program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1.2"}})]
{
    func vision_140<ios18>(tensor<fp16, [1, 1260, 768]> patches, tensor<fp16, [140, 1260]> pool, tensor<fp16, [1, 1260, 768]> position_embeddings, tensor<fp16, [1, 1260, 2]> positions, tensor<fp16, [1, 1260]> valid) {
            fp16 var_1405_to_fp16 = const()[name = string("op_1405_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 1260, 768]> var_1406_cast_fp16 = sub(x = patches, y = var_1405_to_fp16)[name = string("op_1406_cast_fp16")];
            fp16 var_1407_promoted_to_fp16 = const()[name = string("op_1407_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1260, 768]> input_1_cast_fp16 = mul(x = var_1406_cast_fp16, y = var_1407_promoted_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [768, 768]> t_patch_embedder_input_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589952))))[name = string("t_patch_embedder_input_proj_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591552)))];
            tensor<fp16, [1, 1260, 768]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_patch_embedder_input_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [1]> var_1413_axes_0 = const()[name = string("op_1413_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1260, 1]> var_1413_cast_fp16 = expand_dims(axes = var_1413_axes_0, x = valid)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1414_cast_fp16 = mul(x = position_embeddings, y = var_1413_cast_fp16)[name = string("op_1414_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_1_cast_fp16 = add(x = linear_0_cast_fp16, y = var_1414_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [1]> var_1418_axes_0 = const()[name = string("op_1418_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 1260, 2, 1]> var_1418_cast_fp16 = expand_dims(axes = var_1418_axes_0, x = positions)[name = string("op_1418_cast_fp16")];
            tensor<fp16, [16]> inv_freq_to_fp16 = const()[name = string("inv_freq_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593152)))];
            tensor<fp16, [1, 1260, 2, 16]> f_1_cast_fp16 = mul(x = var_1418_cast_fp16, y = inv_freq_to_fp16)[name = string("f_1_cast_fp16")];
            int32 var_1421 = const()[name = string("op_1421"), val = int32(-1)];
            bool var_1422_interleave_0 = const()[name = string("op_1422_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 2, 32]> var_1422_cast_fp16 = concat(axis = var_1421, interleave = var_1422_interleave_0, values = (f_1_cast_fp16, f_1_cast_fp16))[name = string("op_1422_cast_fp16")];
            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1260, 1, 2, 32]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = var_1422_cast_fp16)[name = string("f_cast_fp16")];
            tensor<int32, [5]> var_1452_begin_0 = const()[name = string("op_1452_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1452_end_0 = const()[name = string("op_1452_end_0"), val = tensor<int32, [5]>([1, 1260, 1, 1, 32])];
            tensor<bool, [5]> var_1452_end_mask_0 = const()[name = string("op_1452_end_mask_0"), val = tensor<bool, [5]>([true, true, true, false, true])];
            tensor<bool, [5]> var_1452_squeeze_mask_0 = const()[name = string("op_1452_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, true, false])];
            tensor<fp16, [1, 1260, 1, 32]> var_1452_cast_fp16 = slice_by_index(begin = var_1452_begin_0, end = var_1452_end_0, end_mask = var_1452_end_mask_0, squeeze_mask = var_1452_squeeze_mask_0, x = f_cast_fp16)[name = string("op_1452_cast_fp16")];
            tensor<fp16, [1, 1260, 1, 32]> c_1_cast_fp16 = cos(x = var_1452_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [5]> var_1471_begin_0 = const()[name = string("op_1471_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 1, 0])];
            tensor<int32, [5]> var_1471_end_0 = const()[name = string("op_1471_end_0"), val = tensor<int32, [5]>([1, 1260, 1, 2, 32])];
            tensor<bool, [5]> var_1471_end_mask_0 = const()[name = string("op_1471_end_mask_0"), val = tensor<bool, [5]>([true, true, true, false, true])];
            tensor<bool, [5]> var_1471_squeeze_mask_0 = const()[name = string("op_1471_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, true, false])];
            tensor<fp16, [1, 1260, 1, 32]> var_1471_cast_fp16 = slice_by_index(begin = var_1471_begin_0, end = var_1471_end_0, end_mask = var_1471_end_mask_0, squeeze_mask = var_1471_squeeze_mask_0, x = f_cast_fp16)[name = string("op_1471_cast_fp16")];
            tensor<fp16, [1, 1260, 1, 32]> c_cast_fp16 = cos(x = var_1471_cast_fp16)[name = string("c_cast_fp16")];
            tensor<fp16, [1, 1260, 1, 32]> s_9_cast_fp16 = sin(x = var_1452_cast_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 1260, 1, 32]> s_11_cast_fp16 = sin(x = var_1471_cast_fp16)[name = string("s_11_cast_fp16")];
            fp16 var_1511_to_fp16 = const()[name = string("op_1511_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1260]> var_1513_cast_fp16 = sub(x = var_1511_to_fp16, y = valid)[name = string("op_1513_cast_fp16")];
            fp16 var_1514_to_fp16 = const()[name = string("op_1514_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 1260]> var_1515_cast_fp16 = mul(x = var_1513_cast_fp16, y = var_1514_to_fp16)[name = string("op_1515_cast_fp16")];
            tensor<int32, [1]> var_1522_axes_0 = const()[name = string("op_1522_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1260]> var_1522_cast_fp16 = expand_dims(axes = var_1522_axes_0, x = var_1515_cast_fp16)[name = string("op_1522_cast_fp16")];
            tensor<int32, [1]> var_1524_axes_0 = const()[name = string("op_1524_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 1260]> var_1524_cast_fp16 = expand_dims(axes = var_1524_axes_0, x = var_1522_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1530_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_1530_cast_fp16")];
            bool var_1533 = const()[name = string("op_1533"), val = bool(true)];
            tensor<int32, [1]> var_1534_axes_0 = const()[name = string("op_1534_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_1534_cast_fp16 = reduce_max(axes = var_1534_axes_0, keep_dims = var_1533, x = var_1530_cast_fp16)[name = string("op_1534_cast_fp16")];
            fp16 var_1535_to_fp16 = const()[name = string("op_1535_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_1_cast_fp16 = maximum(x = var_1534_cast_fp16, y = var_1535_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1538_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_1538_cast_fp16")];
            tensor<int32, [1]> var_1543_axes_0 = const()[name = string("op_1543_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1543_keep_dims_0 = const()[name = string("op_1543_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_1543_cast_fp16 = reduce_mean(axes = var_1543_axes_0, keep_dims = var_1543_keep_dims_0, x = var_1538_cast_fp16)[name = string("op_1543_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1544_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_1544_cast_fp16")];
            fp32 var_1545_epsilon_0 = const()[name = string("op_1545_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_1545_cast_fp16 = inverse(epsilon = var_1545_epsilon_0, x = var_1544_cast_fp16)[name = string("op_1545_cast_fp16")];
            fp16 var_1546_to_fp16 = const()[name = string("op_1546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_1547_cast_fp16 = mul(x = var_1545_cast_fp16, y = var_1546_to_fp16)[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1549_cast_fp16 = add(x = var_1543_cast_fp16, y = var_1547_cast_fp16)[name = string("op_1549_cast_fp16")];
            fp32 var_1550_epsilon_0 = const()[name = string("op_1550_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_1550_cast_fp16 = rsqrt(epsilon = var_1550_epsilon_0, x = var_1549_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_1550_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593280)))];
            tensor<fp16, [1, 1260, 768]> input_3_cast_fp16 = mul(x = out_1_cast_fp16, y = t_encoder_layers_0_input_layernorm_weight_to_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1184768))))[name = string("t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_1_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_1560 = const()[name = string("op_1560"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_3_cast_fp16 = reshape(shape = var_1560, x = linear_1_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1562_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_1562_cast_fp16")];
            bool var_1565 = const()[name = string("op_1565"), val = bool(true)];
            tensor<int32, [1]> var_1566_axes_0 = const()[name = string("op_1566_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_1566_cast_fp16 = reduce_max(axes = var_1566_axes_0, keep_dims = var_1565, x = var_1562_cast_fp16)[name = string("op_1566_cast_fp16")];
            fp16 var_1567_to_fp16 = const()[name = string("op_1567_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_3_cast_fp16 = maximum(x = var_1566_cast_fp16, y = var_1567_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_3_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1570_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_1570_cast_fp16")];
            tensor<int32, [1]> var_1575_axes_0 = const()[name = string("op_1575_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1575_keep_dims_0 = const()[name = string("op_1575_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_1575_cast_fp16 = reduce_mean(axes = var_1575_axes_0, keep_dims = var_1575_keep_dims_0, x = var_1570_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1576_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_1576_cast_fp16")];
            fp32 var_1577_epsilon_0 = const()[name = string("op_1577_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_1577_cast_fp16 = inverse(epsilon = var_1577_epsilon_0, x = var_1576_cast_fp16)[name = string("op_1577_cast_fp16")];
            fp16 var_1578_to_fp16 = const()[name = string("op_1578_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_1579_cast_fp16 = mul(x = var_1577_cast_fp16, y = var_1578_to_fp16)[name = string("op_1579_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1581_cast_fp16 = add(x = var_1575_cast_fp16, y = var_1579_cast_fp16)[name = string("op_1581_cast_fp16")];
            fp32 var_1582_epsilon_0 = const()[name = string("op_1582_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_1582_cast_fp16 = rsqrt(epsilon = var_1582_epsilon_0, x = var_1581_cast_fp16)[name = string("op_1582_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_1582_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 1260, 12, 64]> x_9_cast_fp16 = mul(x = out_3_cast_fp16, y = t_encoder_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186560))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1776448))))[name = string("t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1592 = const()[name = string("op_1592"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_5_cast_fp16 = reshape(shape = var_1592, x = linear_2_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1594_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_1594_cast_fp16")];
            bool var_1597 = const()[name = string("op_1597"), val = bool(true)];
            tensor<int32, [1]> var_1598_axes_0 = const()[name = string("op_1598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_1598_cast_fp16 = reduce_max(axes = var_1598_axes_0, keep_dims = var_1597, x = var_1594_cast_fp16)[name = string("op_1598_cast_fp16")];
            fp16 var_1599_to_fp16 = const()[name = string("op_1599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_5_cast_fp16 = maximum(x = var_1598_cast_fp16, y = var_1599_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_5_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1602_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_1602_cast_fp16")];
            tensor<int32, [1]> var_1607_axes_0 = const()[name = string("op_1607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1607_keep_dims_0 = const()[name = string("op_1607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_1607_cast_fp16 = reduce_mean(axes = var_1607_axes_0, keep_dims = var_1607_keep_dims_0, x = var_1602_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1608_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_1608_cast_fp16")];
            fp32 var_1609_epsilon_0 = const()[name = string("op_1609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_1609_cast_fp16 = inverse(epsilon = var_1609_epsilon_0, x = var_1608_cast_fp16)[name = string("op_1609_cast_fp16")];
            fp16 var_1610_to_fp16 = const()[name = string("op_1610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_1611_cast_fp16 = mul(x = var_1609_cast_fp16, y = var_1610_to_fp16)[name = string("op_1611_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1613_cast_fp16 = add(x = var_1607_cast_fp16, y = var_1611_cast_fp16)[name = string("op_1613_cast_fp16")];
            fp32 var_1614_epsilon_0 = const()[name = string("op_1614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_1614_cast_fp16 = rsqrt(epsilon = var_1614_epsilon_0, x = var_1613_cast_fp16)[name = string("op_1614_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_1614_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 1260, 12, 64]> x_11_cast_fp16 = mul(x = out_5_cast_fp16, y = t_encoder_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778240))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2368128))))[name = string("t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_1624 = const()[name = string("op_1624"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_7_cast_fp16 = reshape(shape = var_1624, x = linear_3_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1626_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_1626_cast_fp16")];
            bool var_1629 = const()[name = string("op_1629"), val = bool(true)];
            tensor<int32, [1]> var_1630_axes_0 = const()[name = string("op_1630_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_1630_cast_fp16 = reduce_max(axes = var_1630_axes_0, keep_dims = var_1629, x = var_1626_cast_fp16)[name = string("op_1630_cast_fp16")];
            fp16 var_1631_to_fp16 = const()[name = string("op_1631_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_7_cast_fp16 = maximum(x = var_1630_cast_fp16, y = var_1631_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_7_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1634_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_1634_cast_fp16")];
            tensor<int32, [1]> var_1639_axes_0 = const()[name = string("op_1639_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1639_keep_dims_0 = const()[name = string("op_1639_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_1639_cast_fp16 = reduce_mean(axes = var_1639_axes_0, keep_dims = var_1639_keep_dims_0, x = var_1634_cast_fp16)[name = string("op_1639_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1640_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_1640_cast_fp16")];
            fp32 var_1641_epsilon_0 = const()[name = string("op_1641_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_1641_cast_fp16 = inverse(epsilon = var_1641_epsilon_0, x = var_1640_cast_fp16)[name = string("op_1641_cast_fp16")];
            fp16 var_1642_to_fp16 = const()[name = string("op_1642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_1643_cast_fp16 = mul(x = var_1641_cast_fp16, y = var_1642_to_fp16)[name = string("op_1643_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1645_cast_fp16 = add(x = var_1639_cast_fp16, y = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp32 var_1646_epsilon_0 = const()[name = string("op_1646_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_1646_cast_fp16 = rsqrt(epsilon = var_1646_epsilon_0, x = var_1645_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_1_cast_fp16 = mul(x = y_7_cast_fp16, y = var_1646_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<int32, [4]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = x_9_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [4]> var_1657_begin_0 = const()[name = string("op_1657_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1657_end_0 = const()[name = string("op_1657_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_1657_end_mask_0 = const()[name = string("op_1657_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_1657_cast_fp16 = slice_by_index(begin = var_1657_begin_0, end = var_1657_end_0, end_mask = var_1657_end_mask_0, x = z_1_cast_fp16)[name = string("op_1657_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_1658_cast_fp16 = mul(x = var_1657_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [4]> var_1663_begin_0 = const()[name = string("op_1663_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1663_end_0 = const()[name = string("op_1663_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_1663_end_mask_0 = const()[name = string("op_1663_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_1663_cast_fp16 = slice_by_index(begin = var_1663_begin_0, end = var_1663_end_0, end_mask = var_1663_end_mask_0, x = z_1_cast_fp16)[name = string("op_1663_cast_fp16")];
            int32 var_1665 = const()[name = string("op_1665"), val = int32(-1)];
            bool rot_1_interleave_0 = const()[name = string("rot_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_1_cast_fp16 = concat(axis = var_1665, interleave = rot_1_interleave_0, values = (var_1658_cast_fp16, var_1663_cast_fp16))[name = string("rot_1_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1667_cast_fp16 = mul(x = z_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_1667_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1668_cast_fp16 = mul(x = rot_1_cast_fp16, y = s_9_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1670_cast_fp16 = add(x = var_1667_cast_fp16, y = var_1668_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<int32, [4]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = x_9_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [4]> var_1680_begin_0 = const()[name = string("op_1680_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1680_end_0 = const()[name = string("op_1680_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_1680_end_mask_0 = const()[name = string("op_1680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_1680_cast_fp16 = slice_by_index(begin = var_1680_begin_0, end = var_1680_end_0, end_mask = var_1680_end_mask_0, x = z_3_cast_fp16)[name = string("op_1680_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_1681_cast_fp16 = mul(x = var_1680_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> var_1686_begin_0 = const()[name = string("op_1686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1686_end_0 = const()[name = string("op_1686_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_1686_end_mask_0 = const()[name = string("op_1686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_1686_cast_fp16 = slice_by_index(begin = var_1686_begin_0, end = var_1686_end_0, end_mask = var_1686_end_mask_0, x = z_3_cast_fp16)[name = string("op_1686_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), val = int32(-1)];
            bool rot_3_interleave_0 = const()[name = string("rot_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_3_cast_fp16 = concat(axis = var_1688, interleave = rot_3_interleave_0, values = (var_1681_cast_fp16, var_1686_cast_fp16))[name = string("rot_3_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1690_cast_fp16 = mul(x = z_3_cast_fp16, y = c_cast_fp16)[name = string("op_1690_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1691_cast_fp16 = mul(x = rot_3_cast_fp16, y = s_11_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1693_cast_fp16 = add(x = var_1690_cast_fp16, y = var_1691_cast_fp16)[name = string("op_1693_cast_fp16")];
            int32 var_1695 = const()[name = string("op_1695"), val = int32(-1)];
            bool var_1696_interleave_0 = const()[name = string("op_1696_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_1696_cast_fp16 = concat(axis = var_1695, interleave = var_1696_interleave_0, values = (var_1670_cast_fp16, var_1693_cast_fp16))[name = string("op_1696_cast_fp16")];
            tensor<int32, [4]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = x_11_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [4]> var_1709_begin_0 = const()[name = string("op_1709_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1709_end_0 = const()[name = string("op_1709_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_1709_end_mask_0 = const()[name = string("op_1709_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_1709_cast_fp16 = slice_by_index(begin = var_1709_begin_0, end = var_1709_end_0, end_mask = var_1709_end_mask_0, x = z_5_cast_fp16)[name = string("op_1709_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_1710_cast_fp16 = mul(x = var_1709_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_1710_cast_fp16")];
            tensor<int32, [4]> var_1715_begin_0 = const()[name = string("op_1715_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1715_end_0 = const()[name = string("op_1715_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_1715_end_mask_0 = const()[name = string("op_1715_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_1715_cast_fp16 = slice_by_index(begin = var_1715_begin_0, end = var_1715_end_0, end_mask = var_1715_end_mask_0, x = z_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            int32 var_1717 = const()[name = string("op_1717"), val = int32(-1)];
            bool rot_5_interleave_0 = const()[name = string("rot_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_5_cast_fp16 = concat(axis = var_1717, interleave = rot_5_interleave_0, values = (var_1710_cast_fp16, var_1715_cast_fp16))[name = string("rot_5_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1719_cast_fp16 = mul(x = z_5_cast_fp16, y = c_1_cast_fp16)[name = string("op_1719_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1720_cast_fp16 = mul(x = rot_5_cast_fp16, y = s_9_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1722_cast_fp16 = add(x = var_1719_cast_fp16, y = var_1720_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<int32, [4]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = x_11_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [4]> var_1732_begin_0 = const()[name = string("op_1732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1732_end_0 = const()[name = string("op_1732_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_1732_end_mask_0 = const()[name = string("op_1732_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_1732_cast_fp16 = slice_by_index(begin = var_1732_begin_0, end = var_1732_end_0, end_mask = var_1732_end_mask_0, x = z_7_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_1733_cast_fp16 = mul(x = var_1732_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_1733_cast_fp16")];
            tensor<int32, [4]> var_1738_begin_0 = const()[name = string("op_1738_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1738_end_0 = const()[name = string("op_1738_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_1738_end_mask_0 = const()[name = string("op_1738_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_1738_cast_fp16 = slice_by_index(begin = var_1738_begin_0, end = var_1738_end_0, end_mask = var_1738_end_mask_0, x = z_7_cast_fp16)[name = string("op_1738_cast_fp16")];
            int32 var_1740 = const()[name = string("op_1740"), val = int32(-1)];
            bool rot_7_interleave_0 = const()[name = string("rot_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_7_cast_fp16 = concat(axis = var_1740, interleave = rot_7_interleave_0, values = (var_1733_cast_fp16, var_1738_cast_fp16))[name = string("rot_7_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1742_cast_fp16 = mul(x = z_7_cast_fp16, y = c_cast_fp16)[name = string("op_1742_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1743_cast_fp16 = mul(x = rot_7_cast_fp16, y = s_11_cast_fp16)[name = string("op_1743_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1745_cast_fp16 = add(x = var_1742_cast_fp16, y = var_1743_cast_fp16)[name = string("op_1745_cast_fp16")];
            int32 var_1747 = const()[name = string("op_1747"), val = int32(-1)];
            bool var_1748_interleave_0 = const()[name = string("op_1748_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_1748_cast_fp16 = concat(axis = var_1747, interleave = var_1748_interleave_0, values = (var_1722_cast_fp16, var_1745_cast_fp16))[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_1758_transpose_x_0 = const()[name = string("op_1758_transpose_x_0"), val = bool(false)];
            bool var_1758_transpose_y_0 = const()[name = string("op_1758_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1748_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_64 = transpose(perm = transpose_64_perm_0, x = var_1696_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 12, 1260, 1260]> var_1758_cast_fp16 = matmul(transpose_x = var_1758_transpose_x_0, transpose_y = var_1758_transpose_y_0, x = transpose_64, y = transpose_65)[name = string("op_1758_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_1760_cast_fp16 = add(x = var_1758_cast_fp16, y = var_1524_cast_fp16)[name = string("op_1760_cast_fp16")];
            int32 var_1761 = const()[name = string("op_1761"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_7_cast_fp16 = softmax(axis = var_1761, x = var_1760_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1764_transpose_x_0 = const()[name = string("op_1764_transpose_x_0"), val = bool(false)];
            bool var_1764_transpose_y_0 = const()[name = string("op_1764_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = v_1_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 12, 1260, 64]> var_1764_cast_fp16 = matmul(transpose_x = var_1764_transpose_x_0, transpose_y = var_1764_transpose_y_0, x = w_7_cast_fp16, y = v_3_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<int32, [4]> var_1767_perm_0 = const()[name = string("op_1767_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1771 = const()[name = string("op_1771"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = var_1764_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 1260, 768]> input_5_cast_fp16 = reshape(shape = var_1771, x = var_1767_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2369728))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2959616))))[name = string("t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1776_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_1776_cast_fp16")];
            bool var_1779 = const()[name = string("op_1779"), val = bool(true)];
            tensor<int32, [1]> var_1780_axes_0 = const()[name = string("op_1780_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_1780_cast_fp16 = reduce_max(axes = var_1780_axes_0, keep_dims = var_1779, x = var_1776_cast_fp16)[name = string("op_1780_cast_fp16")];
            fp16 var_1781_to_fp16 = const()[name = string("op_1781_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_13_cast_fp16 = maximum(x = var_1780_cast_fp16, y = var_1781_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_9_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_13_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1784_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_1784_cast_fp16")];
            tensor<int32, [1]> var_1789_axes_0 = const()[name = string("op_1789_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1789_keep_dims_0 = const()[name = string("op_1789_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_1789_cast_fp16 = reduce_mean(axes = var_1789_axes_0, keep_dims = var_1789_keep_dims_0, x = var_1784_cast_fp16)[name = string("op_1789_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1790_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_1790_cast_fp16")];
            fp32 var_1791_epsilon_0 = const()[name = string("op_1791_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_1791_cast_fp16 = inverse(epsilon = var_1791_epsilon_0, x = var_1790_cast_fp16)[name = string("op_1791_cast_fp16")];
            fp16 var_1792_to_fp16 = const()[name = string("op_1792_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_1793_cast_fp16 = mul(x = var_1791_cast_fp16, y = var_1792_to_fp16)[name = string("op_1793_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1795_cast_fp16 = add(x = var_1789_cast_fp16, y = var_1793_cast_fp16)[name = string("op_1795_cast_fp16")];
            fp32 var_1796_epsilon_0 = const()[name = string("op_1796_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_1796_cast_fp16 = rsqrt(epsilon = var_1796_epsilon_0, x = var_1795_cast_fp16)[name = string("op_1796_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_7_cast_fp16 = mul(x = y_9_cast_fp16, y = var_1796_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2961216)))];
            tensor<fp16, [1, 1260, 768]> var_1798_cast_fp16 = mul(x = out_7_cast_fp16, y = t_encoder_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_1798_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_15_cast_fp16 = add(x = x_1_cast_fp16, y = var_1798_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1801_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_1801_cast_fp16")];
            bool var_1804 = const()[name = string("op_1804"), val = bool(true)];
            tensor<int32, [1]> var_1805_axes_0 = const()[name = string("op_1805_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_1805_cast_fp16 = reduce_max(axes = var_1805_axes_0, keep_dims = var_1804, x = var_1801_cast_fp16)[name = string("op_1805_cast_fp16")];
            fp16 var_1806_to_fp16 = const()[name = string("op_1806_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_15_cast_fp16 = maximum(x = var_1805_cast_fp16, y = var_1806_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_11_cast_fp16 = real_div(x = x_15_cast_fp16, y = s_15_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1809_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [1]> var_1814_axes_0 = const()[name = string("op_1814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1814_keep_dims_0 = const()[name = string("op_1814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_1814_cast_fp16 = reduce_mean(axes = var_1814_axes_0, keep_dims = var_1814_keep_dims_0, x = var_1809_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1815_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_1815_cast_fp16")];
            fp32 var_1816_epsilon_0 = const()[name = string("op_1816_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_1816_cast_fp16 = inverse(epsilon = var_1816_epsilon_0, x = var_1815_cast_fp16)[name = string("op_1816_cast_fp16")];
            fp16 var_1817_to_fp16 = const()[name = string("op_1817_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_1818_cast_fp16 = mul(x = var_1816_cast_fp16, y = var_1817_to_fp16)[name = string("op_1818_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1820_cast_fp16 = add(x = var_1814_cast_fp16, y = var_1818_cast_fp16)[name = string("op_1820_cast_fp16")];
            fp32 var_1821_epsilon_0 = const()[name = string("op_1821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_1821_cast_fp16 = rsqrt(epsilon = var_1821_epsilon_0, x = var_1820_cast_fp16)[name = string("op_1821_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_1821_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2962816)))];
            tensor<fp16, [1, 1260, 768]> input_7_cast_fp16 = mul(x = out_9_cast_fp16, y = t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2964416))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5323776))))[name = string("t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5329984)))];
            tensor<fp16, [1, 1260, 3072]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_1828_mode_0 = const()[name = string("op_1828_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_1828_cast_fp16 = gelu(mode = var_1828_mode_0, x = linear_5_cast_fp16)[name = string("op_1828_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5336192))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7695552))))[name = string("t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_11_cast_fp16 = mul(x = var_1828_cast_fp16, y = linear_6_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7701760))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10061120))))[name = string("t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1836_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_1836_cast_fp16")];
            bool var_1839 = const()[name = string("op_1839"), val = bool(true)];
            tensor<int32, [1]> var_1840_axes_0 = const()[name = string("op_1840_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_1840_cast_fp16 = reduce_max(axes = var_1840_axes_0, keep_dims = var_1839, x = var_1836_cast_fp16)[name = string("op_1840_cast_fp16")];
            fp16 var_1841_to_fp16 = const()[name = string("op_1841_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_17_cast_fp16 = maximum(x = var_1840_cast_fp16, y = var_1841_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_13_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_17_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1844_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_1844_cast_fp16")];
            tensor<int32, [1]> var_1849_axes_0 = const()[name = string("op_1849_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1849_keep_dims_0 = const()[name = string("op_1849_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_1849_cast_fp16 = reduce_mean(axes = var_1849_axes_0, keep_dims = var_1849_keep_dims_0, x = var_1844_cast_fp16)[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1850_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_1850_cast_fp16")];
            fp32 var_1851_epsilon_0 = const()[name = string("op_1851_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_1851_cast_fp16 = inverse(epsilon = var_1851_epsilon_0, x = var_1850_cast_fp16)[name = string("op_1851_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_1853_cast_fp16 = mul(x = var_1851_cast_fp16, y = var_1852_to_fp16)[name = string("op_1853_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1855_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1853_cast_fp16)[name = string("op_1855_cast_fp16")];
            fp32 var_1856_epsilon_0 = const()[name = string("op_1856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_1856_cast_fp16 = rsqrt(epsilon = var_1856_epsilon_0, x = var_1855_cast_fp16)[name = string("op_1856_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_1856_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10062720)))];
            tensor<fp16, [1, 1260, 768]> var_1858_cast_fp16 = mul(x = out_11_cast_fp16, y = t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_1858_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_19_cast_fp16 = add(x = x_15_cast_fp16, y = var_1858_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1861_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_1861_cast_fp16")];
            bool var_1864 = const()[name = string("op_1864"), val = bool(true)];
            tensor<int32, [1]> var_1865_axes_0 = const()[name = string("op_1865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_1865_cast_fp16 = reduce_max(axes = var_1865_axes_0, keep_dims = var_1864, x = var_1861_cast_fp16)[name = string("op_1865_cast_fp16")];
            fp16 var_1866_to_fp16 = const()[name = string("op_1866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_19_cast_fp16 = maximum(x = var_1865_cast_fp16, y = var_1866_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_15_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_1869_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_1869_cast_fp16")];
            tensor<int32, [1]> var_1874_axes_0 = const()[name = string("op_1874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1874_keep_dims_0 = const()[name = string("op_1874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_1874_cast_fp16 = reduce_mean(axes = var_1874_axes_0, keep_dims = var_1874_keep_dims_0, x = var_1869_cast_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1875_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_1875_cast_fp16")];
            fp32 var_1876_epsilon_0 = const()[name = string("op_1876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_1876_cast_fp16 = inverse(epsilon = var_1876_epsilon_0, x = var_1875_cast_fp16)[name = string("op_1876_cast_fp16")];
            fp16 var_1877_to_fp16 = const()[name = string("op_1877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_1878_cast_fp16 = mul(x = var_1876_cast_fp16, y = var_1877_to_fp16)[name = string("op_1878_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_1880_cast_fp16 = add(x = var_1874_cast_fp16, y = var_1878_cast_fp16)[name = string("op_1880_cast_fp16")];
            fp32 var_1881_epsilon_0 = const()[name = string("op_1881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_1881_cast_fp16 = rsqrt(epsilon = var_1881_epsilon_0, x = var_1880_cast_fp16)[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_1881_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10064320)))];
            tensor<fp16, [1, 1260, 768]> input_13_cast_fp16 = mul(x = out_13_cast_fp16, y = t_encoder_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10065920))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10655808))))[name = string("t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_1891 = const()[name = string("op_1891"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_21_cast_fp16 = reshape(shape = var_1891, x = linear_8_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1893_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1893_cast_fp16")];
            bool var_1896 = const()[name = string("op_1896"), val = bool(true)];
            tensor<int32, [1]> var_1897_axes_0 = const()[name = string("op_1897_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_1897_cast_fp16 = reduce_max(axes = var_1897_axes_0, keep_dims = var_1896, x = var_1893_cast_fp16)[name = string("op_1897_cast_fp16")];
            fp16 var_1898_to_fp16 = const()[name = string("op_1898_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_21_cast_fp16 = maximum(x = var_1897_cast_fp16, y = var_1898_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_17_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1901_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_1906_axes_0 = const()[name = string("op_1906_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1906_keep_dims_0 = const()[name = string("op_1906_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_1906_cast_fp16 = reduce_mean(axes = var_1906_axes_0, keep_dims = var_1906_keep_dims_0, x = var_1901_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1907_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_1907_cast_fp16")];
            fp32 var_1908_epsilon_0 = const()[name = string("op_1908_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_1908_cast_fp16 = inverse(epsilon = var_1908_epsilon_0, x = var_1907_cast_fp16)[name = string("op_1908_cast_fp16")];
            fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_1910_cast_fp16 = mul(x = var_1908_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1912_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1910_cast_fp16)[name = string("op_1912_cast_fp16")];
            fp32 var_1913_epsilon_0 = const()[name = string("op_1913_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_1913_cast_fp16 = rsqrt(epsilon = var_1913_epsilon_0, x = var_1912_cast_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_1913_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 1260, 12, 64]> x_27_cast_fp16 = mul(x = out_15_cast_fp16, y = t_encoder_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10657408))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11247296))))[name = string("t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_9_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_1923 = const()[name = string("op_1923"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_23_cast_fp16 = reshape(shape = var_1923, x = linear_9_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1925_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1925_cast_fp16")];
            bool var_1928 = const()[name = string("op_1928"), val = bool(true)];
            tensor<int32, [1]> var_1929_axes_0 = const()[name = string("op_1929_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_1929_cast_fp16 = reduce_max(axes = var_1929_axes_0, keep_dims = var_1928, x = var_1925_cast_fp16)[name = string("op_1929_cast_fp16")];
            fp16 var_1930_to_fp16 = const()[name = string("op_1930_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_23_cast_fp16 = maximum(x = var_1929_cast_fp16, y = var_1930_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_19_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1933_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<int32, [1]> var_1938_axes_0 = const()[name = string("op_1938_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1938_keep_dims_0 = const()[name = string("op_1938_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_1938_cast_fp16 = reduce_mean(axes = var_1938_axes_0, keep_dims = var_1938_keep_dims_0, x = var_1933_cast_fp16)[name = string("op_1938_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1939_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_1939_cast_fp16")];
            fp32 var_1940_epsilon_0 = const()[name = string("op_1940_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_1940_cast_fp16 = inverse(epsilon = var_1940_epsilon_0, x = var_1939_cast_fp16)[name = string("op_1940_cast_fp16")];
            fp16 var_1941_to_fp16 = const()[name = string("op_1941_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_1942_cast_fp16 = mul(x = var_1940_cast_fp16, y = var_1941_to_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1944_cast_fp16 = add(x = var_1938_cast_fp16, y = var_1942_cast_fp16)[name = string("op_1944_cast_fp16")];
            fp32 var_1945_epsilon_0 = const()[name = string("op_1945_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_1945_cast_fp16 = rsqrt(epsilon = var_1945_epsilon_0, x = var_1944_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_1945_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 1260, 12, 64]> x_29_cast_fp16 = mul(x = out_17_cast_fp16, y = t_encoder_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11248896))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11838784))))[name = string("t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_25_cast_fp16 = reshape(shape = var_1955, x = linear_10_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1957_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1957_cast_fp16")];
            bool var_1960 = const()[name = string("op_1960"), val = bool(true)];
            tensor<int32, [1]> var_1961_axes_0 = const()[name = string("op_1961_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_1961_cast_fp16 = reduce_max(axes = var_1961_axes_0, keep_dims = var_1960, x = var_1957_cast_fp16)[name = string("op_1961_cast_fp16")];
            fp16 var_1962_to_fp16 = const()[name = string("op_1962_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_25_cast_fp16 = maximum(x = var_1961_cast_fp16, y = var_1962_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_21_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_1965_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_1965_cast_fp16")];
            tensor<int32, [1]> var_1970_axes_0 = const()[name = string("op_1970_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1970_keep_dims_0 = const()[name = string("op_1970_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_1970_cast_fp16 = reduce_mean(axes = var_1970_axes_0, keep_dims = var_1970_keep_dims_0, x = var_1965_cast_fp16)[name = string("op_1970_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1971_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_1971_cast_fp16")];
            fp32 var_1972_epsilon_0 = const()[name = string("op_1972_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_1972_cast_fp16 = inverse(epsilon = var_1972_epsilon_0, x = var_1971_cast_fp16)[name = string("op_1972_cast_fp16")];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_1974_cast_fp16 = mul(x = var_1972_cast_fp16, y = var_1973_to_fp16)[name = string("op_1974_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_1976_cast_fp16 = add(x = var_1970_cast_fp16, y = var_1974_cast_fp16)[name = string("op_1976_cast_fp16")];
            fp32 var_1977_epsilon_0 = const()[name = string("op_1977_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_1977_cast_fp16 = rsqrt(epsilon = var_1977_epsilon_0, x = var_1976_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_5_cast_fp16 = mul(x = y_21_cast_fp16, y = var_1977_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<int32, [4]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = x_27_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [4]> var_1988_begin_0 = const()[name = string("op_1988_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1988_end_0 = const()[name = string("op_1988_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_1988_end_mask_0 = const()[name = string("op_1988_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_1988_cast_fp16 = slice_by_index(begin = var_1988_begin_0, end = var_1988_end_0, end_mask = var_1988_end_mask_0, x = z_9_cast_fp16)[name = string("op_1988_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_1989_cast_fp16 = mul(x = var_1988_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1989_cast_fp16")];
            tensor<int32, [4]> var_1994_begin_0 = const()[name = string("op_1994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1994_end_0 = const()[name = string("op_1994_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_1994_end_mask_0 = const()[name = string("op_1994_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_1994_cast_fp16 = slice_by_index(begin = var_1994_begin_0, end = var_1994_end_0, end_mask = var_1994_end_mask_0, x = z_9_cast_fp16)[name = string("op_1994_cast_fp16")];
            int32 var_1996 = const()[name = string("op_1996"), val = int32(-1)];
            bool rot_9_interleave_0 = const()[name = string("rot_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_9_cast_fp16 = concat(axis = var_1996, interleave = rot_9_interleave_0, values = (var_1989_cast_fp16, var_1994_cast_fp16))[name = string("rot_9_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1998_cast_fp16 = mul(x = z_9_cast_fp16, y = c_1_cast_fp16)[name = string("op_1998_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_1999_cast_fp16 = mul(x = rot_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_1999_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2001_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<int32, [4]> z_11_begin_0 = const()[name = string("z_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_11_end_0 = const()[name = string("z_11_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_11_end_mask_0 = const()[name = string("z_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_11_cast_fp16 = slice_by_index(begin = z_11_begin_0, end = z_11_end_0, end_mask = z_11_end_mask_0, x = x_27_cast_fp16)[name = string("z_11_cast_fp16")];
            tensor<int32, [4]> var_2011_begin_0 = const()[name = string("op_2011_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2011_end_0 = const()[name = string("op_2011_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2011_end_mask_0 = const()[name = string("op_2011_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2011_cast_fp16 = slice_by_index(begin = var_2011_begin_0, end = var_2011_end_0, end_mask = var_2011_end_mask_0, x = z_11_cast_fp16)[name = string("op_2011_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2012_cast_fp16 = mul(x = var_2011_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2012_cast_fp16")];
            tensor<int32, [4]> var_2017_begin_0 = const()[name = string("op_2017_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2017_end_0 = const()[name = string("op_2017_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2017_end_mask_0 = const()[name = string("op_2017_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2017_cast_fp16 = slice_by_index(begin = var_2017_begin_0, end = var_2017_end_0, end_mask = var_2017_end_mask_0, x = z_11_cast_fp16)[name = string("op_2017_cast_fp16")];
            int32 var_2019 = const()[name = string("op_2019"), val = int32(-1)];
            bool rot_11_interleave_0 = const()[name = string("rot_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_11_cast_fp16 = concat(axis = var_2019, interleave = rot_11_interleave_0, values = (var_2012_cast_fp16, var_2017_cast_fp16))[name = string("rot_11_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2021_cast_fp16 = mul(x = z_11_cast_fp16, y = c_cast_fp16)[name = string("op_2021_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2022_cast_fp16 = mul(x = rot_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2022_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2024_cast_fp16 = add(x = var_2021_cast_fp16, y = var_2022_cast_fp16)[name = string("op_2024_cast_fp16")];
            int32 var_2026 = const()[name = string("op_2026"), val = int32(-1)];
            bool var_2027_interleave_0 = const()[name = string("op_2027_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_2027_cast_fp16 = concat(axis = var_2026, interleave = var_2027_interleave_0, values = (var_2001_cast_fp16, var_2024_cast_fp16))[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> z_13_begin_0 = const()[name = string("z_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_13_end_0 = const()[name = string("z_13_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_13_end_mask_0 = const()[name = string("z_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_13_cast_fp16 = slice_by_index(begin = z_13_begin_0, end = z_13_end_0, end_mask = z_13_end_mask_0, x = x_29_cast_fp16)[name = string("z_13_cast_fp16")];
            tensor<int32, [4]> var_2040_begin_0 = const()[name = string("op_2040_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2040_end_0 = const()[name = string("op_2040_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2040_end_mask_0 = const()[name = string("op_2040_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, x = z_13_cast_fp16)[name = string("op_2040_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2041_cast_fp16 = mul(x = var_2040_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> var_2046_begin_0 = const()[name = string("op_2046_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2046_end_0 = const()[name = string("op_2046_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2046_end_mask_0 = const()[name = string("op_2046_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2046_cast_fp16 = slice_by_index(begin = var_2046_begin_0, end = var_2046_end_0, end_mask = var_2046_end_mask_0, x = z_13_cast_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2048 = const()[name = string("op_2048"), val = int32(-1)];
            bool rot_13_interleave_0 = const()[name = string("rot_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_13_cast_fp16 = concat(axis = var_2048, interleave = rot_13_interleave_0, values = (var_2041_cast_fp16, var_2046_cast_fp16))[name = string("rot_13_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2050_cast_fp16 = mul(x = z_13_cast_fp16, y = c_1_cast_fp16)[name = string("op_2050_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2051_cast_fp16 = mul(x = rot_13_cast_fp16, y = s_9_cast_fp16)[name = string("op_2051_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2053_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2051_cast_fp16)[name = string("op_2053_cast_fp16")];
            tensor<int32, [4]> z_15_begin_0 = const()[name = string("z_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_15_end_0 = const()[name = string("z_15_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_15_end_mask_0 = const()[name = string("z_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_15_cast_fp16 = slice_by_index(begin = z_15_begin_0, end = z_15_end_0, end_mask = z_15_end_mask_0, x = x_29_cast_fp16)[name = string("z_15_cast_fp16")];
            tensor<int32, [4]> var_2063_begin_0 = const()[name = string("op_2063_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2063_end_0 = const()[name = string("op_2063_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2063_end_mask_0 = const()[name = string("op_2063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2063_cast_fp16 = slice_by_index(begin = var_2063_begin_0, end = var_2063_end_0, end_mask = var_2063_end_mask_0, x = z_15_cast_fp16)[name = string("op_2063_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2064_cast_fp16 = mul(x = var_2063_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2064_cast_fp16")];
            tensor<int32, [4]> var_2069_begin_0 = const()[name = string("op_2069_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2069_end_0 = const()[name = string("op_2069_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2069_end_mask_0 = const()[name = string("op_2069_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2069_cast_fp16 = slice_by_index(begin = var_2069_begin_0, end = var_2069_end_0, end_mask = var_2069_end_mask_0, x = z_15_cast_fp16)[name = string("op_2069_cast_fp16")];
            int32 var_2071 = const()[name = string("op_2071"), val = int32(-1)];
            bool rot_15_interleave_0 = const()[name = string("rot_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_15_cast_fp16 = concat(axis = var_2071, interleave = rot_15_interleave_0, values = (var_2064_cast_fp16, var_2069_cast_fp16))[name = string("rot_15_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2073_cast_fp16 = mul(x = z_15_cast_fp16, y = c_cast_fp16)[name = string("op_2073_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2074_cast_fp16 = mul(x = rot_15_cast_fp16, y = s_11_cast_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2076_cast_fp16 = add(x = var_2073_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2076_cast_fp16")];
            int32 var_2078 = const()[name = string("op_2078"), val = int32(-1)];
            bool var_2079_interleave_0 = const()[name = string("op_2079_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_2079_cast_fp16 = concat(axis = var_2078, interleave = var_2079_interleave_0, values = (var_2053_cast_fp16, var_2076_cast_fp16))[name = string("op_2079_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2089_transpose_x_0 = const()[name = string("op_2089_transpose_x_0"), val = bool(false)];
            bool var_2089_transpose_y_0 = const()[name = string("op_2089_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_2079_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_66 = transpose(perm = transpose_66_perm_0, x = var_2027_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [1, 12, 1260, 1260]> var_2089_cast_fp16 = matmul(transpose_x = var_2089_transpose_x_0, transpose_y = var_2089_transpose_y_0, x = transpose_66, y = transpose_67)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_2091_cast_fp16 = add(x = var_2089_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2091_cast_fp16")];
            int32 var_2092 = const()[name = string("op_2092"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_21_cast_fp16 = softmax(axis = var_2092, x = var_2091_cast_fp16)[name = string("w_21_cast_fp16")];
            bool var_2095_transpose_x_0 = const()[name = string("op_2095_transpose_x_0"), val = bool(false)];
            bool var_2095_transpose_y_0 = const()[name = string("op_2095_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = v_5_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 12, 1260, 64]> var_2095_cast_fp16 = matmul(transpose_x = var_2095_transpose_x_0, transpose_y = var_2095_transpose_y_0, x = w_21_cast_fp16, y = v_7_cast_fp16)[name = string("op_2095_cast_fp16")];
            tensor<int32, [4]> var_2098_perm_0 = const()[name = string("op_2098_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2102 = const()[name = string("op_2102"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_2098_cast_fp16 = transpose(perm = var_2098_perm_0, x = var_2095_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 1260, 768]> input_15_cast_fp16 = reshape(shape = var_2102, x = var_2098_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11840384))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12430272))))[name = string("t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_11_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2107_cast_fp16 = abs(x = linear_11_cast_fp16)[name = string("op_2107_cast_fp16")];
            bool var_2110 = const()[name = string("op_2110"), val = bool(true)];
            tensor<int32, [1]> var_2111_axes_0 = const()[name = string("op_2111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2111_cast_fp16 = reduce_max(axes = var_2111_axes_0, keep_dims = var_2110, x = var_2107_cast_fp16)[name = string("op_2111_cast_fp16")];
            fp16 var_2112_to_fp16 = const()[name = string("op_2112_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_27_cast_fp16 = maximum(x = var_2111_cast_fp16, y = var_2112_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_23_cast_fp16 = real_div(x = linear_11_cast_fp16, y = s_27_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2115_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<int32, [1]> var_2120_axes_0 = const()[name = string("op_2120_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2120_keep_dims_0 = const()[name = string("op_2120_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2120_cast_fp16 = reduce_mean(axes = var_2120_axes_0, keep_dims = var_2120_keep_dims_0, x = var_2115_cast_fp16)[name = string("op_2120_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2121_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2121_cast_fp16")];
            fp32 var_2122_epsilon_0 = const()[name = string("op_2122_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2122_cast_fp16 = inverse(epsilon = var_2122_epsilon_0, x = var_2121_cast_fp16)[name = string("op_2122_cast_fp16")];
            fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2124_cast_fp16 = mul(x = var_2122_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2126_cast_fp16 = add(x = var_2120_cast_fp16, y = var_2124_cast_fp16)[name = string("op_2126_cast_fp16")];
            fp32 var_2127_epsilon_0 = const()[name = string("op_2127_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2127_cast_fp16 = rsqrt(epsilon = var_2127_epsilon_0, x = var_2126_cast_fp16)[name = string("op_2127_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_19_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2127_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12431872)))];
            tensor<fp16, [1, 1260, 768]> var_2129_cast_fp16 = mul(x = out_19_cast_fp16, y = t_encoder_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2129_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_33_cast_fp16 = add(x = x_19_cast_fp16, y = var_2129_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2132_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_2132_cast_fp16")];
            bool var_2135 = const()[name = string("op_2135"), val = bool(true)];
            tensor<int32, [1]> var_2136_axes_0 = const()[name = string("op_2136_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2136_cast_fp16 = reduce_max(axes = var_2136_axes_0, keep_dims = var_2135, x = var_2132_cast_fp16)[name = string("op_2136_cast_fp16")];
            fp16 var_2137_to_fp16 = const()[name = string("op_2137_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_29_cast_fp16 = maximum(x = var_2136_cast_fp16, y = var_2137_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_25_cast_fp16 = real_div(x = x_33_cast_fp16, y = s_29_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2140_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<int32, [1]> var_2145_axes_0 = const()[name = string("op_2145_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2145_keep_dims_0 = const()[name = string("op_2145_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2145_cast_fp16 = reduce_mean(axes = var_2145_axes_0, keep_dims = var_2145_keep_dims_0, x = var_2140_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2146_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2146_cast_fp16")];
            fp32 var_2147_epsilon_0 = const()[name = string("op_2147_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2147_cast_fp16 = inverse(epsilon = var_2147_epsilon_0, x = var_2146_cast_fp16)[name = string("op_2147_cast_fp16")];
            fp16 var_2148_to_fp16 = const()[name = string("op_2148_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2149_cast_fp16 = mul(x = var_2147_cast_fp16, y = var_2148_to_fp16)[name = string("op_2149_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2151_cast_fp16 = add(x = var_2145_cast_fp16, y = var_2149_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2152_cast_fp16 = rsqrt(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_21_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2152_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12433472)))];
            tensor<fp16, [1, 1260, 768]> input_17_cast_fp16 = mul(x = out_21_cast_fp16, y = t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12435072))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14794432))))[name = string("t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_12_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_12_cast_fp16")];
            string var_2159_mode_0 = const()[name = string("op_2159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_2159_cast_fp16 = gelu(mode = var_2159_mode_0, x = linear_12_cast_fp16)[name = string("op_2159_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14800640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17160000))))[name = string("t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_13_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_21_cast_fp16 = mul(x = var_2159_cast_fp16, y = linear_13_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17166208))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19525568))))[name = string("t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2167_cast_fp16 = abs(x = linear_14_cast_fp16)[name = string("op_2167_cast_fp16")];
            bool var_2170 = const()[name = string("op_2170"), val = bool(true)];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2171_cast_fp16 = reduce_max(axes = var_2171_axes_0, keep_dims = var_2170, x = var_2167_cast_fp16)[name = string("op_2171_cast_fp16")];
            fp16 var_2172_to_fp16 = const()[name = string("op_2172_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_31_cast_fp16 = maximum(x = var_2171_cast_fp16, y = var_2172_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_27_cast_fp16 = real_div(x = linear_14_cast_fp16, y = s_31_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2175_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2175_cast_fp16")];
            tensor<int32, [1]> var_2180_axes_0 = const()[name = string("op_2180_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2180_keep_dims_0 = const()[name = string("op_2180_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2180_cast_fp16 = reduce_mean(axes = var_2180_axes_0, keep_dims = var_2180_keep_dims_0, x = var_2175_cast_fp16)[name = string("op_2180_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2181_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2181_cast_fp16")];
            fp32 var_2182_epsilon_0 = const()[name = string("op_2182_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2182_cast_fp16 = inverse(epsilon = var_2182_epsilon_0, x = var_2181_cast_fp16)[name = string("op_2182_cast_fp16")];
            fp16 var_2183_to_fp16 = const()[name = string("op_2183_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2184_cast_fp16 = mul(x = var_2182_cast_fp16, y = var_2183_to_fp16)[name = string("op_2184_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2186_cast_fp16 = add(x = var_2180_cast_fp16, y = var_2184_cast_fp16)[name = string("op_2186_cast_fp16")];
            fp32 var_2187_epsilon_0 = const()[name = string("op_2187_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2187_cast_fp16 = rsqrt(epsilon = var_2187_epsilon_0, x = var_2186_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2187_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19527168)))];
            tensor<fp16, [1, 1260, 768]> var_2189_cast_fp16 = mul(x = out_23_cast_fp16, y = t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2189_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_37_cast_fp16 = add(x = x_33_cast_fp16, y = var_2189_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2192_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2192_cast_fp16")];
            bool var_2195 = const()[name = string("op_2195"), val = bool(true)];
            tensor<int32, [1]> var_2196_axes_0 = const()[name = string("op_2196_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2196_cast_fp16 = reduce_max(axes = var_2196_axes_0, keep_dims = var_2195, x = var_2192_cast_fp16)[name = string("op_2196_cast_fp16")];
            fp16 var_2197_to_fp16 = const()[name = string("op_2197_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_33_cast_fp16 = maximum(x = var_2196_cast_fp16, y = var_2197_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_29_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_33_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2200_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2200_cast_fp16")];
            tensor<int32, [1]> var_2205_axes_0 = const()[name = string("op_2205_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2205_keep_dims_0 = const()[name = string("op_2205_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2205_cast_fp16 = reduce_mean(axes = var_2205_axes_0, keep_dims = var_2205_keep_dims_0, x = var_2200_cast_fp16)[name = string("op_2205_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2206_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2206_cast_fp16")];
            fp32 var_2207_epsilon_0 = const()[name = string("op_2207_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2207_cast_fp16 = inverse(epsilon = var_2207_epsilon_0, x = var_2206_cast_fp16)[name = string("op_2207_cast_fp16")];
            fp16 var_2208_to_fp16 = const()[name = string("op_2208_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2209_cast_fp16 = mul(x = var_2207_cast_fp16, y = var_2208_to_fp16)[name = string("op_2209_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2211_cast_fp16 = add(x = var_2205_cast_fp16, y = var_2209_cast_fp16)[name = string("op_2211_cast_fp16")];
            fp32 var_2212_epsilon_0 = const()[name = string("op_2212_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2212_cast_fp16 = rsqrt(epsilon = var_2212_epsilon_0, x = var_2211_cast_fp16)[name = string("op_2212_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2212_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19528768)))];
            tensor<fp16, [1, 1260, 768]> input_23_cast_fp16 = mul(x = out_25_cast_fp16, y = t_encoder_layers_2_input_layernorm_weight_to_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19530368))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20120256))))[name = string("t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_15_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_2222 = const()[name = string("op_2222"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_39_cast_fp16 = reshape(shape = var_2222, x = linear_15_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2224_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2224_cast_fp16")];
            bool var_2227 = const()[name = string("op_2227"), val = bool(true)];
            tensor<int32, [1]> var_2228_axes_0 = const()[name = string("op_2228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2228_cast_fp16 = reduce_max(axes = var_2228_axes_0, keep_dims = var_2227, x = var_2224_cast_fp16)[name = string("op_2228_cast_fp16")];
            fp16 var_2229_to_fp16 = const()[name = string("op_2229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_35_cast_fp16 = maximum(x = var_2228_cast_fp16, y = var_2229_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_31_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_35_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2232_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2232_cast_fp16")];
            tensor<int32, [1]> var_2237_axes_0 = const()[name = string("op_2237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2237_keep_dims_0 = const()[name = string("op_2237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2237_cast_fp16 = reduce_mean(axes = var_2237_axes_0, keep_dims = var_2237_keep_dims_0, x = var_2232_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2238_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2238_cast_fp16")];
            fp32 var_2239_epsilon_0 = const()[name = string("op_2239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2239_cast_fp16 = inverse(epsilon = var_2239_epsilon_0, x = var_2238_cast_fp16)[name = string("op_2239_cast_fp16")];
            fp16 var_2240_to_fp16 = const()[name = string("op_2240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2241_cast_fp16 = mul(x = var_2239_cast_fp16, y = var_2240_to_fp16)[name = string("op_2241_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2243_cast_fp16 = add(x = var_2237_cast_fp16, y = var_2241_cast_fp16)[name = string("op_2243_cast_fp16")];
            fp32 var_2244_epsilon_0 = const()[name = string("op_2244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2244_cast_fp16 = rsqrt(epsilon = var_2244_epsilon_0, x = var_2243_cast_fp16)[name = string("op_2244_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2244_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20121856)))];
            tensor<fp16, [1, 1260, 12, 64]> x_45_cast_fp16 = mul(x = out_27_cast_fp16, y = t_encoder_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20122048))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20711936))))[name = string("t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_2254 = const()[name = string("op_2254"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_41_cast_fp16 = reshape(shape = var_2254, x = linear_16_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2256_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2256_cast_fp16")];
            bool var_2259 = const()[name = string("op_2259"), val = bool(true)];
            tensor<int32, [1]> var_2260_axes_0 = const()[name = string("op_2260_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2260_cast_fp16 = reduce_max(axes = var_2260_axes_0, keep_dims = var_2259, x = var_2256_cast_fp16)[name = string("op_2260_cast_fp16")];
            fp16 var_2261_to_fp16 = const()[name = string("op_2261_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_37_cast_fp16 = maximum(x = var_2260_cast_fp16, y = var_2261_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_33_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_37_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2264_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2264_cast_fp16")];
            tensor<int32, [1]> var_2269_axes_0 = const()[name = string("op_2269_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2269_keep_dims_0 = const()[name = string("op_2269_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2269_cast_fp16 = reduce_mean(axes = var_2269_axes_0, keep_dims = var_2269_keep_dims_0, x = var_2264_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2270_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2270_cast_fp16")];
            fp32 var_2271_epsilon_0 = const()[name = string("op_2271_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2271_cast_fp16 = inverse(epsilon = var_2271_epsilon_0, x = var_2270_cast_fp16)[name = string("op_2271_cast_fp16")];
            fp16 var_2272_to_fp16 = const()[name = string("op_2272_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2273_cast_fp16 = mul(x = var_2271_cast_fp16, y = var_2272_to_fp16)[name = string("op_2273_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2275_cast_fp16 = add(x = var_2269_cast_fp16, y = var_2273_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2276_cast_fp16 = rsqrt(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2276_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713536)))];
            tensor<fp16, [1, 1260, 12, 64]> x_47_cast_fp16 = mul(x = out_29_cast_fp16, y = t_encoder_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713728))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21303616))))[name = string("t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_17_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<int32, [4]> var_2286 = const()[name = string("op_2286"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_43_cast_fp16 = reshape(shape = var_2286, x = linear_17_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2288_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2288_cast_fp16")];
            bool var_2291 = const()[name = string("op_2291"), val = bool(true)];
            tensor<int32, [1]> var_2292_axes_0 = const()[name = string("op_2292_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2292_cast_fp16 = reduce_max(axes = var_2292_axes_0, keep_dims = var_2291, x = var_2288_cast_fp16)[name = string("op_2292_cast_fp16")];
            fp16 var_2293_to_fp16 = const()[name = string("op_2293_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_39_cast_fp16 = maximum(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_35_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_39_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2296_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2296_cast_fp16")];
            tensor<int32, [1]> var_2301_axes_0 = const()[name = string("op_2301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2301_keep_dims_0 = const()[name = string("op_2301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2301_cast_fp16 = reduce_mean(axes = var_2301_axes_0, keep_dims = var_2301_keep_dims_0, x = var_2296_cast_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2302_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2302_cast_fp16")];
            fp32 var_2303_epsilon_0 = const()[name = string("op_2303_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2303_cast_fp16 = inverse(epsilon = var_2303_epsilon_0, x = var_2302_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp16 var_2304_to_fp16 = const()[name = string("op_2304_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2305_cast_fp16 = mul(x = var_2303_cast_fp16, y = var_2304_to_fp16)[name = string("op_2305_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2307_cast_fp16 = add(x = var_2301_cast_fp16, y = var_2305_cast_fp16)[name = string("op_2307_cast_fp16")];
            fp32 var_2308_epsilon_0 = const()[name = string("op_2308_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2308_cast_fp16 = rsqrt(epsilon = var_2308_epsilon_0, x = var_2307_cast_fp16)[name = string("op_2308_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_9_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2308_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<int32, [4]> z_17_begin_0 = const()[name = string("z_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_17_end_0 = const()[name = string("z_17_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_17_end_mask_0 = const()[name = string("z_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_17_cast_fp16 = slice_by_index(begin = z_17_begin_0, end = z_17_end_0, end_mask = z_17_end_mask_0, x = x_45_cast_fp16)[name = string("z_17_cast_fp16")];
            tensor<int32, [4]> var_2319_begin_0 = const()[name = string("op_2319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2319_end_0 = const()[name = string("op_2319_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2319_end_mask_0 = const()[name = string("op_2319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2319_cast_fp16 = slice_by_index(begin = var_2319_begin_0, end = var_2319_end_0, end_mask = var_2319_end_mask_0, x = z_17_cast_fp16)[name = string("op_2319_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2320_cast_fp16 = mul(x = var_2319_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_2320_cast_fp16")];
            tensor<int32, [4]> var_2325_begin_0 = const()[name = string("op_2325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2325_end_0 = const()[name = string("op_2325_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2325_end_mask_0 = const()[name = string("op_2325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2325_cast_fp16 = slice_by_index(begin = var_2325_begin_0, end = var_2325_end_0, end_mask = var_2325_end_mask_0, x = z_17_cast_fp16)[name = string("op_2325_cast_fp16")];
            int32 var_2327 = const()[name = string("op_2327"), val = int32(-1)];
            bool rot_17_interleave_0 = const()[name = string("rot_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_17_cast_fp16 = concat(axis = var_2327, interleave = rot_17_interleave_0, values = (var_2320_cast_fp16, var_2325_cast_fp16))[name = string("rot_17_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2329_cast_fp16 = mul(x = z_17_cast_fp16, y = c_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2330_cast_fp16 = mul(x = rot_17_cast_fp16, y = s_9_cast_fp16)[name = string("op_2330_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2332_cast_fp16 = add(x = var_2329_cast_fp16, y = var_2330_cast_fp16)[name = string("op_2332_cast_fp16")];
            tensor<int32, [4]> z_19_begin_0 = const()[name = string("z_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_19_end_0 = const()[name = string("z_19_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_19_end_mask_0 = const()[name = string("z_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_19_cast_fp16 = slice_by_index(begin = z_19_begin_0, end = z_19_end_0, end_mask = z_19_end_mask_0, x = x_45_cast_fp16)[name = string("z_19_cast_fp16")];
            tensor<int32, [4]> var_2342_begin_0 = const()[name = string("op_2342_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2342_end_0 = const()[name = string("op_2342_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2342_end_mask_0 = const()[name = string("op_2342_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2342_cast_fp16 = slice_by_index(begin = var_2342_begin_0, end = var_2342_end_0, end_mask = var_2342_end_mask_0, x = z_19_cast_fp16)[name = string("op_2342_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2343_cast_fp16 = mul(x = var_2342_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2343_cast_fp16")];
            tensor<int32, [4]> var_2348_begin_0 = const()[name = string("op_2348_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2348_end_0 = const()[name = string("op_2348_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2348_end_mask_0 = const()[name = string("op_2348_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2348_cast_fp16 = slice_by_index(begin = var_2348_begin_0, end = var_2348_end_0, end_mask = var_2348_end_mask_0, x = z_19_cast_fp16)[name = string("op_2348_cast_fp16")];
            int32 var_2350 = const()[name = string("op_2350"), val = int32(-1)];
            bool rot_19_interleave_0 = const()[name = string("rot_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_19_cast_fp16 = concat(axis = var_2350, interleave = rot_19_interleave_0, values = (var_2343_cast_fp16, var_2348_cast_fp16))[name = string("rot_19_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2352_cast_fp16 = mul(x = z_19_cast_fp16, y = c_cast_fp16)[name = string("op_2352_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2353_cast_fp16 = mul(x = rot_19_cast_fp16, y = s_11_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2355_cast_fp16 = add(x = var_2352_cast_fp16, y = var_2353_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2357 = const()[name = string("op_2357"), val = int32(-1)];
            bool var_2358_interleave_0 = const()[name = string("op_2358_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_2358_cast_fp16 = concat(axis = var_2357, interleave = var_2358_interleave_0, values = (var_2332_cast_fp16, var_2355_cast_fp16))[name = string("op_2358_cast_fp16")];
            tensor<int32, [4]> z_21_begin_0 = const()[name = string("z_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_21_end_0 = const()[name = string("z_21_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_21_end_mask_0 = const()[name = string("z_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_21_cast_fp16 = slice_by_index(begin = z_21_begin_0, end = z_21_end_0, end_mask = z_21_end_mask_0, x = x_47_cast_fp16)[name = string("z_21_cast_fp16")];
            tensor<int32, [4]> var_2371_begin_0 = const()[name = string("op_2371_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2371_end_0 = const()[name = string("op_2371_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2371_end_mask_0 = const()[name = string("op_2371_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2371_cast_fp16 = slice_by_index(begin = var_2371_begin_0, end = var_2371_end_0, end_mask = var_2371_end_mask_0, x = z_21_cast_fp16)[name = string("op_2371_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2372_cast_fp16 = mul(x = var_2371_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2372_cast_fp16")];
            tensor<int32, [4]> var_2377_begin_0 = const()[name = string("op_2377_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2377_end_0 = const()[name = string("op_2377_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2377_end_mask_0 = const()[name = string("op_2377_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2377_cast_fp16 = slice_by_index(begin = var_2377_begin_0, end = var_2377_end_0, end_mask = var_2377_end_mask_0, x = z_21_cast_fp16)[name = string("op_2377_cast_fp16")];
            int32 var_2379 = const()[name = string("op_2379"), val = int32(-1)];
            bool rot_21_interleave_0 = const()[name = string("rot_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_21_cast_fp16 = concat(axis = var_2379, interleave = rot_21_interleave_0, values = (var_2372_cast_fp16, var_2377_cast_fp16))[name = string("rot_21_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2381_cast_fp16 = mul(x = z_21_cast_fp16, y = c_1_cast_fp16)[name = string("op_2381_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2382_cast_fp16 = mul(x = rot_21_cast_fp16, y = s_9_cast_fp16)[name = string("op_2382_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2384_cast_fp16 = add(x = var_2381_cast_fp16, y = var_2382_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<int32, [4]> z_23_begin_0 = const()[name = string("z_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_23_end_0 = const()[name = string("z_23_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_23_end_mask_0 = const()[name = string("z_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_23_cast_fp16 = slice_by_index(begin = z_23_begin_0, end = z_23_end_0, end_mask = z_23_end_mask_0, x = x_47_cast_fp16)[name = string("z_23_cast_fp16")];
            tensor<int32, [4]> var_2394_begin_0 = const()[name = string("op_2394_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2394_end_0 = const()[name = string("op_2394_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2394_end_mask_0 = const()[name = string("op_2394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> 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 = z_23_cast_fp16)[name = string("op_2394_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2395_cast_fp16 = mul(x = var_2394_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2395_cast_fp16")];
            tensor<int32, [4]> var_2400_begin_0 = const()[name = string("op_2400_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2400_end_0 = const()[name = string("op_2400_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2400_end_mask_0 = const()[name = string("op_2400_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2400_cast_fp16 = slice_by_index(begin = var_2400_begin_0, end = var_2400_end_0, end_mask = var_2400_end_mask_0, x = z_23_cast_fp16)[name = string("op_2400_cast_fp16")];
            int32 var_2402 = const()[name = string("op_2402"), val = int32(-1)];
            bool rot_23_interleave_0 = const()[name = string("rot_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_23_cast_fp16 = concat(axis = var_2402, interleave = rot_23_interleave_0, values = (var_2395_cast_fp16, var_2400_cast_fp16))[name = string("rot_23_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2404_cast_fp16 = mul(x = z_23_cast_fp16, y = c_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2405_cast_fp16 = mul(x = rot_23_cast_fp16, y = s_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2407_cast_fp16 = add(x = var_2404_cast_fp16, y = var_2405_cast_fp16)[name = string("op_2407_cast_fp16")];
            int32 var_2409 = const()[name = string("op_2409"), val = int32(-1)];
            bool var_2410_interleave_0 = const()[name = string("op_2410_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_2410_cast_fp16 = concat(axis = var_2409, interleave = var_2410_interleave_0, values = (var_2384_cast_fp16, var_2407_cast_fp16))[name = string("op_2410_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2420_transpose_x_0 = const()[name = string("op_2420_transpose_x_0"), val = bool(false)];
            bool var_2420_transpose_y_0 = const()[name = string("op_2420_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2410_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_68 = transpose(perm = transpose_68_perm_0, x = var_2358_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 12, 1260, 1260]> var_2420_cast_fp16 = matmul(transpose_x = var_2420_transpose_x_0, transpose_y = var_2420_transpose_y_0, x = transpose_68, y = transpose_69)[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_2422_cast_fp16 = add(x = var_2420_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2422_cast_fp16")];
            int32 var_2423 = const()[name = string("op_2423"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_35_cast_fp16 = softmax(axis = var_2423, x = var_2422_cast_fp16)[name = string("w_35_cast_fp16")];
            bool var_2426_transpose_x_0 = const()[name = string("op_2426_transpose_x_0"), val = bool(false)];
            bool var_2426_transpose_y_0 = const()[name = string("op_2426_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = v_9_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 12, 1260, 64]> var_2426_cast_fp16 = matmul(transpose_x = var_2426_transpose_x_0, transpose_y = var_2426_transpose_y_0, x = w_35_cast_fp16, y = v_11_cast_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [4]> var_2429_perm_0 = const()[name = string("op_2429_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2433 = const()[name = string("op_2433"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_2429_cast_fp16 = transpose(perm = var_2429_perm_0, x = var_2426_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 1260, 768]> input_25_cast_fp16 = reshape(shape = var_2433, x = var_2429_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21305216))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21895104))))[name = string("t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2438_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2438_cast_fp16")];
            bool var_2441 = const()[name = string("op_2441"), val = bool(true)];
            tensor<int32, [1]> var_2442_axes_0 = const()[name = string("op_2442_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2442_cast_fp16 = reduce_max(axes = var_2442_axes_0, keep_dims = var_2441, x = var_2438_cast_fp16)[name = string("op_2442_cast_fp16")];
            fp16 var_2443_to_fp16 = const()[name = string("op_2443_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_41_cast_fp16 = maximum(x = var_2442_cast_fp16, y = var_2443_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_37_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_41_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2446_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2446_cast_fp16")];
            tensor<int32, [1]> var_2451_axes_0 = const()[name = string("op_2451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2451_keep_dims_0 = const()[name = string("op_2451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2451_cast_fp16 = reduce_mean(axes = var_2451_axes_0, keep_dims = var_2451_keep_dims_0, x = var_2446_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2452_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2452_cast_fp16")];
            fp32 var_2453_epsilon_0 = const()[name = string("op_2453_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2453_cast_fp16 = inverse(epsilon = var_2453_epsilon_0, x = var_2452_cast_fp16)[name = string("op_2453_cast_fp16")];
            fp16 var_2454_to_fp16 = const()[name = string("op_2454_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = string("op_2455_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2457_cast_fp16 = add(x = var_2451_cast_fp16, y = var_2455_cast_fp16)[name = string("op_2457_cast_fp16")];
            fp32 var_2458_epsilon_0 = const()[name = string("op_2458_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2458_cast_fp16 = rsqrt(epsilon = var_2458_epsilon_0, x = var_2457_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_31_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2458_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21896704)))];
            tensor<fp16, [1, 1260, 768]> var_2460_cast_fp16 = mul(x = out_31_cast_fp16, y = t_encoder_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_2460_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_51_cast_fp16 = add(x = x_37_cast_fp16, y = var_2460_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2463_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2463_cast_fp16")];
            bool var_2466 = const()[name = string("op_2466"), val = bool(true)];
            tensor<int32, [1]> var_2467_axes_0 = const()[name = string("op_2467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2467_cast_fp16 = reduce_max(axes = var_2467_axes_0, keep_dims = var_2466, x = var_2463_cast_fp16)[name = string("op_2467_cast_fp16")];
            fp16 var_2468_to_fp16 = const()[name = string("op_2468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_43_cast_fp16 = maximum(x = var_2467_cast_fp16, y = var_2468_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_39_cast_fp16 = real_div(x = x_51_cast_fp16, y = s_43_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2471_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2471_cast_fp16")];
            tensor<int32, [1]> var_2476_axes_0 = const()[name = string("op_2476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2476_keep_dims_0 = const()[name = string("op_2476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2476_cast_fp16 = reduce_mean(axes = var_2476_axes_0, keep_dims = var_2476_keep_dims_0, x = var_2471_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2477_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_2477_cast_fp16")];
            fp32 var_2478_epsilon_0 = const()[name = string("op_2478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2478_cast_fp16 = inverse(epsilon = var_2478_epsilon_0, x = var_2477_cast_fp16)[name = string("op_2478_cast_fp16")];
            fp16 var_2479_to_fp16 = const()[name = string("op_2479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2480_cast_fp16 = mul(x = var_2478_cast_fp16, y = var_2479_to_fp16)[name = string("op_2480_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2482_cast_fp16 = add(x = var_2476_cast_fp16, y = var_2480_cast_fp16)[name = string("op_2482_cast_fp16")];
            fp32 var_2483_epsilon_0 = const()[name = string("op_2483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2483_cast_fp16 = rsqrt(epsilon = var_2483_epsilon_0, x = var_2482_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_33_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2483_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21898304)))];
            tensor<fp16, [1, 1260, 768]> input_27_cast_fp16 = mul(x = out_33_cast_fp16, y = t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21899904))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24259264))))[name = string("t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_19_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_19_cast_fp16")];
            string var_2490_mode_0 = const()[name = string("op_2490_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_2490_cast_fp16 = gelu(mode = var_2490_mode_0, x = linear_19_cast_fp16)[name = string("op_2490_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24265472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26624832))))[name = string("t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_20_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_31_cast_fp16 = mul(x = var_2490_cast_fp16, y = linear_20_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26631040))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28990400))))[name = string("t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2498_cast_fp16 = abs(x = linear_21_cast_fp16)[name = string("op_2498_cast_fp16")];
            bool var_2501 = const()[name = string("op_2501"), val = bool(true)];
            tensor<int32, [1]> var_2502_axes_0 = const()[name = string("op_2502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2502_cast_fp16 = reduce_max(axes = var_2502_axes_0, keep_dims = var_2501, x = var_2498_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp16 var_2503_to_fp16 = const()[name = string("op_2503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_45_cast_fp16 = maximum(x = var_2502_cast_fp16, y = var_2503_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_41_cast_fp16 = real_div(x = linear_21_cast_fp16, y = s_45_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2506_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2506_cast_fp16")];
            tensor<int32, [1]> var_2511_axes_0 = const()[name = string("op_2511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2511_keep_dims_0 = const()[name = string("op_2511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2511_cast_fp16 = reduce_mean(axes = var_2511_axes_0, keep_dims = var_2511_keep_dims_0, x = var_2506_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2512_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_2512_cast_fp16")];
            fp32 var_2513_epsilon_0 = const()[name = string("op_2513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2513_cast_fp16 = inverse(epsilon = var_2513_epsilon_0, x = var_2512_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2515_cast_fp16 = mul(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("op_2515_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2517_cast_fp16 = add(x = var_2511_cast_fp16, y = var_2515_cast_fp16)[name = string("op_2517_cast_fp16")];
            fp32 var_2518_epsilon_0 = const()[name = string("op_2518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2518_cast_fp16 = rsqrt(epsilon = var_2518_epsilon_0, x = var_2517_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_35_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2518_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28992000)))];
            tensor<fp16, [1, 1260, 768]> var_2520_cast_fp16 = mul(x = out_35_cast_fp16, y = t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_55_cast_fp16 = add(x = x_51_cast_fp16, y = var_2520_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2523_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2523_cast_fp16")];
            bool var_2526 = const()[name = string("op_2526"), val = bool(true)];
            tensor<int32, [1]> var_2527_axes_0 = const()[name = string("op_2527_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2527_cast_fp16 = reduce_max(axes = var_2527_axes_0, keep_dims = var_2526, x = var_2523_cast_fp16)[name = string("op_2527_cast_fp16")];
            fp16 var_2528_to_fp16 = const()[name = string("op_2528_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_47_cast_fp16 = maximum(x = var_2527_cast_fp16, y = var_2528_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_43_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_47_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2531_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [1]> var_2536_axes_0 = const()[name = string("op_2536_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2536_keep_dims_0 = const()[name = string("op_2536_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2536_cast_fp16 = reduce_mean(axes = var_2536_axes_0, keep_dims = var_2536_keep_dims_0, x = var_2531_cast_fp16)[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2537_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_2537_cast_fp16")];
            fp32 var_2538_epsilon_0 = const()[name = string("op_2538_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2538_cast_fp16 = inverse(epsilon = var_2538_epsilon_0, x = var_2537_cast_fp16)[name = string("op_2538_cast_fp16")];
            fp16 var_2539_to_fp16 = const()[name = string("op_2539_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2540_cast_fp16 = mul(x = var_2538_cast_fp16, y = var_2539_to_fp16)[name = string("op_2540_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2542_cast_fp16 = add(x = var_2536_cast_fp16, y = var_2540_cast_fp16)[name = string("op_2542_cast_fp16")];
            fp32 var_2543_epsilon_0 = const()[name = string("op_2543_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2543_cast_fp16 = rsqrt(epsilon = var_2543_epsilon_0, x = var_2542_cast_fp16)[name = string("op_2543_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_2543_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28993600)))];
            tensor<fp16, [1, 1260, 768]> input_33_cast_fp16 = mul(x = out_37_cast_fp16, y = t_encoder_layers_3_input_layernorm_weight_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28995200))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29585088))))[name = string("t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_2553 = const()[name = string("op_2553"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_57_cast_fp16 = reshape(shape = var_2553, x = linear_22_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2555_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2555_cast_fp16")];
            bool var_2558 = const()[name = string("op_2558"), val = bool(true)];
            tensor<int32, [1]> var_2559_axes_0 = const()[name = string("op_2559_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2559_cast_fp16 = reduce_max(axes = var_2559_axes_0, keep_dims = var_2558, x = var_2555_cast_fp16)[name = string("op_2559_cast_fp16")];
            fp16 var_2560_to_fp16 = const()[name = string("op_2560_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_49_cast_fp16 = maximum(x = var_2559_cast_fp16, y = var_2560_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_45_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_49_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2563_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_2563_cast_fp16")];
            tensor<int32, [1]> var_2568_axes_0 = const()[name = string("op_2568_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2568_keep_dims_0 = const()[name = string("op_2568_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2568_cast_fp16 = reduce_mean(axes = var_2568_axes_0, keep_dims = var_2568_keep_dims_0, x = var_2563_cast_fp16)[name = string("op_2568_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2569_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_2569_cast_fp16")];
            fp32 var_2570_epsilon_0 = const()[name = string("op_2570_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2570_cast_fp16 = inverse(epsilon = var_2570_epsilon_0, x = var_2569_cast_fp16)[name = string("op_2570_cast_fp16")];
            fp16 var_2571_to_fp16 = const()[name = string("op_2571_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2572_cast_fp16 = mul(x = var_2570_cast_fp16, y = var_2571_to_fp16)[name = string("op_2572_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2574_cast_fp16 = add(x = var_2568_cast_fp16, y = var_2572_cast_fp16)[name = string("op_2574_cast_fp16")];
            fp32 var_2575_epsilon_0 = const()[name = string("op_2575_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2575_cast_fp16 = rsqrt(epsilon = var_2575_epsilon_0, x = var_2574_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_2575_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29586688)))];
            tensor<fp16, [1, 1260, 12, 64]> x_63_cast_fp16 = mul(x = out_39_cast_fp16, y = t_encoder_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29586880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30176768))))[name = string("t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_23_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_2585 = const()[name = string("op_2585"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_59_cast_fp16 = reshape(shape = var_2585, x = linear_23_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2587_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2587_cast_fp16")];
            bool var_2590 = const()[name = string("op_2590"), val = bool(true)];
            tensor<int32, [1]> var_2591_axes_0 = const()[name = string("op_2591_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2591_cast_fp16 = reduce_max(axes = var_2591_axes_0, keep_dims = var_2590, x = var_2587_cast_fp16)[name = string("op_2591_cast_fp16")];
            fp16 var_2592_to_fp16 = const()[name = string("op_2592_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_51_cast_fp16 = maximum(x = var_2591_cast_fp16, y = var_2592_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_47_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_51_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2595_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_2595_cast_fp16")];
            tensor<int32, [1]> var_2600_axes_0 = const()[name = string("op_2600_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2600_keep_dims_0 = const()[name = string("op_2600_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2600_cast_fp16 = reduce_mean(axes = var_2600_axes_0, keep_dims = var_2600_keep_dims_0, x = var_2595_cast_fp16)[name = string("op_2600_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2601_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_2601_cast_fp16")];
            fp32 var_2602_epsilon_0 = const()[name = string("op_2602_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2602_cast_fp16 = inverse(epsilon = var_2602_epsilon_0, x = var_2601_cast_fp16)[name = string("op_2602_cast_fp16")];
            fp16 var_2603_to_fp16 = const()[name = string("op_2603_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2604_cast_fp16 = mul(x = var_2602_cast_fp16, y = var_2603_to_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2606_cast_fp16 = add(x = var_2600_cast_fp16, y = var_2604_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp32 var_2607_epsilon_0 = const()[name = string("op_2607_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2607_cast_fp16 = rsqrt(epsilon = var_2607_epsilon_0, x = var_2606_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_2607_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178368)))];
            tensor<fp16, [1, 1260, 12, 64]> x_65_cast_fp16 = mul(x = out_41_cast_fp16, y = t_encoder_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178560))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30768448))))[name = string("t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [4]> var_2617 = const()[name = string("op_2617"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_61_cast_fp16 = reshape(shape = var_2617, x = linear_24_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2619_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2619_cast_fp16")];
            bool var_2622 = const()[name = string("op_2622"), val = bool(true)];
            tensor<int32, [1]> var_2623_axes_0 = const()[name = string("op_2623_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2623_cast_fp16 = reduce_max(axes = var_2623_axes_0, keep_dims = var_2622, x = var_2619_cast_fp16)[name = string("op_2623_cast_fp16")];
            fp16 var_2624_to_fp16 = const()[name = string("op_2624_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_53_cast_fp16 = maximum(x = var_2623_cast_fp16, y = var_2624_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_49_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_53_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2627_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<int32, [1]> var_2632_axes_0 = const()[name = string("op_2632_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2632_keep_dims_0 = const()[name = string("op_2632_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2632_cast_fp16 = reduce_mean(axes = var_2632_axes_0, keep_dims = var_2632_keep_dims_0, x = var_2627_cast_fp16)[name = string("op_2632_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2633_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2634_cast_fp16 = inverse(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            fp16 var_2635_to_fp16 = const()[name = string("op_2635_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2636_cast_fp16 = mul(x = var_2634_cast_fp16, y = var_2635_to_fp16)[name = string("op_2636_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2638_cast_fp16 = add(x = var_2632_cast_fp16, y = var_2636_cast_fp16)[name = string("op_2638_cast_fp16")];
            fp32 var_2639_epsilon_0 = const()[name = string("op_2639_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2639_cast_fp16 = rsqrt(epsilon = var_2639_epsilon_0, x = var_2638_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_13_cast_fp16 = mul(x = y_49_cast_fp16, y = var_2639_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<int32, [4]> z_25_begin_0 = const()[name = string("z_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_25_end_0 = const()[name = string("z_25_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_25_end_mask_0 = const()[name = string("z_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_25_cast_fp16 = slice_by_index(begin = z_25_begin_0, end = z_25_end_0, end_mask = z_25_end_mask_0, x = x_63_cast_fp16)[name = string("z_25_cast_fp16")];
            tensor<int32, [4]> var_2650_begin_0 = const()[name = string("op_2650_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2650_end_0 = const()[name = string("op_2650_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2650_end_mask_0 = const()[name = string("op_2650_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2650_cast_fp16 = slice_by_index(begin = var_2650_begin_0, end = var_2650_end_0, end_mask = var_2650_end_mask_0, x = z_25_cast_fp16)[name = string("op_2650_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2651_cast_fp16")];
            tensor<int32, [4]> var_2656_begin_0 = const()[name = string("op_2656_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2656_end_0 = const()[name = string("op_2656_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2656_end_mask_0 = const()[name = string("op_2656_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2656_cast_fp16 = slice_by_index(begin = var_2656_begin_0, end = var_2656_end_0, end_mask = var_2656_end_mask_0, x = z_25_cast_fp16)[name = string("op_2656_cast_fp16")];
            int32 var_2658 = const()[name = string("op_2658"), val = int32(-1)];
            bool rot_25_interleave_0 = const()[name = string("rot_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_25_cast_fp16 = concat(axis = var_2658, interleave = rot_25_interleave_0, values = (var_2651_cast_fp16, var_2656_cast_fp16))[name = string("rot_25_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2660_cast_fp16 = mul(x = z_25_cast_fp16, y = c_1_cast_fp16)[name = string("op_2660_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2661_cast_fp16 = mul(x = rot_25_cast_fp16, y = s_9_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2663_cast_fp16 = add(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = string("op_2663_cast_fp16")];
            tensor<int32, [4]> z_27_begin_0 = const()[name = string("z_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_27_end_0 = const()[name = string("z_27_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_27_end_mask_0 = const()[name = string("z_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_27_cast_fp16 = slice_by_index(begin = z_27_begin_0, end = z_27_end_0, end_mask = z_27_end_mask_0, x = x_63_cast_fp16)[name = string("z_27_cast_fp16")];
            tensor<int32, [4]> var_2673_begin_0 = const()[name = string("op_2673_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2673_end_0 = const()[name = string("op_2673_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2673_end_mask_0 = const()[name = string("op_2673_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2673_cast_fp16 = slice_by_index(begin = var_2673_begin_0, end = var_2673_end_0, end_mask = var_2673_end_mask_0, x = z_27_cast_fp16)[name = string("op_2673_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2674_cast_fp16 = mul(x = var_2673_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2674_cast_fp16")];
            tensor<int32, [4]> var_2679_begin_0 = const()[name = string("op_2679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2679_end_0 = const()[name = string("op_2679_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2679_end_mask_0 = const()[name = string("op_2679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2679_cast_fp16 = slice_by_index(begin = var_2679_begin_0, end = var_2679_end_0, end_mask = var_2679_end_mask_0, x = z_27_cast_fp16)[name = string("op_2679_cast_fp16")];
            int32 var_2681 = const()[name = string("op_2681"), val = int32(-1)];
            bool rot_27_interleave_0 = const()[name = string("rot_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_27_cast_fp16 = concat(axis = var_2681, interleave = rot_27_interleave_0, values = (var_2674_cast_fp16, var_2679_cast_fp16))[name = string("rot_27_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2683_cast_fp16 = mul(x = z_27_cast_fp16, y = c_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2684_cast_fp16 = mul(x = rot_27_cast_fp16, y = s_11_cast_fp16)[name = string("op_2684_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2686_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2684_cast_fp16)[name = string("op_2686_cast_fp16")];
            int32 var_2688 = const()[name = string("op_2688"), val = int32(-1)];
            bool var_2689_interleave_0 = const()[name = string("op_2689_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_2689_cast_fp16 = concat(axis = var_2688, interleave = var_2689_interleave_0, values = (var_2663_cast_fp16, var_2686_cast_fp16))[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> z_29_begin_0 = const()[name = string("z_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_29_end_0 = const()[name = string("z_29_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_29_end_mask_0 = const()[name = string("z_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_29_cast_fp16 = slice_by_index(begin = z_29_begin_0, end = z_29_end_0, end_mask = z_29_end_mask_0, x = x_65_cast_fp16)[name = string("z_29_cast_fp16")];
            tensor<int32, [4]> var_2702_begin_0 = const()[name = string("op_2702_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2702_end_0 = const()[name = string("op_2702_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2702_end_mask_0 = const()[name = string("op_2702_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2702_cast_fp16 = slice_by_index(begin = var_2702_begin_0, end = var_2702_end_0, end_mask = var_2702_end_mask_0, x = z_29_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2703_cast_fp16 = mul(x = var_2702_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_2703_cast_fp16")];
            tensor<int32, [4]> var_2708_begin_0 = const()[name = string("op_2708_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2708_end_0 = const()[name = string("op_2708_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2708_end_mask_0 = const()[name = string("op_2708_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2708_cast_fp16 = slice_by_index(begin = var_2708_begin_0, end = var_2708_end_0, end_mask = var_2708_end_mask_0, x = z_29_cast_fp16)[name = string("op_2708_cast_fp16")];
            int32 var_2710 = const()[name = string("op_2710"), val = int32(-1)];
            bool rot_29_interleave_0 = const()[name = string("rot_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_29_cast_fp16 = concat(axis = var_2710, interleave = rot_29_interleave_0, values = (var_2703_cast_fp16, var_2708_cast_fp16))[name = string("rot_29_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2712_cast_fp16 = mul(x = z_29_cast_fp16, y = c_1_cast_fp16)[name = string("op_2712_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2713_cast_fp16 = mul(x = rot_29_cast_fp16, y = s_9_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2715_cast_fp16 = add(x = var_2712_cast_fp16, y = var_2713_cast_fp16)[name = string("op_2715_cast_fp16")];
            tensor<int32, [4]> z_31_begin_0 = const()[name = string("z_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_31_end_0 = const()[name = string("z_31_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_31_end_mask_0 = const()[name = string("z_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_31_cast_fp16 = slice_by_index(begin = z_31_begin_0, end = z_31_end_0, end_mask = z_31_end_mask_0, x = x_65_cast_fp16)[name = string("z_31_cast_fp16")];
            tensor<int32, [4]> var_2725_begin_0 = const()[name = string("op_2725_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2725_end_0 = const()[name = string("op_2725_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2725_end_mask_0 = const()[name = string("op_2725_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2725_cast_fp16 = slice_by_index(begin = var_2725_begin_0, end = var_2725_end_0, end_mask = var_2725_end_mask_0, x = z_31_cast_fp16)[name = string("op_2725_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2726_cast_fp16 = mul(x = var_2725_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> var_2731_begin_0 = const()[name = string("op_2731_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2731_end_0 = const()[name = string("op_2731_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2731_end_mask_0 = const()[name = string("op_2731_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2731_cast_fp16 = slice_by_index(begin = var_2731_begin_0, end = var_2731_end_0, end_mask = var_2731_end_mask_0, x = z_31_cast_fp16)[name = string("op_2731_cast_fp16")];
            int32 var_2733 = const()[name = string("op_2733"), val = int32(-1)];
            bool rot_31_interleave_0 = const()[name = string("rot_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_31_cast_fp16 = concat(axis = var_2733, interleave = rot_31_interleave_0, values = (var_2726_cast_fp16, var_2731_cast_fp16))[name = string("rot_31_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2735_cast_fp16 = mul(x = z_31_cast_fp16, y = c_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2736_cast_fp16 = mul(x = rot_31_cast_fp16, y = s_11_cast_fp16)[name = string("op_2736_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2738_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2736_cast_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-1)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2715_cast_fp16, var_2738_cast_fp16))[name = string("op_2741_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2751_transpose_x_0 = const()[name = string("op_2751_transpose_x_0"), val = bool(false)];
            bool var_2751_transpose_y_0 = const()[name = string("op_2751_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2741_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_70 = transpose(perm = transpose_70_perm_0, x = var_2689_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 12, 1260, 1260]> var_2751_cast_fp16 = matmul(transpose_x = var_2751_transpose_x_0, transpose_y = var_2751_transpose_y_0, x = transpose_70, y = transpose_71)[name = string("op_2751_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_2753_cast_fp16 = add(x = var_2751_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2753_cast_fp16")];
            int32 var_2754 = const()[name = string("op_2754"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_49_cast_fp16 = softmax(axis = var_2754, x = var_2753_cast_fp16)[name = string("w_49_cast_fp16")];
            bool var_2757_transpose_x_0 = const()[name = string("op_2757_transpose_x_0"), val = bool(false)];
            bool var_2757_transpose_y_0 = const()[name = string("op_2757_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = v_13_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 12, 1260, 64]> var_2757_cast_fp16 = matmul(transpose_x = var_2757_transpose_x_0, transpose_y = var_2757_transpose_y_0, x = w_49_cast_fp16, y = v_15_cast_fp16)[name = string("op_2757_cast_fp16")];
            tensor<int32, [4]> var_2760_perm_0 = const()[name = string("op_2760_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2764 = const()[name = string("op_2764"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_2760_cast_fp16 = transpose(perm = var_2760_perm_0, x = var_2757_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 1260, 768]> input_35_cast_fp16 = reshape(shape = var_2764, x = var_2760_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30770048))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31359936))))[name = string("t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_25_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2769_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2772 = const()[name = string("op_2772"), val = bool(true)];
            tensor<int32, [1]> var_2773_axes_0 = const()[name = string("op_2773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2773_cast_fp16 = reduce_max(axes = var_2773_axes_0, keep_dims = var_2772, x = var_2769_cast_fp16)[name = string("op_2773_cast_fp16")];
            fp16 var_2774_to_fp16 = const()[name = string("op_2774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_55_cast_fp16 = maximum(x = var_2773_cast_fp16, y = var_2774_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_51_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_55_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2777_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_2777_cast_fp16")];
            tensor<int32, [1]> var_2782_axes_0 = const()[name = string("op_2782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2782_keep_dims_0 = const()[name = string("op_2782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2782_cast_fp16 = reduce_mean(axes = var_2782_axes_0, keep_dims = var_2782_keep_dims_0, x = var_2777_cast_fp16)[name = string("op_2782_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2783_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_2783_cast_fp16")];
            fp32 var_2784_epsilon_0 = const()[name = string("op_2784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2784_cast_fp16 = inverse(epsilon = var_2784_epsilon_0, x = var_2783_cast_fp16)[name = string("op_2784_cast_fp16")];
            fp16 var_2785_to_fp16 = const()[name = string("op_2785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2786_cast_fp16 = mul(x = var_2784_cast_fp16, y = var_2785_to_fp16)[name = string("op_2786_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2788_cast_fp16 = add(x = var_2782_cast_fp16, y = var_2786_cast_fp16)[name = string("op_2788_cast_fp16")];
            fp32 var_2789_epsilon_0 = const()[name = string("op_2789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2789_cast_fp16 = rsqrt(epsilon = var_2789_epsilon_0, x = var_2788_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_43_cast_fp16 = mul(x = y_51_cast_fp16, y = var_2789_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31361536)))];
            tensor<fp16, [1, 1260, 768]> var_2791_cast_fp16 = mul(x = out_43_cast_fp16, y = t_encoder_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_2791_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_69_cast_fp16 = add(x = x_55_cast_fp16, y = var_2791_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2794_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2794_cast_fp16")];
            bool var_2797 = const()[name = string("op_2797"), val = bool(true)];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2798_cast_fp16 = reduce_max(axes = var_2798_axes_0, keep_dims = var_2797, x = var_2794_cast_fp16)[name = string("op_2798_cast_fp16")];
            fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_57_cast_fp16 = maximum(x = var_2798_cast_fp16, y = var_2799_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_53_cast_fp16 = real_div(x = x_69_cast_fp16, y = s_57_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2802_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_2802_cast_fp16")];
            tensor<int32, [1]> var_2807_axes_0 = const()[name = string("op_2807_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2807_keep_dims_0 = const()[name = string("op_2807_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2807_cast_fp16 = reduce_mean(axes = var_2807_axes_0, keep_dims = var_2807_keep_dims_0, x = var_2802_cast_fp16)[name = string("op_2807_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2808_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_2808_cast_fp16")];
            fp32 var_2809_epsilon_0 = const()[name = string("op_2809_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2809_cast_fp16 = inverse(epsilon = var_2809_epsilon_0, x = var_2808_cast_fp16)[name = string("op_2809_cast_fp16")];
            fp16 var_2810_to_fp16 = const()[name = string("op_2810_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2811_cast_fp16 = mul(x = var_2809_cast_fp16, y = var_2810_to_fp16)[name = string("op_2811_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2813_cast_fp16 = add(x = var_2807_cast_fp16, y = var_2811_cast_fp16)[name = string("op_2813_cast_fp16")];
            fp32 var_2814_epsilon_0 = const()[name = string("op_2814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2814_cast_fp16 = rsqrt(epsilon = var_2814_epsilon_0, x = var_2813_cast_fp16)[name = string("op_2814_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_45_cast_fp16 = mul(x = y_53_cast_fp16, y = var_2814_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31363136)))];
            tensor<fp16, [1, 1260, 768]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31364736))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33724096))))[name = string("t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_26_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_26_cast_fp16")];
            string var_2821_mode_0 = const()[name = string("op_2821_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_2821_cast_fp16 = gelu(mode = var_2821_mode_0, x = linear_26_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33730304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36089664))))[name = string("t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_27_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_41_cast_fp16 = mul(x = var_2821_cast_fp16, y = linear_27_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36095872))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38455232))))[name = string("t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2829_cast_fp16 = abs(x = linear_28_cast_fp16)[name = string("op_2829_cast_fp16")];
            bool var_2832 = const()[name = string("op_2832"), val = bool(true)];
            tensor<int32, [1]> var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2833_cast_fp16 = reduce_max(axes = var_2833_axes_0, keep_dims = var_2832, x = var_2829_cast_fp16)[name = string("op_2833_cast_fp16")];
            fp16 var_2834_to_fp16 = const()[name = string("op_2834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_59_cast_fp16 = maximum(x = var_2833_cast_fp16, y = var_2834_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_55_cast_fp16 = real_div(x = linear_28_cast_fp16, y = s_59_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2837_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_2837_cast_fp16")];
            tensor<int32, [1]> var_2842_axes_0 = const()[name = string("op_2842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2842_keep_dims_0 = const()[name = string("op_2842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2842_cast_fp16 = reduce_mean(axes = var_2842_axes_0, keep_dims = var_2842_keep_dims_0, x = var_2837_cast_fp16)[name = string("op_2842_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2843_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp32 var_2844_epsilon_0 = const()[name = string("op_2844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2844_cast_fp16 = inverse(epsilon = var_2844_epsilon_0, x = var_2843_cast_fp16)[name = string("op_2844_cast_fp16")];
            fp16 var_2845_to_fp16 = const()[name = string("op_2845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2846_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2848_cast_fp16 = add(x = var_2842_cast_fp16, y = var_2846_cast_fp16)[name = string("op_2848_cast_fp16")];
            fp32 var_2849_epsilon_0 = const()[name = string("op_2849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2849_cast_fp16 = rsqrt(epsilon = var_2849_epsilon_0, x = var_2848_cast_fp16)[name = string("op_2849_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_47_cast_fp16 = mul(x = y_55_cast_fp16, y = var_2849_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38456832)))];
            tensor<fp16, [1, 1260, 768]> var_2851_cast_fp16 = mul(x = out_47_cast_fp16, y = t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2851_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_73_cast_fp16 = add(x = x_69_cast_fp16, y = var_2851_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2854_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2854_cast_fp16")];
            bool var_2857 = const()[name = string("op_2857"), val = bool(true)];
            tensor<int32, [1]> var_2858_axes_0 = const()[name = string("op_2858_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_2858_cast_fp16 = reduce_max(axes = var_2858_axes_0, keep_dims = var_2857, x = var_2854_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp16 var_2859_to_fp16 = const()[name = string("op_2859_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_61_cast_fp16 = maximum(x = var_2858_cast_fp16, y = var_2859_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_57_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_61_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_2862_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_2862_cast_fp16")];
            tensor<int32, [1]> var_2867_axes_0 = const()[name = string("op_2867_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2867_keep_dims_0 = const()[name = string("op_2867_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_2867_cast_fp16 = reduce_mean(axes = var_2867_axes_0, keep_dims = var_2867_keep_dims_0, x = var_2862_cast_fp16)[name = string("op_2867_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2868_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_2868_cast_fp16")];
            fp32 var_2869_epsilon_0 = const()[name = string("op_2869_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_2869_cast_fp16 = inverse(epsilon = var_2869_epsilon_0, x = var_2868_cast_fp16)[name = string("op_2869_cast_fp16")];
            fp16 var_2870_to_fp16 = const()[name = string("op_2870_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_2871_cast_fp16 = mul(x = var_2869_cast_fp16, y = var_2870_to_fp16)[name = string("op_2871_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_2873_cast_fp16 = add(x = var_2867_cast_fp16, y = var_2871_cast_fp16)[name = string("op_2873_cast_fp16")];
            fp32 var_2874_epsilon_0 = const()[name = string("op_2874_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_2874_cast_fp16 = rsqrt(epsilon = var_2874_epsilon_0, x = var_2873_cast_fp16)[name = string("op_2874_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_49_cast_fp16 = mul(x = y_57_cast_fp16, y = var_2874_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38458432)))];
            tensor<fp16, [1, 1260, 768]> input_43_cast_fp16 = mul(x = out_49_cast_fp16, y = t_encoder_layers_4_input_layernorm_weight_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38460032))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39049920))))[name = string("t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_29_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2884 = const()[name = string("op_2884"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_75_cast_fp16 = reshape(shape = var_2884, x = linear_29_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2886_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_2886_cast_fp16")];
            bool var_2889 = const()[name = string("op_2889"), val = bool(true)];
            tensor<int32, [1]> var_2890_axes_0 = const()[name = string("op_2890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2890_cast_fp16 = reduce_max(axes = var_2890_axes_0, keep_dims = var_2889, x = var_2886_cast_fp16)[name = string("op_2890_cast_fp16")];
            fp16 var_2891_to_fp16 = const()[name = string("op_2891_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_63_cast_fp16 = maximum(x = var_2890_cast_fp16, y = var_2891_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_59_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_63_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2894_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [1]> var_2899_axes_0 = const()[name = string("op_2899_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2899_keep_dims_0 = const()[name = string("op_2899_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2899_cast_fp16 = reduce_mean(axes = var_2899_axes_0, keep_dims = var_2899_keep_dims_0, x = var_2894_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2900_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_2900_cast_fp16")];
            fp32 var_2901_epsilon_0 = const()[name = string("op_2901_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2901_cast_fp16 = inverse(epsilon = var_2901_epsilon_0, x = var_2900_cast_fp16)[name = string("op_2901_cast_fp16")];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2903_cast_fp16 = mul(x = var_2901_cast_fp16, y = var_2902_to_fp16)[name = string("op_2903_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2905_cast_fp16 = add(x = var_2899_cast_fp16, y = var_2903_cast_fp16)[name = string("op_2905_cast_fp16")];
            fp32 var_2906_epsilon_0 = const()[name = string("op_2906_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2906_cast_fp16 = rsqrt(epsilon = var_2906_epsilon_0, x = var_2905_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_2906_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39051520)))];
            tensor<fp16, [1, 1260, 12, 64]> x_81_cast_fp16 = mul(x = out_51_cast_fp16, y = t_encoder_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39051712))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39641600))))[name = string("t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2916 = const()[name = string("op_2916"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_77_cast_fp16 = reshape(shape = var_2916, x = linear_30_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2918_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_2918_cast_fp16")];
            bool var_2921 = const()[name = string("op_2921"), val = bool(true)];
            tensor<int32, [1]> var_2922_axes_0 = const()[name = string("op_2922_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2922_cast_fp16 = reduce_max(axes = var_2922_axes_0, keep_dims = var_2921, x = var_2918_cast_fp16)[name = string("op_2922_cast_fp16")];
            fp16 var_2923_to_fp16 = const()[name = string("op_2923_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_65_cast_fp16 = maximum(x = var_2922_cast_fp16, y = var_2923_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_61_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_65_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2926_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_2926_cast_fp16")];
            tensor<int32, [1]> var_2931_axes_0 = const()[name = string("op_2931_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2931_keep_dims_0 = const()[name = string("op_2931_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2931_cast_fp16 = reduce_mean(axes = var_2931_axes_0, keep_dims = var_2931_keep_dims_0, x = var_2926_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2932_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_2932_cast_fp16")];
            fp32 var_2933_epsilon_0 = const()[name = string("op_2933_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2933_cast_fp16 = inverse(epsilon = var_2933_epsilon_0, x = var_2932_cast_fp16)[name = string("op_2933_cast_fp16")];
            fp16 var_2934_to_fp16 = const()[name = string("op_2934_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2935_cast_fp16 = mul(x = var_2933_cast_fp16, y = var_2934_to_fp16)[name = string("op_2935_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2937_cast_fp16 = add(x = var_2931_cast_fp16, y = var_2935_cast_fp16)[name = string("op_2937_cast_fp16")];
            fp32 var_2938_epsilon_0 = const()[name = string("op_2938_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2938_cast_fp16 = rsqrt(epsilon = var_2938_epsilon_0, x = var_2937_cast_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_2938_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178368)))];
            tensor<fp16, [1, 1260, 12, 64]> x_83_cast_fp16 = mul(x = out_53_cast_fp16, y = t_encoder_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39643200))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40233088))))[name = string("t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<int32, [4]> var_2948 = const()[name = string("op_2948"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_79_cast_fp16 = reshape(shape = var_2948, x = linear_31_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2950_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_2950_cast_fp16")];
            bool var_2953 = const()[name = string("op_2953"), val = bool(true)];
            tensor<int32, [1]> var_2954_axes_0 = const()[name = string("op_2954_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_2954_cast_fp16 = reduce_max(axes = var_2954_axes_0, keep_dims = var_2953, x = var_2950_cast_fp16)[name = string("op_2954_cast_fp16")];
            fp16 var_2955_to_fp16 = const()[name = string("op_2955_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_67_cast_fp16 = maximum(x = var_2954_cast_fp16, y = var_2955_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_63_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_67_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_2958_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_2958_cast_fp16")];
            tensor<int32, [1]> var_2963_axes_0 = const()[name = string("op_2963_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2963_keep_dims_0 = const()[name = string("op_2963_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_2963_cast_fp16 = reduce_mean(axes = var_2963_axes_0, keep_dims = var_2963_keep_dims_0, x = var_2958_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2964_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_2964_cast_fp16")];
            fp32 var_2965_epsilon_0 = const()[name = string("op_2965_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_2965_cast_fp16 = inverse(epsilon = var_2965_epsilon_0, x = var_2964_cast_fp16)[name = string("op_2965_cast_fp16")];
            fp16 var_2966_to_fp16 = const()[name = string("op_2966_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_2967_cast_fp16 = mul(x = var_2965_cast_fp16, y = var_2966_to_fp16)[name = string("op_2967_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_2969_cast_fp16 = add(x = var_2963_cast_fp16, y = var_2967_cast_fp16)[name = string("op_2969_cast_fp16")];
            fp32 var_2970_epsilon_0 = const()[name = string("op_2970_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_2970_cast_fp16 = rsqrt(epsilon = var_2970_epsilon_0, x = var_2969_cast_fp16)[name = string("op_2970_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_17_cast_fp16 = mul(x = y_63_cast_fp16, y = var_2970_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<int32, [4]> z_33_begin_0 = const()[name = string("z_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_33_end_0 = const()[name = string("z_33_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_33_end_mask_0 = const()[name = string("z_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_33_cast_fp16 = slice_by_index(begin = z_33_begin_0, end = z_33_end_0, end_mask = z_33_end_mask_0, x = x_81_cast_fp16)[name = string("z_33_cast_fp16")];
            tensor<int32, [4]> var_2981_begin_0 = const()[name = string("op_2981_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2981_end_0 = const()[name = string("op_2981_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_2981_end_mask_0 = const()[name = string("op_2981_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_2981_cast_fp16 = slice_by_index(begin = var_2981_begin_0, end = var_2981_end_0, end_mask = var_2981_end_mask_0, x = z_33_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_2982_cast_fp16 = mul(x = var_2981_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_2982_cast_fp16")];
            tensor<int32, [4]> var_2987_begin_0 = const()[name = string("op_2987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2987_end_0 = const()[name = string("op_2987_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_2987_end_mask_0 = const()[name = string("op_2987_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_2987_cast_fp16 = slice_by_index(begin = var_2987_begin_0, end = var_2987_end_0, end_mask = var_2987_end_mask_0, x = z_33_cast_fp16)[name = string("op_2987_cast_fp16")];
            int32 var_2989 = const()[name = string("op_2989"), val = int32(-1)];
            bool rot_33_interleave_0 = const()[name = string("rot_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_33_cast_fp16 = concat(axis = var_2989, interleave = rot_33_interleave_0, values = (var_2982_cast_fp16, var_2987_cast_fp16))[name = string("rot_33_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2991_cast_fp16 = mul(x = z_33_cast_fp16, y = c_1_cast_fp16)[name = string("op_2991_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2992_cast_fp16 = mul(x = rot_33_cast_fp16, y = s_9_cast_fp16)[name = string("op_2992_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_2994_cast_fp16 = add(x = var_2991_cast_fp16, y = var_2992_cast_fp16)[name = string("op_2994_cast_fp16")];
            tensor<int32, [4]> z_35_begin_0 = const()[name = string("z_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_35_end_0 = const()[name = string("z_35_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_35_end_mask_0 = const()[name = string("z_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_35_cast_fp16 = slice_by_index(begin = z_35_begin_0, end = z_35_end_0, end_mask = z_35_end_mask_0, x = x_81_cast_fp16)[name = string("z_35_cast_fp16")];
            tensor<int32, [4]> var_3004_begin_0 = const()[name = string("op_3004_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3004_end_0 = const()[name = string("op_3004_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3004_end_mask_0 = const()[name = string("op_3004_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, x = z_35_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3005_cast_fp16 = mul(x = var_3004_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_3005_cast_fp16")];
            tensor<int32, [4]> var_3010_begin_0 = const()[name = string("op_3010_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3010_end_0 = const()[name = string("op_3010_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3010_end_mask_0 = const()[name = string("op_3010_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3010_cast_fp16 = slice_by_index(begin = var_3010_begin_0, end = var_3010_end_0, end_mask = var_3010_end_mask_0, x = z_35_cast_fp16)[name = string("op_3010_cast_fp16")];
            int32 var_3012 = const()[name = string("op_3012"), val = int32(-1)];
            bool rot_35_interleave_0 = const()[name = string("rot_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_35_cast_fp16 = concat(axis = var_3012, interleave = rot_35_interleave_0, values = (var_3005_cast_fp16, var_3010_cast_fp16))[name = string("rot_35_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3014_cast_fp16 = mul(x = z_35_cast_fp16, y = c_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3015_cast_fp16 = mul(x = rot_35_cast_fp16, y = s_11_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3017_cast_fp16 = add(x = var_3014_cast_fp16, y = var_3015_cast_fp16)[name = string("op_3017_cast_fp16")];
            int32 var_3019 = const()[name = string("op_3019"), val = int32(-1)];
            bool var_3020_interleave_0 = const()[name = string("op_3020_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_3020_cast_fp16 = concat(axis = var_3019, interleave = var_3020_interleave_0, values = (var_2994_cast_fp16, var_3017_cast_fp16))[name = string("op_3020_cast_fp16")];
            tensor<int32, [4]> z_37_begin_0 = const()[name = string("z_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_37_end_0 = const()[name = string("z_37_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_37_end_mask_0 = const()[name = string("z_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_37_cast_fp16 = slice_by_index(begin = z_37_begin_0, end = z_37_end_0, end_mask = z_37_end_mask_0, x = x_83_cast_fp16)[name = string("z_37_cast_fp16")];
            tensor<int32, [4]> var_3033_begin_0 = const()[name = string("op_3033_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3033_end_0 = const()[name = string("op_3033_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3033_end_mask_0 = const()[name = string("op_3033_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3033_cast_fp16 = slice_by_index(begin = var_3033_begin_0, end = var_3033_end_0, end_mask = var_3033_end_mask_0, x = z_37_cast_fp16)[name = string("op_3033_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3034_cast_fp16 = mul(x = var_3033_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> var_3039_begin_0 = const()[name = string("op_3039_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3039_end_0 = const()[name = string("op_3039_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3039_end_mask_0 = const()[name = string("op_3039_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3039_cast_fp16 = slice_by_index(begin = var_3039_begin_0, end = var_3039_end_0, end_mask = var_3039_end_mask_0, x = z_37_cast_fp16)[name = string("op_3039_cast_fp16")];
            int32 var_3041 = const()[name = string("op_3041"), val = int32(-1)];
            bool rot_37_interleave_0 = const()[name = string("rot_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_37_cast_fp16 = concat(axis = var_3041, interleave = rot_37_interleave_0, values = (var_3034_cast_fp16, var_3039_cast_fp16))[name = string("rot_37_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3043_cast_fp16 = mul(x = z_37_cast_fp16, y = c_1_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3044_cast_fp16 = mul(x = rot_37_cast_fp16, y = s_9_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3046_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3044_cast_fp16)[name = string("op_3046_cast_fp16")];
            tensor<int32, [4]> z_39_begin_0 = const()[name = string("z_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_39_end_0 = const()[name = string("z_39_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_39_end_mask_0 = const()[name = string("z_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_39_cast_fp16 = slice_by_index(begin = z_39_begin_0, end = z_39_end_0, end_mask = z_39_end_mask_0, x = x_83_cast_fp16)[name = string("z_39_cast_fp16")];
            tensor<int32, [4]> var_3056_begin_0 = const()[name = string("op_3056_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3056_end_0 = const()[name = string("op_3056_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3056_end_mask_0 = const()[name = string("op_3056_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3056_cast_fp16 = slice_by_index(begin = var_3056_begin_0, end = var_3056_end_0, end_mask = var_3056_end_mask_0, x = z_39_cast_fp16)[name = string("op_3056_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3057_cast_fp16 = mul(x = var_3056_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_3057_cast_fp16")];
            tensor<int32, [4]> var_3062_begin_0 = const()[name = string("op_3062_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3062_end_0 = const()[name = string("op_3062_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3062_end_mask_0 = const()[name = string("op_3062_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3062_cast_fp16 = slice_by_index(begin = var_3062_begin_0, end = var_3062_end_0, end_mask = var_3062_end_mask_0, x = z_39_cast_fp16)[name = string("op_3062_cast_fp16")];
            int32 var_3064 = const()[name = string("op_3064"), val = int32(-1)];
            bool rot_39_interleave_0 = const()[name = string("rot_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_39_cast_fp16 = concat(axis = var_3064, interleave = rot_39_interleave_0, values = (var_3057_cast_fp16, var_3062_cast_fp16))[name = string("rot_39_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3066_cast_fp16 = mul(x = z_39_cast_fp16, y = c_cast_fp16)[name = string("op_3066_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3067_cast_fp16 = mul(x = rot_39_cast_fp16, y = s_11_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3069_cast_fp16 = add(x = var_3066_cast_fp16, y = var_3067_cast_fp16)[name = string("op_3069_cast_fp16")];
            int32 var_3071 = const()[name = string("op_3071"), val = int32(-1)];
            bool var_3072_interleave_0 = const()[name = string("op_3072_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_3072_cast_fp16 = concat(axis = var_3071, interleave = var_3072_interleave_0, values = (var_3046_cast_fp16, var_3069_cast_fp16))[name = string("op_3072_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3082_transpose_x_0 = const()[name = string("op_3082_transpose_x_0"), val = bool(false)];
            bool var_3082_transpose_y_0 = const()[name = string("op_3082_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_3072_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_72 = transpose(perm = transpose_72_perm_0, x = var_3020_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 12, 1260, 1260]> var_3082_cast_fp16 = matmul(transpose_x = var_3082_transpose_x_0, transpose_y = var_3082_transpose_y_0, x = transpose_72, y = transpose_73)[name = string("op_3082_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_3084_cast_fp16 = add(x = var_3082_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3084_cast_fp16")];
            int32 var_3085 = const()[name = string("op_3085"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_63_cast_fp16 = softmax(axis = var_3085, x = var_3084_cast_fp16)[name = string("w_63_cast_fp16")];
            bool var_3088_transpose_x_0 = const()[name = string("op_3088_transpose_x_0"), val = bool(false)];
            bool var_3088_transpose_y_0 = const()[name = string("op_3088_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = v_17_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 12, 1260, 64]> var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = w_63_cast_fp16, y = v_19_cast_fp16)[name = string("op_3088_cast_fp16")];
            tensor<int32, [4]> var_3091_perm_0 = const()[name = string("op_3091_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3095 = const()[name = string("op_3095"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_3091_cast_fp16 = transpose(perm = var_3091_perm_0, x = var_3088_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 1260, 768]> input_45_cast_fp16 = reshape(shape = var_3095, x = var_3091_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40234688))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40824576))))[name = string("t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3100_cast_fp16 = abs(x = linear_32_cast_fp16)[name = string("op_3100_cast_fp16")];
            bool var_3103 = const()[name = string("op_3103"), val = bool(true)];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3104_cast_fp16 = reduce_max(axes = var_3104_axes_0, keep_dims = var_3103, x = var_3100_cast_fp16)[name = string("op_3104_cast_fp16")];
            fp16 var_3105_to_fp16 = const()[name = string("op_3105_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_69_cast_fp16 = maximum(x = var_3104_cast_fp16, y = var_3105_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_65_cast_fp16 = real_div(x = linear_32_cast_fp16, y = s_69_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3108_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [1]> var_3113_axes_0 = const()[name = string("op_3113_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3113_keep_dims_0 = const()[name = string("op_3113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3113_cast_fp16 = reduce_mean(axes = var_3113_axes_0, keep_dims = var_3113_keep_dims_0, x = var_3108_cast_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3114_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3114_cast_fp16")];
            fp32 var_3115_epsilon_0 = const()[name = string("op_3115_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3115_cast_fp16 = inverse(epsilon = var_3115_epsilon_0, x = var_3114_cast_fp16)[name = string("op_3115_cast_fp16")];
            fp16 var_3116_to_fp16 = const()[name = string("op_3116_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3117_cast_fp16 = mul(x = var_3115_cast_fp16, y = var_3116_to_fp16)[name = string("op_3117_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3119_cast_fp16 = add(x = var_3113_cast_fp16, y = var_3117_cast_fp16)[name = string("op_3119_cast_fp16")];
            fp32 var_3120_epsilon_0 = const()[name = string("op_3120_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3120_cast_fp16 = rsqrt(epsilon = var_3120_epsilon_0, x = var_3119_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_55_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3120_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40826176)))];
            tensor<fp16, [1, 1260, 768]> var_3122_cast_fp16 = mul(x = out_55_cast_fp16, y = t_encoder_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3122_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_87_cast_fp16 = add(x = x_73_cast_fp16, y = var_3122_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3125_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3125_cast_fp16")];
            bool var_3128 = const()[name = string("op_3128"), val = bool(true)];
            tensor<int32, [1]> var_3129_axes_0 = const()[name = string("op_3129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3129_cast_fp16 = reduce_max(axes = var_3129_axes_0, keep_dims = var_3128, x = var_3125_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_71_cast_fp16 = maximum(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_67_cast_fp16 = real_div(x = x_87_cast_fp16, y = s_71_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3133_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<int32, [1]> var_3138_axes_0 = const()[name = string("op_3138_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3138_keep_dims_0 = const()[name = string("op_3138_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3138_cast_fp16 = reduce_mean(axes = var_3138_axes_0, keep_dims = var_3138_keep_dims_0, x = var_3133_cast_fp16)[name = string("op_3138_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3139_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3139_cast_fp16")];
            fp32 var_3140_epsilon_0 = const()[name = string("op_3140_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3140_cast_fp16 = inverse(epsilon = var_3140_epsilon_0, x = var_3139_cast_fp16)[name = string("op_3140_cast_fp16")];
            fp16 var_3141_to_fp16 = const()[name = string("op_3141_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3142_cast_fp16 = mul(x = var_3140_cast_fp16, y = var_3141_to_fp16)[name = string("op_3142_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3144_cast_fp16 = add(x = var_3138_cast_fp16, y = var_3142_cast_fp16)[name = string("op_3144_cast_fp16")];
            fp32 var_3145_epsilon_0 = const()[name = string("op_3145_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3145_cast_fp16 = rsqrt(epsilon = var_3145_epsilon_0, x = var_3144_cast_fp16)[name = string("op_3145_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_57_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3145_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40827776)))];
            tensor<fp16, [1, 1260, 768]> input_47_cast_fp16 = mul(x = out_57_cast_fp16, y = t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40829376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43188736))))[name = string("t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_33_cast_fp16")];
            string var_3152_mode_0 = const()[name = string("op_3152_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_3152_cast_fp16 = gelu(mode = var_3152_mode_0, x = linear_33_cast_fp16)[name = string("op_3152_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43194944))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45554304))))[name = string("t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_34_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_51_cast_fp16 = mul(x = var_3152_cast_fp16, y = linear_34_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45560512))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47919872))))[name = string("t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3160_cast_fp16 = abs(x = linear_35_cast_fp16)[name = string("op_3160_cast_fp16")];
            bool var_3163 = const()[name = string("op_3163"), val = bool(true)];
            tensor<int32, [1]> var_3164_axes_0 = const()[name = string("op_3164_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3164_cast_fp16 = reduce_max(axes = var_3164_axes_0, keep_dims = var_3163, x = var_3160_cast_fp16)[name = string("op_3164_cast_fp16")];
            fp16 var_3165_to_fp16 = const()[name = string("op_3165_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_73_cast_fp16 = maximum(x = var_3164_cast_fp16, y = var_3165_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_69_cast_fp16 = real_div(x = linear_35_cast_fp16, y = s_73_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3168_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3168_cast_fp16")];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3173_keep_dims_0 = const()[name = string("op_3173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3173_cast_fp16 = reduce_mean(axes = var_3173_axes_0, keep_dims = var_3173_keep_dims_0, x = var_3168_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3174_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3174_cast_fp16")];
            fp32 var_3175_epsilon_0 = const()[name = string("op_3175_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3175_cast_fp16 = inverse(epsilon = var_3175_epsilon_0, x = var_3174_cast_fp16)[name = string("op_3175_cast_fp16")];
            fp16 var_3176_to_fp16 = const()[name = string("op_3176_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3177_cast_fp16 = mul(x = var_3175_cast_fp16, y = var_3176_to_fp16)[name = string("op_3177_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3179_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3177_cast_fp16)[name = string("op_3179_cast_fp16")];
            fp32 var_3180_epsilon_0 = const()[name = string("op_3180_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3180_cast_fp16 = rsqrt(epsilon = var_3180_epsilon_0, x = var_3179_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_59_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3180_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47921472)))];
            tensor<fp16, [1, 1260, 768]> var_3182_cast_fp16 = mul(x = out_59_cast_fp16, y = t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_91_cast_fp16 = add(x = x_87_cast_fp16, y = var_3182_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3185_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3185_cast_fp16")];
            bool var_3188 = const()[name = string("op_3188"), val = bool(true)];
            tensor<int32, [1]> var_3189_axes_0 = const()[name = string("op_3189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3189_cast_fp16 = reduce_max(axes = var_3189_axes_0, keep_dims = var_3188, x = var_3185_cast_fp16)[name = string("op_3189_cast_fp16")];
            fp16 var_3190_to_fp16 = const()[name = string("op_3190_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_75_cast_fp16 = maximum(x = var_3189_cast_fp16, y = var_3190_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_71_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_75_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3193_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3193_cast_fp16")];
            tensor<int32, [1]> var_3198_axes_0 = const()[name = string("op_3198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3198_keep_dims_0 = const()[name = string("op_3198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3198_cast_fp16 = reduce_mean(axes = var_3198_axes_0, keep_dims = var_3198_keep_dims_0, x = var_3193_cast_fp16)[name = string("op_3198_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3199_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3199_cast_fp16")];
            fp32 var_3200_epsilon_0 = const()[name = string("op_3200_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3200_cast_fp16 = inverse(epsilon = var_3200_epsilon_0, x = var_3199_cast_fp16)[name = string("op_3200_cast_fp16")];
            fp16 var_3201_to_fp16 = const()[name = string("op_3201_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3202_cast_fp16 = mul(x = var_3200_cast_fp16, y = var_3201_to_fp16)[name = string("op_3202_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3204_cast_fp16 = add(x = var_3198_cast_fp16, y = var_3202_cast_fp16)[name = string("op_3204_cast_fp16")];
            fp32 var_3205_epsilon_0 = const()[name = string("op_3205_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3205_cast_fp16 = rsqrt(epsilon = var_3205_epsilon_0, x = var_3204_cast_fp16)[name = string("op_3205_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_61_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3205_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47923072)))];
            tensor<fp16, [1, 1260, 768]> input_53_cast_fp16 = mul(x = out_61_cast_fp16, y = t_encoder_layers_5_input_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47924672))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48514560))))[name = string("t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_3215 = const()[name = string("op_3215"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_93_cast_fp16 = reshape(shape = var_3215, x = linear_36_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3217_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3217_cast_fp16")];
            bool var_3220 = const()[name = string("op_3220"), val = bool(true)];
            tensor<int32, [1]> var_3221_axes_0 = const()[name = string("op_3221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3221_cast_fp16 = reduce_max(axes = var_3221_axes_0, keep_dims = var_3220, x = var_3217_cast_fp16)[name = string("op_3221_cast_fp16")];
            fp16 var_3222_to_fp16 = const()[name = string("op_3222_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_77_cast_fp16 = maximum(x = var_3221_cast_fp16, y = var_3222_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_73_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_77_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3225_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<int32, [1]> var_3230_axes_0 = const()[name = string("op_3230_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3230_keep_dims_0 = const()[name = string("op_3230_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3230_cast_fp16 = reduce_mean(axes = var_3230_axes_0, keep_dims = var_3230_keep_dims_0, x = var_3225_cast_fp16)[name = string("op_3230_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3231_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3231_cast_fp16")];
            fp32 var_3232_epsilon_0 = const()[name = string("op_3232_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3232_cast_fp16 = inverse(epsilon = var_3232_epsilon_0, x = var_3231_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3234_cast_fp16 = mul(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3236_cast_fp16 = add(x = var_3230_cast_fp16, y = var_3234_cast_fp16)[name = string("op_3236_cast_fp16")];
            fp32 var_3237_epsilon_0 = const()[name = string("op_3237_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3237_cast_fp16 = rsqrt(epsilon = var_3237_epsilon_0, x = var_3236_cast_fp16)[name = string("op_3237_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_63_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3237_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48516160)))];
            tensor<fp16, [1, 1260, 12, 64]> x_99_cast_fp16 = mul(x = out_63_cast_fp16, y = t_encoder_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_99_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48516352))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49106240))))[name = string("t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_37_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3247 = const()[name = string("op_3247"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_95_cast_fp16 = reshape(shape = var_3247, x = linear_37_cast_fp16)[name = string("x_95_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3249_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3249_cast_fp16")];
            bool var_3252 = const()[name = string("op_3252"), val = bool(true)];
            tensor<int32, [1]> var_3253_axes_0 = const()[name = string("op_3253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3253_cast_fp16 = reduce_max(axes = var_3253_axes_0, keep_dims = var_3252, x = var_3249_cast_fp16)[name = string("op_3253_cast_fp16")];
            fp16 var_3254_to_fp16 = const()[name = string("op_3254_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_79_cast_fp16 = maximum(x = var_3253_cast_fp16, y = var_3254_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_75_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_79_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3257_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3257_cast_fp16")];
            tensor<int32, [1]> var_3262_axes_0 = const()[name = string("op_3262_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3262_keep_dims_0 = const()[name = string("op_3262_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3262_cast_fp16 = reduce_mean(axes = var_3262_axes_0, keep_dims = var_3262_keep_dims_0, x = var_3257_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3263_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3263_cast_fp16")];
            fp32 var_3264_epsilon_0 = const()[name = string("op_3264_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3264_cast_fp16 = inverse(epsilon = var_3264_epsilon_0, x = var_3263_cast_fp16)[name = string("op_3264_cast_fp16")];
            fp16 var_3265_to_fp16 = const()[name = string("op_3265_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3266_cast_fp16 = mul(x = var_3264_cast_fp16, y = var_3265_to_fp16)[name = string("op_3266_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3268_cast_fp16 = add(x = var_3262_cast_fp16, y = var_3266_cast_fp16)[name = string("op_3268_cast_fp16")];
            fp32 var_3269_epsilon_0 = const()[name = string("op_3269_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3269_cast_fp16 = rsqrt(epsilon = var_3269_epsilon_0, x = var_3268_cast_fp16)[name = string("op_3269_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3269_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49107840)))];
            tensor<fp16, [1, 1260, 12, 64]> x_101_cast_fp16 = mul(x = out_65_cast_fp16, y = t_encoder_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49108032))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49697920))))[name = string("t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3279 = const()[name = string("op_3279"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_97_cast_fp16 = reshape(shape = var_3279, x = linear_38_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3281_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3281_cast_fp16")];
            bool var_3284 = const()[name = string("op_3284"), val = bool(true)];
            tensor<int32, [1]> var_3285_axes_0 = const()[name = string("op_3285_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3285_cast_fp16 = reduce_max(axes = var_3285_axes_0, keep_dims = var_3284, x = var_3281_cast_fp16)[name = string("op_3285_cast_fp16")];
            fp16 var_3286_to_fp16 = const()[name = string("op_3286_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_81_cast_fp16 = maximum(x = var_3285_cast_fp16, y = var_3286_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_77_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_81_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3289_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<int32, [1]> var_3294_axes_0 = const()[name = string("op_3294_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3294_keep_dims_0 = const()[name = string("op_3294_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3294_cast_fp16 = reduce_mean(axes = var_3294_axes_0, keep_dims = var_3294_keep_dims_0, x = var_3289_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3295_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3295_cast_fp16")];
            fp32 var_3296_epsilon_0 = const()[name = string("op_3296_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3296_cast_fp16 = inverse(epsilon = var_3296_epsilon_0, x = var_3295_cast_fp16)[name = string("op_3296_cast_fp16")];
            fp16 var_3297_to_fp16 = const()[name = string("op_3297_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3298_cast_fp16 = mul(x = var_3296_cast_fp16, y = var_3297_to_fp16)[name = string("op_3298_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3300_cast_fp16 = add(x = var_3294_cast_fp16, y = var_3298_cast_fp16)[name = string("op_3300_cast_fp16")];
            fp32 var_3301_epsilon_0 = const()[name = string("op_3301_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3301_cast_fp16 = rsqrt(epsilon = var_3301_epsilon_0, x = var_3300_cast_fp16)[name = string("op_3301_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_21_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3301_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<int32, [4]> z_41_begin_0 = const()[name = string("z_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_41_end_0 = const()[name = string("z_41_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_41_end_mask_0 = const()[name = string("z_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_41_cast_fp16 = slice_by_index(begin = z_41_begin_0, end = z_41_end_0, end_mask = z_41_end_mask_0, x = x_99_cast_fp16)[name = string("z_41_cast_fp16")];
