_10559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10559_cast_fp16 = inverse(epsilon = var_10559_epsilon_0, x = var_10558_cast_fp16)[name = string("op_10559_cast_fp16")];
            fp16 var_10560_to_fp16 = const()[name = string("op_10560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10561_cast_fp16 = mul(x = var_10559_cast_fp16, y = var_10560_to_fp16)[name = string("op_10561_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10562_cast_fp16 = add(x = var_10557_cast_fp16, y = var_10561_cast_fp16)[name = string("op_10562_cast_fp16")];
            fp32 var_10563_epsilon_0 = const()[name = string("op_10563_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10563_cast_fp16 = rsqrt(epsilon = var_10563_epsilon_0, x = var_10562_cast_fp16)[name = string("op_10563_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10564_cast_fp16 = mul(x = x_559_cast_fp16, y = var_10563_cast_fp16)[name = string("op_10564_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_10564_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_10569_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_10569_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_10569_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_561_cast_fp16 = add(x = x_557_cast_fp16, y = linear_195_cast_fp16)[name = string("x_561_cast_fp16")];
            int32 var_10579 = const()[name = string("op_10579"), val = int32(1)];
            int32 x_563_batch_dims_0 = const()[name = string("x_563_batch_dims_0"), val = int32(0)];
            bool x_563_validate_indices_0 = const()[name = string("x_563_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_563_cast_fp16_cast_uint16 = gather(axis = var_10579, batch_dims = x_563_batch_dims_0, indices = answerRow_to_uint16, validate_indices = x_563_validate_indices_0, x = x_561_cast_fp16)[name = string("x_563_cast_fp16_cast_uint16")];
            bool var_10584 = const()[name = string("op_10584"), val = bool(true)];
            tensor<fp16, [1, 1, 1024]> var_10587_cast_fp16 = abs(x = x_563_cast_fp16_cast_uint16)[name = string("op_10587_cast_fp16")];
            tensor<int32, [1]> peak_axes_0 = const()[name = string("peak_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 1]> peak_cast_fp16 = reduce_max(axes = peak_axes_0, keep_dims = var_10584, x = var_10587_cast_fp16)[name = string("peak_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> x_cast_fp16 = real_div(x = x_563_cast_fp16_cast_uint16, y = peak_cast_fp16)[name = string("x_cast_fp16")];
            fp16 var_10583_promoted_to_fp16 = const()[name = string("op_10583_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1024]> var_10591_cast_fp16 = pow(x = x_cast_fp16, y = var_10583_promoted_to_fp16)[name = string("op_10591_cast_fp16")];
            tensor<int32, [1]> var_10593_axes_0 = const()[name = string("op_10593_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10593_keep_dims_0 = const()[name = string("op_10593_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1]> var_10593_cast_fp16 = reduce_mean(axes = var_10593_axes_0, keep_dims = var_10593_keep_dims_0, x = var_10591_cast_fp16)[name = string("op_10593_cast_fp16")];
            fp16 var_10583_promoted_1_to_fp16 = const()[name = string("op_10583_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1]> var_10594_cast_fp16 = pow(x = peak_cast_fp16, y = var_10583_promoted_1_to_fp16)[name = string("op_10594_cast_fp16")];
            fp32 var_10595_epsilon_0 = const()[name = string("op_10595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 1]> var_10595_cast_fp16 = inverse(epsilon = var_10595_epsilon_0, x = var_10594_cast_fp16)[name = string("op_10595_cast_fp16")];
            fp16 var_10596_to_fp16 = const()[name = string("op_10596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_10597_cast_fp16 = mul(x = var_10595_cast_fp16, y = var_10596_to_fp16)[name = string("op_10597_cast_fp16")];
            tensor<fp16, [1, 1, 1]> var_10598_cast_fp16 = add(x = var_10593_cast_fp16, y = var_10597_cast_fp16)[name = string("op_10598_cast_fp16")];
            fp32 var_10599_epsilon_0 = const()[name = string("op_10599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1]> var_10599_cast_fp16 = rsqrt(epsilon = var_10599_epsilon_0, x = var_10598_cast_fp16)[name = string("op_10599_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> var_10600_cast_fp16 = mul(x = x_cast_fp16, y = var_10599_cast_fp16)[name = string("op_10600_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_10600_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_10606_axis_0 = const()[name = string("op_10606_axis_0"), val = int32(-1)];
            bool var_10606_keep_dims_0 = const()[name = string("op_10606_keep_dims_0"), val = bool(false)];
            string var_10606_output_dtype_0 = const()[name = string("op_10606_output_dtype_0"), val = string("int32")];
            tensor<int32, [1, 1]> nextToken = reduce_argmax(axis = var_10606_axis_0, keep_dims = var_10606_keep_dims_0, output_dtype = var_10606_output_dtype_0, x = linear_196_cast_fp16)[name = string("op_10606_cast_fp16")];
            int32 var_10613 = const()[name = string("op_10613"), val = int32(0)];
            bool var_10614_interleave_0 = const()[name = string("op_10614_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newKeys = concat(axis = var_10613, interleave = var_10614_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_10614_cast_fp16")];
            int32 var_10616 = const()[name = string("op_10616"), val = int32(0)];
            bool var_10617_interleave_0 = const()[name = string("op_10617_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newValues = concat(axis = var_10616, interleave = var_10617_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_10617_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")];
            bool var_1538 = const()[name = string("op_1538"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1541_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [1]> peak_1_axes_0 = const()[name = string("peak_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_1_cast_fp16 = reduce_max(axes = peak_1_axes_0, keep_dims = var_1538, x = var_1541_cast_fp16)[name = string("peak_1_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_3_cast_fp16 = real_div(x = x_1_cast_fp16, y = peak_1_cast_fp16)[name = string("x_3_cast_fp16")];
            fp16 var_1537_promoted_to_fp16 = const()[name = string("op_1537_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1545_cast_fp16 = pow(x = x_3_cast_fp16, y = var_1537_promoted_to_fp16)[name = string("op_1545_cast_fp16")];
            tensor<int32, [1]> var_1547_axes_0 = const()[name = string("op_1547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1547_keep_dims_0 = const()[name = string("op_1547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1547_cast_fp16 = reduce_mean(axes = var_1547_axes_0, keep_dims = var_1547_keep_dims_0, x = var_1545_cast_fp16)[name = string("op_1547_cast_fp16")];
            fp16 var_1537_promoted_1_to_fp16 = const()[name = string("op_1537_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1548_cast_fp16 = pow(x = peak_1_cast_fp16, y = var_1537_promoted_1_to_fp16)[name = string("op_1548_cast_fp16")];
            fp32 var_1549_epsilon_0 = const()[name = string("op_1549_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1549_cast_fp16 = inverse(epsilon = var_1549_epsilon_0, x = var_1548_cast_fp16)[name = string("op_1549_cast_fp16")];
            fp16 var_1550_to_fp16 = const()[name = string("op_1550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1551_cast_fp16 = mul(x = var_1549_cast_fp16, y = var_1550_to_fp16)[name = string("op_1551_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1552_cast_fp16 = add(x = var_1547_cast_fp16, y = var_1551_cast_fp16)[name = string("op_1552_cast_fp16")];
            fp32 var_1553_epsilon_0 = const()[name = string("op_1553_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1553_cast_fp16 = rsqrt(epsilon = var_1553_epsilon_0, x = var_1552_cast_fp16)[name = string("op_1553_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1554_cast_fp16 = mul(x = x_3_cast_fp16, y = var_1553_cast_fp16)[name = string("op_1554_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_1554_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_1563 = const()[name = string("op_1563"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1564_cast_fp16 = reshape(shape = var_1563, x = linear_0_cast_fp16)[name = string("op_1564_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1571 = const()[name = string("op_1571"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_1564_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 16, 16, 128]> var_1574_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_1574_cast_fp16")];
            tensor<int32, [1]> peak_3_axes_0 = const()[name = string("peak_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_3_cast_fp16 = reduce_max(axes = peak_3_axes_0, keep_dims = var_1571, x = var_1574_cast_fp16)[name = string("peak_3_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_7_cast_fp16 = real_div(x = x_5_cast_fp16, y = peak_3_cast_fp16)[name = string("x_7_cast_fp16")];
            fp16 var_1570_promoted_to_fp16 = const()[name = string("op_1570_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_1578_cast_fp16 = pow(x = x_7_cast_fp16, y = var_1570_promoted_to_fp16)[name = string("op_1578_cast_fp16")];
            tensor<int32, [1]> var_1580_axes_0 = const()[name = string("op_1580_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1580_keep_dims_0 = const()[name = string("op_1580_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1580_cast_fp16 = reduce_mean(axes = var_1580_axes_0, keep_dims = var_1580_keep_dims_0, x = var_1578_cast_fp16)[name = string("op_1580_cast_fp16")];
            fp16 var_1570_promoted_1_to_fp16 = const()[name = string("op_1570_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_1581_cast_fp16 = pow(x = peak_3_cast_fp16, y = var_1570_promoted_1_to_fp16)[name = string("op_1581_cast_fp16")];
            fp32 var_1582_epsilon_0 = const()[name = string("op_1582_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_1582_cast_fp16 = inverse(epsilon = var_1582_epsilon_0, x = var_1581_cast_fp16)[name = string("op_1582_cast_fp16")];
            fp16 var_1583_to_fp16 = const()[name = string("op_1583_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1584_cast_fp16 = mul(x = var_1582_cast_fp16, y = var_1583_to_fp16)[name = string("op_1584_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_1585_cast_fp16 = add(x = var_1580_cast_fp16, y = var_1584_cast_fp16)[name = string("op_1585_cast_fp16")];
            fp32 var_1586_epsilon_0 = const()[name = string("op_1586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1586_cast_fp16 = rsqrt(epsilon = var_1586_epsilon_0, x = var_1585_cast_fp16)[name = string("op_1586_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1587_cast_fp16 = mul(x = x_7_cast_fp16, y = var_1586_cast_fp16)[name = string("op_1587_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_13_cast_fp16 = mul(x = var_1587_cast_fp16, y = model_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_13_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_1596 = const()[name = string("op_1596"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1597_cast_fp16 = reshape(shape = var_1596, x = linear_1_cast_fp16)[name = string("op_1597_cast_fp16")];
            tensor<int32, [4]> x_9_perm_0 = const()[name = string("x_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1604 = const()[name = string("op_1604"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_9_cast_fp16 = transpose(perm = x_9_perm_0, x = var_1597_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 8, 16, 128]> var_1607_cast_fp16 = abs(x = x_9_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<int32, [1]> peak_5_axes_0 = const()[name = string("peak_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_5_cast_fp16 = reduce_max(axes = peak_5_axes_0, keep_dims = var_1604, x = var_1607_cast_fp16)[name = string("peak_5_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_11_cast_fp16 = real_div(x = x_9_cast_fp16, y = peak_5_cast_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1603_promoted_to_fp16 = const()[name = string("op_1603_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_1611_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1603_promoted_to_fp16)[name = string("op_1611_cast_fp16")];
            tensor<int32, [1]> var_1613_axes_0 = const()[name = string("op_1613_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1613_keep_dims_0 = const()[name = string("op_1613_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1613_cast_fp16 = reduce_mean(axes = var_1613_axes_0, keep_dims = var_1613_keep_dims_0, x = var_1611_cast_fp16)[name = string("op_1613_cast_fp16")];
            fp16 var_1603_promoted_1_to_fp16 = const()[name = string("op_1603_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_1614_cast_fp16 = pow(x = peak_5_cast_fp16, y = var_1603_promoted_1_to_fp16)[name = string("op_1614_cast_fp16")];
            fp32 var_1615_epsilon_0 = const()[name = string("op_1615_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_1615_cast_fp16 = inverse(epsilon = var_1615_epsilon_0, x = var_1614_cast_fp16)[name = string("op_1615_cast_fp16")];
            fp16 var_1616_to_fp16 = const()[name = string("op_1616_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1617_cast_fp16 = mul(x = var_1615_cast_fp16, y = var_1616_to_fp16)[name = string("op_1617_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_1618_cast_fp16 = add(x = var_1613_cast_fp16, y = var_1617_cast_fp16)[name = string("op_1618_cast_fp16")];
            fp32 var_1619_epsilon_0 = const()[name = string("op_1619_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1619_cast_fp16 = rsqrt(epsilon = var_1619_epsilon_0, x = var_1618_cast_fp16)[name = string("op_1619_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1620_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1619_cast_fp16)[name = string("op_1620_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_15_cast_fp16 = mul(x = var_1620_cast_fp16, y = model_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_15_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_1629 = const()[name = string("op_1629"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1630_cast_fp16 = reshape(shape = var_1629, x = linear_2_cast_fp16)[name = string("op_1630_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_13_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_13_cast_fp16)[name = string("second_1_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1644_cast_fp16 = mul(x = first_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1644_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1645_cast_fp16 = mul(x = second_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 16, 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, 16, 16, 64]> var_1648_cast_fp16 = mul(x = second_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1649_cast_fp16 = mul(x = first_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1649_cast_fp16")];
            tensor<fp16, [1, 16, 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 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_1653, interleave = q_1_interleave_0, values = (var_1647_cast_fp16, var_1651_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_15_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_15_cast_fp16)[name = string("second_3_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1665_cast_fp16 = mul(x = first_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1665_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1666_cast_fp16 = mul(x = second_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1666_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1668_cast_fp16 = sub(x = var_1665_cast_fp16, y = var_1666_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1669_cast_fp16 = mul(x = second_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1669_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1670_cast_fp16 = mul(x = first_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1672_cast_fp16 = add(x = var_1669_cast_fp16, y = var_1670_cast_fp16)[name = string("op_1672_cast_fp16")];
            int32 var_1674 = const()[name = string("op_1674"), 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_1674, interleave = k_1_interleave_0, values = (var_1668_cast_fp16, var_1672_cast_fp16))[name = string("k_1_cast_fp16")];
            tensor<int32, [4]> var_1685_begin_0 = const()[name = string("op_1685_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1685_end_0 = const()[name = string("op_1685_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1685_end_mask_0 = const()[name = string("op_1685_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1685_stride_0 = const()[name = string("op_1685_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1685_cast_fp16 = slice_by_index(begin = var_1685_begin_0, end = var_1685_end_0, end_mask = var_1685_end_mask_0, stride = var_1685_stride_0, x = q_1_cast_fp16)[name = string("op_1685_cast_fp16")];
            tensor<int32, [4]> var_1695_begin_0 = const()[name = string("op_1695_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1695_end_0 = const()[name = string("op_1695_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1695_end_mask_0 = const()[name = string("op_1695_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1695_stride_0 = const()[name = string("op_1695_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1695_cast_fp16 = slice_by_index(begin = var_1695_begin_0, end = var_1695_end_0, end_mask = var_1695_end_mask_0, stride = var_1695_stride_0, x = q_1_cast_fp16)[name = string("op_1695_cast_fp16")];
            int32 var_1697 = const()[name = string("op_1697"), 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_1697, interleave = stacked_1_interleave_0, values = (var_1685_cast_fp16, var_1695_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_1703_begin_0 = const()[name = string("op_1703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1703_end_0 = const()[name = string("op_1703_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_1703_end_mask_0 = const()[name = string("op_1703_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_1703_cast_fp16 = slice_by_index(begin = var_1703_begin_0, end = var_1703_end_0, end_mask = var_1703_end_mask_0, x = read_state_0)[name = string("op_1703_cast_fp16")];
            tensor<int32, [4]> var_1713_begin_0 = const()[name = string("op_1713_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1713_end_0 = const()[name = string("op_1713_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_1713_end_mask_0 = const()[name = string("op_1713_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_1713_cast_fp16 = slice_by_index(begin = var_1713_begin_0, end = var_1713_end_0, end_mask = var_1713_end_mask_0, x = var_1703_cast_fp16)[name = string("op_1713_cast_fp16")];
            bool var_1722_transpose_x_1 = const()[name = string("op_1722_transpose_x_1"), val = bool(false)];
            bool var_1722_transpose_y_1 = const()[name = string("op_1722_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_1722_cast_fp16 = matmul(transpose_x = var_1722_transpose_x_1, transpose_y = var_1722_transpose_y_1, x = var_1713_cast_fp16, y = stacked_1_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<int32, [4]> var_1725_perm_0 = const()[name = string("op_1725_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_1726_to_fp16 = const()[name = string("op_1726_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_1725_cast_fp16 = transpose(perm = var_1725_perm_0, x = var_1722_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 8, 32, 512]> var_1727_cast_fp16 = mul(x = var_1725_cast_fp16, y = var_1726_to_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_1_cast_fp16 = add(x = var_1727_cast_fp16, y = cacheMask)[name = string("cache_weights_1_cast_fp16")];
            bool var_1733_transpose_x_1 = const()[name = string("op_1733_transpose_x_1"), val = bool(false)];
            bool var_1733_transpose_y_1 = const()[name = string("op_1733_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_1733_cast_fp16 = matmul(transpose_x = var_1733_transpose_x_1, transpose_y = var_1733_transpose_y_1, x = stacked_1_cast_fp16, y = k_1_cast_fp16)[name = string("op_1733_cast_fp16")];
            fp16 var_1734_to_fp16 = const()[name = string("op_1734_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_1735_cast_fp16 = mul(x = var_1733_cast_fp16, y = var_1734_to_fp16)[name = string("op_1735_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_1735_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_1_cast_fp16")];
            int32 var_1739 = const()[name = string("op_1739"), val = int32(-1)];
            bool var_1740_interleave_0 = const()[name = string("op_1740_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_1740_cast_fp16 = concat(axis = var_1739, interleave = var_1740_interleave_0, values = (cache_weights_1_cast_fp16, call_weights_1_cast_fp16))[name = string("op_1740_cast_fp16")];
            int32 var_1741 = const()[name = string("op_1741"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_1_cast_fp16 = softmax(axis = var_1741, x = var_1740_cast_fp16)[name = string("shares_1_cast_fp16")];
            tensor<int32, [4]> var_1748_begin_0 = const()[name = string("op_1748_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1748_end_0 = const()[name = string("op_1748_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_1748_end_mask_0 = const()[name = string("op_1748_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_1748_cast_fp16 = slice_by_index(begin = var_1748_begin_0, end = var_1748_end_0, end_mask = var_1748_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<fp16, [28, 8, 528, 128]> read_state_1 = read_state(input = valueCache)[name = string("read_state_1")];
            tensor<int32, [4]> var_1753_begin_0 = const()[name = string("op_1753_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1753_end_0 = const()[name = string("op_1753_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_1753_end_mask_0 = const()[name = string("op_1753_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> 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 = read_state_1)[name = string("op_1753_cast_fp16")];
            tensor<int32, [4]> var_1763_begin_0 = const()[name = string("op_1763_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1763_end_0 = const()[name = string("op_1763_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_1763_end_mask_0 = const()[name = string("op_1763_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_1763_cast_fp16 = slice_by_index(begin = var_1763_begin_0, end = var_1763_end_0, end_mask = var_1763_end_mask_0, x = var_1753_cast_fp16)[name = string("op_1763_cast_fp16")];
            bool var_1769_transpose_x_0 = const()[name = string("op_1769_transpose_x_0"), val = bool(false)];
            bool var_1769_transpose_y_0 = const()[name = string("op_1769_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1769_cast_fp16 = matmul(transpose_x = var_1769_transpose_x_0, transpose_y = var_1769_transpose_y_0, x = var_1748_cast_fp16, y = var_1763_cast_fp16)[name = string("op_1769_cast_fp16")];
            tensor<int32, [4]> var_1774_begin_0 = const()[name = string("op_1774_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_1774_end_0 = const()[name = string("op_1774_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_1774_end_mask_0 = const()[name = string("op_1774_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_1774_cast_fp16 = slice_by_index(begin = var_1774_begin_0, end = var_1774_end_0, end_mask = var_1774_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1774_cast_fp16")];
            bool var_1775_transpose_x_0 = const()[name = string("op_1775_transpose_x_0"), val = bool(false)];
            bool var_1775_transpose_y_0 = const()[name = string("op_1775_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_1630_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 8, 32, 128]> var_1775_cast_fp16 = matmul(transpose_x = var_1775_transpose_x_0, transpose_y = var_1775_transpose_y_0, x = var_1774_cast_fp16, y = v_1_cast_fp16)[name = string("op_1775_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_1_cast_fp16 = add(x = var_1769_cast_fp16, y = var_1775_cast_fp16)[name = string("mixed_1_cast_fp16")];
            tensor<int32, [4]> var_1792_begin_0 = const()[name = string("op_1792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1792_end_0 = const()[name = string("op_1792_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_1792_end_mask_0 = const()[name = string("op_1792_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1792_cast_fp16 = slice_by_index(begin = var_1792_begin_0, end = var_1792_end_0, end_mask = var_1792_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1792_cast_fp16")];
            tensor<int32, [4]> var_1807_begin_0 = const()[name = string("op_1807_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1807_end_0 = const()[name = string("op_1807_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_1807_end_mask_0 = const()[name = string("op_1807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1807_cast_fp16 = slice_by_index(begin = var_1807_begin_0, end = var_1807_end_0, end_mask = var_1807_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1807_cast_fp16")];
            int32 var_1809 = const()[name = string("op_1809"), 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_1809, interleave = mixed_3_interleave_0, values = (var_1792_cast_fp16, var_1807_cast_fp16))[name = string("mixed_3_cast_fp16")];
            tensor<int32, [4]> var_1813_perm_0 = const()[name = string("op_1813_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1817 = const()[name = string("op_1817"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_1813_cast_fp16 = transpose(perm = var_1813_perm_0, x = mixed_3_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 16, 2048]> input_3_cast_fp16 = reshape(shape = var_1817, x = var_1813_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_17_cast_fp16 = add(x = x_1_cast_fp16, y = linear_3_cast_fp16)[name = string("x_17_cast_fp16")];
            bool var_1827 = const()[name = string("op_1827"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1830_cast_fp16 = abs(x = x_17_cast_fp16)[name = string("op_1830_cast_fp16")];
            tensor<int32, [1]> peak_7_axes_0 = const()[name = string("peak_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_7_cast_fp16 = reduce_max(axes = peak_7_axes_0, keep_dims = var_1827, x = var_1830_cast_fp16)[name = string("peak_7_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_19_cast_fp16 = real_div(x = x_17_cast_fp16, y = peak_7_cast_fp16)[name = string("x_19_cast_fp16")];
            fp16 var_1826_promoted_to_fp16 = const()[name = string("op_1826_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1834_cast_fp16 = pow(x = x_19_cast_fp16, y = var_1826_promoted_to_fp16)[name = string("op_1834_cast_fp16")];
            tensor<int32, [1]> var_1836_axes_0 = const()[name = string("op_1836_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1836_keep_dims_0 = const()[name = string("op_1836_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1836_cast_fp16 = reduce_mean(axes = var_1836_axes_0, keep_dims = var_1836_keep_dims_0, x = var_1834_cast_fp16)[name = string("op_1836_cast_fp16")];
            fp16 var_1826_promoted_1_to_fp16 = const()[name = string("op_1826_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1837_cast_fp16 = pow(x = peak_7_cast_fp16, y = var_1826_promoted_1_to_fp16)[name = string("op_1837_cast_fp16")];
            fp32 var_1838_epsilon_0 = const()[name = string("op_1838_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1838_cast_fp16 = inverse(epsilon = var_1838_epsilon_0, x = var_1837_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 = mul(x = var_1838_cast_fp16, y = var_1839_to_fp16)[name = string("op_1840_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1841_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1840_cast_fp16)[name = string("op_1841_cast_fp16")];
            fp32 var_1842_epsilon_0 = const()[name = string("op_1842_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1842_cast_fp16 = rsqrt(epsilon = var_1842_epsilon_0, x = var_1841_cast_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1843_cast_fp16 = mul(x = x_19_cast_fp16, y = var_1842_cast_fp16)[name = string("op_1843_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_1843_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_1848_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_1848_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_1848_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_21_cast_fp16 = add(x = x_17_cast_fp16, y = linear_6_cast_fp16)[name = string("x_21_cast_fp16")];
            bool var_1861 = const()[name = string("op_1861"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_1864_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1864_cast_fp16")];
            tensor<int32, [1]> peak_9_axes_0 = const()[name = string("peak_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_9_cast_fp16 = reduce_max(axes = peak_9_axes_0, keep_dims = var_1861, x = var_1864_cast_fp16)[name = string("peak_9_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_23_cast_fp16 = real_div(x = x_21_cast_fp16, y = peak_9_cast_fp16)[name = string("x_23_cast_fp16")];
            fp16 var_1860_promoted_to_fp16 = const()[name = string("op_1860_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1868_cast_fp16 = pow(x = x_23_cast_fp16, y = var_1860_promoted_to_fp16)[name = string("op_1868_cast_fp16")];
            tensor<int32, [1]> var_1870_axes_0 = const()[name = string("op_1870_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1870_keep_dims_0 = const()[name = string("op_1870_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1870_cast_fp16 = reduce_mean(axes = var_1870_axes_0, keep_dims = var_1870_keep_dims_0, x = var_1868_cast_fp16)[name = string("op_1870_cast_fp16")];
            fp16 var_1860_promoted_1_to_fp16 = const()[name = string("op_1860_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_1871_cast_fp16 = pow(x = peak_9_cast_fp16, y = var_1860_promoted_1_to_fp16)[name = string("op_1871_cast_fp16")];
            fp32 var_1872_epsilon_0 = const()[name = string("op_1872_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_1872_cast_fp16 = inverse(epsilon = var_1872_epsilon_0, x = var_1871_cast_fp16)[name = string("op_1872_cast_fp16")];
            fp16 var_1873_to_fp16 = const()[name = string("op_1873_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1874_cast_fp16 = mul(x = var_1872_cast_fp16, y = var_1873_to_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_1875_cast_fp16 = add(x = var_1870_cast_fp16, y = var_1874_cast_fp16)[name = string("op_1875_cast_fp16")];
            fp32 var_1876_epsilon_0 = const()[name = string("op_1876_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1876_cast_fp16 = rsqrt(epsilon = var_1876_epsilon_0, x = var_1875_cast_fp16)[name = string("op_1876_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1877_cast_fp16 = mul(x = x_23_cast_fp16, y = var_1876_cast_fp16)[name = string("op_1877_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_1877_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_1886 = const()[name = string("op_1886"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1887_cast_fp16 = reshape(shape = var_1886, x = linear_7_cast_fp16)[name = string("op_1887_cast_fp16")];
            tensor<int32, [4]> x_25_perm_0 = const()[name = string("x_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1894 = const()[name = string("op_1894"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_25_cast_fp16 = transpose(perm = x_25_perm_0, x = var_1887_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 16, 16, 128]> var_1897_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1897_cast_fp16")];
            tensor<int32, [1]> peak_11_axes_0 = const()[name = string("peak_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_11_cast_fp16 = reduce_max(axes = peak_11_axes_0, keep_dims = var_1894, x = var_1897_cast_fp16)[name = string("peak_11_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_27_cast_fp16 = real_div(x = x_25_cast_fp16, y = peak_11_cast_fp16)[name = string("x_27_cast_fp16")];
            fp16 var_1893_promoted_to_fp16 = const()[name = string("op_1893_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_1901_cast_fp16 = pow(x = x_27_cast_fp16, y = var_1893_promoted_to_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_1903_axes_0 = const()[name = string("op_1903_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1903_keep_dims_0 = const()[name = string("op_1903_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1903_cast_fp16 = reduce_mean(axes = var_1903_axes_0, keep_dims = var_1903_keep_dims_0, x = var_1901_cast_fp16)[name = string("op_1903_cast_fp16")];
            fp16 var_1893_promoted_1_to_fp16 = const()[name = string("op_1893_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_1904_cast_fp16 = pow(x = peak_11_cast_fp16, y = var_1893_promoted_1_to_fp16)[name = string("op_1904_cast_fp16")];
            fp32 var_1905_epsilon_0 = const()[name = string("op_1905_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_1905_cast_fp16 = inverse(epsilon = var_1905_epsilon_0, x = var_1904_cast_fp16)[name = string("op_1905_cast_fp16")];
            fp16 var_1906_to_fp16 = const()[name = string("op_1906_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1907_cast_fp16 = mul(x = var_1905_cast_fp16, y = var_1906_to_fp16)[name = string("op_1907_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_1908_cast_fp16 = add(x = var_1903_cast_fp16, y = var_1907_cast_fp16)[name = string("op_1908_cast_fp16")];
            fp32 var_1909_epsilon_0 = const()[name = string("op_1909_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1909_cast_fp16 = rsqrt(epsilon = var_1909_epsilon_0, x = var_1908_cast_fp16)[name = string("op_1909_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1910_cast_fp16 = mul(x = x_27_cast_fp16, y = var_1909_cast_fp16)[name = string("op_1910_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_33_cast_fp16 = mul(x = var_1910_cast_fp16, y = model_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_33_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_1919 = const()[name = string("op_1919"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1920_cast_fp16 = reshape(shape = var_1919, x = linear_8_cast_fp16)[name = string("op_1920_cast_fp16")];
            tensor<int32, [4]> x_29_perm_0 = const()[name = string("x_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_1927 = const()[name = string("op_1927"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_29_cast_fp16 = transpose(perm = x_29_perm_0, x = var_1920_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 8, 16, 128]> var_1930_cast_fp16 = abs(x = x_29_cast_fp16)[name = string("op_1930_cast_fp16")];
            tensor<int32, [1]> peak_13_axes_0 = const()[name = string("peak_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_13_cast_fp16 = reduce_max(axes = peak_13_axes_0, keep_dims = var_1927, x = var_1930_cast_fp16)[name = string("peak_13_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_31_cast_fp16 = real_div(x = x_29_cast_fp16, y = peak_13_cast_fp16)[name = string("x_31_cast_fp16")];
            fp16 var_1926_promoted_to_fp16 = const()[name = string("op_1926_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_1934_cast_fp16 = pow(x = x_31_cast_fp16, y = var_1926_promoted_to_fp16)[name = string("op_1934_cast_fp16")];
            tensor<int32, [1]> var_1936_axes_0 = const()[name = string("op_1936_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1936_keep_dims_0 = const()[name = string("op_1936_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1936_cast_fp16 = reduce_mean(axes = var_1936_axes_0, keep_dims = var_1936_keep_dims_0, x = var_1934_cast_fp16)[name = string("op_1936_cast_fp16")];
            fp16 var_1926_promoted_1_to_fp16 = const()[name = string("op_1926_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_1937_cast_fp16 = pow(x = peak_13_cast_fp16, y = var_1926_promoted_1_to_fp16)[name = string("op_1937_cast_fp16")];
            fp32 var_1938_epsilon_0 = const()[name = string("op_1938_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_1938_cast_fp16 = inverse(epsilon = var_1938_epsilon_0, x = var_1937_cast_fp16)[name = string("op_1938_cast_fp16")];
            fp16 var_1939_to_fp16 = const()[name = string("op_1939_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1940_cast_fp16 = mul(x = var_1938_cast_fp16, y = var_1939_to_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_1941_cast_fp16 = add(x = var_1936_cast_fp16, y = var_1940_cast_fp16)[name = string("op_1941_cast_fp16")];
            fp32 var_1942_epsilon_0 = const()[name = string("op_1942_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1942_cast_fp16 = rsqrt(epsilon = var_1942_epsilon_0, x = var_1941_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1943_cast_fp16 = mul(x = x_31_cast_fp16, y = var_1942_cast_fp16)[name = string("op_1943_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_35_cast_fp16 = mul(x = var_1943_cast_fp16, y = model_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_35_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_1952 = const()[name = string("op_1952"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1953_cast_fp16 = reshape(shape = var_1952, x = linear_9_cast_fp16)[name = string("op_1953_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_33_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_33_cast_fp16)[name = string("second_5_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1967_cast_fp16 = mul(x = first_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1967_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1968_cast_fp16 = mul(x = second_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1968_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1970_cast_fp16 = sub(x = var_1967_cast_fp16, y = var_1968_cast_fp16)[name = string("op_1970_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1971_cast_fp16 = mul(x = second_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1971_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1972_cast_fp16 = mul(x = first_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1972_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1974_cast_fp16 = add(x = var_1971_cast_fp16, y = var_1972_cast_fp16)[name = string("op_1974_cast_fp16")];
            int32 var_1976 = const()[name = string("op_1976"), 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_1976, interleave = q_3_interleave_0, values = (var_1970_cast_fp16, var_1974_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_35_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_35_cast_fp16)[name = string("second_7_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1988_cast_fp16 = mul(x = first_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1988_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1989_cast_fp16 = mul(x = second_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1989_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1991_cast_fp16 = sub(x = var_1988_cast_fp16, y = var_1989_cast_fp16)[name = string("op_1991_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1992_cast_fp16 = mul(x = second_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1992_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1993_cast_fp16 = mul(x = first_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1993_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1995_cast_fp16 = add(x = var_1992_cast_fp16, y = var_1993_cast_fp16)[name = string("op_1995_cast_fp16")];
            int32 var_1997 = const()[name = string("op_1997"), 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_1997, interleave = k_3_interleave_0, values = (var_1991_cast_fp16, var_1995_cast_fp16))[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> var_2008_begin_0 = const()[name = string("op_2008_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2008_end_0 = const()[name = string("op_2008_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2008_end_mask_0 = const()[name = string("op_2008_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2008_stride_0 = const()[name = string("op_2008_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2008_cast_fp16 = slice_by_index(begin = var_2008_begin_0, end = var_2008_end_0, end_mask = var_2008_end_mask_0, stride = var_2008_stride_0, x = q_3_cast_fp16)[name = string("op_2008_cast_fp16")];
            tensor<int32, [4]> var_2018_begin_0 = const()[name = string("op_2018_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2018_end_0 = const()[name = string("op_2018_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2018_end_mask_0 = const()[name = string("op_2018_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2018_stride_0 = const()[name = string("op_2018_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2018_cast_fp16 = slice_by_index(begin = var_2018_begin_0, end = var_2018_end_0, end_mask = var_2018_end_mask_0, stride = var_2018_stride_0, x = q_3_cast_fp16)[name = string("op_2018_cast_fp16")];
            int32 var_2020 = const()[name = string("op_2020"), 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_2020, interleave = stacked_3_interleave_0, values = (var_2008_cast_fp16, var_2018_cast_fp16))[name = string("stacked_3_cast_fp16")];
            tensor<int32, [4]> var_2026_begin_0 = const()[name = string("op_2026_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2026_end_0 = const()[name = string("op_2026_end_0"), val = tensor<int32, [4]>([2, 8, 528, 128])];
            tensor<bool, [4]> var_2026_end_mask_0 = const()[name = string("op_2026_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2026_cast_fp16 = slice_by_index(begin = var_2026_begin_0, end = var_2026_end_0, end_mask = var_2026_end_mask_0, x = read_state_0)[name = string("op_2026_cast_fp16")];
            tensor<int32, [4]> var_2036_begin_0 = const()[name = string("op_2036_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2036_end_0 = const()[name = string("op_2036_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2036_end_mask_0 = const()[name = string("op_2036_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2036_cast_fp16 = slice_by_index(begin = var_2036_begin_0, end = var_2036_end_0, end_mask = var_2036_end_mask_0, x = var_2026_cast_fp16)[name = string("op_2036_cast_fp16")];
            bool var_2045_transpose_x_1 = const()[name = string("op_2045_transpose_x_1"), val = bool(false)];
            bool var_2045_transpose_y_1 = const()[name = string("op_2045_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_2045_cast_fp16 = matmul(transpose_x = var_2045_transpose_x_1, transpose_y = var_2045_transpose_y_1, x = var_2036_cast_fp16, y = stacked_3_cast_fp16)[name = string("op_2045_cast_fp16")];
            tensor<int32, [4]> var_2048_perm_0 = const()[name = string("op_2048_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2049_to_fp16 = const()[name = string("op_2049_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_2048_cast_fp16 = transpose(perm = var_2048_perm_0, x = var_2045_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 32, 512]> var_2050_cast_fp16 = mul(x = var_2048_cast_fp16, y = var_2049_to_fp16)[name = string("op_2050_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_3_cast_fp16 = add(x = var_2050_cast_fp16, y = cacheMask)[name = string("cache_weights_3_cast_fp16")];
            bool var_2056_transpose_x_1 = const()[name = string("op_2056_transpose_x_1"), val = bool(false)];
            bool var_2056_transpose_y_1 = const()[name = string("op_2056_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2056_cast_fp16 = matmul(transpose_x = var_2056_transpose_x_1, transpose_y = var_2056_transpose_y_1, x = stacked_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_2056_cast_fp16")];
            fp16 var_2057_to_fp16 = const()[name = string("op_2057_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2058_cast_fp16 = mul(x = var_2056_cast_fp16, y = var_2057_to_fp16)[name = string("op_2058_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_3_cast_fp16 = add(x = var_2058_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_3_cast_fp16")];
            int32 var_2062 = const()[name = string("op_2062"), val = int32(-1)];
            bool var_2063_interleave_0 = const()[name = string("op_2063_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_2063_cast_fp16 = concat(axis = var_2062, interleave = var_2063_interleave_0, values = (cache_weights_3_cast_fp16, call_weights_3_cast_fp16))[name = string("op_2063_cast_fp16")];
            int32 var_2064 = const()[name = string("op_2064"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_3_cast_fp16 = softmax(axis = var_2064, x = var_2063_cast_fp16)[name = string("shares_3_cast_fp16")];
            tensor<int32, [4]> var_2071_begin_0 = const()[name = string("op_2071_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2071_end_0 = const()[name = string("op_2071_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_2071_end_mask_0 = const()[name = string("op_2071_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_2071_cast_fp16 = slice_by_index(begin = var_2071_begin_0, end = var_2071_end_0, end_mask = var_2071_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2071_cast_fp16")];
            tensor<int32, [4]> var_2076_begin_0 = const()[name = string("op_2076_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2076_end_0 = const()[name = string("op_2076_end_0"), val = tensor<int32, [4]>([2, 8, 528, 128])];
            tensor<bool, [4]> var_2076_end_mask_0 = const()[name = string("op_2076_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2076_cast_fp16 = slice_by_index(begin = var_2076_begin_0, end = var_2076_end_0, end_mask = var_2076_end_mask_0, x = read_state_1)[name = string("op_2076_cast_fp16")];
            tensor<int32, [4]> var_2086_begin_0 = const()[name = string("op_2086_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2086_end_0 = const()[name = string("op_2086_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2086_end_mask_0 = const()[name = string("op_2086_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2086_cast_fp16 = slice_by_index(begin = var_2086_begin_0, end = var_2086_end_0, end_mask = var_2086_end_mask_0, x = var_2076_cast_fp16)[name = string("op_2086_cast_fp16")];
            bool var_2092_transpose_x_0 = const()[name = string("op_2092_transpose_x_0"), val = bool(false)];
            bool var_2092_transpose_y_0 = const()[name = string("op_2092_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2092_cast_fp16 = matmul(transpose_x = var_2092_transpose_x_0, transpose_y = var_2092_transpose_y_0, x = var_2071_cast_fp16, y = var_2086_cast_fp16)[name = string("op_2092_cast_fp16")];
            tensor<int32, [4]> var_2097_begin_0 = const()[name = string("op_2097_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_2097_end_0 = const()[name = string("op_2097_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_2097_end_mask_0 = const()[name = string("op_2097_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2097_cast_fp16 = slice_by_index(begin = var_2097_begin_0, end = var_2097_end_0, end_mask = var_2097_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2097_cast_fp16")];
            bool var_2098_transpose_x_0 = const()[name = string("op_2098_transpose_x_0"), val = bool(false)];
            bool var_2098_transpose_y_0 = const()[name = string("op_2098_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_1953_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 32, 128]> var_2098_cast_fp16 = matmul(transpose_x = var_2098_transpose_x_0, transpose_y = var_2098_transpose_y_0, x = var_2097_cast_fp16, y = v_3_cast_fp16)[name = string("op_2098_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_5_cast_fp16 = add(x = var_2092_cast_fp16, y = var_2098_cast_fp16)[name = string("mixed_5_cast_fp16")];
            tensor<int32, [4]> var_2115_begin_0 = const()[name = string("op_2115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2115_end_0 = const()[name = string("op_2115_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2115_end_mask_0 = const()[name = string("op_2115_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2115_cast_fp16 = slice_by_index(begin = var_2115_begin_0, end = var_2115_end_0, end_mask = var_2115_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<int32, [4]> var_2130_begin_0 = const()[name = string("op_2130_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2130_end_0 = const()[name = string("op_2130_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2130_end_mask_0 = const()[name = string("op_2130_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2130_cast_fp16 = slice_by_index(begin = var_2130_begin_0, end = var_2130_end_0, end_mask = var_2130_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2130_cast_fp16")];
            int32 var_2132 = const()[name = string("op_2132"), 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_2132, interleave = mixed_7_interleave_0, values = (var_2115_cast_fp16, var_2130_cast_fp16))[name = string("mixed_7_cast_fp16")];
            tensor<int32, [4]> var_2136_perm_0 = const()[name = string("op_2136_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2140 = const()[name = string("op_2140"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2136_cast_fp16 = transpose(perm = var_2136_perm_0, x = mixed_7_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 16, 2048]> input_13_cast_fp16 = reshape(shape = var_2140, x = var_2136_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_37_cast_fp16 = add(x = x_21_cast_fp16, y = linear_10_cast_fp16)[name = string("x_37_cast_fp16")];
            bool var_2150 = const()[name = string("op_2150"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2153_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2153_cast_fp16")];
            tensor<int32, [1]> peak_15_axes_0 = const()[name = string("peak_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_15_cast_fp16 = reduce_max(axes = peak_15_axes_0, keep_dims = var_2150, x = var_2153_cast_fp16)[name = string("peak_15_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_39_cast_fp16 = real_div(x = x_37_cast_fp16, y = peak_15_cast_fp16)[name = string("x_39_cast_fp16")];
            fp16 var_2149_promoted_to_fp16 = const()[name = string("op_2149_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2157_cast_fp16 = pow(x = x_39_cast_fp16, y = var_2149_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, 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_2149_promoted_1_to_fp16 = const()[name = string("op_2149_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2160_cast_fp16 = pow(x = peak_15_cast_fp16, y = var_2149_promoted_1_to_fp16)[name = string("op_2160_cast_fp16")];
            fp32 var_2161_epsilon_0 = const()[name = string("op_2161_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2161_cast_fp16 = inverse(epsilon = var_2161_epsilon_0, x = var_2160_cast_fp16)[name = string("op_2161_cast_fp16")];
            fp16 var_2162_to_fp16 = const()[name = string("op_2162_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2163_cast_fp16 = mul(x = var_2161_cast_fp16, y = var_2162_to_fp16)[name = string("op_2163_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2164_cast_fp16 = add(x = var_2159_cast_fp16, y = var_2163_cast_fp16)[name = string("op_2164_cast_fp16")];
            fp32 var_2165_epsilon_0 = const()[name = string("op_2165_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2165_cast_fp16 = rsqrt(epsilon = var_2165_epsilon_0, x = var_2164_cast_fp16)[name = string("op_2165_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2166_cast_fp16 = mul(x = x_39_cast_fp16, y = var_2165_cast_fp16)[name = string("op_2166_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_2166_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_2171_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_2171_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_2171_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_41_cast_fp16 = add(x = x_37_cast_fp16, y = linear_13_cast_fp16)[name = string("x_41_cast_fp16")];
            bool var_2184 = const()[name = string("op_2184"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2187_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<int32, [1]> peak_17_axes_0 = const()[name = string("peak_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_17_cast_fp16 = reduce_max(axes = peak_17_axes_0, keep_dims = var_2184, x = var_2187_cast_fp16)[name = string("peak_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_43_cast_fp16 = real_div(x = x_41_cast_fp16, y = peak_17_cast_fp16)[name = string("x_43_cast_fp16")];
            fp16 var_2183_promoted_to_fp16 = const()[name = string("op_2183_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2191_cast_fp16 = pow(x = x_43_cast_fp16, y = var_2183_promoted_to_fp16)[name = string("op_2191_cast_fp16")];
            tensor<int32, [1]> var_2193_axes_0 = const()[name = string("op_2193_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2193_keep_dims_0 = const()[name = string("op_2193_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2193_cast_fp16 = reduce_mean(axes = var_2193_axes_0, keep_dims = var_2193_keep_dims_0, x = var_2191_cast_fp16)[name = string("op_2193_cast_fp16")];
            fp16 var_2183_promoted_1_to_fp16 = const()[name = string("op_2183_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2194_cast_fp16 = pow(x = peak_17_cast_fp16, y = var_2183_promoted_1_to_fp16)[name = string("op_2194_cast_fp16")];
            fp32 var_2195_epsilon_0 = const()[name = string("op_2195_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2195_cast_fp16 = inverse(epsilon = var_2195_epsilon_0, x = var_2194_cast_fp16)[name = string("op_2195_cast_fp16")];
            fp16 var_2196_to_fp16 = const()[name = string("op_2196_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2197_cast_fp16 = mul(x = var_2195_cast_fp16, y = var_2196_to_fp16)[name = string("op_2197_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2198_cast_fp16 = add(x = var_2193_cast_fp16, y = var_2197_cast_fp16)[name = string("op_2198_cast_fp16")];
            fp32 var_2199_epsilon_0 = const()[name = string("op_2199_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2199_cast_fp16 = rsqrt(epsilon = var_2199_epsilon_0, x = var_2198_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2200_cast_fp16 = mul(x = x_43_cast_fp16, y = var_2199_cast_fp16)[name = string("op_2200_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_2200_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_2209 = const()[name = string("op_2209"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2210_cast_fp16 = reshape(shape = var_2209, x = linear_14_cast_fp16)[name = string("op_2210_cast_fp16")];
            tensor<int32, [4]> x_45_perm_0 = const()[name = string("x_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2217 = const()[name = string("op_2217"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_45_cast_fp16 = transpose(perm = x_45_perm_0, x = var_2210_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 16, 16, 128]> var_2220_cast_fp16 = abs(x = x_45_cast_fp16)[name = string("op_2220_cast_fp16")];
            tensor<int32, [1]> peak_19_axes_0 = const()[name = string("peak_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_19_cast_fp16 = reduce_max(axes = peak_19_axes_0, keep_dims = var_2217, x = var_2220_cast_fp16)[name = string("peak_19_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_47_cast_fp16 = real_div(x = x_45_cast_fp16, y = peak_19_cast_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2216_promoted_to_fp16 = const()[name = string("op_2216_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2224_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2216_promoted_to_fp16)[name = string("op_2224_cast_fp16")];
            tensor<int32, [1]> var_2226_axes_0 = const()[name = string("op_2226_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2226_keep_dims_0 = const()[name = string("op_2226_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2226_cast_fp16 = reduce_mean(axes = var_2226_axes_0, keep_dims = var_2226_keep_dims_0, x = var_2224_cast_fp16)[name = string("op_2226_cast_fp16")];
            fp16 var_2216_promoted_1_to_fp16 = const()[name = string("op_2216_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2227_cast_fp16 = pow(x = peak_19_cast_fp16, y = var_2216_promoted_1_to_fp16)[name = string("op_2227_cast_fp16")];
            fp32 var_2228_epsilon_0 = const()[name = string("op_2228_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2228_cast_fp16 = inverse(epsilon = var_2228_epsilon_0, x = var_2227_cast_fp16)[name = string("op_2228_cast_fp16")];
            fp16 var_2229_to_fp16 = const()[name = string("op_2229_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2230_cast_fp16 = mul(x = var_2228_cast_fp16, y = var_2229_to_fp16)[name = string("op_2230_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2231_cast_fp16 = add(x = var_2226_cast_fp16, y = var_2230_cast_fp16)[name = string("op_2231_cast_fp16")];
            fp32 var_2232_epsilon_0 = const()[name = string("op_2232_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2232_cast_fp16 = rsqrt(epsilon = var_2232_epsilon_0, x = var_2231_cast_fp16)[name = string("op_2232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2233_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2232_cast_fp16)[name = string("op_2233_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_53_cast_fp16 = mul(x = var_2233_cast_fp16, y = model_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_53_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_2242 = const()[name = string("op_2242"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2243_cast_fp16 = reshape(shape = var_2242, x = linear_15_cast_fp16)[name = string("op_2243_cast_fp16")];
            tensor<int32, [4]> x_49_perm_0 = const()[name = string("x_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2250 = const()[name = string("op_2250"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_49_cast_fp16 = transpose(perm = x_49_perm_0, x = var_2243_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 8, 16, 128]> var_2253_cast_fp16 = abs(x = x_49_cast_fp16)[name = string("op_2253_cast_fp16")];
            tensor<int32, [1]> peak_21_axes_0 = const()[name = string("peak_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_21_cast_fp16 = reduce_max(axes = peak_21_axes_0, keep_dims = var_2250, x = var_2253_cast_fp16)[name = string("peak_21_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_51_cast_fp16 = real_div(x = x_49_cast_fp16, y = peak_21_cast_fp16)[name = string("x_51_cast_fp16")];
            fp16 var_2249_promoted_to_fp16 = const()[name = string("op_2249_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2257_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2249_promoted_to_fp16)[name = string("op_2257_cast_fp16")];
            tensor<int32, [1]> var_2259_axes_0 = const()[name = string("op_2259_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2259_keep_dims_0 = const()[name = string("op_2259_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2259_cast_fp16 = reduce_mean(axes = var_2259_axes_0, keep_dims = var_2259_keep_dims_0, x = var_2257_cast_fp16)[name = string("op_2259_cast_fp16")];
            fp16 var_2249_promoted_1_to_fp16 = const()[name = string("op_2249_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2260_cast_fp16 = pow(x = peak_21_cast_fp16, y = var_2249_promoted_1_to_fp16)[name = string("op_2260_cast_fp16")];
            fp32 var_2261_epsilon_0 = const()[name = string("op_2261_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2261_cast_fp16 = inverse(epsilon = var_2261_epsilon_0, x = var_2260_cast_fp16)[name = string("op_2261_cast_fp16")];
            fp16 var_2262_to_fp16 = const()[name = string("op_2262_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2263_cast_fp16 = mul(x = var_2261_cast_fp16, y = var_2262_to_fp16)[name = string("op_2263_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2264_cast_fp16 = add(x = var_2259_cast_fp16, y = var_2263_cast_fp16)[name = string("op_2264_cast_fp16")];
            fp32 var_2265_epsilon_0 = const()[name = string("op_2265_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2265_cast_fp16 = rsqrt(epsilon = var_2265_epsilon_0, x = var_2264_cast_fp16)[name = string("op_2265_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2266_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2265_cast_fp16)[name = string("op_2266_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_55_cast_fp16 = mul(x = var_2266_cast_fp16, y = model_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_55_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_2275 = const()[name = string("op_2275"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2276_cast_fp16 = reshape(shape = var_2275, x = linear_16_cast_fp16)[name = string("op_2276_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_53_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_53_cast_fp16)[name = string("second_9_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2290_cast_fp16 = mul(x = first_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2290_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2291_cast_fp16 = mul(x = second_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2293_cast_fp16 = sub(x = var_2290_cast_fp16, y = var_2291_cast_fp16)[name = string("op_2293_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2294_cast_fp16 = mul(x = second_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2294_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2295_cast_fp16 = mul(x = first_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2295_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2297_cast_fp16 = add(x = var_2294_cast_fp16, y = var_2295_cast_fp16)[name = string("op_2297_cast_fp16")];
            int32 var_2299 = const()[name = string("op_2299"), 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_2299, interleave = q_5_interleave_0, values = (var_2293_cast_fp16, var_2297_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_55_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_55_cast_fp16)[name = string("second_11_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2311_cast_fp16 = mul(x = first_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2311_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2312_cast_fp16 = mul(x = second_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2312_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2314_cast_fp16 = sub(x = var_2311_cast_fp16, y = var_2312_cast_fp16)[name = string("op_2314_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2315_cast_fp16 = mul(x = second_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2315_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2316_cast_fp16 = mul(x = first_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2316_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2318_cast_fp16 = add(x = var_2315_cast_fp16, y = var_2316_cast_fp16)[name = string("op_2318_cast_fp16")];
            int32 var_2320 = const()[name = string("op_2320"), 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_2320, interleave = k_5_interleave_0, values = (var_2314_cast_fp16, var_2318_cast_fp16))[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_2331_begin_0 = const()[name = string("op_2331_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2331_end_0 = const()[name = string("op_2331_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2331_end_mask_0 = const()[name = string("op_2331_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2331_stride_0 = const()[name = string("op_2331_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, stride = var_2331_stride_0, x = q_5_cast_fp16)[name = string("op_2331_cast_fp16")];
            tensor<int32, [4]> var_2341_begin_0 = const()[name = string("op_2341_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2341_end_0 = const()[name = string("op_2341_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2341_end_mask_0 = const()[name = string("op_2341_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2341_stride_0 = const()[name = string("op_2341_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2341_cast_fp16 = slice_by_index(begin = var_2341_begin_0, end = var_2341_end_0, end_mask = var_2341_end_mask_0, stride = var_2341_stride_0, x = q_5_cast_fp16)[name = string("op_2341_cast_fp16")];
            int32 var_2343 = const()[name = string("op_2343"), 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_2343, interleave = stacked_5_interleave_0, values = (var_2331_cast_fp16, var_2341_cast_fp16))[name = string("stacked_5_cast_fp16")];
            tensor<int32, [4]> var_2349_begin_0 = const()[name = string("op_2349_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2349_end_0 = const()[name = string("op_2349_end_0"), val = tensor<int32, [4]>([3, 8, 528, 128])];
            tensor<bool, [4]> var_2349_end_mask_0 = const()[name = string("op_2349_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2349_cast_fp16 = slice_by_index(begin = var_2349_begin_0, end = var_2349_end_0, end_mask = var_2349_end_mask_0, x = read_state_0)[name = string("op_2349_cast_fp16")];
            tensor<int32, [4]> var_2359_begin_0 = const()[name = string("op_2359_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2359_end_0 = const()[name = string("op_2359_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2359_end_mask_0 = const()[name = string("op_2359_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2359_cast_fp16 = slice_by_index(begin = var_2359_begin_0, end = var_2359_end_0, end_mask = var_2359_end_mask_0, x = var_2349_cast_fp16)[name = string("op_2359_cast_fp16")];
            bool var_2368_transpose_x_1 = const()[name = string("op_2368_transpose_x_1"), val = bool(false)];
            bool var_2368_transpose_y_1 = const()[name = string("op_2368_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_2368_cast_fp16 = matmul(transpose_x = var_2368_transpose_x_1, transpose_y = var_2368_transpose_y_1, x = var_2359_cast_fp16, y = stacked_5_cast_fp16)[name = string("op_2368_cast_fp16")];
            tensor<int32, [4]> var_2371_perm_0 = const()[name = string("op_2371_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2372_to_fp16 = const()[name = string("op_2372_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_2371_cast_fp16 = transpose(perm = var_2371_perm_0, x = var_2368_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 8, 32, 512]> var_2373_cast_fp16 = mul(x = var_2371_cast_fp16, y = var_2372_to_fp16)[name = string("op_2373_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_5_cast_fp16 = add(x = var_2373_cast_fp16, y = cacheMask)[name = string("cache_weights_5_cast_fp16")];
            bool var_2379_transpose_x_1 = const()[name = string("op_2379_transpose_x_1"), val = bool(false)];
            bool var_2379_transpose_y_1 = const()[name = string("op_2379_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2379_cast_fp16 = matmul(transpose_x = var_2379_transpose_x_1, transpose_y = var_2379_transpose_y_1, x = stacked_5_cast_fp16, y = k_5_cast_fp16)[name = string("op_2379_cast_fp16")];
            fp16 var_2380_to_fp16 = const()[name = string("op_2380_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2381_cast_fp16 = mul(x = var_2379_cast_fp16, y = var_2380_to_fp16)[name = string("op_2381_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_5_cast_fp16 = add(x = var_2381_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_5_cast_fp16")];
            int32 var_2385 = const()[name = string("op_2385"), val = int32(-1)];
            bool var_2386_interleave_0 = const()[name = string("op_2386_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_2386_cast_fp16 = concat(axis = var_2385, interleave = var_2386_interleave_0, values = (cache_weights_5_cast_fp16, call_weights_5_cast_fp16))[name = string("op_2386_cast_fp16")];
            int32 var_2387 = const()[name = string("op_2387"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_5_cast_fp16 = softmax(axis = var_2387, x = var_2386_cast_fp16)[name = string("shares_5_cast_fp16")];
            tensor<int32, [4]> var_2394_begin_0 = const()[name = string("op_2394_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2394_end_0 = const()[name = string("op_2394_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_2394_end_mask_0 = const()[name = string("op_2394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_2394_cast_fp16 = slice_by_index(begin = var_2394_begin_0, end = var_2394_end_0, end_mask = var_2394_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2394_cast_fp16")];
            tensor<int32, [4]> var_2399_begin_0 = const()[name = string("op_2399_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2399_end_0 = const()[name = string("op_2399_end_0"), val = tensor<int32, [4]>([3, 8, 528, 128])];
            tensor<bool, [4]> var_2399_end_mask_0 = const()[name = string("op_2399_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2399_cast_fp16 = slice_by_index(begin = var_2399_begin_0, end = var_2399_end_0, end_mask = var_2399_end_mask_0, x = read_state_1)[name = string("op_2399_cast_fp16")];
            tensor<int32, [4]> var_2409_begin_0 = const()[name = string("op_2409_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2409_end_0 = const()[name = string("op_2409_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2409_end_mask_0 = const()[name = string("op_2409_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2409_cast_fp16 = slice_by_index(begin = var_2409_begin_0, end = var_2409_end_0, end_mask = var_2409_end_mask_0, x = var_2399_cast_fp16)[name = string("op_2409_cast_fp16")];
            bool var_2415_transpose_x_0 = const()[name = string("op_2415_transpose_x_0"), val = bool(false)];
            bool var_2415_transpose_y_0 = const()[name = string("op_2415_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2415_cast_fp16 = matmul(transpose_x = var_2415_transpose_x_0, transpose_y = var_2415_transpose_y_0, x = var_2394_cast_fp16, y = var_2409_cast_fp16)[name = string("op_2415_cast_fp16")];
            tensor<int32, [4]> var_2420_begin_0 = const()[name = string("op_2420_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_2420_end_0 = const()[name = string("op_2420_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_2420_end_mask_0 = const()[name = string("op_2420_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2420_cast_fp16 = slice_by_index(begin = var_2420_begin_0, end = var_2420_end_0, end_mask = var_2420_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2420_cast_fp16")];
            bool var_2421_transpose_x_0 = const()[name = string("op_2421_transpose_x_0"), val = bool(false)];
            bool var_2421_transpose_y_0 = const()[name = string("op_2421_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_2276_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 8, 32, 128]> var_2421_cast_fp16 = matmul(transpose_x = var_2421_transpose_x_0, transpose_y = var_2421_transpose_y_0, x = var_2420_cast_fp16, y = v_5_cast_fp16)[name = string("op_2421_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_9_cast_fp16 = add(x = var_2415_cast_fp16, y = var_2421_cast_fp16)[name = string("mixed_9_cast_fp16")];
            tensor<int32, [4]> var_2438_begin_0 = const()[name = string("op_2438_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2438_end_0 = const()[name = string("op_2438_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2438_end_mask_0 = const()[name = string("op_2438_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2438_cast_fp16 = slice_by_index(begin = var_2438_begin_0, end = var_2438_end_0, end_mask = var_2438_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2438_cast_fp16")];
            tensor<int32, [4]> var_2453_begin_0 = const()[name = string("op_2453_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2453_end_0 = const()[name = string("op_2453_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2453_end_mask_0 = const()[name = string("op_2453_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2453_cast_fp16 = slice_by_index(begin = var_2453_begin_0, end = var_2453_end_0, end_mask = var_2453_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2453_cast_fp16")];
            int32 var_2455 = const()[name = string("op_2455"), 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_2455, interleave = mixed_11_interleave_0, values = (var_2438_cast_fp16, var_2453_cast_fp16))[name = string("mixed_11_cast_fp16")];
            tensor<int32, [4]> var_2459_perm_0 = const()[name = string("op_2459_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2463 = const()[name = string("op_2463"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2459_cast_fp16 = transpose(perm = var_2459_perm_0, x = mixed_11_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 16, 2048]> input_23_cast_fp16 = reshape(shape = var_2463, x = var_2459_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_57_cast_fp16 = add(x = x_41_cast_fp16, y = linear_17_cast_fp16)[name = string("x_57_cast_fp16")];
            bool var_2473 = const()[name = string("op_2473"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2476_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<int32, [1]> peak_23_axes_0 = const()[name = string("peak_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_23_cast_fp16 = reduce_max(axes = peak_23_axes_0, keep_dims = var_2473, x = var_2476_cast_fp16)[name = string("peak_23_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_59_cast_fp16 = real_div(x = x_57_cast_fp16, y = peak_23_cast_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2472_promoted_to_fp16 = const()[name = string("op_2472_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2480_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2472_promoted_to_fp16)[name = string("op_2480_cast_fp16")];
            tensor<int32, [1]> var_2482_axes_0 = const()[name = string("op_2482_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2482_keep_dims_0 = const()[name = string("op_2482_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2482_cast_fp16 = reduce_mean(axes = var_2482_axes_0, keep_dims = var_2482_keep_dims_0, x = var_2480_cast_fp16)[name = string("op_2482_cast_fp16")];
            fp16 var_2472_promoted_1_to_fp16 = const()[name = string("op_2472_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2483_cast_fp16 = pow(x = peak_23_cast_fp16, y = var_2472_promoted_1_to_fp16)[name = string("op_2483_cast_fp16")];
            fp32 var_2484_epsilon_0 = const()[name = string("op_2484_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2484_cast_fp16 = inverse(epsilon = var_2484_epsilon_0, x = var_2483_cast_fp16)[name = string("op_2484_cast_fp16")];
            fp16 var_2485_to_fp16 = const()[name = string("op_2485_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2486_cast_fp16 = mul(x = var_2484_cast_fp16, y = var_2485_to_fp16)[name = string("op_2486_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2487_cast_fp16 = add(x = var_2482_cast_fp16, y = var_2486_cast_fp16)[name = string("op_2487_cast_fp16")];
            fp32 var_2488_epsilon_0 = const()[name = string("op_2488_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2488_cast_fp16 = rsqrt(epsilon = var_2488_epsilon_0, x = var_2487_cast_fp16)[name = string("op_2488_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2489_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2488_cast_fp16)[name = string("op_2489_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_2489_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_2494_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_2494_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_2494_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_61_cast_fp16 = add(x = x_57_cast_fp16, y = linear_20_cast_fp16)[name = string("x_61_cast_fp16")];
            bool var_2507 = const()[name = string("op_2507"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2510_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2510_cast_fp16")];
            tensor<int32, [1]> peak_25_axes_0 = const()[name = string("peak_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_25_cast_fp16 = reduce_max(axes = peak_25_axes_0, keep_dims = var_2507, x = var_2510_cast_fp16)[name = string("peak_25_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_63_cast_fp16 = real_div(x = x_61_cast_fp16, y = peak_25_cast_fp16)[name = string("x_63_cast_fp16")];
            fp16 var_2506_promoted_to_fp16 = const()[name = string("op_2506_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2514_cast_fp16 = pow(x = x_63_cast_fp16, y = var_2506_promoted_to_fp16)[name = string("op_2514_cast_fp16")];
            tensor<int32, [1]> var_2516_axes_0 = const()[name = string("op_2516_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2516_keep_dims_0 = const()[name = string("op_2516_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2516_cast_fp16 = reduce_mean(axes = var_2516_axes_0, keep_dims = var_2516_keep_dims_0, x = var_2514_cast_fp16)[name = string("op_2516_cast_fp16")];
            fp16 var_2506_promoted_1_to_fp16 = const()[name = string("op_2506_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2517_cast_fp16 = pow(x = peak_25_cast_fp16, y = var_2506_promoted_1_to_fp16)[name = string("op_2517_cast_fp16")];
            fp32 var_2518_epsilon_0 = const()[name = string("op_2518_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2518_cast_fp16 = inverse(epsilon = var_2518_epsilon_0, x = var_2517_cast_fp16)[name = string("op_2518_cast_fp16")];
            fp16 var_2519_to_fp16 = const()[name = string("op_2519_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2520_cast_fp16 = mul(x = var_2518_cast_fp16, y = var_2519_to_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2521_cast_fp16 = add(x = var_2516_cast_fp16, y = var_2520_cast_fp16)[name = string("op_2521_cast_fp16")];
            fp32 var_2522_epsilon_0 = const()[name = string("op_2522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2522_cast_fp16 = rsqrt(epsilon = var_2522_epsilon_0, x = var_2521_cast_fp16)[name = string("op_2522_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2523_cast_fp16 = mul(x = x_63_cast_fp16, y = var_2522_cast_fp16)[name = string("op_2523_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_2523_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_2532 = const()[name = string("op_2532"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2533_cast_fp16 = reshape(shape = var_2532, x = linear_21_cast_fp16)[name = string("op_2533_cast_fp16")];
            tensor<int32, [4]> x_65_perm_0 = const()[name = string("x_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2540 = const()[name = string("op_2540"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_65_cast_fp16 = transpose(perm = x_65_perm_0, x = var_2533_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 16, 128]> var_2543_cast_fp16 = abs(x = x_65_cast_fp16)[name = string("op_2543_cast_fp16")];
            tensor<int32, [1]> peak_27_axes_0 = const()[name = string("peak_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_27_cast_fp16 = reduce_max(axes = peak_27_axes_0, keep_dims = var_2540, x = var_2543_cast_fp16)[name = string("peak_27_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_67_cast_fp16 = real_div(x = x_65_cast_fp16, y = peak_27_cast_fp16)[name = string("x_67_cast_fp16")];
            fp16 var_2539_promoted_to_fp16 = const()[name = string("op_2539_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2547_cast_fp16 = pow(x = x_67_cast_fp16, y = var_2539_promoted_to_fp16)[name = string("op_2547_cast_fp16")];
            tensor<int32, [1]> var_2549_axes_0 = const()[name = string("op_2549_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2549_keep_dims_0 = const()[name = string("op_2549_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2549_cast_fp16 = reduce_mean(axes = var_2549_axes_0, keep_dims = var_2549_keep_dims_0, x = var_2547_cast_fp16)[name = string("op_2549_cast_fp16")];
            fp16 var_2539_promoted_1_to_fp16 = const()[name = string("op_2539_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2550_cast_fp16 = pow(x = peak_27_cast_fp16, y = var_2539_promoted_1_to_fp16)[name = string("op_2550_cast_fp16")];
            fp32 var_2551_epsilon_0 = const()[name = string("op_2551_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2551_cast_fp16 = inverse(epsilon = var_2551_epsilon_0, x = var_2550_cast_fp16)[name = string("op_2551_cast_fp16")];
            fp16 var_2552_to_fp16 = const()[name = string("op_2552_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2553_cast_fp16 = mul(x = var_2551_cast_fp16, y = var_2552_to_fp16)[name = string("op_2553_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2554_cast_fp16 = add(x = var_2549_cast_fp16, y = var_2553_cast_fp16)[name = string("op_2554_cast_fp16")];
            fp32 var_2555_epsilon_0 = const()[name = string("op_2555_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2555_cast_fp16 = rsqrt(epsilon = var_2555_epsilon_0, x = var_2554_cast_fp16)[name = string("op_2555_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2556_cast_fp16 = mul(x = x_67_cast_fp16, y = var_2555_cast_fp16)[name = string("op_2556_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_73_cast_fp16 = mul(x = var_2556_cast_fp16, y = model_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_73_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_2565 = const()[name = string("op_2565"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2566_cast_fp16 = reshape(shape = var_2565, x = linear_22_cast_fp16)[name = string("op_2566_cast_fp16")];
            tensor<int32, [4]> x_69_perm_0 = const()[name = string("x_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2573 = const()[name = string("op_2573"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_69_cast_fp16 = transpose(perm = x_69_perm_0, x = var_2566_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 16, 128]> var_2576_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2576_cast_fp16")];
            tensor<int32, [1]> peak_29_axes_0 = const()[name = string("peak_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_29_cast_fp16 = reduce_max(axes = peak_29_axes_0, keep_dims = var_2573, x = var_2576_cast_fp16)[name = string("peak_29_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_71_cast_fp16 = real_div(x = x_69_cast_fp16, y = peak_29_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_2572_promoted_to_fp16 = const()[name = string("op_2572_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2580_cast_fp16 = pow(x = x_71_cast_fp16, y = var_2572_promoted_to_fp16)[name = string("op_2580_cast_fp16")];
            tensor<int32, [1]> var_2582_axes_0 = const()[name = string("op_2582_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2582_keep_dims_0 = const()[name = string("op_2582_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2582_cast_fp16 = reduce_mean(axes = var_2582_axes_0, keep_dims = var_2582_keep_dims_0, x = var_2580_cast_fp16)[name = string("op_2582_cast_fp16")];
            fp16 var_2572_promoted_1_to_fp16 = const()[name = string("op_2572_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2583_cast_fp16 = pow(x = peak_29_cast_fp16, y = var_2572_promoted_1_to_fp16)[name = string("op_2583_cast_fp16")];
            fp32 var_2584_epsilon_0 = const()[name = string("op_2584_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2584_cast_fp16 = inverse(epsilon = var_2584_epsilon_0, x = var_2583_cast_fp16)[name = string("op_2584_cast_fp16")];
            fp16 var_2585_to_fp16 = const()[name = string("op_2585_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2586_cast_fp16 = mul(x = var_2584_cast_fp16, y = var_2585_to_fp16)[name = string("op_2586_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2587_cast_fp16 = add(x = var_2582_cast_fp16, y = var_2586_cast_fp16)[name = string("op_2587_cast_fp16")];
            fp32 var_2588_epsilon_0 = const()[name = string("op_2588_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2588_cast_fp16 = rsqrt(epsilon = var_2588_epsilon_0, x = var_2587_cast_fp16)[name = string("op_2588_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2589_cast_fp16 = mul(x = x_71_cast_fp16, y = var_2588_cast_fp16)[name = string("op_2589_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_75_cast_fp16 = mul(x = var_2589_cast_fp16, y = model_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_75_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_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2599_cast_fp16 = reshape(shape = var_2598, x = linear_23_cast_fp16)[name = string("op_2599_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_73_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_73_cast_fp16)[name = string("second_13_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2613_cast_fp16 = mul(x = first_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2613_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2614_cast_fp16 = mul(x = second_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2614_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2616_cast_fp16 = sub(x = var_2613_cast_fp16, y = var_2614_cast_fp16)[name = string("op_2616_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2617_cast_fp16 = mul(x = second_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2617_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2618_cast_fp16 = mul(x = first_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2618_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2620_cast_fp16 = add(x = var_2617_cast_fp16, y = var_2618_cast_fp16)[name = string("op_2620_cast_fp16")];
            int32 var_2622 = const()[name = string("op_2622"), 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_2622, interleave = q_7_interleave_0, values = (var_2616_cast_fp16, var_2620_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_75_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_75_cast_fp16)[name = string("second_15_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2634_cast_fp16 = mul(x = first_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2634_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2635_cast_fp16 = mul(x = second_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2635_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2637_cast_fp16 = sub(x = var_2634_cast_fp16, y = var_2635_cast_fp16)[name = string("op_2637_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2638_cast_fp16 = mul(x = second_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2638_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2639_cast_fp16 = mul(x = first_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2641_cast_fp16 = add(x = var_2638_cast_fp16, y = var_2639_cast_fp16)[name = string("op_2641_cast_fp16")];
            int32 var_2643 = const()[name = string("op_2643"), 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_2643, interleave = k_7_interleave_0, values = (var_2637_cast_fp16, var_2641_cast_fp16))[name = string("k_7_cast_fp16")];
            tensor<int32, [4]> var_2654_begin_0 = const()[name = string("op_2654_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2654_end_0 = const()[name = string("op_2654_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2654_end_mask_0 = const()[name = string("op_2654_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2654_stride_0 = const()[name = string("op_2654_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2654_cast_fp16 = slice_by_index(begin = var_2654_begin_0, end = var_2654_end_0, end_mask = var_2654_end_mask_0, stride = var_2654_stride_0, x = q_7_cast_fp16)[name = string("op_2654_cast_fp16")];
            tensor<int32, [4]> var_2664_begin_0 = const()[name = string("op_2664_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2664_end_0 = const()[name = string("op_2664_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2664_end_mask_0 = const()[name = string("op_2664_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2664_stride_0 = const()[name = string("op_2664_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2664_cast_fp16 = slice_by_index(begin = var_2664_begin_0, end = var_2664_end_0, end_mask = var_2664_end_mask_0, stride = var_2664_stride_0, x = q_7_cast_fp16)[name = string("op_2664_cast_fp16")];
            int32 var_2666 = const()[name = string("op_2666"), 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_2666, interleave = stacked_7_interleave_0, values = (var_2654_cast_fp16, var_2664_cast_fp16))[name = string("stacked_7_cast_fp16")];
            tensor<int32, [4]> var_2672_begin_0 = const()[name = string("op_2672_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2672_end_0 = const()[name = string("op_2672_end_0"), val = tensor<int32, [4]>([4, 8, 528, 128])];
            tensor<bool, [4]> var_2672_end_mask_0 = const()[name = string("op_2672_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2672_cast_fp16 = slice_by_index(begin = var_2672_begin_0, end = var_2672_end_0, end_mask = var_2672_end_mask_0, x = read_state_0)[name = string("op_2672_cast_fp16")];
            tensor<int32, [4]> var_2682_begin_0 = const()[name = string("op_2682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2682_end_0 = const()[name = string("op_2682_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2682_end_mask_0 = const()[name = string("op_2682_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2682_cast_fp16 = slice_by_index(begin = var_2682_begin_0, end = var_2682_end_0, end_mask = var_2682_end_mask_0, x = var_2672_cast_fp16)[name = string("op_2682_cast_fp16")];
            bool var_2691_transpose_x_1 = const()[name = string("op_2691_transpose_x_1"), val = bool(false)];
            bool var_2691_transpose_y_1 = const()[name = string("op_2691_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_2691_cast_fp16 = matmul(transpose_x = var_2691_transpose_x_1, transpose_y = var_2691_transpose_y_1, x = var_2682_cast_fp16, y = stacked_7_cast_fp16)[name = string("op_2691_cast_fp16")];
            tensor<int32, [4]> var_2694_perm_0 = const()[name = string("op_2694_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2695_to_fp16 = const()[name = string("op_2695_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_2694_cast_fp16 = transpose(perm = var_2694_perm_0, x = var_2691_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 32, 512]> var_2696_cast_fp16 = mul(x = var_2694_cast_fp16, y = var_2695_to_fp16)[name = string("op_2696_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_7_cast_fp16 = add(x = var_2696_cast_fp16, y = cacheMask)[name = string("cache_weights_7_cast_fp16")];
            bool var_2702_transpose_x_1 = const()[name = string("op_2702_transpose_x_1"), val = bool(false)];
            bool var_2702_transpose_y_1 = const()[name = string("op_2702_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2702_cast_fp16 = matmul(transpose_x = var_2702_transpose_x_1, transpose_y = var_2702_transpose_y_1, x = stacked_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 var_2703_to_fp16 = const()[name = string("op_2703_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2704_cast_fp16 = mul(x = var_2702_cast_fp16, y = var_2703_to_fp16)[name = string("op_2704_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_7_cast_fp16 = add(x = var_2704_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_7_cast_fp16")];
            int32 var_2708 = const()[name = string("op_2708"), val = int32(-1)];
            bool var_2709_interleave_0 = const()[name = string("op_2709_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_2709_cast_fp16 = concat(axis = var_2708, interleave = var_2709_interleave_0, values = (cache_weights_7_cast_fp16, call_weights_7_cast_fp16))[name = string("op_2709_cast_fp16")];
            int32 var_2710 = const()[name = string("op_2710"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_7_cast_fp16 = softmax(axis = var_2710, x = var_2709_cast_fp16)[name = string("shares_7_cast_fp16")];
            tensor<int32, [4]> var_2717_begin_0 = const()[name = string("op_2717_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2717_end_0 = const()[name = string("op_2717_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_2717_end_mask_0 = const()[name = string("op_2717_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_2717_cast_fp16 = slice_by_index(begin = var_2717_begin_0, end = var_2717_end_0, end_mask = var_2717_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2717_cast_fp16")];
            tensor<int32, [4]> var_2722_begin_0 = const()[name = string("op_2722_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2722_end_0 = const()[name = string("op_2722_end_0"), val = tensor<int32, [4]>([4, 8, 528, 128])];
            tensor<bool, [4]> var_2722_end_mask_0 = const()[name = string("op_2722_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2722_cast_fp16 = slice_by_index(begin = var_2722_begin_0, end = var_2722_end_0, end_mask = var_2722_end_mask_0, x = read_state_1)[name = string("op_2722_cast_fp16")];
            tensor<int32, [4]> var_2732_begin_0 = const()[name = string("op_2732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2732_end_0 = const()[name = string("op_2732_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2732_end_mask_0 = const()[name = string("op_2732_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2732_cast_fp16 = slice_by_index(begin = var_2732_begin_0, end = var_2732_end_0, end_mask = var_2732_end_mask_0, x = var_2722_cast_fp16)[name = string("op_2732_cast_fp16")];
            bool var_2738_transpose_x_0 = const()[name = string("op_2738_transpose_x_0"), val = bool(false)];
            bool var_2738_transpose_y_0 = const()[name = string("op_2738_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2738_cast_fp16 = matmul(transpose_x = var_2738_transpose_x_0, transpose_y = var_2738_transpose_y_0, x = var_2717_cast_fp16, y = var_2732_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<int32, [4]> var_2743_begin_0 = const()[name = string("op_2743_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_2743_end_0 = const()[name = string("op_2743_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_2743_end_mask_0 = const()[name = string("op_2743_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2743_cast_fp16 = slice_by_index(begin = var_2743_begin_0, end = var_2743_end_0, end_mask = var_2743_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2743_cast_fp16")];
            bool var_2744_transpose_x_0 = const()[name = string("op_2744_transpose_x_0"), val = bool(false)];
            bool var_2744_transpose_y_0 = const()[name = string("op_2744_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_2599_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 8, 32, 128]> var_2744_cast_fp16 = matmul(transpose_x = var_2744_transpose_x_0, transpose_y = var_2744_transpose_y_0, x = var_2743_cast_fp16, y = v_7_cast_fp16)[name = string("op_2744_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_13_cast_fp16 = add(x = var_2738_cast_fp16, y = var_2744_cast_fp16)[name = string("mixed_13_cast_fp16")];
            tensor<int32, [4]> var_2761_begin_0 = const()[name = string("op_2761_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2761_end_0 = const()[name = string("op_2761_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2761_end_mask_0 = const()[name = string("op_2761_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2761_cast_fp16 = slice_by_index(begin = var_2761_begin_0, end = var_2761_end_0, end_mask = var_2761_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2761_cast_fp16")];
            tensor<int32, [4]> var_2776_begin_0 = const()[name = string("op_2776_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2776_end_0 = const()[name = string("op_2776_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2776_end_mask_0 = const()[name = string("op_2776_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2776_cast_fp16 = slice_by_index(begin = var_2776_begin_0, end = var_2776_end_0, end_mask = var_2776_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2776_cast_fp16")];
            int32 var_2778 = const()[name = string("op_2778"), 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_2778, interleave = mixed_15_interleave_0, values = (var_2761_cast_fp16, var_2776_cast_fp16))[name = string("mixed_15_cast_fp16")];
            tensor<int32, [4]> var_2782_perm_0 = const()[name = string("op_2782_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2786 = const()[name = string("op_2786"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2782_cast_fp16 = transpose(perm = var_2782_perm_0, x = mixed_15_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 16, 2048]> input_33_cast_fp16 = reshape(shape = var_2786, x = var_2782_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_77_cast_fp16 = add(x = x_61_cast_fp16, y = linear_24_cast_fp16)[name = string("x_77_cast_fp16")];
            bool var_2796 = const()[name = string("op_2796"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2799_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_2799_cast_fp16")];
            tensor<int32, [1]> peak_31_axes_0 = const()[name = string("peak_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_31_cast_fp16 = reduce_max(axes = peak_31_axes_0, keep_dims = var_2796, x = var_2799_cast_fp16)[name = string("peak_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_79_cast_fp16 = real_div(x = x_77_cast_fp16, y = peak_31_cast_fp16)[name = string("x_79_cast_fp16")];
            fp16 var_2795_promoted_to_fp16 = const()[name = string("op_2795_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2803_cast_fp16 = pow(x = x_79_cast_fp16, y = var_2795_promoted_to_fp16)[name = string("op_2803_cast_fp16")];
            tensor<int32, [1]> var_2805_axes_0 = const()[name = string("op_2805_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2805_keep_dims_0 = const()[name = string("op_2805_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2805_cast_fp16 = reduce_mean(axes = var_2805_axes_0, keep_dims = var_2805_keep_dims_0, x = var_2803_cast_fp16)[name = string("op_2805_cast_fp16")];
            fp16 var_2795_promoted_1_to_fp16 = const()[name = string("op_2795_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2806_cast_fp16 = pow(x = peak_31_cast_fp16, y = var_2795_promoted_1_to_fp16)[name = string("op_2806_cast_fp16")];
            fp32 var_2807_epsilon_0 = const()[name = string("op_2807_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2807_cast_fp16 = inverse(epsilon = var_2807_epsilon_0, x = var_2806_cast_fp16)[name = string("op_2807_cast_fp16")];
            fp16 var_2808_to_fp16 = const()[name = string("op_2808_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2809_cast_fp16 = mul(x = var_2807_cast_fp16, y = var_2808_to_fp16)[name = string("op_2809_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2810_cast_fp16 = add(x = var_2805_cast_fp16, y = var_2809_cast_fp16)[name = string("op_2810_cast_fp16")];
            fp32 var_2811_epsilon_0 = const()[name = string("op_2811_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2811_cast_fp16 = rsqrt(epsilon = var_2811_epsilon_0, x = var_2810_cast_fp16)[name = string("op_2811_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2812_cast_fp16 = mul(x = x_79_cast_fp16, y = var_2811_cast_fp16)[name = string("op_2812_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_2812_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_2817_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_2817_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_2817_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_81_cast_fp16 = add(x = x_77_cast_fp16, y = linear_27_cast_fp16)[name = string("x_81_cast_fp16")];
            bool var_2830 = const()[name = string("op_2830"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_2833_cast_fp16 = abs(x = x_81_cast_fp16)[name = string("op_2833_cast_fp16")];
            tensor<int32, [1]> peak_33_axes_0 = const()[name = string("peak_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_33_cast_fp16 = reduce_max(axes = peak_33_axes_0, keep_dims = var_2830, x = var_2833_cast_fp16)[name = string("peak_33_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_83_cast_fp16 = real_div(x = x_81_cast_fp16, y = peak_33_cast_fp16)[name = string("x_83_cast_fp16")];
            fp16 var_2829_promoted_to_fp16 = const()[name = string("op_2829_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2837_cast_fp16 = pow(x = x_83_cast_fp16, y = var_2829_promoted_to_fp16)[name = string("op_2837_cast_fp16")];
            tensor<int32, [1]> var_2839_axes_0 = const()[name = string("op_2839_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2839_keep_dims_0 = const()[name = string("op_2839_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2839_cast_fp16 = reduce_mean(axes = var_2839_axes_0, keep_dims = var_2839_keep_dims_0, x = var_2837_cast_fp16)[name = string("op_2839_cast_fp16")];
            fp16 var_2829_promoted_1_to_fp16 = const()[name = string("op_2829_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_2840_cast_fp16 = pow(x = peak_33_cast_fp16, y = var_2829_promoted_1_to_fp16)[name = string("op_2840_cast_fp16")];
            fp32 var_2841_epsilon_0 = const()[name = string("op_2841_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_2841_cast_fp16 = inverse(epsilon = var_2841_epsilon_0, x = var_2840_cast_fp16)[name = string("op_2841_cast_fp16")];
            fp16 var_2842_to_fp16 = const()[name = string("op_2842_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2843_cast_fp16 = mul(x = var_2841_cast_fp16, y = var_2842_to_fp16)[name = string("op_2843_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_2844_cast_fp16 = add(x = var_2839_cast_fp16, y = var_2843_cast_fp16)[name = string("op_2844_cast_fp16")];
            fp32 var_2845_epsilon_0 = const()[name = string("op_2845_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2845_cast_fp16 = rsqrt(epsilon = var_2845_epsilon_0, x = var_2844_cast_fp16)[name = string("op_2845_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2846_cast_fp16 = mul(x = x_83_cast_fp16, y = var_2845_cast_fp16)[name = string("op_2846_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_2846_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_2855 = const()[name = string("op_2855"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2856_cast_fp16 = reshape(shape = var_2855, x = linear_28_cast_fp16)[name = string("op_2856_cast_fp16")];
            tensor<int32, [4]> x_85_perm_0 = const()[name = string("x_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2863 = const()[name = string("op_2863"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_85_cast_fp16 = transpose(perm = x_85_perm_0, x = var_2856_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 16, 16, 128]> var_2866_cast_fp16 = abs(x = x_85_cast_fp16)[name = string("op_2866_cast_fp16")];
            tensor<int32, [1]> peak_35_axes_0 = const()[name = string("peak_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_35_cast_fp16 = reduce_max(axes = peak_35_axes_0, keep_dims = var_2863, x = var_2866_cast_fp16)[name = string("peak_35_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_87_cast_fp16 = real_div(x = x_85_cast_fp16, y = peak_35_cast_fp16)[name = string("x_87_cast_fp16")];
            fp16 var_2862_promoted_to_fp16 = const()[name = string("op_2862_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_2870_cast_fp16 = pow(x = x_87_cast_fp16, y = var_2862_promoted_to_fp16)[name = string("op_2870_cast_fp16")];
            tensor<int32, [1]> var_2872_axes_0 = const()[name = string("op_2872_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2872_keep_dims_0 = const()[name = string("op_2872_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2872_cast_fp16 = reduce_mean(axes = var_2872_axes_0, keep_dims = var_2872_keep_dims_0, x = var_2870_cast_fp16)[name = string("op_2872_cast_fp16")];
            fp16 var_2862_promoted_1_to_fp16 = const()[name = string("op_2862_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_2873_cast_fp16 = pow(x = peak_35_cast_fp16, y = var_2862_promoted_1_to_fp16)[name = string("op_2873_cast_fp16")];
            fp32 var_2874_epsilon_0 = const()[name = string("op_2874_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_2874_cast_fp16 = inverse(epsilon = var_2874_epsilon_0, x = var_2873_cast_fp16)[name = string("op_2874_cast_fp16")];
            fp16 var_2875_to_fp16 = const()[name = string("op_2875_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2876_cast_fp16 = mul(x = var_2874_cast_fp16, y = var_2875_to_fp16)[name = string("op_2876_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_2877_cast_fp16 = add(x = var_2872_cast_fp16, y = var_2876_cast_fp16)[name = string("op_2877_cast_fp16")];
            fp32 var_2878_epsilon_0 = const()[name = string("op_2878_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2878_cast_fp16 = rsqrt(epsilon = var_2878_epsilon_0, x = var_2877_cast_fp16)[name = string("op_2878_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2879_cast_fp16 = mul(x = x_87_cast_fp16, y = var_2878_cast_fp16)[name = string("op_2879_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_93_cast_fp16 = mul(x = var_2879_cast_fp16, y = model_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_93_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_2888 = const()[name = string("op_2888"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2889_cast_fp16 = reshape(shape = var_2888, x = linear_29_cast_fp16)[name = string("op_2889_cast_fp16")];
            tensor<int32, [4]> x_89_perm_0 = const()[name = string("x_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_2896 = const()[name = string("op_2896"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = var_2889_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 8, 16, 128]> var_2899_cast_fp16 = abs(x = x_89_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<int32, [1]> peak_37_axes_0 = const()[name = string("peak_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_37_cast_fp16 = reduce_max(axes = peak_37_axes_0, keep_dims = var_2896, x = var_2899_cast_fp16)[name = string("peak_37_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_91_cast_fp16 = real_div(x = x_89_cast_fp16, y = peak_37_cast_fp16)[name = string("x_91_cast_fp16")];
            fp16 var_2895_promoted_to_fp16 = const()[name = string("op_2895_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_2903_cast_fp16 = pow(x = x_91_cast_fp16, y = var_2895_promoted_to_fp16)[name = string("op_2903_cast_fp16")];
            tensor<int32, [1]> var_2905_axes_0 = const()[name = string("op_2905_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2905_keep_dims_0 = const()[name = string("op_2905_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2905_cast_fp16 = reduce_mean(axes = var_2905_axes_0, keep_dims = var_2905_keep_dims_0, x = var_2903_cast_fp16)[name = string("op_2905_cast_fp16")];
            fp16 var_2895_promoted_1_to_fp16 = const()[name = string("op_2895_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_2906_cast_fp16 = pow(x = peak_37_cast_fp16, y = var_2895_promoted_1_to_fp16)[name = string("op_2906_cast_fp16")];
            fp32 var_2907_epsilon_0 = const()[name = string("op_2907_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_2907_cast_fp16 = inverse(epsilon = var_2907_epsilon_0, x = var_2906_cast_fp16)[name = string("op_2907_cast_fp16")];
            fp16 var_2908_to_fp16 = const()[name = string("op_2908_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2909_cast_fp16 = mul(x = var_2907_cast_fp16, y = var_2908_to_fp16)[name = string("op_2909_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_2910_cast_fp16 = add(x = var_2905_cast_fp16, y = var_2909_cast_fp16)[name = string("op_2910_cast_fp16")];
            fp32 var_2911_epsilon_0 = const()[name = string("op_2911_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2911_cast_fp16 = rsqrt(epsilon = var_2911_epsilon_0, x = var_2910_cast_fp16)[name = string("op_2911_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2912_cast_fp16 = mul(x = x_91_cast_fp16, y = var_2911_cast_fp16)[name = string("op_2912_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_95_cast_fp16 = mul(x = var_2912_cast_fp16, y = model_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_95_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_2921 = const()[name = string("op_2921"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2922_cast_fp16 = reshape(shape = var_2921, x = linear_30_cast_fp16)[name = string("op_2922_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_93_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_93_cast_fp16)[name = string("second_17_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2936_cast_fp16 = mul(x = first_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2936_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2937_cast_fp16 = mul(x = second_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2937_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2939_cast_fp16 = sub(x = var_2936_cast_fp16, y = var_2937_cast_fp16)[name = string("op_2939_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2940_cast_fp16 = mul(x = second_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2940_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2941_cast_fp16 = mul(x = first_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2941_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2943_cast_fp16 = add(x = var_2940_cast_fp16, y = var_2941_cast_fp16)[name = string("op_2943_cast_fp16")];
            int32 var_2945 = const()[name = string("op_2945"), 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_2945, interleave = q_9_interleave_0, values = (var_2939_cast_fp16, var_2943_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_95_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_95_cast_fp16)[name = string("second_19_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2957_cast_fp16 = mul(x = first_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2957_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2958_cast_fp16 = mul(x = second_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2960_cast_fp16 = sub(x = var_2957_cast_fp16, y = var_2958_cast_fp16)[name = string("op_2960_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2961_cast_fp16 = mul(x = second_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2961_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2962_cast_fp16 = mul(x = first_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2962_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2964_cast_fp16 = add(x = var_2961_cast_fp16, y = var_2962_cast_fp16)[name = string("op_2964_cast_fp16")];
            int32 var_2966 = const()[name = string("op_2966"), 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_2966, interleave = k_9_interleave_0, values = (var_2960_cast_fp16, var_2964_cast_fp16))[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> var_2977_begin_0 = const()[name = string("op_2977_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2977_end_0 = const()[name = string("op_2977_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2977_end_mask_0 = const()[name = string("op_2977_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2977_stride_0 = const()[name = string("op_2977_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2977_cast_fp16 = slice_by_index(begin = var_2977_begin_0, end = var_2977_end_0, end_mask = var_2977_end_mask_0, stride = var_2977_stride_0, x = q_9_cast_fp16)[name = string("op_2977_cast_fp16")];
            tensor<int32, [4]> var_2987_begin_0 = const()[name = string("op_2987_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2987_end_0 = const()[name = string("op_2987_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2987_end_mask_0 = const()[name = string("op_2987_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2987_stride_0 = const()[name = string("op_2987_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2987_cast_fp16 = slice_by_index(begin = var_2987_begin_0, end = var_2987_end_0, end_mask = var_2987_end_mask_0, stride = var_2987_stride_0, x = q_9_cast_fp16)[name = string("op_2987_cast_fp16")];
            int32 var_2989 = const()[name = string("op_2989"), 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_2989, interleave = stacked_9_interleave_0, values = (var_2977_cast_fp16, var_2987_cast_fp16))[name = string("stacked_9_cast_fp16")];
            tensor<int32, [4]> var_2995_begin_0 = const()[name = string("op_2995_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2995_end_0 = const()[name = string("op_2995_end_0"), val = tensor<int32, [4]>([5, 8, 528, 128])];
            tensor<bool, [4]> var_2995_end_mask_0 = const()[name = string("op_2995_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2995_cast_fp16 = slice_by_index(begin = var_2995_begin_0, end = var_2995_end_0, end_mask = var_2995_end_mask_0, x = read_state_0)[name = string("op_2995_cast_fp16")];
            tensor<int32, [4]> var_3005_begin_0 = const()[name = string("op_3005_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3005_end_0 = const()[name = string("op_3005_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3005_end_mask_0 = const()[name = string("op_3005_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3005_cast_fp16 = slice_by_index(begin = var_3005_begin_0, end = var_3005_end_0, end_mask = var_3005_end_mask_0, x = var_2995_cast_fp16)[name = string("op_3005_cast_fp16")];
            bool var_3014_transpose_x_1 = const()[name = string("op_3014_transpose_x_1"), val = bool(false)];
            bool var_3014_transpose_y_1 = const()[name = string("op_3014_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_3014_cast_fp16 = matmul(transpose_x = var_3014_transpose_x_1, transpose_y = var_3014_transpose_y_1, x = var_3005_cast_fp16, y = stacked_9_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<int32, [4]> var_3017_perm_0 = const()[name = string("op_3017_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3018_to_fp16 = const()[name = string("op_3018_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_3017_cast_fp16 = transpose(perm = var_3017_perm_0, x = var_3014_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 8, 32, 512]> var_3019_cast_fp16 = mul(x = var_3017_cast_fp16, y = var_3018_to_fp16)[name = string("op_3019_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_9_cast_fp16 = add(x = var_3019_cast_fp16, y = cacheMask)[name = string("cache_weights_9_cast_fp16")];
            bool var_3025_transpose_x_1 = const()[name = string("op_3025_transpose_x_1"), val = bool(false)];
            bool var_3025_transpose_y_1 = const()[name = string("op_3025_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3025_cast_fp16 = matmul(transpose_x = var_3025_transpose_x_1, transpose_y = var_3025_transpose_y_1, x = stacked_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_3025_cast_fp16")];
            fp16 var_3026_to_fp16 = const()[name = string("op_3026_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3027_cast_fp16 = mul(x = var_3025_cast_fp16, y = var_3026_to_fp16)[name = string("op_3027_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_9_cast_fp16 = add(x = var_3027_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_9_cast_fp16")];
            int32 var_3031 = const()[name = string("op_3031"), val = int32(-1)];
            bool var_3032_interleave_0 = const()[name = string("op_3032_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_3032_cast_fp16 = concat(axis = var_3031, interleave = var_3032_interleave_0, values = (cache_weights_9_cast_fp16, call_weights_9_cast_fp16))[name = string("op_3032_cast_fp16")];
            int32 var_3033 = const()[name = string("op_3033"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_9_cast_fp16 = softmax(axis = var_3033, x = var_3032_cast_fp16)[name = string("shares_9_cast_fp16")];
            tensor<int32, [4]> var_3040_begin_0 = const()[name = string("op_3040_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3040_end_0 = const()[name = string("op_3040_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_3040_end_mask_0 = const()[name = string("op_3040_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_3040_cast_fp16 = slice_by_index(begin = var_3040_begin_0, end = var_3040_end_0, end_mask = var_3040_end_mask_0, x = shares_9_cast_fp16)[name = string("op_3040_cast_fp16")];
            tensor<int32, [4]> var_3045_begin_0 = const()[name = string("op_3045_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_3045_end_0 = const()[name = string("op_3045_end_0"), val = tensor<int32, [4]>([5, 8, 528, 128])];
            tensor<bool, [4]> var_3045_end_mask_0 = const()[name = string("op_3045_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3045_cast_fp16 = slice_by_index(begin = var_3045_begin_0, end = var_3045_end_0, end_mask = var_3045_end_mask_0, x = read_state_1)[name = string("op_3045_cast_fp16")];
            tensor<int32, [4]> var_3055_begin_0 = const()[name = string("op_3055_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3055_end_0 = const()[name = string("op_3055_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3055_end_mask_0 = const()[name = string("op_3055_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3055_cast_fp16 = slice_by_index(begin = var_3055_begin_0, end = var_3055_end_0, end_mask = var_3055_end_mask_0, x = var_3045_cast_fp16)[name = string("op_3055_cast_fp16")];
            bool var_3061_transpose_x_0 = const()[name = string("op_3061_transpose_x_0"), val = bool(false)];
            bool var_3061_transpose_y_0 = const()[name = string("op_3061_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3061_cast_fp16 = matmul(transpose_x = var_3061_transpose_x_0, transpose_y = var_3061_transpose_y_0, x = var_3040_cast_fp16, y = var_3055_cast_fp16)[name = string("op_3061_cast_fp16")];
            tensor<int32, [4]> var_3066_begin_0 = const()[name = string("op_3066_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_3066_end_0 = const()[name = string("op_3066_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_3066_end_mask_0 = const()[name = string("op_3066_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3066_cast_fp16 = slice_by_index(begin = var_3066_begin_0, end = var_3066_end_0, end_mask = var_3066_end_mask_0, x = shares_9_cast_fp16)[name = string("op_3066_cast_fp16")];
            bool var_3067_transpose_x_0 = const()[name = string("op_3067_transpose_x_0"), val = bool(false)];
            bool var_3067_transpose_y_0 = const()[name = string("op_3067_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_2922_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 8, 32, 128]> var_3067_cast_fp16 = matmul(transpose_x = var_3067_transpose_x_0, transpose_y = var_3067_transpose_y_0, x = var_3066_cast_fp16, y = v_9_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_17_cast_fp16 = add(x = var_3061_cast_fp16, y = var_3067_cast_fp16)[name = string("mixed_17_cast_fp16")];
            tensor<int32, [4]> var_3084_begin_0 = const()[name = string("op_3084_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3084_end_0 = const()[name = string("op_3084_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3084_end_mask_0 = const()[name = string("op_3084_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3084_cast_fp16 = slice_by_index(begin = var_3084_begin_0, end = var_3084_end_0, end_mask = var_3084_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [4]> var_3099_begin_0 = const()[name = string("op_3099_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3099_end_0 = const()[name = string("op_3099_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3099_end_mask_0 = const()[name = string("op_3099_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3099_cast_fp16 = slice_by_index(begin = var_3099_begin_0, end = var_3099_end_0, end_mask = var_3099_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_3099_cast_fp16")];
            int32 var_3101 = const()[name = string("op_3101"), 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_3101, interleave = mixed_19_interleave_0, values = (var_3084_cast_fp16, var_3099_cast_fp16))[name = string("mixed_19_cast_fp16")];
            tensor<int32, [4]> var_3105_perm_0 = const()[name = string("op_3105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3109 = const()[name = string("op_3109"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3105_cast_fp16 = transpose(perm = var_3105_perm_0, x = mixed_19_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 2048]> input_43_cast_fp16 = reshape(shape = var_3109, x = var_3105_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_97_cast_fp16 = add(x = x_81_cast_fp16, y = linear_31_cast_fp16)[name = string("x_97_cast_fp16")];
            bool var_3119 = const()[name = string("op_3119"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3122_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<int32, [1]> peak_39_axes_0 = const()[name = string("peak_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_39_cast_fp16 = reduce_max(axes = peak_39_axes_0, keep_dims = var_3119, x = var_3122_cast_fp16)[name = string("peak_39_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_99_cast_fp16 = real_div(x = x_97_cast_fp16, y = peak_39_cast_fp16)[name = string("x_99_cast_fp16")];
            fp16 var_3118_promoted_to_fp16 = const()[name = string("op_3118_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3126_cast_fp16 = pow(x = x_99_cast_fp16, y = var_3118_promoted_to_fp16)[name = string("op_3126_cast_fp16")];
            tensor<int32, [1]> var_3128_axes_0 = const()[name = string("op_3128_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3128_keep_dims_0 = const()[name = string("op_3128_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3128_cast_fp16 = reduce_mean(axes = var_3128_axes_0, keep_dims = var_3128_keep_dims_0, x = var_3126_cast_fp16)[name = string("op_3128_cast_fp16")];
            fp16 var_3118_promoted_1_to_fp16 = const()[name = string("op_3118_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3129_cast_fp16 = pow(x = peak_39_cast_fp16, y = var_3118_promoted_1_to_fp16)[name = string("op_3129_cast_fp16")];
            fp32 var_3130_epsilon_0 = const()[name = string("op_3130_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3130_cast_fp16 = inverse(epsilon = var_3130_epsilon_0, x = var_3129_cast_fp16)[name = string("op_3130_cast_fp16")];
            fp16 var_3131_to_fp16 = const()[name = string("op_3131_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3132_cast_fp16 = mul(x = var_3130_cast_fp16, y = var_3131_to_fp16)[name = string("op_3132_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3133_cast_fp16 = add(x = var_3128_cast_fp16, y = var_3132_cast_fp16)[name = string("op_3133_cast_fp16")];
            fp32 var_3134_epsilon_0 = const()[name = string("op_3134_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3134_cast_fp16 = rsqrt(epsilon = var_3134_epsilon_0, x = var_3133_cast_fp16)[name = string("op_3134_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3135_cast_fp16 = mul(x = x_99_cast_fp16, y = var_3134_cast_fp16)[name = string("op_3135_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_3135_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_3140_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_3140_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_3140_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_101_cast_fp16 = add(x = x_97_cast_fp16, y = linear_34_cast_fp16)[name = string("x_101_cast_fp16")];
            bool var_3153 = const()[name = string("op_3153"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3156_cast_fp16 = abs(x = x_101_cast_fp16)[name = string("op_3156_cast_fp16")];
            tensor<int32, [1]> peak_41_axes_0 = const()[name = string("peak_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_41_cast_fp16 = reduce_max(axes = peak_41_axes_0, keep_dims = var_3153, x = var_3156_cast_fp16)[name = string("peak_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_103_cast_fp16 = real_div(x = x_101_cast_fp16, y = peak_41_cast_fp16)[name = string("x_103_cast_fp16")];
            fp16 var_3152_promoted_to_fp16 = const()[name = string("op_3152_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3160_cast_fp16 = pow(x = x_103_cast_fp16, y = var_3152_promoted_to_fp16)[name = string("op_3160_cast_fp16")];
            tensor<int32, [1]> var_3162_axes_0 = const()[name = string("op_3162_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3162_keep_dims_0 = const()[name = string("op_3162_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3162_cast_fp16 = reduce_mean(axes = var_3162_axes_0, keep_dims = var_3162_keep_dims_0, x = var_3160_cast_fp16)[name = string("op_3162_cast_fp16")];
            fp16 var_3152_promoted_1_to_fp16 = const()[name = string("op_3152_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3163_cast_fp16 = pow(x = peak_41_cast_fp16, y = var_3152_promoted_1_to_fp16)[name = string("op_3163_cast_fp16")];
            fp32 var_3164_epsilon_0 = const()[name = string("op_3164_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3164_cast_fp16 = inverse(epsilon = var_3164_epsilon_0, x = var_3163_cast_fp16)[name = string("op_3164_cast_fp16")];
            fp16 var_3165_to_fp16 = const()[name = string("op_3165_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3166_cast_fp16 = mul(x = var_3164_cast_fp16, y = var_3165_to_fp16)[name = string("op_3166_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3167_cast_fp16 = add(x = var_3162_cast_fp16, y = var_3166_cast_fp16)[name = string("op_3167_cast_fp16")];
            fp32 var_3168_epsilon_0 = const()[name = string("op_3168_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3168_cast_fp16 = rsqrt(epsilon = var_3168_epsilon_0, x = var_3167_cast_fp16)[name = string("op_3168_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3169_cast_fp16 = mul(x = x_103_cast_fp16, y = var_3168_cast_fp16)[name = string("op_3169_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_3169_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_3178 = const()[name = string("op_3178"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3179_cast_fp16 = reshape(shape = var_3178, x = linear_35_cast_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [4]> x_105_perm_0 = const()[name = string("x_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3186 = const()[name = string("op_3186"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_105_cast_fp16 = transpose(perm = x_105_perm_0, x = var_3179_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 16, 16, 128]> var_3189_cast_fp16 = abs(x = x_105_cast_fp16)[name = string("op_3189_cast_fp16")];
            tensor<int32, [1]> peak_43_axes_0 = const()[name = string("peak_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_43_cast_fp16 = reduce_max(axes = peak_43_axes_0, keep_dims = var_3186, x = var_3189_cast_fp16)[name = string("peak_43_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_107_cast_fp16 = real_div(x = x_105_cast_fp16, y = peak_43_cast_fp16)[name = string("x_107_cast_fp16")];
            fp16 var_3185_promoted_to_fp16 = const()[name = string("op_3185_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3193_cast_fp16 = pow(x = x_107_cast_fp16, y = var_3185_promoted_to_fp16)[name = string("op_3193_cast_fp16")];
            tensor<int32, [1]> var_3195_axes_0 = const()[name = string("op_3195_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3195_keep_dims_0 = const()[name = string("op_3195_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3195_cast_fp16 = reduce_mean(axes = var_3195_axes_0, keep_dims = var_3195_keep_dims_0, x = var_3193_cast_fp16)[name = string("op_3195_cast_fp16")];
            fp16 var_3185_promoted_1_to_fp16 = const()[name = string("op_3185_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3196_cast_fp16 = pow(x = peak_43_cast_fp16, y = var_3185_promoted_1_to_fp16)[name = string("op_3196_cast_fp16")];
            fp32 var_3197_epsilon_0 = const()[name = string("op_3197_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3197_cast_fp16 = inverse(epsilon = var_3197_epsilon_0, x = var_3196_cast_fp16)[name = string("op_3197_cast_fp16")];
            fp16 var_3198_to_fp16 = const()[name = string("op_3198_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3199_cast_fp16 = mul(x = var_3197_cast_fp16, y = var_3198_to_fp16)[name = string("op_3199_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3200_cast_fp16 = add(x = var_3195_cast_fp16, y = var_3199_cast_fp16)[name = string("op_3200_cast_fp16")];
            fp32 var_3201_epsilon_0 = const()[name = string("op_3201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3201_cast_fp16 = rsqrt(epsilon = var_3201_epsilon_0, x = var_3200_cast_fp16)[name = string("op_3201_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3202_cast_fp16 = mul(x = x_107_cast_fp16, y = var_3201_cast_fp16)[name = string("op_3202_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_113_cast_fp16 = mul(x = var_3202_cast_fp16, y = model_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_113_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_3211 = const()[name = string("op_3211"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3212_cast_fp16 = reshape(shape = var_3211, x = linear_36_cast_fp16)[name = string("op_3212_cast_fp16")];
            tensor<int32, [4]> x_109_perm_0 = const()[name = string("x_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3219 = const()[name = string("op_3219"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_109_cast_fp16 = transpose(perm = x_109_perm_0, x = var_3212_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 16, 128]> var_3222_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_3222_cast_fp16")];
            tensor<int32, [1]> peak_45_axes_0 = const()[name = string("peak_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_45_cast_fp16 = reduce_max(axes = peak_45_axes_0, keep_dims = var_3219, x = var_3222_cast_fp16)[name = string("peak_45_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_111_cast_fp16 = real_div(x = x_109_cast_fp16, y = peak_45_cast_fp16)[name = string("x_111_cast_fp16")];
            fp16 var_3218_promoted_to_fp16 = const()[name = string("op_3218_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3226_cast_fp16 = pow(x = x_111_cast_fp16, y = var_3218_promoted_to_fp16)[name = string("op_3226_cast_fp16")];
            tensor<int32, [1]> var_3228_axes_0 = const()[name = string("op_3228_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3228_keep_dims_0 = const()[name = string("op_3228_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3228_cast_fp16 = reduce_mean(axes = var_3228_axes_0, keep_dims = var_3228_keep_dims_0, x = var_3226_cast_fp16)[name = string("op_3228_cast_fp16")];
            fp16 var_3218_promoted_1_to_fp16 = const()[name = string("op_3218_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3229_cast_fp16 = pow(x = peak_45_cast_fp16, y = var_3218_promoted_1_to_fp16)[name = string("op_3229_cast_fp16")];
            fp32 var_3230_epsilon_0 = const()[name = string("op_3230_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3230_cast_fp16 = inverse(epsilon = var_3230_epsilon_0, x = var_3229_cast_fp16)[name = string("op_3230_cast_fp16")];
            fp16 var_3231_to_fp16 = const()[name = string("op_3231_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3232_cast_fp16 = mul(x = var_3230_cast_fp16, y = var_3231_to_fp16)[name = string("op_3232_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3233_cast_fp16 = add(x = var_3228_cast_fp16, y = var_3232_cast_fp16)[name = string("op_3233_cast_fp16")];
            fp32 var_3234_epsilon_0 = const()[name = string("op_3234_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3234_cast_fp16 = rsqrt(epsilon = var_3234_epsilon_0, x = var_3233_cast_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3235_cast_fp16 = mul(x = x_111_cast_fp16, y = var_3234_cast_fp16)[name = string("op_3235_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_115_cast_fp16 = mul(x = var_3235_cast_fp16, y = model_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_115_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_3244 = const()[name = string("op_3244"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3245_cast_fp16 = reshape(shape = var_3244, x = linear_37_cast_fp16)[name = string("op_3245_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_113_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_113_cast_fp16)[name = string("second_21_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3259_cast_fp16 = mul(x = first_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3259_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3260_cast_fp16 = mul(x = second_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3260_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3262_cast_fp16 = sub(x = var_3259_cast_fp16, y = var_3260_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3263_cast_fp16 = mul(x = second_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3263_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3264_cast_fp16 = mul(x = first_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3264_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3266_cast_fp16 = add(x = var_3263_cast_fp16, y = var_3264_cast_fp16)[name = string("op_3266_cast_fp16")];
            int32 var_3268 = const()[name = string("op_3268"), 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_3268, interleave = q_11_interleave_0, values = (var_3262_cast_fp16, var_3266_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_115_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_115_cast_fp16)[name = string("second_23_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3280_cast_fp16 = mul(x = first_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3280_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3281_cast_fp16 = mul(x = second_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3281_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3283_cast_fp16 = sub(x = var_3280_cast_fp16, y = var_3281_cast_fp16)[name = string("op_3283_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3284_cast_fp16 = mul(x = second_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3284_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3285_cast_fp16 = mul(x = first_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3285_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3287_cast_fp16 = add(x = var_3284_cast_fp16, y = var_3285_cast_fp16)[name = string("op_3287_cast_fp16")];
            int32 var_3289 = const()[name = string("op_3289"), 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_3289, interleave = k_11_interleave_0, values = (var_3283_cast_fp16, var_3287_cast_fp16))[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> var_3300_begin_0 = const()[name = string("op_3300_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3300_end_0 = const()[name = string("op_3300_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3300_end_mask_0 = const()[name = string("op_3300_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3300_stride_0 = const()[name = string("op_3300_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3300_cast_fp16 = slice_by_index(begin = var_3300_begin_0, end = var_3300_end_0, end_mask = var_3300_end_mask_0, stride = var_3300_stride_0, x = q_11_cast_fp16)[name = string("op_3300_cast_fp16")];
            tensor<int32, [4]> var_3310_begin_0 = const()[name = string("op_3310_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3310_end_0 = const()[name = string("op_3310_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3310_end_mask_0 = const()[name = string("op_3310_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3310_stride_0 = const()[name = string("op_3310_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3310_cast_fp16 = slice_by_index(begin = var_3310_begin_0, end = var_3310_end_0, end_mask = var_3310_end_mask_0, stride = var_3310_stride_0, x = q_11_cast_fp16)[name = string("op_3310_cast_fp16")];
            int32 var_3312 = const()[name = string("op_3312"), 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_3312, interleave = stacked_11_interleave_0, values = (var_3300_cast_fp16, var_3310_cast_fp16))[name = string("stacked_11_cast_fp16")];
            tensor<int32, [4]> var_3318_begin_0 = const()[name = string("op_3318_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3318_end_0 = const()[name = string("op_3318_end_0"), val = tensor<int32, [4]>([6, 8, 528, 128])];
            tensor<bool, [4]> var_3318_end_mask_0 = const()[name = string("op_3318_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3318_cast_fp16 = slice_by_index(begin = var_3318_begin_0, end = var_3318_end_0, end_mask = var_3318_end_mask_0, x = read_state_0)[name = string("op_3318_cast_fp16")];
            tensor<int32, [4]> var_3328_begin_0 = const()[name = string("op_3328_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3328_end_0 = const()[name = string("op_3328_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3328_end_mask_0 = const()[name = string("op_3328_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3328_cast_fp16 = slice_by_index(begin = var_3328_begin_0, end = var_3328_end_0, end_mask = var_3328_end_mask_0, x = var_3318_cast_fp16)[name = string("op_3328_cast_fp16")];
            bool var_3337_transpose_x_1 = const()[name = string("op_3337_transpose_x_1"), val = bool(false)];
            bool var_3337_transpose_y_1 = const()[name = string("op_3337_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_3337_cast_fp16 = matmul(transpose_x = var_3337_transpose_x_1, transpose_y = var_3337_transpose_y_1, x = var_3328_cast_fp16, y = stacked_11_cast_fp16)[name = string("op_3337_cast_fp16")];
            tensor<int32, [4]> var_3340_perm_0 = const()[name = string("op_3340_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3341_to_fp16 = const()[name = string("op_3341_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_3340_cast_fp16 = transpose(perm = var_3340_perm_0, x = var_3337_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 8, 32, 512]> var_3342_cast_fp16 = mul(x = var_3340_cast_fp16, y = var_3341_to_fp16)[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_11_cast_fp16 = add(x = var_3342_cast_fp16, y = cacheMask)[name = string("cache_weights_11_cast_fp16")];
            bool var_3348_transpose_x_1 = const()[name = string("op_3348_transpose_x_1"), val = bool(false)];
            bool var_3348_transpose_y_1 = const()[name = string("op_3348_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3348_cast_fp16 = matmul(transpose_x = var_3348_transpose_x_1, transpose_y = var_3348_transpose_y_1, x = stacked_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_3348_cast_fp16")];
            fp16 var_3349_to_fp16 = const()[name = string("op_3349_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3350_cast_fp16 = mul(x = var_3348_cast_fp16, y = var_3349_to_fp16)[name = string("op_3350_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_11_cast_fp16 = add(x = var_3350_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_11_cast_fp16")];
            int32 var_3354 = const()[name = string("op_3354"), val = int32(-1)];
            bool var_3355_interleave_0 = const()[name = string("op_3355_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_3355_cast_fp16 = concat(axis = var_3354, interleave = var_3355_interleave_0, values = (cache_weights_11_cast_fp16, call_weights_11_cast_fp16))[name = string("op_3355_cast_fp16")];
            int32 var_3356 = const()[name = string("op_3356"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_11_cast_fp16 = softmax(axis = var_3356, x = var_3355_cast_fp16)[name = string("shares_11_cast_fp16")];
            tensor<int32, [4]> var_3363_begin_0 = const()[name = string("op_3363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3363_end_0 = const()[name = string("op_3363_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_3363_end_mask_0 = const()[name = string("op_3363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_3363_cast_fp16 = slice_by_index(begin = var_3363_begin_0, end = var_3363_end_0, end_mask = var_3363_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3363_cast_fp16")];
            tensor<int32, [4]> var_3368_begin_0 = const()[name = string("op_3368_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3368_end_0 = const()[name = string("op_3368_end_0"), val = tensor<int32, [4]>([6, 8, 528, 128])];
            tensor<bool, [4]> var_3368_end_mask_0 = const()[name = string("op_3368_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3368_cast_fp16 = slice_by_index(begin = var_3368_begin_0, end = var_3368_end_0, end_mask = var_3368_end_mask_0, x = read_state_1)[name = string("op_3368_cast_fp16")];
            tensor<int32, [4]> var_3378_begin_0 = const()[name = string("op_3378_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3378_end_0 = const()[name = string("op_3378_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3378_end_mask_0 = const()[name = string("op_3378_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3378_cast_fp16 = slice_by_index(begin = var_3378_begin_0, end = var_3378_end_0, end_mask = var_3378_end_mask_0, x = var_3368_cast_fp16)[name = string("op_3378_cast_fp16")];
            bool var_3384_transpose_x_0 = const()[name = string("op_3384_transpose_x_0"), val = bool(false)];
            bool var_3384_transpose_y_0 = const()[name = string("op_3384_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3384_cast_fp16 = matmul(transpose_x = var_3384_transpose_x_0, transpose_y = var_3384_transpose_y_0, x = var_3363_cast_fp16, y = var_3378_cast_fp16)[name = string("op_3384_cast_fp16")];
            tensor<int32, [4]> var_3389_begin_0 = const()[name = string("op_3389_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_3389_end_0 = const()[name = string("op_3389_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_3389_end_mask_0 = const()[name = string("op_3389_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3389_cast_fp16 = slice_by_index(begin = var_3389_begin_0, end = var_3389_end_0, end_mask = var_3389_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3389_cast_fp16")];
            bool var_3390_transpose_x_0 = const()[name = string("op_3390_transpose_x_0"), val = bool(false)];
            bool var_3390_transpose_y_0 = const()[name = string("op_3390_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_3245_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 8, 32, 128]> var_3390_cast_fp16 = matmul(transpose_x = var_3390_transpose_x_0, transpose_y = var_3390_transpose_y_0, x = var_3389_cast_fp16, y = v_11_cast_fp16)[name = string("op_3390_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_21_cast_fp16 = add(x = var_3384_cast_fp16, y = var_3390_cast_fp16)[name = string("mixed_21_cast_fp16")];
            tensor<int32, [4]> var_3407_begin_0 = const()[name = string("op_3407_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3407_end_0 = const()[name = string("op_3407_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3407_end_mask_0 = const()[name = string("op_3407_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3407_cast_fp16")];
            tensor<int32, [4]> var_3422_begin_0 = const()[name = string("op_3422_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3422_end_0 = const()[name = string("op_3422_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3422_end_mask_0 = const()[name = string("op_3422_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3422_cast_fp16 = slice_by_index(begin = var_3422_begin_0, end = var_3422_end_0, end_mask = var_3422_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3422_cast_fp16")];
            int32 var_3424 = const()[name = string("op_3424"), 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_3424, interleave = mixed_23_interleave_0, values = (var_3407_cast_fp16, var_3422_cast_fp16))[name = string("mixed_23_cast_fp16")];
            tensor<int32, [4]> var_3428_perm_0 = const()[name = string("op_3428_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3432 = const()[name = string("op_3432"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3428_cast_fp16 = transpose(perm = var_3428_perm_0, x = mixed_23_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 2048]> input_53_cast_fp16 = reshape(shape = var_3432, x = var_3428_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_117_cast_fp16 = add(x = x_101_cast_fp16, y = linear_38_cast_fp16)[name = string("x_117_cast_fp16")];
            bool var_3442 = const()[name = string("op_3442"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3445_cast_fp16 = abs(x = x_117_cast_fp16)[name = string("op_3445_cast_fp16")];
            tensor<int32, [1]> peak_47_axes_0 = const()[name = string("peak_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_47_cast_fp16 = reduce_max(axes = peak_47_axes_0, keep_dims = var_3442, x = var_3445_cast_fp16)[name = string("peak_47_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_119_cast_fp16 = real_div(x = x_117_cast_fp16, y = peak_47_cast_fp16)[name = string("x_119_cast_fp16")];
            fp16 var_3441_promoted_to_fp16 = const()[name = string("op_3441_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3449_cast_fp16 = pow(x = x_119_cast_fp16, y = var_3441_promoted_to_fp16)[name = string("op_3449_cast_fp16")];
            tensor<int32, [1]> var_3451_axes_0 = const()[name = string("op_3451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3451_keep_dims_0 = const()[name = string("op_3451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3451_cast_fp16 = reduce_mean(axes = var_3451_axes_0, keep_dims = var_3451_keep_dims_0, x = var_3449_cast_fp16)[name = string("op_3451_cast_fp16")];
            fp16 var_3441_promoted_1_to_fp16 = const()[name = string("op_3441_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3452_cast_fp16 = pow(x = peak_47_cast_fp16, y = var_3441_promoted_1_to_fp16)[name = string("op_3452_cast_fp16")];
            fp32 var_3453_epsilon_0 = const()[name = string("op_3453_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3453_cast_fp16 = inverse(epsilon = var_3453_epsilon_0, x = var_3452_cast_fp16)[name = string("op_3453_cast_fp16")];
            fp16 var_3454_to_fp16 = const()[name = string("op_3454_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3455_cast_fp16 = mul(x = var_3453_cast_fp16, y = var_3454_to_fp16)[name = string("op_3455_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3456_cast_fp16 = add(x = var_3451_cast_fp16, y = var_3455_cast_fp16)[name = string("op_3456_cast_fp16")];
            fp32 var_3457_epsilon_0 = const()[name = string("op_3457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3457_cast_fp16 = rsqrt(epsilon = var_3457_epsilon_0, x = var_3456_cast_fp16)[name = string("op_3457_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3458_cast_fp16 = mul(x = x_119_cast_fp16, y = var_3457_cast_fp16)[name = string("op_3458_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_3458_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_3463_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_3463_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_3463_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_121_cast_fp16 = add(x = x_117_cast_fp16, y = linear_41_cast_fp16)[name = string("x_121_cast_fp16")];
            bool var_3476 = const()[name = string("op_3476"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3479_cast_fp16 = abs(x = x_121_cast_fp16)[name = string("op_3479_cast_fp16")];
            tensor<int32, [1]> peak_49_axes_0 = const()[name = string("peak_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_49_cast_fp16 = reduce_max(axes = peak_49_axes_0, keep_dims = var_3476, x = var_3479_cast_fp16)[name = string("peak_49_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_123_cast_fp16 = real_div(x = x_121_cast_fp16, y = peak_49_cast_fp16)[name = string("x_123_cast_fp16")];
            fp16 var_3475_promoted_to_fp16 = const()[name = string("op_3475_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3483_cast_fp16 = pow(x = x_123_cast_fp16, y = var_3475_promoted_to_fp16)[name = string("op_3483_cast_fp16")];
            tensor<int32, [1]> var_3485_axes_0 = const()[name = string("op_3485_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3485_keep_dims_0 = const()[name = string("op_3485_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3485_cast_fp16 = reduce_mean(axes = var_3485_axes_0, keep_dims = var_3485_keep_dims_0, x = var_3483_cast_fp16)[name = string("op_3485_cast_fp16")];
            fp16 var_3475_promoted_1_to_fp16 = const()[name = string("op_3475_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3486_cast_fp16 = pow(x = peak_49_cast_fp16, y = var_3475_promoted_1_to_fp16)[name = string("op_3486_cast_fp16")];
            fp32 var_3487_epsilon_0 = const()[name = string("op_3487_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3487_cast_fp16 = inverse(epsilon = var_3487_epsilon_0, x = var_3486_cast_fp16)[name = string("op_3487_cast_fp16")];
            fp16 var_3488_to_fp16 = const()[name = string("op_3488_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3489_cast_fp16 = mul(x = var_3487_cast_fp16, y = var_3488_to_fp16)[name = string("op_3489_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3490_cast_fp16 = add(x = var_3485_cast_fp16, y = var_3489_cast_fp16)[name = string("op_3490_cast_fp16")];
            fp32 var_3491_epsilon_0 = const()[name = string("op_3491_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3491_cast_fp16 = rsqrt(epsilon = var_3491_epsilon_0, x = var_3490_cast_fp16)[name = string("op_3491_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3492_cast_fp16 = mul(x = x_123_cast_fp16, y = var_3491_cast_fp16)[name = string("op_3492_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_3492_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_3501 = const()[name = string("op_3501"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3502_cast_fp16 = reshape(shape = var_3501, x = linear_42_cast_fp16)[name = string("op_3502_cast_fp16")];
            tensor<int32, [4]> x_125_perm_0 = const()[name = string("x_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3509 = const()[name = string("op_3509"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_125_cast_fp16 = transpose(perm = x_125_perm_0, x = var_3502_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 16, 128]> var_3512_cast_fp16 = abs(x = x_125_cast_fp16)[name = string("op_3512_cast_fp16")];
            tensor<int32, [1]> peak_51_axes_0 = const()[name = string("peak_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_51_cast_fp16 = reduce_max(axes = peak_51_axes_0, keep_dims = var_3509, x = var_3512_cast_fp16)[name = string("peak_51_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_127_cast_fp16 = real_div(x = x_125_cast_fp16, y = peak_51_cast_fp16)[name = string("x_127_cast_fp16")];
            fp16 var_3508_promoted_to_fp16 = const()[name = string("op_3508_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3516_cast_fp16 = pow(x = x_127_cast_fp16, y = var_3508_promoted_to_fp16)[name = string("op_3516_cast_fp16")];
            tensor<int32, [1]> var_3518_axes_0 = const()[name = string("op_3518_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3518_keep_dims_0 = const()[name = string("op_3518_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3518_cast_fp16 = reduce_mean(axes = var_3518_axes_0, keep_dims = var_3518_keep_dims_0, x = var_3516_cast_fp16)[name = string("op_3518_cast_fp16")];
            fp16 var_3508_promoted_1_to_fp16 = const()[name = string("op_3508_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3519_cast_fp16 = pow(x = peak_51_cast_fp16, y = var_3508_promoted_1_to_fp16)[name = string("op_3519_cast_fp16")];
            fp32 var_3520_epsilon_0 = const()[name = string("op_3520_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3520_cast_fp16 = inverse(epsilon = var_3520_epsilon_0, x = var_3519_cast_fp16)[name = string("op_3520_cast_fp16")];
            fp16 var_3521_to_fp16 = const()[name = string("op_3521_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3522_cast_fp16 = mul(x = var_3520_cast_fp16, y = var_3521_to_fp16)[name = string("op_3522_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3523_cast_fp16 = add(x = var_3518_cast_fp16, y = var_3522_cast_fp16)[name = string("op_3523_cast_fp16")];
            fp32 var_3524_epsilon_0 = const()[name = string("op_3524_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3524_cast_fp16 = rsqrt(epsilon = var_3524_epsilon_0, x = var_3523_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3525_cast_fp16 = mul(x = x_127_cast_fp16, y = var_3524_cast_fp16)[name = string("op_3525_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_133_cast_fp16 = mul(x = var_3525_cast_fp16, y = model_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_133_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_3534 = const()[name = string("op_3534"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3535_cast_fp16 = reshape(shape = var_3534, x = linear_43_cast_fp16)[name = string("op_3535_cast_fp16")];
            tensor<int32, [4]> x_129_perm_0 = const()[name = string("x_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3542 = const()[name = string("op_3542"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_129_cast_fp16 = transpose(perm = x_129_perm_0, x = var_3535_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 8, 16, 128]> var_3545_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_3545_cast_fp16")];
            tensor<int32, [1]> peak_53_axes_0 = const()[name = string("peak_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_53_cast_fp16 = reduce_max(axes = peak_53_axes_0, keep_dims = var_3542, x = var_3545_cast_fp16)[name = string("peak_53_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_131_cast_fp16 = real_div(x = x_129_cast_fp16, y = peak_53_cast_fp16)[name = string("x_131_cast_fp16")];
            fp16 var_3541_promoted_to_fp16 = const()[name = string("op_3541_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3549_cast_fp16 = pow(x = x_131_cast_fp16, y = var_3541_promoted_to_fp16)[name = string("op_3549_cast_fp16")];
            tensor<int32, [1]> var_3551_axes_0 = const()[name = string("op_3551_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3551_keep_dims_0 = const()[name = string("op_3551_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3551_cast_fp16 = reduce_mean(axes = var_3551_axes_0, keep_dims = var_3551_keep_dims_0, x = var_3549_cast_fp16)[name = string("op_3551_cast_fp16")];
            fp16 var_3541_promoted_1_to_fp16 = const()[name = string("op_3541_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3552_cast_fp16 = pow(x = peak_53_cast_fp16, y = var_3541_promoted_1_to_fp16)[name = string("op_3552_cast_fp16")];
            fp32 var_3553_epsilon_0 = const()[name = string("op_3553_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3553_cast_fp16 = inverse(epsilon = var_3553_epsilon_0, x = var_3552_cast_fp16)[name = string("op_3553_cast_fp16")];
            fp16 var_3554_to_fp16 = const()[name = string("op_3554_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3555_cast_fp16 = mul(x = var_3553_cast_fp16, y = var_3554_to_fp16)[name = string("op_3555_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3556_cast_fp16 = add(x = var_3551_cast_fp16, y = var_3555_cast_fp16)[name = string("op_3556_cast_fp16")];
            fp32 var_3557_epsilon_0 = const()[name = string("op_3557_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3557_cast_fp16 = rsqrt(epsilon = var_3557_epsilon_0, x = var_3556_cast_fp16)[name = string("op_3557_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3558_cast_fp16 = mul(x = x_131_cast_fp16, y = var_3557_cast_fp16)[name = string("op_3558_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_135_cast_fp16 = mul(x = var_3558_cast_fp16, y = model_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_135_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_3567 = const()[name = string("op_3567"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3568_cast_fp16 = reshape(shape = var_3567, x = linear_44_cast_fp16)[name = string("op_3568_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_133_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_133_cast_fp16)[name = string("second_25_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3582_cast_fp16 = mul(x = first_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3582_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3583_cast_fp16 = mul(x = second_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3583_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3585_cast_fp16 = sub(x = var_3582_cast_fp16, y = var_3583_cast_fp16)[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3586_cast_fp16 = mul(x = second_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3586_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3587_cast_fp16 = mul(x = first_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3587_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3589_cast_fp16 = add(x = var_3586_cast_fp16, y = var_3587_cast_fp16)[name = string("op_3589_cast_fp16")];
            int32 var_3591 = const()[name = string("op_3591"), 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_3591, interleave = q_13_interleave_0, values = (var_3585_cast_fp16, var_3589_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_135_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_135_cast_fp16)[name = string("second_27_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3603_cast_fp16 = mul(x = first_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3603_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3604_cast_fp16 = mul(x = second_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3604_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3606_cast_fp16 = sub(x = var_3603_cast_fp16, y = var_3604_cast_fp16)[name = string("op_3606_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3607_cast_fp16 = mul(x = second_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3607_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3608_cast_fp16 = mul(x = first_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3608_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3610_cast_fp16 = add(x = var_3607_cast_fp16, y = var_3608_cast_fp16)[name = string("op_3610_cast_fp16")];
            int32 var_3612 = const()[name = string("op_3612"), 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_3612, interleave = k_13_interleave_0, values = (var_3606_cast_fp16, var_3610_cast_fp16))[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_3623_begin_0 = const()[name = string("op_3623_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3623_end_0 = const()[name = string("op_3623_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3623_end_mask_0 = const()[name = string("op_3623_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3623_stride_0 = const()[name = string("op_3623_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3623_cast_fp16 = slice_by_index(begin = var_3623_begin_0, end = var_3623_end_0, end_mask = var_3623_end_mask_0, stride = var_3623_stride_0, x = q_13_cast_fp16)[name = string("op_3623_cast_fp16")];
            tensor<int32, [4]> var_3633_begin_0 = const()[name = string("op_3633_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3633_end_0 = const()[name = string("op_3633_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3633_end_mask_0 = const()[name = string("op_3633_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3633_stride_0 = const()[name = string("op_3633_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3633_cast_fp16 = slice_by_index(begin = var_3633_begin_0, end = var_3633_end_0, end_mask = var_3633_end_mask_0, stride = var_3633_stride_0, x = q_13_cast_fp16)[name = string("op_3633_cast_fp16")];
            int32 var_3635 = const()[name = string("op_3635"), 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_3635, interleave = stacked_13_interleave_0, values = (var_3623_cast_fp16, var_3633_cast_fp16))[name = string("stacked_13_cast_fp16")];
            tensor<int32, [4]> var_3641_begin_0 = const()[name = string("op_3641_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3641_end_0 = const()[name = string("op_3641_end_0"), val = tensor<int32, [4]>([7, 8, 528, 128])];
            tensor<bool, [4]> var_3641_end_mask_0 = const()[name = string("op_3641_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3641_cast_fp16 = slice_by_index(begin = var_3641_begin_0, end = var_3641_end_0, end_mask = var_3641_end_mask_0, x = read_state_0)[name = string("op_3641_cast_fp16")];
            tensor<int32, [4]> var_3651_begin_0 = const()[name = string("op_3651_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3651_end_0 = const()[name = string("op_3651_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3651_end_mask_0 = const()[name = string("op_3651_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3651_cast_fp16 = slice_by_index(begin = var_3651_begin_0, end = var_3651_end_0, end_mask = var_3651_end_mask_0, x = var_3641_cast_fp16)[name = string("op_3651_cast_fp16")];
            bool var_3660_transpose_x_1 = const()[name = string("op_3660_transpose_x_1"), val = bool(false)];
            bool var_3660_transpose_y_1 = const()[name = string("op_3660_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_3660_cast_fp16 = matmul(transpose_x = var_3660_transpose_x_1, transpose_y = var_3660_transpose_y_1, x = var_3651_cast_fp16, y = stacked_13_cast_fp16)[name = string("op_3660_cast_fp16")];
            tensor<int32, [4]> var_3663_perm_0 = const()[name = string("op_3663_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3664_to_fp16 = const()[name = string("op_3664_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_3663_cast_fp16 = transpose(perm = var_3663_perm_0, x = var_3660_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 8, 32, 512]> var_3665_cast_fp16 = mul(x = var_3663_cast_fp16, y = var_3664_to_fp16)[name = string("op_3665_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_13_cast_fp16 = add(x = var_3665_cast_fp16, y = cacheMask)[name = string("cache_weights_13_cast_fp16")];
            bool var_3671_transpose_x_1 = const()[name = string("op_3671_transpose_x_1"), val = bool(false)];
            bool var_3671_transpose_y_1 = const()[name = string("op_3671_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3671_cast_fp16 = matmul(transpose_x = var_3671_transpose_x_1, transpose_y = var_3671_transpose_y_1, x = stacked_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_3671_cast_fp16")];
            fp16 var_3672_to_fp16 = const()[name = string("op_3672_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3673_cast_fp16 = mul(x = var_3671_cast_fp16, y = var_3672_to_fp16)[name = string("op_3673_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_13_cast_fp16 = add(x = var_3673_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_13_cast_fp16")];
            int32 var_3677 = const()[name = string("op_3677"), val = int32(-1)];
            bool var_3678_interleave_0 = const()[name = string("op_3678_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_3678_cast_fp16 = concat(axis = var_3677, interleave = var_3678_interleave_0, values = (cache_weights_13_cast_fp16, call_weights_13_cast_fp16))[name = string("op_3678_cast_fp16")];
            int32 var_3679 = const()[name = string("op_3679"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_13_cast_fp16 = softmax(axis = var_3679, x = var_3678_cast_fp16)[name = string("shares_13_cast_fp16")];
            tensor<int32, [4]> var_3686_begin_0 = const()[name = string("op_3686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3686_end_0 = const()[name = string("op_3686_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_3686_end_mask_0 = const()[name = string("op_3686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_3686_cast_fp16 = slice_by_index(begin = var_3686_begin_0, end = var_3686_end_0, end_mask = var_3686_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3686_cast_fp16")];
            tensor<int32, [4]> var_3691_begin_0 = const()[name = string("op_3691_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3691_end_0 = const()[name = string("op_3691_end_0"), val = tensor<int32, [4]>([7, 8, 528, 128])];
            tensor<bool, [4]> var_3691_end_mask_0 = const()[name = string("op_3691_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3691_cast_fp16 = slice_by_index(begin = var_3691_begin_0, end = var_3691_end_0, end_mask = var_3691_end_mask_0, x = read_state_1)[name = string("op_3691_cast_fp16")];
            tensor<int32, [4]> var_3701_begin_0 = const()[name = string("op_3701_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3701_end_0 = const()[name = string("op_3701_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3701_end_mask_0 = const()[name = string("op_3701_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3701_cast_fp16 = slice_by_index(begin = var_3701_begin_0, end = var_3701_end_0, end_mask = var_3701_end_mask_0, x = var_3691_cast_fp16)[name = string("op_3701_cast_fp16")];
            bool var_3707_transpose_x_0 = const()[name = string("op_3707_transpose_x_0"), val = bool(false)];
            bool var_3707_transpose_y_0 = const()[name = string("op_3707_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3707_cast_fp16 = matmul(transpose_x = var_3707_transpose_x_0, transpose_y = var_3707_transpose_y_0, x = var_3686_cast_fp16, y = var_3701_cast_fp16)[name = string("op_3707_cast_fp16")];
            tensor<int32, [4]> var_3712_begin_0 = const()[name = string("op_3712_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_3712_end_0 = const()[name = string("op_3712_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_3712_end_mask_0 = const()[name = string("op_3712_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3712_cast_fp16 = slice_by_index(begin = var_3712_begin_0, end = var_3712_end_0, end_mask = var_3712_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3712_cast_fp16")];
            bool var_3713_transpose_x_0 = const()[name = string("op_3713_transpose_x_0"), val = bool(false)];
            bool var_3713_transpose_y_0 = const()[name = string("op_3713_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3568_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 8, 32, 128]> var_3713_cast_fp16 = matmul(transpose_x = var_3713_transpose_x_0, transpose_y = var_3713_transpose_y_0, x = var_3712_cast_fp16, y = v_13_cast_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_25_cast_fp16 = add(x = var_3707_cast_fp16, y = var_3713_cast_fp16)[name = string("mixed_25_cast_fp16")];
            tensor<int32, [4]> var_3730_begin_0 = const()[name = string("op_3730_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3730_end_0 = const()[name = string("op_3730_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3730_end_mask_0 = const()[name = string("op_3730_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3730_cast_fp16 = slice_by_index(begin = var_3730_begin_0, end = var_3730_end_0, end_mask = var_3730_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3730_cast_fp16")];
            tensor<int32, [4]> var_3745_begin_0 = const()[name = string("op_3745_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3745_end_0 = const()[name = string("op_3745_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3745_end_mask_0 = const()[name = string("op_3745_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3745_cast_fp16 = slice_by_index(begin = var_3745_begin_0, end = var_3745_end_0, end_mask = var_3745_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3745_cast_fp16")];
            int32 var_3747 = const()[name = string("op_3747"), 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_3747, interleave = mixed_27_interleave_0, values = (var_3730_cast_fp16, var_3745_cast_fp16))[name = string("mixed_27_cast_fp16")];
            tensor<int32, [4]> var_3751_perm_0 = const()[name = string("op_3751_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3755 = const()[name = string("op_3755"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3751_cast_fp16 = transpose(perm = var_3751_perm_0, x = mixed_27_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 2048]> input_63_cast_fp16 = reshape(shape = var_3755, x = var_3751_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_137_cast_fp16 = add(x = x_121_cast_fp16, y = linear_45_cast_fp16)[name = string("x_137_cast_fp16")];
            bool var_3765 = const()[name = string("op_3765"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3768_cast_fp16 = abs(x = x_137_cast_fp16)[name = string("op_3768_cast_fp16")];
            tensor<int32, [1]> peak_55_axes_0 = const()[name = string("peak_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_55_cast_fp16 = reduce_max(axes = peak_55_axes_0, keep_dims = var_3765, x = var_3768_cast_fp16)[name = string("peak_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_139_cast_fp16 = real_div(x = x_137_cast_fp16, y = peak_55_cast_fp16)[name = string("x_139_cast_fp16")];
            fp16 var_3764_promoted_to_fp16 = const()[name = string("op_3764_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3772_cast_fp16 = pow(x = x_139_cast_fp16, y = var_3764_promoted_to_fp16)[name = string("op_3772_cast_fp16")];