            tensor<int32, [4]> var_3312_begin_0 = const()[name = string("op_3312_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3312_end_0 = const()[name = string("op_3312_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3312_end_mask_0 = const()[name = string("op_3312_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3312_cast_fp16 = slice_by_index(begin = var_3312_begin_0, end = var_3312_end_0, end_mask = var_3312_end_mask_0, x = z_41_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3313_cast_fp16 = mul(x = var_3312_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_3313_cast_fp16")];
            tensor<int32, [4]> var_3318_begin_0 = const()[name = string("op_3318_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3318_end_0 = const()[name = string("op_3318_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3318_end_mask_0 = const()[name = string("op_3318_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> 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 = z_41_cast_fp16)[name = string("op_3318_cast_fp16")];
            int32 var_3320 = const()[name = string("op_3320"), val = int32(-1)];
            bool rot_41_interleave_0 = const()[name = string("rot_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_41_cast_fp16 = concat(axis = var_3320, interleave = rot_41_interleave_0, values = (var_3313_cast_fp16, var_3318_cast_fp16))[name = string("rot_41_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3322_cast_fp16 = mul(x = z_41_cast_fp16, y = c_1_cast_fp16)[name = string("op_3322_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3323_cast_fp16 = mul(x = rot_41_cast_fp16, y = s_9_cast_fp16)[name = string("op_3323_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3325_cast_fp16 = add(x = var_3322_cast_fp16, y = var_3323_cast_fp16)[name = string("op_3325_cast_fp16")];
            tensor<int32, [4]> z_43_begin_0 = const()[name = string("z_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_43_end_0 = const()[name = string("z_43_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_43_end_mask_0 = const()[name = string("z_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_43_cast_fp16 = slice_by_index(begin = z_43_begin_0, end = z_43_end_0, end_mask = z_43_end_mask_0, x = x_99_cast_fp16)[name = string("z_43_cast_fp16")];
            tensor<int32, [4]> var_3335_begin_0 = const()[name = string("op_3335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3335_end_0 = const()[name = string("op_3335_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3335_end_mask_0 = const()[name = string("op_3335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3335_cast_fp16 = slice_by_index(begin = var_3335_begin_0, end = var_3335_end_0, end_mask = var_3335_end_mask_0, x = z_43_cast_fp16)[name = string("op_3335_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3336_cast_fp16 = mul(x = var_3335_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> var_3341_begin_0 = const()[name = string("op_3341_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3341_end_0 = const()[name = string("op_3341_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3341_end_mask_0 = const()[name = string("op_3341_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3341_cast_fp16 = slice_by_index(begin = var_3341_begin_0, end = var_3341_end_0, end_mask = var_3341_end_mask_0, x = z_43_cast_fp16)[name = string("op_3341_cast_fp16")];
            int32 var_3343 = const()[name = string("op_3343"), val = int32(-1)];
            bool rot_43_interleave_0 = const()[name = string("rot_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_43_cast_fp16 = concat(axis = var_3343, interleave = rot_43_interleave_0, values = (var_3336_cast_fp16, var_3341_cast_fp16))[name = string("rot_43_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3345_cast_fp16 = mul(x = z_43_cast_fp16, y = c_cast_fp16)[name = string("op_3345_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3346_cast_fp16 = mul(x = rot_43_cast_fp16, y = s_11_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3348_cast_fp16 = add(x = var_3345_cast_fp16, y = var_3346_cast_fp16)[name = string("op_3348_cast_fp16")];
            int32 var_3350 = const()[name = string("op_3350"), val = int32(-1)];
            bool var_3351_interleave_0 = const()[name = string("op_3351_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_3351_cast_fp16 = concat(axis = var_3350, interleave = var_3351_interleave_0, values = (var_3325_cast_fp16, var_3348_cast_fp16))[name = string("op_3351_cast_fp16")];
            tensor<int32, [4]> z_45_begin_0 = const()[name = string("z_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_45_end_0 = const()[name = string("z_45_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_45_end_mask_0 = const()[name = string("z_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_45_cast_fp16 = slice_by_index(begin = z_45_begin_0, end = z_45_end_0, end_mask = z_45_end_mask_0, x = x_101_cast_fp16)[name = string("z_45_cast_fp16")];
            tensor<int32, [4]> var_3364_begin_0 = const()[name = string("op_3364_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3364_end_0 = const()[name = string("op_3364_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3364_end_mask_0 = const()[name = string("op_3364_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3364_cast_fp16 = slice_by_index(begin = var_3364_begin_0, end = var_3364_end_0, end_mask = var_3364_end_mask_0, x = z_45_cast_fp16)[name = string("op_3364_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3365_cast_fp16 = mul(x = var_3364_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_3365_cast_fp16")];
            tensor<int32, [4]> var_3370_begin_0 = const()[name = string("op_3370_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3370_end_0 = const()[name = string("op_3370_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3370_end_mask_0 = const()[name = string("op_3370_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3370_cast_fp16 = slice_by_index(begin = var_3370_begin_0, end = var_3370_end_0, end_mask = var_3370_end_mask_0, x = z_45_cast_fp16)[name = string("op_3370_cast_fp16")];
            int32 var_3372 = const()[name = string("op_3372"), val = int32(-1)];
            bool rot_45_interleave_0 = const()[name = string("rot_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_45_cast_fp16 = concat(axis = var_3372, interleave = rot_45_interleave_0, values = (var_3365_cast_fp16, var_3370_cast_fp16))[name = string("rot_45_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3374_cast_fp16 = mul(x = z_45_cast_fp16, y = c_1_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3375_cast_fp16 = mul(x = rot_45_cast_fp16, y = s_9_cast_fp16)[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3377_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3375_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [4]> z_47_begin_0 = const()[name = string("z_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_47_end_0 = const()[name = string("z_47_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_47_end_mask_0 = const()[name = string("z_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_47_cast_fp16 = slice_by_index(begin = z_47_begin_0, end = z_47_end_0, end_mask = z_47_end_mask_0, x = x_101_cast_fp16)[name = string("z_47_cast_fp16")];
            tensor<int32, [4]> var_3387_begin_0 = const()[name = string("op_3387_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3387_end_0 = const()[name = string("op_3387_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3387_end_mask_0 = const()[name = string("op_3387_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3387_cast_fp16 = slice_by_index(begin = var_3387_begin_0, end = var_3387_end_0, end_mask = var_3387_end_mask_0, x = z_47_cast_fp16)[name = string("op_3387_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3388_cast_fp16 = mul(x = var_3387_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_3388_cast_fp16")];
            tensor<int32, [4]> var_3393_begin_0 = const()[name = string("op_3393_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3393_end_0 = const()[name = string("op_3393_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3393_end_mask_0 = const()[name = string("op_3393_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3393_cast_fp16 = slice_by_index(begin = var_3393_begin_0, end = var_3393_end_0, end_mask = var_3393_end_mask_0, x = z_47_cast_fp16)[name = string("op_3393_cast_fp16")];
            int32 var_3395 = const()[name = string("op_3395"), val = int32(-1)];
            bool rot_47_interleave_0 = const()[name = string("rot_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_47_cast_fp16 = concat(axis = var_3395, interleave = rot_47_interleave_0, values = (var_3388_cast_fp16, var_3393_cast_fp16))[name = string("rot_47_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3397_cast_fp16 = mul(x = z_47_cast_fp16, y = c_cast_fp16)[name = string("op_3397_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3398_cast_fp16 = mul(x = rot_47_cast_fp16, y = s_11_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3400_cast_fp16 = add(x = var_3397_cast_fp16, y = var_3398_cast_fp16)[name = string("op_3400_cast_fp16")];
            int32 var_3402 = const()[name = string("op_3402"), val = int32(-1)];
            bool var_3403_interleave_0 = const()[name = string("op_3403_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_3403_cast_fp16 = concat(axis = var_3402, interleave = var_3403_interleave_0, values = (var_3377_cast_fp16, var_3400_cast_fp16))[name = string("op_3403_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3413_transpose_x_0 = const()[name = string("op_3413_transpose_x_0"), val = bool(false)];
            bool var_3413_transpose_y_0 = const()[name = string("op_3413_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_3403_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_74 = transpose(perm = transpose_74_perm_0, x = var_3351_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 12, 1260, 1260]> var_3413_cast_fp16 = matmul(transpose_x = var_3413_transpose_x_0, transpose_y = var_3413_transpose_y_0, x = transpose_74, y = transpose_75)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_3415_cast_fp16 = add(x = var_3413_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3415_cast_fp16")];
            int32 var_3416 = const()[name = string("op_3416"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_77_cast_fp16 = softmax(axis = var_3416, x = var_3415_cast_fp16)[name = string("w_77_cast_fp16")];
            bool var_3419_transpose_x_0 = const()[name = string("op_3419_transpose_x_0"), val = bool(false)];
            bool var_3419_transpose_y_0 = const()[name = string("op_3419_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = v_21_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 12, 1260, 64]> var_3419_cast_fp16 = matmul(transpose_x = var_3419_transpose_x_0, transpose_y = var_3419_transpose_y_0, x = w_77_cast_fp16, y = v_23_cast_fp16)[name = string("op_3419_cast_fp16")];
            tensor<int32, [4]> var_3422_perm_0 = const()[name = string("op_3422_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3426 = const()[name = string("op_3426"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_3422_cast_fp16 = transpose(perm = var_3422_perm_0, x = var_3419_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 1260, 768]> input_55_cast_fp16 = reshape(shape = var_3426, x = var_3422_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49699520))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50289408))))[name = string("t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_39_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3431_cast_fp16 = abs(x = linear_39_cast_fp16)[name = string("op_3431_cast_fp16")];
            bool var_3434 = const()[name = string("op_3434"), val = bool(true)];
            tensor<int32, [1]> var_3435_axes_0 = const()[name = string("op_3435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3435_cast_fp16 = reduce_max(axes = var_3435_axes_0, keep_dims = var_3434, x = var_3431_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp16 var_3436_to_fp16 = const()[name = string("op_3436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_83_cast_fp16 = maximum(x = var_3435_cast_fp16, y = var_3436_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_79_cast_fp16 = real_div(x = linear_39_cast_fp16, y = s_83_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3439_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3439_cast_fp16")];
            tensor<int32, [1]> var_3444_axes_0 = const()[name = string("op_3444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3444_keep_dims_0 = const()[name = string("op_3444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3444_cast_fp16 = reduce_mean(axes = var_3444_axes_0, keep_dims = var_3444_keep_dims_0, x = var_3439_cast_fp16)[name = string("op_3444_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3445_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3445_cast_fp16")];
            fp32 var_3446_epsilon_0 = const()[name = string("op_3446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3446_cast_fp16 = inverse(epsilon = var_3446_epsilon_0, x = var_3445_cast_fp16)[name = string("op_3446_cast_fp16")];
            fp16 var_3447_to_fp16 = const()[name = string("op_3447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3448_cast_fp16 = mul(x = var_3446_cast_fp16, y = var_3447_to_fp16)[name = string("op_3448_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3450_cast_fp16 = add(x = var_3444_cast_fp16, y = var_3448_cast_fp16)[name = string("op_3450_cast_fp16")];
            fp32 var_3451_epsilon_0 = const()[name = string("op_3451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3451_cast_fp16 = rsqrt(epsilon = var_3451_epsilon_0, x = var_3450_cast_fp16)[name = string("op_3451_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_67_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3451_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50291008)))];
            tensor<fp16, [1, 1260, 768]> var_3453_cast_fp16 = mul(x = out_67_cast_fp16, y = t_encoder_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_3453_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_105_cast_fp16 = add(x = x_91_cast_fp16, y = var_3453_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3456_cast_fp16 = abs(x = x_105_cast_fp16)[name = string("op_3456_cast_fp16")];
            bool var_3459 = const()[name = string("op_3459"), val = bool(true)];
            tensor<int32, [1]> var_3460_axes_0 = const()[name = string("op_3460_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3460_cast_fp16 = reduce_max(axes = var_3460_axes_0, keep_dims = var_3459, x = var_3456_cast_fp16)[name = string("op_3460_cast_fp16")];
            fp16 var_3461_to_fp16 = const()[name = string("op_3461_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_85_cast_fp16 = maximum(x = var_3460_cast_fp16, y = var_3461_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_81_cast_fp16 = real_div(x = x_105_cast_fp16, y = s_85_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3464_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3464_cast_fp16")];
            tensor<int32, [1]> var_3469_axes_0 = const()[name = string("op_3469_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3469_keep_dims_0 = const()[name = string("op_3469_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3469_cast_fp16 = reduce_mean(axes = var_3469_axes_0, keep_dims = var_3469_keep_dims_0, x = var_3464_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3470_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3470_cast_fp16")];
            fp32 var_3471_epsilon_0 = const()[name = string("op_3471_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3471_cast_fp16 = inverse(epsilon = var_3471_epsilon_0, x = var_3470_cast_fp16)[name = string("op_3471_cast_fp16")];
            fp16 var_3472_to_fp16 = const()[name = string("op_3472_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3473_cast_fp16 = mul(x = var_3471_cast_fp16, y = var_3472_to_fp16)[name = string("op_3473_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3475_cast_fp16 = add(x = var_3469_cast_fp16, y = var_3473_cast_fp16)[name = string("op_3475_cast_fp16")];
            fp32 var_3476_epsilon_0 = const()[name = string("op_3476_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3476_cast_fp16 = rsqrt(epsilon = var_3476_epsilon_0, x = var_3475_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_69_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3476_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50292608)))];
            tensor<fp16, [1, 1260, 768]> input_57_cast_fp16 = mul(x = out_69_cast_fp16, y = t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50294208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52653568))))[name = string("t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_40_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_40_cast_fp16")];
            string var_3483_mode_0 = const()[name = string("op_3483_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_3483_cast_fp16 = gelu(mode = var_3483_mode_0, x = linear_40_cast_fp16)[name = string("op_3483_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52659776))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55019136))))[name = string("t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_61_cast_fp16 = mul(x = var_3483_cast_fp16, y = linear_41_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55025344))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57384704))))[name = string("t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3491_cast_fp16 = abs(x = linear_42_cast_fp16)[name = string("op_3491_cast_fp16")];
            bool var_3494 = const()[name = string("op_3494"), val = bool(true)];
            tensor<int32, [1]> var_3495_axes_0 = const()[name = string("op_3495_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3495_cast_fp16 = reduce_max(axes = var_3495_axes_0, keep_dims = var_3494, x = var_3491_cast_fp16)[name = string("op_3495_cast_fp16")];
            fp16 var_3496_to_fp16 = const()[name = string("op_3496_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_87_cast_fp16 = maximum(x = var_3495_cast_fp16, y = var_3496_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_83_cast_fp16 = real_div(x = linear_42_cast_fp16, y = s_87_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3499_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3499_cast_fp16")];
            tensor<int32, [1]> var_3504_axes_0 = const()[name = string("op_3504_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3504_keep_dims_0 = const()[name = string("op_3504_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3504_cast_fp16 = reduce_mean(axes = var_3504_axes_0, keep_dims = var_3504_keep_dims_0, x = var_3499_cast_fp16)[name = string("op_3504_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3505_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3505_cast_fp16")];
            fp32 var_3506_epsilon_0 = const()[name = string("op_3506_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3506_cast_fp16 = inverse(epsilon = var_3506_epsilon_0, x = var_3505_cast_fp16)[name = string("op_3506_cast_fp16")];
            fp16 var_3507_to_fp16 = const()[name = string("op_3507_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3508_cast_fp16 = mul(x = var_3506_cast_fp16, y = var_3507_to_fp16)[name = string("op_3508_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3510_cast_fp16 = add(x = var_3504_cast_fp16, y = var_3508_cast_fp16)[name = string("op_3510_cast_fp16")];
            fp32 var_3511_epsilon_0 = const()[name = string("op_3511_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3511_cast_fp16 = rsqrt(epsilon = var_3511_epsilon_0, x = var_3510_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_71_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3511_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57386304)))];
            tensor<fp16, [1, 1260, 768]> var_3513_cast_fp16 = mul(x = out_71_cast_fp16, y = t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3513_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_109_cast_fp16 = add(x = x_105_cast_fp16, y = var_3513_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3516_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_3516_cast_fp16")];
            bool var_3519 = const()[name = string("op_3519"), val = bool(true)];
            tensor<int32, [1]> var_3520_axes_0 = const()[name = string("op_3520_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3520_cast_fp16 = reduce_max(axes = var_3520_axes_0, keep_dims = var_3519, x = var_3516_cast_fp16)[name = string("op_3520_cast_fp16")];
            fp16 var_3521_to_fp16 = const()[name = string("op_3521_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_89_cast_fp16 = maximum(x = var_3520_cast_fp16, y = var_3521_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_85_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_89_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3524_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<int32, [1]> var_3529_axes_0 = const()[name = string("op_3529_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3529_keep_dims_0 = const()[name = string("op_3529_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3529_cast_fp16 = reduce_mean(axes = var_3529_axes_0, keep_dims = var_3529_keep_dims_0, x = var_3524_cast_fp16)[name = string("op_3529_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3530_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3530_cast_fp16")];
            fp32 var_3531_epsilon_0 = const()[name = string("op_3531_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3531_cast_fp16 = inverse(epsilon = var_3531_epsilon_0, x = var_3530_cast_fp16)[name = string("op_3531_cast_fp16")];
            fp16 var_3532_to_fp16 = const()[name = string("op_3532_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3533_cast_fp16 = mul(x = var_3531_cast_fp16, y = var_3532_to_fp16)[name = string("op_3533_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3535_cast_fp16 = add(x = var_3529_cast_fp16, y = var_3533_cast_fp16)[name = string("op_3535_cast_fp16")];
            fp32 var_3536_epsilon_0 = const()[name = string("op_3536_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3536_cast_fp16 = rsqrt(epsilon = var_3536_epsilon_0, x = var_3535_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_73_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3536_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57387904)))];
            tensor<fp16, [1, 1260, 768]> input_63_cast_fp16 = mul(x = out_73_cast_fp16, y = t_encoder_layers_6_input_layernorm_weight_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57389504))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57979392))))[name = string("t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_43_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_3546 = const()[name = string("op_3546"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_111_cast_fp16 = reshape(shape = var_3546, x = linear_43_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3548_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_3548_cast_fp16")];
            bool var_3551 = const()[name = string("op_3551"), val = bool(true)];
            tensor<int32, [1]> var_3552_axes_0 = const()[name = string("op_3552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3552_cast_fp16 = reduce_max(axes = var_3552_axes_0, keep_dims = var_3551, x = var_3548_cast_fp16)[name = string("op_3552_cast_fp16")];
            fp16 var_3553_to_fp16 = const()[name = string("op_3553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_91_cast_fp16 = maximum(x = var_3552_cast_fp16, y = var_3553_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_87_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_91_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3556_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3556_cast_fp16")];
            tensor<int32, [1]> var_3561_axes_0 = const()[name = string("op_3561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3561_keep_dims_0 = const()[name = string("op_3561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3561_cast_fp16 = reduce_mean(axes = var_3561_axes_0, keep_dims = var_3561_keep_dims_0, x = var_3556_cast_fp16)[name = string("op_3561_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3562_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp32 var_3563_epsilon_0 = const()[name = string("op_3563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3563_cast_fp16 = inverse(epsilon = var_3563_epsilon_0, x = var_3562_cast_fp16)[name = string("op_3563_cast_fp16")];
            fp16 var_3564_to_fp16 = const()[name = string("op_3564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3565_cast_fp16 = mul(x = var_3563_cast_fp16, y = var_3564_to_fp16)[name = string("op_3565_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3567_cast_fp16 = add(x = var_3561_cast_fp16, y = var_3565_cast_fp16)[name = string("op_3567_cast_fp16")];
            fp32 var_3568_epsilon_0 = const()[name = string("op_3568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3568_cast_fp16 = rsqrt(epsilon = var_3568_epsilon_0, x = var_3567_cast_fp16)[name = string("op_3568_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_75_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3568_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 1260, 12, 64]> x_117_cast_fp16 = mul(x = out_75_cast_fp16, y = t_encoder_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_117_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57980992))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58570880))))[name = string("t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_3578 = const()[name = string("op_3578"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_113_cast_fp16 = reshape(shape = var_3578, x = linear_44_cast_fp16)[name = string("x_113_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3580_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_3580_cast_fp16")];
            bool var_3583 = const()[name = string("op_3583"), val = bool(true)];
            tensor<int32, [1]> var_3584_axes_0 = const()[name = string("op_3584_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3584_cast_fp16 = reduce_max(axes = var_3584_axes_0, keep_dims = var_3583, x = var_3580_cast_fp16)[name = string("op_3584_cast_fp16")];
            fp16 var_3585_to_fp16 = const()[name = string("op_3585_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_93_cast_fp16 = maximum(x = var_3584_cast_fp16, y = var_3585_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_89_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_93_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3588_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3588_cast_fp16")];
            tensor<int32, [1]> var_3593_axes_0 = const()[name = string("op_3593_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3593_keep_dims_0 = const()[name = string("op_3593_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3593_cast_fp16 = reduce_mean(axes = var_3593_axes_0, keep_dims = var_3593_keep_dims_0, x = var_3588_cast_fp16)[name = string("op_3593_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3594_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_3594_cast_fp16")];
            fp32 var_3595_epsilon_0 = const()[name = string("op_3595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3595_cast_fp16 = inverse(epsilon = var_3595_epsilon_0, x = var_3594_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp16 var_3596_to_fp16 = const()[name = string("op_3596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3597_cast_fp16 = mul(x = var_3595_cast_fp16, y = var_3596_to_fp16)[name = string("op_3597_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3599_cast_fp16 = add(x = var_3593_cast_fp16, y = var_3597_cast_fp16)[name = string("op_3599_cast_fp16")];
            fp32 var_3600_epsilon_0 = const()[name = string("op_3600_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3600_cast_fp16 = rsqrt(epsilon = var_3600_epsilon_0, x = var_3599_cast_fp16)[name = string("op_3600_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_77_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3600_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58572480)))];
            tensor<fp16, [1, 1260, 12, 64]> x_119_cast_fp16 = mul(x = out_77_cast_fp16, y = t_encoder_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58572672))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59162560))))[name = string("t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<int32, [4]> var_3610 = const()[name = string("op_3610"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_115_cast_fp16 = reshape(shape = var_3610, x = linear_45_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3612_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_3612_cast_fp16")];
            bool var_3615 = const()[name = string("op_3615"), val = bool(true)];
            tensor<int32, [1]> var_3616_axes_0 = const()[name = string("op_3616_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3616_cast_fp16 = reduce_max(axes = var_3616_axes_0, keep_dims = var_3615, x = var_3612_cast_fp16)[name = string("op_3616_cast_fp16")];
            fp16 var_3617_to_fp16 = const()[name = string("op_3617_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_95_cast_fp16 = maximum(x = var_3616_cast_fp16, y = var_3617_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_91_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_95_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3620_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_3620_cast_fp16")];
            tensor<int32, [1]> var_3625_axes_0 = const()[name = string("op_3625_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3625_keep_dims_0 = const()[name = string("op_3625_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3625_cast_fp16 = reduce_mean(axes = var_3625_axes_0, keep_dims = var_3625_keep_dims_0, x = var_3620_cast_fp16)[name = string("op_3625_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3626_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_3626_cast_fp16")];
            fp32 var_3627_epsilon_0 = const()[name = string("op_3627_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3627_cast_fp16 = inverse(epsilon = var_3627_epsilon_0, x = var_3626_cast_fp16)[name = string("op_3627_cast_fp16")];
            fp16 var_3628_to_fp16 = const()[name = string("op_3628_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3629_cast_fp16 = mul(x = var_3627_cast_fp16, y = var_3628_to_fp16)[name = string("op_3629_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3631_cast_fp16 = add(x = var_3625_cast_fp16, y = var_3629_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp32 var_3632_epsilon_0 = const()[name = string("op_3632_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3632_cast_fp16 = rsqrt(epsilon = var_3632_epsilon_0, x = var_3631_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_25_cast_fp16 = mul(x = y_91_cast_fp16, y = var_3632_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<int32, [4]> z_49_begin_0 = const()[name = string("z_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_49_end_0 = const()[name = string("z_49_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_49_end_mask_0 = const()[name = string("z_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_49_cast_fp16 = slice_by_index(begin = z_49_begin_0, end = z_49_end_0, end_mask = z_49_end_mask_0, x = x_117_cast_fp16)[name = string("z_49_cast_fp16")];
            tensor<int32, [4]> var_3643_begin_0 = const()[name = string("op_3643_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3643_end_0 = const()[name = string("op_3643_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3643_end_mask_0 = const()[name = string("op_3643_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3643_cast_fp16 = slice_by_index(begin = var_3643_begin_0, end = var_3643_end_0, end_mask = var_3643_end_mask_0, x = z_49_cast_fp16)[name = string("op_3643_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3644_cast_fp16 = mul(x = var_3643_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = z_49_cast_fp16)[name = string("op_3649_cast_fp16")];
            int32 var_3651 = const()[name = string("op_3651"), val = int32(-1)];
            bool rot_49_interleave_0 = const()[name = string("rot_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_49_cast_fp16 = concat(axis = var_3651, interleave = rot_49_interleave_0, values = (var_3644_cast_fp16, var_3649_cast_fp16))[name = string("rot_49_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3653_cast_fp16 = mul(x = z_49_cast_fp16, y = c_1_cast_fp16)[name = string("op_3653_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3654_cast_fp16 = mul(x = rot_49_cast_fp16, y = s_9_cast_fp16)[name = string("op_3654_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3656_cast_fp16 = add(x = var_3653_cast_fp16, y = var_3654_cast_fp16)[name = string("op_3656_cast_fp16")];
            tensor<int32, [4]> z_51_begin_0 = const()[name = string("z_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_51_end_0 = const()[name = string("z_51_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_51_end_mask_0 = const()[name = string("z_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_51_cast_fp16 = slice_by_index(begin = z_51_begin_0, end = z_51_end_0, end_mask = z_51_end_mask_0, x = x_117_cast_fp16)[name = string("z_51_cast_fp16")];
            tensor<int32, [4]> var_3666_begin_0 = const()[name = string("op_3666_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3666_end_0 = const()[name = string("op_3666_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3666_end_mask_0 = const()[name = string("op_3666_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3666_cast_fp16 = slice_by_index(begin = var_3666_begin_0, end = var_3666_end_0, end_mask = var_3666_end_mask_0, x = z_51_cast_fp16)[name = string("op_3666_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3667_cast_fp16 = mul(x = var_3666_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_3667_cast_fp16")];
            tensor<int32, [4]> var_3672_begin_0 = const()[name = string("op_3672_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3672_end_0 = const()[name = string("op_3672_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3672_end_mask_0 = const()[name = string("op_3672_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3672_cast_fp16 = slice_by_index(begin = var_3672_begin_0, end = var_3672_end_0, end_mask = var_3672_end_mask_0, x = z_51_cast_fp16)[name = string("op_3672_cast_fp16")];
            int32 var_3674 = const()[name = string("op_3674"), val = int32(-1)];
            bool rot_51_interleave_0 = const()[name = string("rot_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_51_cast_fp16 = concat(axis = var_3674, interleave = rot_51_interleave_0, values = (var_3667_cast_fp16, var_3672_cast_fp16))[name = string("rot_51_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3676_cast_fp16 = mul(x = z_51_cast_fp16, y = c_cast_fp16)[name = string("op_3676_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3677_cast_fp16 = mul(x = rot_51_cast_fp16, y = s_11_cast_fp16)[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3679_cast_fp16 = add(x = var_3676_cast_fp16, y = var_3677_cast_fp16)[name = string("op_3679_cast_fp16")];
            int32 var_3681 = const()[name = string("op_3681"), val = int32(-1)];
            bool var_3682_interleave_0 = const()[name = string("op_3682_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_3682_cast_fp16 = concat(axis = var_3681, interleave = var_3682_interleave_0, values = (var_3656_cast_fp16, var_3679_cast_fp16))[name = string("op_3682_cast_fp16")];
            tensor<int32, [4]> z_53_begin_0 = const()[name = string("z_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_53_end_0 = const()[name = string("z_53_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_53_end_mask_0 = const()[name = string("z_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_53_cast_fp16 = slice_by_index(begin = z_53_begin_0, end = z_53_end_0, end_mask = z_53_end_mask_0, x = x_119_cast_fp16)[name = string("z_53_cast_fp16")];
            tensor<int32, [4]> var_3695_begin_0 = const()[name = string("op_3695_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3695_end_0 = const()[name = string("op_3695_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3695_end_mask_0 = const()[name = string("op_3695_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3695_cast_fp16 = slice_by_index(begin = var_3695_begin_0, end = var_3695_end_0, end_mask = var_3695_end_mask_0, x = z_53_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3696_cast_fp16 = mul(x = var_3695_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_3696_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, 1260, 12, 16])];
            tensor<bool, [4]> var_3701_end_mask_0 = const()[name = string("op_3701_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> 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 = z_53_cast_fp16)[name = string("op_3701_cast_fp16")];
            int32 var_3703 = const()[name = string("op_3703"), val = int32(-1)];
            bool rot_53_interleave_0 = const()[name = string("rot_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_53_cast_fp16 = concat(axis = var_3703, interleave = rot_53_interleave_0, values = (var_3696_cast_fp16, var_3701_cast_fp16))[name = string("rot_53_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3705_cast_fp16 = mul(x = z_53_cast_fp16, y = c_1_cast_fp16)[name = string("op_3705_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3706_cast_fp16 = mul(x = rot_53_cast_fp16, y = s_9_cast_fp16)[name = string("op_3706_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3708_cast_fp16 = add(x = var_3705_cast_fp16, y = var_3706_cast_fp16)[name = string("op_3708_cast_fp16")];
            tensor<int32, [4]> z_55_begin_0 = const()[name = string("z_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_55_end_0 = const()[name = string("z_55_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_55_end_mask_0 = const()[name = string("z_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_55_cast_fp16 = slice_by_index(begin = z_55_begin_0, end = z_55_end_0, end_mask = z_55_end_mask_0, x = x_119_cast_fp16)[name = string("z_55_cast_fp16")];
            tensor<int32, [4]> var_3718_begin_0 = const()[name = string("op_3718_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3718_end_0 = const()[name = string("op_3718_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3718_end_mask_0 = const()[name = string("op_3718_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3718_cast_fp16 = slice_by_index(begin = var_3718_begin_0, end = var_3718_end_0, end_mask = var_3718_end_mask_0, x = z_55_cast_fp16)[name = string("op_3718_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3719_cast_fp16 = mul(x = var_3718_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_3719_cast_fp16")];
            tensor<int32, [4]> var_3724_begin_0 = const()[name = string("op_3724_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3724_end_0 = const()[name = string("op_3724_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3724_end_mask_0 = const()[name = string("op_3724_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, x = z_55_cast_fp16)[name = string("op_3724_cast_fp16")];
            int32 var_3726 = const()[name = string("op_3726"), val = int32(-1)];
            bool rot_55_interleave_0 = const()[name = string("rot_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_55_cast_fp16 = concat(axis = var_3726, interleave = rot_55_interleave_0, values = (var_3719_cast_fp16, var_3724_cast_fp16))[name = string("rot_55_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3728_cast_fp16 = mul(x = z_55_cast_fp16, y = c_cast_fp16)[name = string("op_3728_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3729_cast_fp16 = mul(x = rot_55_cast_fp16, y = s_11_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3731_cast_fp16 = add(x = var_3728_cast_fp16, y = var_3729_cast_fp16)[name = string("op_3731_cast_fp16")];
            int32 var_3733 = const()[name = string("op_3733"), val = int32(-1)];
            bool var_3734_interleave_0 = const()[name = string("op_3734_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_3734_cast_fp16 = concat(axis = var_3733, interleave = var_3734_interleave_0, values = (var_3708_cast_fp16, var_3731_cast_fp16))[name = string("op_3734_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3744_transpose_x_0 = const()[name = string("op_3744_transpose_x_0"), val = bool(false)];
            bool var_3744_transpose_y_0 = const()[name = string("op_3744_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3734_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_76 = transpose(perm = transpose_76_perm_0, x = var_3682_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 12, 1260, 1260]> var_3744_cast_fp16 = matmul(transpose_x = var_3744_transpose_x_0, transpose_y = var_3744_transpose_y_0, x = transpose_76, y = transpose_77)[name = string("op_3744_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_3746_cast_fp16 = add(x = var_3744_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3746_cast_fp16")];
            int32 var_3747 = const()[name = string("op_3747"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_91_cast_fp16 = softmax(axis = var_3747, x = var_3746_cast_fp16)[name = string("w_91_cast_fp16")];
            bool var_3750_transpose_x_0 = const()[name = string("op_3750_transpose_x_0"), val = bool(false)];
            bool var_3750_transpose_y_0 = const()[name = string("op_3750_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = v_25_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 12, 1260, 64]> var_3750_cast_fp16 = matmul(transpose_x = var_3750_transpose_x_0, transpose_y = var_3750_transpose_y_0, x = w_91_cast_fp16, y = v_27_cast_fp16)[name = string("op_3750_cast_fp16")];
            tensor<int32, [4]> var_3753_perm_0 = const()[name = string("op_3753_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3757 = const()[name = string("op_3757"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_3753_cast_fp16 = transpose(perm = var_3753_perm_0, x = var_3750_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 1260, 768]> input_65_cast_fp16 = reshape(shape = var_3757, x = var_3753_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59164160))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59754048))))[name = string("t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3762_cast_fp16 = abs(x = linear_46_cast_fp16)[name = string("op_3762_cast_fp16")];
            bool var_3765 = const()[name = string("op_3765"), val = bool(true)];
            tensor<int32, [1]> var_3766_axes_0 = const()[name = string("op_3766_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3766_cast_fp16 = reduce_max(axes = var_3766_axes_0, keep_dims = var_3765, x = var_3762_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_97_cast_fp16 = maximum(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_93_cast_fp16 = real_div(x = linear_46_cast_fp16, y = s_97_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3770_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_3770_cast_fp16")];
            tensor<int32, [1]> var_3775_axes_0 = const()[name = string("op_3775_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3775_keep_dims_0 = const()[name = string("op_3775_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3775_cast_fp16 = reduce_mean(axes = var_3775_axes_0, keep_dims = var_3775_keep_dims_0, x = var_3770_cast_fp16)[name = string("op_3775_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3776_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_3776_cast_fp16")];
            fp32 var_3777_epsilon_0 = const()[name = string("op_3777_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3777_cast_fp16 = inverse(epsilon = var_3777_epsilon_0, x = var_3776_cast_fp16)[name = string("op_3777_cast_fp16")];
            fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3779_cast_fp16 = mul(x = var_3777_cast_fp16, y = var_3778_to_fp16)[name = string("op_3779_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3781_cast_fp16 = add(x = var_3775_cast_fp16, y = var_3779_cast_fp16)[name = string("op_3781_cast_fp16")];
            fp32 var_3782_epsilon_0 = const()[name = string("op_3782_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3782_cast_fp16 = rsqrt(epsilon = var_3782_epsilon_0, x = var_3781_cast_fp16)[name = string("op_3782_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_79_cast_fp16 = mul(x = y_93_cast_fp16, y = var_3782_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59755648)))];
            tensor<fp16, [1, 1260, 768]> var_3784_cast_fp16 = mul(x = out_79_cast_fp16, y = t_encoder_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_123_cast_fp16 = add(x = x_109_cast_fp16, y = var_3784_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3787_cast_fp16 = abs(x = x_123_cast_fp16)[name = string("op_3787_cast_fp16")];
            bool var_3790 = const()[name = string("op_3790"), val = bool(true)];
            tensor<int32, [1]> var_3791_axes_0 = const()[name = string("op_3791_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3791_cast_fp16 = reduce_max(axes = var_3791_axes_0, keep_dims = var_3790, x = var_3787_cast_fp16)[name = string("op_3791_cast_fp16")];
            fp16 var_3792_to_fp16 = const()[name = string("op_3792_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_99_cast_fp16 = maximum(x = var_3791_cast_fp16, y = var_3792_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_95_cast_fp16 = real_div(x = x_123_cast_fp16, y = s_99_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3795_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_3795_cast_fp16")];
            tensor<int32, [1]> var_3800_axes_0 = const()[name = string("op_3800_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3800_keep_dims_0 = const()[name = string("op_3800_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3800_cast_fp16 = reduce_mean(axes = var_3800_axes_0, keep_dims = var_3800_keep_dims_0, x = var_3795_cast_fp16)[name = string("op_3800_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3801_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_3801_cast_fp16")];
            fp32 var_3802_epsilon_0 = const()[name = string("op_3802_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3802_cast_fp16 = inverse(epsilon = var_3802_epsilon_0, x = var_3801_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 var_3803_to_fp16 = const()[name = string("op_3803_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16, y = var_3803_to_fp16)[name = string("op_3804_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3806_cast_fp16 = add(x = var_3800_cast_fp16, y = var_3804_cast_fp16)[name = string("op_3806_cast_fp16")];
            fp32 var_3807_epsilon_0 = const()[name = string("op_3807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3807_cast_fp16 = rsqrt(epsilon = var_3807_epsilon_0, x = var_3806_cast_fp16)[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_81_cast_fp16 = mul(x = y_95_cast_fp16, y = var_3807_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59757248)))];
            tensor<fp16, [1, 1260, 768]> input_67_cast_fp16 = mul(x = out_81_cast_fp16, y = t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59758848))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62118208))))[name = string("t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_47_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_47_cast_fp16")];
            string var_3814_mode_0 = const()[name = string("op_3814_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_3814_cast_fp16 = gelu(mode = var_3814_mode_0, x = linear_47_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62124416))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64483776))))[name = string("t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_48_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_71_cast_fp16 = mul(x = var_3814_cast_fp16, y = linear_48_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64489984))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66849344))))[name = string("t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3822_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_3822_cast_fp16")];
            bool var_3825 = const()[name = string("op_3825"), val = bool(true)];
            tensor<int32, [1]> var_3826_axes_0 = const()[name = string("op_3826_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3826_cast_fp16 = reduce_max(axes = var_3826_axes_0, keep_dims = var_3825, x = var_3822_cast_fp16)[name = string("op_3826_cast_fp16")];
            fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_101_cast_fp16 = maximum(x = var_3826_cast_fp16, y = var_3827_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_97_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_101_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3830_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_3830_cast_fp16")];
            tensor<int32, [1]> var_3835_axes_0 = const()[name = string("op_3835_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3835_keep_dims_0 = const()[name = string("op_3835_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3835_cast_fp16 = reduce_mean(axes = var_3835_axes_0, keep_dims = var_3835_keep_dims_0, x = var_3830_cast_fp16)[name = string("op_3835_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3836_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_3836_cast_fp16")];
            fp32 var_3837_epsilon_0 = const()[name = string("op_3837_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3837_cast_fp16 = inverse(epsilon = var_3837_epsilon_0, x = var_3836_cast_fp16)[name = string("op_3837_cast_fp16")];
            fp16 var_3838_to_fp16 = const()[name = string("op_3838_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3839_cast_fp16 = mul(x = var_3837_cast_fp16, y = var_3838_to_fp16)[name = string("op_3839_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3841_cast_fp16 = add(x = var_3835_cast_fp16, y = var_3839_cast_fp16)[name = string("op_3841_cast_fp16")];
            fp32 var_3842_epsilon_0 = const()[name = string("op_3842_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3842_cast_fp16 = rsqrt(epsilon = var_3842_epsilon_0, x = var_3841_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_83_cast_fp16 = mul(x = y_97_cast_fp16, y = var_3842_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66850944)))];
            tensor<fp16, [1, 1260, 768]> var_3844_cast_fp16 = mul(x = out_83_cast_fp16, y = t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3844_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_127_cast_fp16 = add(x = x_123_cast_fp16, y = var_3844_cast_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3847_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_3847_cast_fp16")];
            bool var_3850 = const()[name = string("op_3850"), val = bool(true)];
            tensor<int32, [1]> var_3851_axes_0 = const()[name = string("op_3851_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_3851_cast_fp16 = reduce_max(axes = var_3851_axes_0, keep_dims = var_3850, x = var_3847_cast_fp16)[name = string("op_3851_cast_fp16")];
            fp16 var_3852_to_fp16 = const()[name = string("op_3852_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_103_cast_fp16 = maximum(x = var_3851_cast_fp16, y = var_3852_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_99_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_103_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_3855_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_3855_cast_fp16")];
            tensor<int32, [1]> var_3860_axes_0 = const()[name = string("op_3860_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3860_keep_dims_0 = const()[name = string("op_3860_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_3860_cast_fp16 = reduce_mean(axes = var_3860_axes_0, keep_dims = var_3860_keep_dims_0, x = var_3855_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3861_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_3861_cast_fp16")];
            fp32 var_3862_epsilon_0 = const()[name = string("op_3862_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_3862_cast_fp16 = inverse(epsilon = var_3862_epsilon_0, x = var_3861_cast_fp16)[name = string("op_3862_cast_fp16")];
            fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_3864_cast_fp16 = mul(x = var_3862_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_3866_cast_fp16 = add(x = var_3860_cast_fp16, y = var_3864_cast_fp16)[name = string("op_3866_cast_fp16")];
            fp32 var_3867_epsilon_0 = const()[name = string("op_3867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_3867_cast_fp16 = rsqrt(epsilon = var_3867_epsilon_0, x = var_3866_cast_fp16)[name = string("op_3867_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_85_cast_fp16 = mul(x = y_99_cast_fp16, y = var_3867_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66852544)))];
            tensor<fp16, [1, 1260, 768]> input_73_cast_fp16 = mul(x = out_85_cast_fp16, y = t_encoder_layers_7_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66854144))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67444032))))[name = string("t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3877 = const()[name = string("op_3877"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_129_cast_fp16 = reshape(shape = var_3877, x = linear_50_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3879_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_3879_cast_fp16")];
            bool var_3882 = const()[name = string("op_3882"), val = bool(true)];
            tensor<int32, [1]> var_3883_axes_0 = const()[name = string("op_3883_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3883_cast_fp16 = reduce_max(axes = var_3883_axes_0, keep_dims = var_3882, x = var_3879_cast_fp16)[name = string("op_3883_cast_fp16")];
            fp16 var_3884_to_fp16 = const()[name = string("op_3884_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_105_cast_fp16 = maximum(x = var_3883_cast_fp16, y = var_3884_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_101_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_105_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3887_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_3887_cast_fp16")];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3892_keep_dims_0 = const()[name = string("op_3892_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3892_cast_fp16 = reduce_mean(axes = var_3892_axes_0, keep_dims = var_3892_keep_dims_0, x = var_3887_cast_fp16)[name = string("op_3892_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3893_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_3893_cast_fp16")];
            fp32 var_3894_epsilon_0 = const()[name = string("op_3894_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3894_cast_fp16 = inverse(epsilon = var_3894_epsilon_0, x = var_3893_cast_fp16)[name = string("op_3894_cast_fp16")];
            fp16 var_3895_to_fp16 = const()[name = string("op_3895_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3896_cast_fp16 = mul(x = var_3894_cast_fp16, y = var_3895_to_fp16)[name = string("op_3896_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3898_cast_fp16 = add(x = var_3892_cast_fp16, y = var_3896_cast_fp16)[name = string("op_3898_cast_fp16")];
            fp32 var_3899_epsilon_0 = const()[name = string("op_3899_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3899_cast_fp16 = rsqrt(epsilon = var_3899_epsilon_0, x = var_3898_cast_fp16)[name = string("op_3899_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_87_cast_fp16 = mul(x = y_101_cast_fp16, y = var_3899_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 1260, 12, 64]> x_135_cast_fp16 = mul(x = out_87_cast_fp16, y = t_encoder_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_135_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67445632))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68035520))))[name = string("t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_51_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3909 = const()[name = string("op_3909"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_131_cast_fp16 = reshape(shape = var_3909, x = linear_51_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3911_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_107_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_103_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_107_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3919_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3925_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_89_cast_fp16 = mul(x = y_103_cast_fp16, y = var_3931_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 1260, 12, 64]> x_137_cast_fp16 = mul(x = out_89_cast_fp16, y = t_encoder_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68037120))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68627008))))[name = string("t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [4]> var_3941 = const()[name = string("op_3941"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_133_cast_fp16 = reshape(shape = var_3941, x = linear_52_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3943_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_3943_cast_fp16")];
            bool var_3946 = const()[name = string("op_3946"), val = bool(true)];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_3947_cast_fp16 = reduce_max(axes = var_3947_axes_0, keep_dims = var_3946, x = var_3943_cast_fp16)[name = string("op_3947_cast_fp16")];
            fp16 var_3948_to_fp16 = const()[name = string("op_3948_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_109_cast_fp16 = maximum(x = var_3947_cast_fp16, y = var_3948_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_105_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_109_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_3951_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_3951_cast_fp16")];
            tensor<int32, [1]> var_3956_axes_0 = const()[name = string("op_3956_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3956_keep_dims_0 = const()[name = string("op_3956_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_3956_cast_fp16 = reduce_mean(axes = var_3956_axes_0, keep_dims = var_3956_keep_dims_0, x = var_3951_cast_fp16)[name = string("op_3956_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3957_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_3957_cast_fp16")];
            fp32 var_3958_epsilon_0 = const()[name = string("op_3958_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_3958_cast_fp16 = inverse(epsilon = var_3958_epsilon_0, x = var_3957_cast_fp16)[name = string("op_3958_cast_fp16")];
            fp16 var_3959_to_fp16 = const()[name = string("op_3959_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_3960_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = string("op_3960_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_3962_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3960_cast_fp16)[name = string("op_3962_cast_fp16")];
            fp32 var_3963_epsilon_0 = const()[name = string("op_3963_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_3963_cast_fp16 = rsqrt(epsilon = var_3963_epsilon_0, x = var_3962_cast_fp16)[name = string("op_3963_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_29_cast_fp16 = mul(x = y_105_cast_fp16, y = var_3963_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<int32, [4]> z_57_begin_0 = const()[name = string("z_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_57_end_0 = const()[name = string("z_57_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_57_end_mask_0 = const()[name = string("z_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_57_cast_fp16 = slice_by_index(begin = z_57_begin_0, end = z_57_end_0, end_mask = z_57_end_mask_0, x = x_135_cast_fp16)[name = string("z_57_cast_fp16")];
            tensor<int32, [4]> var_3974_begin_0 = const()[name = string("op_3974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3974_end_0 = const()[name = string("op_3974_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3974_end_mask_0 = const()[name = string("op_3974_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> 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 = z_57_cast_fp16)[name = string("op_3974_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3975_cast_fp16 = mul(x = var_3974_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_3975_cast_fp16")];
            tensor<int32, [4]> var_3980_begin_0 = const()[name = string("op_3980_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3980_end_0 = const()[name = string("op_3980_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_3980_end_mask_0 = const()[name = string("op_3980_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_3980_cast_fp16 = slice_by_index(begin = var_3980_begin_0, end = var_3980_end_0, end_mask = var_3980_end_mask_0, x = z_57_cast_fp16)[name = string("op_3980_cast_fp16")];
            int32 var_3982 = const()[name = string("op_3982"), val = int32(-1)];
            bool rot_57_interleave_0 = const()[name = string("rot_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_57_cast_fp16 = concat(axis = var_3982, interleave = rot_57_interleave_0, values = (var_3975_cast_fp16, var_3980_cast_fp16))[name = string("rot_57_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3984_cast_fp16 = mul(x = z_57_cast_fp16, y = c_1_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3985_cast_fp16 = mul(x = rot_57_cast_fp16, y = s_9_cast_fp16)[name = string("op_3985_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_3987_cast_fp16 = add(x = var_3984_cast_fp16, y = var_3985_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<int32, [4]> z_59_begin_0 = const()[name = string("z_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_59_end_0 = const()[name = string("z_59_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_59_end_mask_0 = const()[name = string("z_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_59_cast_fp16 = slice_by_index(begin = z_59_begin_0, end = z_59_end_0, end_mask = z_59_end_mask_0, x = x_135_cast_fp16)[name = string("z_59_cast_fp16")];
            tensor<int32, [4]> var_3997_begin_0 = const()[name = string("op_3997_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3997_end_0 = const()[name = string("op_3997_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_3997_end_mask_0 = const()[name = string("op_3997_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, x = z_59_cast_fp16)[name = string("op_3997_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_3998_cast_fp16 = mul(x = var_3997_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_3998_cast_fp16")];
            tensor<int32, [4]> var_4003_begin_0 = const()[name = string("op_4003_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4003_end_0 = const()[name = string("op_4003_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4003_end_mask_0 = const()[name = string("op_4003_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4003_cast_fp16 = slice_by_index(begin = var_4003_begin_0, end = var_4003_end_0, end_mask = var_4003_end_mask_0, x = z_59_cast_fp16)[name = string("op_4003_cast_fp16")];
            int32 var_4005 = const()[name = string("op_4005"), val = int32(-1)];
            bool rot_59_interleave_0 = const()[name = string("rot_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_59_cast_fp16 = concat(axis = var_4005, interleave = rot_59_interleave_0, values = (var_3998_cast_fp16, var_4003_cast_fp16))[name = string("rot_59_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4007_cast_fp16 = mul(x = z_59_cast_fp16, y = c_cast_fp16)[name = string("op_4007_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4008_cast_fp16 = mul(x = rot_59_cast_fp16, y = s_11_cast_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4010_cast_fp16 = add(x = var_4007_cast_fp16, y = var_4008_cast_fp16)[name = string("op_4010_cast_fp16")];
            int32 var_4012 = const()[name = string("op_4012"), val = int32(-1)];
            bool var_4013_interleave_0 = const()[name = string("op_4013_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_4013_cast_fp16 = concat(axis = var_4012, interleave = var_4013_interleave_0, values = (var_3987_cast_fp16, var_4010_cast_fp16))[name = string("op_4013_cast_fp16")];
            tensor<int32, [4]> z_61_begin_0 = const()[name = string("z_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_61_end_0 = const()[name = string("z_61_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_61_end_mask_0 = const()[name = string("z_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_61_cast_fp16 = slice_by_index(begin = z_61_begin_0, end = z_61_end_0, end_mask = z_61_end_mask_0, x = x_137_cast_fp16)[name = string("z_61_cast_fp16")];
            tensor<int32, [4]> var_4026_begin_0 = const()[name = string("op_4026_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4026_end_0 = const()[name = string("op_4026_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4026_end_mask_0 = const()[name = string("op_4026_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4026_cast_fp16 = slice_by_index(begin = var_4026_begin_0, end = var_4026_end_0, end_mask = var_4026_end_mask_0, x = z_61_cast_fp16)[name = string("op_4026_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_4027_cast_fp16")];
            tensor<int32, [4]> var_4032_begin_0 = const()[name = string("op_4032_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4032_end_0 = const()[name = string("op_4032_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4032_end_mask_0 = const()[name = string("op_4032_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4032_cast_fp16 = slice_by_index(begin = var_4032_begin_0, end = var_4032_end_0, end_mask = var_4032_end_mask_0, x = z_61_cast_fp16)[name = string("op_4032_cast_fp16")];
            int32 var_4034 = const()[name = string("op_4034"), val = int32(-1)];
            bool rot_61_interleave_0 = const()[name = string("rot_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_61_cast_fp16 = concat(axis = var_4034, interleave = rot_61_interleave_0, values = (var_4027_cast_fp16, var_4032_cast_fp16))[name = string("rot_61_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4036_cast_fp16 = mul(x = z_61_cast_fp16, y = c_1_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4037_cast_fp16 = mul(x = rot_61_cast_fp16, y = s_9_cast_fp16)[name = string("op_4037_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4039_cast_fp16 = add(x = var_4036_cast_fp16, y = var_4037_cast_fp16)[name = string("op_4039_cast_fp16")];
            tensor<int32, [4]> z_63_begin_0 = const()[name = string("z_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_63_end_0 = const()[name = string("z_63_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_63_end_mask_0 = const()[name = string("z_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_63_cast_fp16 = slice_by_index(begin = z_63_begin_0, end = z_63_end_0, end_mask = z_63_end_mask_0, x = x_137_cast_fp16)[name = string("z_63_cast_fp16")];
            tensor<int32, [4]> var_4049_begin_0 = const()[name = string("op_4049_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4049_end_0 = const()[name = string("op_4049_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4049_end_mask_0 = const()[name = string("op_4049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4049_cast_fp16 = slice_by_index(begin = var_4049_begin_0, end = var_4049_end_0, end_mask = var_4049_end_mask_0, x = z_63_cast_fp16)[name = string("op_4049_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4050_cast_fp16 = mul(x = var_4049_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_4050_cast_fp16")];
            tensor<int32, [4]> var_4055_begin_0 = const()[name = string("op_4055_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4055_end_0 = const()[name = string("op_4055_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4055_end_mask_0 = const()[name = string("op_4055_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4055_cast_fp16 = slice_by_index(begin = var_4055_begin_0, end = var_4055_end_0, end_mask = var_4055_end_mask_0, x = z_63_cast_fp16)[name = string("op_4055_cast_fp16")];
            int32 var_4057 = const()[name = string("op_4057"), val = int32(-1)];
            bool rot_63_interleave_0 = const()[name = string("rot_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_63_cast_fp16 = concat(axis = var_4057, interleave = rot_63_interleave_0, values = (var_4050_cast_fp16, var_4055_cast_fp16))[name = string("rot_63_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4059_cast_fp16 = mul(x = z_63_cast_fp16, y = c_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4060_cast_fp16 = mul(x = rot_63_cast_fp16, y = s_11_cast_fp16)[name = string("op_4060_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4062_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4060_cast_fp16)[name = string("op_4062_cast_fp16")];
            int32 var_4064 = const()[name = string("op_4064"), val = int32(-1)];
            bool var_4065_interleave_0 = const()[name = string("op_4065_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_4065_cast_fp16 = concat(axis = var_4064, interleave = var_4065_interleave_0, values = (var_4039_cast_fp16, var_4062_cast_fp16))[name = string("op_4065_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4075_transpose_x_0 = const()[name = string("op_4075_transpose_x_0"), val = bool(false)];
            bool var_4075_transpose_y_0 = const()[name = string("op_4075_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_4065_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_78 = transpose(perm = transpose_78_perm_0, x = var_4013_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 12, 1260, 1260]> var_4075_cast_fp16 = matmul(transpose_x = var_4075_transpose_x_0, transpose_y = var_4075_transpose_y_0, x = transpose_78, y = transpose_79)[name = string("op_4075_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_4077_cast_fp16 = add(x = var_4075_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4077_cast_fp16")];
            int32 var_4078 = const()[name = string("op_4078"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_105_cast_fp16 = softmax(axis = var_4078, x = var_4077_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4081_transpose_x_0 = const()[name = string("op_4081_transpose_x_0"), val = bool(false)];
            bool var_4081_transpose_y_0 = const()[name = string("op_4081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = v_29_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 12, 1260, 64]> var_4081_cast_fp16 = matmul(transpose_x = var_4081_transpose_x_0, transpose_y = var_4081_transpose_y_0, x = w_105_cast_fp16, y = v_31_cast_fp16)[name = string("op_4081_cast_fp16")];
            tensor<int32, [4]> var_4084_perm_0 = const()[name = string("op_4084_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4088 = const()[name = string("op_4088"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_4084_cast_fp16 = transpose(perm = var_4084_perm_0, x = var_4081_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 1260, 768]> input_75_cast_fp16 = reshape(shape = var_4088, x = var_4084_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68628608))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69218496))))[name = string("t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_53_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4093_cast_fp16 = abs(x = linear_53_cast_fp16)[name = string("op_4093_cast_fp16")];
            bool var_4096 = const()[name = string("op_4096"), val = bool(true)];
            tensor<int32, [1]> var_4097_axes_0 = const()[name = string("op_4097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4097_cast_fp16 = reduce_max(axes = var_4097_axes_0, keep_dims = var_4096, x = var_4093_cast_fp16)[name = string("op_4097_cast_fp16")];
            fp16 var_4098_to_fp16 = const()[name = string("op_4098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_111_cast_fp16 = maximum(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_107_cast_fp16 = real_div(x = linear_53_cast_fp16, y = s_111_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4101_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<int32, [1]> var_4106_axes_0 = const()[name = string("op_4106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4106_keep_dims_0 = const()[name = string("op_4106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4106_cast_fp16 = reduce_mean(axes = var_4106_axes_0, keep_dims = var_4106_keep_dims_0, x = var_4101_cast_fp16)[name = string("op_4106_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4107_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4107_cast_fp16")];
            fp32 var_4108_epsilon_0 = const()[name = string("op_4108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4108_cast_fp16 = inverse(epsilon = var_4108_epsilon_0, x = var_4107_cast_fp16)[name = string("op_4108_cast_fp16")];
            fp16 var_4109_to_fp16 = const()[name = string("op_4109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4110_cast_fp16 = mul(x = var_4108_cast_fp16, y = var_4109_to_fp16)[name = string("op_4110_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4112_cast_fp16 = add(x = var_4106_cast_fp16, y = var_4110_cast_fp16)[name = string("op_4112_cast_fp16")];
            fp32 var_4113_epsilon_0 = const()[name = string("op_4113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4113_cast_fp16 = rsqrt(epsilon = var_4113_epsilon_0, x = var_4112_cast_fp16)[name = string("op_4113_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_91_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4113_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69220096)))];
            tensor<fp16, [1, 1260, 768]> var_4115_cast_fp16 = mul(x = out_91_cast_fp16, y = t_encoder_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4115_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_141_cast_fp16 = add(x = x_127_cast_fp16, y = var_4115_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4118_cast_fp16 = abs(x = x_141_cast_fp16)[name = string("op_4118_cast_fp16")];
            bool var_4121 = const()[name = string("op_4121"), val = bool(true)];
            tensor<int32, [1]> var_4122_axes_0 = const()[name = string("op_4122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4122_cast_fp16 = reduce_max(axes = var_4122_axes_0, keep_dims = var_4121, x = var_4118_cast_fp16)[name = string("op_4122_cast_fp16")];
            fp16 var_4123_to_fp16 = const()[name = string("op_4123_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_113_cast_fp16 = maximum(x = var_4122_cast_fp16, y = var_4123_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_109_cast_fp16 = real_div(x = x_141_cast_fp16, y = s_113_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4126_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4126_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, 1260, 1]> var_4131_cast_fp16 = reduce_mean(axes = var_4131_axes_0, keep_dims = var_4131_keep_dims_0, x = var_4126_cast_fp16)[name = string("op_4131_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4132_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_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, 1260, 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, 1260, 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, 1260, 1]> var_4137_cast_fp16 = add(x = var_4131_cast_fp16, y = var_4135_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp32 var_4138_epsilon_0 = const()[name = string("op_4138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4138_cast_fp16 = rsqrt(epsilon = var_4138_epsilon_0, x = var_4137_cast_fp16)[name = string("op_4138_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_93_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4138_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69221696)))];
            tensor<fp16, [1, 1260, 768]> input_77_cast_fp16 = mul(x = out_93_cast_fp16, y = t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69223296))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71582656))))[name = string("t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_54_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_54_cast_fp16")];
            string var_4145_mode_0 = const()[name = string("op_4145_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_4145_cast_fp16 = gelu(mode = var_4145_mode_0, x = linear_54_cast_fp16)[name = string("op_4145_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71588864))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73948224))))[name = string("t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_55_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_81_cast_fp16 = mul(x = var_4145_cast_fp16, y = linear_55_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73954432))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76313792))))[name = string("t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4153_cast_fp16 = abs(x = linear_56_cast_fp16)[name = string("op_4153_cast_fp16")];
            bool var_4156 = const()[name = string("op_4156"), val = bool(true)];
            tensor<int32, [1]> var_4157_axes_0 = const()[name = string("op_4157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4157_cast_fp16 = reduce_max(axes = var_4157_axes_0, keep_dims = var_4156, x = var_4153_cast_fp16)[name = string("op_4157_cast_fp16")];
            fp16 var_4158_to_fp16 = const()[name = string("op_4158_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_115_cast_fp16 = maximum(x = var_4157_cast_fp16, y = var_4158_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_111_cast_fp16 = real_div(x = linear_56_cast_fp16, y = s_115_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4161_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4161_cast_fp16")];
            tensor<int32, [1]> var_4166_axes_0 = const()[name = string("op_4166_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4166_keep_dims_0 = const()[name = string("op_4166_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4166_cast_fp16 = reduce_mean(axes = var_4166_axes_0, keep_dims = var_4166_keep_dims_0, x = var_4161_cast_fp16)[name = string("op_4166_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4167_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4167_cast_fp16")];
            fp32 var_4168_epsilon_0 = const()[name = string("op_4168_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4168_cast_fp16 = inverse(epsilon = var_4168_epsilon_0, x = var_4167_cast_fp16)[name = string("op_4168_cast_fp16")];
            fp16 var_4169_to_fp16 = const()[name = string("op_4169_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4170_cast_fp16 = mul(x = var_4168_cast_fp16, y = var_4169_to_fp16)[name = string("op_4170_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4172_cast_fp16 = add(x = var_4166_cast_fp16, y = var_4170_cast_fp16)[name = string("op_4172_cast_fp16")];
            fp32 var_4173_epsilon_0 = const()[name = string("op_4173_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4173_cast_fp16 = rsqrt(epsilon = var_4173_epsilon_0, x = var_4172_cast_fp16)[name = string("op_4173_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_95_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4173_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76315392)))];
            tensor<fp16, [1, 1260, 768]> var_4175_cast_fp16 = mul(x = out_95_cast_fp16, y = t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4175_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_145_cast_fp16 = add(x = x_141_cast_fp16, y = var_4175_cast_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4178_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4178_cast_fp16")];
            bool var_4181 = const()[name = string("op_4181"), val = bool(true)];
            tensor<int32, [1]> var_4182_axes_0 = const()[name = string("op_4182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4182_cast_fp16 = reduce_max(axes = var_4182_axes_0, keep_dims = var_4181, x = var_4178_cast_fp16)[name = string("op_4182_cast_fp16")];
            fp16 var_4183_to_fp16 = const()[name = string("op_4183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_117_cast_fp16 = maximum(x = var_4182_cast_fp16, y = var_4183_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_113_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_117_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4186_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4186_cast_fp16")];
            tensor<int32, [1]> var_4191_axes_0 = const()[name = string("op_4191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4191_keep_dims_0 = const()[name = string("op_4191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4191_cast_fp16 = reduce_mean(axes = var_4191_axes_0, keep_dims = var_4191_keep_dims_0, x = var_4186_cast_fp16)[name = string("op_4191_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4192_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4192_cast_fp16")];
            fp32 var_4193_epsilon_0 = const()[name = string("op_4193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4193_cast_fp16 = inverse(epsilon = var_4193_epsilon_0, x = var_4192_cast_fp16)[name = string("op_4193_cast_fp16")];
            fp16 var_4194_to_fp16 = const()[name = string("op_4194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4195_cast_fp16 = mul(x = var_4193_cast_fp16, y = var_4194_to_fp16)[name = string("op_4195_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4197_cast_fp16 = add(x = var_4191_cast_fp16, y = var_4195_cast_fp16)[name = string("op_4197_cast_fp16")];
            fp32 var_4198_epsilon_0 = const()[name = string("op_4198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4198_cast_fp16 = rsqrt(epsilon = var_4198_epsilon_0, x = var_4197_cast_fp16)[name = string("op_4198_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_97_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4198_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76316992)))];
            tensor<fp16, [1, 1260, 768]> input_83_cast_fp16 = mul(x = out_97_cast_fp16, y = t_encoder_layers_8_input_layernorm_weight_to_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76318592))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76908480))))[name = string("t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_57_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4208 = const()[name = string("op_4208"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_147_cast_fp16 = reshape(shape = var_4208, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4210_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4210_cast_fp16")];
            bool var_4213 = const()[name = string("op_4213"), val = bool(true)];
            tensor<int32, [1]> var_4214_axes_0 = const()[name = string("op_4214_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4214_cast_fp16 = reduce_max(axes = var_4214_axes_0, keep_dims = var_4213, x = var_4210_cast_fp16)[name = string("op_4214_cast_fp16")];
            fp16 var_4215_to_fp16 = const()[name = string("op_4215_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_119_cast_fp16 = maximum(x = var_4214_cast_fp16, y = var_4215_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_115_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_119_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4218_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4218_cast_fp16")];
            tensor<int32, [1]> var_4223_axes_0 = const()[name = string("op_4223_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4223_keep_dims_0 = const()[name = string("op_4223_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4223_cast_fp16 = reduce_mean(axes = var_4223_axes_0, keep_dims = var_4223_keep_dims_0, x = var_4218_cast_fp16)[name = string("op_4223_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4224_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4224_cast_fp16")];
            fp32 var_4225_epsilon_0 = const()[name = string("op_4225_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4225_cast_fp16 = inverse(epsilon = var_4225_epsilon_0, x = var_4224_cast_fp16)[name = string("op_4225_cast_fp16")];
            fp16 var_4226_to_fp16 = const()[name = string("op_4226_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4227_cast_fp16 = mul(x = var_4225_cast_fp16, y = var_4226_to_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4229_cast_fp16 = add(x = var_4223_cast_fp16, y = var_4227_cast_fp16)[name = string("op_4229_cast_fp16")];
            fp32 var_4230_epsilon_0 = const()[name = string("op_4230_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4230_cast_fp16 = rsqrt(epsilon = var_4230_epsilon_0, x = var_4229_cast_fp16)[name = string("op_4230_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_99_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4230_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910080)))];
            tensor<fp16, [1, 1260, 12, 64]> x_153_cast_fp16 = mul(x = out_99_cast_fp16, y = t_encoder_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910272))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77500160))))[name = string("t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_4240 = const()[name = string("op_4240"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_149_cast_fp16 = reshape(shape = var_4240, x = linear_58_cast_fp16)[name = string("x_149_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4242_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4242_cast_fp16")];
            bool var_4245 = const()[name = string("op_4245"), val = bool(true)];
            tensor<int32, [1]> var_4246_axes_0 = const()[name = string("op_4246_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4246_cast_fp16 = reduce_max(axes = var_4246_axes_0, keep_dims = var_4245, x = var_4242_cast_fp16)[name = string("op_4246_cast_fp16")];
            fp16 var_4247_to_fp16 = const()[name = string("op_4247_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_121_cast_fp16 = maximum(x = var_4246_cast_fp16, y = var_4247_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_117_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_121_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4250_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<int32, [1]> var_4255_axes_0 = const()[name = string("op_4255_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4255_keep_dims_0 = const()[name = string("op_4255_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4255_cast_fp16 = reduce_mean(axes = var_4255_axes_0, keep_dims = var_4255_keep_dims_0, x = var_4250_cast_fp16)[name = string("op_4255_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4256_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp32 var_4257_epsilon_0 = const()[name = string("op_4257_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4257_cast_fp16 = inverse(epsilon = var_4257_epsilon_0, x = var_4256_cast_fp16)[name = string("op_4257_cast_fp16")];
            fp16 var_4258_to_fp16 = const()[name = string("op_4258_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4259_cast_fp16 = mul(x = var_4257_cast_fp16, y = var_4258_to_fp16)[name = string("op_4259_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4261_cast_fp16 = add(x = var_4255_cast_fp16, y = var_4259_cast_fp16)[name = string("op_4261_cast_fp16")];
            fp32 var_4262_epsilon_0 = const()[name = string("op_4262_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4262_cast_fp16 = rsqrt(epsilon = var_4262_epsilon_0, x = var_4261_cast_fp16)[name = string("op_4262_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_101_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4262_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501760)))];
            tensor<fp16, [1, 1260, 12, 64]> x_155_cast_fp16 = mul(x = out_101_cast_fp16, y = t_encoder_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501952))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78091840))))[name = string("t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_59_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<int32, [4]> var_4272 = const()[name = string("op_4272"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_151_cast_fp16 = reshape(shape = var_4272, x = linear_59_cast_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4274_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4274_cast_fp16")];
            bool var_4277 = const()[name = string("op_4277"), val = bool(true)];
            tensor<int32, [1]> var_4278_axes_0 = const()[name = string("op_4278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4278_cast_fp16 = reduce_max(axes = var_4278_axes_0, keep_dims = var_4277, x = var_4274_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp16 var_4279_to_fp16 = const()[name = string("op_4279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_123_cast_fp16 = maximum(x = var_4278_cast_fp16, y = var_4279_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_119_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_123_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4282_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<int32, [1]> var_4287_axes_0 = const()[name = string("op_4287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4287_keep_dims_0 = const()[name = string("op_4287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4287_cast_fp16 = reduce_mean(axes = var_4287_axes_0, keep_dims = var_4287_keep_dims_0, x = var_4282_cast_fp16)[name = string("op_4287_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4288_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4288_cast_fp16")];
            fp32 var_4289_epsilon_0 = const()[name = string("op_4289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4289_cast_fp16 = inverse(epsilon = var_4289_epsilon_0, x = var_4288_cast_fp16)[name = string("op_4289_cast_fp16")];
            fp16 var_4290_to_fp16 = const()[name = string("op_4290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4291_cast_fp16 = mul(x = var_4289_cast_fp16, y = var_4290_to_fp16)[name = string("op_4291_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4293_cast_fp16 = add(x = var_4287_cast_fp16, y = var_4291_cast_fp16)[name = string("op_4293_cast_fp16")];
            fp32 var_4294_epsilon_0 = const()[name = string("op_4294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4294_cast_fp16 = rsqrt(epsilon = var_4294_epsilon_0, x = var_4293_cast_fp16)[name = string("op_4294_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_33_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4294_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<int32, [4]> z_65_begin_0 = const()[name = string("z_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_65_end_0 = const()[name = string("z_65_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_65_end_mask_0 = const()[name = string("z_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_65_cast_fp16 = slice_by_index(begin = z_65_begin_0, end = z_65_end_0, end_mask = z_65_end_mask_0, x = x_153_cast_fp16)[name = string("z_65_cast_fp16")];
            tensor<int32, [4]> var_4305_begin_0 = const()[name = string("op_4305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4305_end_0 = const()[name = string("op_4305_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4305_end_mask_0 = const()[name = string("op_4305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4305_cast_fp16 = slice_by_index(begin = var_4305_begin_0, end = var_4305_end_0, end_mask = var_4305_end_mask_0, x = z_65_cast_fp16)[name = string("op_4305_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4306_cast_fp16 = mul(x = var_4305_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_4306_cast_fp16")];
            tensor<int32, [4]> var_4311_begin_0 = const()[name = string("op_4311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4311_end_0 = const()[name = string("op_4311_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4311_end_mask_0 = const()[name = string("op_4311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4311_cast_fp16 = slice_by_index(begin = var_4311_begin_0, end = var_4311_end_0, end_mask = var_4311_end_mask_0, x = z_65_cast_fp16)[name = string("op_4311_cast_fp16")];
            int32 var_4313 = const()[name = string("op_4313"), val = int32(-1)];
            bool rot_65_interleave_0 = const()[name = string("rot_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_65_cast_fp16 = concat(axis = var_4313, interleave = rot_65_interleave_0, values = (var_4306_cast_fp16, var_4311_cast_fp16))[name = string("rot_65_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4315_cast_fp16 = mul(x = z_65_cast_fp16, y = c_1_cast_fp16)[name = string("op_4315_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4316_cast_fp16 = mul(x = rot_65_cast_fp16, y = s_9_cast_fp16)[name = string("op_4316_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4318_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4316_cast_fp16)[name = string("op_4318_cast_fp16")];
            tensor<int32, [4]> z_67_begin_0 = const()[name = string("z_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_67_end_0 = const()[name = string("z_67_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_67_end_mask_0 = const()[name = string("z_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_67_cast_fp16 = slice_by_index(begin = z_67_begin_0, end = z_67_end_0, end_mask = z_67_end_mask_0, x = x_153_cast_fp16)[name = string("z_67_cast_fp16")];
            tensor<int32, [4]> var_4328_begin_0 = const()[name = string("op_4328_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4328_end_0 = const()[name = string("op_4328_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4328_end_mask_0 = const()[name = string("op_4328_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4328_cast_fp16 = slice_by_index(begin = var_4328_begin_0, end = var_4328_end_0, end_mask = var_4328_end_mask_0, x = z_67_cast_fp16)[name = string("op_4328_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4329_cast_fp16 = mul(x = var_4328_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_4329_cast_fp16")];
            tensor<int32, [4]> var_4334_begin_0 = const()[name = string("op_4334_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4334_end_0 = const()[name = string("op_4334_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4334_end_mask_0 = const()[name = string("op_4334_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4334_cast_fp16 = slice_by_index(begin = var_4334_begin_0, end = var_4334_end_0, end_mask = var_4334_end_mask_0, x = z_67_cast_fp16)[name = string("op_4334_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            bool rot_67_interleave_0 = const()[name = string("rot_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_67_cast_fp16 = concat(axis = var_4336, interleave = rot_67_interleave_0, values = (var_4329_cast_fp16, var_4334_cast_fp16))[name = string("rot_67_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4338_cast_fp16 = mul(x = z_67_cast_fp16, y = c_cast_fp16)[name = string("op_4338_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4339_cast_fp16 = mul(x = rot_67_cast_fp16, y = s_11_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4341_cast_fp16 = add(x = var_4338_cast_fp16, y = var_4339_cast_fp16)[name = string("op_4341_cast_fp16")];
            int32 var_4343 = const()[name = string("op_4343"), val = int32(-1)];
            bool var_4344_interleave_0 = const()[name = string("op_4344_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_4344_cast_fp16 = concat(axis = var_4343, interleave = var_4344_interleave_0, values = (var_4318_cast_fp16, var_4341_cast_fp16))[name = string("op_4344_cast_fp16")];
            tensor<int32, [4]> z_69_begin_0 = const()[name = string("z_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_69_end_0 = const()[name = string("z_69_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_69_end_mask_0 = const()[name = string("z_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_69_cast_fp16 = slice_by_index(begin = z_69_begin_0, end = z_69_end_0, end_mask = z_69_end_mask_0, x = x_155_cast_fp16)[name = string("z_69_cast_fp16")];
            tensor<int32, [4]> var_4357_begin_0 = const()[name = string("op_4357_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4357_end_0 = const()[name = string("op_4357_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4357_end_mask_0 = const()[name = string("op_4357_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4357_cast_fp16 = slice_by_index(begin = var_4357_begin_0, end = var_4357_end_0, end_mask = var_4357_end_mask_0, x = z_69_cast_fp16)[name = string("op_4357_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4358_cast_fp16 = mul(x = var_4357_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_4358_cast_fp16")];
            tensor<int32, [4]> var_4363_begin_0 = const()[name = string("op_4363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4363_end_0 = const()[name = string("op_4363_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4363_end_mask_0 = const()[name = string("op_4363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4363_cast_fp16 = slice_by_index(begin = var_4363_begin_0, end = var_4363_end_0, end_mask = var_4363_end_mask_0, x = z_69_cast_fp16)[name = string("op_4363_cast_fp16")];
            int32 var_4365 = const()[name = string("op_4365"), val = int32(-1)];
            bool rot_69_interleave_0 = const()[name = string("rot_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_69_cast_fp16 = concat(axis = var_4365, interleave = rot_69_interleave_0, values = (var_4358_cast_fp16, var_4363_cast_fp16))[name = string("rot_69_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4367_cast_fp16 = mul(x = z_69_cast_fp16, y = c_1_cast_fp16)[name = string("op_4367_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4368_cast_fp16 = mul(x = rot_69_cast_fp16, y = s_9_cast_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4370_cast_fp16 = add(x = var_4367_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            tensor<int32, [4]> z_71_begin_0 = const()[name = string("z_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_71_end_0 = const()[name = string("z_71_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_71_end_mask_0 = const()[name = string("z_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_71_cast_fp16 = slice_by_index(begin = z_71_begin_0, end = z_71_end_0, end_mask = z_71_end_mask_0, x = x_155_cast_fp16)[name = string("z_71_cast_fp16")];
            tensor<int32, [4]> var_4380_begin_0 = const()[name = string("op_4380_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4380_end_0 = const()[name = string("op_4380_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4380_end_mask_0 = const()[name = string("op_4380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4380_cast_fp16 = slice_by_index(begin = var_4380_begin_0, end = var_4380_end_0, end_mask = var_4380_end_mask_0, x = z_71_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4381_cast_fp16 = mul(x = var_4380_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_4381_cast_fp16")];
            tensor<int32, [4]> var_4386_begin_0 = const()[name = string("op_4386_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4386_end_0 = const()[name = string("op_4386_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4386_end_mask_0 = const()[name = string("op_4386_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4386_cast_fp16 = slice_by_index(begin = var_4386_begin_0, end = var_4386_end_0, end_mask = var_4386_end_mask_0, x = z_71_cast_fp16)[name = string("op_4386_cast_fp16")];
            int32 var_4388 = const()[name = string("op_4388"), val = int32(-1)];
            bool rot_71_interleave_0 = const()[name = string("rot_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_71_cast_fp16 = concat(axis = var_4388, interleave = rot_71_interleave_0, values = (var_4381_cast_fp16, var_4386_cast_fp16))[name = string("rot_71_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4390_cast_fp16 = mul(x = z_71_cast_fp16, y = c_cast_fp16)[name = string("op_4390_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4391_cast_fp16 = mul(x = rot_71_cast_fp16, y = s_11_cast_fp16)[name = string("op_4391_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4393_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4391_cast_fp16)[name = string("op_4393_cast_fp16")];
            int32 var_4395 = const()[name = string("op_4395"), val = int32(-1)];
            bool var_4396_interleave_0 = const()[name = string("op_4396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_4396_cast_fp16 = concat(axis = var_4395, interleave = var_4396_interleave_0, values = (var_4370_cast_fp16, var_4393_cast_fp16))[name = string("op_4396_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4406_transpose_x_0 = const()[name = string("op_4406_transpose_x_0"), val = bool(false)];
            bool var_4406_transpose_y_0 = const()[name = string("op_4406_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_4396_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_80 = transpose(perm = transpose_80_perm_0, x = var_4344_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 12, 1260, 1260]> var_4406_cast_fp16 = matmul(transpose_x = var_4406_transpose_x_0, transpose_y = var_4406_transpose_y_0, x = transpose_80, y = transpose_81)[name = string("op_4406_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_4408_cast_fp16 = add(x = var_4406_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4408_cast_fp16")];
            int32 var_4409 = const()[name = string("op_4409"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_119_cast_fp16 = softmax(axis = var_4409, x = var_4408_cast_fp16)[name = string("w_119_cast_fp16")];
            bool var_4412_transpose_x_0 = const()[name = string("op_4412_transpose_x_0"), val = bool(false)];
            bool var_4412_transpose_y_0 = const()[name = string("op_4412_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = v_33_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 12, 1260, 64]> var_4412_cast_fp16 = matmul(transpose_x = var_4412_transpose_x_0, transpose_y = var_4412_transpose_y_0, x = w_119_cast_fp16, y = v_35_cast_fp16)[name = string("op_4412_cast_fp16")];
            tensor<int32, [4]> var_4415_perm_0 = const()[name = string("op_4415_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4419 = const()[name = string("op_4419"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_4415_cast_fp16 = transpose(perm = var_4415_perm_0, x = var_4412_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 1260, 768]> input_85_cast_fp16 = reshape(shape = var_4419, x = var_4415_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78093440))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78683328))))[name = string("t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4424_cast_fp16 = abs(x = linear_60_cast_fp16)[name = string("op_4424_cast_fp16")];
            bool var_4427 = const()[name = string("op_4427"), val = bool(true)];
            tensor<int32, [1]> var_4428_axes_0 = const()[name = string("op_4428_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4428_cast_fp16 = reduce_max(axes = var_4428_axes_0, keep_dims = var_4427, x = var_4424_cast_fp16)[name = string("op_4428_cast_fp16")];
            fp16 var_4429_to_fp16 = const()[name = string("op_4429_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_125_cast_fp16 = maximum(x = var_4428_cast_fp16, y = var_4429_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_121_cast_fp16 = real_div(x = linear_60_cast_fp16, y = s_125_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4432_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [1]> var_4437_axes_0 = const()[name = string("op_4437_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4437_keep_dims_0 = const()[name = string("op_4437_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4437_cast_fp16 = reduce_mean(axes = var_4437_axes_0, keep_dims = var_4437_keep_dims_0, x = var_4432_cast_fp16)[name = string("op_4437_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4438_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4438_cast_fp16")];
            fp32 var_4439_epsilon_0 = const()[name = string("op_4439_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4439_cast_fp16 = inverse(epsilon = var_4439_epsilon_0, x = var_4438_cast_fp16)[name = string("op_4439_cast_fp16")];
            fp16 var_4440_to_fp16 = const()[name = string("op_4440_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4441_cast_fp16 = mul(x = var_4439_cast_fp16, y = var_4440_to_fp16)[name = string("op_4441_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4443_cast_fp16 = add(x = var_4437_cast_fp16, y = var_4441_cast_fp16)[name = string("op_4443_cast_fp16")];
            fp32 var_4444_epsilon_0 = const()[name = string("op_4444_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4444_cast_fp16 = rsqrt(epsilon = var_4444_epsilon_0, x = var_4443_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_103_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4444_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78684928)))];
            tensor<fp16, [1, 1260, 768]> var_4446_cast_fp16 = mul(x = out_103_cast_fp16, y = t_encoder_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_4446_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_159_cast_fp16 = add(x = x_145_cast_fp16, y = var_4446_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4449_cast_fp16 = abs(x = x_159_cast_fp16)[name = string("op_4449_cast_fp16")];
            bool var_4452 = const()[name = string("op_4452"), val = bool(true)];
            tensor<int32, [1]> var_4453_axes_0 = const()[name = string("op_4453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4453_cast_fp16 = reduce_max(axes = var_4453_axes_0, keep_dims = var_4452, x = var_4449_cast_fp16)[name = string("op_4453_cast_fp16")];
            fp16 var_4454_to_fp16 = const()[name = string("op_4454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_127_cast_fp16 = maximum(x = var_4453_cast_fp16, y = var_4454_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_123_cast_fp16 = real_div(x = x_159_cast_fp16, y = s_127_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4457_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [1]> var_4462_axes_0 = const()[name = string("op_4462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4462_keep_dims_0 = const()[name = string("op_4462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4462_cast_fp16 = reduce_mean(axes = var_4462_axes_0, keep_dims = var_4462_keep_dims_0, x = var_4457_cast_fp16)[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4463_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4463_cast_fp16")];
            fp32 var_4464_epsilon_0 = const()[name = string("op_4464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4464_cast_fp16 = inverse(epsilon = var_4464_epsilon_0, x = var_4463_cast_fp16)[name = string("op_4464_cast_fp16")];
            fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4466_cast_fp16 = mul(x = var_4464_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4468_cast_fp16 = add(x = var_4462_cast_fp16, y = var_4466_cast_fp16)[name = string("op_4468_cast_fp16")];
            fp32 var_4469_epsilon_0 = const()[name = string("op_4469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4469_cast_fp16 = rsqrt(epsilon = var_4469_epsilon_0, x = var_4468_cast_fp16)[name = string("op_4469_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_105_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4469_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78686528)))];
            tensor<fp16, [1, 1260, 768]> input_87_cast_fp16 = mul(x = out_105_cast_fp16, y = t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78688128))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81047488))))[name = string("t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_61_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_61_cast_fp16")];
            string var_4476_mode_0 = const()[name = string("op_4476_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_4476_cast_fp16 = gelu(mode = var_4476_mode_0, x = linear_61_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81053696))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83413056))))[name = string("t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_62_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_91_cast_fp16 = mul(x = var_4476_cast_fp16, y = linear_62_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83419264))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85778624))))[name = string("t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4484_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4484_cast_fp16")];
            bool var_4487 = const()[name = string("op_4487"), val = bool(true)];
            tensor<int32, [1]> var_4488_axes_0 = const()[name = string("op_4488_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4488_cast_fp16 = reduce_max(axes = var_4488_axes_0, keep_dims = var_4487, x = var_4484_cast_fp16)[name = string("op_4488_cast_fp16")];
            fp16 var_4489_to_fp16 = const()[name = string("op_4489_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_129_cast_fp16 = maximum(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_125_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_129_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4492_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4492_cast_fp16")];
            tensor<int32, [1]> var_4497_axes_0 = const()[name = string("op_4497_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4497_keep_dims_0 = const()[name = string("op_4497_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4497_cast_fp16 = reduce_mean(axes = var_4497_axes_0, keep_dims = var_4497_keep_dims_0, x = var_4492_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4498_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4498_cast_fp16")];
            fp32 var_4499_epsilon_0 = const()[name = string("op_4499_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4499_cast_fp16 = inverse(epsilon = var_4499_epsilon_0, x = var_4498_cast_fp16)[name = string("op_4499_cast_fp16")];
            fp16 var_4500_to_fp16 = const()[name = string("op_4500_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4501_cast_fp16 = mul(x = var_4499_cast_fp16, y = var_4500_to_fp16)[name = string("op_4501_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4503_cast_fp16 = add(x = var_4497_cast_fp16, y = var_4501_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4504_cast_fp16 = rsqrt(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_107_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4504_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85780224)))];
            tensor<fp16, [1, 1260, 768]> var_4506_cast_fp16 = mul(x = out_107_cast_fp16, y = t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_163_cast_fp16 = add(x = x_159_cast_fp16, y = var_4506_cast_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4509_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_4509_cast_fp16")];
            bool var_4512 = const()[name = string("op_4512"), val = bool(true)];
            tensor<int32, [1]> var_4513_axes_0 = const()[name = string("op_4513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4513_cast_fp16 = reduce_max(axes = var_4513_axes_0, keep_dims = var_4512, x = var_4509_cast_fp16)[name = string("op_4513_cast_fp16")];
            fp16 var_4514_to_fp16 = const()[name = string("op_4514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_131_cast_fp16 = maximum(x = var_4513_cast_fp16, y = var_4514_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_127_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_131_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4517_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4517_cast_fp16")];
            tensor<int32, [1]> var_4522_axes_0 = const()[name = string("op_4522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4522_keep_dims_0 = const()[name = string("op_4522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4522_cast_fp16 = reduce_mean(axes = var_4522_axes_0, keep_dims = var_4522_keep_dims_0, x = var_4517_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4523_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4523_cast_fp16")];
            fp32 var_4524_epsilon_0 = const()[name = string("op_4524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4524_cast_fp16 = inverse(epsilon = var_4524_epsilon_0, x = var_4523_cast_fp16)[name = string("op_4524_cast_fp16")];
            fp16 var_4525_to_fp16 = const()[name = string("op_4525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4526_cast_fp16 = mul(x = var_4524_cast_fp16, y = var_4525_to_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4528_cast_fp16 = add(x = var_4522_cast_fp16, y = var_4526_cast_fp16)[name = string("op_4528_cast_fp16")];
            fp32 var_4529_epsilon_0 = const()[name = string("op_4529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4529_cast_fp16 = rsqrt(epsilon = var_4529_epsilon_0, x = var_4528_cast_fp16)[name = string("op_4529_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_109_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4529_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85781824)))];
            tensor<fp16, [1, 1260, 768]> input_93_cast_fp16 = mul(x = out_109_cast_fp16, y = t_encoder_layers_9_input_layernorm_weight_to_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85783424))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86373312))))[name = string("t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4539 = const()[name = string("op_4539"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_165_cast_fp16 = reshape(shape = var_4539, x = linear_64_cast_fp16)[name = string("x_165_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4541_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_4541_cast_fp16")];
            bool var_4544 = const()[name = string("op_4544"), val = bool(true)];
            tensor<int32, [1]> var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4545_cast_fp16 = reduce_max(axes = var_4545_axes_0, keep_dims = var_4544, x = var_4541_cast_fp16)[name = string("op_4545_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_133_cast_fp16 = maximum(x = var_4545_cast_fp16, y = var_4546_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_129_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_133_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4549_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<int32, [1]> var_4554_axes_0 = const()[name = string("op_4554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4554_keep_dims_0 = const()[name = string("op_4554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4554_cast_fp16 = reduce_mean(axes = var_4554_axes_0, keep_dims = var_4554_keep_dims_0, x = var_4549_cast_fp16)[name = string("op_4554_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4555_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4555_cast_fp16")];
            fp32 var_4556_epsilon_0 = const()[name = string("op_4556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4556_cast_fp16 = inverse(epsilon = var_4556_epsilon_0, x = var_4555_cast_fp16)[name = string("op_4556_cast_fp16")];
            fp16 var_4557_to_fp16 = const()[name = string("op_4557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4558_cast_fp16 = mul(x = var_4556_cast_fp16, y = var_4557_to_fp16)[name = string("op_4558_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4560_cast_fp16 = add(x = var_4554_cast_fp16, y = var_4558_cast_fp16)[name = string("op_4560_cast_fp16")];
            fp32 var_4561_epsilon_0 = const()[name = string("op_4561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4561_cast_fp16 = rsqrt(epsilon = var_4561_epsilon_0, x = var_4560_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_111_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4561_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86374912)))];
            tensor<fp16, [1, 1260, 12, 64]> x_171_cast_fp16 = mul(x = out_111_cast_fp16, y = t_encoder_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_171_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86375104))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86964992))))[name = string("t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_65_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4571 = const()[name = string("op_4571"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_167_cast_fp16 = reshape(shape = var_4571, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4573_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_4573_cast_fp16")];
            bool var_4576 = const()[name = string("op_4576"), val = bool(true)];
            tensor<int32, [1]> var_4577_axes_0 = const()[name = string("op_4577_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4577_cast_fp16 = reduce_max(axes = var_4577_axes_0, keep_dims = var_4576, x = var_4573_cast_fp16)[name = string("op_4577_cast_fp16")];
            fp16 var_4578_to_fp16 = const()[name = string("op_4578_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_135_cast_fp16 = maximum(x = var_4577_cast_fp16, y = var_4578_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_131_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_135_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4581_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4581_cast_fp16")];
            tensor<int32, [1]> var_4586_axes_0 = const()[name = string("op_4586_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4586_keep_dims_0 = const()[name = string("op_4586_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4586_cast_fp16 = reduce_mean(axes = var_4586_axes_0, keep_dims = var_4586_keep_dims_0, x = var_4581_cast_fp16)[name = string("op_4586_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4587_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4587_cast_fp16")];
            fp32 var_4588_epsilon_0 = const()[name = string("op_4588_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4588_cast_fp16 = inverse(epsilon = var_4588_epsilon_0, x = var_4587_cast_fp16)[name = string("op_4588_cast_fp16")];
            fp16 var_4589_to_fp16 = const()[name = string("op_4589_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4590_cast_fp16 = mul(x = var_4588_cast_fp16, y = var_4589_to_fp16)[name = string("op_4590_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4592_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4590_cast_fp16)[name = string("op_4592_cast_fp16")];
            fp32 var_4593_epsilon_0 = const()[name = string("op_4593_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4593_cast_fp16 = rsqrt(epsilon = var_4593_epsilon_0, x = var_4592_cast_fp16)[name = string("op_4593_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_113_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4593_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966592)))];
            tensor<fp16, [1, 1260, 12, 64]> x_173_cast_fp16 = mul(x = out_113_cast_fp16, y = t_encoder_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966784))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87556672))))[name = string("t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4603 = const()[name = string("op_4603"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_169_cast_fp16 = reshape(shape = var_4603, x = linear_66_cast_fp16)[name = string("x_169_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4605_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_4605_cast_fp16")];
            bool var_4608 = const()[name = string("op_4608"), val = bool(true)];
            tensor<int32, [1]> var_4609_axes_0 = const()[name = string("op_4609_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4609_cast_fp16 = reduce_max(axes = var_4609_axes_0, keep_dims = var_4608, x = var_4605_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_137_cast_fp16 = maximum(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_133_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_137_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4613_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4613_cast_fp16")];
            tensor<int32, [1]> var_4618_axes_0 = const()[name = string("op_4618_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4618_keep_dims_0 = const()[name = string("op_4618_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4618_cast_fp16 = reduce_mean(axes = var_4618_axes_0, keep_dims = var_4618_keep_dims_0, x = var_4613_cast_fp16)[name = string("op_4618_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4619_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4619_cast_fp16")];
            fp32 var_4620_epsilon_0 = const()[name = string("op_4620_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4620_cast_fp16 = inverse(epsilon = var_4620_epsilon_0, x = var_4619_cast_fp16)[name = string("op_4620_cast_fp16")];
            fp16 var_4621_to_fp16 = const()[name = string("op_4621_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4622_cast_fp16 = mul(x = var_4620_cast_fp16, y = var_4621_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4624_cast_fp16 = add(x = var_4618_cast_fp16, y = var_4622_cast_fp16)[name = string("op_4624_cast_fp16")];
            fp32 var_4625_epsilon_0 = const()[name = string("op_4625_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4625_cast_fp16 = rsqrt(epsilon = var_4625_epsilon_0, x = var_4624_cast_fp16)[name = string("op_4625_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_37_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4625_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<int32, [4]> z_73_begin_0 = const()[name = string("z_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_73_end_0 = const()[name = string("z_73_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_73_end_mask_0 = const()[name = string("z_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_73_cast_fp16 = slice_by_index(begin = z_73_begin_0, end = z_73_end_0, end_mask = z_73_end_mask_0, x = x_171_cast_fp16)[name = string("z_73_cast_fp16")];
            tensor<int32, [4]> var_4636_begin_0 = const()[name = string("op_4636_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4636_end_0 = const()[name = string("op_4636_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4636_end_mask_0 = const()[name = string("op_4636_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4636_cast_fp16 = slice_by_index(begin = var_4636_begin_0, end = var_4636_end_0, end_mask = var_4636_end_mask_0, x = z_73_cast_fp16)[name = string("op_4636_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4637_cast_fp16 = mul(x = var_4636_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_4637_cast_fp16")];
            tensor<int32, [4]> var_4642_begin_0 = const()[name = string("op_4642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4642_end_0 = const()[name = string("op_4642_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4642_end_mask_0 = const()[name = string("op_4642_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4642_cast_fp16 = slice_by_index(begin = var_4642_begin_0, end = var_4642_end_0, end_mask = var_4642_end_mask_0, x = z_73_cast_fp16)[name = string("op_4642_cast_fp16")];
            int32 var_4644 = const()[name = string("op_4644"), val = int32(-1)];
            bool rot_73_interleave_0 = const()[name = string("rot_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_73_cast_fp16 = concat(axis = var_4644, interleave = rot_73_interleave_0, values = (var_4637_cast_fp16, var_4642_cast_fp16))[name = string("rot_73_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4646_cast_fp16 = mul(x = z_73_cast_fp16, y = c_1_cast_fp16)[name = string("op_4646_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4647_cast_fp16 = mul(x = rot_73_cast_fp16, y = s_9_cast_fp16)[name = string("op_4647_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4649_cast_fp16 = add(x = var_4646_cast_fp16, y = var_4647_cast_fp16)[name = string("op_4649_cast_fp16")];
            tensor<int32, [4]> z_75_begin_0 = const()[name = string("z_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_75_end_0 = const()[name = string("z_75_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_75_end_mask_0 = const()[name = string("z_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_75_cast_fp16 = slice_by_index(begin = z_75_begin_0, end = z_75_end_0, end_mask = z_75_end_mask_0, x = x_171_cast_fp16)[name = string("z_75_cast_fp16")];
            tensor<int32, [4]> var_4659_begin_0 = const()[name = string("op_4659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4659_end_0 = const()[name = string("op_4659_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4659_end_mask_0 = const()[name = string("op_4659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4659_cast_fp16 = slice_by_index(begin = var_4659_begin_0, end = var_4659_end_0, end_mask = var_4659_end_mask_0, x = z_75_cast_fp16)[name = string("op_4659_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4660_cast_fp16 = mul(x = var_4659_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_4660_cast_fp16")];
            tensor<int32, [4]> var_4665_begin_0 = const()[name = string("op_4665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4665_end_0 = const()[name = string("op_4665_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4665_end_mask_0 = const()[name = string("op_4665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4665_cast_fp16 = slice_by_index(begin = var_4665_begin_0, end = var_4665_end_0, end_mask = var_4665_end_mask_0, x = z_75_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4667 = const()[name = string("op_4667"), val = int32(-1)];
            bool rot_75_interleave_0 = const()[name = string("rot_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_75_cast_fp16 = concat(axis = var_4667, interleave = rot_75_interleave_0, values = (var_4660_cast_fp16, var_4665_cast_fp16))[name = string("rot_75_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4669_cast_fp16 = mul(x = z_75_cast_fp16, y = c_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4670_cast_fp16 = mul(x = rot_75_cast_fp16, y = s_11_cast_fp16)[name = string("op_4670_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4672_cast_fp16 = add(x = var_4669_cast_fp16, y = var_4670_cast_fp16)[name = string("op_4672_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(-1)];
            bool var_4675_interleave_0 = const()[name = string("op_4675_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_4675_cast_fp16 = concat(axis = var_4674, interleave = var_4675_interleave_0, values = (var_4649_cast_fp16, var_4672_cast_fp16))[name = string("op_4675_cast_fp16")];
            tensor<int32, [4]> z_77_begin_0 = const()[name = string("z_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_77_end_0 = const()[name = string("z_77_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_77_end_mask_0 = const()[name = string("z_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_77_cast_fp16 = slice_by_index(begin = z_77_begin_0, end = z_77_end_0, end_mask = z_77_end_mask_0, x = x_173_cast_fp16)[name = string("z_77_cast_fp16")];
            tensor<int32, [4]> var_4688_begin_0 = const()[name = string("op_4688_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4688_end_0 = const()[name = string("op_4688_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4688_end_mask_0 = const()[name = string("op_4688_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4688_cast_fp16 = slice_by_index(begin = var_4688_begin_0, end = var_4688_end_0, end_mask = var_4688_end_mask_0, x = z_77_cast_fp16)[name = string("op_4688_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4689_cast_fp16 = mul(x = var_4688_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [4]> var_4694_begin_0 = const()[name = string("op_4694_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4694_end_0 = const()[name = string("op_4694_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4694_end_mask_0 = const()[name = string("op_4694_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4694_cast_fp16 = slice_by_index(begin = var_4694_begin_0, end = var_4694_end_0, end_mask = var_4694_end_mask_0, x = z_77_cast_fp16)[name = string("op_4694_cast_fp16")];
            int32 var_4696 = const()[name = string("op_4696"), val = int32(-1)];
            bool rot_77_interleave_0 = const()[name = string("rot_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_77_cast_fp16 = concat(axis = var_4696, interleave = rot_77_interleave_0, values = (var_4689_cast_fp16, var_4694_cast_fp16))[name = string("rot_77_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4698_cast_fp16 = mul(x = z_77_cast_fp16, y = c_1_cast_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4699_cast_fp16 = mul(x = rot_77_cast_fp16, y = s_9_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4701_cast_fp16 = add(x = var_4698_cast_fp16, y = var_4699_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<int32, [4]> z_79_begin_0 = const()[name = string("z_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_79_end_0 = const()[name = string("z_79_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_79_end_mask_0 = const()[name = string("z_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_79_cast_fp16 = slice_by_index(begin = z_79_begin_0, end = z_79_end_0, end_mask = z_79_end_mask_0, x = x_173_cast_fp16)[name = string("z_79_cast_fp16")];
            tensor<int32, [4]> var_4711_begin_0 = const()[name = string("op_4711_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4711_end_0 = const()[name = string("op_4711_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4711_end_mask_0 = const()[name = string("op_4711_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4711_cast_fp16 = slice_by_index(begin = var_4711_begin_0, end = var_4711_end_0, end_mask = var_4711_end_mask_0, x = z_79_cast_fp16)[name = string("op_4711_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4712_cast_fp16 = mul(x = var_4711_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_4712_cast_fp16")];
            tensor<int32, [4]> var_4717_begin_0 = const()[name = string("op_4717_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4717_end_0 = const()[name = string("op_4717_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4717_end_mask_0 = const()[name = string("op_4717_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4717_cast_fp16 = slice_by_index(begin = var_4717_begin_0, end = var_4717_end_0, end_mask = var_4717_end_mask_0, x = z_79_cast_fp16)[name = string("op_4717_cast_fp16")];
            int32 var_4719 = const()[name = string("op_4719"), val = int32(-1)];
            bool rot_79_interleave_0 = const()[name = string("rot_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_79_cast_fp16 = concat(axis = var_4719, interleave = rot_79_interleave_0, values = (var_4712_cast_fp16, var_4717_cast_fp16))[name = string("rot_79_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4721_cast_fp16 = mul(x = z_79_cast_fp16, y = c_cast_fp16)[name = string("op_4721_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4722_cast_fp16 = mul(x = rot_79_cast_fp16, y = s_11_cast_fp16)[name = string("op_4722_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4724_cast_fp16 = add(x = var_4721_cast_fp16, y = var_4722_cast_fp16)[name = string("op_4724_cast_fp16")];
            int32 var_4726 = const()[name = string("op_4726"), val = int32(-1)];
            bool var_4727_interleave_0 = const()[name = string("op_4727_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_4727_cast_fp16 = concat(axis = var_4726, interleave = var_4727_interleave_0, values = (var_4701_cast_fp16, var_4724_cast_fp16))[name = string("op_4727_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4737_transpose_x_0 = const()[name = string("op_4737_transpose_x_0"), val = bool(false)];
            bool var_4737_transpose_y_0 = const()[name = string("op_4737_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_4727_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_82 = transpose(perm = transpose_82_perm_0, x = var_4675_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 12, 1260, 1260]> var_4737_cast_fp16 = matmul(transpose_x = var_4737_transpose_x_0, transpose_y = var_4737_transpose_y_0, x = transpose_82, y = transpose_83)[name = string("op_4737_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_4739_cast_fp16 = add(x = var_4737_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4739_cast_fp16")];
            int32 var_4740 = const()[name = string("op_4740"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_133_cast_fp16 = softmax(axis = var_4740, x = var_4739_cast_fp16)[name = string("w_133_cast_fp16")];
            bool var_4743_transpose_x_0 = const()[name = string("op_4743_transpose_x_0"), val = bool(false)];
            bool var_4743_transpose_y_0 = const()[name = string("op_4743_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = v_37_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 12, 1260, 64]> var_4743_cast_fp16 = matmul(transpose_x = var_4743_transpose_x_0, transpose_y = var_4743_transpose_y_0, x = w_133_cast_fp16, y = v_39_cast_fp16)[name = string("op_4743_cast_fp16")];
            tensor<int32, [4]> var_4746_perm_0 = const()[name = string("op_4746_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4750 = const()[name = string("op_4750"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_4746_cast_fp16 = transpose(perm = var_4746_perm_0, x = var_4743_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 1260, 768]> input_95_cast_fp16 = reshape(shape = var_4750, x = var_4746_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87558272))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88148160))))[name = string("t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_67_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4755_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4755_cast_fp16")];
            bool var_4758 = const()[name = string("op_4758"), val = bool(true)];
            tensor<int32, [1]> var_4759_axes_0 = const()[name = string("op_4759_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4759_cast_fp16 = reduce_max(axes = var_4759_axes_0, keep_dims = var_4758, x = var_4755_cast_fp16)[name = string("op_4759_cast_fp16")];
            fp16 var_4760_to_fp16 = const()[name = string("op_4760_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_139_cast_fp16 = maximum(x = var_4759_cast_fp16, y = var_4760_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_135_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_139_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4763_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4763_cast_fp16")];
            tensor<int32, [1]> var_4768_axes_0 = const()[name = string("op_4768_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4768_keep_dims_0 = const()[name = string("op_4768_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4768_cast_fp16 = reduce_mean(axes = var_4768_axes_0, keep_dims = var_4768_keep_dims_0, x = var_4763_cast_fp16)[name = string("op_4768_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4769_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_4769_cast_fp16")];
            fp32 var_4770_epsilon_0 = const()[name = string("op_4770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4770_cast_fp16 = inverse(epsilon = var_4770_epsilon_0, x = var_4769_cast_fp16)[name = string("op_4770_cast_fp16")];
            fp16 var_4771_to_fp16 = const()[name = string("op_4771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4772_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_4771_to_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4774_cast_fp16 = add(x = var_4768_cast_fp16, y = var_4772_cast_fp16)[name = string("op_4774_cast_fp16")];
            fp32 var_4775_epsilon_0 = const()[name = string("op_4775_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4775_cast_fp16 = rsqrt(epsilon = var_4775_epsilon_0, x = var_4774_cast_fp16)[name = string("op_4775_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_115_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4775_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88149760)))];
            tensor<fp16, [1, 1260, 768]> var_4777_cast_fp16 = mul(x = out_115_cast_fp16, y = t_encoder_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_4777_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_177_cast_fp16 = add(x = x_163_cast_fp16, y = var_4777_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4780_cast_fp16 = abs(x = x_177_cast_fp16)[name = string("op_4780_cast_fp16")];
            bool var_4783 = const()[name = string("op_4783"), val = bool(true)];
            tensor<int32, [1]> var_4784_axes_0 = const()[name = string("op_4784_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4784_cast_fp16 = reduce_max(axes = var_4784_axes_0, keep_dims = var_4783, x = var_4780_cast_fp16)[name = string("op_4784_cast_fp16")];
            fp16 var_4785_to_fp16 = const()[name = string("op_4785_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_141_cast_fp16 = maximum(x = var_4784_cast_fp16, y = var_4785_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_137_cast_fp16 = real_div(x = x_177_cast_fp16, y = s_141_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4788_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4788_cast_fp16")];
            tensor<int32, [1]> var_4793_axes_0 = const()[name = string("op_4793_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4793_keep_dims_0 = const()[name = string("op_4793_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4793_cast_fp16 = reduce_mean(axes = var_4793_axes_0, keep_dims = var_4793_keep_dims_0, x = var_4788_cast_fp16)[name = string("op_4793_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4794_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_4794_cast_fp16")];
            fp32 var_4795_epsilon_0 = const()[name = string("op_4795_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4795_cast_fp16 = inverse(epsilon = var_4795_epsilon_0, x = var_4794_cast_fp16)[name = string("op_4795_cast_fp16")];
            fp16 var_4796_to_fp16 = const()[name = string("op_4796_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4797_cast_fp16 = mul(x = var_4795_cast_fp16, y = var_4796_to_fp16)[name = string("op_4797_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4799_cast_fp16 = add(x = var_4793_cast_fp16, y = var_4797_cast_fp16)[name = string("op_4799_cast_fp16")];
            fp32 var_4800_epsilon_0 = const()[name = string("op_4800_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4800_cast_fp16 = rsqrt(epsilon = var_4800_epsilon_0, x = var_4799_cast_fp16)[name = string("op_4800_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_117_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4800_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88151360)))];
            tensor<fp16, [1, 1260, 768]> input_97_cast_fp16 = mul(x = out_117_cast_fp16, y = t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88152960))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90512320))))[name = string("t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4807_mode_0 = const()[name = string("op_4807_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_4807_cast_fp16 = gelu(mode = var_4807_mode_0, x = linear_68_cast_fp16)[name = string("op_4807_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90518528))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92877888))))[name = string("t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_101_cast_fp16 = mul(x = var_4807_cast_fp16, y = linear_69_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92884096))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95243456))))[name = string("t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4815_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4815_cast_fp16")];
            bool var_4818 = const()[name = string("op_4818"), val = bool(true)];
            tensor<int32, [1]> var_4819_axes_0 = const()[name = string("op_4819_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4819_cast_fp16 = reduce_max(axes = var_4819_axes_0, keep_dims = var_4818, x = var_4815_cast_fp16)[name = string("op_4819_cast_fp16")];
            fp16 var_4820_to_fp16 = const()[name = string("op_4820_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_143_cast_fp16 = maximum(x = var_4819_cast_fp16, y = var_4820_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_139_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_143_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4823_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_4823_cast_fp16")];
            tensor<int32, [1]> var_4828_axes_0 = const()[name = string("op_4828_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4828_keep_dims_0 = const()[name = string("op_4828_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4828_cast_fp16 = reduce_mean(axes = var_4828_axes_0, keep_dims = var_4828_keep_dims_0, x = var_4823_cast_fp16)[name = string("op_4828_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4829_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_4829_cast_fp16")];
            fp32 var_4830_epsilon_0 = const()[name = string("op_4830_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4830_cast_fp16 = inverse(epsilon = var_4830_epsilon_0, x = var_4829_cast_fp16)[name = string("op_4830_cast_fp16")];
            fp16 var_4831_to_fp16 = const()[name = string("op_4831_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4832_cast_fp16 = mul(x = var_4830_cast_fp16, y = var_4831_to_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4834_cast_fp16 = add(x = var_4828_cast_fp16, y = var_4832_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp32 var_4835_epsilon_0 = const()[name = string("op_4835_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4835_cast_fp16 = rsqrt(epsilon = var_4835_epsilon_0, x = var_4834_cast_fp16)[name = string("op_4835_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_119_cast_fp16 = mul(x = y_139_cast_fp16, y = var_4835_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95245056)))];
            tensor<fp16, [1, 1260, 768]> var_4837_cast_fp16 = mul(x = out_119_cast_fp16, y = t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4837_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_181_cast_fp16 = add(x = x_177_cast_fp16, y = var_4837_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4840_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_4840_cast_fp16")];
            bool var_4843 = const()[name = string("op_4843"), val = bool(true)];
            tensor<int32, [1]> var_4844_axes_0 = const()[name = string("op_4844_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_4844_cast_fp16 = reduce_max(axes = var_4844_axes_0, keep_dims = var_4843, x = var_4840_cast_fp16)[name = string("op_4844_cast_fp16")];
            fp16 var_4845_to_fp16 = const()[name = string("op_4845_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_145_cast_fp16 = maximum(x = var_4844_cast_fp16, y = var_4845_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_141_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_145_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_4848_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_4848_cast_fp16")];
            tensor<int32, [1]> var_4853_axes_0 = const()[name = string("op_4853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4853_keep_dims_0 = const()[name = string("op_4853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_4853_cast_fp16 = reduce_mean(axes = var_4853_axes_0, keep_dims = var_4853_keep_dims_0, x = var_4848_cast_fp16)[name = string("op_4853_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4854_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_4854_cast_fp16")];
            fp32 var_4855_epsilon_0 = const()[name = string("op_4855_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_4855_cast_fp16 = inverse(epsilon = var_4855_epsilon_0, x = var_4854_cast_fp16)[name = string("op_4855_cast_fp16")];
            fp16 var_4856_to_fp16 = const()[name = string("op_4856_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_4857_cast_fp16 = mul(x = var_4855_cast_fp16, y = var_4856_to_fp16)[name = string("op_4857_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_4859_cast_fp16 = add(x = var_4853_cast_fp16, y = var_4857_cast_fp16)[name = string("op_4859_cast_fp16")];
            fp32 var_4860_epsilon_0 = const()[name = string("op_4860_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_4860_cast_fp16 = rsqrt(epsilon = var_4860_epsilon_0, x = var_4859_cast_fp16)[name = string("op_4860_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_121_cast_fp16 = mul(x = y_141_cast_fp16, y = var_4860_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95246656)))];
            tensor<fp16, [1, 1260, 768]> input_103_cast_fp16 = mul(x = out_121_cast_fp16, y = t_encoder_layers_10_input_layernorm_weight_to_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95248256))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95838144))))[name = string("t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_71_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_4870 = const()[name = string("op_4870"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_183_cast_fp16 = reshape(shape = var_4870, x = linear_71_cast_fp16)[name = string("x_183_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4872_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_4872_cast_fp16")];
            bool var_4875 = const()[name = string("op_4875"), val = bool(true)];
            tensor<int32, [1]> var_4876_axes_0 = const()[name = string("op_4876_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4876_cast_fp16 = reduce_max(axes = var_4876_axes_0, keep_dims = var_4875, x = var_4872_cast_fp16)[name = string("op_4876_cast_fp16")];
            fp16 var_4877_to_fp16 = const()[name = string("op_4877_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_147_cast_fp16 = maximum(x = var_4876_cast_fp16, y = var_4877_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_143_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_147_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4880_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_4880_cast_fp16")];
            tensor<int32, [1]> var_4885_axes_0 = const()[name = string("op_4885_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4885_keep_dims_0 = const()[name = string("op_4885_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4885_cast_fp16 = reduce_mean(axes = var_4885_axes_0, keep_dims = var_4885_keep_dims_0, x = var_4880_cast_fp16)[name = string("op_4885_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4886_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_4886_cast_fp16")];
            fp32 var_4887_epsilon_0 = const()[name = string("op_4887_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4887_cast_fp16 = inverse(epsilon = var_4887_epsilon_0, x = var_4886_cast_fp16)[name = string("op_4887_cast_fp16")];
            fp16 var_4888_to_fp16 = const()[name = string("op_4888_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4889_cast_fp16 = mul(x = var_4887_cast_fp16, y = var_4888_to_fp16)[name = string("op_4889_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4891_cast_fp16 = add(x = var_4885_cast_fp16, y = var_4889_cast_fp16)[name = string("op_4891_cast_fp16")];
            fp32 var_4892_epsilon_0 = const()[name = string("op_4892_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4892_cast_fp16 = rsqrt(epsilon = var_4892_epsilon_0, x = var_4891_cast_fp16)[name = string("op_4892_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_123_cast_fp16 = mul(x = y_143_cast_fp16, y = var_4892_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95839744)))];
            tensor<fp16, [1, 1260, 12, 64]> x_189_cast_fp16 = mul(x = out_123_cast_fp16, y = t_encoder_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_189_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95839936))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96429824))))[name = string("t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_4902 = const()[name = string("op_4902"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_185_cast_fp16 = reshape(shape = var_4902, x = linear_72_cast_fp16)[name = string("x_185_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4904_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_4904_cast_fp16")];
            bool var_4907 = const()[name = string("op_4907"), val = bool(true)];
            tensor<int32, [1]> var_4908_axes_0 = const()[name = string("op_4908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4908_cast_fp16 = reduce_max(axes = var_4908_axes_0, keep_dims = var_4907, x = var_4904_cast_fp16)[name = string("op_4908_cast_fp16")];
            fp16 var_4909_to_fp16 = const()[name = string("op_4909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_149_cast_fp16 = maximum(x = var_4908_cast_fp16, y = var_4909_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_145_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_149_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4912_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_4912_cast_fp16")];
            tensor<int32, [1]> var_4917_axes_0 = const()[name = string("op_4917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4917_keep_dims_0 = const()[name = string("op_4917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4917_cast_fp16 = reduce_mean(axes = var_4917_axes_0, keep_dims = var_4917_keep_dims_0, x = var_4912_cast_fp16)[name = string("op_4917_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4918_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_4918_cast_fp16")];
            fp32 var_4919_epsilon_0 = const()[name = string("op_4919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4919_cast_fp16 = inverse(epsilon = var_4919_epsilon_0, x = var_4918_cast_fp16)[name = string("op_4919_cast_fp16")];
            fp16 var_4920_to_fp16 = const()[name = string("op_4920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4923_cast_fp16 = add(x = var_4917_cast_fp16, y = var_4921_cast_fp16)[name = string("op_4923_cast_fp16")];
            fp32 var_4924_epsilon_0 = const()[name = string("op_4924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4924_cast_fp16 = rsqrt(epsilon = var_4924_epsilon_0, x = var_4923_cast_fp16)[name = string("op_4924_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_125_cast_fp16 = mul(x = y_145_cast_fp16, y = var_4924_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96431424)))];
            tensor<fp16, [1, 1260, 12, 64]> x_191_cast_fp16 = mul(x = out_125_cast_fp16, y = t_encoder_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96431616))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97021504))))[name = string("t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_73_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4934 = const()[name = string("op_4934"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_187_cast_fp16 = reshape(shape = var_4934, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4936_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_4936_cast_fp16")];
            bool var_4939 = const()[name = string("op_4939"), val = bool(true)];
            tensor<int32, [1]> var_4940_axes_0 = const()[name = string("op_4940_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_4940_cast_fp16 = reduce_max(axes = var_4940_axes_0, keep_dims = var_4939, x = var_4936_cast_fp16)[name = string("op_4940_cast_fp16")];
            fp16 var_4941_to_fp16 = const()[name = string("op_4941_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_151_cast_fp16 = maximum(x = var_4940_cast_fp16, y = var_4941_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_147_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_151_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_4944_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<int32, [1]> var_4949_axes_0 = const()[name = string("op_4949_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4949_keep_dims_0 = const()[name = string("op_4949_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_4949_cast_fp16 = reduce_mean(axes = var_4949_axes_0, keep_dims = var_4949_keep_dims_0, x = var_4944_cast_fp16)[name = string("op_4949_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4950_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_4950_cast_fp16")];
            fp32 var_4951_epsilon_0 = const()[name = string("op_4951_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_4951_cast_fp16 = inverse(epsilon = var_4951_epsilon_0, x = var_4950_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 var_4952_to_fp16 = const()[name = string("op_4952_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_4953_cast_fp16 = mul(x = var_4951_cast_fp16, y = var_4952_to_fp16)[name = string("op_4953_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_4955_cast_fp16 = add(x = var_4949_cast_fp16, y = var_4953_cast_fp16)[name = string("op_4955_cast_fp16")];
            fp32 var_4956_epsilon_0 = const()[name = string("op_4956_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_4956_cast_fp16 = rsqrt(epsilon = var_4956_epsilon_0, x = var_4955_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_41_cast_fp16 = mul(x = y_147_cast_fp16, y = var_4956_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<int32, [4]> z_81_begin_0 = const()[name = string("z_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_81_end_0 = const()[name = string("z_81_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_81_end_mask_0 = const()[name = string("z_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_81_cast_fp16 = slice_by_index(begin = z_81_begin_0, end = z_81_end_0, end_mask = z_81_end_mask_0, x = x_189_cast_fp16)[name = string("z_81_cast_fp16")];
            tensor<int32, [4]> var_4967_begin_0 = const()[name = string("op_4967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4967_end_0 = const()[name = string("op_4967_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4967_end_mask_0 = const()[name = string("op_4967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4967_cast_fp16 = slice_by_index(begin = var_4967_begin_0, end = var_4967_end_0, end_mask = var_4967_end_mask_0, x = z_81_cast_fp16)[name = string("op_4967_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4968_cast_fp16 = mul(x = var_4967_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_4968_cast_fp16")];
            tensor<int32, [4]> var_4973_begin_0 = const()[name = string("op_4973_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4973_end_0 = const()[name = string("op_4973_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4973_end_mask_0 = const()[name = string("op_4973_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, x = z_81_cast_fp16)[name = string("op_4973_cast_fp16")];
            int32 var_4975 = const()[name = string("op_4975"), val = int32(-1)];
            bool rot_81_interleave_0 = const()[name = string("rot_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_81_cast_fp16 = concat(axis = var_4975, interleave = rot_81_interleave_0, values = (var_4968_cast_fp16, var_4973_cast_fp16))[name = string("rot_81_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4977_cast_fp16 = mul(x = z_81_cast_fp16, y = c_1_cast_fp16)[name = string("op_4977_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4978_cast_fp16 = mul(x = rot_81_cast_fp16, y = s_9_cast_fp16)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_4980_cast_fp16 = add(x = var_4977_cast_fp16, y = var_4978_cast_fp16)[name = string("op_4980_cast_fp16")];
            tensor<int32, [4]> z_83_begin_0 = const()[name = string("z_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_83_end_0 = const()[name = string("z_83_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_83_end_mask_0 = const()[name = string("z_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_83_cast_fp16 = slice_by_index(begin = z_83_begin_0, end = z_83_end_0, end_mask = z_83_end_mask_0, x = x_189_cast_fp16)[name = string("z_83_cast_fp16")];
            tensor<int32, [4]> var_4990_begin_0 = const()[name = string("op_4990_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4990_end_0 = const()[name = string("op_4990_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_4990_end_mask_0 = const()[name = string("op_4990_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_4990_cast_fp16 = slice_by_index(begin = var_4990_begin_0, end = var_4990_end_0, end_mask = var_4990_end_mask_0, x = z_83_cast_fp16)[name = string("op_4990_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_4991_cast_fp16 = mul(x = var_4990_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_4991_cast_fp16")];
            tensor<int32, [4]> var_4996_begin_0 = const()[name = string("op_4996_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4996_end_0 = const()[name = string("op_4996_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_4996_end_mask_0 = const()[name = string("op_4996_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_4996_cast_fp16 = slice_by_index(begin = var_4996_begin_0, end = var_4996_end_0, end_mask = var_4996_end_mask_0, x = z_83_cast_fp16)[name = string("op_4996_cast_fp16")];
            int32 var_4998 = const()[name = string("op_4998"), val = int32(-1)];
            bool rot_83_interleave_0 = const()[name = string("rot_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_83_cast_fp16 = concat(axis = var_4998, interleave = rot_83_interleave_0, values = (var_4991_cast_fp16, var_4996_cast_fp16))[name = string("rot_83_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5000_cast_fp16 = mul(x = z_83_cast_fp16, y = c_cast_fp16)[name = string("op_5000_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5001_cast_fp16 = mul(x = rot_83_cast_fp16, y = s_11_cast_fp16)[name = string("op_5001_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5003_cast_fp16 = add(x = var_5000_cast_fp16, y = var_5001_cast_fp16)[name = string("op_5003_cast_fp16")];
            int32 var_5005 = const()[name = string("op_5005"), val = int32(-1)];
            bool var_5006_interleave_0 = const()[name = string("op_5006_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_5006_cast_fp16 = concat(axis = var_5005, interleave = var_5006_interleave_0, values = (var_4980_cast_fp16, var_5003_cast_fp16))[name = string("op_5006_cast_fp16")];
            tensor<int32, [4]> z_85_begin_0 = const()[name = string("z_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_85_end_0 = const()[name = string("z_85_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_85_end_mask_0 = const()[name = string("z_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_85_cast_fp16 = slice_by_index(begin = z_85_begin_0, end = z_85_end_0, end_mask = z_85_end_mask_0, x = x_191_cast_fp16)[name = string("z_85_cast_fp16")];
            tensor<int32, [4]> var_5019_begin_0 = const()[name = string("op_5019_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5019_end_0 = const()[name = string("op_5019_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5019_end_mask_0 = const()[name = string("op_5019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5019_cast_fp16 = slice_by_index(begin = var_5019_begin_0, end = var_5019_end_0, end_mask = var_5019_end_mask_0, x = z_85_cast_fp16)[name = string("op_5019_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5020_cast_fp16 = mul(x = var_5019_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_5020_cast_fp16")];
            tensor<int32, [4]> var_5025_begin_0 = const()[name = string("op_5025_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5025_end_0 = const()[name = string("op_5025_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5025_end_mask_0 = const()[name = string("op_5025_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5025_cast_fp16 = slice_by_index(begin = var_5025_begin_0, end = var_5025_end_0, end_mask = var_5025_end_mask_0, x = z_85_cast_fp16)[name = string("op_5025_cast_fp16")];
            int32 var_5027 = const()[name = string("op_5027"), val = int32(-1)];
            bool rot_85_interleave_0 = const()[name = string("rot_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_85_cast_fp16 = concat(axis = var_5027, interleave = rot_85_interleave_0, values = (var_5020_cast_fp16, var_5025_cast_fp16))[name = string("rot_85_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5029_cast_fp16 = mul(x = z_85_cast_fp16, y = c_1_cast_fp16)[name = string("op_5029_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5030_cast_fp16 = mul(x = rot_85_cast_fp16, y = s_9_cast_fp16)[name = string("op_5030_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5032_cast_fp16 = add(x = var_5029_cast_fp16, y = var_5030_cast_fp16)[name = string("op_5032_cast_fp16")];
            tensor<int32, [4]> z_87_begin_0 = const()[name = string("z_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_87_end_0 = const()[name = string("z_87_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_87_end_mask_0 = const()[name = string("z_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_87_cast_fp16 = slice_by_index(begin = z_87_begin_0, end = z_87_end_0, end_mask = z_87_end_mask_0, x = x_191_cast_fp16)[name = string("z_87_cast_fp16")];
            tensor<int32, [4]> var_5042_begin_0 = const()[name = string("op_5042_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5042_end_0 = const()[name = string("op_5042_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5042_end_mask_0 = const()[name = string("op_5042_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5042_cast_fp16 = slice_by_index(begin = var_5042_begin_0, end = var_5042_end_0, end_mask = var_5042_end_mask_0, x = z_87_cast_fp16)[name = string("op_5042_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5043_cast_fp16 = mul(x = var_5042_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_5043_cast_fp16")];
            tensor<int32, [4]> var_5048_begin_0 = const()[name = string("op_5048_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5048_end_0 = const()[name = string("op_5048_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5048_end_mask_0 = const()[name = string("op_5048_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5048_cast_fp16 = slice_by_index(begin = var_5048_begin_0, end = var_5048_end_0, end_mask = var_5048_end_mask_0, x = z_87_cast_fp16)[name = string("op_5048_cast_fp16")];
            int32 var_5050 = const()[name = string("op_5050"), val = int32(-1)];
            bool rot_87_interleave_0 = const()[name = string("rot_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_87_cast_fp16 = concat(axis = var_5050, interleave = rot_87_interleave_0, values = (var_5043_cast_fp16, var_5048_cast_fp16))[name = string("rot_87_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5052_cast_fp16 = mul(x = z_87_cast_fp16, y = c_cast_fp16)[name = string("op_5052_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5053_cast_fp16 = mul(x = rot_87_cast_fp16, y = s_11_cast_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5055_cast_fp16 = add(x = var_5052_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            int32 var_5057 = const()[name = string("op_5057"), val = int32(-1)];
            bool var_5058_interleave_0 = const()[name = string("op_5058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_5058_cast_fp16 = concat(axis = var_5057, interleave = var_5058_interleave_0, values = (var_5032_cast_fp16, var_5055_cast_fp16))[name = string("op_5058_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5068_transpose_x_0 = const()[name = string("op_5068_transpose_x_0"), val = bool(false)];
            bool var_5068_transpose_y_0 = const()[name = string("op_5068_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_85 = transpose(perm = transpose_85_perm_0, x = var_5058_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_84 = transpose(perm = transpose_84_perm_0, x = var_5006_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 12, 1260, 1260]> var_5068_cast_fp16 = matmul(transpose_x = var_5068_transpose_x_0, transpose_y = var_5068_transpose_y_0, x = transpose_84, y = transpose_85)[name = string("op_5068_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_5070_cast_fp16 = add(x = var_5068_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5070_cast_fp16")];
            int32 var_5071 = const()[name = string("op_5071"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_147_cast_fp16 = softmax(axis = var_5071, x = var_5070_cast_fp16)[name = string("w_147_cast_fp16")];
            bool var_5074_transpose_x_0 = const()[name = string("op_5074_transpose_x_0"), val = bool(false)];
            bool var_5074_transpose_y_0 = const()[name = string("op_5074_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = v_41_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 12, 1260, 64]> var_5074_cast_fp16 = matmul(transpose_x = var_5074_transpose_x_0, transpose_y = var_5074_transpose_y_0, x = w_147_cast_fp16, y = v_43_cast_fp16)[name = string("op_5074_cast_fp16")];
            tensor<int32, [4]> var_5077_perm_0 = const()[name = string("op_5077_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5081 = const()[name = string("op_5081"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_5077_cast_fp16 = transpose(perm = var_5077_perm_0, x = var_5074_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 1260, 768]> input_105_cast_fp16 = reshape(shape = var_5081, x = var_5077_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97023104))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97612992))))[name = string("t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5086_cast_fp16 = abs(x = linear_74_cast_fp16)[name = string("op_5086_cast_fp16")];
            bool var_5089 = const()[name = string("op_5089"), val = bool(true)];
            tensor<int32, [1]> var_5090_axes_0 = const()[name = string("op_5090_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5090_cast_fp16 = reduce_max(axes = var_5090_axes_0, keep_dims = var_5089, x = var_5086_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp16 var_5091_to_fp16 = const()[name = string("op_5091_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_153_cast_fp16 = maximum(x = var_5090_cast_fp16, y = var_5091_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_149_cast_fp16 = real_div(x = linear_74_cast_fp16, y = s_153_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5094_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [1]> var_5099_axes_0 = const()[name = string("op_5099_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5099_keep_dims_0 = const()[name = string("op_5099_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5099_cast_fp16 = reduce_mean(axes = var_5099_axes_0, keep_dims = var_5099_keep_dims_0, x = var_5094_cast_fp16)[name = string("op_5099_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5100_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5100_cast_fp16")];
            fp32 var_5101_epsilon_0 = const()[name = string("op_5101_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5101_cast_fp16 = inverse(epsilon = var_5101_epsilon_0, x = var_5100_cast_fp16)[name = string("op_5101_cast_fp16")];
            fp16 var_5102_to_fp16 = const()[name = string("op_5102_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5103_cast_fp16 = mul(x = var_5101_cast_fp16, y = var_5102_to_fp16)[name = string("op_5103_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5105_cast_fp16 = add(x = var_5099_cast_fp16, y = var_5103_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, 1260, 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, 1260, 768]> out_127_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5106_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97614592)))];
            tensor<fp16, [1, 1260, 768]> var_5108_cast_fp16 = mul(x = out_127_cast_fp16, y = t_encoder_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5108_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_195_cast_fp16 = add(x = x_181_cast_fp16, y = var_5108_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5111_cast_fp16 = abs(x = x_195_cast_fp16)[name = string("op_5111_cast_fp16")];
            bool var_5114 = const()[name = string("op_5114"), val = bool(true)];
            tensor<int32, [1]> var_5115_axes_0 = const()[name = string("op_5115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5115_cast_fp16 = reduce_max(axes = var_5115_axes_0, keep_dims = var_5114, x = var_5111_cast_fp16)[name = string("op_5115_cast_fp16")];
            fp16 var_5116_to_fp16 = const()[name = string("op_5116_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_155_cast_fp16 = maximum(x = var_5115_cast_fp16, y = var_5116_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_151_cast_fp16 = real_div(x = x_195_cast_fp16, y = s_155_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5119_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5119_cast_fp16")];
            tensor<int32, [1]> var_5124_axes_0 = const()[name = string("op_5124_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5124_keep_dims_0 = const()[name = string("op_5124_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5124_cast_fp16 = reduce_mean(axes = var_5124_axes_0, keep_dims = var_5124_keep_dims_0, x = var_5119_cast_fp16)[name = string("op_5124_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5125_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5125_cast_fp16")];
            fp32 var_5126_epsilon_0 = const()[name = string("op_5126_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5126_cast_fp16 = inverse(epsilon = var_5126_epsilon_0, x = var_5125_cast_fp16)[name = string("op_5126_cast_fp16")];
            fp16 var_5127_to_fp16 = const()[name = string("op_5127_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5128_cast_fp16 = mul(x = var_5126_cast_fp16, y = var_5127_to_fp16)[name = string("op_5128_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5130_cast_fp16 = add(x = var_5124_cast_fp16, y = var_5128_cast_fp16)[name = string("op_5130_cast_fp16")];
            fp32 var_5131_epsilon_0 = const()[name = string("op_5131_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5131_cast_fp16 = rsqrt(epsilon = var_5131_epsilon_0, x = var_5130_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_129_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5131_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97616192)))];
            tensor<fp16, [1, 1260, 768]> input_107_cast_fp16 = mul(x = out_129_cast_fp16, y = t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97617792))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99977152))))[name = string("t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_75_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_75_cast_fp16")];
            string var_5138_mode_0 = const()[name = string("op_5138_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_5138_cast_fp16 = gelu(mode = var_5138_mode_0, x = linear_75_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99983360))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102342720))))[name = string("t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_76_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_111_cast_fp16 = mul(x = var_5138_cast_fp16, y = linear_76_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102348928))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104708288))))[name = string("t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5146_cast_fp16 = abs(x = linear_77_cast_fp16)[name = string("op_5146_cast_fp16")];
            bool var_5149 = const()[name = string("op_5149"), val = bool(true)];
            tensor<int32, [1]> var_5150_axes_0 = const()[name = string("op_5150_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5150_cast_fp16 = reduce_max(axes = var_5150_axes_0, keep_dims = var_5149, x = var_5146_cast_fp16)[name = string("op_5150_cast_fp16")];
            fp16 var_5151_to_fp16 = const()[name = string("op_5151_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_157_cast_fp16 = maximum(x = var_5150_cast_fp16, y = var_5151_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_153_cast_fp16 = real_div(x = linear_77_cast_fp16, y = s_157_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5154_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5154_cast_fp16")];
            tensor<int32, [1]> var_5159_axes_0 = const()[name = string("op_5159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5159_keep_dims_0 = const()[name = string("op_5159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5159_cast_fp16 = reduce_mean(axes = var_5159_axes_0, keep_dims = var_5159_keep_dims_0, x = var_5154_cast_fp16)[name = string("op_5159_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5160_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5160_cast_fp16")];
            fp32 var_5161_epsilon_0 = const()[name = string("op_5161_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5161_cast_fp16 = inverse(epsilon = var_5161_epsilon_0, x = var_5160_cast_fp16)[name = string("op_5161_cast_fp16")];
            fp16 var_5162_to_fp16 = const()[name = string("op_5162_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5163_cast_fp16 = mul(x = var_5161_cast_fp16, y = var_5162_to_fp16)[name = string("op_5163_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5165_cast_fp16 = add(x = var_5159_cast_fp16, y = var_5163_cast_fp16)[name = string("op_5165_cast_fp16")];
            fp32 var_5166_epsilon_0 = const()[name = string("op_5166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5166_cast_fp16 = rsqrt(epsilon = var_5166_epsilon_0, x = var_5165_cast_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_131_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5166_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104709888)))];
            tensor<fp16, [1, 1260, 768]> var_5168_cast_fp16 = mul(x = out_131_cast_fp16, y = t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5168_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_199_cast_fp16 = add(x = x_195_cast_fp16, y = var_5168_cast_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5171_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5171_cast_fp16")];
            bool var_5174 = const()[name = string("op_5174"), val = bool(true)];
            tensor<int32, [1]> var_5175_axes_0 = const()[name = string("op_5175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5175_cast_fp16 = reduce_max(axes = var_5175_axes_0, keep_dims = var_5174, x = var_5171_cast_fp16)[name = string("op_5175_cast_fp16")];
            fp16 var_5176_to_fp16 = const()[name = string("op_5176_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_159_cast_fp16 = maximum(x = var_5175_cast_fp16, y = var_5176_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_155_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_159_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5179_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5179_cast_fp16")];
            tensor<int32, [1]> var_5184_axes_0 = const()[name = string("op_5184_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5184_keep_dims_0 = const()[name = string("op_5184_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5184_cast_fp16 = reduce_mean(axes = var_5184_axes_0, keep_dims = var_5184_keep_dims_0, x = var_5179_cast_fp16)[name = string("op_5184_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5185_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5185_cast_fp16")];
            fp32 var_5186_epsilon_0 = const()[name = string("op_5186_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5186_cast_fp16 = inverse(epsilon = var_5186_epsilon_0, x = var_5185_cast_fp16)[name = string("op_5186_cast_fp16")];
            fp16 var_5187_to_fp16 = const()[name = string("op_5187_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5188_cast_fp16 = mul(x = var_5186_cast_fp16, y = var_5187_to_fp16)[name = string("op_5188_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5190_cast_fp16 = add(x = var_5184_cast_fp16, y = var_5188_cast_fp16)[name = string("op_5190_cast_fp16")];
            fp32 var_5191_epsilon_0 = const()[name = string("op_5191_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5191_cast_fp16 = rsqrt(epsilon = var_5191_epsilon_0, x = var_5190_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_133_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5191_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104711488)))];
            tensor<fp16, [1, 1260, 768]> input_113_cast_fp16 = mul(x = out_133_cast_fp16, y = t_encoder_layers_11_input_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104713088))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105302976))))[name = string("t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_5201 = const()[name = string("op_5201"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_201_cast_fp16 = reshape(shape = var_5201, x = linear_78_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5203_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5203_cast_fp16")];
            bool var_5206 = const()[name = string("op_5206"), val = bool(true)];
            tensor<int32, [1]> var_5207_axes_0 = const()[name = string("op_5207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5207_cast_fp16 = reduce_max(axes = var_5207_axes_0, keep_dims = var_5206, x = var_5203_cast_fp16)[name = string("op_5207_cast_fp16")];
            fp16 var_5208_to_fp16 = const()[name = string("op_5208_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_161_cast_fp16 = maximum(x = var_5207_cast_fp16, y = var_5208_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_157_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_161_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5211_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5211_cast_fp16")];
            tensor<int32, [1]> var_5216_axes_0 = const()[name = string("op_5216_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5216_keep_dims_0 = const()[name = string("op_5216_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5216_cast_fp16 = reduce_mean(axes = var_5216_axes_0, keep_dims = var_5216_keep_dims_0, x = var_5211_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5217_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5217_cast_fp16")];
            fp32 var_5218_epsilon_0 = const()[name = string("op_5218_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5218_cast_fp16 = inverse(epsilon = var_5218_epsilon_0, x = var_5217_cast_fp16)[name = string("op_5218_cast_fp16")];
            fp16 var_5219_to_fp16 = const()[name = string("op_5219_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5220_cast_fp16 = mul(x = var_5218_cast_fp16, y = var_5219_to_fp16)[name = string("op_5220_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5222_cast_fp16 = add(x = var_5216_cast_fp16, y = var_5220_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5223_cast_fp16 = rsqrt(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_135_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5223_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20121856)))];
            tensor<fp16, [1, 1260, 12, 64]> x_207_cast_fp16 = mul(x = out_135_cast_fp16, y = t_encoder_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_207_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105304576))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105894464))))[name = string("t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_79_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_5233 = const()[name = string("op_5233"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_203_cast_fp16 = reshape(shape = var_5233, x = linear_79_cast_fp16)[name = string("x_203_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5235_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5235_cast_fp16")];
            bool var_5238 = const()[name = string("op_5238"), val = bool(true)];
            tensor<int32, [1]> var_5239_axes_0 = const()[name = string("op_5239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5239_cast_fp16 = reduce_max(axes = var_5239_axes_0, keep_dims = var_5238, x = var_5235_cast_fp16)[name = string("op_5239_cast_fp16")];
            fp16 var_5240_to_fp16 = const()[name = string("op_5240_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_163_cast_fp16 = maximum(x = var_5239_cast_fp16, y = var_5240_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_159_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_163_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5243_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5243_cast_fp16")];
            tensor<int32, [1]> var_5248_axes_0 = const()[name = string("op_5248_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5248_keep_dims_0 = const()[name = string("op_5248_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5248_cast_fp16 = reduce_mean(axes = var_5248_axes_0, keep_dims = var_5248_keep_dims_0, x = var_5243_cast_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5249_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5249_cast_fp16")];
            fp32 var_5250_epsilon_0 = const()[name = string("op_5250_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5250_cast_fp16 = inverse(epsilon = var_5250_epsilon_0, x = var_5249_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp16 var_5251_to_fp16 = const()[name = string("op_5251_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5252_cast_fp16 = mul(x = var_5250_cast_fp16, y = var_5251_to_fp16)[name = string("op_5252_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5254_cast_fp16 = add(x = var_5248_cast_fp16, y = var_5252_cast_fp16)[name = string("op_5254_cast_fp16")];
            fp32 var_5255_epsilon_0 = const()[name = string("op_5255_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5255_cast_fp16 = rsqrt(epsilon = var_5255_epsilon_0, x = var_5254_cast_fp16)[name = string("op_5255_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_137_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5255_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713536)))];
            tensor<fp16, [1, 1260, 12, 64]> x_209_cast_fp16 = mul(x = out_137_cast_fp16, y = t_encoder_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105896064))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106485952))))[name = string("t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [4]> var_5265 = const()[name = string("op_5265"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_205_cast_fp16 = reshape(shape = var_5265, x = linear_80_cast_fp16)[name = string("x_205_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5267_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5267_cast_fp16")];
            bool var_5270 = const()[name = string("op_5270"), val = bool(true)];
            tensor<int32, [1]> var_5271_axes_0 = const()[name = string("op_5271_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5271_cast_fp16 = reduce_max(axes = var_5271_axes_0, keep_dims = var_5270, x = var_5267_cast_fp16)[name = string("op_5271_cast_fp16")];
            fp16 var_5272_to_fp16 = const()[name = string("op_5272_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_165_cast_fp16 = maximum(x = var_5271_cast_fp16, y = var_5272_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_161_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_165_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5275_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5275_cast_fp16")];
            tensor<int32, [1]> var_5280_axes_0 = const()[name = string("op_5280_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5280_keep_dims_0 = const()[name = string("op_5280_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5280_cast_fp16 = reduce_mean(axes = var_5280_axes_0, keep_dims = var_5280_keep_dims_0, x = var_5275_cast_fp16)[name = string("op_5280_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5281_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp32 var_5282_epsilon_0 = const()[name = string("op_5282_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5282_cast_fp16 = inverse(epsilon = var_5282_epsilon_0, x = var_5281_cast_fp16)[name = string("op_5282_cast_fp16")];
            fp16 var_5283_to_fp16 = const()[name = string("op_5283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5284_cast_fp16 = mul(x = var_5282_cast_fp16, y = var_5283_to_fp16)[name = string("op_5284_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5286_cast_fp16 = add(x = var_5280_cast_fp16, y = var_5284_cast_fp16)[name = string("op_5286_cast_fp16")];
            fp32 var_5287_epsilon_0 = const()[name = string("op_5287_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5287_cast_fp16 = rsqrt(epsilon = var_5287_epsilon_0, x = var_5286_cast_fp16)[name = string("op_5287_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_45_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5287_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<int32, [4]> z_89_begin_0 = const()[name = string("z_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_89_end_0 = const()[name = string("z_89_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_89_end_mask_0 = const()[name = string("z_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_89_cast_fp16 = slice_by_index(begin = z_89_begin_0, end = z_89_end_0, end_mask = z_89_end_mask_0, x = x_207_cast_fp16)[name = string("z_89_cast_fp16")];
            tensor<int32, [4]> var_5298_begin_0 = const()[name = string("op_5298_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5298_end_0 = const()[name = string("op_5298_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5298_end_mask_0 = const()[name = string("op_5298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5298_cast_fp16 = slice_by_index(begin = var_5298_begin_0, end = var_5298_end_0, end_mask = var_5298_end_mask_0, x = z_89_cast_fp16)[name = string("op_5298_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5299_cast_fp16 = mul(x = var_5298_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_5299_cast_fp16")];
            tensor<int32, [4]> var_5304_begin_0 = const()[name = string("op_5304_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5304_end_0 = const()[name = string("op_5304_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5304_end_mask_0 = const()[name = string("op_5304_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5304_cast_fp16 = slice_by_index(begin = var_5304_begin_0, end = var_5304_end_0, end_mask = var_5304_end_mask_0, x = z_89_cast_fp16)[name = string("op_5304_cast_fp16")];
            int32 var_5306 = const()[name = string("op_5306"), val = int32(-1)];
            bool rot_89_interleave_0 = const()[name = string("rot_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_89_cast_fp16 = concat(axis = var_5306, interleave = rot_89_interleave_0, values = (var_5299_cast_fp16, var_5304_cast_fp16))[name = string("rot_89_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5308_cast_fp16 = mul(x = z_89_cast_fp16, y = c_1_cast_fp16)[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5309_cast_fp16 = mul(x = rot_89_cast_fp16, y = s_9_cast_fp16)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5311_cast_fp16 = add(x = var_5308_cast_fp16, y = var_5309_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<int32, [4]> z_91_begin_0 = const()[name = string("z_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_91_end_0 = const()[name = string("z_91_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_91_end_mask_0 = const()[name = string("z_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_91_cast_fp16 = slice_by_index(begin = z_91_begin_0, end = z_91_end_0, end_mask = z_91_end_mask_0, x = x_207_cast_fp16)[name = string("z_91_cast_fp16")];
            tensor<int32, [4]> var_5321_begin_0 = const()[name = string("op_5321_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5321_end_0 = const()[name = string("op_5321_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5321_end_mask_0 = const()[name = string("op_5321_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5321_cast_fp16 = slice_by_index(begin = var_5321_begin_0, end = var_5321_end_0, end_mask = var_5321_end_mask_0, x = z_91_cast_fp16)[name = string("op_5321_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5322_cast_fp16 = mul(x = var_5321_cast_fp16, y = const_45_promoted_to_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, 0])];
            tensor<int32, [4]> var_5327_end_0 = const()[name = string("op_5327_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5327_end_mask_0 = const()[name = string("op_5327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 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 = z_91_cast_fp16)[name = string("op_5327_cast_fp16")];
            int32 var_5329 = const()[name = string("op_5329"), val = int32(-1)];
            bool rot_91_interleave_0 = const()[name = string("rot_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_91_cast_fp16 = concat(axis = var_5329, interleave = rot_91_interleave_0, values = (var_5322_cast_fp16, var_5327_cast_fp16))[name = string("rot_91_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5331_cast_fp16 = mul(x = z_91_cast_fp16, y = c_cast_fp16)[name = string("op_5331_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5332_cast_fp16 = mul(x = rot_91_cast_fp16, y = s_11_cast_fp16)[name = string("op_5332_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5334_cast_fp16 = add(x = var_5331_cast_fp16, y = var_5332_cast_fp16)[name = string("op_5334_cast_fp16")];
            int32 var_5336 = const()[name = string("op_5336"), val = int32(-1)];
            bool var_5337_interleave_0 = const()[name = string("op_5337_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_5337_cast_fp16 = concat(axis = var_5336, interleave = var_5337_interleave_0, values = (var_5311_cast_fp16, var_5334_cast_fp16))[name = string("op_5337_cast_fp16")];
            tensor<int32, [4]> z_93_begin_0 = const()[name = string("z_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_93_end_0 = const()[name = string("z_93_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_93_end_mask_0 = const()[name = string("z_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_93_cast_fp16 = slice_by_index(begin = z_93_begin_0, end = z_93_end_0, end_mask = z_93_end_mask_0, x = x_209_cast_fp16)[name = string("z_93_cast_fp16")];
            tensor<int32, [4]> var_5350_begin_0 = const()[name = string("op_5350_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5350_end_0 = const()[name = string("op_5350_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5350_end_mask_0 = const()[name = string("op_5350_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5350_cast_fp16 = slice_by_index(begin = var_5350_begin_0, end = var_5350_end_0, end_mask = var_5350_end_mask_0, x = z_93_cast_fp16)[name = string("op_5350_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5351_cast_fp16 = mul(x = var_5350_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_5351_cast_fp16")];
            tensor<int32, [4]> var_5356_begin_0 = const()[name = string("op_5356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5356_end_0 = const()[name = string("op_5356_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5356_end_mask_0 = const()[name = string("op_5356_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5356_cast_fp16 = slice_by_index(begin = var_5356_begin_0, end = var_5356_end_0, end_mask = var_5356_end_mask_0, x = z_93_cast_fp16)[name = string("op_5356_cast_fp16")];
            int32 var_5358 = const()[name = string("op_5358"), val = int32(-1)];
            bool rot_93_interleave_0 = const()[name = string("rot_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_93_cast_fp16 = concat(axis = var_5358, interleave = rot_93_interleave_0, values = (var_5351_cast_fp16, var_5356_cast_fp16))[name = string("rot_93_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5360_cast_fp16 = mul(x = z_93_cast_fp16, y = c_1_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5361_cast_fp16 = mul(x = rot_93_cast_fp16, y = s_9_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5363_cast_fp16 = add(x = var_5360_cast_fp16, y = var_5361_cast_fp16)[name = string("op_5363_cast_fp16")];
            tensor<int32, [4]> z_95_begin_0 = const()[name = string("z_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_95_end_0 = const()[name = string("z_95_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_95_end_mask_0 = const()[name = string("z_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_95_cast_fp16 = slice_by_index(begin = z_95_begin_0, end = z_95_end_0, end_mask = z_95_end_mask_0, x = x_209_cast_fp16)[name = string("z_95_cast_fp16")];
            tensor<int32, [4]> var_5373_begin_0 = const()[name = string("op_5373_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5373_end_0 = const()[name = string("op_5373_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5373_end_mask_0 = const()[name = string("op_5373_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5373_cast_fp16 = slice_by_index(begin = var_5373_begin_0, end = var_5373_end_0, end_mask = var_5373_end_mask_0, x = z_95_cast_fp16)[name = string("op_5373_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5374_cast_fp16 = mul(x = var_5373_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [4]> var_5379_begin_0 = const()[name = string("op_5379_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5379_end_0 = const()[name = string("op_5379_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5379_end_mask_0 = const()[name = string("op_5379_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5379_cast_fp16 = slice_by_index(begin = var_5379_begin_0, end = var_5379_end_0, end_mask = var_5379_end_mask_0, x = z_95_cast_fp16)[name = string("op_5379_cast_fp16")];
            int32 var_5381 = const()[name = string("op_5381"), val = int32(-1)];
            bool rot_95_interleave_0 = const()[name = string("rot_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_95_cast_fp16 = concat(axis = var_5381, interleave = rot_95_interleave_0, values = (var_5374_cast_fp16, var_5379_cast_fp16))[name = string("rot_95_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5383_cast_fp16 = mul(x = z_95_cast_fp16, y = c_cast_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5384_cast_fp16 = mul(x = rot_95_cast_fp16, y = s_11_cast_fp16)[name = string("op_5384_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5386_cast_fp16 = add(x = var_5383_cast_fp16, y = var_5384_cast_fp16)[name = string("op_5386_cast_fp16")];
            int32 var_5388 = const()[name = string("op_5388"), val = int32(-1)];
            bool var_5389_interleave_0 = const()[name = string("op_5389_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_5389_cast_fp16 = concat(axis = var_5388, interleave = var_5389_interleave_0, values = (var_5363_cast_fp16, var_5386_cast_fp16))[name = string("op_5389_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5399_transpose_x_0 = const()[name = string("op_5399_transpose_x_0"), val = bool(false)];
            bool var_5399_transpose_y_0 = const()[name = string("op_5399_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_87 = transpose(perm = transpose_87_perm_0, x = var_5389_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_86 = transpose(perm = transpose_86_perm_0, x = var_5337_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 12, 1260, 1260]> var_5399_cast_fp16 = matmul(transpose_x = var_5399_transpose_x_0, transpose_y = var_5399_transpose_y_0, x = transpose_86, y = transpose_87)[name = string("op_5399_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_5401_cast_fp16 = add(x = var_5399_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5401_cast_fp16")];
            int32 var_5402 = const()[name = string("op_5402"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_161_cast_fp16 = softmax(axis = var_5402, x = var_5401_cast_fp16)[name = string("w_161_cast_fp16")];
            bool var_5405_transpose_x_0 = const()[name = string("op_5405_transpose_x_0"), val = bool(false)];
            bool var_5405_transpose_y_0 = const()[name = string("op_5405_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = v_45_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 12, 1260, 64]> var_5405_cast_fp16 = matmul(transpose_x = var_5405_transpose_x_0, transpose_y = var_5405_transpose_y_0, x = w_161_cast_fp16, y = v_47_cast_fp16)[name = string("op_5405_cast_fp16")];
            tensor<int32, [4]> var_5408_perm_0 = const()[name = string("op_5408_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5412 = const()[name = string("op_5412"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_5408_cast_fp16 = transpose(perm = var_5408_perm_0, x = var_5405_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 1260, 768]> input_115_cast_fp16 = reshape(shape = var_5412, x = var_5408_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106487552))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107077440))))[name = string("t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_81_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5417_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5417_cast_fp16")];
            bool var_5420 = const()[name = string("op_5420"), val = bool(true)];
            tensor<int32, [1]> var_5421_axes_0 = const()[name = string("op_5421_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5421_cast_fp16 = reduce_max(axes = var_5421_axes_0, keep_dims = var_5420, x = var_5417_cast_fp16)[name = string("op_5421_cast_fp16")];
            fp16 var_5422_to_fp16 = const()[name = string("op_5422_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_167_cast_fp16 = maximum(x = var_5421_cast_fp16, y = var_5422_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_163_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_167_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5425_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5425_cast_fp16")];
            tensor<int32, [1]> var_5430_axes_0 = const()[name = string("op_5430_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5430_keep_dims_0 = const()[name = string("op_5430_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5430_cast_fp16 = reduce_mean(axes = var_5430_axes_0, keep_dims = var_5430_keep_dims_0, x = var_5425_cast_fp16)[name = string("op_5430_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5431_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5431_cast_fp16")];
            fp32 var_5432_epsilon_0 = const()[name = string("op_5432_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5432_cast_fp16 = inverse(epsilon = var_5432_epsilon_0, x = var_5431_cast_fp16)[name = string("op_5432_cast_fp16")];
            fp16 var_5433_to_fp16 = const()[name = string("op_5433_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = string("op_5434_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5436_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5434_cast_fp16)[name = string("op_5436_cast_fp16")];
            fp32 var_5437_epsilon_0 = const()[name = string("op_5437_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5437_cast_fp16 = rsqrt(epsilon = var_5437_epsilon_0, x = var_5436_cast_fp16)[name = string("op_5437_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_139_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5437_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107079040)))];
            tensor<fp16, [1, 1260, 768]> var_5439_cast_fp16 = mul(x = out_139_cast_fp16, y = t_encoder_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_5439_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_213_cast_fp16 = add(x = x_199_cast_fp16, y = var_5439_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5442_cast_fp16 = abs(x = x_213_cast_fp16)[name = string("op_5442_cast_fp16")];
            bool var_5445 = const()[name = string("op_5445"), val = bool(true)];
            tensor<int32, [1]> var_5446_axes_0 = const()[name = string("op_5446_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5446_cast_fp16 = reduce_max(axes = var_5446_axes_0, keep_dims = var_5445, x = var_5442_cast_fp16)[name = string("op_5446_cast_fp16")];
            fp16 var_5447_to_fp16 = const()[name = string("op_5447_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_169_cast_fp16 = maximum(x = var_5446_cast_fp16, y = var_5447_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_165_cast_fp16 = real_div(x = x_213_cast_fp16, y = s_169_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5450_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5450_cast_fp16")];
            tensor<int32, [1]> var_5455_axes_0 = const()[name = string("op_5455_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5455_keep_dims_0 = const()[name = string("op_5455_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5455_cast_fp16 = reduce_mean(axes = var_5455_axes_0, keep_dims = var_5455_keep_dims_0, x = var_5450_cast_fp16)[name = string("op_5455_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5456_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5456_cast_fp16")];
            fp32 var_5457_epsilon_0 = const()[name = string("op_5457_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5457_cast_fp16 = inverse(epsilon = var_5457_epsilon_0, x = var_5456_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5459_cast_fp16 = mul(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("op_5459_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5461_cast_fp16 = add(x = var_5455_cast_fp16, y = var_5459_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, 1260, 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, 1260, 768]> out_141_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5462_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107080640)))];
            tensor<fp16, [1, 1260, 768]> input_117_cast_fp16 = mul(x = out_141_cast_fp16, y = t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107082240))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109441600))))[name = string("t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_82_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_82_cast_fp16")];
            string var_5469_mode_0 = const()[name = string("op_5469_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_5469_cast_fp16 = gelu(mode = var_5469_mode_0, x = linear_82_cast_fp16)[name = string("op_5469_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109447808))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111807168))))[name = string("t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_83_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_121_cast_fp16 = mul(x = var_5469_cast_fp16, y = linear_83_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111813376))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114172736))))[name = string("t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5477_cast_fp16 = abs(x = linear_84_cast_fp16)[name = string("op_5477_cast_fp16")];
            bool var_5480 = const()[name = string("op_5480"), val = bool(true)];
            tensor<int32, [1]> var_5481_axes_0 = const()[name = string("op_5481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5481_cast_fp16 = reduce_max(axes = var_5481_axes_0, keep_dims = var_5480, x = var_5477_cast_fp16)[name = string("op_5481_cast_fp16")];
            fp16 var_5482_to_fp16 = const()[name = string("op_5482_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_171_cast_fp16 = maximum(x = var_5481_cast_fp16, y = var_5482_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_167_cast_fp16 = real_div(x = linear_84_cast_fp16, y = s_171_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5485_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5485_cast_fp16")];
            tensor<int32, [1]> var_5490_axes_0 = const()[name = string("op_5490_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5490_keep_dims_0 = const()[name = string("op_5490_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5490_cast_fp16 = reduce_mean(axes = var_5490_axes_0, keep_dims = var_5490_keep_dims_0, x = var_5485_cast_fp16)[name = string("op_5490_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5491_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5491_cast_fp16")];
            fp32 var_5492_epsilon_0 = const()[name = string("op_5492_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5492_cast_fp16 = inverse(epsilon = var_5492_epsilon_0, x = var_5491_cast_fp16)[name = string("op_5492_cast_fp16")];
            fp16 var_5493_to_fp16 = const()[name = string("op_5493_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5494_cast_fp16 = mul(x = var_5492_cast_fp16, y = var_5493_to_fp16)[name = string("op_5494_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5496_cast_fp16 = add(x = var_5490_cast_fp16, y = var_5494_cast_fp16)[name = string("op_5496_cast_fp16")];
            fp32 var_5497_epsilon_0 = const()[name = string("op_5497_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5497_cast_fp16 = rsqrt(epsilon = var_5497_epsilon_0, x = var_5496_cast_fp16)[name = string("op_5497_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_143_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5497_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114174336)))];
            tensor<fp16, [1, 1260, 768]> var_5499_cast_fp16 = mul(x = out_143_cast_fp16, y = t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_217_cast_fp16 = add(x = x_213_cast_fp16, y = var_5499_cast_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5502_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_5502_cast_fp16")];
            bool var_5505 = const()[name = string("op_5505"), val = bool(true)];
            tensor<int32, [1]> var_5506_axes_0 = const()[name = string("op_5506_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5506_cast_fp16 = reduce_max(axes = var_5506_axes_0, keep_dims = var_5505, x = var_5502_cast_fp16)[name = string("op_5506_cast_fp16")];
            fp16 var_5507_to_fp16 = const()[name = string("op_5507_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_173_cast_fp16 = maximum(x = var_5506_cast_fp16, y = var_5507_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_169_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_173_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5510_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [1]> var_5515_axes_0 = const()[name = string("op_5515_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5515_keep_dims_0 = const()[name = string("op_5515_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5515_cast_fp16 = reduce_mean(axes = var_5515_axes_0, keep_dims = var_5515_keep_dims_0, x = var_5510_cast_fp16)[name = string("op_5515_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5516_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5516_cast_fp16")];
            fp32 var_5517_epsilon_0 = const()[name = string("op_5517_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5517_cast_fp16 = inverse(epsilon = var_5517_epsilon_0, x = var_5516_cast_fp16)[name = string("op_5517_cast_fp16")];
            fp16 var_5518_to_fp16 = const()[name = string("op_5518_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5519_cast_fp16 = mul(x = var_5517_cast_fp16, y = var_5518_to_fp16)[name = string("op_5519_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5521_cast_fp16 = add(x = var_5515_cast_fp16, y = var_5519_cast_fp16)[name = string("op_5521_cast_fp16")];
            fp32 var_5522_epsilon_0 = const()[name = string("op_5522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5522_cast_fp16 = rsqrt(epsilon = var_5522_epsilon_0, x = var_5521_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_145_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5522_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114175936)))];
            tensor<fp16, [1, 1260, 768]> input_123_cast_fp16 = mul(x = out_145_cast_fp16, y = t_encoder_layers_12_input_layernorm_weight_to_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114177536))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114767424))))[name = string("t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_85_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_5532 = const()[name = string("op_5532"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_219_cast_fp16 = reshape(shape = var_5532, x = linear_85_cast_fp16)[name = string("x_219_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5534_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_5534_cast_fp16")];
            bool var_5537 = const()[name = string("op_5537"), val = bool(true)];
            tensor<int32, [1]> var_5538_axes_0 = const()[name = string("op_5538_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5538_cast_fp16 = reduce_max(axes = var_5538_axes_0, keep_dims = var_5537, x = var_5534_cast_fp16)[name = string("op_5538_cast_fp16")];
            fp16 var_5539_to_fp16 = const()[name = string("op_5539_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_175_cast_fp16 = maximum(x = var_5538_cast_fp16, y = var_5539_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_171_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_175_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5542_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5542_cast_fp16")];
            tensor<int32, [1]> var_5547_axes_0 = const()[name = string("op_5547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5547_keep_dims_0 = const()[name = string("op_5547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5547_cast_fp16 = reduce_mean(axes = var_5547_axes_0, keep_dims = var_5547_keep_dims_0, x = var_5542_cast_fp16)[name = string("op_5547_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5548_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5548_cast_fp16")];
            fp32 var_5549_epsilon_0 = const()[name = string("op_5549_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5549_cast_fp16 = inverse(epsilon = var_5549_epsilon_0, x = var_5548_cast_fp16)[name = string("op_5549_cast_fp16")];
            fp16 var_5550_to_fp16 = const()[name = string("op_5550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5553_cast_fp16 = add(x = var_5547_cast_fp16, y = var_5551_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp32 var_5554_epsilon_0 = const()[name = string("op_5554_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5554_cast_fp16 = rsqrt(epsilon = var_5554_epsilon_0, x = var_5553_cast_fp16)[name = string("op_5554_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_147_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5554_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769024)))];
            tensor<fp16, [1, 1260, 12, 64]> x_225_cast_fp16 = mul(x = out_147_cast_fp16, y = t_encoder_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_225_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769216))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115359104))))[name = string("t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_5564 = const()[name = string("op_5564"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_221_cast_fp16 = reshape(shape = var_5564, x = linear_86_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5566_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_5566_cast_fp16")];
            bool var_5569 = const()[name = string("op_5569"), val = bool(true)];
            tensor<int32, [1]> var_5570_axes_0 = const()[name = string("op_5570_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5570_cast_fp16 = reduce_max(axes = var_5570_axes_0, keep_dims = var_5569, x = var_5566_cast_fp16)[name = string("op_5570_cast_fp16")];
            fp16 var_5571_to_fp16 = const()[name = string("op_5571_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_177_cast_fp16 = maximum(x = var_5570_cast_fp16, y = var_5571_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_173_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_177_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5574_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<int32, [1]> var_5579_axes_0 = const()[name = string("op_5579_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5579_keep_dims_0 = const()[name = string("op_5579_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5579_cast_fp16 = reduce_mean(axes = var_5579_axes_0, keep_dims = var_5579_keep_dims_0, x = var_5574_cast_fp16)[name = string("op_5579_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5580_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5581_cast_fp16 = inverse(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            fp16 var_5582_to_fp16 = const()[name = string("op_5582_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5583_cast_fp16 = mul(x = var_5581_cast_fp16, y = var_5582_to_fp16)[name = string("op_5583_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5585_cast_fp16 = add(x = var_5579_cast_fp16, y = var_5583_cast_fp16)[name = string("op_5585_cast_fp16")];
            fp32 var_5586_epsilon_0 = const()[name = string("op_5586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5586_cast_fp16 = rsqrt(epsilon = var_5586_epsilon_0, x = var_5585_cast_fp16)[name = string("op_5586_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_149_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5586_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360704)))];
            tensor<fp16, [1, 1260, 12, 64]> x_227_cast_fp16 = mul(x = out_149_cast_fp16, y = t_encoder_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360896))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115950784))))[name = string("t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_87_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<int32, [4]> var_5596 = const()[name = string("op_5596"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_223_cast_fp16 = reshape(shape = var_5596, x = linear_87_cast_fp16)[name = string("x_223_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5598_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_5598_cast_fp16")];
            bool var_5601 = const()[name = string("op_5601"), val = bool(true)];
            tensor<int32, [1]> var_5602_axes_0 = const()[name = string("op_5602_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5602_cast_fp16 = reduce_max(axes = var_5602_axes_0, keep_dims = var_5601, x = var_5598_cast_fp16)[name = string("op_5602_cast_fp16")];
            fp16 var_5603_to_fp16 = const()[name = string("op_5603_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_179_cast_fp16 = maximum(x = var_5602_cast_fp16, y = var_5603_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_175_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_179_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5606_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<int32, [1]> var_5611_axes_0 = const()[name = string("op_5611_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5611_keep_dims_0 = const()[name = string("op_5611_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5611_cast_fp16 = reduce_mean(axes = var_5611_axes_0, keep_dims = var_5611_keep_dims_0, x = var_5606_cast_fp16)[name = string("op_5611_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5612_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5612_cast_fp16")];
            fp32 var_5613_epsilon_0 = const()[name = string("op_5613_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5613_cast_fp16 = inverse(epsilon = var_5613_epsilon_0, x = var_5612_cast_fp16)[name = string("op_5613_cast_fp16")];
            fp16 var_5614_to_fp16 = const()[name = string("op_5614_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5615_cast_fp16 = mul(x = var_5613_cast_fp16, y = var_5614_to_fp16)[name = string("op_5615_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5617_cast_fp16 = add(x = var_5611_cast_fp16, y = var_5615_cast_fp16)[name = string("op_5617_cast_fp16")];
            fp32 var_5618_epsilon_0 = const()[name = string("op_5618_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5618_cast_fp16 = rsqrt(epsilon = var_5618_epsilon_0, x = var_5617_cast_fp16)[name = string("op_5618_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_49_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5618_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<int32, [4]> z_97_begin_0 = const()[name = string("z_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_97_end_0 = const()[name = string("z_97_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_97_end_mask_0 = const()[name = string("z_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_97_cast_fp16 = slice_by_index(begin = z_97_begin_0, end = z_97_end_0, end_mask = z_97_end_mask_0, x = x_225_cast_fp16)[name = string("z_97_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> 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 = z_97_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = z_97_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool rot_97_interleave_0 = const()[name = string("rot_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_97_cast_fp16 = concat(axis = var_5637, interleave = rot_97_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("rot_97_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5639_cast_fp16 = mul(x = z_97_cast_fp16, y = c_1_cast_fp16)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5640_cast_fp16 = mul(x = rot_97_cast_fp16, y = s_9_cast_fp16)[name = string("op_5640_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5642_cast_fp16 = add(x = var_5639_cast_fp16, y = var_5640_cast_fp16)[name = string("op_5642_cast_fp16")];
            tensor<int32, [4]> z_99_begin_0 = const()[name = string("z_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_99_end_0 = const()[name = string("z_99_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_99_end_mask_0 = const()[name = string("z_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_99_cast_fp16 = slice_by_index(begin = z_99_begin_0, end = z_99_end_0, end_mask = z_99_end_mask_0, x = x_225_cast_fp16)[name = string("z_99_cast_fp16")];
            tensor<int32, [4]> var_5652_begin_0 = const()[name = string("op_5652_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5652_end_0 = const()[name = string("op_5652_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5652_end_mask_0 = const()[name = string("op_5652_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5652_cast_fp16 = slice_by_index(begin = var_5652_begin_0, end = var_5652_end_0, end_mask = var_5652_end_mask_0, x = z_99_cast_fp16)[name = string("op_5652_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5653_cast_fp16 = mul(x = var_5652_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_5653_cast_fp16")];
            tensor<int32, [4]> var_5658_begin_0 = const()[name = string("op_5658_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5658_end_0 = const()[name = string("op_5658_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5658_end_mask_0 = const()[name = string("op_5658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5658_cast_fp16 = slice_by_index(begin = var_5658_begin_0, end = var_5658_end_0, end_mask = var_5658_end_mask_0, x = z_99_cast_fp16)[name = string("op_5658_cast_fp16")];
            int32 var_5660 = const()[name = string("op_5660"), val = int32(-1)];
            bool rot_99_interleave_0 = const()[name = string("rot_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_99_cast_fp16 = concat(axis = var_5660, interleave = rot_99_interleave_0, values = (var_5653_cast_fp16, var_5658_cast_fp16))[name = string("rot_99_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5662_cast_fp16 = mul(x = z_99_cast_fp16, y = c_cast_fp16)[name = string("op_5662_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5663_cast_fp16 = mul(x = rot_99_cast_fp16, y = s_11_cast_fp16)[name = string("op_5663_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5665_cast_fp16 = add(x = var_5662_cast_fp16, y = var_5663_cast_fp16)[name = string("op_5665_cast_fp16")];
            int32 var_5667 = const()[name = string("op_5667"), val = int32(-1)];
            bool var_5668_interleave_0 = const()[name = string("op_5668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_5668_cast_fp16 = concat(axis = var_5667, interleave = var_5668_interleave_0, values = (var_5642_cast_fp16, var_5665_cast_fp16))[name = string("op_5668_cast_fp16")];
            tensor<int32, [4]> z_101_begin_0 = const()[name = string("z_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_101_end_0 = const()[name = string("z_101_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_101_end_mask_0 = const()[name = string("z_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_101_cast_fp16 = slice_by_index(begin = z_101_begin_0, end = z_101_end_0, end_mask = z_101_end_mask_0, x = x_227_cast_fp16)[name = string("z_101_cast_fp16")];
            tensor<int32, [4]> var_5681_begin_0 = const()[name = string("op_5681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5681_end_0 = const()[name = string("op_5681_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5681_end_mask_0 = const()[name = string("op_5681_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5681_cast_fp16 = slice_by_index(begin = var_5681_begin_0, end = var_5681_end_0, end_mask = var_5681_end_mask_0, x = z_101_cast_fp16)[name = string("op_5681_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5682_cast_fp16 = mul(x = var_5681_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_5682_cast_fp16")];
            tensor<int32, [4]> var_5687_begin_0 = const()[name = string("op_5687_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5687_end_0 = const()[name = string("op_5687_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5687_end_mask_0 = const()[name = string("op_5687_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5687_cast_fp16 = slice_by_index(begin = var_5687_begin_0, end = var_5687_end_0, end_mask = var_5687_end_mask_0, x = z_101_cast_fp16)[name = string("op_5687_cast_fp16")];
            int32 var_5689 = const()[name = string("op_5689"), val = int32(-1)];
            bool rot_101_interleave_0 = const()[name = string("rot_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_101_cast_fp16 = concat(axis = var_5689, interleave = rot_101_interleave_0, values = (var_5682_cast_fp16, var_5687_cast_fp16))[name = string("rot_101_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5691_cast_fp16 = mul(x = z_101_cast_fp16, y = c_1_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5692_cast_fp16 = mul(x = rot_101_cast_fp16, y = s_9_cast_fp16)[name = string("op_5692_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5694_cast_fp16 = add(x = var_5691_cast_fp16, y = var_5692_cast_fp16)[name = string("op_5694_cast_fp16")];
            tensor<int32, [4]> z_103_begin_0 = const()[name = string("z_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_103_end_0 = const()[name = string("z_103_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_103_end_mask_0 = const()[name = string("z_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_103_cast_fp16 = slice_by_index(begin = z_103_begin_0, end = z_103_end_0, end_mask = z_103_end_mask_0, x = x_227_cast_fp16)[name = string("z_103_cast_fp16")];
            tensor<int32, [4]> var_5704_begin_0 = const()[name = string("op_5704_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5704_end_0 = const()[name = string("op_5704_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5704_end_mask_0 = const()[name = string("op_5704_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5704_cast_fp16 = slice_by_index(begin = var_5704_begin_0, end = var_5704_end_0, end_mask = var_5704_end_mask_0, x = z_103_cast_fp16)[name = string("op_5704_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5705_cast_fp16 = mul(x = var_5704_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_5705_cast_fp16")];
            tensor<int32, [4]> var_5710_begin_0 = const()[name = string("op_5710_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5710_end_0 = const()[name = string("op_5710_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5710_end_mask_0 = const()[name = string("op_5710_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5710_cast_fp16 = slice_by_index(begin = var_5710_begin_0, end = var_5710_end_0, end_mask = var_5710_end_mask_0, x = z_103_cast_fp16)[name = string("op_5710_cast_fp16")];
            int32 var_5712 = const()[name = string("op_5712"), val = int32(-1)];
            bool rot_103_interleave_0 = const()[name = string("rot_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_103_cast_fp16 = concat(axis = var_5712, interleave = rot_103_interleave_0, values = (var_5705_cast_fp16, var_5710_cast_fp16))[name = string("rot_103_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5714_cast_fp16 = mul(x = z_103_cast_fp16, y = c_cast_fp16)[name = string("op_5714_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5715_cast_fp16 = mul(x = rot_103_cast_fp16, y = s_11_cast_fp16)[name = string("op_5715_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5717_cast_fp16 = add(x = var_5714_cast_fp16, y = var_5715_cast_fp16)[name = string("op_5717_cast_fp16")];
            int32 var_5719 = const()[name = string("op_5719"), val = int32(-1)];
            bool var_5720_interleave_0 = const()[name = string("op_5720_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_5720_cast_fp16 = concat(axis = var_5719, interleave = var_5720_interleave_0, values = (var_5694_cast_fp16, var_5717_cast_fp16))[name = string("op_5720_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5730_transpose_x_0 = const()[name = string("op_5730_transpose_x_0"), val = bool(false)];
            bool var_5730_transpose_y_0 = const()[name = string("op_5730_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_89 = transpose(perm = transpose_89_perm_0, x = var_5720_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_88 = transpose(perm = transpose_88_perm_0, x = var_5668_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 12, 1260, 1260]> var_5730_cast_fp16 = matmul(transpose_x = var_5730_transpose_x_0, transpose_y = var_5730_transpose_y_0, x = transpose_88, y = transpose_89)[name = string("op_5730_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_5732_cast_fp16 = add(x = var_5730_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5732_cast_fp16")];
            int32 var_5733 = const()[name = string("op_5733"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_175_cast_fp16 = softmax(axis = var_5733, x = var_5732_cast_fp16)[name = string("w_175_cast_fp16")];
            bool var_5736_transpose_x_0 = const()[name = string("op_5736_transpose_x_0"), val = bool(false)];
            bool var_5736_transpose_y_0 = const()[name = string("op_5736_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = v_49_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 12, 1260, 64]> var_5736_cast_fp16 = matmul(transpose_x = var_5736_transpose_x_0, transpose_y = var_5736_transpose_y_0, x = w_175_cast_fp16, y = v_51_cast_fp16)[name = string("op_5736_cast_fp16")];
            tensor<int32, [4]> var_5739_perm_0 = const()[name = string("op_5739_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5743 = const()[name = string("op_5743"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_5739_cast_fp16 = transpose(perm = var_5739_perm_0, x = var_5736_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 1260, 768]> input_125_cast_fp16 = reshape(shape = var_5743, x = var_5739_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115952384))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116542272))))[name = string("t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5748_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5748_cast_fp16")];
            bool var_5751 = const()[name = string("op_5751"), val = bool(true)];
            tensor<int32, [1]> var_5752_axes_0 = const()[name = string("op_5752_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5752_cast_fp16 = reduce_max(axes = var_5752_axes_0, keep_dims = var_5751, x = var_5748_cast_fp16)[name = string("op_5752_cast_fp16")];
            fp16 var_5753_to_fp16 = const()[name = string("op_5753_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_181_cast_fp16 = maximum(x = var_5752_cast_fp16, y = var_5753_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_177_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_181_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5756_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5756_cast_fp16")];
            tensor<int32, [1]> var_5761_axes_0 = const()[name = string("op_5761_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5761_keep_dims_0 = const()[name = string("op_5761_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5761_cast_fp16 = reduce_mean(axes = var_5761_axes_0, keep_dims = var_5761_keep_dims_0, x = var_5756_cast_fp16)[name = string("op_5761_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5762_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5762_cast_fp16")];
            fp32 var_5763_epsilon_0 = const()[name = string("op_5763_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5763_cast_fp16 = inverse(epsilon = var_5763_epsilon_0, x = var_5762_cast_fp16)[name = string("op_5763_cast_fp16")];
            fp16 var_5764_to_fp16 = const()[name = string("op_5764_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5765_cast_fp16 = mul(x = var_5763_cast_fp16, y = var_5764_to_fp16)[name = string("op_5765_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5767_cast_fp16 = add(x = var_5761_cast_fp16, y = var_5765_cast_fp16)[name = string("op_5767_cast_fp16")];
            fp32 var_5768_epsilon_0 = const()[name = string("op_5768_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5768_cast_fp16 = rsqrt(epsilon = var_5768_epsilon_0, x = var_5767_cast_fp16)[name = string("op_5768_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_151_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5768_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116543872)))];
            tensor<fp16, [1, 1260, 768]> var_5770_cast_fp16 = mul(x = out_151_cast_fp16, y = t_encoder_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_5770_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_231_cast_fp16 = add(x = x_217_cast_fp16, y = var_5770_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5773_cast_fp16 = abs(x = x_231_cast_fp16)[name = string("op_5773_cast_fp16")];
            bool var_5776 = const()[name = string("op_5776"), val = bool(true)];
            tensor<int32, [1]> var_5777_axes_0 = const()[name = string("op_5777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5777_cast_fp16 = reduce_max(axes = var_5777_axes_0, keep_dims = var_5776, x = var_5773_cast_fp16)[name = string("op_5777_cast_fp16")];
            fp16 var_5778_to_fp16 = const()[name = string("op_5778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_183_cast_fp16 = maximum(x = var_5777_cast_fp16, y = var_5778_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_179_cast_fp16 = real_div(x = x_231_cast_fp16, y = s_183_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5781_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5781_cast_fp16")];
            tensor<int32, [1]> var_5786_axes_0 = const()[name = string("op_5786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5786_keep_dims_0 = const()[name = string("op_5786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5786_cast_fp16 = reduce_mean(axes = var_5786_axes_0, keep_dims = var_5786_keep_dims_0, x = var_5781_cast_fp16)[name = string("op_5786_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5787_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp32 var_5788_epsilon_0 = const()[name = string("op_5788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5788_cast_fp16 = inverse(epsilon = var_5788_epsilon_0, x = var_5787_cast_fp16)[name = string("op_5788_cast_fp16")];
            fp16 var_5789_to_fp16 = const()[name = string("op_5789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5790_cast_fp16 = mul(x = var_5788_cast_fp16, y = var_5789_to_fp16)[name = string("op_5790_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5792_cast_fp16 = add(x = var_5786_cast_fp16, y = var_5790_cast_fp16)[name = string("op_5792_cast_fp16")];
            fp32 var_5793_epsilon_0 = const()[name = string("op_5793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5793_cast_fp16 = rsqrt(epsilon = var_5793_epsilon_0, x = var_5792_cast_fp16)[name = string("op_5793_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_153_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5793_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116545472)))];
            tensor<fp16, [1, 1260, 768]> input_127_cast_fp16 = mul(x = out_153_cast_fp16, y = t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116547072))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118906432))))[name = string("t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_89_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_89_cast_fp16")];
            string var_5800_mode_0 = const()[name = string("op_5800_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_5800_cast_fp16 = gelu(mode = var_5800_mode_0, x = linear_89_cast_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118912640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121272000))))[name = string("t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_90_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_131_cast_fp16 = mul(x = var_5800_cast_fp16, y = linear_90_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121278208))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123637568))))[name = string("t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5808_cast_fp16 = abs(x = linear_91_cast_fp16)[name = string("op_5808_cast_fp16")];
            bool var_5811 = const()[name = string("op_5811"), val = bool(true)];
            tensor<int32, [1]> var_5812_axes_0 = const()[name = string("op_5812_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5812_cast_fp16 = reduce_max(axes = var_5812_axes_0, keep_dims = var_5811, x = var_5808_cast_fp16)[name = string("op_5812_cast_fp16")];
            fp16 var_5813_to_fp16 = const()[name = string("op_5813_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_185_cast_fp16 = maximum(x = var_5812_cast_fp16, y = var_5813_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_181_cast_fp16 = real_div(x = linear_91_cast_fp16, y = s_185_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5816_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5816_cast_fp16")];
            tensor<int32, [1]> var_5821_axes_0 = const()[name = string("op_5821_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5821_keep_dims_0 = const()[name = string("op_5821_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5821_cast_fp16 = reduce_mean(axes = var_5821_axes_0, keep_dims = var_5821_keep_dims_0, x = var_5816_cast_fp16)[name = string("op_5821_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5822_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5822_cast_fp16")];
            fp32 var_5823_epsilon_0 = const()[name = string("op_5823_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5823_cast_fp16 = inverse(epsilon = var_5823_epsilon_0, x = var_5822_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp16 var_5824_to_fp16 = const()[name = string("op_5824_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5825_cast_fp16 = mul(x = var_5823_cast_fp16, y = var_5824_to_fp16)[name = string("op_5825_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5827_cast_fp16 = add(x = var_5821_cast_fp16, y = var_5825_cast_fp16)[name = string("op_5827_cast_fp16")];
            fp32 var_5828_epsilon_0 = const()[name = string("op_5828_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5828_cast_fp16 = rsqrt(epsilon = var_5828_epsilon_0, x = var_5827_cast_fp16)[name = string("op_5828_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_155_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5828_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123639168)))];
            tensor<fp16, [1, 1260, 768]> var_5830_cast_fp16 = mul(x = out_155_cast_fp16, y = t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5830_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_235_cast_fp16 = add(x = x_231_cast_fp16, y = var_5830_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5833_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_5833_cast_fp16")];
            bool var_5836 = const()[name = string("op_5836"), val = bool(true)];
            tensor<int32, [1]> var_5837_axes_0 = const()[name = string("op_5837_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_5837_cast_fp16 = reduce_max(axes = var_5837_axes_0, keep_dims = var_5836, x = var_5833_cast_fp16)[name = string("op_5837_cast_fp16")];
            fp16 var_5838_to_fp16 = const()[name = string("op_5838_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_187_cast_fp16 = maximum(x = var_5837_cast_fp16, y = var_5838_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_183_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_187_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_5841_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5841_cast_fp16")];
            tensor<int32, [1]> var_5846_axes_0 = const()[name = string("op_5846_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5846_keep_dims_0 = const()[name = string("op_5846_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_5846_cast_fp16 = reduce_mean(axes = var_5846_axes_0, keep_dims = var_5846_keep_dims_0, x = var_5841_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5847_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_5847_cast_fp16")];
            fp32 var_5848_epsilon_0 = const()[name = string("op_5848_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_5848_cast_fp16 = inverse(epsilon = var_5848_epsilon_0, x = var_5847_cast_fp16)[name = string("op_5848_cast_fp16")];
            fp16 var_5849_to_fp16 = const()[name = string("op_5849_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_5850_cast_fp16 = mul(x = var_5848_cast_fp16, y = var_5849_to_fp16)[name = string("op_5850_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_5852_cast_fp16 = add(x = var_5846_cast_fp16, y = var_5850_cast_fp16)[name = string("op_5852_cast_fp16")];
            fp32 var_5853_epsilon_0 = const()[name = string("op_5853_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_5853_cast_fp16 = rsqrt(epsilon = var_5853_epsilon_0, x = var_5852_cast_fp16)[name = string("op_5853_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_157_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5853_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123640768)))];
            tensor<fp16, [1, 1260, 768]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = t_encoder_layers_13_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123642368))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124232256))))[name = string("t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_237_cast_fp16 = reshape(shape = var_5863, x = linear_92_cast_fp16)[name = string("x_237_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5865_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_5865_cast_fp16")];
            bool var_5868 = const()[name = string("op_5868"), val = bool(true)];
            tensor<int32, [1]> var_5869_axes_0 = const()[name = string("op_5869_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5869_cast_fp16 = reduce_max(axes = var_5869_axes_0, keep_dims = var_5868, x = var_5865_cast_fp16)[name = string("op_5869_cast_fp16")];
            fp16 var_5870_to_fp16 = const()[name = string("op_5870_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_189_cast_fp16 = maximum(x = var_5869_cast_fp16, y = var_5870_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_185_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_189_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5873_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5873_cast_fp16")];
            tensor<int32, [1]> var_5878_axes_0 = const()[name = string("op_5878_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5878_keep_dims_0 = const()[name = string("op_5878_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5878_cast_fp16 = reduce_mean(axes = var_5878_axes_0, keep_dims = var_5878_keep_dims_0, x = var_5873_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5879_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_5879_cast_fp16")];
            fp32 var_5880_epsilon_0 = const()[name = string("op_5880_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5880_cast_fp16 = inverse(epsilon = var_5880_epsilon_0, x = var_5879_cast_fp16)[name = string("op_5880_cast_fp16")];
            fp16 var_5881_to_fp16 = const()[name = string("op_5881_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5882_cast_fp16 = mul(x = var_5880_cast_fp16, y = var_5881_to_fp16)[name = string("op_5882_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5884_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5882_cast_fp16)[name = string("op_5884_cast_fp16")];
            fp32 var_5885_epsilon_0 = const()[name = string("op_5885_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5885_cast_fp16 = rsqrt(epsilon = var_5885_epsilon_0, x = var_5884_cast_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_159_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5885_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910080)))];
            tensor<fp16, [1, 1260, 12, 64]> x_243_cast_fp16 = mul(x = out_159_cast_fp16, y = t_encoder_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_243_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124233856))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124823744))))[name = string("t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_93_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5895 = const()[name = string("op_5895"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_239_cast_fp16 = reshape(shape = var_5895, x = linear_93_cast_fp16)[name = string("x_239_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5897_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_5897_cast_fp16")];
            bool var_5900 = const()[name = string("op_5900"), val = bool(true)];
            tensor<int32, [1]> var_5901_axes_0 = const()[name = string("op_5901_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5901_cast_fp16 = reduce_max(axes = var_5901_axes_0, keep_dims = var_5900, x = var_5897_cast_fp16)[name = string("op_5901_cast_fp16")];
            fp16 var_5902_to_fp16 = const()[name = string("op_5902_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_191_cast_fp16 = maximum(x = var_5901_cast_fp16, y = var_5902_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_187_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_191_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5905_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_5905_cast_fp16")];
            tensor<int32, [1]> var_5910_axes_0 = const()[name = string("op_5910_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5910_keep_dims_0 = const()[name = string("op_5910_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5910_cast_fp16 = reduce_mean(axes = var_5910_axes_0, keep_dims = var_5910_keep_dims_0, x = var_5905_cast_fp16)[name = string("op_5910_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5911_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_5911_cast_fp16")];
            fp32 var_5912_epsilon_0 = const()[name = string("op_5912_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5912_cast_fp16 = inverse(epsilon = var_5912_epsilon_0, x = var_5911_cast_fp16)[name = string("op_5912_cast_fp16")];
            fp16 var_5913_to_fp16 = const()[name = string("op_5913_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5914_cast_fp16 = mul(x = var_5912_cast_fp16, y = var_5913_to_fp16)[name = string("op_5914_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5916_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5914_cast_fp16)[name = string("op_5916_cast_fp16")];
            fp32 var_5917_epsilon_0 = const()[name = string("op_5917_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5917_cast_fp16 = rsqrt(epsilon = var_5917_epsilon_0, x = var_5916_cast_fp16)[name = string("op_5917_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_161_cast_fp16 = mul(x = y_187_cast_fp16, y = var_5917_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501760)))];
            tensor<fp16, [1, 1260, 12, 64]> x_245_cast_fp16 = mul(x = out_161_cast_fp16, y = t_encoder_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124825344))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125415232))))[name = string("t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [4]> var_5927 = const()[name = string("op_5927"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_241_cast_fp16 = reshape(shape = var_5927, x = linear_94_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5929_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_5929_cast_fp16")];
            bool var_5932 = const()[name = string("op_5932"), val = bool(true)];
            tensor<int32, [1]> var_5933_axes_0 = const()[name = string("op_5933_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_5933_cast_fp16 = reduce_max(axes = var_5933_axes_0, keep_dims = var_5932, x = var_5929_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp16 var_5934_to_fp16 = const()[name = string("op_5934_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_193_cast_fp16 = maximum(x = var_5933_cast_fp16, y = var_5934_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_189_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_193_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_5937_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_5937_cast_fp16")];
            tensor<int32, [1]> var_5942_axes_0 = const()[name = string("op_5942_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5942_keep_dims_0 = const()[name = string("op_5942_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_5942_cast_fp16 = reduce_mean(axes = var_5942_axes_0, keep_dims = var_5942_keep_dims_0, x = var_5937_cast_fp16)[name = string("op_5942_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5943_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_5943_cast_fp16")];
            fp32 var_5944_epsilon_0 = const()[name = string("op_5944_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_5944_cast_fp16 = inverse(epsilon = var_5944_epsilon_0, x = var_5943_cast_fp16)[name = string("op_5944_cast_fp16")];
            fp16 var_5945_to_fp16 = const()[name = string("op_5945_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_5946_cast_fp16 = mul(x = var_5944_cast_fp16, y = var_5945_to_fp16)[name = string("op_5946_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_5948_cast_fp16 = add(x = var_5942_cast_fp16, y = var_5946_cast_fp16)[name = string("op_5948_cast_fp16")];
            fp32 var_5949_epsilon_0 = const()[name = string("op_5949_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_5949_cast_fp16 = rsqrt(epsilon = var_5949_epsilon_0, x = var_5948_cast_fp16)[name = string("op_5949_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_53_cast_fp16 = mul(x = y_189_cast_fp16, y = var_5949_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<int32, [4]> z_105_begin_0 = const()[name = string("z_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_105_end_0 = const()[name = string("z_105_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_105_end_mask_0 = const()[name = string("z_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_105_cast_fp16 = slice_by_index(begin = z_105_begin_0, end = z_105_end_0, end_mask = z_105_end_mask_0, x = x_243_cast_fp16)[name = string("z_105_cast_fp16")];
            tensor<int32, [4]> var_5960_begin_0 = const()[name = string("op_5960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5960_end_0 = const()[name = string("op_5960_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5960_end_mask_0 = const()[name = string("op_5960_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5960_cast_fp16 = slice_by_index(begin = var_5960_begin_0, end = var_5960_end_0, end_mask = var_5960_end_mask_0, x = z_105_cast_fp16)[name = string("op_5960_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5961_cast_fp16 = mul(x = var_5960_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_5961_cast_fp16")];
            tensor<int32, [4]> var_5966_begin_0 = const()[name = string("op_5966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5966_end_0 = const()[name = string("op_5966_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5966_end_mask_0 = const()[name = string("op_5966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5966_cast_fp16 = slice_by_index(begin = var_5966_begin_0, end = var_5966_end_0, end_mask = var_5966_end_mask_0, x = z_105_cast_fp16)[name = string("op_5966_cast_fp16")];
            int32 var_5968 = const()[name = string("op_5968"), val = int32(-1)];
            bool rot_105_interleave_0 = const()[name = string("rot_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_105_cast_fp16 = concat(axis = var_5968, interleave = rot_105_interleave_0, values = (var_5961_cast_fp16, var_5966_cast_fp16))[name = string("rot_105_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5970_cast_fp16 = mul(x = z_105_cast_fp16, y = c_1_cast_fp16)[name = string("op_5970_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5971_cast_fp16 = mul(x = rot_105_cast_fp16, y = s_9_cast_fp16)[name = string("op_5971_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5973_cast_fp16 = add(x = var_5970_cast_fp16, y = var_5971_cast_fp16)[name = string("op_5973_cast_fp16")];
            tensor<int32, [4]> z_107_begin_0 = const()[name = string("z_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_107_end_0 = const()[name = string("z_107_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_107_end_mask_0 = const()[name = string("z_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_107_cast_fp16 = slice_by_index(begin = z_107_begin_0, end = z_107_end_0, end_mask = z_107_end_mask_0, x = x_243_cast_fp16)[name = string("z_107_cast_fp16")];
            tensor<int32, [4]> var_5983_begin_0 = const()[name = string("op_5983_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5983_end_0 = const()[name = string("op_5983_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_5983_end_mask_0 = const()[name = string("op_5983_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_5983_cast_fp16 = slice_by_index(begin = var_5983_begin_0, end = var_5983_end_0, end_mask = var_5983_end_mask_0, x = z_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            fp16 const_53_promoted_to_fp16 = const()[name = string("const_53_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_5984_cast_fp16 = mul(x = var_5983_cast_fp16, y = const_53_promoted_to_fp16)[name = string("op_5984_cast_fp16")];
            tensor<int32, [4]> var_5989_begin_0 = const()[name = string("op_5989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5989_end_0 = const()[name = string("op_5989_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_5989_end_mask_0 = const()[name = string("op_5989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_5989_cast_fp16 = slice_by_index(begin = var_5989_begin_0, end = var_5989_end_0, end_mask = var_5989_end_mask_0, x = z_107_cast_fp16)[name = string("op_5989_cast_fp16")];
            int32 var_5991 = const()[name = string("op_5991"), val = int32(-1)];
            bool rot_107_interleave_0 = const()[name = string("rot_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_107_cast_fp16 = concat(axis = var_5991, interleave = rot_107_interleave_0, values = (var_5984_cast_fp16, var_5989_cast_fp16))[name = string("rot_107_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5993_cast_fp16 = mul(x = z_107_cast_fp16, y = c_cast_fp16)[name = string("op_5993_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5994_cast_fp16 = mul(x = rot_107_cast_fp16, y = s_11_cast_fp16)[name = string("op_5994_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_5996_cast_fp16 = add(x = var_5993_cast_fp16, y = var_5994_cast_fp16)[name = string("op_5996_cast_fp16")];
            int32 var_5998 = const()[name = string("op_5998"), val = int32(-1)];
            bool var_5999_interleave_0 = const()[name = string("op_5999_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_5999_cast_fp16 = concat(axis = var_5998, interleave = var_5999_interleave_0, values = (var_5973_cast_fp16, var_5996_cast_fp16))[name = string("op_5999_cast_fp16")];
            tensor<int32, [4]> z_109_begin_0 = const()[name = string("z_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_109_end_0 = const()[name = string("z_109_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_109_end_mask_0 = const()[name = string("z_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_109_cast_fp16 = slice_by_index(begin = z_109_begin_0, end = z_109_end_0, end_mask = z_109_end_mask_0, x = x_245_cast_fp16)[name = string("z_109_cast_fp16")];
            tensor<int32, [4]> var_6012_begin_0 = const()[name = string("op_6012_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6012_end_0 = const()[name = string("op_6012_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6012_end_mask_0 = const()[name = string("op_6012_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6012_cast_fp16 = slice_by_index(begin = var_6012_begin_0, end = var_6012_end_0, end_mask = var_6012_end_mask_0, x = z_109_cast_fp16)[name = string("op_6012_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6013_cast_fp16 = mul(x = var_6012_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_6013_cast_fp16")];
            tensor<int32, [4]> var_6018_begin_0 = const()[name = string("op_6018_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6018_end_0 = const()[name = string("op_6018_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6018_end_mask_0 = const()[name = string("op_6018_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6018_cast_fp16 = slice_by_index(begin = var_6018_begin_0, end = var_6018_end_0, end_mask = var_6018_end_mask_0, x = z_109_cast_fp16)[name = string("op_6018_cast_fp16")];
            int32 var_6020 = const()[name = string("op_6020"), val = int32(-1)];
            bool rot_109_interleave_0 = const()[name = string("rot_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_109_cast_fp16 = concat(axis = var_6020, interleave = rot_109_interleave_0, values = (var_6013_cast_fp16, var_6018_cast_fp16))[name = string("rot_109_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6022_cast_fp16 = mul(x = z_109_cast_fp16, y = c_1_cast_fp16)[name = string("op_6022_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6023_cast_fp16 = mul(x = rot_109_cast_fp16, y = s_9_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6025_cast_fp16 = add(x = var_6022_cast_fp16, y = var_6023_cast_fp16)[name = string("op_6025_cast_fp16")];
            tensor<int32, [4]> z_111_begin_0 = const()[name = string("z_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_111_end_0 = const()[name = string("z_111_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_111_end_mask_0 = const()[name = string("z_111_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_111_cast_fp16 = slice_by_index(begin = z_111_begin_0, end = z_111_end_0, end_mask = z_111_end_mask_0, x = x_245_cast_fp16)[name = string("z_111_cast_fp16")];
            tensor<int32, [4]> var_6035_begin_0 = const()[name = string("op_6035_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6035_end_0 = const()[name = string("op_6035_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6035_end_mask_0 = const()[name = string("op_6035_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6035_cast_fp16 = slice_by_index(begin = var_6035_begin_0, end = var_6035_end_0, end_mask = var_6035_end_mask_0, x = z_111_cast_fp16)[name = string("op_6035_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6036_cast_fp16 = mul(x = var_6035_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_6036_cast_fp16")];
            tensor<int32, [4]> var_6041_begin_0 = const()[name = string("op_6041_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6041_end_0 = const()[name = string("op_6041_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6041_end_mask_0 = const()[name = string("op_6041_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6041_cast_fp16 = slice_by_index(begin = var_6041_begin_0, end = var_6041_end_0, end_mask = var_6041_end_mask_0, x = z_111_cast_fp16)[name = string("op_6041_cast_fp16")];
            int32 var_6043 = const()[name = string("op_6043"), val = int32(-1)];
            bool rot_111_interleave_0 = const()[name = string("rot_111_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_111_cast_fp16 = concat(axis = var_6043, interleave = rot_111_interleave_0, values = (var_6036_cast_fp16, var_6041_cast_fp16))[name = string("rot_111_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6045_cast_fp16 = mul(x = z_111_cast_fp16, y = c_cast_fp16)[name = string("op_6045_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6046_cast_fp16 = mul(x = rot_111_cast_fp16, y = s_11_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6048_cast_fp16 = add(x = var_6045_cast_fp16, y = var_6046_cast_fp16)[name = string("op_6048_cast_fp16")];
            int32 var_6050 = const()[name = string("op_6050"), val = int32(-1)];
            bool var_6051_interleave_0 = const()[name = string("op_6051_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_6051_cast_fp16 = concat(axis = var_6050, interleave = var_6051_interleave_0, values = (var_6025_cast_fp16, var_6048_cast_fp16))[name = string("op_6051_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6061_transpose_x_0 = const()[name = string("op_6061_transpose_x_0"), val = bool(false)];
            bool var_6061_transpose_y_0 = const()[name = string("op_6061_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_91 = transpose(perm = transpose_91_perm_0, x = var_6051_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_90 = transpose(perm = transpose_90_perm_0, x = var_5999_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 12, 1260, 1260]> var_6061_cast_fp16 = matmul(transpose_x = var_6061_transpose_x_0, transpose_y = var_6061_transpose_y_0, x = transpose_90, y = transpose_91)[name = string("op_6061_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_6063_cast_fp16 = add(x = var_6061_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6063_cast_fp16")];
            int32 var_6064 = const()[name = string("op_6064"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_189_cast_fp16 = softmax(axis = var_6064, x = var_6063_cast_fp16)[name = string("w_189_cast_fp16")];
            bool var_6067_transpose_x_0 = const()[name = string("op_6067_transpose_x_0"), val = bool(false)];
            bool var_6067_transpose_y_0 = const()[name = string("op_6067_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = v_53_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 12, 1260, 64]> var_6067_cast_fp16 = matmul(transpose_x = var_6067_transpose_x_0, transpose_y = var_6067_transpose_y_0, x = w_189_cast_fp16, y = v_55_cast_fp16)[name = string("op_6067_cast_fp16")];
            tensor<int32, [4]> var_6070_perm_0 = const()[name = string("op_6070_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6074 = const()[name = string("op_6074"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_6070_cast_fp16 = transpose(perm = var_6070_perm_0, x = var_6067_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 1260, 768]> input_135_cast_fp16 = reshape(shape = var_6074, x = var_6070_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125416832))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126006720))))[name = string("t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_95_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6079_cast_fp16 = abs(x = linear_95_cast_fp16)[name = string("op_6079_cast_fp16")];
            bool var_6082 = const()[name = string("op_6082"), val = bool(true)];
            tensor<int32, [1]> var_6083_axes_0 = const()[name = string("op_6083_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6083_cast_fp16 = reduce_max(axes = var_6083_axes_0, keep_dims = var_6082, x = var_6079_cast_fp16)[name = string("op_6083_cast_fp16")];
            fp16 var_6084_to_fp16 = const()[name = string("op_6084_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_195_cast_fp16 = maximum(x = var_6083_cast_fp16, y = var_6084_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_191_cast_fp16 = real_div(x = linear_95_cast_fp16, y = s_195_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6087_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6087_cast_fp16")];
            tensor<int32, [1]> var_6092_axes_0 = const()[name = string("op_6092_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6092_keep_dims_0 = const()[name = string("op_6092_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6092_cast_fp16 = reduce_mean(axes = var_6092_axes_0, keep_dims = var_6092_keep_dims_0, x = var_6087_cast_fp16)[name = string("op_6092_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6093_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6093_cast_fp16")];
            fp32 var_6094_epsilon_0 = const()[name = string("op_6094_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6094_cast_fp16 = inverse(epsilon = var_6094_epsilon_0, x = var_6093_cast_fp16)[name = string("op_6094_cast_fp16")];
            fp16 var_6095_to_fp16 = const()[name = string("op_6095_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6096_cast_fp16 = mul(x = var_6094_cast_fp16, y = var_6095_to_fp16)[name = string("op_6096_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6098_cast_fp16 = add(x = var_6092_cast_fp16, y = var_6096_cast_fp16)[name = string("op_6098_cast_fp16")];
            fp32 var_6099_epsilon_0 = const()[name = string("op_6099_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6099_cast_fp16 = rsqrt(epsilon = var_6099_epsilon_0, x = var_6098_cast_fp16)[name = string("op_6099_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_163_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6099_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126008320)))];
            tensor<fp16, [1, 1260, 768]> var_6101_cast_fp16 = mul(x = out_163_cast_fp16, y = t_encoder_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6101_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_249_cast_fp16 = add(x = x_235_cast_fp16, y = var_6101_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6104_cast_fp16 = abs(x = x_249_cast_fp16)[name = string("op_6104_cast_fp16")];
            bool var_6107 = const()[name = string("op_6107"), val = bool(true)];
            tensor<int32, [1]> var_6108_axes_0 = const()[name = string("op_6108_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6108_cast_fp16 = reduce_max(axes = var_6108_axes_0, keep_dims = var_6107, x = var_6104_cast_fp16)[name = string("op_6108_cast_fp16")];
            fp16 var_6109_to_fp16 = const()[name = string("op_6109_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_197_cast_fp16 = maximum(x = var_6108_cast_fp16, y = var_6109_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_193_cast_fp16 = real_div(x = x_249_cast_fp16, y = s_197_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6112_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6112_cast_fp16")];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6117_keep_dims_0 = const()[name = string("op_6117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6117_cast_fp16 = reduce_mean(axes = var_6117_axes_0, keep_dims = var_6117_keep_dims_0, x = var_6112_cast_fp16)[name = string("op_6117_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6118_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6118_cast_fp16")];
            fp32 var_6119_epsilon_0 = const()[name = string("op_6119_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6119_cast_fp16 = inverse(epsilon = var_6119_epsilon_0, x = var_6118_cast_fp16)[name = string("op_6119_cast_fp16")];
            fp16 var_6120_to_fp16 = const()[name = string("op_6120_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6121_cast_fp16 = mul(x = var_6119_cast_fp16, y = var_6120_to_fp16)[name = string("op_6121_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6123_cast_fp16 = add(x = var_6117_cast_fp16, y = var_6121_cast_fp16)[name = string("op_6123_cast_fp16")];
            fp32 var_6124_epsilon_0 = const()[name = string("op_6124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6124_cast_fp16 = rsqrt(epsilon = var_6124_epsilon_0, x = var_6123_cast_fp16)[name = string("op_6124_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_165_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6124_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126009920)))];
            tensor<fp16, [1, 1260, 768]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126011520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128370880))))[name = string("t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_96_cast_fp16")];
            string var_6131_mode_0 = const()[name = string("op_6131_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_6131_cast_fp16 = gelu(mode = var_6131_mode_0, x = linear_96_cast_fp16)[name = string("op_6131_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128377088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130736448))))[name = string("t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_97_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_141_cast_fp16 = mul(x = var_6131_cast_fp16, y = linear_97_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130742656))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133102016))))[name = string("t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6139_cast_fp16 = abs(x = linear_98_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_199_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_195_cast_fp16 = real_div(x = linear_98_cast_fp16, y = s_199_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6153_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_167_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133103616)))];
            tensor<fp16, [1, 1260, 768]> var_6161_cast_fp16 = mul(x = out_167_cast_fp16, y = t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6161_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_253_cast_fp16 = add(x = x_249_cast_fp16, y = var_6161_cast_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6164_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6164_cast_fp16")];
            bool var_6167 = const()[name = string("op_6167"), val = bool(true)];
            tensor<int32, [1]> var_6168_axes_0 = const()[name = string("op_6168_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6168_cast_fp16 = reduce_max(axes = var_6168_axes_0, keep_dims = var_6167, x = var_6164_cast_fp16)[name = string("op_6168_cast_fp16")];
            fp16 var_6169_to_fp16 = const()[name = string("op_6169_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_201_cast_fp16 = maximum(x = var_6168_cast_fp16, y = var_6169_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_197_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_201_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6172_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6172_cast_fp16")];
            tensor<int32, [1]> var_6177_axes_0 = const()[name = string("op_6177_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6177_keep_dims_0 = const()[name = string("op_6177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6177_cast_fp16 = reduce_mean(axes = var_6177_axes_0, keep_dims = var_6177_keep_dims_0, x = var_6172_cast_fp16)[name = string("op_6177_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6178_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6178_cast_fp16")];
            fp32 var_6179_epsilon_0 = const()[name = string("op_6179_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6179_cast_fp16 = inverse(epsilon = var_6179_epsilon_0, x = var_6178_cast_fp16)[name = string("op_6179_cast_fp16")];
            fp16 var_6180_to_fp16 = const()[name = string("op_6180_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6181_cast_fp16 = mul(x = var_6179_cast_fp16, y = var_6180_to_fp16)[name = string("op_6181_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6183_cast_fp16 = add(x = var_6177_cast_fp16, y = var_6181_cast_fp16)[name = string("op_6183_cast_fp16")];
            fp32 var_6184_epsilon_0 = const()[name = string("op_6184_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6184_cast_fp16 = rsqrt(epsilon = var_6184_epsilon_0, x = var_6183_cast_fp16)[name = string("op_6184_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_169_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6184_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133105216)))];
            tensor<fp16, [1, 1260, 768]> input_143_cast_fp16 = mul(x = out_169_cast_fp16, y = t_encoder_layers_14_input_layernorm_weight_to_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133106816))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133696704))))[name = string("t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_99_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_6194 = const()[name = string("op_6194"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_255_cast_fp16 = reshape(shape = var_6194, x = linear_99_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6196_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6196_cast_fp16")];
            bool var_6199 = const()[name = string("op_6199"), val = bool(true)];
            tensor<int32, [1]> var_6200_axes_0 = const()[name = string("op_6200_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_6200_cast_fp16 = reduce_max(axes = var_6200_axes_0, keep_dims = var_6199, x = var_6196_cast_fp16)[name = string("op_6200_cast_fp16")];
            fp16 var_6201_to_fp16 = const()[name = string("op_6201_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_203_cast_fp16 = maximum(x = var_6200_cast_fp16, y = var_6201_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_199_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_203_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6204_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6204_cast_fp16")];
            tensor<int32, [1]> var_6209_axes_0 = const()[name = string("op_6209_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6209_keep_dims_0 = const()[name = string("op_6209_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_6209_cast_fp16 = reduce_mean(axes = var_6209_axes_0, keep_dims = var_6209_keep_dims_0, x = var_6204_cast_fp16)[name = string("op_6209_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6210_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6210_cast_fp16")];
            fp32 var_6211_epsilon_0 = const()[name = string("op_6211_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_6211_cast_fp16 = inverse(epsilon = var_6211_epsilon_0, x = var_6210_cast_fp16)[name = string("op_6211_cast_fp16")];
            fp16 var_6212_to_fp16 = const()[name = string("op_6212_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_6213_cast_fp16 = mul(x = var_6211_cast_fp16, y = var_6212_to_fp16)[name = string("op_6213_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6215_cast_fp16 = add(x = var_6209_cast_fp16, y = var_6213_cast_fp16)[name = string("op_6215_cast_fp16")];
            fp32 var_6216_epsilon_0 = const()[name = string("op_6216_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_6216_cast_fp16 = rsqrt(epsilon = var_6216_epsilon_0, x = var_6215_cast_fp16)[name = string("op_6216_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_171_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6216_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133698304)))];
            tensor<fp16, [1, 1260, 12, 64]> x_261_cast_fp16 = mul(x = out_171_cast_fp16, y = t_encoder_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_261_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133698496))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134288384))))[name = string("t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_100_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_6226 = const()[name = string("op_6226"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_257_cast_fp16 = reshape(shape = var_6226, x = linear_100_cast_fp16)[name = string("x_257_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6228_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6228_cast_fp16")];
            bool var_6231 = const()[name = string("op_6231"), val = bool(true)];
            tensor<int32, [1]> var_6232_axes_0 = const()[name = string("op_6232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_6232_cast_fp16 = reduce_max(axes = var_6232_axes_0, keep_dims = var_6231, x = var_6228_cast_fp16)[name = string("op_6232_cast_fp16")];
            fp16 var_6233_to_fp16 = const()[name = string("op_6233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_205_cast_fp16 = maximum(x = var_6232_cast_fp16, y = var_6233_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_201_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_205_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6236_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6236_cast_fp16")];
            tensor<int32, [1]> var_6241_axes_0 = const()[name = string("op_6241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6241_keep_dims_0 = const()[name = string("op_6241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_6241_cast_fp16 = reduce_mean(axes = var_6241_axes_0, keep_dims = var_6241_keep_dims_0, x = var_6236_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6242_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6242_cast_fp16")];
            fp32 var_6243_epsilon_0 = const()[name = string("op_6243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_6243_cast_fp16 = inverse(epsilon = var_6243_epsilon_0, x = var_6242_cast_fp16)[name = string("op_6243_cast_fp16")];
            fp16 var_6244_to_fp16 = const()[name = string("op_6244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_6245_cast_fp16 = mul(x = var_6243_cast_fp16, y = var_6244_to_fp16)[name = string("op_6245_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6247_cast_fp16 = add(x = var_6241_cast_fp16, y = var_6245_cast_fp16)[name = string("op_6247_cast_fp16")];
            fp32 var_6248_epsilon_0 = const()[name = string("op_6248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_6248_cast_fp16 = rsqrt(epsilon = var_6248_epsilon_0, x = var_6247_cast_fp16)[name = string("op_6248_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_173_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6248_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966592)))];
            tensor<fp16, [1, 1260, 12, 64]> x_263_cast_fp16 = mul(x = out_173_cast_fp16, y = t_encoder_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134289984))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134879872))))[name = string("t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_101_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_6258 = const()[name = string("op_6258"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_259_cast_fp16 = reshape(shape = var_6258, x = linear_101_cast_fp16)[name = string("x_259_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6260_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6260_cast_fp16")];
            bool var_6263 = const()[name = string("op_6263"), val = bool(true)];
            tensor<int32, [1]> var_6264_axes_0 = const()[name = string("op_6264_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_6264_cast_fp16 = reduce_max(axes = var_6264_axes_0, keep_dims = var_6263, x = var_6260_cast_fp16)[name = string("op_6264_cast_fp16")];
            fp16 var_6265_to_fp16 = const()[name = string("op_6265_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_207_cast_fp16 = maximum(x = var_6264_cast_fp16, y = var_6265_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_203_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_207_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6268_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6268_cast_fp16")];
            tensor<int32, [1]> var_6273_axes_0 = const()[name = string("op_6273_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6273_keep_dims_0 = const()[name = string("op_6273_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_6273_cast_fp16 = reduce_mean(axes = var_6273_axes_0, keep_dims = var_6273_keep_dims_0, x = var_6268_cast_fp16)[name = string("op_6273_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6274_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6274_cast_fp16")];
            fp32 var_6275_epsilon_0 = const()[name = string("op_6275_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_6275_cast_fp16 = inverse(epsilon = var_6275_epsilon_0, x = var_6274_cast_fp16)[name = string("op_6275_cast_fp16")];
            fp16 var_6276_to_fp16 = const()[name = string("op_6276_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_6277_cast_fp16 = mul(x = var_6275_cast_fp16, y = var_6276_to_fp16)[name = string("op_6277_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6279_cast_fp16 = add(x = var_6273_cast_fp16, y = var_6277_cast_fp16)[name = string("op_6279_cast_fp16")];
            fp32 var_6280_epsilon_0 = const()[name = string("op_6280_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_6280_cast_fp16 = rsqrt(epsilon = var_6280_epsilon_0, x = var_6279_cast_fp16)[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_57_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6280_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<int32, [4]> z_113_begin_0 = const()[name = string("z_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_113_end_0 = const()[name = string("z_113_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_113_end_mask_0 = const()[name = string("z_113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_113_cast_fp16 = slice_by_index(begin = z_113_begin_0, end = z_113_end_0, end_mask = z_113_end_mask_0, x = x_261_cast_fp16)[name = string("z_113_cast_fp16")];
            tensor<int32, [4]> var_6291_begin_0 = const()[name = string("op_6291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6291_end_0 = const()[name = string("op_6291_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6291_end_mask_0 = const()[name = string("op_6291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6291_cast_fp16 = slice_by_index(begin = var_6291_begin_0, end = var_6291_end_0, end_mask = var_6291_end_mask_0, x = z_113_cast_fp16)[name = string("op_6291_cast_fp16")];
            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6292_cast_fp16 = mul(x = var_6291_cast_fp16, y = const_56_promoted_to_fp16)[name = string("op_6292_cast_fp16")];
            tensor<int32, [4]> var_6297_begin_0 = const()[name = string("op_6297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6297_end_0 = const()[name = string("op_6297_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6297_end_mask_0 = const()[name = string("op_6297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6297_cast_fp16 = slice_by_index(begin = var_6297_begin_0, end = var_6297_end_0, end_mask = var_6297_end_mask_0, x = z_113_cast_fp16)[name = string("op_6297_cast_fp16")];
            int32 var_6299 = const()[name = string("op_6299"), val = int32(-1)];
            bool rot_113_interleave_0 = const()[name = string("rot_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_113_cast_fp16 = concat(axis = var_6299, interleave = rot_113_interleave_0, values = (var_6292_cast_fp16, var_6297_cast_fp16))[name = string("rot_113_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6301_cast_fp16 = mul(x = z_113_cast_fp16, y = c_1_cast_fp16)[name = string("op_6301_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6302_cast_fp16 = mul(x = rot_113_cast_fp16, y = s_9_cast_fp16)[name = string("op_6302_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6304_cast_fp16 = add(x = var_6301_cast_fp16, y = var_6302_cast_fp16)[name = string("op_6304_cast_fp16")];
            tensor<int32, [4]> z_115_begin_0 = const()[name = string("z_115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_115_end_0 = const()[name = string("z_115_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_115_end_mask_0 = const()[name = string("z_115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_115_cast_fp16 = slice_by_index(begin = z_115_begin_0, end = z_115_end_0, end_mask = z_115_end_mask_0, x = x_261_cast_fp16)[name = string("z_115_cast_fp16")];
            tensor<int32, [4]> var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> 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 = z_115_cast_fp16)[name = string("op_6314_cast_fp16")];
            fp16 const_57_promoted_to_fp16 = const()[name = string("const_57_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6315_cast_fp16 = mul(x = var_6314_cast_fp16, y = const_57_promoted_to_fp16)[name = string("op_6315_cast_fp16")];
            tensor<int32, [4]> var_6320_begin_0 = const()[name = string("op_6320_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6320_end_0 = const()[name = string("op_6320_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6320_end_mask_0 = const()[name = string("op_6320_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6320_cast_fp16 = slice_by_index(begin = var_6320_begin_0, end = var_6320_end_0, end_mask = var_6320_end_mask_0, x = z_115_cast_fp16)[name = string("op_6320_cast_fp16")];
            int32 var_6322 = const()[name = string("op_6322"), val = int32(-1)];
            bool rot_115_interleave_0 = const()[name = string("rot_115_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_115_cast_fp16 = concat(axis = var_6322, interleave = rot_115_interleave_0, values = (var_6315_cast_fp16, var_6320_cast_fp16))[name = string("rot_115_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6324_cast_fp16 = mul(x = z_115_cast_fp16, y = c_cast_fp16)[name = string("op_6324_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6325_cast_fp16 = mul(x = rot_115_cast_fp16, y = s_11_cast_fp16)[name = string("op_6325_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6327_cast_fp16 = add(x = var_6324_cast_fp16, y = var_6325_cast_fp16)[name = string("op_6327_cast_fp16")];
            int32 var_6329 = const()[name = string("op_6329"), val = int32(-1)];
            bool var_6330_interleave_0 = const()[name = string("op_6330_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_6330_cast_fp16 = concat(axis = var_6329, interleave = var_6330_interleave_0, values = (var_6304_cast_fp16, var_6327_cast_fp16))[name = string("op_6330_cast_fp16")];
            tensor<int32, [4]> z_117_begin_0 = const()[name = string("z_117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_117_end_0 = const()[name = string("z_117_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_117_end_mask_0 = const()[name = string("z_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_117_cast_fp16 = slice_by_index(begin = z_117_begin_0, end = z_117_end_0, end_mask = z_117_end_mask_0, x = x_263_cast_fp16)[name = string("z_117_cast_fp16")];
            tensor<int32, [4]> var_6343_begin_0 = const()[name = string("op_6343_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6343_end_0 = const()[name = string("op_6343_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6343_end_mask_0 = const()[name = string("op_6343_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6343_cast_fp16 = slice_by_index(begin = var_6343_begin_0, end = var_6343_end_0, end_mask = var_6343_end_mask_0, x = z_117_cast_fp16)[name = string("op_6343_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6344_cast_fp16 = mul(x = var_6343_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_6344_cast_fp16")];
            tensor<int32, [4]> var_6349_begin_0 = const()[name = string("op_6349_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6349_end_0 = const()[name = string("op_6349_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6349_end_mask_0 = const()[name = string("op_6349_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6349_cast_fp16 = slice_by_index(begin = var_6349_begin_0, end = var_6349_end_0, end_mask = var_6349_end_mask_0, x = z_117_cast_fp16)[name = string("op_6349_cast_fp16")];
            int32 var_6351 = const()[name = string("op_6351"), val = int32(-1)];
            bool rot_117_interleave_0 = const()[name = string("rot_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_117_cast_fp16 = concat(axis = var_6351, interleave = rot_117_interleave_0, values = (var_6344_cast_fp16, var_6349_cast_fp16))[name = string("rot_117_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6353_cast_fp16 = mul(x = z_117_cast_fp16, y = c_1_cast_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6354_cast_fp16 = mul(x = rot_117_cast_fp16, y = s_9_cast_fp16)[name = string("op_6354_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6356_cast_fp16 = add(x = var_6353_cast_fp16, y = var_6354_cast_fp16)[name = string("op_6356_cast_fp16")];
            tensor<int32, [4]> z_119_begin_0 = const()[name = string("z_119_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_119_end_0 = const()[name = string("z_119_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_119_end_mask_0 = const()[name = string("z_119_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_119_cast_fp16 = slice_by_index(begin = z_119_begin_0, end = z_119_end_0, end_mask = z_119_end_mask_0, x = x_263_cast_fp16)[name = string("z_119_cast_fp16")];
            tensor<int32, [4]> var_6366_begin_0 = const()[name = string("op_6366_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6366_end_0 = const()[name = string("op_6366_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6366_end_mask_0 = const()[name = string("op_6366_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6366_cast_fp16 = slice_by_index(begin = var_6366_begin_0, end = var_6366_end_0, end_mask = var_6366_end_mask_0, x = z_119_cast_fp16)[name = string("op_6366_cast_fp16")];
            fp16 const_59_promoted_to_fp16 = const()[name = string("const_59_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6367_cast_fp16 = mul(x = var_6366_cast_fp16, y = const_59_promoted_to_fp16)[name = string("op_6367_cast_fp16")];
            tensor<int32, [4]> var_6372_begin_0 = const()[name = string("op_6372_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6372_end_0 = const()[name = string("op_6372_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6372_end_mask_0 = const()[name = string("op_6372_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6372_cast_fp16 = slice_by_index(begin = var_6372_begin_0, end = var_6372_end_0, end_mask = var_6372_end_mask_0, x = z_119_cast_fp16)[name = string("op_6372_cast_fp16")];
            int32 var_6374 = const()[name = string("op_6374"), val = int32(-1)];
            bool rot_119_interleave_0 = const()[name = string("rot_119_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_119_cast_fp16 = concat(axis = var_6374, interleave = rot_119_interleave_0, values = (var_6367_cast_fp16, var_6372_cast_fp16))[name = string("rot_119_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6376_cast_fp16 = mul(x = z_119_cast_fp16, y = c_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6377_cast_fp16 = mul(x = rot_119_cast_fp16, y = s_11_cast_fp16)[name = string("op_6377_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6379_cast_fp16 = add(x = var_6376_cast_fp16, y = var_6377_cast_fp16)[name = string("op_6379_cast_fp16")];
            int32 var_6381 = const()[name = string("op_6381"), val = int32(-1)];
            bool var_6382_interleave_0 = const()[name = string("op_6382_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_6382_cast_fp16 = concat(axis = var_6381, interleave = var_6382_interleave_0, values = (var_6356_cast_fp16, var_6379_cast_fp16))[name = string("op_6382_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6392_transpose_x_0 = const()[name = string("op_6392_transpose_x_0"), val = bool(false)];
            bool var_6392_transpose_y_0 = const()[name = string("op_6392_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_93 = transpose(perm = transpose_93_perm_0, x = var_6382_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_92 = transpose(perm = transpose_92_perm_0, x = var_6330_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 12, 1260, 1260]> var_6392_cast_fp16 = matmul(transpose_x = var_6392_transpose_x_0, transpose_y = var_6392_transpose_y_0, x = transpose_92, y = transpose_93)[name = string("op_6392_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_6394_cast_fp16 = add(x = var_6392_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6394_cast_fp16")];
            int32 var_6395 = const()[name = string("op_6395"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_203_cast_fp16 = softmax(axis = var_6395, x = var_6394_cast_fp16)[name = string("w_203_cast_fp16")];
            bool var_6398_transpose_x_0 = const()[name = string("op_6398_transpose_x_0"), val = bool(false)];
            bool var_6398_transpose_y_0 = const()[name = string("op_6398_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = v_57_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 12, 1260, 64]> var_6398_cast_fp16 = matmul(transpose_x = var_6398_transpose_x_0, transpose_y = var_6398_transpose_y_0, x = w_203_cast_fp16, y = v_59_cast_fp16)[name = string("op_6398_cast_fp16")];
            tensor<int32, [4]> var_6401_perm_0 = const()[name = string("op_6401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6405 = const()[name = string("op_6405"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_6401_cast_fp16 = transpose(perm = var_6401_perm_0, x = var_6398_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 1260, 768]> input_145_cast_fp16 = reshape(shape = var_6405, x = var_6401_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134881472))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135471360))))[name = string("t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_102_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6410_cast_fp16 = abs(x = linear_102_cast_fp16)[name = string("op_6410_cast_fp16")];
            bool var_6413 = const()[name = string("op_6413"), val = bool(true)];
            tensor<int32, [1]> var_6414_axes_0 = const()[name = string("op_6414_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6414_cast_fp16 = reduce_max(axes = var_6414_axes_0, keep_dims = var_6413, x = var_6410_cast_fp16)[name = string("op_6414_cast_fp16")];
            fp16 var_6415_to_fp16 = const()[name = string("op_6415_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_209_cast_fp16 = maximum(x = var_6414_cast_fp16, y = var_6415_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_205_cast_fp16 = real_div(x = linear_102_cast_fp16, y = s_209_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6418_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6418_cast_fp16")];
            tensor<int32, [1]> var_6423_axes_0 = const()[name = string("op_6423_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6423_keep_dims_0 = const()[name = string("op_6423_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6423_cast_fp16 = reduce_mean(axes = var_6423_axes_0, keep_dims = var_6423_keep_dims_0, x = var_6418_cast_fp16)[name = string("op_6423_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6424_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6424_cast_fp16")];
            fp32 var_6425_epsilon_0 = const()[name = string("op_6425_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6425_cast_fp16 = inverse(epsilon = var_6425_epsilon_0, x = var_6424_cast_fp16)[name = string("op_6425_cast_fp16")];
            fp16 var_6426_to_fp16 = const()[name = string("op_6426_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6427_cast_fp16 = mul(x = var_6425_cast_fp16, y = var_6426_to_fp16)[name = string("op_6427_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6429_cast_fp16 = add(x = var_6423_cast_fp16, y = var_6427_cast_fp16)[name = string("op_6429_cast_fp16")];
            fp32 var_6430_epsilon_0 = const()[name = string("op_6430_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6430_cast_fp16 = rsqrt(epsilon = var_6430_epsilon_0, x = var_6429_cast_fp16)[name = string("op_6430_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_175_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6430_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135472960)))];
            tensor<fp16, [1, 1260, 768]> var_6432_cast_fp16 = mul(x = out_175_cast_fp16, y = t_encoder_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_6432_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_267_cast_fp16 = add(x = x_253_cast_fp16, y = var_6432_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6435_cast_fp16 = abs(x = x_267_cast_fp16)[name = string("op_6435_cast_fp16")];
            bool var_6438 = const()[name = string("op_6438"), val = bool(true)];
            tensor<int32, [1]> var_6439_axes_0 = const()[name = string("op_6439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6439_cast_fp16 = reduce_max(axes = var_6439_axes_0, keep_dims = var_6438, x = var_6435_cast_fp16)[name = string("op_6439_cast_fp16")];
            fp16 var_6440_to_fp16 = const()[name = string("op_6440_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_211_cast_fp16 = maximum(x = var_6439_cast_fp16, y = var_6440_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_207_cast_fp16 = real_div(x = x_267_cast_fp16, y = s_211_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6443_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6443_cast_fp16")];
            tensor<int32, [1]> var_6448_axes_0 = const()[name = string("op_6448_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6448_keep_dims_0 = const()[name = string("op_6448_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6448_cast_fp16 = reduce_mean(axes = var_6448_axes_0, keep_dims = var_6448_keep_dims_0, x = var_6443_cast_fp16)[name = string("op_6448_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6449_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6449_cast_fp16")];
            fp32 var_6450_epsilon_0 = const()[name = string("op_6450_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6450_cast_fp16 = inverse(epsilon = var_6450_epsilon_0, x = var_6449_cast_fp16)[name = string("op_6450_cast_fp16")];
            fp16 var_6451_to_fp16 = const()[name = string("op_6451_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6452_cast_fp16 = mul(x = var_6450_cast_fp16, y = var_6451_to_fp16)[name = string("op_6452_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6454_cast_fp16 = add(x = var_6448_cast_fp16, y = var_6452_cast_fp16)[name = string("op_6454_cast_fp16")];
            fp32 var_6455_epsilon_0 = const()[name = string("op_6455_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6455_cast_fp16 = rsqrt(epsilon = var_6455_epsilon_0, x = var_6454_cast_fp16)[name = string("op_6455_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_177_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6455_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135474560)))];
            tensor<fp16, [1, 1260, 768]> input_147_cast_fp16 = mul(x = out_177_cast_fp16, y = t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135476160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137835520))))[name = string("t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_103_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_103_cast_fp16")];
            string var_6462_mode_0 = const()[name = string("op_6462_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_6462_cast_fp16 = gelu(mode = var_6462_mode_0, x = linear_103_cast_fp16)[name = string("op_6462_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137841728))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140201088))))[name = string("t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_151_cast_fp16 = mul(x = var_6462_cast_fp16, y = linear_104_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140207296))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142566656))))[name = string("t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6470_cast_fp16 = abs(x = linear_105_cast_fp16)[name = string("op_6470_cast_fp16")];
            bool var_6473 = const()[name = string("op_6473"), val = bool(true)];
            tensor<int32, [1]> var_6474_axes_0 = const()[name = string("op_6474_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6474_cast_fp16 = reduce_max(axes = var_6474_axes_0, keep_dims = var_6473, x = var_6470_cast_fp16)[name = string("op_6474_cast_fp16")];
            fp16 var_6475_to_fp16 = const()[name = string("op_6475_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_213_cast_fp16 = maximum(x = var_6474_cast_fp16, y = var_6475_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_209_cast_fp16 = real_div(x = linear_105_cast_fp16, y = s_213_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6478_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6478_cast_fp16")];
            tensor<int32, [1]> var_6483_axes_0 = const()[name = string("op_6483_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6483_keep_dims_0 = const()[name = string("op_6483_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6483_cast_fp16 = reduce_mean(axes = var_6483_axes_0, keep_dims = var_6483_keep_dims_0, x = var_6478_cast_fp16)[name = string("op_6483_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6484_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp32 var_6485_epsilon_0 = const()[name = string("op_6485_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6485_cast_fp16 = inverse(epsilon = var_6485_epsilon_0, x = var_6484_cast_fp16)[name = string("op_6485_cast_fp16")];
            fp16 var_6486_to_fp16 = const()[name = string("op_6486_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6487_cast_fp16 = mul(x = var_6485_cast_fp16, y = var_6486_to_fp16)[name = string("op_6487_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6489_cast_fp16 = add(x = var_6483_cast_fp16, y = var_6487_cast_fp16)[name = string("op_6489_cast_fp16")];
            fp32 var_6490_epsilon_0 = const()[name = string("op_6490_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6490_cast_fp16 = rsqrt(epsilon = var_6490_epsilon_0, x = var_6489_cast_fp16)[name = string("op_6490_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_179_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6490_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142568256)))];
            tensor<fp16, [1, 1260, 768]> var_6492_cast_fp16 = mul(x = out_179_cast_fp16, y = t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6492_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_271_cast_fp16 = add(x = x_267_cast_fp16, y = var_6492_cast_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6495_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_6495_cast_fp16")];
            bool var_6498 = const()[name = string("op_6498"), val = bool(true)];
            tensor<int32, [1]> var_6499_axes_0 = const()[name = string("op_6499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6499_cast_fp16 = reduce_max(axes = var_6499_axes_0, keep_dims = var_6498, x = var_6495_cast_fp16)[name = string("op_6499_cast_fp16")];
            fp16 var_6500_to_fp16 = const()[name = string("op_6500_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_215_cast_fp16 = maximum(x = var_6499_cast_fp16, y = var_6500_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_211_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_215_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6503_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6503_cast_fp16")];
            tensor<int32, [1]> var_6508_axes_0 = const()[name = string("op_6508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6508_keep_dims_0 = const()[name = string("op_6508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6508_cast_fp16 = reduce_mean(axes = var_6508_axes_0, keep_dims = var_6508_keep_dims_0, x = var_6503_cast_fp16)[name = string("op_6508_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6509_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6509_cast_fp16")];
            fp32 var_6510_epsilon_0 = const()[name = string("op_6510_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6510_cast_fp16 = inverse(epsilon = var_6510_epsilon_0, x = var_6509_cast_fp16)[name = string("op_6510_cast_fp16")];
            fp16 var_6511_to_fp16 = const()[name = string("op_6511_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6512_cast_fp16 = mul(x = var_6510_cast_fp16, y = var_6511_to_fp16)[name = string("op_6512_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6514_cast_fp16 = add(x = var_6508_cast_fp16, y = var_6512_cast_fp16)[name = string("op_6514_cast_fp16")];
            fp32 var_6515_epsilon_0 = const()[name = string("op_6515_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6515_cast_fp16 = rsqrt(epsilon = var_6515_epsilon_0, x = var_6514_cast_fp16)[name = string("op_6515_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_181_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6515_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142569856)))];
            tensor<fp16, [1, 1260, 768]> input_153_cast_fp16 = mul(x = out_181_cast_fp16, y = t_encoder_layers_15_input_layernorm_weight_to_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142571456))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143161344))))[name = string("t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_106_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_6525 = const()[name = string("op_6525"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_273_cast_fp16 = reshape(shape = var_6525, x = linear_106_cast_fp16)[name = string("x_273_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6527_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_6527_cast_fp16")];
            bool var_6530 = const()[name = string("op_6530"), val = bool(true)];
            tensor<int32, [1]> var_6531_axes_0 = const()[name = string("op_6531_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_6531_cast_fp16 = reduce_max(axes = var_6531_axes_0, keep_dims = var_6530, x = var_6527_cast_fp16)[name = string("op_6531_cast_fp16")];
            fp16 var_6532_to_fp16 = const()[name = string("op_6532_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_217_cast_fp16 = maximum(x = var_6531_cast_fp16, y = var_6532_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_213_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_217_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6535_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6535_cast_fp16")];
            tensor<int32, [1]> var_6540_axes_0 = const()[name = string("op_6540_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6540_keep_dims_0 = const()[name = string("op_6540_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_6540_cast_fp16 = reduce_mean(axes = var_6540_axes_0, keep_dims = var_6540_keep_dims_0, x = var_6535_cast_fp16)[name = string("op_6540_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6541_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6541_cast_fp16")];
            fp32 var_6542_epsilon_0 = const()[name = string("op_6542_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_6542_cast_fp16 = inverse(epsilon = var_6542_epsilon_0, x = var_6541_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp16 var_6543_to_fp16 = const()[name = string("op_6543_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_6544_cast_fp16 = mul(x = var_6542_cast_fp16, y = var_6543_to_fp16)[name = string("op_6544_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6546_cast_fp16 = add(x = var_6540_cast_fp16, y = var_6544_cast_fp16)[name = string("op_6546_cast_fp16")];
            fp32 var_6547_epsilon_0 = const()[name = string("op_6547_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_6547_cast_fp16 = rsqrt(epsilon = var_6547_epsilon_0, x = var_6546_cast_fp16)[name = string("op_6547_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_183_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6547_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769024)))];
            tensor<fp16, [1, 1260, 12, 64]> x_279_cast_fp16 = mul(x = out_183_cast_fp16, y = t_encoder_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_279_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143162944))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143752832))))[name = string("t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_107_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_6557 = const()[name = string("op_6557"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_275_cast_fp16 = reshape(shape = var_6557, x = linear_107_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6559_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_6559_cast_fp16")];
            bool var_6562 = const()[name = string("op_6562"), val = bool(true)];
            tensor<int32, [1]> var_6563_axes_0 = const()[name = string("op_6563_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_6563_cast_fp16 = reduce_max(axes = var_6563_axes_0, keep_dims = var_6562, x = var_6559_cast_fp16)[name = string("op_6563_cast_fp16")];
            fp16 var_6564_to_fp16 = const()[name = string("op_6564_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_219_cast_fp16 = maximum(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_215_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_219_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6567_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6567_cast_fp16")];
            tensor<int32, [1]> var_6572_axes_0 = const()[name = string("op_6572_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6572_keep_dims_0 = const()[name = string("op_6572_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_6572_cast_fp16 = reduce_mean(axes = var_6572_axes_0, keep_dims = var_6572_keep_dims_0, x = var_6567_cast_fp16)[name = string("op_6572_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6573_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6573_cast_fp16")];
            fp32 var_6574_epsilon_0 = const()[name = string("op_6574_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_6574_cast_fp16 = inverse(epsilon = var_6574_epsilon_0, x = var_6573_cast_fp16)[name = string("op_6574_cast_fp16")];
            fp16 var_6575_to_fp16 = const()[name = string("op_6575_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_6576_cast_fp16 = mul(x = var_6574_cast_fp16, y = var_6575_to_fp16)[name = string("op_6576_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6578_cast_fp16 = add(x = var_6572_cast_fp16, y = var_6576_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp32 var_6579_epsilon_0 = const()[name = string("op_6579_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_6579_cast_fp16 = rsqrt(epsilon = var_6579_epsilon_0, x = var_6578_cast_fp16)[name = string("op_6579_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> out_185_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6579_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360704)))];
            tensor<fp16, [1, 1260, 12, 64]> x_281_cast_fp16 = mul(x = out_185_cast_fp16, y = t_encoder_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143754432))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144344320))))[name = string("t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [4]> var_6589 = const()[name = string("op_6589"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 1260, 12, 64]> x_277_cast_fp16 = reshape(shape = var_6589, x = linear_108_cast_fp16)[name = string("x_277_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6591_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_6591_cast_fp16")];
            bool var_6594 = const()[name = string("op_6594"), val = bool(true)];
            tensor<int32, [1]> var_6595_axes_0 = const()[name = string("op_6595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 12, 1]> var_6595_cast_fp16 = reduce_max(axes = var_6595_axes_0, keep_dims = var_6594, x = var_6591_cast_fp16)[name = string("op_6595_cast_fp16")];
            fp16 var_6596_to_fp16 = const()[name = string("op_6596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 12, 1]> s_221_cast_fp16 = maximum(x = var_6595_cast_fp16, y = var_6596_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> y_217_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_221_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> var_6599_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6599_cast_fp16")];
            tensor<int32, [1]> var_6604_axes_0 = const()[name = string("op_6604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6604_keep_dims_0 = const()[name = string("op_6604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 12, 1]> var_6604_cast_fp16 = reduce_mean(axes = var_6604_axes_0, keep_dims = var_6604_keep_dims_0, x = var_6599_cast_fp16)[name = string("op_6604_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6605_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6605_cast_fp16")];
            fp32 var_6606_epsilon_0 = const()[name = string("op_6606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 12, 1]> var_6606_cast_fp16 = inverse(epsilon = var_6606_epsilon_0, x = var_6605_cast_fp16)[name = string("op_6606_cast_fp16")];
            fp16 var_6607_to_fp16 = const()[name = string("op_6607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 12, 1]> var_6608_cast_fp16 = mul(x = var_6606_cast_fp16, y = var_6607_to_fp16)[name = string("op_6608_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 1]> var_6610_cast_fp16 = add(x = var_6604_cast_fp16, y = var_6608_cast_fp16)[name = string("op_6610_cast_fp16")];
            fp32 var_6611_epsilon_0 = const()[name = string("op_6611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 12, 1]> var_6611_cast_fp16 = rsqrt(epsilon = var_6611_epsilon_0, x = var_6610_cast_fp16)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 64]> v_61_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6611_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<int32, [4]> z_121_begin_0 = const()[name = string("z_121_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_121_end_0 = const()[name = string("z_121_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_121_end_mask_0 = const()[name = string("z_121_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_121_cast_fp16 = slice_by_index(begin = z_121_begin_0, end = z_121_end_0, end_mask = z_121_end_mask_0, x = x_279_cast_fp16)[name = string("z_121_cast_fp16")];
            tensor<int32, [4]> var_6622_begin_0 = const()[name = string("op_6622_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6622_end_0 = const()[name = string("op_6622_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6622_end_mask_0 = const()[name = string("op_6622_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6622_cast_fp16 = slice_by_index(begin = var_6622_begin_0, end = var_6622_end_0, end_mask = var_6622_end_mask_0, x = z_121_cast_fp16)[name = string("op_6622_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6623_cast_fp16 = mul(x = var_6622_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_6623_cast_fp16")];
            tensor<int32, [4]> var_6628_begin_0 = const()[name = string("op_6628_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6628_end_0 = const()[name = string("op_6628_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6628_end_mask_0 = const()[name = string("op_6628_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6628_cast_fp16 = slice_by_index(begin = var_6628_begin_0, end = var_6628_end_0, end_mask = var_6628_end_mask_0, x = z_121_cast_fp16)[name = string("op_6628_cast_fp16")];
            int32 var_6630 = const()[name = string("op_6630"), val = int32(-1)];
            bool rot_121_interleave_0 = const()[name = string("rot_121_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_121_cast_fp16 = concat(axis = var_6630, interleave = rot_121_interleave_0, values = (var_6623_cast_fp16, var_6628_cast_fp16))[name = string("rot_121_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6632_cast_fp16 = mul(x = z_121_cast_fp16, y = c_1_cast_fp16)[name = string("op_6632_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6633_cast_fp16 = mul(x = rot_121_cast_fp16, y = s_9_cast_fp16)[name = string("op_6633_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6635_cast_fp16 = add(x = var_6632_cast_fp16, y = var_6633_cast_fp16)[name = string("op_6635_cast_fp16")];
            tensor<int32, [4]> z_123_begin_0 = const()[name = string("z_123_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_123_end_0 = const()[name = string("z_123_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_123_end_mask_0 = const()[name = string("z_123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_123_cast_fp16 = slice_by_index(begin = z_123_begin_0, end = z_123_end_0, end_mask = z_123_end_mask_0, x = x_279_cast_fp16)[name = string("z_123_cast_fp16")];
            tensor<int32, [4]> var_6645_begin_0 = const()[name = string("op_6645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6645_end_0 = const()[name = string("op_6645_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6645_end_mask_0 = const()[name = string("op_6645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6645_cast_fp16 = slice_by_index(begin = var_6645_begin_0, end = var_6645_end_0, end_mask = var_6645_end_mask_0, x = z_123_cast_fp16)[name = string("op_6645_cast_fp16")];
            fp16 const_61_promoted_to_fp16 = const()[name = string("const_61_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6646_cast_fp16 = mul(x = var_6645_cast_fp16, y = const_61_promoted_to_fp16)[name = string("op_6646_cast_fp16")];
            tensor<int32, [4]> var_6651_begin_0 = const()[name = string("op_6651_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6651_end_0 = const()[name = string("op_6651_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6651_end_mask_0 = const()[name = string("op_6651_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6651_cast_fp16 = slice_by_index(begin = var_6651_begin_0, end = var_6651_end_0, end_mask = var_6651_end_mask_0, x = z_123_cast_fp16)[name = string("op_6651_cast_fp16")];
            int32 var_6653 = const()[name = string("op_6653"), val = int32(-1)];
            bool rot_123_interleave_0 = const()[name = string("rot_123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_123_cast_fp16 = concat(axis = var_6653, interleave = rot_123_interleave_0, values = (var_6646_cast_fp16, var_6651_cast_fp16))[name = string("rot_123_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6655_cast_fp16 = mul(x = z_123_cast_fp16, y = c_cast_fp16)[name = string("op_6655_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6656_cast_fp16 = mul(x = rot_123_cast_fp16, y = s_11_cast_fp16)[name = string("op_6656_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6658_cast_fp16 = add(x = var_6655_cast_fp16, y = var_6656_cast_fp16)[name = string("op_6658_cast_fp16")];
            int32 var_6660 = const()[name = string("op_6660"), val = int32(-1)];
            bool var_6661_interleave_0 = const()[name = string("op_6661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_6661_cast_fp16 = concat(axis = var_6660, interleave = var_6661_interleave_0, values = (var_6635_cast_fp16, var_6658_cast_fp16))[name = string("op_6661_cast_fp16")];
            tensor<int32, [4]> z_125_begin_0 = const()[name = string("z_125_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_125_end_0 = const()[name = string("z_125_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> z_125_end_mask_0 = const()[name = string("z_125_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 32]> z_125_cast_fp16 = slice_by_index(begin = z_125_begin_0, end = z_125_end_0, end_mask = z_125_end_mask_0, x = x_281_cast_fp16)[name = string("z_125_cast_fp16")];
            tensor<int32, [4]> var_6674_begin_0 = const()[name = string("op_6674_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6674_end_0 = const()[name = string("op_6674_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6674_end_mask_0 = const()[name = string("op_6674_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6674_cast_fp16 = slice_by_index(begin = var_6674_begin_0, end = var_6674_end_0, end_mask = var_6674_end_mask_0, x = z_125_cast_fp16)[name = string("op_6674_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6675_cast_fp16 = mul(x = var_6674_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_6675_cast_fp16")];
            tensor<int32, [4]> var_6680_begin_0 = const()[name = string("op_6680_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6680_end_0 = const()[name = string("op_6680_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6680_end_mask_0 = const()[name = string("op_6680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6680_cast_fp16 = slice_by_index(begin = var_6680_begin_0, end = var_6680_end_0, end_mask = var_6680_end_mask_0, x = z_125_cast_fp16)[name = string("op_6680_cast_fp16")];
            int32 var_6682 = const()[name = string("op_6682"), val = int32(-1)];
            bool rot_125_interleave_0 = const()[name = string("rot_125_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_125_cast_fp16 = concat(axis = var_6682, interleave = rot_125_interleave_0, values = (var_6675_cast_fp16, var_6680_cast_fp16))[name = string("rot_125_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6684_cast_fp16 = mul(x = z_125_cast_fp16, y = c_1_cast_fp16)[name = string("op_6684_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6685_cast_fp16 = mul(x = rot_125_cast_fp16, y = s_9_cast_fp16)[name = string("op_6685_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6687_cast_fp16 = add(x = var_6684_cast_fp16, y = var_6685_cast_fp16)[name = string("op_6687_cast_fp16")];
            tensor<int32, [4]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 1])];
            tensor<bool, [4]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 32]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = x_281_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [4]> var_6697_begin_0 = const()[name = string("op_6697_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6697_end_0 = const()[name = string("op_6697_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 32])];
            tensor<bool, [4]> var_6697_end_mask_0 = const()[name = string("op_6697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1260, 12, 16]> var_6697_cast_fp16 = slice_by_index(begin = var_6697_begin_0, end = var_6697_end_0, end_mask = var_6697_end_mask_0, x = z_cast_fp16)[name = string("op_6697_cast_fp16")];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 1260, 12, 16]> var_6698_cast_fp16 = mul(x = var_6697_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_6698_cast_fp16")];
            tensor<int32, [4]> var_6703_begin_0 = const()[name = string("op_6703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6703_end_0 = const()[name = string("op_6703_end_0"), val = tensor<int32, [4]>([1, 1260, 12, 16])];
            tensor<bool, [4]> var_6703_end_mask_0 = const()[name = string("op_6703_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1260, 12, 16]> var_6703_cast_fp16 = slice_by_index(begin = var_6703_begin_0, end = var_6703_end_0, end_mask = var_6703_end_mask_0, x = z_cast_fp16)[name = string("op_6703_cast_fp16")];
            int32 var_6705 = const()[name = string("op_6705"), val = int32(-1)];
            bool rot_interleave_0 = const()[name = string("rot_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 32]> rot_cast_fp16 = concat(axis = var_6705, interleave = rot_interleave_0, values = (var_6698_cast_fp16, var_6703_cast_fp16))[name = string("rot_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6707_cast_fp16 = mul(x = z_cast_fp16, y = c_cast_fp16)[name = string("op_6707_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6708_cast_fp16 = mul(x = rot_cast_fp16, y = s_11_cast_fp16)[name = string("op_6708_cast_fp16")];
            tensor<fp16, [1, 1260, 12, 32]> var_6710_cast_fp16 = add(x = var_6707_cast_fp16, y = var_6708_cast_fp16)[name = string("op_6710_cast_fp16")];
            int32 var_6712 = const()[name = string("op_6712"), val = int32(-1)];
            bool var_6713_interleave_0 = const()[name = string("op_6713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 1260, 12, 64]> var_6713_cast_fp16 = concat(axis = var_6712, interleave = var_6713_interleave_0, values = (var_6687_cast_fp16, var_6710_cast_fp16))[name = string("op_6713_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6723_transpose_x_0 = const()[name = string("op_6723_transpose_x_0"), val = bool(false)];
            bool var_6723_transpose_y_0 = const()[name = string("op_6723_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 1260]> transpose_95 = transpose(perm = transpose_95_perm_0, x = var_6713_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 12, 1260, 64]> transpose_94 = transpose(perm = transpose_94_perm_0, x = var_6661_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 12, 1260, 1260]> var_6723_cast_fp16 = matmul(transpose_x = var_6723_transpose_x_0, transpose_y = var_6723_transpose_y_0, x = transpose_94, y = transpose_95)[name = string("op_6723_cast_fp16")];
            tensor<fp16, [1, 12, 1260, 1260]> var_6725_cast_fp16 = add(x = var_6723_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6725_cast_fp16")];
            int32 var_6726 = const()[name = string("op_6726"), val = int32(-1)];
            tensor<fp16, [1, 12, 1260, 1260]> w_217_cast_fp16 = softmax(axis = var_6726, x = var_6725_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6729_transpose_x_0 = const()[name = string("op_6729_transpose_x_0"), val = bool(false)];
            bool var_6729_transpose_y_0 = const()[name = string("op_6729_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 1260, 64]> v_cast_fp16 = transpose(perm = v_perm_0, x = v_61_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 12, 1260, 64]> var_6729_cast_fp16 = matmul(transpose_x = var_6729_transpose_x_0, transpose_y = var_6729_transpose_y_0, x = w_217_cast_fp16, y = v_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [4]> var_6732_perm_0 = const()[name = string("op_6732_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6736 = const()[name = string("op_6736"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 1260, 12, 64]> var_6732_cast_fp16 = transpose(perm = var_6732_perm_0, x = var_6729_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 1260, 768]> input_155_cast_fp16 = reshape(shape = var_6736, x = var_6732_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144345920))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144935808))))[name = string("t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_109_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6741_cast_fp16 = abs(x = linear_109_cast_fp16)[name = string("op_6741_cast_fp16")];
            bool var_6744 = const()[name = string("op_6744"), val = bool(true)];
            tensor<int32, [1]> var_6745_axes_0 = const()[name = string("op_6745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6745_cast_fp16 = reduce_max(axes = var_6745_axes_0, keep_dims = var_6744, x = var_6741_cast_fp16)[name = string("op_6745_cast_fp16")];
            fp16 var_6746_to_fp16 = const()[name = string("op_6746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_223_cast_fp16 = maximum(x = var_6745_cast_fp16, y = var_6746_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_219_cast_fp16 = real_div(x = linear_109_cast_fp16, y = s_223_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6749_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6749_cast_fp16")];
            tensor<int32, [1]> var_6754_axes_0 = const()[name = string("op_6754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6754_keep_dims_0 = const()[name = string("op_6754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6754_cast_fp16 = reduce_mean(axes = var_6754_axes_0, keep_dims = var_6754_keep_dims_0, x = var_6749_cast_fp16)[name = string("op_6754_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6755_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6755_cast_fp16")];
            fp32 var_6756_epsilon_0 = const()[name = string("op_6756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6756_cast_fp16 = inverse(epsilon = var_6756_epsilon_0, x = var_6755_cast_fp16)[name = string("op_6756_cast_fp16")];
            fp16 var_6757_to_fp16 = const()[name = string("op_6757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6758_cast_fp16 = mul(x = var_6756_cast_fp16, y = var_6757_to_fp16)[name = string("op_6758_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6760_cast_fp16 = add(x = var_6754_cast_fp16, y = var_6758_cast_fp16)[name = string("op_6760_cast_fp16")];
            fp32 var_6761_epsilon_0 = const()[name = string("op_6761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6761_cast_fp16 = rsqrt(epsilon = var_6761_epsilon_0, x = var_6760_cast_fp16)[name = string("op_6761_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_187_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6761_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144937408)))];
            tensor<fp16, [1, 1260, 768]> var_6763_cast_fp16 = mul(x = out_187_cast_fp16, y = t_encoder_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> x_285_cast_fp16 = add(x = x_271_cast_fp16, y = var_6763_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6766_cast_fp16 = abs(x = x_285_cast_fp16)[name = string("op_6766_cast_fp16")];
            bool var_6769 = const()[name = string("op_6769"), val = bool(true)];
            tensor<int32, [1]> var_6770_axes_0 = const()[name = string("op_6770_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6770_cast_fp16 = reduce_max(axes = var_6770_axes_0, keep_dims = var_6769, x = var_6766_cast_fp16)[name = string("op_6770_cast_fp16")];
            fp16 var_6771_to_fp16 = const()[name = string("op_6771_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_225_cast_fp16 = maximum(x = var_6770_cast_fp16, y = var_6771_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_221_cast_fp16 = real_div(x = x_285_cast_fp16, y = s_225_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6774_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6774_cast_fp16")];
            tensor<int32, [1]> var_6779_axes_0 = const()[name = string("op_6779_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6779_keep_dims_0 = const()[name = string("op_6779_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6779_cast_fp16 = reduce_mean(axes = var_6779_axes_0, keep_dims = var_6779_keep_dims_0, x = var_6774_cast_fp16)[name = string("op_6779_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6780_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6780_cast_fp16")];
            fp32 var_6781_epsilon_0 = const()[name = string("op_6781_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6781_cast_fp16 = inverse(epsilon = var_6781_epsilon_0, x = var_6780_cast_fp16)[name = string("op_6781_cast_fp16")];
            fp16 var_6782_to_fp16 = const()[name = string("op_6782_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6783_cast_fp16 = mul(x = var_6781_cast_fp16, y = var_6782_to_fp16)[name = string("op_6783_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6785_cast_fp16 = add(x = var_6779_cast_fp16, y = var_6783_cast_fp16)[name = string("op_6785_cast_fp16")];
            fp32 var_6786_epsilon_0 = const()[name = string("op_6786_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6786_cast_fp16 = rsqrt(epsilon = var_6786_epsilon_0, x = var_6785_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_189_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6786_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144939008)))];
            tensor<fp16, [1, 1260, 768]> input_157_cast_fp16 = mul(x = out_189_cast_fp16, y = t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144940608))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147299968))))[name = string("t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_110_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_110_cast_fp16")];
            string var_6793_mode_0 = const()[name = string("op_6793_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 1260, 3072]> var_6793_cast_fp16 = gelu(mode = var_6793_mode_0, x = linear_110_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147306176))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149665536))))[name = string("t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 3072]> linear_111_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 1260, 3072]> input_161_cast_fp16 = mul(x = var_6793_cast_fp16, y = linear_111_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149671744))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152031104))))[name = string("t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 1260, 768]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6801_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6801_cast_fp16")];
            bool var_6804 = const()[name = string("op_6804"), val = bool(true)];
            tensor<int32, [1]> var_6805_axes_0 = const()[name = string("op_6805_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 1260, 1]> var_6805_cast_fp16 = reduce_max(axes = var_6805_axes_0, keep_dims = var_6804, x = var_6801_cast_fp16)[name = string("op_6805_cast_fp16")];
            fp16 var_6806_to_fp16 = const()[name = string("op_6806_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 1260, 1]> s_227_cast_fp16 = maximum(x = var_6805_cast_fp16, y = var_6806_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> y_223_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_227_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6809_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6809_cast_fp16")];
            tensor<int32, [1]> var_6814_axes_0 = const()[name = string("op_6814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6814_keep_dims_0 = const()[name = string("op_6814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1260, 1]> var_6814_cast_fp16 = reduce_mean(axes = var_6814_axes_0, keep_dims = var_6814_keep_dims_0, x = var_6809_cast_fp16)[name = string("op_6814_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6815_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6815_cast_fp16")];
            fp32 var_6816_epsilon_0 = const()[name = string("op_6816_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 1260, 1]> var_6816_cast_fp16 = inverse(epsilon = var_6816_epsilon_0, x = var_6815_cast_fp16)[name = string("op_6816_cast_fp16")];
            fp16 var_6817_to_fp16 = const()[name = string("op_6817_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1260, 1]> var_6818_cast_fp16 = mul(x = var_6816_cast_fp16, y = var_6817_to_fp16)[name = string("op_6818_cast_fp16")];
            tensor<fp16, [1, 1260, 1]> var_6820_cast_fp16 = add(x = var_6814_cast_fp16, y = var_6818_cast_fp16)[name = string("op_6820_cast_fp16")];
            fp32 var_6821_epsilon_0 = const()[name = string("op_6821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1260, 1]> var_6821_cast_fp16 = rsqrt(epsilon = var_6821_epsilon_0, x = var_6820_cast_fp16)[name = string("op_6821_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> out_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6821_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152032704)))];
            tensor<fp16, [1, 1260, 768]> var_6823_cast_fp16 = mul(x = out_cast_fp16, y = t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6823_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> h_cast_fp16 = add(x = x_285_cast_fp16, y = var_6823_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [1, 1260, 768]> var_6828_cast_fp16 = mul(x = h_cast_fp16, y = var_1413_cast_fp16)[name = string("op_6828_cast_fp16")];
            tensor<int32, [3]> var_6831_begin_0 = const()[name = string("op_6831_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_6831_end_0 = const()[name = string("op_6831_end_0"), val = tensor<int32, [3]>([1, 1260, 768])];
            tensor<bool, [3]> var_6831_end_mask_0 = const()[name = string("op_6831_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> var_6831_squeeze_mask_0 = const()[name = string("op_6831_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<fp16, [1260, 768]> var_6831_cast_fp16 = slice_by_index(begin = var_6831_begin_0, end = var_6831_end_0, end_mask = var_6831_end_mask_0, squeeze_mask = var_6831_squeeze_mask_0, x = var_6828_cast_fp16)[name = string("op_6831_cast_fp16")];
            bool x_transpose_x_0 = const()[name = string("x_transpose_x_0"), val = bool(false)];
            bool x_transpose_y_0 = const()[name = string("x_transpose_y_0"), val = bool(false)];
            tensor<fp16, [140, 768]> x_cast_fp16 = matmul(transpose_x = x_transpose_x_0, transpose_y = x_transpose_y_0, x = pool, y = var_6831_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [140, 768]> var_6833_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_6833_cast_fp16")];
            bool var_6836 = const()[name = string("op_6836"), val = bool(true)];
            tensor<int32, [1]> var_6837_axes_0 = const()[name = string("op_6837_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [140, 1]> var_6837_cast_fp16 = reduce_max(axes = var_6837_axes_0, keep_dims = var_6836, x = var_6833_cast_fp16)[name = string("op_6837_cast_fp16")];
            fp16 var_6838_to_fp16 = const()[name = string("op_6838_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [140, 1]> s_cast_fp16 = maximum(x = var_6837_cast_fp16, y = var_6838_to_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [140, 768]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [140, 768]> var_6841_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_6841_cast_fp16")];
            tensor<int32, [1]> var_6846_axes_0 = const()[name = string("op_6846_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6846_keep_dims_0 = const()[name = string("op_6846_keep_dims_0"), val = bool(true)];
            tensor<fp16, [140, 1]> var_6846_cast_fp16 = reduce_mean(axes = var_6846_axes_0, keep_dims = var_6846_keep_dims_0, x = var_6841_cast_fp16)[name = string("op_6846_cast_fp16")];
            tensor<fp16, [140, 1]> var_6847_cast_fp16 = mul(x = s_cast_fp16, y = s_cast_fp16)[name = string("op_6847_cast_fp16")];
            fp32 var_6848_epsilon_0 = const()[name = string("op_6848_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [140, 1]> var_6848_cast_fp16 = inverse(epsilon = var_6848_epsilon_0, x = var_6847_cast_fp16)[name = string("op_6848_cast_fp16")];
            fp16 var_6849_to_fp16 = const()[name = string("op_6849_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [140, 1]> var_6850_cast_fp16 = mul(x = var_6848_cast_fp16, y = var_6849_to_fp16)[name = string("op_6850_cast_fp16")];
            tensor<fp16, [140, 1]> var_6852_cast_fp16 = add(x = var_6846_cast_fp16, y = var_6850_cast_fp16)[name = string("op_6852_cast_fp16")];
            fp32 var_6853_epsilon_0 = const()[name = string("op_6853_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [140, 1]> var_6853_cast_fp16 = rsqrt(epsilon = var_6853_epsilon_0, x = var_6852_cast_fp16)[name = string("op_6853_cast_fp16")];