            tensor<int32, [1]> var_3774_axes_0 = const()[name = string("op_3774_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3774_keep_dims_0 = const()[name = string("op_3774_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3774_cast_fp16 = reduce_mean(axes = var_3774_axes_0, keep_dims = var_3774_keep_dims_0, x = var_3772_cast_fp16)[name = string("op_3774_cast_fp16")];
            fp16 var_3764_promoted_1_to_fp16 = const()[name = string("op_3764_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3775_cast_fp16 = pow(x = peak_55_cast_fp16, y = var_3764_promoted_1_to_fp16)[name = string("op_3775_cast_fp16")];
            fp32 var_3776_epsilon_0 = const()[name = string("op_3776_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3776_cast_fp16 = inverse(epsilon = var_3776_epsilon_0, x = var_3775_cast_fp16)[name = string("op_3776_cast_fp16")];
            fp16 var_3777_to_fp16 = const()[name = string("op_3777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3778_cast_fp16 = mul(x = var_3776_cast_fp16, y = var_3777_to_fp16)[name = string("op_3778_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3779_cast_fp16 = add(x = var_3774_cast_fp16, y = var_3778_cast_fp16)[name = string("op_3779_cast_fp16")];
            fp32 var_3780_epsilon_0 = const()[name = string("op_3780_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3780_cast_fp16 = rsqrt(epsilon = var_3780_epsilon_0, x = var_3779_cast_fp16)[name = string("op_3780_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3781_cast_fp16 = mul(x = x_139_cast_fp16, y = var_3780_cast_fp16)[name = string("op_3781_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_3781_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_3786_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_3786_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_3786_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_141_cast_fp16 = add(x = x_137_cast_fp16, y = linear_48_cast_fp16)[name = string("x_141_cast_fp16")];
            bool var_3799 = const()[name = string("op_3799"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_3802_cast_fp16 = abs(x = x_141_cast_fp16)[name = string("op_3802_cast_fp16")];
            tensor<int32, [1]> peak_57_axes_0 = const()[name = string("peak_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_57_cast_fp16 = reduce_max(axes = peak_57_axes_0, keep_dims = var_3799, x = var_3802_cast_fp16)[name = string("peak_57_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_143_cast_fp16 = real_div(x = x_141_cast_fp16, y = peak_57_cast_fp16)[name = string("x_143_cast_fp16")];
            fp16 var_3798_promoted_to_fp16 = const()[name = string("op_3798_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3806_cast_fp16 = pow(x = x_143_cast_fp16, y = var_3798_promoted_to_fp16)[name = string("op_3806_cast_fp16")];
            tensor<int32, [1]> var_3808_axes_0 = const()[name = string("op_3808_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3808_keep_dims_0 = const()[name = string("op_3808_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3808_cast_fp16 = reduce_mean(axes = var_3808_axes_0, keep_dims = var_3808_keep_dims_0, x = var_3806_cast_fp16)[name = string("op_3808_cast_fp16")];
            fp16 var_3798_promoted_1_to_fp16 = const()[name = string("op_3798_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_3809_cast_fp16 = pow(x = peak_57_cast_fp16, y = var_3798_promoted_1_to_fp16)[name = string("op_3809_cast_fp16")];
            fp32 var_3810_epsilon_0 = const()[name = string("op_3810_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_3810_cast_fp16 = inverse(epsilon = var_3810_epsilon_0, x = var_3809_cast_fp16)[name = string("op_3810_cast_fp16")];
            fp16 var_3811_to_fp16 = const()[name = string("op_3811_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3812_cast_fp16 = mul(x = var_3810_cast_fp16, y = var_3811_to_fp16)[name = string("op_3812_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_3813_cast_fp16 = add(x = var_3808_cast_fp16, y = var_3812_cast_fp16)[name = string("op_3813_cast_fp16")];
            fp32 var_3814_epsilon_0 = const()[name = string("op_3814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3814_cast_fp16 = rsqrt(epsilon = var_3814_epsilon_0, x = var_3813_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3815_cast_fp16 = mul(x = x_143_cast_fp16, y = var_3814_cast_fp16)[name = string("op_3815_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_3815_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_3824 = const()[name = string("op_3824"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3825_cast_fp16 = reshape(shape = var_3824, x = linear_49_cast_fp16)[name = string("op_3825_cast_fp16")];
            tensor<int32, [4]> x_145_perm_0 = const()[name = string("x_145_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3832 = const()[name = string("op_3832"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_145_cast_fp16 = transpose(perm = x_145_perm_0, x = var_3825_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 16, 16, 128]> var_3835_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_3835_cast_fp16")];
            tensor<int32, [1]> peak_59_axes_0 = const()[name = string("peak_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_59_cast_fp16 = reduce_max(axes = peak_59_axes_0, keep_dims = var_3832, x = var_3835_cast_fp16)[name = string("peak_59_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_147_cast_fp16 = real_div(x = x_145_cast_fp16, y = peak_59_cast_fp16)[name = string("x_147_cast_fp16")];
            fp16 var_3831_promoted_to_fp16 = const()[name = string("op_3831_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_3839_cast_fp16 = pow(x = x_147_cast_fp16, y = var_3831_promoted_to_fp16)[name = string("op_3839_cast_fp16")];
            tensor<int32, [1]> var_3841_axes_0 = const()[name = string("op_3841_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3841_keep_dims_0 = const()[name = string("op_3841_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3841_cast_fp16 = reduce_mean(axes = var_3841_axes_0, keep_dims = var_3841_keep_dims_0, x = var_3839_cast_fp16)[name = string("op_3841_cast_fp16")];
            fp16 var_3831_promoted_1_to_fp16 = const()[name = string("op_3831_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_3842_cast_fp16 = pow(x = peak_59_cast_fp16, y = var_3831_promoted_1_to_fp16)[name = string("op_3842_cast_fp16")];
            fp32 var_3843_epsilon_0 = const()[name = string("op_3843_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_3843_cast_fp16 = inverse(epsilon = var_3843_epsilon_0, x = var_3842_cast_fp16)[name = string("op_3843_cast_fp16")];
            fp16 var_3844_to_fp16 = const()[name = string("op_3844_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3845_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3844_to_fp16)[name = string("op_3845_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_3846_cast_fp16 = add(x = var_3841_cast_fp16, y = var_3845_cast_fp16)[name = string("op_3846_cast_fp16")];
            fp32 var_3847_epsilon_0 = const()[name = string("op_3847_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3847_cast_fp16 = rsqrt(epsilon = var_3847_epsilon_0, x = var_3846_cast_fp16)[name = string("op_3847_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3848_cast_fp16 = mul(x = x_147_cast_fp16, y = var_3847_cast_fp16)[name = string("op_3848_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_153_cast_fp16 = mul(x = var_3848_cast_fp16, y = model_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_153_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_3857 = const()[name = string("op_3857"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3858_cast_fp16 = reshape(shape = var_3857, x = linear_50_cast_fp16)[name = string("op_3858_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_3865 = const()[name = string("op_3865"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_3858_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 8, 16, 128]> var_3868_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_3868_cast_fp16")];
            tensor<int32, [1]> peak_61_axes_0 = const()[name = string("peak_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_61_cast_fp16 = reduce_max(axes = peak_61_axes_0, keep_dims = var_3865, x = var_3868_cast_fp16)[name = string("peak_61_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_151_cast_fp16 = real_div(x = x_149_cast_fp16, y = peak_61_cast_fp16)[name = string("x_151_cast_fp16")];
            fp16 var_3864_promoted_to_fp16 = const()[name = string("op_3864_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_3872_cast_fp16 = pow(x = x_151_cast_fp16, y = var_3864_promoted_to_fp16)[name = string("op_3872_cast_fp16")];
            tensor<int32, [1]> var_3874_axes_0 = const()[name = string("op_3874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3874_keep_dims_0 = const()[name = string("op_3874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3874_cast_fp16 = reduce_mean(axes = var_3874_axes_0, keep_dims = var_3874_keep_dims_0, x = var_3872_cast_fp16)[name = string("op_3874_cast_fp16")];
            fp16 var_3864_promoted_1_to_fp16 = const()[name = string("op_3864_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_3875_cast_fp16 = pow(x = peak_61_cast_fp16, y = var_3864_promoted_1_to_fp16)[name = string("op_3875_cast_fp16")];
            fp32 var_3876_epsilon_0 = const()[name = string("op_3876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_3876_cast_fp16 = inverse(epsilon = var_3876_epsilon_0, x = var_3875_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, 8, 16, 1]> var_3878_cast_fp16 = mul(x = var_3876_cast_fp16, y = var_3877_to_fp16)[name = string("op_3878_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_3879_cast_fp16 = add(x = var_3874_cast_fp16, y = var_3878_cast_fp16)[name = string("op_3879_cast_fp16")];
            fp32 var_3880_epsilon_0 = const()[name = string("op_3880_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3880_cast_fp16 = rsqrt(epsilon = var_3880_epsilon_0, x = var_3879_cast_fp16)[name = string("op_3880_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3881_cast_fp16 = mul(x = x_151_cast_fp16, y = var_3880_cast_fp16)[name = string("op_3881_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_155_cast_fp16 = mul(x = var_3881_cast_fp16, y = model_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_155_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_3890 = const()[name = string("op_3890"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3891_cast_fp16 = reshape(shape = var_3890, x = linear_51_cast_fp16)[name = string("op_3891_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_153_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_153_cast_fp16)[name = string("second_29_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3905_cast_fp16 = mul(x = first_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3905_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3906_cast_fp16 = mul(x = second_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3908_cast_fp16 = sub(x = var_3905_cast_fp16, y = var_3906_cast_fp16)[name = string("op_3908_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3909_cast_fp16 = mul(x = second_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3909_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3910_cast_fp16 = mul(x = first_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3910_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3912_cast_fp16 = add(x = var_3909_cast_fp16, y = var_3910_cast_fp16)[name = string("op_3912_cast_fp16")];
            int32 var_3914 = const()[name = string("op_3914"), 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_3914, interleave = q_15_interleave_0, values = (var_3908_cast_fp16, var_3912_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_155_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_155_cast_fp16)[name = string("second_31_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3926_cast_fp16 = mul(x = first_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3926_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3927_cast_fp16 = mul(x = second_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3927_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3929_cast_fp16 = sub(x = var_3926_cast_fp16, y = var_3927_cast_fp16)[name = string("op_3929_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3930_cast_fp16 = mul(x = second_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3930_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3931_cast_fp16 = mul(x = first_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3933_cast_fp16 = add(x = var_3930_cast_fp16, y = var_3931_cast_fp16)[name = string("op_3933_cast_fp16")];
            int32 var_3935 = const()[name = string("op_3935"), 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_3935, interleave = k_15_interleave_0, values = (var_3929_cast_fp16, var_3933_cast_fp16))[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> var_3946_begin_0 = const()[name = string("op_3946_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3946_end_0 = const()[name = string("op_3946_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3946_end_mask_0 = const()[name = string("op_3946_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3946_stride_0 = const()[name = string("op_3946_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3946_cast_fp16 = slice_by_index(begin = var_3946_begin_0, end = var_3946_end_0, end_mask = var_3946_end_mask_0, stride = var_3946_stride_0, x = q_15_cast_fp16)[name = string("op_3946_cast_fp16")];
            tensor<int32, [4]> var_3956_begin_0 = const()[name = string("op_3956_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3956_end_0 = const()[name = string("op_3956_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3956_end_mask_0 = const()[name = string("op_3956_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3956_stride_0 = const()[name = string("op_3956_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3956_cast_fp16 = slice_by_index(begin = var_3956_begin_0, end = var_3956_end_0, end_mask = var_3956_end_mask_0, stride = var_3956_stride_0, x = q_15_cast_fp16)[name = string("op_3956_cast_fp16")];
            int32 var_3958 = const()[name = string("op_3958"), 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_3958, interleave = stacked_15_interleave_0, values = (var_3946_cast_fp16, var_3956_cast_fp16))[name = string("stacked_15_cast_fp16")];
            tensor<int32, [4]> var_3964_begin_0 = const()[name = string("op_3964_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3964_end_0 = const()[name = string("op_3964_end_0"), val = tensor<int32, [4]>([8, 8, 528, 128])];
            tensor<bool, [4]> var_3964_end_mask_0 = const()[name = string("op_3964_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3964_cast_fp16 = slice_by_index(begin = var_3964_begin_0, end = var_3964_end_0, end_mask = var_3964_end_mask_0, x = read_state_0)[name = string("op_3964_cast_fp16")];
            tensor<int32, [4]> var_3974_begin_0 = const()[name = string("op_3974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3974_end_0 = const()[name = string("op_3974_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3974_end_mask_0 = const()[name = string("op_3974_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3974_cast_fp16 = slice_by_index(begin = var_3974_begin_0, end = var_3974_end_0, end_mask = var_3974_end_mask_0, x = var_3964_cast_fp16)[name = string("op_3974_cast_fp16")];
            bool var_3983_transpose_x_1 = const()[name = string("op_3983_transpose_x_1"), val = bool(false)];
            bool var_3983_transpose_y_1 = const()[name = string("op_3983_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_3983_cast_fp16 = matmul(transpose_x = var_3983_transpose_x_1, transpose_y = var_3983_transpose_y_1, x = var_3974_cast_fp16, y = stacked_15_cast_fp16)[name = string("op_3983_cast_fp16")];
            tensor<int32, [4]> var_3986_perm_0 = const()[name = string("op_3986_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3987_to_fp16 = const()[name = string("op_3987_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_3986_cast_fp16 = transpose(perm = var_3986_perm_0, x = var_3983_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 8, 32, 512]> var_3988_cast_fp16 = mul(x = var_3986_cast_fp16, y = var_3987_to_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_15_cast_fp16 = add(x = var_3988_cast_fp16, y = cacheMask)[name = string("cache_weights_15_cast_fp16")];
            bool var_3994_transpose_x_1 = const()[name = string("op_3994_transpose_x_1"), val = bool(false)];
            bool var_3994_transpose_y_1 = const()[name = string("op_3994_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3994_cast_fp16 = matmul(transpose_x = var_3994_transpose_x_1, transpose_y = var_3994_transpose_y_1, x = stacked_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_3994_cast_fp16")];
            fp16 var_3995_to_fp16 = const()[name = string("op_3995_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3996_cast_fp16 = mul(x = var_3994_cast_fp16, y = var_3995_to_fp16)[name = string("op_3996_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_15_cast_fp16 = add(x = var_3996_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_15_cast_fp16")];
            int32 var_4000 = const()[name = string("op_4000"), val = int32(-1)];
            bool var_4001_interleave_0 = const()[name = string("op_4001_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4001_cast_fp16 = concat(axis = var_4000, interleave = var_4001_interleave_0, values = (cache_weights_15_cast_fp16, call_weights_15_cast_fp16))[name = string("op_4001_cast_fp16")];
            int32 var_4002 = const()[name = string("op_4002"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_15_cast_fp16 = softmax(axis = var_4002, x = var_4001_cast_fp16)[name = string("shares_15_cast_fp16")];
            tensor<int32, [4]> var_4009_begin_0 = const()[name = string("op_4009_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4009_end_0 = const()[name = string("op_4009_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4009_end_mask_0 = const()[name = string("op_4009_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4009_cast_fp16 = slice_by_index(begin = var_4009_begin_0, end = var_4009_end_0, end_mask = var_4009_end_mask_0, x = shares_15_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<int32, [4]> var_4014_begin_0 = const()[name = string("op_4014_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_4014_end_0 = const()[name = string("op_4014_end_0"), val = tensor<int32, [4]>([8, 8, 528, 128])];
            tensor<bool, [4]> var_4014_end_mask_0 = const()[name = string("op_4014_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4014_cast_fp16 = slice_by_index(begin = var_4014_begin_0, end = var_4014_end_0, end_mask = var_4014_end_mask_0, x = read_state_1)[name = string("op_4014_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, 8, 512, 128])];
            tensor<bool, [4]> var_4024_end_mask_0 = const()[name = string("op_4024_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 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, x = var_4014_cast_fp16)[name = string("op_4024_cast_fp16")];
            bool var_4030_transpose_x_0 = const()[name = string("op_4030_transpose_x_0"), val = bool(false)];
            bool var_4030_transpose_y_0 = const()[name = string("op_4030_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4030_cast_fp16 = matmul(transpose_x = var_4030_transpose_x_0, transpose_y = var_4030_transpose_y_0, x = var_4009_cast_fp16, y = var_4024_cast_fp16)[name = string("op_4030_cast_fp16")];
            tensor<int32, [4]> var_4035_begin_0 = const()[name = string("op_4035_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_4035_end_0 = const()[name = string("op_4035_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_4035_end_mask_0 = const()[name = string("op_4035_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4035_cast_fp16 = slice_by_index(begin = var_4035_begin_0, end = var_4035_end_0, end_mask = var_4035_end_mask_0, x = shares_15_cast_fp16)[name = string("op_4035_cast_fp16")];
            bool var_4036_transpose_x_0 = const()[name = string("op_4036_transpose_x_0"), val = bool(false)];
            bool var_4036_transpose_y_0 = const()[name = string("op_4036_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_3891_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 8, 32, 128]> var_4036_cast_fp16 = matmul(transpose_x = var_4036_transpose_x_0, transpose_y = var_4036_transpose_y_0, x = var_4035_cast_fp16, y = v_15_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_29_cast_fp16 = add(x = var_4030_cast_fp16, y = var_4036_cast_fp16)[name = string("mixed_29_cast_fp16")];
            tensor<int32, [4]> var_4053_begin_0 = const()[name = string("op_4053_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4053_end_0 = const()[name = string("op_4053_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4053_end_mask_0 = const()[name = string("op_4053_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4053_cast_fp16 = slice_by_index(begin = var_4053_begin_0, end = var_4053_end_0, end_mask = var_4053_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_4053_cast_fp16")];
            tensor<int32, [4]> var_4068_begin_0 = const()[name = string("op_4068_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4068_end_0 = const()[name = string("op_4068_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4068_end_mask_0 = const()[name = string("op_4068_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4068_cast_fp16 = slice_by_index(begin = var_4068_begin_0, end = var_4068_end_0, end_mask = var_4068_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_4068_cast_fp16")];
            int32 var_4070 = const()[name = string("op_4070"), 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_4070, interleave = mixed_31_interleave_0, values = (var_4053_cast_fp16, var_4068_cast_fp16))[name = string("mixed_31_cast_fp16")];
            tensor<int32, [4]> var_4074_perm_0 = const()[name = string("op_4074_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4078 = const()[name = string("op_4078"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4074_cast_fp16 = transpose(perm = var_4074_perm_0, x = mixed_31_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 16, 2048]> input_73_cast_fp16 = reshape(shape = var_4078, x = var_4074_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_157_cast_fp16 = add(x = x_141_cast_fp16, y = linear_52_cast_fp16)[name = string("x_157_cast_fp16")];
            bool var_4088 = const()[name = string("op_4088"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4091_cast_fp16 = abs(x = x_157_cast_fp16)[name = string("op_4091_cast_fp16")];
            tensor<int32, [1]> peak_63_axes_0 = const()[name = string("peak_63_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_63_cast_fp16 = reduce_max(axes = peak_63_axes_0, keep_dims = var_4088, x = var_4091_cast_fp16)[name = string("peak_63_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_159_cast_fp16 = real_div(x = x_157_cast_fp16, y = peak_63_cast_fp16)[name = string("x_159_cast_fp16")];
            fp16 var_4087_promoted_to_fp16 = const()[name = string("op_4087_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4095_cast_fp16 = pow(x = x_159_cast_fp16, y = var_4087_promoted_to_fp16)[name = string("op_4095_cast_fp16")];
            tensor<int32, [1]> var_4097_axes_0 = const()[name = string("op_4097_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4097_keep_dims_0 = const()[name = string("op_4097_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4097_cast_fp16 = reduce_mean(axes = var_4097_axes_0, keep_dims = var_4097_keep_dims_0, x = var_4095_cast_fp16)[name = string("op_4097_cast_fp16")];
            fp16 var_4087_promoted_1_to_fp16 = const()[name = string("op_4087_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4098_cast_fp16 = pow(x = peak_63_cast_fp16, y = var_4087_promoted_1_to_fp16)[name = string("op_4098_cast_fp16")];
            fp32 var_4099_epsilon_0 = const()[name = string("op_4099_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4099_cast_fp16 = inverse(epsilon = var_4099_epsilon_0, x = var_4098_cast_fp16)[name = string("op_4099_cast_fp16")];
            fp16 var_4100_to_fp16 = const()[name = string("op_4100_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4101_cast_fp16 = mul(x = var_4099_cast_fp16, y = var_4100_to_fp16)[name = string("op_4101_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4102_cast_fp16 = add(x = var_4097_cast_fp16, y = var_4101_cast_fp16)[name = string("op_4102_cast_fp16")];
            fp32 var_4103_epsilon_0 = const()[name = string("op_4103_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4103_cast_fp16 = rsqrt(epsilon = var_4103_epsilon_0, x = var_4102_cast_fp16)[name = string("op_4103_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4104_cast_fp16 = mul(x = x_159_cast_fp16, y = var_4103_cast_fp16)[name = string("op_4104_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_4104_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_4109_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_4109_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_4109_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_161_cast_fp16 = add(x = x_157_cast_fp16, y = linear_55_cast_fp16)[name = string("x_161_cast_fp16")];
            bool var_4122 = const()[name = string("op_4122"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4125_cast_fp16 = abs(x = x_161_cast_fp16)[name = string("op_4125_cast_fp16")];
            tensor<int32, [1]> peak_65_axes_0 = const()[name = string("peak_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_65_cast_fp16 = reduce_max(axes = peak_65_axes_0, keep_dims = var_4122, x = var_4125_cast_fp16)[name = string("peak_65_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_163_cast_fp16 = real_div(x = x_161_cast_fp16, y = peak_65_cast_fp16)[name = string("x_163_cast_fp16")];
            fp16 var_4121_promoted_to_fp16 = const()[name = string("op_4121_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4129_cast_fp16 = pow(x = x_163_cast_fp16, y = var_4121_promoted_to_fp16)[name = string("op_4129_cast_fp16")];
            tensor<int32, [1]> var_4131_axes_0 = const()[name = string("op_4131_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4131_keep_dims_0 = const()[name = string("op_4131_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4131_cast_fp16 = reduce_mean(axes = var_4131_axes_0, keep_dims = var_4131_keep_dims_0, x = var_4129_cast_fp16)[name = string("op_4131_cast_fp16")];
            fp16 var_4121_promoted_1_to_fp16 = const()[name = string("op_4121_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4132_cast_fp16 = pow(x = peak_65_cast_fp16, y = var_4121_promoted_1_to_fp16)[name = string("op_4132_cast_fp16")];
            fp32 var_4133_epsilon_0 = const()[name = string("op_4133_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4133_cast_fp16 = inverse(epsilon = var_4133_epsilon_0, x = var_4132_cast_fp16)[name = string("op_4133_cast_fp16")];
            fp16 var_4134_to_fp16 = const()[name = string("op_4134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4135_cast_fp16 = mul(x = var_4133_cast_fp16, y = var_4134_to_fp16)[name = string("op_4135_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4136_cast_fp16 = add(x = var_4131_cast_fp16, y = var_4135_cast_fp16)[name = string("op_4136_cast_fp16")];
            fp32 var_4137_epsilon_0 = const()[name = string("op_4137_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4137_cast_fp16 = rsqrt(epsilon = var_4137_epsilon_0, x = var_4136_cast_fp16)[name = string("op_4137_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4138_cast_fp16 = mul(x = x_163_cast_fp16, y = var_4137_cast_fp16)[name = string("op_4138_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_4138_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_4147 = const()[name = string("op_4147"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4148_cast_fp16 = reshape(shape = var_4147, x = linear_56_cast_fp16)[name = string("op_4148_cast_fp16")];
            tensor<int32, [4]> x_165_perm_0 = const()[name = string("x_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4155 = const()[name = string("op_4155"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_165_cast_fp16 = transpose(perm = x_165_perm_0, x = var_4148_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 16, 128]> var_4158_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_4158_cast_fp16")];
            tensor<int32, [1]> peak_67_axes_0 = const()[name = string("peak_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_67_cast_fp16 = reduce_max(axes = peak_67_axes_0, keep_dims = var_4155, x = var_4158_cast_fp16)[name = string("peak_67_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_167_cast_fp16 = real_div(x = x_165_cast_fp16, y = peak_67_cast_fp16)[name = string("x_167_cast_fp16")];
            fp16 var_4154_promoted_to_fp16 = const()[name = string("op_4154_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4162_cast_fp16 = pow(x = x_167_cast_fp16, y = var_4154_promoted_to_fp16)[name = string("op_4162_cast_fp16")];
            tensor<int32, [1]> var_4164_axes_0 = const()[name = string("op_4164_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4164_keep_dims_0 = const()[name = string("op_4164_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4164_cast_fp16 = reduce_mean(axes = var_4164_axes_0, keep_dims = var_4164_keep_dims_0, x = var_4162_cast_fp16)[name = string("op_4164_cast_fp16")];
            fp16 var_4154_promoted_1_to_fp16 = const()[name = string("op_4154_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4165_cast_fp16 = pow(x = peak_67_cast_fp16, y = var_4154_promoted_1_to_fp16)[name = string("op_4165_cast_fp16")];
            fp32 var_4166_epsilon_0 = const()[name = string("op_4166_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4166_cast_fp16 = inverse(epsilon = var_4166_epsilon_0, x = var_4165_cast_fp16)[name = string("op_4166_cast_fp16")];
            fp16 var_4167_to_fp16 = const()[name = string("op_4167_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4168_cast_fp16 = mul(x = var_4166_cast_fp16, y = var_4167_to_fp16)[name = string("op_4168_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4169_cast_fp16 = add(x = var_4164_cast_fp16, y = var_4168_cast_fp16)[name = string("op_4169_cast_fp16")];
            fp32 var_4170_epsilon_0 = const()[name = string("op_4170_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4170_cast_fp16 = rsqrt(epsilon = var_4170_epsilon_0, x = var_4169_cast_fp16)[name = string("op_4170_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4171_cast_fp16 = mul(x = x_167_cast_fp16, y = var_4170_cast_fp16)[name = string("op_4171_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_173_cast_fp16 = mul(x = var_4171_cast_fp16, y = model_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_173_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_4180 = const()[name = string("op_4180"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4181_cast_fp16 = reshape(shape = var_4180, x = linear_57_cast_fp16)[name = string("op_4181_cast_fp16")];
            tensor<int32, [4]> x_169_perm_0 = const()[name = string("x_169_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4188 = const()[name = string("op_4188"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_169_cast_fp16 = transpose(perm = x_169_perm_0, x = var_4181_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 8, 16, 128]> var_4191_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_4191_cast_fp16")];
            tensor<int32, [1]> peak_69_axes_0 = const()[name = string("peak_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_69_cast_fp16 = reduce_max(axes = peak_69_axes_0, keep_dims = var_4188, x = var_4191_cast_fp16)[name = string("peak_69_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_171_cast_fp16 = real_div(x = x_169_cast_fp16, y = peak_69_cast_fp16)[name = string("x_171_cast_fp16")];
            fp16 var_4187_promoted_to_fp16 = const()[name = string("op_4187_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4195_cast_fp16 = pow(x = x_171_cast_fp16, y = var_4187_promoted_to_fp16)[name = string("op_4195_cast_fp16")];
            tensor<int32, [1]> var_4197_axes_0 = const()[name = string("op_4197_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4197_keep_dims_0 = const()[name = string("op_4197_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4197_cast_fp16 = reduce_mean(axes = var_4197_axes_0, keep_dims = var_4197_keep_dims_0, x = var_4195_cast_fp16)[name = string("op_4197_cast_fp16")];
            fp16 var_4187_promoted_1_to_fp16 = const()[name = string("op_4187_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4198_cast_fp16 = pow(x = peak_69_cast_fp16, y = var_4187_promoted_1_to_fp16)[name = string("op_4198_cast_fp16")];
            fp32 var_4199_epsilon_0 = const()[name = string("op_4199_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4199_cast_fp16 = inverse(epsilon = var_4199_epsilon_0, x = var_4198_cast_fp16)[name = string("op_4199_cast_fp16")];
            fp16 var_4200_to_fp16 = const()[name = string("op_4200_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4201_cast_fp16 = mul(x = var_4199_cast_fp16, y = var_4200_to_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4202_cast_fp16 = add(x = var_4197_cast_fp16, y = var_4201_cast_fp16)[name = string("op_4202_cast_fp16")];
            fp32 var_4203_epsilon_0 = const()[name = string("op_4203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4203_cast_fp16 = rsqrt(epsilon = var_4203_epsilon_0, x = var_4202_cast_fp16)[name = string("op_4203_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4204_cast_fp16 = mul(x = x_171_cast_fp16, y = var_4203_cast_fp16)[name = string("op_4204_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_175_cast_fp16 = mul(x = var_4204_cast_fp16, y = model_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_175_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_4213 = const()[name = string("op_4213"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4214_cast_fp16 = reshape(shape = var_4213, x = linear_58_cast_fp16)[name = string("op_4214_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_173_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_173_cast_fp16)[name = string("second_33_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4228_cast_fp16 = mul(x = first_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4229_cast_fp16 = mul(x = second_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4229_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4231_cast_fp16 = sub(x = var_4228_cast_fp16, y = var_4229_cast_fp16)[name = string("op_4231_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4232_cast_fp16 = mul(x = second_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4233_cast_fp16 = mul(x = first_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4233_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4235_cast_fp16 = add(x = var_4232_cast_fp16, y = var_4233_cast_fp16)[name = string("op_4235_cast_fp16")];
            int32 var_4237 = const()[name = string("op_4237"), 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_4237, interleave = q_17_interleave_0, values = (var_4231_cast_fp16, var_4235_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_175_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_175_cast_fp16)[name = string("second_35_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4249_cast_fp16 = mul(x = first_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4249_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4250_cast_fp16 = mul(x = second_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4252_cast_fp16 = sub(x = var_4249_cast_fp16, y = var_4250_cast_fp16)[name = string("op_4252_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4253_cast_fp16 = mul(x = second_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4254_cast_fp16 = mul(x = first_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4256_cast_fp16 = add(x = var_4253_cast_fp16, y = var_4254_cast_fp16)[name = string("op_4256_cast_fp16")];
            int32 var_4258 = const()[name = string("op_4258"), 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_4258, interleave = k_17_interleave_0, values = (var_4252_cast_fp16, var_4256_cast_fp16))[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_4269_begin_0 = const()[name = string("op_4269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4269_end_0 = const()[name = string("op_4269_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4269_end_mask_0 = const()[name = string("op_4269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4269_stride_0 = const()[name = string("op_4269_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4269_cast_fp16 = slice_by_index(begin = var_4269_begin_0, end = var_4269_end_0, end_mask = var_4269_end_mask_0, stride = var_4269_stride_0, x = q_17_cast_fp16)[name = string("op_4269_cast_fp16")];
            tensor<int32, [4]> var_4279_begin_0 = const()[name = string("op_4279_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4279_end_0 = const()[name = string("op_4279_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4279_end_mask_0 = const()[name = string("op_4279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4279_stride_0 = const()[name = string("op_4279_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4279_cast_fp16 = slice_by_index(begin = var_4279_begin_0, end = var_4279_end_0, end_mask = var_4279_end_mask_0, stride = var_4279_stride_0, x = q_17_cast_fp16)[name = string("op_4279_cast_fp16")];
            int32 var_4281 = const()[name = string("op_4281"), 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_4281, interleave = stacked_17_interleave_0, values = (var_4269_cast_fp16, var_4279_cast_fp16))[name = string("stacked_17_cast_fp16")];
            tensor<int32, [4]> var_4287_begin_0 = const()[name = string("op_4287_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4287_end_0 = const()[name = string("op_4287_end_0"), val = tensor<int32, [4]>([9, 8, 528, 128])];
            tensor<bool, [4]> var_4287_end_mask_0 = const()[name = string("op_4287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4287_cast_fp16 = slice_by_index(begin = var_4287_begin_0, end = var_4287_end_0, end_mask = var_4287_end_mask_0, x = read_state_0)[name = string("op_4287_cast_fp16")];
            tensor<int32, [4]> var_4297_begin_0 = const()[name = string("op_4297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4297_end_0 = const()[name = string("op_4297_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4297_end_mask_0 = const()[name = string("op_4297_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4297_cast_fp16 = slice_by_index(begin = var_4297_begin_0, end = var_4297_end_0, end_mask = var_4297_end_mask_0, x = var_4287_cast_fp16)[name = string("op_4297_cast_fp16")];
            bool var_4306_transpose_x_1 = const()[name = string("op_4306_transpose_x_1"), val = bool(false)];
            bool var_4306_transpose_y_1 = const()[name = string("op_4306_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_4306_cast_fp16 = matmul(transpose_x = var_4306_transpose_x_1, transpose_y = var_4306_transpose_y_1, x = var_4297_cast_fp16, y = stacked_17_cast_fp16)[name = string("op_4306_cast_fp16")];
            tensor<int32, [4]> var_4309_perm_0 = const()[name = string("op_4309_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4310_to_fp16 = const()[name = string("op_4310_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_4309_cast_fp16 = transpose(perm = var_4309_perm_0, x = var_4306_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 32, 512]> var_4311_cast_fp16 = mul(x = var_4309_cast_fp16, y = var_4310_to_fp16)[name = string("op_4311_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_17_cast_fp16 = add(x = var_4311_cast_fp16, y = cacheMask)[name = string("cache_weights_17_cast_fp16")];
            bool var_4317_transpose_x_1 = const()[name = string("op_4317_transpose_x_1"), val = bool(false)];
            bool var_4317_transpose_y_1 = const()[name = string("op_4317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4317_cast_fp16 = matmul(transpose_x = var_4317_transpose_x_1, transpose_y = var_4317_transpose_y_1, x = stacked_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_4317_cast_fp16")];
            fp16 var_4318_to_fp16 = const()[name = string("op_4318_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4319_cast_fp16 = mul(x = var_4317_cast_fp16, y = var_4318_to_fp16)[name = string("op_4319_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_17_cast_fp16 = add(x = var_4319_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_17_cast_fp16")];
            int32 var_4323 = const()[name = string("op_4323"), val = int32(-1)];
            bool var_4324_interleave_0 = const()[name = string("op_4324_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4324_cast_fp16 = concat(axis = var_4323, interleave = var_4324_interleave_0, values = (cache_weights_17_cast_fp16, call_weights_17_cast_fp16))[name = string("op_4324_cast_fp16")];
            int32 var_4325 = const()[name = string("op_4325"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_17_cast_fp16 = softmax(axis = var_4325, x = var_4324_cast_fp16)[name = string("shares_17_cast_fp16")];
            tensor<int32, [4]> var_4332_begin_0 = const()[name = string("op_4332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4332_end_0 = const()[name = string("op_4332_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4332_end_mask_0 = const()[name = string("op_4332_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4332_cast_fp16 = slice_by_index(begin = var_4332_begin_0, end = var_4332_end_0, end_mask = var_4332_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4332_cast_fp16")];
            tensor<int32, [4]> var_4337_begin_0 = const()[name = string("op_4337_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4337_end_0 = const()[name = string("op_4337_end_0"), val = tensor<int32, [4]>([9, 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_1)[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_4353_transpose_x_0 = const()[name = string("op_4353_transpose_x_0"), val = bool(false)];
            bool var_4353_transpose_y_0 = const()[name = string("op_4353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4353_cast_fp16 = matmul(transpose_x = var_4353_transpose_x_0, transpose_y = var_4353_transpose_y_0, x = var_4332_cast_fp16, y = var_4347_cast_fp16)[name = string("op_4353_cast_fp16")];
            tensor<int32, [4]> var_4358_begin_0 = const()[name = string("op_4358_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_4358_end_0 = const()[name = string("op_4358_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_4358_end_mask_0 = const()[name = string("op_4358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4358_cast_fp16 = slice_by_index(begin = var_4358_begin_0, end = var_4358_end_0, end_mask = var_4358_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4358_cast_fp16")];
            bool var_4359_transpose_x_0 = const()[name = string("op_4359_transpose_x_0"), val = bool(false)];
            bool var_4359_transpose_y_0 = const()[name = string("op_4359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_4214_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 32, 128]> var_4359_cast_fp16 = matmul(transpose_x = var_4359_transpose_x_0, transpose_y = var_4359_transpose_y_0, x = var_4358_cast_fp16, y = v_17_cast_fp16)[name = string("op_4359_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_33_cast_fp16 = add(x = var_4353_cast_fp16, y = var_4359_cast_fp16)[name = string("mixed_33_cast_fp16")];
            tensor<int32, [4]> var_4376_begin_0 = const()[name = string("op_4376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4376_end_0 = const()[name = string("op_4376_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4376_end_mask_0 = const()[name = string("op_4376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4376_cast_fp16 = slice_by_index(begin = var_4376_begin_0, end = var_4376_end_0, end_mask = var_4376_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4376_cast_fp16")];
            tensor<int32, [4]> var_4391_begin_0 = const()[name = string("op_4391_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4391_end_0 = const()[name = string("op_4391_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4391_end_mask_0 = const()[name = string("op_4391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4391_cast_fp16 = slice_by_index(begin = var_4391_begin_0, end = var_4391_end_0, end_mask = var_4391_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4391_cast_fp16")];
            int32 var_4393 = const()[name = string("op_4393"), 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_4393, interleave = mixed_35_interleave_0, values = (var_4376_cast_fp16, var_4391_cast_fp16))[name = string("mixed_35_cast_fp16")];
            tensor<int32, [4]> var_4397_perm_0 = const()[name = string("op_4397_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4401 = const()[name = string("op_4401"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4397_cast_fp16 = transpose(perm = var_4397_perm_0, x = mixed_35_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 2048]> input_83_cast_fp16 = reshape(shape = var_4401, x = var_4397_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_177_cast_fp16 = add(x = x_161_cast_fp16, y = linear_59_cast_fp16)[name = string("x_177_cast_fp16")];
            bool var_4411 = const()[name = string("op_4411"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4414_cast_fp16 = abs(x = x_177_cast_fp16)[name = string("op_4414_cast_fp16")];
            tensor<int32, [1]> peak_71_axes_0 = const()[name = string("peak_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_71_cast_fp16 = reduce_max(axes = peak_71_axes_0, keep_dims = var_4411, x = var_4414_cast_fp16)[name = string("peak_71_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_179_cast_fp16 = real_div(x = x_177_cast_fp16, y = peak_71_cast_fp16)[name = string("x_179_cast_fp16")];
            fp16 var_4410_promoted_to_fp16 = const()[name = string("op_4410_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4418_cast_fp16 = pow(x = x_179_cast_fp16, y = var_4410_promoted_to_fp16)[name = string("op_4418_cast_fp16")];
            tensor<int32, [1]> var_4420_axes_0 = const()[name = string("op_4420_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4420_keep_dims_0 = const()[name = string("op_4420_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4420_cast_fp16 = reduce_mean(axes = var_4420_axes_0, keep_dims = var_4420_keep_dims_0, x = var_4418_cast_fp16)[name = string("op_4420_cast_fp16")];
            fp16 var_4410_promoted_1_to_fp16 = const()[name = string("op_4410_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4421_cast_fp16 = pow(x = peak_71_cast_fp16, y = var_4410_promoted_1_to_fp16)[name = string("op_4421_cast_fp16")];
            fp32 var_4422_epsilon_0 = const()[name = string("op_4422_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4422_cast_fp16 = inverse(epsilon = var_4422_epsilon_0, x = var_4421_cast_fp16)[name = string("op_4422_cast_fp16")];
            fp16 var_4423_to_fp16 = const()[name = string("op_4423_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4424_cast_fp16 = mul(x = var_4422_cast_fp16, y = var_4423_to_fp16)[name = string("op_4424_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4425_cast_fp16 = add(x = var_4420_cast_fp16, y = var_4424_cast_fp16)[name = string("op_4425_cast_fp16")];
            fp32 var_4426_epsilon_0 = const()[name = string("op_4426_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4426_cast_fp16 = rsqrt(epsilon = var_4426_epsilon_0, x = var_4425_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4427_cast_fp16 = mul(x = x_179_cast_fp16, y = var_4426_cast_fp16)[name = string("op_4427_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_4427_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_4432_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_4432_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_4432_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_181_cast_fp16 = add(x = x_177_cast_fp16, y = linear_62_cast_fp16)[name = string("x_181_cast_fp16")];
            bool var_4445 = const()[name = string("op_4445"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4448_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_4448_cast_fp16")];
            tensor<int32, [1]> peak_73_axes_0 = const()[name = string("peak_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_73_cast_fp16 = reduce_max(axes = peak_73_axes_0, keep_dims = var_4445, x = var_4448_cast_fp16)[name = string("peak_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_183_cast_fp16 = real_div(x = x_181_cast_fp16, y = peak_73_cast_fp16)[name = string("x_183_cast_fp16")];
            fp16 var_4444_promoted_to_fp16 = const()[name = string("op_4444_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4452_cast_fp16 = pow(x = x_183_cast_fp16, y = var_4444_promoted_to_fp16)[name = string("op_4452_cast_fp16")];
            tensor<int32, [1]> var_4454_axes_0 = const()[name = string("op_4454_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4454_keep_dims_0 = const()[name = string("op_4454_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4454_cast_fp16 = reduce_mean(axes = var_4454_axes_0, keep_dims = var_4454_keep_dims_0, x = var_4452_cast_fp16)[name = string("op_4454_cast_fp16")];
            fp16 var_4444_promoted_1_to_fp16 = const()[name = string("op_4444_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4455_cast_fp16 = pow(x = peak_73_cast_fp16, y = var_4444_promoted_1_to_fp16)[name = string("op_4455_cast_fp16")];
            fp32 var_4456_epsilon_0 = const()[name = string("op_4456_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4456_cast_fp16 = inverse(epsilon = var_4456_epsilon_0, x = var_4455_cast_fp16)[name = string("op_4456_cast_fp16")];
            fp16 var_4457_to_fp16 = const()[name = string("op_4457_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4458_cast_fp16 = mul(x = var_4456_cast_fp16, y = var_4457_to_fp16)[name = string("op_4458_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4459_cast_fp16 = add(x = var_4454_cast_fp16, y = var_4458_cast_fp16)[name = string("op_4459_cast_fp16")];
            fp32 var_4460_epsilon_0 = const()[name = string("op_4460_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4460_cast_fp16 = rsqrt(epsilon = var_4460_epsilon_0, x = var_4459_cast_fp16)[name = string("op_4460_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4461_cast_fp16 = mul(x = x_183_cast_fp16, y = var_4460_cast_fp16)[name = string("op_4461_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_4461_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_4470 = const()[name = string("op_4470"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4471_cast_fp16 = reshape(shape = var_4470, x = linear_63_cast_fp16)[name = string("op_4471_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4478 = const()[name = string("op_4478"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_4471_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 16, 128]> var_4481_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_4481_cast_fp16")];
            tensor<int32, [1]> peak_75_axes_0 = const()[name = string("peak_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_75_cast_fp16 = reduce_max(axes = peak_75_axes_0, keep_dims = var_4478, x = var_4481_cast_fp16)[name = string("peak_75_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_187_cast_fp16 = real_div(x = x_185_cast_fp16, y = peak_75_cast_fp16)[name = string("x_187_cast_fp16")];
            fp16 var_4477_promoted_to_fp16 = const()[name = string("op_4477_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4485_cast_fp16 = pow(x = x_187_cast_fp16, y = var_4477_promoted_to_fp16)[name = string("op_4485_cast_fp16")];
            tensor<int32, [1]> var_4487_axes_0 = const()[name = string("op_4487_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4487_keep_dims_0 = const()[name = string("op_4487_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4487_cast_fp16 = reduce_mean(axes = var_4487_axes_0, keep_dims = var_4487_keep_dims_0, x = var_4485_cast_fp16)[name = string("op_4487_cast_fp16")];
            fp16 var_4477_promoted_1_to_fp16 = const()[name = string("op_4477_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4488_cast_fp16 = pow(x = peak_75_cast_fp16, y = var_4477_promoted_1_to_fp16)[name = string("op_4488_cast_fp16")];
            fp32 var_4489_epsilon_0 = const()[name = string("op_4489_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4489_cast_fp16 = inverse(epsilon = var_4489_epsilon_0, x = var_4488_cast_fp16)[name = string("op_4489_cast_fp16")];
            fp16 var_4490_to_fp16 = const()[name = string("op_4490_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4491_cast_fp16 = mul(x = var_4489_cast_fp16, y = var_4490_to_fp16)[name = string("op_4491_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4492_cast_fp16 = add(x = var_4487_cast_fp16, y = var_4491_cast_fp16)[name = string("op_4492_cast_fp16")];
            fp32 var_4493_epsilon_0 = const()[name = string("op_4493_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4493_cast_fp16 = rsqrt(epsilon = var_4493_epsilon_0, x = var_4492_cast_fp16)[name = string("op_4493_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4494_cast_fp16 = mul(x = x_187_cast_fp16, y = var_4493_cast_fp16)[name = string("op_4494_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_193_cast_fp16 = mul(x = var_4494_cast_fp16, y = model_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_193_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_4503 = const()[name = string("op_4503"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4504_cast_fp16 = reshape(shape = var_4503, x = linear_64_cast_fp16)[name = string("op_4504_cast_fp16")];
            tensor<int32, [4]> x_189_perm_0 = const()[name = string("x_189_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4511 = const()[name = string("op_4511"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_189_cast_fp16 = transpose(perm = x_189_perm_0, x = var_4504_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 16, 128]> var_4514_cast_fp16 = abs(x = x_189_cast_fp16)[name = string("op_4514_cast_fp16")];
            tensor<int32, [1]> peak_77_axes_0 = const()[name = string("peak_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_77_cast_fp16 = reduce_max(axes = peak_77_axes_0, keep_dims = var_4511, x = var_4514_cast_fp16)[name = string("peak_77_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_191_cast_fp16 = real_div(x = x_189_cast_fp16, y = peak_77_cast_fp16)[name = string("x_191_cast_fp16")];
            fp16 var_4510_promoted_to_fp16 = const()[name = string("op_4510_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4518_cast_fp16 = pow(x = x_191_cast_fp16, y = var_4510_promoted_to_fp16)[name = string("op_4518_cast_fp16")];
            tensor<int32, [1]> var_4520_axes_0 = const()[name = string("op_4520_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4520_keep_dims_0 = const()[name = string("op_4520_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4520_cast_fp16 = reduce_mean(axes = var_4520_axes_0, keep_dims = var_4520_keep_dims_0, x = var_4518_cast_fp16)[name = string("op_4520_cast_fp16")];
            fp16 var_4510_promoted_1_to_fp16 = const()[name = string("op_4510_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4521_cast_fp16 = pow(x = peak_77_cast_fp16, y = var_4510_promoted_1_to_fp16)[name = string("op_4521_cast_fp16")];
            fp32 var_4522_epsilon_0 = const()[name = string("op_4522_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4522_cast_fp16 = inverse(epsilon = var_4522_epsilon_0, x = var_4521_cast_fp16)[name = string("op_4522_cast_fp16")];
            fp16 var_4523_to_fp16 = const()[name = string("op_4523_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4524_cast_fp16 = mul(x = var_4522_cast_fp16, y = var_4523_to_fp16)[name = string("op_4524_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4525_cast_fp16 = add(x = var_4520_cast_fp16, y = var_4524_cast_fp16)[name = string("op_4525_cast_fp16")];
            fp32 var_4526_epsilon_0 = const()[name = string("op_4526_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4526_cast_fp16 = rsqrt(epsilon = var_4526_epsilon_0, x = var_4525_cast_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4527_cast_fp16 = mul(x = x_191_cast_fp16, y = var_4526_cast_fp16)[name = string("op_4527_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_195_cast_fp16 = mul(x = var_4527_cast_fp16, y = model_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_195_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_4536 = const()[name = string("op_4536"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4537_cast_fp16 = reshape(shape = var_4536, x = linear_65_cast_fp16)[name = string("op_4537_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_193_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_193_cast_fp16)[name = string("second_37_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4551_cast_fp16 = mul(x = first_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4551_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4552_cast_fp16 = mul(x = second_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4552_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4554_cast_fp16 = sub(x = var_4551_cast_fp16, y = var_4552_cast_fp16)[name = string("op_4554_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4555_cast_fp16 = mul(x = second_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4555_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4556_cast_fp16 = mul(x = first_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4556_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4558_cast_fp16 = add(x = var_4555_cast_fp16, y = var_4556_cast_fp16)[name = string("op_4558_cast_fp16")];
            int32 var_4560 = const()[name = string("op_4560"), 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_4560, interleave = q_19_interleave_0, values = (var_4554_cast_fp16, var_4558_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_195_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_195_cast_fp16)[name = string("second_39_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4572_cast_fp16 = mul(x = first_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4572_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4573_cast_fp16 = mul(x = second_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4573_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4575_cast_fp16 = sub(x = var_4572_cast_fp16, y = var_4573_cast_fp16)[name = string("op_4575_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4576_cast_fp16 = mul(x = second_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4576_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4577_cast_fp16 = mul(x = first_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4579_cast_fp16 = add(x = var_4576_cast_fp16, y = var_4577_cast_fp16)[name = string("op_4579_cast_fp16")];
            int32 var_4581 = const()[name = string("op_4581"), 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_4581, interleave = k_19_interleave_0, values = (var_4575_cast_fp16, var_4579_cast_fp16))[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_4592_begin_0 = const()[name = string("op_4592_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4592_end_0 = const()[name = string("op_4592_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4592_end_mask_0 = const()[name = string("op_4592_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4592_stride_0 = const()[name = string("op_4592_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4592_cast_fp16 = slice_by_index(begin = var_4592_begin_0, end = var_4592_end_0, end_mask = var_4592_end_mask_0, stride = var_4592_stride_0, x = q_19_cast_fp16)[name = string("op_4592_cast_fp16")];
            tensor<int32, [4]> var_4602_begin_0 = const()[name = string("op_4602_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4602_end_0 = const()[name = string("op_4602_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4602_end_mask_0 = const()[name = string("op_4602_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4602_stride_0 = const()[name = string("op_4602_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4602_cast_fp16 = slice_by_index(begin = var_4602_begin_0, end = var_4602_end_0, end_mask = var_4602_end_mask_0, stride = var_4602_stride_0, x = q_19_cast_fp16)[name = string("op_4602_cast_fp16")];
            int32 var_4604 = const()[name = string("op_4604"), 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_4604, interleave = stacked_19_interleave_0, values = (var_4592_cast_fp16, var_4602_cast_fp16))[name = string("stacked_19_cast_fp16")];
            tensor<int32, [4]> var_4610_begin_0 = const()[name = string("op_4610_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4610_end_0 = const()[name = string("op_4610_end_0"), val = tensor<int32, [4]>([10, 8, 528, 128])];
            tensor<bool, [4]> var_4610_end_mask_0 = const()[name = string("op_4610_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4610_cast_fp16 = slice_by_index(begin = var_4610_begin_0, end = var_4610_end_0, end_mask = var_4610_end_mask_0, x = read_state_0)[name = string("op_4610_cast_fp16")];
            tensor<int32, [4]> var_4620_begin_0 = const()[name = string("op_4620_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4620_end_0 = const()[name = string("op_4620_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4620_end_mask_0 = const()[name = string("op_4620_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4620_cast_fp16 = slice_by_index(begin = var_4620_begin_0, end = var_4620_end_0, end_mask = var_4620_end_mask_0, x = var_4610_cast_fp16)[name = string("op_4620_cast_fp16")];
            bool var_4629_transpose_x_1 = const()[name = string("op_4629_transpose_x_1"), val = bool(false)];
            bool var_4629_transpose_y_1 = const()[name = string("op_4629_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_4629_cast_fp16 = matmul(transpose_x = var_4629_transpose_x_1, transpose_y = var_4629_transpose_y_1, x = var_4620_cast_fp16, y = stacked_19_cast_fp16)[name = string("op_4629_cast_fp16")];
            tensor<int32, [4]> var_4632_perm_0 = const()[name = string("op_4632_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4633_to_fp16 = const()[name = string("op_4633_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_4632_cast_fp16 = transpose(perm = var_4632_perm_0, x = var_4629_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 8, 32, 512]> var_4634_cast_fp16 = mul(x = var_4632_cast_fp16, y = var_4633_to_fp16)[name = string("op_4634_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_19_cast_fp16 = add(x = var_4634_cast_fp16, y = cacheMask)[name = string("cache_weights_19_cast_fp16")];
            bool var_4640_transpose_x_1 = const()[name = string("op_4640_transpose_x_1"), val = bool(false)];
            bool var_4640_transpose_y_1 = const()[name = string("op_4640_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4640_cast_fp16 = matmul(transpose_x = var_4640_transpose_x_1, transpose_y = var_4640_transpose_y_1, x = stacked_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_4640_cast_fp16")];
            fp16 var_4641_to_fp16 = const()[name = string("op_4641_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4642_cast_fp16 = mul(x = var_4640_cast_fp16, y = var_4641_to_fp16)[name = string("op_4642_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_19_cast_fp16 = add(x = var_4642_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_19_cast_fp16")];
            int32 var_4646 = const()[name = string("op_4646"), val = int32(-1)];
            bool var_4647_interleave_0 = const()[name = string("op_4647_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4647_cast_fp16 = concat(axis = var_4646, interleave = var_4647_interleave_0, values = (cache_weights_19_cast_fp16, call_weights_19_cast_fp16))[name = string("op_4647_cast_fp16")];
            int32 var_4648 = const()[name = string("op_4648"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_19_cast_fp16 = softmax(axis = var_4648, x = var_4647_cast_fp16)[name = string("shares_19_cast_fp16")];
            tensor<int32, [4]> var_4655_begin_0 = const()[name = string("op_4655_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4655_end_0 = const()[name = string("op_4655_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4655_end_mask_0 = const()[name = string("op_4655_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4655_cast_fp16 = slice_by_index(begin = var_4655_begin_0, end = var_4655_end_0, end_mask = var_4655_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4655_cast_fp16")];
            tensor<int32, [4]> var_4660_begin_0 = const()[name = string("op_4660_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4660_end_0 = const()[name = string("op_4660_end_0"), val = tensor<int32, [4]>([10, 8, 528, 128])];
            tensor<bool, [4]> var_4660_end_mask_0 = const()[name = string("op_4660_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4660_cast_fp16 = slice_by_index(begin = var_4660_begin_0, end = var_4660_end_0, end_mask = var_4660_end_mask_0, x = read_state_1)[name = string("op_4660_cast_fp16")];
            tensor<int32, [4]> var_4670_begin_0 = const()[name = string("op_4670_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4670_end_0 = const()[name = string("op_4670_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4670_end_mask_0 = const()[name = string("op_4670_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4670_cast_fp16 = slice_by_index(begin = var_4670_begin_0, end = var_4670_end_0, end_mask = var_4670_end_mask_0, x = var_4660_cast_fp16)[name = string("op_4670_cast_fp16")];
            bool var_4676_transpose_x_0 = const()[name = string("op_4676_transpose_x_0"), val = bool(false)];
            bool var_4676_transpose_y_0 = const()[name = string("op_4676_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4676_cast_fp16 = matmul(transpose_x = var_4676_transpose_x_0, transpose_y = var_4676_transpose_y_0, x = var_4655_cast_fp16, y = var_4670_cast_fp16)[name = string("op_4676_cast_fp16")];
            tensor<int32, [4]> var_4681_begin_0 = const()[name = string("op_4681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_4681_end_0 = const()[name = string("op_4681_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_4681_end_mask_0 = const()[name = string("op_4681_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4681_cast_fp16 = slice_by_index(begin = var_4681_begin_0, end = var_4681_end_0, end_mask = var_4681_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4681_cast_fp16")];
            bool var_4682_transpose_x_0 = const()[name = string("op_4682_transpose_x_0"), val = bool(false)];
            bool var_4682_transpose_y_0 = const()[name = string("op_4682_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_4537_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 8, 32, 128]> var_4682_cast_fp16 = matmul(transpose_x = var_4682_transpose_x_0, transpose_y = var_4682_transpose_y_0, x = var_4681_cast_fp16, y = v_19_cast_fp16)[name = string("op_4682_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_37_cast_fp16 = add(x = var_4676_cast_fp16, y = var_4682_cast_fp16)[name = string("mixed_37_cast_fp16")];
            tensor<int32, [4]> var_4699_begin_0 = const()[name = string("op_4699_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4699_end_0 = const()[name = string("op_4699_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4699_end_mask_0 = const()[name = string("op_4699_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4699_cast_fp16 = slice_by_index(begin = var_4699_begin_0, end = var_4699_end_0, end_mask = var_4699_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<int32, [4]> var_4714_begin_0 = const()[name = string("op_4714_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4714_end_0 = const()[name = string("op_4714_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4714_end_mask_0 = const()[name = string("op_4714_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4714_cast_fp16 = slice_by_index(begin = var_4714_begin_0, end = var_4714_end_0, end_mask = var_4714_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4714_cast_fp16")];
            int32 var_4716 = const()[name = string("op_4716"), 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_4716, interleave = mixed_39_interleave_0, values = (var_4699_cast_fp16, var_4714_cast_fp16))[name = string("mixed_39_cast_fp16")];
            tensor<int32, [4]> var_4720_perm_0 = const()[name = string("op_4720_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4724 = const()[name = string("op_4724"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4720_cast_fp16 = transpose(perm = var_4720_perm_0, x = mixed_39_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 2048]> input_93_cast_fp16 = reshape(shape = var_4724, x = var_4720_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_197_cast_fp16 = add(x = x_181_cast_fp16, y = linear_66_cast_fp16)[name = string("x_197_cast_fp16")];
            bool var_4734 = const()[name = string("op_4734"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4737_cast_fp16 = abs(x = x_197_cast_fp16)[name = string("op_4737_cast_fp16")];
            tensor<int32, [1]> peak_79_axes_0 = const()[name = string("peak_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_79_cast_fp16 = reduce_max(axes = peak_79_axes_0, keep_dims = var_4734, x = var_4737_cast_fp16)[name = string("peak_79_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_199_cast_fp16 = real_div(x = x_197_cast_fp16, y = peak_79_cast_fp16)[name = string("x_199_cast_fp16")];
            fp16 var_4733_promoted_to_fp16 = const()[name = string("op_4733_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4741_cast_fp16 = pow(x = x_199_cast_fp16, y = var_4733_promoted_to_fp16)[name = string("op_4741_cast_fp16")];
            tensor<int32, [1]> var_4743_axes_0 = const()[name = string("op_4743_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4743_keep_dims_0 = const()[name = string("op_4743_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4743_cast_fp16 = reduce_mean(axes = var_4743_axes_0, keep_dims = var_4743_keep_dims_0, x = var_4741_cast_fp16)[name = string("op_4743_cast_fp16")];
            fp16 var_4733_promoted_1_to_fp16 = const()[name = string("op_4733_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4744_cast_fp16 = pow(x = peak_79_cast_fp16, y = var_4733_promoted_1_to_fp16)[name = string("op_4744_cast_fp16")];
            fp32 var_4745_epsilon_0 = const()[name = string("op_4745_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4745_cast_fp16 = inverse(epsilon = var_4745_epsilon_0, x = var_4744_cast_fp16)[name = string("op_4745_cast_fp16")];
            fp16 var_4746_to_fp16 = const()[name = string("op_4746_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4747_cast_fp16 = mul(x = var_4745_cast_fp16, y = var_4746_to_fp16)[name = string("op_4747_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4748_cast_fp16 = add(x = var_4743_cast_fp16, y = var_4747_cast_fp16)[name = string("op_4748_cast_fp16")];
            fp32 var_4749_epsilon_0 = const()[name = string("op_4749_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4749_cast_fp16 = rsqrt(epsilon = var_4749_epsilon_0, x = var_4748_cast_fp16)[name = string("op_4749_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4750_cast_fp16 = mul(x = x_199_cast_fp16, y = var_4749_cast_fp16)[name = string("op_4750_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_4750_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_4755_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_4755_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_4755_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_201_cast_fp16 = add(x = x_197_cast_fp16, y = linear_69_cast_fp16)[name = string("x_201_cast_fp16")];
            bool var_4768 = const()[name = string("op_4768"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_4771_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_4771_cast_fp16")];
            tensor<int32, [1]> peak_81_axes_0 = const()[name = string("peak_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_81_cast_fp16 = reduce_max(axes = peak_81_axes_0, keep_dims = var_4768, x = var_4771_cast_fp16)[name = string("peak_81_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_203_cast_fp16 = real_div(x = x_201_cast_fp16, y = peak_81_cast_fp16)[name = string("x_203_cast_fp16")];
            fp16 var_4767_promoted_to_fp16 = const()[name = string("op_4767_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4775_cast_fp16 = pow(x = x_203_cast_fp16, y = var_4767_promoted_to_fp16)[name = string("op_4775_cast_fp16")];
            tensor<int32, [1]> var_4777_axes_0 = const()[name = string("op_4777_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4777_keep_dims_0 = const()[name = string("op_4777_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4777_cast_fp16 = reduce_mean(axes = var_4777_axes_0, keep_dims = var_4777_keep_dims_0, x = var_4775_cast_fp16)[name = string("op_4777_cast_fp16")];
            fp16 var_4767_promoted_1_to_fp16 = const()[name = string("op_4767_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_4778_cast_fp16 = pow(x = peak_81_cast_fp16, y = var_4767_promoted_1_to_fp16)[name = string("op_4778_cast_fp16")];
            fp32 var_4779_epsilon_0 = const()[name = string("op_4779_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_4779_cast_fp16 = inverse(epsilon = var_4779_epsilon_0, x = var_4778_cast_fp16)[name = string("op_4779_cast_fp16")];
            fp16 var_4780_to_fp16 = const()[name = string("op_4780_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4781_cast_fp16 = mul(x = var_4779_cast_fp16, y = var_4780_to_fp16)[name = string("op_4781_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_4782_cast_fp16 = add(x = var_4777_cast_fp16, y = var_4781_cast_fp16)[name = string("op_4782_cast_fp16")];
            fp32 var_4783_epsilon_0 = const()[name = string("op_4783_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4783_cast_fp16 = rsqrt(epsilon = var_4783_epsilon_0, x = var_4782_cast_fp16)[name = string("op_4783_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4784_cast_fp16 = mul(x = x_203_cast_fp16, y = var_4783_cast_fp16)[name = string("op_4784_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_4784_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_4793 = const()[name = string("op_4793"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4794_cast_fp16 = reshape(shape = var_4793, x = linear_70_cast_fp16)[name = string("op_4794_cast_fp16")];
            tensor<int32, [4]> x_205_perm_0 = const()[name = string("x_205_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4801 = const()[name = string("op_4801"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_205_cast_fp16 = transpose(perm = x_205_perm_0, x = var_4794_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 16, 128]> var_4804_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_4804_cast_fp16")];
            tensor<int32, [1]> peak_83_axes_0 = const()[name = string("peak_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_83_cast_fp16 = reduce_max(axes = peak_83_axes_0, keep_dims = var_4801, x = var_4804_cast_fp16)[name = string("peak_83_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_207_cast_fp16 = real_div(x = x_205_cast_fp16, y = peak_83_cast_fp16)[name = string("x_207_cast_fp16")];
            fp16 var_4800_promoted_to_fp16 = const()[name = string("op_4800_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_4808_cast_fp16 = pow(x = x_207_cast_fp16, y = var_4800_promoted_to_fp16)[name = string("op_4808_cast_fp16")];
            tensor<int32, [1]> var_4810_axes_0 = const()[name = string("op_4810_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4810_keep_dims_0 = const()[name = string("op_4810_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4810_cast_fp16 = reduce_mean(axes = var_4810_axes_0, keep_dims = var_4810_keep_dims_0, x = var_4808_cast_fp16)[name = string("op_4810_cast_fp16")];
            fp16 var_4800_promoted_1_to_fp16 = const()[name = string("op_4800_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_4811_cast_fp16 = pow(x = peak_83_cast_fp16, y = var_4800_promoted_1_to_fp16)[name = string("op_4811_cast_fp16")];
            fp32 var_4812_epsilon_0 = const()[name = string("op_4812_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_4812_cast_fp16 = inverse(epsilon = var_4812_epsilon_0, x = var_4811_cast_fp16)[name = string("op_4812_cast_fp16")];
            fp16 var_4813_to_fp16 = const()[name = string("op_4813_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4814_cast_fp16 = mul(x = var_4812_cast_fp16, y = var_4813_to_fp16)[name = string("op_4814_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_4815_cast_fp16 = add(x = var_4810_cast_fp16, y = var_4814_cast_fp16)[name = string("op_4815_cast_fp16")];
            fp32 var_4816_epsilon_0 = const()[name = string("op_4816_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4816_cast_fp16 = rsqrt(epsilon = var_4816_epsilon_0, x = var_4815_cast_fp16)[name = string("op_4816_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4817_cast_fp16 = mul(x = x_207_cast_fp16, y = var_4816_cast_fp16)[name = string("op_4817_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_213_cast_fp16 = mul(x = var_4817_cast_fp16, y = model_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_213_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_4826 = const()[name = string("op_4826"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4827_cast_fp16 = reshape(shape = var_4826, x = linear_71_cast_fp16)[name = string("op_4827_cast_fp16")];
            tensor<int32, [4]> x_209_perm_0 = const()[name = string("x_209_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_4834 = const()[name = string("op_4834"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_209_cast_fp16 = transpose(perm = x_209_perm_0, x = var_4827_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 8, 16, 128]> var_4837_cast_fp16 = abs(x = x_209_cast_fp16)[name = string("op_4837_cast_fp16")];
            tensor<int32, [1]> peak_85_axes_0 = const()[name = string("peak_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_85_cast_fp16 = reduce_max(axes = peak_85_axes_0, keep_dims = var_4834, x = var_4837_cast_fp16)[name = string("peak_85_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_211_cast_fp16 = real_div(x = x_209_cast_fp16, y = peak_85_cast_fp16)[name = string("x_211_cast_fp16")];
            fp16 var_4833_promoted_to_fp16 = const()[name = string("op_4833_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_4841_cast_fp16 = pow(x = x_211_cast_fp16, y = var_4833_promoted_to_fp16)[name = string("op_4841_cast_fp16")];
            tensor<int32, [1]> var_4843_axes_0 = const()[name = string("op_4843_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4843_keep_dims_0 = const()[name = string("op_4843_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4843_cast_fp16 = reduce_mean(axes = var_4843_axes_0, keep_dims = var_4843_keep_dims_0, x = var_4841_cast_fp16)[name = string("op_4843_cast_fp16")];
            fp16 var_4833_promoted_1_to_fp16 = const()[name = string("op_4833_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_4844_cast_fp16 = pow(x = peak_85_cast_fp16, y = var_4833_promoted_1_to_fp16)[name = string("op_4844_cast_fp16")];
            fp32 var_4845_epsilon_0 = const()[name = string("op_4845_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_4845_cast_fp16 = inverse(epsilon = var_4845_epsilon_0, x = var_4844_cast_fp16)[name = string("op_4845_cast_fp16")];
            fp16 var_4846_to_fp16 = const()[name = string("op_4846_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4847_cast_fp16 = mul(x = var_4845_cast_fp16, y = var_4846_to_fp16)[name = string("op_4847_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_4848_cast_fp16 = add(x = var_4843_cast_fp16, y = var_4847_cast_fp16)[name = string("op_4848_cast_fp16")];
            fp32 var_4849_epsilon_0 = const()[name = string("op_4849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4849_cast_fp16 = rsqrt(epsilon = var_4849_epsilon_0, x = var_4848_cast_fp16)[name = string("op_4849_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4850_cast_fp16 = mul(x = x_211_cast_fp16, y = var_4849_cast_fp16)[name = string("op_4850_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_215_cast_fp16 = mul(x = var_4850_cast_fp16, y = model_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_215_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_4859 = const()[name = string("op_4859"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4860_cast_fp16 = reshape(shape = var_4859, x = linear_72_cast_fp16)[name = string("op_4860_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_213_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_213_cast_fp16)[name = string("second_41_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4874_cast_fp16 = mul(x = first_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4874_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4875_cast_fp16 = mul(x = second_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4875_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4877_cast_fp16 = sub(x = var_4874_cast_fp16, y = var_4875_cast_fp16)[name = string("op_4877_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4878_cast_fp16 = mul(x = second_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4878_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4879_cast_fp16 = mul(x = first_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4879_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4881_cast_fp16 = add(x = var_4878_cast_fp16, y = var_4879_cast_fp16)[name = string("op_4881_cast_fp16")];
            int32 var_4883 = const()[name = string("op_4883"), 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_4883, interleave = q_21_interleave_0, values = (var_4877_cast_fp16, var_4881_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_215_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_215_cast_fp16)[name = string("second_43_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4895_cast_fp16 = mul(x = first_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4895_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4896_cast_fp16 = mul(x = second_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4896_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4898_cast_fp16 = sub(x = var_4895_cast_fp16, y = var_4896_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4899_cast_fp16 = mul(x = second_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4899_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4900_cast_fp16 = mul(x = first_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4900_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4902_cast_fp16 = add(x = var_4899_cast_fp16, y = var_4900_cast_fp16)[name = string("op_4902_cast_fp16")];
            int32 var_4904 = const()[name = string("op_4904"), 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_4904, interleave = k_21_interleave_0, values = (var_4898_cast_fp16, var_4902_cast_fp16))[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_4915_begin_0 = const()[name = string("op_4915_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4915_end_0 = const()[name = string("op_4915_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4915_end_mask_0 = const()[name = string("op_4915_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4915_stride_0 = const()[name = string("op_4915_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4915_cast_fp16 = slice_by_index(begin = var_4915_begin_0, end = var_4915_end_0, end_mask = var_4915_end_mask_0, stride = var_4915_stride_0, x = q_21_cast_fp16)[name = string("op_4915_cast_fp16")];
            tensor<int32, [4]> var_4925_begin_0 = const()[name = string("op_4925_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4925_end_0 = const()[name = string("op_4925_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4925_end_mask_0 = const()[name = string("op_4925_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4925_stride_0 = const()[name = string("op_4925_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4925_cast_fp16 = slice_by_index(begin = var_4925_begin_0, end = var_4925_end_0, end_mask = var_4925_end_mask_0, stride = var_4925_stride_0, x = q_21_cast_fp16)[name = string("op_4925_cast_fp16")];