            tensor<fp16, [140, 768]> input_cast_fp16 = mul(x = y_cast_fp16, y = var_6853_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [512, 768]> e_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152034304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152427584))))[name = string("e_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_113_bias_0_to_fp16 = const()[name = string("linear_113_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152428672)))];
            tensor<fp16, [140, 512]> linear_113_cast_fp16 = linear(bias = linear_113_bias_0_to_fp16, weight = e_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [1]> var_6859_axes_0 = const()[name = string("op_6859_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 140, 512]> image_tokens = expand_dims(axes = var_6859_axes_0, x = linear_113_cast_fp16)[name = string("op_6859_cast_fp16")];
        } -> (image_tokens);
    func vision_280<ios18>(tensor<fp16, [1, 2520, 768]> patches, tensor<fp16, [280, 2520]> pool, tensor<fp16, [1, 2520, 768]> position_embeddings, tensor<fp16, [1, 2520, 2]> positions, tensor<fp16, [1, 2520]> valid) {
            fp16 var_1405_to_fp16 = const()[name = string("op_1405_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 2520, 768]> var_1406_cast_fp16 = sub(x = patches, y = var_1405_to_fp16)[name = string("op_1406_cast_fp16")];
            fp16 var_1407_promoted_to_fp16 = const()[name = string("op_1407_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 2520, 768]> input_1_cast_fp16 = mul(x = var_1406_cast_fp16, y = var_1407_promoted_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [768, 768]> t_patch_embedder_input_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589952))))[name = string("t_patch_embedder_input_proj_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591552)))];
            tensor<fp16, [1, 2520, 768]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_patch_embedder_input_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [1]> var_1413_axes_0 = const()[name = string("op_1413_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2520, 1]> var_1413_cast_fp16 = expand_dims(axes = var_1413_axes_0, x = valid)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1414_cast_fp16 = mul(x = position_embeddings, y = var_1413_cast_fp16)[name = string("op_1414_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_1_cast_fp16 = add(x = linear_0_cast_fp16, y = var_1414_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [1]> var_1418_axes_0 = const()[name = string("op_1418_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 2520, 2, 1]> var_1418_cast_fp16 = expand_dims(axes = var_1418_axes_0, x = positions)[name = string("op_1418_cast_fp16")];
            tensor<fp16, [16]> inv_freq_to_fp16 = const()[name = string("inv_freq_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593152)))];
            tensor<fp16, [1, 2520, 2, 16]> f_1_cast_fp16 = mul(x = var_1418_cast_fp16, y = inv_freq_to_fp16)[name = string("f_1_cast_fp16")];
            int32 var_1421 = const()[name = string("op_1421"), val = int32(-1)];
            bool var_1422_interleave_0 = const()[name = string("op_1422_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 2, 32]> var_1422_cast_fp16 = concat(axis = var_1421, interleave = var_1422_interleave_0, values = (f_1_cast_fp16, f_1_cast_fp16))[name = string("op_1422_cast_fp16")];
            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 2520, 1, 2, 32]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = var_1422_cast_fp16)[name = string("f_cast_fp16")];
            tensor<int32, [5]> var_1452_begin_0 = const()[name = string("op_1452_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1452_end_0 = const()[name = string("op_1452_end_0"), val = tensor<int32, [5]>([1, 2520, 1, 1, 32])];
            tensor<bool, [5]> var_1452_end_mask_0 = const()[name = string("op_1452_end_mask_0"), val = tensor<bool, [5]>([true, true, true, false, true])];
            tensor<bool, [5]> var_1452_squeeze_mask_0 = const()[name = string("op_1452_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, true, false])];
            tensor<fp16, [1, 2520, 1, 32]> var_1452_cast_fp16 = slice_by_index(begin = var_1452_begin_0, end = var_1452_end_0, end_mask = var_1452_end_mask_0, squeeze_mask = var_1452_squeeze_mask_0, x = f_cast_fp16)[name = string("op_1452_cast_fp16")];
            tensor<fp16, [1, 2520, 1, 32]> c_1_cast_fp16 = cos(x = var_1452_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [5]> var_1471_begin_0 = const()[name = string("op_1471_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 1, 0])];
            tensor<int32, [5]> var_1471_end_0 = const()[name = string("op_1471_end_0"), val = tensor<int32, [5]>([1, 2520, 1, 2, 32])];
            tensor<bool, [5]> var_1471_end_mask_0 = const()[name = string("op_1471_end_mask_0"), val = tensor<bool, [5]>([true, true, true, false, true])];
            tensor<bool, [5]> var_1471_squeeze_mask_0 = const()[name = string("op_1471_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, true, false])];
            tensor<fp16, [1, 2520, 1, 32]> var_1471_cast_fp16 = slice_by_index(begin = var_1471_begin_0, end = var_1471_end_0, end_mask = var_1471_end_mask_0, squeeze_mask = var_1471_squeeze_mask_0, x = f_cast_fp16)[name = string("op_1471_cast_fp16")];
            tensor<fp16, [1, 2520, 1, 32]> c_cast_fp16 = cos(x = var_1471_cast_fp16)[name = string("c_cast_fp16")];
            tensor<fp16, [1, 2520, 1, 32]> s_9_cast_fp16 = sin(x = var_1452_cast_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 2520, 1, 32]> s_11_cast_fp16 = sin(x = var_1471_cast_fp16)[name = string("s_11_cast_fp16")];
            fp16 var_1511_to_fp16 = const()[name = string("op_1511_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 2520]> var_1513_cast_fp16 = sub(x = var_1511_to_fp16, y = valid)[name = string("op_1513_cast_fp16")];
            fp16 var_1514_to_fp16 = const()[name = string("op_1514_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 2520]> var_1515_cast_fp16 = mul(x = var_1513_cast_fp16, y = var_1514_to_fp16)[name = string("op_1515_cast_fp16")];
            tensor<int32, [1]> var_1522_axes_0 = const()[name = string("op_1522_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 2520]> var_1522_cast_fp16 = expand_dims(axes = var_1522_axes_0, x = var_1515_cast_fp16)[name = string("op_1522_cast_fp16")];
            tensor<int32, [1]> var_1524_axes_0 = const()[name = string("op_1524_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 2520]> var_1524_cast_fp16 = expand_dims(axes = var_1524_axes_0, x = var_1522_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1530_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_1530_cast_fp16")];
            bool var_1533 = const()[name = string("op_1533"), val = bool(true)];
            tensor<int32, [1]> var_1534_axes_0 = const()[name = string("op_1534_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_1534_cast_fp16 = reduce_max(axes = var_1534_axes_0, keep_dims = var_1533, x = var_1530_cast_fp16)[name = string("op_1534_cast_fp16")];
            fp16 var_1535_to_fp16 = const()[name = string("op_1535_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_1_cast_fp16 = maximum(x = var_1534_cast_fp16, y = var_1535_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1538_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_1538_cast_fp16")];
            tensor<int32, [1]> var_1543_axes_0 = const()[name = string("op_1543_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1543_keep_dims_0 = const()[name = string("op_1543_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_1543_cast_fp16 = reduce_mean(axes = var_1543_axes_0, keep_dims = var_1543_keep_dims_0, x = var_1538_cast_fp16)[name = string("op_1543_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1544_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_1544_cast_fp16")];
            fp32 var_1545_epsilon_0 = const()[name = string("op_1545_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_1545_cast_fp16 = inverse(epsilon = var_1545_epsilon_0, x = var_1544_cast_fp16)[name = string("op_1545_cast_fp16")];
            fp16 var_1546_to_fp16 = const()[name = string("op_1546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_1547_cast_fp16 = mul(x = var_1545_cast_fp16, y = var_1546_to_fp16)[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1549_cast_fp16 = add(x = var_1543_cast_fp16, y = var_1547_cast_fp16)[name = string("op_1549_cast_fp16")];
            fp32 var_1550_epsilon_0 = const()[name = string("op_1550_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_1550_cast_fp16 = rsqrt(epsilon = var_1550_epsilon_0, x = var_1549_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_1550_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593280)))];
            tensor<fp16, [1, 2520, 768]> input_3_cast_fp16 = mul(x = out_1_cast_fp16, y = t_encoder_layers_0_input_layernorm_weight_to_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1184768))))[name = string("t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_1_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_1560 = const()[name = string("op_1560"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_3_cast_fp16 = reshape(shape = var_1560, x = linear_1_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1562_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_1562_cast_fp16")];
            bool var_1565 = const()[name = string("op_1565"), val = bool(true)];
            tensor<int32, [1]> var_1566_axes_0 = const()[name = string("op_1566_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_1566_cast_fp16 = reduce_max(axes = var_1566_axes_0, keep_dims = var_1565, x = var_1562_cast_fp16)[name = string("op_1566_cast_fp16")];
            fp16 var_1567_to_fp16 = const()[name = string("op_1567_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_3_cast_fp16 = maximum(x = var_1566_cast_fp16, y = var_1567_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_3_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1570_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_1570_cast_fp16")];
            tensor<int32, [1]> var_1575_axes_0 = const()[name = string("op_1575_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1575_keep_dims_0 = const()[name = string("op_1575_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_1575_cast_fp16 = reduce_mean(axes = var_1575_axes_0, keep_dims = var_1575_keep_dims_0, x = var_1570_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1576_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_1576_cast_fp16")];
            fp32 var_1577_epsilon_0 = const()[name = string("op_1577_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_1577_cast_fp16 = inverse(epsilon = var_1577_epsilon_0, x = var_1576_cast_fp16)[name = string("op_1577_cast_fp16")];
            fp16 var_1578_to_fp16 = const()[name = string("op_1578_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_1579_cast_fp16 = mul(x = var_1577_cast_fp16, y = var_1578_to_fp16)[name = string("op_1579_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1581_cast_fp16 = add(x = var_1575_cast_fp16, y = var_1579_cast_fp16)[name = string("op_1581_cast_fp16")];
            fp32 var_1582_epsilon_0 = const()[name = string("op_1582_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_1582_cast_fp16 = rsqrt(epsilon = var_1582_epsilon_0, x = var_1581_cast_fp16)[name = string("op_1582_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_1582_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 2520, 12, 64]> x_9_cast_fp16 = mul(x = out_3_cast_fp16, y = t_encoder_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186560))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1776448))))[name = string("t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1592 = const()[name = string("op_1592"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_5_cast_fp16 = reshape(shape = var_1592, x = linear_2_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1594_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_1594_cast_fp16")];
            bool var_1597 = const()[name = string("op_1597"), val = bool(true)];
            tensor<int32, [1]> var_1598_axes_0 = const()[name = string("op_1598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_1598_cast_fp16 = reduce_max(axes = var_1598_axes_0, keep_dims = var_1597, x = var_1594_cast_fp16)[name = string("op_1598_cast_fp16")];
            fp16 var_1599_to_fp16 = const()[name = string("op_1599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_5_cast_fp16 = maximum(x = var_1598_cast_fp16, y = var_1599_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_5_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1602_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_1602_cast_fp16")];
            tensor<int32, [1]> var_1607_axes_0 = const()[name = string("op_1607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1607_keep_dims_0 = const()[name = string("op_1607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_1607_cast_fp16 = reduce_mean(axes = var_1607_axes_0, keep_dims = var_1607_keep_dims_0, x = var_1602_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1608_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_1608_cast_fp16")];
            fp32 var_1609_epsilon_0 = const()[name = string("op_1609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_1609_cast_fp16 = inverse(epsilon = var_1609_epsilon_0, x = var_1608_cast_fp16)[name = string("op_1609_cast_fp16")];
            fp16 var_1610_to_fp16 = const()[name = string("op_1610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_1611_cast_fp16 = mul(x = var_1609_cast_fp16, y = var_1610_to_fp16)[name = string("op_1611_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1613_cast_fp16 = add(x = var_1607_cast_fp16, y = var_1611_cast_fp16)[name = string("op_1613_cast_fp16")];
            fp32 var_1614_epsilon_0 = const()[name = string("op_1614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_1614_cast_fp16 = rsqrt(epsilon = var_1614_epsilon_0, x = var_1613_cast_fp16)[name = string("op_1614_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_1614_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 2520, 12, 64]> x_11_cast_fp16 = mul(x = out_5_cast_fp16, y = t_encoder_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778240))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2368128))))[name = string("t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_1624 = const()[name = string("op_1624"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_7_cast_fp16 = reshape(shape = var_1624, x = linear_3_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1626_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_1626_cast_fp16")];
            bool var_1629 = const()[name = string("op_1629"), val = bool(true)];
            tensor<int32, [1]> var_1630_axes_0 = const()[name = string("op_1630_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_1630_cast_fp16 = reduce_max(axes = var_1630_axes_0, keep_dims = var_1629, x = var_1626_cast_fp16)[name = string("op_1630_cast_fp16")];
            fp16 var_1631_to_fp16 = const()[name = string("op_1631_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_7_cast_fp16 = maximum(x = var_1630_cast_fp16, y = var_1631_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_7_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1634_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_1634_cast_fp16")];
            tensor<int32, [1]> var_1639_axes_0 = const()[name = string("op_1639_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1639_keep_dims_0 = const()[name = string("op_1639_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_1639_cast_fp16 = reduce_mean(axes = var_1639_axes_0, keep_dims = var_1639_keep_dims_0, x = var_1634_cast_fp16)[name = string("op_1639_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1640_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_1640_cast_fp16")];
            fp32 var_1641_epsilon_0 = const()[name = string("op_1641_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_1641_cast_fp16 = inverse(epsilon = var_1641_epsilon_0, x = var_1640_cast_fp16)[name = string("op_1641_cast_fp16")];
            fp16 var_1642_to_fp16 = const()[name = string("op_1642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_1643_cast_fp16 = mul(x = var_1641_cast_fp16, y = var_1642_to_fp16)[name = string("op_1643_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1645_cast_fp16 = add(x = var_1639_cast_fp16, y = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp32 var_1646_epsilon_0 = const()[name = string("op_1646_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_1646_cast_fp16 = rsqrt(epsilon = var_1646_epsilon_0, x = var_1645_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_1_cast_fp16 = mul(x = y_7_cast_fp16, y = var_1646_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<int32, [4]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = x_9_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [4]> var_1657_begin_0 = const()[name = string("op_1657_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1657_end_0 = const()[name = string("op_1657_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_1657_end_mask_0 = const()[name = string("op_1657_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_1657_cast_fp16 = slice_by_index(begin = var_1657_begin_0, end = var_1657_end_0, end_mask = var_1657_end_mask_0, x = z_1_cast_fp16)[name = string("op_1657_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_1658_cast_fp16 = mul(x = var_1657_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [4]> var_1663_begin_0 = const()[name = string("op_1663_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1663_end_0 = const()[name = string("op_1663_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_1663_end_mask_0 = const()[name = string("op_1663_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_1663_cast_fp16 = slice_by_index(begin = var_1663_begin_0, end = var_1663_end_0, end_mask = var_1663_end_mask_0, x = z_1_cast_fp16)[name = string("op_1663_cast_fp16")];
            int32 var_1665 = const()[name = string("op_1665"), val = int32(-1)];
            bool rot_1_interleave_0 = const()[name = string("rot_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_1_cast_fp16 = concat(axis = var_1665, interleave = rot_1_interleave_0, values = (var_1658_cast_fp16, var_1663_cast_fp16))[name = string("rot_1_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1667_cast_fp16 = mul(x = z_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_1667_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1668_cast_fp16 = mul(x = rot_1_cast_fp16, y = s_9_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1670_cast_fp16 = add(x = var_1667_cast_fp16, y = var_1668_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<int32, [4]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = x_9_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [4]> var_1680_begin_0 = const()[name = string("op_1680_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1680_end_0 = const()[name = string("op_1680_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_1680_end_mask_0 = const()[name = string("op_1680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_1680_cast_fp16 = slice_by_index(begin = var_1680_begin_0, end = var_1680_end_0, end_mask = var_1680_end_mask_0, x = z_3_cast_fp16)[name = string("op_1680_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_1681_cast_fp16 = mul(x = var_1680_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> var_1686_begin_0 = const()[name = string("op_1686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1686_end_0 = const()[name = string("op_1686_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_1686_end_mask_0 = const()[name = string("op_1686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_1686_cast_fp16 = slice_by_index(begin = var_1686_begin_0, end = var_1686_end_0, end_mask = var_1686_end_mask_0, x = z_3_cast_fp16)[name = string("op_1686_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), val = int32(-1)];
            bool rot_3_interleave_0 = const()[name = string("rot_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_3_cast_fp16 = concat(axis = var_1688, interleave = rot_3_interleave_0, values = (var_1681_cast_fp16, var_1686_cast_fp16))[name = string("rot_3_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1690_cast_fp16 = mul(x = z_3_cast_fp16, y = c_cast_fp16)[name = string("op_1690_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1691_cast_fp16 = mul(x = rot_3_cast_fp16, y = s_11_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1693_cast_fp16 = add(x = var_1690_cast_fp16, y = var_1691_cast_fp16)[name = string("op_1693_cast_fp16")];
            int32 var_1695 = const()[name = string("op_1695"), val = int32(-1)];
            bool var_1696_interleave_0 = const()[name = string("op_1696_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_1696_cast_fp16 = concat(axis = var_1695, interleave = var_1696_interleave_0, values = (var_1670_cast_fp16, var_1693_cast_fp16))[name = string("op_1696_cast_fp16")];
            tensor<int32, [4]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = x_11_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [4]> var_1709_begin_0 = const()[name = string("op_1709_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1709_end_0 = const()[name = string("op_1709_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_1709_end_mask_0 = const()[name = string("op_1709_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_1709_cast_fp16 = slice_by_index(begin = var_1709_begin_0, end = var_1709_end_0, end_mask = var_1709_end_mask_0, x = z_5_cast_fp16)[name = string("op_1709_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_1710_cast_fp16 = mul(x = var_1709_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_1710_cast_fp16")];
            tensor<int32, [4]> var_1715_begin_0 = const()[name = string("op_1715_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1715_end_0 = const()[name = string("op_1715_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_1715_end_mask_0 = const()[name = string("op_1715_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_1715_cast_fp16 = slice_by_index(begin = var_1715_begin_0, end = var_1715_end_0, end_mask = var_1715_end_mask_0, x = z_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            int32 var_1717 = const()[name = string("op_1717"), val = int32(-1)];
            bool rot_5_interleave_0 = const()[name = string("rot_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_5_cast_fp16 = concat(axis = var_1717, interleave = rot_5_interleave_0, values = (var_1710_cast_fp16, var_1715_cast_fp16))[name = string("rot_5_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1719_cast_fp16 = mul(x = z_5_cast_fp16, y = c_1_cast_fp16)[name = string("op_1719_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1720_cast_fp16 = mul(x = rot_5_cast_fp16, y = s_9_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1722_cast_fp16 = add(x = var_1719_cast_fp16, y = var_1720_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<int32, [4]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = x_11_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [4]> var_1732_begin_0 = const()[name = string("op_1732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1732_end_0 = const()[name = string("op_1732_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_1732_end_mask_0 = const()[name = string("op_1732_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_1732_cast_fp16 = slice_by_index(begin = var_1732_begin_0, end = var_1732_end_0, end_mask = var_1732_end_mask_0, x = z_7_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_1733_cast_fp16 = mul(x = var_1732_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_1733_cast_fp16")];
            tensor<int32, [4]> var_1738_begin_0 = const()[name = string("op_1738_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1738_end_0 = const()[name = string("op_1738_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_1738_end_mask_0 = const()[name = string("op_1738_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_1738_cast_fp16 = slice_by_index(begin = var_1738_begin_0, end = var_1738_end_0, end_mask = var_1738_end_mask_0, x = z_7_cast_fp16)[name = string("op_1738_cast_fp16")];
            int32 var_1740 = const()[name = string("op_1740"), val = int32(-1)];
            bool rot_7_interleave_0 = const()[name = string("rot_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_7_cast_fp16 = concat(axis = var_1740, interleave = rot_7_interleave_0, values = (var_1733_cast_fp16, var_1738_cast_fp16))[name = string("rot_7_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1742_cast_fp16 = mul(x = z_7_cast_fp16, y = c_cast_fp16)[name = string("op_1742_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1743_cast_fp16 = mul(x = rot_7_cast_fp16, y = s_11_cast_fp16)[name = string("op_1743_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1745_cast_fp16 = add(x = var_1742_cast_fp16, y = var_1743_cast_fp16)[name = string("op_1745_cast_fp16")];
            int32 var_1747 = const()[name = string("op_1747"), val = int32(-1)];
            bool var_1748_interleave_0 = const()[name = string("op_1748_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_1748_cast_fp16 = concat(axis = var_1747, interleave = var_1748_interleave_0, values = (var_1722_cast_fp16, var_1745_cast_fp16))[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_1758_transpose_x_0 = const()[name = string("op_1758_transpose_x_0"), val = bool(false)];
            bool var_1758_transpose_y_0 = const()[name = string("op_1758_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1748_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_64 = transpose(perm = transpose_64_perm_0, x = var_1696_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 12, 2520, 2520]> var_1758_cast_fp16 = matmul(transpose_x = var_1758_transpose_x_0, transpose_y = var_1758_transpose_y_0, x = transpose_64, y = transpose_65)[name = string("op_1758_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_1760_cast_fp16 = add(x = var_1758_cast_fp16, y = var_1524_cast_fp16)[name = string("op_1760_cast_fp16")];
            int32 var_1761 = const()[name = string("op_1761"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_7_cast_fp16 = softmax(axis = var_1761, x = var_1760_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1764_transpose_x_0 = const()[name = string("op_1764_transpose_x_0"), val = bool(false)];
            bool var_1764_transpose_y_0 = const()[name = string("op_1764_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = v_1_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 12, 2520, 64]> var_1764_cast_fp16 = matmul(transpose_x = var_1764_transpose_x_0, transpose_y = var_1764_transpose_y_0, x = w_7_cast_fp16, y = v_3_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<int32, [4]> var_1767_perm_0 = const()[name = string("op_1767_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1771 = const()[name = string("op_1771"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = var_1764_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 2520, 768]> input_5_cast_fp16 = reshape(shape = var_1771, x = var_1767_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2369728))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2959616))))[name = string("t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1776_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_1776_cast_fp16")];
            bool var_1779 = const()[name = string("op_1779"), val = bool(true)];
            tensor<int32, [1]> var_1780_axes_0 = const()[name = string("op_1780_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_1780_cast_fp16 = reduce_max(axes = var_1780_axes_0, keep_dims = var_1779, x = var_1776_cast_fp16)[name = string("op_1780_cast_fp16")];
            fp16 var_1781_to_fp16 = const()[name = string("op_1781_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_13_cast_fp16 = maximum(x = var_1780_cast_fp16, y = var_1781_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_9_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_13_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1784_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_1784_cast_fp16")];
            tensor<int32, [1]> var_1789_axes_0 = const()[name = string("op_1789_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1789_keep_dims_0 = const()[name = string("op_1789_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_1789_cast_fp16 = reduce_mean(axes = var_1789_axes_0, keep_dims = var_1789_keep_dims_0, x = var_1784_cast_fp16)[name = string("op_1789_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1790_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_1790_cast_fp16")];
            fp32 var_1791_epsilon_0 = const()[name = string("op_1791_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_1791_cast_fp16 = inverse(epsilon = var_1791_epsilon_0, x = var_1790_cast_fp16)[name = string("op_1791_cast_fp16")];
            fp16 var_1792_to_fp16 = const()[name = string("op_1792_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_1793_cast_fp16 = mul(x = var_1791_cast_fp16, y = var_1792_to_fp16)[name = string("op_1793_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1795_cast_fp16 = add(x = var_1789_cast_fp16, y = var_1793_cast_fp16)[name = string("op_1795_cast_fp16")];
            fp32 var_1796_epsilon_0 = const()[name = string("op_1796_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_1796_cast_fp16 = rsqrt(epsilon = var_1796_epsilon_0, x = var_1795_cast_fp16)[name = string("op_1796_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_7_cast_fp16 = mul(x = y_9_cast_fp16, y = var_1796_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2961216)))];
            tensor<fp16, [1, 2520, 768]> var_1798_cast_fp16 = mul(x = out_7_cast_fp16, y = t_encoder_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_1798_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_15_cast_fp16 = add(x = x_1_cast_fp16, y = var_1798_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1801_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_1801_cast_fp16")];
            bool var_1804 = const()[name = string("op_1804"), val = bool(true)];
            tensor<int32, [1]> var_1805_axes_0 = const()[name = string("op_1805_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_1805_cast_fp16 = reduce_max(axes = var_1805_axes_0, keep_dims = var_1804, x = var_1801_cast_fp16)[name = string("op_1805_cast_fp16")];
            fp16 var_1806_to_fp16 = const()[name = string("op_1806_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_15_cast_fp16 = maximum(x = var_1805_cast_fp16, y = var_1806_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_11_cast_fp16 = real_div(x = x_15_cast_fp16, y = s_15_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1809_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [1]> var_1814_axes_0 = const()[name = string("op_1814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1814_keep_dims_0 = const()[name = string("op_1814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_1814_cast_fp16 = reduce_mean(axes = var_1814_axes_0, keep_dims = var_1814_keep_dims_0, x = var_1809_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1815_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_1815_cast_fp16")];
            fp32 var_1816_epsilon_0 = const()[name = string("op_1816_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_1816_cast_fp16 = inverse(epsilon = var_1816_epsilon_0, x = var_1815_cast_fp16)[name = string("op_1816_cast_fp16")];
            fp16 var_1817_to_fp16 = const()[name = string("op_1817_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_1818_cast_fp16 = mul(x = var_1816_cast_fp16, y = var_1817_to_fp16)[name = string("op_1818_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1820_cast_fp16 = add(x = var_1814_cast_fp16, y = var_1818_cast_fp16)[name = string("op_1820_cast_fp16")];
            fp32 var_1821_epsilon_0 = const()[name = string("op_1821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_1821_cast_fp16 = rsqrt(epsilon = var_1821_epsilon_0, x = var_1820_cast_fp16)[name = string("op_1821_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_1821_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2962816)))];
            tensor<fp16, [1, 2520, 768]> input_7_cast_fp16 = mul(x = out_9_cast_fp16, y = t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2964416))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5323776))))[name = string("t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5329984)))];
            tensor<fp16, [1, 2520, 3072]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_1828_mode_0 = const()[name = string("op_1828_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_1828_cast_fp16 = gelu(mode = var_1828_mode_0, x = linear_5_cast_fp16)[name = string("op_1828_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5336192))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7695552))))[name = string("t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_11_cast_fp16 = mul(x = var_1828_cast_fp16, y = linear_6_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7701760))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10061120))))[name = string("t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1836_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_1836_cast_fp16")];
            bool var_1839 = const()[name = string("op_1839"), val = bool(true)];
            tensor<int32, [1]> var_1840_axes_0 = const()[name = string("op_1840_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_1840_cast_fp16 = reduce_max(axes = var_1840_axes_0, keep_dims = var_1839, x = var_1836_cast_fp16)[name = string("op_1840_cast_fp16")];
            fp16 var_1841_to_fp16 = const()[name = string("op_1841_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_17_cast_fp16 = maximum(x = var_1840_cast_fp16, y = var_1841_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_13_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_17_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1844_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_1844_cast_fp16")];
            tensor<int32, [1]> var_1849_axes_0 = const()[name = string("op_1849_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1849_keep_dims_0 = const()[name = string("op_1849_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_1849_cast_fp16 = reduce_mean(axes = var_1849_axes_0, keep_dims = var_1849_keep_dims_0, x = var_1844_cast_fp16)[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1850_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_1850_cast_fp16")];
            fp32 var_1851_epsilon_0 = const()[name = string("op_1851_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_1851_cast_fp16 = inverse(epsilon = var_1851_epsilon_0, x = var_1850_cast_fp16)[name = string("op_1851_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_1853_cast_fp16 = mul(x = var_1851_cast_fp16, y = var_1852_to_fp16)[name = string("op_1853_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1855_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1853_cast_fp16)[name = string("op_1855_cast_fp16")];
            fp32 var_1856_epsilon_0 = const()[name = string("op_1856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_1856_cast_fp16 = rsqrt(epsilon = var_1856_epsilon_0, x = var_1855_cast_fp16)[name = string("op_1856_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_1856_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10062720)))];
            tensor<fp16, [1, 2520, 768]> var_1858_cast_fp16 = mul(x = out_11_cast_fp16, y = t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_1858_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_19_cast_fp16 = add(x = x_15_cast_fp16, y = var_1858_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1861_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_1861_cast_fp16")];
            bool var_1864 = const()[name = string("op_1864"), val = bool(true)];
            tensor<int32, [1]> var_1865_axes_0 = const()[name = string("op_1865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_1865_cast_fp16 = reduce_max(axes = var_1865_axes_0, keep_dims = var_1864, x = var_1861_cast_fp16)[name = string("op_1865_cast_fp16")];
            fp16 var_1866_to_fp16 = const()[name = string("op_1866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_19_cast_fp16 = maximum(x = var_1865_cast_fp16, y = var_1866_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_15_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_1869_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_1869_cast_fp16")];
            tensor<int32, [1]> var_1874_axes_0 = const()[name = string("op_1874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1874_keep_dims_0 = const()[name = string("op_1874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_1874_cast_fp16 = reduce_mean(axes = var_1874_axes_0, keep_dims = var_1874_keep_dims_0, x = var_1869_cast_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1875_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_1875_cast_fp16")];
            fp32 var_1876_epsilon_0 = const()[name = string("op_1876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_1876_cast_fp16 = inverse(epsilon = var_1876_epsilon_0, x = var_1875_cast_fp16)[name = string("op_1876_cast_fp16")];
            fp16 var_1877_to_fp16 = const()[name = string("op_1877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_1878_cast_fp16 = mul(x = var_1876_cast_fp16, y = var_1877_to_fp16)[name = string("op_1878_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_1880_cast_fp16 = add(x = var_1874_cast_fp16, y = var_1878_cast_fp16)[name = string("op_1880_cast_fp16")];
            fp32 var_1881_epsilon_0 = const()[name = string("op_1881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_1881_cast_fp16 = rsqrt(epsilon = var_1881_epsilon_0, x = var_1880_cast_fp16)[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_1881_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10064320)))];
            tensor<fp16, [1, 2520, 768]> input_13_cast_fp16 = mul(x = out_13_cast_fp16, y = t_encoder_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10065920))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10655808))))[name = string("t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_1891 = const()[name = string("op_1891"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_21_cast_fp16 = reshape(shape = var_1891, x = linear_8_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1893_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1893_cast_fp16")];
            bool var_1896 = const()[name = string("op_1896"), val = bool(true)];
            tensor<int32, [1]> var_1897_axes_0 = const()[name = string("op_1897_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_1897_cast_fp16 = reduce_max(axes = var_1897_axes_0, keep_dims = var_1896, x = var_1893_cast_fp16)[name = string("op_1897_cast_fp16")];
            fp16 var_1898_to_fp16 = const()[name = string("op_1898_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_21_cast_fp16 = maximum(x = var_1897_cast_fp16, y = var_1898_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_17_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1901_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_1906_axes_0 = const()[name = string("op_1906_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1906_keep_dims_0 = const()[name = string("op_1906_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_1906_cast_fp16 = reduce_mean(axes = var_1906_axes_0, keep_dims = var_1906_keep_dims_0, x = var_1901_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1907_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_1907_cast_fp16")];
            fp32 var_1908_epsilon_0 = const()[name = string("op_1908_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_1908_cast_fp16 = inverse(epsilon = var_1908_epsilon_0, x = var_1907_cast_fp16)[name = string("op_1908_cast_fp16")];
            fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_1910_cast_fp16 = mul(x = var_1908_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1912_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1910_cast_fp16)[name = string("op_1912_cast_fp16")];
            fp32 var_1913_epsilon_0 = const()[name = string("op_1913_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_1913_cast_fp16 = rsqrt(epsilon = var_1913_epsilon_0, x = var_1912_cast_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_1913_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 2520, 12, 64]> x_27_cast_fp16 = mul(x = out_15_cast_fp16, y = t_encoder_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10657408))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11247296))))[name = string("t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_9_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_1923 = const()[name = string("op_1923"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_23_cast_fp16 = reshape(shape = var_1923, x = linear_9_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1925_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1925_cast_fp16")];
            bool var_1928 = const()[name = string("op_1928"), val = bool(true)];
            tensor<int32, [1]> var_1929_axes_0 = const()[name = string("op_1929_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_1929_cast_fp16 = reduce_max(axes = var_1929_axes_0, keep_dims = var_1928, x = var_1925_cast_fp16)[name = string("op_1929_cast_fp16")];
            fp16 var_1930_to_fp16 = const()[name = string("op_1930_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_23_cast_fp16 = maximum(x = var_1929_cast_fp16, y = var_1930_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_19_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1933_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<int32, [1]> var_1938_axes_0 = const()[name = string("op_1938_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1938_keep_dims_0 = const()[name = string("op_1938_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_1938_cast_fp16 = reduce_mean(axes = var_1938_axes_0, keep_dims = var_1938_keep_dims_0, x = var_1933_cast_fp16)[name = string("op_1938_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1939_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_1939_cast_fp16")];
            fp32 var_1940_epsilon_0 = const()[name = string("op_1940_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_1940_cast_fp16 = inverse(epsilon = var_1940_epsilon_0, x = var_1939_cast_fp16)[name = string("op_1940_cast_fp16")];
            fp16 var_1941_to_fp16 = const()[name = string("op_1941_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_1942_cast_fp16 = mul(x = var_1940_cast_fp16, y = var_1941_to_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1944_cast_fp16 = add(x = var_1938_cast_fp16, y = var_1942_cast_fp16)[name = string("op_1944_cast_fp16")];
            fp32 var_1945_epsilon_0 = const()[name = string("op_1945_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_1945_cast_fp16 = rsqrt(epsilon = var_1945_epsilon_0, x = var_1944_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_1945_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 2520, 12, 64]> x_29_cast_fp16 = mul(x = out_17_cast_fp16, y = t_encoder_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11248896))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11838784))))[name = string("t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_25_cast_fp16 = reshape(shape = var_1955, x = linear_10_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1957_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1957_cast_fp16")];
            bool var_1960 = const()[name = string("op_1960"), val = bool(true)];
            tensor<int32, [1]> var_1961_axes_0 = const()[name = string("op_1961_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_1961_cast_fp16 = reduce_max(axes = var_1961_axes_0, keep_dims = var_1960, x = var_1957_cast_fp16)[name = string("op_1961_cast_fp16")];
            fp16 var_1962_to_fp16 = const()[name = string("op_1962_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_25_cast_fp16 = maximum(x = var_1961_cast_fp16, y = var_1962_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_21_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_1965_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_1965_cast_fp16")];
            tensor<int32, [1]> var_1970_axes_0 = const()[name = string("op_1970_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1970_keep_dims_0 = const()[name = string("op_1970_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_1970_cast_fp16 = reduce_mean(axes = var_1970_axes_0, keep_dims = var_1970_keep_dims_0, x = var_1965_cast_fp16)[name = string("op_1970_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1971_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_1971_cast_fp16")];
            fp32 var_1972_epsilon_0 = const()[name = string("op_1972_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_1972_cast_fp16 = inverse(epsilon = var_1972_epsilon_0, x = var_1971_cast_fp16)[name = string("op_1972_cast_fp16")];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_1974_cast_fp16 = mul(x = var_1972_cast_fp16, y = var_1973_to_fp16)[name = string("op_1974_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_1976_cast_fp16 = add(x = var_1970_cast_fp16, y = var_1974_cast_fp16)[name = string("op_1976_cast_fp16")];
            fp32 var_1977_epsilon_0 = const()[name = string("op_1977_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_1977_cast_fp16 = rsqrt(epsilon = var_1977_epsilon_0, x = var_1976_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_5_cast_fp16 = mul(x = y_21_cast_fp16, y = var_1977_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<int32, [4]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = x_27_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [4]> var_1988_begin_0 = const()[name = string("op_1988_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1988_end_0 = const()[name = string("op_1988_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_1988_end_mask_0 = const()[name = string("op_1988_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_1988_cast_fp16 = slice_by_index(begin = var_1988_begin_0, end = var_1988_end_0, end_mask = var_1988_end_mask_0, x = z_9_cast_fp16)[name = string("op_1988_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_1989_cast_fp16 = mul(x = var_1988_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1989_cast_fp16")];
            tensor<int32, [4]> var_1994_begin_0 = const()[name = string("op_1994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1994_end_0 = const()[name = string("op_1994_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_1994_end_mask_0 = const()[name = string("op_1994_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_1994_cast_fp16 = slice_by_index(begin = var_1994_begin_0, end = var_1994_end_0, end_mask = var_1994_end_mask_0, x = z_9_cast_fp16)[name = string("op_1994_cast_fp16")];
            int32 var_1996 = const()[name = string("op_1996"), val = int32(-1)];
            bool rot_9_interleave_0 = const()[name = string("rot_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_9_cast_fp16 = concat(axis = var_1996, interleave = rot_9_interleave_0, values = (var_1989_cast_fp16, var_1994_cast_fp16))[name = string("rot_9_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1998_cast_fp16 = mul(x = z_9_cast_fp16, y = c_1_cast_fp16)[name = string("op_1998_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_1999_cast_fp16 = mul(x = rot_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_1999_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2001_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<int32, [4]> z_11_begin_0 = const()[name = string("z_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_11_end_0 = const()[name = string("z_11_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_11_end_mask_0 = const()[name = string("z_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_11_cast_fp16 = slice_by_index(begin = z_11_begin_0, end = z_11_end_0, end_mask = z_11_end_mask_0, x = x_27_cast_fp16)[name = string("z_11_cast_fp16")];
            tensor<int32, [4]> var_2011_begin_0 = const()[name = string("op_2011_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2011_end_0 = const()[name = string("op_2011_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2011_end_mask_0 = const()[name = string("op_2011_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2011_cast_fp16 = slice_by_index(begin = var_2011_begin_0, end = var_2011_end_0, end_mask = var_2011_end_mask_0, x = z_11_cast_fp16)[name = string("op_2011_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2012_cast_fp16 = mul(x = var_2011_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2012_cast_fp16")];
            tensor<int32, [4]> var_2017_begin_0 = const()[name = string("op_2017_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2017_end_0 = const()[name = string("op_2017_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2017_end_mask_0 = const()[name = string("op_2017_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2017_cast_fp16 = slice_by_index(begin = var_2017_begin_0, end = var_2017_end_0, end_mask = var_2017_end_mask_0, x = z_11_cast_fp16)[name = string("op_2017_cast_fp16")];
            int32 var_2019 = const()[name = string("op_2019"), val = int32(-1)];
            bool rot_11_interleave_0 = const()[name = string("rot_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_11_cast_fp16 = concat(axis = var_2019, interleave = rot_11_interleave_0, values = (var_2012_cast_fp16, var_2017_cast_fp16))[name = string("rot_11_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2021_cast_fp16 = mul(x = z_11_cast_fp16, y = c_cast_fp16)[name = string("op_2021_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2022_cast_fp16 = mul(x = rot_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2022_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2024_cast_fp16 = add(x = var_2021_cast_fp16, y = var_2022_cast_fp16)[name = string("op_2024_cast_fp16")];
            int32 var_2026 = const()[name = string("op_2026"), val = int32(-1)];
            bool var_2027_interleave_0 = const()[name = string("op_2027_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_2027_cast_fp16 = concat(axis = var_2026, interleave = var_2027_interleave_0, values = (var_2001_cast_fp16, var_2024_cast_fp16))[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> z_13_begin_0 = const()[name = string("z_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_13_end_0 = const()[name = string("z_13_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_13_end_mask_0 = const()[name = string("z_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_13_cast_fp16 = slice_by_index(begin = z_13_begin_0, end = z_13_end_0, end_mask = z_13_end_mask_0, x = x_29_cast_fp16)[name = string("z_13_cast_fp16")];
            tensor<int32, [4]> var_2040_begin_0 = const()[name = string("op_2040_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2040_end_0 = const()[name = string("op_2040_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2040_end_mask_0 = const()[name = string("op_2040_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, x = z_13_cast_fp16)[name = string("op_2040_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2041_cast_fp16 = mul(x = var_2040_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> var_2046_begin_0 = const()[name = string("op_2046_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2046_end_0 = const()[name = string("op_2046_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2046_end_mask_0 = const()[name = string("op_2046_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2046_cast_fp16 = slice_by_index(begin = var_2046_begin_0, end = var_2046_end_0, end_mask = var_2046_end_mask_0, x = z_13_cast_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2048 = const()[name = string("op_2048"), val = int32(-1)];
            bool rot_13_interleave_0 = const()[name = string("rot_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_13_cast_fp16 = concat(axis = var_2048, interleave = rot_13_interleave_0, values = (var_2041_cast_fp16, var_2046_cast_fp16))[name = string("rot_13_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2050_cast_fp16 = mul(x = z_13_cast_fp16, y = c_1_cast_fp16)[name = string("op_2050_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2051_cast_fp16 = mul(x = rot_13_cast_fp16, y = s_9_cast_fp16)[name = string("op_2051_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2053_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2051_cast_fp16)[name = string("op_2053_cast_fp16")];
            tensor<int32, [4]> z_15_begin_0 = const()[name = string("z_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_15_end_0 = const()[name = string("z_15_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_15_end_mask_0 = const()[name = string("z_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_15_cast_fp16 = slice_by_index(begin = z_15_begin_0, end = z_15_end_0, end_mask = z_15_end_mask_0, x = x_29_cast_fp16)[name = string("z_15_cast_fp16")];
            tensor<int32, [4]> var_2063_begin_0 = const()[name = string("op_2063_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2063_end_0 = const()[name = string("op_2063_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2063_end_mask_0 = const()[name = string("op_2063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2063_cast_fp16 = slice_by_index(begin = var_2063_begin_0, end = var_2063_end_0, end_mask = var_2063_end_mask_0, x = z_15_cast_fp16)[name = string("op_2063_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2064_cast_fp16 = mul(x = var_2063_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2064_cast_fp16")];
            tensor<int32, [4]> var_2069_begin_0 = const()[name = string("op_2069_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2069_end_0 = const()[name = string("op_2069_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2069_end_mask_0 = const()[name = string("op_2069_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2069_cast_fp16 = slice_by_index(begin = var_2069_begin_0, end = var_2069_end_0, end_mask = var_2069_end_mask_0, x = z_15_cast_fp16)[name = string("op_2069_cast_fp16")];
            int32 var_2071 = const()[name = string("op_2071"), val = int32(-1)];
            bool rot_15_interleave_0 = const()[name = string("rot_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_15_cast_fp16 = concat(axis = var_2071, interleave = rot_15_interleave_0, values = (var_2064_cast_fp16, var_2069_cast_fp16))[name = string("rot_15_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2073_cast_fp16 = mul(x = z_15_cast_fp16, y = c_cast_fp16)[name = string("op_2073_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2074_cast_fp16 = mul(x = rot_15_cast_fp16, y = s_11_cast_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2076_cast_fp16 = add(x = var_2073_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2076_cast_fp16")];
            int32 var_2078 = const()[name = string("op_2078"), val = int32(-1)];
            bool var_2079_interleave_0 = const()[name = string("op_2079_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_2079_cast_fp16 = concat(axis = var_2078, interleave = var_2079_interleave_0, values = (var_2053_cast_fp16, var_2076_cast_fp16))[name = string("op_2079_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2089_transpose_x_0 = const()[name = string("op_2089_transpose_x_0"), val = bool(false)];
            bool var_2089_transpose_y_0 = const()[name = string("op_2089_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_2079_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_66 = transpose(perm = transpose_66_perm_0, x = var_2027_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [1, 12, 2520, 2520]> var_2089_cast_fp16 = matmul(transpose_x = var_2089_transpose_x_0, transpose_y = var_2089_transpose_y_0, x = transpose_66, y = transpose_67)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_2091_cast_fp16 = add(x = var_2089_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2091_cast_fp16")];
            int32 var_2092 = const()[name = string("op_2092"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_21_cast_fp16 = softmax(axis = var_2092, x = var_2091_cast_fp16)[name = string("w_21_cast_fp16")];
            bool var_2095_transpose_x_0 = const()[name = string("op_2095_transpose_x_0"), val = bool(false)];
            bool var_2095_transpose_y_0 = const()[name = string("op_2095_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = v_5_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 12, 2520, 64]> var_2095_cast_fp16 = matmul(transpose_x = var_2095_transpose_x_0, transpose_y = var_2095_transpose_y_0, x = w_21_cast_fp16, y = v_7_cast_fp16)[name = string("op_2095_cast_fp16")];
            tensor<int32, [4]> var_2098_perm_0 = const()[name = string("op_2098_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2102 = const()[name = string("op_2102"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_2098_cast_fp16 = transpose(perm = var_2098_perm_0, x = var_2095_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 2520, 768]> input_15_cast_fp16 = reshape(shape = var_2102, x = var_2098_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11840384))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12430272))))[name = string("t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_11_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2107_cast_fp16 = abs(x = linear_11_cast_fp16)[name = string("op_2107_cast_fp16")];
            bool var_2110 = const()[name = string("op_2110"), val = bool(true)];
            tensor<int32, [1]> var_2111_axes_0 = const()[name = string("op_2111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2111_cast_fp16 = reduce_max(axes = var_2111_axes_0, keep_dims = var_2110, x = var_2107_cast_fp16)[name = string("op_2111_cast_fp16")];
            fp16 var_2112_to_fp16 = const()[name = string("op_2112_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_27_cast_fp16 = maximum(x = var_2111_cast_fp16, y = var_2112_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_23_cast_fp16 = real_div(x = linear_11_cast_fp16, y = s_27_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2115_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<int32, [1]> var_2120_axes_0 = const()[name = string("op_2120_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2120_keep_dims_0 = const()[name = string("op_2120_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2120_cast_fp16 = reduce_mean(axes = var_2120_axes_0, keep_dims = var_2120_keep_dims_0, x = var_2115_cast_fp16)[name = string("op_2120_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2121_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2121_cast_fp16")];
            fp32 var_2122_epsilon_0 = const()[name = string("op_2122_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2122_cast_fp16 = inverse(epsilon = var_2122_epsilon_0, x = var_2121_cast_fp16)[name = string("op_2122_cast_fp16")];
            fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2124_cast_fp16 = mul(x = var_2122_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2126_cast_fp16 = add(x = var_2120_cast_fp16, y = var_2124_cast_fp16)[name = string("op_2126_cast_fp16")];
            fp32 var_2127_epsilon_0 = const()[name = string("op_2127_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2127_cast_fp16 = rsqrt(epsilon = var_2127_epsilon_0, x = var_2126_cast_fp16)[name = string("op_2127_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_19_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2127_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12431872)))];
            tensor<fp16, [1, 2520, 768]> var_2129_cast_fp16 = mul(x = out_19_cast_fp16, y = t_encoder_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2129_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_33_cast_fp16 = add(x = x_19_cast_fp16, y = var_2129_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2132_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_2132_cast_fp16")];
            bool var_2135 = const()[name = string("op_2135"), val = bool(true)];
            tensor<int32, [1]> var_2136_axes_0 = const()[name = string("op_2136_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2136_cast_fp16 = reduce_max(axes = var_2136_axes_0, keep_dims = var_2135, x = var_2132_cast_fp16)[name = string("op_2136_cast_fp16")];
            fp16 var_2137_to_fp16 = const()[name = string("op_2137_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_29_cast_fp16 = maximum(x = var_2136_cast_fp16, y = var_2137_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_25_cast_fp16 = real_div(x = x_33_cast_fp16, y = s_29_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2140_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<int32, [1]> var_2145_axes_0 = const()[name = string("op_2145_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2145_keep_dims_0 = const()[name = string("op_2145_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2145_cast_fp16 = reduce_mean(axes = var_2145_axes_0, keep_dims = var_2145_keep_dims_0, x = var_2140_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2146_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2146_cast_fp16")];
            fp32 var_2147_epsilon_0 = const()[name = string("op_2147_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2147_cast_fp16 = inverse(epsilon = var_2147_epsilon_0, x = var_2146_cast_fp16)[name = string("op_2147_cast_fp16")];
            fp16 var_2148_to_fp16 = const()[name = string("op_2148_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2149_cast_fp16 = mul(x = var_2147_cast_fp16, y = var_2148_to_fp16)[name = string("op_2149_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2151_cast_fp16 = add(x = var_2145_cast_fp16, y = var_2149_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2152_cast_fp16 = rsqrt(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_21_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2152_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12433472)))];
            tensor<fp16, [1, 2520, 768]> input_17_cast_fp16 = mul(x = out_21_cast_fp16, y = t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12435072))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14794432))))[name = string("t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_12_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_12_cast_fp16")];
            string var_2159_mode_0 = const()[name = string("op_2159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_2159_cast_fp16 = gelu(mode = var_2159_mode_0, x = linear_12_cast_fp16)[name = string("op_2159_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14800640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17160000))))[name = string("t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_13_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_21_cast_fp16 = mul(x = var_2159_cast_fp16, y = linear_13_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17166208))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19525568))))[name = string("t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2167_cast_fp16 = abs(x = linear_14_cast_fp16)[name = string("op_2167_cast_fp16")];
            bool var_2170 = const()[name = string("op_2170"), val = bool(true)];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2171_cast_fp16 = reduce_max(axes = var_2171_axes_0, keep_dims = var_2170, x = var_2167_cast_fp16)[name = string("op_2171_cast_fp16")];
            fp16 var_2172_to_fp16 = const()[name = string("op_2172_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_31_cast_fp16 = maximum(x = var_2171_cast_fp16, y = var_2172_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_27_cast_fp16 = real_div(x = linear_14_cast_fp16, y = s_31_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2175_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2175_cast_fp16")];
            tensor<int32, [1]> var_2180_axes_0 = const()[name = string("op_2180_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2180_keep_dims_0 = const()[name = string("op_2180_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2180_cast_fp16 = reduce_mean(axes = var_2180_axes_0, keep_dims = var_2180_keep_dims_0, x = var_2175_cast_fp16)[name = string("op_2180_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2181_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2181_cast_fp16")];
            fp32 var_2182_epsilon_0 = const()[name = string("op_2182_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2182_cast_fp16 = inverse(epsilon = var_2182_epsilon_0, x = var_2181_cast_fp16)[name = string("op_2182_cast_fp16")];
            fp16 var_2183_to_fp16 = const()[name = string("op_2183_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2184_cast_fp16 = mul(x = var_2182_cast_fp16, y = var_2183_to_fp16)[name = string("op_2184_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2186_cast_fp16 = add(x = var_2180_cast_fp16, y = var_2184_cast_fp16)[name = string("op_2186_cast_fp16")];
            fp32 var_2187_epsilon_0 = const()[name = string("op_2187_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2187_cast_fp16 = rsqrt(epsilon = var_2187_epsilon_0, x = var_2186_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2187_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19527168)))];
            tensor<fp16, [1, 2520, 768]> var_2189_cast_fp16 = mul(x = out_23_cast_fp16, y = t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2189_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_37_cast_fp16 = add(x = x_33_cast_fp16, y = var_2189_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2192_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2192_cast_fp16")];
            bool var_2195 = const()[name = string("op_2195"), val = bool(true)];
            tensor<int32, [1]> var_2196_axes_0 = const()[name = string("op_2196_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2196_cast_fp16 = reduce_max(axes = var_2196_axes_0, keep_dims = var_2195, x = var_2192_cast_fp16)[name = string("op_2196_cast_fp16")];
            fp16 var_2197_to_fp16 = const()[name = string("op_2197_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_33_cast_fp16 = maximum(x = var_2196_cast_fp16, y = var_2197_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_29_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_33_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2200_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2200_cast_fp16")];
            tensor<int32, [1]> var_2205_axes_0 = const()[name = string("op_2205_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2205_keep_dims_0 = const()[name = string("op_2205_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2205_cast_fp16 = reduce_mean(axes = var_2205_axes_0, keep_dims = var_2205_keep_dims_0, x = var_2200_cast_fp16)[name = string("op_2205_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2206_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2206_cast_fp16")];
            fp32 var_2207_epsilon_0 = const()[name = string("op_2207_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2207_cast_fp16 = inverse(epsilon = var_2207_epsilon_0, x = var_2206_cast_fp16)[name = string("op_2207_cast_fp16")];
            fp16 var_2208_to_fp16 = const()[name = string("op_2208_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2209_cast_fp16 = mul(x = var_2207_cast_fp16, y = var_2208_to_fp16)[name = string("op_2209_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2211_cast_fp16 = add(x = var_2205_cast_fp16, y = var_2209_cast_fp16)[name = string("op_2211_cast_fp16")];
            fp32 var_2212_epsilon_0 = const()[name = string("op_2212_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2212_cast_fp16 = rsqrt(epsilon = var_2212_epsilon_0, x = var_2211_cast_fp16)[name = string("op_2212_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2212_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19528768)))];
            tensor<fp16, [1, 2520, 768]> input_23_cast_fp16 = mul(x = out_25_cast_fp16, y = t_encoder_layers_2_input_layernorm_weight_to_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19530368))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20120256))))[name = string("t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_15_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_2222 = const()[name = string("op_2222"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_39_cast_fp16 = reshape(shape = var_2222, x = linear_15_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2224_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2224_cast_fp16")];
            bool var_2227 = const()[name = string("op_2227"), val = bool(true)];
            tensor<int32, [1]> var_2228_axes_0 = const()[name = string("op_2228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2228_cast_fp16 = reduce_max(axes = var_2228_axes_0, keep_dims = var_2227, x = var_2224_cast_fp16)[name = string("op_2228_cast_fp16")];
            fp16 var_2229_to_fp16 = const()[name = string("op_2229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_35_cast_fp16 = maximum(x = var_2228_cast_fp16, y = var_2229_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_31_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_35_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2232_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2232_cast_fp16")];
            tensor<int32, [1]> var_2237_axes_0 = const()[name = string("op_2237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2237_keep_dims_0 = const()[name = string("op_2237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2237_cast_fp16 = reduce_mean(axes = var_2237_axes_0, keep_dims = var_2237_keep_dims_0, x = var_2232_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2238_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2238_cast_fp16")];
            fp32 var_2239_epsilon_0 = const()[name = string("op_2239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2239_cast_fp16 = inverse(epsilon = var_2239_epsilon_0, x = var_2238_cast_fp16)[name = string("op_2239_cast_fp16")];
            fp16 var_2240_to_fp16 = const()[name = string("op_2240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2241_cast_fp16 = mul(x = var_2239_cast_fp16, y = var_2240_to_fp16)[name = string("op_2241_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2243_cast_fp16 = add(x = var_2237_cast_fp16, y = var_2241_cast_fp16)[name = string("op_2243_cast_fp16")];
            fp32 var_2244_epsilon_0 = const()[name = string("op_2244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2244_cast_fp16 = rsqrt(epsilon = var_2244_epsilon_0, x = var_2243_cast_fp16)[name = string("op_2244_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2244_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20121856)))];
            tensor<fp16, [1, 2520, 12, 64]> x_45_cast_fp16 = mul(x = out_27_cast_fp16, y = t_encoder_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20122048))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20711936))))[name = string("t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_2254 = const()[name = string("op_2254"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_41_cast_fp16 = reshape(shape = var_2254, x = linear_16_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2256_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2256_cast_fp16")];
            bool var_2259 = const()[name = string("op_2259"), val = bool(true)];
            tensor<int32, [1]> var_2260_axes_0 = const()[name = string("op_2260_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2260_cast_fp16 = reduce_max(axes = var_2260_axes_0, keep_dims = var_2259, x = var_2256_cast_fp16)[name = string("op_2260_cast_fp16")];
            fp16 var_2261_to_fp16 = const()[name = string("op_2261_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_37_cast_fp16 = maximum(x = var_2260_cast_fp16, y = var_2261_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_33_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_37_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2264_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2264_cast_fp16")];
            tensor<int32, [1]> var_2269_axes_0 = const()[name = string("op_2269_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2269_keep_dims_0 = const()[name = string("op_2269_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2269_cast_fp16 = reduce_mean(axes = var_2269_axes_0, keep_dims = var_2269_keep_dims_0, x = var_2264_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2270_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2270_cast_fp16")];
            fp32 var_2271_epsilon_0 = const()[name = string("op_2271_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2271_cast_fp16 = inverse(epsilon = var_2271_epsilon_0, x = var_2270_cast_fp16)[name = string("op_2271_cast_fp16")];
            fp16 var_2272_to_fp16 = const()[name = string("op_2272_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2273_cast_fp16 = mul(x = var_2271_cast_fp16, y = var_2272_to_fp16)[name = string("op_2273_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2275_cast_fp16 = add(x = var_2269_cast_fp16, y = var_2273_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2276_cast_fp16 = rsqrt(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2276_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713536)))];
            tensor<fp16, [1, 2520, 12, 64]> x_47_cast_fp16 = mul(x = out_29_cast_fp16, y = t_encoder_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713728))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21303616))))[name = string("t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_17_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<int32, [4]> var_2286 = const()[name = string("op_2286"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_43_cast_fp16 = reshape(shape = var_2286, x = linear_17_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2288_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2288_cast_fp16")];
            bool var_2291 = const()[name = string("op_2291"), val = bool(true)];
            tensor<int32, [1]> var_2292_axes_0 = const()[name = string("op_2292_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2292_cast_fp16 = reduce_max(axes = var_2292_axes_0, keep_dims = var_2291, x = var_2288_cast_fp16)[name = string("op_2292_cast_fp16")];
            fp16 var_2293_to_fp16 = const()[name = string("op_2293_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_39_cast_fp16 = maximum(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_35_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_39_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2296_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2296_cast_fp16")];
            tensor<int32, [1]> var_2301_axes_0 = const()[name = string("op_2301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2301_keep_dims_0 = const()[name = string("op_2301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2301_cast_fp16 = reduce_mean(axes = var_2301_axes_0, keep_dims = var_2301_keep_dims_0, x = var_2296_cast_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2302_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2302_cast_fp16")];
            fp32 var_2303_epsilon_0 = const()[name = string("op_2303_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2303_cast_fp16 = inverse(epsilon = var_2303_epsilon_0, x = var_2302_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp16 var_2304_to_fp16 = const()[name = string("op_2304_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2305_cast_fp16 = mul(x = var_2303_cast_fp16, y = var_2304_to_fp16)[name = string("op_2305_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2307_cast_fp16 = add(x = var_2301_cast_fp16, y = var_2305_cast_fp16)[name = string("op_2307_cast_fp16")];
            fp32 var_2308_epsilon_0 = const()[name = string("op_2308_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2308_cast_fp16 = rsqrt(epsilon = var_2308_epsilon_0, x = var_2307_cast_fp16)[name = string("op_2308_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_9_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2308_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<int32, [4]> z_17_begin_0 = const()[name = string("z_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_17_end_0 = const()[name = string("z_17_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_17_end_mask_0 = const()[name = string("z_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_17_cast_fp16 = slice_by_index(begin = z_17_begin_0, end = z_17_end_0, end_mask = z_17_end_mask_0, x = x_45_cast_fp16)[name = string("z_17_cast_fp16")];
            tensor<int32, [4]> var_2319_begin_0 = const()[name = string("op_2319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2319_end_0 = const()[name = string("op_2319_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2319_end_mask_0 = const()[name = string("op_2319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2319_cast_fp16 = slice_by_index(begin = var_2319_begin_0, end = var_2319_end_0, end_mask = var_2319_end_mask_0, x = z_17_cast_fp16)[name = string("op_2319_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2320_cast_fp16 = mul(x = var_2319_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_2320_cast_fp16")];
            tensor<int32, [4]> var_2325_begin_0 = const()[name = string("op_2325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2325_end_0 = const()[name = string("op_2325_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2325_end_mask_0 = const()[name = string("op_2325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2325_cast_fp16 = slice_by_index(begin = var_2325_begin_0, end = var_2325_end_0, end_mask = var_2325_end_mask_0, x = z_17_cast_fp16)[name = string("op_2325_cast_fp16")];
            int32 var_2327 = const()[name = string("op_2327"), val = int32(-1)];
            bool rot_17_interleave_0 = const()[name = string("rot_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_17_cast_fp16 = concat(axis = var_2327, interleave = rot_17_interleave_0, values = (var_2320_cast_fp16, var_2325_cast_fp16))[name = string("rot_17_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2329_cast_fp16 = mul(x = z_17_cast_fp16, y = c_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2330_cast_fp16 = mul(x = rot_17_cast_fp16, y = s_9_cast_fp16)[name = string("op_2330_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2332_cast_fp16 = add(x = var_2329_cast_fp16, y = var_2330_cast_fp16)[name = string("op_2332_cast_fp16")];
            tensor<int32, [4]> z_19_begin_0 = const()[name = string("z_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_19_end_0 = const()[name = string("z_19_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_19_end_mask_0 = const()[name = string("z_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_19_cast_fp16 = slice_by_index(begin = z_19_begin_0, end = z_19_end_0, end_mask = z_19_end_mask_0, x = x_45_cast_fp16)[name = string("z_19_cast_fp16")];
            tensor<int32, [4]> var_2342_begin_0 = const()[name = string("op_2342_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2342_end_0 = const()[name = string("op_2342_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2342_end_mask_0 = const()[name = string("op_2342_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2342_cast_fp16 = slice_by_index(begin = var_2342_begin_0, end = var_2342_end_0, end_mask = var_2342_end_mask_0, x = z_19_cast_fp16)[name = string("op_2342_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2343_cast_fp16 = mul(x = var_2342_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2343_cast_fp16")];
            tensor<int32, [4]> var_2348_begin_0 = const()[name = string("op_2348_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2348_end_0 = const()[name = string("op_2348_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2348_end_mask_0 = const()[name = string("op_2348_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2348_cast_fp16 = slice_by_index(begin = var_2348_begin_0, end = var_2348_end_0, end_mask = var_2348_end_mask_0, x = z_19_cast_fp16)[name = string("op_2348_cast_fp16")];
            int32 var_2350 = const()[name = string("op_2350"), val = int32(-1)];
            bool rot_19_interleave_0 = const()[name = string("rot_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_19_cast_fp16 = concat(axis = var_2350, interleave = rot_19_interleave_0, values = (var_2343_cast_fp16, var_2348_cast_fp16))[name = string("rot_19_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2352_cast_fp16 = mul(x = z_19_cast_fp16, y = c_cast_fp16)[name = string("op_2352_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2353_cast_fp16 = mul(x = rot_19_cast_fp16, y = s_11_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2355_cast_fp16 = add(x = var_2352_cast_fp16, y = var_2353_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2357 = const()[name = string("op_2357"), val = int32(-1)];
            bool var_2358_interleave_0 = const()[name = string("op_2358_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_2358_cast_fp16 = concat(axis = var_2357, interleave = var_2358_interleave_0, values = (var_2332_cast_fp16, var_2355_cast_fp16))[name = string("op_2358_cast_fp16")];
            tensor<int32, [4]> z_21_begin_0 = const()[name = string("z_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_21_end_0 = const()[name = string("z_21_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_21_end_mask_0 = const()[name = string("z_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_21_cast_fp16 = slice_by_index(begin = z_21_begin_0, end = z_21_end_0, end_mask = z_21_end_mask_0, x = x_47_cast_fp16)[name = string("z_21_cast_fp16")];
            tensor<int32, [4]> var_2371_begin_0 = const()[name = string("op_2371_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2371_end_0 = const()[name = string("op_2371_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2371_end_mask_0 = const()[name = string("op_2371_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2371_cast_fp16 = slice_by_index(begin = var_2371_begin_0, end = var_2371_end_0, end_mask = var_2371_end_mask_0, x = z_21_cast_fp16)[name = string("op_2371_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2372_cast_fp16 = mul(x = var_2371_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2372_cast_fp16")];
            tensor<int32, [4]> var_2377_begin_0 = const()[name = string("op_2377_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2377_end_0 = const()[name = string("op_2377_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2377_end_mask_0 = const()[name = string("op_2377_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2377_cast_fp16 = slice_by_index(begin = var_2377_begin_0, end = var_2377_end_0, end_mask = var_2377_end_mask_0, x = z_21_cast_fp16)[name = string("op_2377_cast_fp16")];
            int32 var_2379 = const()[name = string("op_2379"), val = int32(-1)];
            bool rot_21_interleave_0 = const()[name = string("rot_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_21_cast_fp16 = concat(axis = var_2379, interleave = rot_21_interleave_0, values = (var_2372_cast_fp16, var_2377_cast_fp16))[name = string("rot_21_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2381_cast_fp16 = mul(x = z_21_cast_fp16, y = c_1_cast_fp16)[name = string("op_2381_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2382_cast_fp16 = mul(x = rot_21_cast_fp16, y = s_9_cast_fp16)[name = string("op_2382_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2384_cast_fp16 = add(x = var_2381_cast_fp16, y = var_2382_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<int32, [4]> z_23_begin_0 = const()[name = string("z_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_23_end_0 = const()[name = string("z_23_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_23_end_mask_0 = const()[name = string("z_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_23_cast_fp16 = slice_by_index(begin = z_23_begin_0, end = z_23_end_0, end_mask = z_23_end_mask_0, x = x_47_cast_fp16)[name = string("z_23_cast_fp16")];
            tensor<int32, [4]> var_2394_begin_0 = const()[name = string("op_2394_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2394_end_0 = const()[name = string("op_2394_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2394_end_mask_0 = const()[name = string("op_2394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> 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 = z_23_cast_fp16)[name = string("op_2394_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2395_cast_fp16 = mul(x = var_2394_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2395_cast_fp16")];
            tensor<int32, [4]> var_2400_begin_0 = const()[name = string("op_2400_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2400_end_0 = const()[name = string("op_2400_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2400_end_mask_0 = const()[name = string("op_2400_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2400_cast_fp16 = slice_by_index(begin = var_2400_begin_0, end = var_2400_end_0, end_mask = var_2400_end_mask_0, x = z_23_cast_fp16)[name = string("op_2400_cast_fp16")];
            int32 var_2402 = const()[name = string("op_2402"), val = int32(-1)];
            bool rot_23_interleave_0 = const()[name = string("rot_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_23_cast_fp16 = concat(axis = var_2402, interleave = rot_23_interleave_0, values = (var_2395_cast_fp16, var_2400_cast_fp16))[name = string("rot_23_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2404_cast_fp16 = mul(x = z_23_cast_fp16, y = c_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2405_cast_fp16 = mul(x = rot_23_cast_fp16, y = s_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2407_cast_fp16 = add(x = var_2404_cast_fp16, y = var_2405_cast_fp16)[name = string("op_2407_cast_fp16")];
            int32 var_2409 = const()[name = string("op_2409"), val = int32(-1)];
            bool var_2410_interleave_0 = const()[name = string("op_2410_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_2410_cast_fp16 = concat(axis = var_2409, interleave = var_2410_interleave_0, values = (var_2384_cast_fp16, var_2407_cast_fp16))[name = string("op_2410_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2420_transpose_x_0 = const()[name = string("op_2420_transpose_x_0"), val = bool(false)];
            bool var_2420_transpose_y_0 = const()[name = string("op_2420_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2410_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_68 = transpose(perm = transpose_68_perm_0, x = var_2358_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 12, 2520, 2520]> var_2420_cast_fp16 = matmul(transpose_x = var_2420_transpose_x_0, transpose_y = var_2420_transpose_y_0, x = transpose_68, y = transpose_69)[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_2422_cast_fp16 = add(x = var_2420_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2422_cast_fp16")];
            int32 var_2423 = const()[name = string("op_2423"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_35_cast_fp16 = softmax(axis = var_2423, x = var_2422_cast_fp16)[name = string("w_35_cast_fp16")];
            bool var_2426_transpose_x_0 = const()[name = string("op_2426_transpose_x_0"), val = bool(false)];
            bool var_2426_transpose_y_0 = const()[name = string("op_2426_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = v_9_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 12, 2520, 64]> var_2426_cast_fp16 = matmul(transpose_x = var_2426_transpose_x_0, transpose_y = var_2426_transpose_y_0, x = w_35_cast_fp16, y = v_11_cast_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [4]> var_2429_perm_0 = const()[name = string("op_2429_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2433 = const()[name = string("op_2433"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_2429_cast_fp16 = transpose(perm = var_2429_perm_0, x = var_2426_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 2520, 768]> input_25_cast_fp16 = reshape(shape = var_2433, x = var_2429_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21305216))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21895104))))[name = string("t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2438_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2438_cast_fp16")];
            bool var_2441 = const()[name = string("op_2441"), val = bool(true)];
            tensor<int32, [1]> var_2442_axes_0 = const()[name = string("op_2442_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2442_cast_fp16 = reduce_max(axes = var_2442_axes_0, keep_dims = var_2441, x = var_2438_cast_fp16)[name = string("op_2442_cast_fp16")];
            fp16 var_2443_to_fp16 = const()[name = string("op_2443_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_41_cast_fp16 = maximum(x = var_2442_cast_fp16, y = var_2443_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_37_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_41_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2446_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2446_cast_fp16")];
            tensor<int32, [1]> var_2451_axes_0 = const()[name = string("op_2451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2451_keep_dims_0 = const()[name = string("op_2451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2451_cast_fp16 = reduce_mean(axes = var_2451_axes_0, keep_dims = var_2451_keep_dims_0, x = var_2446_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2452_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2452_cast_fp16")];
            fp32 var_2453_epsilon_0 = const()[name = string("op_2453_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2453_cast_fp16 = inverse(epsilon = var_2453_epsilon_0, x = var_2452_cast_fp16)[name = string("op_2453_cast_fp16")];
            fp16 var_2454_to_fp16 = const()[name = string("op_2454_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = string("op_2455_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2457_cast_fp16 = add(x = var_2451_cast_fp16, y = var_2455_cast_fp16)[name = string("op_2457_cast_fp16")];
            fp32 var_2458_epsilon_0 = const()[name = string("op_2458_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2458_cast_fp16 = rsqrt(epsilon = var_2458_epsilon_0, x = var_2457_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_31_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2458_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21896704)))];
            tensor<fp16, [1, 2520, 768]> var_2460_cast_fp16 = mul(x = out_31_cast_fp16, y = t_encoder_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_2460_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_51_cast_fp16 = add(x = x_37_cast_fp16, y = var_2460_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2463_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2463_cast_fp16")];
            bool var_2466 = const()[name = string("op_2466"), val = bool(true)];
            tensor<int32, [1]> var_2467_axes_0 = const()[name = string("op_2467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2467_cast_fp16 = reduce_max(axes = var_2467_axes_0, keep_dims = var_2466, x = var_2463_cast_fp16)[name = string("op_2467_cast_fp16")];
            fp16 var_2468_to_fp16 = const()[name = string("op_2468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_43_cast_fp16 = maximum(x = var_2467_cast_fp16, y = var_2468_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_39_cast_fp16 = real_div(x = x_51_cast_fp16, y = s_43_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2471_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2471_cast_fp16")];
            tensor<int32, [1]> var_2476_axes_0 = const()[name = string("op_2476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2476_keep_dims_0 = const()[name = string("op_2476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2476_cast_fp16 = reduce_mean(axes = var_2476_axes_0, keep_dims = var_2476_keep_dims_0, x = var_2471_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2477_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_2477_cast_fp16")];
            fp32 var_2478_epsilon_0 = const()[name = string("op_2478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2478_cast_fp16 = inverse(epsilon = var_2478_epsilon_0, x = var_2477_cast_fp16)[name = string("op_2478_cast_fp16")];
            fp16 var_2479_to_fp16 = const()[name = string("op_2479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2480_cast_fp16 = mul(x = var_2478_cast_fp16, y = var_2479_to_fp16)[name = string("op_2480_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2482_cast_fp16 = add(x = var_2476_cast_fp16, y = var_2480_cast_fp16)[name = string("op_2482_cast_fp16")];
            fp32 var_2483_epsilon_0 = const()[name = string("op_2483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2483_cast_fp16 = rsqrt(epsilon = var_2483_epsilon_0, x = var_2482_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_33_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2483_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21898304)))];
            tensor<fp16, [1, 2520, 768]> input_27_cast_fp16 = mul(x = out_33_cast_fp16, y = t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21899904))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24259264))))[name = string("t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_19_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_19_cast_fp16")];
            string var_2490_mode_0 = const()[name = string("op_2490_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_2490_cast_fp16 = gelu(mode = var_2490_mode_0, x = linear_19_cast_fp16)[name = string("op_2490_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24265472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26624832))))[name = string("t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_20_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_31_cast_fp16 = mul(x = var_2490_cast_fp16, y = linear_20_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26631040))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28990400))))[name = string("t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2498_cast_fp16 = abs(x = linear_21_cast_fp16)[name = string("op_2498_cast_fp16")];
            bool var_2501 = const()[name = string("op_2501"), val = bool(true)];
            tensor<int32, [1]> var_2502_axes_0 = const()[name = string("op_2502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2502_cast_fp16 = reduce_max(axes = var_2502_axes_0, keep_dims = var_2501, x = var_2498_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp16 var_2503_to_fp16 = const()[name = string("op_2503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_45_cast_fp16 = maximum(x = var_2502_cast_fp16, y = var_2503_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_41_cast_fp16 = real_div(x = linear_21_cast_fp16, y = s_45_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2506_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2506_cast_fp16")];
            tensor<int32, [1]> var_2511_axes_0 = const()[name = string("op_2511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2511_keep_dims_0 = const()[name = string("op_2511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2511_cast_fp16 = reduce_mean(axes = var_2511_axes_0, keep_dims = var_2511_keep_dims_0, x = var_2506_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2512_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_2512_cast_fp16")];
            fp32 var_2513_epsilon_0 = const()[name = string("op_2513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2513_cast_fp16 = inverse(epsilon = var_2513_epsilon_0, x = var_2512_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2515_cast_fp16 = mul(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("op_2515_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2517_cast_fp16 = add(x = var_2511_cast_fp16, y = var_2515_cast_fp16)[name = string("op_2517_cast_fp16")];
            fp32 var_2518_epsilon_0 = const()[name = string("op_2518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2518_cast_fp16 = rsqrt(epsilon = var_2518_epsilon_0, x = var_2517_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_35_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2518_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28992000)))];
            tensor<fp16, [1, 2520, 768]> var_2520_cast_fp16 = mul(x = out_35_cast_fp16, y = t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_55_cast_fp16 = add(x = x_51_cast_fp16, y = var_2520_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2523_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2523_cast_fp16")];
            bool var_2526 = const()[name = string("op_2526"), val = bool(true)];
            tensor<int32, [1]> var_2527_axes_0 = const()[name = string("op_2527_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2527_cast_fp16 = reduce_max(axes = var_2527_axes_0, keep_dims = var_2526, x = var_2523_cast_fp16)[name = string("op_2527_cast_fp16")];
            fp16 var_2528_to_fp16 = const()[name = string("op_2528_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_47_cast_fp16 = maximum(x = var_2527_cast_fp16, y = var_2528_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_43_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_47_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2531_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [1]> var_2536_axes_0 = const()[name = string("op_2536_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2536_keep_dims_0 = const()[name = string("op_2536_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2536_cast_fp16 = reduce_mean(axes = var_2536_axes_0, keep_dims = var_2536_keep_dims_0, x = var_2531_cast_fp16)[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2537_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_2537_cast_fp16")];
            fp32 var_2538_epsilon_0 = const()[name = string("op_2538_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2538_cast_fp16 = inverse(epsilon = var_2538_epsilon_0, x = var_2537_cast_fp16)[name = string("op_2538_cast_fp16")];
            fp16 var_2539_to_fp16 = const()[name = string("op_2539_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2540_cast_fp16 = mul(x = var_2538_cast_fp16, y = var_2539_to_fp16)[name = string("op_2540_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2542_cast_fp16 = add(x = var_2536_cast_fp16, y = var_2540_cast_fp16)[name = string("op_2542_cast_fp16")];
            fp32 var_2543_epsilon_0 = const()[name = string("op_2543_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2543_cast_fp16 = rsqrt(epsilon = var_2543_epsilon_0, x = var_2542_cast_fp16)[name = string("op_2543_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_2543_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28993600)))];
            tensor<fp16, [1, 2520, 768]> input_33_cast_fp16 = mul(x = out_37_cast_fp16, y = t_encoder_layers_3_input_layernorm_weight_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28995200))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29585088))))[name = string("t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_2553 = const()[name = string("op_2553"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_57_cast_fp16 = reshape(shape = var_2553, x = linear_22_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2555_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2555_cast_fp16")];
            bool var_2558 = const()[name = string("op_2558"), val = bool(true)];
            tensor<int32, [1]> var_2559_axes_0 = const()[name = string("op_2559_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2559_cast_fp16 = reduce_max(axes = var_2559_axes_0, keep_dims = var_2558, x = var_2555_cast_fp16)[name = string("op_2559_cast_fp16")];
            fp16 var_2560_to_fp16 = const()[name = string("op_2560_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_49_cast_fp16 = maximum(x = var_2559_cast_fp16, y = var_2560_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_45_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_49_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2563_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_2563_cast_fp16")];
            tensor<int32, [1]> var_2568_axes_0 = const()[name = string("op_2568_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2568_keep_dims_0 = const()[name = string("op_2568_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2568_cast_fp16 = reduce_mean(axes = var_2568_axes_0, keep_dims = var_2568_keep_dims_0, x = var_2563_cast_fp16)[name = string("op_2568_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2569_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_2569_cast_fp16")];
            fp32 var_2570_epsilon_0 = const()[name = string("op_2570_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2570_cast_fp16 = inverse(epsilon = var_2570_epsilon_0, x = var_2569_cast_fp16)[name = string("op_2570_cast_fp16")];
            fp16 var_2571_to_fp16 = const()[name = string("op_2571_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2572_cast_fp16 = mul(x = var_2570_cast_fp16, y = var_2571_to_fp16)[name = string("op_2572_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2574_cast_fp16 = add(x = var_2568_cast_fp16, y = var_2572_cast_fp16)[name = string("op_2574_cast_fp16")];
            fp32 var_2575_epsilon_0 = const()[name = string("op_2575_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2575_cast_fp16 = rsqrt(epsilon = var_2575_epsilon_0, x = var_2574_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_2575_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29586688)))];
            tensor<fp16, [1, 2520, 12, 64]> x_63_cast_fp16 = mul(x = out_39_cast_fp16, y = t_encoder_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29586880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30176768))))[name = string("t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_23_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_2585 = const()[name = string("op_2585"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_59_cast_fp16 = reshape(shape = var_2585, x = linear_23_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2587_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2587_cast_fp16")];
            bool var_2590 = const()[name = string("op_2590"), val = bool(true)];
            tensor<int32, [1]> var_2591_axes_0 = const()[name = string("op_2591_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2591_cast_fp16 = reduce_max(axes = var_2591_axes_0, keep_dims = var_2590, x = var_2587_cast_fp16)[name = string("op_2591_cast_fp16")];
            fp16 var_2592_to_fp16 = const()[name = string("op_2592_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_51_cast_fp16 = maximum(x = var_2591_cast_fp16, y = var_2592_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_47_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_51_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2595_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_2595_cast_fp16")];
            tensor<int32, [1]> var_2600_axes_0 = const()[name = string("op_2600_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2600_keep_dims_0 = const()[name = string("op_2600_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2600_cast_fp16 = reduce_mean(axes = var_2600_axes_0, keep_dims = var_2600_keep_dims_0, x = var_2595_cast_fp16)[name = string("op_2600_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2601_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_2601_cast_fp16")];
            fp32 var_2602_epsilon_0 = const()[name = string("op_2602_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2602_cast_fp16 = inverse(epsilon = var_2602_epsilon_0, x = var_2601_cast_fp16)[name = string("op_2602_cast_fp16")];
            fp16 var_2603_to_fp16 = const()[name = string("op_2603_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2604_cast_fp16 = mul(x = var_2602_cast_fp16, y = var_2603_to_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2606_cast_fp16 = add(x = var_2600_cast_fp16, y = var_2604_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp32 var_2607_epsilon_0 = const()[name = string("op_2607_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2607_cast_fp16 = rsqrt(epsilon = var_2607_epsilon_0, x = var_2606_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_2607_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178368)))];
            tensor<fp16, [1, 2520, 12, 64]> x_65_cast_fp16 = mul(x = out_41_cast_fp16, y = t_encoder_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178560))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30768448))))[name = string("t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [4]> var_2617 = const()[name = string("op_2617"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_61_cast_fp16 = reshape(shape = var_2617, x = linear_24_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2619_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2619_cast_fp16")];
            bool var_2622 = const()[name = string("op_2622"), val = bool(true)];
            tensor<int32, [1]> var_2623_axes_0 = const()[name = string("op_2623_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2623_cast_fp16 = reduce_max(axes = var_2623_axes_0, keep_dims = var_2622, x = var_2619_cast_fp16)[name = string("op_2623_cast_fp16")];
            fp16 var_2624_to_fp16 = const()[name = string("op_2624_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_53_cast_fp16 = maximum(x = var_2623_cast_fp16, y = var_2624_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_49_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_53_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2627_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<int32, [1]> var_2632_axes_0 = const()[name = string("op_2632_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2632_keep_dims_0 = const()[name = string("op_2632_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2632_cast_fp16 = reduce_mean(axes = var_2632_axes_0, keep_dims = var_2632_keep_dims_0, x = var_2627_cast_fp16)[name = string("op_2632_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2633_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2634_cast_fp16 = inverse(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            fp16 var_2635_to_fp16 = const()[name = string("op_2635_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2636_cast_fp16 = mul(x = var_2634_cast_fp16, y = var_2635_to_fp16)[name = string("op_2636_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2638_cast_fp16 = add(x = var_2632_cast_fp16, y = var_2636_cast_fp16)[name = string("op_2638_cast_fp16")];
            fp32 var_2639_epsilon_0 = const()[name = string("op_2639_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2639_cast_fp16 = rsqrt(epsilon = var_2639_epsilon_0, x = var_2638_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_13_cast_fp16 = mul(x = y_49_cast_fp16, y = var_2639_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<int32, [4]> z_25_begin_0 = const()[name = string("z_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_25_end_0 = const()[name = string("z_25_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_25_end_mask_0 = const()[name = string("z_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_25_cast_fp16 = slice_by_index(begin = z_25_begin_0, end = z_25_end_0, end_mask = z_25_end_mask_0, x = x_63_cast_fp16)[name = string("z_25_cast_fp16")];
            tensor<int32, [4]> var_2650_begin_0 = const()[name = string("op_2650_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2650_end_0 = const()[name = string("op_2650_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2650_end_mask_0 = const()[name = string("op_2650_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2650_cast_fp16 = slice_by_index(begin = var_2650_begin_0, end = var_2650_end_0, end_mask = var_2650_end_mask_0, x = z_25_cast_fp16)[name = string("op_2650_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2651_cast_fp16")];
            tensor<int32, [4]> var_2656_begin_0 = const()[name = string("op_2656_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2656_end_0 = const()[name = string("op_2656_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2656_end_mask_0 = const()[name = string("op_2656_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2656_cast_fp16 = slice_by_index(begin = var_2656_begin_0, end = var_2656_end_0, end_mask = var_2656_end_mask_0, x = z_25_cast_fp16)[name = string("op_2656_cast_fp16")];
            int32 var_2658 = const()[name = string("op_2658"), val = int32(-1)];
            bool rot_25_interleave_0 = const()[name = string("rot_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_25_cast_fp16 = concat(axis = var_2658, interleave = rot_25_interleave_0, values = (var_2651_cast_fp16, var_2656_cast_fp16))[name = string("rot_25_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2660_cast_fp16 = mul(x = z_25_cast_fp16, y = c_1_cast_fp16)[name = string("op_2660_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2661_cast_fp16 = mul(x = rot_25_cast_fp16, y = s_9_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2663_cast_fp16 = add(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = string("op_2663_cast_fp16")];
            tensor<int32, [4]> z_27_begin_0 = const()[name = string("z_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_27_end_0 = const()[name = string("z_27_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_27_end_mask_0 = const()[name = string("z_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_27_cast_fp16 = slice_by_index(begin = z_27_begin_0, end = z_27_end_0, end_mask = z_27_end_mask_0, x = x_63_cast_fp16)[name = string("z_27_cast_fp16")];
            tensor<int32, [4]> var_2673_begin_0 = const()[name = string("op_2673_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2673_end_0 = const()[name = string("op_2673_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2673_end_mask_0 = const()[name = string("op_2673_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2673_cast_fp16 = slice_by_index(begin = var_2673_begin_0, end = var_2673_end_0, end_mask = var_2673_end_mask_0, x = z_27_cast_fp16)[name = string("op_2673_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2674_cast_fp16 = mul(x = var_2673_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2674_cast_fp16")];
            tensor<int32, [4]> var_2679_begin_0 = const()[name = string("op_2679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2679_end_0 = const()[name = string("op_2679_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2679_end_mask_0 = const()[name = string("op_2679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2679_cast_fp16 = slice_by_index(begin = var_2679_begin_0, end = var_2679_end_0, end_mask = var_2679_end_mask_0, x = z_27_cast_fp16)[name = string("op_2679_cast_fp16")];
            int32 var_2681 = const()[name = string("op_2681"), val = int32(-1)];
            bool rot_27_interleave_0 = const()[name = string("rot_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_27_cast_fp16 = concat(axis = var_2681, interleave = rot_27_interleave_0, values = (var_2674_cast_fp16, var_2679_cast_fp16))[name = string("rot_27_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2683_cast_fp16 = mul(x = z_27_cast_fp16, y = c_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2684_cast_fp16 = mul(x = rot_27_cast_fp16, y = s_11_cast_fp16)[name = string("op_2684_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2686_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2684_cast_fp16)[name = string("op_2686_cast_fp16")];
            int32 var_2688 = const()[name = string("op_2688"), val = int32(-1)];
            bool var_2689_interleave_0 = const()[name = string("op_2689_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_2689_cast_fp16 = concat(axis = var_2688, interleave = var_2689_interleave_0, values = (var_2663_cast_fp16, var_2686_cast_fp16))[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> z_29_begin_0 = const()[name = string("z_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_29_end_0 = const()[name = string("z_29_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_29_end_mask_0 = const()[name = string("z_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_29_cast_fp16 = slice_by_index(begin = z_29_begin_0, end = z_29_end_0, end_mask = z_29_end_mask_0, x = x_65_cast_fp16)[name = string("z_29_cast_fp16")];
            tensor<int32, [4]> var_2702_begin_0 = const()[name = string("op_2702_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2702_end_0 = const()[name = string("op_2702_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2702_end_mask_0 = const()[name = string("op_2702_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2702_cast_fp16 = slice_by_index(begin = var_2702_begin_0, end = var_2702_end_0, end_mask = var_2702_end_mask_0, x = z_29_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2703_cast_fp16 = mul(x = var_2702_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_2703_cast_fp16")];
            tensor<int32, [4]> var_2708_begin_0 = const()[name = string("op_2708_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2708_end_0 = const()[name = string("op_2708_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2708_end_mask_0 = const()[name = string("op_2708_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2708_cast_fp16 = slice_by_index(begin = var_2708_begin_0, end = var_2708_end_0, end_mask = var_2708_end_mask_0, x = z_29_cast_fp16)[name = string("op_2708_cast_fp16")];
            int32 var_2710 = const()[name = string("op_2710"), val = int32(-1)];
            bool rot_29_interleave_0 = const()[name = string("rot_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_29_cast_fp16 = concat(axis = var_2710, interleave = rot_29_interleave_0, values = (var_2703_cast_fp16, var_2708_cast_fp16))[name = string("rot_29_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2712_cast_fp16 = mul(x = z_29_cast_fp16, y = c_1_cast_fp16)[name = string("op_2712_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2713_cast_fp16 = mul(x = rot_29_cast_fp16, y = s_9_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2715_cast_fp16 = add(x = var_2712_cast_fp16, y = var_2713_cast_fp16)[name = string("op_2715_cast_fp16")];
            tensor<int32, [4]> z_31_begin_0 = const()[name = string("z_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_31_end_0 = const()[name = string("z_31_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_31_end_mask_0 = const()[name = string("z_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_31_cast_fp16 = slice_by_index(begin = z_31_begin_0, end = z_31_end_0, end_mask = z_31_end_mask_0, x = x_65_cast_fp16)[name = string("z_31_cast_fp16")];
            tensor<int32, [4]> var_2725_begin_0 = const()[name = string("op_2725_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2725_end_0 = const()[name = string("op_2725_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2725_end_mask_0 = const()[name = string("op_2725_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2725_cast_fp16 = slice_by_index(begin = var_2725_begin_0, end = var_2725_end_0, end_mask = var_2725_end_mask_0, x = z_31_cast_fp16)[name = string("op_2725_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2726_cast_fp16 = mul(x = var_2725_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> var_2731_begin_0 = const()[name = string("op_2731_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2731_end_0 = const()[name = string("op_2731_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2731_end_mask_0 = const()[name = string("op_2731_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2731_cast_fp16 = slice_by_index(begin = var_2731_begin_0, end = var_2731_end_0, end_mask = var_2731_end_mask_0, x = z_31_cast_fp16)[name = string("op_2731_cast_fp16")];
            int32 var_2733 = const()[name = string("op_2733"), val = int32(-1)];
            bool rot_31_interleave_0 = const()[name = string("rot_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_31_cast_fp16 = concat(axis = var_2733, interleave = rot_31_interleave_0, values = (var_2726_cast_fp16, var_2731_cast_fp16))[name = string("rot_31_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2735_cast_fp16 = mul(x = z_31_cast_fp16, y = c_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2736_cast_fp16 = mul(x = rot_31_cast_fp16, y = s_11_cast_fp16)[name = string("op_2736_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2738_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2736_cast_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-1)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2715_cast_fp16, var_2738_cast_fp16))[name = string("op_2741_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2751_transpose_x_0 = const()[name = string("op_2751_transpose_x_0"), val = bool(false)];
            bool var_2751_transpose_y_0 = const()[name = string("op_2751_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2741_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_70 = transpose(perm = transpose_70_perm_0, x = var_2689_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 12, 2520, 2520]> var_2751_cast_fp16 = matmul(transpose_x = var_2751_transpose_x_0, transpose_y = var_2751_transpose_y_0, x = transpose_70, y = transpose_71)[name = string("op_2751_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_2753_cast_fp16 = add(x = var_2751_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2753_cast_fp16")];
            int32 var_2754 = const()[name = string("op_2754"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_49_cast_fp16 = softmax(axis = var_2754, x = var_2753_cast_fp16)[name = string("w_49_cast_fp16")];
            bool var_2757_transpose_x_0 = const()[name = string("op_2757_transpose_x_0"), val = bool(false)];
            bool var_2757_transpose_y_0 = const()[name = string("op_2757_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = v_13_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 12, 2520, 64]> var_2757_cast_fp16 = matmul(transpose_x = var_2757_transpose_x_0, transpose_y = var_2757_transpose_y_0, x = w_49_cast_fp16, y = v_15_cast_fp16)[name = string("op_2757_cast_fp16")];
            tensor<int32, [4]> var_2760_perm_0 = const()[name = string("op_2760_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2764 = const()[name = string("op_2764"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_2760_cast_fp16 = transpose(perm = var_2760_perm_0, x = var_2757_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 2520, 768]> input_35_cast_fp16 = reshape(shape = var_2764, x = var_2760_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30770048))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31359936))))[name = string("t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_25_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2769_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2772 = const()[name = string("op_2772"), val = bool(true)];
            tensor<int32, [1]> var_2773_axes_0 = const()[name = string("op_2773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2773_cast_fp16 = reduce_max(axes = var_2773_axes_0, keep_dims = var_2772, x = var_2769_cast_fp16)[name = string("op_2773_cast_fp16")];
            fp16 var_2774_to_fp16 = const()[name = string("op_2774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_55_cast_fp16 = maximum(x = var_2773_cast_fp16, y = var_2774_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_51_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_55_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2777_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_2777_cast_fp16")];
            tensor<int32, [1]> var_2782_axes_0 = const()[name = string("op_2782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2782_keep_dims_0 = const()[name = string("op_2782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2782_cast_fp16 = reduce_mean(axes = var_2782_axes_0, keep_dims = var_2782_keep_dims_0, x = var_2777_cast_fp16)[name = string("op_2782_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2783_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_2783_cast_fp16")];
            fp32 var_2784_epsilon_0 = const()[name = string("op_2784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2784_cast_fp16 = inverse(epsilon = var_2784_epsilon_0, x = var_2783_cast_fp16)[name = string("op_2784_cast_fp16")];
            fp16 var_2785_to_fp16 = const()[name = string("op_2785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2786_cast_fp16 = mul(x = var_2784_cast_fp16, y = var_2785_to_fp16)[name = string("op_2786_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2788_cast_fp16 = add(x = var_2782_cast_fp16, y = var_2786_cast_fp16)[name = string("op_2788_cast_fp16")];
            fp32 var_2789_epsilon_0 = const()[name = string("op_2789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2789_cast_fp16 = rsqrt(epsilon = var_2789_epsilon_0, x = var_2788_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_43_cast_fp16 = mul(x = y_51_cast_fp16, y = var_2789_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31361536)))];
            tensor<fp16, [1, 2520, 768]> var_2791_cast_fp16 = mul(x = out_43_cast_fp16, y = t_encoder_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_2791_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_69_cast_fp16 = add(x = x_55_cast_fp16, y = var_2791_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2794_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2794_cast_fp16")];
            bool var_2797 = const()[name = string("op_2797"), val = bool(true)];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2798_cast_fp16 = reduce_max(axes = var_2798_axes_0, keep_dims = var_2797, x = var_2794_cast_fp16)[name = string("op_2798_cast_fp16")];
            fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_57_cast_fp16 = maximum(x = var_2798_cast_fp16, y = var_2799_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_53_cast_fp16 = real_div(x = x_69_cast_fp16, y = s_57_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2802_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_2802_cast_fp16")];
            tensor<int32, [1]> var_2807_axes_0 = const()[name = string("op_2807_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2807_keep_dims_0 = const()[name = string("op_2807_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2807_cast_fp16 = reduce_mean(axes = var_2807_axes_0, keep_dims = var_2807_keep_dims_0, x = var_2802_cast_fp16)[name = string("op_2807_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2808_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_2808_cast_fp16")];
            fp32 var_2809_epsilon_0 = const()[name = string("op_2809_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2809_cast_fp16 = inverse(epsilon = var_2809_epsilon_0, x = var_2808_cast_fp16)[name = string("op_2809_cast_fp16")];
            fp16 var_2810_to_fp16 = const()[name = string("op_2810_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2811_cast_fp16 = mul(x = var_2809_cast_fp16, y = var_2810_to_fp16)[name = string("op_2811_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2813_cast_fp16 = add(x = var_2807_cast_fp16, y = var_2811_cast_fp16)[name = string("op_2813_cast_fp16")];
            fp32 var_2814_epsilon_0 = const()[name = string("op_2814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2814_cast_fp16 = rsqrt(epsilon = var_2814_epsilon_0, x = var_2813_cast_fp16)[name = string("op_2814_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_45_cast_fp16 = mul(x = y_53_cast_fp16, y = var_2814_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31363136)))];
            tensor<fp16, [1, 2520, 768]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31364736))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33724096))))[name = string("t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_26_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_26_cast_fp16")];
            string var_2821_mode_0 = const()[name = string("op_2821_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_2821_cast_fp16 = gelu(mode = var_2821_mode_0, x = linear_26_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33730304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36089664))))[name = string("t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_27_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_41_cast_fp16 = mul(x = var_2821_cast_fp16, y = linear_27_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36095872))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38455232))))[name = string("t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2829_cast_fp16 = abs(x = linear_28_cast_fp16)[name = string("op_2829_cast_fp16")];
            bool var_2832 = const()[name = string("op_2832"), val = bool(true)];
            tensor<int32, [1]> var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2833_cast_fp16 = reduce_max(axes = var_2833_axes_0, keep_dims = var_2832, x = var_2829_cast_fp16)[name = string("op_2833_cast_fp16")];
            fp16 var_2834_to_fp16 = const()[name = string("op_2834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_59_cast_fp16 = maximum(x = var_2833_cast_fp16, y = var_2834_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_55_cast_fp16 = real_div(x = linear_28_cast_fp16, y = s_59_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2837_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_2837_cast_fp16")];
            tensor<int32, [1]> var_2842_axes_0 = const()[name = string("op_2842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2842_keep_dims_0 = const()[name = string("op_2842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2842_cast_fp16 = reduce_mean(axes = var_2842_axes_0, keep_dims = var_2842_keep_dims_0, x = var_2837_cast_fp16)[name = string("op_2842_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2843_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp32 var_2844_epsilon_0 = const()[name = string("op_2844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2844_cast_fp16 = inverse(epsilon = var_2844_epsilon_0, x = var_2843_cast_fp16)[name = string("op_2844_cast_fp16")];
            fp16 var_2845_to_fp16 = const()[name = string("op_2845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2846_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2848_cast_fp16 = add(x = var_2842_cast_fp16, y = var_2846_cast_fp16)[name = string("op_2848_cast_fp16")];
            fp32 var_2849_epsilon_0 = const()[name = string("op_2849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2849_cast_fp16 = rsqrt(epsilon = var_2849_epsilon_0, x = var_2848_cast_fp16)[name = string("op_2849_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_47_cast_fp16 = mul(x = y_55_cast_fp16, y = var_2849_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38456832)))];
            tensor<fp16, [1, 2520, 768]> var_2851_cast_fp16 = mul(x = out_47_cast_fp16, y = t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2851_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_73_cast_fp16 = add(x = x_69_cast_fp16, y = var_2851_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2854_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2854_cast_fp16")];
            bool var_2857 = const()[name = string("op_2857"), val = bool(true)];
            tensor<int32, [1]> var_2858_axes_0 = const()[name = string("op_2858_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_2858_cast_fp16 = reduce_max(axes = var_2858_axes_0, keep_dims = var_2857, x = var_2854_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp16 var_2859_to_fp16 = const()[name = string("op_2859_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_61_cast_fp16 = maximum(x = var_2858_cast_fp16, y = var_2859_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_57_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_61_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_2862_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_2862_cast_fp16")];
            tensor<int32, [1]> var_2867_axes_0 = const()[name = string("op_2867_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2867_keep_dims_0 = const()[name = string("op_2867_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_2867_cast_fp16 = reduce_mean(axes = var_2867_axes_0, keep_dims = var_2867_keep_dims_0, x = var_2862_cast_fp16)[name = string("op_2867_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2868_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_2868_cast_fp16")];
            fp32 var_2869_epsilon_0 = const()[name = string("op_2869_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_2869_cast_fp16 = inverse(epsilon = var_2869_epsilon_0, x = var_2868_cast_fp16)[name = string("op_2869_cast_fp16")];
            fp16 var_2870_to_fp16 = const()[name = string("op_2870_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_2871_cast_fp16 = mul(x = var_2869_cast_fp16, y = var_2870_to_fp16)[name = string("op_2871_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_2873_cast_fp16 = add(x = var_2867_cast_fp16, y = var_2871_cast_fp16)[name = string("op_2873_cast_fp16")];
            fp32 var_2874_epsilon_0 = const()[name = string("op_2874_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_2874_cast_fp16 = rsqrt(epsilon = var_2874_epsilon_0, x = var_2873_cast_fp16)[name = string("op_2874_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_49_cast_fp16 = mul(x = y_57_cast_fp16, y = var_2874_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38458432)))];
            tensor<fp16, [1, 2520, 768]> input_43_cast_fp16 = mul(x = out_49_cast_fp16, y = t_encoder_layers_4_input_layernorm_weight_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38460032))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39049920))))[name = string("t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_29_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2884 = const()[name = string("op_2884"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_75_cast_fp16 = reshape(shape = var_2884, x = linear_29_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2886_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_2886_cast_fp16")];
            bool var_2889 = const()[name = string("op_2889"), val = bool(true)];
            tensor<int32, [1]> var_2890_axes_0 = const()[name = string("op_2890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2890_cast_fp16 = reduce_max(axes = var_2890_axes_0, keep_dims = var_2889, x = var_2886_cast_fp16)[name = string("op_2890_cast_fp16")];
            fp16 var_2891_to_fp16 = const()[name = string("op_2891_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_63_cast_fp16 = maximum(x = var_2890_cast_fp16, y = var_2891_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_59_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_63_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2894_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [1]> var_2899_axes_0 = const()[name = string("op_2899_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2899_keep_dims_0 = const()[name = string("op_2899_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2899_cast_fp16 = reduce_mean(axes = var_2899_axes_0, keep_dims = var_2899_keep_dims_0, x = var_2894_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2900_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_2900_cast_fp16")];
            fp32 var_2901_epsilon_0 = const()[name = string("op_2901_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2901_cast_fp16 = inverse(epsilon = var_2901_epsilon_0, x = var_2900_cast_fp16)[name = string("op_2901_cast_fp16")];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2903_cast_fp16 = mul(x = var_2901_cast_fp16, y = var_2902_to_fp16)[name = string("op_2903_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2905_cast_fp16 = add(x = var_2899_cast_fp16, y = var_2903_cast_fp16)[name = string("op_2905_cast_fp16")];
            fp32 var_2906_epsilon_0 = const()[name = string("op_2906_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2906_cast_fp16 = rsqrt(epsilon = var_2906_epsilon_0, x = var_2905_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_2906_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39051520)))];
            tensor<fp16, [1, 2520, 12, 64]> x_81_cast_fp16 = mul(x = out_51_cast_fp16, y = t_encoder_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39051712))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39641600))))[name = string("t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2916 = const()[name = string("op_2916"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_77_cast_fp16 = reshape(shape = var_2916, x = linear_30_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2918_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_2918_cast_fp16")];
            bool var_2921 = const()[name = string("op_2921"), val = bool(true)];
            tensor<int32, [1]> var_2922_axes_0 = const()[name = string("op_2922_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2922_cast_fp16 = reduce_max(axes = var_2922_axes_0, keep_dims = var_2921, x = var_2918_cast_fp16)[name = string("op_2922_cast_fp16")];
            fp16 var_2923_to_fp16 = const()[name = string("op_2923_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_65_cast_fp16 = maximum(x = var_2922_cast_fp16, y = var_2923_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_61_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_65_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2926_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_2926_cast_fp16")];
            tensor<int32, [1]> var_2931_axes_0 = const()[name = string("op_2931_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2931_keep_dims_0 = const()[name = string("op_2931_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2931_cast_fp16 = reduce_mean(axes = var_2931_axes_0, keep_dims = var_2931_keep_dims_0, x = var_2926_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2932_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_2932_cast_fp16")];
            fp32 var_2933_epsilon_0 = const()[name = string("op_2933_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2933_cast_fp16 = inverse(epsilon = var_2933_epsilon_0, x = var_2932_cast_fp16)[name = string("op_2933_cast_fp16")];
            fp16 var_2934_to_fp16 = const()[name = string("op_2934_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2935_cast_fp16 = mul(x = var_2933_cast_fp16, y = var_2934_to_fp16)[name = string("op_2935_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2937_cast_fp16 = add(x = var_2931_cast_fp16, y = var_2935_cast_fp16)[name = string("op_2937_cast_fp16")];
            fp32 var_2938_epsilon_0 = const()[name = string("op_2938_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2938_cast_fp16 = rsqrt(epsilon = var_2938_epsilon_0, x = var_2937_cast_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_2938_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178368)))];
            tensor<fp16, [1, 2520, 12, 64]> x_83_cast_fp16 = mul(x = out_53_cast_fp16, y = t_encoder_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39643200))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40233088))))[name = string("t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<int32, [4]> var_2948 = const()[name = string("op_2948"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_79_cast_fp16 = reshape(shape = var_2948, x = linear_31_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2950_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_2950_cast_fp16")];
            bool var_2953 = const()[name = string("op_2953"), val = bool(true)];
            tensor<int32, [1]> var_2954_axes_0 = const()[name = string("op_2954_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_2954_cast_fp16 = reduce_max(axes = var_2954_axes_0, keep_dims = var_2953, x = var_2950_cast_fp16)[name = string("op_2954_cast_fp16")];
            fp16 var_2955_to_fp16 = const()[name = string("op_2955_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_67_cast_fp16 = maximum(x = var_2954_cast_fp16, y = var_2955_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_63_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_67_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_2958_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_2958_cast_fp16")];
            tensor<int32, [1]> var_2963_axes_0 = const()[name = string("op_2963_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2963_keep_dims_0 = const()[name = string("op_2963_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_2963_cast_fp16 = reduce_mean(axes = var_2963_axes_0, keep_dims = var_2963_keep_dims_0, x = var_2958_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2964_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_2964_cast_fp16")];
            fp32 var_2965_epsilon_0 = const()[name = string("op_2965_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_2965_cast_fp16 = inverse(epsilon = var_2965_epsilon_0, x = var_2964_cast_fp16)[name = string("op_2965_cast_fp16")];
            fp16 var_2966_to_fp16 = const()[name = string("op_2966_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_2967_cast_fp16 = mul(x = var_2965_cast_fp16, y = var_2966_to_fp16)[name = string("op_2967_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_2969_cast_fp16 = add(x = var_2963_cast_fp16, y = var_2967_cast_fp16)[name = string("op_2969_cast_fp16")];
            fp32 var_2970_epsilon_0 = const()[name = string("op_2970_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_2970_cast_fp16 = rsqrt(epsilon = var_2970_epsilon_0, x = var_2969_cast_fp16)[name = string("op_2970_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_17_cast_fp16 = mul(x = y_63_cast_fp16, y = var_2970_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<int32, [4]> z_33_begin_0 = const()[name = string("z_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_33_end_0 = const()[name = string("z_33_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_33_end_mask_0 = const()[name = string("z_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_33_cast_fp16 = slice_by_index(begin = z_33_begin_0, end = z_33_end_0, end_mask = z_33_end_mask_0, x = x_81_cast_fp16)[name = string("z_33_cast_fp16")];
            tensor<int32, [4]> var_2981_begin_0 = const()[name = string("op_2981_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2981_end_0 = const()[name = string("op_2981_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_2981_end_mask_0 = const()[name = string("op_2981_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_2981_cast_fp16 = slice_by_index(begin = var_2981_begin_0, end = var_2981_end_0, end_mask = var_2981_end_mask_0, x = z_33_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_2982_cast_fp16 = mul(x = var_2981_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_2982_cast_fp16")];
            tensor<int32, [4]> var_2987_begin_0 = const()[name = string("op_2987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2987_end_0 = const()[name = string("op_2987_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_2987_end_mask_0 = const()[name = string("op_2987_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_2987_cast_fp16 = slice_by_index(begin = var_2987_begin_0, end = var_2987_end_0, end_mask = var_2987_end_mask_0, x = z_33_cast_fp16)[name = string("op_2987_cast_fp16")];
            int32 var_2989 = const()[name = string("op_2989"), val = int32(-1)];
            bool rot_33_interleave_0 = const()[name = string("rot_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_33_cast_fp16 = concat(axis = var_2989, interleave = rot_33_interleave_0, values = (var_2982_cast_fp16, var_2987_cast_fp16))[name = string("rot_33_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2991_cast_fp16 = mul(x = z_33_cast_fp16, y = c_1_cast_fp16)[name = string("op_2991_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2992_cast_fp16 = mul(x = rot_33_cast_fp16, y = s_9_cast_fp16)[name = string("op_2992_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_2994_cast_fp16 = add(x = var_2991_cast_fp16, y = var_2992_cast_fp16)[name = string("op_2994_cast_fp16")];
            tensor<int32, [4]> z_35_begin_0 = const()[name = string("z_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_35_end_0 = const()[name = string("z_35_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_35_end_mask_0 = const()[name = string("z_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_35_cast_fp16 = slice_by_index(begin = z_35_begin_0, end = z_35_end_0, end_mask = z_35_end_mask_0, x = x_81_cast_fp16)[name = string("z_35_cast_fp16")];
            tensor<int32, [4]> var_3004_begin_0 = const()[name = string("op_3004_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3004_end_0 = const()[name = string("op_3004_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3004_end_mask_0 = const()[name = string("op_3004_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, x = z_35_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3005_cast_fp16 = mul(x = var_3004_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_3005_cast_fp16")];
            tensor<int32, [4]> var_3010_begin_0 = const()[name = string("op_3010_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3010_end_0 = const()[name = string("op_3010_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3010_end_mask_0 = const()[name = string("op_3010_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3010_cast_fp16 = slice_by_index(begin = var_3010_begin_0, end = var_3010_end_0, end_mask = var_3010_end_mask_0, x = z_35_cast_fp16)[name = string("op_3010_cast_fp16")];
            int32 var_3012 = const()[name = string("op_3012"), val = int32(-1)];
            bool rot_35_interleave_0 = const()[name = string("rot_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_35_cast_fp16 = concat(axis = var_3012, interleave = rot_35_interleave_0, values = (var_3005_cast_fp16, var_3010_cast_fp16))[name = string("rot_35_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3014_cast_fp16 = mul(x = z_35_cast_fp16, y = c_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3015_cast_fp16 = mul(x = rot_35_cast_fp16, y = s_11_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3017_cast_fp16 = add(x = var_3014_cast_fp16, y = var_3015_cast_fp16)[name = string("op_3017_cast_fp16")];
            int32 var_3019 = const()[name = string("op_3019"), val = int32(-1)];
            bool var_3020_interleave_0 = const()[name = string("op_3020_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_3020_cast_fp16 = concat(axis = var_3019, interleave = var_3020_interleave_0, values = (var_2994_cast_fp16, var_3017_cast_fp16))[name = string("op_3020_cast_fp16")];
            tensor<int32, [4]> z_37_begin_0 = const()[name = string("z_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_37_end_0 = const()[name = string("z_37_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_37_end_mask_0 = const()[name = string("z_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_37_cast_fp16 = slice_by_index(begin = z_37_begin_0, end = z_37_end_0, end_mask = z_37_end_mask_0, x = x_83_cast_fp16)[name = string("z_37_cast_fp16")];
            tensor<int32, [4]> var_3033_begin_0 = const()[name = string("op_3033_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3033_end_0 = const()[name = string("op_3033_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3033_end_mask_0 = const()[name = string("op_3033_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3033_cast_fp16 = slice_by_index(begin = var_3033_begin_0, end = var_3033_end_0, end_mask = var_3033_end_mask_0, x = z_37_cast_fp16)[name = string("op_3033_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3034_cast_fp16 = mul(x = var_3033_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> var_3039_begin_0 = const()[name = string("op_3039_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3039_end_0 = const()[name = string("op_3039_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3039_end_mask_0 = const()[name = string("op_3039_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3039_cast_fp16 = slice_by_index(begin = var_3039_begin_0, end = var_3039_end_0, end_mask = var_3039_end_mask_0, x = z_37_cast_fp16)[name = string("op_3039_cast_fp16")];
            int32 var_3041 = const()[name = string("op_3041"), val = int32(-1)];
            bool rot_37_interleave_0 = const()[name = string("rot_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_37_cast_fp16 = concat(axis = var_3041, interleave = rot_37_interleave_0, values = (var_3034_cast_fp16, var_3039_cast_fp16))[name = string("rot_37_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3043_cast_fp16 = mul(x = z_37_cast_fp16, y = c_1_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3044_cast_fp16 = mul(x = rot_37_cast_fp16, y = s_9_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3046_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3044_cast_fp16)[name = string("op_3046_cast_fp16")];
            tensor<int32, [4]> z_39_begin_0 = const()[name = string("z_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_39_end_0 = const()[name = string("z_39_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_39_end_mask_0 = const()[name = string("z_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_39_cast_fp16 = slice_by_index(begin = z_39_begin_0, end = z_39_end_0, end_mask = z_39_end_mask_0, x = x_83_cast_fp16)[name = string("z_39_cast_fp16")];
            tensor<int32, [4]> var_3056_begin_0 = const()[name = string("op_3056_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3056_end_0 = const()[name = string("op_3056_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3056_end_mask_0 = const()[name = string("op_3056_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3056_cast_fp16 = slice_by_index(begin = var_3056_begin_0, end = var_3056_end_0, end_mask = var_3056_end_mask_0, x = z_39_cast_fp16)[name = string("op_3056_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3057_cast_fp16 = mul(x = var_3056_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_3057_cast_fp16")];
            tensor<int32, [4]> var_3062_begin_0 = const()[name = string("op_3062_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3062_end_0 = const()[name = string("op_3062_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3062_end_mask_0 = const()[name = string("op_3062_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3062_cast_fp16 = slice_by_index(begin = var_3062_begin_0, end = var_3062_end_0, end_mask = var_3062_end_mask_0, x = z_39_cast_fp16)[name = string("op_3062_cast_fp16")];
            int32 var_3064 = const()[name = string("op_3064"), val = int32(-1)];
            bool rot_39_interleave_0 = const()[name = string("rot_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_39_cast_fp16 = concat(axis = var_3064, interleave = rot_39_interleave_0, values = (var_3057_cast_fp16, var_3062_cast_fp16))[name = string("rot_39_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3066_cast_fp16 = mul(x = z_39_cast_fp16, y = c_cast_fp16)[name = string("op_3066_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3067_cast_fp16 = mul(x = rot_39_cast_fp16, y = s_11_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3069_cast_fp16 = add(x = var_3066_cast_fp16, y = var_3067_cast_fp16)[name = string("op_3069_cast_fp16")];
            int32 var_3071 = const()[name = string("op_3071"), val = int32(-1)];
            bool var_3072_interleave_0 = const()[name = string("op_3072_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_3072_cast_fp16 = concat(axis = var_3071, interleave = var_3072_interleave_0, values = (var_3046_cast_fp16, var_3069_cast_fp16))[name = string("op_3072_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3082_transpose_x_0 = const()[name = string("op_3082_transpose_x_0"), val = bool(false)];
            bool var_3082_transpose_y_0 = const()[name = string("op_3082_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_3072_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_72 = transpose(perm = transpose_72_perm_0, x = var_3020_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 12, 2520, 2520]> var_3082_cast_fp16 = matmul(transpose_x = var_3082_transpose_x_0, transpose_y = var_3082_transpose_y_0, x = transpose_72, y = transpose_73)[name = string("op_3082_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_3084_cast_fp16 = add(x = var_3082_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3084_cast_fp16")];
            int32 var_3085 = const()[name = string("op_3085"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_63_cast_fp16 = softmax(axis = var_3085, x = var_3084_cast_fp16)[name = string("w_63_cast_fp16")];
            bool var_3088_transpose_x_0 = const()[name = string("op_3088_transpose_x_0"), val = bool(false)];
            bool var_3088_transpose_y_0 = const()[name = string("op_3088_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = v_17_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 12, 2520, 64]> var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = w_63_cast_fp16, y = v_19_cast_fp16)[name = string("op_3088_cast_fp16")];
            tensor<int32, [4]> var_3091_perm_0 = const()[name = string("op_3091_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3095 = const()[name = string("op_3095"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_3091_cast_fp16 = transpose(perm = var_3091_perm_0, x = var_3088_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 2520, 768]> input_45_cast_fp16 = reshape(shape = var_3095, x = var_3091_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40234688))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40824576))))[name = string("t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3100_cast_fp16 = abs(x = linear_32_cast_fp16)[name = string("op_3100_cast_fp16")];
            bool var_3103 = const()[name = string("op_3103"), val = bool(true)];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3104_cast_fp16 = reduce_max(axes = var_3104_axes_0, keep_dims = var_3103, x = var_3100_cast_fp16)[name = string("op_3104_cast_fp16")];
            fp16 var_3105_to_fp16 = const()[name = string("op_3105_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_69_cast_fp16 = maximum(x = var_3104_cast_fp16, y = var_3105_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_65_cast_fp16 = real_div(x = linear_32_cast_fp16, y = s_69_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3108_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [1]> var_3113_axes_0 = const()[name = string("op_3113_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3113_keep_dims_0 = const()[name = string("op_3113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3113_cast_fp16 = reduce_mean(axes = var_3113_axes_0, keep_dims = var_3113_keep_dims_0, x = var_3108_cast_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3114_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3114_cast_fp16")];
            fp32 var_3115_epsilon_0 = const()[name = string("op_3115_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3115_cast_fp16 = inverse(epsilon = var_3115_epsilon_0, x = var_3114_cast_fp16)[name = string("op_3115_cast_fp16")];
            fp16 var_3116_to_fp16 = const()[name = string("op_3116_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3117_cast_fp16 = mul(x = var_3115_cast_fp16, y = var_3116_to_fp16)[name = string("op_3117_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3119_cast_fp16 = add(x = var_3113_cast_fp16, y = var_3117_cast_fp16)[name = string("op_3119_cast_fp16")];
            fp32 var_3120_epsilon_0 = const()[name = string("op_3120_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3120_cast_fp16 = rsqrt(epsilon = var_3120_epsilon_0, x = var_3119_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_55_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3120_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40826176)))];
            tensor<fp16, [1, 2520, 768]> var_3122_cast_fp16 = mul(x = out_55_cast_fp16, y = t_encoder_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3122_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_87_cast_fp16 = add(x = x_73_cast_fp16, y = var_3122_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3125_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3125_cast_fp16")];
            bool var_3128 = const()[name = string("op_3128"), val = bool(true)];
            tensor<int32, [1]> var_3129_axes_0 = const()[name = string("op_3129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3129_cast_fp16 = reduce_max(axes = var_3129_axes_0, keep_dims = var_3128, x = var_3125_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_71_cast_fp16 = maximum(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_67_cast_fp16 = real_div(x = x_87_cast_fp16, y = s_71_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3133_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<int32, [1]> var_3138_axes_0 = const()[name = string("op_3138_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3138_keep_dims_0 = const()[name = string("op_3138_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3138_cast_fp16 = reduce_mean(axes = var_3138_axes_0, keep_dims = var_3138_keep_dims_0, x = var_3133_cast_fp16)[name = string("op_3138_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3139_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3139_cast_fp16")];
            fp32 var_3140_epsilon_0 = const()[name = string("op_3140_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3140_cast_fp16 = inverse(epsilon = var_3140_epsilon_0, x = var_3139_cast_fp16)[name = string("op_3140_cast_fp16")];
            fp16 var_3141_to_fp16 = const()[name = string("op_3141_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3142_cast_fp16 = mul(x = var_3140_cast_fp16, y = var_3141_to_fp16)[name = string("op_3142_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3144_cast_fp16 = add(x = var_3138_cast_fp16, y = var_3142_cast_fp16)[name = string("op_3144_cast_fp16")];
            fp32 var_3145_epsilon_0 = const()[name = string("op_3145_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3145_cast_fp16 = rsqrt(epsilon = var_3145_epsilon_0, x = var_3144_cast_fp16)[name = string("op_3145_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_57_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3145_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40827776)))];
            tensor<fp16, [1, 2520, 768]> input_47_cast_fp16 = mul(x = out_57_cast_fp16, y = t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40829376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43188736))))[name = string("t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_33_cast_fp16")];
            string var_3152_mode_0 = const()[name = string("op_3152_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_3152_cast_fp16 = gelu(mode = var_3152_mode_0, x = linear_33_cast_fp16)[name = string("op_3152_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43194944))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45554304))))[name = string("t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_34_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_51_cast_fp16 = mul(x = var_3152_cast_fp16, y = linear_34_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45560512))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47919872))))[name = string("t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3160_cast_fp16 = abs(x = linear_35_cast_fp16)[name = string("op_3160_cast_fp16")];
            bool var_3163 = const()[name = string("op_3163"), val = bool(true)];
            tensor<int32, [1]> var_3164_axes_0 = const()[name = string("op_3164_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3164_cast_fp16 = reduce_max(axes = var_3164_axes_0, keep_dims = var_3163, x = var_3160_cast_fp16)[name = string("op_3164_cast_fp16")];
            fp16 var_3165_to_fp16 = const()[name = string("op_3165_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_73_cast_fp16 = maximum(x = var_3164_cast_fp16, y = var_3165_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_69_cast_fp16 = real_div(x = linear_35_cast_fp16, y = s_73_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3168_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3168_cast_fp16")];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3173_keep_dims_0 = const()[name = string("op_3173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3173_cast_fp16 = reduce_mean(axes = var_3173_axes_0, keep_dims = var_3173_keep_dims_0, x = var_3168_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3174_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3174_cast_fp16")];
            fp32 var_3175_epsilon_0 = const()[name = string("op_3175_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3175_cast_fp16 = inverse(epsilon = var_3175_epsilon_0, x = var_3174_cast_fp16)[name = string("op_3175_cast_fp16")];
            fp16 var_3176_to_fp16 = const()[name = string("op_3176_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3177_cast_fp16 = mul(x = var_3175_cast_fp16, y = var_3176_to_fp16)[name = string("op_3177_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3179_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3177_cast_fp16)[name = string("op_3179_cast_fp16")];
            fp32 var_3180_epsilon_0 = const()[name = string("op_3180_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3180_cast_fp16 = rsqrt(epsilon = var_3180_epsilon_0, x = var_3179_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_59_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3180_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47921472)))];
            tensor<fp16, [1, 2520, 768]> var_3182_cast_fp16 = mul(x = out_59_cast_fp16, y = t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_91_cast_fp16 = add(x = x_87_cast_fp16, y = var_3182_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3185_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3185_cast_fp16")];
            bool var_3188 = const()[name = string("op_3188"), val = bool(true)];
            tensor<int32, [1]> var_3189_axes_0 = const()[name = string("op_3189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3189_cast_fp16 = reduce_max(axes = var_3189_axes_0, keep_dims = var_3188, x = var_3185_cast_fp16)[name = string("op_3189_cast_fp16")];
            fp16 var_3190_to_fp16 = const()[name = string("op_3190_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_75_cast_fp16 = maximum(x = var_3189_cast_fp16, y = var_3190_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_71_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_75_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3193_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3193_cast_fp16")];
            tensor<int32, [1]> var_3198_axes_0 = const()[name = string("op_3198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3198_keep_dims_0 = const()[name = string("op_3198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3198_cast_fp16 = reduce_mean(axes = var_3198_axes_0, keep_dims = var_3198_keep_dims_0, x = var_3193_cast_fp16)[name = string("op_3198_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3199_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3199_cast_fp16")];
            fp32 var_3200_epsilon_0 = const()[name = string("op_3200_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3200_cast_fp16 = inverse(epsilon = var_3200_epsilon_0, x = var_3199_cast_fp16)[name = string("op_3200_cast_fp16")];
            fp16 var_3201_to_fp16 = const()[name = string("op_3201_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3202_cast_fp16 = mul(x = var_3200_cast_fp16, y = var_3201_to_fp16)[name = string("op_3202_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3204_cast_fp16 = add(x = var_3198_cast_fp16, y = var_3202_cast_fp16)[name = string("op_3204_cast_fp16")];
            fp32 var_3205_epsilon_0 = const()[name = string("op_3205_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3205_cast_fp16 = rsqrt(epsilon = var_3205_epsilon_0, x = var_3204_cast_fp16)[name = string("op_3205_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_61_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3205_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47923072)))];
            tensor<fp16, [1, 2520, 768]> input_53_cast_fp16 = mul(x = out_61_cast_fp16, y = t_encoder_layers_5_input_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47924672))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48514560))))[name = string("t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_3215 = const()[name = string("op_3215"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_93_cast_fp16 = reshape(shape = var_3215, x = linear_36_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3217_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3217_cast_fp16")];
            bool var_3220 = const()[name = string("op_3220"), val = bool(true)];
            tensor<int32, [1]> var_3221_axes_0 = const()[name = string("op_3221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3221_cast_fp16 = reduce_max(axes = var_3221_axes_0, keep_dims = var_3220, x = var_3217_cast_fp16)[name = string("op_3221_cast_fp16")];
            fp16 var_3222_to_fp16 = const()[name = string("op_3222_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_77_cast_fp16 = maximum(x = var_3221_cast_fp16, y = var_3222_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_73_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_77_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3225_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<int32, [1]> var_3230_axes_0 = const()[name = string("op_3230_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3230_keep_dims_0 = const()[name = string("op_3230_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3230_cast_fp16 = reduce_mean(axes = var_3230_axes_0, keep_dims = var_3230_keep_dims_0, x = var_3225_cast_fp16)[name = string("op_3230_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3231_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3231_cast_fp16")];
            fp32 var_3232_epsilon_0 = const()[name = string("op_3232_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3232_cast_fp16 = inverse(epsilon = var_3232_epsilon_0, x = var_3231_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3234_cast_fp16 = mul(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3236_cast_fp16 = add(x = var_3230_cast_fp16, y = var_3234_cast_fp16)[name = string("op_3236_cast_fp16")];
            fp32 var_3237_epsilon_0 = const()[name = string("op_3237_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3237_cast_fp16 = rsqrt(epsilon = var_3237_epsilon_0, x = var_3236_cast_fp16)[name = string("op_3237_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_63_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3237_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48516160)))];
            tensor<fp16, [1, 2520, 12, 64]> x_99_cast_fp16 = mul(x = out_63_cast_fp16, y = t_encoder_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_99_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48516352))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49106240))))[name = string("t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_37_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3247 = const()[name = string("op_3247"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_95_cast_fp16 = reshape(shape = var_3247, x = linear_37_cast_fp16)[name = string("x_95_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3249_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3249_cast_fp16")];
            bool var_3252 = const()[name = string("op_3252"), val = bool(true)];
            tensor<int32, [1]> var_3253_axes_0 = const()[name = string("op_3253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3253_cast_fp16 = reduce_max(axes = var_3253_axes_0, keep_dims = var_3252, x = var_3249_cast_fp16)[name = string("op_3253_cast_fp16")];
            fp16 var_3254_to_fp16 = const()[name = string("op_3254_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_79_cast_fp16 = maximum(x = var_3253_cast_fp16, y = var_3254_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_75_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_79_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3257_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3257_cast_fp16")];
            tensor<int32, [1]> var_3262_axes_0 = const()[name = string("op_3262_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3262_keep_dims_0 = const()[name = string("op_3262_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3262_cast_fp16 = reduce_mean(axes = var_3262_axes_0, keep_dims = var_3262_keep_dims_0, x = var_3257_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3263_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3263_cast_fp16")];
            fp32 var_3264_epsilon_0 = const()[name = string("op_3264_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3264_cast_fp16 = inverse(epsilon = var_3264_epsilon_0, x = var_3263_cast_fp16)[name = string("op_3264_cast_fp16")];
            fp16 var_3265_to_fp16 = const()[name = string("op_3265_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3266_cast_fp16 = mul(x = var_3264_cast_fp16, y = var_3265_to_fp16)[name = string("op_3266_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3268_cast_fp16 = add(x = var_3262_cast_fp16, y = var_3266_cast_fp16)[name = string("op_3268_cast_fp16")];
            fp32 var_3269_epsilon_0 = const()[name = string("op_3269_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3269_cast_fp16 = rsqrt(epsilon = var_3269_epsilon_0, x = var_3268_cast_fp16)[name = string("op_3269_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3269_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49107840)))];
            tensor<fp16, [1, 2520, 12, 64]> x_101_cast_fp16 = mul(x = out_65_cast_fp16, y = t_encoder_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49108032))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49697920))))[name = string("t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3279 = const()[name = string("op_3279"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_97_cast_fp16 = reshape(shape = var_3279, x = linear_38_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3281_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3281_cast_fp16")];
            bool var_3284 = const()[name = string("op_3284"), val = bool(true)];
            tensor<int32, [1]> var_3285_axes_0 = const()[name = string("op_3285_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3285_cast_fp16 = reduce_max(axes = var_3285_axes_0, keep_dims = var_3284, x = var_3281_cast_fp16)[name = string("op_3285_cast_fp16")];
            fp16 var_3286_to_fp16 = const()[name = string("op_3286_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_81_cast_fp16 = maximum(x = var_3285_cast_fp16, y = var_3286_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_77_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_81_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3289_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<int32, [1]> var_3294_axes_0 = const()[name = string("op_3294_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3294_keep_dims_0 = const()[name = string("op_3294_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3294_cast_fp16 = reduce_mean(axes = var_3294_axes_0, keep_dims = var_3294_keep_dims_0, x = var_3289_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3295_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3295_cast_fp16")];
            fp32 var_3296_epsilon_0 = const()[name = string("op_3296_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3296_cast_fp16 = inverse(epsilon = var_3296_epsilon_0, x = var_3295_cast_fp16)[name = string("op_3296_cast_fp16")];