            int32 var_4927 = const()[name = string("op_4927"), 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_4927, interleave = stacked_21_interleave_0, values = (var_4915_cast_fp16, var_4925_cast_fp16))[name = string("stacked_21_cast_fp16")];
            tensor<int32, [4]> var_4933_begin_0 = const()[name = string("op_4933_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4933_end_0 = const()[name = string("op_4933_end_0"), val = tensor<int32, [4]>([11, 8, 528, 128])];
            tensor<bool, [4]> var_4933_end_mask_0 = const()[name = string("op_4933_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4933_cast_fp16 = slice_by_index(begin = var_4933_begin_0, end = var_4933_end_0, end_mask = var_4933_end_mask_0, x = read_state_0)[name = string("op_4933_cast_fp16")];
            tensor<int32, [4]> var_4943_begin_0 = const()[name = string("op_4943_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4943_end_0 = const()[name = string("op_4943_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4943_end_mask_0 = const()[name = string("op_4943_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4943_cast_fp16 = slice_by_index(begin = var_4943_begin_0, end = var_4943_end_0, end_mask = var_4943_end_mask_0, x = var_4933_cast_fp16)[name = string("op_4943_cast_fp16")];
            bool var_4952_transpose_x_1 = const()[name = string("op_4952_transpose_x_1"), val = bool(false)];
            bool var_4952_transpose_y_1 = const()[name = string("op_4952_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_4952_cast_fp16 = matmul(transpose_x = var_4952_transpose_x_1, transpose_y = var_4952_transpose_y_1, x = var_4943_cast_fp16, y = stacked_21_cast_fp16)[name = string("op_4952_cast_fp16")];
            tensor<int32, [4]> var_4955_perm_0 = const()[name = string("op_4955_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4956_to_fp16 = const()[name = string("op_4956_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_4955_cast_fp16 = transpose(perm = var_4955_perm_0, x = var_4952_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 32, 512]> var_4957_cast_fp16 = mul(x = var_4955_cast_fp16, y = var_4956_to_fp16)[name = string("op_4957_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_21_cast_fp16 = add(x = var_4957_cast_fp16, y = cacheMask)[name = string("cache_weights_21_cast_fp16")];
            bool var_4963_transpose_x_1 = const()[name = string("op_4963_transpose_x_1"), val = bool(false)];
            bool var_4963_transpose_y_1 = const()[name = string("op_4963_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4963_cast_fp16 = matmul(transpose_x = var_4963_transpose_x_1, transpose_y = var_4963_transpose_y_1, x = stacked_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_4963_cast_fp16")];
            fp16 var_4964_to_fp16 = const()[name = string("op_4964_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4965_cast_fp16 = mul(x = var_4963_cast_fp16, y = var_4964_to_fp16)[name = string("op_4965_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_21_cast_fp16 = add(x = var_4965_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_21_cast_fp16")];
            int32 var_4969 = const()[name = string("op_4969"), val = int32(-1)];
            bool var_4970_interleave_0 = const()[name = string("op_4970_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4970_cast_fp16 = concat(axis = var_4969, interleave = var_4970_interleave_0, values = (cache_weights_21_cast_fp16, call_weights_21_cast_fp16))[name = string("op_4970_cast_fp16")];
            int32 var_4971 = const()[name = string("op_4971"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_21_cast_fp16 = softmax(axis = var_4971, x = var_4970_cast_fp16)[name = string("shares_21_cast_fp16")];
            tensor<int32, [4]> var_4978_begin_0 = const()[name = string("op_4978_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4978_end_0 = const()[name = string("op_4978_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4978_end_mask_0 = const()[name = string("op_4978_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4978_cast_fp16 = slice_by_index(begin = var_4978_begin_0, end = var_4978_end_0, end_mask = var_4978_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4978_cast_fp16")];
            tensor<int32, [4]> var_4983_begin_0 = const()[name = string("op_4983_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4983_end_0 = const()[name = string("op_4983_end_0"), val = tensor<int32, [4]>([11, 8, 528, 128])];
            tensor<bool, [4]> var_4983_end_mask_0 = const()[name = string("op_4983_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4983_cast_fp16 = slice_by_index(begin = var_4983_begin_0, end = var_4983_end_0, end_mask = var_4983_end_mask_0, x = read_state_1)[name = string("op_4983_cast_fp16")];
            tensor<int32, [4]> var_4993_begin_0 = const()[name = string("op_4993_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4993_end_0 = const()[name = string("op_4993_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4993_end_mask_0 = const()[name = string("op_4993_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4993_cast_fp16 = slice_by_index(begin = var_4993_begin_0, end = var_4993_end_0, end_mask = var_4993_end_mask_0, x = var_4983_cast_fp16)[name = string("op_4993_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, 32, 128]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = var_4978_cast_fp16, y = var_4993_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<int32, [4]> var_5004_begin_0 = const()[name = string("op_5004_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_5004_end_0 = const()[name = string("op_5004_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_5004_end_mask_0 = const()[name = string("op_5004_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5004_cast_fp16 = slice_by_index(begin = var_5004_begin_0, end = var_5004_end_0, end_mask = var_5004_end_mask_0, x = shares_21_cast_fp16)[name = string("op_5004_cast_fp16")];
            bool var_5005_transpose_x_0 = const()[name = string("op_5005_transpose_x_0"), val = bool(false)];
            bool var_5005_transpose_y_0 = const()[name = string("op_5005_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_4860_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 32, 128]> var_5005_cast_fp16 = matmul(transpose_x = var_5005_transpose_x_0, transpose_y = var_5005_transpose_y_0, x = var_5004_cast_fp16, y = v_21_cast_fp16)[name = string("op_5005_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_41_cast_fp16 = add(x = var_4999_cast_fp16, y = var_5005_cast_fp16)[name = string("mixed_41_cast_fp16")];
            tensor<int32, [4]> var_5022_begin_0 = const()[name = string("op_5022_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5022_end_0 = const()[name = string("op_5022_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5022_end_mask_0 = const()[name = string("op_5022_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5022_cast_fp16 = slice_by_index(begin = var_5022_begin_0, end = var_5022_end_0, end_mask = var_5022_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_5022_cast_fp16")];
            tensor<int32, [4]> var_5037_begin_0 = const()[name = string("op_5037_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5037_end_0 = const()[name = string("op_5037_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5037_end_mask_0 = const()[name = string("op_5037_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5037_cast_fp16 = slice_by_index(begin = var_5037_begin_0, end = var_5037_end_0, end_mask = var_5037_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_5037_cast_fp16")];
            int32 var_5039 = const()[name = string("op_5039"), 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_5039, interleave = mixed_43_interleave_0, values = (var_5022_cast_fp16, var_5037_cast_fp16))[name = string("mixed_43_cast_fp16")];
            tensor<int32, [4]> var_5043_perm_0 = const()[name = string("op_5043_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5047 = const()[name = string("op_5047"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5043_cast_fp16 = transpose(perm = var_5043_perm_0, x = mixed_43_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 2048]> input_103_cast_fp16 = reshape(shape = var_5047, x = var_5043_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_217_cast_fp16 = add(x = x_201_cast_fp16, y = linear_73_cast_fp16)[name = string("x_217_cast_fp16")];
            bool var_5057 = const()[name = string("op_5057"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5060_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<int32, [1]> peak_87_axes_0 = const()[name = string("peak_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_87_cast_fp16 = reduce_max(axes = peak_87_axes_0, keep_dims = var_5057, x = var_5060_cast_fp16)[name = string("peak_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_219_cast_fp16 = real_div(x = x_217_cast_fp16, y = peak_87_cast_fp16)[name = string("x_219_cast_fp16")];
            fp16 var_5056_promoted_to_fp16 = const()[name = string("op_5056_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5064_cast_fp16 = pow(x = x_219_cast_fp16, y = var_5056_promoted_to_fp16)[name = string("op_5064_cast_fp16")];
            tensor<int32, [1]> var_5066_axes_0 = const()[name = string("op_5066_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5066_keep_dims_0 = const()[name = string("op_5066_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5066_cast_fp16 = reduce_mean(axes = var_5066_axes_0, keep_dims = var_5066_keep_dims_0, x = var_5064_cast_fp16)[name = string("op_5066_cast_fp16")];
            fp16 var_5056_promoted_1_to_fp16 = const()[name = string("op_5056_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5067_cast_fp16 = pow(x = peak_87_cast_fp16, y = var_5056_promoted_1_to_fp16)[name = string("op_5067_cast_fp16")];
            fp32 var_5068_epsilon_0 = const()[name = string("op_5068_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5068_cast_fp16 = inverse(epsilon = var_5068_epsilon_0, x = var_5067_cast_fp16)[name = string("op_5068_cast_fp16")];
            fp16 var_5069_to_fp16 = const()[name = string("op_5069_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5070_cast_fp16 = mul(x = var_5068_cast_fp16, y = var_5069_to_fp16)[name = string("op_5070_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5071_cast_fp16 = add(x = var_5066_cast_fp16, y = var_5070_cast_fp16)[name = string("op_5071_cast_fp16")];
            fp32 var_5072_epsilon_0 = const()[name = string("op_5072_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5072_cast_fp16 = rsqrt(epsilon = var_5072_epsilon_0, x = var_5071_cast_fp16)[name = string("op_5072_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5073_cast_fp16 = mul(x = x_219_cast_fp16, y = var_5072_cast_fp16)[name = string("op_5073_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_5073_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_5078_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_5078_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_5078_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_221_cast_fp16 = add(x = x_217_cast_fp16, y = linear_76_cast_fp16)[name = string("x_221_cast_fp16")];
            bool var_5091 = const()[name = string("op_5091"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5094_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [1]> peak_89_axes_0 = const()[name = string("peak_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_89_cast_fp16 = reduce_max(axes = peak_89_axes_0, keep_dims = var_5091, x = var_5094_cast_fp16)[name = string("peak_89_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_223_cast_fp16 = real_div(x = x_221_cast_fp16, y = peak_89_cast_fp16)[name = string("x_223_cast_fp16")];
            fp16 var_5090_promoted_to_fp16 = const()[name = string("op_5090_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5098_cast_fp16 = pow(x = x_223_cast_fp16, y = var_5090_promoted_to_fp16)[name = string("op_5098_cast_fp16")];
            tensor<int32, [1]> var_5100_axes_0 = const()[name = string("op_5100_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5100_keep_dims_0 = const()[name = string("op_5100_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5100_cast_fp16 = reduce_mean(axes = var_5100_axes_0, keep_dims = var_5100_keep_dims_0, x = var_5098_cast_fp16)[name = string("op_5100_cast_fp16")];
            fp16 var_5090_promoted_1_to_fp16 = const()[name = string("op_5090_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5101_cast_fp16 = pow(x = peak_89_cast_fp16, y = var_5090_promoted_1_to_fp16)[name = string("op_5101_cast_fp16")];
            fp32 var_5102_epsilon_0 = const()[name = string("op_5102_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5102_cast_fp16 = inverse(epsilon = var_5102_epsilon_0, x = var_5101_cast_fp16)[name = string("op_5102_cast_fp16")];
            fp16 var_5103_to_fp16 = const()[name = string("op_5103_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5104_cast_fp16 = mul(x = var_5102_cast_fp16, y = var_5103_to_fp16)[name = string("op_5104_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5105_cast_fp16 = add(x = var_5100_cast_fp16, y = var_5104_cast_fp16)[name = string("op_5105_cast_fp16")];
            fp32 var_5106_epsilon_0 = const()[name = string("op_5106_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5106_cast_fp16 = rsqrt(epsilon = var_5106_epsilon_0, x = var_5105_cast_fp16)[name = string("op_5106_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5107_cast_fp16 = mul(x = x_223_cast_fp16, y = var_5106_cast_fp16)[name = string("op_5107_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_5107_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_5116 = const()[name = string("op_5116"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5117_cast_fp16 = reshape(shape = var_5116, x = linear_77_cast_fp16)[name = string("op_5117_cast_fp16")];
            tensor<int32, [4]> x_225_perm_0 = const()[name = string("x_225_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5124 = const()[name = string("op_5124"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_225_cast_fp16 = transpose(perm = x_225_perm_0, x = var_5117_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 16, 16, 128]> var_5127_cast_fp16 = abs(x = x_225_cast_fp16)[name = string("op_5127_cast_fp16")];
            tensor<int32, [1]> peak_91_axes_0 = const()[name = string("peak_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_91_cast_fp16 = reduce_max(axes = peak_91_axes_0, keep_dims = var_5124, x = var_5127_cast_fp16)[name = string("peak_91_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_227_cast_fp16 = real_div(x = x_225_cast_fp16, y = peak_91_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_5123_promoted_to_fp16 = const()[name = string("op_5123_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5131_cast_fp16 = pow(x = x_227_cast_fp16, y = var_5123_promoted_to_fp16)[name = string("op_5131_cast_fp16")];
            tensor<int32, [1]> var_5133_axes_0 = const()[name = string("op_5133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5133_keep_dims_0 = const()[name = string("op_5133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5133_cast_fp16 = reduce_mean(axes = var_5133_axes_0, keep_dims = var_5133_keep_dims_0, x = var_5131_cast_fp16)[name = string("op_5133_cast_fp16")];
            fp16 var_5123_promoted_1_to_fp16 = const()[name = string("op_5123_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5134_cast_fp16 = pow(x = peak_91_cast_fp16, y = var_5123_promoted_1_to_fp16)[name = string("op_5134_cast_fp16")];
            fp32 var_5135_epsilon_0 = const()[name = string("op_5135_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5135_cast_fp16 = inverse(epsilon = var_5135_epsilon_0, x = var_5134_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, 16, 16, 1]> var_5137_cast_fp16 = mul(x = var_5135_cast_fp16, y = var_5136_to_fp16)[name = string("op_5137_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5138_cast_fp16 = add(x = var_5133_cast_fp16, y = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            fp32 var_5139_epsilon_0 = const()[name = string("op_5139_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5139_cast_fp16 = rsqrt(epsilon = var_5139_epsilon_0, x = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5140_cast_fp16 = mul(x = x_227_cast_fp16, y = var_5139_cast_fp16)[name = string("op_5140_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_233_cast_fp16 = mul(x = var_5140_cast_fp16, y = model_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_233_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_5149 = const()[name = string("op_5149"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5150_cast_fp16 = reshape(shape = var_5149, x = linear_78_cast_fp16)[name = string("op_5150_cast_fp16")];
            tensor<int32, [4]> x_229_perm_0 = const()[name = string("x_229_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5157 = const()[name = string("op_5157"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_229_cast_fp16 = transpose(perm = x_229_perm_0, x = var_5150_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 8, 16, 128]> var_5160_cast_fp16 = abs(x = x_229_cast_fp16)[name = string("op_5160_cast_fp16")];
            tensor<int32, [1]> peak_93_axes_0 = const()[name = string("peak_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_93_cast_fp16 = reduce_max(axes = peak_93_axes_0, keep_dims = var_5157, x = var_5160_cast_fp16)[name = string("peak_93_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_231_cast_fp16 = real_div(x = x_229_cast_fp16, y = peak_93_cast_fp16)[name = string("x_231_cast_fp16")];
            fp16 var_5156_promoted_to_fp16 = const()[name = string("op_5156_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5164_cast_fp16 = pow(x = x_231_cast_fp16, y = var_5156_promoted_to_fp16)[name = string("op_5164_cast_fp16")];
            tensor<int32, [1]> var_5166_axes_0 = const()[name = string("op_5166_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5166_keep_dims_0 = const()[name = string("op_5166_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5166_cast_fp16 = reduce_mean(axes = var_5166_axes_0, keep_dims = var_5166_keep_dims_0, x = var_5164_cast_fp16)[name = string("op_5166_cast_fp16")];
            fp16 var_5156_promoted_1_to_fp16 = const()[name = string("op_5156_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5167_cast_fp16 = pow(x = peak_93_cast_fp16, y = var_5156_promoted_1_to_fp16)[name = string("op_5167_cast_fp16")];
            fp32 var_5168_epsilon_0 = const()[name = string("op_5168_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5168_cast_fp16 = inverse(epsilon = var_5168_epsilon_0, x = var_5167_cast_fp16)[name = string("op_5168_cast_fp16")];
            fp16 var_5169_to_fp16 = const()[name = string("op_5169_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5170_cast_fp16 = mul(x = var_5168_cast_fp16, y = var_5169_to_fp16)[name = string("op_5170_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5171_cast_fp16 = add(x = var_5166_cast_fp16, y = var_5170_cast_fp16)[name = string("op_5171_cast_fp16")];
            fp32 var_5172_epsilon_0 = const()[name = string("op_5172_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5172_cast_fp16 = rsqrt(epsilon = var_5172_epsilon_0, x = var_5171_cast_fp16)[name = string("op_5172_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5173_cast_fp16 = mul(x = x_231_cast_fp16, y = var_5172_cast_fp16)[name = string("op_5173_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_235_cast_fp16 = mul(x = var_5173_cast_fp16, y = model_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_235_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_5182 = const()[name = string("op_5182"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5183_cast_fp16 = reshape(shape = var_5182, x = linear_79_cast_fp16)[name = string("op_5183_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_233_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_233_cast_fp16)[name = string("second_45_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5197_cast_fp16 = mul(x = first_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5197_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5198_cast_fp16 = mul(x = second_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5198_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5200_cast_fp16 = sub(x = var_5197_cast_fp16, y = var_5198_cast_fp16)[name = string("op_5200_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5201_cast_fp16 = mul(x = second_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5201_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5202_cast_fp16 = mul(x = first_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5202_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5204_cast_fp16 = add(x = var_5201_cast_fp16, y = var_5202_cast_fp16)[name = string("op_5204_cast_fp16")];
            int32 var_5206 = const()[name = string("op_5206"), 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_5206, interleave = q_23_interleave_0, values = (var_5200_cast_fp16, var_5204_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_235_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_235_cast_fp16)[name = string("second_47_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5218_cast_fp16 = mul(x = first_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5218_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5219_cast_fp16 = mul(x = second_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5219_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5221_cast_fp16 = sub(x = var_5218_cast_fp16, y = var_5219_cast_fp16)[name = string("op_5221_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5222_cast_fp16 = mul(x = second_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5222_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5223_cast_fp16 = mul(x = first_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5225_cast_fp16 = add(x = var_5222_cast_fp16, y = var_5223_cast_fp16)[name = string("op_5225_cast_fp16")];
            int32 var_5227 = const()[name = string("op_5227"), 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_5227, interleave = k_23_interleave_0, values = (var_5221_cast_fp16, var_5225_cast_fp16))[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> var_5238_begin_0 = const()[name = string("op_5238_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5238_end_0 = const()[name = string("op_5238_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5238_end_mask_0 = const()[name = string("op_5238_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5238_stride_0 = const()[name = string("op_5238_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, stride = var_5238_stride_0, x = q_23_cast_fp16)[name = string("op_5238_cast_fp16")];
            tensor<int32, [4]> var_5248_begin_0 = const()[name = string("op_5248_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5248_end_0 = const()[name = string("op_5248_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5248_end_mask_0 = const()[name = string("op_5248_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5248_stride_0 = const()[name = string("op_5248_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5248_cast_fp16 = slice_by_index(begin = var_5248_begin_0, end = var_5248_end_0, end_mask = var_5248_end_mask_0, stride = var_5248_stride_0, x = q_23_cast_fp16)[name = string("op_5248_cast_fp16")];
            int32 var_5250 = const()[name = string("op_5250"), 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_5250, interleave = stacked_23_interleave_0, values = (var_5238_cast_fp16, var_5248_cast_fp16))[name = string("stacked_23_cast_fp16")];
            tensor<int32, [4]> var_5256_begin_0 = const()[name = string("op_5256_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_5256_end_0 = const()[name = string("op_5256_end_0"), val = tensor<int32, [4]>([12, 8, 528, 128])];
            tensor<bool, [4]> var_5256_end_mask_0 = const()[name = string("op_5256_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5256_cast_fp16 = slice_by_index(begin = var_5256_begin_0, end = var_5256_end_0, end_mask = var_5256_end_mask_0, x = read_state_0)[name = string("op_5256_cast_fp16")];
            tensor<int32, [4]> var_5266_begin_0 = const()[name = string("op_5266_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5266_end_0 = const()[name = string("op_5266_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5266_end_mask_0 = const()[name = string("op_5266_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5266_cast_fp16 = slice_by_index(begin = var_5266_begin_0, end = var_5266_end_0, end_mask = var_5266_end_mask_0, x = var_5256_cast_fp16)[name = string("op_5266_cast_fp16")];
            bool var_5275_transpose_x_1 = const()[name = string("op_5275_transpose_x_1"), val = bool(false)];
            bool var_5275_transpose_y_1 = const()[name = string("op_5275_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_5275_cast_fp16 = matmul(transpose_x = var_5275_transpose_x_1, transpose_y = var_5275_transpose_y_1, x = var_5266_cast_fp16, y = stacked_23_cast_fp16)[name = string("op_5275_cast_fp16")];
            tensor<int32, [4]> var_5278_perm_0 = const()[name = string("op_5278_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5279_to_fp16 = const()[name = string("op_5279_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_5278_cast_fp16 = transpose(perm = var_5278_perm_0, x = var_5275_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, 32, 512]> var_5280_cast_fp16 = mul(x = var_5278_cast_fp16, y = var_5279_to_fp16)[name = string("op_5280_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_23_cast_fp16 = add(x = var_5280_cast_fp16, y = cacheMask)[name = string("cache_weights_23_cast_fp16")];
            bool var_5286_transpose_x_1 = const()[name = string("op_5286_transpose_x_1"), val = bool(false)];
            bool var_5286_transpose_y_1 = const()[name = string("op_5286_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5286_cast_fp16 = matmul(transpose_x = var_5286_transpose_x_1, transpose_y = var_5286_transpose_y_1, x = stacked_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_5286_cast_fp16")];
            fp16 var_5287_to_fp16 = const()[name = string("op_5287_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5288_cast_fp16 = mul(x = var_5286_cast_fp16, y = var_5287_to_fp16)[name = string("op_5288_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_23_cast_fp16 = add(x = var_5288_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_23_cast_fp16")];
            int32 var_5292 = const()[name = string("op_5292"), val = int32(-1)];
            bool var_5293_interleave_0 = const()[name = string("op_5293_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_5293_cast_fp16 = concat(axis = var_5292, interleave = var_5293_interleave_0, values = (cache_weights_23_cast_fp16, call_weights_23_cast_fp16))[name = string("op_5293_cast_fp16")];
            int32 var_5294 = const()[name = string("op_5294"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_23_cast_fp16 = softmax(axis = var_5294, x = var_5293_cast_fp16)[name = string("shares_23_cast_fp16")];
            tensor<int32, [4]> var_5301_begin_0 = const()[name = string("op_5301_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5301_end_0 = const()[name = string("op_5301_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_5301_end_mask_0 = const()[name = string("op_5301_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_5301_cast_fp16 = slice_by_index(begin = var_5301_begin_0, end = var_5301_end_0, end_mask = var_5301_end_mask_0, x = shares_23_cast_fp16)[name = string("op_5301_cast_fp16")];
            tensor<int32, [4]> var_5306_begin_0 = const()[name = string("op_5306_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_5306_end_0 = const()[name = string("op_5306_end_0"), val = tensor<int32, [4]>([12, 8, 528, 128])];
            tensor<bool, [4]> var_5306_end_mask_0 = const()[name = string("op_5306_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5306_cast_fp16 = slice_by_index(begin = var_5306_begin_0, end = var_5306_end_0, end_mask = var_5306_end_mask_0, x = read_state_1)[name = string("op_5306_cast_fp16")];
            tensor<int32, [4]> var_5316_begin_0 = const()[name = string("op_5316_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5316_end_0 = const()[name = string("op_5316_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5316_end_mask_0 = const()[name = string("op_5316_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5316_cast_fp16 = slice_by_index(begin = var_5316_begin_0, end = var_5316_end_0, end_mask = var_5316_end_mask_0, x = var_5306_cast_fp16)[name = string("op_5316_cast_fp16")];
            bool var_5322_transpose_x_0 = const()[name = string("op_5322_transpose_x_0"), val = bool(false)];
            bool var_5322_transpose_y_0 = const()[name = string("op_5322_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5322_cast_fp16 = matmul(transpose_x = var_5322_transpose_x_0, transpose_y = var_5322_transpose_y_0, x = var_5301_cast_fp16, y = var_5316_cast_fp16)[name = string("op_5322_cast_fp16")];
            tensor<int32, [4]> var_5327_begin_0 = const()[name = string("op_5327_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_5327_end_0 = const()[name = string("op_5327_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_5327_end_mask_0 = const()[name = string("op_5327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5327_cast_fp16 = slice_by_index(begin = var_5327_begin_0, end = var_5327_end_0, end_mask = var_5327_end_mask_0, x = shares_23_cast_fp16)[name = string("op_5327_cast_fp16")];
            bool var_5328_transpose_x_0 = const()[name = string("op_5328_transpose_x_0"), val = bool(false)];
            bool var_5328_transpose_y_0 = const()[name = string("op_5328_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_5183_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 8, 32, 128]> var_5328_cast_fp16 = matmul(transpose_x = var_5328_transpose_x_0, transpose_y = var_5328_transpose_y_0, x = var_5327_cast_fp16, y = v_23_cast_fp16)[name = string("op_5328_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_45_cast_fp16 = add(x = var_5322_cast_fp16, y = var_5328_cast_fp16)[name = string("mixed_45_cast_fp16")];
            tensor<int32, [4]> var_5345_begin_0 = const()[name = string("op_5345_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5345_end_0 = const()[name = string("op_5345_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5345_end_mask_0 = const()[name = string("op_5345_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5345_cast_fp16 = slice_by_index(begin = var_5345_begin_0, end = var_5345_end_0, end_mask = var_5345_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5345_cast_fp16")];
            tensor<int32, [4]> var_5360_begin_0 = const()[name = string("op_5360_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5360_end_0 = const()[name = string("op_5360_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5360_end_mask_0 = const()[name = string("op_5360_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5360_cast_fp16 = slice_by_index(begin = var_5360_begin_0, end = var_5360_end_0, end_mask = var_5360_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5360_cast_fp16")];
            int32 var_5362 = const()[name = string("op_5362"), 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_5362, interleave = mixed_47_interleave_0, values = (var_5345_cast_fp16, var_5360_cast_fp16))[name = string("mixed_47_cast_fp16")];
            tensor<int32, [4]> var_5366_perm_0 = const()[name = string("op_5366_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5370 = const()[name = string("op_5370"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5366_cast_fp16 = transpose(perm = var_5366_perm_0, x = mixed_47_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 16, 2048]> input_113_cast_fp16 = reshape(shape = var_5370, x = var_5366_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_237_cast_fp16 = add(x = x_221_cast_fp16, y = linear_80_cast_fp16)[name = string("x_237_cast_fp16")];
            bool var_5380 = const()[name = string("op_5380"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5383_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_5383_cast_fp16")];
            tensor<int32, [1]> peak_95_axes_0 = const()[name = string("peak_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_95_cast_fp16 = reduce_max(axes = peak_95_axes_0, keep_dims = var_5380, x = var_5383_cast_fp16)[name = string("peak_95_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_239_cast_fp16 = real_div(x = x_237_cast_fp16, y = peak_95_cast_fp16)[name = string("x_239_cast_fp16")];
            fp16 var_5379_promoted_to_fp16 = const()[name = string("op_5379_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5387_cast_fp16 = pow(x = x_239_cast_fp16, y = var_5379_promoted_to_fp16)[name = string("op_5387_cast_fp16")];
            tensor<int32, [1]> var_5389_axes_0 = const()[name = string("op_5389_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5389_keep_dims_0 = const()[name = string("op_5389_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5389_cast_fp16 = reduce_mean(axes = var_5389_axes_0, keep_dims = var_5389_keep_dims_0, x = var_5387_cast_fp16)[name = string("op_5389_cast_fp16")];
            fp16 var_5379_promoted_1_to_fp16 = const()[name = string("op_5379_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5390_cast_fp16 = pow(x = peak_95_cast_fp16, y = var_5379_promoted_1_to_fp16)[name = string("op_5390_cast_fp16")];
            fp32 var_5391_epsilon_0 = const()[name = string("op_5391_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5391_cast_fp16 = inverse(epsilon = var_5391_epsilon_0, x = var_5390_cast_fp16)[name = string("op_5391_cast_fp16")];
            fp16 var_5392_to_fp16 = const()[name = string("op_5392_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5393_cast_fp16 = mul(x = var_5391_cast_fp16, y = var_5392_to_fp16)[name = string("op_5393_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5394_cast_fp16 = add(x = var_5389_cast_fp16, y = var_5393_cast_fp16)[name = string("op_5394_cast_fp16")];
            fp32 var_5395_epsilon_0 = const()[name = string("op_5395_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5395_cast_fp16 = rsqrt(epsilon = var_5395_epsilon_0, x = var_5394_cast_fp16)[name = string("op_5395_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5396_cast_fp16 = mul(x = x_239_cast_fp16, y = var_5395_cast_fp16)[name = string("op_5396_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_5396_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_5401_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_5401_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_5401_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_241_cast_fp16 = add(x = x_237_cast_fp16, y = linear_83_cast_fp16)[name = string("x_241_cast_fp16")];
            bool var_5414 = const()[name = string("op_5414"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5417_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_5417_cast_fp16")];
            tensor<int32, [1]> peak_97_axes_0 = const()[name = string("peak_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_97_cast_fp16 = reduce_max(axes = peak_97_axes_0, keep_dims = var_5414, x = var_5417_cast_fp16)[name = string("peak_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_243_cast_fp16 = real_div(x = x_241_cast_fp16, y = peak_97_cast_fp16)[name = string("x_243_cast_fp16")];
            fp16 var_5413_promoted_to_fp16 = const()[name = string("op_5413_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5421_cast_fp16 = pow(x = x_243_cast_fp16, y = var_5413_promoted_to_fp16)[name = string("op_5421_cast_fp16")];
            tensor<int32, [1]> var_5423_axes_0 = const()[name = string("op_5423_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5423_keep_dims_0 = const()[name = string("op_5423_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5423_cast_fp16 = reduce_mean(axes = var_5423_axes_0, keep_dims = var_5423_keep_dims_0, x = var_5421_cast_fp16)[name = string("op_5423_cast_fp16")];
            fp16 var_5413_promoted_1_to_fp16 = const()[name = string("op_5413_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5424_cast_fp16 = pow(x = peak_97_cast_fp16, y = var_5413_promoted_1_to_fp16)[name = string("op_5424_cast_fp16")];
            fp32 var_5425_epsilon_0 = const()[name = string("op_5425_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5425_cast_fp16 = inverse(epsilon = var_5425_epsilon_0, x = var_5424_cast_fp16)[name = string("op_5425_cast_fp16")];
            fp16 var_5426_to_fp16 = const()[name = string("op_5426_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5427_cast_fp16 = mul(x = var_5425_cast_fp16, y = var_5426_to_fp16)[name = string("op_5427_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5428_cast_fp16 = add(x = var_5423_cast_fp16, y = var_5427_cast_fp16)[name = string("op_5428_cast_fp16")];
            fp32 var_5429_epsilon_0 = const()[name = string("op_5429_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5429_cast_fp16 = rsqrt(epsilon = var_5429_epsilon_0, x = var_5428_cast_fp16)[name = string("op_5429_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5430_cast_fp16 = mul(x = x_243_cast_fp16, y = var_5429_cast_fp16)[name = string("op_5430_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_5430_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_5439 = const()[name = string("op_5439"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5440_cast_fp16 = reshape(shape = var_5439, x = linear_84_cast_fp16)[name = string("op_5440_cast_fp16")];
            tensor<int32, [4]> x_245_perm_0 = const()[name = string("x_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5447 = const()[name = string("op_5447"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_245_cast_fp16 = transpose(perm = x_245_perm_0, x = var_5440_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 16, 128]> var_5450_cast_fp16 = abs(x = x_245_cast_fp16)[name = string("op_5450_cast_fp16")];
            tensor<int32, [1]> peak_99_axes_0 = const()[name = string("peak_99_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_99_cast_fp16 = reduce_max(axes = peak_99_axes_0, keep_dims = var_5447, x = var_5450_cast_fp16)[name = string("peak_99_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_247_cast_fp16 = real_div(x = x_245_cast_fp16, y = peak_99_cast_fp16)[name = string("x_247_cast_fp16")];
            fp16 var_5446_promoted_to_fp16 = const()[name = string("op_5446_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5454_cast_fp16 = pow(x = x_247_cast_fp16, y = var_5446_promoted_to_fp16)[name = string("op_5454_cast_fp16")];
            tensor<int32, [1]> var_5456_axes_0 = const()[name = string("op_5456_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5456_keep_dims_0 = const()[name = string("op_5456_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5456_cast_fp16 = reduce_mean(axes = var_5456_axes_0, keep_dims = var_5456_keep_dims_0, x = var_5454_cast_fp16)[name = string("op_5456_cast_fp16")];
            fp16 var_5446_promoted_1_to_fp16 = const()[name = string("op_5446_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5457_cast_fp16 = pow(x = peak_99_cast_fp16, y = var_5446_promoted_1_to_fp16)[name = string("op_5457_cast_fp16")];
            fp32 var_5458_epsilon_0 = const()[name = string("op_5458_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5458_cast_fp16 = inverse(epsilon = var_5458_epsilon_0, x = var_5457_cast_fp16)[name = string("op_5458_cast_fp16")];
            fp16 var_5459_to_fp16 = const()[name = string("op_5459_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5460_cast_fp16 = mul(x = var_5458_cast_fp16, y = var_5459_to_fp16)[name = string("op_5460_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5461_cast_fp16 = add(x = var_5456_cast_fp16, y = var_5460_cast_fp16)[name = string("op_5461_cast_fp16")];
            fp32 var_5462_epsilon_0 = const()[name = string("op_5462_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5462_cast_fp16 = rsqrt(epsilon = var_5462_epsilon_0, x = var_5461_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5463_cast_fp16 = mul(x = x_247_cast_fp16, y = var_5462_cast_fp16)[name = string("op_5463_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_253_cast_fp16 = mul(x = var_5463_cast_fp16, y = model_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_253_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_5472 = const()[name = string("op_5472"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5473_cast_fp16 = reshape(shape = var_5472, x = linear_85_cast_fp16)[name = string("op_5473_cast_fp16")];
            tensor<int32, [4]> x_249_perm_0 = const()[name = string("x_249_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5480 = const()[name = string("op_5480"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_249_cast_fp16 = transpose(perm = x_249_perm_0, x = var_5473_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 16, 128]> var_5483_cast_fp16 = abs(x = x_249_cast_fp16)[name = string("op_5483_cast_fp16")];
            tensor<int32, [1]> peak_101_axes_0 = const()[name = string("peak_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_101_cast_fp16 = reduce_max(axes = peak_101_axes_0, keep_dims = var_5480, x = var_5483_cast_fp16)[name = string("peak_101_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_251_cast_fp16 = real_div(x = x_249_cast_fp16, y = peak_101_cast_fp16)[name = string("x_251_cast_fp16")];
            fp16 var_5479_promoted_to_fp16 = const()[name = string("op_5479_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5487_cast_fp16 = pow(x = x_251_cast_fp16, y = var_5479_promoted_to_fp16)[name = string("op_5487_cast_fp16")];
            tensor<int32, [1]> var_5489_axes_0 = const()[name = string("op_5489_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5489_keep_dims_0 = const()[name = string("op_5489_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5489_cast_fp16 = reduce_mean(axes = var_5489_axes_0, keep_dims = var_5489_keep_dims_0, x = var_5487_cast_fp16)[name = string("op_5489_cast_fp16")];
            fp16 var_5479_promoted_1_to_fp16 = const()[name = string("op_5479_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5490_cast_fp16 = pow(x = peak_101_cast_fp16, y = var_5479_promoted_1_to_fp16)[name = string("op_5490_cast_fp16")];
            fp32 var_5491_epsilon_0 = const()[name = string("op_5491_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5491_cast_fp16 = inverse(epsilon = var_5491_epsilon_0, x = var_5490_cast_fp16)[name = string("op_5491_cast_fp16")];
            fp16 var_5492_to_fp16 = const()[name = string("op_5492_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5493_cast_fp16 = mul(x = var_5491_cast_fp16, y = var_5492_to_fp16)[name = string("op_5493_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5494_cast_fp16 = add(x = var_5489_cast_fp16, y = var_5493_cast_fp16)[name = string("op_5494_cast_fp16")];
            fp32 var_5495_epsilon_0 = const()[name = string("op_5495_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5495_cast_fp16 = rsqrt(epsilon = var_5495_epsilon_0, x = var_5494_cast_fp16)[name = string("op_5495_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5496_cast_fp16 = mul(x = x_251_cast_fp16, y = var_5495_cast_fp16)[name = string("op_5496_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_255_cast_fp16 = mul(x = var_5496_cast_fp16, y = model_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_255_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_5505 = const()[name = string("op_5505"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5506_cast_fp16 = reshape(shape = var_5505, x = linear_86_cast_fp16)[name = string("op_5506_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_253_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_253_cast_fp16)[name = string("second_49_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5520_cast_fp16 = mul(x = first_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5520_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5521_cast_fp16 = mul(x = second_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5521_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5523_cast_fp16 = sub(x = var_5520_cast_fp16, y = var_5521_cast_fp16)[name = string("op_5523_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5524_cast_fp16 = mul(x = second_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5525_cast_fp16 = mul(x = first_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5525_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5527_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5525_cast_fp16)[name = string("op_5527_cast_fp16")];
            int32 var_5529 = const()[name = string("op_5529"), 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_5529, interleave = q_25_interleave_0, values = (var_5523_cast_fp16, var_5527_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_255_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_255_cast_fp16)[name = string("second_51_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5541_cast_fp16 = mul(x = first_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5541_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5542_cast_fp16 = mul(x = second_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5542_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5544_cast_fp16 = sub(x = var_5541_cast_fp16, y = var_5542_cast_fp16)[name = string("op_5544_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5545_cast_fp16 = mul(x = second_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5545_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5546_cast_fp16 = mul(x = first_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5546_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5548_cast_fp16 = add(x = var_5545_cast_fp16, y = var_5546_cast_fp16)[name = string("op_5548_cast_fp16")];
            int32 var_5550 = const()[name = string("op_5550"), 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_5550, interleave = k_25_interleave_0, values = (var_5544_cast_fp16, var_5548_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_5561_begin_0 = const()[name = string("op_5561_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5561_end_0 = const()[name = string("op_5561_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5561_end_mask_0 = const()[name = string("op_5561_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5561_stride_0 = const()[name = string("op_5561_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5561_cast_fp16 = slice_by_index(begin = var_5561_begin_0, end = var_5561_end_0, end_mask = var_5561_end_mask_0, stride = var_5561_stride_0, x = q_25_cast_fp16)[name = string("op_5561_cast_fp16")];
            tensor<int32, [4]> var_5571_begin_0 = const()[name = string("op_5571_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5571_end_0 = const()[name = string("op_5571_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5571_end_mask_0 = const()[name = string("op_5571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5571_stride_0 = const()[name = string("op_5571_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5571_cast_fp16 = slice_by_index(begin = var_5571_begin_0, end = var_5571_end_0, end_mask = var_5571_end_mask_0, stride = var_5571_stride_0, x = q_25_cast_fp16)[name = string("op_5571_cast_fp16")];
            int32 var_5573 = const()[name = string("op_5573"), 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_5573, interleave = stacked_25_interleave_0, values = (var_5561_cast_fp16, var_5571_cast_fp16))[name = string("stacked_25_cast_fp16")];
            tensor<int32, [4]> var_5579_begin_0 = const()[name = string("op_5579_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5579_end_0 = const()[name = string("op_5579_end_0"), val = tensor<int32, [4]>([13, 8, 528, 128])];
            tensor<bool, [4]> var_5579_end_mask_0 = const()[name = string("op_5579_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5579_cast_fp16 = slice_by_index(begin = var_5579_begin_0, end = var_5579_end_0, end_mask = var_5579_end_mask_0, x = read_state_0)[name = string("op_5579_cast_fp16")];
            tensor<int32, [4]> var_5589_begin_0 = const()[name = string("op_5589_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5589_end_0 = const()[name = string("op_5589_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5589_end_mask_0 = const()[name = string("op_5589_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5589_cast_fp16 = slice_by_index(begin = var_5589_begin_0, end = var_5589_end_0, end_mask = var_5589_end_mask_0, x = var_5579_cast_fp16)[name = string("op_5589_cast_fp16")];
            bool var_5598_transpose_x_1 = const()[name = string("op_5598_transpose_x_1"), val = bool(false)];
            bool var_5598_transpose_y_1 = const()[name = string("op_5598_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_5598_cast_fp16 = matmul(transpose_x = var_5598_transpose_x_1, transpose_y = var_5598_transpose_y_1, x = var_5589_cast_fp16, y = stacked_25_cast_fp16)[name = string("op_5598_cast_fp16")];
            tensor<int32, [4]> var_5601_perm_0 = const()[name = string("op_5601_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5602_to_fp16 = const()[name = string("op_5602_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_5601_cast_fp16 = transpose(perm = var_5601_perm_0, x = var_5598_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 32, 512]> var_5603_cast_fp16 = mul(x = var_5601_cast_fp16, y = var_5602_to_fp16)[name = string("op_5603_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_25_cast_fp16 = add(x = var_5603_cast_fp16, y = cacheMask)[name = string("cache_weights_25_cast_fp16")];
            bool var_5609_transpose_x_1 = const()[name = string("op_5609_transpose_x_1"), val = bool(false)];
            bool var_5609_transpose_y_1 = const()[name = string("op_5609_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5609_cast_fp16 = matmul(transpose_x = var_5609_transpose_x_1, transpose_y = var_5609_transpose_y_1, x = stacked_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_5609_cast_fp16")];
            fp16 var_5610_to_fp16 = const()[name = string("op_5610_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5611_cast_fp16 = mul(x = var_5609_cast_fp16, y = var_5610_to_fp16)[name = string("op_5611_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_25_cast_fp16 = add(x = var_5611_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_25_cast_fp16")];
            int32 var_5615 = const()[name = string("op_5615"), val = int32(-1)];
            bool var_5616_interleave_0 = const()[name = string("op_5616_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_5616_cast_fp16 = concat(axis = var_5615, interleave = var_5616_interleave_0, values = (cache_weights_25_cast_fp16, call_weights_25_cast_fp16))[name = string("op_5616_cast_fp16")];
            int32 var_5617 = const()[name = string("op_5617"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_25_cast_fp16 = softmax(axis = var_5617, x = var_5616_cast_fp16)[name = string("shares_25_cast_fp16")];
            tensor<int32, [4]> var_5624_begin_0 = const()[name = string("op_5624_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5624_end_0 = const()[name = string("op_5624_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_5624_end_mask_0 = const()[name = string("op_5624_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_5624_cast_fp16 = slice_by_index(begin = var_5624_begin_0, end = var_5624_end_0, end_mask = var_5624_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5624_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([13, 8, 528, 128])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5629_cast_fp16 = slice_by_index(begin = var_5629_begin_0, end = var_5629_end_0, end_mask = var_5629_end_mask_0, x = read_state_1)[name = string("op_5629_cast_fp16")];
            tensor<int32, [4]> var_5639_begin_0 = const()[name = string("op_5639_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5639_end_0 = const()[name = string("op_5639_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5639_end_mask_0 = const()[name = string("op_5639_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5639_cast_fp16 = slice_by_index(begin = var_5639_begin_0, end = var_5639_end_0, end_mask = var_5639_end_mask_0, x = var_5629_cast_fp16)[name = string("op_5639_cast_fp16")];
            bool var_5645_transpose_x_0 = const()[name = string("op_5645_transpose_x_0"), val = bool(false)];
            bool var_5645_transpose_y_0 = const()[name = string("op_5645_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5645_cast_fp16 = matmul(transpose_x = var_5645_transpose_x_0, transpose_y = var_5645_transpose_y_0, x = var_5624_cast_fp16, y = var_5639_cast_fp16)[name = string("op_5645_cast_fp16")];
            tensor<int32, [4]> var_5650_begin_0 = const()[name = string("op_5650_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_5650_end_0 = const()[name = string("op_5650_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_5650_end_mask_0 = const()[name = string("op_5650_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5650_cast_fp16 = slice_by_index(begin = var_5650_begin_0, end = var_5650_end_0, end_mask = var_5650_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5650_cast_fp16")];
            bool var_5651_transpose_x_0 = const()[name = string("op_5651_transpose_x_0"), val = bool(false)];
            bool var_5651_transpose_y_0 = const()[name = string("op_5651_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_5506_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 8, 32, 128]> var_5651_cast_fp16 = matmul(transpose_x = var_5651_transpose_x_0, transpose_y = var_5651_transpose_y_0, x = var_5650_cast_fp16, y = v_25_cast_fp16)[name = string("op_5651_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_49_cast_fp16 = add(x = var_5645_cast_fp16, y = var_5651_cast_fp16)[name = string("mixed_49_cast_fp16")];
            tensor<int32, [4]> var_5668_begin_0 = const()[name = string("op_5668_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5668_end_0 = const()[name = string("op_5668_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5668_end_mask_0 = const()[name = string("op_5668_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5668_cast_fp16 = slice_by_index(begin = var_5668_begin_0, end = var_5668_end_0, end_mask = var_5668_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5668_cast_fp16")];
            tensor<int32, [4]> var_5683_begin_0 = const()[name = string("op_5683_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5683_end_0 = const()[name = string("op_5683_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5683_end_mask_0 = const()[name = string("op_5683_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5683_cast_fp16 = slice_by_index(begin = var_5683_begin_0, end = var_5683_end_0, end_mask = var_5683_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5683_cast_fp16")];
            int32 var_5685 = const()[name = string("op_5685"), 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_5685, interleave = mixed_51_interleave_0, values = (var_5668_cast_fp16, var_5683_cast_fp16))[name = string("mixed_51_cast_fp16")];
            tensor<int32, [4]> var_5689_perm_0 = const()[name = string("op_5689_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5693 = const()[name = string("op_5693"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5689_cast_fp16 = transpose(perm = var_5689_perm_0, x = mixed_51_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 2048]> input_123_cast_fp16 = reshape(shape = var_5693, x = var_5689_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_257_cast_fp16 = add(x = x_241_cast_fp16, y = linear_87_cast_fp16)[name = string("x_257_cast_fp16")];
            bool var_5703 = const()[name = string("op_5703"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5706_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_5706_cast_fp16")];
            tensor<int32, [1]> peak_103_axes_0 = const()[name = string("peak_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_103_cast_fp16 = reduce_max(axes = peak_103_axes_0, keep_dims = var_5703, x = var_5706_cast_fp16)[name = string("peak_103_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_259_cast_fp16 = real_div(x = x_257_cast_fp16, y = peak_103_cast_fp16)[name = string("x_259_cast_fp16")];
            fp16 var_5702_promoted_to_fp16 = const()[name = string("op_5702_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5710_cast_fp16 = pow(x = x_259_cast_fp16, y = var_5702_promoted_to_fp16)[name = string("op_5710_cast_fp16")];
            tensor<int32, [1]> var_5712_axes_0 = const()[name = string("op_5712_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5712_keep_dims_0 = const()[name = string("op_5712_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5712_cast_fp16 = reduce_mean(axes = var_5712_axes_0, keep_dims = var_5712_keep_dims_0, x = var_5710_cast_fp16)[name = string("op_5712_cast_fp16")];
            fp16 var_5702_promoted_1_to_fp16 = const()[name = string("op_5702_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5713_cast_fp16 = pow(x = peak_103_cast_fp16, y = var_5702_promoted_1_to_fp16)[name = string("op_5713_cast_fp16")];
            fp32 var_5714_epsilon_0 = const()[name = string("op_5714_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5714_cast_fp16 = inverse(epsilon = var_5714_epsilon_0, x = var_5713_cast_fp16)[name = string("op_5714_cast_fp16")];
            fp16 var_5715_to_fp16 = const()[name = string("op_5715_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5716_cast_fp16 = mul(x = var_5714_cast_fp16, y = var_5715_to_fp16)[name = string("op_5716_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5717_cast_fp16 = add(x = var_5712_cast_fp16, y = var_5716_cast_fp16)[name = string("op_5717_cast_fp16")];
            fp32 var_5718_epsilon_0 = const()[name = string("op_5718_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5718_cast_fp16 = rsqrt(epsilon = var_5718_epsilon_0, x = var_5717_cast_fp16)[name = string("op_5718_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5719_cast_fp16 = mul(x = x_259_cast_fp16, y = var_5718_cast_fp16)[name = string("op_5719_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_5719_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_5724_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_5724_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_5724_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_261_cast_fp16 = add(x = x_257_cast_fp16, y = linear_90_cast_fp16)[name = string("x_261_cast_fp16")];
            bool var_5737 = const()[name = string("op_5737"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_5740_cast_fp16 = abs(x = x_261_cast_fp16)[name = string("op_5740_cast_fp16")];
            tensor<int32, [1]> peak_105_axes_0 = const()[name = string("peak_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_105_cast_fp16 = reduce_max(axes = peak_105_axes_0, keep_dims = var_5737, x = var_5740_cast_fp16)[name = string("peak_105_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_263_cast_fp16 = real_div(x = x_261_cast_fp16, y = peak_105_cast_fp16)[name = string("x_263_cast_fp16")];
            fp16 var_5736_promoted_to_fp16 = const()[name = string("op_5736_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5744_cast_fp16 = pow(x = x_263_cast_fp16, y = var_5736_promoted_to_fp16)[name = string("op_5744_cast_fp16")];
            tensor<int32, [1]> var_5746_axes_0 = const()[name = string("op_5746_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5746_keep_dims_0 = const()[name = string("op_5746_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5746_cast_fp16 = reduce_mean(axes = var_5746_axes_0, keep_dims = var_5746_keep_dims_0, x = var_5744_cast_fp16)[name = string("op_5746_cast_fp16")];
            fp16 var_5736_promoted_1_to_fp16 = const()[name = string("op_5736_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_5747_cast_fp16 = pow(x = peak_105_cast_fp16, y = var_5736_promoted_1_to_fp16)[name = string("op_5747_cast_fp16")];
            fp32 var_5748_epsilon_0 = const()[name = string("op_5748_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_5748_cast_fp16 = inverse(epsilon = var_5748_epsilon_0, x = var_5747_cast_fp16)[name = string("op_5748_cast_fp16")];
            fp16 var_5749_to_fp16 = const()[name = string("op_5749_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5750_cast_fp16 = mul(x = var_5748_cast_fp16, y = var_5749_to_fp16)[name = string("op_5750_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_5751_cast_fp16 = add(x = var_5746_cast_fp16, y = var_5750_cast_fp16)[name = string("op_5751_cast_fp16")];
            fp32 var_5752_epsilon_0 = const()[name = string("op_5752_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5752_cast_fp16 = rsqrt(epsilon = var_5752_epsilon_0, x = var_5751_cast_fp16)[name = string("op_5752_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5753_cast_fp16 = mul(x = x_263_cast_fp16, y = var_5752_cast_fp16)[name = string("op_5753_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_5753_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_5762 = const()[name = string("op_5762"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5763_cast_fp16 = reshape(shape = var_5762, x = linear_91_cast_fp16)[name = string("op_5763_cast_fp16")];
            tensor<int32, [4]> x_265_perm_0 = const()[name = string("x_265_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5770 = const()[name = string("op_5770"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_265_cast_fp16 = transpose(perm = x_265_perm_0, x = var_5763_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 16, 128]> var_5773_cast_fp16 = abs(x = x_265_cast_fp16)[name = string("op_5773_cast_fp16")];
            tensor<int32, [1]> peak_107_axes_0 = const()[name = string("peak_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_107_cast_fp16 = reduce_max(axes = peak_107_axes_0, keep_dims = var_5770, x = var_5773_cast_fp16)[name = string("peak_107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_267_cast_fp16 = real_div(x = x_265_cast_fp16, y = peak_107_cast_fp16)[name = string("x_267_cast_fp16")];
            fp16 var_5769_promoted_to_fp16 = const()[name = string("op_5769_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_5777_cast_fp16 = pow(x = x_267_cast_fp16, y = var_5769_promoted_to_fp16)[name = string("op_5777_cast_fp16")];
            tensor<int32, [1]> var_5779_axes_0 = const()[name = string("op_5779_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5779_keep_dims_0 = const()[name = string("op_5779_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5779_cast_fp16 = reduce_mean(axes = var_5779_axes_0, keep_dims = var_5779_keep_dims_0, x = var_5777_cast_fp16)[name = string("op_5779_cast_fp16")];
            fp16 var_5769_promoted_1_to_fp16 = const()[name = string("op_5769_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_5780_cast_fp16 = pow(x = peak_107_cast_fp16, y = var_5769_promoted_1_to_fp16)[name = string("op_5780_cast_fp16")];
            fp32 var_5781_epsilon_0 = const()[name = string("op_5781_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_5781_cast_fp16 = inverse(epsilon = var_5781_epsilon_0, x = var_5780_cast_fp16)[name = string("op_5781_cast_fp16")];
            fp16 var_5782_to_fp16 = const()[name = string("op_5782_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5783_cast_fp16 = mul(x = var_5781_cast_fp16, y = var_5782_to_fp16)[name = string("op_5783_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_5784_cast_fp16 = add(x = var_5779_cast_fp16, y = var_5783_cast_fp16)[name = string("op_5784_cast_fp16")];
            fp32 var_5785_epsilon_0 = const()[name = string("op_5785_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5785_cast_fp16 = rsqrt(epsilon = var_5785_epsilon_0, x = var_5784_cast_fp16)[name = string("op_5785_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5786_cast_fp16 = mul(x = x_267_cast_fp16, y = var_5785_cast_fp16)[name = string("op_5786_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_273_cast_fp16 = mul(x = var_5786_cast_fp16, y = model_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_273_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_5795 = const()[name = string("op_5795"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5796_cast_fp16 = reshape(shape = var_5795, x = linear_92_cast_fp16)[name = string("op_5796_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_5803 = const()[name = string("op_5803"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_5796_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 8, 16, 128]> var_5806_cast_fp16 = abs(x = x_269_cast_fp16)[name = string("op_5806_cast_fp16")];
            tensor<int32, [1]> peak_109_axes_0 = const()[name = string("peak_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_109_cast_fp16 = reduce_max(axes = peak_109_axes_0, keep_dims = var_5803, x = var_5806_cast_fp16)[name = string("peak_109_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_271_cast_fp16 = real_div(x = x_269_cast_fp16, y = peak_109_cast_fp16)[name = string("x_271_cast_fp16")];
            fp16 var_5802_promoted_to_fp16 = const()[name = string("op_5802_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_5810_cast_fp16 = pow(x = x_271_cast_fp16, y = var_5802_promoted_to_fp16)[name = string("op_5810_cast_fp16")];
            tensor<int32, [1]> var_5812_axes_0 = const()[name = string("op_5812_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5812_keep_dims_0 = const()[name = string("op_5812_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5812_cast_fp16 = reduce_mean(axes = var_5812_axes_0, keep_dims = var_5812_keep_dims_0, x = var_5810_cast_fp16)[name = string("op_5812_cast_fp16")];
            fp16 var_5802_promoted_1_to_fp16 = const()[name = string("op_5802_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_5813_cast_fp16 = pow(x = peak_109_cast_fp16, y = var_5802_promoted_1_to_fp16)[name = string("op_5813_cast_fp16")];
            fp32 var_5814_epsilon_0 = const()[name = string("op_5814_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_5814_cast_fp16 = inverse(epsilon = var_5814_epsilon_0, x = var_5813_cast_fp16)[name = string("op_5814_cast_fp16")];
            fp16 var_5815_to_fp16 = const()[name = string("op_5815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5816_cast_fp16 = mul(x = var_5814_cast_fp16, y = var_5815_to_fp16)[name = string("op_5816_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_5817_cast_fp16 = add(x = var_5812_cast_fp16, y = var_5816_cast_fp16)[name = string("op_5817_cast_fp16")];
            fp32 var_5818_epsilon_0 = const()[name = string("op_5818_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5818_cast_fp16 = rsqrt(epsilon = var_5818_epsilon_0, x = var_5817_cast_fp16)[name = string("op_5818_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5819_cast_fp16 = mul(x = x_271_cast_fp16, y = var_5818_cast_fp16)[name = string("op_5819_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_275_cast_fp16 = mul(x = var_5819_cast_fp16, y = model_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_275_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_5828 = const()[name = string("op_5828"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5829_cast_fp16 = reshape(shape = var_5828, x = linear_93_cast_fp16)[name = string("op_5829_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_273_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_273_cast_fp16)[name = string("second_53_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5843_cast_fp16 = mul(x = first_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5843_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5844_cast_fp16 = mul(x = second_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5844_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5846_cast_fp16 = sub(x = var_5843_cast_fp16, y = var_5844_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5847_cast_fp16 = mul(x = second_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5847_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5848_cast_fp16 = mul(x = first_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5848_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5850_cast_fp16 = add(x = var_5847_cast_fp16, y = var_5848_cast_fp16)[name = string("op_5850_cast_fp16")];
            int32 var_5852 = const()[name = string("op_5852"), 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_5852, interleave = q_27_interleave_0, values = (var_5846_cast_fp16, var_5850_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_275_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_275_cast_fp16)[name = string("second_55_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5864_cast_fp16 = mul(x = first_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5864_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5865_cast_fp16 = mul(x = second_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5865_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5867_cast_fp16 = sub(x = var_5864_cast_fp16, y = var_5865_cast_fp16)[name = string("op_5867_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5868_cast_fp16 = mul(x = second_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5868_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5869_cast_fp16 = mul(x = first_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5869_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5871_cast_fp16 = add(x = var_5868_cast_fp16, y = var_5869_cast_fp16)[name = string("op_5871_cast_fp16")];
            int32 var_5873 = const()[name = string("op_5873"), 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_5873, interleave = k_27_interleave_0, values = (var_5867_cast_fp16, var_5871_cast_fp16))[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> var_5884_begin_0 = const()[name = string("op_5884_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5884_end_0 = const()[name = string("op_5884_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5884_end_mask_0 = const()[name = string("op_5884_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5884_stride_0 = const()[name = string("op_5884_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5884_cast_fp16 = slice_by_index(begin = var_5884_begin_0, end = var_5884_end_0, end_mask = var_5884_end_mask_0, stride = var_5884_stride_0, x = q_27_cast_fp16)[name = string("op_5884_cast_fp16")];
            tensor<int32, [4]> var_5894_begin_0 = const()[name = string("op_5894_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5894_end_0 = const()[name = string("op_5894_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5894_end_mask_0 = const()[name = string("op_5894_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5894_stride_0 = const()[name = string("op_5894_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5894_cast_fp16 = slice_by_index(begin = var_5894_begin_0, end = var_5894_end_0, end_mask = var_5894_end_mask_0, stride = var_5894_stride_0, x = q_27_cast_fp16)[name = string("op_5894_cast_fp16")];
            int32 var_5896 = const()[name = string("op_5896"), 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_5896, interleave = stacked_27_interleave_0, values = (var_5884_cast_fp16, var_5894_cast_fp16))[name = string("stacked_27_cast_fp16")];
            tensor<int32, [4]> var_5902_begin_0 = const()[name = string("op_5902_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5902_end_0 = const()[name = string("op_5902_end_0"), val = tensor<int32, [4]>([14, 8, 528, 128])];
            tensor<bool, [4]> var_5902_end_mask_0 = const()[name = string("op_5902_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5902_cast_fp16 = slice_by_index(begin = var_5902_begin_0, end = var_5902_end_0, end_mask = var_5902_end_mask_0, x = read_state_0)[name = string("op_5902_cast_fp16")];
            tensor<int32, [4]> var_5912_begin_0 = const()[name = string("op_5912_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5912_end_0 = const()[name = string("op_5912_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5912_end_mask_0 = const()[name = string("op_5912_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5912_cast_fp16 = slice_by_index(begin = var_5912_begin_0, end = var_5912_end_0, end_mask = var_5912_end_mask_0, x = var_5902_cast_fp16)[name = string("op_5912_cast_fp16")];
            bool var_5921_transpose_x_1 = const()[name = string("op_5921_transpose_x_1"), val = bool(false)];
            bool var_5921_transpose_y_1 = const()[name = string("op_5921_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_5921_cast_fp16 = matmul(transpose_x = var_5921_transpose_x_1, transpose_y = var_5921_transpose_y_1, x = var_5912_cast_fp16, y = stacked_27_cast_fp16)[name = string("op_5921_cast_fp16")];
            tensor<int32, [4]> var_5924_perm_0 = const()[name = string("op_5924_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5925_to_fp16 = const()[name = string("op_5925_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_5924_cast_fp16 = transpose(perm = var_5924_perm_0, x = var_5921_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 8, 32, 512]> var_5926_cast_fp16 = mul(x = var_5924_cast_fp16, y = var_5925_to_fp16)[name = string("op_5926_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_27_cast_fp16 = add(x = var_5926_cast_fp16, y = cacheMask)[name = string("cache_weights_27_cast_fp16")];
            bool var_5932_transpose_x_1 = const()[name = string("op_5932_transpose_x_1"), val = bool(false)];
            bool var_5932_transpose_y_1 = const()[name = string("op_5932_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5932_cast_fp16 = matmul(transpose_x = var_5932_transpose_x_1, transpose_y = var_5932_transpose_y_1, x = stacked_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_5932_cast_fp16")];
            fp16 var_5933_to_fp16 = const()[name = string("op_5933_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5934_cast_fp16 = mul(x = var_5932_cast_fp16, y = var_5933_to_fp16)[name = string("op_5934_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_27_cast_fp16 = add(x = var_5934_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_27_cast_fp16")];
            int32 var_5938 = const()[name = string("op_5938"), val = int32(-1)];
            bool var_5939_interleave_0 = const()[name = string("op_5939_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_5939_cast_fp16 = concat(axis = var_5938, interleave = var_5939_interleave_0, values = (cache_weights_27_cast_fp16, call_weights_27_cast_fp16))[name = string("op_5939_cast_fp16")];
            int32 var_5940 = const()[name = string("op_5940"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_27_cast_fp16 = softmax(axis = var_5940, x = var_5939_cast_fp16)[name = string("shares_27_cast_fp16")];
            tensor<int32, [4]> var_5947_begin_0 = const()[name = string("op_5947_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5947_end_0 = const()[name = string("op_5947_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_5947_end_mask_0 = const()[name = string("op_5947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_5947_cast_fp16 = slice_by_index(begin = var_5947_begin_0, end = var_5947_end_0, end_mask = var_5947_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5947_cast_fp16")];
            tensor<int32, [4]> var_5952_begin_0 = const()[name = string("op_5952_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5952_end_0 = const()[name = string("op_5952_end_0"), val = tensor<int32, [4]>([14, 8, 528, 128])];
            tensor<bool, [4]> var_5952_end_mask_0 = const()[name = string("op_5952_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5952_cast_fp16 = slice_by_index(begin = var_5952_begin_0, end = var_5952_end_0, end_mask = var_5952_end_mask_0, x = read_state_1)[name = string("op_5952_cast_fp16")];
            tensor<int32, [4]> var_5962_begin_0 = const()[name = string("op_5962_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5962_end_0 = const()[name = string("op_5962_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5962_end_mask_0 = const()[name = string("op_5962_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5962_cast_fp16 = slice_by_index(begin = var_5962_begin_0, end = var_5962_end_0, end_mask = var_5962_end_mask_0, x = var_5952_cast_fp16)[name = string("op_5962_cast_fp16")];
            bool var_5968_transpose_x_0 = const()[name = string("op_5968_transpose_x_0"), val = bool(false)];
            bool var_5968_transpose_y_0 = const()[name = string("op_5968_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5968_cast_fp16 = matmul(transpose_x = var_5968_transpose_x_0, transpose_y = var_5968_transpose_y_0, x = var_5947_cast_fp16, y = var_5962_cast_fp16)[name = string("op_5968_cast_fp16")];
            tensor<int32, [4]> var_5973_begin_0 = const()[name = string("op_5973_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_5973_end_0 = const()[name = string("op_5973_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_5973_end_mask_0 = const()[name = string("op_5973_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5973_cast_fp16 = slice_by_index(begin = var_5973_begin_0, end = var_5973_end_0, end_mask = var_5973_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5973_cast_fp16")];
            bool var_5974_transpose_x_0 = const()[name = string("op_5974_transpose_x_0"), val = bool(false)];
            bool var_5974_transpose_y_0 = const()[name = string("op_5974_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_5829_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 8, 32, 128]> var_5974_cast_fp16 = matmul(transpose_x = var_5974_transpose_x_0, transpose_y = var_5974_transpose_y_0, x = var_5973_cast_fp16, y = v_27_cast_fp16)[name = string("op_5974_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_53_cast_fp16 = add(x = var_5968_cast_fp16, y = var_5974_cast_fp16)[name = string("mixed_53_cast_fp16")];
            tensor<int32, [4]> var_5991_begin_0 = const()[name = string("op_5991_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5991_end_0 = const()[name = string("op_5991_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5991_end_mask_0 = const()[name = string("op_5991_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5991_cast_fp16 = slice_by_index(begin = var_5991_begin_0, end = var_5991_end_0, end_mask = var_5991_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5991_cast_fp16")];
            tensor<int32, [4]> var_6006_begin_0 = const()[name = string("op_6006_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6006_end_0 = const()[name = string("op_6006_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6006_end_mask_0 = const()[name = string("op_6006_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6006_cast_fp16 = slice_by_index(begin = var_6006_begin_0, end = var_6006_end_0, end_mask = var_6006_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_6006_cast_fp16")];
            int32 var_6008 = const()[name = string("op_6008"), 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_6008, interleave = mixed_55_interleave_0, values = (var_5991_cast_fp16, var_6006_cast_fp16))[name = string("mixed_55_cast_fp16")];
            tensor<int32, [4]> var_6012_perm_0 = const()[name = string("op_6012_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6016 = const()[name = string("op_6016"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6012_cast_fp16 = transpose(perm = var_6012_perm_0, x = mixed_55_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 2048]> input_133_cast_fp16 = reshape(shape = var_6016, x = var_6012_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_277_cast_fp16 = add(x = x_261_cast_fp16, y = linear_94_cast_fp16)[name = string("x_277_cast_fp16")];
            bool var_6026 = const()[name = string("op_6026"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6029_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_6029_cast_fp16")];
            tensor<int32, [1]> peak_111_axes_0 = const()[name = string("peak_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_111_cast_fp16 = reduce_max(axes = peak_111_axes_0, keep_dims = var_6026, x = var_6029_cast_fp16)[name = string("peak_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_279_cast_fp16 = real_div(x = x_277_cast_fp16, y = peak_111_cast_fp16)[name = string("x_279_cast_fp16")];
            fp16 var_6025_promoted_to_fp16 = const()[name = string("op_6025_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6033_cast_fp16 = pow(x = x_279_cast_fp16, y = var_6025_promoted_to_fp16)[name = string("op_6033_cast_fp16")];
            tensor<int32, [1]> var_6035_axes_0 = const()[name = string("op_6035_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6035_keep_dims_0 = const()[name = string("op_6035_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6035_cast_fp16 = reduce_mean(axes = var_6035_axes_0, keep_dims = var_6035_keep_dims_0, x = var_6033_cast_fp16)[name = string("op_6035_cast_fp16")];
            fp16 var_6025_promoted_1_to_fp16 = const()[name = string("op_6025_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6036_cast_fp16 = pow(x = peak_111_cast_fp16, y = var_6025_promoted_1_to_fp16)[name = string("op_6036_cast_fp16")];
            fp32 var_6037_epsilon_0 = const()[name = string("op_6037_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6037_cast_fp16 = inverse(epsilon = var_6037_epsilon_0, x = var_6036_cast_fp16)[name = string("op_6037_cast_fp16")];
            fp16 var_6038_to_fp16 = const()[name = string("op_6038_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6039_cast_fp16 = mul(x = var_6037_cast_fp16, y = var_6038_to_fp16)[name = string("op_6039_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6040_cast_fp16 = add(x = var_6035_cast_fp16, y = var_6039_cast_fp16)[name = string("op_6040_cast_fp16")];
            fp32 var_6041_epsilon_0 = const()[name = string("op_6041_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6041_cast_fp16 = rsqrt(epsilon = var_6041_epsilon_0, x = var_6040_cast_fp16)[name = string("op_6041_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6042_cast_fp16 = mul(x = x_279_cast_fp16, y = var_6041_cast_fp16)[name = string("op_6042_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_6042_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_6047_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_6047_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_6047_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_281_cast_fp16 = add(x = x_277_cast_fp16, y = linear_97_cast_fp16)[name = string("x_281_cast_fp16")];
            bool var_6060 = const()[name = string("op_6060"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6063_cast_fp16 = abs(x = x_281_cast_fp16)[name = string("op_6063_cast_fp16")];
            tensor<int32, [1]> peak_113_axes_0 = const()[name = string("peak_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_113_cast_fp16 = reduce_max(axes = peak_113_axes_0, keep_dims = var_6060, x = var_6063_cast_fp16)[name = string("peak_113_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_283_cast_fp16 = real_div(x = x_281_cast_fp16, y = peak_113_cast_fp16)[name = string("x_283_cast_fp16")];
            fp16 var_6059_promoted_to_fp16 = const()[name = string("op_6059_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6067_cast_fp16 = pow(x = x_283_cast_fp16, y = var_6059_promoted_to_fp16)[name = string("op_6067_cast_fp16")];
            tensor<int32, [1]> var_6069_axes_0 = const()[name = string("op_6069_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6069_keep_dims_0 = const()[name = string("op_6069_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6069_cast_fp16 = reduce_mean(axes = var_6069_axes_0, keep_dims = var_6069_keep_dims_0, x = var_6067_cast_fp16)[name = string("op_6069_cast_fp16")];
            fp16 var_6059_promoted_1_to_fp16 = const()[name = string("op_6059_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6070_cast_fp16 = pow(x = peak_113_cast_fp16, y = var_6059_promoted_1_to_fp16)[name = string("op_6070_cast_fp16")];
            fp32 var_6071_epsilon_0 = const()[name = string("op_6071_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6071_cast_fp16 = inverse(epsilon = var_6071_epsilon_0, x = var_6070_cast_fp16)[name = string("op_6071_cast_fp16")];
            fp16 var_6072_to_fp16 = const()[name = string("op_6072_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6073_cast_fp16 = mul(x = var_6071_cast_fp16, y = var_6072_to_fp16)[name = string("op_6073_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6074_cast_fp16 = add(x = var_6069_cast_fp16, y = var_6073_cast_fp16)[name = string("op_6074_cast_fp16")];