            fp16 var_3297_to_fp16 = const()[name = string("op_3297_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3298_cast_fp16 = mul(x = var_3296_cast_fp16, y = var_3297_to_fp16)[name = string("op_3298_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3300_cast_fp16 = add(x = var_3294_cast_fp16, y = var_3298_cast_fp16)[name = string("op_3300_cast_fp16")];
            fp32 var_3301_epsilon_0 = const()[name = string("op_3301_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3301_cast_fp16 = rsqrt(epsilon = var_3301_epsilon_0, x = var_3300_cast_fp16)[name = string("op_3301_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_21_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3301_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<int32, [4]> z_41_begin_0 = const()[name = string("z_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_41_end_0 = const()[name = string("z_41_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_41_end_mask_0 = const()[name = string("z_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_41_cast_fp16 = slice_by_index(begin = z_41_begin_0, end = z_41_end_0, end_mask = z_41_end_mask_0, x = x_99_cast_fp16)[name = string("z_41_cast_fp16")];
            tensor<int32, [4]> var_3312_begin_0 = const()[name = string("op_3312_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3312_end_0 = const()[name = string("op_3312_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3312_end_mask_0 = const()[name = string("op_3312_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3312_cast_fp16 = slice_by_index(begin = var_3312_begin_0, end = var_3312_end_0, end_mask = var_3312_end_mask_0, x = z_41_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3313_cast_fp16 = mul(x = var_3312_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_3313_cast_fp16")];
            tensor<int32, [4]> var_3318_begin_0 = const()[name = string("op_3318_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3318_end_0 = const()[name = string("op_3318_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3318_end_mask_0 = const()[name = string("op_3318_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> 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 = z_41_cast_fp16)[name = string("op_3318_cast_fp16")];
            int32 var_3320 = const()[name = string("op_3320"), val = int32(-1)];
            bool rot_41_interleave_0 = const()[name = string("rot_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_41_cast_fp16 = concat(axis = var_3320, interleave = rot_41_interleave_0, values = (var_3313_cast_fp16, var_3318_cast_fp16))[name = string("rot_41_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3322_cast_fp16 = mul(x = z_41_cast_fp16, y = c_1_cast_fp16)[name = string("op_3322_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3323_cast_fp16 = mul(x = rot_41_cast_fp16, y = s_9_cast_fp16)[name = string("op_3323_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3325_cast_fp16 = add(x = var_3322_cast_fp16, y = var_3323_cast_fp16)[name = string("op_3325_cast_fp16")];
            tensor<int32, [4]> z_43_begin_0 = const()[name = string("z_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_43_end_0 = const()[name = string("z_43_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_43_end_mask_0 = const()[name = string("z_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_43_cast_fp16 = slice_by_index(begin = z_43_begin_0, end = z_43_end_0, end_mask = z_43_end_mask_0, x = x_99_cast_fp16)[name = string("z_43_cast_fp16")];
            tensor<int32, [4]> var_3335_begin_0 = const()[name = string("op_3335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3335_end_0 = const()[name = string("op_3335_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3335_end_mask_0 = const()[name = string("op_3335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3335_cast_fp16 = slice_by_index(begin = var_3335_begin_0, end = var_3335_end_0, end_mask = var_3335_end_mask_0, x = z_43_cast_fp16)[name = string("op_3335_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3336_cast_fp16 = mul(x = var_3335_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> var_3341_begin_0 = const()[name = string("op_3341_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3341_end_0 = const()[name = string("op_3341_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3341_end_mask_0 = const()[name = string("op_3341_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3341_cast_fp16 = slice_by_index(begin = var_3341_begin_0, end = var_3341_end_0, end_mask = var_3341_end_mask_0, x = z_43_cast_fp16)[name = string("op_3341_cast_fp16")];
            int32 var_3343 = const()[name = string("op_3343"), val = int32(-1)];
            bool rot_43_interleave_0 = const()[name = string("rot_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_43_cast_fp16 = concat(axis = var_3343, interleave = rot_43_interleave_0, values = (var_3336_cast_fp16, var_3341_cast_fp16))[name = string("rot_43_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3345_cast_fp16 = mul(x = z_43_cast_fp16, y = c_cast_fp16)[name = string("op_3345_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3346_cast_fp16 = mul(x = rot_43_cast_fp16, y = s_11_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3348_cast_fp16 = add(x = var_3345_cast_fp16, y = var_3346_cast_fp16)[name = string("op_3348_cast_fp16")];
            int32 var_3350 = const()[name = string("op_3350"), val = int32(-1)];
            bool var_3351_interleave_0 = const()[name = string("op_3351_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_3351_cast_fp16 = concat(axis = var_3350, interleave = var_3351_interleave_0, values = (var_3325_cast_fp16, var_3348_cast_fp16))[name = string("op_3351_cast_fp16")];
            tensor<int32, [4]> z_45_begin_0 = const()[name = string("z_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_45_end_0 = const()[name = string("z_45_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_45_end_mask_0 = const()[name = string("z_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_45_cast_fp16 = slice_by_index(begin = z_45_begin_0, end = z_45_end_0, end_mask = z_45_end_mask_0, x = x_101_cast_fp16)[name = string("z_45_cast_fp16")];
            tensor<int32, [4]> var_3364_begin_0 = const()[name = string("op_3364_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3364_end_0 = const()[name = string("op_3364_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3364_end_mask_0 = const()[name = string("op_3364_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3364_cast_fp16 = slice_by_index(begin = var_3364_begin_0, end = var_3364_end_0, end_mask = var_3364_end_mask_0, x = z_45_cast_fp16)[name = string("op_3364_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3365_cast_fp16 = mul(x = var_3364_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_3365_cast_fp16")];
            tensor<int32, [4]> var_3370_begin_0 = const()[name = string("op_3370_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3370_end_0 = const()[name = string("op_3370_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3370_end_mask_0 = const()[name = string("op_3370_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3370_cast_fp16 = slice_by_index(begin = var_3370_begin_0, end = var_3370_end_0, end_mask = var_3370_end_mask_0, x = z_45_cast_fp16)[name = string("op_3370_cast_fp16")];
            int32 var_3372 = const()[name = string("op_3372"), val = int32(-1)];
            bool rot_45_interleave_0 = const()[name = string("rot_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_45_cast_fp16 = concat(axis = var_3372, interleave = rot_45_interleave_0, values = (var_3365_cast_fp16, var_3370_cast_fp16))[name = string("rot_45_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3374_cast_fp16 = mul(x = z_45_cast_fp16, y = c_1_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3375_cast_fp16 = mul(x = rot_45_cast_fp16, y = s_9_cast_fp16)[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3377_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3375_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [4]> z_47_begin_0 = const()[name = string("z_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_47_end_0 = const()[name = string("z_47_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_47_end_mask_0 = const()[name = string("z_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_47_cast_fp16 = slice_by_index(begin = z_47_begin_0, end = z_47_end_0, end_mask = z_47_end_mask_0, x = x_101_cast_fp16)[name = string("z_47_cast_fp16")];
            tensor<int32, [4]> var_3387_begin_0 = const()[name = string("op_3387_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3387_end_0 = const()[name = string("op_3387_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3387_end_mask_0 = const()[name = string("op_3387_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3387_cast_fp16 = slice_by_index(begin = var_3387_begin_0, end = var_3387_end_0, end_mask = var_3387_end_mask_0, x = z_47_cast_fp16)[name = string("op_3387_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3388_cast_fp16 = mul(x = var_3387_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_3388_cast_fp16")];
            tensor<int32, [4]> var_3393_begin_0 = const()[name = string("op_3393_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3393_end_0 = const()[name = string("op_3393_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3393_end_mask_0 = const()[name = string("op_3393_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3393_cast_fp16 = slice_by_index(begin = var_3393_begin_0, end = var_3393_end_0, end_mask = var_3393_end_mask_0, x = z_47_cast_fp16)[name = string("op_3393_cast_fp16")];
            int32 var_3395 = const()[name = string("op_3395"), val = int32(-1)];
            bool rot_47_interleave_0 = const()[name = string("rot_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_47_cast_fp16 = concat(axis = var_3395, interleave = rot_47_interleave_0, values = (var_3388_cast_fp16, var_3393_cast_fp16))[name = string("rot_47_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3397_cast_fp16 = mul(x = z_47_cast_fp16, y = c_cast_fp16)[name = string("op_3397_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3398_cast_fp16 = mul(x = rot_47_cast_fp16, y = s_11_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3400_cast_fp16 = add(x = var_3397_cast_fp16, y = var_3398_cast_fp16)[name = string("op_3400_cast_fp16")];
            int32 var_3402 = const()[name = string("op_3402"), val = int32(-1)];
            bool var_3403_interleave_0 = const()[name = string("op_3403_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_3403_cast_fp16 = concat(axis = var_3402, interleave = var_3403_interleave_0, values = (var_3377_cast_fp16, var_3400_cast_fp16))[name = string("op_3403_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3413_transpose_x_0 = const()[name = string("op_3413_transpose_x_0"), val = bool(false)];
            bool var_3413_transpose_y_0 = const()[name = string("op_3413_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_3403_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_74 = transpose(perm = transpose_74_perm_0, x = var_3351_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 12, 2520, 2520]> var_3413_cast_fp16 = matmul(transpose_x = var_3413_transpose_x_0, transpose_y = var_3413_transpose_y_0, x = transpose_74, y = transpose_75)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_3415_cast_fp16 = add(x = var_3413_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3415_cast_fp16")];
            int32 var_3416 = const()[name = string("op_3416"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_77_cast_fp16 = softmax(axis = var_3416, x = var_3415_cast_fp16)[name = string("w_77_cast_fp16")];
            bool var_3419_transpose_x_0 = const()[name = string("op_3419_transpose_x_0"), val = bool(false)];
            bool var_3419_transpose_y_0 = const()[name = string("op_3419_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = v_21_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 12, 2520, 64]> var_3419_cast_fp16 = matmul(transpose_x = var_3419_transpose_x_0, transpose_y = var_3419_transpose_y_0, x = w_77_cast_fp16, y = v_23_cast_fp16)[name = string("op_3419_cast_fp16")];
            tensor<int32, [4]> var_3422_perm_0 = const()[name = string("op_3422_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3426 = const()[name = string("op_3426"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_3422_cast_fp16 = transpose(perm = var_3422_perm_0, x = var_3419_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 2520, 768]> input_55_cast_fp16 = reshape(shape = var_3426, x = var_3422_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49699520))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50289408))))[name = string("t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_39_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3431_cast_fp16 = abs(x = linear_39_cast_fp16)[name = string("op_3431_cast_fp16")];
            bool var_3434 = const()[name = string("op_3434"), val = bool(true)];
            tensor<int32, [1]> var_3435_axes_0 = const()[name = string("op_3435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3435_cast_fp16 = reduce_max(axes = var_3435_axes_0, keep_dims = var_3434, x = var_3431_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp16 var_3436_to_fp16 = const()[name = string("op_3436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_83_cast_fp16 = maximum(x = var_3435_cast_fp16, y = var_3436_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_79_cast_fp16 = real_div(x = linear_39_cast_fp16, y = s_83_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3439_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3439_cast_fp16")];
            tensor<int32, [1]> var_3444_axes_0 = const()[name = string("op_3444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3444_keep_dims_0 = const()[name = string("op_3444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3444_cast_fp16 = reduce_mean(axes = var_3444_axes_0, keep_dims = var_3444_keep_dims_0, x = var_3439_cast_fp16)[name = string("op_3444_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3445_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3445_cast_fp16")];
            fp32 var_3446_epsilon_0 = const()[name = string("op_3446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3446_cast_fp16 = inverse(epsilon = var_3446_epsilon_0, x = var_3445_cast_fp16)[name = string("op_3446_cast_fp16")];
            fp16 var_3447_to_fp16 = const()[name = string("op_3447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3448_cast_fp16 = mul(x = var_3446_cast_fp16, y = var_3447_to_fp16)[name = string("op_3448_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3450_cast_fp16 = add(x = var_3444_cast_fp16, y = var_3448_cast_fp16)[name = string("op_3450_cast_fp16")];
            fp32 var_3451_epsilon_0 = const()[name = string("op_3451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3451_cast_fp16 = rsqrt(epsilon = var_3451_epsilon_0, x = var_3450_cast_fp16)[name = string("op_3451_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_67_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3451_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50291008)))];
            tensor<fp16, [1, 2520, 768]> var_3453_cast_fp16 = mul(x = out_67_cast_fp16, y = t_encoder_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_3453_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_105_cast_fp16 = add(x = x_91_cast_fp16, y = var_3453_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3456_cast_fp16 = abs(x = x_105_cast_fp16)[name = string("op_3456_cast_fp16")];
            bool var_3459 = const()[name = string("op_3459"), val = bool(true)];
            tensor<int32, [1]> var_3460_axes_0 = const()[name = string("op_3460_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3460_cast_fp16 = reduce_max(axes = var_3460_axes_0, keep_dims = var_3459, x = var_3456_cast_fp16)[name = string("op_3460_cast_fp16")];
            fp16 var_3461_to_fp16 = const()[name = string("op_3461_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_85_cast_fp16 = maximum(x = var_3460_cast_fp16, y = var_3461_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_81_cast_fp16 = real_div(x = x_105_cast_fp16, y = s_85_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3464_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3464_cast_fp16")];
            tensor<int32, [1]> var_3469_axes_0 = const()[name = string("op_3469_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3469_keep_dims_0 = const()[name = string("op_3469_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3469_cast_fp16 = reduce_mean(axes = var_3469_axes_0, keep_dims = var_3469_keep_dims_0, x = var_3464_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3470_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3470_cast_fp16")];
            fp32 var_3471_epsilon_0 = const()[name = string("op_3471_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3471_cast_fp16 = inverse(epsilon = var_3471_epsilon_0, x = var_3470_cast_fp16)[name = string("op_3471_cast_fp16")];
            fp16 var_3472_to_fp16 = const()[name = string("op_3472_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3473_cast_fp16 = mul(x = var_3471_cast_fp16, y = var_3472_to_fp16)[name = string("op_3473_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3475_cast_fp16 = add(x = var_3469_cast_fp16, y = var_3473_cast_fp16)[name = string("op_3475_cast_fp16")];
            fp32 var_3476_epsilon_0 = const()[name = string("op_3476_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3476_cast_fp16 = rsqrt(epsilon = var_3476_epsilon_0, x = var_3475_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_69_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3476_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50292608)))];
            tensor<fp16, [1, 2520, 768]> input_57_cast_fp16 = mul(x = out_69_cast_fp16, y = t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50294208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52653568))))[name = string("t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_40_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_40_cast_fp16")];
            string var_3483_mode_0 = const()[name = string("op_3483_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_3483_cast_fp16 = gelu(mode = var_3483_mode_0, x = linear_40_cast_fp16)[name = string("op_3483_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52659776))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55019136))))[name = string("t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_61_cast_fp16 = mul(x = var_3483_cast_fp16, y = linear_41_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55025344))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57384704))))[name = string("t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3491_cast_fp16 = abs(x = linear_42_cast_fp16)[name = string("op_3491_cast_fp16")];
            bool var_3494 = const()[name = string("op_3494"), val = bool(true)];
            tensor<int32, [1]> var_3495_axes_0 = const()[name = string("op_3495_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3495_cast_fp16 = reduce_max(axes = var_3495_axes_0, keep_dims = var_3494, x = var_3491_cast_fp16)[name = string("op_3495_cast_fp16")];
            fp16 var_3496_to_fp16 = const()[name = string("op_3496_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_87_cast_fp16 = maximum(x = var_3495_cast_fp16, y = var_3496_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_83_cast_fp16 = real_div(x = linear_42_cast_fp16, y = s_87_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3499_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3499_cast_fp16")];
            tensor<int32, [1]> var_3504_axes_0 = const()[name = string("op_3504_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3504_keep_dims_0 = const()[name = string("op_3504_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3504_cast_fp16 = reduce_mean(axes = var_3504_axes_0, keep_dims = var_3504_keep_dims_0, x = var_3499_cast_fp16)[name = string("op_3504_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3505_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3505_cast_fp16")];
            fp32 var_3506_epsilon_0 = const()[name = string("op_3506_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3506_cast_fp16 = inverse(epsilon = var_3506_epsilon_0, x = var_3505_cast_fp16)[name = string("op_3506_cast_fp16")];
            fp16 var_3507_to_fp16 = const()[name = string("op_3507_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3508_cast_fp16 = mul(x = var_3506_cast_fp16, y = var_3507_to_fp16)[name = string("op_3508_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3510_cast_fp16 = add(x = var_3504_cast_fp16, y = var_3508_cast_fp16)[name = string("op_3510_cast_fp16")];
            fp32 var_3511_epsilon_0 = const()[name = string("op_3511_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3511_cast_fp16 = rsqrt(epsilon = var_3511_epsilon_0, x = var_3510_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_71_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3511_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57386304)))];
            tensor<fp16, [1, 2520, 768]> var_3513_cast_fp16 = mul(x = out_71_cast_fp16, y = t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3513_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_109_cast_fp16 = add(x = x_105_cast_fp16, y = var_3513_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3516_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_3516_cast_fp16")];
            bool var_3519 = const()[name = string("op_3519"), val = bool(true)];
            tensor<int32, [1]> var_3520_axes_0 = const()[name = string("op_3520_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3520_cast_fp16 = reduce_max(axes = var_3520_axes_0, keep_dims = var_3519, x = var_3516_cast_fp16)[name = string("op_3520_cast_fp16")];
            fp16 var_3521_to_fp16 = const()[name = string("op_3521_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_89_cast_fp16 = maximum(x = var_3520_cast_fp16, y = var_3521_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_85_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_89_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3524_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<int32, [1]> var_3529_axes_0 = const()[name = string("op_3529_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3529_keep_dims_0 = const()[name = string("op_3529_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3529_cast_fp16 = reduce_mean(axes = var_3529_axes_0, keep_dims = var_3529_keep_dims_0, x = var_3524_cast_fp16)[name = string("op_3529_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3530_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3530_cast_fp16")];
            fp32 var_3531_epsilon_0 = const()[name = string("op_3531_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3531_cast_fp16 = inverse(epsilon = var_3531_epsilon_0, x = var_3530_cast_fp16)[name = string("op_3531_cast_fp16")];
            fp16 var_3532_to_fp16 = const()[name = string("op_3532_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3533_cast_fp16 = mul(x = var_3531_cast_fp16, y = var_3532_to_fp16)[name = string("op_3533_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3535_cast_fp16 = add(x = var_3529_cast_fp16, y = var_3533_cast_fp16)[name = string("op_3535_cast_fp16")];
            fp32 var_3536_epsilon_0 = const()[name = string("op_3536_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3536_cast_fp16 = rsqrt(epsilon = var_3536_epsilon_0, x = var_3535_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_73_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3536_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57387904)))];
            tensor<fp16, [1, 2520, 768]> input_63_cast_fp16 = mul(x = out_73_cast_fp16, y = t_encoder_layers_6_input_layernorm_weight_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57389504))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57979392))))[name = string("t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_43_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_3546 = const()[name = string("op_3546"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_111_cast_fp16 = reshape(shape = var_3546, x = linear_43_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3548_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_3548_cast_fp16")];
            bool var_3551 = const()[name = string("op_3551"), val = bool(true)];
            tensor<int32, [1]> var_3552_axes_0 = const()[name = string("op_3552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3552_cast_fp16 = reduce_max(axes = var_3552_axes_0, keep_dims = var_3551, x = var_3548_cast_fp16)[name = string("op_3552_cast_fp16")];
            fp16 var_3553_to_fp16 = const()[name = string("op_3553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_91_cast_fp16 = maximum(x = var_3552_cast_fp16, y = var_3553_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_87_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_91_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3556_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3556_cast_fp16")];
            tensor<int32, [1]> var_3561_axes_0 = const()[name = string("op_3561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3561_keep_dims_0 = const()[name = string("op_3561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3561_cast_fp16 = reduce_mean(axes = var_3561_axes_0, keep_dims = var_3561_keep_dims_0, x = var_3556_cast_fp16)[name = string("op_3561_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3562_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp32 var_3563_epsilon_0 = const()[name = string("op_3563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3563_cast_fp16 = inverse(epsilon = var_3563_epsilon_0, x = var_3562_cast_fp16)[name = string("op_3563_cast_fp16")];
            fp16 var_3564_to_fp16 = const()[name = string("op_3564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3565_cast_fp16 = mul(x = var_3563_cast_fp16, y = var_3564_to_fp16)[name = string("op_3565_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3567_cast_fp16 = add(x = var_3561_cast_fp16, y = var_3565_cast_fp16)[name = string("op_3567_cast_fp16")];
            fp32 var_3568_epsilon_0 = const()[name = string("op_3568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3568_cast_fp16 = rsqrt(epsilon = var_3568_epsilon_0, x = var_3567_cast_fp16)[name = string("op_3568_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_75_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3568_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 2520, 12, 64]> x_117_cast_fp16 = mul(x = out_75_cast_fp16, y = t_encoder_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_117_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57980992))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58570880))))[name = string("t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_3578 = const()[name = string("op_3578"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_113_cast_fp16 = reshape(shape = var_3578, x = linear_44_cast_fp16)[name = string("x_113_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3580_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_3580_cast_fp16")];
            bool var_3583 = const()[name = string("op_3583"), val = bool(true)];
            tensor<int32, [1]> var_3584_axes_0 = const()[name = string("op_3584_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3584_cast_fp16 = reduce_max(axes = var_3584_axes_0, keep_dims = var_3583, x = var_3580_cast_fp16)[name = string("op_3584_cast_fp16")];
            fp16 var_3585_to_fp16 = const()[name = string("op_3585_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_93_cast_fp16 = maximum(x = var_3584_cast_fp16, y = var_3585_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_89_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_93_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3588_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3588_cast_fp16")];
            tensor<int32, [1]> var_3593_axes_0 = const()[name = string("op_3593_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3593_keep_dims_0 = const()[name = string("op_3593_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3593_cast_fp16 = reduce_mean(axes = var_3593_axes_0, keep_dims = var_3593_keep_dims_0, x = var_3588_cast_fp16)[name = string("op_3593_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3594_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_3594_cast_fp16")];
            fp32 var_3595_epsilon_0 = const()[name = string("op_3595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3595_cast_fp16 = inverse(epsilon = var_3595_epsilon_0, x = var_3594_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp16 var_3596_to_fp16 = const()[name = string("op_3596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3597_cast_fp16 = mul(x = var_3595_cast_fp16, y = var_3596_to_fp16)[name = string("op_3597_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3599_cast_fp16 = add(x = var_3593_cast_fp16, y = var_3597_cast_fp16)[name = string("op_3599_cast_fp16")];
            fp32 var_3600_epsilon_0 = const()[name = string("op_3600_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3600_cast_fp16 = rsqrt(epsilon = var_3600_epsilon_0, x = var_3599_cast_fp16)[name = string("op_3600_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_77_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3600_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58572480)))];
            tensor<fp16, [1, 2520, 12, 64]> x_119_cast_fp16 = mul(x = out_77_cast_fp16, y = t_encoder_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58572672))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59162560))))[name = string("t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<int32, [4]> var_3610 = const()[name = string("op_3610"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_115_cast_fp16 = reshape(shape = var_3610, x = linear_45_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3612_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_3612_cast_fp16")];
            bool var_3615 = const()[name = string("op_3615"), val = bool(true)];
            tensor<int32, [1]> var_3616_axes_0 = const()[name = string("op_3616_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3616_cast_fp16 = reduce_max(axes = var_3616_axes_0, keep_dims = var_3615, x = var_3612_cast_fp16)[name = string("op_3616_cast_fp16")];
            fp16 var_3617_to_fp16 = const()[name = string("op_3617_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_95_cast_fp16 = maximum(x = var_3616_cast_fp16, y = var_3617_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_91_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_95_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3620_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_3620_cast_fp16")];
            tensor<int32, [1]> var_3625_axes_0 = const()[name = string("op_3625_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3625_keep_dims_0 = const()[name = string("op_3625_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3625_cast_fp16 = reduce_mean(axes = var_3625_axes_0, keep_dims = var_3625_keep_dims_0, x = var_3620_cast_fp16)[name = string("op_3625_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3626_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_3626_cast_fp16")];
            fp32 var_3627_epsilon_0 = const()[name = string("op_3627_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3627_cast_fp16 = inverse(epsilon = var_3627_epsilon_0, x = var_3626_cast_fp16)[name = string("op_3627_cast_fp16")];
            fp16 var_3628_to_fp16 = const()[name = string("op_3628_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3629_cast_fp16 = mul(x = var_3627_cast_fp16, y = var_3628_to_fp16)[name = string("op_3629_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3631_cast_fp16 = add(x = var_3625_cast_fp16, y = var_3629_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp32 var_3632_epsilon_0 = const()[name = string("op_3632_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3632_cast_fp16 = rsqrt(epsilon = var_3632_epsilon_0, x = var_3631_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_25_cast_fp16 = mul(x = y_91_cast_fp16, y = var_3632_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<int32, [4]> z_49_begin_0 = const()[name = string("z_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_49_end_0 = const()[name = string("z_49_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_49_end_mask_0 = const()[name = string("z_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_49_cast_fp16 = slice_by_index(begin = z_49_begin_0, end = z_49_end_0, end_mask = z_49_end_mask_0, x = x_117_cast_fp16)[name = string("z_49_cast_fp16")];
            tensor<int32, [4]> var_3643_begin_0 = const()[name = string("op_3643_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3643_end_0 = const()[name = string("op_3643_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3643_end_mask_0 = const()[name = string("op_3643_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3643_cast_fp16 = slice_by_index(begin = var_3643_begin_0, end = var_3643_end_0, end_mask = var_3643_end_mask_0, x = z_49_cast_fp16)[name = string("op_3643_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3644_cast_fp16 = mul(x = var_3643_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = z_49_cast_fp16)[name = string("op_3649_cast_fp16")];
            int32 var_3651 = const()[name = string("op_3651"), val = int32(-1)];
            bool rot_49_interleave_0 = const()[name = string("rot_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_49_cast_fp16 = concat(axis = var_3651, interleave = rot_49_interleave_0, values = (var_3644_cast_fp16, var_3649_cast_fp16))[name = string("rot_49_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3653_cast_fp16 = mul(x = z_49_cast_fp16, y = c_1_cast_fp16)[name = string("op_3653_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3654_cast_fp16 = mul(x = rot_49_cast_fp16, y = s_9_cast_fp16)[name = string("op_3654_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3656_cast_fp16 = add(x = var_3653_cast_fp16, y = var_3654_cast_fp16)[name = string("op_3656_cast_fp16")];
            tensor<int32, [4]> z_51_begin_0 = const()[name = string("z_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_51_end_0 = const()[name = string("z_51_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_51_end_mask_0 = const()[name = string("z_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_51_cast_fp16 = slice_by_index(begin = z_51_begin_0, end = z_51_end_0, end_mask = z_51_end_mask_0, x = x_117_cast_fp16)[name = string("z_51_cast_fp16")];
            tensor<int32, [4]> var_3666_begin_0 = const()[name = string("op_3666_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3666_end_0 = const()[name = string("op_3666_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3666_end_mask_0 = const()[name = string("op_3666_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3666_cast_fp16 = slice_by_index(begin = var_3666_begin_0, end = var_3666_end_0, end_mask = var_3666_end_mask_0, x = z_51_cast_fp16)[name = string("op_3666_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3667_cast_fp16 = mul(x = var_3666_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_3667_cast_fp16")];
            tensor<int32, [4]> var_3672_begin_0 = const()[name = string("op_3672_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3672_end_0 = const()[name = string("op_3672_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3672_end_mask_0 = const()[name = string("op_3672_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3672_cast_fp16 = slice_by_index(begin = var_3672_begin_0, end = var_3672_end_0, end_mask = var_3672_end_mask_0, x = z_51_cast_fp16)[name = string("op_3672_cast_fp16")];
            int32 var_3674 = const()[name = string("op_3674"), val = int32(-1)];
            bool rot_51_interleave_0 = const()[name = string("rot_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_51_cast_fp16 = concat(axis = var_3674, interleave = rot_51_interleave_0, values = (var_3667_cast_fp16, var_3672_cast_fp16))[name = string("rot_51_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3676_cast_fp16 = mul(x = z_51_cast_fp16, y = c_cast_fp16)[name = string("op_3676_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3677_cast_fp16 = mul(x = rot_51_cast_fp16, y = s_11_cast_fp16)[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3679_cast_fp16 = add(x = var_3676_cast_fp16, y = var_3677_cast_fp16)[name = string("op_3679_cast_fp16")];
            int32 var_3681 = const()[name = string("op_3681"), val = int32(-1)];
            bool var_3682_interleave_0 = const()[name = string("op_3682_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_3682_cast_fp16 = concat(axis = var_3681, interleave = var_3682_interleave_0, values = (var_3656_cast_fp16, var_3679_cast_fp16))[name = string("op_3682_cast_fp16")];
            tensor<int32, [4]> z_53_begin_0 = const()[name = string("z_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_53_end_0 = const()[name = string("z_53_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_53_end_mask_0 = const()[name = string("z_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_53_cast_fp16 = slice_by_index(begin = z_53_begin_0, end = z_53_end_0, end_mask = z_53_end_mask_0, x = x_119_cast_fp16)[name = string("z_53_cast_fp16")];
            tensor<int32, [4]> var_3695_begin_0 = const()[name = string("op_3695_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3695_end_0 = const()[name = string("op_3695_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3695_end_mask_0 = const()[name = string("op_3695_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3695_cast_fp16 = slice_by_index(begin = var_3695_begin_0, end = var_3695_end_0, end_mask = var_3695_end_mask_0, x = z_53_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3696_cast_fp16 = mul(x = var_3695_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_3696_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, 2520, 12, 16])];
            tensor<bool, [4]> var_3701_end_mask_0 = const()[name = string("op_3701_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> 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 = z_53_cast_fp16)[name = string("op_3701_cast_fp16")];
            int32 var_3703 = const()[name = string("op_3703"), val = int32(-1)];
            bool rot_53_interleave_0 = const()[name = string("rot_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_53_cast_fp16 = concat(axis = var_3703, interleave = rot_53_interleave_0, values = (var_3696_cast_fp16, var_3701_cast_fp16))[name = string("rot_53_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3705_cast_fp16 = mul(x = z_53_cast_fp16, y = c_1_cast_fp16)[name = string("op_3705_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3706_cast_fp16 = mul(x = rot_53_cast_fp16, y = s_9_cast_fp16)[name = string("op_3706_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3708_cast_fp16 = add(x = var_3705_cast_fp16, y = var_3706_cast_fp16)[name = string("op_3708_cast_fp16")];
            tensor<int32, [4]> z_55_begin_0 = const()[name = string("z_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_55_end_0 = const()[name = string("z_55_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_55_end_mask_0 = const()[name = string("z_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_55_cast_fp16 = slice_by_index(begin = z_55_begin_0, end = z_55_end_0, end_mask = z_55_end_mask_0, x = x_119_cast_fp16)[name = string("z_55_cast_fp16")];
            tensor<int32, [4]> var_3718_begin_0 = const()[name = string("op_3718_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3718_end_0 = const()[name = string("op_3718_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3718_end_mask_0 = const()[name = string("op_3718_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3718_cast_fp16 = slice_by_index(begin = var_3718_begin_0, end = var_3718_end_0, end_mask = var_3718_end_mask_0, x = z_55_cast_fp16)[name = string("op_3718_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3719_cast_fp16 = mul(x = var_3718_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_3719_cast_fp16")];
            tensor<int32, [4]> var_3724_begin_0 = const()[name = string("op_3724_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3724_end_0 = const()[name = string("op_3724_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3724_end_mask_0 = const()[name = string("op_3724_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, x = z_55_cast_fp16)[name = string("op_3724_cast_fp16")];
            int32 var_3726 = const()[name = string("op_3726"), val = int32(-1)];
            bool rot_55_interleave_0 = const()[name = string("rot_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_55_cast_fp16 = concat(axis = var_3726, interleave = rot_55_interleave_0, values = (var_3719_cast_fp16, var_3724_cast_fp16))[name = string("rot_55_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3728_cast_fp16 = mul(x = z_55_cast_fp16, y = c_cast_fp16)[name = string("op_3728_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3729_cast_fp16 = mul(x = rot_55_cast_fp16, y = s_11_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3731_cast_fp16 = add(x = var_3728_cast_fp16, y = var_3729_cast_fp16)[name = string("op_3731_cast_fp16")];
            int32 var_3733 = const()[name = string("op_3733"), val = int32(-1)];
            bool var_3734_interleave_0 = const()[name = string("op_3734_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_3734_cast_fp16 = concat(axis = var_3733, interleave = var_3734_interleave_0, values = (var_3708_cast_fp16, var_3731_cast_fp16))[name = string("op_3734_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3744_transpose_x_0 = const()[name = string("op_3744_transpose_x_0"), val = bool(false)];
            bool var_3744_transpose_y_0 = const()[name = string("op_3744_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3734_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_76 = transpose(perm = transpose_76_perm_0, x = var_3682_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 12, 2520, 2520]> var_3744_cast_fp16 = matmul(transpose_x = var_3744_transpose_x_0, transpose_y = var_3744_transpose_y_0, x = transpose_76, y = transpose_77)[name = string("op_3744_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_3746_cast_fp16 = add(x = var_3744_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3746_cast_fp16")];
            int32 var_3747 = const()[name = string("op_3747"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_91_cast_fp16 = softmax(axis = var_3747, x = var_3746_cast_fp16)[name = string("w_91_cast_fp16")];
            bool var_3750_transpose_x_0 = const()[name = string("op_3750_transpose_x_0"), val = bool(false)];
            bool var_3750_transpose_y_0 = const()[name = string("op_3750_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = v_25_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 12, 2520, 64]> var_3750_cast_fp16 = matmul(transpose_x = var_3750_transpose_x_0, transpose_y = var_3750_transpose_y_0, x = w_91_cast_fp16, y = v_27_cast_fp16)[name = string("op_3750_cast_fp16")];
            tensor<int32, [4]> var_3753_perm_0 = const()[name = string("op_3753_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3757 = const()[name = string("op_3757"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_3753_cast_fp16 = transpose(perm = var_3753_perm_0, x = var_3750_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 2520, 768]> input_65_cast_fp16 = reshape(shape = var_3757, x = var_3753_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59164160))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59754048))))[name = string("t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3762_cast_fp16 = abs(x = linear_46_cast_fp16)[name = string("op_3762_cast_fp16")];
            bool var_3765 = const()[name = string("op_3765"), val = bool(true)];
            tensor<int32, [1]> var_3766_axes_0 = const()[name = string("op_3766_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3766_cast_fp16 = reduce_max(axes = var_3766_axes_0, keep_dims = var_3765, x = var_3762_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_97_cast_fp16 = maximum(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_93_cast_fp16 = real_div(x = linear_46_cast_fp16, y = s_97_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3770_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_3770_cast_fp16")];
            tensor<int32, [1]> var_3775_axes_0 = const()[name = string("op_3775_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3775_keep_dims_0 = const()[name = string("op_3775_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3775_cast_fp16 = reduce_mean(axes = var_3775_axes_0, keep_dims = var_3775_keep_dims_0, x = var_3770_cast_fp16)[name = string("op_3775_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3776_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_3776_cast_fp16")];
            fp32 var_3777_epsilon_0 = const()[name = string("op_3777_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3777_cast_fp16 = inverse(epsilon = var_3777_epsilon_0, x = var_3776_cast_fp16)[name = string("op_3777_cast_fp16")];
            fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3779_cast_fp16 = mul(x = var_3777_cast_fp16, y = var_3778_to_fp16)[name = string("op_3779_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3781_cast_fp16 = add(x = var_3775_cast_fp16, y = var_3779_cast_fp16)[name = string("op_3781_cast_fp16")];
            fp32 var_3782_epsilon_0 = const()[name = string("op_3782_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3782_cast_fp16 = rsqrt(epsilon = var_3782_epsilon_0, x = var_3781_cast_fp16)[name = string("op_3782_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_79_cast_fp16 = mul(x = y_93_cast_fp16, y = var_3782_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59755648)))];
            tensor<fp16, [1, 2520, 768]> var_3784_cast_fp16 = mul(x = out_79_cast_fp16, y = t_encoder_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_123_cast_fp16 = add(x = x_109_cast_fp16, y = var_3784_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3787_cast_fp16 = abs(x = x_123_cast_fp16)[name = string("op_3787_cast_fp16")];
            bool var_3790 = const()[name = string("op_3790"), val = bool(true)];
            tensor<int32, [1]> var_3791_axes_0 = const()[name = string("op_3791_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3791_cast_fp16 = reduce_max(axes = var_3791_axes_0, keep_dims = var_3790, x = var_3787_cast_fp16)[name = string("op_3791_cast_fp16")];
            fp16 var_3792_to_fp16 = const()[name = string("op_3792_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_99_cast_fp16 = maximum(x = var_3791_cast_fp16, y = var_3792_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_95_cast_fp16 = real_div(x = x_123_cast_fp16, y = s_99_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3795_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_3795_cast_fp16")];
            tensor<int32, [1]> var_3800_axes_0 = const()[name = string("op_3800_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3800_keep_dims_0 = const()[name = string("op_3800_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3800_cast_fp16 = reduce_mean(axes = var_3800_axes_0, keep_dims = var_3800_keep_dims_0, x = var_3795_cast_fp16)[name = string("op_3800_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3801_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_3801_cast_fp16")];
            fp32 var_3802_epsilon_0 = const()[name = string("op_3802_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3802_cast_fp16 = inverse(epsilon = var_3802_epsilon_0, x = var_3801_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 var_3803_to_fp16 = const()[name = string("op_3803_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16, y = var_3803_to_fp16)[name = string("op_3804_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3806_cast_fp16 = add(x = var_3800_cast_fp16, y = var_3804_cast_fp16)[name = string("op_3806_cast_fp16")];
            fp32 var_3807_epsilon_0 = const()[name = string("op_3807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3807_cast_fp16 = rsqrt(epsilon = var_3807_epsilon_0, x = var_3806_cast_fp16)[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_81_cast_fp16 = mul(x = y_95_cast_fp16, y = var_3807_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59757248)))];
            tensor<fp16, [1, 2520, 768]> input_67_cast_fp16 = mul(x = out_81_cast_fp16, y = t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59758848))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62118208))))[name = string("t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_47_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_47_cast_fp16")];
            string var_3814_mode_0 = const()[name = string("op_3814_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_3814_cast_fp16 = gelu(mode = var_3814_mode_0, x = linear_47_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62124416))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64483776))))[name = string("t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_48_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_71_cast_fp16 = mul(x = var_3814_cast_fp16, y = linear_48_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64489984))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66849344))))[name = string("t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3822_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_3822_cast_fp16")];
            bool var_3825 = const()[name = string("op_3825"), val = bool(true)];
            tensor<int32, [1]> var_3826_axes_0 = const()[name = string("op_3826_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3826_cast_fp16 = reduce_max(axes = var_3826_axes_0, keep_dims = var_3825, x = var_3822_cast_fp16)[name = string("op_3826_cast_fp16")];
            fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_101_cast_fp16 = maximum(x = var_3826_cast_fp16, y = var_3827_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_97_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_101_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3830_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_3830_cast_fp16")];
            tensor<int32, [1]> var_3835_axes_0 = const()[name = string("op_3835_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3835_keep_dims_0 = const()[name = string("op_3835_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3835_cast_fp16 = reduce_mean(axes = var_3835_axes_0, keep_dims = var_3835_keep_dims_0, x = var_3830_cast_fp16)[name = string("op_3835_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3836_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_3836_cast_fp16")];
            fp32 var_3837_epsilon_0 = const()[name = string("op_3837_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3837_cast_fp16 = inverse(epsilon = var_3837_epsilon_0, x = var_3836_cast_fp16)[name = string("op_3837_cast_fp16")];
            fp16 var_3838_to_fp16 = const()[name = string("op_3838_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3839_cast_fp16 = mul(x = var_3837_cast_fp16, y = var_3838_to_fp16)[name = string("op_3839_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3841_cast_fp16 = add(x = var_3835_cast_fp16, y = var_3839_cast_fp16)[name = string("op_3841_cast_fp16")];
            fp32 var_3842_epsilon_0 = const()[name = string("op_3842_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3842_cast_fp16 = rsqrt(epsilon = var_3842_epsilon_0, x = var_3841_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_83_cast_fp16 = mul(x = y_97_cast_fp16, y = var_3842_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66850944)))];
            tensor<fp16, [1, 2520, 768]> var_3844_cast_fp16 = mul(x = out_83_cast_fp16, y = t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3844_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_127_cast_fp16 = add(x = x_123_cast_fp16, y = var_3844_cast_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3847_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_3847_cast_fp16")];
            bool var_3850 = const()[name = string("op_3850"), val = bool(true)];
            tensor<int32, [1]> var_3851_axes_0 = const()[name = string("op_3851_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_3851_cast_fp16 = reduce_max(axes = var_3851_axes_0, keep_dims = var_3850, x = var_3847_cast_fp16)[name = string("op_3851_cast_fp16")];
            fp16 var_3852_to_fp16 = const()[name = string("op_3852_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_103_cast_fp16 = maximum(x = var_3851_cast_fp16, y = var_3852_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_99_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_103_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_3855_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_3855_cast_fp16")];
            tensor<int32, [1]> var_3860_axes_0 = const()[name = string("op_3860_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3860_keep_dims_0 = const()[name = string("op_3860_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_3860_cast_fp16 = reduce_mean(axes = var_3860_axes_0, keep_dims = var_3860_keep_dims_0, x = var_3855_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3861_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_3861_cast_fp16")];
            fp32 var_3862_epsilon_0 = const()[name = string("op_3862_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_3862_cast_fp16 = inverse(epsilon = var_3862_epsilon_0, x = var_3861_cast_fp16)[name = string("op_3862_cast_fp16")];
            fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_3864_cast_fp16 = mul(x = var_3862_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_3866_cast_fp16 = add(x = var_3860_cast_fp16, y = var_3864_cast_fp16)[name = string("op_3866_cast_fp16")];
            fp32 var_3867_epsilon_0 = const()[name = string("op_3867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_3867_cast_fp16 = rsqrt(epsilon = var_3867_epsilon_0, x = var_3866_cast_fp16)[name = string("op_3867_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_85_cast_fp16 = mul(x = y_99_cast_fp16, y = var_3867_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66852544)))];
            tensor<fp16, [1, 2520, 768]> input_73_cast_fp16 = mul(x = out_85_cast_fp16, y = t_encoder_layers_7_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66854144))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67444032))))[name = string("t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3877 = const()[name = string("op_3877"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_129_cast_fp16 = reshape(shape = var_3877, x = linear_50_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3879_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_3879_cast_fp16")];
            bool var_3882 = const()[name = string("op_3882"), val = bool(true)];
            tensor<int32, [1]> var_3883_axes_0 = const()[name = string("op_3883_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3883_cast_fp16 = reduce_max(axes = var_3883_axes_0, keep_dims = var_3882, x = var_3879_cast_fp16)[name = string("op_3883_cast_fp16")];
            fp16 var_3884_to_fp16 = const()[name = string("op_3884_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_105_cast_fp16 = maximum(x = var_3883_cast_fp16, y = var_3884_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_101_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_105_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3887_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_3887_cast_fp16")];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3892_keep_dims_0 = const()[name = string("op_3892_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3892_cast_fp16 = reduce_mean(axes = var_3892_axes_0, keep_dims = var_3892_keep_dims_0, x = var_3887_cast_fp16)[name = string("op_3892_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3893_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_3893_cast_fp16")];
            fp32 var_3894_epsilon_0 = const()[name = string("op_3894_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3894_cast_fp16 = inverse(epsilon = var_3894_epsilon_0, x = var_3893_cast_fp16)[name = string("op_3894_cast_fp16")];
            fp16 var_3895_to_fp16 = const()[name = string("op_3895_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3896_cast_fp16 = mul(x = var_3894_cast_fp16, y = var_3895_to_fp16)[name = string("op_3896_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3898_cast_fp16 = add(x = var_3892_cast_fp16, y = var_3896_cast_fp16)[name = string("op_3898_cast_fp16")];
            fp32 var_3899_epsilon_0 = const()[name = string("op_3899_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3899_cast_fp16 = rsqrt(epsilon = var_3899_epsilon_0, x = var_3898_cast_fp16)[name = string("op_3899_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_87_cast_fp16 = mul(x = y_101_cast_fp16, y = var_3899_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 2520, 12, 64]> x_135_cast_fp16 = mul(x = out_87_cast_fp16, y = t_encoder_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_135_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67445632))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68035520))))[name = string("t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_51_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3909 = const()[name = string("op_3909"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_131_cast_fp16 = reshape(shape = var_3909, x = linear_51_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3911_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_107_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_103_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_107_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3919_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3925_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_89_cast_fp16 = mul(x = y_103_cast_fp16, y = var_3931_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 2520, 12, 64]> x_137_cast_fp16 = mul(x = out_89_cast_fp16, y = t_encoder_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68037120))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68627008))))[name = string("t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [4]> var_3941 = const()[name = string("op_3941"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_133_cast_fp16 = reshape(shape = var_3941, x = linear_52_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3943_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_3943_cast_fp16")];
            bool var_3946 = const()[name = string("op_3946"), val = bool(true)];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_3947_cast_fp16 = reduce_max(axes = var_3947_axes_0, keep_dims = var_3946, x = var_3943_cast_fp16)[name = string("op_3947_cast_fp16")];
            fp16 var_3948_to_fp16 = const()[name = string("op_3948_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_109_cast_fp16 = maximum(x = var_3947_cast_fp16, y = var_3948_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_105_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_109_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_3951_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_3951_cast_fp16")];
            tensor<int32, [1]> var_3956_axes_0 = const()[name = string("op_3956_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3956_keep_dims_0 = const()[name = string("op_3956_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_3956_cast_fp16 = reduce_mean(axes = var_3956_axes_0, keep_dims = var_3956_keep_dims_0, x = var_3951_cast_fp16)[name = string("op_3956_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3957_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_3957_cast_fp16")];
            fp32 var_3958_epsilon_0 = const()[name = string("op_3958_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_3958_cast_fp16 = inverse(epsilon = var_3958_epsilon_0, x = var_3957_cast_fp16)[name = string("op_3958_cast_fp16")];
            fp16 var_3959_to_fp16 = const()[name = string("op_3959_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_3960_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = string("op_3960_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_3962_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3960_cast_fp16)[name = string("op_3962_cast_fp16")];
            fp32 var_3963_epsilon_0 = const()[name = string("op_3963_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_3963_cast_fp16 = rsqrt(epsilon = var_3963_epsilon_0, x = var_3962_cast_fp16)[name = string("op_3963_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_29_cast_fp16 = mul(x = y_105_cast_fp16, y = var_3963_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<int32, [4]> z_57_begin_0 = const()[name = string("z_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_57_end_0 = const()[name = string("z_57_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_57_end_mask_0 = const()[name = string("z_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_57_cast_fp16 = slice_by_index(begin = z_57_begin_0, end = z_57_end_0, end_mask = z_57_end_mask_0, x = x_135_cast_fp16)[name = string("z_57_cast_fp16")];
            tensor<int32, [4]> var_3974_begin_0 = const()[name = string("op_3974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3974_end_0 = const()[name = string("op_3974_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3974_end_mask_0 = const()[name = string("op_3974_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> 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 = z_57_cast_fp16)[name = string("op_3974_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3975_cast_fp16 = mul(x = var_3974_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_3975_cast_fp16")];
            tensor<int32, [4]> var_3980_begin_0 = const()[name = string("op_3980_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3980_end_0 = const()[name = string("op_3980_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_3980_end_mask_0 = const()[name = string("op_3980_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_3980_cast_fp16 = slice_by_index(begin = var_3980_begin_0, end = var_3980_end_0, end_mask = var_3980_end_mask_0, x = z_57_cast_fp16)[name = string("op_3980_cast_fp16")];
            int32 var_3982 = const()[name = string("op_3982"), val = int32(-1)];
            bool rot_57_interleave_0 = const()[name = string("rot_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_57_cast_fp16 = concat(axis = var_3982, interleave = rot_57_interleave_0, values = (var_3975_cast_fp16, var_3980_cast_fp16))[name = string("rot_57_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3984_cast_fp16 = mul(x = z_57_cast_fp16, y = c_1_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3985_cast_fp16 = mul(x = rot_57_cast_fp16, y = s_9_cast_fp16)[name = string("op_3985_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_3987_cast_fp16 = add(x = var_3984_cast_fp16, y = var_3985_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<int32, [4]> z_59_begin_0 = const()[name = string("z_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_59_end_0 = const()[name = string("z_59_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_59_end_mask_0 = const()[name = string("z_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_59_cast_fp16 = slice_by_index(begin = z_59_begin_0, end = z_59_end_0, end_mask = z_59_end_mask_0, x = x_135_cast_fp16)[name = string("z_59_cast_fp16")];
            tensor<int32, [4]> var_3997_begin_0 = const()[name = string("op_3997_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3997_end_0 = const()[name = string("op_3997_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_3997_end_mask_0 = const()[name = string("op_3997_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, x = z_59_cast_fp16)[name = string("op_3997_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_3998_cast_fp16 = mul(x = var_3997_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_3998_cast_fp16")];
            tensor<int32, [4]> var_4003_begin_0 = const()[name = string("op_4003_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4003_end_0 = const()[name = string("op_4003_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4003_end_mask_0 = const()[name = string("op_4003_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4003_cast_fp16 = slice_by_index(begin = var_4003_begin_0, end = var_4003_end_0, end_mask = var_4003_end_mask_0, x = z_59_cast_fp16)[name = string("op_4003_cast_fp16")];
            int32 var_4005 = const()[name = string("op_4005"), val = int32(-1)];
            bool rot_59_interleave_0 = const()[name = string("rot_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_59_cast_fp16 = concat(axis = var_4005, interleave = rot_59_interleave_0, values = (var_3998_cast_fp16, var_4003_cast_fp16))[name = string("rot_59_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4007_cast_fp16 = mul(x = z_59_cast_fp16, y = c_cast_fp16)[name = string("op_4007_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4008_cast_fp16 = mul(x = rot_59_cast_fp16, y = s_11_cast_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4010_cast_fp16 = add(x = var_4007_cast_fp16, y = var_4008_cast_fp16)[name = string("op_4010_cast_fp16")];
            int32 var_4012 = const()[name = string("op_4012"), val = int32(-1)];
            bool var_4013_interleave_0 = const()[name = string("op_4013_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_4013_cast_fp16 = concat(axis = var_4012, interleave = var_4013_interleave_0, values = (var_3987_cast_fp16, var_4010_cast_fp16))[name = string("op_4013_cast_fp16")];
            tensor<int32, [4]> z_61_begin_0 = const()[name = string("z_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_61_end_0 = const()[name = string("z_61_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_61_end_mask_0 = const()[name = string("z_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_61_cast_fp16 = slice_by_index(begin = z_61_begin_0, end = z_61_end_0, end_mask = z_61_end_mask_0, x = x_137_cast_fp16)[name = string("z_61_cast_fp16")];
            tensor<int32, [4]> var_4026_begin_0 = const()[name = string("op_4026_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4026_end_0 = const()[name = string("op_4026_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4026_end_mask_0 = const()[name = string("op_4026_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4026_cast_fp16 = slice_by_index(begin = var_4026_begin_0, end = var_4026_end_0, end_mask = var_4026_end_mask_0, x = z_61_cast_fp16)[name = string("op_4026_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_4027_cast_fp16")];
            tensor<int32, [4]> var_4032_begin_0 = const()[name = string("op_4032_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4032_end_0 = const()[name = string("op_4032_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4032_end_mask_0 = const()[name = string("op_4032_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4032_cast_fp16 = slice_by_index(begin = var_4032_begin_0, end = var_4032_end_0, end_mask = var_4032_end_mask_0, x = z_61_cast_fp16)[name = string("op_4032_cast_fp16")];
            int32 var_4034 = const()[name = string("op_4034"), val = int32(-1)];
            bool rot_61_interleave_0 = const()[name = string("rot_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_61_cast_fp16 = concat(axis = var_4034, interleave = rot_61_interleave_0, values = (var_4027_cast_fp16, var_4032_cast_fp16))[name = string("rot_61_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4036_cast_fp16 = mul(x = z_61_cast_fp16, y = c_1_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4037_cast_fp16 = mul(x = rot_61_cast_fp16, y = s_9_cast_fp16)[name = string("op_4037_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4039_cast_fp16 = add(x = var_4036_cast_fp16, y = var_4037_cast_fp16)[name = string("op_4039_cast_fp16")];
            tensor<int32, [4]> z_63_begin_0 = const()[name = string("z_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_63_end_0 = const()[name = string("z_63_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_63_end_mask_0 = const()[name = string("z_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_63_cast_fp16 = slice_by_index(begin = z_63_begin_0, end = z_63_end_0, end_mask = z_63_end_mask_0, x = x_137_cast_fp16)[name = string("z_63_cast_fp16")];
            tensor<int32, [4]> var_4049_begin_0 = const()[name = string("op_4049_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4049_end_0 = const()[name = string("op_4049_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4049_end_mask_0 = const()[name = string("op_4049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4049_cast_fp16 = slice_by_index(begin = var_4049_begin_0, end = var_4049_end_0, end_mask = var_4049_end_mask_0, x = z_63_cast_fp16)[name = string("op_4049_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4050_cast_fp16 = mul(x = var_4049_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_4050_cast_fp16")];
            tensor<int32, [4]> var_4055_begin_0 = const()[name = string("op_4055_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4055_end_0 = const()[name = string("op_4055_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4055_end_mask_0 = const()[name = string("op_4055_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4055_cast_fp16 = slice_by_index(begin = var_4055_begin_0, end = var_4055_end_0, end_mask = var_4055_end_mask_0, x = z_63_cast_fp16)[name = string("op_4055_cast_fp16")];
            int32 var_4057 = const()[name = string("op_4057"), val = int32(-1)];
            bool rot_63_interleave_0 = const()[name = string("rot_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_63_cast_fp16 = concat(axis = var_4057, interleave = rot_63_interleave_0, values = (var_4050_cast_fp16, var_4055_cast_fp16))[name = string("rot_63_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4059_cast_fp16 = mul(x = z_63_cast_fp16, y = c_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4060_cast_fp16 = mul(x = rot_63_cast_fp16, y = s_11_cast_fp16)[name = string("op_4060_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4062_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4060_cast_fp16)[name = string("op_4062_cast_fp16")];
            int32 var_4064 = const()[name = string("op_4064"), val = int32(-1)];
            bool var_4065_interleave_0 = const()[name = string("op_4065_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_4065_cast_fp16 = concat(axis = var_4064, interleave = var_4065_interleave_0, values = (var_4039_cast_fp16, var_4062_cast_fp16))[name = string("op_4065_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4075_transpose_x_0 = const()[name = string("op_4075_transpose_x_0"), val = bool(false)];
            bool var_4075_transpose_y_0 = const()[name = string("op_4075_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_4065_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_78 = transpose(perm = transpose_78_perm_0, x = var_4013_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 12, 2520, 2520]> var_4075_cast_fp16 = matmul(transpose_x = var_4075_transpose_x_0, transpose_y = var_4075_transpose_y_0, x = transpose_78, y = transpose_79)[name = string("op_4075_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_4077_cast_fp16 = add(x = var_4075_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4077_cast_fp16")];
            int32 var_4078 = const()[name = string("op_4078"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_105_cast_fp16 = softmax(axis = var_4078, x = var_4077_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4081_transpose_x_0 = const()[name = string("op_4081_transpose_x_0"), val = bool(false)];
            bool var_4081_transpose_y_0 = const()[name = string("op_4081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = v_29_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 12, 2520, 64]> var_4081_cast_fp16 = matmul(transpose_x = var_4081_transpose_x_0, transpose_y = var_4081_transpose_y_0, x = w_105_cast_fp16, y = v_31_cast_fp16)[name = string("op_4081_cast_fp16")];
            tensor<int32, [4]> var_4084_perm_0 = const()[name = string("op_4084_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4088 = const()[name = string("op_4088"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_4084_cast_fp16 = transpose(perm = var_4084_perm_0, x = var_4081_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 2520, 768]> input_75_cast_fp16 = reshape(shape = var_4088, x = var_4084_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68628608))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69218496))))[name = string("t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_53_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4093_cast_fp16 = abs(x = linear_53_cast_fp16)[name = string("op_4093_cast_fp16")];
            bool var_4096 = const()[name = string("op_4096"), val = bool(true)];
            tensor<int32, [1]> var_4097_axes_0 = const()[name = string("op_4097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4097_cast_fp16 = reduce_max(axes = var_4097_axes_0, keep_dims = var_4096, x = var_4093_cast_fp16)[name = string("op_4097_cast_fp16")];
            fp16 var_4098_to_fp16 = const()[name = string("op_4098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_111_cast_fp16 = maximum(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_107_cast_fp16 = real_div(x = linear_53_cast_fp16, y = s_111_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4101_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<int32, [1]> var_4106_axes_0 = const()[name = string("op_4106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4106_keep_dims_0 = const()[name = string("op_4106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4106_cast_fp16 = reduce_mean(axes = var_4106_axes_0, keep_dims = var_4106_keep_dims_0, x = var_4101_cast_fp16)[name = string("op_4106_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4107_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4107_cast_fp16")];
            fp32 var_4108_epsilon_0 = const()[name = string("op_4108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4108_cast_fp16 = inverse(epsilon = var_4108_epsilon_0, x = var_4107_cast_fp16)[name = string("op_4108_cast_fp16")];
            fp16 var_4109_to_fp16 = const()[name = string("op_4109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4110_cast_fp16 = mul(x = var_4108_cast_fp16, y = var_4109_to_fp16)[name = string("op_4110_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4112_cast_fp16 = add(x = var_4106_cast_fp16, y = var_4110_cast_fp16)[name = string("op_4112_cast_fp16")];
            fp32 var_4113_epsilon_0 = const()[name = string("op_4113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4113_cast_fp16 = rsqrt(epsilon = var_4113_epsilon_0, x = var_4112_cast_fp16)[name = string("op_4113_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_91_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4113_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69220096)))];
            tensor<fp16, [1, 2520, 768]> var_4115_cast_fp16 = mul(x = out_91_cast_fp16, y = t_encoder_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4115_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_141_cast_fp16 = add(x = x_127_cast_fp16, y = var_4115_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4118_cast_fp16 = abs(x = x_141_cast_fp16)[name = string("op_4118_cast_fp16")];
            bool var_4121 = const()[name = string("op_4121"), val = bool(true)];
            tensor<int32, [1]> var_4122_axes_0 = const()[name = string("op_4122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4122_cast_fp16 = reduce_max(axes = var_4122_axes_0, keep_dims = var_4121, x = var_4118_cast_fp16)[name = string("op_4122_cast_fp16")];
            fp16 var_4123_to_fp16 = const()[name = string("op_4123_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_113_cast_fp16 = maximum(x = var_4122_cast_fp16, y = var_4123_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_109_cast_fp16 = real_div(x = x_141_cast_fp16, y = s_113_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4126_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4126_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, 2520, 1]> var_4131_cast_fp16 = reduce_mean(axes = var_4131_axes_0, keep_dims = var_4131_keep_dims_0, x = var_4126_cast_fp16)[name = string("op_4131_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4132_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_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, 2520, 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, 2520, 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, 2520, 1]> var_4137_cast_fp16 = add(x = var_4131_cast_fp16, y = var_4135_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp32 var_4138_epsilon_0 = const()[name = string("op_4138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4138_cast_fp16 = rsqrt(epsilon = var_4138_epsilon_0, x = var_4137_cast_fp16)[name = string("op_4138_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_93_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4138_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69221696)))];
            tensor<fp16, [1, 2520, 768]> input_77_cast_fp16 = mul(x = out_93_cast_fp16, y = t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69223296))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71582656))))[name = string("t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_54_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_54_cast_fp16")];
            string var_4145_mode_0 = const()[name = string("op_4145_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_4145_cast_fp16 = gelu(mode = var_4145_mode_0, x = linear_54_cast_fp16)[name = string("op_4145_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71588864))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73948224))))[name = string("t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_55_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_81_cast_fp16 = mul(x = var_4145_cast_fp16, y = linear_55_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73954432))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76313792))))[name = string("t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4153_cast_fp16 = abs(x = linear_56_cast_fp16)[name = string("op_4153_cast_fp16")];
            bool var_4156 = const()[name = string("op_4156"), val = bool(true)];
            tensor<int32, [1]> var_4157_axes_0 = const()[name = string("op_4157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4157_cast_fp16 = reduce_max(axes = var_4157_axes_0, keep_dims = var_4156, x = var_4153_cast_fp16)[name = string("op_4157_cast_fp16")];
            fp16 var_4158_to_fp16 = const()[name = string("op_4158_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_115_cast_fp16 = maximum(x = var_4157_cast_fp16, y = var_4158_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_111_cast_fp16 = real_div(x = linear_56_cast_fp16, y = s_115_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4161_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4161_cast_fp16")];
            tensor<int32, [1]> var_4166_axes_0 = const()[name = string("op_4166_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4166_keep_dims_0 = const()[name = string("op_4166_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4166_cast_fp16 = reduce_mean(axes = var_4166_axes_0, keep_dims = var_4166_keep_dims_0, x = var_4161_cast_fp16)[name = string("op_4166_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4167_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4167_cast_fp16")];
            fp32 var_4168_epsilon_0 = const()[name = string("op_4168_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4168_cast_fp16 = inverse(epsilon = var_4168_epsilon_0, x = var_4167_cast_fp16)[name = string("op_4168_cast_fp16")];
            fp16 var_4169_to_fp16 = const()[name = string("op_4169_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4170_cast_fp16 = mul(x = var_4168_cast_fp16, y = var_4169_to_fp16)[name = string("op_4170_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4172_cast_fp16 = add(x = var_4166_cast_fp16, y = var_4170_cast_fp16)[name = string("op_4172_cast_fp16")];
            fp32 var_4173_epsilon_0 = const()[name = string("op_4173_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4173_cast_fp16 = rsqrt(epsilon = var_4173_epsilon_0, x = var_4172_cast_fp16)[name = string("op_4173_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_95_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4173_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76315392)))];
            tensor<fp16, [1, 2520, 768]> var_4175_cast_fp16 = mul(x = out_95_cast_fp16, y = t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4175_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_145_cast_fp16 = add(x = x_141_cast_fp16, y = var_4175_cast_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4178_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4178_cast_fp16")];
            bool var_4181 = const()[name = string("op_4181"), val = bool(true)];
            tensor<int32, [1]> var_4182_axes_0 = const()[name = string("op_4182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4182_cast_fp16 = reduce_max(axes = var_4182_axes_0, keep_dims = var_4181, x = var_4178_cast_fp16)[name = string("op_4182_cast_fp16")];
            fp16 var_4183_to_fp16 = const()[name = string("op_4183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_117_cast_fp16 = maximum(x = var_4182_cast_fp16, y = var_4183_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_113_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_117_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4186_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4186_cast_fp16")];
            tensor<int32, [1]> var_4191_axes_0 = const()[name = string("op_4191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4191_keep_dims_0 = const()[name = string("op_4191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4191_cast_fp16 = reduce_mean(axes = var_4191_axes_0, keep_dims = var_4191_keep_dims_0, x = var_4186_cast_fp16)[name = string("op_4191_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4192_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4192_cast_fp16")];
            fp32 var_4193_epsilon_0 = const()[name = string("op_4193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4193_cast_fp16 = inverse(epsilon = var_4193_epsilon_0, x = var_4192_cast_fp16)[name = string("op_4193_cast_fp16")];
            fp16 var_4194_to_fp16 = const()[name = string("op_4194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4195_cast_fp16 = mul(x = var_4193_cast_fp16, y = var_4194_to_fp16)[name = string("op_4195_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4197_cast_fp16 = add(x = var_4191_cast_fp16, y = var_4195_cast_fp16)[name = string("op_4197_cast_fp16")];
            fp32 var_4198_epsilon_0 = const()[name = string("op_4198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4198_cast_fp16 = rsqrt(epsilon = var_4198_epsilon_0, x = var_4197_cast_fp16)[name = string("op_4198_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_97_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4198_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76316992)))];
            tensor<fp16, [1, 2520, 768]> input_83_cast_fp16 = mul(x = out_97_cast_fp16, y = t_encoder_layers_8_input_layernorm_weight_to_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76318592))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76908480))))[name = string("t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_57_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4208 = const()[name = string("op_4208"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_147_cast_fp16 = reshape(shape = var_4208, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4210_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4210_cast_fp16")];
            bool var_4213 = const()[name = string("op_4213"), val = bool(true)];
            tensor<int32, [1]> var_4214_axes_0 = const()[name = string("op_4214_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4214_cast_fp16 = reduce_max(axes = var_4214_axes_0, keep_dims = var_4213, x = var_4210_cast_fp16)[name = string("op_4214_cast_fp16")];
            fp16 var_4215_to_fp16 = const()[name = string("op_4215_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_119_cast_fp16 = maximum(x = var_4214_cast_fp16, y = var_4215_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_115_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_119_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4218_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4218_cast_fp16")];
            tensor<int32, [1]> var_4223_axes_0 = const()[name = string("op_4223_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4223_keep_dims_0 = const()[name = string("op_4223_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4223_cast_fp16 = reduce_mean(axes = var_4223_axes_0, keep_dims = var_4223_keep_dims_0, x = var_4218_cast_fp16)[name = string("op_4223_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4224_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4224_cast_fp16")];
            fp32 var_4225_epsilon_0 = const()[name = string("op_4225_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4225_cast_fp16 = inverse(epsilon = var_4225_epsilon_0, x = var_4224_cast_fp16)[name = string("op_4225_cast_fp16")];
            fp16 var_4226_to_fp16 = const()[name = string("op_4226_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4227_cast_fp16 = mul(x = var_4225_cast_fp16, y = var_4226_to_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4229_cast_fp16 = add(x = var_4223_cast_fp16, y = var_4227_cast_fp16)[name = string("op_4229_cast_fp16")];
            fp32 var_4230_epsilon_0 = const()[name = string("op_4230_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4230_cast_fp16 = rsqrt(epsilon = var_4230_epsilon_0, x = var_4229_cast_fp16)[name = string("op_4230_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_99_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4230_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910080)))];
            tensor<fp16, [1, 2520, 12, 64]> x_153_cast_fp16 = mul(x = out_99_cast_fp16, y = t_encoder_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910272))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77500160))))[name = string("t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_4240 = const()[name = string("op_4240"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_149_cast_fp16 = reshape(shape = var_4240, x = linear_58_cast_fp16)[name = string("x_149_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4242_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4242_cast_fp16")];
            bool var_4245 = const()[name = string("op_4245"), val = bool(true)];
            tensor<int32, [1]> var_4246_axes_0 = const()[name = string("op_4246_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4246_cast_fp16 = reduce_max(axes = var_4246_axes_0, keep_dims = var_4245, x = var_4242_cast_fp16)[name = string("op_4246_cast_fp16")];
            fp16 var_4247_to_fp16 = const()[name = string("op_4247_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_121_cast_fp16 = maximum(x = var_4246_cast_fp16, y = var_4247_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_117_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_121_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4250_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<int32, [1]> var_4255_axes_0 = const()[name = string("op_4255_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4255_keep_dims_0 = const()[name = string("op_4255_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4255_cast_fp16 = reduce_mean(axes = var_4255_axes_0, keep_dims = var_4255_keep_dims_0, x = var_4250_cast_fp16)[name = string("op_4255_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4256_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp32 var_4257_epsilon_0 = const()[name = string("op_4257_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4257_cast_fp16 = inverse(epsilon = var_4257_epsilon_0, x = var_4256_cast_fp16)[name = string("op_4257_cast_fp16")];
            fp16 var_4258_to_fp16 = const()[name = string("op_4258_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4259_cast_fp16 = mul(x = var_4257_cast_fp16, y = var_4258_to_fp16)[name = string("op_4259_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4261_cast_fp16 = add(x = var_4255_cast_fp16, y = var_4259_cast_fp16)[name = string("op_4261_cast_fp16")];
            fp32 var_4262_epsilon_0 = const()[name = string("op_4262_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4262_cast_fp16 = rsqrt(epsilon = var_4262_epsilon_0, x = var_4261_cast_fp16)[name = string("op_4262_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_101_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4262_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501760)))];
            tensor<fp16, [1, 2520, 12, 64]> x_155_cast_fp16 = mul(x = out_101_cast_fp16, y = t_encoder_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501952))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78091840))))[name = string("t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_59_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<int32, [4]> var_4272 = const()[name = string("op_4272"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_151_cast_fp16 = reshape(shape = var_4272, x = linear_59_cast_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4274_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4274_cast_fp16")];
            bool var_4277 = const()[name = string("op_4277"), val = bool(true)];
            tensor<int32, [1]> var_4278_axes_0 = const()[name = string("op_4278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4278_cast_fp16 = reduce_max(axes = var_4278_axes_0, keep_dims = var_4277, x = var_4274_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp16 var_4279_to_fp16 = const()[name = string("op_4279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_123_cast_fp16 = maximum(x = var_4278_cast_fp16, y = var_4279_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_119_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_123_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4282_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<int32, [1]> var_4287_axes_0 = const()[name = string("op_4287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4287_keep_dims_0 = const()[name = string("op_4287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4287_cast_fp16 = reduce_mean(axes = var_4287_axes_0, keep_dims = var_4287_keep_dims_0, x = var_4282_cast_fp16)[name = string("op_4287_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4288_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4288_cast_fp16")];
            fp32 var_4289_epsilon_0 = const()[name = string("op_4289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4289_cast_fp16 = inverse(epsilon = var_4289_epsilon_0, x = var_4288_cast_fp16)[name = string("op_4289_cast_fp16")];
            fp16 var_4290_to_fp16 = const()[name = string("op_4290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4291_cast_fp16 = mul(x = var_4289_cast_fp16, y = var_4290_to_fp16)[name = string("op_4291_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4293_cast_fp16 = add(x = var_4287_cast_fp16, y = var_4291_cast_fp16)[name = string("op_4293_cast_fp16")];
            fp32 var_4294_epsilon_0 = const()[name = string("op_4294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4294_cast_fp16 = rsqrt(epsilon = var_4294_epsilon_0, x = var_4293_cast_fp16)[name = string("op_4294_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_33_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4294_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<int32, [4]> z_65_begin_0 = const()[name = string("z_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_65_end_0 = const()[name = string("z_65_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_65_end_mask_0 = const()[name = string("z_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_65_cast_fp16 = slice_by_index(begin = z_65_begin_0, end = z_65_end_0, end_mask = z_65_end_mask_0, x = x_153_cast_fp16)[name = string("z_65_cast_fp16")];
            tensor<int32, [4]> var_4305_begin_0 = const()[name = string("op_4305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4305_end_0 = const()[name = string("op_4305_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4305_end_mask_0 = const()[name = string("op_4305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4305_cast_fp16 = slice_by_index(begin = var_4305_begin_0, end = var_4305_end_0, end_mask = var_4305_end_mask_0, x = z_65_cast_fp16)[name = string("op_4305_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4306_cast_fp16 = mul(x = var_4305_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_4306_cast_fp16")];
            tensor<int32, [4]> var_4311_begin_0 = const()[name = string("op_4311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4311_end_0 = const()[name = string("op_4311_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4311_end_mask_0 = const()[name = string("op_4311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4311_cast_fp16 = slice_by_index(begin = var_4311_begin_0, end = var_4311_end_0, end_mask = var_4311_end_mask_0, x = z_65_cast_fp16)[name = string("op_4311_cast_fp16")];
            int32 var_4313 = const()[name = string("op_4313"), val = int32(-1)];
            bool rot_65_interleave_0 = const()[name = string("rot_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_65_cast_fp16 = concat(axis = var_4313, interleave = rot_65_interleave_0, values = (var_4306_cast_fp16, var_4311_cast_fp16))[name = string("rot_65_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4315_cast_fp16 = mul(x = z_65_cast_fp16, y = c_1_cast_fp16)[name = string("op_4315_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4316_cast_fp16 = mul(x = rot_65_cast_fp16, y = s_9_cast_fp16)[name = string("op_4316_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4318_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4316_cast_fp16)[name = string("op_4318_cast_fp16")];
            tensor<int32, [4]> z_67_begin_0 = const()[name = string("z_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_67_end_0 = const()[name = string("z_67_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_67_end_mask_0 = const()[name = string("z_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_67_cast_fp16 = slice_by_index(begin = z_67_begin_0, end = z_67_end_0, end_mask = z_67_end_mask_0, x = x_153_cast_fp16)[name = string("z_67_cast_fp16")];
            tensor<int32, [4]> var_4328_begin_0 = const()[name = string("op_4328_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4328_end_0 = const()[name = string("op_4328_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4328_end_mask_0 = const()[name = string("op_4328_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4328_cast_fp16 = slice_by_index(begin = var_4328_begin_0, end = var_4328_end_0, end_mask = var_4328_end_mask_0, x = z_67_cast_fp16)[name = string("op_4328_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4329_cast_fp16 = mul(x = var_4328_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_4329_cast_fp16")];
            tensor<int32, [4]> var_4334_begin_0 = const()[name = string("op_4334_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4334_end_0 = const()[name = string("op_4334_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4334_end_mask_0 = const()[name = string("op_4334_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4334_cast_fp16 = slice_by_index(begin = var_4334_begin_0, end = var_4334_end_0, end_mask = var_4334_end_mask_0, x = z_67_cast_fp16)[name = string("op_4334_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            bool rot_67_interleave_0 = const()[name = string("rot_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_67_cast_fp16 = concat(axis = var_4336, interleave = rot_67_interleave_0, values = (var_4329_cast_fp16, var_4334_cast_fp16))[name = string("rot_67_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4338_cast_fp16 = mul(x = z_67_cast_fp16, y = c_cast_fp16)[name = string("op_4338_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4339_cast_fp16 = mul(x = rot_67_cast_fp16, y = s_11_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4341_cast_fp16 = add(x = var_4338_cast_fp16, y = var_4339_cast_fp16)[name = string("op_4341_cast_fp16")];
            int32 var_4343 = const()[name = string("op_4343"), val = int32(-1)];
            bool var_4344_interleave_0 = const()[name = string("op_4344_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_4344_cast_fp16 = concat(axis = var_4343, interleave = var_4344_interleave_0, values = (var_4318_cast_fp16, var_4341_cast_fp16))[name = string("op_4344_cast_fp16")];
            tensor<int32, [4]> z_69_begin_0 = const()[name = string("z_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_69_end_0 = const()[name = string("z_69_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_69_end_mask_0 = const()[name = string("z_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_69_cast_fp16 = slice_by_index(begin = z_69_begin_0, end = z_69_end_0, end_mask = z_69_end_mask_0, x = x_155_cast_fp16)[name = string("z_69_cast_fp16")];
            tensor<int32, [4]> var_4357_begin_0 = const()[name = string("op_4357_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4357_end_0 = const()[name = string("op_4357_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4357_end_mask_0 = const()[name = string("op_4357_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4357_cast_fp16 = slice_by_index(begin = var_4357_begin_0, end = var_4357_end_0, end_mask = var_4357_end_mask_0, x = z_69_cast_fp16)[name = string("op_4357_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4358_cast_fp16 = mul(x = var_4357_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_4358_cast_fp16")];
            tensor<int32, [4]> var_4363_begin_0 = const()[name = string("op_4363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4363_end_0 = const()[name = string("op_4363_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4363_end_mask_0 = const()[name = string("op_4363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4363_cast_fp16 = slice_by_index(begin = var_4363_begin_0, end = var_4363_end_0, end_mask = var_4363_end_mask_0, x = z_69_cast_fp16)[name = string("op_4363_cast_fp16")];
            int32 var_4365 = const()[name = string("op_4365"), val = int32(-1)];
            bool rot_69_interleave_0 = const()[name = string("rot_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_69_cast_fp16 = concat(axis = var_4365, interleave = rot_69_interleave_0, values = (var_4358_cast_fp16, var_4363_cast_fp16))[name = string("rot_69_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4367_cast_fp16 = mul(x = z_69_cast_fp16, y = c_1_cast_fp16)[name = string("op_4367_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4368_cast_fp16 = mul(x = rot_69_cast_fp16, y = s_9_cast_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4370_cast_fp16 = add(x = var_4367_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            tensor<int32, [4]> z_71_begin_0 = const()[name = string("z_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_71_end_0 = const()[name = string("z_71_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_71_end_mask_0 = const()[name = string("z_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_71_cast_fp16 = slice_by_index(begin = z_71_begin_0, end = z_71_end_0, end_mask = z_71_end_mask_0, x = x_155_cast_fp16)[name = string("z_71_cast_fp16")];
            tensor<int32, [4]> var_4380_begin_0 = const()[name = string("op_4380_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4380_end_0 = const()[name = string("op_4380_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4380_end_mask_0 = const()[name = string("op_4380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4380_cast_fp16 = slice_by_index(begin = var_4380_begin_0, end = var_4380_end_0, end_mask = var_4380_end_mask_0, x = z_71_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4381_cast_fp16 = mul(x = var_4380_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_4381_cast_fp16")];
            tensor<int32, [4]> var_4386_begin_0 = const()[name = string("op_4386_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4386_end_0 = const()[name = string("op_4386_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4386_end_mask_0 = const()[name = string("op_4386_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4386_cast_fp16 = slice_by_index(begin = var_4386_begin_0, end = var_4386_end_0, end_mask = var_4386_end_mask_0, x = z_71_cast_fp16)[name = string("op_4386_cast_fp16")];
            int32 var_4388 = const()[name = string("op_4388"), val = int32(-1)];
            bool rot_71_interleave_0 = const()[name = string("rot_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_71_cast_fp16 = concat(axis = var_4388, interleave = rot_71_interleave_0, values = (var_4381_cast_fp16, var_4386_cast_fp16))[name = string("rot_71_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4390_cast_fp16 = mul(x = z_71_cast_fp16, y = c_cast_fp16)[name = string("op_4390_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4391_cast_fp16 = mul(x = rot_71_cast_fp16, y = s_11_cast_fp16)[name = string("op_4391_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4393_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4391_cast_fp16)[name = string("op_4393_cast_fp16")];
            int32 var_4395 = const()[name = string("op_4395"), val = int32(-1)];
            bool var_4396_interleave_0 = const()[name = string("op_4396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_4396_cast_fp16 = concat(axis = var_4395, interleave = var_4396_interleave_0, values = (var_4370_cast_fp16, var_4393_cast_fp16))[name = string("op_4396_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4406_transpose_x_0 = const()[name = string("op_4406_transpose_x_0"), val = bool(false)];
            bool var_4406_transpose_y_0 = const()[name = string("op_4406_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_4396_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_80 = transpose(perm = transpose_80_perm_0, x = var_4344_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 12, 2520, 2520]> var_4406_cast_fp16 = matmul(transpose_x = var_4406_transpose_x_0, transpose_y = var_4406_transpose_y_0, x = transpose_80, y = transpose_81)[name = string("op_4406_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_4408_cast_fp16 = add(x = var_4406_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4408_cast_fp16")];
            int32 var_4409 = const()[name = string("op_4409"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_119_cast_fp16 = softmax(axis = var_4409, x = var_4408_cast_fp16)[name = string("w_119_cast_fp16")];
            bool var_4412_transpose_x_0 = const()[name = string("op_4412_transpose_x_0"), val = bool(false)];
            bool var_4412_transpose_y_0 = const()[name = string("op_4412_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = v_33_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 12, 2520, 64]> var_4412_cast_fp16 = matmul(transpose_x = var_4412_transpose_x_0, transpose_y = var_4412_transpose_y_0, x = w_119_cast_fp16, y = v_35_cast_fp16)[name = string("op_4412_cast_fp16")];
            tensor<int32, [4]> var_4415_perm_0 = const()[name = string("op_4415_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4419 = const()[name = string("op_4419"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_4415_cast_fp16 = transpose(perm = var_4415_perm_0, x = var_4412_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 2520, 768]> input_85_cast_fp16 = reshape(shape = var_4419, x = var_4415_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78093440))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78683328))))[name = string("t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4424_cast_fp16 = abs(x = linear_60_cast_fp16)[name = string("op_4424_cast_fp16")];
            bool var_4427 = const()[name = string("op_4427"), val = bool(true)];
            tensor<int32, [1]> var_4428_axes_0 = const()[name = string("op_4428_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4428_cast_fp16 = reduce_max(axes = var_4428_axes_0, keep_dims = var_4427, x = var_4424_cast_fp16)[name = string("op_4428_cast_fp16")];
            fp16 var_4429_to_fp16 = const()[name = string("op_4429_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_125_cast_fp16 = maximum(x = var_4428_cast_fp16, y = var_4429_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_121_cast_fp16 = real_div(x = linear_60_cast_fp16, y = s_125_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4432_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [1]> var_4437_axes_0 = const()[name = string("op_4437_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4437_keep_dims_0 = const()[name = string("op_4437_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4437_cast_fp16 = reduce_mean(axes = var_4437_axes_0, keep_dims = var_4437_keep_dims_0, x = var_4432_cast_fp16)[name = string("op_4437_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4438_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4438_cast_fp16")];
            fp32 var_4439_epsilon_0 = const()[name = string("op_4439_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4439_cast_fp16 = inverse(epsilon = var_4439_epsilon_0, x = var_4438_cast_fp16)[name = string("op_4439_cast_fp16")];
            fp16 var_4440_to_fp16 = const()[name = string("op_4440_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4441_cast_fp16 = mul(x = var_4439_cast_fp16, y = var_4440_to_fp16)[name = string("op_4441_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4443_cast_fp16 = add(x = var_4437_cast_fp16, y = var_4441_cast_fp16)[name = string("op_4443_cast_fp16")];
            fp32 var_4444_epsilon_0 = const()[name = string("op_4444_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4444_cast_fp16 = rsqrt(epsilon = var_4444_epsilon_0, x = var_4443_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_103_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4444_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78684928)))];
            tensor<fp16, [1, 2520, 768]> var_4446_cast_fp16 = mul(x = out_103_cast_fp16, y = t_encoder_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_4446_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_159_cast_fp16 = add(x = x_145_cast_fp16, y = var_4446_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4449_cast_fp16 = abs(x = x_159_cast_fp16)[name = string("op_4449_cast_fp16")];
            bool var_4452 = const()[name = string("op_4452"), val = bool(true)];
            tensor<int32, [1]> var_4453_axes_0 = const()[name = string("op_4453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4453_cast_fp16 = reduce_max(axes = var_4453_axes_0, keep_dims = var_4452, x = var_4449_cast_fp16)[name = string("op_4453_cast_fp16")];
            fp16 var_4454_to_fp16 = const()[name = string("op_4454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_127_cast_fp16 = maximum(x = var_4453_cast_fp16, y = var_4454_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_123_cast_fp16 = real_div(x = x_159_cast_fp16, y = s_127_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4457_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [1]> var_4462_axes_0 = const()[name = string("op_4462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4462_keep_dims_0 = const()[name = string("op_4462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4462_cast_fp16 = reduce_mean(axes = var_4462_axes_0, keep_dims = var_4462_keep_dims_0, x = var_4457_cast_fp16)[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4463_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4463_cast_fp16")];
            fp32 var_4464_epsilon_0 = const()[name = string("op_4464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4464_cast_fp16 = inverse(epsilon = var_4464_epsilon_0, x = var_4463_cast_fp16)[name = string("op_4464_cast_fp16")];
            fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4466_cast_fp16 = mul(x = var_4464_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4468_cast_fp16 = add(x = var_4462_cast_fp16, y = var_4466_cast_fp16)[name = string("op_4468_cast_fp16")];
            fp32 var_4469_epsilon_0 = const()[name = string("op_4469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4469_cast_fp16 = rsqrt(epsilon = var_4469_epsilon_0, x = var_4468_cast_fp16)[name = string("op_4469_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_105_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4469_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78686528)))];
            tensor<fp16, [1, 2520, 768]> input_87_cast_fp16 = mul(x = out_105_cast_fp16, y = t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78688128))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81047488))))[name = string("t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_61_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_61_cast_fp16")];
            string var_4476_mode_0 = const()[name = string("op_4476_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_4476_cast_fp16 = gelu(mode = var_4476_mode_0, x = linear_61_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81053696))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83413056))))[name = string("t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_62_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_91_cast_fp16 = mul(x = var_4476_cast_fp16, y = linear_62_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83419264))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85778624))))[name = string("t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4484_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4484_cast_fp16")];
            bool var_4487 = const()[name = string("op_4487"), val = bool(true)];
            tensor<int32, [1]> var_4488_axes_0 = const()[name = string("op_4488_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4488_cast_fp16 = reduce_max(axes = var_4488_axes_0, keep_dims = var_4487, x = var_4484_cast_fp16)[name = string("op_4488_cast_fp16")];
            fp16 var_4489_to_fp16 = const()[name = string("op_4489_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_129_cast_fp16 = maximum(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_125_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_129_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4492_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4492_cast_fp16")];
            tensor<int32, [1]> var_4497_axes_0 = const()[name = string("op_4497_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4497_keep_dims_0 = const()[name = string("op_4497_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4497_cast_fp16 = reduce_mean(axes = var_4497_axes_0, keep_dims = var_4497_keep_dims_0, x = var_4492_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4498_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4498_cast_fp16")];
            fp32 var_4499_epsilon_0 = const()[name = string("op_4499_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4499_cast_fp16 = inverse(epsilon = var_4499_epsilon_0, x = var_4498_cast_fp16)[name = string("op_4499_cast_fp16")];
            fp16 var_4500_to_fp16 = const()[name = string("op_4500_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4501_cast_fp16 = mul(x = var_4499_cast_fp16, y = var_4500_to_fp16)[name = string("op_4501_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4503_cast_fp16 = add(x = var_4497_cast_fp16, y = var_4501_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4504_cast_fp16 = rsqrt(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_107_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4504_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85780224)))];
            tensor<fp16, [1, 2520, 768]> var_4506_cast_fp16 = mul(x = out_107_cast_fp16, y = t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_163_cast_fp16 = add(x = x_159_cast_fp16, y = var_4506_cast_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4509_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_4509_cast_fp16")];
            bool var_4512 = const()[name = string("op_4512"), val = bool(true)];
            tensor<int32, [1]> var_4513_axes_0 = const()[name = string("op_4513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4513_cast_fp16 = reduce_max(axes = var_4513_axes_0, keep_dims = var_4512, x = var_4509_cast_fp16)[name = string("op_4513_cast_fp16")];
            fp16 var_4514_to_fp16 = const()[name = string("op_4514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_131_cast_fp16 = maximum(x = var_4513_cast_fp16, y = var_4514_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_127_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_131_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4517_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4517_cast_fp16")];
            tensor<int32, [1]> var_4522_axes_0 = const()[name = string("op_4522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4522_keep_dims_0 = const()[name = string("op_4522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4522_cast_fp16 = reduce_mean(axes = var_4522_axes_0, keep_dims = var_4522_keep_dims_0, x = var_4517_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4523_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4523_cast_fp16")];
            fp32 var_4524_epsilon_0 = const()[name = string("op_4524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4524_cast_fp16 = inverse(epsilon = var_4524_epsilon_0, x = var_4523_cast_fp16)[name = string("op_4524_cast_fp16")];
            fp16 var_4525_to_fp16 = const()[name = string("op_4525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4526_cast_fp16 = mul(x = var_4524_cast_fp16, y = var_4525_to_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4528_cast_fp16 = add(x = var_4522_cast_fp16, y = var_4526_cast_fp16)[name = string("op_4528_cast_fp16")];
            fp32 var_4529_epsilon_0 = const()[name = string("op_4529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4529_cast_fp16 = rsqrt(epsilon = var_4529_epsilon_0, x = var_4528_cast_fp16)[name = string("op_4529_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_109_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4529_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85781824)))];
            tensor<fp16, [1, 2520, 768]> input_93_cast_fp16 = mul(x = out_109_cast_fp16, y = t_encoder_layers_9_input_layernorm_weight_to_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85783424))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86373312))))[name = string("t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4539 = const()[name = string("op_4539"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_165_cast_fp16 = reshape(shape = var_4539, x = linear_64_cast_fp16)[name = string("x_165_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4541_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_4541_cast_fp16")];
            bool var_4544 = const()[name = string("op_4544"), val = bool(true)];
            tensor<int32, [1]> var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4545_cast_fp16 = reduce_max(axes = var_4545_axes_0, keep_dims = var_4544, x = var_4541_cast_fp16)[name = string("op_4545_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_133_cast_fp16 = maximum(x = var_4545_cast_fp16, y = var_4546_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_129_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_133_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4549_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<int32, [1]> var_4554_axes_0 = const()[name = string("op_4554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4554_keep_dims_0 = const()[name = string("op_4554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4554_cast_fp16 = reduce_mean(axes = var_4554_axes_0, keep_dims = var_4554_keep_dims_0, x = var_4549_cast_fp16)[name = string("op_4554_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4555_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4555_cast_fp16")];
            fp32 var_4556_epsilon_0 = const()[name = string("op_4556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4556_cast_fp16 = inverse(epsilon = var_4556_epsilon_0, x = var_4555_cast_fp16)[name = string("op_4556_cast_fp16")];
            fp16 var_4557_to_fp16 = const()[name = string("op_4557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4558_cast_fp16 = mul(x = var_4556_cast_fp16, y = var_4557_to_fp16)[name = string("op_4558_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4560_cast_fp16 = add(x = var_4554_cast_fp16, y = var_4558_cast_fp16)[name = string("op_4560_cast_fp16")];
            fp32 var_4561_epsilon_0 = const()[name = string("op_4561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4561_cast_fp16 = rsqrt(epsilon = var_4561_epsilon_0, x = var_4560_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_111_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4561_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86374912)))];
            tensor<fp16, [1, 2520, 12, 64]> x_171_cast_fp16 = mul(x = out_111_cast_fp16, y = t_encoder_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_171_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86375104))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86964992))))[name = string("t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_65_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4571 = const()[name = string("op_4571"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_167_cast_fp16 = reshape(shape = var_4571, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4573_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_4573_cast_fp16")];
            bool var_4576 = const()[name = string("op_4576"), val = bool(true)];
            tensor<int32, [1]> var_4577_axes_0 = const()[name = string("op_4577_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4577_cast_fp16 = reduce_max(axes = var_4577_axes_0, keep_dims = var_4576, x = var_4573_cast_fp16)[name = string("op_4577_cast_fp16")];
            fp16 var_4578_to_fp16 = const()[name = string("op_4578_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_135_cast_fp16 = maximum(x = var_4577_cast_fp16, y = var_4578_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_131_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_135_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4581_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4581_cast_fp16")];
            tensor<int32, [1]> var_4586_axes_0 = const()[name = string("op_4586_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4586_keep_dims_0 = const()[name = string("op_4586_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4586_cast_fp16 = reduce_mean(axes = var_4586_axes_0, keep_dims = var_4586_keep_dims_0, x = var_4581_cast_fp16)[name = string("op_4586_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4587_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4587_cast_fp16")];
            fp32 var_4588_epsilon_0 = const()[name = string("op_4588_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4588_cast_fp16 = inverse(epsilon = var_4588_epsilon_0, x = var_4587_cast_fp16)[name = string("op_4588_cast_fp16")];
            fp16 var_4589_to_fp16 = const()[name = string("op_4589_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4590_cast_fp16 = mul(x = var_4588_cast_fp16, y = var_4589_to_fp16)[name = string("op_4590_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4592_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4590_cast_fp16)[name = string("op_4592_cast_fp16")];
            fp32 var_4593_epsilon_0 = const()[name = string("op_4593_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4593_cast_fp16 = rsqrt(epsilon = var_4593_epsilon_0, x = var_4592_cast_fp16)[name = string("op_4593_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_113_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4593_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966592)))];
            tensor<fp16, [1, 2520, 12, 64]> x_173_cast_fp16 = mul(x = out_113_cast_fp16, y = t_encoder_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966784))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87556672))))[name = string("t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4603 = const()[name = string("op_4603"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_169_cast_fp16 = reshape(shape = var_4603, x = linear_66_cast_fp16)[name = string("x_169_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4605_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_4605_cast_fp16")];
            bool var_4608 = const()[name = string("op_4608"), val = bool(true)];
            tensor<int32, [1]> var_4609_axes_0 = const()[name = string("op_4609_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4609_cast_fp16 = reduce_max(axes = var_4609_axes_0, keep_dims = var_4608, x = var_4605_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_137_cast_fp16 = maximum(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_133_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_137_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4613_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4613_cast_fp16")];
            tensor<int32, [1]> var_4618_axes_0 = const()[name = string("op_4618_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4618_keep_dims_0 = const()[name = string("op_4618_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4618_cast_fp16 = reduce_mean(axes = var_4618_axes_0, keep_dims = var_4618_keep_dims_0, x = var_4613_cast_fp16)[name = string("op_4618_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4619_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4619_cast_fp16")];
            fp32 var_4620_epsilon_0 = const()[name = string("op_4620_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4620_cast_fp16 = inverse(epsilon = var_4620_epsilon_0, x = var_4619_cast_fp16)[name = string("op_4620_cast_fp16")];
            fp16 var_4621_to_fp16 = const()[name = string("op_4621_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4622_cast_fp16 = mul(x = var_4620_cast_fp16, y = var_4621_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4624_cast_fp16 = add(x = var_4618_cast_fp16, y = var_4622_cast_fp16)[name = string("op_4624_cast_fp16")];
            fp32 var_4625_epsilon_0 = const()[name = string("op_4625_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4625_cast_fp16 = rsqrt(epsilon = var_4625_epsilon_0, x = var_4624_cast_fp16)[name = string("op_4625_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_37_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4625_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<int32, [4]> z_73_begin_0 = const()[name = string("z_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_73_end_0 = const()[name = string("z_73_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_73_end_mask_0 = const()[name = string("z_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_73_cast_fp16 = slice_by_index(begin = z_73_begin_0, end = z_73_end_0, end_mask = z_73_end_mask_0, x = x_171_cast_fp16)[name = string("z_73_cast_fp16")];
            tensor<int32, [4]> var_4636_begin_0 = const()[name = string("op_4636_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4636_end_0 = const()[name = string("op_4636_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4636_end_mask_0 = const()[name = string("op_4636_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4636_cast_fp16 = slice_by_index(begin = var_4636_begin_0, end = var_4636_end_0, end_mask = var_4636_end_mask_0, x = z_73_cast_fp16)[name = string("op_4636_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4637_cast_fp16 = mul(x = var_4636_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_4637_cast_fp16")];
            tensor<int32, [4]> var_4642_begin_0 = const()[name = string("op_4642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4642_end_0 = const()[name = string("op_4642_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4642_end_mask_0 = const()[name = string("op_4642_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4642_cast_fp16 = slice_by_index(begin = var_4642_begin_0, end = var_4642_end_0, end_mask = var_4642_end_mask_0, x = z_73_cast_fp16)[name = string("op_4642_cast_fp16")];
            int32 var_4644 = const()[name = string("op_4644"), val = int32(-1)];
            bool rot_73_interleave_0 = const()[name = string("rot_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_73_cast_fp16 = concat(axis = var_4644, interleave = rot_73_interleave_0, values = (var_4637_cast_fp16, var_4642_cast_fp16))[name = string("rot_73_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4646_cast_fp16 = mul(x = z_73_cast_fp16, y = c_1_cast_fp16)[name = string("op_4646_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4647_cast_fp16 = mul(x = rot_73_cast_fp16, y = s_9_cast_fp16)[name = string("op_4647_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4649_cast_fp16 = add(x = var_4646_cast_fp16, y = var_4647_cast_fp16)[name = string("op_4649_cast_fp16")];
            tensor<int32, [4]> z_75_begin_0 = const()[name = string("z_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_75_end_0 = const()[name = string("z_75_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_75_end_mask_0 = const()[name = string("z_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_75_cast_fp16 = slice_by_index(begin = z_75_begin_0, end = z_75_end_0, end_mask = z_75_end_mask_0, x = x_171_cast_fp16)[name = string("z_75_cast_fp16")];
            tensor<int32, [4]> var_4659_begin_0 = const()[name = string("op_4659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4659_end_0 = const()[name = string("op_4659_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4659_end_mask_0 = const()[name = string("op_4659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4659_cast_fp16 = slice_by_index(begin = var_4659_begin_0, end = var_4659_end_0, end_mask = var_4659_end_mask_0, x = z_75_cast_fp16)[name = string("op_4659_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4660_cast_fp16 = mul(x = var_4659_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_4660_cast_fp16")];
            tensor<int32, [4]> var_4665_begin_0 = const()[name = string("op_4665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4665_end_0 = const()[name = string("op_4665_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4665_end_mask_0 = const()[name = string("op_4665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4665_cast_fp16 = slice_by_index(begin = var_4665_begin_0, end = var_4665_end_0, end_mask = var_4665_end_mask_0, x = z_75_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4667 = const()[name = string("op_4667"), val = int32(-1)];
            bool rot_75_interleave_0 = const()[name = string("rot_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_75_cast_fp16 = concat(axis = var_4667, interleave = rot_75_interleave_0, values = (var_4660_cast_fp16, var_4665_cast_fp16))[name = string("rot_75_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4669_cast_fp16 = mul(x = z_75_cast_fp16, y = c_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4670_cast_fp16 = mul(x = rot_75_cast_fp16, y = s_11_cast_fp16)[name = string("op_4670_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4672_cast_fp16 = add(x = var_4669_cast_fp16, y = var_4670_cast_fp16)[name = string("op_4672_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(-1)];
            bool var_4675_interleave_0 = const()[name = string("op_4675_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_4675_cast_fp16 = concat(axis = var_4674, interleave = var_4675_interleave_0, values = (var_4649_cast_fp16, var_4672_cast_fp16))[name = string("op_4675_cast_fp16")];
            tensor<int32, [4]> z_77_begin_0 = const()[name = string("z_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_77_end_0 = const()[name = string("z_77_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_77_end_mask_0 = const()[name = string("z_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_77_cast_fp16 = slice_by_index(begin = z_77_begin_0, end = z_77_end_0, end_mask = z_77_end_mask_0, x = x_173_cast_fp16)[name = string("z_77_cast_fp16")];
            tensor<int32, [4]> var_4688_begin_0 = const()[name = string("op_4688_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4688_end_0 = const()[name = string("op_4688_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4688_end_mask_0 = const()[name = string("op_4688_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4688_cast_fp16 = slice_by_index(begin = var_4688_begin_0, end = var_4688_end_0, end_mask = var_4688_end_mask_0, x = z_77_cast_fp16)[name = string("op_4688_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4689_cast_fp16 = mul(x = var_4688_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [4]> var_4694_begin_0 = const()[name = string("op_4694_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4694_end_0 = const()[name = string("op_4694_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4694_end_mask_0 = const()[name = string("op_4694_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4694_cast_fp16 = slice_by_index(begin = var_4694_begin_0, end = var_4694_end_0, end_mask = var_4694_end_mask_0, x = z_77_cast_fp16)[name = string("op_4694_cast_fp16")];
            int32 var_4696 = const()[name = string("op_4696"), val = int32(-1)];
            bool rot_77_interleave_0 = const()[name = string("rot_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_77_cast_fp16 = concat(axis = var_4696, interleave = rot_77_interleave_0, values = (var_4689_cast_fp16, var_4694_cast_fp16))[name = string("rot_77_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4698_cast_fp16 = mul(x = z_77_cast_fp16, y = c_1_cast_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4699_cast_fp16 = mul(x = rot_77_cast_fp16, y = s_9_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4701_cast_fp16 = add(x = var_4698_cast_fp16, y = var_4699_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<int32, [4]> z_79_begin_0 = const()[name = string("z_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_79_end_0 = const()[name = string("z_79_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_79_end_mask_0 = const()[name = string("z_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_79_cast_fp16 = slice_by_index(begin = z_79_begin_0, end = z_79_end_0, end_mask = z_79_end_mask_0, x = x_173_cast_fp16)[name = string("z_79_cast_fp16")];
            tensor<int32, [4]> var_4711_begin_0 = const()[name = string("op_4711_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4711_end_0 = const()[name = string("op_4711_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4711_end_mask_0 = const()[name = string("op_4711_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4711_cast_fp16 = slice_by_index(begin = var_4711_begin_0, end = var_4711_end_0, end_mask = var_4711_end_mask_0, x = z_79_cast_fp16)[name = string("op_4711_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4712_cast_fp16 = mul(x = var_4711_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_4712_cast_fp16")];
            tensor<int32, [4]> var_4717_begin_0 = const()[name = string("op_4717_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4717_end_0 = const()[name = string("op_4717_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4717_end_mask_0 = const()[name = string("op_4717_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4717_cast_fp16 = slice_by_index(begin = var_4717_begin_0, end = var_4717_end_0, end_mask = var_4717_end_mask_0, x = z_79_cast_fp16)[name = string("op_4717_cast_fp16")];
            int32 var_4719 = const()[name = string("op_4719"), val = int32(-1)];
            bool rot_79_interleave_0 = const()[name = string("rot_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_79_cast_fp16 = concat(axis = var_4719, interleave = rot_79_interleave_0, values = (var_4712_cast_fp16, var_4717_cast_fp16))[name = string("rot_79_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4721_cast_fp16 = mul(x = z_79_cast_fp16, y = c_cast_fp16)[name = string("op_4721_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4722_cast_fp16 = mul(x = rot_79_cast_fp16, y = s_11_cast_fp16)[name = string("op_4722_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4724_cast_fp16 = add(x = var_4721_cast_fp16, y = var_4722_cast_fp16)[name = string("op_4724_cast_fp16")];
            int32 var_4726 = const()[name = string("op_4726"), val = int32(-1)];
            bool var_4727_interleave_0 = const()[name = string("op_4727_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_4727_cast_fp16 = concat(axis = var_4726, interleave = var_4727_interleave_0, values = (var_4701_cast_fp16, var_4724_cast_fp16))[name = string("op_4727_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4737_transpose_x_0 = const()[name = string("op_4737_transpose_x_0"), val = bool(false)];
            bool var_4737_transpose_y_0 = const()[name = string("op_4737_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_4727_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_82 = transpose(perm = transpose_82_perm_0, x = var_4675_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 12, 2520, 2520]> var_4737_cast_fp16 = matmul(transpose_x = var_4737_transpose_x_0, transpose_y = var_4737_transpose_y_0, x = transpose_82, y = transpose_83)[name = string("op_4737_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_4739_cast_fp16 = add(x = var_4737_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4739_cast_fp16")];
            int32 var_4740 = const()[name = string("op_4740"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_133_cast_fp16 = softmax(axis = var_4740, x = var_4739_cast_fp16)[name = string("w_133_cast_fp16")];
            bool var_4743_transpose_x_0 = const()[name = string("op_4743_transpose_x_0"), val = bool(false)];
            bool var_4743_transpose_y_0 = const()[name = string("op_4743_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = v_37_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 12, 2520, 64]> var_4743_cast_fp16 = matmul(transpose_x = var_4743_transpose_x_0, transpose_y = var_4743_transpose_y_0, x = w_133_cast_fp16, y = v_39_cast_fp16)[name = string("op_4743_cast_fp16")];
            tensor<int32, [4]> var_4746_perm_0 = const()[name = string("op_4746_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4750 = const()[name = string("op_4750"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_4746_cast_fp16 = transpose(perm = var_4746_perm_0, x = var_4743_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 2520, 768]> input_95_cast_fp16 = reshape(shape = var_4750, x = var_4746_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87558272))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88148160))))[name = string("t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_67_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4755_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4755_cast_fp16")];
            bool var_4758 = const()[name = string("op_4758"), val = bool(true)];
            tensor<int32, [1]> var_4759_axes_0 = const()[name = string("op_4759_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4759_cast_fp16 = reduce_max(axes = var_4759_axes_0, keep_dims = var_4758, x = var_4755_cast_fp16)[name = string("op_4759_cast_fp16")];
            fp16 var_4760_to_fp16 = const()[name = string("op_4760_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_139_cast_fp16 = maximum(x = var_4759_cast_fp16, y = var_4760_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_135_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_139_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4763_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4763_cast_fp16")];
            tensor<int32, [1]> var_4768_axes_0 = const()[name = string("op_4768_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4768_keep_dims_0 = const()[name = string("op_4768_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4768_cast_fp16 = reduce_mean(axes = var_4768_axes_0, keep_dims = var_4768_keep_dims_0, x = var_4763_cast_fp16)[name = string("op_4768_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4769_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_4769_cast_fp16")];
            fp32 var_4770_epsilon_0 = const()[name = string("op_4770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4770_cast_fp16 = inverse(epsilon = var_4770_epsilon_0, x = var_4769_cast_fp16)[name = string("op_4770_cast_fp16")];
            fp16 var_4771_to_fp16 = const()[name = string("op_4771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4772_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_4771_to_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4774_cast_fp16 = add(x = var_4768_cast_fp16, y = var_4772_cast_fp16)[name = string("op_4774_cast_fp16")];
            fp32 var_4775_epsilon_0 = const()[name = string("op_4775_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4775_cast_fp16 = rsqrt(epsilon = var_4775_epsilon_0, x = var_4774_cast_fp16)[name = string("op_4775_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_115_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4775_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88149760)))];
            tensor<fp16, [1, 2520, 768]> var_4777_cast_fp16 = mul(x = out_115_cast_fp16, y = t_encoder_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_4777_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_177_cast_fp16 = add(x = x_163_cast_fp16, y = var_4777_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4780_cast_fp16 = abs(x = x_177_cast_fp16)[name = string("op_4780_cast_fp16")];
            bool var_4783 = const()[name = string("op_4783"), val = bool(true)];
            tensor<int32, [1]> var_4784_axes_0 = const()[name = string("op_4784_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4784_cast_fp16 = reduce_max(axes = var_4784_axes_0, keep_dims = var_4783, x = var_4780_cast_fp16)[name = string("op_4784_cast_fp16")];
            fp16 var_4785_to_fp16 = const()[name = string("op_4785_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_141_cast_fp16 = maximum(x = var_4784_cast_fp16, y = var_4785_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_137_cast_fp16 = real_div(x = x_177_cast_fp16, y = s_141_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4788_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4788_cast_fp16")];
            tensor<int32, [1]> var_4793_axes_0 = const()[name = string("op_4793_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4793_keep_dims_0 = const()[name = string("op_4793_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4793_cast_fp16 = reduce_mean(axes = var_4793_axes_0, keep_dims = var_4793_keep_dims_0, x = var_4788_cast_fp16)[name = string("op_4793_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4794_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_4794_cast_fp16")];
            fp32 var_4795_epsilon_0 = const()[name = string("op_4795_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4795_cast_fp16 = inverse(epsilon = var_4795_epsilon_0, x = var_4794_cast_fp16)[name = string("op_4795_cast_fp16")];
            fp16 var_4796_to_fp16 = const()[name = string("op_4796_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4797_cast_fp16 = mul(x = var_4795_cast_fp16, y = var_4796_to_fp16)[name = string("op_4797_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4799_cast_fp16 = add(x = var_4793_cast_fp16, y = var_4797_cast_fp16)[name = string("op_4799_cast_fp16")];
            fp32 var_4800_epsilon_0 = const()[name = string("op_4800_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4800_cast_fp16 = rsqrt(epsilon = var_4800_epsilon_0, x = var_4799_cast_fp16)[name = string("op_4800_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_117_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4800_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88151360)))];
            tensor<fp16, [1, 2520, 768]> input_97_cast_fp16 = mul(x = out_117_cast_fp16, y = t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88152960))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90512320))))[name = string("t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4807_mode_0 = const()[name = string("op_4807_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_4807_cast_fp16 = gelu(mode = var_4807_mode_0, x = linear_68_cast_fp16)[name = string("op_4807_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90518528))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92877888))))[name = string("t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_101_cast_fp16 = mul(x = var_4807_cast_fp16, y = linear_69_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92884096))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95243456))))[name = string("t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4815_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4815_cast_fp16")];
            bool var_4818 = const()[name = string("op_4818"), val = bool(true)];
            tensor<int32, [1]> var_4819_axes_0 = const()[name = string("op_4819_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4819_cast_fp16 = reduce_max(axes = var_4819_axes_0, keep_dims = var_4818, x = var_4815_cast_fp16)[name = string("op_4819_cast_fp16")];
            fp16 var_4820_to_fp16 = const()[name = string("op_4820_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_143_cast_fp16 = maximum(x = var_4819_cast_fp16, y = var_4820_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_139_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_143_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4823_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_4823_cast_fp16")];
            tensor<int32, [1]> var_4828_axes_0 = const()[name = string("op_4828_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4828_keep_dims_0 = const()[name = string("op_4828_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4828_cast_fp16 = reduce_mean(axes = var_4828_axes_0, keep_dims = var_4828_keep_dims_0, x = var_4823_cast_fp16)[name = string("op_4828_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4829_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_4829_cast_fp16")];
            fp32 var_4830_epsilon_0 = const()[name = string("op_4830_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4830_cast_fp16 = inverse(epsilon = var_4830_epsilon_0, x = var_4829_cast_fp16)[name = string("op_4830_cast_fp16")];
            fp16 var_4831_to_fp16 = const()[name = string("op_4831_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4832_cast_fp16 = mul(x = var_4830_cast_fp16, y = var_4831_to_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4834_cast_fp16 = add(x = var_4828_cast_fp16, y = var_4832_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp32 var_4835_epsilon_0 = const()[name = string("op_4835_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4835_cast_fp16 = rsqrt(epsilon = var_4835_epsilon_0, x = var_4834_cast_fp16)[name = string("op_4835_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_119_cast_fp16 = mul(x = y_139_cast_fp16, y = var_4835_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95245056)))];
            tensor<fp16, [1, 2520, 768]> var_4837_cast_fp16 = mul(x = out_119_cast_fp16, y = t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4837_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_181_cast_fp16 = add(x = x_177_cast_fp16, y = var_4837_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4840_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_4840_cast_fp16")];
            bool var_4843 = const()[name = string("op_4843"), val = bool(true)];
            tensor<int32, [1]> var_4844_axes_0 = const()[name = string("op_4844_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_4844_cast_fp16 = reduce_max(axes = var_4844_axes_0, keep_dims = var_4843, x = var_4840_cast_fp16)[name = string("op_4844_cast_fp16")];
            fp16 var_4845_to_fp16 = const()[name = string("op_4845_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_145_cast_fp16 = maximum(x = var_4844_cast_fp16, y = var_4845_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_141_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_145_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_4848_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_4848_cast_fp16")];
            tensor<int32, [1]> var_4853_axes_0 = const()[name = string("op_4853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4853_keep_dims_0 = const()[name = string("op_4853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_4853_cast_fp16 = reduce_mean(axes = var_4853_axes_0, keep_dims = var_4853_keep_dims_0, x = var_4848_cast_fp16)[name = string("op_4853_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4854_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_4854_cast_fp16")];
            fp32 var_4855_epsilon_0 = const()[name = string("op_4855_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_4855_cast_fp16 = inverse(epsilon = var_4855_epsilon_0, x = var_4854_cast_fp16)[name = string("op_4855_cast_fp16")];
            fp16 var_4856_to_fp16 = const()[name = string("op_4856_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_4857_cast_fp16 = mul(x = var_4855_cast_fp16, y = var_4856_to_fp16)[name = string("op_4857_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_4859_cast_fp16 = add(x = var_4853_cast_fp16, y = var_4857_cast_fp16)[name = string("op_4859_cast_fp16")];
            fp32 var_4860_epsilon_0 = const()[name = string("op_4860_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_4860_cast_fp16 = rsqrt(epsilon = var_4860_epsilon_0, x = var_4859_cast_fp16)[name = string("op_4860_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_121_cast_fp16 = mul(x = y_141_cast_fp16, y = var_4860_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95246656)))];
            tensor<fp16, [1, 2520, 768]> input_103_cast_fp16 = mul(x = out_121_cast_fp16, y = t_encoder_layers_10_input_layernorm_weight_to_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95248256))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95838144))))[name = string("t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_71_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_4870 = const()[name = string("op_4870"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_183_cast_fp16 = reshape(shape = var_4870, x = linear_71_cast_fp16)[name = string("x_183_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4872_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_4872_cast_fp16")];
            bool var_4875 = const()[name = string("op_4875"), val = bool(true)];
            tensor<int32, [1]> var_4876_axes_0 = const()[name = string("op_4876_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4876_cast_fp16 = reduce_max(axes = var_4876_axes_0, keep_dims = var_4875, x = var_4872_cast_fp16)[name = string("op_4876_cast_fp16")];
            fp16 var_4877_to_fp16 = const()[name = string("op_4877_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_147_cast_fp16 = maximum(x = var_4876_cast_fp16, y = var_4877_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_143_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_147_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4880_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_4880_cast_fp16")];
            tensor<int32, [1]> var_4885_axes_0 = const()[name = string("op_4885_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4885_keep_dims_0 = const()[name = string("op_4885_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4885_cast_fp16 = reduce_mean(axes = var_4885_axes_0, keep_dims = var_4885_keep_dims_0, x = var_4880_cast_fp16)[name = string("op_4885_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4886_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_4886_cast_fp16")];
            fp32 var_4887_epsilon_0 = const()[name = string("op_4887_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4887_cast_fp16 = inverse(epsilon = var_4887_epsilon_0, x = var_4886_cast_fp16)[name = string("op_4887_cast_fp16")];
            fp16 var_4888_to_fp16 = const()[name = string("op_4888_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4889_cast_fp16 = mul(x = var_4887_cast_fp16, y = var_4888_to_fp16)[name = string("op_4889_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4891_cast_fp16 = add(x = var_4885_cast_fp16, y = var_4889_cast_fp16)[name = string("op_4891_cast_fp16")];
            fp32 var_4892_epsilon_0 = const()[name = string("op_4892_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4892_cast_fp16 = rsqrt(epsilon = var_4892_epsilon_0, x = var_4891_cast_fp16)[name = string("op_4892_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_123_cast_fp16 = mul(x = y_143_cast_fp16, y = var_4892_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95839744)))];
            tensor<fp16, [1, 2520, 12, 64]> x_189_cast_fp16 = mul(x = out_123_cast_fp16, y = t_encoder_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_189_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95839936))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96429824))))[name = string("t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_4902 = const()[name = string("op_4902"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_185_cast_fp16 = reshape(shape = var_4902, x = linear_72_cast_fp16)[name = string("x_185_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4904_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_4904_cast_fp16")];
            bool var_4907 = const()[name = string("op_4907"), val = bool(true)];
            tensor<int32, [1]> var_4908_axes_0 = const()[name = string("op_4908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4908_cast_fp16 = reduce_max(axes = var_4908_axes_0, keep_dims = var_4907, x = var_4904_cast_fp16)[name = string("op_4908_cast_fp16")];
            fp16 var_4909_to_fp16 = const()[name = string("op_4909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_149_cast_fp16 = maximum(x = var_4908_cast_fp16, y = var_4909_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_145_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_149_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4912_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_4912_cast_fp16")];
            tensor<int32, [1]> var_4917_axes_0 = const()[name = string("op_4917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4917_keep_dims_0 = const()[name = string("op_4917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4917_cast_fp16 = reduce_mean(axes = var_4917_axes_0, keep_dims = var_4917_keep_dims_0, x = var_4912_cast_fp16)[name = string("op_4917_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4918_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_4918_cast_fp16")];
            fp32 var_4919_epsilon_0 = const()[name = string("op_4919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4919_cast_fp16 = inverse(epsilon = var_4919_epsilon_0, x = var_4918_cast_fp16)[name = string("op_4919_cast_fp16")];
            fp16 var_4920_to_fp16 = const()[name = string("op_4920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4923_cast_fp16 = add(x = var_4917_cast_fp16, y = var_4921_cast_fp16)[name = string("op_4923_cast_fp16")];
            fp32 var_4924_epsilon_0 = const()[name = string("op_4924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4924_cast_fp16 = rsqrt(epsilon = var_4924_epsilon_0, x = var_4923_cast_fp16)[name = string("op_4924_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_125_cast_fp16 = mul(x = y_145_cast_fp16, y = var_4924_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96431424)))];
            tensor<fp16, [1, 2520, 12, 64]> x_191_cast_fp16 = mul(x = out_125_cast_fp16, y = t_encoder_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96431616))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97021504))))[name = string("t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_73_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4934 = const()[name = string("op_4934"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_187_cast_fp16 = reshape(shape = var_4934, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4936_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_4936_cast_fp16")];
            bool var_4939 = const()[name = string("op_4939"), val = bool(true)];
            tensor<int32, [1]> var_4940_axes_0 = const()[name = string("op_4940_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_4940_cast_fp16 = reduce_max(axes = var_4940_axes_0, keep_dims = var_4939, x = var_4936_cast_fp16)[name = string("op_4940_cast_fp16")];
            fp16 var_4941_to_fp16 = const()[name = string("op_4941_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_151_cast_fp16 = maximum(x = var_4940_cast_fp16, y = var_4941_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_147_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_151_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_4944_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<int32, [1]> var_4949_axes_0 = const()[name = string("op_4949_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4949_keep_dims_0 = const()[name = string("op_4949_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_4949_cast_fp16 = reduce_mean(axes = var_4949_axes_0, keep_dims = var_4949_keep_dims_0, x = var_4944_cast_fp16)[name = string("op_4949_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4950_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_4950_cast_fp16")];
            fp32 var_4951_epsilon_0 = const()[name = string("op_4951_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_4951_cast_fp16 = inverse(epsilon = var_4951_epsilon_0, x = var_4950_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 var_4952_to_fp16 = const()[name = string("op_4952_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_4953_cast_fp16 = mul(x = var_4951_cast_fp16, y = var_4952_to_fp16)[name = string("op_4953_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_4955_cast_fp16 = add(x = var_4949_cast_fp16, y = var_4953_cast_fp16)[name = string("op_4955_cast_fp16")];
            fp32 var_4956_epsilon_0 = const()[name = string("op_4956_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_4956_cast_fp16 = rsqrt(epsilon = var_4956_epsilon_0, x = var_4955_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_41_cast_fp16 = mul(x = y_147_cast_fp16, y = var_4956_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<int32, [4]> z_81_begin_0 = const()[name = string("z_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_81_end_0 = const()[name = string("z_81_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_81_end_mask_0 = const()[name = string("z_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_81_cast_fp16 = slice_by_index(begin = z_81_begin_0, end = z_81_end_0, end_mask = z_81_end_mask_0, x = x_189_cast_fp16)[name = string("z_81_cast_fp16")];
            tensor<int32, [4]> var_4967_begin_0 = const()[name = string("op_4967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4967_end_0 = const()[name = string("op_4967_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4967_end_mask_0 = const()[name = string("op_4967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4967_cast_fp16 = slice_by_index(begin = var_4967_begin_0, end = var_4967_end_0, end_mask = var_4967_end_mask_0, x = z_81_cast_fp16)[name = string("op_4967_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4968_cast_fp16 = mul(x = var_4967_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_4968_cast_fp16")];
            tensor<int32, [4]> var_4973_begin_0 = const()[name = string("op_4973_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4973_end_0 = const()[name = string("op_4973_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4973_end_mask_0 = const()[name = string("op_4973_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, x = z_81_cast_fp16)[name = string("op_4973_cast_fp16")];
            int32 var_4975 = const()[name = string("op_4975"), val = int32(-1)];
            bool rot_81_interleave_0 = const()[name = string("rot_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_81_cast_fp16 = concat(axis = var_4975, interleave = rot_81_interleave_0, values = (var_4968_cast_fp16, var_4973_cast_fp16))[name = string("rot_81_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4977_cast_fp16 = mul(x = z_81_cast_fp16, y = c_1_cast_fp16)[name = string("op_4977_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4978_cast_fp16 = mul(x = rot_81_cast_fp16, y = s_9_cast_fp16)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_4980_cast_fp16 = add(x = var_4977_cast_fp16, y = var_4978_cast_fp16)[name = string("op_4980_cast_fp16")];
            tensor<int32, [4]> z_83_begin_0 = const()[name = string("z_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_83_end_0 = const()[name = string("z_83_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_83_end_mask_0 = const()[name = string("z_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_83_cast_fp16 = slice_by_index(begin = z_83_begin_0, end = z_83_end_0, end_mask = z_83_end_mask_0, x = x_189_cast_fp16)[name = string("z_83_cast_fp16")];
            tensor<int32, [4]> var_4990_begin_0 = const()[name = string("op_4990_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4990_end_0 = const()[name = string("op_4990_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_4990_end_mask_0 = const()[name = string("op_4990_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_4990_cast_fp16 = slice_by_index(begin = var_4990_begin_0, end = var_4990_end_0, end_mask = var_4990_end_mask_0, x = z_83_cast_fp16)[name = string("op_4990_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_4991_cast_fp16 = mul(x = var_4990_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_4991_cast_fp16")];
            tensor<int32, [4]> var_4996_begin_0 = const()[name = string("op_4996_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4996_end_0 = const()[name = string("op_4996_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_4996_end_mask_0 = const()[name = string("op_4996_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_4996_cast_fp16 = slice_by_index(begin = var_4996_begin_0, end = var_4996_end_0, end_mask = var_4996_end_mask_0, x = z_83_cast_fp16)[name = string("op_4996_cast_fp16")];
            int32 var_4998 = const()[name = string("op_4998"), val = int32(-1)];
            bool rot_83_interleave_0 = const()[name = string("rot_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_83_cast_fp16 = concat(axis = var_4998, interleave = rot_83_interleave_0, values = (var_4991_cast_fp16, var_4996_cast_fp16))[name = string("rot_83_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5000_cast_fp16 = mul(x = z_83_cast_fp16, y = c_cast_fp16)[name = string("op_5000_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5001_cast_fp16 = mul(x = rot_83_cast_fp16, y = s_11_cast_fp16)[name = string("op_5001_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5003_cast_fp16 = add(x = var_5000_cast_fp16, y = var_5001_cast_fp16)[name = string("op_5003_cast_fp16")];
            int32 var_5005 = const()[name = string("op_5005"), val = int32(-1)];
            bool var_5006_interleave_0 = const()[name = string("op_5006_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_5006_cast_fp16 = concat(axis = var_5005, interleave = var_5006_interleave_0, values = (var_4980_cast_fp16, var_5003_cast_fp16))[name = string("op_5006_cast_fp16")];
            tensor<int32, [4]> z_85_begin_0 = const()[name = string("z_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_85_end_0 = const()[name = string("z_85_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_85_end_mask_0 = const()[name = string("z_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_85_cast_fp16 = slice_by_index(begin = z_85_begin_0, end = z_85_end_0, end_mask = z_85_end_mask_0, x = x_191_cast_fp16)[name = string("z_85_cast_fp16")];
            tensor<int32, [4]> var_5019_begin_0 = const()[name = string("op_5019_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5019_end_0 = const()[name = string("op_5019_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5019_end_mask_0 = const()[name = string("op_5019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5019_cast_fp16 = slice_by_index(begin = var_5019_begin_0, end = var_5019_end_0, end_mask = var_5019_end_mask_0, x = z_85_cast_fp16)[name = string("op_5019_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5020_cast_fp16 = mul(x = var_5019_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_5020_cast_fp16")];
            tensor<int32, [4]> var_5025_begin_0 = const()[name = string("op_5025_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5025_end_0 = const()[name = string("op_5025_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5025_end_mask_0 = const()[name = string("op_5025_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5025_cast_fp16 = slice_by_index(begin = var_5025_begin_0, end = var_5025_end_0, end_mask = var_5025_end_mask_0, x = z_85_cast_fp16)[name = string("op_5025_cast_fp16")];
            int32 var_5027 = const()[name = string("op_5027"), val = int32(-1)];
            bool rot_85_interleave_0 = const()[name = string("rot_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_85_cast_fp16 = concat(axis = var_5027, interleave = rot_85_interleave_0, values = (var_5020_cast_fp16, var_5025_cast_fp16))[name = string("rot_85_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5029_cast_fp16 = mul(x = z_85_cast_fp16, y = c_1_cast_fp16)[name = string("op_5029_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5030_cast_fp16 = mul(x = rot_85_cast_fp16, y = s_9_cast_fp16)[name = string("op_5030_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5032_cast_fp16 = add(x = var_5029_cast_fp16, y = var_5030_cast_fp16)[name = string("op_5032_cast_fp16")];
            tensor<int32, [4]> z_87_begin_0 = const()[name = string("z_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_87_end_0 = const()[name = string("z_87_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_87_end_mask_0 = const()[name = string("z_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_87_cast_fp16 = slice_by_index(begin = z_87_begin_0, end = z_87_end_0, end_mask = z_87_end_mask_0, x = x_191_cast_fp16)[name = string("z_87_cast_fp16")];
            tensor<int32, [4]> var_5042_begin_0 = const()[name = string("op_5042_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5042_end_0 = const()[name = string("op_5042_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5042_end_mask_0 = const()[name = string("op_5042_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5042_cast_fp16 = slice_by_index(begin = var_5042_begin_0, end = var_5042_end_0, end_mask = var_5042_end_mask_0, x = z_87_cast_fp16)[name = string("op_5042_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5043_cast_fp16 = mul(x = var_5042_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_5043_cast_fp16")];
            tensor<int32, [4]> var_5048_begin_0 = const()[name = string("op_5048_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5048_end_0 = const()[name = string("op_5048_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5048_end_mask_0 = const()[name = string("op_5048_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5048_cast_fp16 = slice_by_index(begin = var_5048_begin_0, end = var_5048_end_0, end_mask = var_5048_end_mask_0, x = z_87_cast_fp16)[name = string("op_5048_cast_fp16")];
            int32 var_5050 = const()[name = string("op_5050"), val = int32(-1)];
            bool rot_87_interleave_0 = const()[name = string("rot_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_87_cast_fp16 = concat(axis = var_5050, interleave = rot_87_interleave_0, values = (var_5043_cast_fp16, var_5048_cast_fp16))[name = string("rot_87_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5052_cast_fp16 = mul(x = z_87_cast_fp16, y = c_cast_fp16)[name = string("op_5052_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5053_cast_fp16 = mul(x = rot_87_cast_fp16, y = s_11_cast_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5055_cast_fp16 = add(x = var_5052_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            int32 var_5057 = const()[name = string("op_5057"), val = int32(-1)];
            bool var_5058_interleave_0 = const()[name = string("op_5058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_5058_cast_fp16 = concat(axis = var_5057, interleave = var_5058_interleave_0, values = (var_5032_cast_fp16, var_5055_cast_fp16))[name = string("op_5058_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5068_transpose_x_0 = const()[name = string("op_5068_transpose_x_0"), val = bool(false)];
            bool var_5068_transpose_y_0 = const()[name = string("op_5068_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_85 = transpose(perm = transpose_85_perm_0, x = var_5058_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_84 = transpose(perm = transpose_84_perm_0, x = var_5006_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 12, 2520, 2520]> var_5068_cast_fp16 = matmul(transpose_x = var_5068_transpose_x_0, transpose_y = var_5068_transpose_y_0, x = transpose_84, y = transpose_85)[name = string("op_5068_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_5070_cast_fp16 = add(x = var_5068_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5070_cast_fp16")];
            int32 var_5071 = const()[name = string("op_5071"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_147_cast_fp16 = softmax(axis = var_5071, x = var_5070_cast_fp16)[name = string("w_147_cast_fp16")];
            bool var_5074_transpose_x_0 = const()[name = string("op_5074_transpose_x_0"), val = bool(false)];
            bool var_5074_transpose_y_0 = const()[name = string("op_5074_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = v_41_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 12, 2520, 64]> var_5074_cast_fp16 = matmul(transpose_x = var_5074_transpose_x_0, transpose_y = var_5074_transpose_y_0, x = w_147_cast_fp16, y = v_43_cast_fp16)[name = string("op_5074_cast_fp16")];
            tensor<int32, [4]> var_5077_perm_0 = const()[name = string("op_5077_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5081 = const()[name = string("op_5081"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_5077_cast_fp16 = transpose(perm = var_5077_perm_0, x = var_5074_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 2520, 768]> input_105_cast_fp16 = reshape(shape = var_5081, x = var_5077_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97023104))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97612992))))[name = string("t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5086_cast_fp16 = abs(x = linear_74_cast_fp16)[name = string("op_5086_cast_fp16")];
            bool var_5089 = const()[name = string("op_5089"), val = bool(true)];
            tensor<int32, [1]> var_5090_axes_0 = const()[name = string("op_5090_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5090_cast_fp16 = reduce_max(axes = var_5090_axes_0, keep_dims = var_5089, x = var_5086_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp16 var_5091_to_fp16 = const()[name = string("op_5091_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_153_cast_fp16 = maximum(x = var_5090_cast_fp16, y = var_5091_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_149_cast_fp16 = real_div(x = linear_74_cast_fp16, y = s_153_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5094_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [1]> var_5099_axes_0 = const()[name = string("op_5099_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5099_keep_dims_0 = const()[name = string("op_5099_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5099_cast_fp16 = reduce_mean(axes = var_5099_axes_0, keep_dims = var_5099_keep_dims_0, x = var_5094_cast_fp16)[name = string("op_5099_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5100_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5100_cast_fp16")];
            fp32 var_5101_epsilon_0 = const()[name = string("op_5101_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5101_cast_fp16 = inverse(epsilon = var_5101_epsilon_0, x = var_5100_cast_fp16)[name = string("op_5101_cast_fp16")];
            fp16 var_5102_to_fp16 = const()[name = string("op_5102_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5103_cast_fp16 = mul(x = var_5101_cast_fp16, y = var_5102_to_fp16)[name = string("op_5103_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5105_cast_fp16 = add(x = var_5099_cast_fp16, y = var_5103_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, 2520, 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, 2520, 768]> out_127_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5106_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97614592)))];
            tensor<fp16, [1, 2520, 768]> var_5108_cast_fp16 = mul(x = out_127_cast_fp16, y = t_encoder_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5108_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_195_cast_fp16 = add(x = x_181_cast_fp16, y = var_5108_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5111_cast_fp16 = abs(x = x_195_cast_fp16)[name = string("op_5111_cast_fp16")];
            bool var_5114 = const()[name = string("op_5114"), val = bool(true)];
            tensor<int32, [1]> var_5115_axes_0 = const()[name = string("op_5115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5115_cast_fp16 = reduce_max(axes = var_5115_axes_0, keep_dims = var_5114, x = var_5111_cast_fp16)[name = string("op_5115_cast_fp16")];
            fp16 var_5116_to_fp16 = const()[name = string("op_5116_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_155_cast_fp16 = maximum(x = var_5115_cast_fp16, y = var_5116_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_151_cast_fp16 = real_div(x = x_195_cast_fp16, y = s_155_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5119_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5119_cast_fp16")];
            tensor<int32, [1]> var_5124_axes_0 = const()[name = string("op_5124_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5124_keep_dims_0 = const()[name = string("op_5124_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5124_cast_fp16 = reduce_mean(axes = var_5124_axes_0, keep_dims = var_5124_keep_dims_0, x = var_5119_cast_fp16)[name = string("op_5124_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5125_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5125_cast_fp16")];
            fp32 var_5126_epsilon_0 = const()[name = string("op_5126_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5126_cast_fp16 = inverse(epsilon = var_5126_epsilon_0, x = var_5125_cast_fp16)[name = string("op_5126_cast_fp16")];
            fp16 var_5127_to_fp16 = const()[name = string("op_5127_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5128_cast_fp16 = mul(x = var_5126_cast_fp16, y = var_5127_to_fp16)[name = string("op_5128_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5130_cast_fp16 = add(x = var_5124_cast_fp16, y = var_5128_cast_fp16)[name = string("op_5130_cast_fp16")];
            fp32 var_5131_epsilon_0 = const()[name = string("op_5131_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5131_cast_fp16 = rsqrt(epsilon = var_5131_epsilon_0, x = var_5130_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_129_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5131_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97616192)))];
            tensor<fp16, [1, 2520, 768]> input_107_cast_fp16 = mul(x = out_129_cast_fp16, y = t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97617792))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99977152))))[name = string("t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_75_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_75_cast_fp16")];
            string var_5138_mode_0 = const()[name = string("op_5138_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_5138_cast_fp16 = gelu(mode = var_5138_mode_0, x = linear_75_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99983360))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102342720))))[name = string("t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_76_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_111_cast_fp16 = mul(x = var_5138_cast_fp16, y = linear_76_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102348928))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104708288))))[name = string("t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5146_cast_fp16 = abs(x = linear_77_cast_fp16)[name = string("op_5146_cast_fp16")];
            bool var_5149 = const()[name = string("op_5149"), val = bool(true)];
            tensor<int32, [1]> var_5150_axes_0 = const()[name = string("op_5150_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5150_cast_fp16 = reduce_max(axes = var_5150_axes_0, keep_dims = var_5149, x = var_5146_cast_fp16)[name = string("op_5150_cast_fp16")];
            fp16 var_5151_to_fp16 = const()[name = string("op_5151_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_157_cast_fp16 = maximum(x = var_5150_cast_fp16, y = var_5151_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_153_cast_fp16 = real_div(x = linear_77_cast_fp16, y = s_157_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5154_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5154_cast_fp16")];
            tensor<int32, [1]> var_5159_axes_0 = const()[name = string("op_5159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5159_keep_dims_0 = const()[name = string("op_5159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5159_cast_fp16 = reduce_mean(axes = var_5159_axes_0, keep_dims = var_5159_keep_dims_0, x = var_5154_cast_fp16)[name = string("op_5159_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5160_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5160_cast_fp16")];
            fp32 var_5161_epsilon_0 = const()[name = string("op_5161_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5161_cast_fp16 = inverse(epsilon = var_5161_epsilon_0, x = var_5160_cast_fp16)[name = string("op_5161_cast_fp16")];
            fp16 var_5162_to_fp16 = const()[name = string("op_5162_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5163_cast_fp16 = mul(x = var_5161_cast_fp16, y = var_5162_to_fp16)[name = string("op_5163_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5165_cast_fp16 = add(x = var_5159_cast_fp16, y = var_5163_cast_fp16)[name = string("op_5165_cast_fp16")];
            fp32 var_5166_epsilon_0 = const()[name = string("op_5166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5166_cast_fp16 = rsqrt(epsilon = var_5166_epsilon_0, x = var_5165_cast_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_131_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5166_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104709888)))];
            tensor<fp16, [1, 2520, 768]> var_5168_cast_fp16 = mul(x = out_131_cast_fp16, y = t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5168_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_199_cast_fp16 = add(x = x_195_cast_fp16, y = var_5168_cast_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5171_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5171_cast_fp16")];
            bool var_5174 = const()[name = string("op_5174"), val = bool(true)];
            tensor<int32, [1]> var_5175_axes_0 = const()[name = string("op_5175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5175_cast_fp16 = reduce_max(axes = var_5175_axes_0, keep_dims = var_5174, x = var_5171_cast_fp16)[name = string("op_5175_cast_fp16")];
            fp16 var_5176_to_fp16 = const()[name = string("op_5176_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_159_cast_fp16 = maximum(x = var_5175_cast_fp16, y = var_5176_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_155_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_159_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5179_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5179_cast_fp16")];
            tensor<int32, [1]> var_5184_axes_0 = const()[name = string("op_5184_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5184_keep_dims_0 = const()[name = string("op_5184_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5184_cast_fp16 = reduce_mean(axes = var_5184_axes_0, keep_dims = var_5184_keep_dims_0, x = var_5179_cast_fp16)[name = string("op_5184_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5185_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5185_cast_fp16")];
            fp32 var_5186_epsilon_0 = const()[name = string("op_5186_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5186_cast_fp16 = inverse(epsilon = var_5186_epsilon_0, x = var_5185_cast_fp16)[name = string("op_5186_cast_fp16")];
            fp16 var_5187_to_fp16 = const()[name = string("op_5187_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5188_cast_fp16 = mul(x = var_5186_cast_fp16, y = var_5187_to_fp16)[name = string("op_5188_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5190_cast_fp16 = add(x = var_5184_cast_fp16, y = var_5188_cast_fp16)[name = string("op_5190_cast_fp16")];
            fp32 var_5191_epsilon_0 = const()[name = string("op_5191_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5191_cast_fp16 = rsqrt(epsilon = var_5191_epsilon_0, x = var_5190_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_133_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5191_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104711488)))];
            tensor<fp16, [1, 2520, 768]> input_113_cast_fp16 = mul(x = out_133_cast_fp16, y = t_encoder_layers_11_input_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104713088))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105302976))))[name = string("t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_5201 = const()[name = string("op_5201"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_201_cast_fp16 = reshape(shape = var_5201, x = linear_78_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5203_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5203_cast_fp16")];
            bool var_5206 = const()[name = string("op_5206"), val = bool(true)];
            tensor<int32, [1]> var_5207_axes_0 = const()[name = string("op_5207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5207_cast_fp16 = reduce_max(axes = var_5207_axes_0, keep_dims = var_5206, x = var_5203_cast_fp16)[name = string("op_5207_cast_fp16")];
            fp16 var_5208_to_fp16 = const()[name = string("op_5208_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_161_cast_fp16 = maximum(x = var_5207_cast_fp16, y = var_5208_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_157_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_161_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5211_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5211_cast_fp16")];
            tensor<int32, [1]> var_5216_axes_0 = const()[name = string("op_5216_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5216_keep_dims_0 = const()[name = string("op_5216_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5216_cast_fp16 = reduce_mean(axes = var_5216_axes_0, keep_dims = var_5216_keep_dims_0, x = var_5211_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5217_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5217_cast_fp16")];
            fp32 var_5218_epsilon_0 = const()[name = string("op_5218_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5218_cast_fp16 = inverse(epsilon = var_5218_epsilon_0, x = var_5217_cast_fp16)[name = string("op_5218_cast_fp16")];
            fp16 var_5219_to_fp16 = const()[name = string("op_5219_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5220_cast_fp16 = mul(x = var_5218_cast_fp16, y = var_5219_to_fp16)[name = string("op_5220_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5222_cast_fp16 = add(x = var_5216_cast_fp16, y = var_5220_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5223_cast_fp16 = rsqrt(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_135_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5223_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20121856)))];
            tensor<fp16, [1, 2520, 12, 64]> x_207_cast_fp16 = mul(x = out_135_cast_fp16, y = t_encoder_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_207_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105304576))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105894464))))[name = string("t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_79_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_5233 = const()[name = string("op_5233"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_203_cast_fp16 = reshape(shape = var_5233, x = linear_79_cast_fp16)[name = string("x_203_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5235_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5235_cast_fp16")];
            bool var_5238 = const()[name = string("op_5238"), val = bool(true)];
            tensor<int32, [1]> var_5239_axes_0 = const()[name = string("op_5239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5239_cast_fp16 = reduce_max(axes = var_5239_axes_0, keep_dims = var_5238, x = var_5235_cast_fp16)[name = string("op_5239_cast_fp16")];
            fp16 var_5240_to_fp16 = const()[name = string("op_5240_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_163_cast_fp16 = maximum(x = var_5239_cast_fp16, y = var_5240_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_159_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_163_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5243_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5243_cast_fp16")];
            tensor<int32, [1]> var_5248_axes_0 = const()[name = string("op_5248_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5248_keep_dims_0 = const()[name = string("op_5248_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5248_cast_fp16 = reduce_mean(axes = var_5248_axes_0, keep_dims = var_5248_keep_dims_0, x = var_5243_cast_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5249_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5249_cast_fp16")];
            fp32 var_5250_epsilon_0 = const()[name = string("op_5250_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5250_cast_fp16 = inverse(epsilon = var_5250_epsilon_0, x = var_5249_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp16 var_5251_to_fp16 = const()[name = string("op_5251_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5252_cast_fp16 = mul(x = var_5250_cast_fp16, y = var_5251_to_fp16)[name = string("op_5252_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5254_cast_fp16 = add(x = var_5248_cast_fp16, y = var_5252_cast_fp16)[name = string("op_5254_cast_fp16")];
            fp32 var_5255_epsilon_0 = const()[name = string("op_5255_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5255_cast_fp16 = rsqrt(epsilon = var_5255_epsilon_0, x = var_5254_cast_fp16)[name = string("op_5255_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_137_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5255_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713536)))];
            tensor<fp16, [1, 2520, 12, 64]> x_209_cast_fp16 = mul(x = out_137_cast_fp16, y = t_encoder_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105896064))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106485952))))[name = string("t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [4]> var_5265 = const()[name = string("op_5265"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_205_cast_fp16 = reshape(shape = var_5265, x = linear_80_cast_fp16)[name = string("x_205_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5267_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5267_cast_fp16")];
            bool var_5270 = const()[name = string("op_5270"), val = bool(true)];
            tensor<int32, [1]> var_5271_axes_0 = const()[name = string("op_5271_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5271_cast_fp16 = reduce_max(axes = var_5271_axes_0, keep_dims = var_5270, x = var_5267_cast_fp16)[name = string("op_5271_cast_fp16")];
            fp16 var_5272_to_fp16 = const()[name = string("op_5272_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_165_cast_fp16 = maximum(x = var_5271_cast_fp16, y = var_5272_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_161_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_165_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5275_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5275_cast_fp16")];
            tensor<int32, [1]> var_5280_axes_0 = const()[name = string("op_5280_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5280_keep_dims_0 = const()[name = string("op_5280_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5280_cast_fp16 = reduce_mean(axes = var_5280_axes_0, keep_dims = var_5280_keep_dims_0, x = var_5275_cast_fp16)[name = string("op_5280_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5281_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp32 var_5282_epsilon_0 = const()[name = string("op_5282_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5282_cast_fp16 = inverse(epsilon = var_5282_epsilon_0, x = var_5281_cast_fp16)[name = string("op_5282_cast_fp16")];
            fp16 var_5283_to_fp16 = const()[name = string("op_5283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5284_cast_fp16 = mul(x = var_5282_cast_fp16, y = var_5283_to_fp16)[name = string("op_5284_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5286_cast_fp16 = add(x = var_5280_cast_fp16, y = var_5284_cast_fp16)[name = string("op_5286_cast_fp16")];
            fp32 var_5287_epsilon_0 = const()[name = string("op_5287_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5287_cast_fp16 = rsqrt(epsilon = var_5287_epsilon_0, x = var_5286_cast_fp16)[name = string("op_5287_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_45_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5287_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<int32, [4]> z_89_begin_0 = const()[name = string("z_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_89_end_0 = const()[name = string("z_89_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_89_end_mask_0 = const()[name = string("z_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_89_cast_fp16 = slice_by_index(begin = z_89_begin_0, end = z_89_end_0, end_mask = z_89_end_mask_0, x = x_207_cast_fp16)[name = string("z_89_cast_fp16")];
            tensor<int32, [4]> var_5298_begin_0 = const()[name = string("op_5298_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5298_end_0 = const()[name = string("op_5298_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5298_end_mask_0 = const()[name = string("op_5298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5298_cast_fp16 = slice_by_index(begin = var_5298_begin_0, end = var_5298_end_0, end_mask = var_5298_end_mask_0, x = z_89_cast_fp16)[name = string("op_5298_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5299_cast_fp16 = mul(x = var_5298_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_5299_cast_fp16")];
            tensor<int32, [4]> var_5304_begin_0 = const()[name = string("op_5304_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5304_end_0 = const()[name = string("op_5304_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5304_end_mask_0 = const()[name = string("op_5304_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5304_cast_fp16 = slice_by_index(begin = var_5304_begin_0, end = var_5304_end_0, end_mask = var_5304_end_mask_0, x = z_89_cast_fp16)[name = string("op_5304_cast_fp16")];
            int32 var_5306 = const()[name = string("op_5306"), val = int32(-1)];
            bool rot_89_interleave_0 = const()[name = string("rot_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_89_cast_fp16 = concat(axis = var_5306, interleave = rot_89_interleave_0, values = (var_5299_cast_fp16, var_5304_cast_fp16))[name = string("rot_89_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5308_cast_fp16 = mul(x = z_89_cast_fp16, y = c_1_cast_fp16)[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5309_cast_fp16 = mul(x = rot_89_cast_fp16, y = s_9_cast_fp16)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5311_cast_fp16 = add(x = var_5308_cast_fp16, y = var_5309_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<int32, [4]> z_91_begin_0 = const()[name = string("z_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_91_end_0 = const()[name = string("z_91_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_91_end_mask_0 = const()[name = string("z_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_91_cast_fp16 = slice_by_index(begin = z_91_begin_0, end = z_91_end_0, end_mask = z_91_end_mask_0, x = x_207_cast_fp16)[name = string("z_91_cast_fp16")];
            tensor<int32, [4]> var_5321_begin_0 = const()[name = string("op_5321_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5321_end_0 = const()[name = string("op_5321_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5321_end_mask_0 = const()[name = string("op_5321_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5321_cast_fp16 = slice_by_index(begin = var_5321_begin_0, end = var_5321_end_0, end_mask = var_5321_end_mask_0, x = z_91_cast_fp16)[name = string("op_5321_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5322_cast_fp16 = mul(x = var_5321_cast_fp16, y = const_45_promoted_to_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, 0])];
            tensor<int32, [4]> var_5327_end_0 = const()[name = string("op_5327_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5327_end_mask_0 = const()[name = string("op_5327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 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 = z_91_cast_fp16)[name = string("op_5327_cast_fp16")];
            int32 var_5329 = const()[name = string("op_5329"), val = int32(-1)];
            bool rot_91_interleave_0 = const()[name = string("rot_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_91_cast_fp16 = concat(axis = var_5329, interleave = rot_91_interleave_0, values = (var_5322_cast_fp16, var_5327_cast_fp16))[name = string("rot_91_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5331_cast_fp16 = mul(x = z_91_cast_fp16, y = c_cast_fp16)[name = string("op_5331_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5332_cast_fp16 = mul(x = rot_91_cast_fp16, y = s_11_cast_fp16)[name = string("op_5332_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5334_cast_fp16 = add(x = var_5331_cast_fp16, y = var_5332_cast_fp16)[name = string("op_5334_cast_fp16")];
            int32 var_5336 = const()[name = string("op_5336"), val = int32(-1)];
            bool var_5337_interleave_0 = const()[name = string("op_5337_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_5337_cast_fp16 = concat(axis = var_5336, interleave = var_5337_interleave_0, values = (var_5311_cast_fp16, var_5334_cast_fp16))[name = string("op_5337_cast_fp16")];
            tensor<int32, [4]> z_93_begin_0 = const()[name = string("z_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_93_end_0 = const()[name = string("z_93_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_93_end_mask_0 = const()[name = string("z_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_93_cast_fp16 = slice_by_index(begin = z_93_begin_0, end = z_93_end_0, end_mask = z_93_end_mask_0, x = x_209_cast_fp16)[name = string("z_93_cast_fp16")];
            tensor<int32, [4]> var_5350_begin_0 = const()[name = string("op_5350_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5350_end_0 = const()[name = string("op_5350_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5350_end_mask_0 = const()[name = string("op_5350_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5350_cast_fp16 = slice_by_index(begin = var_5350_begin_0, end = var_5350_end_0, end_mask = var_5350_end_mask_0, x = z_93_cast_fp16)[name = string("op_5350_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5351_cast_fp16 = mul(x = var_5350_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_5351_cast_fp16")];
            tensor<int32, [4]> var_5356_begin_0 = const()[name = string("op_5356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5356_end_0 = const()[name = string("op_5356_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5356_end_mask_0 = const()[name = string("op_5356_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5356_cast_fp16 = slice_by_index(begin = var_5356_begin_0, end = var_5356_end_0, end_mask = var_5356_end_mask_0, x = z_93_cast_fp16)[name = string("op_5356_cast_fp16")];
            int32 var_5358 = const()[name = string("op_5358"), val = int32(-1)];
            bool rot_93_interleave_0 = const()[name = string("rot_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_93_cast_fp16 = concat(axis = var_5358, interleave = rot_93_interleave_0, values = (var_5351_cast_fp16, var_5356_cast_fp16))[name = string("rot_93_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5360_cast_fp16 = mul(x = z_93_cast_fp16, y = c_1_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5361_cast_fp16 = mul(x = rot_93_cast_fp16, y = s_9_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5363_cast_fp16 = add(x = var_5360_cast_fp16, y = var_5361_cast_fp16)[name = string("op_5363_cast_fp16")];
            tensor<int32, [4]> z_95_begin_0 = const()[name = string("z_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_95_end_0 = const()[name = string("z_95_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_95_end_mask_0 = const()[name = string("z_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_95_cast_fp16 = slice_by_index(begin = z_95_begin_0, end = z_95_end_0, end_mask = z_95_end_mask_0, x = x_209_cast_fp16)[name = string("z_95_cast_fp16")];
            tensor<int32, [4]> var_5373_begin_0 = const()[name = string("op_5373_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5373_end_0 = const()[name = string("op_5373_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5373_end_mask_0 = const()[name = string("op_5373_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5373_cast_fp16 = slice_by_index(begin = var_5373_begin_0, end = var_5373_end_0, end_mask = var_5373_end_mask_0, x = z_95_cast_fp16)[name = string("op_5373_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5374_cast_fp16 = mul(x = var_5373_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [4]> var_5379_begin_0 = const()[name = string("op_5379_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5379_end_0 = const()[name = string("op_5379_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5379_end_mask_0 = const()[name = string("op_5379_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5379_cast_fp16 = slice_by_index(begin = var_5379_begin_0, end = var_5379_end_0, end_mask = var_5379_end_mask_0, x = z_95_cast_fp16)[name = string("op_5379_cast_fp16")];
            int32 var_5381 = const()[name = string("op_5381"), val = int32(-1)];
            bool rot_95_interleave_0 = const()[name = string("rot_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_95_cast_fp16 = concat(axis = var_5381, interleave = rot_95_interleave_0, values = (var_5374_cast_fp16, var_5379_cast_fp16))[name = string("rot_95_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5383_cast_fp16 = mul(x = z_95_cast_fp16, y = c_cast_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5384_cast_fp16 = mul(x = rot_95_cast_fp16, y = s_11_cast_fp16)[name = string("op_5384_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5386_cast_fp16 = add(x = var_5383_cast_fp16, y = var_5384_cast_fp16)[name = string("op_5386_cast_fp16")];
            int32 var_5388 = const()[name = string("op_5388"), val = int32(-1)];
            bool var_5389_interleave_0 = const()[name = string("op_5389_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_5389_cast_fp16 = concat(axis = var_5388, interleave = var_5389_interleave_0, values = (var_5363_cast_fp16, var_5386_cast_fp16))[name = string("op_5389_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5399_transpose_x_0 = const()[name = string("op_5399_transpose_x_0"), val = bool(false)];
            bool var_5399_transpose_y_0 = const()[name = string("op_5399_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_87 = transpose(perm = transpose_87_perm_0, x = var_5389_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_86 = transpose(perm = transpose_86_perm_0, x = var_5337_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 12, 2520, 2520]> var_5399_cast_fp16 = matmul(transpose_x = var_5399_transpose_x_0, transpose_y = var_5399_transpose_y_0, x = transpose_86, y = transpose_87)[name = string("op_5399_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_5401_cast_fp16 = add(x = var_5399_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5401_cast_fp16")];
            int32 var_5402 = const()[name = string("op_5402"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_161_cast_fp16 = softmax(axis = var_5402, x = var_5401_cast_fp16)[name = string("w_161_cast_fp16")];
            bool var_5405_transpose_x_0 = const()[name = string("op_5405_transpose_x_0"), val = bool(false)];
            bool var_5405_transpose_y_0 = const()[name = string("op_5405_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = v_45_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 12, 2520, 64]> var_5405_cast_fp16 = matmul(transpose_x = var_5405_transpose_x_0, transpose_y = var_5405_transpose_y_0, x = w_161_cast_fp16, y = v_47_cast_fp16)[name = string("op_5405_cast_fp16")];
            tensor<int32, [4]> var_5408_perm_0 = const()[name = string("op_5408_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5412 = const()[name = string("op_5412"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_5408_cast_fp16 = transpose(perm = var_5408_perm_0, x = var_5405_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 2520, 768]> input_115_cast_fp16 = reshape(shape = var_5412, x = var_5408_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106487552))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107077440))))[name = string("t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_81_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5417_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5417_cast_fp16")];
            bool var_5420 = const()[name = string("op_5420"), val = bool(true)];
            tensor<int32, [1]> var_5421_axes_0 = const()[name = string("op_5421_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5421_cast_fp16 = reduce_max(axes = var_5421_axes_0, keep_dims = var_5420, x = var_5417_cast_fp16)[name = string("op_5421_cast_fp16")];
            fp16 var_5422_to_fp16 = const()[name = string("op_5422_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_167_cast_fp16 = maximum(x = var_5421_cast_fp16, y = var_5422_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_163_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_167_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5425_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5425_cast_fp16")];
            tensor<int32, [1]> var_5430_axes_0 = const()[name = string("op_5430_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5430_keep_dims_0 = const()[name = string("op_5430_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5430_cast_fp16 = reduce_mean(axes = var_5430_axes_0, keep_dims = var_5430_keep_dims_0, x = var_5425_cast_fp16)[name = string("op_5430_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5431_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5431_cast_fp16")];
            fp32 var_5432_epsilon_0 = const()[name = string("op_5432_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5432_cast_fp16 = inverse(epsilon = var_5432_epsilon_0, x = var_5431_cast_fp16)[name = string("op_5432_cast_fp16")];
            fp16 var_5433_to_fp16 = const()[name = string("op_5433_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = string("op_5434_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5436_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5434_cast_fp16)[name = string("op_5436_cast_fp16")];
            fp32 var_5437_epsilon_0 = const()[name = string("op_5437_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5437_cast_fp16 = rsqrt(epsilon = var_5437_epsilon_0, x = var_5436_cast_fp16)[name = string("op_5437_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_139_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5437_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107079040)))];
            tensor<fp16, [1, 2520, 768]> var_5439_cast_fp16 = mul(x = out_139_cast_fp16, y = t_encoder_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_5439_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_213_cast_fp16 = add(x = x_199_cast_fp16, y = var_5439_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5442_cast_fp16 = abs(x = x_213_cast_fp16)[name = string("op_5442_cast_fp16")];
            bool var_5445 = const()[name = string("op_5445"), val = bool(true)];
            tensor<int32, [1]> var_5446_axes_0 = const()[name = string("op_5446_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5446_cast_fp16 = reduce_max(axes = var_5446_axes_0, keep_dims = var_5445, x = var_5442_cast_fp16)[name = string("op_5446_cast_fp16")];
            fp16 var_5447_to_fp16 = const()[name = string("op_5447_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_169_cast_fp16 = maximum(x = var_5446_cast_fp16, y = var_5447_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_165_cast_fp16 = real_div(x = x_213_cast_fp16, y = s_169_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5450_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5450_cast_fp16")];
            tensor<int32, [1]> var_5455_axes_0 = const()[name = string("op_5455_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5455_keep_dims_0 = const()[name = string("op_5455_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5455_cast_fp16 = reduce_mean(axes = var_5455_axes_0, keep_dims = var_5455_keep_dims_0, x = var_5450_cast_fp16)[name = string("op_5455_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5456_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5456_cast_fp16")];
            fp32 var_5457_epsilon_0 = const()[name = string("op_5457_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5457_cast_fp16 = inverse(epsilon = var_5457_epsilon_0, x = var_5456_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5459_cast_fp16 = mul(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("op_5459_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5461_cast_fp16 = add(x = var_5455_cast_fp16, y = var_5459_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, 2520, 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, 2520, 768]> out_141_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5462_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107080640)))];
            tensor<fp16, [1, 2520, 768]> input_117_cast_fp16 = mul(x = out_141_cast_fp16, y = t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107082240))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109441600))))[name = string("t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_82_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_82_cast_fp16")];
            string var_5469_mode_0 = const()[name = string("op_5469_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_5469_cast_fp16 = gelu(mode = var_5469_mode_0, x = linear_82_cast_fp16)[name = string("op_5469_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109447808))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111807168))))[name = string("t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_83_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_121_cast_fp16 = mul(x = var_5469_cast_fp16, y = linear_83_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111813376))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114172736))))[name = string("t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5477_cast_fp16 = abs(x = linear_84_cast_fp16)[name = string("op_5477_cast_fp16")];
            bool var_5480 = const()[name = string("op_5480"), val = bool(true)];
            tensor<int32, [1]> var_5481_axes_0 = const()[name = string("op_5481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5481_cast_fp16 = reduce_max(axes = var_5481_axes_0, keep_dims = var_5480, x = var_5477_cast_fp16)[name = string("op_5481_cast_fp16")];
            fp16 var_5482_to_fp16 = const()[name = string("op_5482_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_171_cast_fp16 = maximum(x = var_5481_cast_fp16, y = var_5482_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_167_cast_fp16 = real_div(x = linear_84_cast_fp16, y = s_171_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5485_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5485_cast_fp16")];
            tensor<int32, [1]> var_5490_axes_0 = const()[name = string("op_5490_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5490_keep_dims_0 = const()[name = string("op_5490_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5490_cast_fp16 = reduce_mean(axes = var_5490_axes_0, keep_dims = var_5490_keep_dims_0, x = var_5485_cast_fp16)[name = string("op_5490_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5491_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5491_cast_fp16")];
            fp32 var_5492_epsilon_0 = const()[name = string("op_5492_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5492_cast_fp16 = inverse(epsilon = var_5492_epsilon_0, x = var_5491_cast_fp16)[name = string("op_5492_cast_fp16")];
            fp16 var_5493_to_fp16 = const()[name = string("op_5493_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5494_cast_fp16 = mul(x = var_5492_cast_fp16, y = var_5493_to_fp16)[name = string("op_5494_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5496_cast_fp16 = add(x = var_5490_cast_fp16, y = var_5494_cast_fp16)[name = string("op_5496_cast_fp16")];
            fp32 var_5497_epsilon_0 = const()[name = string("op_5497_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5497_cast_fp16 = rsqrt(epsilon = var_5497_epsilon_0, x = var_5496_cast_fp16)[name = string("op_5497_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_143_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5497_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114174336)))];
            tensor<fp16, [1, 2520, 768]> var_5499_cast_fp16 = mul(x = out_143_cast_fp16, y = t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_217_cast_fp16 = add(x = x_213_cast_fp16, y = var_5499_cast_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5502_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_5502_cast_fp16")];
            bool var_5505 = const()[name = string("op_5505"), val = bool(true)];
            tensor<int32, [1]> var_5506_axes_0 = const()[name = string("op_5506_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5506_cast_fp16 = reduce_max(axes = var_5506_axes_0, keep_dims = var_5505, x = var_5502_cast_fp16)[name = string("op_5506_cast_fp16")];
            fp16 var_5507_to_fp16 = const()[name = string("op_5507_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_173_cast_fp16 = maximum(x = var_5506_cast_fp16, y = var_5507_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_169_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_173_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5510_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [1]> var_5515_axes_0 = const()[name = string("op_5515_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5515_keep_dims_0 = const()[name = string("op_5515_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5515_cast_fp16 = reduce_mean(axes = var_5515_axes_0, keep_dims = var_5515_keep_dims_0, x = var_5510_cast_fp16)[name = string("op_5515_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5516_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5516_cast_fp16")];
            fp32 var_5517_epsilon_0 = const()[name = string("op_5517_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5517_cast_fp16 = inverse(epsilon = var_5517_epsilon_0, x = var_5516_cast_fp16)[name = string("op_5517_cast_fp16")];
            fp16 var_5518_to_fp16 = const()[name = string("op_5518_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5519_cast_fp16 = mul(x = var_5517_cast_fp16, y = var_5518_to_fp16)[name = string("op_5519_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5521_cast_fp16 = add(x = var_5515_cast_fp16, y = var_5519_cast_fp16)[name = string("op_5521_cast_fp16")];
            fp32 var_5522_epsilon_0 = const()[name = string("op_5522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5522_cast_fp16 = rsqrt(epsilon = var_5522_epsilon_0, x = var_5521_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_145_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5522_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114175936)))];
            tensor<fp16, [1, 2520, 768]> input_123_cast_fp16 = mul(x = out_145_cast_fp16, y = t_encoder_layers_12_input_layernorm_weight_to_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114177536))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114767424))))[name = string("t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_85_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_5532 = const()[name = string("op_5532"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_219_cast_fp16 = reshape(shape = var_5532, x = linear_85_cast_fp16)[name = string("x_219_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5534_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_5534_cast_fp16")];
            bool var_5537 = const()[name = string("op_5537"), val = bool(true)];
            tensor<int32, [1]> var_5538_axes_0 = const()[name = string("op_5538_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5538_cast_fp16 = reduce_max(axes = var_5538_axes_0, keep_dims = var_5537, x = var_5534_cast_fp16)[name = string("op_5538_cast_fp16")];
            fp16 var_5539_to_fp16 = const()[name = string("op_5539_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_175_cast_fp16 = maximum(x = var_5538_cast_fp16, y = var_5539_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_171_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_175_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5542_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5542_cast_fp16")];
            tensor<int32, [1]> var_5547_axes_0 = const()[name = string("op_5547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5547_keep_dims_0 = const()[name = string("op_5547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5547_cast_fp16 = reduce_mean(axes = var_5547_axes_0, keep_dims = var_5547_keep_dims_0, x = var_5542_cast_fp16)[name = string("op_5547_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5548_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5548_cast_fp16")];
            fp32 var_5549_epsilon_0 = const()[name = string("op_5549_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5549_cast_fp16 = inverse(epsilon = var_5549_epsilon_0, x = var_5548_cast_fp16)[name = string("op_5549_cast_fp16")];
            fp16 var_5550_to_fp16 = const()[name = string("op_5550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5553_cast_fp16 = add(x = var_5547_cast_fp16, y = var_5551_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp32 var_5554_epsilon_0 = const()[name = string("op_5554_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5554_cast_fp16 = rsqrt(epsilon = var_5554_epsilon_0, x = var_5553_cast_fp16)[name = string("op_5554_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_147_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5554_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769024)))];
            tensor<fp16, [1, 2520, 12, 64]> x_225_cast_fp16 = mul(x = out_147_cast_fp16, y = t_encoder_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_225_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769216))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115359104))))[name = string("t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_5564 = const()[name = string("op_5564"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_221_cast_fp16 = reshape(shape = var_5564, x = linear_86_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5566_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_5566_cast_fp16")];
            bool var_5569 = const()[name = string("op_5569"), val = bool(true)];
            tensor<int32, [1]> var_5570_axes_0 = const()[name = string("op_5570_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5570_cast_fp16 = reduce_max(axes = var_5570_axes_0, keep_dims = var_5569, x = var_5566_cast_fp16)[name = string("op_5570_cast_fp16")];
            fp16 var_5571_to_fp16 = const()[name = string("op_5571_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_177_cast_fp16 = maximum(x = var_5570_cast_fp16, y = var_5571_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_173_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_177_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5574_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<int32, [1]> var_5579_axes_0 = const()[name = string("op_5579_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5579_keep_dims_0 = const()[name = string("op_5579_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5579_cast_fp16 = reduce_mean(axes = var_5579_axes_0, keep_dims = var_5579_keep_dims_0, x = var_5574_cast_fp16)[name = string("op_5579_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5580_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5581_cast_fp16 = inverse(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            fp16 var_5582_to_fp16 = const()[name = string("op_5582_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5583_cast_fp16 = mul(x = var_5581_cast_fp16, y = var_5582_to_fp16)[name = string("op_5583_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5585_cast_fp16 = add(x = var_5579_cast_fp16, y = var_5583_cast_fp16)[name = string("op_5585_cast_fp16")];
            fp32 var_5586_epsilon_0 = const()[name = string("op_5586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5586_cast_fp16 = rsqrt(epsilon = var_5586_epsilon_0, x = var_5585_cast_fp16)[name = string("op_5586_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_149_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5586_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360704)))];
            tensor<fp16, [1, 2520, 12, 64]> x_227_cast_fp16 = mul(x = out_149_cast_fp16, y = t_encoder_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360896))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115950784))))[name = string("t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_87_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<int32, [4]> var_5596 = const()[name = string("op_5596"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_223_cast_fp16 = reshape(shape = var_5596, x = linear_87_cast_fp16)[name = string("x_223_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5598_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_5598_cast_fp16")];
            bool var_5601 = const()[name = string("op_5601"), val = bool(true)];
            tensor<int32, [1]> var_5602_axes_0 = const()[name = string("op_5602_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5602_cast_fp16 = reduce_max(axes = var_5602_axes_0, keep_dims = var_5601, x = var_5598_cast_fp16)[name = string("op_5602_cast_fp16")];
            fp16 var_5603_to_fp16 = const()[name = string("op_5603_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_179_cast_fp16 = maximum(x = var_5602_cast_fp16, y = var_5603_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_175_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_179_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5606_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<int32, [1]> var_5611_axes_0 = const()[name = string("op_5611_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5611_keep_dims_0 = const()[name = string("op_5611_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5611_cast_fp16 = reduce_mean(axes = var_5611_axes_0, keep_dims = var_5611_keep_dims_0, x = var_5606_cast_fp16)[name = string("op_5611_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5612_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5612_cast_fp16")];
            fp32 var_5613_epsilon_0 = const()[name = string("op_5613_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5613_cast_fp16 = inverse(epsilon = var_5613_epsilon_0, x = var_5612_cast_fp16)[name = string("op_5613_cast_fp16")];
            fp16 var_5614_to_fp16 = const()[name = string("op_5614_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5615_cast_fp16 = mul(x = var_5613_cast_fp16, y = var_5614_to_fp16)[name = string("op_5615_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5617_cast_fp16 = add(x = var_5611_cast_fp16, y = var_5615_cast_fp16)[name = string("op_5617_cast_fp16")];
            fp32 var_5618_epsilon_0 = const()[name = string("op_5618_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5618_cast_fp16 = rsqrt(epsilon = var_5618_epsilon_0, x = var_5617_cast_fp16)[name = string("op_5618_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_49_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5618_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<int32, [4]> z_97_begin_0 = const()[name = string("z_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_97_end_0 = const()[name = string("z_97_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_97_end_mask_0 = const()[name = string("z_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_97_cast_fp16 = slice_by_index(begin = z_97_begin_0, end = z_97_end_0, end_mask = z_97_end_mask_0, x = x_225_cast_fp16)[name = string("z_97_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> 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 = z_97_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = z_97_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool rot_97_interleave_0 = const()[name = string("rot_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_97_cast_fp16 = concat(axis = var_5637, interleave = rot_97_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("rot_97_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5639_cast_fp16 = mul(x = z_97_cast_fp16, y = c_1_cast_fp16)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5640_cast_fp16 = mul(x = rot_97_cast_fp16, y = s_9_cast_fp16)[name = string("op_5640_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5642_cast_fp16 = add(x = var_5639_cast_fp16, y = var_5640_cast_fp16)[name = string("op_5642_cast_fp16")];
            tensor<int32, [4]> z_99_begin_0 = const()[name = string("z_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_99_end_0 = const()[name = string("z_99_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_99_end_mask_0 = const()[name = string("z_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_99_cast_fp16 = slice_by_index(begin = z_99_begin_0, end = z_99_end_0, end_mask = z_99_end_mask_0, x = x_225_cast_fp16)[name = string("z_99_cast_fp16")];
            tensor<int32, [4]> var_5652_begin_0 = const()[name = string("op_5652_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5652_end_0 = const()[name = string("op_5652_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5652_end_mask_0 = const()[name = string("op_5652_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5652_cast_fp16 = slice_by_index(begin = var_5652_begin_0, end = var_5652_end_0, end_mask = var_5652_end_mask_0, x = z_99_cast_fp16)[name = string("op_5652_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5653_cast_fp16 = mul(x = var_5652_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_5653_cast_fp16")];
            tensor<int32, [4]> var_5658_begin_0 = const()[name = string("op_5658_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5658_end_0 = const()[name = string("op_5658_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5658_end_mask_0 = const()[name = string("op_5658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5658_cast_fp16 = slice_by_index(begin = var_5658_begin_0, end = var_5658_end_0, end_mask = var_5658_end_mask_0, x = z_99_cast_fp16)[name = string("op_5658_cast_fp16")];
            int32 var_5660 = const()[name = string("op_5660"), val = int32(-1)];
            bool rot_99_interleave_0 = const()[name = string("rot_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_99_cast_fp16 = concat(axis = var_5660, interleave = rot_99_interleave_0, values = (var_5653_cast_fp16, var_5658_cast_fp16))[name = string("rot_99_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5662_cast_fp16 = mul(x = z_99_cast_fp16, y = c_cast_fp16)[name = string("op_5662_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5663_cast_fp16 = mul(x = rot_99_cast_fp16, y = s_11_cast_fp16)[name = string("op_5663_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5665_cast_fp16 = add(x = var_5662_cast_fp16, y = var_5663_cast_fp16)[name = string("op_5665_cast_fp16")];
            int32 var_5667 = const()[name = string("op_5667"), val = int32(-1)];
            bool var_5668_interleave_0 = const()[name = string("op_5668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_5668_cast_fp16 = concat(axis = var_5667, interleave = var_5668_interleave_0, values = (var_5642_cast_fp16, var_5665_cast_fp16))[name = string("op_5668_cast_fp16")];
            tensor<int32, [4]> z_101_begin_0 = const()[name = string("z_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_101_end_0 = const()[name = string("z_101_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_101_end_mask_0 = const()[name = string("z_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_101_cast_fp16 = slice_by_index(begin = z_101_begin_0, end = z_101_end_0, end_mask = z_101_end_mask_0, x = x_227_cast_fp16)[name = string("z_101_cast_fp16")];
            tensor<int32, [4]> var_5681_begin_0 = const()[name = string("op_5681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5681_end_0 = const()[name = string("op_5681_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5681_end_mask_0 = const()[name = string("op_5681_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5681_cast_fp16 = slice_by_index(begin = var_5681_begin_0, end = var_5681_end_0, end_mask = var_5681_end_mask_0, x = z_101_cast_fp16)[name = string("op_5681_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5682_cast_fp16 = mul(x = var_5681_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_5682_cast_fp16")];
            tensor<int32, [4]> var_5687_begin_0 = const()[name = string("op_5687_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5687_end_0 = const()[name = string("op_5687_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5687_end_mask_0 = const()[name = string("op_5687_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5687_cast_fp16 = slice_by_index(begin = var_5687_begin_0, end = var_5687_end_0, end_mask = var_5687_end_mask_0, x = z_101_cast_fp16)[name = string("op_5687_cast_fp16")];
            int32 var_5689 = const()[name = string("op_5689"), val = int32(-1)];
            bool rot_101_interleave_0 = const()[name = string("rot_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_101_cast_fp16 = concat(axis = var_5689, interleave = rot_101_interleave_0, values = (var_5682_cast_fp16, var_5687_cast_fp16))[name = string("rot_101_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5691_cast_fp16 = mul(x = z_101_cast_fp16, y = c_1_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5692_cast_fp16 = mul(x = rot_101_cast_fp16, y = s_9_cast_fp16)[name = string("op_5692_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5694_cast_fp16 = add(x = var_5691_cast_fp16, y = var_5692_cast_fp16)[name = string("op_5694_cast_fp16")];
            tensor<int32, [4]> z_103_begin_0 = const()[name = string("z_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_103_end_0 = const()[name = string("z_103_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_103_end_mask_0 = const()[name = string("z_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_103_cast_fp16 = slice_by_index(begin = z_103_begin_0, end = z_103_end_0, end_mask = z_103_end_mask_0, x = x_227_cast_fp16)[name = string("z_103_cast_fp16")];
            tensor<int32, [4]> var_5704_begin_0 = const()[name = string("op_5704_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5704_end_0 = const()[name = string("op_5704_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5704_end_mask_0 = const()[name = string("op_5704_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5704_cast_fp16 = slice_by_index(begin = var_5704_begin_0, end = var_5704_end_0, end_mask = var_5704_end_mask_0, x = z_103_cast_fp16)[name = string("op_5704_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5705_cast_fp16 = mul(x = var_5704_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_5705_cast_fp16")];
            tensor<int32, [4]> var_5710_begin_0 = const()[name = string("op_5710_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5710_end_0 = const()[name = string("op_5710_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5710_end_mask_0 = const()[name = string("op_5710_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5710_cast_fp16 = slice_by_index(begin = var_5710_begin_0, end = var_5710_end_0, end_mask = var_5710_end_mask_0, x = z_103_cast_fp16)[name = string("op_5710_cast_fp16")];
            int32 var_5712 = const()[name = string("op_5712"), val = int32(-1)];
            bool rot_103_interleave_0 = const()[name = string("rot_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_103_cast_fp16 = concat(axis = var_5712, interleave = rot_103_interleave_0, values = (var_5705_cast_fp16, var_5710_cast_fp16))[name = string("rot_103_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5714_cast_fp16 = mul(x = z_103_cast_fp16, y = c_cast_fp16)[name = string("op_5714_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5715_cast_fp16 = mul(x = rot_103_cast_fp16, y = s_11_cast_fp16)[name = string("op_5715_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5717_cast_fp16 = add(x = var_5714_cast_fp16, y = var_5715_cast_fp16)[name = string("op_5717_cast_fp16")];
            int32 var_5719 = const()[name = string("op_5719"), val = int32(-1)];
            bool var_5720_interleave_0 = const()[name = string("op_5720_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_5720_cast_fp16 = concat(axis = var_5719, interleave = var_5720_interleave_0, values = (var_5694_cast_fp16, var_5717_cast_fp16))[name = string("op_5720_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5730_transpose_x_0 = const()[name = string("op_5730_transpose_x_0"), val = bool(false)];
            bool var_5730_transpose_y_0 = const()[name = string("op_5730_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_89 = transpose(perm = transpose_89_perm_0, x = var_5720_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_88 = transpose(perm = transpose_88_perm_0, x = var_5668_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 12, 2520, 2520]> var_5730_cast_fp16 = matmul(transpose_x = var_5730_transpose_x_0, transpose_y = var_5730_transpose_y_0, x = transpose_88, y = transpose_89)[name = string("op_5730_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_5732_cast_fp16 = add(x = var_5730_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5732_cast_fp16")];
            int32 var_5733 = const()[name = string("op_5733"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_175_cast_fp16 = softmax(axis = var_5733, x = var_5732_cast_fp16)[name = string("w_175_cast_fp16")];
            bool var_5736_transpose_x_0 = const()[name = string("op_5736_transpose_x_0"), val = bool(false)];
            bool var_5736_transpose_y_0 = const()[name = string("op_5736_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = v_49_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 12, 2520, 64]> var_5736_cast_fp16 = matmul(transpose_x = var_5736_transpose_x_0, transpose_y = var_5736_transpose_y_0, x = w_175_cast_fp16, y = v_51_cast_fp16)[name = string("op_5736_cast_fp16")];
            tensor<int32, [4]> var_5739_perm_0 = const()[name = string("op_5739_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5743 = const()[name = string("op_5743"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_5739_cast_fp16 = transpose(perm = var_5739_perm_0, x = var_5736_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 2520, 768]> input_125_cast_fp16 = reshape(shape = var_5743, x = var_5739_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115952384))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116542272))))[name = string("t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5748_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5748_cast_fp16")];
            bool var_5751 = const()[name = string("op_5751"), val = bool(true)];
            tensor<int32, [1]> var_5752_axes_0 = const()[name = string("op_5752_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5752_cast_fp16 = reduce_max(axes = var_5752_axes_0, keep_dims = var_5751, x = var_5748_cast_fp16)[name = string("op_5752_cast_fp16")];
            fp16 var_5753_to_fp16 = const()[name = string("op_5753_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_181_cast_fp16 = maximum(x = var_5752_cast_fp16, y = var_5753_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_177_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_181_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5756_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5756_cast_fp16")];
            tensor<int32, [1]> var_5761_axes_0 = const()[name = string("op_5761_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5761_keep_dims_0 = const()[name = string("op_5761_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5761_cast_fp16 = reduce_mean(axes = var_5761_axes_0, keep_dims = var_5761_keep_dims_0, x = var_5756_cast_fp16)[name = string("op_5761_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5762_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5762_cast_fp16")];
            fp32 var_5763_epsilon_0 = const()[name = string("op_5763_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5763_cast_fp16 = inverse(epsilon = var_5763_epsilon_0, x = var_5762_cast_fp16)[name = string("op_5763_cast_fp16")];
            fp16 var_5764_to_fp16 = const()[name = string("op_5764_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5765_cast_fp16 = mul(x = var_5763_cast_fp16, y = var_5764_to_fp16)[name = string("op_5765_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5767_cast_fp16 = add(x = var_5761_cast_fp16, y = var_5765_cast_fp16)[name = string("op_5767_cast_fp16")];
            fp32 var_5768_epsilon_0 = const()[name = string("op_5768_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5768_cast_fp16 = rsqrt(epsilon = var_5768_epsilon_0, x = var_5767_cast_fp16)[name = string("op_5768_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_151_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5768_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116543872)))];
            tensor<fp16, [1, 2520, 768]> var_5770_cast_fp16 = mul(x = out_151_cast_fp16, y = t_encoder_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_5770_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_231_cast_fp16 = add(x = x_217_cast_fp16, y = var_5770_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5773_cast_fp16 = abs(x = x_231_cast_fp16)[name = string("op_5773_cast_fp16")];
            bool var_5776 = const()[name = string("op_5776"), val = bool(true)];
            tensor<int32, [1]> var_5777_axes_0 = const()[name = string("op_5777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5777_cast_fp16 = reduce_max(axes = var_5777_axes_0, keep_dims = var_5776, x = var_5773_cast_fp16)[name = string("op_5777_cast_fp16")];
            fp16 var_5778_to_fp16 = const()[name = string("op_5778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_183_cast_fp16 = maximum(x = var_5777_cast_fp16, y = var_5778_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_179_cast_fp16 = real_div(x = x_231_cast_fp16, y = s_183_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5781_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5781_cast_fp16")];
            tensor<int32, [1]> var_5786_axes_0 = const()[name = string("op_5786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5786_keep_dims_0 = const()[name = string("op_5786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5786_cast_fp16 = reduce_mean(axes = var_5786_axes_0, keep_dims = var_5786_keep_dims_0, x = var_5781_cast_fp16)[name = string("op_5786_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5787_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp32 var_5788_epsilon_0 = const()[name = string("op_5788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5788_cast_fp16 = inverse(epsilon = var_5788_epsilon_0, x = var_5787_cast_fp16)[name = string("op_5788_cast_fp16")];
            fp16 var_5789_to_fp16 = const()[name = string("op_5789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5790_cast_fp16 = mul(x = var_5788_cast_fp16, y = var_5789_to_fp16)[name = string("op_5790_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5792_cast_fp16 = add(x = var_5786_cast_fp16, y = var_5790_cast_fp16)[name = string("op_5792_cast_fp16")];
            fp32 var_5793_epsilon_0 = const()[name = string("op_5793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5793_cast_fp16 = rsqrt(epsilon = var_5793_epsilon_0, x = var_5792_cast_fp16)[name = string("op_5793_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_153_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5793_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116545472)))];
            tensor<fp16, [1, 2520, 768]> input_127_cast_fp16 = mul(x = out_153_cast_fp16, y = t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116547072))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118906432))))[name = string("t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_89_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_89_cast_fp16")];
            string var_5800_mode_0 = const()[name = string("op_5800_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_5800_cast_fp16 = gelu(mode = var_5800_mode_0, x = linear_89_cast_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118912640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121272000))))[name = string("t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_90_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_131_cast_fp16 = mul(x = var_5800_cast_fp16, y = linear_90_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121278208))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123637568))))[name = string("t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5808_cast_fp16 = abs(x = linear_91_cast_fp16)[name = string("op_5808_cast_fp16")];
            bool var_5811 = const()[name = string("op_5811"), val = bool(true)];
            tensor<int32, [1]> var_5812_axes_0 = const()[name = string("op_5812_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5812_cast_fp16 = reduce_max(axes = var_5812_axes_0, keep_dims = var_5811, x = var_5808_cast_fp16)[name = string("op_5812_cast_fp16")];
            fp16 var_5813_to_fp16 = const()[name = string("op_5813_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_185_cast_fp16 = maximum(x = var_5812_cast_fp16, y = var_5813_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_181_cast_fp16 = real_div(x = linear_91_cast_fp16, y = s_185_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5816_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5816_cast_fp16")];
            tensor<int32, [1]> var_5821_axes_0 = const()[name = string("op_5821_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5821_keep_dims_0 = const()[name = string("op_5821_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5821_cast_fp16 = reduce_mean(axes = var_5821_axes_0, keep_dims = var_5821_keep_dims_0, x = var_5816_cast_fp16)[name = string("op_5821_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5822_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5822_cast_fp16")];
            fp32 var_5823_epsilon_0 = const()[name = string("op_5823_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5823_cast_fp16 = inverse(epsilon = var_5823_epsilon_0, x = var_5822_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp16 var_5824_to_fp16 = const()[name = string("op_5824_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5825_cast_fp16 = mul(x = var_5823_cast_fp16, y = var_5824_to_fp16)[name = string("op_5825_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5827_cast_fp16 = add(x = var_5821_cast_fp16, y = var_5825_cast_fp16)[name = string("op_5827_cast_fp16")];
            fp32 var_5828_epsilon_0 = const()[name = string("op_5828_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5828_cast_fp16 = rsqrt(epsilon = var_5828_epsilon_0, x = var_5827_cast_fp16)[name = string("op_5828_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_155_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5828_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123639168)))];
            tensor<fp16, [1, 2520, 768]> var_5830_cast_fp16 = mul(x = out_155_cast_fp16, y = t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5830_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_235_cast_fp16 = add(x = x_231_cast_fp16, y = var_5830_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5833_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_5833_cast_fp16")];
            bool var_5836 = const()[name = string("op_5836"), val = bool(true)];
            tensor<int32, [1]> var_5837_axes_0 = const()[name = string("op_5837_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_5837_cast_fp16 = reduce_max(axes = var_5837_axes_0, keep_dims = var_5836, x = var_5833_cast_fp16)[name = string("op_5837_cast_fp16")];
            fp16 var_5838_to_fp16 = const()[name = string("op_5838_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_187_cast_fp16 = maximum(x = var_5837_cast_fp16, y = var_5838_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_183_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_187_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_5841_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5841_cast_fp16")];
            tensor<int32, [1]> var_5846_axes_0 = const()[name = string("op_5846_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5846_keep_dims_0 = const()[name = string("op_5846_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_5846_cast_fp16 = reduce_mean(axes = var_5846_axes_0, keep_dims = var_5846_keep_dims_0, x = var_5841_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5847_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_5847_cast_fp16")];
            fp32 var_5848_epsilon_0 = const()[name = string("op_5848_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_5848_cast_fp16 = inverse(epsilon = var_5848_epsilon_0, x = var_5847_cast_fp16)[name = string("op_5848_cast_fp16")];
            fp16 var_5849_to_fp16 = const()[name = string("op_5849_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_5850_cast_fp16 = mul(x = var_5848_cast_fp16, y = var_5849_to_fp16)[name = string("op_5850_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_5852_cast_fp16 = add(x = var_5846_cast_fp16, y = var_5850_cast_fp16)[name = string("op_5852_cast_fp16")];
            fp32 var_5853_epsilon_0 = const()[name = string("op_5853_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_5853_cast_fp16 = rsqrt(epsilon = var_5853_epsilon_0, x = var_5852_cast_fp16)[name = string("op_5853_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_157_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5853_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123640768)))];
            tensor<fp16, [1, 2520, 768]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = t_encoder_layers_13_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123642368))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124232256))))[name = string("t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_237_cast_fp16 = reshape(shape = var_5863, x = linear_92_cast_fp16)[name = string("x_237_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5865_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_5865_cast_fp16")];
            bool var_5868 = const()[name = string("op_5868"), val = bool(true)];
            tensor<int32, [1]> var_5869_axes_0 = const()[name = string("op_5869_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5869_cast_fp16 = reduce_max(axes = var_5869_axes_0, keep_dims = var_5868, x = var_5865_cast_fp16)[name = string("op_5869_cast_fp16")];
            fp16 var_5870_to_fp16 = const()[name = string("op_5870_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_189_cast_fp16 = maximum(x = var_5869_cast_fp16, y = var_5870_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_185_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_189_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5873_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5873_cast_fp16")];
            tensor<int32, [1]> var_5878_axes_0 = const()[name = string("op_5878_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5878_keep_dims_0 = const()[name = string("op_5878_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5878_cast_fp16 = reduce_mean(axes = var_5878_axes_0, keep_dims = var_5878_keep_dims_0, x = var_5873_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5879_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_5879_cast_fp16")];
            fp32 var_5880_epsilon_0 = const()[name = string("op_5880_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5880_cast_fp16 = inverse(epsilon = var_5880_epsilon_0, x = var_5879_cast_fp16)[name = string("op_5880_cast_fp16")];
            fp16 var_5881_to_fp16 = const()[name = string("op_5881_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5882_cast_fp16 = mul(x = var_5880_cast_fp16, y = var_5881_to_fp16)[name = string("op_5882_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5884_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5882_cast_fp16)[name = string("op_5884_cast_fp16")];
            fp32 var_5885_epsilon_0 = const()[name = string("op_5885_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5885_cast_fp16 = rsqrt(epsilon = var_5885_epsilon_0, x = var_5884_cast_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_159_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5885_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910080)))];
            tensor<fp16, [1, 2520, 12, 64]> x_243_cast_fp16 = mul(x = out_159_cast_fp16, y = t_encoder_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_243_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124233856))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124823744))))[name = string("t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_93_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5895 = const()[name = string("op_5895"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_239_cast_fp16 = reshape(shape = var_5895, x = linear_93_cast_fp16)[name = string("x_239_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5897_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_5897_cast_fp16")];
            bool var_5900 = const()[name = string("op_5900"), val = bool(true)];
            tensor<int32, [1]> var_5901_axes_0 = const()[name = string("op_5901_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5901_cast_fp16 = reduce_max(axes = var_5901_axes_0, keep_dims = var_5900, x = var_5897_cast_fp16)[name = string("op_5901_cast_fp16")];
            fp16 var_5902_to_fp16 = const()[name = string("op_5902_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_191_cast_fp16 = maximum(x = var_5901_cast_fp16, y = var_5902_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_187_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_191_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5905_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_5905_cast_fp16")];
            tensor<int32, [1]> var_5910_axes_0 = const()[name = string("op_5910_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5910_keep_dims_0 = const()[name = string("op_5910_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5910_cast_fp16 = reduce_mean(axes = var_5910_axes_0, keep_dims = var_5910_keep_dims_0, x = var_5905_cast_fp16)[name = string("op_5910_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5911_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_5911_cast_fp16")];
            fp32 var_5912_epsilon_0 = const()[name = string("op_5912_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5912_cast_fp16 = inverse(epsilon = var_5912_epsilon_0, x = var_5911_cast_fp16)[name = string("op_5912_cast_fp16")];
            fp16 var_5913_to_fp16 = const()[name = string("op_5913_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5914_cast_fp16 = mul(x = var_5912_cast_fp16, y = var_5913_to_fp16)[name = string("op_5914_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5916_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5914_cast_fp16)[name = string("op_5916_cast_fp16")];
            fp32 var_5917_epsilon_0 = const()[name = string("op_5917_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5917_cast_fp16 = rsqrt(epsilon = var_5917_epsilon_0, x = var_5916_cast_fp16)[name = string("op_5917_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_161_cast_fp16 = mul(x = y_187_cast_fp16, y = var_5917_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501760)))];
            tensor<fp16, [1, 2520, 12, 64]> x_245_cast_fp16 = mul(x = out_161_cast_fp16, y = t_encoder_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124825344))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125415232))))[name = string("t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [4]> var_5927 = const()[name = string("op_5927"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_241_cast_fp16 = reshape(shape = var_5927, x = linear_94_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5929_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_5929_cast_fp16")];
            bool var_5932 = const()[name = string("op_5932"), val = bool(true)];
            tensor<int32, [1]> var_5933_axes_0 = const()[name = string("op_5933_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_5933_cast_fp16 = reduce_max(axes = var_5933_axes_0, keep_dims = var_5932, x = var_5929_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp16 var_5934_to_fp16 = const()[name = string("op_5934_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_193_cast_fp16 = maximum(x = var_5933_cast_fp16, y = var_5934_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_189_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_193_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_5937_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_5937_cast_fp16")];
            tensor<int32, [1]> var_5942_axes_0 = const()[name = string("op_5942_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5942_keep_dims_0 = const()[name = string("op_5942_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_5942_cast_fp16 = reduce_mean(axes = var_5942_axes_0, keep_dims = var_5942_keep_dims_0, x = var_5937_cast_fp16)[name = string("op_5942_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5943_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_5943_cast_fp16")];
            fp32 var_5944_epsilon_0 = const()[name = string("op_5944_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_5944_cast_fp16 = inverse(epsilon = var_5944_epsilon_0, x = var_5943_cast_fp16)[name = string("op_5944_cast_fp16")];
            fp16 var_5945_to_fp16 = const()[name = string("op_5945_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_5946_cast_fp16 = mul(x = var_5944_cast_fp16, y = var_5945_to_fp16)[name = string("op_5946_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_5948_cast_fp16 = add(x = var_5942_cast_fp16, y = var_5946_cast_fp16)[name = string("op_5948_cast_fp16")];
            fp32 var_5949_epsilon_0 = const()[name = string("op_5949_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_5949_cast_fp16 = rsqrt(epsilon = var_5949_epsilon_0, x = var_5948_cast_fp16)[name = string("op_5949_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_53_cast_fp16 = mul(x = y_189_cast_fp16, y = var_5949_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<int32, [4]> z_105_begin_0 = const()[name = string("z_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_105_end_0 = const()[name = string("z_105_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_105_end_mask_0 = const()[name = string("z_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_105_cast_fp16 = slice_by_index(begin = z_105_begin_0, end = z_105_end_0, end_mask = z_105_end_mask_0, x = x_243_cast_fp16)[name = string("z_105_cast_fp16")];
            tensor<int32, [4]> var_5960_begin_0 = const()[name = string("op_5960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5960_end_0 = const()[name = string("op_5960_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5960_end_mask_0 = const()[name = string("op_5960_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5960_cast_fp16 = slice_by_index(begin = var_5960_begin_0, end = var_5960_end_0, end_mask = var_5960_end_mask_0, x = z_105_cast_fp16)[name = string("op_5960_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5961_cast_fp16 = mul(x = var_5960_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_5961_cast_fp16")];
            tensor<int32, [4]> var_5966_begin_0 = const()[name = string("op_5966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5966_end_0 = const()[name = string("op_5966_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5966_end_mask_0 = const()[name = string("op_5966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5966_cast_fp16 = slice_by_index(begin = var_5966_begin_0, end = var_5966_end_0, end_mask = var_5966_end_mask_0, x = z_105_cast_fp16)[name = string("op_5966_cast_fp16")];
            int32 var_5968 = const()[name = string("op_5968"), val = int32(-1)];
            bool rot_105_interleave_0 = const()[name = string("rot_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_105_cast_fp16 = concat(axis = var_5968, interleave = rot_105_interleave_0, values = (var_5961_cast_fp16, var_5966_cast_fp16))[name = string("rot_105_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5970_cast_fp16 = mul(x = z_105_cast_fp16, y = c_1_cast_fp16)[name = string("op_5970_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5971_cast_fp16 = mul(x = rot_105_cast_fp16, y = s_9_cast_fp16)[name = string("op_5971_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5973_cast_fp16 = add(x = var_5970_cast_fp16, y = var_5971_cast_fp16)[name = string("op_5973_cast_fp16")];
            tensor<int32, [4]> z_107_begin_0 = const()[name = string("z_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_107_end_0 = const()[name = string("z_107_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_107_end_mask_0 = const()[name = string("z_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_107_cast_fp16 = slice_by_index(begin = z_107_begin_0, end = z_107_end_0, end_mask = z_107_end_mask_0, x = x_243_cast_fp16)[name = string("z_107_cast_fp16")];
            tensor<int32, [4]> var_5983_begin_0 = const()[name = string("op_5983_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5983_end_0 = const()[name = string("op_5983_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_5983_end_mask_0 = const()[name = string("op_5983_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_5983_cast_fp16 = slice_by_index(begin = var_5983_begin_0, end = var_5983_end_0, end_mask = var_5983_end_mask_0, x = z_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            fp16 const_53_promoted_to_fp16 = const()[name = string("const_53_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_5984_cast_fp16 = mul(x = var_5983_cast_fp16, y = const_53_promoted_to_fp16)[name = string("op_5984_cast_fp16")];
            tensor<int32, [4]> var_5989_begin_0 = const()[name = string("op_5989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5989_end_0 = const()[name = string("op_5989_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_5989_end_mask_0 = const()[name = string("op_5989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_5989_cast_fp16 = slice_by_index(begin = var_5989_begin_0, end = var_5989_end_0, end_mask = var_5989_end_mask_0, x = z_107_cast_fp16)[name = string("op_5989_cast_fp16")];
            int32 var_5991 = const()[name = string("op_5991"), val = int32(-1)];
            bool rot_107_interleave_0 = const()[name = string("rot_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_107_cast_fp16 = concat(axis = var_5991, interleave = rot_107_interleave_0, values = (var_5984_cast_fp16, var_5989_cast_fp16))[name = string("rot_107_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5993_cast_fp16 = mul(x = z_107_cast_fp16, y = c_cast_fp16)[name = string("op_5993_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5994_cast_fp16 = mul(x = rot_107_cast_fp16, y = s_11_cast_fp16)[name = string("op_5994_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_5996_cast_fp16 = add(x = var_5993_cast_fp16, y = var_5994_cast_fp16)[name = string("op_5996_cast_fp16")];
            int32 var_5998 = const()[name = string("op_5998"), val = int32(-1)];
            bool var_5999_interleave_0 = const()[name = string("op_5999_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_5999_cast_fp16 = concat(axis = var_5998, interleave = var_5999_interleave_0, values = (var_5973_cast_fp16, var_5996_cast_fp16))[name = string("op_5999_cast_fp16")];
            tensor<int32, [4]> z_109_begin_0 = const()[name = string("z_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_109_end_0 = const()[name = string("z_109_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_109_end_mask_0 = const()[name = string("z_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_109_cast_fp16 = slice_by_index(begin = z_109_begin_0, end = z_109_end_0, end_mask = z_109_end_mask_0, x = x_245_cast_fp16)[name = string("z_109_cast_fp16")];
            tensor<int32, [4]> var_6012_begin_0 = const()[name = string("op_6012_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6012_end_0 = const()[name = string("op_6012_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6012_end_mask_0 = const()[name = string("op_6012_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6012_cast_fp16 = slice_by_index(begin = var_6012_begin_0, end = var_6012_end_0, end_mask = var_6012_end_mask_0, x = z_109_cast_fp16)[name = string("op_6012_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6013_cast_fp16 = mul(x = var_6012_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_6013_cast_fp16")];
            tensor<int32, [4]> var_6018_begin_0 = const()[name = string("op_6018_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6018_end_0 = const()[name = string("op_6018_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6018_end_mask_0 = const()[name = string("op_6018_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6018_cast_fp16 = slice_by_index(begin = var_6018_begin_0, end = var_6018_end_0, end_mask = var_6018_end_mask_0, x = z_109_cast_fp16)[name = string("op_6018_cast_fp16")];
            int32 var_6020 = const()[name = string("op_6020"), val = int32(-1)];
            bool rot_109_interleave_0 = const()[name = string("rot_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_109_cast_fp16 = concat(axis = var_6020, interleave = rot_109_interleave_0, values = (var_6013_cast_fp16, var_6018_cast_fp16))[name = string("rot_109_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6022_cast_fp16 = mul(x = z_109_cast_fp16, y = c_1_cast_fp16)[name = string("op_6022_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6023_cast_fp16 = mul(x = rot_109_cast_fp16, y = s_9_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6025_cast_fp16 = add(x = var_6022_cast_fp16, y = var_6023_cast_fp16)[name = string("op_6025_cast_fp16")];
            tensor<int32, [4]> z_111_begin_0 = const()[name = string("z_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_111_end_0 = const()[name = string("z_111_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_111_end_mask_0 = const()[name = string("z_111_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_111_cast_fp16 = slice_by_index(begin = z_111_begin_0, end = z_111_end_0, end_mask = z_111_end_mask_0, x = x_245_cast_fp16)[name = string("z_111_cast_fp16")];
            tensor<int32, [4]> var_6035_begin_0 = const()[name = string("op_6035_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6035_end_0 = const()[name = string("op_6035_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6035_end_mask_0 = const()[name = string("op_6035_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6035_cast_fp16 = slice_by_index(begin = var_6035_begin_0, end = var_6035_end_0, end_mask = var_6035_end_mask_0, x = z_111_cast_fp16)[name = string("op_6035_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6036_cast_fp16 = mul(x = var_6035_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_6036_cast_fp16")];
            tensor<int32, [4]> var_6041_begin_0 = const()[name = string("op_6041_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6041_end_0 = const()[name = string("op_6041_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6041_end_mask_0 = const()[name = string("op_6041_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6041_cast_fp16 = slice_by_index(begin = var_6041_begin_0, end = var_6041_end_0, end_mask = var_6041_end_mask_0, x = z_111_cast_fp16)[name = string("op_6041_cast_fp16")];
            int32 var_6043 = const()[name = string("op_6043"), val = int32(-1)];
            bool rot_111_interleave_0 = const()[name = string("rot_111_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_111_cast_fp16 = concat(axis = var_6043, interleave = rot_111_interleave_0, values = (var_6036_cast_fp16, var_6041_cast_fp16))[name = string("rot_111_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6045_cast_fp16 = mul(x = z_111_cast_fp16, y = c_cast_fp16)[name = string("op_6045_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6046_cast_fp16 = mul(x = rot_111_cast_fp16, y = s_11_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6048_cast_fp16 = add(x = var_6045_cast_fp16, y = var_6046_cast_fp16)[name = string("op_6048_cast_fp16")];
            int32 var_6050 = const()[name = string("op_6050"), val = int32(-1)];
            bool var_6051_interleave_0 = const()[name = string("op_6051_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_6051_cast_fp16 = concat(axis = var_6050, interleave = var_6051_interleave_0, values = (var_6025_cast_fp16, var_6048_cast_fp16))[name = string("op_6051_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6061_transpose_x_0 = const()[name = string("op_6061_transpose_x_0"), val = bool(false)];
            bool var_6061_transpose_y_0 = const()[name = string("op_6061_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_91 = transpose(perm = transpose_91_perm_0, x = var_6051_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_90 = transpose(perm = transpose_90_perm_0, x = var_5999_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 12, 2520, 2520]> var_6061_cast_fp16 = matmul(transpose_x = var_6061_transpose_x_0, transpose_y = var_6061_transpose_y_0, x = transpose_90, y = transpose_91)[name = string("op_6061_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_6063_cast_fp16 = add(x = var_6061_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6063_cast_fp16")];
            int32 var_6064 = const()[name = string("op_6064"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_189_cast_fp16 = softmax(axis = var_6064, x = var_6063_cast_fp16)[name = string("w_189_cast_fp16")];
            bool var_6067_transpose_x_0 = const()[name = string("op_6067_transpose_x_0"), val = bool(false)];
            bool var_6067_transpose_y_0 = const()[name = string("op_6067_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = v_53_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 12, 2520, 64]> var_6067_cast_fp16 = matmul(transpose_x = var_6067_transpose_x_0, transpose_y = var_6067_transpose_y_0, x = w_189_cast_fp16, y = v_55_cast_fp16)[name = string("op_6067_cast_fp16")];
            tensor<int32, [4]> var_6070_perm_0 = const()[name = string("op_6070_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6074 = const()[name = string("op_6074"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_6070_cast_fp16 = transpose(perm = var_6070_perm_0, x = var_6067_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 2520, 768]> input_135_cast_fp16 = reshape(shape = var_6074, x = var_6070_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125416832))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126006720))))[name = string("t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_95_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6079_cast_fp16 = abs(x = linear_95_cast_fp16)[name = string("op_6079_cast_fp16")];
            bool var_6082 = const()[name = string("op_6082"), val = bool(true)];
            tensor<int32, [1]> var_6083_axes_0 = const()[name = string("op_6083_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6083_cast_fp16 = reduce_max(axes = var_6083_axes_0, keep_dims = var_6082, x = var_6079_cast_fp16)[name = string("op_6083_cast_fp16")];
            fp16 var_6084_to_fp16 = const()[name = string("op_6084_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_195_cast_fp16 = maximum(x = var_6083_cast_fp16, y = var_6084_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_191_cast_fp16 = real_div(x = linear_95_cast_fp16, y = s_195_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6087_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6087_cast_fp16")];
            tensor<int32, [1]> var_6092_axes_0 = const()[name = string("op_6092_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6092_keep_dims_0 = const()[name = string("op_6092_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6092_cast_fp16 = reduce_mean(axes = var_6092_axes_0, keep_dims = var_6092_keep_dims_0, x = var_6087_cast_fp16)[name = string("op_6092_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6093_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6093_cast_fp16")];
            fp32 var_6094_epsilon_0 = const()[name = string("op_6094_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6094_cast_fp16 = inverse(epsilon = var_6094_epsilon_0, x = var_6093_cast_fp16)[name = string("op_6094_cast_fp16")];
            fp16 var_6095_to_fp16 = const()[name = string("op_6095_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6096_cast_fp16 = mul(x = var_6094_cast_fp16, y = var_6095_to_fp16)[name = string("op_6096_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6098_cast_fp16 = add(x = var_6092_cast_fp16, y = var_6096_cast_fp16)[name = string("op_6098_cast_fp16")];
            fp32 var_6099_epsilon_0 = const()[name = string("op_6099_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6099_cast_fp16 = rsqrt(epsilon = var_6099_epsilon_0, x = var_6098_cast_fp16)[name = string("op_6099_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_163_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6099_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126008320)))];
            tensor<fp16, [1, 2520, 768]> var_6101_cast_fp16 = mul(x = out_163_cast_fp16, y = t_encoder_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6101_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_249_cast_fp16 = add(x = x_235_cast_fp16, y = var_6101_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6104_cast_fp16 = abs(x = x_249_cast_fp16)[name = string("op_6104_cast_fp16")];
            bool var_6107 = const()[name = string("op_6107"), val = bool(true)];
            tensor<int32, [1]> var_6108_axes_0 = const()[name = string("op_6108_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6108_cast_fp16 = reduce_max(axes = var_6108_axes_0, keep_dims = var_6107, x = var_6104_cast_fp16)[name = string("op_6108_cast_fp16")];
            fp16 var_6109_to_fp16 = const()[name = string("op_6109_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_197_cast_fp16 = maximum(x = var_6108_cast_fp16, y = var_6109_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_193_cast_fp16 = real_div(x = x_249_cast_fp16, y = s_197_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6112_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6112_cast_fp16")];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6117_keep_dims_0 = const()[name = string("op_6117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6117_cast_fp16 = reduce_mean(axes = var_6117_axes_0, keep_dims = var_6117_keep_dims_0, x = var_6112_cast_fp16)[name = string("op_6117_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6118_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6118_cast_fp16")];
            fp32 var_6119_epsilon_0 = const()[name = string("op_6119_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6119_cast_fp16 = inverse(epsilon = var_6119_epsilon_0, x = var_6118_cast_fp16)[name = string("op_6119_cast_fp16")];
            fp16 var_6120_to_fp16 = const()[name = string("op_6120_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6121_cast_fp16 = mul(x = var_6119_cast_fp16, y = var_6120_to_fp16)[name = string("op_6121_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6123_cast_fp16 = add(x = var_6117_cast_fp16, y = var_6121_cast_fp16)[name = string("op_6123_cast_fp16")];
            fp32 var_6124_epsilon_0 = const()[name = string("op_6124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6124_cast_fp16 = rsqrt(epsilon = var_6124_epsilon_0, x = var_6123_cast_fp16)[name = string("op_6124_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_165_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6124_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126009920)))];
            tensor<fp16, [1, 2520, 768]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126011520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128370880))))[name = string("t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_96_cast_fp16")];
            string var_6131_mode_0 = const()[name = string("op_6131_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_6131_cast_fp16 = gelu(mode = var_6131_mode_0, x = linear_96_cast_fp16)[name = string("op_6131_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128377088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130736448))))[name = string("t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_97_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_141_cast_fp16 = mul(x = var_6131_cast_fp16, y = linear_97_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130742656))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133102016))))[name = string("t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6139_cast_fp16 = abs(x = linear_98_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_199_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_195_cast_fp16 = real_div(x = linear_98_cast_fp16, y = s_199_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6153_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_167_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133103616)))];
            tensor<fp16, [1, 2520, 768]> var_6161_cast_fp16 = mul(x = out_167_cast_fp16, y = t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6161_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_253_cast_fp16 = add(x = x_249_cast_fp16, y = var_6161_cast_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6164_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6164_cast_fp16")];
            bool var_6167 = const()[name = string("op_6167"), val = bool(true)];
            tensor<int32, [1]> var_6168_axes_0 = const()[name = string("op_6168_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6168_cast_fp16 = reduce_max(axes = var_6168_axes_0, keep_dims = var_6167, x = var_6164_cast_fp16)[name = string("op_6168_cast_fp16")];
            fp16 var_6169_to_fp16 = const()[name = string("op_6169_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_201_cast_fp16 = maximum(x = var_6168_cast_fp16, y = var_6169_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_197_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_201_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6172_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6172_cast_fp16")];
            tensor<int32, [1]> var_6177_axes_0 = const()[name = string("op_6177_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6177_keep_dims_0 = const()[name = string("op_6177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6177_cast_fp16 = reduce_mean(axes = var_6177_axes_0, keep_dims = var_6177_keep_dims_0, x = var_6172_cast_fp16)[name = string("op_6177_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6178_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6178_cast_fp16")];
            fp32 var_6179_epsilon_0 = const()[name = string("op_6179_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6179_cast_fp16 = inverse(epsilon = var_6179_epsilon_0, x = var_6178_cast_fp16)[name = string("op_6179_cast_fp16")];
            fp16 var_6180_to_fp16 = const()[name = string("op_6180_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6181_cast_fp16 = mul(x = var_6179_cast_fp16, y = var_6180_to_fp16)[name = string("op_6181_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6183_cast_fp16 = add(x = var_6177_cast_fp16, y = var_6181_cast_fp16)[name = string("op_6183_cast_fp16")];
            fp32 var_6184_epsilon_0 = const()[name = string("op_6184_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6184_cast_fp16 = rsqrt(epsilon = var_6184_epsilon_0, x = var_6183_cast_fp16)[name = string("op_6184_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_169_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6184_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133105216)))];
            tensor<fp16, [1, 2520, 768]> input_143_cast_fp16 = mul(x = out_169_cast_fp16, y = t_encoder_layers_14_input_layernorm_weight_to_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133106816))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133696704))))[name = string("t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_99_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_6194 = const()[name = string("op_6194"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_255_cast_fp16 = reshape(shape = var_6194, x = linear_99_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6196_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6196_cast_fp16")];
            bool var_6199 = const()[name = string("op_6199"), val = bool(true)];
            tensor<int32, [1]> var_6200_axes_0 = const()[name = string("op_6200_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_6200_cast_fp16 = reduce_max(axes = var_6200_axes_0, keep_dims = var_6199, x = var_6196_cast_fp16)[name = string("op_6200_cast_fp16")];
            fp16 var_6201_to_fp16 = const()[name = string("op_6201_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_203_cast_fp16 = maximum(x = var_6200_cast_fp16, y = var_6201_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_199_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_203_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6204_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6204_cast_fp16")];
            tensor<int32, [1]> var_6209_axes_0 = const()[name = string("op_6209_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6209_keep_dims_0 = const()[name = string("op_6209_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_6209_cast_fp16 = reduce_mean(axes = var_6209_axes_0, keep_dims = var_6209_keep_dims_0, x = var_6204_cast_fp16)[name = string("op_6209_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6210_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6210_cast_fp16")];
            fp32 var_6211_epsilon_0 = const()[name = string("op_6211_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_6211_cast_fp16 = inverse(epsilon = var_6211_epsilon_0, x = var_6210_cast_fp16)[name = string("op_6211_cast_fp16")];
            fp16 var_6212_to_fp16 = const()[name = string("op_6212_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_6213_cast_fp16 = mul(x = var_6211_cast_fp16, y = var_6212_to_fp16)[name = string("op_6213_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6215_cast_fp16 = add(x = var_6209_cast_fp16, y = var_6213_cast_fp16)[name = string("op_6215_cast_fp16")];
            fp32 var_6216_epsilon_0 = const()[name = string("op_6216_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_6216_cast_fp16 = rsqrt(epsilon = var_6216_epsilon_0, x = var_6215_cast_fp16)[name = string("op_6216_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_171_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6216_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133698304)))];
            tensor<fp16, [1, 2520, 12, 64]> x_261_cast_fp16 = mul(x = out_171_cast_fp16, y = t_encoder_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_261_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133698496))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134288384))))[name = string("t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_100_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_6226 = const()[name = string("op_6226"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_257_cast_fp16 = reshape(shape = var_6226, x = linear_100_cast_fp16)[name = string("x_257_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6228_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6228_cast_fp16")];
            bool var_6231 = const()[name = string("op_6231"), val = bool(true)];
            tensor<int32, [1]> var_6232_axes_0 = const()[name = string("op_6232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_6232_cast_fp16 = reduce_max(axes = var_6232_axes_0, keep_dims = var_6231, x = var_6228_cast_fp16)[name = string("op_6232_cast_fp16")];
            fp16 var_6233_to_fp16 = const()[name = string("op_6233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_205_cast_fp16 = maximum(x = var_6232_cast_fp16, y = var_6233_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_201_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_205_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6236_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6236_cast_fp16")];
            tensor<int32, [1]> var_6241_axes_0 = const()[name = string("op_6241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6241_keep_dims_0 = const()[name = string("op_6241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_6241_cast_fp16 = reduce_mean(axes = var_6241_axes_0, keep_dims = var_6241_keep_dims_0, x = var_6236_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6242_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6242_cast_fp16")];
            fp32 var_6243_epsilon_0 = const()[name = string("op_6243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_6243_cast_fp16 = inverse(epsilon = var_6243_epsilon_0, x = var_6242_cast_fp16)[name = string("op_6243_cast_fp16")];
            fp16 var_6244_to_fp16 = const()[name = string("op_6244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_6245_cast_fp16 = mul(x = var_6243_cast_fp16, y = var_6244_to_fp16)[name = string("op_6245_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6247_cast_fp16 = add(x = var_6241_cast_fp16, y = var_6245_cast_fp16)[name = string("op_6247_cast_fp16")];
            fp32 var_6248_epsilon_0 = const()[name = string("op_6248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_6248_cast_fp16 = rsqrt(epsilon = var_6248_epsilon_0, x = var_6247_cast_fp16)[name = string("op_6248_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_173_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6248_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966592)))];
            tensor<fp16, [1, 2520, 12, 64]> x_263_cast_fp16 = mul(x = out_173_cast_fp16, y = t_encoder_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134289984))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134879872))))[name = string("t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_101_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_6258 = const()[name = string("op_6258"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_259_cast_fp16 = reshape(shape = var_6258, x = linear_101_cast_fp16)[name = string("x_259_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6260_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6260_cast_fp16")];
            bool var_6263 = const()[name = string("op_6263"), val = bool(true)];
            tensor<int32, [1]> var_6264_axes_0 = const()[name = string("op_6264_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_6264_cast_fp16 = reduce_max(axes = var_6264_axes_0, keep_dims = var_6263, x = var_6260_cast_fp16)[name = string("op_6264_cast_fp16")];
            fp16 var_6265_to_fp16 = const()[name = string("op_6265_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_207_cast_fp16 = maximum(x = var_6264_cast_fp16, y = var_6265_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_203_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_207_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6268_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6268_cast_fp16")];
            tensor<int32, [1]> var_6273_axes_0 = const()[name = string("op_6273_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6273_keep_dims_0 = const()[name = string("op_6273_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_6273_cast_fp16 = reduce_mean(axes = var_6273_axes_0, keep_dims = var_6273_keep_dims_0, x = var_6268_cast_fp16)[name = string("op_6273_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6274_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6274_cast_fp16")];
            fp32 var_6275_epsilon_0 = const()[name = string("op_6275_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_6275_cast_fp16 = inverse(epsilon = var_6275_epsilon_0, x = var_6274_cast_fp16)[name = string("op_6275_cast_fp16")];
            fp16 var_6276_to_fp16 = const()[name = string("op_6276_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_6277_cast_fp16 = mul(x = var_6275_cast_fp16, y = var_6276_to_fp16)[name = string("op_6277_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6279_cast_fp16 = add(x = var_6273_cast_fp16, y = var_6277_cast_fp16)[name = string("op_6279_cast_fp16")];
            fp32 var_6280_epsilon_0 = const()[name = string("op_6280_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_6280_cast_fp16 = rsqrt(epsilon = var_6280_epsilon_0, x = var_6279_cast_fp16)[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_57_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6280_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<int32, [4]> z_113_begin_0 = const()[name = string("z_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_113_end_0 = const()[name = string("z_113_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_113_end_mask_0 = const()[name = string("z_113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_113_cast_fp16 = slice_by_index(begin = z_113_begin_0, end = z_113_end_0, end_mask = z_113_end_mask_0, x = x_261_cast_fp16)[name = string("z_113_cast_fp16")];
            tensor<int32, [4]> var_6291_begin_0 = const()[name = string("op_6291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6291_end_0 = const()[name = string("op_6291_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6291_end_mask_0 = const()[name = string("op_6291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6291_cast_fp16 = slice_by_index(begin = var_6291_begin_0, end = var_6291_end_0, end_mask = var_6291_end_mask_0, x = z_113_cast_fp16)[name = string("op_6291_cast_fp16")];
            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6292_cast_fp16 = mul(x = var_6291_cast_fp16, y = const_56_promoted_to_fp16)[name = string("op_6292_cast_fp16")];
            tensor<int32, [4]> var_6297_begin_0 = const()[name = string("op_6297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6297_end_0 = const()[name = string("op_6297_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6297_end_mask_0 = const()[name = string("op_6297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6297_cast_fp16 = slice_by_index(begin = var_6297_begin_0, end = var_6297_end_0, end_mask = var_6297_end_mask_0, x = z_113_cast_fp16)[name = string("op_6297_cast_fp16")];
            int32 var_6299 = const()[name = string("op_6299"), val = int32(-1)];
            bool rot_113_interleave_0 = const()[name = string("rot_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_113_cast_fp16 = concat(axis = var_6299, interleave = rot_113_interleave_0, values = (var_6292_cast_fp16, var_6297_cast_fp16))[name = string("rot_113_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6301_cast_fp16 = mul(x = z_113_cast_fp16, y = c_1_cast_fp16)[name = string("op_6301_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6302_cast_fp16 = mul(x = rot_113_cast_fp16, y = s_9_cast_fp16)[name = string("op_6302_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6304_cast_fp16 = add(x = var_6301_cast_fp16, y = var_6302_cast_fp16)[name = string("op_6304_cast_fp16")];
            tensor<int32, [4]> z_115_begin_0 = const()[name = string("z_115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_115_end_0 = const()[name = string("z_115_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_115_end_mask_0 = const()[name = string("z_115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_115_cast_fp16 = slice_by_index(begin = z_115_begin_0, end = z_115_end_0, end_mask = z_115_end_mask_0, x = x_261_cast_fp16)[name = string("z_115_cast_fp16")];
            tensor<int32, [4]> var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> 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 = z_115_cast_fp16)[name = string("op_6314_cast_fp16")];
            fp16 const_57_promoted_to_fp16 = const()[name = string("const_57_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6315_cast_fp16 = mul(x = var_6314_cast_fp16, y = const_57_promoted_to_fp16)[name = string("op_6315_cast_fp16")];
            tensor<int32, [4]> var_6320_begin_0 = const()[name = string("op_6320_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6320_end_0 = const()[name = string("op_6320_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6320_end_mask_0 = const()[name = string("op_6320_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6320_cast_fp16 = slice_by_index(begin = var_6320_begin_0, end = var_6320_end_0, end_mask = var_6320_end_mask_0, x = z_115_cast_fp16)[name = string("op_6320_cast_fp16")];
            int32 var_6322 = const()[name = string("op_6322"), val = int32(-1)];
            bool rot_115_interleave_0 = const()[name = string("rot_115_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_115_cast_fp16 = concat(axis = var_6322, interleave = rot_115_interleave_0, values = (var_6315_cast_fp16, var_6320_cast_fp16))[name = string("rot_115_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6324_cast_fp16 = mul(x = z_115_cast_fp16, y = c_cast_fp16)[name = string("op_6324_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6325_cast_fp16 = mul(x = rot_115_cast_fp16, y = s_11_cast_fp16)[name = string("op_6325_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6327_cast_fp16 = add(x = var_6324_cast_fp16, y = var_6325_cast_fp16)[name = string("op_6327_cast_fp16")];
            int32 var_6329 = const()[name = string("op_6329"), val = int32(-1)];
            bool var_6330_interleave_0 = const()[name = string("op_6330_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_6330_cast_fp16 = concat(axis = var_6329, interleave = var_6330_interleave_0, values = (var_6304_cast_fp16, var_6327_cast_fp16))[name = string("op_6330_cast_fp16")];
            tensor<int32, [4]> z_117_begin_0 = const()[name = string("z_117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_117_end_0 = const()[name = string("z_117_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_117_end_mask_0 = const()[name = string("z_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_117_cast_fp16 = slice_by_index(begin = z_117_begin_0, end = z_117_end_0, end_mask = z_117_end_mask_0, x = x_263_cast_fp16)[name = string("z_117_cast_fp16")];
            tensor<int32, [4]> var_6343_begin_0 = const()[name = string("op_6343_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6343_end_0 = const()[name = string("op_6343_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6343_end_mask_0 = const()[name = string("op_6343_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6343_cast_fp16 = slice_by_index(begin = var_6343_begin_0, end = var_6343_end_0, end_mask = var_6343_end_mask_0, x = z_117_cast_fp16)[name = string("op_6343_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6344_cast_fp16 = mul(x = var_6343_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_6344_cast_fp16")];
            tensor<int32, [4]> var_6349_begin_0 = const()[name = string("op_6349_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6349_end_0 = const()[name = string("op_6349_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6349_end_mask_0 = const()[name = string("op_6349_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6349_cast_fp16 = slice_by_index(begin = var_6349_begin_0, end = var_6349_end_0, end_mask = var_6349_end_mask_0, x = z_117_cast_fp16)[name = string("op_6349_cast_fp16")];
            int32 var_6351 = const()[name = string("op_6351"), val = int32(-1)];
            bool rot_117_interleave_0 = const()[name = string("rot_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_117_cast_fp16 = concat(axis = var_6351, interleave = rot_117_interleave_0, values = (var_6344_cast_fp16, var_6349_cast_fp16))[name = string("rot_117_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6353_cast_fp16 = mul(x = z_117_cast_fp16, y = c_1_cast_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6354_cast_fp16 = mul(x = rot_117_cast_fp16, y = s_9_cast_fp16)[name = string("op_6354_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6356_cast_fp16 = add(x = var_6353_cast_fp16, y = var_6354_cast_fp16)[name = string("op_6356_cast_fp16")];
            tensor<int32, [4]> z_119_begin_0 = const()[name = string("z_119_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_119_end_0 = const()[name = string("z_119_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_119_end_mask_0 = const()[name = string("z_119_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_119_cast_fp16 = slice_by_index(begin = z_119_begin_0, end = z_119_end_0, end_mask = z_119_end_mask_0, x = x_263_cast_fp16)[name = string("z_119_cast_fp16")];
            tensor<int32, [4]> var_6366_begin_0 = const()[name = string("op_6366_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6366_end_0 = const()[name = string("op_6366_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6366_end_mask_0 = const()[name = string("op_6366_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6366_cast_fp16 = slice_by_index(begin = var_6366_begin_0, end = var_6366_end_0, end_mask = var_6366_end_mask_0, x = z_119_cast_fp16)[name = string("op_6366_cast_fp16")];
            fp16 const_59_promoted_to_fp16 = const()[name = string("const_59_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6367_cast_fp16 = mul(x = var_6366_cast_fp16, y = const_59_promoted_to_fp16)[name = string("op_6367_cast_fp16")];
            tensor<int32, [4]> var_6372_begin_0 = const()[name = string("op_6372_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6372_end_0 = const()[name = string("op_6372_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6372_end_mask_0 = const()[name = string("op_6372_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6372_cast_fp16 = slice_by_index(begin = var_6372_begin_0, end = var_6372_end_0, end_mask = var_6372_end_mask_0, x = z_119_cast_fp16)[name = string("op_6372_cast_fp16")];
            int32 var_6374 = const()[name = string("op_6374"), val = int32(-1)];
            bool rot_119_interleave_0 = const()[name = string("rot_119_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_119_cast_fp16 = concat(axis = var_6374, interleave = rot_119_interleave_0, values = (var_6367_cast_fp16, var_6372_cast_fp16))[name = string("rot_119_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6376_cast_fp16 = mul(x = z_119_cast_fp16, y = c_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6377_cast_fp16 = mul(x = rot_119_cast_fp16, y = s_11_cast_fp16)[name = string("op_6377_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6379_cast_fp16 = add(x = var_6376_cast_fp16, y = var_6377_cast_fp16)[name = string("op_6379_cast_fp16")];
            int32 var_6381 = const()[name = string("op_6381"), val = int32(-1)];
            bool var_6382_interleave_0 = const()[name = string("op_6382_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_6382_cast_fp16 = concat(axis = var_6381, interleave = var_6382_interleave_0, values = (var_6356_cast_fp16, var_6379_cast_fp16))[name = string("op_6382_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6392_transpose_x_0 = const()[name = string("op_6392_transpose_x_0"), val = bool(false)];
            bool var_6392_transpose_y_0 = const()[name = string("op_6392_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_93 = transpose(perm = transpose_93_perm_0, x = var_6382_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_92 = transpose(perm = transpose_92_perm_0, x = var_6330_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 12, 2520, 2520]> var_6392_cast_fp16 = matmul(transpose_x = var_6392_transpose_x_0, transpose_y = var_6392_transpose_y_0, x = transpose_92, y = transpose_93)[name = string("op_6392_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_6394_cast_fp16 = add(x = var_6392_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6394_cast_fp16")];
            int32 var_6395 = const()[name = string("op_6395"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_203_cast_fp16 = softmax(axis = var_6395, x = var_6394_cast_fp16)[name = string("w_203_cast_fp16")];
            bool var_6398_transpose_x_0 = const()[name = string("op_6398_transpose_x_0"), val = bool(false)];
            bool var_6398_transpose_y_0 = const()[name = string("op_6398_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = v_57_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 12, 2520, 64]> var_6398_cast_fp16 = matmul(transpose_x = var_6398_transpose_x_0, transpose_y = var_6398_transpose_y_0, x = w_203_cast_fp16, y = v_59_cast_fp16)[name = string("op_6398_cast_fp16")];
            tensor<int32, [4]> var_6401_perm_0 = const()[name = string("op_6401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6405 = const()[name = string("op_6405"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_6401_cast_fp16 = transpose(perm = var_6401_perm_0, x = var_6398_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 2520, 768]> input_145_cast_fp16 = reshape(shape = var_6405, x = var_6401_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134881472))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135471360))))[name = string("t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_102_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6410_cast_fp16 = abs(x = linear_102_cast_fp16)[name = string("op_6410_cast_fp16")];
            bool var_6413 = const()[name = string("op_6413"), val = bool(true)];
            tensor<int32, [1]> var_6414_axes_0 = const()[name = string("op_6414_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6414_cast_fp16 = reduce_max(axes = var_6414_axes_0, keep_dims = var_6413, x = var_6410_cast_fp16)[name = string("op_6414_cast_fp16")];
            fp16 var_6415_to_fp16 = const()[name = string("op_6415_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_209_cast_fp16 = maximum(x = var_6414_cast_fp16, y = var_6415_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_205_cast_fp16 = real_div(x = linear_102_cast_fp16, y = s_209_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6418_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6418_cast_fp16")];
            tensor<int32, [1]> var_6423_axes_0 = const()[name = string("op_6423_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6423_keep_dims_0 = const()[name = string("op_6423_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6423_cast_fp16 = reduce_mean(axes = var_6423_axes_0, keep_dims = var_6423_keep_dims_0, x = var_6418_cast_fp16)[name = string("op_6423_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6424_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6424_cast_fp16")];
            fp32 var_6425_epsilon_0 = const()[name = string("op_6425_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6425_cast_fp16 = inverse(epsilon = var_6425_epsilon_0, x = var_6424_cast_fp16)[name = string("op_6425_cast_fp16")];
            fp16 var_6426_to_fp16 = const()[name = string("op_6426_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6427_cast_fp16 = mul(x = var_6425_cast_fp16, y = var_6426_to_fp16)[name = string("op_6427_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6429_cast_fp16 = add(x = var_6423_cast_fp16, y = var_6427_cast_fp16)[name = string("op_6429_cast_fp16")];
            fp32 var_6430_epsilon_0 = const()[name = string("op_6430_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6430_cast_fp16 = rsqrt(epsilon = var_6430_epsilon_0, x = var_6429_cast_fp16)[name = string("op_6430_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_175_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6430_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135472960)))];
            tensor<fp16, [1, 2520, 768]> var_6432_cast_fp16 = mul(x = out_175_cast_fp16, y = t_encoder_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_6432_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_267_cast_fp16 = add(x = x_253_cast_fp16, y = var_6432_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6435_cast_fp16 = abs(x = x_267_cast_fp16)[name = string("op_6435_cast_fp16")];
            bool var_6438 = const()[name = string("op_6438"), val = bool(true)];
            tensor<int32, [1]> var_6439_axes_0 = const()[name = string("op_6439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6439_cast_fp16 = reduce_max(axes = var_6439_axes_0, keep_dims = var_6438, x = var_6435_cast_fp16)[name = string("op_6439_cast_fp16")];
            fp16 var_6440_to_fp16 = const()[name = string("op_6440_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_211_cast_fp16 = maximum(x = var_6439_cast_fp16, y = var_6440_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_207_cast_fp16 = real_div(x = x_267_cast_fp16, y = s_211_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6443_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6443_cast_fp16")];
            tensor<int32, [1]> var_6448_axes_0 = const()[name = string("op_6448_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6448_keep_dims_0 = const()[name = string("op_6448_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6448_cast_fp16 = reduce_mean(axes = var_6448_axes_0, keep_dims = var_6448_keep_dims_0, x = var_6443_cast_fp16)[name = string("op_6448_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6449_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6449_cast_fp16")];
            fp32 var_6450_epsilon_0 = const()[name = string("op_6450_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6450_cast_fp16 = inverse(epsilon = var_6450_epsilon_0, x = var_6449_cast_fp16)[name = string("op_6450_cast_fp16")];
            fp16 var_6451_to_fp16 = const()[name = string("op_6451_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6452_cast_fp16 = mul(x = var_6450_cast_fp16, y = var_6451_to_fp16)[name = string("op_6452_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6454_cast_fp16 = add(x = var_6448_cast_fp16, y = var_6452_cast_fp16)[name = string("op_6454_cast_fp16")];
            fp32 var_6455_epsilon_0 = const()[name = string("op_6455_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6455_cast_fp16 = rsqrt(epsilon = var_6455_epsilon_0, x = var_6454_cast_fp16)[name = string("op_6455_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_177_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6455_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135474560)))];
            tensor<fp16, [1, 2520, 768]> input_147_cast_fp16 = mul(x = out_177_cast_fp16, y = t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135476160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137835520))))[name = string("t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_103_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_103_cast_fp16")];
            string var_6462_mode_0 = const()[name = string("op_6462_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_6462_cast_fp16 = gelu(mode = var_6462_mode_0, x = linear_103_cast_fp16)[name = string("op_6462_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137841728))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140201088))))[name = string("t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_151_cast_fp16 = mul(x = var_6462_cast_fp16, y = linear_104_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140207296))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142566656))))[name = string("t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6470_cast_fp16 = abs(x = linear_105_cast_fp16)[name = string("op_6470_cast_fp16")];
            bool var_6473 = const()[name = string("op_6473"), val = bool(true)];
            tensor<int32, [1]> var_6474_axes_0 = const()[name = string("op_6474_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6474_cast_fp16 = reduce_max(axes = var_6474_axes_0, keep_dims = var_6473, x = var_6470_cast_fp16)[name = string("op_6474_cast_fp16")];
            fp16 var_6475_to_fp16 = const()[name = string("op_6475_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_213_cast_fp16 = maximum(x = var_6474_cast_fp16, y = var_6475_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_209_cast_fp16 = real_div(x = linear_105_cast_fp16, y = s_213_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6478_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6478_cast_fp16")];
            tensor<int32, [1]> var_6483_axes_0 = const()[name = string("op_6483_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6483_keep_dims_0 = const()[name = string("op_6483_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6483_cast_fp16 = reduce_mean(axes = var_6483_axes_0, keep_dims = var_6483_keep_dims_0, x = var_6478_cast_fp16)[name = string("op_6483_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6484_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp32 var_6485_epsilon_0 = const()[name = string("op_6485_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6485_cast_fp16 = inverse(epsilon = var_6485_epsilon_0, x = var_6484_cast_fp16)[name = string("op_6485_cast_fp16")];
            fp16 var_6486_to_fp16 = const()[name = string("op_6486_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6487_cast_fp16 = mul(x = var_6485_cast_fp16, y = var_6486_to_fp16)[name = string("op_6487_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6489_cast_fp16 = add(x = var_6483_cast_fp16, y = var_6487_cast_fp16)[name = string("op_6489_cast_fp16")];
            fp32 var_6490_epsilon_0 = const()[name = string("op_6490_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6490_cast_fp16 = rsqrt(epsilon = var_6490_epsilon_0, x = var_6489_cast_fp16)[name = string("op_6490_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_179_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6490_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142568256)))];
            tensor<fp16, [1, 2520, 768]> var_6492_cast_fp16 = mul(x = out_179_cast_fp16, y = t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6492_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_271_cast_fp16 = add(x = x_267_cast_fp16, y = var_6492_cast_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6495_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_6495_cast_fp16")];
            bool var_6498 = const()[name = string("op_6498"), val = bool(true)];
            tensor<int32, [1]> var_6499_axes_0 = const()[name = string("op_6499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6499_cast_fp16 = reduce_max(axes = var_6499_axes_0, keep_dims = var_6498, x = var_6495_cast_fp16)[name = string("op_6499_cast_fp16")];
            fp16 var_6500_to_fp16 = const()[name = string("op_6500_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_215_cast_fp16 = maximum(x = var_6499_cast_fp16, y = var_6500_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_211_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_215_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6503_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6503_cast_fp16")];
            tensor<int32, [1]> var_6508_axes_0 = const()[name = string("op_6508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6508_keep_dims_0 = const()[name = string("op_6508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6508_cast_fp16 = reduce_mean(axes = var_6508_axes_0, keep_dims = var_6508_keep_dims_0, x = var_6503_cast_fp16)[name = string("op_6508_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6509_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6509_cast_fp16")];
            fp32 var_6510_epsilon_0 = const()[name = string("op_6510_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6510_cast_fp16 = inverse(epsilon = var_6510_epsilon_0, x = var_6509_cast_fp16)[name = string("op_6510_cast_fp16")];
            fp16 var_6511_to_fp16 = const()[name = string("op_6511_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6512_cast_fp16 = mul(x = var_6510_cast_fp16, y = var_6511_to_fp16)[name = string("op_6512_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6514_cast_fp16 = add(x = var_6508_cast_fp16, y = var_6512_cast_fp16)[name = string("op_6514_cast_fp16")];
            fp32 var_6515_epsilon_0 = const()[name = string("op_6515_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6515_cast_fp16 = rsqrt(epsilon = var_6515_epsilon_0, x = var_6514_cast_fp16)[name = string("op_6515_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_181_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6515_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142569856)))];
            tensor<fp16, [1, 2520, 768]> input_153_cast_fp16 = mul(x = out_181_cast_fp16, y = t_encoder_layers_15_input_layernorm_weight_to_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142571456))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143161344))))[name = string("t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_106_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_6525 = const()[name = string("op_6525"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_273_cast_fp16 = reshape(shape = var_6525, x = linear_106_cast_fp16)[name = string("x_273_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6527_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_6527_cast_fp16")];
            bool var_6530 = const()[name = string("op_6530"), val = bool(true)];
            tensor<int32, [1]> var_6531_axes_0 = const()[name = string("op_6531_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_6531_cast_fp16 = reduce_max(axes = var_6531_axes_0, keep_dims = var_6530, x = var_6527_cast_fp16)[name = string("op_6531_cast_fp16")];
            fp16 var_6532_to_fp16 = const()[name = string("op_6532_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_217_cast_fp16 = maximum(x = var_6531_cast_fp16, y = var_6532_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_213_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_217_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6535_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6535_cast_fp16")];
            tensor<int32, [1]> var_6540_axes_0 = const()[name = string("op_6540_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6540_keep_dims_0 = const()[name = string("op_6540_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_6540_cast_fp16 = reduce_mean(axes = var_6540_axes_0, keep_dims = var_6540_keep_dims_0, x = var_6535_cast_fp16)[name = string("op_6540_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6541_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6541_cast_fp16")];
            fp32 var_6542_epsilon_0 = const()[name = string("op_6542_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_6542_cast_fp16 = inverse(epsilon = var_6542_epsilon_0, x = var_6541_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp16 var_6543_to_fp16 = const()[name = string("op_6543_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_6544_cast_fp16 = mul(x = var_6542_cast_fp16, y = var_6543_to_fp16)[name = string("op_6544_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6546_cast_fp16 = add(x = var_6540_cast_fp16, y = var_6544_cast_fp16)[name = string("op_6546_cast_fp16")];
            fp32 var_6547_epsilon_0 = const()[name = string("op_6547_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_6547_cast_fp16 = rsqrt(epsilon = var_6547_epsilon_0, x = var_6546_cast_fp16)[name = string("op_6547_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_183_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6547_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769024)))];
            tensor<fp16, [1, 2520, 12, 64]> x_279_cast_fp16 = mul(x = out_183_cast_fp16, y = t_encoder_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_279_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143162944))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143752832))))[name = string("t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_107_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_6557 = const()[name = string("op_6557"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_275_cast_fp16 = reshape(shape = var_6557, x = linear_107_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6559_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_6559_cast_fp16")];
            bool var_6562 = const()[name = string("op_6562"), val = bool(true)];
            tensor<int32, [1]> var_6563_axes_0 = const()[name = string("op_6563_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_6563_cast_fp16 = reduce_max(axes = var_6563_axes_0, keep_dims = var_6562, x = var_6559_cast_fp16)[name = string("op_6563_cast_fp16")];
            fp16 var_6564_to_fp16 = const()[name = string("op_6564_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_219_cast_fp16 = maximum(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_215_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_219_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6567_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6567_cast_fp16")];
            tensor<int32, [1]> var_6572_axes_0 = const()[name = string("op_6572_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6572_keep_dims_0 = const()[name = string("op_6572_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_6572_cast_fp16 = reduce_mean(axes = var_6572_axes_0, keep_dims = var_6572_keep_dims_0, x = var_6567_cast_fp16)[name = string("op_6572_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6573_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6573_cast_fp16")];
            fp32 var_6574_epsilon_0 = const()[name = string("op_6574_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_6574_cast_fp16 = inverse(epsilon = var_6574_epsilon_0, x = var_6573_cast_fp16)[name = string("op_6574_cast_fp16")];
            fp16 var_6575_to_fp16 = const()[name = string("op_6575_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_6576_cast_fp16 = mul(x = var_6574_cast_fp16, y = var_6575_to_fp16)[name = string("op_6576_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6578_cast_fp16 = add(x = var_6572_cast_fp16, y = var_6576_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp32 var_6579_epsilon_0 = const()[name = string("op_6579_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_6579_cast_fp16 = rsqrt(epsilon = var_6579_epsilon_0, x = var_6578_cast_fp16)[name = string("op_6579_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> out_185_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6579_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360704)))];
            tensor<fp16, [1, 2520, 12, 64]> x_281_cast_fp16 = mul(x = out_185_cast_fp16, y = t_encoder_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143754432))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144344320))))[name = string("t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [4]> var_6589 = const()[name = string("op_6589"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 2520, 12, 64]> x_277_cast_fp16 = reshape(shape = var_6589, x = linear_108_cast_fp16)[name = string("x_277_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6591_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_6591_cast_fp16")];
            bool var_6594 = const()[name = string("op_6594"), val = bool(true)];
            tensor<int32, [1]> var_6595_axes_0 = const()[name = string("op_6595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 12, 1]> var_6595_cast_fp16 = reduce_max(axes = var_6595_axes_0, keep_dims = var_6594, x = var_6591_cast_fp16)[name = string("op_6595_cast_fp16")];
            fp16 var_6596_to_fp16 = const()[name = string("op_6596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 12, 1]> s_221_cast_fp16 = maximum(x = var_6595_cast_fp16, y = var_6596_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> y_217_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_221_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> var_6599_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6599_cast_fp16")];
            tensor<int32, [1]> var_6604_axes_0 = const()[name = string("op_6604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6604_keep_dims_0 = const()[name = string("op_6604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 12, 1]> var_6604_cast_fp16 = reduce_mean(axes = var_6604_axes_0, keep_dims = var_6604_keep_dims_0, x = var_6599_cast_fp16)[name = string("op_6604_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6605_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6605_cast_fp16")];
            fp32 var_6606_epsilon_0 = const()[name = string("op_6606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 12, 1]> var_6606_cast_fp16 = inverse(epsilon = var_6606_epsilon_0, x = var_6605_cast_fp16)[name = string("op_6606_cast_fp16")];
            fp16 var_6607_to_fp16 = const()[name = string("op_6607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 12, 1]> var_6608_cast_fp16 = mul(x = var_6606_cast_fp16, y = var_6607_to_fp16)[name = string("op_6608_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 1]> var_6610_cast_fp16 = add(x = var_6604_cast_fp16, y = var_6608_cast_fp16)[name = string("op_6610_cast_fp16")];
            fp32 var_6611_epsilon_0 = const()[name = string("op_6611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 12, 1]> var_6611_cast_fp16 = rsqrt(epsilon = var_6611_epsilon_0, x = var_6610_cast_fp16)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 64]> v_61_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6611_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<int32, [4]> z_121_begin_0 = const()[name = string("z_121_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_121_end_0 = const()[name = string("z_121_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_121_end_mask_0 = const()[name = string("z_121_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_121_cast_fp16 = slice_by_index(begin = z_121_begin_0, end = z_121_end_0, end_mask = z_121_end_mask_0, x = x_279_cast_fp16)[name = string("z_121_cast_fp16")];
            tensor<int32, [4]> var_6622_begin_0 = const()[name = string("op_6622_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6622_end_0 = const()[name = string("op_6622_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6622_end_mask_0 = const()[name = string("op_6622_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6622_cast_fp16 = slice_by_index(begin = var_6622_begin_0, end = var_6622_end_0, end_mask = var_6622_end_mask_0, x = z_121_cast_fp16)[name = string("op_6622_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6623_cast_fp16 = mul(x = var_6622_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_6623_cast_fp16")];
            tensor<int32, [4]> var_6628_begin_0 = const()[name = string("op_6628_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6628_end_0 = const()[name = string("op_6628_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6628_end_mask_0 = const()[name = string("op_6628_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6628_cast_fp16 = slice_by_index(begin = var_6628_begin_0, end = var_6628_end_0, end_mask = var_6628_end_mask_0, x = z_121_cast_fp16)[name = string("op_6628_cast_fp16")];
            int32 var_6630 = const()[name = string("op_6630"), val = int32(-1)];
            bool rot_121_interleave_0 = const()[name = string("rot_121_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_121_cast_fp16 = concat(axis = var_6630, interleave = rot_121_interleave_0, values = (var_6623_cast_fp16, var_6628_cast_fp16))[name = string("rot_121_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6632_cast_fp16 = mul(x = z_121_cast_fp16, y = c_1_cast_fp16)[name = string("op_6632_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6633_cast_fp16 = mul(x = rot_121_cast_fp16, y = s_9_cast_fp16)[name = string("op_6633_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6635_cast_fp16 = add(x = var_6632_cast_fp16, y = var_6633_cast_fp16)[name = string("op_6635_cast_fp16")];
            tensor<int32, [4]> z_123_begin_0 = const()[name = string("z_123_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_123_end_0 = const()[name = string("z_123_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_123_end_mask_0 = const()[name = string("z_123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_123_cast_fp16 = slice_by_index(begin = z_123_begin_0, end = z_123_end_0, end_mask = z_123_end_mask_0, x = x_279_cast_fp16)[name = string("z_123_cast_fp16")];
            tensor<int32, [4]> var_6645_begin_0 = const()[name = string("op_6645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6645_end_0 = const()[name = string("op_6645_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6645_end_mask_0 = const()[name = string("op_6645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6645_cast_fp16 = slice_by_index(begin = var_6645_begin_0, end = var_6645_end_0, end_mask = var_6645_end_mask_0, x = z_123_cast_fp16)[name = string("op_6645_cast_fp16")];
            fp16 const_61_promoted_to_fp16 = const()[name = string("const_61_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6646_cast_fp16 = mul(x = var_6645_cast_fp16, y = const_61_promoted_to_fp16)[name = string("op_6646_cast_fp16")];
            tensor<int32, [4]> var_6651_begin_0 = const()[name = string("op_6651_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6651_end_0 = const()[name = string("op_6651_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6651_end_mask_0 = const()[name = string("op_6651_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6651_cast_fp16 = slice_by_index(begin = var_6651_begin_0, end = var_6651_end_0, end_mask = var_6651_end_mask_0, x = z_123_cast_fp16)[name = string("op_6651_cast_fp16")];
            int32 var_6653 = const()[name = string("op_6653"), val = int32(-1)];
            bool rot_123_interleave_0 = const()[name = string("rot_123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_123_cast_fp16 = concat(axis = var_6653, interleave = rot_123_interleave_0, values = (var_6646_cast_fp16, var_6651_cast_fp16))[name = string("rot_123_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6655_cast_fp16 = mul(x = z_123_cast_fp16, y = c_cast_fp16)[name = string("op_6655_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6656_cast_fp16 = mul(x = rot_123_cast_fp16, y = s_11_cast_fp16)[name = string("op_6656_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6658_cast_fp16 = add(x = var_6655_cast_fp16, y = var_6656_cast_fp16)[name = string("op_6658_cast_fp16")];
            int32 var_6660 = const()[name = string("op_6660"), val = int32(-1)];
            bool var_6661_interleave_0 = const()[name = string("op_6661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_6661_cast_fp16 = concat(axis = var_6660, interleave = var_6661_interleave_0, values = (var_6635_cast_fp16, var_6658_cast_fp16))[name = string("op_6661_cast_fp16")];
            tensor<int32, [4]> z_125_begin_0 = const()[name = string("z_125_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_125_end_0 = const()[name = string("z_125_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> z_125_end_mask_0 = const()[name = string("z_125_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 32]> z_125_cast_fp16 = slice_by_index(begin = z_125_begin_0, end = z_125_end_0, end_mask = z_125_end_mask_0, x = x_281_cast_fp16)[name = string("z_125_cast_fp16")];
            tensor<int32, [4]> var_6674_begin_0 = const()[name = string("op_6674_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6674_end_0 = const()[name = string("op_6674_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6674_end_mask_0 = const()[name = string("op_6674_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6674_cast_fp16 = slice_by_index(begin = var_6674_begin_0, end = var_6674_end_0, end_mask = var_6674_end_mask_0, x = z_125_cast_fp16)[name = string("op_6674_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6675_cast_fp16 = mul(x = var_6674_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_6675_cast_fp16")];
            tensor<int32, [4]> var_6680_begin_0 = const()[name = string("op_6680_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6680_end_0 = const()[name = string("op_6680_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6680_end_mask_0 = const()[name = string("op_6680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6680_cast_fp16 = slice_by_index(begin = var_6680_begin_0, end = var_6680_end_0, end_mask = var_6680_end_mask_0, x = z_125_cast_fp16)[name = string("op_6680_cast_fp16")];
            int32 var_6682 = const()[name = string("op_6682"), val = int32(-1)];
            bool rot_125_interleave_0 = const()[name = string("rot_125_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_125_cast_fp16 = concat(axis = var_6682, interleave = rot_125_interleave_0, values = (var_6675_cast_fp16, var_6680_cast_fp16))[name = string("rot_125_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6684_cast_fp16 = mul(x = z_125_cast_fp16, y = c_1_cast_fp16)[name = string("op_6684_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6685_cast_fp16 = mul(x = rot_125_cast_fp16, y = s_9_cast_fp16)[name = string("op_6685_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6687_cast_fp16 = add(x = var_6684_cast_fp16, y = var_6685_cast_fp16)[name = string("op_6687_cast_fp16")];
            tensor<int32, [4]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 1])];
            tensor<bool, [4]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 32]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = x_281_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [4]> var_6697_begin_0 = const()[name = string("op_6697_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6697_end_0 = const()[name = string("op_6697_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 32])];
            tensor<bool, [4]> var_6697_end_mask_0 = const()[name = string("op_6697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 2520, 12, 16]> var_6697_cast_fp16 = slice_by_index(begin = var_6697_begin_0, end = var_6697_end_0, end_mask = var_6697_end_mask_0, x = z_cast_fp16)[name = string("op_6697_cast_fp16")];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 2520, 12, 16]> var_6698_cast_fp16 = mul(x = var_6697_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_6698_cast_fp16")];
            tensor<int32, [4]> var_6703_begin_0 = const()[name = string("op_6703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6703_end_0 = const()[name = string("op_6703_end_0"), val = tensor<int32, [4]>([1, 2520, 12, 16])];
            tensor<bool, [4]> var_6703_end_mask_0 = const()[name = string("op_6703_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 2520, 12, 16]> var_6703_cast_fp16 = slice_by_index(begin = var_6703_begin_0, end = var_6703_end_0, end_mask = var_6703_end_mask_0, x = z_cast_fp16)[name = string("op_6703_cast_fp16")];
            int32 var_6705 = const()[name = string("op_6705"), val = int32(-1)];
            bool rot_interleave_0 = const()[name = string("rot_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 32]> rot_cast_fp16 = concat(axis = var_6705, interleave = rot_interleave_0, values = (var_6698_cast_fp16, var_6703_cast_fp16))[name = string("rot_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6707_cast_fp16 = mul(x = z_cast_fp16, y = c_cast_fp16)[name = string("op_6707_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6708_cast_fp16 = mul(x = rot_cast_fp16, y = s_11_cast_fp16)[name = string("op_6708_cast_fp16")];
            tensor<fp16, [1, 2520, 12, 32]> var_6710_cast_fp16 = add(x = var_6707_cast_fp16, y = var_6708_cast_fp16)[name = string("op_6710_cast_fp16")];
            int32 var_6712 = const()[name = string("op_6712"), val = int32(-1)];
            bool var_6713_interleave_0 = const()[name = string("op_6713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 2520, 12, 64]> var_6713_cast_fp16 = concat(axis = var_6712, interleave = var_6713_interleave_0, values = (var_6687_cast_fp16, var_6710_cast_fp16))[name = string("op_6713_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6723_transpose_x_0 = const()[name = string("op_6723_transpose_x_0"), val = bool(false)];
            bool var_6723_transpose_y_0 = const()[name = string("op_6723_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 2520]> transpose_95 = transpose(perm = transpose_95_perm_0, x = var_6713_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 12, 2520, 64]> transpose_94 = transpose(perm = transpose_94_perm_0, x = var_6661_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 12, 2520, 2520]> var_6723_cast_fp16 = matmul(transpose_x = var_6723_transpose_x_0, transpose_y = var_6723_transpose_y_0, x = transpose_94, y = transpose_95)[name = string("op_6723_cast_fp16")];
            tensor<fp16, [1, 12, 2520, 2520]> var_6725_cast_fp16 = add(x = var_6723_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6725_cast_fp16")];
            int32 var_6726 = const()[name = string("op_6726"), val = int32(-1)];
            tensor<fp16, [1, 12, 2520, 2520]> w_217_cast_fp16 = softmax(axis = var_6726, x = var_6725_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6729_transpose_x_0 = const()[name = string("op_6729_transpose_x_0"), val = bool(false)];
            bool var_6729_transpose_y_0 = const()[name = string("op_6729_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 2520, 64]> v_cast_fp16 = transpose(perm = v_perm_0, x = v_61_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 12, 2520, 64]> var_6729_cast_fp16 = matmul(transpose_x = var_6729_transpose_x_0, transpose_y = var_6729_transpose_y_0, x = w_217_cast_fp16, y = v_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [4]> var_6732_perm_0 = const()[name = string("op_6732_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6736 = const()[name = string("op_6736"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 2520, 12, 64]> var_6732_cast_fp16 = transpose(perm = var_6732_perm_0, x = var_6729_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 2520, 768]> input_155_cast_fp16 = reshape(shape = var_6736, x = var_6732_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144345920))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144935808))))[name = string("t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_109_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6741_cast_fp16 = abs(x = linear_109_cast_fp16)[name = string("op_6741_cast_fp16")];
            bool var_6744 = const()[name = string("op_6744"), val = bool(true)];
            tensor<int32, [1]> var_6745_axes_0 = const()[name = string("op_6745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6745_cast_fp16 = reduce_max(axes = var_6745_axes_0, keep_dims = var_6744, x = var_6741_cast_fp16)[name = string("op_6745_cast_fp16")];
            fp16 var_6746_to_fp16 = const()[name = string("op_6746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_223_cast_fp16 = maximum(x = var_6745_cast_fp16, y = var_6746_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_219_cast_fp16 = real_div(x = linear_109_cast_fp16, y = s_223_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6749_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6749_cast_fp16")];
            tensor<int32, [1]> var_6754_axes_0 = const()[name = string("op_6754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6754_keep_dims_0 = const()[name = string("op_6754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6754_cast_fp16 = reduce_mean(axes = var_6754_axes_0, keep_dims = var_6754_keep_dims_0, x = var_6749_cast_fp16)[name = string("op_6754_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6755_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6755_cast_fp16")];
            fp32 var_6756_epsilon_0 = const()[name = string("op_6756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6756_cast_fp16 = inverse(epsilon = var_6756_epsilon_0, x = var_6755_cast_fp16)[name = string("op_6756_cast_fp16")];
            fp16 var_6757_to_fp16 = const()[name = string("op_6757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6758_cast_fp16 = mul(x = var_6756_cast_fp16, y = var_6757_to_fp16)[name = string("op_6758_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6760_cast_fp16 = add(x = var_6754_cast_fp16, y = var_6758_cast_fp16)[name = string("op_6760_cast_fp16")];
            fp32 var_6761_epsilon_0 = const()[name = string("op_6761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6761_cast_fp16 = rsqrt(epsilon = var_6761_epsilon_0, x = var_6760_cast_fp16)[name = string("op_6761_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_187_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6761_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144937408)))];
            tensor<fp16, [1, 2520, 768]> var_6763_cast_fp16 = mul(x = out_187_cast_fp16, y = t_encoder_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> x_285_cast_fp16 = add(x = x_271_cast_fp16, y = var_6763_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6766_cast_fp16 = abs(x = x_285_cast_fp16)[name = string("op_6766_cast_fp16")];
            bool var_6769 = const()[name = string("op_6769"), val = bool(true)];
            tensor<int32, [1]> var_6770_axes_0 = const()[name = string("op_6770_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6770_cast_fp16 = reduce_max(axes = var_6770_axes_0, keep_dims = var_6769, x = var_6766_cast_fp16)[name = string("op_6770_cast_fp16")];
            fp16 var_6771_to_fp16 = const()[name = string("op_6771_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_225_cast_fp16 = maximum(x = var_6770_cast_fp16, y = var_6771_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_221_cast_fp16 = real_div(x = x_285_cast_fp16, y = s_225_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6774_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6774_cast_fp16")];
            tensor<int32, [1]> var_6779_axes_0 = const()[name = string("op_6779_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6779_keep_dims_0 = const()[name = string("op_6779_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6779_cast_fp16 = reduce_mean(axes = var_6779_axes_0, keep_dims = var_6779_keep_dims_0, x = var_6774_cast_fp16)[name = string("op_6779_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6780_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6780_cast_fp16")];
            fp32 var_6781_epsilon_0 = const()[name = string("op_6781_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6781_cast_fp16 = inverse(epsilon = var_6781_epsilon_0, x = var_6780_cast_fp16)[name = string("op_6781_cast_fp16")];
            fp16 var_6782_to_fp16 = const()[name = string("op_6782_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6783_cast_fp16 = mul(x = var_6781_cast_fp16, y = var_6782_to_fp16)[name = string("op_6783_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6785_cast_fp16 = add(x = var_6779_cast_fp16, y = var_6783_cast_fp16)[name = string("op_6785_cast_fp16")];
            fp32 var_6786_epsilon_0 = const()[name = string("op_6786_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6786_cast_fp16 = rsqrt(epsilon = var_6786_epsilon_0, x = var_6785_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_189_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6786_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144939008)))];
            tensor<fp16, [1, 2520, 768]> input_157_cast_fp16 = mul(x = out_189_cast_fp16, y = t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144940608))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147299968))))[name = string("t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_110_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_110_cast_fp16")];
            string var_6793_mode_0 = const()[name = string("op_6793_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 2520, 3072]> var_6793_cast_fp16 = gelu(mode = var_6793_mode_0, x = linear_110_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147306176))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149665536))))[name = string("t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 3072]> linear_111_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 2520, 3072]> input_161_cast_fp16 = mul(x = var_6793_cast_fp16, y = linear_111_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149671744))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152031104))))[name = string("t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 2520, 768]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6801_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6801_cast_fp16")];
            bool var_6804 = const()[name = string("op_6804"), val = bool(true)];
            tensor<int32, [1]> var_6805_axes_0 = const()[name = string("op_6805_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 2520, 1]> var_6805_cast_fp16 = reduce_max(axes = var_6805_axes_0, keep_dims = var_6804, x = var_6801_cast_fp16)[name = string("op_6805_cast_fp16")];
            fp16 var_6806_to_fp16 = const()[name = string("op_6806_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 2520, 1]> s_227_cast_fp16 = maximum(x = var_6805_cast_fp16, y = var_6806_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> y_223_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_227_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6809_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6809_cast_fp16")];
            tensor<int32, [1]> var_6814_axes_0 = const()[name = string("op_6814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6814_keep_dims_0 = const()[name = string("op_6814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 2520, 1]> var_6814_cast_fp16 = reduce_mean(axes = var_6814_axes_0, keep_dims = var_6814_keep_dims_0, x = var_6809_cast_fp16)[name = string("op_6814_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6815_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6815_cast_fp16")];
            fp32 var_6816_epsilon_0 = const()[name = string("op_6816_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 2520, 1]> var_6816_cast_fp16 = inverse(epsilon = var_6816_epsilon_0, x = var_6815_cast_fp16)[name = string("op_6816_cast_fp16")];
            fp16 var_6817_to_fp16 = const()[name = string("op_6817_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 2520, 1]> var_6818_cast_fp16 = mul(x = var_6816_cast_fp16, y = var_6817_to_fp16)[name = string("op_6818_cast_fp16")];
            tensor<fp16, [1, 2520, 1]> var_6820_cast_fp16 = add(x = var_6814_cast_fp16, y = var_6818_cast_fp16)[name = string("op_6820_cast_fp16")];
            fp32 var_6821_epsilon_0 = const()[name = string("op_6821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 2520, 1]> var_6821_cast_fp16 = rsqrt(epsilon = var_6821_epsilon_0, x = var_6820_cast_fp16)[name = string("op_6821_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> out_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6821_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152032704)))];
            tensor<fp16, [1, 2520, 768]> var_6823_cast_fp16 = mul(x = out_cast_fp16, y = t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6823_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> h_cast_fp16 = add(x = x_285_cast_fp16, y = var_6823_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [1, 2520, 768]> var_6828_cast_fp16 = mul(x = h_cast_fp16, y = var_1413_cast_fp16)[name = string("op_6828_cast_fp16")];
            tensor<int32, [3]> var_6831_begin_0 = const()[name = string("op_6831_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_6831_end_0 = const()[name = string("op_6831_end_0"), val = tensor<int32, [3]>([1, 2520, 768])];
            tensor<bool, [3]> var_6831_end_mask_0 = const()[name = string("op_6831_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> var_6831_squeeze_mask_0 = const()[name = string("op_6831_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<fp16, [2520, 768]> var_6831_cast_fp16 = slice_by_index(begin = var_6831_begin_0, end = var_6831_end_0, end_mask = var_6831_end_mask_0, squeeze_mask = var_6831_squeeze_mask_0, x = var_6828_cast_fp16)[name = string("op_6831_cast_fp16")];
            bool x_transpose_x_0 = const()[name = string("x_transpose_x_0"), val = bool(false)];
            bool x_transpose_y_0 = const()[name = string("x_transpose_y_0"), val = bool(false)];
            tensor<fp16, [280, 768]> x_cast_fp16 = matmul(transpose_x = x_transpose_x_0, transpose_y = x_transpose_y_0, x = pool, y = var_6831_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [280, 768]> var_6833_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_6833_cast_fp16")];
            bool var_6836 = const()[name = string("op_6836"), val = bool(true)];
            tensor<int32, [1]> var_6837_axes_0 = const()[name = string("op_6837_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [280, 1]> var_6837_cast_fp16 = reduce_max(axes = var_6837_axes_0, keep_dims = var_6836, x = var_6833_cast_fp16)[name = string("op_6837_cast_fp16")];
            fp16 var_6838_to_fp16 = const()[name = string("op_6838_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [280, 1]> s_cast_fp16 = maximum(x = var_6837_cast_fp16, y = var_6838_to_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [280, 768]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [280, 768]> var_6841_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_6841_cast_fp16")];
            tensor<int32, [1]> var_6846_axes_0 = const()[name = string("op_6846_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6846_keep_dims_0 = const()[name = string("op_6846_keep_dims_0"), val = bool(true)];
            tensor<fp16, [280, 1]> var_6846_cast_fp16 = reduce_mean(axes = var_6846_axes_0, keep_dims = var_6846_keep_dims_0, x = var_6841_cast_fp16)[name = string("op_6846_cast_fp16")];