            fp32 var_6075_epsilon_0 = const()[name = string("op_6075_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6075_cast_fp16 = rsqrt(epsilon = var_6075_epsilon_0, x = var_6074_cast_fp16)[name = string("op_6075_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6076_cast_fp16 = mul(x = x_283_cast_fp16, y = var_6075_cast_fp16)[name = string("op_6076_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_6076_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_6085 = const()[name = string("op_6085"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6086_cast_fp16 = reshape(shape = var_6085, x = linear_98_cast_fp16)[name = string("op_6086_cast_fp16")];
            tensor<int32, [4]> x_285_perm_0 = const()[name = string("x_285_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6093 = const()[name = string("op_6093"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_285_cast_fp16 = transpose(perm = x_285_perm_0, x = var_6086_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 16, 128]> var_6096_cast_fp16 = abs(x = x_285_cast_fp16)[name = string("op_6096_cast_fp16")];
            tensor<int32, [1]> peak_115_axes_0 = const()[name = string("peak_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_115_cast_fp16 = reduce_max(axes = peak_115_axes_0, keep_dims = var_6093, x = var_6096_cast_fp16)[name = string("peak_115_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_287_cast_fp16 = real_div(x = x_285_cast_fp16, y = peak_115_cast_fp16)[name = string("x_287_cast_fp16")];
            fp16 var_6092_promoted_to_fp16 = const()[name = string("op_6092_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6100_cast_fp16 = pow(x = x_287_cast_fp16, y = var_6092_promoted_to_fp16)[name = string("op_6100_cast_fp16")];
            tensor<int32, [1]> var_6102_axes_0 = const()[name = string("op_6102_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6102_keep_dims_0 = const()[name = string("op_6102_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6102_cast_fp16 = reduce_mean(axes = var_6102_axes_0, keep_dims = var_6102_keep_dims_0, x = var_6100_cast_fp16)[name = string("op_6102_cast_fp16")];
            fp16 var_6092_promoted_1_to_fp16 = const()[name = string("op_6092_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6103_cast_fp16 = pow(x = peak_115_cast_fp16, y = var_6092_promoted_1_to_fp16)[name = string("op_6103_cast_fp16")];
            fp32 var_6104_epsilon_0 = const()[name = string("op_6104_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6104_cast_fp16 = inverse(epsilon = var_6104_epsilon_0, x = var_6103_cast_fp16)[name = string("op_6104_cast_fp16")];
            fp16 var_6105_to_fp16 = const()[name = string("op_6105_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6106_cast_fp16 = mul(x = var_6104_cast_fp16, y = var_6105_to_fp16)[name = string("op_6106_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6107_cast_fp16 = add(x = var_6102_cast_fp16, y = var_6106_cast_fp16)[name = string("op_6107_cast_fp16")];
            fp32 var_6108_epsilon_0 = const()[name = string("op_6108_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6108_cast_fp16 = rsqrt(epsilon = var_6108_epsilon_0, x = var_6107_cast_fp16)[name = string("op_6108_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6109_cast_fp16 = mul(x = x_287_cast_fp16, y = var_6108_cast_fp16)[name = string("op_6109_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_293_cast_fp16 = mul(x = var_6109_cast_fp16, y = model_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_293_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_6118 = const()[name = string("op_6118"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6119_cast_fp16 = reshape(shape = var_6118, x = linear_99_cast_fp16)[name = string("op_6119_cast_fp16")];
            tensor<int32, [4]> x_289_perm_0 = const()[name = string("x_289_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6126 = const()[name = string("op_6126"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_289_cast_fp16 = transpose(perm = x_289_perm_0, x = var_6119_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 16, 128]> var_6129_cast_fp16 = abs(x = x_289_cast_fp16)[name = string("op_6129_cast_fp16")];
            tensor<int32, [1]> peak_117_axes_0 = const()[name = string("peak_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_117_cast_fp16 = reduce_max(axes = peak_117_axes_0, keep_dims = var_6126, x = var_6129_cast_fp16)[name = string("peak_117_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_291_cast_fp16 = real_div(x = x_289_cast_fp16, y = peak_117_cast_fp16)[name = string("x_291_cast_fp16")];
            fp16 var_6125_promoted_to_fp16 = const()[name = string("op_6125_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6133_cast_fp16 = pow(x = x_291_cast_fp16, y = var_6125_promoted_to_fp16)[name = string("op_6133_cast_fp16")];
            tensor<int32, [1]> var_6135_axes_0 = const()[name = string("op_6135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6135_keep_dims_0 = const()[name = string("op_6135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6135_cast_fp16 = reduce_mean(axes = var_6135_axes_0, keep_dims = var_6135_keep_dims_0, x = var_6133_cast_fp16)[name = string("op_6135_cast_fp16")];
            fp16 var_6125_promoted_1_to_fp16 = const()[name = string("op_6125_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6136_cast_fp16 = pow(x = peak_117_cast_fp16, y = var_6125_promoted_1_to_fp16)[name = string("op_6136_cast_fp16")];
            fp32 var_6137_epsilon_0 = const()[name = string("op_6137_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6137_cast_fp16 = inverse(epsilon = var_6137_epsilon_0, x = var_6136_cast_fp16)[name = string("op_6137_cast_fp16")];
            fp16 var_6138_to_fp16 = const()[name = string("op_6138_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> 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, 16, 1]> var_6140_cast_fp16 = add(x = var_6135_cast_fp16, y = var_6139_cast_fp16)[name = string("op_6140_cast_fp16")];
            fp32 var_6141_epsilon_0 = const()[name = string("op_6141_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6141_cast_fp16 = rsqrt(epsilon = var_6141_epsilon_0, x = var_6140_cast_fp16)[name = string("op_6141_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6142_cast_fp16 = mul(x = x_291_cast_fp16, y = var_6141_cast_fp16)[name = string("op_6142_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_295_cast_fp16 = mul(x = var_6142_cast_fp16, y = model_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_295_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_6151 = const()[name = string("op_6151"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6152_cast_fp16 = reshape(shape = var_6151, x = linear_100_cast_fp16)[name = string("op_6152_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_293_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_293_cast_fp16)[name = string("second_57_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6166_cast_fp16 = mul(x = first_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6166_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6167_cast_fp16 = mul(x = second_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6167_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6169_cast_fp16 = sub(x = var_6166_cast_fp16, y = var_6167_cast_fp16)[name = string("op_6169_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6170_cast_fp16 = mul(x = second_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6170_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6171_cast_fp16 = mul(x = first_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6171_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6173_cast_fp16 = add(x = var_6170_cast_fp16, y = var_6171_cast_fp16)[name = string("op_6173_cast_fp16")];
            int32 var_6175 = const()[name = string("op_6175"), 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_6175, interleave = q_29_interleave_0, values = (var_6169_cast_fp16, var_6173_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_295_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_295_cast_fp16)[name = string("second_59_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6187_cast_fp16 = mul(x = first_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6187_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6188_cast_fp16 = mul(x = second_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6190_cast_fp16 = sub(x = var_6187_cast_fp16, y = var_6188_cast_fp16)[name = string("op_6190_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6191_cast_fp16 = mul(x = second_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6191_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6192_cast_fp16 = mul(x = first_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6192_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6194_cast_fp16 = add(x = var_6191_cast_fp16, y = var_6192_cast_fp16)[name = string("op_6194_cast_fp16")];
            int32 var_6196 = const()[name = string("op_6196"), 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_6196, interleave = k_29_interleave_0, values = (var_6190_cast_fp16, var_6194_cast_fp16))[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_6207_begin_0 = const()[name = string("op_6207_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6207_end_0 = const()[name = string("op_6207_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6207_end_mask_0 = const()[name = string("op_6207_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6207_stride_0 = const()[name = string("op_6207_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6207_cast_fp16 = slice_by_index(begin = var_6207_begin_0, end = var_6207_end_0, end_mask = var_6207_end_mask_0, stride = var_6207_stride_0, x = q_29_cast_fp16)[name = string("op_6207_cast_fp16")];
            tensor<int32, [4]> var_6217_begin_0 = const()[name = string("op_6217_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6217_end_0 = const()[name = string("op_6217_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6217_end_mask_0 = const()[name = string("op_6217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6217_stride_0 = const()[name = string("op_6217_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6217_cast_fp16 = slice_by_index(begin = var_6217_begin_0, end = var_6217_end_0, end_mask = var_6217_end_mask_0, stride = var_6217_stride_0, x = q_29_cast_fp16)[name = string("op_6217_cast_fp16")];
            int32 var_6219 = const()[name = string("op_6219"), 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_6219, interleave = stacked_29_interleave_0, values = (var_6207_cast_fp16, var_6217_cast_fp16))[name = string("stacked_29_cast_fp16")];
            tensor<int32, [4]> var_6225_begin_0 = const()[name = string("op_6225_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_6225_end_0 = const()[name = string("op_6225_end_0"), val = tensor<int32, [4]>([15, 8, 528, 128])];
            tensor<bool, [4]> var_6225_end_mask_0 = const()[name = string("op_6225_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6225_cast_fp16 = slice_by_index(begin = var_6225_begin_0, end = var_6225_end_0, end_mask = var_6225_end_mask_0, x = read_state_0)[name = string("op_6225_cast_fp16")];
            tensor<int32, [4]> var_6235_begin_0 = const()[name = string("op_6235_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6235_end_0 = const()[name = string("op_6235_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6235_end_mask_0 = const()[name = string("op_6235_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6235_cast_fp16 = slice_by_index(begin = var_6235_begin_0, end = var_6235_end_0, end_mask = var_6235_end_mask_0, x = var_6225_cast_fp16)[name = string("op_6235_cast_fp16")];
            bool var_6244_transpose_x_1 = const()[name = string("op_6244_transpose_x_1"), val = bool(false)];
            bool var_6244_transpose_y_1 = const()[name = string("op_6244_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_6244_cast_fp16 = matmul(transpose_x = var_6244_transpose_x_1, transpose_y = var_6244_transpose_y_1, x = var_6235_cast_fp16, y = stacked_29_cast_fp16)[name = string("op_6244_cast_fp16")];
            tensor<int32, [4]> var_6247_perm_0 = const()[name = string("op_6247_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6248_to_fp16 = const()[name = string("op_6248_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_6247_cast_fp16 = transpose(perm = var_6247_perm_0, x = var_6244_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 8, 32, 512]> var_6249_cast_fp16 = mul(x = var_6247_cast_fp16, y = var_6248_to_fp16)[name = string("op_6249_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_29_cast_fp16 = add(x = var_6249_cast_fp16, y = cacheMask)[name = string("cache_weights_29_cast_fp16")];
            bool var_6255_transpose_x_1 = const()[name = string("op_6255_transpose_x_1"), val = bool(false)];
            bool var_6255_transpose_y_1 = const()[name = string("op_6255_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6255_cast_fp16 = matmul(transpose_x = var_6255_transpose_x_1, transpose_y = var_6255_transpose_y_1, x = stacked_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_6255_cast_fp16")];
            fp16 var_6256_to_fp16 = const()[name = string("op_6256_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6257_cast_fp16 = mul(x = var_6255_cast_fp16, y = var_6256_to_fp16)[name = string("op_6257_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_29_cast_fp16 = add(x = var_6257_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_29_cast_fp16")];
            int32 var_6261 = const()[name = string("op_6261"), val = int32(-1)];
            bool var_6262_interleave_0 = const()[name = string("op_6262_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_6262_cast_fp16 = concat(axis = var_6261, interleave = var_6262_interleave_0, values = (cache_weights_29_cast_fp16, call_weights_29_cast_fp16))[name = string("op_6262_cast_fp16")];
            int32 var_6263 = const()[name = string("op_6263"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_29_cast_fp16 = softmax(axis = var_6263, x = var_6262_cast_fp16)[name = string("shares_29_cast_fp16")];
            tensor<int32, [4]> var_6270_begin_0 = const()[name = string("op_6270_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6270_end_0 = const()[name = string("op_6270_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_6270_end_mask_0 = const()[name = string("op_6270_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_6270_cast_fp16 = slice_by_index(begin = var_6270_begin_0, end = var_6270_end_0, end_mask = var_6270_end_mask_0, x = shares_29_cast_fp16)[name = string("op_6270_cast_fp16")];
            tensor<int32, [4]> var_6275_begin_0 = const()[name = string("op_6275_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_6275_end_0 = const()[name = string("op_6275_end_0"), val = tensor<int32, [4]>([15, 8, 528, 128])];
            tensor<bool, [4]> var_6275_end_mask_0 = const()[name = string("op_6275_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6275_cast_fp16 = slice_by_index(begin = var_6275_begin_0, end = var_6275_end_0, end_mask = var_6275_end_mask_0, x = read_state_1)[name = string("op_6275_cast_fp16")];
            tensor<int32, [4]> var_6285_begin_0 = const()[name = string("op_6285_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6285_end_0 = const()[name = string("op_6285_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6285_end_mask_0 = const()[name = string("op_6285_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6285_cast_fp16 = slice_by_index(begin = var_6285_begin_0, end = var_6285_end_0, end_mask = var_6285_end_mask_0, x = var_6275_cast_fp16)[name = string("op_6285_cast_fp16")];
            bool var_6291_transpose_x_0 = const()[name = string("op_6291_transpose_x_0"), val = bool(false)];
            bool var_6291_transpose_y_0 = const()[name = string("op_6291_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6291_cast_fp16 = matmul(transpose_x = var_6291_transpose_x_0, transpose_y = var_6291_transpose_y_0, x = var_6270_cast_fp16, y = var_6285_cast_fp16)[name = string("op_6291_cast_fp16")];
            tensor<int32, [4]> var_6296_begin_0 = const()[name = string("op_6296_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_6296_end_0 = const()[name = string("op_6296_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_6296_end_mask_0 = const()[name = string("op_6296_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6296_cast_fp16 = slice_by_index(begin = var_6296_begin_0, end = var_6296_end_0, end_mask = var_6296_end_mask_0, x = shares_29_cast_fp16)[name = string("op_6296_cast_fp16")];
            bool var_6297_transpose_x_0 = const()[name = string("op_6297_transpose_x_0"), val = bool(false)];
            bool var_6297_transpose_y_0 = const()[name = string("op_6297_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_6152_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 8, 32, 128]> var_6297_cast_fp16 = matmul(transpose_x = var_6297_transpose_x_0, transpose_y = var_6297_transpose_y_0, x = var_6296_cast_fp16, y = v_29_cast_fp16)[name = string("op_6297_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_57_cast_fp16 = add(x = var_6291_cast_fp16, y = var_6297_cast_fp16)[name = string("mixed_57_cast_fp16")];
            tensor<int32, [4]> var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6314_cast_fp16 = slice_by_index(begin = var_6314_begin_0, end = var_6314_end_0, end_mask = var_6314_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_6314_cast_fp16")];
            tensor<int32, [4]> var_6329_begin_0 = const()[name = string("op_6329_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6329_end_0 = const()[name = string("op_6329_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6329_end_mask_0 = const()[name = string("op_6329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6329_cast_fp16 = slice_by_index(begin = var_6329_begin_0, end = var_6329_end_0, end_mask = var_6329_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_6329_cast_fp16")];
            int32 var_6331 = const()[name = string("op_6331"), 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_6331, interleave = mixed_59_interleave_0, values = (var_6314_cast_fp16, var_6329_cast_fp16))[name = string("mixed_59_cast_fp16")];
            tensor<int32, [4]> var_6335_perm_0 = const()[name = string("op_6335_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6339 = const()[name = string("op_6339"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6335_cast_fp16 = transpose(perm = var_6335_perm_0, x = mixed_59_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 2048]> input_143_cast_fp16 = reshape(shape = var_6339, x = var_6335_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_297_cast_fp16 = add(x = x_281_cast_fp16, y = linear_101_cast_fp16)[name = string("x_297_cast_fp16")];
            bool var_6349 = const()[name = string("op_6349"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6352_cast_fp16 = abs(x = x_297_cast_fp16)[name = string("op_6352_cast_fp16")];
            tensor<int32, [1]> peak_119_axes_0 = const()[name = string("peak_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_119_cast_fp16 = reduce_max(axes = peak_119_axes_0, keep_dims = var_6349, x = var_6352_cast_fp16)[name = string("peak_119_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_299_cast_fp16 = real_div(x = x_297_cast_fp16, y = peak_119_cast_fp16)[name = string("x_299_cast_fp16")];
            fp16 var_6348_promoted_to_fp16 = const()[name = string("op_6348_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6356_cast_fp16 = pow(x = x_299_cast_fp16, y = var_6348_promoted_to_fp16)[name = string("op_6356_cast_fp16")];
            tensor<int32, [1]> var_6358_axes_0 = const()[name = string("op_6358_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6358_keep_dims_0 = const()[name = string("op_6358_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6358_cast_fp16 = reduce_mean(axes = var_6358_axes_0, keep_dims = var_6358_keep_dims_0, x = var_6356_cast_fp16)[name = string("op_6358_cast_fp16")];
            fp16 var_6348_promoted_1_to_fp16 = const()[name = string("op_6348_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6359_cast_fp16 = pow(x = peak_119_cast_fp16, y = var_6348_promoted_1_to_fp16)[name = string("op_6359_cast_fp16")];
            fp32 var_6360_epsilon_0 = const()[name = string("op_6360_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6360_cast_fp16 = inverse(epsilon = var_6360_epsilon_0, x = var_6359_cast_fp16)[name = string("op_6360_cast_fp16")];
            fp16 var_6361_to_fp16 = const()[name = string("op_6361_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6362_cast_fp16 = mul(x = var_6360_cast_fp16, y = var_6361_to_fp16)[name = string("op_6362_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6363_cast_fp16 = add(x = var_6358_cast_fp16, y = var_6362_cast_fp16)[name = string("op_6363_cast_fp16")];
            fp32 var_6364_epsilon_0 = const()[name = string("op_6364_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6364_cast_fp16 = rsqrt(epsilon = var_6364_epsilon_0, x = var_6363_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6365_cast_fp16 = mul(x = x_299_cast_fp16, y = var_6364_cast_fp16)[name = string("op_6365_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_6365_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_6370_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_6370_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_6370_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_301_cast_fp16 = add(x = x_297_cast_fp16, y = linear_104_cast_fp16)[name = string("x_301_cast_fp16")];
            bool var_6383 = const()[name = string("op_6383"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6386_cast_fp16 = abs(x = x_301_cast_fp16)[name = string("op_6386_cast_fp16")];
            tensor<int32, [1]> peak_121_axes_0 = const()[name = string("peak_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_121_cast_fp16 = reduce_max(axes = peak_121_axes_0, keep_dims = var_6383, x = var_6386_cast_fp16)[name = string("peak_121_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_303_cast_fp16 = real_div(x = x_301_cast_fp16, y = peak_121_cast_fp16)[name = string("x_303_cast_fp16")];
            fp16 var_6382_promoted_to_fp16 = const()[name = string("op_6382_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6390_cast_fp16 = pow(x = x_303_cast_fp16, y = var_6382_promoted_to_fp16)[name = string("op_6390_cast_fp16")];
            tensor<int32, [1]> var_6392_axes_0 = const()[name = string("op_6392_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6392_keep_dims_0 = const()[name = string("op_6392_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6392_cast_fp16 = reduce_mean(axes = var_6392_axes_0, keep_dims = var_6392_keep_dims_0, x = var_6390_cast_fp16)[name = string("op_6392_cast_fp16")];
            fp16 var_6382_promoted_1_to_fp16 = const()[name = string("op_6382_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6393_cast_fp16 = pow(x = peak_121_cast_fp16, y = var_6382_promoted_1_to_fp16)[name = string("op_6393_cast_fp16")];
            fp32 var_6394_epsilon_0 = const()[name = string("op_6394_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6394_cast_fp16 = inverse(epsilon = var_6394_epsilon_0, x = var_6393_cast_fp16)[name = string("op_6394_cast_fp16")];
            fp16 var_6395_to_fp16 = const()[name = string("op_6395_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6396_cast_fp16 = mul(x = var_6394_cast_fp16, y = var_6395_to_fp16)[name = string("op_6396_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6397_cast_fp16 = add(x = var_6392_cast_fp16, y = var_6396_cast_fp16)[name = string("op_6397_cast_fp16")];
            fp32 var_6398_epsilon_0 = const()[name = string("op_6398_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6398_cast_fp16 = rsqrt(epsilon = var_6398_epsilon_0, x = var_6397_cast_fp16)[name = string("op_6398_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6399_cast_fp16 = mul(x = x_303_cast_fp16, y = var_6398_cast_fp16)[name = string("op_6399_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_6399_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_6408 = const()[name = string("op_6408"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6409_cast_fp16 = reshape(shape = var_6408, x = linear_105_cast_fp16)[name = string("op_6409_cast_fp16")];
            tensor<int32, [4]> x_305_perm_0 = const()[name = string("x_305_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6416 = const()[name = string("op_6416"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_305_cast_fp16 = transpose(perm = x_305_perm_0, x = var_6409_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 16, 16, 128]> var_6419_cast_fp16 = abs(x = x_305_cast_fp16)[name = string("op_6419_cast_fp16")];
            tensor<int32, [1]> peak_123_axes_0 = const()[name = string("peak_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_123_cast_fp16 = reduce_max(axes = peak_123_axes_0, keep_dims = var_6416, x = var_6419_cast_fp16)[name = string("peak_123_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_307_cast_fp16 = real_div(x = x_305_cast_fp16, y = peak_123_cast_fp16)[name = string("x_307_cast_fp16")];
            fp16 var_6415_promoted_to_fp16 = const()[name = string("op_6415_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6423_cast_fp16 = pow(x = x_307_cast_fp16, y = var_6415_promoted_to_fp16)[name = string("op_6423_cast_fp16")];
            tensor<int32, [1]> var_6425_axes_0 = const()[name = string("op_6425_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6425_keep_dims_0 = const()[name = string("op_6425_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6425_cast_fp16 = reduce_mean(axes = var_6425_axes_0, keep_dims = var_6425_keep_dims_0, x = var_6423_cast_fp16)[name = string("op_6425_cast_fp16")];
            fp16 var_6415_promoted_1_to_fp16 = const()[name = string("op_6415_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6426_cast_fp16 = pow(x = peak_123_cast_fp16, y = var_6415_promoted_1_to_fp16)[name = string("op_6426_cast_fp16")];
            fp32 var_6427_epsilon_0 = const()[name = string("op_6427_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6427_cast_fp16 = inverse(epsilon = var_6427_epsilon_0, x = var_6426_cast_fp16)[name = string("op_6427_cast_fp16")];
            fp16 var_6428_to_fp16 = const()[name = string("op_6428_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6429_cast_fp16 = mul(x = var_6427_cast_fp16, y = var_6428_to_fp16)[name = string("op_6429_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6430_cast_fp16 = add(x = var_6425_cast_fp16, y = var_6429_cast_fp16)[name = string("op_6430_cast_fp16")];
            fp32 var_6431_epsilon_0 = const()[name = string("op_6431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6431_cast_fp16 = rsqrt(epsilon = var_6431_epsilon_0, x = var_6430_cast_fp16)[name = string("op_6431_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6432_cast_fp16 = mul(x = x_307_cast_fp16, y = var_6431_cast_fp16)[name = string("op_6432_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_313_cast_fp16 = mul(x = var_6432_cast_fp16, y = model_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_313_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_6441 = const()[name = string("op_6441"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6442_cast_fp16 = reshape(shape = var_6441, x = linear_106_cast_fp16)[name = string("op_6442_cast_fp16")];
            tensor<int32, [4]> x_309_perm_0 = const()[name = string("x_309_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6449 = const()[name = string("op_6449"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_309_cast_fp16 = transpose(perm = x_309_perm_0, x = var_6442_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 8, 16, 128]> var_6452_cast_fp16 = abs(x = x_309_cast_fp16)[name = string("op_6452_cast_fp16")];
            tensor<int32, [1]> peak_125_axes_0 = const()[name = string("peak_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_125_cast_fp16 = reduce_max(axes = peak_125_axes_0, keep_dims = var_6449, x = var_6452_cast_fp16)[name = string("peak_125_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_311_cast_fp16 = real_div(x = x_309_cast_fp16, y = peak_125_cast_fp16)[name = string("x_311_cast_fp16")];
            fp16 var_6448_promoted_to_fp16 = const()[name = string("op_6448_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6456_cast_fp16 = pow(x = x_311_cast_fp16, y = var_6448_promoted_to_fp16)[name = string("op_6456_cast_fp16")];
            tensor<int32, [1]> var_6458_axes_0 = const()[name = string("op_6458_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6458_keep_dims_0 = const()[name = string("op_6458_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6458_cast_fp16 = reduce_mean(axes = var_6458_axes_0, keep_dims = var_6458_keep_dims_0, x = var_6456_cast_fp16)[name = string("op_6458_cast_fp16")];
            fp16 var_6448_promoted_1_to_fp16 = const()[name = string("op_6448_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6459_cast_fp16 = pow(x = peak_125_cast_fp16, y = var_6448_promoted_1_to_fp16)[name = string("op_6459_cast_fp16")];
            fp32 var_6460_epsilon_0 = const()[name = string("op_6460_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6460_cast_fp16 = inverse(epsilon = var_6460_epsilon_0, x = var_6459_cast_fp16)[name = string("op_6460_cast_fp16")];
            fp16 var_6461_to_fp16 = const()[name = string("op_6461_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6462_cast_fp16 = mul(x = var_6460_cast_fp16, y = var_6461_to_fp16)[name = string("op_6462_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_6463_cast_fp16 = add(x = var_6458_cast_fp16, y = var_6462_cast_fp16)[name = string("op_6463_cast_fp16")];
            fp32 var_6464_epsilon_0 = const()[name = string("op_6464_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6464_cast_fp16 = rsqrt(epsilon = var_6464_epsilon_0, x = var_6463_cast_fp16)[name = string("op_6464_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6465_cast_fp16 = mul(x = x_311_cast_fp16, y = var_6464_cast_fp16)[name = string("op_6465_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_315_cast_fp16 = mul(x = var_6465_cast_fp16, y = model_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_315_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_6474 = const()[name = string("op_6474"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6475_cast_fp16 = reshape(shape = var_6474, x = linear_107_cast_fp16)[name = string("op_6475_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_313_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_313_cast_fp16)[name = string("second_61_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6489_cast_fp16 = mul(x = first_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6489_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6490_cast_fp16 = mul(x = second_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6490_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6492_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = string("op_6492_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6493_cast_fp16 = mul(x = second_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6493_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6494_cast_fp16 = mul(x = first_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6494_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6496_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = string("op_6496_cast_fp16")];
            int32 var_6498 = const()[name = string("op_6498"), 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_6498, interleave = q_31_interleave_0, values = (var_6492_cast_fp16, var_6496_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_315_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_315_cast_fp16)[name = string("second_63_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6510_cast_fp16 = mul(x = first_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6510_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6511_cast_fp16 = mul(x = second_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6511_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6513_cast_fp16 = sub(x = var_6510_cast_fp16, y = var_6511_cast_fp16)[name = string("op_6513_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6514_cast_fp16 = mul(x = second_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6514_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6515_cast_fp16 = mul(x = first_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6515_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6517_cast_fp16 = add(x = var_6514_cast_fp16, y = var_6515_cast_fp16)[name = string("op_6517_cast_fp16")];
            int32 var_6519 = const()[name = string("op_6519"), 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_6519, interleave = k_31_interleave_0, values = (var_6513_cast_fp16, var_6517_cast_fp16))[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> var_6530_begin_0 = const()[name = string("op_6530_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6530_end_0 = const()[name = string("op_6530_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6530_end_mask_0 = const()[name = string("op_6530_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6530_stride_0 = const()[name = string("op_6530_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6530_cast_fp16 = slice_by_index(begin = var_6530_begin_0, end = var_6530_end_0, end_mask = var_6530_end_mask_0, stride = var_6530_stride_0, x = q_31_cast_fp16)[name = string("op_6530_cast_fp16")];
            tensor<int32, [4]> var_6540_begin_0 = const()[name = string("op_6540_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6540_end_0 = const()[name = string("op_6540_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6540_end_mask_0 = const()[name = string("op_6540_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6540_stride_0 = const()[name = string("op_6540_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6540_cast_fp16 = slice_by_index(begin = var_6540_begin_0, end = var_6540_end_0, end_mask = var_6540_end_mask_0, stride = var_6540_stride_0, x = q_31_cast_fp16)[name = string("op_6540_cast_fp16")];
            int32 var_6542 = const()[name = string("op_6542"), 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_6542, interleave = stacked_31_interleave_0, values = (var_6530_cast_fp16, var_6540_cast_fp16))[name = string("stacked_31_cast_fp16")];
            tensor<int32, [4]> var_6548_begin_0 = const()[name = string("op_6548_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6548_end_0 = const()[name = string("op_6548_end_0"), val = tensor<int32, [4]>([16, 8, 528, 128])];
            tensor<bool, [4]> var_6548_end_mask_0 = const()[name = string("op_6548_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6548_cast_fp16 = slice_by_index(begin = var_6548_begin_0, end = var_6548_end_0, end_mask = var_6548_end_mask_0, x = read_state_0)[name = string("op_6548_cast_fp16")];
            tensor<int32, [4]> var_6558_begin_0 = const()[name = string("op_6558_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6558_end_0 = const()[name = string("op_6558_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6558_end_mask_0 = const()[name = string("op_6558_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6558_cast_fp16 = slice_by_index(begin = var_6558_begin_0, end = var_6558_end_0, end_mask = var_6558_end_mask_0, x = var_6548_cast_fp16)[name = string("op_6558_cast_fp16")];
            bool var_6567_transpose_x_1 = const()[name = string("op_6567_transpose_x_1"), val = bool(false)];
            bool var_6567_transpose_y_1 = const()[name = string("op_6567_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_6567_cast_fp16 = matmul(transpose_x = var_6567_transpose_x_1, transpose_y = var_6567_transpose_y_1, x = var_6558_cast_fp16, y = stacked_31_cast_fp16)[name = string("op_6567_cast_fp16")];
            tensor<int32, [4]> var_6570_perm_0 = const()[name = string("op_6570_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6571_to_fp16 = const()[name = string("op_6571_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_6570_cast_fp16 = transpose(perm = var_6570_perm_0, x = var_6567_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, 32, 512]> var_6572_cast_fp16 = mul(x = var_6570_cast_fp16, y = var_6571_to_fp16)[name = string("op_6572_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_31_cast_fp16 = add(x = var_6572_cast_fp16, y = cacheMask)[name = string("cache_weights_31_cast_fp16")];
            bool var_6578_transpose_x_1 = const()[name = string("op_6578_transpose_x_1"), val = bool(false)];
            bool var_6578_transpose_y_1 = const()[name = string("op_6578_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6578_cast_fp16 = matmul(transpose_x = var_6578_transpose_x_1, transpose_y = var_6578_transpose_y_1, x = stacked_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp16 var_6579_to_fp16 = const()[name = string("op_6579_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6580_cast_fp16 = mul(x = var_6578_cast_fp16, y = var_6579_to_fp16)[name = string("op_6580_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_31_cast_fp16 = add(x = var_6580_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_31_cast_fp16")];
            int32 var_6584 = const()[name = string("op_6584"), val = int32(-1)];
            bool var_6585_interleave_0 = const()[name = string("op_6585_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_6585_cast_fp16 = concat(axis = var_6584, interleave = var_6585_interleave_0, values = (cache_weights_31_cast_fp16, call_weights_31_cast_fp16))[name = string("op_6585_cast_fp16")];
            int32 var_6586 = const()[name = string("op_6586"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_31_cast_fp16 = softmax(axis = var_6586, x = var_6585_cast_fp16)[name = string("shares_31_cast_fp16")];
            tensor<int32, [4]> var_6593_begin_0 = const()[name = string("op_6593_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6593_end_0 = const()[name = string("op_6593_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_6593_end_mask_0 = const()[name = string("op_6593_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_6593_cast_fp16 = slice_by_index(begin = var_6593_begin_0, end = var_6593_end_0, end_mask = var_6593_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6593_cast_fp16")];
            tensor<int32, [4]> var_6598_begin_0 = const()[name = string("op_6598_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6598_end_0 = const()[name = string("op_6598_end_0"), val = tensor<int32, [4]>([16, 8, 528, 128])];
            tensor<bool, [4]> var_6598_end_mask_0 = const()[name = string("op_6598_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6598_cast_fp16 = slice_by_index(begin = var_6598_begin_0, end = var_6598_end_0, end_mask = var_6598_end_mask_0, x = read_state_1)[name = string("op_6598_cast_fp16")];
            tensor<int32, [4]> var_6608_begin_0 = const()[name = string("op_6608_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6608_end_0 = const()[name = string("op_6608_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6608_end_mask_0 = const()[name = string("op_6608_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6608_cast_fp16 = slice_by_index(begin = var_6608_begin_0, end = var_6608_end_0, end_mask = var_6608_end_mask_0, x = var_6598_cast_fp16)[name = string("op_6608_cast_fp16")];
            bool var_6614_transpose_x_0 = const()[name = string("op_6614_transpose_x_0"), val = bool(false)];
            bool var_6614_transpose_y_0 = const()[name = string("op_6614_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6614_cast_fp16 = matmul(transpose_x = var_6614_transpose_x_0, transpose_y = var_6614_transpose_y_0, x = var_6593_cast_fp16, y = var_6608_cast_fp16)[name = string("op_6614_cast_fp16")];
            tensor<int32, [4]> var_6619_begin_0 = const()[name = string("op_6619_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_6619_end_0 = const()[name = string("op_6619_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_6619_end_mask_0 = const()[name = string("op_6619_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6619_cast_fp16")];
            bool var_6620_transpose_x_0 = const()[name = string("op_6620_transpose_x_0"), val = bool(false)];
            bool var_6620_transpose_y_0 = const()[name = string("op_6620_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_6475_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 32, 128]> var_6620_cast_fp16 = matmul(transpose_x = var_6620_transpose_x_0, transpose_y = var_6620_transpose_y_0, x = var_6619_cast_fp16, y = v_31_cast_fp16)[name = string("op_6620_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_61_cast_fp16 = add(x = var_6614_cast_fp16, y = var_6620_cast_fp16)[name = string("mixed_61_cast_fp16")];
            tensor<int32, [4]> var_6637_begin_0 = const()[name = string("op_6637_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6637_end_0 = const()[name = string("op_6637_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6637_end_mask_0 = const()[name = string("op_6637_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6637_cast_fp16 = slice_by_index(begin = var_6637_begin_0, end = var_6637_end_0, end_mask = var_6637_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6637_cast_fp16")];
            tensor<int32, [4]> var_6652_begin_0 = const()[name = string("op_6652_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6652_end_0 = const()[name = string("op_6652_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6652_end_mask_0 = const()[name = string("op_6652_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6652_cast_fp16 = slice_by_index(begin = var_6652_begin_0, end = var_6652_end_0, end_mask = var_6652_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6652_cast_fp16")];
            int32 var_6654 = const()[name = string("op_6654"), 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_6654, interleave = mixed_63_interleave_0, values = (var_6637_cast_fp16, var_6652_cast_fp16))[name = string("mixed_63_cast_fp16")];
            tensor<int32, [4]> var_6658_perm_0 = const()[name = string("op_6658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6662 = const()[name = string("op_6662"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6658_cast_fp16 = transpose(perm = var_6658_perm_0, x = mixed_63_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 16, 2048]> input_153_cast_fp16 = reshape(shape = var_6662, x = var_6658_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_317_cast_fp16 = add(x = x_301_cast_fp16, y = linear_108_cast_fp16)[name = string("x_317_cast_fp16")];
            bool var_6672 = const()[name = string("op_6672"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6675_cast_fp16 = abs(x = x_317_cast_fp16)[name = string("op_6675_cast_fp16")];
            tensor<int32, [1]> peak_127_axes_0 = const()[name = string("peak_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_127_cast_fp16 = reduce_max(axes = peak_127_axes_0, keep_dims = var_6672, x = var_6675_cast_fp16)[name = string("peak_127_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_319_cast_fp16 = real_div(x = x_317_cast_fp16, y = peak_127_cast_fp16)[name = string("x_319_cast_fp16")];
            fp16 var_6671_promoted_to_fp16 = const()[name = string("op_6671_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6679_cast_fp16 = pow(x = x_319_cast_fp16, y = var_6671_promoted_to_fp16)[name = string("op_6679_cast_fp16")];
            tensor<int32, [1]> var_6681_axes_0 = const()[name = string("op_6681_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6681_keep_dims_0 = const()[name = string("op_6681_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6681_cast_fp16 = reduce_mean(axes = var_6681_axes_0, keep_dims = var_6681_keep_dims_0, x = var_6679_cast_fp16)[name = string("op_6681_cast_fp16")];
            fp16 var_6671_promoted_1_to_fp16 = const()[name = string("op_6671_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6682_cast_fp16 = pow(x = peak_127_cast_fp16, y = var_6671_promoted_1_to_fp16)[name = string("op_6682_cast_fp16")];
            fp32 var_6683_epsilon_0 = const()[name = string("op_6683_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6683_cast_fp16 = inverse(epsilon = var_6683_epsilon_0, x = var_6682_cast_fp16)[name = string("op_6683_cast_fp16")];
            fp16 var_6684_to_fp16 = const()[name = string("op_6684_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6685_cast_fp16 = mul(x = var_6683_cast_fp16, y = var_6684_to_fp16)[name = string("op_6685_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6686_cast_fp16 = add(x = var_6681_cast_fp16, y = var_6685_cast_fp16)[name = string("op_6686_cast_fp16")];
            fp32 var_6687_epsilon_0 = const()[name = string("op_6687_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6687_cast_fp16 = rsqrt(epsilon = var_6687_epsilon_0, x = var_6686_cast_fp16)[name = string("op_6687_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6688_cast_fp16 = mul(x = x_319_cast_fp16, y = var_6687_cast_fp16)[name = string("op_6688_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_6688_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_6693_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_6693_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_6693_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_321_cast_fp16 = add(x = x_317_cast_fp16, y = linear_111_cast_fp16)[name = string("x_321_cast_fp16")];
            bool var_6706 = const()[name = string("op_6706"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6709_cast_fp16 = abs(x = x_321_cast_fp16)[name = string("op_6709_cast_fp16")];
            tensor<int32, [1]> peak_129_axes_0 = const()[name = string("peak_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_129_cast_fp16 = reduce_max(axes = peak_129_axes_0, keep_dims = var_6706, x = var_6709_cast_fp16)[name = string("peak_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_323_cast_fp16 = real_div(x = x_321_cast_fp16, y = peak_129_cast_fp16)[name = string("x_323_cast_fp16")];
            fp16 var_6705_promoted_to_fp16 = const()[name = string("op_6705_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6713_cast_fp16 = pow(x = x_323_cast_fp16, y = var_6705_promoted_to_fp16)[name = string("op_6713_cast_fp16")];
            tensor<int32, [1]> var_6715_axes_0 = const()[name = string("op_6715_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6715_keep_dims_0 = const()[name = string("op_6715_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6715_cast_fp16 = reduce_mean(axes = var_6715_axes_0, keep_dims = var_6715_keep_dims_0, x = var_6713_cast_fp16)[name = string("op_6715_cast_fp16")];
            fp16 var_6705_promoted_1_to_fp16 = const()[name = string("op_6705_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_6716_cast_fp16 = pow(x = peak_129_cast_fp16, y = var_6705_promoted_1_to_fp16)[name = string("op_6716_cast_fp16")];
            fp32 var_6717_epsilon_0 = const()[name = string("op_6717_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_6717_cast_fp16 = inverse(epsilon = var_6717_epsilon_0, x = var_6716_cast_fp16)[name = string("op_6717_cast_fp16")];
            fp16 var_6718_to_fp16 = const()[name = string("op_6718_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6719_cast_fp16 = mul(x = var_6717_cast_fp16, y = var_6718_to_fp16)[name = string("op_6719_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_6720_cast_fp16 = add(x = var_6715_cast_fp16, y = var_6719_cast_fp16)[name = string("op_6720_cast_fp16")];
            fp32 var_6721_epsilon_0 = const()[name = string("op_6721_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6721_cast_fp16 = rsqrt(epsilon = var_6721_epsilon_0, x = var_6720_cast_fp16)[name = string("op_6721_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6722_cast_fp16 = mul(x = x_323_cast_fp16, y = var_6721_cast_fp16)[name = string("op_6722_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_6722_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_6731 = const()[name = string("op_6731"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6732_cast_fp16 = reshape(shape = var_6731, x = linear_112_cast_fp16)[name = string("op_6732_cast_fp16")];
            tensor<int32, [4]> x_325_perm_0 = const()[name = string("x_325_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6739 = const()[name = string("op_6739"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_325_cast_fp16 = transpose(perm = x_325_perm_0, x = var_6732_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 16, 128]> var_6742_cast_fp16 = abs(x = x_325_cast_fp16)[name = string("op_6742_cast_fp16")];
            tensor<int32, [1]> peak_131_axes_0 = const()[name = string("peak_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_131_cast_fp16 = reduce_max(axes = peak_131_axes_0, keep_dims = var_6739, x = var_6742_cast_fp16)[name = string("peak_131_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_327_cast_fp16 = real_div(x = x_325_cast_fp16, y = peak_131_cast_fp16)[name = string("x_327_cast_fp16")];
            fp16 var_6738_promoted_to_fp16 = const()[name = string("op_6738_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_6746_cast_fp16 = pow(x = x_327_cast_fp16, y = var_6738_promoted_to_fp16)[name = string("op_6746_cast_fp16")];
            tensor<int32, [1]> var_6748_axes_0 = const()[name = string("op_6748_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6748_keep_dims_0 = const()[name = string("op_6748_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6748_cast_fp16 = reduce_mean(axes = var_6748_axes_0, keep_dims = var_6748_keep_dims_0, x = var_6746_cast_fp16)[name = string("op_6748_cast_fp16")];
            fp16 var_6738_promoted_1_to_fp16 = const()[name = string("op_6738_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_6749_cast_fp16 = pow(x = peak_131_cast_fp16, y = var_6738_promoted_1_to_fp16)[name = string("op_6749_cast_fp16")];
            fp32 var_6750_epsilon_0 = const()[name = string("op_6750_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_6750_cast_fp16 = inverse(epsilon = var_6750_epsilon_0, x = var_6749_cast_fp16)[name = string("op_6750_cast_fp16")];
            fp16 var_6751_to_fp16 = const()[name = string("op_6751_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6752_cast_fp16 = mul(x = var_6750_cast_fp16, y = var_6751_to_fp16)[name = string("op_6752_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_6753_cast_fp16 = add(x = var_6748_cast_fp16, y = var_6752_cast_fp16)[name = string("op_6753_cast_fp16")];
            fp32 var_6754_epsilon_0 = const()[name = string("op_6754_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6754_cast_fp16 = rsqrt(epsilon = var_6754_epsilon_0, x = var_6753_cast_fp16)[name = string("op_6754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6755_cast_fp16 = mul(x = x_327_cast_fp16, y = var_6754_cast_fp16)[name = string("op_6755_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_333_cast_fp16 = mul(x = var_6755_cast_fp16, y = model_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("x_333_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_6764 = const()[name = string("op_6764"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6765_cast_fp16 = reshape(shape = var_6764, x = linear_113_cast_fp16)[name = string("op_6765_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_6772 = const()[name = string("op_6772"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_6765_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 16, 128]> var_6775_cast_fp16 = abs(x = x_329_cast_fp16)[name = string("op_6775_cast_fp16")];
            tensor<int32, [1]> peak_133_axes_0 = const()[name = string("peak_133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_133_cast_fp16 = reduce_max(axes = peak_133_axes_0, keep_dims = var_6772, x = var_6775_cast_fp16)[name = string("peak_133_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_331_cast_fp16 = real_div(x = x_329_cast_fp16, y = peak_133_cast_fp16)[name = string("x_331_cast_fp16")];
            fp16 var_6771_promoted_to_fp16 = const()[name = string("op_6771_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_6779_cast_fp16 = pow(x = x_331_cast_fp16, y = var_6771_promoted_to_fp16)[name = string("op_6779_cast_fp16")];
            tensor<int32, [1]> var_6781_axes_0 = const()[name = string("op_6781_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6781_keep_dims_0 = const()[name = string("op_6781_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6781_cast_fp16 = reduce_mean(axes = var_6781_axes_0, keep_dims = var_6781_keep_dims_0, x = var_6779_cast_fp16)[name = string("op_6781_cast_fp16")];
            fp16 var_6771_promoted_1_to_fp16 = const()[name = string("op_6771_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_6782_cast_fp16 = pow(x = peak_133_cast_fp16, y = var_6771_promoted_1_to_fp16)[name = string("op_6782_cast_fp16")];
            fp32 var_6783_epsilon_0 = const()[name = string("op_6783_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_6783_cast_fp16 = inverse(epsilon = var_6783_epsilon_0, x = var_6782_cast_fp16)[name = string("op_6783_cast_fp16")];
            fp16 var_6784_to_fp16 = const()[name = string("op_6784_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6785_cast_fp16 = mul(x = var_6783_cast_fp16, y = var_6784_to_fp16)[name = string("op_6785_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_6786_cast_fp16 = add(x = var_6781_cast_fp16, y = var_6785_cast_fp16)[name = string("op_6786_cast_fp16")];
            fp32 var_6787_epsilon_0 = const()[name = string("op_6787_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6787_cast_fp16 = rsqrt(epsilon = var_6787_epsilon_0, x = var_6786_cast_fp16)[name = string("op_6787_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6788_cast_fp16 = mul(x = x_331_cast_fp16, y = var_6787_cast_fp16)[name = string("op_6788_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_335_cast_fp16 = mul(x = var_6788_cast_fp16, y = model_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("x_335_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_6797 = const()[name = string("op_6797"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6798_cast_fp16 = reshape(shape = var_6797, x = linear_114_cast_fp16)[name = string("op_6798_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_333_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_333_cast_fp16)[name = string("second_65_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6812_cast_fp16 = mul(x = first_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6812_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6813_cast_fp16 = mul(x = second_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6813_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6815_cast_fp16 = sub(x = var_6812_cast_fp16, y = var_6813_cast_fp16)[name = string("op_6815_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6816_cast_fp16 = mul(x = second_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6816_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6817_cast_fp16 = mul(x = first_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6817_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6819_cast_fp16 = add(x = var_6816_cast_fp16, y = var_6817_cast_fp16)[name = string("op_6819_cast_fp16")];
            int32 var_6821 = const()[name = string("op_6821"), 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_6821, interleave = q_33_interleave_0, values = (var_6815_cast_fp16, var_6819_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_335_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_335_cast_fp16)[name = string("second_67_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6833_cast_fp16 = mul(x = first_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6833_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6834_cast_fp16 = mul(x = second_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6834_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6836_cast_fp16 = sub(x = var_6833_cast_fp16, y = var_6834_cast_fp16)[name = string("op_6836_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6837_cast_fp16 = mul(x = second_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6837_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6838_cast_fp16 = mul(x = first_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6838_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6840_cast_fp16 = add(x = var_6837_cast_fp16, y = var_6838_cast_fp16)[name = string("op_6840_cast_fp16")];
            int32 var_6842 = const()[name = string("op_6842"), 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_6842, interleave = k_33_interleave_0, values = (var_6836_cast_fp16, var_6840_cast_fp16))[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_6853_begin_0 = const()[name = string("op_6853_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6853_end_0 = const()[name = string("op_6853_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6853_end_mask_0 = const()[name = string("op_6853_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6853_stride_0 = const()[name = string("op_6853_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6853_cast_fp16 = slice_by_index(begin = var_6853_begin_0, end = var_6853_end_0, end_mask = var_6853_end_mask_0, stride = var_6853_stride_0, x = q_33_cast_fp16)[name = string("op_6853_cast_fp16")];
            tensor<int32, [4]> var_6863_begin_0 = const()[name = string("op_6863_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6863_end_0 = const()[name = string("op_6863_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6863_end_mask_0 = const()[name = string("op_6863_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6863_stride_0 = const()[name = string("op_6863_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6863_cast_fp16 = slice_by_index(begin = var_6863_begin_0, end = var_6863_end_0, end_mask = var_6863_end_mask_0, stride = var_6863_stride_0, x = q_33_cast_fp16)[name = string("op_6863_cast_fp16")];
            int32 var_6865 = const()[name = string("op_6865"), 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_6865, interleave = stacked_33_interleave_0, values = (var_6853_cast_fp16, var_6863_cast_fp16))[name = string("stacked_33_cast_fp16")];
            tensor<int32, [4]> var_6871_begin_0 = const()[name = string("op_6871_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6871_end_0 = const()[name = string("op_6871_end_0"), val = tensor<int32, [4]>([17, 8, 528, 128])];
            tensor<bool, [4]> var_6871_end_mask_0 = const()[name = string("op_6871_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6871_cast_fp16 = slice_by_index(begin = var_6871_begin_0, end = var_6871_end_0, end_mask = var_6871_end_mask_0, x = read_state_0)[name = string("op_6871_cast_fp16")];
            tensor<int32, [4]> var_6881_begin_0 = const()[name = string("op_6881_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6881_end_0 = const()[name = string("op_6881_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6881_end_mask_0 = const()[name = string("op_6881_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6881_cast_fp16 = slice_by_index(begin = var_6881_begin_0, end = var_6881_end_0, end_mask = var_6881_end_mask_0, x = var_6871_cast_fp16)[name = string("op_6881_cast_fp16")];
            bool var_6890_transpose_x_1 = const()[name = string("op_6890_transpose_x_1"), val = bool(false)];
            bool var_6890_transpose_y_1 = const()[name = string("op_6890_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_6890_cast_fp16 = matmul(transpose_x = var_6890_transpose_x_1, transpose_y = var_6890_transpose_y_1, x = var_6881_cast_fp16, y = stacked_33_cast_fp16)[name = string("op_6890_cast_fp16")];
            tensor<int32, [4]> var_6893_perm_0 = const()[name = string("op_6893_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6894_to_fp16 = const()[name = string("op_6894_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_6893_cast_fp16 = transpose(perm = var_6893_perm_0, x = var_6890_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 32, 512]> var_6895_cast_fp16 = mul(x = var_6893_cast_fp16, y = var_6894_to_fp16)[name = string("op_6895_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_33_cast_fp16 = add(x = var_6895_cast_fp16, y = cacheMask)[name = string("cache_weights_33_cast_fp16")];
            bool var_6901_transpose_x_1 = const()[name = string("op_6901_transpose_x_1"), val = bool(false)];
            bool var_6901_transpose_y_1 = const()[name = string("op_6901_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6901_cast_fp16 = matmul(transpose_x = var_6901_transpose_x_1, transpose_y = var_6901_transpose_y_1, x = stacked_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_6901_cast_fp16")];
            fp16 var_6902_to_fp16 = const()[name = string("op_6902_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6903_cast_fp16 = mul(x = var_6901_cast_fp16, y = var_6902_to_fp16)[name = string("op_6903_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_33_cast_fp16 = add(x = var_6903_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_33_cast_fp16")];
            int32 var_6907 = const()[name = string("op_6907"), val = int32(-1)];
            bool var_6908_interleave_0 = const()[name = string("op_6908_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_6908_cast_fp16 = concat(axis = var_6907, interleave = var_6908_interleave_0, values = (cache_weights_33_cast_fp16, call_weights_33_cast_fp16))[name = string("op_6908_cast_fp16")];
            int32 var_6909 = const()[name = string("op_6909"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_33_cast_fp16 = softmax(axis = var_6909, x = var_6908_cast_fp16)[name = string("shares_33_cast_fp16")];
            tensor<int32, [4]> var_6916_begin_0 = const()[name = string("op_6916_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6916_end_0 = const()[name = string("op_6916_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_6916_end_mask_0 = const()[name = string("op_6916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_6916_cast_fp16 = slice_by_index(begin = var_6916_begin_0, end = var_6916_end_0, end_mask = var_6916_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6916_cast_fp16")];
            tensor<int32, [4]> var_6921_begin_0 = const()[name = string("op_6921_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6921_end_0 = const()[name = string("op_6921_end_0"), val = tensor<int32, [4]>([17, 8, 528, 128])];
            tensor<bool, [4]> var_6921_end_mask_0 = const()[name = string("op_6921_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6921_cast_fp16 = slice_by_index(begin = var_6921_begin_0, end = var_6921_end_0, end_mask = var_6921_end_mask_0, x = read_state_1)[name = string("op_6921_cast_fp16")];
            tensor<int32, [4]> var_6931_begin_0 = const()[name = string("op_6931_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6931_end_0 = const()[name = string("op_6931_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6931_end_mask_0 = const()[name = string("op_6931_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6931_cast_fp16 = slice_by_index(begin = var_6931_begin_0, end = var_6931_end_0, end_mask = var_6931_end_mask_0, x = var_6921_cast_fp16)[name = string("op_6931_cast_fp16")];
            bool var_6937_transpose_x_0 = const()[name = string("op_6937_transpose_x_0"), val = bool(false)];
            bool var_6937_transpose_y_0 = const()[name = string("op_6937_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6937_cast_fp16 = matmul(transpose_x = var_6937_transpose_x_0, transpose_y = var_6937_transpose_y_0, x = var_6916_cast_fp16, y = var_6931_cast_fp16)[name = string("op_6937_cast_fp16")];
            tensor<int32, [4]> var_6942_begin_0 = const()[name = string("op_6942_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_6942_end_0 = const()[name = string("op_6942_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_6942_end_mask_0 = const()[name = string("op_6942_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6942_cast_fp16 = slice_by_index(begin = var_6942_begin_0, end = var_6942_end_0, end_mask = var_6942_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6942_cast_fp16")];
            bool var_6943_transpose_x_0 = const()[name = string("op_6943_transpose_x_0"), val = bool(false)];
            bool var_6943_transpose_y_0 = const()[name = string("op_6943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_6798_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 8, 32, 128]> var_6943_cast_fp16 = matmul(transpose_x = var_6943_transpose_x_0, transpose_y = var_6943_transpose_y_0, x = var_6942_cast_fp16, y = v_33_cast_fp16)[name = string("op_6943_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_65_cast_fp16 = add(x = var_6937_cast_fp16, y = var_6943_cast_fp16)[name = string("mixed_65_cast_fp16")];
            tensor<int32, [4]> var_6960_begin_0 = const()[name = string("op_6960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6960_end_0 = const()[name = string("op_6960_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6960_end_mask_0 = const()[name = string("op_6960_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6960_cast_fp16 = slice_by_index(begin = var_6960_begin_0, end = var_6960_end_0, end_mask = var_6960_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6960_cast_fp16")];
            tensor<int32, [4]> var_6975_begin_0 = const()[name = string("op_6975_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6975_end_0 = const()[name = string("op_6975_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6975_end_mask_0 = const()[name = string("op_6975_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6975_cast_fp16 = slice_by_index(begin = var_6975_begin_0, end = var_6975_end_0, end_mask = var_6975_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6975_cast_fp16")];
            int32 var_6977 = const()[name = string("op_6977"), 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_6977, interleave = mixed_67_interleave_0, values = (var_6960_cast_fp16, var_6975_cast_fp16))[name = string("mixed_67_cast_fp16")];
            tensor<int32, [4]> var_6981_perm_0 = const()[name = string("op_6981_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6985 = const()[name = string("op_6985"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6981_cast_fp16 = transpose(perm = var_6981_perm_0, x = mixed_67_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 2048]> input_163_cast_fp16 = reshape(shape = var_6985, x = var_6981_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_337_cast_fp16 = add(x = x_321_cast_fp16, y = linear_115_cast_fp16)[name = string("x_337_cast_fp16")];
            bool var_6995 = const()[name = string("op_6995"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_6998_cast_fp16 = abs(x = x_337_cast_fp16)[name = string("op_6998_cast_fp16")];
            tensor<int32, [1]> peak_135_axes_0 = const()[name = string("peak_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_135_cast_fp16 = reduce_max(axes = peak_135_axes_0, keep_dims = var_6995, x = var_6998_cast_fp16)[name = string("peak_135_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_339_cast_fp16 = real_div(x = x_337_cast_fp16, y = peak_135_cast_fp16)[name = string("x_339_cast_fp16")];
            fp16 var_6994_promoted_to_fp16 = const()[name = string("op_6994_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7002_cast_fp16 = pow(x = x_339_cast_fp16, y = var_6994_promoted_to_fp16)[name = string("op_7002_cast_fp16")];
            tensor<int32, [1]> var_7004_axes_0 = const()[name = string("op_7004_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7004_keep_dims_0 = const()[name = string("op_7004_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7004_cast_fp16 = reduce_mean(axes = var_7004_axes_0, keep_dims = var_7004_keep_dims_0, x = var_7002_cast_fp16)[name = string("op_7004_cast_fp16")];
            fp16 var_6994_promoted_1_to_fp16 = const()[name = string("op_6994_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7005_cast_fp16 = pow(x = peak_135_cast_fp16, y = var_6994_promoted_1_to_fp16)[name = string("op_7005_cast_fp16")];
            fp32 var_7006_epsilon_0 = const()[name = string("op_7006_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7006_cast_fp16 = inverse(epsilon = var_7006_epsilon_0, x = var_7005_cast_fp16)[name = string("op_7006_cast_fp16")];
            fp16 var_7007_to_fp16 = const()[name = string("op_7007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7008_cast_fp16 = mul(x = var_7006_cast_fp16, y = var_7007_to_fp16)[name = string("op_7008_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7009_cast_fp16 = add(x = var_7004_cast_fp16, y = var_7008_cast_fp16)[name = string("op_7009_cast_fp16")];
            fp32 var_7010_epsilon_0 = const()[name = string("op_7010_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7010_cast_fp16 = rsqrt(epsilon = var_7010_epsilon_0, x = var_7009_cast_fp16)[name = string("op_7010_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7011_cast_fp16 = mul(x = x_339_cast_fp16, y = var_7010_cast_fp16)[name = string("op_7011_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_7011_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_7016_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_7016_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_7016_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_341_cast_fp16 = add(x = x_337_cast_fp16, y = linear_118_cast_fp16)[name = string("x_341_cast_fp16")];
            bool var_7029 = const()[name = string("op_7029"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7032_cast_fp16 = abs(x = x_341_cast_fp16)[name = string("op_7032_cast_fp16")];
            tensor<int32, [1]> peak_137_axes_0 = const()[name = string("peak_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_137_cast_fp16 = reduce_max(axes = peak_137_axes_0, keep_dims = var_7029, x = var_7032_cast_fp16)[name = string("peak_137_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_343_cast_fp16 = real_div(x = x_341_cast_fp16, y = peak_137_cast_fp16)[name = string("x_343_cast_fp16")];
            fp16 var_7028_promoted_to_fp16 = const()[name = string("op_7028_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7036_cast_fp16 = pow(x = x_343_cast_fp16, y = var_7028_promoted_to_fp16)[name = string("op_7036_cast_fp16")];
            tensor<int32, [1]> var_7038_axes_0 = const()[name = string("op_7038_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7038_keep_dims_0 = const()[name = string("op_7038_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7038_cast_fp16 = reduce_mean(axes = var_7038_axes_0, keep_dims = var_7038_keep_dims_0, x = var_7036_cast_fp16)[name = string("op_7038_cast_fp16")];
            fp16 var_7028_promoted_1_to_fp16 = const()[name = string("op_7028_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7039_cast_fp16 = pow(x = peak_137_cast_fp16, y = var_7028_promoted_1_to_fp16)[name = string("op_7039_cast_fp16")];
            fp32 var_7040_epsilon_0 = const()[name = string("op_7040_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7040_cast_fp16 = inverse(epsilon = var_7040_epsilon_0, x = var_7039_cast_fp16)[name = string("op_7040_cast_fp16")];
            fp16 var_7041_to_fp16 = const()[name = string("op_7041_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7042_cast_fp16 = mul(x = var_7040_cast_fp16, y = var_7041_to_fp16)[name = string("op_7042_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7043_cast_fp16 = add(x = var_7038_cast_fp16, y = var_7042_cast_fp16)[name = string("op_7043_cast_fp16")];
            fp32 var_7044_epsilon_0 = const()[name = string("op_7044_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7044_cast_fp16 = rsqrt(epsilon = var_7044_epsilon_0, x = var_7043_cast_fp16)[name = string("op_7044_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7045_cast_fp16 = mul(x = x_343_cast_fp16, y = var_7044_cast_fp16)[name = string("op_7045_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_7045_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_7054 = const()[name = string("op_7054"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7055_cast_fp16 = reshape(shape = var_7054, x = linear_119_cast_fp16)[name = string("op_7055_cast_fp16")];
            tensor<int32, [4]> x_345_perm_0 = const()[name = string("x_345_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7062 = const()[name = string("op_7062"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_345_cast_fp16 = transpose(perm = x_345_perm_0, x = var_7055_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 16, 128]> var_7065_cast_fp16 = abs(x = x_345_cast_fp16)[name = string("op_7065_cast_fp16")];
            tensor<int32, [1]> peak_139_axes_0 = const()[name = string("peak_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_139_cast_fp16 = reduce_max(axes = peak_139_axes_0, keep_dims = var_7062, x = var_7065_cast_fp16)[name = string("peak_139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_347_cast_fp16 = real_div(x = x_345_cast_fp16, y = peak_139_cast_fp16)[name = string("x_347_cast_fp16")];
            fp16 var_7061_promoted_to_fp16 = const()[name = string("op_7061_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7069_cast_fp16 = pow(x = x_347_cast_fp16, y = var_7061_promoted_to_fp16)[name = string("op_7069_cast_fp16")];
            tensor<int32, [1]> var_7071_axes_0 = const()[name = string("op_7071_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7071_keep_dims_0 = const()[name = string("op_7071_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7071_cast_fp16 = reduce_mean(axes = var_7071_axes_0, keep_dims = var_7071_keep_dims_0, x = var_7069_cast_fp16)[name = string("op_7071_cast_fp16")];
            fp16 var_7061_promoted_1_to_fp16 = const()[name = string("op_7061_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7072_cast_fp16 = pow(x = peak_139_cast_fp16, y = var_7061_promoted_1_to_fp16)[name = string("op_7072_cast_fp16")];
            fp32 var_7073_epsilon_0 = const()[name = string("op_7073_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7073_cast_fp16 = inverse(epsilon = var_7073_epsilon_0, x = var_7072_cast_fp16)[name = string("op_7073_cast_fp16")];
            fp16 var_7074_to_fp16 = const()[name = string("op_7074_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7075_cast_fp16 = mul(x = var_7073_cast_fp16, y = var_7074_to_fp16)[name = string("op_7075_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7076_cast_fp16 = add(x = var_7071_cast_fp16, y = var_7075_cast_fp16)[name = string("op_7076_cast_fp16")];
            fp32 var_7077_epsilon_0 = const()[name = string("op_7077_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7077_cast_fp16 = rsqrt(epsilon = var_7077_epsilon_0, x = var_7076_cast_fp16)[name = string("op_7077_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7078_cast_fp16 = mul(x = x_347_cast_fp16, y = var_7077_cast_fp16)[name = string("op_7078_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_353_cast_fp16 = mul(x = var_7078_cast_fp16, y = model_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("x_353_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_7087 = const()[name = string("op_7087"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7088_cast_fp16 = reshape(shape = var_7087, x = linear_120_cast_fp16)[name = string("op_7088_cast_fp16")];
            tensor<int32, [4]> x_349_perm_0 = const()[name = string("x_349_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7095 = const()[name = string("op_7095"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_349_cast_fp16 = transpose(perm = x_349_perm_0, x = var_7088_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 8, 16, 128]> var_7098_cast_fp16 = abs(x = x_349_cast_fp16)[name = string("op_7098_cast_fp16")];
            tensor<int32, [1]> peak_141_axes_0 = const()[name = string("peak_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_141_cast_fp16 = reduce_max(axes = peak_141_axes_0, keep_dims = var_7095, x = var_7098_cast_fp16)[name = string("peak_141_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_351_cast_fp16 = real_div(x = x_349_cast_fp16, y = peak_141_cast_fp16)[name = string("x_351_cast_fp16")];
            fp16 var_7094_promoted_to_fp16 = const()[name = string("op_7094_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7102_cast_fp16 = pow(x = x_351_cast_fp16, y = var_7094_promoted_to_fp16)[name = string("op_7102_cast_fp16")];
            tensor<int32, [1]> var_7104_axes_0 = const()[name = string("op_7104_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7104_keep_dims_0 = const()[name = string("op_7104_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7104_cast_fp16 = reduce_mean(axes = var_7104_axes_0, keep_dims = var_7104_keep_dims_0, x = var_7102_cast_fp16)[name = string("op_7104_cast_fp16")];
            fp16 var_7094_promoted_1_to_fp16 = const()[name = string("op_7094_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7105_cast_fp16 = pow(x = peak_141_cast_fp16, y = var_7094_promoted_1_to_fp16)[name = string("op_7105_cast_fp16")];
            fp32 var_7106_epsilon_0 = const()[name = string("op_7106_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7106_cast_fp16 = inverse(epsilon = var_7106_epsilon_0, x = var_7105_cast_fp16)[name = string("op_7106_cast_fp16")];
            fp16 var_7107_to_fp16 = const()[name = string("op_7107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7108_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7107_to_fp16)[name = string("op_7108_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7109_cast_fp16 = add(x = var_7104_cast_fp16, y = var_7108_cast_fp16)[name = string("op_7109_cast_fp16")];
            fp32 var_7110_epsilon_0 = const()[name = string("op_7110_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7110_cast_fp16 = rsqrt(epsilon = var_7110_epsilon_0, x = var_7109_cast_fp16)[name = string("op_7110_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7111_cast_fp16 = mul(x = x_351_cast_fp16, y = var_7110_cast_fp16)[name = string("op_7111_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_355_cast_fp16 = mul(x = var_7111_cast_fp16, y = model_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("x_355_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_7120 = const()[name = string("op_7120"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7121_cast_fp16 = reshape(shape = var_7120, x = linear_121_cast_fp16)[name = string("op_7121_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_353_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_353_cast_fp16)[name = string("second_69_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7135_cast_fp16 = mul(x = first_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7135_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7136_cast_fp16 = mul(x = second_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7136_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7138_cast_fp16 = sub(x = var_7135_cast_fp16, y = var_7136_cast_fp16)[name = string("op_7138_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7139_cast_fp16 = mul(x = second_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7140_cast_fp16 = mul(x = first_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7140_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7142_cast_fp16 = add(x = var_7139_cast_fp16, y = var_7140_cast_fp16)[name = string("op_7142_cast_fp16")];
            int32 var_7144 = const()[name = string("op_7144"), 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_7144, interleave = q_35_interleave_0, values = (var_7138_cast_fp16, var_7142_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_355_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_355_cast_fp16)[name = string("second_71_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7156_cast_fp16 = mul(x = first_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7156_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7157_cast_fp16 = mul(x = second_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7157_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7159_cast_fp16 = sub(x = var_7156_cast_fp16, y = var_7157_cast_fp16)[name = string("op_7159_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7160_cast_fp16 = mul(x = second_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7160_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7161_cast_fp16 = mul(x = first_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7161_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7163_cast_fp16 = add(x = var_7160_cast_fp16, y = var_7161_cast_fp16)[name = string("op_7163_cast_fp16")];
            int32 var_7165 = const()[name = string("op_7165"), 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_7165, interleave = k_35_interleave_0, values = (var_7159_cast_fp16, var_7163_cast_fp16))[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> var_7176_begin_0 = const()[name = string("op_7176_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7176_end_0 = const()[name = string("op_7176_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7176_end_mask_0 = const()[name = string("op_7176_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7176_stride_0 = const()[name = string("op_7176_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7176_cast_fp16 = slice_by_index(begin = var_7176_begin_0, end = var_7176_end_0, end_mask = var_7176_end_mask_0, stride = var_7176_stride_0, x = q_35_cast_fp16)[name = string("op_7176_cast_fp16")];
            tensor<int32, [4]> var_7186_begin_0 = const()[name = string("op_7186_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7186_end_0 = const()[name = string("op_7186_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7186_end_mask_0 = const()[name = string("op_7186_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7186_stride_0 = const()[name = string("op_7186_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7186_cast_fp16 = slice_by_index(begin = var_7186_begin_0, end = var_7186_end_0, end_mask = var_7186_end_mask_0, stride = var_7186_stride_0, x = q_35_cast_fp16)[name = string("op_7186_cast_fp16")];
            int32 var_7188 = const()[name = string("op_7188"), 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_7188, interleave = stacked_35_interleave_0, values = (var_7176_cast_fp16, var_7186_cast_fp16))[name = string("stacked_35_cast_fp16")];
            tensor<int32, [4]> var_7194_begin_0 = const()[name = string("op_7194_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_7194_end_0 = const()[name = string("op_7194_end_0"), val = tensor<int32, [4]>([18, 8, 528, 128])];
            tensor<bool, [4]> var_7194_end_mask_0 = const()[name = string("op_7194_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7194_cast_fp16 = slice_by_index(begin = var_7194_begin_0, end = var_7194_end_0, end_mask = var_7194_end_mask_0, x = read_state_0)[name = string("op_7194_cast_fp16")];
            tensor<int32, [4]> var_7204_begin_0 = const()[name = string("op_7204_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7204_end_0 = const()[name = string("op_7204_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7204_end_mask_0 = const()[name = string("op_7204_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7204_cast_fp16 = slice_by_index(begin = var_7204_begin_0, end = var_7204_end_0, end_mask = var_7204_end_mask_0, x = var_7194_cast_fp16)[name = string("op_7204_cast_fp16")];
            bool var_7213_transpose_x_1 = const()[name = string("op_7213_transpose_x_1"), val = bool(false)];
            bool var_7213_transpose_y_1 = const()[name = string("op_7213_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_7213_cast_fp16 = matmul(transpose_x = var_7213_transpose_x_1, transpose_y = var_7213_transpose_y_1, x = var_7204_cast_fp16, y = stacked_35_cast_fp16)[name = string("op_7213_cast_fp16")];
            tensor<int32, [4]> var_7216_perm_0 = const()[name = string("op_7216_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7217_to_fp16 = const()[name = string("op_7217_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_7216_cast_fp16 = transpose(perm = var_7216_perm_0, x = var_7213_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 8, 32, 512]> var_7218_cast_fp16 = mul(x = var_7216_cast_fp16, y = var_7217_to_fp16)[name = string("op_7218_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_35_cast_fp16 = add(x = var_7218_cast_fp16, y = cacheMask)[name = string("cache_weights_35_cast_fp16")];
            bool var_7224_transpose_x_1 = const()[name = string("op_7224_transpose_x_1"), val = bool(false)];
            bool var_7224_transpose_y_1 = const()[name = string("op_7224_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7224_cast_fp16 = matmul(transpose_x = var_7224_transpose_x_1, transpose_y = var_7224_transpose_y_1, x = stacked_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_7224_cast_fp16")];
            fp16 var_7225_to_fp16 = const()[name = string("op_7225_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7226_cast_fp16 = mul(x = var_7224_cast_fp16, y = var_7225_to_fp16)[name = string("op_7226_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_35_cast_fp16 = add(x = var_7226_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_35_cast_fp16")];
            int32 var_7230 = const()[name = string("op_7230"), val = int32(-1)];
            bool var_7231_interleave_0 = const()[name = string("op_7231_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_7231_cast_fp16 = concat(axis = var_7230, interleave = var_7231_interleave_0, values = (cache_weights_35_cast_fp16, call_weights_35_cast_fp16))[name = string("op_7231_cast_fp16")];
            int32 var_7232 = const()[name = string("op_7232"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_35_cast_fp16 = softmax(axis = var_7232, x = var_7231_cast_fp16)[name = string("shares_35_cast_fp16")];
            tensor<int32, [4]> var_7239_begin_0 = const()[name = string("op_7239_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7239_end_0 = const()[name = string("op_7239_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_7239_end_mask_0 = const()[name = string("op_7239_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_7239_cast_fp16 = slice_by_index(begin = var_7239_begin_0, end = var_7239_end_0, end_mask = var_7239_end_mask_0, x = shares_35_cast_fp16)[name = string("op_7239_cast_fp16")];
            tensor<int32, [4]> var_7244_begin_0 = const()[name = string("op_7244_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_7244_end_0 = const()[name = string("op_7244_end_0"), val = tensor<int32, [4]>([18, 8, 528, 128])];
            tensor<bool, [4]> var_7244_end_mask_0 = const()[name = string("op_7244_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7244_cast_fp16 = slice_by_index(begin = var_7244_begin_0, end = var_7244_end_0, end_mask = var_7244_end_mask_0, x = read_state_1)[name = string("op_7244_cast_fp16")];
            tensor<int32, [4]> var_7254_begin_0 = const()[name = string("op_7254_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7254_end_0 = const()[name = string("op_7254_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7254_end_mask_0 = const()[name = string("op_7254_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7254_cast_fp16 = slice_by_index(begin = var_7254_begin_0, end = var_7254_end_0, end_mask = var_7254_end_mask_0, x = var_7244_cast_fp16)[name = string("op_7254_cast_fp16")];
            bool var_7260_transpose_x_0 = const()[name = string("op_7260_transpose_x_0"), val = bool(false)];
            bool var_7260_transpose_y_0 = const()[name = string("op_7260_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7260_cast_fp16 = matmul(transpose_x = var_7260_transpose_x_0, transpose_y = var_7260_transpose_y_0, x = var_7239_cast_fp16, y = var_7254_cast_fp16)[name = string("op_7260_cast_fp16")];
            tensor<int32, [4]> var_7265_begin_0 = const()[name = string("op_7265_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_7265_end_0 = const()[name = string("op_7265_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_7265_end_mask_0 = const()[name = string("op_7265_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7265_cast_fp16 = slice_by_index(begin = var_7265_begin_0, end = var_7265_end_0, end_mask = var_7265_end_mask_0, x = shares_35_cast_fp16)[name = string("op_7265_cast_fp16")];
            bool var_7266_transpose_x_0 = const()[name = string("op_7266_transpose_x_0"), val = bool(false)];
            bool var_7266_transpose_y_0 = const()[name = string("op_7266_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_7121_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 32, 128]> var_7266_cast_fp16 = matmul(transpose_x = var_7266_transpose_x_0, transpose_y = var_7266_transpose_y_0, x = var_7265_cast_fp16, y = v_35_cast_fp16)[name = string("op_7266_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_69_cast_fp16 = add(x = var_7260_cast_fp16, y = var_7266_cast_fp16)[name = string("mixed_69_cast_fp16")];
            tensor<int32, [4]> var_7283_begin_0 = const()[name = string("op_7283_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7283_end_0 = const()[name = string("op_7283_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7283_end_mask_0 = const()[name = string("op_7283_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7283_cast_fp16 = slice_by_index(begin = var_7283_begin_0, end = var_7283_end_0, end_mask = var_7283_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_7283_cast_fp16")];
            tensor<int32, [4]> var_7298_begin_0 = const()[name = string("op_7298_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7298_end_0 = const()[name = string("op_7298_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7298_end_mask_0 = const()[name = string("op_7298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7298_cast_fp16 = slice_by_index(begin = var_7298_begin_0, end = var_7298_end_0, end_mask = var_7298_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_7298_cast_fp16")];
            int32 var_7300 = const()[name = string("op_7300"), 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_7300, interleave = mixed_71_interleave_0, values = (var_7283_cast_fp16, var_7298_cast_fp16))[name = string("mixed_71_cast_fp16")];
            tensor<int32, [4]> var_7304_perm_0 = const()[name = string("op_7304_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7308 = const()[name = string("op_7308"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7304_cast_fp16 = transpose(perm = var_7304_perm_0, x = mixed_71_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 2048]> input_173_cast_fp16 = reshape(shape = var_7308, x = var_7304_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_357_cast_fp16 = add(x = x_341_cast_fp16, y = linear_122_cast_fp16)[name = string("x_357_cast_fp16")];
            bool var_7318 = const()[name = string("op_7318"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7321_cast_fp16 = abs(x = x_357_cast_fp16)[name = string("op_7321_cast_fp16")];
            tensor<int32, [1]> peak_143_axes_0 = const()[name = string("peak_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_143_cast_fp16 = reduce_max(axes = peak_143_axes_0, keep_dims = var_7318, x = var_7321_cast_fp16)[name = string("peak_143_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_359_cast_fp16 = real_div(x = x_357_cast_fp16, y = peak_143_cast_fp16)[name = string("x_359_cast_fp16")];
            fp16 var_7317_promoted_to_fp16 = const()[name = string("op_7317_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7325_cast_fp16 = pow(x = x_359_cast_fp16, y = var_7317_promoted_to_fp16)[name = string("op_7325_cast_fp16")];
            tensor<int32, [1]> var_7327_axes_0 = const()[name = string("op_7327_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7327_keep_dims_0 = const()[name = string("op_7327_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7327_cast_fp16 = reduce_mean(axes = var_7327_axes_0, keep_dims = var_7327_keep_dims_0, x = var_7325_cast_fp16)[name = string("op_7327_cast_fp16")];
            fp16 var_7317_promoted_1_to_fp16 = const()[name = string("op_7317_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7328_cast_fp16 = pow(x = peak_143_cast_fp16, y = var_7317_promoted_1_to_fp16)[name = string("op_7328_cast_fp16")];
            fp32 var_7329_epsilon_0 = const()[name = string("op_7329_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7329_cast_fp16 = inverse(epsilon = var_7329_epsilon_0, x = var_7328_cast_fp16)[name = string("op_7329_cast_fp16")];
            fp16 var_7330_to_fp16 = const()[name = string("op_7330_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7331_cast_fp16 = mul(x = var_7329_cast_fp16, y = var_7330_to_fp16)[name = string("op_7331_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7332_cast_fp16 = add(x = var_7327_cast_fp16, y = var_7331_cast_fp16)[name = string("op_7332_cast_fp16")];
            fp32 var_7333_epsilon_0 = const()[name = string("op_7333_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7333_cast_fp16 = rsqrt(epsilon = var_7333_epsilon_0, x = var_7332_cast_fp16)[name = string("op_7333_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7334_cast_fp16 = mul(x = x_359_cast_fp16, y = var_7333_cast_fp16)[name = string("op_7334_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_7334_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_7339_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_7339_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_7339_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_361_cast_fp16 = add(x = x_357_cast_fp16, y = linear_125_cast_fp16)[name = string("x_361_cast_fp16")];
            bool var_7352 = const()[name = string("op_7352"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7355_cast_fp16 = abs(x = x_361_cast_fp16)[name = string("op_7355_cast_fp16")];
            tensor<int32, [1]> peak_145_axes_0 = const()[name = string("peak_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_145_cast_fp16 = reduce_max(axes = peak_145_axes_0, keep_dims = var_7352, x = var_7355_cast_fp16)[name = string("peak_145_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_363_cast_fp16 = real_div(x = x_361_cast_fp16, y = peak_145_cast_fp16)[name = string("x_363_cast_fp16")];
            fp16 var_7351_promoted_to_fp16 = const()[name = string("op_7351_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7359_cast_fp16 = pow(x = x_363_cast_fp16, y = var_7351_promoted_to_fp16)[name = string("op_7359_cast_fp16")];
            tensor<int32, [1]> var_7361_axes_0 = const()[name = string("op_7361_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7361_keep_dims_0 = const()[name = string("op_7361_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7361_cast_fp16 = reduce_mean(axes = var_7361_axes_0, keep_dims = var_7361_keep_dims_0, x = var_7359_cast_fp16)[name = string("op_7361_cast_fp16")];
            fp16 var_7351_promoted_1_to_fp16 = const()[name = string("op_7351_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7362_cast_fp16 = pow(x = peak_145_cast_fp16, y = var_7351_promoted_1_to_fp16)[name = string("op_7362_cast_fp16")];
            fp32 var_7363_epsilon_0 = const()[name = string("op_7363_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7363_cast_fp16 = inverse(epsilon = var_7363_epsilon_0, x = var_7362_cast_fp16)[name = string("op_7363_cast_fp16")];
            fp16 var_7364_to_fp16 = const()[name = string("op_7364_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7365_cast_fp16 = mul(x = var_7363_cast_fp16, y = var_7364_to_fp16)[name = string("op_7365_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7366_cast_fp16 = add(x = var_7361_cast_fp16, y = var_7365_cast_fp16)[name = string("op_7366_cast_fp16")];
            fp32 var_7367_epsilon_0 = const()[name = string("op_7367_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7367_cast_fp16 = rsqrt(epsilon = var_7367_epsilon_0, x = var_7366_cast_fp16)[name = string("op_7367_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7368_cast_fp16 = mul(x = x_363_cast_fp16, y = var_7367_cast_fp16)[name = string("op_7368_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_7368_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_7377 = const()[name = string("op_7377"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7378_cast_fp16 = reshape(shape = var_7377, x = linear_126_cast_fp16)[name = string("op_7378_cast_fp16")];
            tensor<int32, [4]> x_365_perm_0 = const()[name = string("x_365_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7385 = const()[name = string("op_7385"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_365_cast_fp16 = transpose(perm = x_365_perm_0, x = var_7378_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 16, 128]> var_7388_cast_fp16 = abs(x = x_365_cast_fp16)[name = string("op_7388_cast_fp16")];
            tensor<int32, [1]> peak_147_axes_0 = const()[name = string("peak_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_147_cast_fp16 = reduce_max(axes = peak_147_axes_0, keep_dims = var_7385, x = var_7388_cast_fp16)[name = string("peak_147_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_367_cast_fp16 = real_div(x = x_365_cast_fp16, y = peak_147_cast_fp16)[name = string("x_367_cast_fp16")];
            fp16 var_7384_promoted_to_fp16 = const()[name = string("op_7384_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7392_cast_fp16 = pow(x = x_367_cast_fp16, y = var_7384_promoted_to_fp16)[name = string("op_7392_cast_fp16")];
            tensor<int32, [1]> var_7394_axes_0 = const()[name = string("op_7394_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7394_keep_dims_0 = const()[name = string("op_7394_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7394_cast_fp16 = reduce_mean(axes = var_7394_axes_0, keep_dims = var_7394_keep_dims_0, x = var_7392_cast_fp16)[name = string("op_7394_cast_fp16")];
            fp16 var_7384_promoted_1_to_fp16 = const()[name = string("op_7384_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7395_cast_fp16 = pow(x = peak_147_cast_fp16, y = var_7384_promoted_1_to_fp16)[name = string("op_7395_cast_fp16")];
            fp32 var_7396_epsilon_0 = const()[name = string("op_7396_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7396_cast_fp16 = inverse(epsilon = var_7396_epsilon_0, x = var_7395_cast_fp16)[name = string("op_7396_cast_fp16")];
            fp16 var_7397_to_fp16 = const()[name = string("op_7397_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7398_cast_fp16 = mul(x = var_7396_cast_fp16, y = var_7397_to_fp16)[name = string("op_7398_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7399_cast_fp16 = add(x = var_7394_cast_fp16, y = var_7398_cast_fp16)[name = string("op_7399_cast_fp16")];
            fp32 var_7400_epsilon_0 = const()[name = string("op_7400_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7400_cast_fp16 = rsqrt(epsilon = var_7400_epsilon_0, x = var_7399_cast_fp16)[name = string("op_7400_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7401_cast_fp16 = mul(x = x_367_cast_fp16, y = var_7400_cast_fp16)[name = string("op_7401_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_373_cast_fp16 = mul(x = var_7401_cast_fp16, y = model_layers_18_self_attn_q_norm_weight_to_fp16)[name = string("x_373_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_7410 = const()[name = string("op_7410"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7411_cast_fp16 = reshape(shape = var_7410, x = linear_127_cast_fp16)[name = string("op_7411_cast_fp16")];
            tensor<int32, [4]> x_369_perm_0 = const()[name = string("x_369_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7418 = const()[name = string("op_7418"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_369_cast_fp16 = transpose(perm = x_369_perm_0, x = var_7411_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 8, 16, 128]> var_7421_cast_fp16 = abs(x = x_369_cast_fp16)[name = string("op_7421_cast_fp16")];
            tensor<int32, [1]> peak_149_axes_0 = const()[name = string("peak_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_149_cast_fp16 = reduce_max(axes = peak_149_axes_0, keep_dims = var_7418, x = var_7421_cast_fp16)[name = string("peak_149_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_371_cast_fp16 = real_div(x = x_369_cast_fp16, y = peak_149_cast_fp16)[name = string("x_371_cast_fp16")];
            fp16 var_7417_promoted_to_fp16 = const()[name = string("op_7417_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7425_cast_fp16 = pow(x = x_371_cast_fp16, y = var_7417_promoted_to_fp16)[name = string("op_7425_cast_fp16")];
            tensor<int32, [1]> var_7427_axes_0 = const()[name = string("op_7427_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7427_keep_dims_0 = const()[name = string("op_7427_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7427_cast_fp16 = reduce_mean(axes = var_7427_axes_0, keep_dims = var_7427_keep_dims_0, x = var_7425_cast_fp16)[name = string("op_7427_cast_fp16")];
            fp16 var_7417_promoted_1_to_fp16 = const()[name = string("op_7417_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7428_cast_fp16 = pow(x = peak_149_cast_fp16, y = var_7417_promoted_1_to_fp16)[name = string("op_7428_cast_fp16")];
            fp32 var_7429_epsilon_0 = const()[name = string("op_7429_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7429_cast_fp16 = inverse(epsilon = var_7429_epsilon_0, x = var_7428_cast_fp16)[name = string("op_7429_cast_fp16")];
            fp16 var_7430_to_fp16 = const()[name = string("op_7430_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7431_cast_fp16 = mul(x = var_7429_cast_fp16, y = var_7430_to_fp16)[name = string("op_7431_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7432_cast_fp16 = add(x = var_7427_cast_fp16, y = var_7431_cast_fp16)[name = string("op_7432_cast_fp16")];
            fp32 var_7433_epsilon_0 = const()[name = string("op_7433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7433_cast_fp16 = rsqrt(epsilon = var_7433_epsilon_0, x = var_7432_cast_fp16)[name = string("op_7433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7434_cast_fp16 = mul(x = x_371_cast_fp16, y = var_7433_cast_fp16)[name = string("op_7434_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_375_cast_fp16 = mul(x = var_7434_cast_fp16, y = model_layers_18_self_attn_k_norm_weight_to_fp16)[name = string("x_375_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_7443 = const()[name = string("op_7443"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7444_cast_fp16 = reshape(shape = var_7443, x = linear_128_cast_fp16)[name = string("op_7444_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_373_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_373_cast_fp16)[name = string("second_73_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7458_cast_fp16 = mul(x = first_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7458_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7459_cast_fp16 = mul(x = second_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7459_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7461_cast_fp16 = sub(x = var_7458_cast_fp16, y = var_7459_cast_fp16)[name = string("op_7461_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7462_cast_fp16 = mul(x = second_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7463_cast_fp16 = mul(x = first_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7463_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7465_cast_fp16 = add(x = var_7462_cast_fp16, y = var_7463_cast_fp16)[name = string("op_7465_cast_fp16")];
            int32 var_7467 = const()[name = string("op_7467"), 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_7467, interleave = q_37_interleave_0, values = (var_7461_cast_fp16, var_7465_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_375_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_375_cast_fp16)[name = string("second_75_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7479_cast_fp16 = mul(x = first_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7479_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7480_cast_fp16 = mul(x = second_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7480_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7482_cast_fp16 = sub(x = var_7479_cast_fp16, y = var_7480_cast_fp16)[name = string("op_7482_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7483_cast_fp16 = mul(x = second_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7484_cast_fp16 = mul(x = first_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7484_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7486_cast_fp16 = add(x = var_7483_cast_fp16, y = var_7484_cast_fp16)[name = string("op_7486_cast_fp16")];
            int32 var_7488 = const()[name = string("op_7488"), 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_7488, interleave = k_37_interleave_0, values = (var_7482_cast_fp16, var_7486_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_7499_begin_0 = const()[name = string("op_7499_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7499_end_0 = const()[name = string("op_7499_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7499_end_mask_0 = const()[name = string("op_7499_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7499_stride_0 = const()[name = string("op_7499_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7499_cast_fp16 = slice_by_index(begin = var_7499_begin_0, end = var_7499_end_0, end_mask = var_7499_end_mask_0, stride = var_7499_stride_0, x = q_37_cast_fp16)[name = string("op_7499_cast_fp16")];
            tensor<int32, [4]> var_7509_begin_0 = const()[name = string("op_7509_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7509_end_0 = const()[name = string("op_7509_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7509_end_mask_0 = const()[name = string("op_7509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7509_stride_0 = const()[name = string("op_7509_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7509_cast_fp16 = slice_by_index(begin = var_7509_begin_0, end = var_7509_end_0, end_mask = var_7509_end_mask_0, stride = var_7509_stride_0, x = q_37_cast_fp16)[name = string("op_7509_cast_fp16")];
            int32 var_7511 = const()[name = string("op_7511"), 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_7511, interleave = stacked_37_interleave_0, values = (var_7499_cast_fp16, var_7509_cast_fp16))[name = string("stacked_37_cast_fp16")];
            tensor<int32, [4]> var_7517_begin_0 = const()[name = string("op_7517_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7517_end_0 = const()[name = string("op_7517_end_0"), val = tensor<int32, [4]>([19, 8, 528, 128])];
            tensor<bool, [4]> var_7517_end_mask_0 = const()[name = string("op_7517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7517_cast_fp16 = slice_by_index(begin = var_7517_begin_0, end = var_7517_end_0, end_mask = var_7517_end_mask_0, x = read_state_0)[name = string("op_7517_cast_fp16")];
            tensor<int32, [4]> var_7527_begin_0 = const()[name = string("op_7527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7527_end_0 = const()[name = string("op_7527_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7527_end_mask_0 = const()[name = string("op_7527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7527_cast_fp16 = slice_by_index(begin = var_7527_begin_0, end = var_7527_end_0, end_mask = var_7527_end_mask_0, x = var_7517_cast_fp16)[name = string("op_7527_cast_fp16")];
            bool var_7536_transpose_x_1 = const()[name = string("op_7536_transpose_x_1"), val = bool(false)];
            bool var_7536_transpose_y_1 = const()[name = string("op_7536_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_7536_cast_fp16 = matmul(transpose_x = var_7536_transpose_x_1, transpose_y = var_7536_transpose_y_1, x = var_7527_cast_fp16, y = stacked_37_cast_fp16)[name = string("op_7536_cast_fp16")];
            tensor<int32, [4]> var_7539_perm_0 = const()[name = string("op_7539_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7540_to_fp16 = const()[name = string("op_7540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_7539_cast_fp16 = transpose(perm = var_7539_perm_0, x = var_7536_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 8, 32, 512]> var_7541_cast_fp16 = mul(x = var_7539_cast_fp16, y = var_7540_to_fp16)[name = string("op_7541_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_37_cast_fp16 = add(x = var_7541_cast_fp16, y = cacheMask)[name = string("cache_weights_37_cast_fp16")];
            bool var_7547_transpose_x_1 = const()[name = string("op_7547_transpose_x_1"), val = bool(false)];
            bool var_7547_transpose_y_1 = const()[name = string("op_7547_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7547_cast_fp16 = matmul(transpose_x = var_7547_transpose_x_1, transpose_y = var_7547_transpose_y_1, x = stacked_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_7547_cast_fp16")];
            fp16 var_7548_to_fp16 = const()[name = string("op_7548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7549_cast_fp16 = mul(x = var_7547_cast_fp16, y = var_7548_to_fp16)[name = string("op_7549_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_37_cast_fp16 = add(x = var_7549_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_37_cast_fp16")];
            int32 var_7553 = const()[name = string("op_7553"), val = int32(-1)];
            bool var_7554_interleave_0 = const()[name = string("op_7554_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_7554_cast_fp16 = concat(axis = var_7553, interleave = var_7554_interleave_0, values = (cache_weights_37_cast_fp16, call_weights_37_cast_fp16))[name = string("op_7554_cast_fp16")];
            int32 var_7555 = const()[name = string("op_7555"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_37_cast_fp16 = softmax(axis = var_7555, x = var_7554_cast_fp16)[name = string("shares_37_cast_fp16")];
            tensor<int32, [4]> var_7562_begin_0 = const()[name = string("op_7562_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7562_end_0 = const()[name = string("op_7562_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_7562_end_mask_0 = const()[name = string("op_7562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_7562_cast_fp16 = slice_by_index(begin = var_7562_begin_0, end = var_7562_end_0, end_mask = var_7562_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7562_cast_fp16")];
            tensor<int32, [4]> var_7567_begin_0 = const()[name = string("op_7567_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7567_end_0 = const()[name = string("op_7567_end_0"), val = tensor<int32, [4]>([19, 8, 528, 128])];
            tensor<bool, [4]> var_7567_end_mask_0 = const()[name = string("op_7567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7567_cast_fp16 = slice_by_index(begin = var_7567_begin_0, end = var_7567_end_0, end_mask = var_7567_end_mask_0, x = read_state_1)[name = string("op_7567_cast_fp16")];
            tensor<int32, [4]> var_7577_begin_0 = const()[name = string("op_7577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7577_end_0 = const()[name = string("op_7577_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7577_end_mask_0 = const()[name = string("op_7577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7577_cast_fp16 = slice_by_index(begin = var_7577_begin_0, end = var_7577_end_0, end_mask = var_7577_end_mask_0, x = var_7567_cast_fp16)[name = string("op_7577_cast_fp16")];
            bool var_7583_transpose_x_0 = const()[name = string("op_7583_transpose_x_0"), val = bool(false)];
            bool var_7583_transpose_y_0 = const()[name = string("op_7583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7583_cast_fp16 = matmul(transpose_x = var_7583_transpose_x_0, transpose_y = var_7583_transpose_y_0, x = var_7562_cast_fp16, y = var_7577_cast_fp16)[name = string("op_7583_cast_fp16")];
            tensor<int32, [4]> var_7588_begin_0 = const()[name = string("op_7588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_7588_end_0 = const()[name = string("op_7588_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_7588_end_mask_0 = const()[name = string("op_7588_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7588_cast_fp16 = slice_by_index(begin = var_7588_begin_0, end = var_7588_end_0, end_mask = var_7588_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7588_cast_fp16")];
            bool var_7589_transpose_x_0 = const()[name = string("op_7589_transpose_x_0"), val = bool(false)];
            bool var_7589_transpose_y_0 = const()[name = string("op_7589_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_7444_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, 32, 128]> var_7589_cast_fp16 = matmul(transpose_x = var_7589_transpose_x_0, transpose_y = var_7589_transpose_y_0, x = var_7588_cast_fp16, y = v_37_cast_fp16)[name = string("op_7589_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_73_cast_fp16 = add(x = var_7583_cast_fp16, y = var_7589_cast_fp16)[name = string("mixed_73_cast_fp16")];
            tensor<int32, [4]> var_7606_begin_0 = const()[name = string("op_7606_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7606_end_0 = const()[name = string("op_7606_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7606_end_mask_0 = const()[name = string("op_7606_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7606_cast_fp16 = slice_by_index(begin = var_7606_begin_0, end = var_7606_end_0, end_mask = var_7606_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7606_cast_fp16")];
            tensor<int32, [4]> var_7621_begin_0 = const()[name = string("op_7621_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7621_end_0 = const()[name = string("op_7621_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7621_end_mask_0 = const()[name = string("op_7621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7621_cast_fp16 = slice_by_index(begin = var_7621_begin_0, end = var_7621_end_0, end_mask = var_7621_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7621_cast_fp16")];
            int32 var_7623 = const()[name = string("op_7623"), 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_7623, interleave = mixed_75_interleave_0, values = (var_7606_cast_fp16, var_7621_cast_fp16))[name = string("mixed_75_cast_fp16")];
            tensor<int32, [4]> var_7627_perm_0 = const()[name = string("op_7627_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7631 = const()[name = string("op_7631"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7627_cast_fp16 = transpose(perm = var_7627_perm_0, x = mixed_75_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 2048]> input_183_cast_fp16 = reshape(shape = var_7631, x = var_7627_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_377_cast_fp16 = add(x = x_361_cast_fp16, y = linear_129_cast_fp16)[name = string("x_377_cast_fp16")];
            bool var_7641 = const()[name = string("op_7641"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7644_cast_fp16 = abs(x = x_377_cast_fp16)[name = string("op_7644_cast_fp16")];
            tensor<int32, [1]> peak_151_axes_0 = const()[name = string("peak_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_151_cast_fp16 = reduce_max(axes = peak_151_axes_0, keep_dims = var_7641, x = var_7644_cast_fp16)[name = string("peak_151_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_379_cast_fp16 = real_div(x = x_377_cast_fp16, y = peak_151_cast_fp16)[name = string("x_379_cast_fp16")];
            fp16 var_7640_promoted_to_fp16 = const()[name = string("op_7640_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7648_cast_fp16 = pow(x = x_379_cast_fp16, y = var_7640_promoted_to_fp16)[name = string("op_7648_cast_fp16")];
            tensor<int32, [1]> var_7650_axes_0 = const()[name = string("op_7650_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7650_keep_dims_0 = const()[name = string("op_7650_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7650_cast_fp16 = reduce_mean(axes = var_7650_axes_0, keep_dims = var_7650_keep_dims_0, x = var_7648_cast_fp16)[name = string("op_7650_cast_fp16")];
            fp16 var_7640_promoted_1_to_fp16 = const()[name = string("op_7640_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7651_cast_fp16 = pow(x = peak_151_cast_fp16, y = var_7640_promoted_1_to_fp16)[name = string("op_7651_cast_fp16")];
            fp32 var_7652_epsilon_0 = const()[name = string("op_7652_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7652_cast_fp16 = inverse(epsilon = var_7652_epsilon_0, x = var_7651_cast_fp16)[name = string("op_7652_cast_fp16")];
            fp16 var_7653_to_fp16 = const()[name = string("op_7653_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7654_cast_fp16 = mul(x = var_7652_cast_fp16, y = var_7653_to_fp16)[name = string("op_7654_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7655_cast_fp16 = add(x = var_7650_cast_fp16, y = var_7654_cast_fp16)[name = string("op_7655_cast_fp16")];
            fp32 var_7656_epsilon_0 = const()[name = string("op_7656_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7656_cast_fp16 = rsqrt(epsilon = var_7656_epsilon_0, x = var_7655_cast_fp16)[name = string("op_7656_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7657_cast_fp16 = mul(x = x_379_cast_fp16, y = var_7656_cast_fp16)[name = string("op_7657_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_7657_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_7662_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_7662_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_7662_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_381_cast_fp16 = add(x = x_377_cast_fp16, y = linear_132_cast_fp16)[name = string("x_381_cast_fp16")];
            bool var_7675 = const()[name = string("op_7675"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7678_cast_fp16 = abs(x = x_381_cast_fp16)[name = string("op_7678_cast_fp16")];
            tensor<int32, [1]> peak_153_axes_0 = const()[name = string("peak_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_153_cast_fp16 = reduce_max(axes = peak_153_axes_0, keep_dims = var_7675, x = var_7678_cast_fp16)[name = string("peak_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_383_cast_fp16 = real_div(x = x_381_cast_fp16, y = peak_153_cast_fp16)[name = string("x_383_cast_fp16")];
            fp16 var_7674_promoted_to_fp16 = const()[name = string("op_7674_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7682_cast_fp16 = pow(x = x_383_cast_fp16, y = var_7674_promoted_to_fp16)[name = string("op_7682_cast_fp16")];
            tensor<int32, [1]> var_7684_axes_0 = const()[name = string("op_7684_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7684_keep_dims_0 = const()[name = string("op_7684_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7684_cast_fp16 = reduce_mean(axes = var_7684_axes_0, keep_dims = var_7684_keep_dims_0, x = var_7682_cast_fp16)[name = string("op_7684_cast_fp16")];
            fp16 var_7674_promoted_1_to_fp16 = const()[name = string("op_7674_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7685_cast_fp16 = pow(x = peak_153_cast_fp16, y = var_7674_promoted_1_to_fp16)[name = string("op_7685_cast_fp16")];
            fp32 var_7686_epsilon_0 = const()[name = string("op_7686_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7686_cast_fp16 = inverse(epsilon = var_7686_epsilon_0, x = var_7685_cast_fp16)[name = string("op_7686_cast_fp16")];
            fp16 var_7687_to_fp16 = const()[name = string("op_7687_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7688_cast_fp16 = mul(x = var_7686_cast_fp16, y = var_7687_to_fp16)[name = string("op_7688_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7689_cast_fp16 = add(x = var_7684_cast_fp16, y = var_7688_cast_fp16)[name = string("op_7689_cast_fp16")];
            fp32 var_7690_epsilon_0 = const()[name = string("op_7690_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7690_cast_fp16 = rsqrt(epsilon = var_7690_epsilon_0, x = var_7689_cast_fp16)[name = string("op_7690_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7691_cast_fp16 = mul(x = x_383_cast_fp16, y = var_7690_cast_fp16)[name = string("op_7691_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_7691_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_7700 = const()[name = string("op_7700"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7701_cast_fp16 = reshape(shape = var_7700, x = linear_133_cast_fp16)[name = string("op_7701_cast_fp16")];
            tensor<int32, [4]> x_385_perm_0 = const()[name = string("x_385_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7708 = const()[name = string("op_7708"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_385_cast_fp16 = transpose(perm = x_385_perm_0, x = var_7701_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 16, 16, 128]> var_7711_cast_fp16 = abs(x = x_385_cast_fp16)[name = string("op_7711_cast_fp16")];
            tensor<int32, [1]> peak_155_axes_0 = const()[name = string("peak_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_155_cast_fp16 = reduce_max(axes = peak_155_axes_0, keep_dims = var_7708, x = var_7711_cast_fp16)[name = string("peak_155_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_387_cast_fp16 = real_div(x = x_385_cast_fp16, y = peak_155_cast_fp16)[name = string("x_387_cast_fp16")];
            fp16 var_7707_promoted_to_fp16 = const()[name = string("op_7707_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_7715_cast_fp16 = pow(x = x_387_cast_fp16, y = var_7707_promoted_to_fp16)[name = string("op_7715_cast_fp16")];
            tensor<int32, [1]> var_7717_axes_0 = const()[name = string("op_7717_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7717_keep_dims_0 = const()[name = string("op_7717_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7717_cast_fp16 = reduce_mean(axes = var_7717_axes_0, keep_dims = var_7717_keep_dims_0, x = var_7715_cast_fp16)[name = string("op_7717_cast_fp16")];
            fp16 var_7707_promoted_1_to_fp16 = const()[name = string("op_7707_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_7718_cast_fp16 = pow(x = peak_155_cast_fp16, y = var_7707_promoted_1_to_fp16)[name = string("op_7718_cast_fp16")];
            fp32 var_7719_epsilon_0 = const()[name = string("op_7719_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_7719_cast_fp16 = inverse(epsilon = var_7719_epsilon_0, x = var_7718_cast_fp16)[name = string("op_7719_cast_fp16")];
            fp16 var_7720_to_fp16 = const()[name = string("op_7720_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7721_cast_fp16 = mul(x = var_7719_cast_fp16, y = var_7720_to_fp16)[name = string("op_7721_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_7722_cast_fp16 = add(x = var_7717_cast_fp16, y = var_7721_cast_fp16)[name = string("op_7722_cast_fp16")];
            fp32 var_7723_epsilon_0 = const()[name = string("op_7723_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7723_cast_fp16 = rsqrt(epsilon = var_7723_epsilon_0, x = var_7722_cast_fp16)[name = string("op_7723_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7724_cast_fp16 = mul(x = x_387_cast_fp16, y = var_7723_cast_fp16)[name = string("op_7724_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_393_cast_fp16 = mul(x = var_7724_cast_fp16, y = model_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("x_393_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_7733 = const()[name = string("op_7733"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7734_cast_fp16 = reshape(shape = var_7733, x = linear_134_cast_fp16)[name = string("op_7734_cast_fp16")];
            tensor<int32, [4]> x_389_perm_0 = const()[name = string("x_389_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_7741 = const()[name = string("op_7741"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_389_cast_fp16 = transpose(perm = x_389_perm_0, x = var_7734_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 8, 16, 128]> var_7744_cast_fp16 = abs(x = x_389_cast_fp16)[name = string("op_7744_cast_fp16")];
            tensor<int32, [1]> peak_157_axes_0 = const()[name = string("peak_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_157_cast_fp16 = reduce_max(axes = peak_157_axes_0, keep_dims = var_7741, x = var_7744_cast_fp16)[name = string("peak_157_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_391_cast_fp16 = real_div(x = x_389_cast_fp16, y = peak_157_cast_fp16)[name = string("x_391_cast_fp16")];
            fp16 var_7740_promoted_to_fp16 = const()[name = string("op_7740_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_7748_cast_fp16 = pow(x = x_391_cast_fp16, y = var_7740_promoted_to_fp16)[name = string("op_7748_cast_fp16")];
            tensor<int32, [1]> var_7750_axes_0 = const()[name = string("op_7750_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7750_keep_dims_0 = const()[name = string("op_7750_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7750_cast_fp16 = reduce_mean(axes = var_7750_axes_0, keep_dims = var_7750_keep_dims_0, x = var_7748_cast_fp16)[name = string("op_7750_cast_fp16")];
            fp16 var_7740_promoted_1_to_fp16 = const()[name = string("op_7740_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_7751_cast_fp16 = pow(x = peak_157_cast_fp16, y = var_7740_promoted_1_to_fp16)[name = string("op_7751_cast_fp16")];
            fp32 var_7752_epsilon_0 = const()[name = string("op_7752_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_7752_cast_fp16 = inverse(epsilon = var_7752_epsilon_0, x = var_7751_cast_fp16)[name = string("op_7752_cast_fp16")];
            fp16 var_7753_to_fp16 = const()[name = string("op_7753_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7754_cast_fp16 = mul(x = var_7752_cast_fp16, y = var_7753_to_fp16)[name = string("op_7754_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_7755_cast_fp16 = add(x = var_7750_cast_fp16, y = var_7754_cast_fp16)[name = string("op_7755_cast_fp16")];
            fp32 var_7756_epsilon_0 = const()[name = string("op_7756_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7756_cast_fp16 = rsqrt(epsilon = var_7756_epsilon_0, x = var_7755_cast_fp16)[name = string("op_7756_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7757_cast_fp16 = mul(x = x_391_cast_fp16, y = var_7756_cast_fp16)[name = string("op_7757_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_395_cast_fp16 = mul(x = var_7757_cast_fp16, y = model_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("x_395_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_7766 = const()[name = string("op_7766"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7767_cast_fp16 = reshape(shape = var_7766, x = linear_135_cast_fp16)[name = string("op_7767_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_393_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_393_cast_fp16)[name = string("second_77_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7781_cast_fp16 = mul(x = first_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7781_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7782_cast_fp16 = mul(x = second_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7782_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7784_cast_fp16 = sub(x = var_7781_cast_fp16, y = var_7782_cast_fp16)[name = string("op_7784_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7785_cast_fp16 = mul(x = second_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7785_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7786_cast_fp16 = mul(x = first_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7786_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7788_cast_fp16 = add(x = var_7785_cast_fp16, y = var_7786_cast_fp16)[name = string("op_7788_cast_fp16")];
            int32 var_7790 = const()[name = string("op_7790"), 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_7790, interleave = q_39_interleave_0, values = (var_7784_cast_fp16, var_7788_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_395_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_395_cast_fp16)[name = string("second_79_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7802_cast_fp16 = mul(x = first_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7802_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7803_cast_fp16 = mul(x = second_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7803_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7805_cast_fp16 = sub(x = var_7802_cast_fp16, y = var_7803_cast_fp16)[name = string("op_7805_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7806_cast_fp16 = mul(x = second_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7806_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7807_cast_fp16 = mul(x = first_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7807_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7809_cast_fp16 = add(x = var_7806_cast_fp16, y = var_7807_cast_fp16)[name = string("op_7809_cast_fp16")];
            int32 var_7811 = const()[name = string("op_7811"), 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_7811, interleave = k_39_interleave_0, values = (var_7805_cast_fp16, var_7809_cast_fp16))[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> var_7822_begin_0 = const()[name = string("op_7822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7822_end_0 = const()[name = string("op_7822_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7822_end_mask_0 = const()[name = string("op_7822_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7822_stride_0 = const()[name = string("op_7822_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7822_cast_fp16 = slice_by_index(begin = var_7822_begin_0, end = var_7822_end_0, end_mask = var_7822_end_mask_0, stride = var_7822_stride_0, x = q_39_cast_fp16)[name = string("op_7822_cast_fp16")];
            tensor<int32, [4]> var_7832_begin_0 = const()[name = string("op_7832_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7832_end_0 = const()[name = string("op_7832_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7832_end_mask_0 = const()[name = string("op_7832_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7832_stride_0 = const()[name = string("op_7832_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7832_cast_fp16 = slice_by_index(begin = var_7832_begin_0, end = var_7832_end_0, end_mask = var_7832_end_mask_0, stride = var_7832_stride_0, x = q_39_cast_fp16)[name = string("op_7832_cast_fp16")];
            int32 var_7834 = const()[name = string("op_7834"), 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_7834, interleave = stacked_39_interleave_0, values = (var_7822_cast_fp16, var_7832_cast_fp16))[name = string("stacked_39_cast_fp16")];
            tensor<int32, [4]> var_7840_begin_0 = const()[name = string("op_7840_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7840_end_0 = const()[name = string("op_7840_end_0"), val = tensor<int32, [4]>([20, 8, 528, 128])];
            tensor<bool, [4]> var_7840_end_mask_0 = const()[name = string("op_7840_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7840_cast_fp16 = slice_by_index(begin = var_7840_begin_0, end = var_7840_end_0, end_mask = var_7840_end_mask_0, x = read_state_0)[name = string("op_7840_cast_fp16")];
            tensor<int32, [4]> var_7850_begin_0 = const()[name = string("op_7850_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7850_end_0 = const()[name = string("op_7850_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7850_end_mask_0 = const()[name = string("op_7850_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7850_cast_fp16 = slice_by_index(begin = var_7850_begin_0, end = var_7850_end_0, end_mask = var_7850_end_mask_0, x = var_7840_cast_fp16)[name = string("op_7850_cast_fp16")];
            bool var_7859_transpose_x_1 = const()[name = string("op_7859_transpose_x_1"), val = bool(false)];
            bool var_7859_transpose_y_1 = const()[name = string("op_7859_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_7859_cast_fp16 = matmul(transpose_x = var_7859_transpose_x_1, transpose_y = var_7859_transpose_y_1, x = var_7850_cast_fp16, y = stacked_39_cast_fp16)[name = string("op_7859_cast_fp16")];
            tensor<int32, [4]> var_7862_perm_0 = const()[name = string("op_7862_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7863_to_fp16 = const()[name = string("op_7863_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_7862_cast_fp16 = transpose(perm = var_7862_perm_0, x = var_7859_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 32, 512]> var_7864_cast_fp16 = mul(x = var_7862_cast_fp16, y = var_7863_to_fp16)[name = string("op_7864_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_39_cast_fp16 = add(x = var_7864_cast_fp16, y = cacheMask)[name = string("cache_weights_39_cast_fp16")];
            bool var_7870_transpose_x_1 = const()[name = string("op_7870_transpose_x_1"), val = bool(false)];
            bool var_7870_transpose_y_1 = const()[name = string("op_7870_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7870_cast_fp16 = matmul(transpose_x = var_7870_transpose_x_1, transpose_y = var_7870_transpose_y_1, x = stacked_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_7870_cast_fp16")];
            fp16 var_7871_to_fp16 = const()[name = string("op_7871_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7872_cast_fp16 = mul(x = var_7870_cast_fp16, y = var_7871_to_fp16)[name = string("op_7872_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_39_cast_fp16 = add(x = var_7872_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_39_cast_fp16")];
            int32 var_7876 = const()[name = string("op_7876"), val = int32(-1)];
            bool var_7877_interleave_0 = const()[name = string("op_7877_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_7877_cast_fp16 = concat(axis = var_7876, interleave = var_7877_interleave_0, values = (cache_weights_39_cast_fp16, call_weights_39_cast_fp16))[name = string("op_7877_cast_fp16")];
            int32 var_7878 = const()[name = string("op_7878"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_39_cast_fp16 = softmax(axis = var_7878, x = var_7877_cast_fp16)[name = string("shares_39_cast_fp16")];
            tensor<int32, [4]> var_7885_begin_0 = const()[name = string("op_7885_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7885_end_0 = const()[name = string("op_7885_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_7885_end_mask_0 = const()[name = string("op_7885_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_7885_cast_fp16 = slice_by_index(begin = var_7885_begin_0, end = var_7885_end_0, end_mask = var_7885_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7885_cast_fp16")];
            tensor<int32, [4]> var_7890_begin_0 = const()[name = string("op_7890_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7890_end_0 = const()[name = string("op_7890_end_0"), val = tensor<int32, [4]>([20, 8, 528, 128])];
            tensor<bool, [4]> var_7890_end_mask_0 = const()[name = string("op_7890_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7890_cast_fp16 = slice_by_index(begin = var_7890_begin_0, end = var_7890_end_0, end_mask = var_7890_end_mask_0, x = read_state_1)[name = string("op_7890_cast_fp16")];
            tensor<int32, [4]> var_7900_begin_0 = const()[name = string("op_7900_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7900_end_0 = const()[name = string("op_7900_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7900_end_mask_0 = const()[name = string("op_7900_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7900_cast_fp16 = slice_by_index(begin = var_7900_begin_0, end = var_7900_end_0, end_mask = var_7900_end_mask_0, x = var_7890_cast_fp16)[name = string("op_7900_cast_fp16")];
            bool var_7906_transpose_x_0 = const()[name = string("op_7906_transpose_x_0"), val = bool(false)];
            bool var_7906_transpose_y_0 = const()[name = string("op_7906_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7906_cast_fp16 = matmul(transpose_x = var_7906_transpose_x_0, transpose_y = var_7906_transpose_y_0, x = var_7885_cast_fp16, y = var_7900_cast_fp16)[name = string("op_7906_cast_fp16")];
            tensor<int32, [4]> var_7911_begin_0 = const()[name = string("op_7911_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_7911_end_0 = const()[name = string("op_7911_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_7911_end_mask_0 = const()[name = string("op_7911_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7911_cast_fp16 = slice_by_index(begin = var_7911_begin_0, end = var_7911_end_0, end_mask = var_7911_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7911_cast_fp16")];
            bool var_7912_transpose_x_0 = const()[name = string("op_7912_transpose_x_0"), val = bool(false)];
            bool var_7912_transpose_y_0 = const()[name = string("op_7912_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_7767_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 32, 128]> var_7912_cast_fp16 = matmul(transpose_x = var_7912_transpose_x_0, transpose_y = var_7912_transpose_y_0, x = var_7911_cast_fp16, y = v_39_cast_fp16)[name = string("op_7912_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_77_cast_fp16 = add(x = var_7906_cast_fp16, y = var_7912_cast_fp16)[name = string("mixed_77_cast_fp16")];
            tensor<int32, [4]> var_7929_begin_0 = const()[name = string("op_7929_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7929_end_0 = const()[name = string("op_7929_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7929_end_mask_0 = const()[name = string("op_7929_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7929_cast_fp16 = slice_by_index(begin = var_7929_begin_0, end = var_7929_end_0, end_mask = var_7929_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7929_cast_fp16")];
            tensor<int32, [4]> var_7944_begin_0 = const()[name = string("op_7944_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7944_end_0 = const()[name = string("op_7944_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7944_end_mask_0 = const()[name = string("op_7944_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7944_cast_fp16 = slice_by_index(begin = var_7944_begin_0, end = var_7944_end_0, end_mask = var_7944_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7944_cast_fp16")];
            int32 var_7946 = const()[name = string("op_7946"), 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_7946, interleave = mixed_79_interleave_0, values = (var_7929_cast_fp16, var_7944_cast_fp16))[name = string("mixed_79_cast_fp16")];
            tensor<int32, [4]> var_7950_perm_0 = const()[name = string("op_7950_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7954 = const()[name = string("op_7954"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7950_cast_fp16 = transpose(perm = var_7950_perm_0, x = mixed_79_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 16, 2048]> input_193_cast_fp16 = reshape(shape = var_7954, x = var_7950_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_397_cast_fp16 = add(x = x_381_cast_fp16, y = linear_136_cast_fp16)[name = string("x_397_cast_fp16")];
            bool var_7964 = const()[name = string("op_7964"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_7967_cast_fp16 = abs(x = x_397_cast_fp16)[name = string("op_7967_cast_fp16")];
            tensor<int32, [1]> peak_159_axes_0 = const()[name = string("peak_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_159_cast_fp16 = reduce_max(axes = peak_159_axes_0, keep_dims = var_7964, x = var_7967_cast_fp16)[name = string("peak_159_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_399_cast_fp16 = real_div(x = x_397_cast_fp16, y = peak_159_cast_fp16)[name = string("x_399_cast_fp16")];
            fp16 var_7963_promoted_to_fp16 = const()[name = string("op_7963_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7971_cast_fp16 = pow(x = x_399_cast_fp16, y = var_7963_promoted_to_fp16)[name = string("op_7971_cast_fp16")];
            tensor<int32, [1]> var_7973_axes_0 = const()[name = string("op_7973_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7973_keep_dims_0 = const()[name = string("op_7973_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7973_cast_fp16 = reduce_mean(axes = var_7973_axes_0, keep_dims = var_7973_keep_dims_0, x = var_7971_cast_fp16)[name = string("op_7973_cast_fp16")];
            fp16 var_7963_promoted_1_to_fp16 = const()[name = string("op_7963_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_7974_cast_fp16 = pow(x = peak_159_cast_fp16, y = var_7963_promoted_1_to_fp16)[name = string("op_7974_cast_fp16")];
            fp32 var_7975_epsilon_0 = const()[name = string("op_7975_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_7975_cast_fp16 = inverse(epsilon = var_7975_epsilon_0, x = var_7974_cast_fp16)[name = string("op_7975_cast_fp16")];
            fp16 var_7976_to_fp16 = const()[name = string("op_7976_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7977_cast_fp16 = mul(x = var_7975_cast_fp16, y = var_7976_to_fp16)[name = string("op_7977_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_7978_cast_fp16 = add(x = var_7973_cast_fp16, y = var_7977_cast_fp16)[name = string("op_7978_cast_fp16")];
            fp32 var_7979_epsilon_0 = const()[name = string("op_7979_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7979_cast_fp16 = rsqrt(epsilon = var_7979_epsilon_0, x = var_7978_cast_fp16)[name = string("op_7979_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7980_cast_fp16 = mul(x = x_399_cast_fp16, y = var_7979_cast_fp16)[name = string("op_7980_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_7980_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_7985_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_7985_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_7985_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_401_cast_fp16 = add(x = x_397_cast_fp16, y = linear_139_cast_fp16)[name = string("x_401_cast_fp16")];
            bool var_7998 = const()[name = string("op_7998"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8001_cast_fp16 = abs(x = x_401_cast_fp16)[name = string("op_8001_cast_fp16")];
            tensor<int32, [1]> peak_161_axes_0 = const()[name = string("peak_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_161_cast_fp16 = reduce_max(axes = peak_161_axes_0, keep_dims = var_7998, x = var_8001_cast_fp16)[name = string("peak_161_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_403_cast_fp16 = real_div(x = x_401_cast_fp16, y = peak_161_cast_fp16)[name = string("x_403_cast_fp16")];
            fp16 var_7997_promoted_to_fp16 = const()[name = string("op_7997_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8005_cast_fp16 = pow(x = x_403_cast_fp16, y = var_7997_promoted_to_fp16)[name = string("op_8005_cast_fp16")];
            tensor<int32, [1]> var_8007_axes_0 = const()[name = string("op_8007_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8007_keep_dims_0 = const()[name = string("op_8007_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8007_cast_fp16 = reduce_mean(axes = var_8007_axes_0, keep_dims = var_8007_keep_dims_0, x = var_8005_cast_fp16)[name = string("op_8007_cast_fp16")];
            fp16 var_7997_promoted_1_to_fp16 = const()[name = string("op_7997_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8008_cast_fp16 = pow(x = peak_161_cast_fp16, y = var_7997_promoted_1_to_fp16)[name = string("op_8008_cast_fp16")];
            fp32 var_8009_epsilon_0 = const()[name = string("op_8009_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8009_cast_fp16 = inverse(epsilon = var_8009_epsilon_0, x = var_8008_cast_fp16)[name = string("op_8009_cast_fp16")];
            fp16 var_8010_to_fp16 = const()[name = string("op_8010_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8011_cast_fp16 = mul(x = var_8009_cast_fp16, y = var_8010_to_fp16)[name = string("op_8011_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8012_cast_fp16 = add(x = var_8007_cast_fp16, y = var_8011_cast_fp16)[name = string("op_8012_cast_fp16")];
            fp32 var_8013_epsilon_0 = const()[name = string("op_8013_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8013_cast_fp16 = rsqrt(epsilon = var_8013_epsilon_0, x = var_8012_cast_fp16)[name = string("op_8013_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8014_cast_fp16 = mul(x = x_403_cast_fp16, y = var_8013_cast_fp16)[name = string("op_8014_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_8014_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_8023 = const()[name = string("op_8023"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8024_cast_fp16 = reshape(shape = var_8023, x = linear_140_cast_fp16)[name = string("op_8024_cast_fp16")];
            tensor<int32, [4]> x_405_perm_0 = const()[name = string("x_405_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8031 = const()[name = string("op_8031"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_405_cast_fp16 = transpose(perm = x_405_perm_0, x = var_8024_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 16, 128]> var_8034_cast_fp16 = abs(x = x_405_cast_fp16)[name = string("op_8034_cast_fp16")];
            tensor<int32, [1]> peak_163_axes_0 = const()[name = string("peak_163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_163_cast_fp16 = reduce_max(axes = peak_163_axes_0, keep_dims = var_8031, x = var_8034_cast_fp16)[name = string("peak_163_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_407_cast_fp16 = real_div(x = x_405_cast_fp16, y = peak_163_cast_fp16)[name = string("x_407_cast_fp16")];
            fp16 var_8030_promoted_to_fp16 = const()[name = string("op_8030_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8038_cast_fp16 = pow(x = x_407_cast_fp16, y = var_8030_promoted_to_fp16)[name = string("op_8038_cast_fp16")];
            tensor<int32, [1]> var_8040_axes_0 = const()[name = string("op_8040_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8040_keep_dims_0 = const()[name = string("op_8040_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8040_cast_fp16 = reduce_mean(axes = var_8040_axes_0, keep_dims = var_8040_keep_dims_0, x = var_8038_cast_fp16)[name = string("op_8040_cast_fp16")];
            fp16 var_8030_promoted_1_to_fp16 = const()[name = string("op_8030_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8041_cast_fp16 = pow(x = peak_163_cast_fp16, y = var_8030_promoted_1_to_fp16)[name = string("op_8041_cast_fp16")];
            fp32 var_8042_epsilon_0 = const()[name = string("op_8042_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8042_cast_fp16 = inverse(epsilon = var_8042_epsilon_0, x = var_8041_cast_fp16)[name = string("op_8042_cast_fp16")];
            fp16 var_8043_to_fp16 = const()[name = string("op_8043_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8044_cast_fp16 = mul(x = var_8042_cast_fp16, y = var_8043_to_fp16)[name = string("op_8044_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8045_cast_fp16 = add(x = var_8040_cast_fp16, y = var_8044_cast_fp16)[name = string("op_8045_cast_fp16")];
            fp32 var_8046_epsilon_0 = const()[name = string("op_8046_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8046_cast_fp16 = rsqrt(epsilon = var_8046_epsilon_0, x = var_8045_cast_fp16)[name = string("op_8046_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8047_cast_fp16 = mul(x = x_407_cast_fp16, y = var_8046_cast_fp16)[name = string("op_8047_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_413_cast_fp16 = mul(x = var_8047_cast_fp16, y = model_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("x_413_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_8056 = const()[name = string("op_8056"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8057_cast_fp16 = reshape(shape = var_8056, x = linear_141_cast_fp16)[name = string("op_8057_cast_fp16")];
            tensor<int32, [4]> x_409_perm_0 = const()[name = string("x_409_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8064 = const()[name = string("op_8064"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_409_cast_fp16 = transpose(perm = x_409_perm_0, x = var_8057_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, 16, 128]> var_8067_cast_fp16 = abs(x = x_409_cast_fp16)[name = string("op_8067_cast_fp16")];
            tensor<int32, [1]> peak_165_axes_0 = const()[name = string("peak_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_165_cast_fp16 = reduce_max(axes = peak_165_axes_0, keep_dims = var_8064, x = var_8067_cast_fp16)[name = string("peak_165_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_411_cast_fp16 = real_div(x = x_409_cast_fp16, y = peak_165_cast_fp16)[name = string("x_411_cast_fp16")];
            fp16 var_8063_promoted_to_fp16 = const()[name = string("op_8063_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8071_cast_fp16 = pow(x = x_411_cast_fp16, y = var_8063_promoted_to_fp16)[name = string("op_8071_cast_fp16")];
            tensor<int32, [1]> var_8073_axes_0 = const()[name = string("op_8073_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8073_keep_dims_0 = const()[name = string("op_8073_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8073_cast_fp16 = reduce_mean(axes = var_8073_axes_0, keep_dims = var_8073_keep_dims_0, x = var_8071_cast_fp16)[name = string("op_8073_cast_fp16")];
            fp16 var_8063_promoted_1_to_fp16 = const()[name = string("op_8063_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8074_cast_fp16 = pow(x = peak_165_cast_fp16, y = var_8063_promoted_1_to_fp16)[name = string("op_8074_cast_fp16")];
            fp32 var_8075_epsilon_0 = const()[name = string("op_8075_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8075_cast_fp16 = inverse(epsilon = var_8075_epsilon_0, x = var_8074_cast_fp16)[name = string("op_8075_cast_fp16")];
            fp16 var_8076_to_fp16 = const()[name = string("op_8076_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8077_cast_fp16 = mul(x = var_8075_cast_fp16, y = var_8076_to_fp16)[name = string("op_8077_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8078_cast_fp16 = add(x = var_8073_cast_fp16, y = var_8077_cast_fp16)[name = string("op_8078_cast_fp16")];
            fp32 var_8079_epsilon_0 = const()[name = string("op_8079_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8079_cast_fp16 = rsqrt(epsilon = var_8079_epsilon_0, x = var_8078_cast_fp16)[name = string("op_8079_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8080_cast_fp16 = mul(x = x_411_cast_fp16, y = var_8079_cast_fp16)[name = string("op_8080_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_415_cast_fp16 = mul(x = var_8080_cast_fp16, y = model_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("x_415_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_8089 = const()[name = string("op_8089"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8090_cast_fp16 = reshape(shape = var_8089, x = linear_142_cast_fp16)[name = string("op_8090_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_413_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_413_cast_fp16)[name = string("second_81_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8104_cast_fp16 = mul(x = first_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8104_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8105_cast_fp16 = mul(x = second_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8107_cast_fp16 = sub(x = var_8104_cast_fp16, y = var_8105_cast_fp16)[name = string("op_8107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8108_cast_fp16 = mul(x = second_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8108_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8109_cast_fp16 = mul(x = first_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8111_cast_fp16 = add(x = var_8108_cast_fp16, y = var_8109_cast_fp16)[name = string("op_8111_cast_fp16")];
            int32 var_8113 = const()[name = string("op_8113"), 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_8113, interleave = q_41_interleave_0, values = (var_8107_cast_fp16, var_8111_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_415_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_415_cast_fp16)[name = string("second_83_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8125_cast_fp16 = mul(x = first_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8125_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8126_cast_fp16 = mul(x = second_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8126_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8128_cast_fp16 = sub(x = var_8125_cast_fp16, y = var_8126_cast_fp16)[name = string("op_8128_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8129_cast_fp16 = mul(x = second_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8129_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8130_cast_fp16 = mul(x = first_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8130_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8132_cast_fp16 = add(x = var_8129_cast_fp16, y = var_8130_cast_fp16)[name = string("op_8132_cast_fp16")];
            int32 var_8134 = const()[name = string("op_8134"), 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_8134, interleave = k_41_interleave_0, values = (var_8128_cast_fp16, var_8132_cast_fp16))[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_8145_begin_0 = const()[name = string("op_8145_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8145_end_0 = const()[name = string("op_8145_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8145_end_mask_0 = const()[name = string("op_8145_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8145_stride_0 = const()[name = string("op_8145_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8145_cast_fp16 = slice_by_index(begin = var_8145_begin_0, end = var_8145_end_0, end_mask = var_8145_end_mask_0, stride = var_8145_stride_0, x = q_41_cast_fp16)[name = string("op_8145_cast_fp16")];
            tensor<int32, [4]> var_8155_begin_0 = const()[name = string("op_8155_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8155_end_0 = const()[name = string("op_8155_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8155_end_mask_0 = const()[name = string("op_8155_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8155_stride_0 = const()[name = string("op_8155_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8155_cast_fp16 = slice_by_index(begin = var_8155_begin_0, end = var_8155_end_0, end_mask = var_8155_end_mask_0, stride = var_8155_stride_0, x = q_41_cast_fp16)[name = string("op_8155_cast_fp16")];
            int32 var_8157 = const()[name = string("op_8157"), 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_8157, interleave = stacked_41_interleave_0, values = (var_8145_cast_fp16, var_8155_cast_fp16))[name = string("stacked_41_cast_fp16")];
            tensor<int32, [4]> var_8163_begin_0 = const()[name = string("op_8163_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_8163_end_0 = const()[name = string("op_8163_end_0"), val = tensor<int32, [4]>([21, 8, 528, 128])];
            tensor<bool, [4]> var_8163_end_mask_0 = const()[name = string("op_8163_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, x = read_state_0)[name = string("op_8163_cast_fp16")];
            tensor<int32, [4]> var_8173_begin_0 = const()[name = string("op_8173_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8173_end_0 = const()[name = string("op_8173_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8173_end_mask_0 = const()[name = string("op_8173_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8173_cast_fp16 = slice_by_index(begin = var_8173_begin_0, end = var_8173_end_0, end_mask = var_8173_end_mask_0, x = var_8163_cast_fp16)[name = string("op_8173_cast_fp16")];
            bool var_8182_transpose_x_1 = const()[name = string("op_8182_transpose_x_1"), val = bool(false)];
            bool var_8182_transpose_y_1 = const()[name = string("op_8182_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_8182_cast_fp16 = matmul(transpose_x = var_8182_transpose_x_1, transpose_y = var_8182_transpose_y_1, x = var_8173_cast_fp16, y = stacked_41_cast_fp16)[name = string("op_8182_cast_fp16")];
            tensor<int32, [4]> var_8185_perm_0 = const()[name = string("op_8185_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8186_to_fp16 = const()[name = string("op_8186_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_8185_cast_fp16 = transpose(perm = var_8185_perm_0, x = var_8182_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 32, 512]> var_8187_cast_fp16 = mul(x = var_8185_cast_fp16, y = var_8186_to_fp16)[name = string("op_8187_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_41_cast_fp16 = add(x = var_8187_cast_fp16, y = cacheMask)[name = string("cache_weights_41_cast_fp16")];
            bool var_8193_transpose_x_1 = const()[name = string("op_8193_transpose_x_1"), val = bool(false)];
            bool var_8193_transpose_y_1 = const()[name = string("op_8193_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8193_cast_fp16 = matmul(transpose_x = var_8193_transpose_x_1, transpose_y = var_8193_transpose_y_1, x = stacked_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_8193_cast_fp16")];
            fp16 var_8194_to_fp16 = const()[name = string("op_8194_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8195_cast_fp16 = mul(x = var_8193_cast_fp16, y = var_8194_to_fp16)[name = string("op_8195_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_41_cast_fp16 = add(x = var_8195_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_41_cast_fp16")];
            int32 var_8199 = const()[name = string("op_8199"), val = int32(-1)];
            bool var_8200_interleave_0 = const()[name = string("op_8200_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_8200_cast_fp16 = concat(axis = var_8199, interleave = var_8200_interleave_0, values = (cache_weights_41_cast_fp16, call_weights_41_cast_fp16))[name = string("op_8200_cast_fp16")];
            int32 var_8201 = const()[name = string("op_8201"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_41_cast_fp16 = softmax(axis = var_8201, x = var_8200_cast_fp16)[name = string("shares_41_cast_fp16")];
            tensor<int32, [4]> var_8208_begin_0 = const()[name = string("op_8208_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8208_end_0 = const()[name = string("op_8208_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_8208_end_mask_0 = const()[name = string("op_8208_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_8208_cast_fp16 = slice_by_index(begin = var_8208_begin_0, end = var_8208_end_0, end_mask = var_8208_end_mask_0, x = shares_41_cast_fp16)[name = string("op_8208_cast_fp16")];
            tensor<int32, [4]> var_8213_begin_0 = const()[name = string("op_8213_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_8213_end_0 = const()[name = string("op_8213_end_0"), val = tensor<int32, [4]>([21, 8, 528, 128])];
            tensor<bool, [4]> var_8213_end_mask_0 = const()[name = string("op_8213_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8213_cast_fp16 = slice_by_index(begin = var_8213_begin_0, end = var_8213_end_0, end_mask = var_8213_end_mask_0, x = read_state_1)[name = string("op_8213_cast_fp16")];
            tensor<int32, [4]> var_8223_begin_0 = const()[name = string("op_8223_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8223_end_0 = const()[name = string("op_8223_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8223_end_mask_0 = const()[name = string("op_8223_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8223_cast_fp16 = slice_by_index(begin = var_8223_begin_0, end = var_8223_end_0, end_mask = var_8223_end_mask_0, x = var_8213_cast_fp16)[name = string("op_8223_cast_fp16")];
            bool var_8229_transpose_x_0 = const()[name = string("op_8229_transpose_x_0"), val = bool(false)];
            bool var_8229_transpose_y_0 = const()[name = string("op_8229_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8229_cast_fp16 = matmul(transpose_x = var_8229_transpose_x_0, transpose_y = var_8229_transpose_y_0, x = var_8208_cast_fp16, y = var_8223_cast_fp16)[name = string("op_8229_cast_fp16")];
            tensor<int32, [4]> var_8234_begin_0 = const()[name = string("op_8234_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_8234_end_0 = const()[name = string("op_8234_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_8234_end_mask_0 = const()[name = string("op_8234_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8234_cast_fp16 = slice_by_index(begin = var_8234_begin_0, end = var_8234_end_0, end_mask = var_8234_end_mask_0, x = shares_41_cast_fp16)[name = string("op_8234_cast_fp16")];
            bool var_8235_transpose_x_0 = const()[name = string("op_8235_transpose_x_0"), val = bool(false)];
            bool var_8235_transpose_y_0 = const()[name = string("op_8235_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_8090_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 8, 32, 128]> var_8235_cast_fp16 = matmul(transpose_x = var_8235_transpose_x_0, transpose_y = var_8235_transpose_y_0, x = var_8234_cast_fp16, y = v_41_cast_fp16)[name = string("op_8235_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_81_cast_fp16 = add(x = var_8229_cast_fp16, y = var_8235_cast_fp16)[name = string("mixed_81_cast_fp16")];
            tensor<int32, [4]> var_8252_begin_0 = const()[name = string("op_8252_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8252_end_0 = const()[name = string("op_8252_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8252_end_mask_0 = const()[name = string("op_8252_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8252_cast_fp16 = slice_by_index(begin = var_8252_begin_0, end = var_8252_end_0, end_mask = var_8252_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_8252_cast_fp16")];
            tensor<int32, [4]> var_8267_begin_0 = const()[name = string("op_8267_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8267_end_0 = const()[name = string("op_8267_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8267_end_mask_0 = const()[name = string("op_8267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8267_cast_fp16 = slice_by_index(begin = var_8267_begin_0, end = var_8267_end_0, end_mask = var_8267_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_8267_cast_fp16")];
            int32 var_8269 = const()[name = string("op_8269"), 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_8269, interleave = mixed_83_interleave_0, values = (var_8252_cast_fp16, var_8267_cast_fp16))[name = string("mixed_83_cast_fp16")];
            tensor<int32, [4]> var_8273_perm_0 = const()[name = string("op_8273_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8277 = const()[name = string("op_8277"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8273_cast_fp16 = transpose(perm = var_8273_perm_0, x = mixed_83_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 2048]> input_203_cast_fp16 = reshape(shape = var_8277, x = var_8273_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_417_cast_fp16 = add(x = x_401_cast_fp16, y = linear_143_cast_fp16)[name = string("x_417_cast_fp16")];
            bool var_8287 = const()[name = string("op_8287"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8290_cast_fp16 = abs(x = x_417_cast_fp16)[name = string("op_8290_cast_fp16")];
            tensor<int32, [1]> peak_167_axes_0 = const()[name = string("peak_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_167_cast_fp16 = reduce_max(axes = peak_167_axes_0, keep_dims = var_8287, x = var_8290_cast_fp16)[name = string("peak_167_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_419_cast_fp16 = real_div(x = x_417_cast_fp16, y = peak_167_cast_fp16)[name = string("x_419_cast_fp16")];
            fp16 var_8286_promoted_to_fp16 = const()[name = string("op_8286_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8294_cast_fp16 = pow(x = x_419_cast_fp16, y = var_8286_promoted_to_fp16)[name = string("op_8294_cast_fp16")];
            tensor<int32, [1]> var_8296_axes_0 = const()[name = string("op_8296_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8296_keep_dims_0 = const()[name = string("op_8296_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8296_cast_fp16 = reduce_mean(axes = var_8296_axes_0, keep_dims = var_8296_keep_dims_0, x = var_8294_cast_fp16)[name = string("op_8296_cast_fp16")];
            fp16 var_8286_promoted_1_to_fp16 = const()[name = string("op_8286_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8297_cast_fp16 = pow(x = peak_167_cast_fp16, y = var_8286_promoted_1_to_fp16)[name = string("op_8297_cast_fp16")];
            fp32 var_8298_epsilon_0 = const()[name = string("op_8298_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8298_cast_fp16 = inverse(epsilon = var_8298_epsilon_0, x = var_8297_cast_fp16)[name = string("op_8298_cast_fp16")];
            fp16 var_8299_to_fp16 = const()[name = string("op_8299_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8300_cast_fp16 = mul(x = var_8298_cast_fp16, y = var_8299_to_fp16)[name = string("op_8300_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8301_cast_fp16 = add(x = var_8296_cast_fp16, y = var_8300_cast_fp16)[name = string("op_8301_cast_fp16")];
            fp32 var_8302_epsilon_0 = const()[name = string("op_8302_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8302_cast_fp16 = rsqrt(epsilon = var_8302_epsilon_0, x = var_8301_cast_fp16)[name = string("op_8302_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8303_cast_fp16 = mul(x = x_419_cast_fp16, y = var_8302_cast_fp16)[name = string("op_8303_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_8303_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_8308_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_8308_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_8308_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_421_cast_fp16 = add(x = x_417_cast_fp16, y = linear_146_cast_fp16)[name = string("x_421_cast_fp16")];
            bool var_8321 = const()[name = string("op_8321"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8324_cast_fp16 = abs(x = x_421_cast_fp16)[name = string("op_8324_cast_fp16")];
            tensor<int32, [1]> peak_169_axes_0 = const()[name = string("peak_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_169_cast_fp16 = reduce_max(axes = peak_169_axes_0, keep_dims = var_8321, x = var_8324_cast_fp16)[name = string("peak_169_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_423_cast_fp16 = real_div(x = x_421_cast_fp16, y = peak_169_cast_fp16)[name = string("x_423_cast_fp16")];
            fp16 var_8320_promoted_to_fp16 = const()[name = string("op_8320_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8328_cast_fp16 = pow(x = x_423_cast_fp16, y = var_8320_promoted_to_fp16)[name = string("op_8328_cast_fp16")];
            tensor<int32, [1]> var_8330_axes_0 = const()[name = string("op_8330_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8330_keep_dims_0 = const()[name = string("op_8330_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8330_cast_fp16 = reduce_mean(axes = var_8330_axes_0, keep_dims = var_8330_keep_dims_0, x = var_8328_cast_fp16)[name = string("op_8330_cast_fp16")];
            fp16 var_8320_promoted_1_to_fp16 = const()[name = string("op_8320_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8331_cast_fp16 = pow(x = peak_169_cast_fp16, y = var_8320_promoted_1_to_fp16)[name = string("op_8331_cast_fp16")];
            fp32 var_8332_epsilon_0 = const()[name = string("op_8332_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8332_cast_fp16 = inverse(epsilon = var_8332_epsilon_0, x = var_8331_cast_fp16)[name = string("op_8332_cast_fp16")];
            fp16 var_8333_to_fp16 = const()[name = string("op_8333_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8334_cast_fp16 = mul(x = var_8332_cast_fp16, y = var_8333_to_fp16)[name = string("op_8334_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8335_cast_fp16 = add(x = var_8330_cast_fp16, y = var_8334_cast_fp16)[name = string("op_8335_cast_fp16")];
            fp32 var_8336_epsilon_0 = const()[name = string("op_8336_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8336_cast_fp16 = rsqrt(epsilon = var_8336_epsilon_0, x = var_8335_cast_fp16)[name = string("op_8336_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8337_cast_fp16 = mul(x = x_423_cast_fp16, y = var_8336_cast_fp16)[name = string("op_8337_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_8337_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_8346 = const()[name = string("op_8346"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8347_cast_fp16 = reshape(shape = var_8346, x = linear_147_cast_fp16)[name = string("op_8347_cast_fp16")];
            tensor<int32, [4]> x_425_perm_0 = const()[name = string("x_425_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8354 = const()[name = string("op_8354"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_425_cast_fp16 = transpose(perm = x_425_perm_0, x = var_8347_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 16, 128]> var_8357_cast_fp16 = abs(x = x_425_cast_fp16)[name = string("op_8357_cast_fp16")];
            tensor<int32, [1]> peak_171_axes_0 = const()[name = string("peak_171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_171_cast_fp16 = reduce_max(axes = peak_171_axes_0, keep_dims = var_8354, x = var_8357_cast_fp16)[name = string("peak_171_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_427_cast_fp16 = real_div(x = x_425_cast_fp16, y = peak_171_cast_fp16)[name = string("x_427_cast_fp16")];
            fp16 var_8353_promoted_to_fp16 = const()[name = string("op_8353_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8361_cast_fp16 = pow(x = x_427_cast_fp16, y = var_8353_promoted_to_fp16)[name = string("op_8361_cast_fp16")];