            tensor<fp16, [280, 1]> var_6847_cast_fp16 = mul(x = s_cast_fp16, y = s_cast_fp16)[name = string("op_6847_cast_fp16")];
            fp32 var_6848_epsilon_0 = const()[name = string("op_6848_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [280, 1]> var_6848_cast_fp16 = inverse(epsilon = var_6848_epsilon_0, x = var_6847_cast_fp16)[name = string("op_6848_cast_fp16")];
            fp16 var_6849_to_fp16 = const()[name = string("op_6849_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [280, 1]> var_6850_cast_fp16 = mul(x = var_6848_cast_fp16, y = var_6849_to_fp16)[name = string("op_6850_cast_fp16")];
            tensor<fp16, [280, 1]> var_6852_cast_fp16 = add(x = var_6846_cast_fp16, y = var_6850_cast_fp16)[name = string("op_6852_cast_fp16")];
            fp32 var_6853_epsilon_0 = const()[name = string("op_6853_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [280, 1]> var_6853_cast_fp16 = rsqrt(epsilon = var_6853_epsilon_0, x = var_6852_cast_fp16)[name = string("op_6853_cast_fp16")];
            tensor<fp16, [280, 768]> input_cast_fp16 = mul(x = y_cast_fp16, y = var_6853_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [512, 768]> e_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152034304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152427584))))[name = string("e_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_113_bias_0_to_fp16 = const()[name = string("linear_113_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152428672)))];
            tensor<fp16, [280, 512]> linear_113_cast_fp16 = linear(bias = linear_113_bias_0_to_fp16, weight = e_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [1]> var_6859_axes_0 = const()[name = string("op_6859_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 280, 512]> image_tokens = expand_dims(axes = var_6859_axes_0, x = linear_113_cast_fp16)[name = string("op_6859_cast_fp16")];
        } -> (image_tokens);
    func vision_70<ios18>(tensor<fp16, [1, 630, 768]> patches, tensor<fp16, [70, 630]> pool, tensor<fp16, [1, 630, 768]> position_embeddings, tensor<fp16, [1, 630, 2]> positions, tensor<fp16, [1, 630]> valid) {
            fp16 var_1405_to_fp16 = const()[name = string("op_1405_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [1, 630, 768]> var_1406_cast_fp16 = sub(x = patches, y = var_1405_to_fp16)[name = string("op_1406_cast_fp16")];
            fp16 var_1407_promoted_to_fp16 = const()[name = string("op_1407_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 630, 768]> input_1_cast_fp16 = mul(x = var_1406_cast_fp16, y = var_1407_promoted_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [768, 768]> t_patch_embedder_input_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(589952))))[name = string("t_patch_embedder_input_proj_weight_to_fp16_quantized")];
            tensor<fp16, [768]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591552)))];
            tensor<fp16, [1, 630, 768]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_patch_embedder_input_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [1]> var_1413_axes_0 = const()[name = string("op_1413_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 630, 1]> var_1413_cast_fp16 = expand_dims(axes = var_1413_axes_0, x = valid)[name = string("op_1413_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1414_cast_fp16 = mul(x = position_embeddings, y = var_1413_cast_fp16)[name = string("op_1414_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_1_cast_fp16 = add(x = linear_0_cast_fp16, y = var_1414_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [1]> var_1418_axes_0 = const()[name = string("op_1418_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 630, 2, 1]> var_1418_cast_fp16 = expand_dims(axes = var_1418_axes_0, x = positions)[name = string("op_1418_cast_fp16")];
            tensor<fp16, [16]> inv_freq_to_fp16 = const()[name = string("inv_freq_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593152)))];
            tensor<fp16, [1, 630, 2, 16]> f_1_cast_fp16 = mul(x = var_1418_cast_fp16, y = inv_freq_to_fp16)[name = string("f_1_cast_fp16")];
            int32 var_1421 = const()[name = string("op_1421"), val = int32(-1)];
            bool var_1422_interleave_0 = const()[name = string("op_1422_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 2, 32]> var_1422_cast_fp16 = concat(axis = var_1421, interleave = var_1422_interleave_0, values = (f_1_cast_fp16, f_1_cast_fp16))[name = string("op_1422_cast_fp16")];
            tensor<int32, [1]> f_axes_0 = const()[name = string("f_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 630, 1, 2, 32]> f_cast_fp16 = expand_dims(axes = f_axes_0, x = var_1422_cast_fp16)[name = string("f_cast_fp16")];
            tensor<int32, [5]> var_1452_begin_0 = const()[name = string("op_1452_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1452_end_0 = const()[name = string("op_1452_end_0"), val = tensor<int32, [5]>([1, 630, 1, 1, 32])];
            tensor<bool, [5]> var_1452_end_mask_0 = const()[name = string("op_1452_end_mask_0"), val = tensor<bool, [5]>([true, true, true, false, true])];
            tensor<bool, [5]> var_1452_squeeze_mask_0 = const()[name = string("op_1452_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, true, false])];
            tensor<fp16, [1, 630, 1, 32]> var_1452_cast_fp16 = slice_by_index(begin = var_1452_begin_0, end = var_1452_end_0, end_mask = var_1452_end_mask_0, squeeze_mask = var_1452_squeeze_mask_0, x = f_cast_fp16)[name = string("op_1452_cast_fp16")];
            tensor<fp16, [1, 630, 1, 32]> c_1_cast_fp16 = cos(x = var_1452_cast_fp16)[name = string("c_1_cast_fp16")];
            tensor<int32, [5]> var_1471_begin_0 = const()[name = string("op_1471_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 1, 0])];
            tensor<int32, [5]> var_1471_end_0 = const()[name = string("op_1471_end_0"), val = tensor<int32, [5]>([1, 630, 1, 2, 32])];
            tensor<bool, [5]> var_1471_end_mask_0 = const()[name = string("op_1471_end_mask_0"), val = tensor<bool, [5]>([true, true, true, false, true])];
            tensor<bool, [5]> var_1471_squeeze_mask_0 = const()[name = string("op_1471_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, true, false])];
            tensor<fp16, [1, 630, 1, 32]> var_1471_cast_fp16 = slice_by_index(begin = var_1471_begin_0, end = var_1471_end_0, end_mask = var_1471_end_mask_0, squeeze_mask = var_1471_squeeze_mask_0, x = f_cast_fp16)[name = string("op_1471_cast_fp16")];
            tensor<fp16, [1, 630, 1, 32]> c_cast_fp16 = cos(x = var_1471_cast_fp16)[name = string("c_cast_fp16")];
            tensor<fp16, [1, 630, 1, 32]> s_9_cast_fp16 = sin(x = var_1452_cast_fp16)[name = string("s_9_cast_fp16")];
            tensor<fp16, [1, 630, 1, 32]> s_11_cast_fp16 = sin(x = var_1471_cast_fp16)[name = string("s_11_cast_fp16")];
            fp16 var_1511_to_fp16 = const()[name = string("op_1511_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 630]> var_1513_cast_fp16 = sub(x = var_1511_to_fp16, y = valid)[name = string("op_1513_cast_fp16")];
            fp16 var_1514_to_fp16 = const()[name = string("op_1514_to_fp16"), val = fp16(-0x1.388p+13)];
            tensor<fp16, [1, 630]> var_1515_cast_fp16 = mul(x = var_1513_cast_fp16, y = var_1514_to_fp16)[name = string("op_1515_cast_fp16")];
            tensor<int32, [1]> var_1522_axes_0 = const()[name = string("op_1522_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 630]> var_1522_cast_fp16 = expand_dims(axes = var_1522_axes_0, x = var_1515_cast_fp16)[name = string("op_1522_cast_fp16")];
            tensor<int32, [1]> var_1524_axes_0 = const()[name = string("op_1524_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 630]> var_1524_cast_fp16 = expand_dims(axes = var_1524_axes_0, x = var_1522_cast_fp16)[name = string("op_1524_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1530_cast_fp16 = abs(x = x_1_cast_fp16)[name = string("op_1530_cast_fp16")];
            bool var_1533 = const()[name = string("op_1533"), val = bool(true)];
            tensor<int32, [1]> var_1534_axes_0 = const()[name = string("op_1534_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_1534_cast_fp16 = reduce_max(axes = var_1534_axes_0, keep_dims = var_1533, x = var_1530_cast_fp16)[name = string("op_1534_cast_fp16")];
            fp16 var_1535_to_fp16 = const()[name = string("op_1535_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_1_cast_fp16 = maximum(x = var_1534_cast_fp16, y = var_1535_to_fp16)[name = string("s_1_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_1_cast_fp16 = real_div(x = x_1_cast_fp16, y = s_1_cast_fp16)[name = string("y_1_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1538_cast_fp16 = mul(x = y_1_cast_fp16, y = y_1_cast_fp16)[name = string("op_1538_cast_fp16")];
            tensor<int32, [1]> var_1543_axes_0 = const()[name = string("op_1543_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1543_keep_dims_0 = const()[name = string("op_1543_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_1543_cast_fp16 = reduce_mean(axes = var_1543_axes_0, keep_dims = var_1543_keep_dims_0, x = var_1538_cast_fp16)[name = string("op_1543_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1544_cast_fp16 = mul(x = s_1_cast_fp16, y = s_1_cast_fp16)[name = string("op_1544_cast_fp16")];
            fp32 var_1545_epsilon_0 = const()[name = string("op_1545_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_1545_cast_fp16 = inverse(epsilon = var_1545_epsilon_0, x = var_1544_cast_fp16)[name = string("op_1545_cast_fp16")];
            fp16 var_1546_to_fp16 = const()[name = string("op_1546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_1547_cast_fp16 = mul(x = var_1545_cast_fp16, y = var_1546_to_fp16)[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1549_cast_fp16 = add(x = var_1543_cast_fp16, y = var_1547_cast_fp16)[name = string("op_1549_cast_fp16")];
            fp32 var_1550_epsilon_0 = const()[name = string("op_1550_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_1550_cast_fp16 = rsqrt(epsilon = var_1550_epsilon_0, x = var_1549_cast_fp16)[name = string("op_1550_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_1_cast_fp16 = mul(x = y_1_cast_fp16, y = var_1550_cast_fp16)[name = string("out_1_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(593280)))];
            tensor<fp16, [1, 630, 768]> input_3_cast_fp16 = mul(x = out_1_cast_fp16, y = t_encoder_layers_0_input_layernorm_weight_to_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1184768))))[name = string("t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_1_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_1560 = const()[name = string("op_1560"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_3_cast_fp16 = reshape(shape = var_1560, x = linear_1_cast_fp16)[name = string("x_3_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1562_cast_fp16 = abs(x = x_3_cast_fp16)[name = string("op_1562_cast_fp16")];
            bool var_1565 = const()[name = string("op_1565"), val = bool(true)];
            tensor<int32, [1]> var_1566_axes_0 = const()[name = string("op_1566_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_1566_cast_fp16 = reduce_max(axes = var_1566_axes_0, keep_dims = var_1565, x = var_1562_cast_fp16)[name = string("op_1566_cast_fp16")];
            fp16 var_1567_to_fp16 = const()[name = string("op_1567_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_3_cast_fp16 = maximum(x = var_1566_cast_fp16, y = var_1567_to_fp16)[name = string("s_3_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_3_cast_fp16 = real_div(x = x_3_cast_fp16, y = s_3_cast_fp16)[name = string("y_3_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1570_cast_fp16 = mul(x = y_3_cast_fp16, y = y_3_cast_fp16)[name = string("op_1570_cast_fp16")];
            tensor<int32, [1]> var_1575_axes_0 = const()[name = string("op_1575_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1575_keep_dims_0 = const()[name = string("op_1575_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_1575_cast_fp16 = reduce_mean(axes = var_1575_axes_0, keep_dims = var_1575_keep_dims_0, x = var_1570_cast_fp16)[name = string("op_1575_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1576_cast_fp16 = mul(x = s_3_cast_fp16, y = s_3_cast_fp16)[name = string("op_1576_cast_fp16")];
            fp32 var_1577_epsilon_0 = const()[name = string("op_1577_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_1577_cast_fp16 = inverse(epsilon = var_1577_epsilon_0, x = var_1576_cast_fp16)[name = string("op_1577_cast_fp16")];
            fp16 var_1578_to_fp16 = const()[name = string("op_1578_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_1579_cast_fp16 = mul(x = var_1577_cast_fp16, y = var_1578_to_fp16)[name = string("op_1579_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1581_cast_fp16 = add(x = var_1575_cast_fp16, y = var_1579_cast_fp16)[name = string("op_1581_cast_fp16")];
            fp32 var_1582_epsilon_0 = const()[name = string("op_1582_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_1582_cast_fp16 = rsqrt(epsilon = var_1582_epsilon_0, x = var_1581_cast_fp16)[name = string("op_1582_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_3_cast_fp16 = mul(x = y_3_cast_fp16, y = var_1582_cast_fp16)[name = string("out_3_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 630, 12, 64]> x_9_cast_fp16 = mul(x = out_3_cast_fp16, y = t_encoder_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186560))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1776448))))[name = string("t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1592 = const()[name = string("op_1592"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_5_cast_fp16 = reshape(shape = var_1592, x = linear_2_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1594_cast_fp16 = abs(x = x_5_cast_fp16)[name = string("op_1594_cast_fp16")];
            bool var_1597 = const()[name = string("op_1597"), val = bool(true)];
            tensor<int32, [1]> var_1598_axes_0 = const()[name = string("op_1598_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_1598_cast_fp16 = reduce_max(axes = var_1598_axes_0, keep_dims = var_1597, x = var_1594_cast_fp16)[name = string("op_1598_cast_fp16")];
            fp16 var_1599_to_fp16 = const()[name = string("op_1599_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_5_cast_fp16 = maximum(x = var_1598_cast_fp16, y = var_1599_to_fp16)[name = string("s_5_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_5_cast_fp16 = real_div(x = x_5_cast_fp16, y = s_5_cast_fp16)[name = string("y_5_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1602_cast_fp16 = mul(x = y_5_cast_fp16, y = y_5_cast_fp16)[name = string("op_1602_cast_fp16")];
            tensor<int32, [1]> var_1607_axes_0 = const()[name = string("op_1607_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1607_keep_dims_0 = const()[name = string("op_1607_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_1607_cast_fp16 = reduce_mean(axes = var_1607_axes_0, keep_dims = var_1607_keep_dims_0, x = var_1602_cast_fp16)[name = string("op_1607_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1608_cast_fp16 = mul(x = s_5_cast_fp16, y = s_5_cast_fp16)[name = string("op_1608_cast_fp16")];
            fp32 var_1609_epsilon_0 = const()[name = string("op_1609_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_1609_cast_fp16 = inverse(epsilon = var_1609_epsilon_0, x = var_1608_cast_fp16)[name = string("op_1609_cast_fp16")];
            fp16 var_1610_to_fp16 = const()[name = string("op_1610_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_1611_cast_fp16 = mul(x = var_1609_cast_fp16, y = var_1610_to_fp16)[name = string("op_1611_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1613_cast_fp16 = add(x = var_1607_cast_fp16, y = var_1611_cast_fp16)[name = string("op_1613_cast_fp16")];
            fp32 var_1614_epsilon_0 = const()[name = string("op_1614_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_1614_cast_fp16 = rsqrt(epsilon = var_1614_epsilon_0, x = var_1613_cast_fp16)[name = string("op_1614_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_5_cast_fp16 = mul(x = y_5_cast_fp16, y = var_1614_cast_fp16)[name = string("out_5_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 630, 12, 64]> x_11_cast_fp16 = mul(x = out_5_cast_fp16, y = t_encoder_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_11_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778240))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2368128))))[name = string("t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<int32, [4]> var_1624 = const()[name = string("op_1624"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_7_cast_fp16 = reshape(shape = var_1624, x = linear_3_cast_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1626_cast_fp16 = abs(x = x_7_cast_fp16)[name = string("op_1626_cast_fp16")];
            bool var_1629 = const()[name = string("op_1629"), val = bool(true)];
            tensor<int32, [1]> var_1630_axes_0 = const()[name = string("op_1630_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_1630_cast_fp16 = reduce_max(axes = var_1630_axes_0, keep_dims = var_1629, x = var_1626_cast_fp16)[name = string("op_1630_cast_fp16")];
            fp16 var_1631_to_fp16 = const()[name = string("op_1631_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_7_cast_fp16 = maximum(x = var_1630_cast_fp16, y = var_1631_to_fp16)[name = string("s_7_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_7_cast_fp16 = real_div(x = x_7_cast_fp16, y = s_7_cast_fp16)[name = string("y_7_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1634_cast_fp16 = mul(x = y_7_cast_fp16, y = y_7_cast_fp16)[name = string("op_1634_cast_fp16")];
            tensor<int32, [1]> var_1639_axes_0 = const()[name = string("op_1639_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1639_keep_dims_0 = const()[name = string("op_1639_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_1639_cast_fp16 = reduce_mean(axes = var_1639_axes_0, keep_dims = var_1639_keep_dims_0, x = var_1634_cast_fp16)[name = string("op_1639_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1640_cast_fp16 = mul(x = s_7_cast_fp16, y = s_7_cast_fp16)[name = string("op_1640_cast_fp16")];
            fp32 var_1641_epsilon_0 = const()[name = string("op_1641_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_1641_cast_fp16 = inverse(epsilon = var_1641_epsilon_0, x = var_1640_cast_fp16)[name = string("op_1641_cast_fp16")];
            fp16 var_1642_to_fp16 = const()[name = string("op_1642_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_1643_cast_fp16 = mul(x = var_1641_cast_fp16, y = var_1642_to_fp16)[name = string("op_1643_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1645_cast_fp16 = add(x = var_1639_cast_fp16, y = var_1643_cast_fp16)[name = string("op_1645_cast_fp16")];
            fp32 var_1646_epsilon_0 = const()[name = string("op_1646_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_1646_cast_fp16 = rsqrt(epsilon = var_1646_epsilon_0, x = var_1645_cast_fp16)[name = string("op_1646_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_1_cast_fp16 = mul(x = y_7_cast_fp16, y = var_1646_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<int32, [4]> z_1_begin_0 = const()[name = string("z_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_1_end_0 = const()[name = string("z_1_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_1_end_mask_0 = const()[name = string("z_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_1_cast_fp16 = slice_by_index(begin = z_1_begin_0, end = z_1_end_0, end_mask = z_1_end_mask_0, x = x_9_cast_fp16)[name = string("z_1_cast_fp16")];
            tensor<int32, [4]> var_1657_begin_0 = const()[name = string("op_1657_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1657_end_0 = const()[name = string("op_1657_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_1657_end_mask_0 = const()[name = string("op_1657_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_1657_cast_fp16 = slice_by_index(begin = var_1657_begin_0, end = var_1657_end_0, end_mask = var_1657_end_mask_0, x = z_1_cast_fp16)[name = string("op_1657_cast_fp16")];
            fp16 const_0_promoted_to_fp16 = const()[name = string("const_0_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_1658_cast_fp16 = mul(x = var_1657_cast_fp16, y = const_0_promoted_to_fp16)[name = string("op_1658_cast_fp16")];
            tensor<int32, [4]> var_1663_begin_0 = const()[name = string("op_1663_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1663_end_0 = const()[name = string("op_1663_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_1663_end_mask_0 = const()[name = string("op_1663_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_1663_cast_fp16 = slice_by_index(begin = var_1663_begin_0, end = var_1663_end_0, end_mask = var_1663_end_mask_0, x = z_1_cast_fp16)[name = string("op_1663_cast_fp16")];
            int32 var_1665 = const()[name = string("op_1665"), val = int32(-1)];
            bool rot_1_interleave_0 = const()[name = string("rot_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_1_cast_fp16 = concat(axis = var_1665, interleave = rot_1_interleave_0, values = (var_1658_cast_fp16, var_1663_cast_fp16))[name = string("rot_1_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1667_cast_fp16 = mul(x = z_1_cast_fp16, y = c_1_cast_fp16)[name = string("op_1667_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1668_cast_fp16 = mul(x = rot_1_cast_fp16, y = s_9_cast_fp16)[name = string("op_1668_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1670_cast_fp16 = add(x = var_1667_cast_fp16, y = var_1668_cast_fp16)[name = string("op_1670_cast_fp16")];
            tensor<int32, [4]> z_3_begin_0 = const()[name = string("z_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_3_end_0 = const()[name = string("z_3_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_3_end_mask_0 = const()[name = string("z_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_3_cast_fp16 = slice_by_index(begin = z_3_begin_0, end = z_3_end_0, end_mask = z_3_end_mask_0, x = x_9_cast_fp16)[name = string("z_3_cast_fp16")];
            tensor<int32, [4]> var_1680_begin_0 = const()[name = string("op_1680_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1680_end_0 = const()[name = string("op_1680_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_1680_end_mask_0 = const()[name = string("op_1680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_1680_cast_fp16 = slice_by_index(begin = var_1680_begin_0, end = var_1680_end_0, end_mask = var_1680_end_mask_0, x = z_3_cast_fp16)[name = string("op_1680_cast_fp16")];
            fp16 const_1_promoted_to_fp16 = const()[name = string("const_1_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_1681_cast_fp16 = mul(x = var_1680_cast_fp16, y = const_1_promoted_to_fp16)[name = string("op_1681_cast_fp16")];
            tensor<int32, [4]> var_1686_begin_0 = const()[name = string("op_1686_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1686_end_0 = const()[name = string("op_1686_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_1686_end_mask_0 = const()[name = string("op_1686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_1686_cast_fp16 = slice_by_index(begin = var_1686_begin_0, end = var_1686_end_0, end_mask = var_1686_end_mask_0, x = z_3_cast_fp16)[name = string("op_1686_cast_fp16")];
            int32 var_1688 = const()[name = string("op_1688"), val = int32(-1)];
            bool rot_3_interleave_0 = const()[name = string("rot_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_3_cast_fp16 = concat(axis = var_1688, interleave = rot_3_interleave_0, values = (var_1681_cast_fp16, var_1686_cast_fp16))[name = string("rot_3_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1690_cast_fp16 = mul(x = z_3_cast_fp16, y = c_cast_fp16)[name = string("op_1690_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1691_cast_fp16 = mul(x = rot_3_cast_fp16, y = s_11_cast_fp16)[name = string("op_1691_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1693_cast_fp16 = add(x = var_1690_cast_fp16, y = var_1691_cast_fp16)[name = string("op_1693_cast_fp16")];
            int32 var_1695 = const()[name = string("op_1695"), val = int32(-1)];
            bool var_1696_interleave_0 = const()[name = string("op_1696_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_1696_cast_fp16 = concat(axis = var_1695, interleave = var_1696_interleave_0, values = (var_1670_cast_fp16, var_1693_cast_fp16))[name = string("op_1696_cast_fp16")];
            tensor<int32, [4]> z_5_begin_0 = const()[name = string("z_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_5_end_0 = const()[name = string("z_5_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_5_end_mask_0 = const()[name = string("z_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_5_cast_fp16 = slice_by_index(begin = z_5_begin_0, end = z_5_end_0, end_mask = z_5_end_mask_0, x = x_11_cast_fp16)[name = string("z_5_cast_fp16")];
            tensor<int32, [4]> var_1709_begin_0 = const()[name = string("op_1709_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1709_end_0 = const()[name = string("op_1709_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_1709_end_mask_0 = const()[name = string("op_1709_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_1709_cast_fp16 = slice_by_index(begin = var_1709_begin_0, end = var_1709_end_0, end_mask = var_1709_end_mask_0, x = z_5_cast_fp16)[name = string("op_1709_cast_fp16")];
            fp16 const_2_promoted_to_fp16 = const()[name = string("const_2_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_1710_cast_fp16 = mul(x = var_1709_cast_fp16, y = const_2_promoted_to_fp16)[name = string("op_1710_cast_fp16")];
            tensor<int32, [4]> var_1715_begin_0 = const()[name = string("op_1715_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1715_end_0 = const()[name = string("op_1715_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_1715_end_mask_0 = const()[name = string("op_1715_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_1715_cast_fp16 = slice_by_index(begin = var_1715_begin_0, end = var_1715_end_0, end_mask = var_1715_end_mask_0, x = z_5_cast_fp16)[name = string("op_1715_cast_fp16")];
            int32 var_1717 = const()[name = string("op_1717"), val = int32(-1)];
            bool rot_5_interleave_0 = const()[name = string("rot_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_5_cast_fp16 = concat(axis = var_1717, interleave = rot_5_interleave_0, values = (var_1710_cast_fp16, var_1715_cast_fp16))[name = string("rot_5_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1719_cast_fp16 = mul(x = z_5_cast_fp16, y = c_1_cast_fp16)[name = string("op_1719_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1720_cast_fp16 = mul(x = rot_5_cast_fp16, y = s_9_cast_fp16)[name = string("op_1720_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1722_cast_fp16 = add(x = var_1719_cast_fp16, y = var_1720_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<int32, [4]> z_7_begin_0 = const()[name = string("z_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_7_end_0 = const()[name = string("z_7_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_7_end_mask_0 = const()[name = string("z_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_7_cast_fp16 = slice_by_index(begin = z_7_begin_0, end = z_7_end_0, end_mask = z_7_end_mask_0, x = x_11_cast_fp16)[name = string("z_7_cast_fp16")];
            tensor<int32, [4]> var_1732_begin_0 = const()[name = string("op_1732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1732_end_0 = const()[name = string("op_1732_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_1732_end_mask_0 = const()[name = string("op_1732_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_1732_cast_fp16 = slice_by_index(begin = var_1732_begin_0, end = var_1732_end_0, end_mask = var_1732_end_mask_0, x = z_7_cast_fp16)[name = string("op_1732_cast_fp16")];
            fp16 const_3_promoted_to_fp16 = const()[name = string("const_3_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_1733_cast_fp16 = mul(x = var_1732_cast_fp16, y = const_3_promoted_to_fp16)[name = string("op_1733_cast_fp16")];
            tensor<int32, [4]> var_1738_begin_0 = const()[name = string("op_1738_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1738_end_0 = const()[name = string("op_1738_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_1738_end_mask_0 = const()[name = string("op_1738_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_1738_cast_fp16 = slice_by_index(begin = var_1738_begin_0, end = var_1738_end_0, end_mask = var_1738_end_mask_0, x = z_7_cast_fp16)[name = string("op_1738_cast_fp16")];
            int32 var_1740 = const()[name = string("op_1740"), val = int32(-1)];
            bool rot_7_interleave_0 = const()[name = string("rot_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_7_cast_fp16 = concat(axis = var_1740, interleave = rot_7_interleave_0, values = (var_1733_cast_fp16, var_1738_cast_fp16))[name = string("rot_7_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1742_cast_fp16 = mul(x = z_7_cast_fp16, y = c_cast_fp16)[name = string("op_1742_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1743_cast_fp16 = mul(x = rot_7_cast_fp16, y = s_11_cast_fp16)[name = string("op_1743_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1745_cast_fp16 = add(x = var_1742_cast_fp16, y = var_1743_cast_fp16)[name = string("op_1745_cast_fp16")];
            int32 var_1747 = const()[name = string("op_1747"), val = int32(-1)];
            bool var_1748_interleave_0 = const()[name = string("op_1748_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_1748_cast_fp16 = concat(axis = var_1747, interleave = var_1748_interleave_0, values = (var_1722_cast_fp16, var_1745_cast_fp16))[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_1758_transpose_x_0 = const()[name = string("op_1758_transpose_x_0"), val = bool(false)];
            bool var_1758_transpose_y_0 = const()[name = string("op_1758_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_64_perm_0 = const()[name = string("transpose_64_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_65_perm_0 = const()[name = string("transpose_65_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_65 = transpose(perm = transpose_65_perm_0, x = var_1748_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 12, 630, 64]> transpose_64 = transpose(perm = transpose_64_perm_0, x = var_1696_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 12, 630, 630]> var_1758_cast_fp16 = matmul(transpose_x = var_1758_transpose_x_0, transpose_y = var_1758_transpose_y_0, x = transpose_64, y = transpose_65)[name = string("op_1758_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_1760_cast_fp16 = add(x = var_1758_cast_fp16, y = var_1524_cast_fp16)[name = string("op_1760_cast_fp16")];
            int32 var_1761 = const()[name = string("op_1761"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_7_cast_fp16 = softmax(axis = var_1761, x = var_1760_cast_fp16)[name = string("w_7_cast_fp16")];
            bool var_1764_transpose_x_0 = const()[name = string("op_1764_transpose_x_0"), val = bool(false)];
            bool var_1764_transpose_y_0 = const()[name = string("op_1764_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = v_1_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 12, 630, 64]> var_1764_cast_fp16 = matmul(transpose_x = var_1764_transpose_x_0, transpose_y = var_1764_transpose_y_0, x = w_7_cast_fp16, y = v_3_cast_fp16)[name = string("op_1764_cast_fp16")];
            tensor<int32, [4]> var_1767_perm_0 = const()[name = string("op_1767_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1771 = const()[name = string("op_1771"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = var_1764_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 630, 768]> input_5_cast_fp16 = reshape(shape = var_1771, x = var_1767_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2369728))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2959616))))[name = string("t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1776_cast_fp16 = abs(x = linear_4_cast_fp16)[name = string("op_1776_cast_fp16")];
            bool var_1779 = const()[name = string("op_1779"), val = bool(true)];
            tensor<int32, [1]> var_1780_axes_0 = const()[name = string("op_1780_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_1780_cast_fp16 = reduce_max(axes = var_1780_axes_0, keep_dims = var_1779, x = var_1776_cast_fp16)[name = string("op_1780_cast_fp16")];
            fp16 var_1781_to_fp16 = const()[name = string("op_1781_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_13_cast_fp16 = maximum(x = var_1780_cast_fp16, y = var_1781_to_fp16)[name = string("s_13_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_9_cast_fp16 = real_div(x = linear_4_cast_fp16, y = s_13_cast_fp16)[name = string("y_9_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1784_cast_fp16 = mul(x = y_9_cast_fp16, y = y_9_cast_fp16)[name = string("op_1784_cast_fp16")];
            tensor<int32, [1]> var_1789_axes_0 = const()[name = string("op_1789_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1789_keep_dims_0 = const()[name = string("op_1789_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_1789_cast_fp16 = reduce_mean(axes = var_1789_axes_0, keep_dims = var_1789_keep_dims_0, x = var_1784_cast_fp16)[name = string("op_1789_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1790_cast_fp16 = mul(x = s_13_cast_fp16, y = s_13_cast_fp16)[name = string("op_1790_cast_fp16")];
            fp32 var_1791_epsilon_0 = const()[name = string("op_1791_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_1791_cast_fp16 = inverse(epsilon = var_1791_epsilon_0, x = var_1790_cast_fp16)[name = string("op_1791_cast_fp16")];
            fp16 var_1792_to_fp16 = const()[name = string("op_1792_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_1793_cast_fp16 = mul(x = var_1791_cast_fp16, y = var_1792_to_fp16)[name = string("op_1793_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1795_cast_fp16 = add(x = var_1789_cast_fp16, y = var_1793_cast_fp16)[name = string("op_1795_cast_fp16")];
            fp32 var_1796_epsilon_0 = const()[name = string("op_1796_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_1796_cast_fp16 = rsqrt(epsilon = var_1796_epsilon_0, x = var_1795_cast_fp16)[name = string("op_1796_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_7_cast_fp16 = mul(x = y_9_cast_fp16, y = var_1796_cast_fp16)[name = string("out_7_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2961216)))];
            tensor<fp16, [1, 630, 768]> var_1798_cast_fp16 = mul(x = out_7_cast_fp16, y = t_encoder_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("op_1798_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_15_cast_fp16 = add(x = x_1_cast_fp16, y = var_1798_cast_fp16)[name = string("x_15_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1801_cast_fp16 = abs(x = x_15_cast_fp16)[name = string("op_1801_cast_fp16")];
            bool var_1804 = const()[name = string("op_1804"), val = bool(true)];
            tensor<int32, [1]> var_1805_axes_0 = const()[name = string("op_1805_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_1805_cast_fp16 = reduce_max(axes = var_1805_axes_0, keep_dims = var_1804, x = var_1801_cast_fp16)[name = string("op_1805_cast_fp16")];
            fp16 var_1806_to_fp16 = const()[name = string("op_1806_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_15_cast_fp16 = maximum(x = var_1805_cast_fp16, y = var_1806_to_fp16)[name = string("s_15_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_11_cast_fp16 = real_div(x = x_15_cast_fp16, y = s_15_cast_fp16)[name = string("y_11_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1809_cast_fp16 = mul(x = y_11_cast_fp16, y = y_11_cast_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [1]> var_1814_axes_0 = const()[name = string("op_1814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1814_keep_dims_0 = const()[name = string("op_1814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_1814_cast_fp16 = reduce_mean(axes = var_1814_axes_0, keep_dims = var_1814_keep_dims_0, x = var_1809_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1815_cast_fp16 = mul(x = s_15_cast_fp16, y = s_15_cast_fp16)[name = string("op_1815_cast_fp16")];
            fp32 var_1816_epsilon_0 = const()[name = string("op_1816_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_1816_cast_fp16 = inverse(epsilon = var_1816_epsilon_0, x = var_1815_cast_fp16)[name = string("op_1816_cast_fp16")];
            fp16 var_1817_to_fp16 = const()[name = string("op_1817_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_1818_cast_fp16 = mul(x = var_1816_cast_fp16, y = var_1817_to_fp16)[name = string("op_1818_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1820_cast_fp16 = add(x = var_1814_cast_fp16, y = var_1818_cast_fp16)[name = string("op_1820_cast_fp16")];
            fp32 var_1821_epsilon_0 = const()[name = string("op_1821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_1821_cast_fp16 = rsqrt(epsilon = var_1821_epsilon_0, x = var_1820_cast_fp16)[name = string("op_1821_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_9_cast_fp16 = mul(x = y_11_cast_fp16, y = var_1821_cast_fp16)[name = string("out_9_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2962816)))];
            tensor<fp16, [1, 630, 768]> input_7_cast_fp16 = mul(x = out_9_cast_fp16, y = t_encoder_layers_0_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2964416))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5323776))))[name = string("t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_5_bias_0_to_fp16 = const()[name = string("linear_5_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5329984)))];
            tensor<fp16, [1, 630, 3072]> linear_5_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_5_cast_fp16")];
            string var_1828_mode_0 = const()[name = string("op_1828_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_1828_cast_fp16 = gelu(mode = var_1828_mode_0, x = linear_5_cast_fp16)[name = string("op_1828_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5336192))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7695552))))[name = string("t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_6_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_7_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_11_cast_fp16 = mul(x = var_1828_cast_fp16, y = linear_6_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7701760))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10061120))))[name = string("t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_0_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1836_cast_fp16 = abs(x = linear_7_cast_fp16)[name = string("op_1836_cast_fp16")];
            bool var_1839 = const()[name = string("op_1839"), val = bool(true)];
            tensor<int32, [1]> var_1840_axes_0 = const()[name = string("op_1840_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_1840_cast_fp16 = reduce_max(axes = var_1840_axes_0, keep_dims = var_1839, x = var_1836_cast_fp16)[name = string("op_1840_cast_fp16")];
            fp16 var_1841_to_fp16 = const()[name = string("op_1841_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_17_cast_fp16 = maximum(x = var_1840_cast_fp16, y = var_1841_to_fp16)[name = string("s_17_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_13_cast_fp16 = real_div(x = linear_7_cast_fp16, y = s_17_cast_fp16)[name = string("y_13_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1844_cast_fp16 = mul(x = y_13_cast_fp16, y = y_13_cast_fp16)[name = string("op_1844_cast_fp16")];
            tensor<int32, [1]> var_1849_axes_0 = const()[name = string("op_1849_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1849_keep_dims_0 = const()[name = string("op_1849_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_1849_cast_fp16 = reduce_mean(axes = var_1849_axes_0, keep_dims = var_1849_keep_dims_0, x = var_1844_cast_fp16)[name = string("op_1849_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1850_cast_fp16 = mul(x = s_17_cast_fp16, y = s_17_cast_fp16)[name = string("op_1850_cast_fp16")];
            fp32 var_1851_epsilon_0 = const()[name = string("op_1851_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_1851_cast_fp16 = inverse(epsilon = var_1851_epsilon_0, x = var_1850_cast_fp16)[name = string("op_1851_cast_fp16")];
            fp16 var_1852_to_fp16 = const()[name = string("op_1852_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_1853_cast_fp16 = mul(x = var_1851_cast_fp16, y = var_1852_to_fp16)[name = string("op_1853_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1855_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1853_cast_fp16)[name = string("op_1855_cast_fp16")];
            fp32 var_1856_epsilon_0 = const()[name = string("op_1856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_1856_cast_fp16 = rsqrt(epsilon = var_1856_epsilon_0, x = var_1855_cast_fp16)[name = string("op_1856_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_11_cast_fp16 = mul(x = y_13_cast_fp16, y = var_1856_cast_fp16)[name = string("out_11_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10062720)))];
            tensor<fp16, [1, 630, 768]> var_1858_cast_fp16 = mul(x = out_11_cast_fp16, y = t_encoder_layers_0_post_feedforward_layernorm_weight_to_fp16)[name = string("op_1858_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_19_cast_fp16 = add(x = x_15_cast_fp16, y = var_1858_cast_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1861_cast_fp16 = abs(x = x_19_cast_fp16)[name = string("op_1861_cast_fp16")];
            bool var_1864 = const()[name = string("op_1864"), val = bool(true)];
            tensor<int32, [1]> var_1865_axes_0 = const()[name = string("op_1865_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_1865_cast_fp16 = reduce_max(axes = var_1865_axes_0, keep_dims = var_1864, x = var_1861_cast_fp16)[name = string("op_1865_cast_fp16")];
            fp16 var_1866_to_fp16 = const()[name = string("op_1866_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_19_cast_fp16 = maximum(x = var_1865_cast_fp16, y = var_1866_to_fp16)[name = string("s_19_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_15_cast_fp16 = real_div(x = x_19_cast_fp16, y = s_19_cast_fp16)[name = string("y_15_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_1869_cast_fp16 = mul(x = y_15_cast_fp16, y = y_15_cast_fp16)[name = string("op_1869_cast_fp16")];
            tensor<int32, [1]> var_1874_axes_0 = const()[name = string("op_1874_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1874_keep_dims_0 = const()[name = string("op_1874_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_1874_cast_fp16 = reduce_mean(axes = var_1874_axes_0, keep_dims = var_1874_keep_dims_0, x = var_1869_cast_fp16)[name = string("op_1874_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1875_cast_fp16 = mul(x = s_19_cast_fp16, y = s_19_cast_fp16)[name = string("op_1875_cast_fp16")];
            fp32 var_1876_epsilon_0 = const()[name = string("op_1876_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_1876_cast_fp16 = inverse(epsilon = var_1876_epsilon_0, x = var_1875_cast_fp16)[name = string("op_1876_cast_fp16")];
            fp16 var_1877_to_fp16 = const()[name = string("op_1877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_1878_cast_fp16 = mul(x = var_1876_cast_fp16, y = var_1877_to_fp16)[name = string("op_1878_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_1880_cast_fp16 = add(x = var_1874_cast_fp16, y = var_1878_cast_fp16)[name = string("op_1880_cast_fp16")];
            fp32 var_1881_epsilon_0 = const()[name = string("op_1881_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_1881_cast_fp16 = rsqrt(epsilon = var_1881_epsilon_0, x = var_1880_cast_fp16)[name = string("op_1881_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_13_cast_fp16 = mul(x = y_15_cast_fp16, y = var_1881_cast_fp16)[name = string("out_13_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10064320)))];
            tensor<fp16, [1, 630, 768]> input_13_cast_fp16 = mul(x = out_13_cast_fp16, y = t_encoder_layers_1_input_layernorm_weight_to_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10065920))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10655808))))[name = string("t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_1891 = const()[name = string("op_1891"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_21_cast_fp16 = reshape(shape = var_1891, x = linear_8_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1893_cast_fp16 = abs(x = x_21_cast_fp16)[name = string("op_1893_cast_fp16")];
            bool var_1896 = const()[name = string("op_1896"), val = bool(true)];
            tensor<int32, [1]> var_1897_axes_0 = const()[name = string("op_1897_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_1897_cast_fp16 = reduce_max(axes = var_1897_axes_0, keep_dims = var_1896, x = var_1893_cast_fp16)[name = string("op_1897_cast_fp16")];
            fp16 var_1898_to_fp16 = const()[name = string("op_1898_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_21_cast_fp16 = maximum(x = var_1897_cast_fp16, y = var_1898_to_fp16)[name = string("s_21_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_17_cast_fp16 = real_div(x = x_21_cast_fp16, y = s_21_cast_fp16)[name = string("y_17_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1901_cast_fp16 = mul(x = y_17_cast_fp16, y = y_17_cast_fp16)[name = string("op_1901_cast_fp16")];
            tensor<int32, [1]> var_1906_axes_0 = const()[name = string("op_1906_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1906_keep_dims_0 = const()[name = string("op_1906_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_1906_cast_fp16 = reduce_mean(axes = var_1906_axes_0, keep_dims = var_1906_keep_dims_0, x = var_1901_cast_fp16)[name = string("op_1906_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1907_cast_fp16 = mul(x = s_21_cast_fp16, y = s_21_cast_fp16)[name = string("op_1907_cast_fp16")];
            fp32 var_1908_epsilon_0 = const()[name = string("op_1908_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_1908_cast_fp16 = inverse(epsilon = var_1908_epsilon_0, x = var_1907_cast_fp16)[name = string("op_1908_cast_fp16")];
            fp16 var_1909_to_fp16 = const()[name = string("op_1909_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_1910_cast_fp16 = mul(x = var_1908_cast_fp16, y = var_1909_to_fp16)[name = string("op_1910_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1912_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1910_cast_fp16)[name = string("op_1912_cast_fp16")];
            fp32 var_1913_epsilon_0 = const()[name = string("op_1913_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_1913_cast_fp16 = rsqrt(epsilon = var_1913_epsilon_0, x = var_1912_cast_fp16)[name = string("op_1913_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_15_cast_fp16 = mul(x = y_17_cast_fp16, y = var_1913_cast_fp16)[name = string("out_15_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 630, 12, 64]> x_27_cast_fp16 = mul(x = out_15_cast_fp16, y = t_encoder_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_27_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10657408))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11247296))))[name = string("t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_9_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_1923 = const()[name = string("op_1923"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_23_cast_fp16 = reshape(shape = var_1923, x = linear_9_cast_fp16)[name = string("x_23_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1925_cast_fp16 = abs(x = x_23_cast_fp16)[name = string("op_1925_cast_fp16")];
            bool var_1928 = const()[name = string("op_1928"), val = bool(true)];
            tensor<int32, [1]> var_1929_axes_0 = const()[name = string("op_1929_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_1929_cast_fp16 = reduce_max(axes = var_1929_axes_0, keep_dims = var_1928, x = var_1925_cast_fp16)[name = string("op_1929_cast_fp16")];
            fp16 var_1930_to_fp16 = const()[name = string("op_1930_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_23_cast_fp16 = maximum(x = var_1929_cast_fp16, y = var_1930_to_fp16)[name = string("s_23_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_19_cast_fp16 = real_div(x = x_23_cast_fp16, y = s_23_cast_fp16)[name = string("y_19_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1933_cast_fp16 = mul(x = y_19_cast_fp16, y = y_19_cast_fp16)[name = string("op_1933_cast_fp16")];
            tensor<int32, [1]> var_1938_axes_0 = const()[name = string("op_1938_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1938_keep_dims_0 = const()[name = string("op_1938_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_1938_cast_fp16 = reduce_mean(axes = var_1938_axes_0, keep_dims = var_1938_keep_dims_0, x = var_1933_cast_fp16)[name = string("op_1938_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1939_cast_fp16 = mul(x = s_23_cast_fp16, y = s_23_cast_fp16)[name = string("op_1939_cast_fp16")];
            fp32 var_1940_epsilon_0 = const()[name = string("op_1940_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_1940_cast_fp16 = inverse(epsilon = var_1940_epsilon_0, x = var_1939_cast_fp16)[name = string("op_1940_cast_fp16")];
            fp16 var_1941_to_fp16 = const()[name = string("op_1941_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_1942_cast_fp16 = mul(x = var_1940_cast_fp16, y = var_1941_to_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1944_cast_fp16 = add(x = var_1938_cast_fp16, y = var_1942_cast_fp16)[name = string("op_1944_cast_fp16")];
            fp32 var_1945_epsilon_0 = const()[name = string("op_1945_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_1945_cast_fp16 = rsqrt(epsilon = var_1945_epsilon_0, x = var_1944_cast_fp16)[name = string("op_1945_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_17_cast_fp16 = mul(x = y_19_cast_fp16, y = var_1945_cast_fp16)[name = string("out_17_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 630, 12, 64]> x_29_cast_fp16 = mul(x = out_17_cast_fp16, y = t_encoder_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_29_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11248896))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11838784))))[name = string("t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_25_cast_fp16 = reshape(shape = var_1955, x = linear_10_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1957_cast_fp16 = abs(x = x_25_cast_fp16)[name = string("op_1957_cast_fp16")];
            bool var_1960 = const()[name = string("op_1960"), val = bool(true)];
            tensor<int32, [1]> var_1961_axes_0 = const()[name = string("op_1961_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_1961_cast_fp16 = reduce_max(axes = var_1961_axes_0, keep_dims = var_1960, x = var_1957_cast_fp16)[name = string("op_1961_cast_fp16")];
            fp16 var_1962_to_fp16 = const()[name = string("op_1962_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_25_cast_fp16 = maximum(x = var_1961_cast_fp16, y = var_1962_to_fp16)[name = string("s_25_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_21_cast_fp16 = real_div(x = x_25_cast_fp16, y = s_25_cast_fp16)[name = string("y_21_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_1965_cast_fp16 = mul(x = y_21_cast_fp16, y = y_21_cast_fp16)[name = string("op_1965_cast_fp16")];
            tensor<int32, [1]> var_1970_axes_0 = const()[name = string("op_1970_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1970_keep_dims_0 = const()[name = string("op_1970_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_1970_cast_fp16 = reduce_mean(axes = var_1970_axes_0, keep_dims = var_1970_keep_dims_0, x = var_1965_cast_fp16)[name = string("op_1970_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1971_cast_fp16 = mul(x = s_25_cast_fp16, y = s_25_cast_fp16)[name = string("op_1971_cast_fp16")];
            fp32 var_1972_epsilon_0 = const()[name = string("op_1972_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_1972_cast_fp16 = inverse(epsilon = var_1972_epsilon_0, x = var_1971_cast_fp16)[name = string("op_1972_cast_fp16")];
            fp16 var_1973_to_fp16 = const()[name = string("op_1973_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_1974_cast_fp16 = mul(x = var_1972_cast_fp16, y = var_1973_to_fp16)[name = string("op_1974_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_1976_cast_fp16 = add(x = var_1970_cast_fp16, y = var_1974_cast_fp16)[name = string("op_1976_cast_fp16")];
            fp32 var_1977_epsilon_0 = const()[name = string("op_1977_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_1977_cast_fp16 = rsqrt(epsilon = var_1977_epsilon_0, x = var_1976_cast_fp16)[name = string("op_1977_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_5_cast_fp16 = mul(x = y_21_cast_fp16, y = var_1977_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<int32, [4]> z_9_begin_0 = const()[name = string("z_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_9_end_0 = const()[name = string("z_9_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_9_end_mask_0 = const()[name = string("z_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_9_cast_fp16 = slice_by_index(begin = z_9_begin_0, end = z_9_end_0, end_mask = z_9_end_mask_0, x = x_27_cast_fp16)[name = string("z_9_cast_fp16")];
            tensor<int32, [4]> var_1988_begin_0 = const()[name = string("op_1988_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_1988_end_0 = const()[name = string("op_1988_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_1988_end_mask_0 = const()[name = string("op_1988_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_1988_cast_fp16 = slice_by_index(begin = var_1988_begin_0, end = var_1988_end_0, end_mask = var_1988_end_mask_0, x = z_9_cast_fp16)[name = string("op_1988_cast_fp16")];
            fp16 const_4_promoted_to_fp16 = const()[name = string("const_4_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_1989_cast_fp16 = mul(x = var_1988_cast_fp16, y = const_4_promoted_to_fp16)[name = string("op_1989_cast_fp16")];
            tensor<int32, [4]> var_1994_begin_0 = const()[name = string("op_1994_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1994_end_0 = const()[name = string("op_1994_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_1994_end_mask_0 = const()[name = string("op_1994_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_1994_cast_fp16 = slice_by_index(begin = var_1994_begin_0, end = var_1994_end_0, end_mask = var_1994_end_mask_0, x = z_9_cast_fp16)[name = string("op_1994_cast_fp16")];
            int32 var_1996 = const()[name = string("op_1996"), val = int32(-1)];
            bool rot_9_interleave_0 = const()[name = string("rot_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_9_cast_fp16 = concat(axis = var_1996, interleave = rot_9_interleave_0, values = (var_1989_cast_fp16, var_1994_cast_fp16))[name = string("rot_9_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1998_cast_fp16 = mul(x = z_9_cast_fp16, y = c_1_cast_fp16)[name = string("op_1998_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_1999_cast_fp16 = mul(x = rot_9_cast_fp16, y = s_9_cast_fp16)[name = string("op_1999_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2001_cast_fp16 = add(x = var_1998_cast_fp16, y = var_1999_cast_fp16)[name = string("op_2001_cast_fp16")];
            tensor<int32, [4]> z_11_begin_0 = const()[name = string("z_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_11_end_0 = const()[name = string("z_11_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_11_end_mask_0 = const()[name = string("z_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_11_cast_fp16 = slice_by_index(begin = z_11_begin_0, end = z_11_end_0, end_mask = z_11_end_mask_0, x = x_27_cast_fp16)[name = string("z_11_cast_fp16")];
            tensor<int32, [4]> var_2011_begin_0 = const()[name = string("op_2011_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2011_end_0 = const()[name = string("op_2011_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2011_end_mask_0 = const()[name = string("op_2011_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2011_cast_fp16 = slice_by_index(begin = var_2011_begin_0, end = var_2011_end_0, end_mask = var_2011_end_mask_0, x = z_11_cast_fp16)[name = string("op_2011_cast_fp16")];
            fp16 const_5_promoted_to_fp16 = const()[name = string("const_5_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2012_cast_fp16 = mul(x = var_2011_cast_fp16, y = const_5_promoted_to_fp16)[name = string("op_2012_cast_fp16")];
            tensor<int32, [4]> var_2017_begin_0 = const()[name = string("op_2017_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2017_end_0 = const()[name = string("op_2017_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2017_end_mask_0 = const()[name = string("op_2017_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2017_cast_fp16 = slice_by_index(begin = var_2017_begin_0, end = var_2017_end_0, end_mask = var_2017_end_mask_0, x = z_11_cast_fp16)[name = string("op_2017_cast_fp16")];
            int32 var_2019 = const()[name = string("op_2019"), val = int32(-1)];
            bool rot_11_interleave_0 = const()[name = string("rot_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_11_cast_fp16 = concat(axis = var_2019, interleave = rot_11_interleave_0, values = (var_2012_cast_fp16, var_2017_cast_fp16))[name = string("rot_11_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2021_cast_fp16 = mul(x = z_11_cast_fp16, y = c_cast_fp16)[name = string("op_2021_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2022_cast_fp16 = mul(x = rot_11_cast_fp16, y = s_11_cast_fp16)[name = string("op_2022_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2024_cast_fp16 = add(x = var_2021_cast_fp16, y = var_2022_cast_fp16)[name = string("op_2024_cast_fp16")];
            int32 var_2026 = const()[name = string("op_2026"), val = int32(-1)];
            bool var_2027_interleave_0 = const()[name = string("op_2027_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_2027_cast_fp16 = concat(axis = var_2026, interleave = var_2027_interleave_0, values = (var_2001_cast_fp16, var_2024_cast_fp16))[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> z_13_begin_0 = const()[name = string("z_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_13_end_0 = const()[name = string("z_13_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_13_end_mask_0 = const()[name = string("z_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_13_cast_fp16 = slice_by_index(begin = z_13_begin_0, end = z_13_end_0, end_mask = z_13_end_mask_0, x = x_29_cast_fp16)[name = string("z_13_cast_fp16")];
            tensor<int32, [4]> var_2040_begin_0 = const()[name = string("op_2040_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2040_end_0 = const()[name = string("op_2040_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2040_end_mask_0 = const()[name = string("op_2040_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, x = z_13_cast_fp16)[name = string("op_2040_cast_fp16")];
            fp16 const_6_promoted_to_fp16 = const()[name = string("const_6_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2041_cast_fp16 = mul(x = var_2040_cast_fp16, y = const_6_promoted_to_fp16)[name = string("op_2041_cast_fp16")];
            tensor<int32, [4]> var_2046_begin_0 = const()[name = string("op_2046_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2046_end_0 = const()[name = string("op_2046_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2046_end_mask_0 = const()[name = string("op_2046_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2046_cast_fp16 = slice_by_index(begin = var_2046_begin_0, end = var_2046_end_0, end_mask = var_2046_end_mask_0, x = z_13_cast_fp16)[name = string("op_2046_cast_fp16")];
            int32 var_2048 = const()[name = string("op_2048"), val = int32(-1)];
            bool rot_13_interleave_0 = const()[name = string("rot_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_13_cast_fp16 = concat(axis = var_2048, interleave = rot_13_interleave_0, values = (var_2041_cast_fp16, var_2046_cast_fp16))[name = string("rot_13_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2050_cast_fp16 = mul(x = z_13_cast_fp16, y = c_1_cast_fp16)[name = string("op_2050_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2051_cast_fp16 = mul(x = rot_13_cast_fp16, y = s_9_cast_fp16)[name = string("op_2051_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2053_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2051_cast_fp16)[name = string("op_2053_cast_fp16")];
            tensor<int32, [4]> z_15_begin_0 = const()[name = string("z_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_15_end_0 = const()[name = string("z_15_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_15_end_mask_0 = const()[name = string("z_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_15_cast_fp16 = slice_by_index(begin = z_15_begin_0, end = z_15_end_0, end_mask = z_15_end_mask_0, x = x_29_cast_fp16)[name = string("z_15_cast_fp16")];
            tensor<int32, [4]> var_2063_begin_0 = const()[name = string("op_2063_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2063_end_0 = const()[name = string("op_2063_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2063_end_mask_0 = const()[name = string("op_2063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2063_cast_fp16 = slice_by_index(begin = var_2063_begin_0, end = var_2063_end_0, end_mask = var_2063_end_mask_0, x = z_15_cast_fp16)[name = string("op_2063_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2064_cast_fp16 = mul(x = var_2063_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_2064_cast_fp16")];
            tensor<int32, [4]> var_2069_begin_0 = const()[name = string("op_2069_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2069_end_0 = const()[name = string("op_2069_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2069_end_mask_0 = const()[name = string("op_2069_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2069_cast_fp16 = slice_by_index(begin = var_2069_begin_0, end = var_2069_end_0, end_mask = var_2069_end_mask_0, x = z_15_cast_fp16)[name = string("op_2069_cast_fp16")];
            int32 var_2071 = const()[name = string("op_2071"), val = int32(-1)];
            bool rot_15_interleave_0 = const()[name = string("rot_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_15_cast_fp16 = concat(axis = var_2071, interleave = rot_15_interleave_0, values = (var_2064_cast_fp16, var_2069_cast_fp16))[name = string("rot_15_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2073_cast_fp16 = mul(x = z_15_cast_fp16, y = c_cast_fp16)[name = string("op_2073_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2074_cast_fp16 = mul(x = rot_15_cast_fp16, y = s_11_cast_fp16)[name = string("op_2074_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2076_cast_fp16 = add(x = var_2073_cast_fp16, y = var_2074_cast_fp16)[name = string("op_2076_cast_fp16")];
            int32 var_2078 = const()[name = string("op_2078"), val = int32(-1)];
            bool var_2079_interleave_0 = const()[name = string("op_2079_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_2079_cast_fp16 = concat(axis = var_2078, interleave = var_2079_interleave_0, values = (var_2053_cast_fp16, var_2076_cast_fp16))[name = string("op_2079_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2089_transpose_x_0 = const()[name = string("op_2089_transpose_x_0"), val = bool(false)];
            bool var_2089_transpose_y_0 = const()[name = string("op_2089_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_66_perm_0 = const()[name = string("transpose_66_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_67_perm_0 = const()[name = string("transpose_67_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_67 = transpose(perm = transpose_67_perm_0, x = var_2079_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 12, 630, 64]> transpose_66 = transpose(perm = transpose_66_perm_0, x = var_2027_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [1, 12, 630, 630]> var_2089_cast_fp16 = matmul(transpose_x = var_2089_transpose_x_0, transpose_y = var_2089_transpose_y_0, x = transpose_66, y = transpose_67)[name = string("op_2089_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_2091_cast_fp16 = add(x = var_2089_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2091_cast_fp16")];
            int32 var_2092 = const()[name = string("op_2092"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_21_cast_fp16 = softmax(axis = var_2092, x = var_2091_cast_fp16)[name = string("w_21_cast_fp16")];
            bool var_2095_transpose_x_0 = const()[name = string("op_2095_transpose_x_0"), val = bool(false)];
            bool var_2095_transpose_y_0 = const()[name = string("op_2095_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = v_5_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 12, 630, 64]> var_2095_cast_fp16 = matmul(transpose_x = var_2095_transpose_x_0, transpose_y = var_2095_transpose_y_0, x = w_21_cast_fp16, y = v_7_cast_fp16)[name = string("op_2095_cast_fp16")];
            tensor<int32, [4]> var_2098_perm_0 = const()[name = string("op_2098_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2102 = const()[name = string("op_2102"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_2098_cast_fp16 = transpose(perm = var_2098_perm_0, x = var_2095_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 630, 768]> input_15_cast_fp16 = reshape(shape = var_2102, x = var_2098_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11840384))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12430272))))[name = string("t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_11_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2107_cast_fp16 = abs(x = linear_11_cast_fp16)[name = string("op_2107_cast_fp16")];
            bool var_2110 = const()[name = string("op_2110"), val = bool(true)];
            tensor<int32, [1]> var_2111_axes_0 = const()[name = string("op_2111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2111_cast_fp16 = reduce_max(axes = var_2111_axes_0, keep_dims = var_2110, x = var_2107_cast_fp16)[name = string("op_2111_cast_fp16")];
            fp16 var_2112_to_fp16 = const()[name = string("op_2112_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_27_cast_fp16 = maximum(x = var_2111_cast_fp16, y = var_2112_to_fp16)[name = string("s_27_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_23_cast_fp16 = real_div(x = linear_11_cast_fp16, y = s_27_cast_fp16)[name = string("y_23_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2115_cast_fp16 = mul(x = y_23_cast_fp16, y = y_23_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<int32, [1]> var_2120_axes_0 = const()[name = string("op_2120_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2120_keep_dims_0 = const()[name = string("op_2120_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2120_cast_fp16 = reduce_mean(axes = var_2120_axes_0, keep_dims = var_2120_keep_dims_0, x = var_2115_cast_fp16)[name = string("op_2120_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2121_cast_fp16 = mul(x = s_27_cast_fp16, y = s_27_cast_fp16)[name = string("op_2121_cast_fp16")];
            fp32 var_2122_epsilon_0 = const()[name = string("op_2122_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2122_cast_fp16 = inverse(epsilon = var_2122_epsilon_0, x = var_2121_cast_fp16)[name = string("op_2122_cast_fp16")];
            fp16 var_2123_to_fp16 = const()[name = string("op_2123_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2124_cast_fp16 = mul(x = var_2122_cast_fp16, y = var_2123_to_fp16)[name = string("op_2124_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2126_cast_fp16 = add(x = var_2120_cast_fp16, y = var_2124_cast_fp16)[name = string("op_2126_cast_fp16")];
            fp32 var_2127_epsilon_0 = const()[name = string("op_2127_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2127_cast_fp16 = rsqrt(epsilon = var_2127_epsilon_0, x = var_2126_cast_fp16)[name = string("op_2127_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_19_cast_fp16 = mul(x = y_23_cast_fp16, y = var_2127_cast_fp16)[name = string("out_19_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12431872)))];
            tensor<fp16, [1, 630, 768]> var_2129_cast_fp16 = mul(x = out_19_cast_fp16, y = t_encoder_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("op_2129_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_33_cast_fp16 = add(x = x_19_cast_fp16, y = var_2129_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2132_cast_fp16 = abs(x = x_33_cast_fp16)[name = string("op_2132_cast_fp16")];
            bool var_2135 = const()[name = string("op_2135"), val = bool(true)];
            tensor<int32, [1]> var_2136_axes_0 = const()[name = string("op_2136_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2136_cast_fp16 = reduce_max(axes = var_2136_axes_0, keep_dims = var_2135, x = var_2132_cast_fp16)[name = string("op_2136_cast_fp16")];
            fp16 var_2137_to_fp16 = const()[name = string("op_2137_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_29_cast_fp16 = maximum(x = var_2136_cast_fp16, y = var_2137_to_fp16)[name = string("s_29_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_25_cast_fp16 = real_div(x = x_33_cast_fp16, y = s_29_cast_fp16)[name = string("y_25_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2140_cast_fp16 = mul(x = y_25_cast_fp16, y = y_25_cast_fp16)[name = string("op_2140_cast_fp16")];
            tensor<int32, [1]> var_2145_axes_0 = const()[name = string("op_2145_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2145_keep_dims_0 = const()[name = string("op_2145_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2145_cast_fp16 = reduce_mean(axes = var_2145_axes_0, keep_dims = var_2145_keep_dims_0, x = var_2140_cast_fp16)[name = string("op_2145_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2146_cast_fp16 = mul(x = s_29_cast_fp16, y = s_29_cast_fp16)[name = string("op_2146_cast_fp16")];
            fp32 var_2147_epsilon_0 = const()[name = string("op_2147_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2147_cast_fp16 = inverse(epsilon = var_2147_epsilon_0, x = var_2146_cast_fp16)[name = string("op_2147_cast_fp16")];
            fp16 var_2148_to_fp16 = const()[name = string("op_2148_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2149_cast_fp16 = mul(x = var_2147_cast_fp16, y = var_2148_to_fp16)[name = string("op_2149_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2151_cast_fp16 = add(x = var_2145_cast_fp16, y = var_2149_cast_fp16)[name = string("op_2151_cast_fp16")];
            fp32 var_2152_epsilon_0 = const()[name = string("op_2152_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2152_cast_fp16 = rsqrt(epsilon = var_2152_epsilon_0, x = var_2151_cast_fp16)[name = string("op_2152_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_21_cast_fp16 = mul(x = y_25_cast_fp16, y = var_2152_cast_fp16)[name = string("out_21_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12433472)))];
            tensor<fp16, [1, 630, 768]> input_17_cast_fp16 = mul(x = out_21_cast_fp16, y = t_encoder_layers_1_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_17_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12435072))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14794432))))[name = string("t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_12_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_12_cast_fp16")];
            string var_2159_mode_0 = const()[name = string("op_2159_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_2159_cast_fp16 = gelu(mode = var_2159_mode_0, x = linear_12_cast_fp16)[name = string("op_2159_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(14800640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17160000))))[name = string("t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_13_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_17_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_21_cast_fp16 = mul(x = var_2159_cast_fp16, y = linear_13_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(17166208))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19525568))))[name = string("t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_1_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2167_cast_fp16 = abs(x = linear_14_cast_fp16)[name = string("op_2167_cast_fp16")];
            bool var_2170 = const()[name = string("op_2170"), val = bool(true)];
            tensor<int32, [1]> var_2171_axes_0 = const()[name = string("op_2171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2171_cast_fp16 = reduce_max(axes = var_2171_axes_0, keep_dims = var_2170, x = var_2167_cast_fp16)[name = string("op_2171_cast_fp16")];
            fp16 var_2172_to_fp16 = const()[name = string("op_2172_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_31_cast_fp16 = maximum(x = var_2171_cast_fp16, y = var_2172_to_fp16)[name = string("s_31_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_27_cast_fp16 = real_div(x = linear_14_cast_fp16, y = s_31_cast_fp16)[name = string("y_27_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2175_cast_fp16 = mul(x = y_27_cast_fp16, y = y_27_cast_fp16)[name = string("op_2175_cast_fp16")];
            tensor<int32, [1]> var_2180_axes_0 = const()[name = string("op_2180_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2180_keep_dims_0 = const()[name = string("op_2180_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2180_cast_fp16 = reduce_mean(axes = var_2180_axes_0, keep_dims = var_2180_keep_dims_0, x = var_2175_cast_fp16)[name = string("op_2180_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2181_cast_fp16 = mul(x = s_31_cast_fp16, y = s_31_cast_fp16)[name = string("op_2181_cast_fp16")];
            fp32 var_2182_epsilon_0 = const()[name = string("op_2182_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2182_cast_fp16 = inverse(epsilon = var_2182_epsilon_0, x = var_2181_cast_fp16)[name = string("op_2182_cast_fp16")];
            fp16 var_2183_to_fp16 = const()[name = string("op_2183_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2184_cast_fp16 = mul(x = var_2182_cast_fp16, y = var_2183_to_fp16)[name = string("op_2184_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2186_cast_fp16 = add(x = var_2180_cast_fp16, y = var_2184_cast_fp16)[name = string("op_2186_cast_fp16")];
            fp32 var_2187_epsilon_0 = const()[name = string("op_2187_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2187_cast_fp16 = rsqrt(epsilon = var_2187_epsilon_0, x = var_2186_cast_fp16)[name = string("op_2187_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_23_cast_fp16 = mul(x = y_27_cast_fp16, y = var_2187_cast_fp16)[name = string("out_23_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19527168)))];
            tensor<fp16, [1, 630, 768]> var_2189_cast_fp16 = mul(x = out_23_cast_fp16, y = t_encoder_layers_1_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2189_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_37_cast_fp16 = add(x = x_33_cast_fp16, y = var_2189_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2192_cast_fp16 = abs(x = x_37_cast_fp16)[name = string("op_2192_cast_fp16")];
            bool var_2195 = const()[name = string("op_2195"), val = bool(true)];
            tensor<int32, [1]> var_2196_axes_0 = const()[name = string("op_2196_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2196_cast_fp16 = reduce_max(axes = var_2196_axes_0, keep_dims = var_2195, x = var_2192_cast_fp16)[name = string("op_2196_cast_fp16")];
            fp16 var_2197_to_fp16 = const()[name = string("op_2197_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_33_cast_fp16 = maximum(x = var_2196_cast_fp16, y = var_2197_to_fp16)[name = string("s_33_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_29_cast_fp16 = real_div(x = x_37_cast_fp16, y = s_33_cast_fp16)[name = string("y_29_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2200_cast_fp16 = mul(x = y_29_cast_fp16, y = y_29_cast_fp16)[name = string("op_2200_cast_fp16")];
            tensor<int32, [1]> var_2205_axes_0 = const()[name = string("op_2205_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2205_keep_dims_0 = const()[name = string("op_2205_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2205_cast_fp16 = reduce_mean(axes = var_2205_axes_0, keep_dims = var_2205_keep_dims_0, x = var_2200_cast_fp16)[name = string("op_2205_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2206_cast_fp16 = mul(x = s_33_cast_fp16, y = s_33_cast_fp16)[name = string("op_2206_cast_fp16")];
            fp32 var_2207_epsilon_0 = const()[name = string("op_2207_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2207_cast_fp16 = inverse(epsilon = var_2207_epsilon_0, x = var_2206_cast_fp16)[name = string("op_2207_cast_fp16")];
            fp16 var_2208_to_fp16 = const()[name = string("op_2208_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2209_cast_fp16 = mul(x = var_2207_cast_fp16, y = var_2208_to_fp16)[name = string("op_2209_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2211_cast_fp16 = add(x = var_2205_cast_fp16, y = var_2209_cast_fp16)[name = string("op_2211_cast_fp16")];
            fp32 var_2212_epsilon_0 = const()[name = string("op_2212_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2212_cast_fp16 = rsqrt(epsilon = var_2212_epsilon_0, x = var_2211_cast_fp16)[name = string("op_2212_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_25_cast_fp16 = mul(x = y_29_cast_fp16, y = var_2212_cast_fp16)[name = string("out_25_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19528768)))];
            tensor<fp16, [1, 630, 768]> input_23_cast_fp16 = mul(x = out_25_cast_fp16, y = t_encoder_layers_2_input_layernorm_weight_to_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(19530368))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20120256))))[name = string("t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_15_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_2222 = const()[name = string("op_2222"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_39_cast_fp16 = reshape(shape = var_2222, x = linear_15_cast_fp16)[name = string("x_39_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2224_cast_fp16 = abs(x = x_39_cast_fp16)[name = string("op_2224_cast_fp16")];
            bool var_2227 = const()[name = string("op_2227"), val = bool(true)];
            tensor<int32, [1]> var_2228_axes_0 = const()[name = string("op_2228_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2228_cast_fp16 = reduce_max(axes = var_2228_axes_0, keep_dims = var_2227, x = var_2224_cast_fp16)[name = string("op_2228_cast_fp16")];
            fp16 var_2229_to_fp16 = const()[name = string("op_2229_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_35_cast_fp16 = maximum(x = var_2228_cast_fp16, y = var_2229_to_fp16)[name = string("s_35_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_31_cast_fp16 = real_div(x = x_39_cast_fp16, y = s_35_cast_fp16)[name = string("y_31_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2232_cast_fp16 = mul(x = y_31_cast_fp16, y = y_31_cast_fp16)[name = string("op_2232_cast_fp16")];
            tensor<int32, [1]> var_2237_axes_0 = const()[name = string("op_2237_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2237_keep_dims_0 = const()[name = string("op_2237_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2237_cast_fp16 = reduce_mean(axes = var_2237_axes_0, keep_dims = var_2237_keep_dims_0, x = var_2232_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2238_cast_fp16 = mul(x = s_35_cast_fp16, y = s_35_cast_fp16)[name = string("op_2238_cast_fp16")];
            fp32 var_2239_epsilon_0 = const()[name = string("op_2239_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2239_cast_fp16 = inverse(epsilon = var_2239_epsilon_0, x = var_2238_cast_fp16)[name = string("op_2239_cast_fp16")];
            fp16 var_2240_to_fp16 = const()[name = string("op_2240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2241_cast_fp16 = mul(x = var_2239_cast_fp16, y = var_2240_to_fp16)[name = string("op_2241_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2243_cast_fp16 = add(x = var_2237_cast_fp16, y = var_2241_cast_fp16)[name = string("op_2243_cast_fp16")];
            fp32 var_2244_epsilon_0 = const()[name = string("op_2244_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2244_cast_fp16 = rsqrt(epsilon = var_2244_epsilon_0, x = var_2243_cast_fp16)[name = string("op_2244_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_27_cast_fp16 = mul(x = y_31_cast_fp16, y = var_2244_cast_fp16)[name = string("out_27_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20121856)))];
            tensor<fp16, [1, 630, 12, 64]> x_45_cast_fp16 = mul(x = out_27_cast_fp16, y = t_encoder_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20122048))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20711936))))[name = string("t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_2254 = const()[name = string("op_2254"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_41_cast_fp16 = reshape(shape = var_2254, x = linear_16_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2256_cast_fp16 = abs(x = x_41_cast_fp16)[name = string("op_2256_cast_fp16")];
            bool var_2259 = const()[name = string("op_2259"), val = bool(true)];
            tensor<int32, [1]> var_2260_axes_0 = const()[name = string("op_2260_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2260_cast_fp16 = reduce_max(axes = var_2260_axes_0, keep_dims = var_2259, x = var_2256_cast_fp16)[name = string("op_2260_cast_fp16")];
            fp16 var_2261_to_fp16 = const()[name = string("op_2261_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_37_cast_fp16 = maximum(x = var_2260_cast_fp16, y = var_2261_to_fp16)[name = string("s_37_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_33_cast_fp16 = real_div(x = x_41_cast_fp16, y = s_37_cast_fp16)[name = string("y_33_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2264_cast_fp16 = mul(x = y_33_cast_fp16, y = y_33_cast_fp16)[name = string("op_2264_cast_fp16")];
            tensor<int32, [1]> var_2269_axes_0 = const()[name = string("op_2269_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2269_keep_dims_0 = const()[name = string("op_2269_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2269_cast_fp16 = reduce_mean(axes = var_2269_axes_0, keep_dims = var_2269_keep_dims_0, x = var_2264_cast_fp16)[name = string("op_2269_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2270_cast_fp16 = mul(x = s_37_cast_fp16, y = s_37_cast_fp16)[name = string("op_2270_cast_fp16")];
            fp32 var_2271_epsilon_0 = const()[name = string("op_2271_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2271_cast_fp16 = inverse(epsilon = var_2271_epsilon_0, x = var_2270_cast_fp16)[name = string("op_2271_cast_fp16")];
            fp16 var_2272_to_fp16 = const()[name = string("op_2272_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2273_cast_fp16 = mul(x = var_2271_cast_fp16, y = var_2272_to_fp16)[name = string("op_2273_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2275_cast_fp16 = add(x = var_2269_cast_fp16, y = var_2273_cast_fp16)[name = string("op_2275_cast_fp16")];
            fp32 var_2276_epsilon_0 = const()[name = string("op_2276_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2276_cast_fp16 = rsqrt(epsilon = var_2276_epsilon_0, x = var_2275_cast_fp16)[name = string("op_2276_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_29_cast_fp16 = mul(x = y_33_cast_fp16, y = var_2276_cast_fp16)[name = string("out_29_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713536)))];
            tensor<fp16, [1, 630, 12, 64]> x_47_cast_fp16 = mul(x = out_29_cast_fp16, y = t_encoder_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_47_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713728))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21303616))))[name = string("t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_17_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<int32, [4]> var_2286 = const()[name = string("op_2286"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_43_cast_fp16 = reshape(shape = var_2286, x = linear_17_cast_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2288_cast_fp16 = abs(x = x_43_cast_fp16)[name = string("op_2288_cast_fp16")];
            bool var_2291 = const()[name = string("op_2291"), val = bool(true)];
            tensor<int32, [1]> var_2292_axes_0 = const()[name = string("op_2292_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2292_cast_fp16 = reduce_max(axes = var_2292_axes_0, keep_dims = var_2291, x = var_2288_cast_fp16)[name = string("op_2292_cast_fp16")];
            fp16 var_2293_to_fp16 = const()[name = string("op_2293_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_39_cast_fp16 = maximum(x = var_2292_cast_fp16, y = var_2293_to_fp16)[name = string("s_39_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_35_cast_fp16 = real_div(x = x_43_cast_fp16, y = s_39_cast_fp16)[name = string("y_35_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2296_cast_fp16 = mul(x = y_35_cast_fp16, y = y_35_cast_fp16)[name = string("op_2296_cast_fp16")];
            tensor<int32, [1]> var_2301_axes_0 = const()[name = string("op_2301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2301_keep_dims_0 = const()[name = string("op_2301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2301_cast_fp16 = reduce_mean(axes = var_2301_axes_0, keep_dims = var_2301_keep_dims_0, x = var_2296_cast_fp16)[name = string("op_2301_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2302_cast_fp16 = mul(x = s_39_cast_fp16, y = s_39_cast_fp16)[name = string("op_2302_cast_fp16")];
            fp32 var_2303_epsilon_0 = const()[name = string("op_2303_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2303_cast_fp16 = inverse(epsilon = var_2303_epsilon_0, x = var_2302_cast_fp16)[name = string("op_2303_cast_fp16")];
            fp16 var_2304_to_fp16 = const()[name = string("op_2304_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2305_cast_fp16 = mul(x = var_2303_cast_fp16, y = var_2304_to_fp16)[name = string("op_2305_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2307_cast_fp16 = add(x = var_2301_cast_fp16, y = var_2305_cast_fp16)[name = string("op_2307_cast_fp16")];
            fp32 var_2308_epsilon_0 = const()[name = string("op_2308_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2308_cast_fp16 = rsqrt(epsilon = var_2308_epsilon_0, x = var_2307_cast_fp16)[name = string("op_2308_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_9_cast_fp16 = mul(x = y_35_cast_fp16, y = var_2308_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<int32, [4]> z_17_begin_0 = const()[name = string("z_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_17_end_0 = const()[name = string("z_17_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_17_end_mask_0 = const()[name = string("z_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_17_cast_fp16 = slice_by_index(begin = z_17_begin_0, end = z_17_end_0, end_mask = z_17_end_mask_0, x = x_45_cast_fp16)[name = string("z_17_cast_fp16")];
            tensor<int32, [4]> var_2319_begin_0 = const()[name = string("op_2319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2319_end_0 = const()[name = string("op_2319_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2319_end_mask_0 = const()[name = string("op_2319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2319_cast_fp16 = slice_by_index(begin = var_2319_begin_0, end = var_2319_end_0, end_mask = var_2319_end_mask_0, x = z_17_cast_fp16)[name = string("op_2319_cast_fp16")];
            fp16 const_8_promoted_to_fp16 = const()[name = string("const_8_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2320_cast_fp16 = mul(x = var_2319_cast_fp16, y = const_8_promoted_to_fp16)[name = string("op_2320_cast_fp16")];
            tensor<int32, [4]> var_2325_begin_0 = const()[name = string("op_2325_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2325_end_0 = const()[name = string("op_2325_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2325_end_mask_0 = const()[name = string("op_2325_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2325_cast_fp16 = slice_by_index(begin = var_2325_begin_0, end = var_2325_end_0, end_mask = var_2325_end_mask_0, x = z_17_cast_fp16)[name = string("op_2325_cast_fp16")];
            int32 var_2327 = const()[name = string("op_2327"), val = int32(-1)];
            bool rot_17_interleave_0 = const()[name = string("rot_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_17_cast_fp16 = concat(axis = var_2327, interleave = rot_17_interleave_0, values = (var_2320_cast_fp16, var_2325_cast_fp16))[name = string("rot_17_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2329_cast_fp16 = mul(x = z_17_cast_fp16, y = c_1_cast_fp16)[name = string("op_2329_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2330_cast_fp16 = mul(x = rot_17_cast_fp16, y = s_9_cast_fp16)[name = string("op_2330_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2332_cast_fp16 = add(x = var_2329_cast_fp16, y = var_2330_cast_fp16)[name = string("op_2332_cast_fp16")];
            tensor<int32, [4]> z_19_begin_0 = const()[name = string("z_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_19_end_0 = const()[name = string("z_19_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_19_end_mask_0 = const()[name = string("z_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_19_cast_fp16 = slice_by_index(begin = z_19_begin_0, end = z_19_end_0, end_mask = z_19_end_mask_0, x = x_45_cast_fp16)[name = string("z_19_cast_fp16")];
            tensor<int32, [4]> var_2342_begin_0 = const()[name = string("op_2342_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2342_end_0 = const()[name = string("op_2342_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2342_end_mask_0 = const()[name = string("op_2342_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2342_cast_fp16 = slice_by_index(begin = var_2342_begin_0, end = var_2342_end_0, end_mask = var_2342_end_mask_0, x = z_19_cast_fp16)[name = string("op_2342_cast_fp16")];
            fp16 const_9_promoted_to_fp16 = const()[name = string("const_9_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2343_cast_fp16 = mul(x = var_2342_cast_fp16, y = const_9_promoted_to_fp16)[name = string("op_2343_cast_fp16")];
            tensor<int32, [4]> var_2348_begin_0 = const()[name = string("op_2348_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2348_end_0 = const()[name = string("op_2348_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2348_end_mask_0 = const()[name = string("op_2348_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2348_cast_fp16 = slice_by_index(begin = var_2348_begin_0, end = var_2348_end_0, end_mask = var_2348_end_mask_0, x = z_19_cast_fp16)[name = string("op_2348_cast_fp16")];
            int32 var_2350 = const()[name = string("op_2350"), val = int32(-1)];
            bool rot_19_interleave_0 = const()[name = string("rot_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_19_cast_fp16 = concat(axis = var_2350, interleave = rot_19_interleave_0, values = (var_2343_cast_fp16, var_2348_cast_fp16))[name = string("rot_19_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2352_cast_fp16 = mul(x = z_19_cast_fp16, y = c_cast_fp16)[name = string("op_2352_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2353_cast_fp16 = mul(x = rot_19_cast_fp16, y = s_11_cast_fp16)[name = string("op_2353_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2355_cast_fp16 = add(x = var_2352_cast_fp16, y = var_2353_cast_fp16)[name = string("op_2355_cast_fp16")];
            int32 var_2357 = const()[name = string("op_2357"), val = int32(-1)];
            bool var_2358_interleave_0 = const()[name = string("op_2358_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_2358_cast_fp16 = concat(axis = var_2357, interleave = var_2358_interleave_0, values = (var_2332_cast_fp16, var_2355_cast_fp16))[name = string("op_2358_cast_fp16")];
            tensor<int32, [4]> z_21_begin_0 = const()[name = string("z_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_21_end_0 = const()[name = string("z_21_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_21_end_mask_0 = const()[name = string("z_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_21_cast_fp16 = slice_by_index(begin = z_21_begin_0, end = z_21_end_0, end_mask = z_21_end_mask_0, x = x_47_cast_fp16)[name = string("z_21_cast_fp16")];
            tensor<int32, [4]> var_2371_begin_0 = const()[name = string("op_2371_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2371_end_0 = const()[name = string("op_2371_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2371_end_mask_0 = const()[name = string("op_2371_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2371_cast_fp16 = slice_by_index(begin = var_2371_begin_0, end = var_2371_end_0, end_mask = var_2371_end_mask_0, x = z_21_cast_fp16)[name = string("op_2371_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2372_cast_fp16 = mul(x = var_2371_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_2372_cast_fp16")];
            tensor<int32, [4]> var_2377_begin_0 = const()[name = string("op_2377_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2377_end_0 = const()[name = string("op_2377_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2377_end_mask_0 = const()[name = string("op_2377_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2377_cast_fp16 = slice_by_index(begin = var_2377_begin_0, end = var_2377_end_0, end_mask = var_2377_end_mask_0, x = z_21_cast_fp16)[name = string("op_2377_cast_fp16")];
            int32 var_2379 = const()[name = string("op_2379"), val = int32(-1)];
            bool rot_21_interleave_0 = const()[name = string("rot_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_21_cast_fp16 = concat(axis = var_2379, interleave = rot_21_interleave_0, values = (var_2372_cast_fp16, var_2377_cast_fp16))[name = string("rot_21_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2381_cast_fp16 = mul(x = z_21_cast_fp16, y = c_1_cast_fp16)[name = string("op_2381_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2382_cast_fp16 = mul(x = rot_21_cast_fp16, y = s_9_cast_fp16)[name = string("op_2382_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2384_cast_fp16 = add(x = var_2381_cast_fp16, y = var_2382_cast_fp16)[name = string("op_2384_cast_fp16")];
            tensor<int32, [4]> z_23_begin_0 = const()[name = string("z_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_23_end_0 = const()[name = string("z_23_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_23_end_mask_0 = const()[name = string("z_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_23_cast_fp16 = slice_by_index(begin = z_23_begin_0, end = z_23_end_0, end_mask = z_23_end_mask_0, x = x_47_cast_fp16)[name = string("z_23_cast_fp16")];
            tensor<int32, [4]> var_2394_begin_0 = const()[name = string("op_2394_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2394_end_0 = const()[name = string("op_2394_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2394_end_mask_0 = const()[name = string("op_2394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> 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 = z_23_cast_fp16)[name = string("op_2394_cast_fp16")];
            fp16 const_11_promoted_to_fp16 = const()[name = string("const_11_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2395_cast_fp16 = mul(x = var_2394_cast_fp16, y = const_11_promoted_to_fp16)[name = string("op_2395_cast_fp16")];
            tensor<int32, [4]> var_2400_begin_0 = const()[name = string("op_2400_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2400_end_0 = const()[name = string("op_2400_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2400_end_mask_0 = const()[name = string("op_2400_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2400_cast_fp16 = slice_by_index(begin = var_2400_begin_0, end = var_2400_end_0, end_mask = var_2400_end_mask_0, x = z_23_cast_fp16)[name = string("op_2400_cast_fp16")];
            int32 var_2402 = const()[name = string("op_2402"), val = int32(-1)];
            bool rot_23_interleave_0 = const()[name = string("rot_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_23_cast_fp16 = concat(axis = var_2402, interleave = rot_23_interleave_0, values = (var_2395_cast_fp16, var_2400_cast_fp16))[name = string("rot_23_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2404_cast_fp16 = mul(x = z_23_cast_fp16, y = c_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2405_cast_fp16 = mul(x = rot_23_cast_fp16, y = s_11_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2407_cast_fp16 = add(x = var_2404_cast_fp16, y = var_2405_cast_fp16)[name = string("op_2407_cast_fp16")];
            int32 var_2409 = const()[name = string("op_2409"), val = int32(-1)];
            bool var_2410_interleave_0 = const()[name = string("op_2410_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_2410_cast_fp16 = concat(axis = var_2409, interleave = var_2410_interleave_0, values = (var_2384_cast_fp16, var_2407_cast_fp16))[name = string("op_2410_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2420_transpose_x_0 = const()[name = string("op_2420_transpose_x_0"), val = bool(false)];
            bool var_2420_transpose_y_0 = const()[name = string("op_2420_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_68_perm_0 = const()[name = string("transpose_68_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_69_perm_0 = const()[name = string("transpose_69_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_69 = transpose(perm = transpose_69_perm_0, x = var_2410_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 12, 630, 64]> transpose_68 = transpose(perm = transpose_68_perm_0, x = var_2358_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 12, 630, 630]> var_2420_cast_fp16 = matmul(transpose_x = var_2420_transpose_x_0, transpose_y = var_2420_transpose_y_0, x = transpose_68, y = transpose_69)[name = string("op_2420_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_2422_cast_fp16 = add(x = var_2420_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2422_cast_fp16")];
            int32 var_2423 = const()[name = string("op_2423"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_35_cast_fp16 = softmax(axis = var_2423, x = var_2422_cast_fp16)[name = string("w_35_cast_fp16")];
            bool var_2426_transpose_x_0 = const()[name = string("op_2426_transpose_x_0"), val = bool(false)];
            bool var_2426_transpose_y_0 = const()[name = string("op_2426_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = v_9_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 12, 630, 64]> var_2426_cast_fp16 = matmul(transpose_x = var_2426_transpose_x_0, transpose_y = var_2426_transpose_y_0, x = w_35_cast_fp16, y = v_11_cast_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [4]> var_2429_perm_0 = const()[name = string("op_2429_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2433 = const()[name = string("op_2433"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_2429_cast_fp16 = transpose(perm = var_2429_perm_0, x = var_2426_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 630, 768]> input_25_cast_fp16 = reshape(shape = var_2433, x = var_2429_cast_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21305216))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21895104))))[name = string("t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2438_cast_fp16 = abs(x = linear_18_cast_fp16)[name = string("op_2438_cast_fp16")];
            bool var_2441 = const()[name = string("op_2441"), val = bool(true)];
            tensor<int32, [1]> var_2442_axes_0 = const()[name = string("op_2442_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2442_cast_fp16 = reduce_max(axes = var_2442_axes_0, keep_dims = var_2441, x = var_2438_cast_fp16)[name = string("op_2442_cast_fp16")];
            fp16 var_2443_to_fp16 = const()[name = string("op_2443_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_41_cast_fp16 = maximum(x = var_2442_cast_fp16, y = var_2443_to_fp16)[name = string("s_41_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_37_cast_fp16 = real_div(x = linear_18_cast_fp16, y = s_41_cast_fp16)[name = string("y_37_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2446_cast_fp16 = mul(x = y_37_cast_fp16, y = y_37_cast_fp16)[name = string("op_2446_cast_fp16")];
            tensor<int32, [1]> var_2451_axes_0 = const()[name = string("op_2451_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2451_keep_dims_0 = const()[name = string("op_2451_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2451_cast_fp16 = reduce_mean(axes = var_2451_axes_0, keep_dims = var_2451_keep_dims_0, x = var_2446_cast_fp16)[name = string("op_2451_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2452_cast_fp16 = mul(x = s_41_cast_fp16, y = s_41_cast_fp16)[name = string("op_2452_cast_fp16")];
            fp32 var_2453_epsilon_0 = const()[name = string("op_2453_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2453_cast_fp16 = inverse(epsilon = var_2453_epsilon_0, x = var_2452_cast_fp16)[name = string("op_2453_cast_fp16")];
            fp16 var_2454_to_fp16 = const()[name = string("op_2454_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = string("op_2455_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2457_cast_fp16 = add(x = var_2451_cast_fp16, y = var_2455_cast_fp16)[name = string("op_2457_cast_fp16")];
            fp32 var_2458_epsilon_0 = const()[name = string("op_2458_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2458_cast_fp16 = rsqrt(epsilon = var_2458_epsilon_0, x = var_2457_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_31_cast_fp16 = mul(x = y_37_cast_fp16, y = var_2458_cast_fp16)[name = string("out_31_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21896704)))];
            tensor<fp16, [1, 630, 768]> var_2460_cast_fp16 = mul(x = out_31_cast_fp16, y = t_encoder_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("op_2460_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_51_cast_fp16 = add(x = x_37_cast_fp16, y = var_2460_cast_fp16)[name = string("x_51_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2463_cast_fp16 = abs(x = x_51_cast_fp16)[name = string("op_2463_cast_fp16")];
            bool var_2466 = const()[name = string("op_2466"), val = bool(true)];
            tensor<int32, [1]> var_2467_axes_0 = const()[name = string("op_2467_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2467_cast_fp16 = reduce_max(axes = var_2467_axes_0, keep_dims = var_2466, x = var_2463_cast_fp16)[name = string("op_2467_cast_fp16")];
            fp16 var_2468_to_fp16 = const()[name = string("op_2468_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_43_cast_fp16 = maximum(x = var_2467_cast_fp16, y = var_2468_to_fp16)[name = string("s_43_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_39_cast_fp16 = real_div(x = x_51_cast_fp16, y = s_43_cast_fp16)[name = string("y_39_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2471_cast_fp16 = mul(x = y_39_cast_fp16, y = y_39_cast_fp16)[name = string("op_2471_cast_fp16")];
            tensor<int32, [1]> var_2476_axes_0 = const()[name = string("op_2476_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2476_keep_dims_0 = const()[name = string("op_2476_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2476_cast_fp16 = reduce_mean(axes = var_2476_axes_0, keep_dims = var_2476_keep_dims_0, x = var_2471_cast_fp16)[name = string("op_2476_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2477_cast_fp16 = mul(x = s_43_cast_fp16, y = s_43_cast_fp16)[name = string("op_2477_cast_fp16")];
            fp32 var_2478_epsilon_0 = const()[name = string("op_2478_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2478_cast_fp16 = inverse(epsilon = var_2478_epsilon_0, x = var_2477_cast_fp16)[name = string("op_2478_cast_fp16")];
            fp16 var_2479_to_fp16 = const()[name = string("op_2479_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2480_cast_fp16 = mul(x = var_2478_cast_fp16, y = var_2479_to_fp16)[name = string("op_2480_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2482_cast_fp16 = add(x = var_2476_cast_fp16, y = var_2480_cast_fp16)[name = string("op_2482_cast_fp16")];
            fp32 var_2483_epsilon_0 = const()[name = string("op_2483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2483_cast_fp16 = rsqrt(epsilon = var_2483_epsilon_0, x = var_2482_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_33_cast_fp16 = mul(x = y_39_cast_fp16, y = var_2483_cast_fp16)[name = string("out_33_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21898304)))];
            tensor<fp16, [1, 630, 768]> input_27_cast_fp16 = mul(x = out_33_cast_fp16, y = t_encoder_layers_2_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(21899904))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24259264))))[name = string("t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_19_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_19_cast_fp16")];
            string var_2490_mode_0 = const()[name = string("op_2490_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_2490_cast_fp16 = gelu(mode = var_2490_mode_0, x = linear_19_cast_fp16)[name = string("op_2490_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(24265472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26624832))))[name = string("t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_20_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_27_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_31_cast_fp16 = mul(x = var_2490_cast_fp16, y = linear_20_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(26631040))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28990400))))[name = string("t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_2_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2498_cast_fp16 = abs(x = linear_21_cast_fp16)[name = string("op_2498_cast_fp16")];
            bool var_2501 = const()[name = string("op_2501"), val = bool(true)];
            tensor<int32, [1]> var_2502_axes_0 = const()[name = string("op_2502_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2502_cast_fp16 = reduce_max(axes = var_2502_axes_0, keep_dims = var_2501, x = var_2498_cast_fp16)[name = string("op_2502_cast_fp16")];
            fp16 var_2503_to_fp16 = const()[name = string("op_2503_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_45_cast_fp16 = maximum(x = var_2502_cast_fp16, y = var_2503_to_fp16)[name = string("s_45_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_41_cast_fp16 = real_div(x = linear_21_cast_fp16, y = s_45_cast_fp16)[name = string("y_41_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2506_cast_fp16 = mul(x = y_41_cast_fp16, y = y_41_cast_fp16)[name = string("op_2506_cast_fp16")];
            tensor<int32, [1]> var_2511_axes_0 = const()[name = string("op_2511_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2511_keep_dims_0 = const()[name = string("op_2511_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2511_cast_fp16 = reduce_mean(axes = var_2511_axes_0, keep_dims = var_2511_keep_dims_0, x = var_2506_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2512_cast_fp16 = mul(x = s_45_cast_fp16, y = s_45_cast_fp16)[name = string("op_2512_cast_fp16")];
            fp32 var_2513_epsilon_0 = const()[name = string("op_2513_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2513_cast_fp16 = inverse(epsilon = var_2513_epsilon_0, x = var_2512_cast_fp16)[name = string("op_2513_cast_fp16")];
            fp16 var_2514_to_fp16 = const()[name = string("op_2514_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2515_cast_fp16 = mul(x = var_2513_cast_fp16, y = var_2514_to_fp16)[name = string("op_2515_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2517_cast_fp16 = add(x = var_2511_cast_fp16, y = var_2515_cast_fp16)[name = string("op_2517_cast_fp16")];
            fp32 var_2518_epsilon_0 = const()[name = string("op_2518_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2518_cast_fp16 = rsqrt(epsilon = var_2518_epsilon_0, x = var_2517_cast_fp16)[name = string("op_2518_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_35_cast_fp16 = mul(x = y_41_cast_fp16, y = var_2518_cast_fp16)[name = string("out_35_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28992000)))];
            tensor<fp16, [1, 630, 768]> var_2520_cast_fp16 = mul(x = out_35_cast_fp16, y = t_encoder_layers_2_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2520_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_55_cast_fp16 = add(x = x_51_cast_fp16, y = var_2520_cast_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2523_cast_fp16 = abs(x = x_55_cast_fp16)[name = string("op_2523_cast_fp16")];
            bool var_2526 = const()[name = string("op_2526"), val = bool(true)];
            tensor<int32, [1]> var_2527_axes_0 = const()[name = string("op_2527_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2527_cast_fp16 = reduce_max(axes = var_2527_axes_0, keep_dims = var_2526, x = var_2523_cast_fp16)[name = string("op_2527_cast_fp16")];
            fp16 var_2528_to_fp16 = const()[name = string("op_2528_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_47_cast_fp16 = maximum(x = var_2527_cast_fp16, y = var_2528_to_fp16)[name = string("s_47_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_43_cast_fp16 = real_div(x = x_55_cast_fp16, y = s_47_cast_fp16)[name = string("y_43_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2531_cast_fp16 = mul(x = y_43_cast_fp16, y = y_43_cast_fp16)[name = string("op_2531_cast_fp16")];
            tensor<int32, [1]> var_2536_axes_0 = const()[name = string("op_2536_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2536_keep_dims_0 = const()[name = string("op_2536_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2536_cast_fp16 = reduce_mean(axes = var_2536_axes_0, keep_dims = var_2536_keep_dims_0, x = var_2531_cast_fp16)[name = string("op_2536_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2537_cast_fp16 = mul(x = s_47_cast_fp16, y = s_47_cast_fp16)[name = string("op_2537_cast_fp16")];
            fp32 var_2538_epsilon_0 = const()[name = string("op_2538_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2538_cast_fp16 = inverse(epsilon = var_2538_epsilon_0, x = var_2537_cast_fp16)[name = string("op_2538_cast_fp16")];
            fp16 var_2539_to_fp16 = const()[name = string("op_2539_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2540_cast_fp16 = mul(x = var_2538_cast_fp16, y = var_2539_to_fp16)[name = string("op_2540_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2542_cast_fp16 = add(x = var_2536_cast_fp16, y = var_2540_cast_fp16)[name = string("op_2542_cast_fp16")];
            fp32 var_2543_epsilon_0 = const()[name = string("op_2543_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2543_cast_fp16 = rsqrt(epsilon = var_2543_epsilon_0, x = var_2542_cast_fp16)[name = string("op_2543_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_37_cast_fp16 = mul(x = y_43_cast_fp16, y = var_2543_cast_fp16)[name = string("out_37_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28993600)))];
            tensor<fp16, [1, 630, 768]> input_33_cast_fp16 = mul(x = out_37_cast_fp16, y = t_encoder_layers_3_input_layernorm_weight_to_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(28995200))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29585088))))[name = string("t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_2553 = const()[name = string("op_2553"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_57_cast_fp16 = reshape(shape = var_2553, x = linear_22_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2555_cast_fp16 = abs(x = x_57_cast_fp16)[name = string("op_2555_cast_fp16")];
            bool var_2558 = const()[name = string("op_2558"), val = bool(true)];
            tensor<int32, [1]> var_2559_axes_0 = const()[name = string("op_2559_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2559_cast_fp16 = reduce_max(axes = var_2559_axes_0, keep_dims = var_2558, x = var_2555_cast_fp16)[name = string("op_2559_cast_fp16")];
            fp16 var_2560_to_fp16 = const()[name = string("op_2560_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_49_cast_fp16 = maximum(x = var_2559_cast_fp16, y = var_2560_to_fp16)[name = string("s_49_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_45_cast_fp16 = real_div(x = x_57_cast_fp16, y = s_49_cast_fp16)[name = string("y_45_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2563_cast_fp16 = mul(x = y_45_cast_fp16, y = y_45_cast_fp16)[name = string("op_2563_cast_fp16")];
            tensor<int32, [1]> var_2568_axes_0 = const()[name = string("op_2568_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2568_keep_dims_0 = const()[name = string("op_2568_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2568_cast_fp16 = reduce_mean(axes = var_2568_axes_0, keep_dims = var_2568_keep_dims_0, x = var_2563_cast_fp16)[name = string("op_2568_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2569_cast_fp16 = mul(x = s_49_cast_fp16, y = s_49_cast_fp16)[name = string("op_2569_cast_fp16")];
            fp32 var_2570_epsilon_0 = const()[name = string("op_2570_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2570_cast_fp16 = inverse(epsilon = var_2570_epsilon_0, x = var_2569_cast_fp16)[name = string("op_2570_cast_fp16")];
            fp16 var_2571_to_fp16 = const()[name = string("op_2571_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2572_cast_fp16 = mul(x = var_2570_cast_fp16, y = var_2571_to_fp16)[name = string("op_2572_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2574_cast_fp16 = add(x = var_2568_cast_fp16, y = var_2572_cast_fp16)[name = string("op_2574_cast_fp16")];
            fp32 var_2575_epsilon_0 = const()[name = string("op_2575_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2575_cast_fp16 = rsqrt(epsilon = var_2575_epsilon_0, x = var_2574_cast_fp16)[name = string("op_2575_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_39_cast_fp16 = mul(x = y_45_cast_fp16, y = var_2575_cast_fp16)[name = string("out_39_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29586688)))];
            tensor<fp16, [1, 630, 12, 64]> x_63_cast_fp16 = mul(x = out_39_cast_fp16, y = t_encoder_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_63_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29586880))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30176768))))[name = string("t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_23_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_2585 = const()[name = string("op_2585"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_59_cast_fp16 = reshape(shape = var_2585, x = linear_23_cast_fp16)[name = string("x_59_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2587_cast_fp16 = abs(x = x_59_cast_fp16)[name = string("op_2587_cast_fp16")];
            bool var_2590 = const()[name = string("op_2590"), val = bool(true)];
            tensor<int32, [1]> var_2591_axes_0 = const()[name = string("op_2591_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2591_cast_fp16 = reduce_max(axes = var_2591_axes_0, keep_dims = var_2590, x = var_2587_cast_fp16)[name = string("op_2591_cast_fp16")];
            fp16 var_2592_to_fp16 = const()[name = string("op_2592_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_51_cast_fp16 = maximum(x = var_2591_cast_fp16, y = var_2592_to_fp16)[name = string("s_51_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_47_cast_fp16 = real_div(x = x_59_cast_fp16, y = s_51_cast_fp16)[name = string("y_47_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2595_cast_fp16 = mul(x = y_47_cast_fp16, y = y_47_cast_fp16)[name = string("op_2595_cast_fp16")];
            tensor<int32, [1]> var_2600_axes_0 = const()[name = string("op_2600_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2600_keep_dims_0 = const()[name = string("op_2600_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2600_cast_fp16 = reduce_mean(axes = var_2600_axes_0, keep_dims = var_2600_keep_dims_0, x = var_2595_cast_fp16)[name = string("op_2600_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2601_cast_fp16 = mul(x = s_51_cast_fp16, y = s_51_cast_fp16)[name = string("op_2601_cast_fp16")];
            fp32 var_2602_epsilon_0 = const()[name = string("op_2602_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2602_cast_fp16 = inverse(epsilon = var_2602_epsilon_0, x = var_2601_cast_fp16)[name = string("op_2602_cast_fp16")];
            fp16 var_2603_to_fp16 = const()[name = string("op_2603_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2604_cast_fp16 = mul(x = var_2602_cast_fp16, y = var_2603_to_fp16)[name = string("op_2604_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2606_cast_fp16 = add(x = var_2600_cast_fp16, y = var_2604_cast_fp16)[name = string("op_2606_cast_fp16")];
            fp32 var_2607_epsilon_0 = const()[name = string("op_2607_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2607_cast_fp16 = rsqrt(epsilon = var_2607_epsilon_0, x = var_2606_cast_fp16)[name = string("op_2607_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_41_cast_fp16 = mul(x = y_47_cast_fp16, y = var_2607_cast_fp16)[name = string("out_41_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178368)))];
            tensor<fp16, [1, 630, 12, 64]> x_65_cast_fp16 = mul(x = out_41_cast_fp16, y = t_encoder_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_65_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178560))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30768448))))[name = string("t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<int32, [4]> var_2617 = const()[name = string("op_2617"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_61_cast_fp16 = reshape(shape = var_2617, x = linear_24_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2619_cast_fp16 = abs(x = x_61_cast_fp16)[name = string("op_2619_cast_fp16")];
            bool var_2622 = const()[name = string("op_2622"), val = bool(true)];
            tensor<int32, [1]> var_2623_axes_0 = const()[name = string("op_2623_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2623_cast_fp16 = reduce_max(axes = var_2623_axes_0, keep_dims = var_2622, x = var_2619_cast_fp16)[name = string("op_2623_cast_fp16")];
            fp16 var_2624_to_fp16 = const()[name = string("op_2624_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_53_cast_fp16 = maximum(x = var_2623_cast_fp16, y = var_2624_to_fp16)[name = string("s_53_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_49_cast_fp16 = real_div(x = x_61_cast_fp16, y = s_53_cast_fp16)[name = string("y_49_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2627_cast_fp16 = mul(x = y_49_cast_fp16, y = y_49_cast_fp16)[name = string("op_2627_cast_fp16")];
            tensor<int32, [1]> var_2632_axes_0 = const()[name = string("op_2632_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2632_keep_dims_0 = const()[name = string("op_2632_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2632_cast_fp16 = reduce_mean(axes = var_2632_axes_0, keep_dims = var_2632_keep_dims_0, x = var_2627_cast_fp16)[name = string("op_2632_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2633_cast_fp16 = mul(x = s_53_cast_fp16, y = s_53_cast_fp16)[name = string("op_2633_cast_fp16")];
            fp32 var_2634_epsilon_0 = const()[name = string("op_2634_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2634_cast_fp16 = inverse(epsilon = var_2634_epsilon_0, x = var_2633_cast_fp16)[name = string("op_2634_cast_fp16")];
            fp16 var_2635_to_fp16 = const()[name = string("op_2635_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2636_cast_fp16 = mul(x = var_2634_cast_fp16, y = var_2635_to_fp16)[name = string("op_2636_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2638_cast_fp16 = add(x = var_2632_cast_fp16, y = var_2636_cast_fp16)[name = string("op_2638_cast_fp16")];
            fp32 var_2639_epsilon_0 = const()[name = string("op_2639_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2639_cast_fp16 = rsqrt(epsilon = var_2639_epsilon_0, x = var_2638_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_13_cast_fp16 = mul(x = y_49_cast_fp16, y = var_2639_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<int32, [4]> z_25_begin_0 = const()[name = string("z_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_25_end_0 = const()[name = string("z_25_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_25_end_mask_0 = const()[name = string("z_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_25_cast_fp16 = slice_by_index(begin = z_25_begin_0, end = z_25_end_0, end_mask = z_25_end_mask_0, x = x_63_cast_fp16)[name = string("z_25_cast_fp16")];
            tensor<int32, [4]> var_2650_begin_0 = const()[name = string("op_2650_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2650_end_0 = const()[name = string("op_2650_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2650_end_mask_0 = const()[name = string("op_2650_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2650_cast_fp16 = slice_by_index(begin = var_2650_begin_0, end = var_2650_end_0, end_mask = var_2650_end_mask_0, x = z_25_cast_fp16)[name = string("op_2650_cast_fp16")];
            fp16 const_12_promoted_to_fp16 = const()[name = string("const_12_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2651_cast_fp16 = mul(x = var_2650_cast_fp16, y = const_12_promoted_to_fp16)[name = string("op_2651_cast_fp16")];
            tensor<int32, [4]> var_2656_begin_0 = const()[name = string("op_2656_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2656_end_0 = const()[name = string("op_2656_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2656_end_mask_0 = const()[name = string("op_2656_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2656_cast_fp16 = slice_by_index(begin = var_2656_begin_0, end = var_2656_end_0, end_mask = var_2656_end_mask_0, x = z_25_cast_fp16)[name = string("op_2656_cast_fp16")];
            int32 var_2658 = const()[name = string("op_2658"), val = int32(-1)];
            bool rot_25_interleave_0 = const()[name = string("rot_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_25_cast_fp16 = concat(axis = var_2658, interleave = rot_25_interleave_0, values = (var_2651_cast_fp16, var_2656_cast_fp16))[name = string("rot_25_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2660_cast_fp16 = mul(x = z_25_cast_fp16, y = c_1_cast_fp16)[name = string("op_2660_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2661_cast_fp16 = mul(x = rot_25_cast_fp16, y = s_9_cast_fp16)[name = string("op_2661_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2663_cast_fp16 = add(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = string("op_2663_cast_fp16")];
            tensor<int32, [4]> z_27_begin_0 = const()[name = string("z_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_27_end_0 = const()[name = string("z_27_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_27_end_mask_0 = const()[name = string("z_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_27_cast_fp16 = slice_by_index(begin = z_27_begin_0, end = z_27_end_0, end_mask = z_27_end_mask_0, x = x_63_cast_fp16)[name = string("z_27_cast_fp16")];
            tensor<int32, [4]> var_2673_begin_0 = const()[name = string("op_2673_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2673_end_0 = const()[name = string("op_2673_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2673_end_mask_0 = const()[name = string("op_2673_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2673_cast_fp16 = slice_by_index(begin = var_2673_begin_0, end = var_2673_end_0, end_mask = var_2673_end_mask_0, x = z_27_cast_fp16)[name = string("op_2673_cast_fp16")];
            fp16 const_13_promoted_to_fp16 = const()[name = string("const_13_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2674_cast_fp16 = mul(x = var_2673_cast_fp16, y = const_13_promoted_to_fp16)[name = string("op_2674_cast_fp16")];
            tensor<int32, [4]> var_2679_begin_0 = const()[name = string("op_2679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2679_end_0 = const()[name = string("op_2679_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2679_end_mask_0 = const()[name = string("op_2679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2679_cast_fp16 = slice_by_index(begin = var_2679_begin_0, end = var_2679_end_0, end_mask = var_2679_end_mask_0, x = z_27_cast_fp16)[name = string("op_2679_cast_fp16")];
            int32 var_2681 = const()[name = string("op_2681"), val = int32(-1)];
            bool rot_27_interleave_0 = const()[name = string("rot_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_27_cast_fp16 = concat(axis = var_2681, interleave = rot_27_interleave_0, values = (var_2674_cast_fp16, var_2679_cast_fp16))[name = string("rot_27_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2683_cast_fp16 = mul(x = z_27_cast_fp16, y = c_cast_fp16)[name = string("op_2683_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2684_cast_fp16 = mul(x = rot_27_cast_fp16, y = s_11_cast_fp16)[name = string("op_2684_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2686_cast_fp16 = add(x = var_2683_cast_fp16, y = var_2684_cast_fp16)[name = string("op_2686_cast_fp16")];
            int32 var_2688 = const()[name = string("op_2688"), val = int32(-1)];
            bool var_2689_interleave_0 = const()[name = string("op_2689_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_2689_cast_fp16 = concat(axis = var_2688, interleave = var_2689_interleave_0, values = (var_2663_cast_fp16, var_2686_cast_fp16))[name = string("op_2689_cast_fp16")];
            tensor<int32, [4]> z_29_begin_0 = const()[name = string("z_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_29_end_0 = const()[name = string("z_29_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_29_end_mask_0 = const()[name = string("z_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_29_cast_fp16 = slice_by_index(begin = z_29_begin_0, end = z_29_end_0, end_mask = z_29_end_mask_0, x = x_65_cast_fp16)[name = string("z_29_cast_fp16")];
            tensor<int32, [4]> var_2702_begin_0 = const()[name = string("op_2702_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2702_end_0 = const()[name = string("op_2702_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2702_end_mask_0 = const()[name = string("op_2702_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2702_cast_fp16 = slice_by_index(begin = var_2702_begin_0, end = var_2702_end_0, end_mask = var_2702_end_mask_0, x = z_29_cast_fp16)[name = string("op_2702_cast_fp16")];
            fp16 const_14_promoted_to_fp16 = const()[name = string("const_14_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2703_cast_fp16 = mul(x = var_2702_cast_fp16, y = const_14_promoted_to_fp16)[name = string("op_2703_cast_fp16")];
            tensor<int32, [4]> var_2708_begin_0 = const()[name = string("op_2708_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2708_end_0 = const()[name = string("op_2708_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2708_end_mask_0 = const()[name = string("op_2708_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2708_cast_fp16 = slice_by_index(begin = var_2708_begin_0, end = var_2708_end_0, end_mask = var_2708_end_mask_0, x = z_29_cast_fp16)[name = string("op_2708_cast_fp16")];
            int32 var_2710 = const()[name = string("op_2710"), val = int32(-1)];
            bool rot_29_interleave_0 = const()[name = string("rot_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_29_cast_fp16 = concat(axis = var_2710, interleave = rot_29_interleave_0, values = (var_2703_cast_fp16, var_2708_cast_fp16))[name = string("rot_29_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2712_cast_fp16 = mul(x = z_29_cast_fp16, y = c_1_cast_fp16)[name = string("op_2712_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2713_cast_fp16 = mul(x = rot_29_cast_fp16, y = s_9_cast_fp16)[name = string("op_2713_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2715_cast_fp16 = add(x = var_2712_cast_fp16, y = var_2713_cast_fp16)[name = string("op_2715_cast_fp16")];
            tensor<int32, [4]> z_31_begin_0 = const()[name = string("z_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_31_end_0 = const()[name = string("z_31_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_31_end_mask_0 = const()[name = string("z_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_31_cast_fp16 = slice_by_index(begin = z_31_begin_0, end = z_31_end_0, end_mask = z_31_end_mask_0, x = x_65_cast_fp16)[name = string("z_31_cast_fp16")];
            tensor<int32, [4]> var_2725_begin_0 = const()[name = string("op_2725_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2725_end_0 = const()[name = string("op_2725_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2725_end_mask_0 = const()[name = string("op_2725_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2725_cast_fp16 = slice_by_index(begin = var_2725_begin_0, end = var_2725_end_0, end_mask = var_2725_end_mask_0, x = z_31_cast_fp16)[name = string("op_2725_cast_fp16")];
            fp16 const_15_promoted_to_fp16 = const()[name = string("const_15_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2726_cast_fp16 = mul(x = var_2725_cast_fp16, y = const_15_promoted_to_fp16)[name = string("op_2726_cast_fp16")];
            tensor<int32, [4]> var_2731_begin_0 = const()[name = string("op_2731_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2731_end_0 = const()[name = string("op_2731_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2731_end_mask_0 = const()[name = string("op_2731_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2731_cast_fp16 = slice_by_index(begin = var_2731_begin_0, end = var_2731_end_0, end_mask = var_2731_end_mask_0, x = z_31_cast_fp16)[name = string("op_2731_cast_fp16")];
            int32 var_2733 = const()[name = string("op_2733"), val = int32(-1)];
            bool rot_31_interleave_0 = const()[name = string("rot_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_31_cast_fp16 = concat(axis = var_2733, interleave = rot_31_interleave_0, values = (var_2726_cast_fp16, var_2731_cast_fp16))[name = string("rot_31_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2735_cast_fp16 = mul(x = z_31_cast_fp16, y = c_cast_fp16)[name = string("op_2735_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2736_cast_fp16 = mul(x = rot_31_cast_fp16, y = s_11_cast_fp16)[name = string("op_2736_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2738_cast_fp16 = add(x = var_2735_cast_fp16, y = var_2736_cast_fp16)[name = string("op_2738_cast_fp16")];
            int32 var_2740 = const()[name = string("op_2740"), val = int32(-1)];
            bool var_2741_interleave_0 = const()[name = string("op_2741_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_2741_cast_fp16 = concat(axis = var_2740, interleave = var_2741_interleave_0, values = (var_2715_cast_fp16, var_2738_cast_fp16))[name = string("op_2741_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_2751_transpose_x_0 = const()[name = string("op_2751_transpose_x_0"), val = bool(false)];
            bool var_2751_transpose_y_0 = const()[name = string("op_2751_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_70_perm_0 = const()[name = string("transpose_70_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_71_perm_0 = const()[name = string("transpose_71_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_71 = transpose(perm = transpose_71_perm_0, x = var_2741_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [1, 12, 630, 64]> transpose_70 = transpose(perm = transpose_70_perm_0, x = var_2689_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 12, 630, 630]> var_2751_cast_fp16 = matmul(transpose_x = var_2751_transpose_x_0, transpose_y = var_2751_transpose_y_0, x = transpose_70, y = transpose_71)[name = string("op_2751_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_2753_cast_fp16 = add(x = var_2751_cast_fp16, y = var_1524_cast_fp16)[name = string("op_2753_cast_fp16")];
            int32 var_2754 = const()[name = string("op_2754"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_49_cast_fp16 = softmax(axis = var_2754, x = var_2753_cast_fp16)[name = string("w_49_cast_fp16")];
            bool var_2757_transpose_x_0 = const()[name = string("op_2757_transpose_x_0"), val = bool(false)];
            bool var_2757_transpose_y_0 = const()[name = string("op_2757_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = v_13_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 12, 630, 64]> var_2757_cast_fp16 = matmul(transpose_x = var_2757_transpose_x_0, transpose_y = var_2757_transpose_y_0, x = w_49_cast_fp16, y = v_15_cast_fp16)[name = string("op_2757_cast_fp16")];
            tensor<int32, [4]> var_2760_perm_0 = const()[name = string("op_2760_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2764 = const()[name = string("op_2764"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_2760_cast_fp16 = transpose(perm = var_2760_perm_0, x = var_2757_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 630, 768]> input_35_cast_fp16 = reshape(shape = var_2764, x = var_2760_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30770048))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31359936))))[name = string("t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_25_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2769_cast_fp16 = abs(x = linear_25_cast_fp16)[name = string("op_2769_cast_fp16")];
            bool var_2772 = const()[name = string("op_2772"), val = bool(true)];
            tensor<int32, [1]> var_2773_axes_0 = const()[name = string("op_2773_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2773_cast_fp16 = reduce_max(axes = var_2773_axes_0, keep_dims = var_2772, x = var_2769_cast_fp16)[name = string("op_2773_cast_fp16")];
            fp16 var_2774_to_fp16 = const()[name = string("op_2774_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_55_cast_fp16 = maximum(x = var_2773_cast_fp16, y = var_2774_to_fp16)[name = string("s_55_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_51_cast_fp16 = real_div(x = linear_25_cast_fp16, y = s_55_cast_fp16)[name = string("y_51_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2777_cast_fp16 = mul(x = y_51_cast_fp16, y = y_51_cast_fp16)[name = string("op_2777_cast_fp16")];
            tensor<int32, [1]> var_2782_axes_0 = const()[name = string("op_2782_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2782_keep_dims_0 = const()[name = string("op_2782_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2782_cast_fp16 = reduce_mean(axes = var_2782_axes_0, keep_dims = var_2782_keep_dims_0, x = var_2777_cast_fp16)[name = string("op_2782_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2783_cast_fp16 = mul(x = s_55_cast_fp16, y = s_55_cast_fp16)[name = string("op_2783_cast_fp16")];
            fp32 var_2784_epsilon_0 = const()[name = string("op_2784_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2784_cast_fp16 = inverse(epsilon = var_2784_epsilon_0, x = var_2783_cast_fp16)[name = string("op_2784_cast_fp16")];
            fp16 var_2785_to_fp16 = const()[name = string("op_2785_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2786_cast_fp16 = mul(x = var_2784_cast_fp16, y = var_2785_to_fp16)[name = string("op_2786_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2788_cast_fp16 = add(x = var_2782_cast_fp16, y = var_2786_cast_fp16)[name = string("op_2788_cast_fp16")];
            fp32 var_2789_epsilon_0 = const()[name = string("op_2789_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2789_cast_fp16 = rsqrt(epsilon = var_2789_epsilon_0, x = var_2788_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_43_cast_fp16 = mul(x = y_51_cast_fp16, y = var_2789_cast_fp16)[name = string("out_43_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31361536)))];
            tensor<fp16, [1, 630, 768]> var_2791_cast_fp16 = mul(x = out_43_cast_fp16, y = t_encoder_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("op_2791_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_69_cast_fp16 = add(x = x_55_cast_fp16, y = var_2791_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2794_cast_fp16 = abs(x = x_69_cast_fp16)[name = string("op_2794_cast_fp16")];
            bool var_2797 = const()[name = string("op_2797"), val = bool(true)];
            tensor<int32, [1]> var_2798_axes_0 = const()[name = string("op_2798_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2798_cast_fp16 = reduce_max(axes = var_2798_axes_0, keep_dims = var_2797, x = var_2794_cast_fp16)[name = string("op_2798_cast_fp16")];
            fp16 var_2799_to_fp16 = const()[name = string("op_2799_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_57_cast_fp16 = maximum(x = var_2798_cast_fp16, y = var_2799_to_fp16)[name = string("s_57_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_53_cast_fp16 = real_div(x = x_69_cast_fp16, y = s_57_cast_fp16)[name = string("y_53_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2802_cast_fp16 = mul(x = y_53_cast_fp16, y = y_53_cast_fp16)[name = string("op_2802_cast_fp16")];
            tensor<int32, [1]> var_2807_axes_0 = const()[name = string("op_2807_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2807_keep_dims_0 = const()[name = string("op_2807_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2807_cast_fp16 = reduce_mean(axes = var_2807_axes_0, keep_dims = var_2807_keep_dims_0, x = var_2802_cast_fp16)[name = string("op_2807_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2808_cast_fp16 = mul(x = s_57_cast_fp16, y = s_57_cast_fp16)[name = string("op_2808_cast_fp16")];
            fp32 var_2809_epsilon_0 = const()[name = string("op_2809_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2809_cast_fp16 = inverse(epsilon = var_2809_epsilon_0, x = var_2808_cast_fp16)[name = string("op_2809_cast_fp16")];
            fp16 var_2810_to_fp16 = const()[name = string("op_2810_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2811_cast_fp16 = mul(x = var_2809_cast_fp16, y = var_2810_to_fp16)[name = string("op_2811_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2813_cast_fp16 = add(x = var_2807_cast_fp16, y = var_2811_cast_fp16)[name = string("op_2813_cast_fp16")];
            fp32 var_2814_epsilon_0 = const()[name = string("op_2814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2814_cast_fp16 = rsqrt(epsilon = var_2814_epsilon_0, x = var_2813_cast_fp16)[name = string("op_2814_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_45_cast_fp16 = mul(x = y_53_cast_fp16, y = var_2814_cast_fp16)[name = string("out_45_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31363136)))];
            tensor<fp16, [1, 630, 768]> input_37_cast_fp16 = mul(x = out_45_cast_fp16, y = t_encoder_layers_3_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31364736))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33724096))))[name = string("t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_26_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_26_cast_fp16")];
            string var_2821_mode_0 = const()[name = string("op_2821_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_2821_cast_fp16 = gelu(mode = var_2821_mode_0, x = linear_26_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33730304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36089664))))[name = string("t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_27_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_37_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_41_cast_fp16 = mul(x = var_2821_cast_fp16, y = linear_27_cast_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(36095872))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38455232))))[name = string("t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_3_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2829_cast_fp16 = abs(x = linear_28_cast_fp16)[name = string("op_2829_cast_fp16")];
            bool var_2832 = const()[name = string("op_2832"), val = bool(true)];
            tensor<int32, [1]> var_2833_axes_0 = const()[name = string("op_2833_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2833_cast_fp16 = reduce_max(axes = var_2833_axes_0, keep_dims = var_2832, x = var_2829_cast_fp16)[name = string("op_2833_cast_fp16")];
            fp16 var_2834_to_fp16 = const()[name = string("op_2834_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_59_cast_fp16 = maximum(x = var_2833_cast_fp16, y = var_2834_to_fp16)[name = string("s_59_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_55_cast_fp16 = real_div(x = linear_28_cast_fp16, y = s_59_cast_fp16)[name = string("y_55_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2837_cast_fp16 = mul(x = y_55_cast_fp16, y = y_55_cast_fp16)[name = string("op_2837_cast_fp16")];
            tensor<int32, [1]> var_2842_axes_0 = const()[name = string("op_2842_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2842_keep_dims_0 = const()[name = string("op_2842_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2842_cast_fp16 = reduce_mean(axes = var_2842_axes_0, keep_dims = var_2842_keep_dims_0, x = var_2837_cast_fp16)[name = string("op_2842_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2843_cast_fp16 = mul(x = s_59_cast_fp16, y = s_59_cast_fp16)[name = string("op_2843_cast_fp16")];
            fp32 var_2844_epsilon_0 = const()[name = string("op_2844_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2844_cast_fp16 = inverse(epsilon = var_2844_epsilon_0, x = var_2843_cast_fp16)[name = string("op_2844_cast_fp16")];
            fp16 var_2845_to_fp16 = const()[name = string("op_2845_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2846_cast_fp16 = mul(x = var_2844_cast_fp16, y = var_2845_to_fp16)[name = string("op_2846_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2848_cast_fp16 = add(x = var_2842_cast_fp16, y = var_2846_cast_fp16)[name = string("op_2848_cast_fp16")];
            fp32 var_2849_epsilon_0 = const()[name = string("op_2849_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2849_cast_fp16 = rsqrt(epsilon = var_2849_epsilon_0, x = var_2848_cast_fp16)[name = string("op_2849_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_47_cast_fp16 = mul(x = y_55_cast_fp16, y = var_2849_cast_fp16)[name = string("out_47_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38456832)))];
            tensor<fp16, [1, 630, 768]> var_2851_cast_fp16 = mul(x = out_47_cast_fp16, y = t_encoder_layers_3_post_feedforward_layernorm_weight_to_fp16)[name = string("op_2851_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_73_cast_fp16 = add(x = x_69_cast_fp16, y = var_2851_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2854_cast_fp16 = abs(x = x_73_cast_fp16)[name = string("op_2854_cast_fp16")];
            bool var_2857 = const()[name = string("op_2857"), val = bool(true)];
            tensor<int32, [1]> var_2858_axes_0 = const()[name = string("op_2858_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_2858_cast_fp16 = reduce_max(axes = var_2858_axes_0, keep_dims = var_2857, x = var_2854_cast_fp16)[name = string("op_2858_cast_fp16")];
            fp16 var_2859_to_fp16 = const()[name = string("op_2859_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_61_cast_fp16 = maximum(x = var_2858_cast_fp16, y = var_2859_to_fp16)[name = string("s_61_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_57_cast_fp16 = real_div(x = x_73_cast_fp16, y = s_61_cast_fp16)[name = string("y_57_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_2862_cast_fp16 = mul(x = y_57_cast_fp16, y = y_57_cast_fp16)[name = string("op_2862_cast_fp16")];
            tensor<int32, [1]> var_2867_axes_0 = const()[name = string("op_2867_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2867_keep_dims_0 = const()[name = string("op_2867_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_2867_cast_fp16 = reduce_mean(axes = var_2867_axes_0, keep_dims = var_2867_keep_dims_0, x = var_2862_cast_fp16)[name = string("op_2867_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2868_cast_fp16 = mul(x = s_61_cast_fp16, y = s_61_cast_fp16)[name = string("op_2868_cast_fp16")];
            fp32 var_2869_epsilon_0 = const()[name = string("op_2869_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_2869_cast_fp16 = inverse(epsilon = var_2869_epsilon_0, x = var_2868_cast_fp16)[name = string("op_2869_cast_fp16")];
            fp16 var_2870_to_fp16 = const()[name = string("op_2870_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_2871_cast_fp16 = mul(x = var_2869_cast_fp16, y = var_2870_to_fp16)[name = string("op_2871_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_2873_cast_fp16 = add(x = var_2867_cast_fp16, y = var_2871_cast_fp16)[name = string("op_2873_cast_fp16")];
            fp32 var_2874_epsilon_0 = const()[name = string("op_2874_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_2874_cast_fp16 = rsqrt(epsilon = var_2874_epsilon_0, x = var_2873_cast_fp16)[name = string("op_2874_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_49_cast_fp16 = mul(x = y_57_cast_fp16, y = var_2874_cast_fp16)[name = string("out_49_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38458432)))];
            tensor<fp16, [1, 630, 768]> input_43_cast_fp16 = mul(x = out_49_cast_fp16, y = t_encoder_layers_4_input_layernorm_weight_to_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(38460032))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39049920))))[name = string("t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_29_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2884 = const()[name = string("op_2884"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_75_cast_fp16 = reshape(shape = var_2884, x = linear_29_cast_fp16)[name = string("x_75_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2886_cast_fp16 = abs(x = x_75_cast_fp16)[name = string("op_2886_cast_fp16")];
            bool var_2889 = const()[name = string("op_2889"), val = bool(true)];
            tensor<int32, [1]> var_2890_axes_0 = const()[name = string("op_2890_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2890_cast_fp16 = reduce_max(axes = var_2890_axes_0, keep_dims = var_2889, x = var_2886_cast_fp16)[name = string("op_2890_cast_fp16")];
            fp16 var_2891_to_fp16 = const()[name = string("op_2891_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_63_cast_fp16 = maximum(x = var_2890_cast_fp16, y = var_2891_to_fp16)[name = string("s_63_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_59_cast_fp16 = real_div(x = x_75_cast_fp16, y = s_63_cast_fp16)[name = string("y_59_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2894_cast_fp16 = mul(x = y_59_cast_fp16, y = y_59_cast_fp16)[name = string("op_2894_cast_fp16")];
            tensor<int32, [1]> var_2899_axes_0 = const()[name = string("op_2899_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2899_keep_dims_0 = const()[name = string("op_2899_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2899_cast_fp16 = reduce_mean(axes = var_2899_axes_0, keep_dims = var_2899_keep_dims_0, x = var_2894_cast_fp16)[name = string("op_2899_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2900_cast_fp16 = mul(x = s_63_cast_fp16, y = s_63_cast_fp16)[name = string("op_2900_cast_fp16")];
            fp32 var_2901_epsilon_0 = const()[name = string("op_2901_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2901_cast_fp16 = inverse(epsilon = var_2901_epsilon_0, x = var_2900_cast_fp16)[name = string("op_2901_cast_fp16")];
            fp16 var_2902_to_fp16 = const()[name = string("op_2902_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2903_cast_fp16 = mul(x = var_2901_cast_fp16, y = var_2902_to_fp16)[name = string("op_2903_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2905_cast_fp16 = add(x = var_2899_cast_fp16, y = var_2903_cast_fp16)[name = string("op_2905_cast_fp16")];
            fp32 var_2906_epsilon_0 = const()[name = string("op_2906_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2906_cast_fp16 = rsqrt(epsilon = var_2906_epsilon_0, x = var_2905_cast_fp16)[name = string("op_2906_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_51_cast_fp16 = mul(x = y_59_cast_fp16, y = var_2906_cast_fp16)[name = string("out_51_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39051520)))];
            tensor<fp16, [1, 630, 12, 64]> x_81_cast_fp16 = mul(x = out_51_cast_fp16, y = t_encoder_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39051712))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39641600))))[name = string("t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2916 = const()[name = string("op_2916"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_77_cast_fp16 = reshape(shape = var_2916, x = linear_30_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2918_cast_fp16 = abs(x = x_77_cast_fp16)[name = string("op_2918_cast_fp16")];
            bool var_2921 = const()[name = string("op_2921"), val = bool(true)];
            tensor<int32, [1]> var_2922_axes_0 = const()[name = string("op_2922_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2922_cast_fp16 = reduce_max(axes = var_2922_axes_0, keep_dims = var_2921, x = var_2918_cast_fp16)[name = string("op_2922_cast_fp16")];
            fp16 var_2923_to_fp16 = const()[name = string("op_2923_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_65_cast_fp16 = maximum(x = var_2922_cast_fp16, y = var_2923_to_fp16)[name = string("s_65_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_61_cast_fp16 = real_div(x = x_77_cast_fp16, y = s_65_cast_fp16)[name = string("y_61_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2926_cast_fp16 = mul(x = y_61_cast_fp16, y = y_61_cast_fp16)[name = string("op_2926_cast_fp16")];
            tensor<int32, [1]> var_2931_axes_0 = const()[name = string("op_2931_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2931_keep_dims_0 = const()[name = string("op_2931_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2931_cast_fp16 = reduce_mean(axes = var_2931_axes_0, keep_dims = var_2931_keep_dims_0, x = var_2926_cast_fp16)[name = string("op_2931_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2932_cast_fp16 = mul(x = s_65_cast_fp16, y = s_65_cast_fp16)[name = string("op_2932_cast_fp16")];
            fp32 var_2933_epsilon_0 = const()[name = string("op_2933_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2933_cast_fp16 = inverse(epsilon = var_2933_epsilon_0, x = var_2932_cast_fp16)[name = string("op_2933_cast_fp16")];
            fp16 var_2934_to_fp16 = const()[name = string("op_2934_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2935_cast_fp16 = mul(x = var_2933_cast_fp16, y = var_2934_to_fp16)[name = string("op_2935_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2937_cast_fp16 = add(x = var_2931_cast_fp16, y = var_2935_cast_fp16)[name = string("op_2937_cast_fp16")];
            fp32 var_2938_epsilon_0 = const()[name = string("op_2938_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2938_cast_fp16 = rsqrt(epsilon = var_2938_epsilon_0, x = var_2937_cast_fp16)[name = string("op_2938_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_53_cast_fp16 = mul(x = y_61_cast_fp16, y = var_2938_cast_fp16)[name = string("out_53_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30178368)))];
            tensor<fp16, [1, 630, 12, 64]> x_83_cast_fp16 = mul(x = out_53_cast_fp16, y = t_encoder_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_83_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(39643200))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40233088))))[name = string("t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<int32, [4]> var_2948 = const()[name = string("op_2948"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_79_cast_fp16 = reshape(shape = var_2948, x = linear_31_cast_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2950_cast_fp16 = abs(x = x_79_cast_fp16)[name = string("op_2950_cast_fp16")];
            bool var_2953 = const()[name = string("op_2953"), val = bool(true)];
            tensor<int32, [1]> var_2954_axes_0 = const()[name = string("op_2954_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_2954_cast_fp16 = reduce_max(axes = var_2954_axes_0, keep_dims = var_2953, x = var_2950_cast_fp16)[name = string("op_2954_cast_fp16")];
            fp16 var_2955_to_fp16 = const()[name = string("op_2955_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_67_cast_fp16 = maximum(x = var_2954_cast_fp16, y = var_2955_to_fp16)[name = string("s_67_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_63_cast_fp16 = real_div(x = x_79_cast_fp16, y = s_67_cast_fp16)[name = string("y_63_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_2958_cast_fp16 = mul(x = y_63_cast_fp16, y = y_63_cast_fp16)[name = string("op_2958_cast_fp16")];
            tensor<int32, [1]> var_2963_axes_0 = const()[name = string("op_2963_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2963_keep_dims_0 = const()[name = string("op_2963_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_2963_cast_fp16 = reduce_mean(axes = var_2963_axes_0, keep_dims = var_2963_keep_dims_0, x = var_2958_cast_fp16)[name = string("op_2963_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2964_cast_fp16 = mul(x = s_67_cast_fp16, y = s_67_cast_fp16)[name = string("op_2964_cast_fp16")];
            fp32 var_2965_epsilon_0 = const()[name = string("op_2965_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_2965_cast_fp16 = inverse(epsilon = var_2965_epsilon_0, x = var_2964_cast_fp16)[name = string("op_2965_cast_fp16")];
            fp16 var_2966_to_fp16 = const()[name = string("op_2966_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_2967_cast_fp16 = mul(x = var_2965_cast_fp16, y = var_2966_to_fp16)[name = string("op_2967_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_2969_cast_fp16 = add(x = var_2963_cast_fp16, y = var_2967_cast_fp16)[name = string("op_2969_cast_fp16")];
            fp32 var_2970_epsilon_0 = const()[name = string("op_2970_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_2970_cast_fp16 = rsqrt(epsilon = var_2970_epsilon_0, x = var_2969_cast_fp16)[name = string("op_2970_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_17_cast_fp16 = mul(x = y_63_cast_fp16, y = var_2970_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<int32, [4]> z_33_begin_0 = const()[name = string("z_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_33_end_0 = const()[name = string("z_33_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_33_end_mask_0 = const()[name = string("z_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_33_cast_fp16 = slice_by_index(begin = z_33_begin_0, end = z_33_end_0, end_mask = z_33_end_mask_0, x = x_81_cast_fp16)[name = string("z_33_cast_fp16")];
            tensor<int32, [4]> var_2981_begin_0 = const()[name = string("op_2981_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_2981_end_0 = const()[name = string("op_2981_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_2981_end_mask_0 = const()[name = string("op_2981_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_2981_cast_fp16 = slice_by_index(begin = var_2981_begin_0, end = var_2981_end_0, end_mask = var_2981_end_mask_0, x = z_33_cast_fp16)[name = string("op_2981_cast_fp16")];
            fp16 const_16_promoted_to_fp16 = const()[name = string("const_16_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_2982_cast_fp16 = mul(x = var_2981_cast_fp16, y = const_16_promoted_to_fp16)[name = string("op_2982_cast_fp16")];
            tensor<int32, [4]> var_2987_begin_0 = const()[name = string("op_2987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2987_end_0 = const()[name = string("op_2987_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_2987_end_mask_0 = const()[name = string("op_2987_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_2987_cast_fp16 = slice_by_index(begin = var_2987_begin_0, end = var_2987_end_0, end_mask = var_2987_end_mask_0, x = z_33_cast_fp16)[name = string("op_2987_cast_fp16")];
            int32 var_2989 = const()[name = string("op_2989"), val = int32(-1)];
            bool rot_33_interleave_0 = const()[name = string("rot_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_33_cast_fp16 = concat(axis = var_2989, interleave = rot_33_interleave_0, values = (var_2982_cast_fp16, var_2987_cast_fp16))[name = string("rot_33_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2991_cast_fp16 = mul(x = z_33_cast_fp16, y = c_1_cast_fp16)[name = string("op_2991_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2992_cast_fp16 = mul(x = rot_33_cast_fp16, y = s_9_cast_fp16)[name = string("op_2992_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_2994_cast_fp16 = add(x = var_2991_cast_fp16, y = var_2992_cast_fp16)[name = string("op_2994_cast_fp16")];
            tensor<int32, [4]> z_35_begin_0 = const()[name = string("z_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_35_end_0 = const()[name = string("z_35_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_35_end_mask_0 = const()[name = string("z_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_35_cast_fp16 = slice_by_index(begin = z_35_begin_0, end = z_35_end_0, end_mask = z_35_end_mask_0, x = x_81_cast_fp16)[name = string("z_35_cast_fp16")];
            tensor<int32, [4]> var_3004_begin_0 = const()[name = string("op_3004_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3004_end_0 = const()[name = string("op_3004_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3004_end_mask_0 = const()[name = string("op_3004_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, x = z_35_cast_fp16)[name = string("op_3004_cast_fp16")];
            fp16 const_17_promoted_to_fp16 = const()[name = string("const_17_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3005_cast_fp16 = mul(x = var_3004_cast_fp16, y = const_17_promoted_to_fp16)[name = string("op_3005_cast_fp16")];
            tensor<int32, [4]> var_3010_begin_0 = const()[name = string("op_3010_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3010_end_0 = const()[name = string("op_3010_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3010_end_mask_0 = const()[name = string("op_3010_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3010_cast_fp16 = slice_by_index(begin = var_3010_begin_0, end = var_3010_end_0, end_mask = var_3010_end_mask_0, x = z_35_cast_fp16)[name = string("op_3010_cast_fp16")];
            int32 var_3012 = const()[name = string("op_3012"), val = int32(-1)];
            bool rot_35_interleave_0 = const()[name = string("rot_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_35_cast_fp16 = concat(axis = var_3012, interleave = rot_35_interleave_0, values = (var_3005_cast_fp16, var_3010_cast_fp16))[name = string("rot_35_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3014_cast_fp16 = mul(x = z_35_cast_fp16, y = c_cast_fp16)[name = string("op_3014_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3015_cast_fp16 = mul(x = rot_35_cast_fp16, y = s_11_cast_fp16)[name = string("op_3015_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3017_cast_fp16 = add(x = var_3014_cast_fp16, y = var_3015_cast_fp16)[name = string("op_3017_cast_fp16")];
            int32 var_3019 = const()[name = string("op_3019"), val = int32(-1)];
            bool var_3020_interleave_0 = const()[name = string("op_3020_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_3020_cast_fp16 = concat(axis = var_3019, interleave = var_3020_interleave_0, values = (var_2994_cast_fp16, var_3017_cast_fp16))[name = string("op_3020_cast_fp16")];
            tensor<int32, [4]> z_37_begin_0 = const()[name = string("z_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_37_end_0 = const()[name = string("z_37_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_37_end_mask_0 = const()[name = string("z_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_37_cast_fp16 = slice_by_index(begin = z_37_begin_0, end = z_37_end_0, end_mask = z_37_end_mask_0, x = x_83_cast_fp16)[name = string("z_37_cast_fp16")];
            tensor<int32, [4]> var_3033_begin_0 = const()[name = string("op_3033_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3033_end_0 = const()[name = string("op_3033_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3033_end_mask_0 = const()[name = string("op_3033_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3033_cast_fp16 = slice_by_index(begin = var_3033_begin_0, end = var_3033_end_0, end_mask = var_3033_end_mask_0, x = z_37_cast_fp16)[name = string("op_3033_cast_fp16")];
            fp16 const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3034_cast_fp16 = mul(x = var_3033_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_3034_cast_fp16")];
            tensor<int32, [4]> var_3039_begin_0 = const()[name = string("op_3039_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3039_end_0 = const()[name = string("op_3039_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3039_end_mask_0 = const()[name = string("op_3039_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3039_cast_fp16 = slice_by_index(begin = var_3039_begin_0, end = var_3039_end_0, end_mask = var_3039_end_mask_0, x = z_37_cast_fp16)[name = string("op_3039_cast_fp16")];
            int32 var_3041 = const()[name = string("op_3041"), val = int32(-1)];
            bool rot_37_interleave_0 = const()[name = string("rot_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_37_cast_fp16 = concat(axis = var_3041, interleave = rot_37_interleave_0, values = (var_3034_cast_fp16, var_3039_cast_fp16))[name = string("rot_37_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3043_cast_fp16 = mul(x = z_37_cast_fp16, y = c_1_cast_fp16)[name = string("op_3043_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3044_cast_fp16 = mul(x = rot_37_cast_fp16, y = s_9_cast_fp16)[name = string("op_3044_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3046_cast_fp16 = add(x = var_3043_cast_fp16, y = var_3044_cast_fp16)[name = string("op_3046_cast_fp16")];
            tensor<int32, [4]> z_39_begin_0 = const()[name = string("z_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_39_end_0 = const()[name = string("z_39_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_39_end_mask_0 = const()[name = string("z_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_39_cast_fp16 = slice_by_index(begin = z_39_begin_0, end = z_39_end_0, end_mask = z_39_end_mask_0, x = x_83_cast_fp16)[name = string("z_39_cast_fp16")];
            tensor<int32, [4]> var_3056_begin_0 = const()[name = string("op_3056_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3056_end_0 = const()[name = string("op_3056_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3056_end_mask_0 = const()[name = string("op_3056_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3056_cast_fp16 = slice_by_index(begin = var_3056_begin_0, end = var_3056_end_0, end_mask = var_3056_end_mask_0, x = z_39_cast_fp16)[name = string("op_3056_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3057_cast_fp16 = mul(x = var_3056_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_3057_cast_fp16")];
            tensor<int32, [4]> var_3062_begin_0 = const()[name = string("op_3062_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3062_end_0 = const()[name = string("op_3062_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3062_end_mask_0 = const()[name = string("op_3062_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3062_cast_fp16 = slice_by_index(begin = var_3062_begin_0, end = var_3062_end_0, end_mask = var_3062_end_mask_0, x = z_39_cast_fp16)[name = string("op_3062_cast_fp16")];
            int32 var_3064 = const()[name = string("op_3064"), val = int32(-1)];
            bool rot_39_interleave_0 = const()[name = string("rot_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_39_cast_fp16 = concat(axis = var_3064, interleave = rot_39_interleave_0, values = (var_3057_cast_fp16, var_3062_cast_fp16))[name = string("rot_39_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3066_cast_fp16 = mul(x = z_39_cast_fp16, y = c_cast_fp16)[name = string("op_3066_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3067_cast_fp16 = mul(x = rot_39_cast_fp16, y = s_11_cast_fp16)[name = string("op_3067_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3069_cast_fp16 = add(x = var_3066_cast_fp16, y = var_3067_cast_fp16)[name = string("op_3069_cast_fp16")];
            int32 var_3071 = const()[name = string("op_3071"), val = int32(-1)];
            bool var_3072_interleave_0 = const()[name = string("op_3072_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_3072_cast_fp16 = concat(axis = var_3071, interleave = var_3072_interleave_0, values = (var_3046_cast_fp16, var_3069_cast_fp16))[name = string("op_3072_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3082_transpose_x_0 = const()[name = string("op_3082_transpose_x_0"), val = bool(false)];
            bool var_3082_transpose_y_0 = const()[name = string("op_3082_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_3072_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 12, 630, 64]> transpose_72 = transpose(perm = transpose_72_perm_0, x = var_3020_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 12, 630, 630]> var_3082_cast_fp16 = matmul(transpose_x = var_3082_transpose_x_0, transpose_y = var_3082_transpose_y_0, x = transpose_72, y = transpose_73)[name = string("op_3082_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_3084_cast_fp16 = add(x = var_3082_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3084_cast_fp16")];
            int32 var_3085 = const()[name = string("op_3085"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_63_cast_fp16 = softmax(axis = var_3085, x = var_3084_cast_fp16)[name = string("w_63_cast_fp16")];
            bool var_3088_transpose_x_0 = const()[name = string("op_3088_transpose_x_0"), val = bool(false)];
            bool var_3088_transpose_y_0 = const()[name = string("op_3088_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = v_17_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 12, 630, 64]> var_3088_cast_fp16 = matmul(transpose_x = var_3088_transpose_x_0, transpose_y = var_3088_transpose_y_0, x = w_63_cast_fp16, y = v_19_cast_fp16)[name = string("op_3088_cast_fp16")];
            tensor<int32, [4]> var_3091_perm_0 = const()[name = string("op_3091_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3095 = const()[name = string("op_3095"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_3091_cast_fp16 = transpose(perm = var_3091_perm_0, x = var_3088_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 630, 768]> input_45_cast_fp16 = reshape(shape = var_3095, x = var_3091_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40234688))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40824576))))[name = string("t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3100_cast_fp16 = abs(x = linear_32_cast_fp16)[name = string("op_3100_cast_fp16")];
            bool var_3103 = const()[name = string("op_3103"), val = bool(true)];
            tensor<int32, [1]> var_3104_axes_0 = const()[name = string("op_3104_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3104_cast_fp16 = reduce_max(axes = var_3104_axes_0, keep_dims = var_3103, x = var_3100_cast_fp16)[name = string("op_3104_cast_fp16")];
            fp16 var_3105_to_fp16 = const()[name = string("op_3105_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_69_cast_fp16 = maximum(x = var_3104_cast_fp16, y = var_3105_to_fp16)[name = string("s_69_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_65_cast_fp16 = real_div(x = linear_32_cast_fp16, y = s_69_cast_fp16)[name = string("y_65_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3108_cast_fp16 = mul(x = y_65_cast_fp16, y = y_65_cast_fp16)[name = string("op_3108_cast_fp16")];
            tensor<int32, [1]> var_3113_axes_0 = const()[name = string("op_3113_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3113_keep_dims_0 = const()[name = string("op_3113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3113_cast_fp16 = reduce_mean(axes = var_3113_axes_0, keep_dims = var_3113_keep_dims_0, x = var_3108_cast_fp16)[name = string("op_3113_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3114_cast_fp16 = mul(x = s_69_cast_fp16, y = s_69_cast_fp16)[name = string("op_3114_cast_fp16")];