            tensor<int32, [1]> var_8363_axes_0 = const()[name = string("op_8363_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8363_keep_dims_0 = const()[name = string("op_8363_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8363_cast_fp16 = reduce_mean(axes = var_8363_axes_0, keep_dims = var_8363_keep_dims_0, x = var_8361_cast_fp16)[name = string("op_8363_cast_fp16")];
            fp16 var_8353_promoted_1_to_fp16 = const()[name = string("op_8353_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8364_cast_fp16 = pow(x = peak_171_cast_fp16, y = var_8353_promoted_1_to_fp16)[name = string("op_8364_cast_fp16")];
            fp32 var_8365_epsilon_0 = const()[name = string("op_8365_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8365_cast_fp16 = inverse(epsilon = var_8365_epsilon_0, x = var_8364_cast_fp16)[name = string("op_8365_cast_fp16")];
            fp16 var_8366_to_fp16 = const()[name = string("op_8366_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8367_cast_fp16 = mul(x = var_8365_cast_fp16, y = var_8366_to_fp16)[name = string("op_8367_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8368_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8367_cast_fp16)[name = string("op_8368_cast_fp16")];
            fp32 var_8369_epsilon_0 = const()[name = string("op_8369_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8369_cast_fp16 = rsqrt(epsilon = var_8369_epsilon_0, x = var_8368_cast_fp16)[name = string("op_8369_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8370_cast_fp16 = mul(x = x_427_cast_fp16, y = var_8369_cast_fp16)[name = string("op_8370_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_433_cast_fp16 = mul(x = var_8370_cast_fp16, y = model_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("x_433_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_8379 = const()[name = string("op_8379"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8380_cast_fp16 = reshape(shape = var_8379, x = linear_148_cast_fp16)[name = string("op_8380_cast_fp16")];
            tensor<int32, [4]> x_429_perm_0 = const()[name = string("x_429_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8387 = const()[name = string("op_8387"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_429_cast_fp16 = transpose(perm = x_429_perm_0, x = var_8380_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 16, 128]> var_8390_cast_fp16 = abs(x = x_429_cast_fp16)[name = string("op_8390_cast_fp16")];
            tensor<int32, [1]> peak_173_axes_0 = const()[name = string("peak_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_173_cast_fp16 = reduce_max(axes = peak_173_axes_0, keep_dims = var_8387, x = var_8390_cast_fp16)[name = string("peak_173_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_431_cast_fp16 = real_div(x = x_429_cast_fp16, y = peak_173_cast_fp16)[name = string("x_431_cast_fp16")];
            fp16 var_8386_promoted_to_fp16 = const()[name = string("op_8386_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8394_cast_fp16 = pow(x = x_431_cast_fp16, y = var_8386_promoted_to_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [1]> var_8396_axes_0 = const()[name = string("op_8396_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8396_keep_dims_0 = const()[name = string("op_8396_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8396_cast_fp16 = reduce_mean(axes = var_8396_axes_0, keep_dims = var_8396_keep_dims_0, x = var_8394_cast_fp16)[name = string("op_8396_cast_fp16")];
            fp16 var_8386_promoted_1_to_fp16 = const()[name = string("op_8386_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8397_cast_fp16 = pow(x = peak_173_cast_fp16, y = var_8386_promoted_1_to_fp16)[name = string("op_8397_cast_fp16")];
            fp32 var_8398_epsilon_0 = const()[name = string("op_8398_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8398_cast_fp16 = inverse(epsilon = var_8398_epsilon_0, x = var_8397_cast_fp16)[name = string("op_8398_cast_fp16")];
            fp16 var_8399_to_fp16 = const()[name = string("op_8399_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8400_cast_fp16 = mul(x = var_8398_cast_fp16, y = var_8399_to_fp16)[name = string("op_8400_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8401_cast_fp16 = add(x = var_8396_cast_fp16, y = var_8400_cast_fp16)[name = string("op_8401_cast_fp16")];
            fp32 var_8402_epsilon_0 = const()[name = string("op_8402_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8402_cast_fp16 = rsqrt(epsilon = var_8402_epsilon_0, x = var_8401_cast_fp16)[name = string("op_8402_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8403_cast_fp16 = mul(x = x_431_cast_fp16, y = var_8402_cast_fp16)[name = string("op_8403_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_435_cast_fp16 = mul(x = var_8403_cast_fp16, y = model_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("x_435_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_8412 = const()[name = string("op_8412"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8413_cast_fp16 = reshape(shape = var_8412, x = linear_149_cast_fp16)[name = string("op_8413_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_433_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_433_cast_fp16)[name = string("second_85_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8427_cast_fp16 = mul(x = first_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8427_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8428_cast_fp16 = mul(x = second_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8428_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8430_cast_fp16 = sub(x = var_8427_cast_fp16, y = var_8428_cast_fp16)[name = string("op_8430_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8431_cast_fp16 = mul(x = second_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8431_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8432_cast_fp16 = mul(x = first_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8432_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8434_cast_fp16 = add(x = var_8431_cast_fp16, y = var_8432_cast_fp16)[name = string("op_8434_cast_fp16")];
            int32 var_8436 = const()[name = string("op_8436"), 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_8436, interleave = q_43_interleave_0, values = (var_8430_cast_fp16, var_8434_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_435_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_435_cast_fp16)[name = string("second_87_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8448_cast_fp16 = mul(x = first_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8448_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8449_cast_fp16 = mul(x = second_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8449_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8451_cast_fp16 = sub(x = var_8448_cast_fp16, y = var_8449_cast_fp16)[name = string("op_8451_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8452_cast_fp16 = mul(x = second_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8452_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8453_cast_fp16 = mul(x = first_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8453_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8455_cast_fp16 = add(x = var_8452_cast_fp16, y = var_8453_cast_fp16)[name = string("op_8455_cast_fp16")];
            int32 var_8457 = const()[name = string("op_8457"), 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_8457, interleave = k_43_interleave_0, values = (var_8451_cast_fp16, var_8455_cast_fp16))[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> var_8468_begin_0 = const()[name = string("op_8468_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8468_end_0 = const()[name = string("op_8468_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8468_end_mask_0 = const()[name = string("op_8468_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8468_stride_0 = const()[name = string("op_8468_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8468_cast_fp16 = slice_by_index(begin = var_8468_begin_0, end = var_8468_end_0, end_mask = var_8468_end_mask_0, stride = var_8468_stride_0, x = q_43_cast_fp16)[name = string("op_8468_cast_fp16")];
            tensor<int32, [4]> var_8478_begin_0 = const()[name = string("op_8478_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8478_end_0 = const()[name = string("op_8478_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8478_end_mask_0 = const()[name = string("op_8478_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8478_stride_0 = const()[name = string("op_8478_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8478_cast_fp16 = slice_by_index(begin = var_8478_begin_0, end = var_8478_end_0, end_mask = var_8478_end_mask_0, stride = var_8478_stride_0, x = q_43_cast_fp16)[name = string("op_8478_cast_fp16")];
            int32 var_8480 = const()[name = string("op_8480"), 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_8480, interleave = stacked_43_interleave_0, values = (var_8468_cast_fp16, var_8478_cast_fp16))[name = string("stacked_43_cast_fp16")];
            tensor<int32, [4]> var_8486_begin_0 = const()[name = string("op_8486_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_8486_end_0 = const()[name = string("op_8486_end_0"), val = tensor<int32, [4]>([22, 8, 528, 128])];
            tensor<bool, [4]> var_8486_end_mask_0 = const()[name = string("op_8486_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8486_cast_fp16 = slice_by_index(begin = var_8486_begin_0, end = var_8486_end_0, end_mask = var_8486_end_mask_0, x = read_state_0)[name = string("op_8486_cast_fp16")];
            tensor<int32, [4]> var_8496_begin_0 = const()[name = string("op_8496_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8496_end_0 = const()[name = string("op_8496_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8496_end_mask_0 = const()[name = string("op_8496_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8496_cast_fp16 = slice_by_index(begin = var_8496_begin_0, end = var_8496_end_0, end_mask = var_8496_end_mask_0, x = var_8486_cast_fp16)[name = string("op_8496_cast_fp16")];
            bool var_8505_transpose_x_1 = const()[name = string("op_8505_transpose_x_1"), val = bool(false)];
            bool var_8505_transpose_y_1 = const()[name = string("op_8505_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_8505_cast_fp16 = matmul(transpose_x = var_8505_transpose_x_1, transpose_y = var_8505_transpose_y_1, x = var_8496_cast_fp16, y = stacked_43_cast_fp16)[name = string("op_8505_cast_fp16")];
            tensor<int32, [4]> var_8508_perm_0 = const()[name = string("op_8508_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8509_to_fp16 = const()[name = string("op_8509_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_8508_cast_fp16 = transpose(perm = var_8508_perm_0, x = var_8505_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 32, 512]> var_8510_cast_fp16 = mul(x = var_8508_cast_fp16, y = var_8509_to_fp16)[name = string("op_8510_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_43_cast_fp16 = add(x = var_8510_cast_fp16, y = cacheMask)[name = string("cache_weights_43_cast_fp16")];
            bool var_8516_transpose_x_1 = const()[name = string("op_8516_transpose_x_1"), val = bool(false)];
            bool var_8516_transpose_y_1 = const()[name = string("op_8516_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8516_cast_fp16 = matmul(transpose_x = var_8516_transpose_x_1, transpose_y = var_8516_transpose_y_1, x = stacked_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_8516_cast_fp16")];
            fp16 var_8517_to_fp16 = const()[name = string("op_8517_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8518_cast_fp16 = mul(x = var_8516_cast_fp16, y = var_8517_to_fp16)[name = string("op_8518_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_43_cast_fp16 = add(x = var_8518_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_43_cast_fp16")];
            int32 var_8522 = const()[name = string("op_8522"), val = int32(-1)];
            bool var_8523_interleave_0 = const()[name = string("op_8523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_8523_cast_fp16 = concat(axis = var_8522, interleave = var_8523_interleave_0, values = (cache_weights_43_cast_fp16, call_weights_43_cast_fp16))[name = string("op_8523_cast_fp16")];
            int32 var_8524 = const()[name = string("op_8524"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_43_cast_fp16 = softmax(axis = var_8524, x = var_8523_cast_fp16)[name = string("shares_43_cast_fp16")];
            tensor<int32, [4]> var_8531_begin_0 = const()[name = string("op_8531_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8531_end_0 = const()[name = string("op_8531_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_8531_end_mask_0 = const()[name = string("op_8531_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_8531_cast_fp16 = slice_by_index(begin = var_8531_begin_0, end = var_8531_end_0, end_mask = var_8531_end_mask_0, x = shares_43_cast_fp16)[name = string("op_8531_cast_fp16")];
            tensor<int32, [4]> var_8536_begin_0 = const()[name = string("op_8536_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_8536_end_0 = const()[name = string("op_8536_end_0"), val = tensor<int32, [4]>([22, 8, 528, 128])];
            tensor<bool, [4]> var_8536_end_mask_0 = const()[name = string("op_8536_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8536_cast_fp16 = slice_by_index(begin = var_8536_begin_0, end = var_8536_end_0, end_mask = var_8536_end_mask_0, x = read_state_1)[name = string("op_8536_cast_fp16")];
            tensor<int32, [4]> var_8546_begin_0 = const()[name = string("op_8546_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8546_end_0 = const()[name = string("op_8546_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8546_end_mask_0 = const()[name = string("op_8546_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8546_cast_fp16 = slice_by_index(begin = var_8546_begin_0, end = var_8546_end_0, end_mask = var_8546_end_mask_0, x = var_8536_cast_fp16)[name = string("op_8546_cast_fp16")];
            bool var_8552_transpose_x_0 = const()[name = string("op_8552_transpose_x_0"), val = bool(false)];
            bool var_8552_transpose_y_0 = const()[name = string("op_8552_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8552_cast_fp16 = matmul(transpose_x = var_8552_transpose_x_0, transpose_y = var_8552_transpose_y_0, x = var_8531_cast_fp16, y = var_8546_cast_fp16)[name = string("op_8552_cast_fp16")];
            tensor<int32, [4]> var_8557_begin_0 = const()[name = string("op_8557_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_8557_end_0 = const()[name = string("op_8557_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_8557_end_mask_0 = const()[name = string("op_8557_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8557_cast_fp16 = slice_by_index(begin = var_8557_begin_0, end = var_8557_end_0, end_mask = var_8557_end_mask_0, x = shares_43_cast_fp16)[name = string("op_8557_cast_fp16")];
            bool var_8558_transpose_x_0 = const()[name = string("op_8558_transpose_x_0"), val = bool(false)];
            bool var_8558_transpose_y_0 = const()[name = string("op_8558_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_8413_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 8, 32, 128]> var_8558_cast_fp16 = matmul(transpose_x = var_8558_transpose_x_0, transpose_y = var_8558_transpose_y_0, x = var_8557_cast_fp16, y = v_43_cast_fp16)[name = string("op_8558_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_85_cast_fp16 = add(x = var_8552_cast_fp16, y = var_8558_cast_fp16)[name = string("mixed_85_cast_fp16")];
            tensor<int32, [4]> var_8575_begin_0 = const()[name = string("op_8575_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8575_end_0 = const()[name = string("op_8575_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8575_end_mask_0 = const()[name = string("op_8575_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8575_cast_fp16 = slice_by_index(begin = var_8575_begin_0, end = var_8575_end_0, end_mask = var_8575_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_8575_cast_fp16")];
            tensor<int32, [4]> var_8590_begin_0 = const()[name = string("op_8590_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8590_end_0 = const()[name = string("op_8590_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8590_end_mask_0 = const()[name = string("op_8590_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8590_cast_fp16 = slice_by_index(begin = var_8590_begin_0, end = var_8590_end_0, end_mask = var_8590_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_8590_cast_fp16")];
            int32 var_8592 = const()[name = string("op_8592"), 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_8592, interleave = mixed_87_interleave_0, values = (var_8575_cast_fp16, var_8590_cast_fp16))[name = string("mixed_87_cast_fp16")];
            tensor<int32, [4]> var_8596_perm_0 = const()[name = string("op_8596_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8600 = const()[name = string("op_8600"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8596_cast_fp16 = transpose(perm = var_8596_perm_0, x = mixed_87_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 2048]> input_213_cast_fp16 = reshape(shape = var_8600, x = var_8596_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_437_cast_fp16 = add(x = x_421_cast_fp16, y = linear_150_cast_fp16)[name = string("x_437_cast_fp16")];
            bool var_8610 = const()[name = string("op_8610"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8613_cast_fp16 = abs(x = x_437_cast_fp16)[name = string("op_8613_cast_fp16")];
            tensor<int32, [1]> peak_175_axes_0 = const()[name = string("peak_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_175_cast_fp16 = reduce_max(axes = peak_175_axes_0, keep_dims = var_8610, x = var_8613_cast_fp16)[name = string("peak_175_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_439_cast_fp16 = real_div(x = x_437_cast_fp16, y = peak_175_cast_fp16)[name = string("x_439_cast_fp16")];
            fp16 var_8609_promoted_to_fp16 = const()[name = string("op_8609_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8617_cast_fp16 = pow(x = x_439_cast_fp16, y = var_8609_promoted_to_fp16)[name = string("op_8617_cast_fp16")];
            tensor<int32, [1]> var_8619_axes_0 = const()[name = string("op_8619_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8619_keep_dims_0 = const()[name = string("op_8619_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8619_cast_fp16 = reduce_mean(axes = var_8619_axes_0, keep_dims = var_8619_keep_dims_0, x = var_8617_cast_fp16)[name = string("op_8619_cast_fp16")];
            fp16 var_8609_promoted_1_to_fp16 = const()[name = string("op_8609_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8620_cast_fp16 = pow(x = peak_175_cast_fp16, y = var_8609_promoted_1_to_fp16)[name = string("op_8620_cast_fp16")];
            fp32 var_8621_epsilon_0 = const()[name = string("op_8621_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8621_cast_fp16 = inverse(epsilon = var_8621_epsilon_0, x = var_8620_cast_fp16)[name = string("op_8621_cast_fp16")];
            fp16 var_8622_to_fp16 = const()[name = string("op_8622_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8623_cast_fp16 = mul(x = var_8621_cast_fp16, y = var_8622_to_fp16)[name = string("op_8623_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8624_cast_fp16 = add(x = var_8619_cast_fp16, y = var_8623_cast_fp16)[name = string("op_8624_cast_fp16")];
            fp32 var_8625_epsilon_0 = const()[name = string("op_8625_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8625_cast_fp16 = rsqrt(epsilon = var_8625_epsilon_0, x = var_8624_cast_fp16)[name = string("op_8625_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8626_cast_fp16 = mul(x = x_439_cast_fp16, y = var_8625_cast_fp16)[name = string("op_8626_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_8626_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_8631_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_8631_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_8631_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_441_cast_fp16 = add(x = x_437_cast_fp16, y = linear_153_cast_fp16)[name = string("x_441_cast_fp16")];
            bool var_8644 = const()[name = string("op_8644"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8647_cast_fp16 = abs(x = x_441_cast_fp16)[name = string("op_8647_cast_fp16")];
            tensor<int32, [1]> peak_177_axes_0 = const()[name = string("peak_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_177_cast_fp16 = reduce_max(axes = peak_177_axes_0, keep_dims = var_8644, x = var_8647_cast_fp16)[name = string("peak_177_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_443_cast_fp16 = real_div(x = x_441_cast_fp16, y = peak_177_cast_fp16)[name = string("x_443_cast_fp16")];
            fp16 var_8643_promoted_to_fp16 = const()[name = string("op_8643_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8651_cast_fp16 = pow(x = x_443_cast_fp16, y = var_8643_promoted_to_fp16)[name = string("op_8651_cast_fp16")];
            tensor<int32, [1]> var_8653_axes_0 = const()[name = string("op_8653_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8653_keep_dims_0 = const()[name = string("op_8653_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8653_cast_fp16 = reduce_mean(axes = var_8653_axes_0, keep_dims = var_8653_keep_dims_0, x = var_8651_cast_fp16)[name = string("op_8653_cast_fp16")];
            fp16 var_8643_promoted_1_to_fp16 = const()[name = string("op_8643_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8654_cast_fp16 = pow(x = peak_177_cast_fp16, y = var_8643_promoted_1_to_fp16)[name = string("op_8654_cast_fp16")];
            fp32 var_8655_epsilon_0 = const()[name = string("op_8655_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8655_cast_fp16 = inverse(epsilon = var_8655_epsilon_0, x = var_8654_cast_fp16)[name = string("op_8655_cast_fp16")];
            fp16 var_8656_to_fp16 = const()[name = string("op_8656_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8657_cast_fp16 = mul(x = var_8655_cast_fp16, y = var_8656_to_fp16)[name = string("op_8657_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8658_cast_fp16 = add(x = var_8653_cast_fp16, y = var_8657_cast_fp16)[name = string("op_8658_cast_fp16")];
            fp32 var_8659_epsilon_0 = const()[name = string("op_8659_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8659_cast_fp16 = rsqrt(epsilon = var_8659_epsilon_0, x = var_8658_cast_fp16)[name = string("op_8659_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8660_cast_fp16 = mul(x = x_443_cast_fp16, y = var_8659_cast_fp16)[name = string("op_8660_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_8660_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_8669 = const()[name = string("op_8669"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8670_cast_fp16 = reshape(shape = var_8669, x = linear_154_cast_fp16)[name = string("op_8670_cast_fp16")];
            tensor<int32, [4]> x_445_perm_0 = const()[name = string("x_445_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8677 = const()[name = string("op_8677"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_445_cast_fp16 = transpose(perm = x_445_perm_0, x = var_8670_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 16, 128]> var_8680_cast_fp16 = abs(x = x_445_cast_fp16)[name = string("op_8680_cast_fp16")];
            tensor<int32, [1]> peak_179_axes_0 = const()[name = string("peak_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_179_cast_fp16 = reduce_max(axes = peak_179_axes_0, keep_dims = var_8677, x = var_8680_cast_fp16)[name = string("peak_179_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_447_cast_fp16 = real_div(x = x_445_cast_fp16, y = peak_179_cast_fp16)[name = string("x_447_cast_fp16")];
            fp16 var_8676_promoted_to_fp16 = const()[name = string("op_8676_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_8684_cast_fp16 = pow(x = x_447_cast_fp16, y = var_8676_promoted_to_fp16)[name = string("op_8684_cast_fp16")];
            tensor<int32, [1]> var_8686_axes_0 = const()[name = string("op_8686_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8686_keep_dims_0 = const()[name = string("op_8686_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8686_cast_fp16 = reduce_mean(axes = var_8686_axes_0, keep_dims = var_8686_keep_dims_0, x = var_8684_cast_fp16)[name = string("op_8686_cast_fp16")];
            fp16 var_8676_promoted_1_to_fp16 = const()[name = string("op_8676_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_8687_cast_fp16 = pow(x = peak_179_cast_fp16, y = var_8676_promoted_1_to_fp16)[name = string("op_8687_cast_fp16")];
            fp32 var_8688_epsilon_0 = const()[name = string("op_8688_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_8688_cast_fp16 = inverse(epsilon = var_8688_epsilon_0, x = var_8687_cast_fp16)[name = string("op_8688_cast_fp16")];
            fp16 var_8689_to_fp16 = const()[name = string("op_8689_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8690_cast_fp16 = mul(x = var_8688_cast_fp16, y = var_8689_to_fp16)[name = string("op_8690_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_8691_cast_fp16 = add(x = var_8686_cast_fp16, y = var_8690_cast_fp16)[name = string("op_8691_cast_fp16")];
            fp32 var_8692_epsilon_0 = const()[name = string("op_8692_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8692_cast_fp16 = rsqrt(epsilon = var_8692_epsilon_0, x = var_8691_cast_fp16)[name = string("op_8692_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8693_cast_fp16 = mul(x = x_447_cast_fp16, y = var_8692_cast_fp16)[name = string("op_8693_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_453_cast_fp16 = mul(x = var_8693_cast_fp16, y = model_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("x_453_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_8702 = const()[name = string("op_8702"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8703_cast_fp16 = reshape(shape = var_8702, x = linear_155_cast_fp16)[name = string("op_8703_cast_fp16")];
            tensor<int32, [4]> x_449_perm_0 = const()[name = string("x_449_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_8710 = const()[name = string("op_8710"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_449_cast_fp16 = transpose(perm = x_449_perm_0, x = var_8703_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 8, 16, 128]> var_8713_cast_fp16 = abs(x = x_449_cast_fp16)[name = string("op_8713_cast_fp16")];
            tensor<int32, [1]> peak_181_axes_0 = const()[name = string("peak_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_181_cast_fp16 = reduce_max(axes = peak_181_axes_0, keep_dims = var_8710, x = var_8713_cast_fp16)[name = string("peak_181_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_451_cast_fp16 = real_div(x = x_449_cast_fp16, y = peak_181_cast_fp16)[name = string("x_451_cast_fp16")];
            fp16 var_8709_promoted_to_fp16 = const()[name = string("op_8709_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_8717_cast_fp16 = pow(x = x_451_cast_fp16, y = var_8709_promoted_to_fp16)[name = string("op_8717_cast_fp16")];
            tensor<int32, [1]> var_8719_axes_0 = const()[name = string("op_8719_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8719_keep_dims_0 = const()[name = string("op_8719_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8719_cast_fp16 = reduce_mean(axes = var_8719_axes_0, keep_dims = var_8719_keep_dims_0, x = var_8717_cast_fp16)[name = string("op_8719_cast_fp16")];
            fp16 var_8709_promoted_1_to_fp16 = const()[name = string("op_8709_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_8720_cast_fp16 = pow(x = peak_181_cast_fp16, y = var_8709_promoted_1_to_fp16)[name = string("op_8720_cast_fp16")];
            fp32 var_8721_epsilon_0 = const()[name = string("op_8721_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_8721_cast_fp16 = inverse(epsilon = var_8721_epsilon_0, x = var_8720_cast_fp16)[name = string("op_8721_cast_fp16")];
            fp16 var_8722_to_fp16 = const()[name = string("op_8722_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8723_cast_fp16 = mul(x = var_8721_cast_fp16, y = var_8722_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_8724_cast_fp16 = add(x = var_8719_cast_fp16, y = var_8723_cast_fp16)[name = string("op_8724_cast_fp16")];
            fp32 var_8725_epsilon_0 = const()[name = string("op_8725_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8725_cast_fp16 = rsqrt(epsilon = var_8725_epsilon_0, x = var_8724_cast_fp16)[name = string("op_8725_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8726_cast_fp16 = mul(x = x_451_cast_fp16, y = var_8725_cast_fp16)[name = string("op_8726_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_455_cast_fp16 = mul(x = var_8726_cast_fp16, y = model_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("x_455_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_8735 = const()[name = string("op_8735"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8736_cast_fp16 = reshape(shape = var_8735, x = linear_156_cast_fp16)[name = string("op_8736_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_453_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_453_cast_fp16)[name = string("second_89_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8750_cast_fp16 = mul(x = first_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8750_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8751_cast_fp16 = mul(x = second_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8751_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8753_cast_fp16 = sub(x = var_8750_cast_fp16, y = var_8751_cast_fp16)[name = string("op_8753_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8754_cast_fp16 = mul(x = second_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8755_cast_fp16 = mul(x = first_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8755_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8757_cast_fp16 = add(x = var_8754_cast_fp16, y = var_8755_cast_fp16)[name = string("op_8757_cast_fp16")];
            int32 var_8759 = const()[name = string("op_8759"), 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_8759, interleave = q_45_interleave_0, values = (var_8753_cast_fp16, var_8757_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_455_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_455_cast_fp16)[name = string("second_91_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8771_cast_fp16 = mul(x = first_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8771_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8772_cast_fp16 = mul(x = second_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8772_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8774_cast_fp16 = sub(x = var_8771_cast_fp16, y = var_8772_cast_fp16)[name = string("op_8774_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8775_cast_fp16 = mul(x = second_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8775_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8776_cast_fp16 = mul(x = first_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8776_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8778_cast_fp16 = add(x = var_8775_cast_fp16, y = var_8776_cast_fp16)[name = string("op_8778_cast_fp16")];
            int32 var_8780 = const()[name = string("op_8780"), 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_8780, interleave = k_45_interleave_0, values = (var_8774_cast_fp16, var_8778_cast_fp16))[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_8791_begin_0 = const()[name = string("op_8791_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8791_end_0 = const()[name = string("op_8791_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8791_end_mask_0 = const()[name = string("op_8791_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8791_stride_0 = const()[name = string("op_8791_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8791_cast_fp16 = slice_by_index(begin = var_8791_begin_0, end = var_8791_end_0, end_mask = var_8791_end_mask_0, stride = var_8791_stride_0, x = q_45_cast_fp16)[name = string("op_8791_cast_fp16")];
            tensor<int32, [4]> var_8801_begin_0 = const()[name = string("op_8801_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8801_end_0 = const()[name = string("op_8801_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8801_end_mask_0 = const()[name = string("op_8801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8801_stride_0 = const()[name = string("op_8801_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8801_cast_fp16 = slice_by_index(begin = var_8801_begin_0, end = var_8801_end_0, end_mask = var_8801_end_mask_0, stride = var_8801_stride_0, x = q_45_cast_fp16)[name = string("op_8801_cast_fp16")];
            int32 var_8803 = const()[name = string("op_8803"), 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_8803, interleave = stacked_45_interleave_0, values = (var_8791_cast_fp16, var_8801_cast_fp16))[name = string("stacked_45_cast_fp16")];
            tensor<int32, [4]> var_8809_begin_0 = const()[name = string("op_8809_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8809_end_0 = const()[name = string("op_8809_end_0"), val = tensor<int32, [4]>([23, 8, 528, 128])];
            tensor<bool, [4]> var_8809_end_mask_0 = const()[name = string("op_8809_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8809_cast_fp16 = slice_by_index(begin = var_8809_begin_0, end = var_8809_end_0, end_mask = var_8809_end_mask_0, x = read_state_0)[name = string("op_8809_cast_fp16")];
            tensor<int32, [4]> var_8819_begin_0 = const()[name = string("op_8819_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8819_end_0 = const()[name = string("op_8819_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8819_end_mask_0 = const()[name = string("op_8819_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8819_cast_fp16 = slice_by_index(begin = var_8819_begin_0, end = var_8819_end_0, end_mask = var_8819_end_mask_0, x = var_8809_cast_fp16)[name = string("op_8819_cast_fp16")];
            bool var_8828_transpose_x_1 = const()[name = string("op_8828_transpose_x_1"), val = bool(false)];
            bool var_8828_transpose_y_1 = const()[name = string("op_8828_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_8828_cast_fp16 = matmul(transpose_x = var_8828_transpose_x_1, transpose_y = var_8828_transpose_y_1, x = var_8819_cast_fp16, y = stacked_45_cast_fp16)[name = string("op_8828_cast_fp16")];
            tensor<int32, [4]> var_8831_perm_0 = const()[name = string("op_8831_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8832_to_fp16 = const()[name = string("op_8832_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_8831_cast_fp16 = transpose(perm = var_8831_perm_0, x = var_8828_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 8, 32, 512]> var_8833_cast_fp16 = mul(x = var_8831_cast_fp16, y = var_8832_to_fp16)[name = string("op_8833_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_45_cast_fp16 = add(x = var_8833_cast_fp16, y = cacheMask)[name = string("cache_weights_45_cast_fp16")];
            bool var_8839_transpose_x_1 = const()[name = string("op_8839_transpose_x_1"), val = bool(false)];
            bool var_8839_transpose_y_1 = const()[name = string("op_8839_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8839_cast_fp16 = matmul(transpose_x = var_8839_transpose_x_1, transpose_y = var_8839_transpose_y_1, x = stacked_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_8839_cast_fp16")];
            fp16 var_8840_to_fp16 = const()[name = string("op_8840_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8841_cast_fp16 = mul(x = var_8839_cast_fp16, y = var_8840_to_fp16)[name = string("op_8841_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_45_cast_fp16 = add(x = var_8841_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_45_cast_fp16")];
            int32 var_8845 = const()[name = string("op_8845"), val = int32(-1)];
            bool var_8846_interleave_0 = const()[name = string("op_8846_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_8846_cast_fp16 = concat(axis = var_8845, interleave = var_8846_interleave_0, values = (cache_weights_45_cast_fp16, call_weights_45_cast_fp16))[name = string("op_8846_cast_fp16")];
            int32 var_8847 = const()[name = string("op_8847"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_45_cast_fp16 = softmax(axis = var_8847, x = var_8846_cast_fp16)[name = string("shares_45_cast_fp16")];
            tensor<int32, [4]> var_8854_begin_0 = const()[name = string("op_8854_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8854_end_0 = const()[name = string("op_8854_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_8854_end_mask_0 = const()[name = string("op_8854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8854_cast_fp16")];
            tensor<int32, [4]> var_8859_begin_0 = const()[name = string("op_8859_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8859_end_0 = const()[name = string("op_8859_end_0"), val = tensor<int32, [4]>([23, 8, 528, 128])];
            tensor<bool, [4]> var_8859_end_mask_0 = const()[name = string("op_8859_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8859_cast_fp16 = slice_by_index(begin = var_8859_begin_0, end = var_8859_end_0, end_mask = var_8859_end_mask_0, x = read_state_1)[name = string("op_8859_cast_fp16")];
            tensor<int32, [4]> var_8869_begin_0 = const()[name = string("op_8869_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8869_end_0 = const()[name = string("op_8869_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8869_end_mask_0 = const()[name = string("op_8869_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8869_cast_fp16 = slice_by_index(begin = var_8869_begin_0, end = var_8869_end_0, end_mask = var_8869_end_mask_0, x = var_8859_cast_fp16)[name = string("op_8869_cast_fp16")];
            bool var_8875_transpose_x_0 = const()[name = string("op_8875_transpose_x_0"), val = bool(false)];
            bool var_8875_transpose_y_0 = const()[name = string("op_8875_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8875_cast_fp16 = matmul(transpose_x = var_8875_transpose_x_0, transpose_y = var_8875_transpose_y_0, x = var_8854_cast_fp16, y = var_8869_cast_fp16)[name = string("op_8875_cast_fp16")];
            tensor<int32, [4]> var_8880_begin_0 = const()[name = string("op_8880_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_8880_end_0 = const()[name = string("op_8880_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_8880_end_mask_0 = const()[name = string("op_8880_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8880_cast_fp16 = slice_by_index(begin = var_8880_begin_0, end = var_8880_end_0, end_mask = var_8880_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8880_cast_fp16")];
            bool var_8881_transpose_x_0 = const()[name = string("op_8881_transpose_x_0"), val = bool(false)];
            bool var_8881_transpose_y_0 = const()[name = string("op_8881_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_8736_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, 32, 128]> var_8881_cast_fp16 = matmul(transpose_x = var_8881_transpose_x_0, transpose_y = var_8881_transpose_y_0, x = var_8880_cast_fp16, y = v_45_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_89_cast_fp16 = add(x = var_8875_cast_fp16, y = var_8881_cast_fp16)[name = string("mixed_89_cast_fp16")];
            tensor<int32, [4]> var_8898_begin_0 = const()[name = string("op_8898_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8898_end_0 = const()[name = string("op_8898_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8898_end_mask_0 = const()[name = string("op_8898_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8898_cast_fp16 = slice_by_index(begin = var_8898_begin_0, end = var_8898_end_0, end_mask = var_8898_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8898_cast_fp16")];
            tensor<int32, [4]> var_8913_begin_0 = const()[name = string("op_8913_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8913_end_0 = const()[name = string("op_8913_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8913_end_mask_0 = const()[name = string("op_8913_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8913_cast_fp16 = slice_by_index(begin = var_8913_begin_0, end = var_8913_end_0, end_mask = var_8913_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8913_cast_fp16")];
            int32 var_8915 = const()[name = string("op_8915"), 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_8915, interleave = mixed_91_interleave_0, values = (var_8898_cast_fp16, var_8913_cast_fp16))[name = string("mixed_91_cast_fp16")];
            tensor<int32, [4]> var_8919_perm_0 = const()[name = string("op_8919_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8923 = const()[name = string("op_8923"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8919_cast_fp16 = transpose(perm = var_8919_perm_0, x = mixed_91_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 2048]> input_223_cast_fp16 = reshape(shape = var_8923, x = var_8919_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_457_cast_fp16 = add(x = x_441_cast_fp16, y = linear_157_cast_fp16)[name = string("x_457_cast_fp16")];
            bool var_8933 = const()[name = string("op_8933"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8936_cast_fp16 = abs(x = x_457_cast_fp16)[name = string("op_8936_cast_fp16")];
            tensor<int32, [1]> peak_183_axes_0 = const()[name = string("peak_183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_183_cast_fp16 = reduce_max(axes = peak_183_axes_0, keep_dims = var_8933, x = var_8936_cast_fp16)[name = string("peak_183_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_459_cast_fp16 = real_div(x = x_457_cast_fp16, y = peak_183_cast_fp16)[name = string("x_459_cast_fp16")];
            fp16 var_8932_promoted_to_fp16 = const()[name = string("op_8932_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8940_cast_fp16 = pow(x = x_459_cast_fp16, y = var_8932_promoted_to_fp16)[name = string("op_8940_cast_fp16")];
            tensor<int32, [1]> var_8942_axes_0 = const()[name = string("op_8942_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8942_keep_dims_0 = const()[name = string("op_8942_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8942_cast_fp16 = reduce_mean(axes = var_8942_axes_0, keep_dims = var_8942_keep_dims_0, x = var_8940_cast_fp16)[name = string("op_8942_cast_fp16")];
            fp16 var_8932_promoted_1_to_fp16 = const()[name = string("op_8932_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8943_cast_fp16 = pow(x = peak_183_cast_fp16, y = var_8932_promoted_1_to_fp16)[name = string("op_8943_cast_fp16")];
            fp32 var_8944_epsilon_0 = const()[name = string("op_8944_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8944_cast_fp16 = inverse(epsilon = var_8944_epsilon_0, x = var_8943_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, 1]> var_8946_cast_fp16 = mul(x = var_8944_cast_fp16, y = var_8945_to_fp16)[name = string("op_8946_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8947_cast_fp16 = add(x = var_8942_cast_fp16, y = var_8946_cast_fp16)[name = string("op_8947_cast_fp16")];
            fp32 var_8948_epsilon_0 = const()[name = string("op_8948_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8948_cast_fp16 = rsqrt(epsilon = var_8948_epsilon_0, x = var_8947_cast_fp16)[name = string("op_8948_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8949_cast_fp16 = mul(x = x_459_cast_fp16, y = var_8948_cast_fp16)[name = string("op_8949_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_8949_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_8954_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_8954_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_8954_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_461_cast_fp16 = add(x = x_457_cast_fp16, y = linear_160_cast_fp16)[name = string("x_461_cast_fp16")];
            bool var_8967 = const()[name = string("op_8967"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_8970_cast_fp16 = abs(x = x_461_cast_fp16)[name = string("op_8970_cast_fp16")];
            tensor<int32, [1]> peak_185_axes_0 = const()[name = string("peak_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_185_cast_fp16 = reduce_max(axes = peak_185_axes_0, keep_dims = var_8967, x = var_8970_cast_fp16)[name = string("peak_185_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_463_cast_fp16 = real_div(x = x_461_cast_fp16, y = peak_185_cast_fp16)[name = string("x_463_cast_fp16")];
            fp16 var_8966_promoted_to_fp16 = const()[name = string("op_8966_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8974_cast_fp16 = pow(x = x_463_cast_fp16, y = var_8966_promoted_to_fp16)[name = string("op_8974_cast_fp16")];
            tensor<int32, [1]> var_8976_axes_0 = const()[name = string("op_8976_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8976_keep_dims_0 = const()[name = string("op_8976_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8976_cast_fp16 = reduce_mean(axes = var_8976_axes_0, keep_dims = var_8976_keep_dims_0, x = var_8974_cast_fp16)[name = string("op_8976_cast_fp16")];
            fp16 var_8966_promoted_1_to_fp16 = const()[name = string("op_8966_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_8977_cast_fp16 = pow(x = peak_185_cast_fp16, y = var_8966_promoted_1_to_fp16)[name = string("op_8977_cast_fp16")];
            fp32 var_8978_epsilon_0 = const()[name = string("op_8978_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_8978_cast_fp16 = inverse(epsilon = var_8978_epsilon_0, x = var_8977_cast_fp16)[name = string("op_8978_cast_fp16")];
            fp16 var_8979_to_fp16 = const()[name = string("op_8979_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8980_cast_fp16 = mul(x = var_8978_cast_fp16, y = var_8979_to_fp16)[name = string("op_8980_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_8981_cast_fp16 = add(x = var_8976_cast_fp16, y = var_8980_cast_fp16)[name = string("op_8981_cast_fp16")];
            fp32 var_8982_epsilon_0 = const()[name = string("op_8982_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8982_cast_fp16 = rsqrt(epsilon = var_8982_epsilon_0, x = var_8981_cast_fp16)[name = string("op_8982_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8983_cast_fp16 = mul(x = x_463_cast_fp16, y = var_8982_cast_fp16)[name = string("op_8983_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_8983_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_8992 = const()[name = string("op_8992"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8993_cast_fp16 = reshape(shape = var_8992, x = linear_161_cast_fp16)[name = string("op_8993_cast_fp16")];
            tensor<int32, [4]> x_465_perm_0 = const()[name = string("x_465_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9000 = const()[name = string("op_9000"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_465_cast_fp16 = transpose(perm = x_465_perm_0, x = var_8993_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 16, 16, 128]> var_9003_cast_fp16 = abs(x = x_465_cast_fp16)[name = string("op_9003_cast_fp16")];
            tensor<int32, [1]> peak_187_axes_0 = const()[name = string("peak_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_187_cast_fp16 = reduce_max(axes = peak_187_axes_0, keep_dims = var_9000, x = var_9003_cast_fp16)[name = string("peak_187_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_467_cast_fp16 = real_div(x = x_465_cast_fp16, y = peak_187_cast_fp16)[name = string("x_467_cast_fp16")];
            fp16 var_8999_promoted_to_fp16 = const()[name = string("op_8999_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9007_cast_fp16 = pow(x = x_467_cast_fp16, y = var_8999_promoted_to_fp16)[name = string("op_9007_cast_fp16")];
            tensor<int32, [1]> var_9009_axes_0 = const()[name = string("op_9009_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9009_keep_dims_0 = const()[name = string("op_9009_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9009_cast_fp16 = reduce_mean(axes = var_9009_axes_0, keep_dims = var_9009_keep_dims_0, x = var_9007_cast_fp16)[name = string("op_9009_cast_fp16")];
            fp16 var_8999_promoted_1_to_fp16 = const()[name = string("op_8999_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9010_cast_fp16 = pow(x = peak_187_cast_fp16, y = var_8999_promoted_1_to_fp16)[name = string("op_9010_cast_fp16")];
            fp32 var_9011_epsilon_0 = const()[name = string("op_9011_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9011_cast_fp16 = inverse(epsilon = var_9011_epsilon_0, x = var_9010_cast_fp16)[name = string("op_9011_cast_fp16")];
            fp16 var_9012_to_fp16 = const()[name = string("op_9012_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9013_cast_fp16 = mul(x = var_9011_cast_fp16, y = var_9012_to_fp16)[name = string("op_9013_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9014_cast_fp16 = add(x = var_9009_cast_fp16, y = var_9013_cast_fp16)[name = string("op_9014_cast_fp16")];
            fp32 var_9015_epsilon_0 = const()[name = string("op_9015_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9015_cast_fp16 = rsqrt(epsilon = var_9015_epsilon_0, x = var_9014_cast_fp16)[name = string("op_9015_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9016_cast_fp16 = mul(x = x_467_cast_fp16, y = var_9015_cast_fp16)[name = string("op_9016_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_473_cast_fp16 = mul(x = var_9016_cast_fp16, y = model_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("x_473_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_9025 = const()[name = string("op_9025"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9026_cast_fp16 = reshape(shape = var_9025, x = linear_162_cast_fp16)[name = string("op_9026_cast_fp16")];
            tensor<int32, [4]> x_469_perm_0 = const()[name = string("x_469_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9033 = const()[name = string("op_9033"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_469_cast_fp16 = transpose(perm = x_469_perm_0, x = var_9026_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 16, 128]> var_9036_cast_fp16 = abs(x = x_469_cast_fp16)[name = string("op_9036_cast_fp16")];
            tensor<int32, [1]> peak_189_axes_0 = const()[name = string("peak_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_189_cast_fp16 = reduce_max(axes = peak_189_axes_0, keep_dims = var_9033, x = var_9036_cast_fp16)[name = string("peak_189_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_471_cast_fp16 = real_div(x = x_469_cast_fp16, y = peak_189_cast_fp16)[name = string("x_471_cast_fp16")];
            fp16 var_9032_promoted_to_fp16 = const()[name = string("op_9032_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9040_cast_fp16 = pow(x = x_471_cast_fp16, y = var_9032_promoted_to_fp16)[name = string("op_9040_cast_fp16")];
            tensor<int32, [1]> var_9042_axes_0 = const()[name = string("op_9042_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9042_keep_dims_0 = const()[name = string("op_9042_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9042_cast_fp16 = reduce_mean(axes = var_9042_axes_0, keep_dims = var_9042_keep_dims_0, x = var_9040_cast_fp16)[name = string("op_9042_cast_fp16")];
            fp16 var_9032_promoted_1_to_fp16 = const()[name = string("op_9032_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9043_cast_fp16 = pow(x = peak_189_cast_fp16, y = var_9032_promoted_1_to_fp16)[name = string("op_9043_cast_fp16")];
            fp32 var_9044_epsilon_0 = const()[name = string("op_9044_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9044_cast_fp16 = inverse(epsilon = var_9044_epsilon_0, x = var_9043_cast_fp16)[name = string("op_9044_cast_fp16")];
            fp16 var_9045_to_fp16 = const()[name = string("op_9045_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9046_cast_fp16 = mul(x = var_9044_cast_fp16, y = var_9045_to_fp16)[name = string("op_9046_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9047_cast_fp16 = add(x = var_9042_cast_fp16, y = var_9046_cast_fp16)[name = string("op_9047_cast_fp16")];
            fp32 var_9048_epsilon_0 = const()[name = string("op_9048_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9048_cast_fp16 = rsqrt(epsilon = var_9048_epsilon_0, x = var_9047_cast_fp16)[name = string("op_9048_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9049_cast_fp16 = mul(x = x_471_cast_fp16, y = var_9048_cast_fp16)[name = string("op_9049_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_475_cast_fp16 = mul(x = var_9049_cast_fp16, y = model_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("x_475_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_9058 = const()[name = string("op_9058"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9059_cast_fp16 = reshape(shape = var_9058, x = linear_163_cast_fp16)[name = string("op_9059_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_473_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_473_cast_fp16)[name = string("second_93_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9073_cast_fp16 = mul(x = first_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9073_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9074_cast_fp16 = mul(x = second_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9074_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9076_cast_fp16 = sub(x = var_9073_cast_fp16, y = var_9074_cast_fp16)[name = string("op_9076_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9077_cast_fp16 = mul(x = second_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9077_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9078_cast_fp16 = mul(x = first_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9078_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9080_cast_fp16 = add(x = var_9077_cast_fp16, y = var_9078_cast_fp16)[name = string("op_9080_cast_fp16")];
            int32 var_9082 = const()[name = string("op_9082"), 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_9082, interleave = q_47_interleave_0, values = (var_9076_cast_fp16, var_9080_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_475_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_475_cast_fp16)[name = string("second_95_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9094_cast_fp16 = mul(x = first_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9094_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9095_cast_fp16 = mul(x = second_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9095_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9097_cast_fp16 = sub(x = var_9094_cast_fp16, y = var_9095_cast_fp16)[name = string("op_9097_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9098_cast_fp16 = mul(x = second_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9098_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9099_cast_fp16 = mul(x = first_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9099_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9101_cast_fp16 = add(x = var_9098_cast_fp16, y = var_9099_cast_fp16)[name = string("op_9101_cast_fp16")];
            int32 var_9103 = const()[name = string("op_9103"), 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_9103, interleave = k_47_interleave_0, values = (var_9097_cast_fp16, var_9101_cast_fp16))[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> var_9114_begin_0 = const()[name = string("op_9114_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9114_end_0 = const()[name = string("op_9114_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9114_end_mask_0 = const()[name = string("op_9114_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9114_stride_0 = const()[name = string("op_9114_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9114_cast_fp16 = slice_by_index(begin = var_9114_begin_0, end = var_9114_end_0, end_mask = var_9114_end_mask_0, stride = var_9114_stride_0, x = q_47_cast_fp16)[name = string("op_9114_cast_fp16")];
            tensor<int32, [4]> var_9124_begin_0 = const()[name = string("op_9124_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9124_end_0 = const()[name = string("op_9124_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9124_end_mask_0 = const()[name = string("op_9124_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9124_stride_0 = const()[name = string("op_9124_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9124_cast_fp16 = slice_by_index(begin = var_9124_begin_0, end = var_9124_end_0, end_mask = var_9124_end_mask_0, stride = var_9124_stride_0, x = q_47_cast_fp16)[name = string("op_9124_cast_fp16")];
            int32 var_9126 = const()[name = string("op_9126"), 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_9126, interleave = stacked_47_interleave_0, values = (var_9114_cast_fp16, var_9124_cast_fp16))[name = string("stacked_47_cast_fp16")];
            tensor<int32, [4]> var_9132_begin_0 = const()[name = string("op_9132_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_9132_end_0 = const()[name = string("op_9132_end_0"), val = tensor<int32, [4]>([24, 8, 528, 128])];
            tensor<bool, [4]> var_9132_end_mask_0 = const()[name = string("op_9132_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9132_cast_fp16 = slice_by_index(begin = var_9132_begin_0, end = var_9132_end_0, end_mask = var_9132_end_mask_0, x = read_state_0)[name = string("op_9132_cast_fp16")];
            tensor<int32, [4]> var_9142_begin_0 = const()[name = string("op_9142_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9142_end_0 = const()[name = string("op_9142_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9142_end_mask_0 = const()[name = string("op_9142_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9142_cast_fp16 = slice_by_index(begin = var_9142_begin_0, end = var_9142_end_0, end_mask = var_9142_end_mask_0, x = var_9132_cast_fp16)[name = string("op_9142_cast_fp16")];
            bool var_9151_transpose_x_1 = const()[name = string("op_9151_transpose_x_1"), val = bool(false)];
            bool var_9151_transpose_y_1 = const()[name = string("op_9151_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_9151_cast_fp16 = matmul(transpose_x = var_9151_transpose_x_1, transpose_y = var_9151_transpose_y_1, x = var_9142_cast_fp16, y = stacked_47_cast_fp16)[name = string("op_9151_cast_fp16")];
            tensor<int32, [4]> var_9154_perm_0 = const()[name = string("op_9154_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9155_to_fp16 = const()[name = string("op_9155_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_9154_cast_fp16 = transpose(perm = var_9154_perm_0, x = var_9151_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, 32, 512]> var_9156_cast_fp16 = mul(x = var_9154_cast_fp16, y = var_9155_to_fp16)[name = string("op_9156_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_47_cast_fp16 = add(x = var_9156_cast_fp16, y = cacheMask)[name = string("cache_weights_47_cast_fp16")];
            bool var_9162_transpose_x_1 = const()[name = string("op_9162_transpose_x_1"), val = bool(false)];
            bool var_9162_transpose_y_1 = const()[name = string("op_9162_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9162_cast_fp16 = matmul(transpose_x = var_9162_transpose_x_1, transpose_y = var_9162_transpose_y_1, x = stacked_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_9162_cast_fp16")];
            fp16 var_9163_to_fp16 = const()[name = string("op_9163_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9164_cast_fp16 = mul(x = var_9162_cast_fp16, y = var_9163_to_fp16)[name = string("op_9164_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_47_cast_fp16 = add(x = var_9164_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_47_cast_fp16")];
            int32 var_9168 = const()[name = string("op_9168"), val = int32(-1)];
            bool var_9169_interleave_0 = const()[name = string("op_9169_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_9169_cast_fp16 = concat(axis = var_9168, interleave = var_9169_interleave_0, values = (cache_weights_47_cast_fp16, call_weights_47_cast_fp16))[name = string("op_9169_cast_fp16")];
            int32 var_9170 = const()[name = string("op_9170"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_47_cast_fp16 = softmax(axis = var_9170, x = var_9169_cast_fp16)[name = string("shares_47_cast_fp16")];
            tensor<int32, [4]> var_9177_begin_0 = const()[name = string("op_9177_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9177_end_0 = const()[name = string("op_9177_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_9177_end_mask_0 = const()[name = string("op_9177_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_9177_cast_fp16 = slice_by_index(begin = var_9177_begin_0, end = var_9177_end_0, end_mask = var_9177_end_mask_0, x = shares_47_cast_fp16)[name = string("op_9177_cast_fp16")];
            tensor<int32, [4]> var_9182_begin_0 = const()[name = string("op_9182_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_9182_end_0 = const()[name = string("op_9182_end_0"), val = tensor<int32, [4]>([24, 8, 528, 128])];
            tensor<bool, [4]> var_9182_end_mask_0 = const()[name = string("op_9182_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9182_cast_fp16 = slice_by_index(begin = var_9182_begin_0, end = var_9182_end_0, end_mask = var_9182_end_mask_0, x = read_state_1)[name = string("op_9182_cast_fp16")];
            tensor<int32, [4]> var_9192_begin_0 = const()[name = string("op_9192_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9192_end_0 = const()[name = string("op_9192_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9192_end_mask_0 = const()[name = string("op_9192_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9192_cast_fp16 = slice_by_index(begin = var_9192_begin_0, end = var_9192_end_0, end_mask = var_9192_end_mask_0, x = var_9182_cast_fp16)[name = string("op_9192_cast_fp16")];
            bool var_9198_transpose_x_0 = const()[name = string("op_9198_transpose_x_0"), val = bool(false)];
            bool var_9198_transpose_y_0 = const()[name = string("op_9198_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9198_cast_fp16 = matmul(transpose_x = var_9198_transpose_x_0, transpose_y = var_9198_transpose_y_0, x = var_9177_cast_fp16, y = var_9192_cast_fp16)[name = string("op_9198_cast_fp16")];
            tensor<int32, [4]> var_9203_begin_0 = const()[name = string("op_9203_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_9203_end_0 = const()[name = string("op_9203_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_9203_end_mask_0 = const()[name = string("op_9203_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9203_cast_fp16 = slice_by_index(begin = var_9203_begin_0, end = var_9203_end_0, end_mask = var_9203_end_mask_0, x = shares_47_cast_fp16)[name = string("op_9203_cast_fp16")];
            bool var_9204_transpose_x_0 = const()[name = string("op_9204_transpose_x_0"), val = bool(false)];
            bool var_9204_transpose_y_0 = const()[name = string("op_9204_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_9059_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 8, 32, 128]> var_9204_cast_fp16 = matmul(transpose_x = var_9204_transpose_x_0, transpose_y = var_9204_transpose_y_0, x = var_9203_cast_fp16, y = v_47_cast_fp16)[name = string("op_9204_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_93_cast_fp16 = add(x = var_9198_cast_fp16, y = var_9204_cast_fp16)[name = string("mixed_93_cast_fp16")];
            tensor<int32, [4]> var_9221_begin_0 = const()[name = string("op_9221_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9221_end_0 = const()[name = string("op_9221_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9221_end_mask_0 = const()[name = string("op_9221_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9221_cast_fp16 = slice_by_index(begin = var_9221_begin_0, end = var_9221_end_0, end_mask = var_9221_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_9221_cast_fp16")];
            tensor<int32, [4]> var_9236_begin_0 = const()[name = string("op_9236_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9236_end_0 = const()[name = string("op_9236_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9236_end_mask_0 = const()[name = string("op_9236_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9236_cast_fp16 = slice_by_index(begin = var_9236_begin_0, end = var_9236_end_0, end_mask = var_9236_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_9236_cast_fp16")];
            int32 var_9238 = const()[name = string("op_9238"), 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_9238, interleave = mixed_95_interleave_0, values = (var_9221_cast_fp16, var_9236_cast_fp16))[name = string("mixed_95_cast_fp16")];
            tensor<int32, [4]> var_9242_perm_0 = const()[name = string("op_9242_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9246 = const()[name = string("op_9246"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9242_cast_fp16 = transpose(perm = var_9242_perm_0, x = mixed_95_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 16, 2048]> input_233_cast_fp16 = reshape(shape = var_9246, x = var_9242_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_477_cast_fp16 = add(x = x_461_cast_fp16, y = linear_164_cast_fp16)[name = string("x_477_cast_fp16")];
            bool var_9256 = const()[name = string("op_9256"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9259_cast_fp16 = abs(x = x_477_cast_fp16)[name = string("op_9259_cast_fp16")];
            tensor<int32, [1]> peak_191_axes_0 = const()[name = string("peak_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_191_cast_fp16 = reduce_max(axes = peak_191_axes_0, keep_dims = var_9256, x = var_9259_cast_fp16)[name = string("peak_191_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_479_cast_fp16 = real_div(x = x_477_cast_fp16, y = peak_191_cast_fp16)[name = string("x_479_cast_fp16")];
            fp16 var_9255_promoted_to_fp16 = const()[name = string("op_9255_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9263_cast_fp16 = pow(x = x_479_cast_fp16, y = var_9255_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, 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_9255_promoted_1_to_fp16 = const()[name = string("op_9255_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9266_cast_fp16 = pow(x = peak_191_cast_fp16, y = var_9255_promoted_1_to_fp16)[name = string("op_9266_cast_fp16")];
            fp32 var_9267_epsilon_0 = const()[name = string("op_9267_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9267_cast_fp16 = inverse(epsilon = var_9267_epsilon_0, x = var_9266_cast_fp16)[name = string("op_9267_cast_fp16")];
            fp16 var_9268_to_fp16 = const()[name = string("op_9268_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9269_cast_fp16 = mul(x = var_9267_cast_fp16, y = var_9268_to_fp16)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9270_cast_fp16 = add(x = var_9265_cast_fp16, y = var_9269_cast_fp16)[name = string("op_9270_cast_fp16")];
            fp32 var_9271_epsilon_0 = const()[name = string("op_9271_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9271_cast_fp16 = rsqrt(epsilon = var_9271_epsilon_0, x = var_9270_cast_fp16)[name = string("op_9271_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9272_cast_fp16 = mul(x = x_479_cast_fp16, y = var_9271_cast_fp16)[name = string("op_9272_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_9272_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_9277_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_9277_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_9277_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_481_cast_fp16 = add(x = x_477_cast_fp16, y = linear_167_cast_fp16)[name = string("x_481_cast_fp16")];
            bool var_9290 = const()[name = string("op_9290"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9293_cast_fp16 = abs(x = x_481_cast_fp16)[name = string("op_9293_cast_fp16")];
            tensor<int32, [1]> peak_193_axes_0 = const()[name = string("peak_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_193_cast_fp16 = reduce_max(axes = peak_193_axes_0, keep_dims = var_9290, x = var_9293_cast_fp16)[name = string("peak_193_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_483_cast_fp16 = real_div(x = x_481_cast_fp16, y = peak_193_cast_fp16)[name = string("x_483_cast_fp16")];
            fp16 var_9289_promoted_to_fp16 = const()[name = string("op_9289_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9297_cast_fp16 = pow(x = x_483_cast_fp16, y = var_9289_promoted_to_fp16)[name = string("op_9297_cast_fp16")];
            tensor<int32, [1]> var_9299_axes_0 = const()[name = string("op_9299_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9299_keep_dims_0 = const()[name = string("op_9299_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9299_cast_fp16 = reduce_mean(axes = var_9299_axes_0, keep_dims = var_9299_keep_dims_0, x = var_9297_cast_fp16)[name = string("op_9299_cast_fp16")];
            fp16 var_9289_promoted_1_to_fp16 = const()[name = string("op_9289_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9300_cast_fp16 = pow(x = peak_193_cast_fp16, y = var_9289_promoted_1_to_fp16)[name = string("op_9300_cast_fp16")];
            fp32 var_9301_epsilon_0 = const()[name = string("op_9301_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9301_cast_fp16 = inverse(epsilon = var_9301_epsilon_0, x = var_9300_cast_fp16)[name = string("op_9301_cast_fp16")];
            fp16 var_9302_to_fp16 = const()[name = string("op_9302_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9303_cast_fp16 = mul(x = var_9301_cast_fp16, y = var_9302_to_fp16)[name = string("op_9303_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9304_cast_fp16 = add(x = var_9299_cast_fp16, y = var_9303_cast_fp16)[name = string("op_9304_cast_fp16")];
            fp32 var_9305_epsilon_0 = const()[name = string("op_9305_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9305_cast_fp16 = rsqrt(epsilon = var_9305_epsilon_0, x = var_9304_cast_fp16)[name = string("op_9305_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9306_cast_fp16 = mul(x = x_483_cast_fp16, y = var_9305_cast_fp16)[name = string("op_9306_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_9306_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_9315 = const()[name = string("op_9315"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9316_cast_fp16 = reshape(shape = var_9315, x = linear_168_cast_fp16)[name = string("op_9316_cast_fp16")];
            tensor<int32, [4]> x_485_perm_0 = const()[name = string("x_485_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9323 = const()[name = string("op_9323"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_485_cast_fp16 = transpose(perm = x_485_perm_0, x = var_9316_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 16, 128]> var_9326_cast_fp16 = abs(x = x_485_cast_fp16)[name = string("op_9326_cast_fp16")];
            tensor<int32, [1]> peak_195_axes_0 = const()[name = string("peak_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_195_cast_fp16 = reduce_max(axes = peak_195_axes_0, keep_dims = var_9323, x = var_9326_cast_fp16)[name = string("peak_195_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_487_cast_fp16 = real_div(x = x_485_cast_fp16, y = peak_195_cast_fp16)[name = string("x_487_cast_fp16")];
            fp16 var_9322_promoted_to_fp16 = const()[name = string("op_9322_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9330_cast_fp16 = pow(x = x_487_cast_fp16, y = var_9322_promoted_to_fp16)[name = string("op_9330_cast_fp16")];
            tensor<int32, [1]> var_9332_axes_0 = const()[name = string("op_9332_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9332_keep_dims_0 = const()[name = string("op_9332_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9332_cast_fp16 = reduce_mean(axes = var_9332_axes_0, keep_dims = var_9332_keep_dims_0, x = var_9330_cast_fp16)[name = string("op_9332_cast_fp16")];
            fp16 var_9322_promoted_1_to_fp16 = const()[name = string("op_9322_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9333_cast_fp16 = pow(x = peak_195_cast_fp16, y = var_9322_promoted_1_to_fp16)[name = string("op_9333_cast_fp16")];
            fp32 var_9334_epsilon_0 = const()[name = string("op_9334_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9334_cast_fp16 = inverse(epsilon = var_9334_epsilon_0, x = var_9333_cast_fp16)[name = string("op_9334_cast_fp16")];
            fp16 var_9335_to_fp16 = const()[name = string("op_9335_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9336_cast_fp16 = mul(x = var_9334_cast_fp16, y = var_9335_to_fp16)[name = string("op_9336_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9337_cast_fp16 = add(x = var_9332_cast_fp16, y = var_9336_cast_fp16)[name = string("op_9337_cast_fp16")];
            fp32 var_9338_epsilon_0 = const()[name = string("op_9338_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9338_cast_fp16 = rsqrt(epsilon = var_9338_epsilon_0, x = var_9337_cast_fp16)[name = string("op_9338_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9339_cast_fp16 = mul(x = x_487_cast_fp16, y = var_9338_cast_fp16)[name = string("op_9339_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_493_cast_fp16 = mul(x = var_9339_cast_fp16, y = model_layers_24_self_attn_q_norm_weight_to_fp16)[name = string("x_493_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_9348 = const()[name = string("op_9348"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9349_cast_fp16 = reshape(shape = var_9348, x = linear_169_cast_fp16)[name = string("op_9349_cast_fp16")];
            tensor<int32, [4]> x_489_perm_0 = const()[name = string("x_489_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9356 = const()[name = string("op_9356"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_489_cast_fp16 = transpose(perm = x_489_perm_0, x = var_9349_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, 16, 128]> var_9359_cast_fp16 = abs(x = x_489_cast_fp16)[name = string("op_9359_cast_fp16")];
            tensor<int32, [1]> peak_197_axes_0 = const()[name = string("peak_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_197_cast_fp16 = reduce_max(axes = peak_197_axes_0, keep_dims = var_9356, x = var_9359_cast_fp16)[name = string("peak_197_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_491_cast_fp16 = real_div(x = x_489_cast_fp16, y = peak_197_cast_fp16)[name = string("x_491_cast_fp16")];
            fp16 var_9355_promoted_to_fp16 = const()[name = string("op_9355_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9363_cast_fp16 = pow(x = x_491_cast_fp16, y = var_9355_promoted_to_fp16)[name = string("op_9363_cast_fp16")];
            tensor<int32, [1]> var_9365_axes_0 = const()[name = string("op_9365_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9365_keep_dims_0 = const()[name = string("op_9365_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9365_cast_fp16 = reduce_mean(axes = var_9365_axes_0, keep_dims = var_9365_keep_dims_0, x = var_9363_cast_fp16)[name = string("op_9365_cast_fp16")];
            fp16 var_9355_promoted_1_to_fp16 = const()[name = string("op_9355_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9366_cast_fp16 = pow(x = peak_197_cast_fp16, y = var_9355_promoted_1_to_fp16)[name = string("op_9366_cast_fp16")];
            fp32 var_9367_epsilon_0 = const()[name = string("op_9367_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9367_cast_fp16 = inverse(epsilon = var_9367_epsilon_0, x = var_9366_cast_fp16)[name = string("op_9367_cast_fp16")];
            fp16 var_9368_to_fp16 = const()[name = string("op_9368_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9369_cast_fp16 = mul(x = var_9367_cast_fp16, y = var_9368_to_fp16)[name = string("op_9369_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9370_cast_fp16 = add(x = var_9365_cast_fp16, y = var_9369_cast_fp16)[name = string("op_9370_cast_fp16")];
            fp32 var_9371_epsilon_0 = const()[name = string("op_9371_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9371_cast_fp16 = rsqrt(epsilon = var_9371_epsilon_0, x = var_9370_cast_fp16)[name = string("op_9371_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9372_cast_fp16 = mul(x = x_491_cast_fp16, y = var_9371_cast_fp16)[name = string("op_9372_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_495_cast_fp16 = mul(x = var_9372_cast_fp16, y = model_layers_24_self_attn_k_norm_weight_to_fp16)[name = string("x_495_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_9381 = const()[name = string("op_9381"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9382_cast_fp16 = reshape(shape = var_9381, x = linear_170_cast_fp16)[name = string("op_9382_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_493_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_493_cast_fp16)[name = string("second_97_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9396_cast_fp16 = mul(x = first_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9396_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9397_cast_fp16 = mul(x = second_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9397_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9399_cast_fp16 = sub(x = var_9396_cast_fp16, y = var_9397_cast_fp16)[name = string("op_9399_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9400_cast_fp16 = mul(x = second_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9400_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9401_cast_fp16 = mul(x = first_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9401_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9403_cast_fp16 = add(x = var_9400_cast_fp16, y = var_9401_cast_fp16)[name = string("op_9403_cast_fp16")];
            int32 var_9405 = const()[name = string("op_9405"), 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_9405, interleave = q_49_interleave_0, values = (var_9399_cast_fp16, var_9403_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_495_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_495_cast_fp16)[name = string("second_99_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9417_cast_fp16 = mul(x = first_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9417_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9418_cast_fp16 = mul(x = second_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9418_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9420_cast_fp16 = sub(x = var_9417_cast_fp16, y = var_9418_cast_fp16)[name = string("op_9420_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9421_cast_fp16 = mul(x = second_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9421_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9422_cast_fp16 = mul(x = first_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9422_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9424_cast_fp16 = add(x = var_9421_cast_fp16, y = var_9422_cast_fp16)[name = string("op_9424_cast_fp16")];
            int32 var_9426 = const()[name = string("op_9426"), 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_9426, interleave = k_49_interleave_0, values = (var_9420_cast_fp16, var_9424_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_9437_begin_0 = const()[name = string("op_9437_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9437_end_0 = const()[name = string("op_9437_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9437_end_mask_0 = const()[name = string("op_9437_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9437_stride_0 = const()[name = string("op_9437_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9437_cast_fp16 = slice_by_index(begin = var_9437_begin_0, end = var_9437_end_0, end_mask = var_9437_end_mask_0, stride = var_9437_stride_0, x = q_49_cast_fp16)[name = string("op_9437_cast_fp16")];
            tensor<int32, [4]> var_9447_begin_0 = const()[name = string("op_9447_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9447_end_0 = const()[name = string("op_9447_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9447_end_mask_0 = const()[name = string("op_9447_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9447_stride_0 = const()[name = string("op_9447_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9447_cast_fp16 = slice_by_index(begin = var_9447_begin_0, end = var_9447_end_0, end_mask = var_9447_end_mask_0, stride = var_9447_stride_0, x = q_49_cast_fp16)[name = string("op_9447_cast_fp16")];
            int32 var_9449 = const()[name = string("op_9449"), 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_9449, interleave = stacked_49_interleave_0, values = (var_9437_cast_fp16, var_9447_cast_fp16))[name = string("stacked_49_cast_fp16")];
            tensor<int32, [4]> var_9455_begin_0 = const()[name = string("op_9455_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_9455_end_0 = const()[name = string("op_9455_end_0"), val = tensor<int32, [4]>([25, 8, 528, 128])];
            tensor<bool, [4]> var_9455_end_mask_0 = const()[name = string("op_9455_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9455_cast_fp16 = slice_by_index(begin = var_9455_begin_0, end = var_9455_end_0, end_mask = var_9455_end_mask_0, x = read_state_0)[name = string("op_9455_cast_fp16")];
            tensor<int32, [4]> var_9465_begin_0 = const()[name = string("op_9465_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9465_end_0 = const()[name = string("op_9465_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9465_end_mask_0 = const()[name = string("op_9465_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9465_cast_fp16 = slice_by_index(begin = var_9465_begin_0, end = var_9465_end_0, end_mask = var_9465_end_mask_0, x = var_9455_cast_fp16)[name = string("op_9465_cast_fp16")];
            bool var_9474_transpose_x_1 = const()[name = string("op_9474_transpose_x_1"), val = bool(false)];
            bool var_9474_transpose_y_1 = const()[name = string("op_9474_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_9474_cast_fp16 = matmul(transpose_x = var_9474_transpose_x_1, transpose_y = var_9474_transpose_y_1, x = var_9465_cast_fp16, y = stacked_49_cast_fp16)[name = string("op_9474_cast_fp16")];
            tensor<int32, [4]> var_9477_perm_0 = const()[name = string("op_9477_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9478_to_fp16 = const()[name = string("op_9478_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_9477_cast_fp16 = transpose(perm = var_9477_perm_0, x = var_9474_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 32, 512]> var_9479_cast_fp16 = mul(x = var_9477_cast_fp16, y = var_9478_to_fp16)[name = string("op_9479_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_49_cast_fp16 = add(x = var_9479_cast_fp16, y = cacheMask)[name = string("cache_weights_49_cast_fp16")];
            bool var_9485_transpose_x_1 = const()[name = string("op_9485_transpose_x_1"), val = bool(false)];
            bool var_9485_transpose_y_1 = const()[name = string("op_9485_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9485_cast_fp16 = matmul(transpose_x = var_9485_transpose_x_1, transpose_y = var_9485_transpose_y_1, x = stacked_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_9485_cast_fp16")];
            fp16 var_9486_to_fp16 = const()[name = string("op_9486_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9487_cast_fp16 = mul(x = var_9485_cast_fp16, y = var_9486_to_fp16)[name = string("op_9487_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_49_cast_fp16 = add(x = var_9487_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_49_cast_fp16")];
            int32 var_9491 = const()[name = string("op_9491"), val = int32(-1)];
            bool var_9492_interleave_0 = const()[name = string("op_9492_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_9492_cast_fp16 = concat(axis = var_9491, interleave = var_9492_interleave_0, values = (cache_weights_49_cast_fp16, call_weights_49_cast_fp16))[name = string("op_9492_cast_fp16")];
            int32 var_9493 = const()[name = string("op_9493"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_49_cast_fp16 = softmax(axis = var_9493, x = var_9492_cast_fp16)[name = string("shares_49_cast_fp16")];
            tensor<int32, [4]> var_9500_begin_0 = const()[name = string("op_9500_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9500_end_0 = const()[name = string("op_9500_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_9500_end_mask_0 = const()[name = string("op_9500_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_9500_cast_fp16 = slice_by_index(begin = var_9500_begin_0, end = var_9500_end_0, end_mask = var_9500_end_mask_0, x = shares_49_cast_fp16)[name = string("op_9500_cast_fp16")];
            tensor<int32, [4]> var_9505_begin_0 = const()[name = string("op_9505_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_9505_end_0 = const()[name = string("op_9505_end_0"), val = tensor<int32, [4]>([25, 8, 528, 128])];
            tensor<bool, [4]> var_9505_end_mask_0 = const()[name = string("op_9505_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9505_cast_fp16 = slice_by_index(begin = var_9505_begin_0, end = var_9505_end_0, end_mask = var_9505_end_mask_0, x = read_state_1)[name = string("op_9505_cast_fp16")];
            tensor<int32, [4]> var_9515_begin_0 = const()[name = string("op_9515_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9515_end_0 = const()[name = string("op_9515_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9515_end_mask_0 = const()[name = string("op_9515_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9515_cast_fp16 = slice_by_index(begin = var_9515_begin_0, end = var_9515_end_0, end_mask = var_9515_end_mask_0, x = var_9505_cast_fp16)[name = string("op_9515_cast_fp16")];
            bool var_9521_transpose_x_0 = const()[name = string("op_9521_transpose_x_0"), val = bool(false)];
            bool var_9521_transpose_y_0 = const()[name = string("op_9521_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9521_cast_fp16 = matmul(transpose_x = var_9521_transpose_x_0, transpose_y = var_9521_transpose_y_0, x = var_9500_cast_fp16, y = var_9515_cast_fp16)[name = string("op_9521_cast_fp16")];
            tensor<int32, [4]> var_9526_begin_0 = const()[name = string("op_9526_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_9526_end_0 = const()[name = string("op_9526_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_9526_end_mask_0 = const()[name = string("op_9526_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9526_cast_fp16 = slice_by_index(begin = var_9526_begin_0, end = var_9526_end_0, end_mask = var_9526_end_mask_0, x = shares_49_cast_fp16)[name = string("op_9526_cast_fp16")];
            bool var_9527_transpose_x_0 = const()[name = string("op_9527_transpose_x_0"), val = bool(false)];
            bool var_9527_transpose_y_0 = const()[name = string("op_9527_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_9382_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 32, 128]> var_9527_cast_fp16 = matmul(transpose_x = var_9527_transpose_x_0, transpose_y = var_9527_transpose_y_0, x = var_9526_cast_fp16, y = v_49_cast_fp16)[name = string("op_9527_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_97_cast_fp16 = add(x = var_9521_cast_fp16, y = var_9527_cast_fp16)[name = string("mixed_97_cast_fp16")];
            tensor<int32, [4]> var_9544_begin_0 = const()[name = string("op_9544_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9544_end_0 = const()[name = string("op_9544_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9544_end_mask_0 = const()[name = string("op_9544_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9544_cast_fp16 = slice_by_index(begin = var_9544_begin_0, end = var_9544_end_0, end_mask = var_9544_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_9544_cast_fp16")];
            tensor<int32, [4]> var_9559_begin_0 = const()[name = string("op_9559_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9559_end_0 = const()[name = string("op_9559_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9559_end_mask_0 = const()[name = string("op_9559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9559_cast_fp16 = slice_by_index(begin = var_9559_begin_0, end = var_9559_end_0, end_mask = var_9559_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_9559_cast_fp16")];
            int32 var_9561 = const()[name = string("op_9561"), 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_9561, interleave = mixed_99_interleave_0, values = (var_9544_cast_fp16, var_9559_cast_fp16))[name = string("mixed_99_cast_fp16")];
            tensor<int32, [4]> var_9565_perm_0 = const()[name = string("op_9565_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9569 = const()[name = string("op_9569"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9565_cast_fp16 = transpose(perm = var_9565_perm_0, x = mixed_99_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 2048]> input_243_cast_fp16 = reshape(shape = var_9569, x = var_9565_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_497_cast_fp16 = add(x = x_481_cast_fp16, y = linear_171_cast_fp16)[name = string("x_497_cast_fp16")];
            bool var_9579 = const()[name = string("op_9579"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9582_cast_fp16 = abs(x = x_497_cast_fp16)[name = string("op_9582_cast_fp16")];
            tensor<int32, [1]> peak_199_axes_0 = const()[name = string("peak_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_199_cast_fp16 = reduce_max(axes = peak_199_axes_0, keep_dims = var_9579, x = var_9582_cast_fp16)[name = string("peak_199_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_499_cast_fp16 = real_div(x = x_497_cast_fp16, y = peak_199_cast_fp16)[name = string("x_499_cast_fp16")];
            fp16 var_9578_promoted_to_fp16 = const()[name = string("op_9578_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9586_cast_fp16 = pow(x = x_499_cast_fp16, y = var_9578_promoted_to_fp16)[name = string("op_9586_cast_fp16")];
            tensor<int32, [1]> var_9588_axes_0 = const()[name = string("op_9588_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9588_keep_dims_0 = const()[name = string("op_9588_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9588_cast_fp16 = reduce_mean(axes = var_9588_axes_0, keep_dims = var_9588_keep_dims_0, x = var_9586_cast_fp16)[name = string("op_9588_cast_fp16")];
            fp16 var_9578_promoted_1_to_fp16 = const()[name = string("op_9578_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9589_cast_fp16 = pow(x = peak_199_cast_fp16, y = var_9578_promoted_1_to_fp16)[name = string("op_9589_cast_fp16")];
            fp32 var_9590_epsilon_0 = const()[name = string("op_9590_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9590_cast_fp16 = inverse(epsilon = var_9590_epsilon_0, x = var_9589_cast_fp16)[name = string("op_9590_cast_fp16")];
            fp16 var_9591_to_fp16 = const()[name = string("op_9591_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9592_cast_fp16 = mul(x = var_9590_cast_fp16, y = var_9591_to_fp16)[name = string("op_9592_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9593_cast_fp16 = add(x = var_9588_cast_fp16, y = var_9592_cast_fp16)[name = string("op_9593_cast_fp16")];
            fp32 var_9594_epsilon_0 = const()[name = string("op_9594_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9594_cast_fp16 = rsqrt(epsilon = var_9594_epsilon_0, x = var_9593_cast_fp16)[name = string("op_9594_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9595_cast_fp16 = mul(x = x_499_cast_fp16, y = var_9594_cast_fp16)[name = string("op_9595_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_9595_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_9600_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_9600_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_9600_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_501_cast_fp16 = add(x = x_497_cast_fp16, y = linear_174_cast_fp16)[name = string("x_501_cast_fp16")];
            bool var_9613 = const()[name = string("op_9613"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9616_cast_fp16 = abs(x = x_501_cast_fp16)[name = string("op_9616_cast_fp16")];
            tensor<int32, [1]> peak_201_axes_0 = const()[name = string("peak_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_201_cast_fp16 = reduce_max(axes = peak_201_axes_0, keep_dims = var_9613, x = var_9616_cast_fp16)[name = string("peak_201_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_503_cast_fp16 = real_div(x = x_501_cast_fp16, y = peak_201_cast_fp16)[name = string("x_503_cast_fp16")];
            fp16 var_9612_promoted_to_fp16 = const()[name = string("op_9612_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9620_cast_fp16 = pow(x = x_503_cast_fp16, y = var_9612_promoted_to_fp16)[name = string("op_9620_cast_fp16")];
            tensor<int32, [1]> var_9622_axes_0 = const()[name = string("op_9622_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9622_keep_dims_0 = const()[name = string("op_9622_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9622_cast_fp16 = reduce_mean(axes = var_9622_axes_0, keep_dims = var_9622_keep_dims_0, x = var_9620_cast_fp16)[name = string("op_9622_cast_fp16")];
            fp16 var_9612_promoted_1_to_fp16 = const()[name = string("op_9612_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9623_cast_fp16 = pow(x = peak_201_cast_fp16, y = var_9612_promoted_1_to_fp16)[name = string("op_9623_cast_fp16")];
            fp32 var_9624_epsilon_0 = const()[name = string("op_9624_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9624_cast_fp16 = inverse(epsilon = var_9624_epsilon_0, x = var_9623_cast_fp16)[name = string("op_9624_cast_fp16")];
            fp16 var_9625_to_fp16 = const()[name = string("op_9625_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9626_cast_fp16 = mul(x = var_9624_cast_fp16, y = var_9625_to_fp16)[name = string("op_9626_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9627_cast_fp16 = add(x = var_9622_cast_fp16, y = var_9626_cast_fp16)[name = string("op_9627_cast_fp16")];
            fp32 var_9628_epsilon_0 = const()[name = string("op_9628_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9628_cast_fp16 = rsqrt(epsilon = var_9628_epsilon_0, x = var_9627_cast_fp16)[name = string("op_9628_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9629_cast_fp16 = mul(x = x_503_cast_fp16, y = var_9628_cast_fp16)[name = string("op_9629_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_9629_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_9638 = const()[name = string("op_9638"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9639_cast_fp16 = reshape(shape = var_9638, x = linear_175_cast_fp16)[name = string("op_9639_cast_fp16")];
            tensor<int32, [4]> x_505_perm_0 = const()[name = string("x_505_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9646 = const()[name = string("op_9646"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_505_cast_fp16 = transpose(perm = x_505_perm_0, x = var_9639_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 16, 128]> var_9649_cast_fp16 = abs(x = x_505_cast_fp16)[name = string("op_9649_cast_fp16")];
            tensor<int32, [1]> peak_203_axes_0 = const()[name = string("peak_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_203_cast_fp16 = reduce_max(axes = peak_203_axes_0, keep_dims = var_9646, x = var_9649_cast_fp16)[name = string("peak_203_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_507_cast_fp16 = real_div(x = x_505_cast_fp16, y = peak_203_cast_fp16)[name = string("x_507_cast_fp16")];
            fp16 var_9645_promoted_to_fp16 = const()[name = string("op_9645_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9653_cast_fp16 = pow(x = x_507_cast_fp16, y = var_9645_promoted_to_fp16)[name = string("op_9653_cast_fp16")];
            tensor<int32, [1]> var_9655_axes_0 = const()[name = string("op_9655_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9655_keep_dims_0 = const()[name = string("op_9655_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9655_cast_fp16 = reduce_mean(axes = var_9655_axes_0, keep_dims = var_9655_keep_dims_0, x = var_9653_cast_fp16)[name = string("op_9655_cast_fp16")];
            fp16 var_9645_promoted_1_to_fp16 = const()[name = string("op_9645_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9656_cast_fp16 = pow(x = peak_203_cast_fp16, y = var_9645_promoted_1_to_fp16)[name = string("op_9656_cast_fp16")];
            fp32 var_9657_epsilon_0 = const()[name = string("op_9657_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9657_cast_fp16 = inverse(epsilon = var_9657_epsilon_0, x = var_9656_cast_fp16)[name = string("op_9657_cast_fp16")];
            fp16 var_9658_to_fp16 = const()[name = string("op_9658_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9659_cast_fp16 = mul(x = var_9657_cast_fp16, y = var_9658_to_fp16)[name = string("op_9659_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9660_cast_fp16 = add(x = var_9655_cast_fp16, y = var_9659_cast_fp16)[name = string("op_9660_cast_fp16")];
            fp32 var_9661_epsilon_0 = const()[name = string("op_9661_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9661_cast_fp16 = rsqrt(epsilon = var_9661_epsilon_0, x = var_9660_cast_fp16)[name = string("op_9661_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9662_cast_fp16 = mul(x = x_507_cast_fp16, y = var_9661_cast_fp16)[name = string("op_9662_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_513_cast_fp16 = mul(x = var_9662_cast_fp16, y = model_layers_25_self_attn_q_norm_weight_to_fp16)[name = string("x_513_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_9671 = const()[name = string("op_9671"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9672_cast_fp16 = reshape(shape = var_9671, x = linear_176_cast_fp16)[name = string("op_9672_cast_fp16")];
            tensor<int32, [4]> x_509_perm_0 = const()[name = string("x_509_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9679 = const()[name = string("op_9679"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_509_cast_fp16 = transpose(perm = x_509_perm_0, x = var_9672_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 16, 128]> var_9682_cast_fp16 = abs(x = x_509_cast_fp16)[name = string("op_9682_cast_fp16")];
            tensor<int32, [1]> peak_205_axes_0 = const()[name = string("peak_205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_205_cast_fp16 = reduce_max(axes = peak_205_axes_0, keep_dims = var_9679, x = var_9682_cast_fp16)[name = string("peak_205_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_511_cast_fp16 = real_div(x = x_509_cast_fp16, y = peak_205_cast_fp16)[name = string("x_511_cast_fp16")];
            fp16 var_9678_promoted_to_fp16 = const()[name = string("op_9678_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_9686_cast_fp16 = pow(x = x_511_cast_fp16, y = var_9678_promoted_to_fp16)[name = string("op_9686_cast_fp16")];
            tensor<int32, [1]> var_9688_axes_0 = const()[name = string("op_9688_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9688_keep_dims_0 = const()[name = string("op_9688_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9688_cast_fp16 = reduce_mean(axes = var_9688_axes_0, keep_dims = var_9688_keep_dims_0, x = var_9686_cast_fp16)[name = string("op_9688_cast_fp16")];
            fp16 var_9678_promoted_1_to_fp16 = const()[name = string("op_9678_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_9689_cast_fp16 = pow(x = peak_205_cast_fp16, y = var_9678_promoted_1_to_fp16)[name = string("op_9689_cast_fp16")];
            fp32 var_9690_epsilon_0 = const()[name = string("op_9690_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_9690_cast_fp16 = inverse(epsilon = var_9690_epsilon_0, x = var_9689_cast_fp16)[name = string("op_9690_cast_fp16")];
            fp16 var_9691_to_fp16 = const()[name = string("op_9691_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9692_cast_fp16 = mul(x = var_9690_cast_fp16, y = var_9691_to_fp16)[name = string("op_9692_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_9693_cast_fp16 = add(x = var_9688_cast_fp16, y = var_9692_cast_fp16)[name = string("op_9693_cast_fp16")];
            fp32 var_9694_epsilon_0 = const()[name = string("op_9694_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9694_cast_fp16 = rsqrt(epsilon = var_9694_epsilon_0, x = var_9693_cast_fp16)[name = string("op_9694_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9695_cast_fp16 = mul(x = x_511_cast_fp16, y = var_9694_cast_fp16)[name = string("op_9695_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_515_cast_fp16 = mul(x = var_9695_cast_fp16, y = model_layers_25_self_attn_k_norm_weight_to_fp16)[name = string("x_515_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_9704 = const()[name = string("op_9704"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9705_cast_fp16 = reshape(shape = var_9704, x = linear_177_cast_fp16)[name = string("op_9705_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_513_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_513_cast_fp16)[name = string("second_101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9719_cast_fp16 = mul(x = first_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9719_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9720_cast_fp16 = mul(x = second_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9720_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9722_cast_fp16 = sub(x = var_9719_cast_fp16, y = var_9720_cast_fp16)[name = string("op_9722_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9723_cast_fp16 = mul(x = second_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9723_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9724_cast_fp16 = mul(x = first_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9724_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9726_cast_fp16 = add(x = var_9723_cast_fp16, y = var_9724_cast_fp16)[name = string("op_9726_cast_fp16")];
            int32 var_9728 = const()[name = string("op_9728"), 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_9728, interleave = q_51_interleave_0, values = (var_9722_cast_fp16, var_9726_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_515_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_515_cast_fp16)[name = string("second_103_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9740_cast_fp16 = mul(x = first_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9740_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9741_cast_fp16 = mul(x = second_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9741_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9743_cast_fp16 = sub(x = var_9740_cast_fp16, y = var_9741_cast_fp16)[name = string("op_9743_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9744_cast_fp16 = mul(x = second_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9744_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9745_cast_fp16 = mul(x = first_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9745_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9747_cast_fp16 = add(x = var_9744_cast_fp16, y = var_9745_cast_fp16)[name = string("op_9747_cast_fp16")];
            int32 var_9749 = const()[name = string("op_9749"), 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_9749, interleave = k_51_interleave_0, values = (var_9743_cast_fp16, var_9747_cast_fp16))[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> var_9760_begin_0 = const()[name = string("op_9760_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9760_end_0 = const()[name = string("op_9760_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9760_end_mask_0 = const()[name = string("op_9760_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9760_stride_0 = const()[name = string("op_9760_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9760_cast_fp16 = slice_by_index(begin = var_9760_begin_0, end = var_9760_end_0, end_mask = var_9760_end_mask_0, stride = var_9760_stride_0, x = q_51_cast_fp16)[name = string("op_9760_cast_fp16")];
            tensor<int32, [4]> var_9770_begin_0 = const()[name = string("op_9770_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9770_end_0 = const()[name = string("op_9770_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9770_end_mask_0 = const()[name = string("op_9770_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9770_stride_0 = const()[name = string("op_9770_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9770_cast_fp16 = slice_by_index(begin = var_9770_begin_0, end = var_9770_end_0, end_mask = var_9770_end_mask_0, stride = var_9770_stride_0, x = q_51_cast_fp16)[name = string("op_9770_cast_fp16")];
            int32 var_9772 = const()[name = string("op_9772"), 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_9772, interleave = stacked_51_interleave_0, values = (var_9760_cast_fp16, var_9770_cast_fp16))[name = string("stacked_51_cast_fp16")];
            tensor<int32, [4]> var_9778_begin_0 = const()[name = string("op_9778_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9778_end_0 = const()[name = string("op_9778_end_0"), val = tensor<int32, [4]>([26, 8, 528, 128])];
            tensor<bool, [4]> var_9778_end_mask_0 = const()[name = string("op_9778_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9778_cast_fp16 = slice_by_index(begin = var_9778_begin_0, end = var_9778_end_0, end_mask = var_9778_end_mask_0, x = read_state_0)[name = string("op_9778_cast_fp16")];
            tensor<int32, [4]> var_9788_begin_0 = const()[name = string("op_9788_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9788_end_0 = const()[name = string("op_9788_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9788_end_mask_0 = const()[name = string("op_9788_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9788_cast_fp16 = slice_by_index(begin = var_9788_begin_0, end = var_9788_end_0, end_mask = var_9788_end_mask_0, x = var_9778_cast_fp16)[name = string("op_9788_cast_fp16")];
            bool var_9797_transpose_x_1 = const()[name = string("op_9797_transpose_x_1"), val = bool(false)];
            bool var_9797_transpose_y_1 = const()[name = string("op_9797_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_1, transpose_y = var_9797_transpose_y_1, x = var_9788_cast_fp16, y = stacked_51_cast_fp16)[name = string("op_9797_cast_fp16")];
            tensor<int32, [4]> var_9800_perm_0 = const()[name = string("op_9800_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9801_to_fp16 = const()[name = string("op_9801_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_9800_cast_fp16 = transpose(perm = var_9800_perm_0, x = var_9797_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 32, 512]> var_9802_cast_fp16 = mul(x = var_9800_cast_fp16, y = var_9801_to_fp16)[name = string("op_9802_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_51_cast_fp16 = add(x = var_9802_cast_fp16, y = cacheMask)[name = string("cache_weights_51_cast_fp16")];
            bool var_9808_transpose_x_1 = const()[name = string("op_9808_transpose_x_1"), val = bool(false)];
            bool var_9808_transpose_y_1 = const()[name = string("op_9808_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9808_cast_fp16 = matmul(transpose_x = var_9808_transpose_x_1, transpose_y = var_9808_transpose_y_1, x = stacked_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_9808_cast_fp16")];
            fp16 var_9809_to_fp16 = const()[name = string("op_9809_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9810_cast_fp16 = mul(x = var_9808_cast_fp16, y = var_9809_to_fp16)[name = string("op_9810_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_51_cast_fp16 = add(x = var_9810_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_51_cast_fp16")];
            int32 var_9814 = const()[name = string("op_9814"), val = int32(-1)];
            bool var_9815_interleave_0 = const()[name = string("op_9815_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_9815_cast_fp16 = concat(axis = var_9814, interleave = var_9815_interleave_0, values = (cache_weights_51_cast_fp16, call_weights_51_cast_fp16))[name = string("op_9815_cast_fp16")];
            int32 var_9816 = const()[name = string("op_9816"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_51_cast_fp16 = softmax(axis = var_9816, x = var_9815_cast_fp16)[name = string("shares_51_cast_fp16")];
            tensor<int32, [4]> var_9823_begin_0 = const()[name = string("op_9823_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9823_end_0 = const()[name = string("op_9823_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_9823_end_mask_0 = const()[name = string("op_9823_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_9823_cast_fp16 = slice_by_index(begin = var_9823_begin_0, end = var_9823_end_0, end_mask = var_9823_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9823_cast_fp16")];
            tensor<int32, [4]> var_9828_begin_0 = const()[name = string("op_9828_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9828_end_0 = const()[name = string("op_9828_end_0"), val = tensor<int32, [4]>([26, 8, 528, 128])];
            tensor<bool, [4]> var_9828_end_mask_0 = const()[name = string("op_9828_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9828_cast_fp16 = slice_by_index(begin = var_9828_begin_0, end = var_9828_end_0, end_mask = var_9828_end_mask_0, x = read_state_1)[name = string("op_9828_cast_fp16")];
            tensor<int32, [4]> var_9838_begin_0 = const()[name = string("op_9838_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9838_end_0 = const()[name = string("op_9838_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9838_end_mask_0 = const()[name = string("op_9838_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9838_cast_fp16 = slice_by_index(begin = var_9838_begin_0, end = var_9838_end_0, end_mask = var_9838_end_mask_0, x = var_9828_cast_fp16)[name = string("op_9838_cast_fp16")];
            bool var_9844_transpose_x_0 = const()[name = string("op_9844_transpose_x_0"), val = bool(false)];
            bool var_9844_transpose_y_0 = const()[name = string("op_9844_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9844_cast_fp16 = matmul(transpose_x = var_9844_transpose_x_0, transpose_y = var_9844_transpose_y_0, x = var_9823_cast_fp16, y = var_9838_cast_fp16)[name = string("op_9844_cast_fp16")];
            tensor<int32, [4]> var_9849_begin_0 = const()[name = string("op_9849_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_9849_end_0 = const()[name = string("op_9849_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_9849_end_mask_0 = const()[name = string("op_9849_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9849_cast_fp16 = slice_by_index(begin = var_9849_begin_0, end = var_9849_end_0, end_mask = var_9849_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9849_cast_fp16")];
            bool var_9850_transpose_x_0 = const()[name = string("op_9850_transpose_x_0"), val = bool(false)];
            bool var_9850_transpose_y_0 = const()[name = string("op_9850_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_9705_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 8, 32, 128]> var_9850_cast_fp16 = matmul(transpose_x = var_9850_transpose_x_0, transpose_y = var_9850_transpose_y_0, x = var_9849_cast_fp16, y = v_51_cast_fp16)[name = string("op_9850_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_101_cast_fp16 = add(x = var_9844_cast_fp16, y = var_9850_cast_fp16)[name = string("mixed_101_cast_fp16")];
            tensor<int32, [4]> var_9867_begin_0 = const()[name = string("op_9867_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9867_end_0 = const()[name = string("op_9867_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9867_end_mask_0 = const()[name = string("op_9867_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9867_cast_fp16 = slice_by_index(begin = var_9867_begin_0, end = var_9867_end_0, end_mask = var_9867_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9867_cast_fp16")];
            tensor<int32, [4]> var_9882_begin_0 = const()[name = string("op_9882_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9882_end_0 = const()[name = string("op_9882_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9882_end_mask_0 = const()[name = string("op_9882_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9882_cast_fp16 = slice_by_index(begin = var_9882_begin_0, end = var_9882_end_0, end_mask = var_9882_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9882_cast_fp16")];
            int32 var_9884 = const()[name = string("op_9884"), 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_9884, interleave = mixed_103_interleave_0, values = (var_9867_cast_fp16, var_9882_cast_fp16))[name = string("mixed_103_cast_fp16")];
            tensor<int32, [4]> var_9888_perm_0 = const()[name = string("op_9888_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9892 = const()[name = string("op_9892"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9888_cast_fp16 = transpose(perm = var_9888_perm_0, x = mixed_103_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 2048]> input_253_cast_fp16 = reshape(shape = var_9892, x = var_9888_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_517_cast_fp16 = add(x = x_501_cast_fp16, y = linear_178_cast_fp16)[name = string("x_517_cast_fp16")];
            bool var_9902 = const()[name = string("op_9902"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9905_cast_fp16 = abs(x = x_517_cast_fp16)[name = string("op_9905_cast_fp16")];
            tensor<int32, [1]> peak_207_axes_0 = const()[name = string("peak_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_207_cast_fp16 = reduce_max(axes = peak_207_axes_0, keep_dims = var_9902, x = var_9905_cast_fp16)[name = string("peak_207_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_519_cast_fp16 = real_div(x = x_517_cast_fp16, y = peak_207_cast_fp16)[name = string("x_519_cast_fp16")];
            fp16 var_9901_promoted_to_fp16 = const()[name = string("op_9901_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9909_cast_fp16 = pow(x = x_519_cast_fp16, y = var_9901_promoted_to_fp16)[name = string("op_9909_cast_fp16")];
            tensor<int32, [1]> var_9911_axes_0 = const()[name = string("op_9911_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9911_keep_dims_0 = const()[name = string("op_9911_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9911_cast_fp16 = reduce_mean(axes = var_9911_axes_0, keep_dims = var_9911_keep_dims_0, x = var_9909_cast_fp16)[name = string("op_9911_cast_fp16")];
            fp16 var_9901_promoted_1_to_fp16 = const()[name = string("op_9901_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9912_cast_fp16 = pow(x = peak_207_cast_fp16, y = var_9901_promoted_1_to_fp16)[name = string("op_9912_cast_fp16")];
            fp32 var_9913_epsilon_0 = const()[name = string("op_9913_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9913_cast_fp16 = inverse(epsilon = var_9913_epsilon_0, x = var_9912_cast_fp16)[name = string("op_9913_cast_fp16")];
            fp16 var_9914_to_fp16 = const()[name = string("op_9914_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9915_cast_fp16 = mul(x = var_9913_cast_fp16, y = var_9914_to_fp16)[name = string("op_9915_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9916_cast_fp16 = add(x = var_9911_cast_fp16, y = var_9915_cast_fp16)[name = string("op_9916_cast_fp16")];
            fp32 var_9917_epsilon_0 = const()[name = string("op_9917_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9917_cast_fp16 = rsqrt(epsilon = var_9917_epsilon_0, x = var_9916_cast_fp16)[name = string("op_9917_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9918_cast_fp16 = mul(x = x_519_cast_fp16, y = var_9917_cast_fp16)[name = string("op_9918_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_9918_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_9923_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_9923_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_9923_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_521_cast_fp16 = add(x = x_517_cast_fp16, y = linear_181_cast_fp16)[name = string("x_521_cast_fp16")];
            bool var_9936 = const()[name = string("op_9936"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_9939_cast_fp16 = abs(x = x_521_cast_fp16)[name = string("op_9939_cast_fp16")];
            tensor<int32, [1]> peak_209_axes_0 = const()[name = string("peak_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_209_cast_fp16 = reduce_max(axes = peak_209_axes_0, keep_dims = var_9936, x = var_9939_cast_fp16)[name = string("peak_209_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_523_cast_fp16 = real_div(x = x_521_cast_fp16, y = peak_209_cast_fp16)[name = string("x_523_cast_fp16")];
            fp16 var_9935_promoted_to_fp16 = const()[name = string("op_9935_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9943_cast_fp16 = pow(x = x_523_cast_fp16, y = var_9935_promoted_to_fp16)[name = string("op_9943_cast_fp16")];
            tensor<int32, [1]> var_9945_axes_0 = const()[name = string("op_9945_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9945_keep_dims_0 = const()[name = string("op_9945_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9945_cast_fp16 = reduce_mean(axes = var_9945_axes_0, keep_dims = var_9945_keep_dims_0, x = var_9943_cast_fp16)[name = string("op_9945_cast_fp16")];
            fp16 var_9935_promoted_1_to_fp16 = const()[name = string("op_9935_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_9946_cast_fp16 = pow(x = peak_209_cast_fp16, y = var_9935_promoted_1_to_fp16)[name = string("op_9946_cast_fp16")];
            fp32 var_9947_epsilon_0 = const()[name = string("op_9947_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_9947_cast_fp16 = inverse(epsilon = var_9947_epsilon_0, x = var_9946_cast_fp16)[name = string("op_9947_cast_fp16")];
            fp16 var_9948_to_fp16 = const()[name = string("op_9948_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9949_cast_fp16 = mul(x = var_9947_cast_fp16, y = var_9948_to_fp16)[name = string("op_9949_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_9950_cast_fp16 = add(x = var_9945_cast_fp16, y = var_9949_cast_fp16)[name = string("op_9950_cast_fp16")];
            fp32 var_9951_epsilon_0 = const()[name = string("op_9951_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9951_cast_fp16 = rsqrt(epsilon = var_9951_epsilon_0, x = var_9950_cast_fp16)[name = string("op_9951_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9952_cast_fp16 = mul(x = x_523_cast_fp16, y = var_9951_cast_fp16)[name = string("op_9952_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_9952_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_9961 = const()[name = string("op_9961"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9962_cast_fp16 = reshape(shape = var_9961, x = linear_182_cast_fp16)[name = string("op_9962_cast_fp16")];
            tensor<int32, [4]> x_525_perm_0 = const()[name = string("x_525_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_9969 = const()[name = string("op_9969"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_525_cast_fp16 = transpose(perm = x_525_perm_0, x = var_9962_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 16, 128]> var_9972_cast_fp16 = abs(x = x_525_cast_fp16)[name = string("op_9972_cast_fp16")];
            tensor<int32, [1]> peak_211_axes_0 = const()[name = string("peak_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_211_cast_fp16 = reduce_max(axes = peak_211_axes_0, keep_dims = var_9969, x = var_9972_cast_fp16)[name = string("peak_211_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_527_cast_fp16 = real_div(x = x_525_cast_fp16, y = peak_211_cast_fp16)[name = string("x_527_cast_fp16")];
            fp16 var_9968_promoted_to_fp16 = const()[name = string("op_9968_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_9976_cast_fp16 = pow(x = x_527_cast_fp16, y = var_9968_promoted_to_fp16)[name = string("op_9976_cast_fp16")];
            tensor<int32, [1]> var_9978_axes_0 = const()[name = string("op_9978_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9978_keep_dims_0 = const()[name = string("op_9978_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9978_cast_fp16 = reduce_mean(axes = var_9978_axes_0, keep_dims = var_9978_keep_dims_0, x = var_9976_cast_fp16)[name = string("op_9978_cast_fp16")];
            fp16 var_9968_promoted_1_to_fp16 = const()[name = string("op_9968_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_9979_cast_fp16 = pow(x = peak_211_cast_fp16, y = var_9968_promoted_1_to_fp16)[name = string("op_9979_cast_fp16")];
            fp32 var_9980_epsilon_0 = const()[name = string("op_9980_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_9980_cast_fp16 = inverse(epsilon = var_9980_epsilon_0, x = var_9979_cast_fp16)[name = string("op_9980_cast_fp16")];
            fp16 var_9981_to_fp16 = const()[name = string("op_9981_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9982_cast_fp16 = mul(x = var_9980_cast_fp16, y = var_9981_to_fp16)[name = string("op_9982_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_9983_cast_fp16 = add(x = var_9978_cast_fp16, y = var_9982_cast_fp16)[name = string("op_9983_cast_fp16")];
            fp32 var_9984_epsilon_0 = const()[name = string("op_9984_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9984_cast_fp16 = rsqrt(epsilon = var_9984_epsilon_0, x = var_9983_cast_fp16)[name = string("op_9984_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9985_cast_fp16 = mul(x = x_527_cast_fp16, y = var_9984_cast_fp16)[name = string("op_9985_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_533_cast_fp16 = mul(x = var_9985_cast_fp16, y = model_layers_26_self_attn_q_norm_weight_to_fp16)[name = string("x_533_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_9994 = const()[name = string("op_9994"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9995_cast_fp16 = reshape(shape = var_9994, x = linear_183_cast_fp16)[name = string("op_9995_cast_fp16")];
            tensor<int32, [4]> x_529_perm_0 = const()[name = string("x_529_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10002 = const()[name = string("op_10002"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_529_cast_fp16 = transpose(perm = x_529_perm_0, x = var_9995_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 16, 128]> var_10005_cast_fp16 = abs(x = x_529_cast_fp16)[name = string("op_10005_cast_fp16")];
            tensor<int32, [1]> peak_213_axes_0 = const()[name = string("peak_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_213_cast_fp16 = reduce_max(axes = peak_213_axes_0, keep_dims = var_10002, x = var_10005_cast_fp16)[name = string("peak_213_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_531_cast_fp16 = real_div(x = x_529_cast_fp16, y = peak_213_cast_fp16)[name = string("x_531_cast_fp16")];
            fp16 var_10001_promoted_to_fp16 = const()[name = string("op_10001_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_10009_cast_fp16 = pow(x = x_531_cast_fp16, y = var_10001_promoted_to_fp16)[name = string("op_10009_cast_fp16")];
            tensor<int32, [1]> var_10011_axes_0 = const()[name = string("op_10011_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10011_keep_dims_0 = const()[name = string("op_10011_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_10011_cast_fp16 = reduce_mean(axes = var_10011_axes_0, keep_dims = var_10011_keep_dims_0, x = var_10009_cast_fp16)[name = string("op_10011_cast_fp16")];
            fp16 var_10001_promoted_1_to_fp16 = const()[name = string("op_10001_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_10012_cast_fp16 = pow(x = peak_213_cast_fp16, y = var_10001_promoted_1_to_fp16)[name = string("op_10012_cast_fp16")];
            fp32 var_10013_epsilon_0 = const()[name = string("op_10013_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_10013_cast_fp16 = inverse(epsilon = var_10013_epsilon_0, x = var_10012_cast_fp16)[name = string("op_10013_cast_fp16")];
            fp16 var_10014_to_fp16 = const()[name = string("op_10014_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_10015_cast_fp16 = mul(x = var_10013_cast_fp16, y = var_10014_to_fp16)[name = string("op_10015_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_10016_cast_fp16 = add(x = var_10011_cast_fp16, y = var_10015_cast_fp16)[name = string("op_10016_cast_fp16")];
            fp32 var_10017_epsilon_0 = const()[name = string("op_10017_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_10017_cast_fp16 = rsqrt(epsilon = var_10017_epsilon_0, x = var_10016_cast_fp16)[name = string("op_10017_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_10018_cast_fp16 = mul(x = x_531_cast_fp16, y = var_10017_cast_fp16)[name = string("op_10018_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_535_cast_fp16 = mul(x = var_10018_cast_fp16, y = model_layers_26_self_attn_k_norm_weight_to_fp16)[name = string("x_535_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_10027 = const()[name = string("op_10027"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10028_cast_fp16 = reshape(shape = var_10027, x = linear_184_cast_fp16)[name = string("op_10028_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_533_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_533_cast_fp16)[name = string("second_105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10042_cast_fp16 = mul(x = first_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10042_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10043_cast_fp16 = mul(x = second_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10043_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10045_cast_fp16 = sub(x = var_10042_cast_fp16, y = var_10043_cast_fp16)[name = string("op_10045_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10046_cast_fp16 = mul(x = second_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10046_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10047_cast_fp16 = mul(x = first_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10047_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10049_cast_fp16 = add(x = var_10046_cast_fp16, y = var_10047_cast_fp16)[name = string("op_10049_cast_fp16")];
            int32 var_10051 = const()[name = string("op_10051"), 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_10051, interleave = q_53_interleave_0, values = (var_10045_cast_fp16, var_10049_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_535_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_535_cast_fp16)[name = string("second_107_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10063_cast_fp16 = mul(x = first_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10063_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10064_cast_fp16 = mul(x = second_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10064_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10066_cast_fp16 = sub(x = var_10063_cast_fp16, y = var_10064_cast_fp16)[name = string("op_10066_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10067_cast_fp16 = mul(x = second_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10067_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10068_cast_fp16 = mul(x = first_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10068_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10070_cast_fp16 = add(x = var_10067_cast_fp16, y = var_10068_cast_fp16)[name = string("op_10070_cast_fp16")];
            int32 var_10072 = const()[name = string("op_10072"), 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_10072, interleave = k_53_interleave_0, values = (var_10066_cast_fp16, var_10070_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> var_10083_begin_0 = const()[name = string("op_10083_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10083_end_0 = const()[name = string("op_10083_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10083_end_mask_0 = const()[name = string("op_10083_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10083_stride_0 = const()[name = string("op_10083_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10083_cast_fp16 = slice_by_index(begin = var_10083_begin_0, end = var_10083_end_0, end_mask = var_10083_end_mask_0, stride = var_10083_stride_0, x = q_53_cast_fp16)[name = string("op_10083_cast_fp16")];
            tensor<int32, [4]> var_10093_begin_0 = const()[name = string("op_10093_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_10093_end_0 = const()[name = string("op_10093_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10093_end_mask_0 = const()[name = string("op_10093_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10093_stride_0 = const()[name = string("op_10093_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10093_cast_fp16 = slice_by_index(begin = var_10093_begin_0, end = var_10093_end_0, end_mask = var_10093_end_mask_0, stride = var_10093_stride_0, x = q_53_cast_fp16)[name = string("op_10093_cast_fp16")];
            int32 var_10095 = const()[name = string("op_10095"), 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_10095, interleave = stacked_53_interleave_0, values = (var_10083_cast_fp16, var_10093_cast_fp16))[name = string("stacked_53_cast_fp16")];
            tensor<int32, [4]> var_10101_begin_0 = const()[name = string("op_10101_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_10101_end_0 = const()[name = string("op_10101_end_0"), val = tensor<int32, [4]>([27, 8, 528, 128])];
            tensor<bool, [4]> var_10101_end_mask_0 = const()[name = string("op_10101_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_10101_cast_fp16 = slice_by_index(begin = var_10101_begin_0, end = var_10101_end_0, end_mask = var_10101_end_mask_0, x = read_state_0)[name = string("op_10101_cast_fp16")];
            tensor<int32, [4]> var_10111_begin_0 = const()[name = string("op_10111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10111_end_0 = const()[name = string("op_10111_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_10111_end_mask_0 = const()[name = string("op_10111_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, x = var_10101_cast_fp16)[name = string("op_10111_cast_fp16")];
            bool var_10120_transpose_x_1 = const()[name = string("op_10120_transpose_x_1"), val = bool(false)];
            bool var_10120_transpose_y_1 = const()[name = string("op_10120_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_10120_cast_fp16 = matmul(transpose_x = var_10120_transpose_x_1, transpose_y = var_10120_transpose_y_1, x = var_10111_cast_fp16, y = stacked_53_cast_fp16)[name = string("op_10120_cast_fp16")];
            tensor<int32, [4]> var_10123_perm_0 = const()[name = string("op_10123_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_10124_to_fp16 = const()[name = string("op_10124_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_10123_cast_fp16 = transpose(perm = var_10123_perm_0, x = var_10120_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 8, 32, 512]> var_10125_cast_fp16 = mul(x = var_10123_cast_fp16, y = var_10124_to_fp16)[name = string("op_10125_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_53_cast_fp16 = add(x = var_10125_cast_fp16, y = cacheMask)[name = string("cache_weights_53_cast_fp16")];
            bool var_10131_transpose_x_1 = const()[name = string("op_10131_transpose_x_1"), val = bool(false)];
            bool var_10131_transpose_y_1 = const()[name = string("op_10131_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_10131_cast_fp16 = matmul(transpose_x = var_10131_transpose_x_1, transpose_y = var_10131_transpose_y_1, x = stacked_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_10131_cast_fp16")];
            fp16 var_10132_to_fp16 = const()[name = string("op_10132_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_10133_cast_fp16 = mul(x = var_10131_cast_fp16, y = var_10132_to_fp16)[name = string("op_10133_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_53_cast_fp16 = add(x = var_10133_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_53_cast_fp16")];
            int32 var_10137 = const()[name = string("op_10137"), val = int32(-1)];
            bool var_10138_interleave_0 = const()[name = string("op_10138_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_10138_cast_fp16 = concat(axis = var_10137, interleave = var_10138_interleave_0, values = (cache_weights_53_cast_fp16, call_weights_53_cast_fp16))[name = string("op_10138_cast_fp16")];
            int32 var_10139 = const()[name = string("op_10139"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_53_cast_fp16 = softmax(axis = var_10139, x = var_10138_cast_fp16)[name = string("shares_53_cast_fp16")];
            tensor<int32, [4]> var_10146_begin_0 = const()[name = string("op_10146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10146_end_0 = const()[name = string("op_10146_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_10146_end_mask_0 = const()[name = string("op_10146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_10146_cast_fp16 = slice_by_index(begin = var_10146_begin_0, end = var_10146_end_0, end_mask = var_10146_end_mask_0, x = shares_53_cast_fp16)[name = string("op_10146_cast_fp16")];
            tensor<int32, [4]> var_10151_begin_0 = const()[name = string("op_10151_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_10151_end_0 = const()[name = string("op_10151_end_0"), val = tensor<int32, [4]>([27, 8, 528, 128])];
            tensor<bool, [4]> var_10151_end_mask_0 = const()[name = string("op_10151_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_10151_cast_fp16 = slice_by_index(begin = var_10151_begin_0, end = var_10151_end_0, end_mask = var_10151_end_mask_0, x = read_state_1)[name = string("op_10151_cast_fp16")];
            tensor<int32, [4]> var_10161_begin_0 = const()[name = string("op_10161_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10161_end_0 = const()[name = string("op_10161_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_10161_end_mask_0 = const()[name = string("op_10161_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_10161_cast_fp16 = slice_by_index(begin = var_10161_begin_0, end = var_10161_end_0, end_mask = var_10161_end_mask_0, x = var_10151_cast_fp16)[name = string("op_10161_cast_fp16")];
            bool var_10167_transpose_x_0 = const()[name = string("op_10167_transpose_x_0"), val = bool(false)];
            bool var_10167_transpose_y_0 = const()[name = string("op_10167_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_10167_cast_fp16 = matmul(transpose_x = var_10167_transpose_x_0, transpose_y = var_10167_transpose_y_0, x = var_10146_cast_fp16, y = var_10161_cast_fp16)[name = string("op_10167_cast_fp16")];
            tensor<int32, [4]> var_10172_begin_0 = const()[name = string("op_10172_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_10172_end_0 = const()[name = string("op_10172_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_10172_end_mask_0 = const()[name = string("op_10172_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_10172_cast_fp16 = slice_by_index(begin = var_10172_begin_0, end = var_10172_end_0, end_mask = var_10172_end_mask_0, x = shares_53_cast_fp16)[name = string("op_10172_cast_fp16")];
            bool var_10173_transpose_x_0 = const()[name = string("op_10173_transpose_x_0"), val = bool(false)];
            bool var_10173_transpose_y_0 = const()[name = string("op_10173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_10028_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 32, 128]> var_10173_cast_fp16 = matmul(transpose_x = var_10173_transpose_x_0, transpose_y = var_10173_transpose_y_0, x = var_10172_cast_fp16, y = v_53_cast_fp16)[name = string("op_10173_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_105_cast_fp16 = add(x = var_10167_cast_fp16, y = var_10173_cast_fp16)[name = string("mixed_105_cast_fp16")];
            tensor<int32, [4]> var_10190_begin_0 = const()[name = string("op_10190_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10190_end_0 = const()[name = string("op_10190_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_10190_end_mask_0 = const()[name = string("op_10190_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10190_cast_fp16 = slice_by_index(begin = var_10190_begin_0, end = var_10190_end_0, end_mask = var_10190_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_10190_cast_fp16")];
            tensor<int32, [4]> var_10205_begin_0 = const()[name = string("op_10205_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_10205_end_0 = const()[name = string("op_10205_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_10205_end_mask_0 = const()[name = string("op_10205_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10205_cast_fp16 = slice_by_index(begin = var_10205_begin_0, end = var_10205_end_0, end_mask = var_10205_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_10205_cast_fp16")];
            int32 var_10207 = const()[name = string("op_10207"), 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_10207, interleave = mixed_107_interleave_0, values = (var_10190_cast_fp16, var_10205_cast_fp16))[name = string("mixed_107_cast_fp16")];
            tensor<int32, [4]> var_10211_perm_0 = const()[name = string("op_10211_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_10215 = const()[name = string("op_10215"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_10211_cast_fp16 = transpose(perm = var_10211_perm_0, x = mixed_107_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 2048]> input_263_cast_fp16 = reshape(shape = var_10215, x = var_10211_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_537_cast_fp16 = add(x = x_521_cast_fp16, y = linear_185_cast_fp16)[name = string("x_537_cast_fp16")];
            bool var_10225 = const()[name = string("op_10225"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10228_cast_fp16 = abs(x = x_537_cast_fp16)[name = string("op_10228_cast_fp16")];
            tensor<int32, [1]> peak_215_axes_0 = const()[name = string("peak_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_215_cast_fp16 = reduce_max(axes = peak_215_axes_0, keep_dims = var_10225, x = var_10228_cast_fp16)[name = string("peak_215_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_539_cast_fp16 = real_div(x = x_537_cast_fp16, y = peak_215_cast_fp16)[name = string("x_539_cast_fp16")];
            fp16 var_10224_promoted_to_fp16 = const()[name = string("op_10224_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10232_cast_fp16 = pow(x = x_539_cast_fp16, y = var_10224_promoted_to_fp16)[name = string("op_10232_cast_fp16")];
            tensor<int32, [1]> var_10234_axes_0 = const()[name = string("op_10234_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10234_keep_dims_0 = const()[name = string("op_10234_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10234_cast_fp16 = reduce_mean(axes = var_10234_axes_0, keep_dims = var_10234_keep_dims_0, x = var_10232_cast_fp16)[name = string("op_10234_cast_fp16")];
            fp16 var_10224_promoted_1_to_fp16 = const()[name = string("op_10224_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10235_cast_fp16 = pow(x = peak_215_cast_fp16, y = var_10224_promoted_1_to_fp16)[name = string("op_10235_cast_fp16")];
            fp32 var_10236_epsilon_0 = const()[name = string("op_10236_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10236_cast_fp16 = inverse(epsilon = var_10236_epsilon_0, x = var_10235_cast_fp16)[name = string("op_10236_cast_fp16")];
            fp16 var_10237_to_fp16 = const()[name = string("op_10237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10238_cast_fp16 = mul(x = var_10236_cast_fp16, y = var_10237_to_fp16)[name = string("op_10238_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10239_cast_fp16 = add(x = var_10234_cast_fp16, y = var_10238_cast_fp16)[name = string("op_10239_cast_fp16")];
            fp32 var_10240_epsilon_0 = const()[name = string("op_10240_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10240_cast_fp16 = rsqrt(epsilon = var_10240_epsilon_0, x = var_10239_cast_fp16)[name = string("op_10240_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10241_cast_fp16 = mul(x = x_539_cast_fp16, y = var_10240_cast_fp16)[name = string("op_10241_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_10241_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_10246_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_10246_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_10246_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_541_cast_fp16 = add(x = x_537_cast_fp16, y = linear_188_cast_fp16)[name = string("x_541_cast_fp16")];
            bool var_10259 = const()[name = string("op_10259"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10262_cast_fp16 = abs(x = x_541_cast_fp16)[name = string("op_10262_cast_fp16")];
            tensor<int32, [1]> peak_217_axes_0 = const()[name = string("peak_217_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_217_cast_fp16 = reduce_max(axes = peak_217_axes_0, keep_dims = var_10259, x = var_10262_cast_fp16)[name = string("peak_217_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_543_cast_fp16 = real_div(x = x_541_cast_fp16, y = peak_217_cast_fp16)[name = string("x_543_cast_fp16")];
            fp16 var_10258_promoted_to_fp16 = const()[name = string("op_10258_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10266_cast_fp16 = pow(x = x_543_cast_fp16, y = var_10258_promoted_to_fp16)[name = string("op_10266_cast_fp16")];
            tensor<int32, [1]> var_10268_axes_0 = const()[name = string("op_10268_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10268_keep_dims_0 = const()[name = string("op_10268_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10268_cast_fp16 = reduce_mean(axes = var_10268_axes_0, keep_dims = var_10268_keep_dims_0, x = var_10266_cast_fp16)[name = string("op_10268_cast_fp16")];
            fp16 var_10258_promoted_1_to_fp16 = const()[name = string("op_10258_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10269_cast_fp16 = pow(x = peak_217_cast_fp16, y = var_10258_promoted_1_to_fp16)[name = string("op_10269_cast_fp16")];
            fp32 var_10270_epsilon_0 = const()[name = string("op_10270_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10270_cast_fp16 = inverse(epsilon = var_10270_epsilon_0, x = var_10269_cast_fp16)[name = string("op_10270_cast_fp16")];
            fp16 var_10271_to_fp16 = const()[name = string("op_10271_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10272_cast_fp16 = mul(x = var_10270_cast_fp16, y = var_10271_to_fp16)[name = string("op_10272_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10273_cast_fp16 = add(x = var_10268_cast_fp16, y = var_10272_cast_fp16)[name = string("op_10273_cast_fp16")];
            fp32 var_10274_epsilon_0 = const()[name = string("op_10274_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10274_cast_fp16 = rsqrt(epsilon = var_10274_epsilon_0, x = var_10273_cast_fp16)[name = string("op_10274_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10275_cast_fp16 = mul(x = x_543_cast_fp16, y = var_10274_cast_fp16)[name = string("op_10275_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_10275_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_10284 = const()[name = string("op_10284"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_10285_cast_fp16 = reshape(shape = var_10284, x = linear_189_cast_fp16)[name = string("op_10285_cast_fp16")];
            tensor<int32, [4]> x_545_perm_0 = const()[name = string("x_545_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10292 = const()[name = string("op_10292"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 128]> x_545_cast_fp16 = transpose(perm = x_545_perm_0, x = var_10285_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 16, 16, 128]> var_10295_cast_fp16 = abs(x = x_545_cast_fp16)[name = string("op_10295_cast_fp16")];
            tensor<int32, [1]> peak_219_axes_0 = const()[name = string("peak_219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 16, 1]> peak_219_cast_fp16 = reduce_max(axes = peak_219_axes_0, keep_dims = var_10292, x = var_10295_cast_fp16)[name = string("peak_219_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> x_547_cast_fp16 = real_div(x = x_545_cast_fp16, y = peak_219_cast_fp16)[name = string("x_547_cast_fp16")];
            fp16 var_10291_promoted_to_fp16 = const()[name = string("op_10291_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> var_10299_cast_fp16 = pow(x = x_547_cast_fp16, y = var_10291_promoted_to_fp16)[name = string("op_10299_cast_fp16")];
            tensor<int32, [1]> var_10301_axes_0 = const()[name = string("op_10301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10301_keep_dims_0 = const()[name = string("op_10301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_10301_cast_fp16 = reduce_mean(axes = var_10301_axes_0, keep_dims = var_10301_keep_dims_0, x = var_10299_cast_fp16)[name = string("op_10301_cast_fp16")];
            fp16 var_10291_promoted_1_to_fp16 = const()[name = string("op_10291_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 1]> var_10302_cast_fp16 = pow(x = peak_219_cast_fp16, y = var_10291_promoted_1_to_fp16)[name = string("op_10302_cast_fp16")];
            fp32 var_10303_epsilon_0 = const()[name = string("op_10303_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 16, 1]> var_10303_cast_fp16 = inverse(epsilon = var_10303_epsilon_0, x = var_10302_cast_fp16)[name = string("op_10303_cast_fp16")];
            fp16 var_10304_to_fp16 = const()[name = string("op_10304_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_10305_cast_fp16 = mul(x = var_10303_cast_fp16, y = var_10304_to_fp16)[name = string("op_10305_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1]> var_10306_cast_fp16 = add(x = var_10301_cast_fp16, y = var_10305_cast_fp16)[name = string("op_10306_cast_fp16")];
            fp32 var_10307_epsilon_0 = const()[name = string("op_10307_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_10307_cast_fp16 = rsqrt(epsilon = var_10307_epsilon_0, x = var_10306_cast_fp16)[name = string("op_10307_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_10308_cast_fp16 = mul(x = x_547_cast_fp16, y = var_10307_cast_fp16)[name = string("op_10308_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_553_cast_fp16 = mul(x = var_10308_cast_fp16, y = model_layers_27_self_attn_q_norm_weight_to_fp16)[name = string("x_553_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_10317 = const()[name = string("op_10317"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10318_cast_fp16 = reshape(shape = var_10317, x = linear_190_cast_fp16)[name = string("op_10318_cast_fp16")];
            tensor<int32, [4]> x_549_perm_0 = const()[name = string("x_549_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            bool var_10325 = const()[name = string("op_10325"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 128]> x_549_cast_fp16 = transpose(perm = x_549_perm_0, x = var_10318_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 8, 16, 128]> var_10328_cast_fp16 = abs(x = x_549_cast_fp16)[name = string("op_10328_cast_fp16")];
            tensor<int32, [1]> peak_221_axes_0 = const()[name = string("peak_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 8, 16, 1]> peak_221_cast_fp16 = reduce_max(axes = peak_221_axes_0, keep_dims = var_10325, x = var_10328_cast_fp16)[name = string("peak_221_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> x_551_cast_fp16 = real_div(x = x_549_cast_fp16, y = peak_221_cast_fp16)[name = string("x_551_cast_fp16")];
            fp16 var_10324_promoted_to_fp16 = const()[name = string("op_10324_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> var_10332_cast_fp16 = pow(x = x_551_cast_fp16, y = var_10324_promoted_to_fp16)[name = string("op_10332_cast_fp16")];
            tensor<int32, [1]> var_10334_axes_0 = const()[name = string("op_10334_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10334_keep_dims_0 = const()[name = string("op_10334_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_10334_cast_fp16 = reduce_mean(axes = var_10334_axes_0, keep_dims = var_10334_keep_dims_0, x = var_10332_cast_fp16)[name = string("op_10334_cast_fp16")];
            fp16 var_10324_promoted_1_to_fp16 = const()[name = string("op_10324_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 1]> var_10335_cast_fp16 = pow(x = peak_221_cast_fp16, y = var_10324_promoted_1_to_fp16)[name = string("op_10335_cast_fp16")];
            fp32 var_10336_epsilon_0 = const()[name = string("op_10336_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 8, 16, 1]> var_10336_cast_fp16 = inverse(epsilon = var_10336_epsilon_0, x = var_10335_cast_fp16)[name = string("op_10336_cast_fp16")];
            fp16 var_10337_to_fp16 = const()[name = string("op_10337_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_10338_cast_fp16 = mul(x = var_10336_cast_fp16, y = var_10337_to_fp16)[name = string("op_10338_cast_fp16")];
            tensor<fp16, [1, 8, 16, 1]> var_10339_cast_fp16 = add(x = var_10334_cast_fp16, y = var_10338_cast_fp16)[name = string("op_10339_cast_fp16")];
            fp32 var_10340_epsilon_0 = const()[name = string("op_10340_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_10340_cast_fp16 = rsqrt(epsilon = var_10340_epsilon_0, x = var_10339_cast_fp16)[name = string("op_10340_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_10341_cast_fp16 = mul(x = x_551_cast_fp16, y = var_10340_cast_fp16)[name = string("op_10341_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_555_cast_fp16 = mul(x = var_10341_cast_fp16, y = model_layers_27_self_attn_k_norm_weight_to_fp16)[name = string("x_555_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_10350 = const()[name = string("op_10350"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_10351_cast_fp16 = reshape(shape = var_10350, x = linear_191_cast_fp16)[name = string("op_10351_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_553_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_553_cast_fp16)[name = string("second_109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10365_cast_fp16 = mul(x = first_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10365_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10366_cast_fp16 = mul(x = second_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10366_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10368_cast_fp16 = sub(x = var_10365_cast_fp16, y = var_10366_cast_fp16)[name = string("op_10368_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10369_cast_fp16 = mul(x = second_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10369_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10370_cast_fp16 = mul(x = first_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10370_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10372_cast_fp16 = add(x = var_10369_cast_fp16, y = var_10370_cast_fp16)[name = string("op_10372_cast_fp16")];
            int32 var_10374 = const()[name = string("op_10374"), 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_10374, interleave = q_interleave_0, values = (var_10368_cast_fp16, var_10372_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_555_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_555_cast_fp16)[name = string("second_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10386_cast_fp16 = mul(x = first_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10386_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10387_cast_fp16 = mul(x = second_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10387_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10389_cast_fp16 = sub(x = var_10386_cast_fp16, y = var_10387_cast_fp16)[name = string("op_10389_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10390_cast_fp16 = mul(x = second_cast_fp16, y = cos_1_cast_fp16)[name = string("op_10390_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10391_cast_fp16 = mul(x = first_cast_fp16, y = sin_1_cast_fp16)[name = string("op_10391_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_10393_cast_fp16 = add(x = var_10390_cast_fp16, y = var_10391_cast_fp16)[name = string("op_10393_cast_fp16")];
            int32 var_10395 = const()[name = string("op_10395"), 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_10395, interleave = k_interleave_0, values = (var_10389_cast_fp16, var_10393_cast_fp16))[name = string("k_cast_fp16")];
            tensor<int32, [4]> var_10406_begin_0 = const()[name = string("op_10406_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10406_end_0 = const()[name = string("op_10406_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10406_end_mask_0 = const()[name = string("op_10406_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10406_stride_0 = const()[name = string("op_10406_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10406_cast_fp16 = slice_by_index(begin = var_10406_begin_0, end = var_10406_end_0, end_mask = var_10406_end_mask_0, stride = var_10406_stride_0, x = q_cast_fp16)[name = string("op_10406_cast_fp16")];
            tensor<int32, [4]> var_10416_begin_0 = const()[name = string("op_10416_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_10416_end_0 = const()[name = string("op_10416_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_10416_end_mask_0 = const()[name = string("op_10416_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_10416_stride_0 = const()[name = string("op_10416_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_10416_cast_fp16 = slice_by_index(begin = var_10416_begin_0, end = var_10416_end_0, end_mask = var_10416_end_mask_0, stride = var_10416_stride_0, x = q_cast_fp16)[name = string("op_10416_cast_fp16")];
            int32 var_10418 = const()[name = string("op_10418"), 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_10418, interleave = stacked_interleave_0, values = (var_10406_cast_fp16, var_10416_cast_fp16))[name = string("stacked_cast_fp16")];
            tensor<int32, [4]> var_10424_begin_0 = const()[name = string("op_10424_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_10424_end_0 = const()[name = string("op_10424_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_10424_end_mask_0 = const()[name = string("op_10424_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_10424_cast_fp16 = slice_by_index(begin = var_10424_begin_0, end = var_10424_end_0, end_mask = var_10424_end_mask_0, x = read_state_0)[name = string("op_10424_cast_fp16")];
            tensor<int32, [4]> var_10434_begin_0 = const()[name = string("op_10434_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10434_end_0 = const()[name = string("op_10434_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_10434_end_mask_0 = const()[name = string("op_10434_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_10434_cast_fp16 = slice_by_index(begin = var_10434_begin_0, end = var_10434_end_0, end_mask = var_10434_end_mask_0, x = var_10424_cast_fp16)[name = string("op_10434_cast_fp16")];
            bool var_10443_transpose_x_1 = const()[name = string("op_10443_transpose_x_1"), val = bool(false)];
            bool var_10443_transpose_y_1 = const()[name = string("op_10443_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_10443_cast_fp16 = matmul(transpose_x = var_10443_transpose_x_1, transpose_y = var_10443_transpose_y_1, x = var_10434_cast_fp16, y = stacked_cast_fp16)[name = string("op_10443_cast_fp16")];
            tensor<int32, [4]> var_10446_perm_0 = const()[name = string("op_10446_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_10447_to_fp16 = const()[name = string("op_10447_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_10446_cast_fp16 = transpose(perm = var_10446_perm_0, x = var_10443_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, 32, 512]> var_10448_cast_fp16 = mul(x = var_10446_cast_fp16, y = var_10447_to_fp16)[name = string("op_10448_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_cast_fp16 = add(x = var_10448_cast_fp16, y = cacheMask)[name = string("cache_weights_cast_fp16")];
            bool var_10454_transpose_x_1 = const()[name = string("op_10454_transpose_x_1"), val = bool(false)];
            bool var_10454_transpose_y_1 = const()[name = string("op_10454_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_10454_cast_fp16 = matmul(transpose_x = var_10454_transpose_x_1, transpose_y = var_10454_transpose_y_1, x = stacked_cast_fp16, y = k_cast_fp16)[name = string("op_10454_cast_fp16")];
            fp16 var_10455_to_fp16 = const()[name = string("op_10455_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_10456_cast_fp16 = mul(x = var_10454_cast_fp16, y = var_10455_to_fp16)[name = string("op_10456_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_cast_fp16 = add(x = var_10456_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_cast_fp16")];
            int32 var_10460 = const()[name = string("op_10460"), val = int32(-1)];
            bool var_10461_interleave_0 = const()[name = string("op_10461_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_10461_cast_fp16 = concat(axis = var_10460, interleave = var_10461_interleave_0, values = (cache_weights_cast_fp16, call_weights_cast_fp16))[name = string("op_10461_cast_fp16")];
            int32 var_10462 = const()[name = string("op_10462"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_cast_fp16 = softmax(axis = var_10462, x = var_10461_cast_fp16)[name = string("shares_cast_fp16")];
            tensor<int32, [4]> var_10469_begin_0 = const()[name = string("op_10469_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10469_end_0 = const()[name = string("op_10469_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_10469_end_mask_0 = const()[name = string("op_10469_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_10469_cast_fp16 = slice_by_index(begin = var_10469_begin_0, end = var_10469_end_0, end_mask = var_10469_end_mask_0, x = shares_cast_fp16)[name = string("op_10469_cast_fp16")];
            tensor<int32, [4]> var_10474_begin_0 = const()[name = string("op_10474_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_10474_end_0 = const()[name = string("op_10474_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_10474_end_mask_0 = const()[name = string("op_10474_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_10474_cast_fp16 = slice_by_index(begin = var_10474_begin_0, end = var_10474_end_0, end_mask = var_10474_end_mask_0, x = read_state_1)[name = string("op_10474_cast_fp16")];
            tensor<int32, [4]> var_10484_begin_0 = const()[name = string("op_10484_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10484_end_0 = const()[name = string("op_10484_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_10484_end_mask_0 = const()[name = string("op_10484_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_10484_cast_fp16 = slice_by_index(begin = var_10484_begin_0, end = var_10484_end_0, end_mask = var_10484_end_mask_0, x = var_10474_cast_fp16)[name = string("op_10484_cast_fp16")];
            bool var_10490_transpose_x_0 = const()[name = string("op_10490_transpose_x_0"), val = bool(false)];
            bool var_10490_transpose_y_0 = const()[name = string("op_10490_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_10490_cast_fp16 = matmul(transpose_x = var_10490_transpose_x_0, transpose_y = var_10490_transpose_y_0, x = var_10469_cast_fp16, y = var_10484_cast_fp16)[name = string("op_10490_cast_fp16")];
            tensor<int32, [4]> var_10495_begin_0 = const()[name = string("op_10495_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_10495_end_0 = const()[name = string("op_10495_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_10495_end_mask_0 = const()[name = string("op_10495_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_10495_cast_fp16 = slice_by_index(begin = var_10495_begin_0, end = var_10495_end_0, end_mask = var_10495_end_mask_0, x = shares_cast_fp16)[name = string("op_10495_cast_fp16")];
            bool var_10496_transpose_x_0 = const()[name = string("op_10496_transpose_x_0"), val = bool(false)];
            bool var_10496_transpose_y_0 = const()[name = string("op_10496_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_10351_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 32, 128]> var_10496_cast_fp16 = matmul(transpose_x = var_10496_transpose_x_0, transpose_y = var_10496_transpose_y_0, x = var_10495_cast_fp16, y = v_cast_fp16)[name = string("op_10496_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_109_cast_fp16 = add(x = var_10490_cast_fp16, y = var_10496_cast_fp16)[name = string("mixed_109_cast_fp16")];
            tensor<int32, [4]> var_10513_begin_0 = const()[name = string("op_10513_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_10513_end_0 = const()[name = string("op_10513_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_10513_end_mask_0 = const()[name = string("op_10513_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10513_cast_fp16 = slice_by_index(begin = var_10513_begin_0, end = var_10513_end_0, end_mask = var_10513_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_10513_cast_fp16")];
            tensor<int32, [4]> var_10528_begin_0 = const()[name = string("op_10528_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_10528_end_0 = const()[name = string("op_10528_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_10528_end_mask_0 = const()[name = string("op_10528_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_10528_cast_fp16 = slice_by_index(begin = var_10528_begin_0, end = var_10528_end_0, end_mask = var_10528_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_10528_cast_fp16")];
            int32 var_10530 = const()[name = string("op_10530"), 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_10530, interleave = mixed_interleave_0, values = (var_10513_cast_fp16, var_10528_cast_fp16))[name = string("mixed_cast_fp16")];
            tensor<int32, [4]> var_10534_perm_0 = const()[name = string("op_10534_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_10538 = const()[name = string("op_10538"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_10534_cast_fp16 = transpose(perm = var_10534_perm_0, x = mixed_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 16, 2048]> input_273_cast_fp16 = reshape(shape = var_10538, x = var_10534_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_557_cast_fp16 = add(x = x_541_cast_fp16, y = linear_192_cast_fp16)[name = string("x_557_cast_fp16")];
            bool var_10548 = const()[name = string("op_10548"), val = bool(true)];
            tensor<fp16, [1, 16, 1024]> var_10551_cast_fp16 = abs(x = x_557_cast_fp16)[name = string("op_10551_cast_fp16")];
            tensor<int32, [1]> peak_223_axes_0 = const()[name = string("peak_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 16, 1]> peak_223_cast_fp16 = reduce_max(axes = peak_223_axes_0, keep_dims = var_10548, x = var_10551_cast_fp16)[name = string("peak_223_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_559_cast_fp16 = real_div(x = x_557_cast_fp16, y = peak_223_cast_fp16)[name = string("x_559_cast_fp16")];
            fp16 var_10547_promoted_to_fp16 = const()[name = string("op_10547_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10555_cast_fp16 = pow(x = x_559_cast_fp16, y = var_10547_promoted_to_fp16)[name = string("op_10555_cast_fp16")];
            tensor<int32, [1]> var_10557_axes_0 = const()[name = string("op_10557_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10557_keep_dims_0 = const()[name = string("op_10557_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_10557_cast_fp16 = reduce_mean(axes = var_10557_axes_0, keep_dims = var_10557_keep_dims_0, x = var_10555_cast_fp16)[name = string("op_10557_cast_fp16")];
            fp16 var_10547_promoted_1_to_fp16 = const()[name = string("op_10547_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1]> var_10558_cast_fp16 = pow(x = peak_223_cast_fp16, y = var_10547_promoted_1_to_fp16)[name = string("op_10558_cast_fp16")];
            fp32 var_10559_epsilon_0 = const()[name = string("op_10559_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 16, 1]> var_10559_cast_fp16 = inverse(epsilon = var_10559_epsilon_0, x = var_10558_cast_fp16)[name = string("op_10559_cast_fp16")];
            fp16 var_10560_to_fp16 = const()[name = string("op_10560_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_10561_cast_fp16 = mul(x = var_10559_cast_fp16, y = var_10560_to_fp16)[name = string("op_10561_cast_fp16")];
            tensor<fp16, [1, 16, 1]> var_10562_cast_fp16 = add(x = var_10557_cast_fp16, y = var_10561_cast_fp16)[name = string("op_10562_cast_fp16")];
            fp32 var_10563_epsilon_0 = const()[name = string("op_10563_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10563_cast_fp16 = rsqrt(epsilon = var_10563_epsilon_0, x = var_10562_cast_fp16)[name = string("op_10563_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_10564_cast_fp16 = mul(x = x_559_cast_fp16, y = var_10563_cast_fp16)[name = string("op_10564_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_10564_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_10569_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_10569_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_10569_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_561_cast_fp16 = add(x = x_557_cast_fp16, y = linear_195_cast_fp16)[name = string("x_561_cast_fp16")];
            int32 var_10579 = const()[name = string("op_10579"), val = int32(1)];
            int32 x_563_batch_dims_0 = const()[name = string("x_563_batch_dims_0"), val = int32(0)];
            bool x_563_validate_indices_0 = const()[name = string("x_563_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_563_cast_fp16_cast_uint16 = gather(axis = var_10579, batch_dims = x_563_batch_dims_0, indices = answerRow_to_uint16, validate_indices = x_563_validate_indices_0, x = x_561_cast_fp16)[name = string("x_563_cast_fp16_cast_uint16")];
            bool var_10584 = const()[name = string("op_10584"), val = bool(true)];
            tensor<fp16, [1, 1, 1024]> var_10587_cast_fp16 = abs(x = x_563_cast_fp16_cast_uint16)[name = string("op_10587_cast_fp16")];
            tensor<int32, [1]> peak_axes_0 = const()[name = string("peak_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1, 1]> peak_cast_fp16 = reduce_max(axes = peak_axes_0, keep_dims = var_10584, x = var_10587_cast_fp16)[name = string("peak_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> x_cast_fp16 = real_div(x = x_563_cast_fp16_cast_uint16, y = peak_cast_fp16)[name = string("x_cast_fp16")];
            fp16 var_10583_promoted_to_fp16 = const()[name = string("op_10583_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1024]> var_10591_cast_fp16 = pow(x = x_cast_fp16, y = var_10583_promoted_to_fp16)[name = string("op_10591_cast_fp16")];
            tensor<int32, [1]> var_10593_axes_0 = const()[name = string("op_10593_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_10593_keep_dims_0 = const()[name = string("op_10593_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1]> var_10593_cast_fp16 = reduce_mean(axes = var_10593_axes_0, keep_dims = var_10593_keep_dims_0, x = var_10591_cast_fp16)[name = string("op_10593_cast_fp16")];
            fp16 var_10583_promoted_1_to_fp16 = const()[name = string("op_10583_promoted_1_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1]> var_10594_cast_fp16 = pow(x = peak_cast_fp16, y = var_10583_promoted_1_to_fp16)[name = string("op_10594_cast_fp16")];
            fp32 var_10595_epsilon_0 = const()[name = string("op_10595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1, 1]> var_10595_cast_fp16 = inverse(epsilon = var_10595_epsilon_0, x = var_10594_cast_fp16)[name = string("op_10595_cast_fp16")];
            fp16 var_10596_to_fp16 = const()[name = string("op_10596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_10597_cast_fp16 = mul(x = var_10595_cast_fp16, y = var_10596_to_fp16)[name = string("op_10597_cast_fp16")];
            tensor<fp16, [1, 1, 1]> var_10598_cast_fp16 = add(x = var_10593_cast_fp16, y = var_10597_cast_fp16)[name = string("op_10598_cast_fp16")];
            fp32 var_10599_epsilon_0 = const()[name = string("op_10599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1]> var_10599_cast_fp16 = rsqrt(epsilon = var_10599_epsilon_0, x = var_10598_cast_fp16)[name = string("op_10599_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> var_10600_cast_fp16 = mul(x = x_cast_fp16, y = var_10599_cast_fp16)[name = string("op_10600_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_10600_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_10606_axis_0 = const()[name = string("op_10606_axis_0"), val = int32(-1)];
            bool var_10606_keep_dims_0 = const()[name = string("op_10606_keep_dims_0"), val = bool(false)];
            string var_10606_output_dtype_0 = const()[name = string("op_10606_output_dtype_0"), val = string("int32")];
            tensor<int32, [1, 1]> nextToken = reduce_argmax(axis = var_10606_axis_0, keep_dims = var_10606_keep_dims_0, output_dtype = var_10606_output_dtype_0, x = linear_196_cast_fp16)[name = string("op_10606_cast_fp16")];
            int32 var_10613 = const()[name = string("op_10613"), val = int32(0)];
            bool var_10614_interleave_0 = const()[name = string("op_10614_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newKeys = concat(axis = var_10613, interleave = var_10614_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_10614_cast_fp16")];
            int32 var_10616 = const()[name = string("op_10616"), val = int32(0)];
            bool var_10617_interleave_0 = const()[name = string("op_10617_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newValues = concat(axis = var_10616, interleave = var_10617_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_10617_cast_fp16")];
        } -> (nextToken, newKeys, newValues);
}