            fp32 var_3115_epsilon_0 = const()[name = string("op_3115_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3115_cast_fp16 = inverse(epsilon = var_3115_epsilon_0, x = var_3114_cast_fp16)[name = string("op_3115_cast_fp16")];
            fp16 var_3116_to_fp16 = const()[name = string("op_3116_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3117_cast_fp16 = mul(x = var_3115_cast_fp16, y = var_3116_to_fp16)[name = string("op_3117_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3119_cast_fp16 = add(x = var_3113_cast_fp16, y = var_3117_cast_fp16)[name = string("op_3119_cast_fp16")];
            fp32 var_3120_epsilon_0 = const()[name = string("op_3120_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3120_cast_fp16 = rsqrt(epsilon = var_3120_epsilon_0, x = var_3119_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_55_cast_fp16 = mul(x = y_65_cast_fp16, y = var_3120_cast_fp16)[name = string("out_55_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40826176)))];
            tensor<fp16, [1, 630, 768]> var_3122_cast_fp16 = mul(x = out_55_cast_fp16, y = t_encoder_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("op_3122_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_87_cast_fp16 = add(x = x_73_cast_fp16, y = var_3122_cast_fp16)[name = string("x_87_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3125_cast_fp16 = abs(x = x_87_cast_fp16)[name = string("op_3125_cast_fp16")];
            bool var_3128 = const()[name = string("op_3128"), val = bool(true)];
            tensor<int32, [1]> var_3129_axes_0 = const()[name = string("op_3129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3129_cast_fp16 = reduce_max(axes = var_3129_axes_0, keep_dims = var_3128, x = var_3125_cast_fp16)[name = string("op_3129_cast_fp16")];
            fp16 var_3130_to_fp16 = const()[name = string("op_3130_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_71_cast_fp16 = maximum(x = var_3129_cast_fp16, y = var_3130_to_fp16)[name = string("s_71_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_67_cast_fp16 = real_div(x = x_87_cast_fp16, y = s_71_cast_fp16)[name = string("y_67_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3133_cast_fp16 = mul(x = y_67_cast_fp16, y = y_67_cast_fp16)[name = string("op_3133_cast_fp16")];
            tensor<int32, [1]> var_3138_axes_0 = const()[name = string("op_3138_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3138_keep_dims_0 = const()[name = string("op_3138_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3138_cast_fp16 = reduce_mean(axes = var_3138_axes_0, keep_dims = var_3138_keep_dims_0, x = var_3133_cast_fp16)[name = string("op_3138_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3139_cast_fp16 = mul(x = s_71_cast_fp16, y = s_71_cast_fp16)[name = string("op_3139_cast_fp16")];
            fp32 var_3140_epsilon_0 = const()[name = string("op_3140_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3140_cast_fp16 = inverse(epsilon = var_3140_epsilon_0, x = var_3139_cast_fp16)[name = string("op_3140_cast_fp16")];
            fp16 var_3141_to_fp16 = const()[name = string("op_3141_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3142_cast_fp16 = mul(x = var_3140_cast_fp16, y = var_3141_to_fp16)[name = string("op_3142_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3144_cast_fp16 = add(x = var_3138_cast_fp16, y = var_3142_cast_fp16)[name = string("op_3144_cast_fp16")];
            fp32 var_3145_epsilon_0 = const()[name = string("op_3145_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3145_cast_fp16 = rsqrt(epsilon = var_3145_epsilon_0, x = var_3144_cast_fp16)[name = string("op_3145_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_57_cast_fp16 = mul(x = y_67_cast_fp16, y = var_3145_cast_fp16)[name = string("out_57_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40827776)))];
            tensor<fp16, [1, 630, 768]> input_47_cast_fp16 = mul(x = out_57_cast_fp16, y = t_encoder_layers_4_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40829376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43188736))))[name = string("t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_33_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_33_cast_fp16")];
            string var_3152_mode_0 = const()[name = string("op_3152_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_3152_cast_fp16 = gelu(mode = var_3152_mode_0, x = linear_33_cast_fp16)[name = string("op_3152_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(43194944))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45554304))))[name = string("t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_34_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_47_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_51_cast_fp16 = mul(x = var_3152_cast_fp16, y = linear_34_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(45560512))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47919872))))[name = string("t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_4_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3160_cast_fp16 = abs(x = linear_35_cast_fp16)[name = string("op_3160_cast_fp16")];
            bool var_3163 = const()[name = string("op_3163"), val = bool(true)];
            tensor<int32, [1]> var_3164_axes_0 = const()[name = string("op_3164_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3164_cast_fp16 = reduce_max(axes = var_3164_axes_0, keep_dims = var_3163, x = var_3160_cast_fp16)[name = string("op_3164_cast_fp16")];
            fp16 var_3165_to_fp16 = const()[name = string("op_3165_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_73_cast_fp16 = maximum(x = var_3164_cast_fp16, y = var_3165_to_fp16)[name = string("s_73_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_69_cast_fp16 = real_div(x = linear_35_cast_fp16, y = s_73_cast_fp16)[name = string("y_69_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3168_cast_fp16 = mul(x = y_69_cast_fp16, y = y_69_cast_fp16)[name = string("op_3168_cast_fp16")];
            tensor<int32, [1]> var_3173_axes_0 = const()[name = string("op_3173_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3173_keep_dims_0 = const()[name = string("op_3173_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3173_cast_fp16 = reduce_mean(axes = var_3173_axes_0, keep_dims = var_3173_keep_dims_0, x = var_3168_cast_fp16)[name = string("op_3173_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3174_cast_fp16 = mul(x = s_73_cast_fp16, y = s_73_cast_fp16)[name = string("op_3174_cast_fp16")];
            fp32 var_3175_epsilon_0 = const()[name = string("op_3175_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3175_cast_fp16 = inverse(epsilon = var_3175_epsilon_0, x = var_3174_cast_fp16)[name = string("op_3175_cast_fp16")];
            fp16 var_3176_to_fp16 = const()[name = string("op_3176_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3177_cast_fp16 = mul(x = var_3175_cast_fp16, y = var_3176_to_fp16)[name = string("op_3177_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3179_cast_fp16 = add(x = var_3173_cast_fp16, y = var_3177_cast_fp16)[name = string("op_3179_cast_fp16")];
            fp32 var_3180_epsilon_0 = const()[name = string("op_3180_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3180_cast_fp16 = rsqrt(epsilon = var_3180_epsilon_0, x = var_3179_cast_fp16)[name = string("op_3180_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_59_cast_fp16 = mul(x = y_69_cast_fp16, y = var_3180_cast_fp16)[name = string("out_59_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47921472)))];
            tensor<fp16, [1, 630, 768]> var_3182_cast_fp16 = mul(x = out_59_cast_fp16, y = t_encoder_layers_4_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3182_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_91_cast_fp16 = add(x = x_87_cast_fp16, y = var_3182_cast_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3185_cast_fp16 = abs(x = x_91_cast_fp16)[name = string("op_3185_cast_fp16")];
            bool var_3188 = const()[name = string("op_3188"), val = bool(true)];
            tensor<int32, [1]> var_3189_axes_0 = const()[name = string("op_3189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3189_cast_fp16 = reduce_max(axes = var_3189_axes_0, keep_dims = var_3188, x = var_3185_cast_fp16)[name = string("op_3189_cast_fp16")];
            fp16 var_3190_to_fp16 = const()[name = string("op_3190_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_75_cast_fp16 = maximum(x = var_3189_cast_fp16, y = var_3190_to_fp16)[name = string("s_75_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_71_cast_fp16 = real_div(x = x_91_cast_fp16, y = s_75_cast_fp16)[name = string("y_71_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3193_cast_fp16 = mul(x = y_71_cast_fp16, y = y_71_cast_fp16)[name = string("op_3193_cast_fp16")];
            tensor<int32, [1]> var_3198_axes_0 = const()[name = string("op_3198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3198_keep_dims_0 = const()[name = string("op_3198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3198_cast_fp16 = reduce_mean(axes = var_3198_axes_0, keep_dims = var_3198_keep_dims_0, x = var_3193_cast_fp16)[name = string("op_3198_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3199_cast_fp16 = mul(x = s_75_cast_fp16, y = s_75_cast_fp16)[name = string("op_3199_cast_fp16")];
            fp32 var_3200_epsilon_0 = const()[name = string("op_3200_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3200_cast_fp16 = inverse(epsilon = var_3200_epsilon_0, x = var_3199_cast_fp16)[name = string("op_3200_cast_fp16")];
            fp16 var_3201_to_fp16 = const()[name = string("op_3201_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3202_cast_fp16 = mul(x = var_3200_cast_fp16, y = var_3201_to_fp16)[name = string("op_3202_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3204_cast_fp16 = add(x = var_3198_cast_fp16, y = var_3202_cast_fp16)[name = string("op_3204_cast_fp16")];
            fp32 var_3205_epsilon_0 = const()[name = string("op_3205_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3205_cast_fp16 = rsqrt(epsilon = var_3205_epsilon_0, x = var_3204_cast_fp16)[name = string("op_3205_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_61_cast_fp16 = mul(x = y_71_cast_fp16, y = var_3205_cast_fp16)[name = string("out_61_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47923072)))];
            tensor<fp16, [1, 630, 768]> input_53_cast_fp16 = mul(x = out_61_cast_fp16, y = t_encoder_layers_5_input_layernorm_weight_to_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(47924672))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48514560))))[name = string("t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_3215 = const()[name = string("op_3215"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_93_cast_fp16 = reshape(shape = var_3215, x = linear_36_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3217_cast_fp16 = abs(x = x_93_cast_fp16)[name = string("op_3217_cast_fp16")];
            bool var_3220 = const()[name = string("op_3220"), val = bool(true)];
            tensor<int32, [1]> var_3221_axes_0 = const()[name = string("op_3221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3221_cast_fp16 = reduce_max(axes = var_3221_axes_0, keep_dims = var_3220, x = var_3217_cast_fp16)[name = string("op_3221_cast_fp16")];
            fp16 var_3222_to_fp16 = const()[name = string("op_3222_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_77_cast_fp16 = maximum(x = var_3221_cast_fp16, y = var_3222_to_fp16)[name = string("s_77_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_73_cast_fp16 = real_div(x = x_93_cast_fp16, y = s_77_cast_fp16)[name = string("y_73_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3225_cast_fp16 = mul(x = y_73_cast_fp16, y = y_73_cast_fp16)[name = string("op_3225_cast_fp16")];
            tensor<int32, [1]> var_3230_axes_0 = const()[name = string("op_3230_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3230_keep_dims_0 = const()[name = string("op_3230_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3230_cast_fp16 = reduce_mean(axes = var_3230_axes_0, keep_dims = var_3230_keep_dims_0, x = var_3225_cast_fp16)[name = string("op_3230_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3231_cast_fp16 = mul(x = s_77_cast_fp16, y = s_77_cast_fp16)[name = string("op_3231_cast_fp16")];
            fp32 var_3232_epsilon_0 = const()[name = string("op_3232_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3232_cast_fp16 = inverse(epsilon = var_3232_epsilon_0, x = var_3231_cast_fp16)[name = string("op_3232_cast_fp16")];
            fp16 var_3233_to_fp16 = const()[name = string("op_3233_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3234_cast_fp16 = mul(x = var_3232_cast_fp16, y = var_3233_to_fp16)[name = string("op_3234_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3236_cast_fp16 = add(x = var_3230_cast_fp16, y = var_3234_cast_fp16)[name = string("op_3236_cast_fp16")];
            fp32 var_3237_epsilon_0 = const()[name = string("op_3237_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3237_cast_fp16 = rsqrt(epsilon = var_3237_epsilon_0, x = var_3236_cast_fp16)[name = string("op_3237_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_63_cast_fp16 = mul(x = y_73_cast_fp16, y = var_3237_cast_fp16)[name = string("out_63_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48516160)))];
            tensor<fp16, [1, 630, 12, 64]> x_99_cast_fp16 = mul(x = out_63_cast_fp16, y = t_encoder_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_99_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(48516352))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49106240))))[name = string("t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_37_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3247 = const()[name = string("op_3247"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_95_cast_fp16 = reshape(shape = var_3247, x = linear_37_cast_fp16)[name = string("x_95_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3249_cast_fp16 = abs(x = x_95_cast_fp16)[name = string("op_3249_cast_fp16")];
            bool var_3252 = const()[name = string("op_3252"), val = bool(true)];
            tensor<int32, [1]> var_3253_axes_0 = const()[name = string("op_3253_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3253_cast_fp16 = reduce_max(axes = var_3253_axes_0, keep_dims = var_3252, x = var_3249_cast_fp16)[name = string("op_3253_cast_fp16")];
            fp16 var_3254_to_fp16 = const()[name = string("op_3254_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_79_cast_fp16 = maximum(x = var_3253_cast_fp16, y = var_3254_to_fp16)[name = string("s_79_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_75_cast_fp16 = real_div(x = x_95_cast_fp16, y = s_79_cast_fp16)[name = string("y_75_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3257_cast_fp16 = mul(x = y_75_cast_fp16, y = y_75_cast_fp16)[name = string("op_3257_cast_fp16")];
            tensor<int32, [1]> var_3262_axes_0 = const()[name = string("op_3262_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3262_keep_dims_0 = const()[name = string("op_3262_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3262_cast_fp16 = reduce_mean(axes = var_3262_axes_0, keep_dims = var_3262_keep_dims_0, x = var_3257_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3263_cast_fp16 = mul(x = s_79_cast_fp16, y = s_79_cast_fp16)[name = string("op_3263_cast_fp16")];
            fp32 var_3264_epsilon_0 = const()[name = string("op_3264_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3264_cast_fp16 = inverse(epsilon = var_3264_epsilon_0, x = var_3263_cast_fp16)[name = string("op_3264_cast_fp16")];
            fp16 var_3265_to_fp16 = const()[name = string("op_3265_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3266_cast_fp16 = mul(x = var_3264_cast_fp16, y = var_3265_to_fp16)[name = string("op_3266_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3268_cast_fp16 = add(x = var_3262_cast_fp16, y = var_3266_cast_fp16)[name = string("op_3268_cast_fp16")];
            fp32 var_3269_epsilon_0 = const()[name = string("op_3269_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3269_cast_fp16 = rsqrt(epsilon = var_3269_epsilon_0, x = var_3268_cast_fp16)[name = string("op_3269_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_65_cast_fp16 = mul(x = y_75_cast_fp16, y = var_3269_cast_fp16)[name = string("out_65_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49107840)))];
            tensor<fp16, [1, 630, 12, 64]> x_101_cast_fp16 = mul(x = out_65_cast_fp16, y = t_encoder_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_101_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49108032))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49697920))))[name = string("t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<int32, [4]> var_3279 = const()[name = string("op_3279"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_97_cast_fp16 = reshape(shape = var_3279, x = linear_38_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3281_cast_fp16 = abs(x = x_97_cast_fp16)[name = string("op_3281_cast_fp16")];
            bool var_3284 = const()[name = string("op_3284"), val = bool(true)];
            tensor<int32, [1]> var_3285_axes_0 = const()[name = string("op_3285_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3285_cast_fp16 = reduce_max(axes = var_3285_axes_0, keep_dims = var_3284, x = var_3281_cast_fp16)[name = string("op_3285_cast_fp16")];
            fp16 var_3286_to_fp16 = const()[name = string("op_3286_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_81_cast_fp16 = maximum(x = var_3285_cast_fp16, y = var_3286_to_fp16)[name = string("s_81_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_77_cast_fp16 = real_div(x = x_97_cast_fp16, y = s_81_cast_fp16)[name = string("y_77_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3289_cast_fp16 = mul(x = y_77_cast_fp16, y = y_77_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<int32, [1]> var_3294_axes_0 = const()[name = string("op_3294_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3294_keep_dims_0 = const()[name = string("op_3294_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3294_cast_fp16 = reduce_mean(axes = var_3294_axes_0, keep_dims = var_3294_keep_dims_0, x = var_3289_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3295_cast_fp16 = mul(x = s_81_cast_fp16, y = s_81_cast_fp16)[name = string("op_3295_cast_fp16")];
            fp32 var_3296_epsilon_0 = const()[name = string("op_3296_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3296_cast_fp16 = inverse(epsilon = var_3296_epsilon_0, x = var_3295_cast_fp16)[name = string("op_3296_cast_fp16")];
            fp16 var_3297_to_fp16 = const()[name = string("op_3297_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3298_cast_fp16 = mul(x = var_3296_cast_fp16, y = var_3297_to_fp16)[name = string("op_3298_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3300_cast_fp16 = add(x = var_3294_cast_fp16, y = var_3298_cast_fp16)[name = string("op_3300_cast_fp16")];
            fp32 var_3301_epsilon_0 = const()[name = string("op_3301_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3301_cast_fp16 = rsqrt(epsilon = var_3301_epsilon_0, x = var_3300_cast_fp16)[name = string("op_3301_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_21_cast_fp16 = mul(x = y_77_cast_fp16, y = var_3301_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<int32, [4]> z_41_begin_0 = const()[name = string("z_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_41_end_0 = const()[name = string("z_41_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_41_end_mask_0 = const()[name = string("z_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_41_cast_fp16 = slice_by_index(begin = z_41_begin_0, end = z_41_end_0, end_mask = z_41_end_mask_0, x = x_99_cast_fp16)[name = string("z_41_cast_fp16")];
            tensor<int32, [4]> var_3312_begin_0 = const()[name = string("op_3312_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3312_end_0 = const()[name = string("op_3312_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3312_end_mask_0 = const()[name = string("op_3312_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3312_cast_fp16 = slice_by_index(begin = var_3312_begin_0, end = var_3312_end_0, end_mask = var_3312_end_mask_0, x = z_41_cast_fp16)[name = string("op_3312_cast_fp16")];
            fp16 const_20_promoted_to_fp16 = const()[name = string("const_20_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3313_cast_fp16 = mul(x = var_3312_cast_fp16, y = const_20_promoted_to_fp16)[name = string("op_3313_cast_fp16")];
            tensor<int32, [4]> var_3318_begin_0 = const()[name = string("op_3318_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3318_end_0 = const()[name = string("op_3318_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3318_end_mask_0 = const()[name = string("op_3318_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> 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 = z_41_cast_fp16)[name = string("op_3318_cast_fp16")];
            int32 var_3320 = const()[name = string("op_3320"), val = int32(-1)];
            bool rot_41_interleave_0 = const()[name = string("rot_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_41_cast_fp16 = concat(axis = var_3320, interleave = rot_41_interleave_0, values = (var_3313_cast_fp16, var_3318_cast_fp16))[name = string("rot_41_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3322_cast_fp16 = mul(x = z_41_cast_fp16, y = c_1_cast_fp16)[name = string("op_3322_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3323_cast_fp16 = mul(x = rot_41_cast_fp16, y = s_9_cast_fp16)[name = string("op_3323_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3325_cast_fp16 = add(x = var_3322_cast_fp16, y = var_3323_cast_fp16)[name = string("op_3325_cast_fp16")];
            tensor<int32, [4]> z_43_begin_0 = const()[name = string("z_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_43_end_0 = const()[name = string("z_43_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_43_end_mask_0 = const()[name = string("z_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_43_cast_fp16 = slice_by_index(begin = z_43_begin_0, end = z_43_end_0, end_mask = z_43_end_mask_0, x = x_99_cast_fp16)[name = string("z_43_cast_fp16")];
            tensor<int32, [4]> var_3335_begin_0 = const()[name = string("op_3335_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3335_end_0 = const()[name = string("op_3335_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3335_end_mask_0 = const()[name = string("op_3335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3335_cast_fp16 = slice_by_index(begin = var_3335_begin_0, end = var_3335_end_0, end_mask = var_3335_end_mask_0, x = z_43_cast_fp16)[name = string("op_3335_cast_fp16")];
            fp16 const_21_promoted_to_fp16 = const()[name = string("const_21_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3336_cast_fp16 = mul(x = var_3335_cast_fp16, y = const_21_promoted_to_fp16)[name = string("op_3336_cast_fp16")];
            tensor<int32, [4]> var_3341_begin_0 = const()[name = string("op_3341_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3341_end_0 = const()[name = string("op_3341_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3341_end_mask_0 = const()[name = string("op_3341_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3341_cast_fp16 = slice_by_index(begin = var_3341_begin_0, end = var_3341_end_0, end_mask = var_3341_end_mask_0, x = z_43_cast_fp16)[name = string("op_3341_cast_fp16")];
            int32 var_3343 = const()[name = string("op_3343"), val = int32(-1)];
            bool rot_43_interleave_0 = const()[name = string("rot_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_43_cast_fp16 = concat(axis = var_3343, interleave = rot_43_interleave_0, values = (var_3336_cast_fp16, var_3341_cast_fp16))[name = string("rot_43_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3345_cast_fp16 = mul(x = z_43_cast_fp16, y = c_cast_fp16)[name = string("op_3345_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3346_cast_fp16 = mul(x = rot_43_cast_fp16, y = s_11_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3348_cast_fp16 = add(x = var_3345_cast_fp16, y = var_3346_cast_fp16)[name = string("op_3348_cast_fp16")];
            int32 var_3350 = const()[name = string("op_3350"), val = int32(-1)];
            bool var_3351_interleave_0 = const()[name = string("op_3351_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_3351_cast_fp16 = concat(axis = var_3350, interleave = var_3351_interleave_0, values = (var_3325_cast_fp16, var_3348_cast_fp16))[name = string("op_3351_cast_fp16")];
            tensor<int32, [4]> z_45_begin_0 = const()[name = string("z_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_45_end_0 = const()[name = string("z_45_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_45_end_mask_0 = const()[name = string("z_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_45_cast_fp16 = slice_by_index(begin = z_45_begin_0, end = z_45_end_0, end_mask = z_45_end_mask_0, x = x_101_cast_fp16)[name = string("z_45_cast_fp16")];
            tensor<int32, [4]> var_3364_begin_0 = const()[name = string("op_3364_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3364_end_0 = const()[name = string("op_3364_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3364_end_mask_0 = const()[name = string("op_3364_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3364_cast_fp16 = slice_by_index(begin = var_3364_begin_0, end = var_3364_end_0, end_mask = var_3364_end_mask_0, x = z_45_cast_fp16)[name = string("op_3364_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3365_cast_fp16 = mul(x = var_3364_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_3365_cast_fp16")];
            tensor<int32, [4]> var_3370_begin_0 = const()[name = string("op_3370_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3370_end_0 = const()[name = string("op_3370_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3370_end_mask_0 = const()[name = string("op_3370_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3370_cast_fp16 = slice_by_index(begin = var_3370_begin_0, end = var_3370_end_0, end_mask = var_3370_end_mask_0, x = z_45_cast_fp16)[name = string("op_3370_cast_fp16")];
            int32 var_3372 = const()[name = string("op_3372"), val = int32(-1)];
            bool rot_45_interleave_0 = const()[name = string("rot_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_45_cast_fp16 = concat(axis = var_3372, interleave = rot_45_interleave_0, values = (var_3365_cast_fp16, var_3370_cast_fp16))[name = string("rot_45_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3374_cast_fp16 = mul(x = z_45_cast_fp16, y = c_1_cast_fp16)[name = string("op_3374_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3375_cast_fp16 = mul(x = rot_45_cast_fp16, y = s_9_cast_fp16)[name = string("op_3375_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3377_cast_fp16 = add(x = var_3374_cast_fp16, y = var_3375_cast_fp16)[name = string("op_3377_cast_fp16")];
            tensor<int32, [4]> z_47_begin_0 = const()[name = string("z_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_47_end_0 = const()[name = string("z_47_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_47_end_mask_0 = const()[name = string("z_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_47_cast_fp16 = slice_by_index(begin = z_47_begin_0, end = z_47_end_0, end_mask = z_47_end_mask_0, x = x_101_cast_fp16)[name = string("z_47_cast_fp16")];
            tensor<int32, [4]> var_3387_begin_0 = const()[name = string("op_3387_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3387_end_0 = const()[name = string("op_3387_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3387_end_mask_0 = const()[name = string("op_3387_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3387_cast_fp16 = slice_by_index(begin = var_3387_begin_0, end = var_3387_end_0, end_mask = var_3387_end_mask_0, x = z_47_cast_fp16)[name = string("op_3387_cast_fp16")];
            fp16 const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3388_cast_fp16 = mul(x = var_3387_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_3388_cast_fp16")];
            tensor<int32, [4]> var_3393_begin_0 = const()[name = string("op_3393_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3393_end_0 = const()[name = string("op_3393_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3393_end_mask_0 = const()[name = string("op_3393_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3393_cast_fp16 = slice_by_index(begin = var_3393_begin_0, end = var_3393_end_0, end_mask = var_3393_end_mask_0, x = z_47_cast_fp16)[name = string("op_3393_cast_fp16")];
            int32 var_3395 = const()[name = string("op_3395"), val = int32(-1)];
            bool rot_47_interleave_0 = const()[name = string("rot_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_47_cast_fp16 = concat(axis = var_3395, interleave = rot_47_interleave_0, values = (var_3388_cast_fp16, var_3393_cast_fp16))[name = string("rot_47_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3397_cast_fp16 = mul(x = z_47_cast_fp16, y = c_cast_fp16)[name = string("op_3397_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3398_cast_fp16 = mul(x = rot_47_cast_fp16, y = s_11_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3400_cast_fp16 = add(x = var_3397_cast_fp16, y = var_3398_cast_fp16)[name = string("op_3400_cast_fp16")];
            int32 var_3402 = const()[name = string("op_3402"), val = int32(-1)];
            bool var_3403_interleave_0 = const()[name = string("op_3403_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_3403_cast_fp16 = concat(axis = var_3402, interleave = var_3403_interleave_0, values = (var_3377_cast_fp16, var_3400_cast_fp16))[name = string("op_3403_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3413_transpose_x_0 = const()[name = string("op_3413_transpose_x_0"), val = bool(false)];
            bool var_3413_transpose_y_0 = const()[name = string("op_3413_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_3403_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 12, 630, 64]> transpose_74 = transpose(perm = transpose_74_perm_0, x = var_3351_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 12, 630, 630]> var_3413_cast_fp16 = matmul(transpose_x = var_3413_transpose_x_0, transpose_y = var_3413_transpose_y_0, x = transpose_74, y = transpose_75)[name = string("op_3413_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_3415_cast_fp16 = add(x = var_3413_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3415_cast_fp16")];
            int32 var_3416 = const()[name = string("op_3416"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_77_cast_fp16 = softmax(axis = var_3416, x = var_3415_cast_fp16)[name = string("w_77_cast_fp16")];
            bool var_3419_transpose_x_0 = const()[name = string("op_3419_transpose_x_0"), val = bool(false)];
            bool var_3419_transpose_y_0 = const()[name = string("op_3419_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = v_21_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 12, 630, 64]> var_3419_cast_fp16 = matmul(transpose_x = var_3419_transpose_x_0, transpose_y = var_3419_transpose_y_0, x = w_77_cast_fp16, y = v_23_cast_fp16)[name = string("op_3419_cast_fp16")];
            tensor<int32, [4]> var_3422_perm_0 = const()[name = string("op_3422_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3426 = const()[name = string("op_3426"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_3422_cast_fp16 = transpose(perm = var_3422_perm_0, x = var_3419_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 630, 768]> input_55_cast_fp16 = reshape(shape = var_3426, x = var_3422_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(49699520))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50289408))))[name = string("t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_39_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3431_cast_fp16 = abs(x = linear_39_cast_fp16)[name = string("op_3431_cast_fp16")];
            bool var_3434 = const()[name = string("op_3434"), val = bool(true)];
            tensor<int32, [1]> var_3435_axes_0 = const()[name = string("op_3435_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3435_cast_fp16 = reduce_max(axes = var_3435_axes_0, keep_dims = var_3434, x = var_3431_cast_fp16)[name = string("op_3435_cast_fp16")];
            fp16 var_3436_to_fp16 = const()[name = string("op_3436_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_83_cast_fp16 = maximum(x = var_3435_cast_fp16, y = var_3436_to_fp16)[name = string("s_83_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_79_cast_fp16 = real_div(x = linear_39_cast_fp16, y = s_83_cast_fp16)[name = string("y_79_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3439_cast_fp16 = mul(x = y_79_cast_fp16, y = y_79_cast_fp16)[name = string("op_3439_cast_fp16")];
            tensor<int32, [1]> var_3444_axes_0 = const()[name = string("op_3444_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3444_keep_dims_0 = const()[name = string("op_3444_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3444_cast_fp16 = reduce_mean(axes = var_3444_axes_0, keep_dims = var_3444_keep_dims_0, x = var_3439_cast_fp16)[name = string("op_3444_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3445_cast_fp16 = mul(x = s_83_cast_fp16, y = s_83_cast_fp16)[name = string("op_3445_cast_fp16")];
            fp32 var_3446_epsilon_0 = const()[name = string("op_3446_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3446_cast_fp16 = inverse(epsilon = var_3446_epsilon_0, x = var_3445_cast_fp16)[name = string("op_3446_cast_fp16")];
            fp16 var_3447_to_fp16 = const()[name = string("op_3447_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3448_cast_fp16 = mul(x = var_3446_cast_fp16, y = var_3447_to_fp16)[name = string("op_3448_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3450_cast_fp16 = add(x = var_3444_cast_fp16, y = var_3448_cast_fp16)[name = string("op_3450_cast_fp16")];
            fp32 var_3451_epsilon_0 = const()[name = string("op_3451_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3451_cast_fp16 = rsqrt(epsilon = var_3451_epsilon_0, x = var_3450_cast_fp16)[name = string("op_3451_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_67_cast_fp16 = mul(x = y_79_cast_fp16, y = var_3451_cast_fp16)[name = string("out_67_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50291008)))];
            tensor<fp16, [1, 630, 768]> var_3453_cast_fp16 = mul(x = out_67_cast_fp16, y = t_encoder_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("op_3453_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_105_cast_fp16 = add(x = x_91_cast_fp16, y = var_3453_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3456_cast_fp16 = abs(x = x_105_cast_fp16)[name = string("op_3456_cast_fp16")];
            bool var_3459 = const()[name = string("op_3459"), val = bool(true)];
            tensor<int32, [1]> var_3460_axes_0 = const()[name = string("op_3460_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3460_cast_fp16 = reduce_max(axes = var_3460_axes_0, keep_dims = var_3459, x = var_3456_cast_fp16)[name = string("op_3460_cast_fp16")];
            fp16 var_3461_to_fp16 = const()[name = string("op_3461_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_85_cast_fp16 = maximum(x = var_3460_cast_fp16, y = var_3461_to_fp16)[name = string("s_85_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_81_cast_fp16 = real_div(x = x_105_cast_fp16, y = s_85_cast_fp16)[name = string("y_81_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3464_cast_fp16 = mul(x = y_81_cast_fp16, y = y_81_cast_fp16)[name = string("op_3464_cast_fp16")];
            tensor<int32, [1]> var_3469_axes_0 = const()[name = string("op_3469_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3469_keep_dims_0 = const()[name = string("op_3469_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3469_cast_fp16 = reduce_mean(axes = var_3469_axes_0, keep_dims = var_3469_keep_dims_0, x = var_3464_cast_fp16)[name = string("op_3469_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3470_cast_fp16 = mul(x = s_85_cast_fp16, y = s_85_cast_fp16)[name = string("op_3470_cast_fp16")];
            fp32 var_3471_epsilon_0 = const()[name = string("op_3471_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3471_cast_fp16 = inverse(epsilon = var_3471_epsilon_0, x = var_3470_cast_fp16)[name = string("op_3471_cast_fp16")];
            fp16 var_3472_to_fp16 = const()[name = string("op_3472_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3473_cast_fp16 = mul(x = var_3471_cast_fp16, y = var_3472_to_fp16)[name = string("op_3473_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3475_cast_fp16 = add(x = var_3469_cast_fp16, y = var_3473_cast_fp16)[name = string("op_3475_cast_fp16")];
            fp32 var_3476_epsilon_0 = const()[name = string("op_3476_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3476_cast_fp16 = rsqrt(epsilon = var_3476_epsilon_0, x = var_3475_cast_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_69_cast_fp16 = mul(x = y_81_cast_fp16, y = var_3476_cast_fp16)[name = string("out_69_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50292608)))];
            tensor<fp16, [1, 630, 768]> input_57_cast_fp16 = mul(x = out_69_cast_fp16, y = t_encoder_layers_5_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_57_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50294208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52653568))))[name = string("t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_40_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_40_cast_fp16")];
            string var_3483_mode_0 = const()[name = string("op_3483_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_3483_cast_fp16 = gelu(mode = var_3483_mode_0, x = linear_40_cast_fp16)[name = string("op_3483_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52659776))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55019136))))[name = string("t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_41_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_57_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_61_cast_fp16 = mul(x = var_3483_cast_fp16, y = linear_41_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(55025344))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57384704))))[name = string("t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_5_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3491_cast_fp16 = abs(x = linear_42_cast_fp16)[name = string("op_3491_cast_fp16")];
            bool var_3494 = const()[name = string("op_3494"), val = bool(true)];
            tensor<int32, [1]> var_3495_axes_0 = const()[name = string("op_3495_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3495_cast_fp16 = reduce_max(axes = var_3495_axes_0, keep_dims = var_3494, x = var_3491_cast_fp16)[name = string("op_3495_cast_fp16")];
            fp16 var_3496_to_fp16 = const()[name = string("op_3496_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_87_cast_fp16 = maximum(x = var_3495_cast_fp16, y = var_3496_to_fp16)[name = string("s_87_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_83_cast_fp16 = real_div(x = linear_42_cast_fp16, y = s_87_cast_fp16)[name = string("y_83_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3499_cast_fp16 = mul(x = y_83_cast_fp16, y = y_83_cast_fp16)[name = string("op_3499_cast_fp16")];
            tensor<int32, [1]> var_3504_axes_0 = const()[name = string("op_3504_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3504_keep_dims_0 = const()[name = string("op_3504_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3504_cast_fp16 = reduce_mean(axes = var_3504_axes_0, keep_dims = var_3504_keep_dims_0, x = var_3499_cast_fp16)[name = string("op_3504_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3505_cast_fp16 = mul(x = s_87_cast_fp16, y = s_87_cast_fp16)[name = string("op_3505_cast_fp16")];
            fp32 var_3506_epsilon_0 = const()[name = string("op_3506_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3506_cast_fp16 = inverse(epsilon = var_3506_epsilon_0, x = var_3505_cast_fp16)[name = string("op_3506_cast_fp16")];
            fp16 var_3507_to_fp16 = const()[name = string("op_3507_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3508_cast_fp16 = mul(x = var_3506_cast_fp16, y = var_3507_to_fp16)[name = string("op_3508_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3510_cast_fp16 = add(x = var_3504_cast_fp16, y = var_3508_cast_fp16)[name = string("op_3510_cast_fp16")];
            fp32 var_3511_epsilon_0 = const()[name = string("op_3511_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3511_cast_fp16 = rsqrt(epsilon = var_3511_epsilon_0, x = var_3510_cast_fp16)[name = string("op_3511_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_71_cast_fp16 = mul(x = y_83_cast_fp16, y = var_3511_cast_fp16)[name = string("out_71_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57386304)))];
            tensor<fp16, [1, 630, 768]> var_3513_cast_fp16 = mul(x = out_71_cast_fp16, y = t_encoder_layers_5_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3513_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_109_cast_fp16 = add(x = x_105_cast_fp16, y = var_3513_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3516_cast_fp16 = abs(x = x_109_cast_fp16)[name = string("op_3516_cast_fp16")];
            bool var_3519 = const()[name = string("op_3519"), val = bool(true)];
            tensor<int32, [1]> var_3520_axes_0 = const()[name = string("op_3520_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3520_cast_fp16 = reduce_max(axes = var_3520_axes_0, keep_dims = var_3519, x = var_3516_cast_fp16)[name = string("op_3520_cast_fp16")];
            fp16 var_3521_to_fp16 = const()[name = string("op_3521_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_89_cast_fp16 = maximum(x = var_3520_cast_fp16, y = var_3521_to_fp16)[name = string("s_89_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_85_cast_fp16 = real_div(x = x_109_cast_fp16, y = s_89_cast_fp16)[name = string("y_85_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3524_cast_fp16 = mul(x = y_85_cast_fp16, y = y_85_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<int32, [1]> var_3529_axes_0 = const()[name = string("op_3529_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3529_keep_dims_0 = const()[name = string("op_3529_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3529_cast_fp16 = reduce_mean(axes = var_3529_axes_0, keep_dims = var_3529_keep_dims_0, x = var_3524_cast_fp16)[name = string("op_3529_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3530_cast_fp16 = mul(x = s_89_cast_fp16, y = s_89_cast_fp16)[name = string("op_3530_cast_fp16")];
            fp32 var_3531_epsilon_0 = const()[name = string("op_3531_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3531_cast_fp16 = inverse(epsilon = var_3531_epsilon_0, x = var_3530_cast_fp16)[name = string("op_3531_cast_fp16")];
            fp16 var_3532_to_fp16 = const()[name = string("op_3532_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3533_cast_fp16 = mul(x = var_3531_cast_fp16, y = var_3532_to_fp16)[name = string("op_3533_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3535_cast_fp16 = add(x = var_3529_cast_fp16, y = var_3533_cast_fp16)[name = string("op_3535_cast_fp16")];
            fp32 var_3536_epsilon_0 = const()[name = string("op_3536_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3536_cast_fp16 = rsqrt(epsilon = var_3536_epsilon_0, x = var_3535_cast_fp16)[name = string("op_3536_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_73_cast_fp16 = mul(x = y_85_cast_fp16, y = var_3536_cast_fp16)[name = string("out_73_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57387904)))];
            tensor<fp16, [1, 630, 768]> input_63_cast_fp16 = mul(x = out_73_cast_fp16, y = t_encoder_layers_6_input_layernorm_weight_to_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57389504))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57979392))))[name = string("t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_43_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_3546 = const()[name = string("op_3546"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_111_cast_fp16 = reshape(shape = var_3546, x = linear_43_cast_fp16)[name = string("x_111_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3548_cast_fp16 = abs(x = x_111_cast_fp16)[name = string("op_3548_cast_fp16")];
            bool var_3551 = const()[name = string("op_3551"), val = bool(true)];
            tensor<int32, [1]> var_3552_axes_0 = const()[name = string("op_3552_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3552_cast_fp16 = reduce_max(axes = var_3552_axes_0, keep_dims = var_3551, x = var_3548_cast_fp16)[name = string("op_3552_cast_fp16")];
            fp16 var_3553_to_fp16 = const()[name = string("op_3553_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_91_cast_fp16 = maximum(x = var_3552_cast_fp16, y = var_3553_to_fp16)[name = string("s_91_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_87_cast_fp16 = real_div(x = x_111_cast_fp16, y = s_91_cast_fp16)[name = string("y_87_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3556_cast_fp16 = mul(x = y_87_cast_fp16, y = y_87_cast_fp16)[name = string("op_3556_cast_fp16")];
            tensor<int32, [1]> var_3561_axes_0 = const()[name = string("op_3561_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3561_keep_dims_0 = const()[name = string("op_3561_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3561_cast_fp16 = reduce_mean(axes = var_3561_axes_0, keep_dims = var_3561_keep_dims_0, x = var_3556_cast_fp16)[name = string("op_3561_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3562_cast_fp16 = mul(x = s_91_cast_fp16, y = s_91_cast_fp16)[name = string("op_3562_cast_fp16")];
            fp32 var_3563_epsilon_0 = const()[name = string("op_3563_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3563_cast_fp16 = inverse(epsilon = var_3563_epsilon_0, x = var_3562_cast_fp16)[name = string("op_3563_cast_fp16")];
            fp16 var_3564_to_fp16 = const()[name = string("op_3564_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3565_cast_fp16 = mul(x = var_3563_cast_fp16, y = var_3564_to_fp16)[name = string("op_3565_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3567_cast_fp16 = add(x = var_3561_cast_fp16, y = var_3565_cast_fp16)[name = string("op_3567_cast_fp16")];
            fp32 var_3568_epsilon_0 = const()[name = string("op_3568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3568_cast_fp16 = rsqrt(epsilon = var_3568_epsilon_0, x = var_3567_cast_fp16)[name = string("op_3568_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_75_cast_fp16 = mul(x = y_87_cast_fp16, y = var_3568_cast_fp16)[name = string("out_75_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 630, 12, 64]> x_117_cast_fp16 = mul(x = out_75_cast_fp16, y = t_encoder_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_117_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(57980992))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58570880))))[name = string("t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_3578 = const()[name = string("op_3578"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_113_cast_fp16 = reshape(shape = var_3578, x = linear_44_cast_fp16)[name = string("x_113_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3580_cast_fp16 = abs(x = x_113_cast_fp16)[name = string("op_3580_cast_fp16")];
            bool var_3583 = const()[name = string("op_3583"), val = bool(true)];
            tensor<int32, [1]> var_3584_axes_0 = const()[name = string("op_3584_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3584_cast_fp16 = reduce_max(axes = var_3584_axes_0, keep_dims = var_3583, x = var_3580_cast_fp16)[name = string("op_3584_cast_fp16")];
            fp16 var_3585_to_fp16 = const()[name = string("op_3585_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_93_cast_fp16 = maximum(x = var_3584_cast_fp16, y = var_3585_to_fp16)[name = string("s_93_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_89_cast_fp16 = real_div(x = x_113_cast_fp16, y = s_93_cast_fp16)[name = string("y_89_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3588_cast_fp16 = mul(x = y_89_cast_fp16, y = y_89_cast_fp16)[name = string("op_3588_cast_fp16")];
            tensor<int32, [1]> var_3593_axes_0 = const()[name = string("op_3593_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3593_keep_dims_0 = const()[name = string("op_3593_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3593_cast_fp16 = reduce_mean(axes = var_3593_axes_0, keep_dims = var_3593_keep_dims_0, x = var_3588_cast_fp16)[name = string("op_3593_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3594_cast_fp16 = mul(x = s_93_cast_fp16, y = s_93_cast_fp16)[name = string("op_3594_cast_fp16")];
            fp32 var_3595_epsilon_0 = const()[name = string("op_3595_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3595_cast_fp16 = inverse(epsilon = var_3595_epsilon_0, x = var_3594_cast_fp16)[name = string("op_3595_cast_fp16")];
            fp16 var_3596_to_fp16 = const()[name = string("op_3596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3597_cast_fp16 = mul(x = var_3595_cast_fp16, y = var_3596_to_fp16)[name = string("op_3597_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3599_cast_fp16 = add(x = var_3593_cast_fp16, y = var_3597_cast_fp16)[name = string("op_3599_cast_fp16")];
            fp32 var_3600_epsilon_0 = const()[name = string("op_3600_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3600_cast_fp16 = rsqrt(epsilon = var_3600_epsilon_0, x = var_3599_cast_fp16)[name = string("op_3600_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_77_cast_fp16 = mul(x = y_89_cast_fp16, y = var_3600_cast_fp16)[name = string("out_77_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58572480)))];
            tensor<fp16, [1, 630, 12, 64]> x_119_cast_fp16 = mul(x = out_77_cast_fp16, y = t_encoder_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_119_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58572672))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59162560))))[name = string("t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<int32, [4]> var_3610 = const()[name = string("op_3610"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_115_cast_fp16 = reshape(shape = var_3610, x = linear_45_cast_fp16)[name = string("x_115_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3612_cast_fp16 = abs(x = x_115_cast_fp16)[name = string("op_3612_cast_fp16")];
            bool var_3615 = const()[name = string("op_3615"), val = bool(true)];
            tensor<int32, [1]> var_3616_axes_0 = const()[name = string("op_3616_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3616_cast_fp16 = reduce_max(axes = var_3616_axes_0, keep_dims = var_3615, x = var_3612_cast_fp16)[name = string("op_3616_cast_fp16")];
            fp16 var_3617_to_fp16 = const()[name = string("op_3617_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_95_cast_fp16 = maximum(x = var_3616_cast_fp16, y = var_3617_to_fp16)[name = string("s_95_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_91_cast_fp16 = real_div(x = x_115_cast_fp16, y = s_95_cast_fp16)[name = string("y_91_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3620_cast_fp16 = mul(x = y_91_cast_fp16, y = y_91_cast_fp16)[name = string("op_3620_cast_fp16")];
            tensor<int32, [1]> var_3625_axes_0 = const()[name = string("op_3625_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3625_keep_dims_0 = const()[name = string("op_3625_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3625_cast_fp16 = reduce_mean(axes = var_3625_axes_0, keep_dims = var_3625_keep_dims_0, x = var_3620_cast_fp16)[name = string("op_3625_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3626_cast_fp16 = mul(x = s_95_cast_fp16, y = s_95_cast_fp16)[name = string("op_3626_cast_fp16")];
            fp32 var_3627_epsilon_0 = const()[name = string("op_3627_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3627_cast_fp16 = inverse(epsilon = var_3627_epsilon_0, x = var_3626_cast_fp16)[name = string("op_3627_cast_fp16")];
            fp16 var_3628_to_fp16 = const()[name = string("op_3628_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3629_cast_fp16 = mul(x = var_3627_cast_fp16, y = var_3628_to_fp16)[name = string("op_3629_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3631_cast_fp16 = add(x = var_3625_cast_fp16, y = var_3629_cast_fp16)[name = string("op_3631_cast_fp16")];
            fp32 var_3632_epsilon_0 = const()[name = string("op_3632_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3632_cast_fp16 = rsqrt(epsilon = var_3632_epsilon_0, x = var_3631_cast_fp16)[name = string("op_3632_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_25_cast_fp16 = mul(x = y_91_cast_fp16, y = var_3632_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<int32, [4]> z_49_begin_0 = const()[name = string("z_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_49_end_0 = const()[name = string("z_49_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_49_end_mask_0 = const()[name = string("z_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_49_cast_fp16 = slice_by_index(begin = z_49_begin_0, end = z_49_end_0, end_mask = z_49_end_mask_0, x = x_117_cast_fp16)[name = string("z_49_cast_fp16")];
            tensor<int32, [4]> var_3643_begin_0 = const()[name = string("op_3643_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3643_end_0 = const()[name = string("op_3643_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3643_end_mask_0 = const()[name = string("op_3643_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3643_cast_fp16 = slice_by_index(begin = var_3643_begin_0, end = var_3643_end_0, end_mask = var_3643_end_mask_0, x = z_49_cast_fp16)[name = string("op_3643_cast_fp16")];
            fp16 const_24_promoted_to_fp16 = const()[name = string("const_24_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3644_cast_fp16 = mul(x = var_3643_cast_fp16, y = const_24_promoted_to_fp16)[name = string("op_3644_cast_fp16")];
            tensor<int32, [4]> var_3649_begin_0 = const()[name = string("op_3649_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3649_end_0 = const()[name = string("op_3649_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3649_end_mask_0 = const()[name = string("op_3649_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3649_cast_fp16 = slice_by_index(begin = var_3649_begin_0, end = var_3649_end_0, end_mask = var_3649_end_mask_0, x = z_49_cast_fp16)[name = string("op_3649_cast_fp16")];
            int32 var_3651 = const()[name = string("op_3651"), val = int32(-1)];
            bool rot_49_interleave_0 = const()[name = string("rot_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_49_cast_fp16 = concat(axis = var_3651, interleave = rot_49_interleave_0, values = (var_3644_cast_fp16, var_3649_cast_fp16))[name = string("rot_49_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3653_cast_fp16 = mul(x = z_49_cast_fp16, y = c_1_cast_fp16)[name = string("op_3653_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3654_cast_fp16 = mul(x = rot_49_cast_fp16, y = s_9_cast_fp16)[name = string("op_3654_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3656_cast_fp16 = add(x = var_3653_cast_fp16, y = var_3654_cast_fp16)[name = string("op_3656_cast_fp16")];
            tensor<int32, [4]> z_51_begin_0 = const()[name = string("z_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_51_end_0 = const()[name = string("z_51_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_51_end_mask_0 = const()[name = string("z_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_51_cast_fp16 = slice_by_index(begin = z_51_begin_0, end = z_51_end_0, end_mask = z_51_end_mask_0, x = x_117_cast_fp16)[name = string("z_51_cast_fp16")];
            tensor<int32, [4]> var_3666_begin_0 = const()[name = string("op_3666_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3666_end_0 = const()[name = string("op_3666_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3666_end_mask_0 = const()[name = string("op_3666_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3666_cast_fp16 = slice_by_index(begin = var_3666_begin_0, end = var_3666_end_0, end_mask = var_3666_end_mask_0, x = z_51_cast_fp16)[name = string("op_3666_cast_fp16")];
            fp16 const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3667_cast_fp16 = mul(x = var_3666_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_3667_cast_fp16")];
            tensor<int32, [4]> var_3672_begin_0 = const()[name = string("op_3672_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3672_end_0 = const()[name = string("op_3672_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3672_end_mask_0 = const()[name = string("op_3672_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3672_cast_fp16 = slice_by_index(begin = var_3672_begin_0, end = var_3672_end_0, end_mask = var_3672_end_mask_0, x = z_51_cast_fp16)[name = string("op_3672_cast_fp16")];
            int32 var_3674 = const()[name = string("op_3674"), val = int32(-1)];
            bool rot_51_interleave_0 = const()[name = string("rot_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_51_cast_fp16 = concat(axis = var_3674, interleave = rot_51_interleave_0, values = (var_3667_cast_fp16, var_3672_cast_fp16))[name = string("rot_51_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3676_cast_fp16 = mul(x = z_51_cast_fp16, y = c_cast_fp16)[name = string("op_3676_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3677_cast_fp16 = mul(x = rot_51_cast_fp16, y = s_11_cast_fp16)[name = string("op_3677_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3679_cast_fp16 = add(x = var_3676_cast_fp16, y = var_3677_cast_fp16)[name = string("op_3679_cast_fp16")];
            int32 var_3681 = const()[name = string("op_3681"), val = int32(-1)];
            bool var_3682_interleave_0 = const()[name = string("op_3682_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_3682_cast_fp16 = concat(axis = var_3681, interleave = var_3682_interleave_0, values = (var_3656_cast_fp16, var_3679_cast_fp16))[name = string("op_3682_cast_fp16")];
            tensor<int32, [4]> z_53_begin_0 = const()[name = string("z_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_53_end_0 = const()[name = string("z_53_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_53_end_mask_0 = const()[name = string("z_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_53_cast_fp16 = slice_by_index(begin = z_53_begin_0, end = z_53_end_0, end_mask = z_53_end_mask_0, x = x_119_cast_fp16)[name = string("z_53_cast_fp16")];
            tensor<int32, [4]> var_3695_begin_0 = const()[name = string("op_3695_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3695_end_0 = const()[name = string("op_3695_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3695_end_mask_0 = const()[name = string("op_3695_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3695_cast_fp16 = slice_by_index(begin = var_3695_begin_0, end = var_3695_end_0, end_mask = var_3695_end_mask_0, x = z_53_cast_fp16)[name = string("op_3695_cast_fp16")];
            fp16 const_26_promoted_to_fp16 = const()[name = string("const_26_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3696_cast_fp16 = mul(x = var_3695_cast_fp16, y = const_26_promoted_to_fp16)[name = string("op_3696_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, 630, 12, 16])];
            tensor<bool, [4]> var_3701_end_mask_0 = const()[name = string("op_3701_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> 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 = z_53_cast_fp16)[name = string("op_3701_cast_fp16")];
            int32 var_3703 = const()[name = string("op_3703"), val = int32(-1)];
            bool rot_53_interleave_0 = const()[name = string("rot_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_53_cast_fp16 = concat(axis = var_3703, interleave = rot_53_interleave_0, values = (var_3696_cast_fp16, var_3701_cast_fp16))[name = string("rot_53_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3705_cast_fp16 = mul(x = z_53_cast_fp16, y = c_1_cast_fp16)[name = string("op_3705_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3706_cast_fp16 = mul(x = rot_53_cast_fp16, y = s_9_cast_fp16)[name = string("op_3706_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3708_cast_fp16 = add(x = var_3705_cast_fp16, y = var_3706_cast_fp16)[name = string("op_3708_cast_fp16")];
            tensor<int32, [4]> z_55_begin_0 = const()[name = string("z_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_55_end_0 = const()[name = string("z_55_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_55_end_mask_0 = const()[name = string("z_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_55_cast_fp16 = slice_by_index(begin = z_55_begin_0, end = z_55_end_0, end_mask = z_55_end_mask_0, x = x_119_cast_fp16)[name = string("z_55_cast_fp16")];
            tensor<int32, [4]> var_3718_begin_0 = const()[name = string("op_3718_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3718_end_0 = const()[name = string("op_3718_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3718_end_mask_0 = const()[name = string("op_3718_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3718_cast_fp16 = slice_by_index(begin = var_3718_begin_0, end = var_3718_end_0, end_mask = var_3718_end_mask_0, x = z_55_cast_fp16)[name = string("op_3718_cast_fp16")];
            fp16 const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3719_cast_fp16 = mul(x = var_3718_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_3719_cast_fp16")];
            tensor<int32, [4]> var_3724_begin_0 = const()[name = string("op_3724_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3724_end_0 = const()[name = string("op_3724_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3724_end_mask_0 = const()[name = string("op_3724_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, x = z_55_cast_fp16)[name = string("op_3724_cast_fp16")];
            int32 var_3726 = const()[name = string("op_3726"), val = int32(-1)];
            bool rot_55_interleave_0 = const()[name = string("rot_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_55_cast_fp16 = concat(axis = var_3726, interleave = rot_55_interleave_0, values = (var_3719_cast_fp16, var_3724_cast_fp16))[name = string("rot_55_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3728_cast_fp16 = mul(x = z_55_cast_fp16, y = c_cast_fp16)[name = string("op_3728_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3729_cast_fp16 = mul(x = rot_55_cast_fp16, y = s_11_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3731_cast_fp16 = add(x = var_3728_cast_fp16, y = var_3729_cast_fp16)[name = string("op_3731_cast_fp16")];
            int32 var_3733 = const()[name = string("op_3733"), val = int32(-1)];
            bool var_3734_interleave_0 = const()[name = string("op_3734_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_3734_cast_fp16 = concat(axis = var_3733, interleave = var_3734_interleave_0, values = (var_3708_cast_fp16, var_3731_cast_fp16))[name = string("op_3734_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_3744_transpose_x_0 = const()[name = string("op_3744_transpose_x_0"), val = bool(false)];
            bool var_3744_transpose_y_0 = const()[name = string("op_3744_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_3734_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 12, 630, 64]> transpose_76 = transpose(perm = transpose_76_perm_0, x = var_3682_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 12, 630, 630]> var_3744_cast_fp16 = matmul(transpose_x = var_3744_transpose_x_0, transpose_y = var_3744_transpose_y_0, x = transpose_76, y = transpose_77)[name = string("op_3744_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_3746_cast_fp16 = add(x = var_3744_cast_fp16, y = var_1524_cast_fp16)[name = string("op_3746_cast_fp16")];
            int32 var_3747 = const()[name = string("op_3747"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_91_cast_fp16 = softmax(axis = var_3747, x = var_3746_cast_fp16)[name = string("w_91_cast_fp16")];
            bool var_3750_transpose_x_0 = const()[name = string("op_3750_transpose_x_0"), val = bool(false)];
            bool var_3750_transpose_y_0 = const()[name = string("op_3750_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = v_25_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 12, 630, 64]> var_3750_cast_fp16 = matmul(transpose_x = var_3750_transpose_x_0, transpose_y = var_3750_transpose_y_0, x = w_91_cast_fp16, y = v_27_cast_fp16)[name = string("op_3750_cast_fp16")];
            tensor<int32, [4]> var_3753_perm_0 = const()[name = string("op_3753_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3757 = const()[name = string("op_3757"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_3753_cast_fp16 = transpose(perm = var_3753_perm_0, x = var_3750_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 630, 768]> input_65_cast_fp16 = reshape(shape = var_3757, x = var_3753_cast_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59164160))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59754048))))[name = string("t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3762_cast_fp16 = abs(x = linear_46_cast_fp16)[name = string("op_3762_cast_fp16")];
            bool var_3765 = const()[name = string("op_3765"), val = bool(true)];
            tensor<int32, [1]> var_3766_axes_0 = const()[name = string("op_3766_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3766_cast_fp16 = reduce_max(axes = var_3766_axes_0, keep_dims = var_3765, x = var_3762_cast_fp16)[name = string("op_3766_cast_fp16")];
            fp16 var_3767_to_fp16 = const()[name = string("op_3767_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_97_cast_fp16 = maximum(x = var_3766_cast_fp16, y = var_3767_to_fp16)[name = string("s_97_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_93_cast_fp16 = real_div(x = linear_46_cast_fp16, y = s_97_cast_fp16)[name = string("y_93_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3770_cast_fp16 = mul(x = y_93_cast_fp16, y = y_93_cast_fp16)[name = string("op_3770_cast_fp16")];
            tensor<int32, [1]> var_3775_axes_0 = const()[name = string("op_3775_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3775_keep_dims_0 = const()[name = string("op_3775_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3775_cast_fp16 = reduce_mean(axes = var_3775_axes_0, keep_dims = var_3775_keep_dims_0, x = var_3770_cast_fp16)[name = string("op_3775_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3776_cast_fp16 = mul(x = s_97_cast_fp16, y = s_97_cast_fp16)[name = string("op_3776_cast_fp16")];
            fp32 var_3777_epsilon_0 = const()[name = string("op_3777_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3777_cast_fp16 = inverse(epsilon = var_3777_epsilon_0, x = var_3776_cast_fp16)[name = string("op_3777_cast_fp16")];
            fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3779_cast_fp16 = mul(x = var_3777_cast_fp16, y = var_3778_to_fp16)[name = string("op_3779_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3781_cast_fp16 = add(x = var_3775_cast_fp16, y = var_3779_cast_fp16)[name = string("op_3781_cast_fp16")];
            fp32 var_3782_epsilon_0 = const()[name = string("op_3782_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3782_cast_fp16 = rsqrt(epsilon = var_3782_epsilon_0, x = var_3781_cast_fp16)[name = string("op_3782_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_79_cast_fp16 = mul(x = y_93_cast_fp16, y = var_3782_cast_fp16)[name = string("out_79_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59755648)))];
            tensor<fp16, [1, 630, 768]> var_3784_cast_fp16 = mul(x = out_79_cast_fp16, y = t_encoder_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("op_3784_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_123_cast_fp16 = add(x = x_109_cast_fp16, y = var_3784_cast_fp16)[name = string("x_123_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3787_cast_fp16 = abs(x = x_123_cast_fp16)[name = string("op_3787_cast_fp16")];
            bool var_3790 = const()[name = string("op_3790"), val = bool(true)];
            tensor<int32, [1]> var_3791_axes_0 = const()[name = string("op_3791_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3791_cast_fp16 = reduce_max(axes = var_3791_axes_0, keep_dims = var_3790, x = var_3787_cast_fp16)[name = string("op_3791_cast_fp16")];
            fp16 var_3792_to_fp16 = const()[name = string("op_3792_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_99_cast_fp16 = maximum(x = var_3791_cast_fp16, y = var_3792_to_fp16)[name = string("s_99_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_95_cast_fp16 = real_div(x = x_123_cast_fp16, y = s_99_cast_fp16)[name = string("y_95_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3795_cast_fp16 = mul(x = y_95_cast_fp16, y = y_95_cast_fp16)[name = string("op_3795_cast_fp16")];
            tensor<int32, [1]> var_3800_axes_0 = const()[name = string("op_3800_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3800_keep_dims_0 = const()[name = string("op_3800_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3800_cast_fp16 = reduce_mean(axes = var_3800_axes_0, keep_dims = var_3800_keep_dims_0, x = var_3795_cast_fp16)[name = string("op_3800_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3801_cast_fp16 = mul(x = s_99_cast_fp16, y = s_99_cast_fp16)[name = string("op_3801_cast_fp16")];
            fp32 var_3802_epsilon_0 = const()[name = string("op_3802_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3802_cast_fp16 = inverse(epsilon = var_3802_epsilon_0, x = var_3801_cast_fp16)[name = string("op_3802_cast_fp16")];
            fp16 var_3803_to_fp16 = const()[name = string("op_3803_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3804_cast_fp16 = mul(x = var_3802_cast_fp16, y = var_3803_to_fp16)[name = string("op_3804_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3806_cast_fp16 = add(x = var_3800_cast_fp16, y = var_3804_cast_fp16)[name = string("op_3806_cast_fp16")];
            fp32 var_3807_epsilon_0 = const()[name = string("op_3807_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3807_cast_fp16 = rsqrt(epsilon = var_3807_epsilon_0, x = var_3806_cast_fp16)[name = string("op_3807_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_81_cast_fp16 = mul(x = y_95_cast_fp16, y = var_3807_cast_fp16)[name = string("out_81_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59757248)))];
            tensor<fp16, [1, 630, 768]> input_67_cast_fp16 = mul(x = out_81_cast_fp16, y = t_encoder_layers_6_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(59758848))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62118208))))[name = string("t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_47_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_47_cast_fp16")];
            string var_3814_mode_0 = const()[name = string("op_3814_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_3814_cast_fp16 = gelu(mode = var_3814_mode_0, x = linear_47_cast_fp16)[name = string("op_3814_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62124416))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64483776))))[name = string("t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_48_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_67_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_71_cast_fp16 = mul(x = var_3814_cast_fp16, y = linear_48_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64489984))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66849344))))[name = string("t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_6_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3822_cast_fp16 = abs(x = linear_49_cast_fp16)[name = string("op_3822_cast_fp16")];
            bool var_3825 = const()[name = string("op_3825"), val = bool(true)];
            tensor<int32, [1]> var_3826_axes_0 = const()[name = string("op_3826_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3826_cast_fp16 = reduce_max(axes = var_3826_axes_0, keep_dims = var_3825, x = var_3822_cast_fp16)[name = string("op_3826_cast_fp16")];
            fp16 var_3827_to_fp16 = const()[name = string("op_3827_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_101_cast_fp16 = maximum(x = var_3826_cast_fp16, y = var_3827_to_fp16)[name = string("s_101_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_97_cast_fp16 = real_div(x = linear_49_cast_fp16, y = s_101_cast_fp16)[name = string("y_97_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3830_cast_fp16 = mul(x = y_97_cast_fp16, y = y_97_cast_fp16)[name = string("op_3830_cast_fp16")];
            tensor<int32, [1]> var_3835_axes_0 = const()[name = string("op_3835_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3835_keep_dims_0 = const()[name = string("op_3835_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3835_cast_fp16 = reduce_mean(axes = var_3835_axes_0, keep_dims = var_3835_keep_dims_0, x = var_3830_cast_fp16)[name = string("op_3835_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3836_cast_fp16 = mul(x = s_101_cast_fp16, y = s_101_cast_fp16)[name = string("op_3836_cast_fp16")];
            fp32 var_3837_epsilon_0 = const()[name = string("op_3837_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3837_cast_fp16 = inverse(epsilon = var_3837_epsilon_0, x = var_3836_cast_fp16)[name = string("op_3837_cast_fp16")];
            fp16 var_3838_to_fp16 = const()[name = string("op_3838_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3839_cast_fp16 = mul(x = var_3837_cast_fp16, y = var_3838_to_fp16)[name = string("op_3839_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3841_cast_fp16 = add(x = var_3835_cast_fp16, y = var_3839_cast_fp16)[name = string("op_3841_cast_fp16")];
            fp32 var_3842_epsilon_0 = const()[name = string("op_3842_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3842_cast_fp16 = rsqrt(epsilon = var_3842_epsilon_0, x = var_3841_cast_fp16)[name = string("op_3842_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_83_cast_fp16 = mul(x = y_97_cast_fp16, y = var_3842_cast_fp16)[name = string("out_83_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66850944)))];
            tensor<fp16, [1, 630, 768]> var_3844_cast_fp16 = mul(x = out_83_cast_fp16, y = t_encoder_layers_6_post_feedforward_layernorm_weight_to_fp16)[name = string("op_3844_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_127_cast_fp16 = add(x = x_123_cast_fp16, y = var_3844_cast_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3847_cast_fp16 = abs(x = x_127_cast_fp16)[name = string("op_3847_cast_fp16")];
            bool var_3850 = const()[name = string("op_3850"), val = bool(true)];
            tensor<int32, [1]> var_3851_axes_0 = const()[name = string("op_3851_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_3851_cast_fp16 = reduce_max(axes = var_3851_axes_0, keep_dims = var_3850, x = var_3847_cast_fp16)[name = string("op_3851_cast_fp16")];
            fp16 var_3852_to_fp16 = const()[name = string("op_3852_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_103_cast_fp16 = maximum(x = var_3851_cast_fp16, y = var_3852_to_fp16)[name = string("s_103_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_99_cast_fp16 = real_div(x = x_127_cast_fp16, y = s_103_cast_fp16)[name = string("y_99_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_3855_cast_fp16 = mul(x = y_99_cast_fp16, y = y_99_cast_fp16)[name = string("op_3855_cast_fp16")];
            tensor<int32, [1]> var_3860_axes_0 = const()[name = string("op_3860_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3860_keep_dims_0 = const()[name = string("op_3860_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_3860_cast_fp16 = reduce_mean(axes = var_3860_axes_0, keep_dims = var_3860_keep_dims_0, x = var_3855_cast_fp16)[name = string("op_3860_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3861_cast_fp16 = mul(x = s_103_cast_fp16, y = s_103_cast_fp16)[name = string("op_3861_cast_fp16")];
            fp32 var_3862_epsilon_0 = const()[name = string("op_3862_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_3862_cast_fp16 = inverse(epsilon = var_3862_epsilon_0, x = var_3861_cast_fp16)[name = string("op_3862_cast_fp16")];
            fp16 var_3863_to_fp16 = const()[name = string("op_3863_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_3864_cast_fp16 = mul(x = var_3862_cast_fp16, y = var_3863_to_fp16)[name = string("op_3864_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_3866_cast_fp16 = add(x = var_3860_cast_fp16, y = var_3864_cast_fp16)[name = string("op_3866_cast_fp16")];
            fp32 var_3867_epsilon_0 = const()[name = string("op_3867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_3867_cast_fp16 = rsqrt(epsilon = var_3867_epsilon_0, x = var_3866_cast_fp16)[name = string("op_3867_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_85_cast_fp16 = mul(x = y_99_cast_fp16, y = var_3867_cast_fp16)[name = string("out_85_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66852544)))];
            tensor<fp16, [1, 630, 768]> input_73_cast_fp16 = mul(x = out_85_cast_fp16, y = t_encoder_layers_7_input_layernorm_weight_to_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66854144))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67444032))))[name = string("t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3877 = const()[name = string("op_3877"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_129_cast_fp16 = reshape(shape = var_3877, x = linear_50_cast_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3879_cast_fp16 = abs(x = x_129_cast_fp16)[name = string("op_3879_cast_fp16")];
            bool var_3882 = const()[name = string("op_3882"), val = bool(true)];
            tensor<int32, [1]> var_3883_axes_0 = const()[name = string("op_3883_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3883_cast_fp16 = reduce_max(axes = var_3883_axes_0, keep_dims = var_3882, x = var_3879_cast_fp16)[name = string("op_3883_cast_fp16")];
            fp16 var_3884_to_fp16 = const()[name = string("op_3884_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_105_cast_fp16 = maximum(x = var_3883_cast_fp16, y = var_3884_to_fp16)[name = string("s_105_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_101_cast_fp16 = real_div(x = x_129_cast_fp16, y = s_105_cast_fp16)[name = string("y_101_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3887_cast_fp16 = mul(x = y_101_cast_fp16, y = y_101_cast_fp16)[name = string("op_3887_cast_fp16")];
            tensor<int32, [1]> var_3892_axes_0 = const()[name = string("op_3892_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3892_keep_dims_0 = const()[name = string("op_3892_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3892_cast_fp16 = reduce_mean(axes = var_3892_axes_0, keep_dims = var_3892_keep_dims_0, x = var_3887_cast_fp16)[name = string("op_3892_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3893_cast_fp16 = mul(x = s_105_cast_fp16, y = s_105_cast_fp16)[name = string("op_3893_cast_fp16")];
            fp32 var_3894_epsilon_0 = const()[name = string("op_3894_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3894_cast_fp16 = inverse(epsilon = var_3894_epsilon_0, x = var_3893_cast_fp16)[name = string("op_3894_cast_fp16")];
            fp16 var_3895_to_fp16 = const()[name = string("op_3895_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3896_cast_fp16 = mul(x = var_3894_cast_fp16, y = var_3895_to_fp16)[name = string("op_3896_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3898_cast_fp16 = add(x = var_3892_cast_fp16, y = var_3896_cast_fp16)[name = string("op_3898_cast_fp16")];
            fp32 var_3899_epsilon_0 = const()[name = string("op_3899_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3899_cast_fp16 = rsqrt(epsilon = var_3899_epsilon_0, x = var_3898_cast_fp16)[name = string("op_3899_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_87_cast_fp16 = mul(x = y_101_cast_fp16, y = var_3899_cast_fp16)[name = string("out_87_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1186368)))];
            tensor<fp16, [1, 630, 12, 64]> x_135_cast_fp16 = mul(x = out_87_cast_fp16, y = t_encoder_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_135_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67445632))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68035520))))[name = string("t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_51_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3909 = const()[name = string("op_3909"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_131_cast_fp16 = reshape(shape = var_3909, x = linear_51_cast_fp16)[name = string("x_131_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3911_cast_fp16 = abs(x = x_131_cast_fp16)[name = string("op_3911_cast_fp16")];
            bool var_3914 = const()[name = string("op_3914"), val = bool(true)];
            tensor<int32, [1]> var_3915_axes_0 = const()[name = string("op_3915_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3915_cast_fp16 = reduce_max(axes = var_3915_axes_0, keep_dims = var_3914, x = var_3911_cast_fp16)[name = string("op_3915_cast_fp16")];
            fp16 var_3916_to_fp16 = const()[name = string("op_3916_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_107_cast_fp16 = maximum(x = var_3915_cast_fp16, y = var_3916_to_fp16)[name = string("s_107_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_103_cast_fp16 = real_div(x = x_131_cast_fp16, y = s_107_cast_fp16)[name = string("y_103_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3919_cast_fp16 = mul(x = y_103_cast_fp16, y = y_103_cast_fp16)[name = string("op_3919_cast_fp16")];
            tensor<int32, [1]> var_3924_axes_0 = const()[name = string("op_3924_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3924_keep_dims_0 = const()[name = string("op_3924_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3924_cast_fp16 = reduce_mean(axes = var_3924_axes_0, keep_dims = var_3924_keep_dims_0, x = var_3919_cast_fp16)[name = string("op_3924_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3925_cast_fp16 = mul(x = s_107_cast_fp16, y = s_107_cast_fp16)[name = string("op_3925_cast_fp16")];
            fp32 var_3926_epsilon_0 = const()[name = string("op_3926_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3926_cast_fp16 = inverse(epsilon = var_3926_epsilon_0, x = var_3925_cast_fp16)[name = string("op_3926_cast_fp16")];
            fp16 var_3927_to_fp16 = const()[name = string("op_3927_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3928_cast_fp16 = mul(x = var_3926_cast_fp16, y = var_3927_to_fp16)[name = string("op_3928_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3930_cast_fp16 = add(x = var_3924_cast_fp16, y = var_3928_cast_fp16)[name = string("op_3930_cast_fp16")];
            fp32 var_3931_epsilon_0 = const()[name = string("op_3931_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3931_cast_fp16 = rsqrt(epsilon = var_3931_epsilon_0, x = var_3930_cast_fp16)[name = string("op_3931_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_89_cast_fp16 = mul(x = y_103_cast_fp16, y = var_3931_cast_fp16)[name = string("out_89_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1778048)))];
            tensor<fp16, [1, 630, 12, 64]> x_137_cast_fp16 = mul(x = out_89_cast_fp16, y = t_encoder_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_137_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68037120))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68627008))))[name = string("t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<int32, [4]> var_3941 = const()[name = string("op_3941"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_133_cast_fp16 = reshape(shape = var_3941, x = linear_52_cast_fp16)[name = string("x_133_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3943_cast_fp16 = abs(x = x_133_cast_fp16)[name = string("op_3943_cast_fp16")];
            bool var_3946 = const()[name = string("op_3946"), val = bool(true)];
            tensor<int32, [1]> var_3947_axes_0 = const()[name = string("op_3947_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_3947_cast_fp16 = reduce_max(axes = var_3947_axes_0, keep_dims = var_3946, x = var_3943_cast_fp16)[name = string("op_3947_cast_fp16")];
            fp16 var_3948_to_fp16 = const()[name = string("op_3948_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_109_cast_fp16 = maximum(x = var_3947_cast_fp16, y = var_3948_to_fp16)[name = string("s_109_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_105_cast_fp16 = real_div(x = x_133_cast_fp16, y = s_109_cast_fp16)[name = string("y_105_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_3951_cast_fp16 = mul(x = y_105_cast_fp16, y = y_105_cast_fp16)[name = string("op_3951_cast_fp16")];
            tensor<int32, [1]> var_3956_axes_0 = const()[name = string("op_3956_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3956_keep_dims_0 = const()[name = string("op_3956_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_3956_cast_fp16 = reduce_mean(axes = var_3956_axes_0, keep_dims = var_3956_keep_dims_0, x = var_3951_cast_fp16)[name = string("op_3956_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3957_cast_fp16 = mul(x = s_109_cast_fp16, y = s_109_cast_fp16)[name = string("op_3957_cast_fp16")];
            fp32 var_3958_epsilon_0 = const()[name = string("op_3958_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_3958_cast_fp16 = inverse(epsilon = var_3958_epsilon_0, x = var_3957_cast_fp16)[name = string("op_3958_cast_fp16")];
            fp16 var_3959_to_fp16 = const()[name = string("op_3959_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_3960_cast_fp16 = mul(x = var_3958_cast_fp16, y = var_3959_to_fp16)[name = string("op_3960_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_3962_cast_fp16 = add(x = var_3956_cast_fp16, y = var_3960_cast_fp16)[name = string("op_3962_cast_fp16")];
            fp32 var_3963_epsilon_0 = const()[name = string("op_3963_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_3963_cast_fp16 = rsqrt(epsilon = var_3963_epsilon_0, x = var_3962_cast_fp16)[name = string("op_3963_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_29_cast_fp16 = mul(x = y_105_cast_fp16, y = var_3963_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<int32, [4]> z_57_begin_0 = const()[name = string("z_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_57_end_0 = const()[name = string("z_57_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_57_end_mask_0 = const()[name = string("z_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_57_cast_fp16 = slice_by_index(begin = z_57_begin_0, end = z_57_end_0, end_mask = z_57_end_mask_0, x = x_135_cast_fp16)[name = string("z_57_cast_fp16")];
            tensor<int32, [4]> var_3974_begin_0 = const()[name = string("op_3974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3974_end_0 = const()[name = string("op_3974_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3974_end_mask_0 = const()[name = string("op_3974_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> 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 = z_57_cast_fp16)[name = string("op_3974_cast_fp16")];
            fp16 const_28_promoted_to_fp16 = const()[name = string("const_28_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3975_cast_fp16 = mul(x = var_3974_cast_fp16, y = const_28_promoted_to_fp16)[name = string("op_3975_cast_fp16")];
            tensor<int32, [4]> var_3980_begin_0 = const()[name = string("op_3980_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3980_end_0 = const()[name = string("op_3980_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_3980_end_mask_0 = const()[name = string("op_3980_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_3980_cast_fp16 = slice_by_index(begin = var_3980_begin_0, end = var_3980_end_0, end_mask = var_3980_end_mask_0, x = z_57_cast_fp16)[name = string("op_3980_cast_fp16")];
            int32 var_3982 = const()[name = string("op_3982"), val = int32(-1)];
            bool rot_57_interleave_0 = const()[name = string("rot_57_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_57_cast_fp16 = concat(axis = var_3982, interleave = rot_57_interleave_0, values = (var_3975_cast_fp16, var_3980_cast_fp16))[name = string("rot_57_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3984_cast_fp16 = mul(x = z_57_cast_fp16, y = c_1_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3985_cast_fp16 = mul(x = rot_57_cast_fp16, y = s_9_cast_fp16)[name = string("op_3985_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_3987_cast_fp16 = add(x = var_3984_cast_fp16, y = var_3985_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<int32, [4]> z_59_begin_0 = const()[name = string("z_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_59_end_0 = const()[name = string("z_59_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_59_end_mask_0 = const()[name = string("z_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_59_cast_fp16 = slice_by_index(begin = z_59_begin_0, end = z_59_end_0, end_mask = z_59_end_mask_0, x = x_135_cast_fp16)[name = string("z_59_cast_fp16")];
            tensor<int32, [4]> var_3997_begin_0 = const()[name = string("op_3997_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_3997_end_0 = const()[name = string("op_3997_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_3997_end_mask_0 = const()[name = string("op_3997_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, x = z_59_cast_fp16)[name = string("op_3997_cast_fp16")];
            fp16 const_29_promoted_to_fp16 = const()[name = string("const_29_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_3998_cast_fp16 = mul(x = var_3997_cast_fp16, y = const_29_promoted_to_fp16)[name = string("op_3998_cast_fp16")];
            tensor<int32, [4]> var_4003_begin_0 = const()[name = string("op_4003_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4003_end_0 = const()[name = string("op_4003_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4003_end_mask_0 = const()[name = string("op_4003_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4003_cast_fp16 = slice_by_index(begin = var_4003_begin_0, end = var_4003_end_0, end_mask = var_4003_end_mask_0, x = z_59_cast_fp16)[name = string("op_4003_cast_fp16")];
            int32 var_4005 = const()[name = string("op_4005"), val = int32(-1)];
            bool rot_59_interleave_0 = const()[name = string("rot_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_59_cast_fp16 = concat(axis = var_4005, interleave = rot_59_interleave_0, values = (var_3998_cast_fp16, var_4003_cast_fp16))[name = string("rot_59_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4007_cast_fp16 = mul(x = z_59_cast_fp16, y = c_cast_fp16)[name = string("op_4007_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4008_cast_fp16 = mul(x = rot_59_cast_fp16, y = s_11_cast_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4010_cast_fp16 = add(x = var_4007_cast_fp16, y = var_4008_cast_fp16)[name = string("op_4010_cast_fp16")];
            int32 var_4012 = const()[name = string("op_4012"), val = int32(-1)];
            bool var_4013_interleave_0 = const()[name = string("op_4013_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_4013_cast_fp16 = concat(axis = var_4012, interleave = var_4013_interleave_0, values = (var_3987_cast_fp16, var_4010_cast_fp16))[name = string("op_4013_cast_fp16")];
            tensor<int32, [4]> z_61_begin_0 = const()[name = string("z_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_61_end_0 = const()[name = string("z_61_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_61_end_mask_0 = const()[name = string("z_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_61_cast_fp16 = slice_by_index(begin = z_61_begin_0, end = z_61_end_0, end_mask = z_61_end_mask_0, x = x_137_cast_fp16)[name = string("z_61_cast_fp16")];
            tensor<int32, [4]> var_4026_begin_0 = const()[name = string("op_4026_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4026_end_0 = const()[name = string("op_4026_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4026_end_mask_0 = const()[name = string("op_4026_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4026_cast_fp16 = slice_by_index(begin = var_4026_begin_0, end = var_4026_end_0, end_mask = var_4026_end_mask_0, x = z_61_cast_fp16)[name = string("op_4026_cast_fp16")];
            fp16 const_30_promoted_to_fp16 = const()[name = string("const_30_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4027_cast_fp16 = mul(x = var_4026_cast_fp16, y = const_30_promoted_to_fp16)[name = string("op_4027_cast_fp16")];
            tensor<int32, [4]> var_4032_begin_0 = const()[name = string("op_4032_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4032_end_0 = const()[name = string("op_4032_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4032_end_mask_0 = const()[name = string("op_4032_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4032_cast_fp16 = slice_by_index(begin = var_4032_begin_0, end = var_4032_end_0, end_mask = var_4032_end_mask_0, x = z_61_cast_fp16)[name = string("op_4032_cast_fp16")];
            int32 var_4034 = const()[name = string("op_4034"), val = int32(-1)];
            bool rot_61_interleave_0 = const()[name = string("rot_61_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_61_cast_fp16 = concat(axis = var_4034, interleave = rot_61_interleave_0, values = (var_4027_cast_fp16, var_4032_cast_fp16))[name = string("rot_61_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4036_cast_fp16 = mul(x = z_61_cast_fp16, y = c_1_cast_fp16)[name = string("op_4036_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4037_cast_fp16 = mul(x = rot_61_cast_fp16, y = s_9_cast_fp16)[name = string("op_4037_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4039_cast_fp16 = add(x = var_4036_cast_fp16, y = var_4037_cast_fp16)[name = string("op_4039_cast_fp16")];
            tensor<int32, [4]> z_63_begin_0 = const()[name = string("z_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_63_end_0 = const()[name = string("z_63_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_63_end_mask_0 = const()[name = string("z_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_63_cast_fp16 = slice_by_index(begin = z_63_begin_0, end = z_63_end_0, end_mask = z_63_end_mask_0, x = x_137_cast_fp16)[name = string("z_63_cast_fp16")];
            tensor<int32, [4]> var_4049_begin_0 = const()[name = string("op_4049_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4049_end_0 = const()[name = string("op_4049_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4049_end_mask_0 = const()[name = string("op_4049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4049_cast_fp16 = slice_by_index(begin = var_4049_begin_0, end = var_4049_end_0, end_mask = var_4049_end_mask_0, x = z_63_cast_fp16)[name = string("op_4049_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4050_cast_fp16 = mul(x = var_4049_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_4050_cast_fp16")];
            tensor<int32, [4]> var_4055_begin_0 = const()[name = string("op_4055_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4055_end_0 = const()[name = string("op_4055_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4055_end_mask_0 = const()[name = string("op_4055_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4055_cast_fp16 = slice_by_index(begin = var_4055_begin_0, end = var_4055_end_0, end_mask = var_4055_end_mask_0, x = z_63_cast_fp16)[name = string("op_4055_cast_fp16")];
            int32 var_4057 = const()[name = string("op_4057"), val = int32(-1)];
            bool rot_63_interleave_0 = const()[name = string("rot_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_63_cast_fp16 = concat(axis = var_4057, interleave = rot_63_interleave_0, values = (var_4050_cast_fp16, var_4055_cast_fp16))[name = string("rot_63_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4059_cast_fp16 = mul(x = z_63_cast_fp16, y = c_cast_fp16)[name = string("op_4059_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4060_cast_fp16 = mul(x = rot_63_cast_fp16, y = s_11_cast_fp16)[name = string("op_4060_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4062_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4060_cast_fp16)[name = string("op_4062_cast_fp16")];
            int32 var_4064 = const()[name = string("op_4064"), val = int32(-1)];
            bool var_4065_interleave_0 = const()[name = string("op_4065_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_4065_cast_fp16 = concat(axis = var_4064, interleave = var_4065_interleave_0, values = (var_4039_cast_fp16, var_4062_cast_fp16))[name = string("op_4065_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4075_transpose_x_0 = const()[name = string("op_4075_transpose_x_0"), val = bool(false)];
            bool var_4075_transpose_y_0 = const()[name = string("op_4075_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_4065_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 12, 630, 64]> transpose_78 = transpose(perm = transpose_78_perm_0, x = var_4013_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 12, 630, 630]> var_4075_cast_fp16 = matmul(transpose_x = var_4075_transpose_x_0, transpose_y = var_4075_transpose_y_0, x = transpose_78, y = transpose_79)[name = string("op_4075_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_4077_cast_fp16 = add(x = var_4075_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4077_cast_fp16")];
            int32 var_4078 = const()[name = string("op_4078"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_105_cast_fp16 = softmax(axis = var_4078, x = var_4077_cast_fp16)[name = string("w_105_cast_fp16")];
            bool var_4081_transpose_x_0 = const()[name = string("op_4081_transpose_x_0"), val = bool(false)];
            bool var_4081_transpose_y_0 = const()[name = string("op_4081_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = v_29_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 12, 630, 64]> var_4081_cast_fp16 = matmul(transpose_x = var_4081_transpose_x_0, transpose_y = var_4081_transpose_y_0, x = w_105_cast_fp16, y = v_31_cast_fp16)[name = string("op_4081_cast_fp16")];
            tensor<int32, [4]> var_4084_perm_0 = const()[name = string("op_4084_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4088 = const()[name = string("op_4088"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_4084_cast_fp16 = transpose(perm = var_4084_perm_0, x = var_4081_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 630, 768]> input_75_cast_fp16 = reshape(shape = var_4088, x = var_4084_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(68628608))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69218496))))[name = string("t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_53_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4093_cast_fp16 = abs(x = linear_53_cast_fp16)[name = string("op_4093_cast_fp16")];
            bool var_4096 = const()[name = string("op_4096"), val = bool(true)];
            tensor<int32, [1]> var_4097_axes_0 = const()[name = string("op_4097_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4097_cast_fp16 = reduce_max(axes = var_4097_axes_0, keep_dims = var_4096, x = var_4093_cast_fp16)[name = string("op_4097_cast_fp16")];
            fp16 var_4098_to_fp16 = const()[name = string("op_4098_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_111_cast_fp16 = maximum(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = string("s_111_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_107_cast_fp16 = real_div(x = linear_53_cast_fp16, y = s_111_cast_fp16)[name = string("y_107_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4101_cast_fp16 = mul(x = y_107_cast_fp16, y = y_107_cast_fp16)[name = string("op_4101_cast_fp16")];
            tensor<int32, [1]> var_4106_axes_0 = const()[name = string("op_4106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4106_keep_dims_0 = const()[name = string("op_4106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4106_cast_fp16 = reduce_mean(axes = var_4106_axes_0, keep_dims = var_4106_keep_dims_0, x = var_4101_cast_fp16)[name = string("op_4106_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4107_cast_fp16 = mul(x = s_111_cast_fp16, y = s_111_cast_fp16)[name = string("op_4107_cast_fp16")];
            fp32 var_4108_epsilon_0 = const()[name = string("op_4108_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4108_cast_fp16 = inverse(epsilon = var_4108_epsilon_0, x = var_4107_cast_fp16)[name = string("op_4108_cast_fp16")];
            fp16 var_4109_to_fp16 = const()[name = string("op_4109_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4110_cast_fp16 = mul(x = var_4108_cast_fp16, y = var_4109_to_fp16)[name = string("op_4110_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4112_cast_fp16 = add(x = var_4106_cast_fp16, y = var_4110_cast_fp16)[name = string("op_4112_cast_fp16")];
            fp32 var_4113_epsilon_0 = const()[name = string("op_4113_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4113_cast_fp16 = rsqrt(epsilon = var_4113_epsilon_0, x = var_4112_cast_fp16)[name = string("op_4113_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_91_cast_fp16 = mul(x = y_107_cast_fp16, y = var_4113_cast_fp16)[name = string("out_91_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69220096)))];
            tensor<fp16, [1, 630, 768]> var_4115_cast_fp16 = mul(x = out_91_cast_fp16, y = t_encoder_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("op_4115_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_141_cast_fp16 = add(x = x_127_cast_fp16, y = var_4115_cast_fp16)[name = string("x_141_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4118_cast_fp16 = abs(x = x_141_cast_fp16)[name = string("op_4118_cast_fp16")];
            bool var_4121 = const()[name = string("op_4121"), val = bool(true)];
            tensor<int32, [1]> var_4122_axes_0 = const()[name = string("op_4122_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4122_cast_fp16 = reduce_max(axes = var_4122_axes_0, keep_dims = var_4121, x = var_4118_cast_fp16)[name = string("op_4122_cast_fp16")];
            fp16 var_4123_to_fp16 = const()[name = string("op_4123_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_113_cast_fp16 = maximum(x = var_4122_cast_fp16, y = var_4123_to_fp16)[name = string("s_113_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_109_cast_fp16 = real_div(x = x_141_cast_fp16, y = s_113_cast_fp16)[name = string("y_109_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4126_cast_fp16 = mul(x = y_109_cast_fp16, y = y_109_cast_fp16)[name = string("op_4126_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, 630, 1]> var_4131_cast_fp16 = reduce_mean(axes = var_4131_axes_0, keep_dims = var_4131_keep_dims_0, x = var_4126_cast_fp16)[name = string("op_4131_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4132_cast_fp16 = mul(x = s_113_cast_fp16, y = s_113_cast_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, 630, 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, 630, 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, 630, 1]> var_4137_cast_fp16 = add(x = var_4131_cast_fp16, y = var_4135_cast_fp16)[name = string("op_4137_cast_fp16")];
            fp32 var_4138_epsilon_0 = const()[name = string("op_4138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4138_cast_fp16 = rsqrt(epsilon = var_4138_epsilon_0, x = var_4137_cast_fp16)[name = string("op_4138_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_93_cast_fp16 = mul(x = y_109_cast_fp16, y = var_4138_cast_fp16)[name = string("out_93_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69221696)))];
            tensor<fp16, [1, 630, 768]> input_77_cast_fp16 = mul(x = out_93_cast_fp16, y = t_encoder_layers_7_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69223296))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71582656))))[name = string("t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_54_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_54_cast_fp16")];
            string var_4145_mode_0 = const()[name = string("op_4145_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_4145_cast_fp16 = gelu(mode = var_4145_mode_0, x = linear_54_cast_fp16)[name = string("op_4145_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71588864))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73948224))))[name = string("t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_55_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_77_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_81_cast_fp16 = mul(x = var_4145_cast_fp16, y = linear_55_cast_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(73954432))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76313792))))[name = string("t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_7_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4153_cast_fp16 = abs(x = linear_56_cast_fp16)[name = string("op_4153_cast_fp16")];
            bool var_4156 = const()[name = string("op_4156"), val = bool(true)];
            tensor<int32, [1]> var_4157_axes_0 = const()[name = string("op_4157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4157_cast_fp16 = reduce_max(axes = var_4157_axes_0, keep_dims = var_4156, x = var_4153_cast_fp16)[name = string("op_4157_cast_fp16")];
            fp16 var_4158_to_fp16 = const()[name = string("op_4158_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_115_cast_fp16 = maximum(x = var_4157_cast_fp16, y = var_4158_to_fp16)[name = string("s_115_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_111_cast_fp16 = real_div(x = linear_56_cast_fp16, y = s_115_cast_fp16)[name = string("y_111_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4161_cast_fp16 = mul(x = y_111_cast_fp16, y = y_111_cast_fp16)[name = string("op_4161_cast_fp16")];
            tensor<int32, [1]> var_4166_axes_0 = const()[name = string("op_4166_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4166_keep_dims_0 = const()[name = string("op_4166_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4166_cast_fp16 = reduce_mean(axes = var_4166_axes_0, keep_dims = var_4166_keep_dims_0, x = var_4161_cast_fp16)[name = string("op_4166_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4167_cast_fp16 = mul(x = s_115_cast_fp16, y = s_115_cast_fp16)[name = string("op_4167_cast_fp16")];
            fp32 var_4168_epsilon_0 = const()[name = string("op_4168_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4168_cast_fp16 = inverse(epsilon = var_4168_epsilon_0, x = var_4167_cast_fp16)[name = string("op_4168_cast_fp16")];
            fp16 var_4169_to_fp16 = const()[name = string("op_4169_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4170_cast_fp16 = mul(x = var_4168_cast_fp16, y = var_4169_to_fp16)[name = string("op_4170_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4172_cast_fp16 = add(x = var_4166_cast_fp16, y = var_4170_cast_fp16)[name = string("op_4172_cast_fp16")];
            fp32 var_4173_epsilon_0 = const()[name = string("op_4173_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4173_cast_fp16 = rsqrt(epsilon = var_4173_epsilon_0, x = var_4172_cast_fp16)[name = string("op_4173_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_95_cast_fp16 = mul(x = y_111_cast_fp16, y = var_4173_cast_fp16)[name = string("out_95_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76315392)))];
            tensor<fp16, [1, 630, 768]> var_4175_cast_fp16 = mul(x = out_95_cast_fp16, y = t_encoder_layers_7_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4175_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_145_cast_fp16 = add(x = x_141_cast_fp16, y = var_4175_cast_fp16)[name = string("x_145_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4178_cast_fp16 = abs(x = x_145_cast_fp16)[name = string("op_4178_cast_fp16")];
            bool var_4181 = const()[name = string("op_4181"), val = bool(true)];
            tensor<int32, [1]> var_4182_axes_0 = const()[name = string("op_4182_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4182_cast_fp16 = reduce_max(axes = var_4182_axes_0, keep_dims = var_4181, x = var_4178_cast_fp16)[name = string("op_4182_cast_fp16")];
            fp16 var_4183_to_fp16 = const()[name = string("op_4183_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_117_cast_fp16 = maximum(x = var_4182_cast_fp16, y = var_4183_to_fp16)[name = string("s_117_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_113_cast_fp16 = real_div(x = x_145_cast_fp16, y = s_117_cast_fp16)[name = string("y_113_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4186_cast_fp16 = mul(x = y_113_cast_fp16, y = y_113_cast_fp16)[name = string("op_4186_cast_fp16")];
            tensor<int32, [1]> var_4191_axes_0 = const()[name = string("op_4191_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4191_keep_dims_0 = const()[name = string("op_4191_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4191_cast_fp16 = reduce_mean(axes = var_4191_axes_0, keep_dims = var_4191_keep_dims_0, x = var_4186_cast_fp16)[name = string("op_4191_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4192_cast_fp16 = mul(x = s_117_cast_fp16, y = s_117_cast_fp16)[name = string("op_4192_cast_fp16")];
            fp32 var_4193_epsilon_0 = const()[name = string("op_4193_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4193_cast_fp16 = inverse(epsilon = var_4193_epsilon_0, x = var_4192_cast_fp16)[name = string("op_4193_cast_fp16")];
            fp16 var_4194_to_fp16 = const()[name = string("op_4194_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4195_cast_fp16 = mul(x = var_4193_cast_fp16, y = var_4194_to_fp16)[name = string("op_4195_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4197_cast_fp16 = add(x = var_4191_cast_fp16, y = var_4195_cast_fp16)[name = string("op_4197_cast_fp16")];
            fp32 var_4198_epsilon_0 = const()[name = string("op_4198_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4198_cast_fp16 = rsqrt(epsilon = var_4198_epsilon_0, x = var_4197_cast_fp16)[name = string("op_4198_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_97_cast_fp16 = mul(x = y_113_cast_fp16, y = var_4198_cast_fp16)[name = string("out_97_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76316992)))];
            tensor<fp16, [1, 630, 768]> input_83_cast_fp16 = mul(x = out_97_cast_fp16, y = t_encoder_layers_8_input_layernorm_weight_to_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76318592))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76908480))))[name = string("t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_57_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_4208 = const()[name = string("op_4208"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_147_cast_fp16 = reshape(shape = var_4208, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4210_cast_fp16 = abs(x = x_147_cast_fp16)[name = string("op_4210_cast_fp16")];
            bool var_4213 = const()[name = string("op_4213"), val = bool(true)];
            tensor<int32, [1]> var_4214_axes_0 = const()[name = string("op_4214_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4214_cast_fp16 = reduce_max(axes = var_4214_axes_0, keep_dims = var_4213, x = var_4210_cast_fp16)[name = string("op_4214_cast_fp16")];
            fp16 var_4215_to_fp16 = const()[name = string("op_4215_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_119_cast_fp16 = maximum(x = var_4214_cast_fp16, y = var_4215_to_fp16)[name = string("s_119_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_115_cast_fp16 = real_div(x = x_147_cast_fp16, y = s_119_cast_fp16)[name = string("y_115_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4218_cast_fp16 = mul(x = y_115_cast_fp16, y = y_115_cast_fp16)[name = string("op_4218_cast_fp16")];
            tensor<int32, [1]> var_4223_axes_0 = const()[name = string("op_4223_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4223_keep_dims_0 = const()[name = string("op_4223_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4223_cast_fp16 = reduce_mean(axes = var_4223_axes_0, keep_dims = var_4223_keep_dims_0, x = var_4218_cast_fp16)[name = string("op_4223_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4224_cast_fp16 = mul(x = s_119_cast_fp16, y = s_119_cast_fp16)[name = string("op_4224_cast_fp16")];
            fp32 var_4225_epsilon_0 = const()[name = string("op_4225_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4225_cast_fp16 = inverse(epsilon = var_4225_epsilon_0, x = var_4224_cast_fp16)[name = string("op_4225_cast_fp16")];
            fp16 var_4226_to_fp16 = const()[name = string("op_4226_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4227_cast_fp16 = mul(x = var_4225_cast_fp16, y = var_4226_to_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4229_cast_fp16 = add(x = var_4223_cast_fp16, y = var_4227_cast_fp16)[name = string("op_4229_cast_fp16")];
            fp32 var_4230_epsilon_0 = const()[name = string("op_4230_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4230_cast_fp16 = rsqrt(epsilon = var_4230_epsilon_0, x = var_4229_cast_fp16)[name = string("op_4230_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_99_cast_fp16 = mul(x = y_115_cast_fp16, y = var_4230_cast_fp16)[name = string("out_99_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910080)))];
            tensor<fp16, [1, 630, 12, 64]> x_153_cast_fp16 = mul(x = out_99_cast_fp16, y = t_encoder_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910272))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77500160))))[name = string("t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_4240 = const()[name = string("op_4240"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_149_cast_fp16 = reshape(shape = var_4240, x = linear_58_cast_fp16)[name = string("x_149_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4242_cast_fp16 = abs(x = x_149_cast_fp16)[name = string("op_4242_cast_fp16")];
            bool var_4245 = const()[name = string("op_4245"), val = bool(true)];
            tensor<int32, [1]> var_4246_axes_0 = const()[name = string("op_4246_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4246_cast_fp16 = reduce_max(axes = var_4246_axes_0, keep_dims = var_4245, x = var_4242_cast_fp16)[name = string("op_4246_cast_fp16")];
            fp16 var_4247_to_fp16 = const()[name = string("op_4247_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_121_cast_fp16 = maximum(x = var_4246_cast_fp16, y = var_4247_to_fp16)[name = string("s_121_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_117_cast_fp16 = real_div(x = x_149_cast_fp16, y = s_121_cast_fp16)[name = string("y_117_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4250_cast_fp16 = mul(x = y_117_cast_fp16, y = y_117_cast_fp16)[name = string("op_4250_cast_fp16")];
            tensor<int32, [1]> var_4255_axes_0 = const()[name = string("op_4255_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4255_keep_dims_0 = const()[name = string("op_4255_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4255_cast_fp16 = reduce_mean(axes = var_4255_axes_0, keep_dims = var_4255_keep_dims_0, x = var_4250_cast_fp16)[name = string("op_4255_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4256_cast_fp16 = mul(x = s_121_cast_fp16, y = s_121_cast_fp16)[name = string("op_4256_cast_fp16")];
            fp32 var_4257_epsilon_0 = const()[name = string("op_4257_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4257_cast_fp16 = inverse(epsilon = var_4257_epsilon_0, x = var_4256_cast_fp16)[name = string("op_4257_cast_fp16")];
            fp16 var_4258_to_fp16 = const()[name = string("op_4258_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4259_cast_fp16 = mul(x = var_4257_cast_fp16, y = var_4258_to_fp16)[name = string("op_4259_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4261_cast_fp16 = add(x = var_4255_cast_fp16, y = var_4259_cast_fp16)[name = string("op_4261_cast_fp16")];
            fp32 var_4262_epsilon_0 = const()[name = string("op_4262_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4262_cast_fp16 = rsqrt(epsilon = var_4262_epsilon_0, x = var_4261_cast_fp16)[name = string("op_4262_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_101_cast_fp16 = mul(x = y_117_cast_fp16, y = var_4262_cast_fp16)[name = string("out_101_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501760)))];
            tensor<fp16, [1, 630, 12, 64]> x_155_cast_fp16 = mul(x = out_101_cast_fp16, y = t_encoder_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_155_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501952))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78091840))))[name = string("t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_59_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<int32, [4]> var_4272 = const()[name = string("op_4272"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_151_cast_fp16 = reshape(shape = var_4272, x = linear_59_cast_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4274_cast_fp16 = abs(x = x_151_cast_fp16)[name = string("op_4274_cast_fp16")];
            bool var_4277 = const()[name = string("op_4277"), val = bool(true)];
            tensor<int32, [1]> var_4278_axes_0 = const()[name = string("op_4278_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4278_cast_fp16 = reduce_max(axes = var_4278_axes_0, keep_dims = var_4277, x = var_4274_cast_fp16)[name = string("op_4278_cast_fp16")];
            fp16 var_4279_to_fp16 = const()[name = string("op_4279_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_123_cast_fp16 = maximum(x = var_4278_cast_fp16, y = var_4279_to_fp16)[name = string("s_123_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_119_cast_fp16 = real_div(x = x_151_cast_fp16, y = s_123_cast_fp16)[name = string("y_119_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4282_cast_fp16 = mul(x = y_119_cast_fp16, y = y_119_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<int32, [1]> var_4287_axes_0 = const()[name = string("op_4287_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4287_keep_dims_0 = const()[name = string("op_4287_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4287_cast_fp16 = reduce_mean(axes = var_4287_axes_0, keep_dims = var_4287_keep_dims_0, x = var_4282_cast_fp16)[name = string("op_4287_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4288_cast_fp16 = mul(x = s_123_cast_fp16, y = s_123_cast_fp16)[name = string("op_4288_cast_fp16")];
            fp32 var_4289_epsilon_0 = const()[name = string("op_4289_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4289_cast_fp16 = inverse(epsilon = var_4289_epsilon_0, x = var_4288_cast_fp16)[name = string("op_4289_cast_fp16")];
            fp16 var_4290_to_fp16 = const()[name = string("op_4290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4291_cast_fp16 = mul(x = var_4289_cast_fp16, y = var_4290_to_fp16)[name = string("op_4291_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4293_cast_fp16 = add(x = var_4287_cast_fp16, y = var_4291_cast_fp16)[name = string("op_4293_cast_fp16")];
            fp32 var_4294_epsilon_0 = const()[name = string("op_4294_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4294_cast_fp16 = rsqrt(epsilon = var_4294_epsilon_0, x = var_4293_cast_fp16)[name = string("op_4294_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_33_cast_fp16 = mul(x = y_119_cast_fp16, y = var_4294_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<int32, [4]> z_65_begin_0 = const()[name = string("z_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_65_end_0 = const()[name = string("z_65_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_65_end_mask_0 = const()[name = string("z_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_65_cast_fp16 = slice_by_index(begin = z_65_begin_0, end = z_65_end_0, end_mask = z_65_end_mask_0, x = x_153_cast_fp16)[name = string("z_65_cast_fp16")];
            tensor<int32, [4]> var_4305_begin_0 = const()[name = string("op_4305_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4305_end_0 = const()[name = string("op_4305_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4305_end_mask_0 = const()[name = string("op_4305_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4305_cast_fp16 = slice_by_index(begin = var_4305_begin_0, end = var_4305_end_0, end_mask = var_4305_end_mask_0, x = z_65_cast_fp16)[name = string("op_4305_cast_fp16")];
            fp16 const_32_promoted_to_fp16 = const()[name = string("const_32_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4306_cast_fp16 = mul(x = var_4305_cast_fp16, y = const_32_promoted_to_fp16)[name = string("op_4306_cast_fp16")];
            tensor<int32, [4]> var_4311_begin_0 = const()[name = string("op_4311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4311_end_0 = const()[name = string("op_4311_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4311_end_mask_0 = const()[name = string("op_4311_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4311_cast_fp16 = slice_by_index(begin = var_4311_begin_0, end = var_4311_end_0, end_mask = var_4311_end_mask_0, x = z_65_cast_fp16)[name = string("op_4311_cast_fp16")];
            int32 var_4313 = const()[name = string("op_4313"), val = int32(-1)];
            bool rot_65_interleave_0 = const()[name = string("rot_65_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_65_cast_fp16 = concat(axis = var_4313, interleave = rot_65_interleave_0, values = (var_4306_cast_fp16, var_4311_cast_fp16))[name = string("rot_65_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4315_cast_fp16 = mul(x = z_65_cast_fp16, y = c_1_cast_fp16)[name = string("op_4315_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4316_cast_fp16 = mul(x = rot_65_cast_fp16, y = s_9_cast_fp16)[name = string("op_4316_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4318_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4316_cast_fp16)[name = string("op_4318_cast_fp16")];
            tensor<int32, [4]> z_67_begin_0 = const()[name = string("z_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_67_end_0 = const()[name = string("z_67_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_67_end_mask_0 = const()[name = string("z_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_67_cast_fp16 = slice_by_index(begin = z_67_begin_0, end = z_67_end_0, end_mask = z_67_end_mask_0, x = x_153_cast_fp16)[name = string("z_67_cast_fp16")];
            tensor<int32, [4]> var_4328_begin_0 = const()[name = string("op_4328_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4328_end_0 = const()[name = string("op_4328_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4328_end_mask_0 = const()[name = string("op_4328_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4328_cast_fp16 = slice_by_index(begin = var_4328_begin_0, end = var_4328_end_0, end_mask = var_4328_end_mask_0, x = z_67_cast_fp16)[name = string("op_4328_cast_fp16")];
            fp16 const_33_promoted_to_fp16 = const()[name = string("const_33_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4329_cast_fp16 = mul(x = var_4328_cast_fp16, y = const_33_promoted_to_fp16)[name = string("op_4329_cast_fp16")];
            tensor<int32, [4]> var_4334_begin_0 = const()[name = string("op_4334_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4334_end_0 = const()[name = string("op_4334_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4334_end_mask_0 = const()[name = string("op_4334_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4334_cast_fp16 = slice_by_index(begin = var_4334_begin_0, end = var_4334_end_0, end_mask = var_4334_end_mask_0, x = z_67_cast_fp16)[name = string("op_4334_cast_fp16")];
            int32 var_4336 = const()[name = string("op_4336"), val = int32(-1)];
            bool rot_67_interleave_0 = const()[name = string("rot_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_67_cast_fp16 = concat(axis = var_4336, interleave = rot_67_interleave_0, values = (var_4329_cast_fp16, var_4334_cast_fp16))[name = string("rot_67_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4338_cast_fp16 = mul(x = z_67_cast_fp16, y = c_cast_fp16)[name = string("op_4338_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4339_cast_fp16 = mul(x = rot_67_cast_fp16, y = s_11_cast_fp16)[name = string("op_4339_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4341_cast_fp16 = add(x = var_4338_cast_fp16, y = var_4339_cast_fp16)[name = string("op_4341_cast_fp16")];
            int32 var_4343 = const()[name = string("op_4343"), val = int32(-1)];
            bool var_4344_interleave_0 = const()[name = string("op_4344_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_4344_cast_fp16 = concat(axis = var_4343, interleave = var_4344_interleave_0, values = (var_4318_cast_fp16, var_4341_cast_fp16))[name = string("op_4344_cast_fp16")];
            tensor<int32, [4]> z_69_begin_0 = const()[name = string("z_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_69_end_0 = const()[name = string("z_69_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_69_end_mask_0 = const()[name = string("z_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_69_cast_fp16 = slice_by_index(begin = z_69_begin_0, end = z_69_end_0, end_mask = z_69_end_mask_0, x = x_155_cast_fp16)[name = string("z_69_cast_fp16")];
            tensor<int32, [4]> var_4357_begin_0 = const()[name = string("op_4357_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4357_end_0 = const()[name = string("op_4357_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4357_end_mask_0 = const()[name = string("op_4357_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4357_cast_fp16 = slice_by_index(begin = var_4357_begin_0, end = var_4357_end_0, end_mask = var_4357_end_mask_0, x = z_69_cast_fp16)[name = string("op_4357_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4358_cast_fp16 = mul(x = var_4357_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_4358_cast_fp16")];
            tensor<int32, [4]> var_4363_begin_0 = const()[name = string("op_4363_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4363_end_0 = const()[name = string("op_4363_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4363_end_mask_0 = const()[name = string("op_4363_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4363_cast_fp16 = slice_by_index(begin = var_4363_begin_0, end = var_4363_end_0, end_mask = var_4363_end_mask_0, x = z_69_cast_fp16)[name = string("op_4363_cast_fp16")];
            int32 var_4365 = const()[name = string("op_4365"), val = int32(-1)];
            bool rot_69_interleave_0 = const()[name = string("rot_69_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_69_cast_fp16 = concat(axis = var_4365, interleave = rot_69_interleave_0, values = (var_4358_cast_fp16, var_4363_cast_fp16))[name = string("rot_69_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4367_cast_fp16 = mul(x = z_69_cast_fp16, y = c_1_cast_fp16)[name = string("op_4367_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4368_cast_fp16 = mul(x = rot_69_cast_fp16, y = s_9_cast_fp16)[name = string("op_4368_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4370_cast_fp16 = add(x = var_4367_cast_fp16, y = var_4368_cast_fp16)[name = string("op_4370_cast_fp16")];
            tensor<int32, [4]> z_71_begin_0 = const()[name = string("z_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_71_end_0 = const()[name = string("z_71_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_71_end_mask_0 = const()[name = string("z_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_71_cast_fp16 = slice_by_index(begin = z_71_begin_0, end = z_71_end_0, end_mask = z_71_end_mask_0, x = x_155_cast_fp16)[name = string("z_71_cast_fp16")];
            tensor<int32, [4]> var_4380_begin_0 = const()[name = string("op_4380_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4380_end_0 = const()[name = string("op_4380_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4380_end_mask_0 = const()[name = string("op_4380_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4380_cast_fp16 = slice_by_index(begin = var_4380_begin_0, end = var_4380_end_0, end_mask = var_4380_end_mask_0, x = z_71_cast_fp16)[name = string("op_4380_cast_fp16")];
            fp16 const_35_promoted_to_fp16 = const()[name = string("const_35_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4381_cast_fp16 = mul(x = var_4380_cast_fp16, y = const_35_promoted_to_fp16)[name = string("op_4381_cast_fp16")];
            tensor<int32, [4]> var_4386_begin_0 = const()[name = string("op_4386_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4386_end_0 = const()[name = string("op_4386_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4386_end_mask_0 = const()[name = string("op_4386_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4386_cast_fp16 = slice_by_index(begin = var_4386_begin_0, end = var_4386_end_0, end_mask = var_4386_end_mask_0, x = z_71_cast_fp16)[name = string("op_4386_cast_fp16")];
            int32 var_4388 = const()[name = string("op_4388"), val = int32(-1)];
            bool rot_71_interleave_0 = const()[name = string("rot_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_71_cast_fp16 = concat(axis = var_4388, interleave = rot_71_interleave_0, values = (var_4381_cast_fp16, var_4386_cast_fp16))[name = string("rot_71_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4390_cast_fp16 = mul(x = z_71_cast_fp16, y = c_cast_fp16)[name = string("op_4390_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4391_cast_fp16 = mul(x = rot_71_cast_fp16, y = s_11_cast_fp16)[name = string("op_4391_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4393_cast_fp16 = add(x = var_4390_cast_fp16, y = var_4391_cast_fp16)[name = string("op_4393_cast_fp16")];
            int32 var_4395 = const()[name = string("op_4395"), val = int32(-1)];
            bool var_4396_interleave_0 = const()[name = string("op_4396_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_4396_cast_fp16 = concat(axis = var_4395, interleave = var_4396_interleave_0, values = (var_4370_cast_fp16, var_4393_cast_fp16))[name = string("op_4396_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4406_transpose_x_0 = const()[name = string("op_4406_transpose_x_0"), val = bool(false)];
            bool var_4406_transpose_y_0 = const()[name = string("op_4406_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_4396_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 12, 630, 64]> transpose_80 = transpose(perm = transpose_80_perm_0, x = var_4344_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 12, 630, 630]> var_4406_cast_fp16 = matmul(transpose_x = var_4406_transpose_x_0, transpose_y = var_4406_transpose_y_0, x = transpose_80, y = transpose_81)[name = string("op_4406_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_4408_cast_fp16 = add(x = var_4406_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4408_cast_fp16")];
            int32 var_4409 = const()[name = string("op_4409"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_119_cast_fp16 = softmax(axis = var_4409, x = var_4408_cast_fp16)[name = string("w_119_cast_fp16")];
            bool var_4412_transpose_x_0 = const()[name = string("op_4412_transpose_x_0"), val = bool(false)];
            bool var_4412_transpose_y_0 = const()[name = string("op_4412_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = v_33_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 12, 630, 64]> var_4412_cast_fp16 = matmul(transpose_x = var_4412_transpose_x_0, transpose_y = var_4412_transpose_y_0, x = w_119_cast_fp16, y = v_35_cast_fp16)[name = string("op_4412_cast_fp16")];
            tensor<int32, [4]> var_4415_perm_0 = const()[name = string("op_4415_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4419 = const()[name = string("op_4419"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_4415_cast_fp16 = transpose(perm = var_4415_perm_0, x = var_4412_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 630, 768]> input_85_cast_fp16 = reshape(shape = var_4419, x = var_4415_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78093440))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78683328))))[name = string("t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4424_cast_fp16 = abs(x = linear_60_cast_fp16)[name = string("op_4424_cast_fp16")];
            bool var_4427 = const()[name = string("op_4427"), val = bool(true)];
            tensor<int32, [1]> var_4428_axes_0 = const()[name = string("op_4428_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4428_cast_fp16 = reduce_max(axes = var_4428_axes_0, keep_dims = var_4427, x = var_4424_cast_fp16)[name = string("op_4428_cast_fp16")];
            fp16 var_4429_to_fp16 = const()[name = string("op_4429_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_125_cast_fp16 = maximum(x = var_4428_cast_fp16, y = var_4429_to_fp16)[name = string("s_125_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_121_cast_fp16 = real_div(x = linear_60_cast_fp16, y = s_125_cast_fp16)[name = string("y_121_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4432_cast_fp16 = mul(x = y_121_cast_fp16, y = y_121_cast_fp16)[name = string("op_4432_cast_fp16")];
            tensor<int32, [1]> var_4437_axes_0 = const()[name = string("op_4437_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4437_keep_dims_0 = const()[name = string("op_4437_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4437_cast_fp16 = reduce_mean(axes = var_4437_axes_0, keep_dims = var_4437_keep_dims_0, x = var_4432_cast_fp16)[name = string("op_4437_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4438_cast_fp16 = mul(x = s_125_cast_fp16, y = s_125_cast_fp16)[name = string("op_4438_cast_fp16")];
            fp32 var_4439_epsilon_0 = const()[name = string("op_4439_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4439_cast_fp16 = inverse(epsilon = var_4439_epsilon_0, x = var_4438_cast_fp16)[name = string("op_4439_cast_fp16")];
            fp16 var_4440_to_fp16 = const()[name = string("op_4440_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4441_cast_fp16 = mul(x = var_4439_cast_fp16, y = var_4440_to_fp16)[name = string("op_4441_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4443_cast_fp16 = add(x = var_4437_cast_fp16, y = var_4441_cast_fp16)[name = string("op_4443_cast_fp16")];
            fp32 var_4444_epsilon_0 = const()[name = string("op_4444_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4444_cast_fp16 = rsqrt(epsilon = var_4444_epsilon_0, x = var_4443_cast_fp16)[name = string("op_4444_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_103_cast_fp16 = mul(x = y_121_cast_fp16, y = var_4444_cast_fp16)[name = string("out_103_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78684928)))];
            tensor<fp16, [1, 630, 768]> var_4446_cast_fp16 = mul(x = out_103_cast_fp16, y = t_encoder_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("op_4446_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_159_cast_fp16 = add(x = x_145_cast_fp16, y = var_4446_cast_fp16)[name = string("x_159_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4449_cast_fp16 = abs(x = x_159_cast_fp16)[name = string("op_4449_cast_fp16")];
            bool var_4452 = const()[name = string("op_4452"), val = bool(true)];
            tensor<int32, [1]> var_4453_axes_0 = const()[name = string("op_4453_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4453_cast_fp16 = reduce_max(axes = var_4453_axes_0, keep_dims = var_4452, x = var_4449_cast_fp16)[name = string("op_4453_cast_fp16")];
            fp16 var_4454_to_fp16 = const()[name = string("op_4454_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_127_cast_fp16 = maximum(x = var_4453_cast_fp16, y = var_4454_to_fp16)[name = string("s_127_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_123_cast_fp16 = real_div(x = x_159_cast_fp16, y = s_127_cast_fp16)[name = string("y_123_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4457_cast_fp16 = mul(x = y_123_cast_fp16, y = y_123_cast_fp16)[name = string("op_4457_cast_fp16")];
            tensor<int32, [1]> var_4462_axes_0 = const()[name = string("op_4462_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4462_keep_dims_0 = const()[name = string("op_4462_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4462_cast_fp16 = reduce_mean(axes = var_4462_axes_0, keep_dims = var_4462_keep_dims_0, x = var_4457_cast_fp16)[name = string("op_4462_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4463_cast_fp16 = mul(x = s_127_cast_fp16, y = s_127_cast_fp16)[name = string("op_4463_cast_fp16")];
            fp32 var_4464_epsilon_0 = const()[name = string("op_4464_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4464_cast_fp16 = inverse(epsilon = var_4464_epsilon_0, x = var_4463_cast_fp16)[name = string("op_4464_cast_fp16")];
            fp16 var_4465_to_fp16 = const()[name = string("op_4465_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4466_cast_fp16 = mul(x = var_4464_cast_fp16, y = var_4465_to_fp16)[name = string("op_4466_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4468_cast_fp16 = add(x = var_4462_cast_fp16, y = var_4466_cast_fp16)[name = string("op_4468_cast_fp16")];
            fp32 var_4469_epsilon_0 = const()[name = string("op_4469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4469_cast_fp16 = rsqrt(epsilon = var_4469_epsilon_0, x = var_4468_cast_fp16)[name = string("op_4469_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_105_cast_fp16 = mul(x = y_123_cast_fp16, y = var_4469_cast_fp16)[name = string("out_105_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78686528)))];
            tensor<fp16, [1, 630, 768]> input_87_cast_fp16 = mul(x = out_105_cast_fp16, y = t_encoder_layers_8_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(78688128))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81047488))))[name = string("t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_61_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_61_cast_fp16")];
            string var_4476_mode_0 = const()[name = string("op_4476_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_4476_cast_fp16 = gelu(mode = var_4476_mode_0, x = linear_61_cast_fp16)[name = string("op_4476_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81053696))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83413056))))[name = string("t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_62_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_87_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_91_cast_fp16 = mul(x = var_4476_cast_fp16, y = linear_62_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83419264))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85778624))))[name = string("t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_8_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4484_cast_fp16 = abs(x = linear_63_cast_fp16)[name = string("op_4484_cast_fp16")];
            bool var_4487 = const()[name = string("op_4487"), val = bool(true)];
            tensor<int32, [1]> var_4488_axes_0 = const()[name = string("op_4488_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4488_cast_fp16 = reduce_max(axes = var_4488_axes_0, keep_dims = var_4487, x = var_4484_cast_fp16)[name = string("op_4488_cast_fp16")];
            fp16 var_4489_to_fp16 = const()[name = string("op_4489_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_129_cast_fp16 = maximum(x = var_4488_cast_fp16, y = var_4489_to_fp16)[name = string("s_129_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_125_cast_fp16 = real_div(x = linear_63_cast_fp16, y = s_129_cast_fp16)[name = string("y_125_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4492_cast_fp16 = mul(x = y_125_cast_fp16, y = y_125_cast_fp16)[name = string("op_4492_cast_fp16")];
            tensor<int32, [1]> var_4497_axes_0 = const()[name = string("op_4497_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4497_keep_dims_0 = const()[name = string("op_4497_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4497_cast_fp16 = reduce_mean(axes = var_4497_axes_0, keep_dims = var_4497_keep_dims_0, x = var_4492_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4498_cast_fp16 = mul(x = s_129_cast_fp16, y = s_129_cast_fp16)[name = string("op_4498_cast_fp16")];
            fp32 var_4499_epsilon_0 = const()[name = string("op_4499_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4499_cast_fp16 = inverse(epsilon = var_4499_epsilon_0, x = var_4498_cast_fp16)[name = string("op_4499_cast_fp16")];
            fp16 var_4500_to_fp16 = const()[name = string("op_4500_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4501_cast_fp16 = mul(x = var_4499_cast_fp16, y = var_4500_to_fp16)[name = string("op_4501_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4503_cast_fp16 = add(x = var_4497_cast_fp16, y = var_4501_cast_fp16)[name = string("op_4503_cast_fp16")];
            fp32 var_4504_epsilon_0 = const()[name = string("op_4504_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4504_cast_fp16 = rsqrt(epsilon = var_4504_epsilon_0, x = var_4503_cast_fp16)[name = string("op_4504_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_107_cast_fp16 = mul(x = y_125_cast_fp16, y = var_4504_cast_fp16)[name = string("out_107_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85780224)))];
            tensor<fp16, [1, 630, 768]> var_4506_cast_fp16 = mul(x = out_107_cast_fp16, y = t_encoder_layers_8_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4506_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_163_cast_fp16 = add(x = x_159_cast_fp16, y = var_4506_cast_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4509_cast_fp16 = abs(x = x_163_cast_fp16)[name = string("op_4509_cast_fp16")];
            bool var_4512 = const()[name = string("op_4512"), val = bool(true)];
            tensor<int32, [1]> var_4513_axes_0 = const()[name = string("op_4513_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4513_cast_fp16 = reduce_max(axes = var_4513_axes_0, keep_dims = var_4512, x = var_4509_cast_fp16)[name = string("op_4513_cast_fp16")];
            fp16 var_4514_to_fp16 = const()[name = string("op_4514_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_131_cast_fp16 = maximum(x = var_4513_cast_fp16, y = var_4514_to_fp16)[name = string("s_131_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_127_cast_fp16 = real_div(x = x_163_cast_fp16, y = s_131_cast_fp16)[name = string("y_127_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4517_cast_fp16 = mul(x = y_127_cast_fp16, y = y_127_cast_fp16)[name = string("op_4517_cast_fp16")];
            tensor<int32, [1]> var_4522_axes_0 = const()[name = string("op_4522_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4522_keep_dims_0 = const()[name = string("op_4522_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4522_cast_fp16 = reduce_mean(axes = var_4522_axes_0, keep_dims = var_4522_keep_dims_0, x = var_4517_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4523_cast_fp16 = mul(x = s_131_cast_fp16, y = s_131_cast_fp16)[name = string("op_4523_cast_fp16")];
            fp32 var_4524_epsilon_0 = const()[name = string("op_4524_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4524_cast_fp16 = inverse(epsilon = var_4524_epsilon_0, x = var_4523_cast_fp16)[name = string("op_4524_cast_fp16")];
            fp16 var_4525_to_fp16 = const()[name = string("op_4525_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4526_cast_fp16 = mul(x = var_4524_cast_fp16, y = var_4525_to_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4528_cast_fp16 = add(x = var_4522_cast_fp16, y = var_4526_cast_fp16)[name = string("op_4528_cast_fp16")];
            fp32 var_4529_epsilon_0 = const()[name = string("op_4529_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4529_cast_fp16 = rsqrt(epsilon = var_4529_epsilon_0, x = var_4528_cast_fp16)[name = string("op_4529_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_109_cast_fp16 = mul(x = y_127_cast_fp16, y = var_4529_cast_fp16)[name = string("out_109_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85781824)))];
            tensor<fp16, [1, 630, 768]> input_93_cast_fp16 = mul(x = out_109_cast_fp16, y = t_encoder_layers_9_input_layernorm_weight_to_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(85783424))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86373312))))[name = string("t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4539 = const()[name = string("op_4539"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_165_cast_fp16 = reshape(shape = var_4539, x = linear_64_cast_fp16)[name = string("x_165_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4541_cast_fp16 = abs(x = x_165_cast_fp16)[name = string("op_4541_cast_fp16")];
            bool var_4544 = const()[name = string("op_4544"), val = bool(true)];
            tensor<int32, [1]> var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4545_cast_fp16 = reduce_max(axes = var_4545_axes_0, keep_dims = var_4544, x = var_4541_cast_fp16)[name = string("op_4545_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_133_cast_fp16 = maximum(x = var_4545_cast_fp16, y = var_4546_to_fp16)[name = string("s_133_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_129_cast_fp16 = real_div(x = x_165_cast_fp16, y = s_133_cast_fp16)[name = string("y_129_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4549_cast_fp16 = mul(x = y_129_cast_fp16, y = y_129_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<int32, [1]> var_4554_axes_0 = const()[name = string("op_4554_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4554_keep_dims_0 = const()[name = string("op_4554_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4554_cast_fp16 = reduce_mean(axes = var_4554_axes_0, keep_dims = var_4554_keep_dims_0, x = var_4549_cast_fp16)[name = string("op_4554_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4555_cast_fp16 = mul(x = s_133_cast_fp16, y = s_133_cast_fp16)[name = string("op_4555_cast_fp16")];
            fp32 var_4556_epsilon_0 = const()[name = string("op_4556_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4556_cast_fp16 = inverse(epsilon = var_4556_epsilon_0, x = var_4555_cast_fp16)[name = string("op_4556_cast_fp16")];
            fp16 var_4557_to_fp16 = const()[name = string("op_4557_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4558_cast_fp16 = mul(x = var_4556_cast_fp16, y = var_4557_to_fp16)[name = string("op_4558_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4560_cast_fp16 = add(x = var_4554_cast_fp16, y = var_4558_cast_fp16)[name = string("op_4560_cast_fp16")];
            fp32 var_4561_epsilon_0 = const()[name = string("op_4561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4561_cast_fp16 = rsqrt(epsilon = var_4561_epsilon_0, x = var_4560_cast_fp16)[name = string("op_4561_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_111_cast_fp16 = mul(x = y_129_cast_fp16, y = var_4561_cast_fp16)[name = string("out_111_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86374912)))];
            tensor<fp16, [1, 630, 12, 64]> x_171_cast_fp16 = mul(x = out_111_cast_fp16, y = t_encoder_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_171_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86375104))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86964992))))[name = string("t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_65_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4571 = const()[name = string("op_4571"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_167_cast_fp16 = reshape(shape = var_4571, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4573_cast_fp16 = abs(x = x_167_cast_fp16)[name = string("op_4573_cast_fp16")];
            bool var_4576 = const()[name = string("op_4576"), val = bool(true)];
            tensor<int32, [1]> var_4577_axes_0 = const()[name = string("op_4577_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4577_cast_fp16 = reduce_max(axes = var_4577_axes_0, keep_dims = var_4576, x = var_4573_cast_fp16)[name = string("op_4577_cast_fp16")];
            fp16 var_4578_to_fp16 = const()[name = string("op_4578_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_135_cast_fp16 = maximum(x = var_4577_cast_fp16, y = var_4578_to_fp16)[name = string("s_135_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_131_cast_fp16 = real_div(x = x_167_cast_fp16, y = s_135_cast_fp16)[name = string("y_131_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4581_cast_fp16 = mul(x = y_131_cast_fp16, y = y_131_cast_fp16)[name = string("op_4581_cast_fp16")];
            tensor<int32, [1]> var_4586_axes_0 = const()[name = string("op_4586_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4586_keep_dims_0 = const()[name = string("op_4586_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4586_cast_fp16 = reduce_mean(axes = var_4586_axes_0, keep_dims = var_4586_keep_dims_0, x = var_4581_cast_fp16)[name = string("op_4586_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4587_cast_fp16 = mul(x = s_135_cast_fp16, y = s_135_cast_fp16)[name = string("op_4587_cast_fp16")];
            fp32 var_4588_epsilon_0 = const()[name = string("op_4588_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4588_cast_fp16 = inverse(epsilon = var_4588_epsilon_0, x = var_4587_cast_fp16)[name = string("op_4588_cast_fp16")];
            fp16 var_4589_to_fp16 = const()[name = string("op_4589_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4590_cast_fp16 = mul(x = var_4588_cast_fp16, y = var_4589_to_fp16)[name = string("op_4590_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4592_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4590_cast_fp16)[name = string("op_4592_cast_fp16")];
            fp32 var_4593_epsilon_0 = const()[name = string("op_4593_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4593_cast_fp16 = rsqrt(epsilon = var_4593_epsilon_0, x = var_4592_cast_fp16)[name = string("op_4593_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_113_cast_fp16 = mul(x = y_131_cast_fp16, y = var_4593_cast_fp16)[name = string("out_113_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966592)))];
            tensor<fp16, [1, 630, 12, 64]> x_173_cast_fp16 = mul(x = out_113_cast_fp16, y = t_encoder_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_173_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966784))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87556672))))[name = string("t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<int32, [4]> var_4603 = const()[name = string("op_4603"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_169_cast_fp16 = reshape(shape = var_4603, x = linear_66_cast_fp16)[name = string("x_169_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4605_cast_fp16 = abs(x = x_169_cast_fp16)[name = string("op_4605_cast_fp16")];
            bool var_4608 = const()[name = string("op_4608"), val = bool(true)];
            tensor<int32, [1]> var_4609_axes_0 = const()[name = string("op_4609_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4609_cast_fp16 = reduce_max(axes = var_4609_axes_0, keep_dims = var_4608, x = var_4605_cast_fp16)[name = string("op_4609_cast_fp16")];
            fp16 var_4610_to_fp16 = const()[name = string("op_4610_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_137_cast_fp16 = maximum(x = var_4609_cast_fp16, y = var_4610_to_fp16)[name = string("s_137_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_133_cast_fp16 = real_div(x = x_169_cast_fp16, y = s_137_cast_fp16)[name = string("y_133_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4613_cast_fp16 = mul(x = y_133_cast_fp16, y = y_133_cast_fp16)[name = string("op_4613_cast_fp16")];
            tensor<int32, [1]> var_4618_axes_0 = const()[name = string("op_4618_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4618_keep_dims_0 = const()[name = string("op_4618_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4618_cast_fp16 = reduce_mean(axes = var_4618_axes_0, keep_dims = var_4618_keep_dims_0, x = var_4613_cast_fp16)[name = string("op_4618_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4619_cast_fp16 = mul(x = s_137_cast_fp16, y = s_137_cast_fp16)[name = string("op_4619_cast_fp16")];
            fp32 var_4620_epsilon_0 = const()[name = string("op_4620_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4620_cast_fp16 = inverse(epsilon = var_4620_epsilon_0, x = var_4619_cast_fp16)[name = string("op_4620_cast_fp16")];
            fp16 var_4621_to_fp16 = const()[name = string("op_4621_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4622_cast_fp16 = mul(x = var_4620_cast_fp16, y = var_4621_to_fp16)[name = string("op_4622_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4624_cast_fp16 = add(x = var_4618_cast_fp16, y = var_4622_cast_fp16)[name = string("op_4624_cast_fp16")];
            fp32 var_4625_epsilon_0 = const()[name = string("op_4625_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4625_cast_fp16 = rsqrt(epsilon = var_4625_epsilon_0, x = var_4624_cast_fp16)[name = string("op_4625_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_37_cast_fp16 = mul(x = y_133_cast_fp16, y = var_4625_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<int32, [4]> z_73_begin_0 = const()[name = string("z_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_73_end_0 = const()[name = string("z_73_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_73_end_mask_0 = const()[name = string("z_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_73_cast_fp16 = slice_by_index(begin = z_73_begin_0, end = z_73_end_0, end_mask = z_73_end_mask_0, x = x_171_cast_fp16)[name = string("z_73_cast_fp16")];
            tensor<int32, [4]> var_4636_begin_0 = const()[name = string("op_4636_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4636_end_0 = const()[name = string("op_4636_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4636_end_mask_0 = const()[name = string("op_4636_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4636_cast_fp16 = slice_by_index(begin = var_4636_begin_0, end = var_4636_end_0, end_mask = var_4636_end_mask_0, x = z_73_cast_fp16)[name = string("op_4636_cast_fp16")];
            fp16 const_36_promoted_to_fp16 = const()[name = string("const_36_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4637_cast_fp16 = mul(x = var_4636_cast_fp16, y = const_36_promoted_to_fp16)[name = string("op_4637_cast_fp16")];
            tensor<int32, [4]> var_4642_begin_0 = const()[name = string("op_4642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4642_end_0 = const()[name = string("op_4642_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4642_end_mask_0 = const()[name = string("op_4642_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4642_cast_fp16 = slice_by_index(begin = var_4642_begin_0, end = var_4642_end_0, end_mask = var_4642_end_mask_0, x = z_73_cast_fp16)[name = string("op_4642_cast_fp16")];
            int32 var_4644 = const()[name = string("op_4644"), val = int32(-1)];
            bool rot_73_interleave_0 = const()[name = string("rot_73_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_73_cast_fp16 = concat(axis = var_4644, interleave = rot_73_interleave_0, values = (var_4637_cast_fp16, var_4642_cast_fp16))[name = string("rot_73_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4646_cast_fp16 = mul(x = z_73_cast_fp16, y = c_1_cast_fp16)[name = string("op_4646_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4647_cast_fp16 = mul(x = rot_73_cast_fp16, y = s_9_cast_fp16)[name = string("op_4647_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4649_cast_fp16 = add(x = var_4646_cast_fp16, y = var_4647_cast_fp16)[name = string("op_4649_cast_fp16")];
            tensor<int32, [4]> z_75_begin_0 = const()[name = string("z_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_75_end_0 = const()[name = string("z_75_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_75_end_mask_0 = const()[name = string("z_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_75_cast_fp16 = slice_by_index(begin = z_75_begin_0, end = z_75_end_0, end_mask = z_75_end_mask_0, x = x_171_cast_fp16)[name = string("z_75_cast_fp16")];
            tensor<int32, [4]> var_4659_begin_0 = const()[name = string("op_4659_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4659_end_0 = const()[name = string("op_4659_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4659_end_mask_0 = const()[name = string("op_4659_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4659_cast_fp16 = slice_by_index(begin = var_4659_begin_0, end = var_4659_end_0, end_mask = var_4659_end_mask_0, x = z_75_cast_fp16)[name = string("op_4659_cast_fp16")];
            fp16 const_37_promoted_to_fp16 = const()[name = string("const_37_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4660_cast_fp16 = mul(x = var_4659_cast_fp16, y = const_37_promoted_to_fp16)[name = string("op_4660_cast_fp16")];
            tensor<int32, [4]> var_4665_begin_0 = const()[name = string("op_4665_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4665_end_0 = const()[name = string("op_4665_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4665_end_mask_0 = const()[name = string("op_4665_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4665_cast_fp16 = slice_by_index(begin = var_4665_begin_0, end = var_4665_end_0, end_mask = var_4665_end_mask_0, x = z_75_cast_fp16)[name = string("op_4665_cast_fp16")];
            int32 var_4667 = const()[name = string("op_4667"), val = int32(-1)];
            bool rot_75_interleave_0 = const()[name = string("rot_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_75_cast_fp16 = concat(axis = var_4667, interleave = rot_75_interleave_0, values = (var_4660_cast_fp16, var_4665_cast_fp16))[name = string("rot_75_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4669_cast_fp16 = mul(x = z_75_cast_fp16, y = c_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4670_cast_fp16 = mul(x = rot_75_cast_fp16, y = s_11_cast_fp16)[name = string("op_4670_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4672_cast_fp16 = add(x = var_4669_cast_fp16, y = var_4670_cast_fp16)[name = string("op_4672_cast_fp16")];
            int32 var_4674 = const()[name = string("op_4674"), val = int32(-1)];
            bool var_4675_interleave_0 = const()[name = string("op_4675_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_4675_cast_fp16 = concat(axis = var_4674, interleave = var_4675_interleave_0, values = (var_4649_cast_fp16, var_4672_cast_fp16))[name = string("op_4675_cast_fp16")];
            tensor<int32, [4]> z_77_begin_0 = const()[name = string("z_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_77_end_0 = const()[name = string("z_77_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_77_end_mask_0 = const()[name = string("z_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_77_cast_fp16 = slice_by_index(begin = z_77_begin_0, end = z_77_end_0, end_mask = z_77_end_mask_0, x = x_173_cast_fp16)[name = string("z_77_cast_fp16")];
            tensor<int32, [4]> var_4688_begin_0 = const()[name = string("op_4688_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4688_end_0 = const()[name = string("op_4688_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4688_end_mask_0 = const()[name = string("op_4688_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4688_cast_fp16 = slice_by_index(begin = var_4688_begin_0, end = var_4688_end_0, end_mask = var_4688_end_mask_0, x = z_77_cast_fp16)[name = string("op_4688_cast_fp16")];
            fp16 const_38_promoted_to_fp16 = const()[name = string("const_38_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4689_cast_fp16 = mul(x = var_4688_cast_fp16, y = const_38_promoted_to_fp16)[name = string("op_4689_cast_fp16")];
            tensor<int32, [4]> var_4694_begin_0 = const()[name = string("op_4694_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4694_end_0 = const()[name = string("op_4694_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4694_end_mask_0 = const()[name = string("op_4694_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4694_cast_fp16 = slice_by_index(begin = var_4694_begin_0, end = var_4694_end_0, end_mask = var_4694_end_mask_0, x = z_77_cast_fp16)[name = string("op_4694_cast_fp16")];
            int32 var_4696 = const()[name = string("op_4696"), val = int32(-1)];
            bool rot_77_interleave_0 = const()[name = string("rot_77_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_77_cast_fp16 = concat(axis = var_4696, interleave = rot_77_interleave_0, values = (var_4689_cast_fp16, var_4694_cast_fp16))[name = string("rot_77_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4698_cast_fp16 = mul(x = z_77_cast_fp16, y = c_1_cast_fp16)[name = string("op_4698_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4699_cast_fp16 = mul(x = rot_77_cast_fp16, y = s_9_cast_fp16)[name = string("op_4699_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4701_cast_fp16 = add(x = var_4698_cast_fp16, y = var_4699_cast_fp16)[name = string("op_4701_cast_fp16")];
            tensor<int32, [4]> z_79_begin_0 = const()[name = string("z_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_79_end_0 = const()[name = string("z_79_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_79_end_mask_0 = const()[name = string("z_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_79_cast_fp16 = slice_by_index(begin = z_79_begin_0, end = z_79_end_0, end_mask = z_79_end_mask_0, x = x_173_cast_fp16)[name = string("z_79_cast_fp16")];
            tensor<int32, [4]> var_4711_begin_0 = const()[name = string("op_4711_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4711_end_0 = const()[name = string("op_4711_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4711_end_mask_0 = const()[name = string("op_4711_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4711_cast_fp16 = slice_by_index(begin = var_4711_begin_0, end = var_4711_end_0, end_mask = var_4711_end_mask_0, x = z_79_cast_fp16)[name = string("op_4711_cast_fp16")];
            fp16 const_39_promoted_to_fp16 = const()[name = string("const_39_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4712_cast_fp16 = mul(x = var_4711_cast_fp16, y = const_39_promoted_to_fp16)[name = string("op_4712_cast_fp16")];
            tensor<int32, [4]> var_4717_begin_0 = const()[name = string("op_4717_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4717_end_0 = const()[name = string("op_4717_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4717_end_mask_0 = const()[name = string("op_4717_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4717_cast_fp16 = slice_by_index(begin = var_4717_begin_0, end = var_4717_end_0, end_mask = var_4717_end_mask_0, x = z_79_cast_fp16)[name = string("op_4717_cast_fp16")];
            int32 var_4719 = const()[name = string("op_4719"), val = int32(-1)];
            bool rot_79_interleave_0 = const()[name = string("rot_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_79_cast_fp16 = concat(axis = var_4719, interleave = rot_79_interleave_0, values = (var_4712_cast_fp16, var_4717_cast_fp16))[name = string("rot_79_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4721_cast_fp16 = mul(x = z_79_cast_fp16, y = c_cast_fp16)[name = string("op_4721_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4722_cast_fp16 = mul(x = rot_79_cast_fp16, y = s_11_cast_fp16)[name = string("op_4722_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4724_cast_fp16 = add(x = var_4721_cast_fp16, y = var_4722_cast_fp16)[name = string("op_4724_cast_fp16")];
            int32 var_4726 = const()[name = string("op_4726"), val = int32(-1)];
            bool var_4727_interleave_0 = const()[name = string("op_4727_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_4727_cast_fp16 = concat(axis = var_4726, interleave = var_4727_interleave_0, values = (var_4701_cast_fp16, var_4724_cast_fp16))[name = string("op_4727_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_4737_transpose_x_0 = const()[name = string("op_4737_transpose_x_0"), val = bool(false)];
            bool var_4737_transpose_y_0 = const()[name = string("op_4737_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_4727_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 12, 630, 64]> transpose_82 = transpose(perm = transpose_82_perm_0, x = var_4675_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 12, 630, 630]> var_4737_cast_fp16 = matmul(transpose_x = var_4737_transpose_x_0, transpose_y = var_4737_transpose_y_0, x = transpose_82, y = transpose_83)[name = string("op_4737_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_4739_cast_fp16 = add(x = var_4737_cast_fp16, y = var_1524_cast_fp16)[name = string("op_4739_cast_fp16")];
            int32 var_4740 = const()[name = string("op_4740"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_133_cast_fp16 = softmax(axis = var_4740, x = var_4739_cast_fp16)[name = string("w_133_cast_fp16")];
            bool var_4743_transpose_x_0 = const()[name = string("op_4743_transpose_x_0"), val = bool(false)];
            bool var_4743_transpose_y_0 = const()[name = string("op_4743_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = v_37_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 12, 630, 64]> var_4743_cast_fp16 = matmul(transpose_x = var_4743_transpose_x_0, transpose_y = var_4743_transpose_y_0, x = w_133_cast_fp16, y = v_39_cast_fp16)[name = string("op_4743_cast_fp16")];
            tensor<int32, [4]> var_4746_perm_0 = const()[name = string("op_4746_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4750 = const()[name = string("op_4750"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_4746_cast_fp16 = transpose(perm = var_4746_perm_0, x = var_4743_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 630, 768]> input_95_cast_fp16 = reshape(shape = var_4750, x = var_4746_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(87558272))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88148160))))[name = string("t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_67_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4755_cast_fp16 = abs(x = linear_67_cast_fp16)[name = string("op_4755_cast_fp16")];
            bool var_4758 = const()[name = string("op_4758"), val = bool(true)];
            tensor<int32, [1]> var_4759_axes_0 = const()[name = string("op_4759_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4759_cast_fp16 = reduce_max(axes = var_4759_axes_0, keep_dims = var_4758, x = var_4755_cast_fp16)[name = string("op_4759_cast_fp16")];
            fp16 var_4760_to_fp16 = const()[name = string("op_4760_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_139_cast_fp16 = maximum(x = var_4759_cast_fp16, y = var_4760_to_fp16)[name = string("s_139_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_135_cast_fp16 = real_div(x = linear_67_cast_fp16, y = s_139_cast_fp16)[name = string("y_135_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4763_cast_fp16 = mul(x = y_135_cast_fp16, y = y_135_cast_fp16)[name = string("op_4763_cast_fp16")];
            tensor<int32, [1]> var_4768_axes_0 = const()[name = string("op_4768_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4768_keep_dims_0 = const()[name = string("op_4768_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4768_cast_fp16 = reduce_mean(axes = var_4768_axes_0, keep_dims = var_4768_keep_dims_0, x = var_4763_cast_fp16)[name = string("op_4768_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4769_cast_fp16 = mul(x = s_139_cast_fp16, y = s_139_cast_fp16)[name = string("op_4769_cast_fp16")];
            fp32 var_4770_epsilon_0 = const()[name = string("op_4770_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4770_cast_fp16 = inverse(epsilon = var_4770_epsilon_0, x = var_4769_cast_fp16)[name = string("op_4770_cast_fp16")];
            fp16 var_4771_to_fp16 = const()[name = string("op_4771_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4772_cast_fp16 = mul(x = var_4770_cast_fp16, y = var_4771_to_fp16)[name = string("op_4772_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4774_cast_fp16 = add(x = var_4768_cast_fp16, y = var_4772_cast_fp16)[name = string("op_4774_cast_fp16")];
            fp32 var_4775_epsilon_0 = const()[name = string("op_4775_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4775_cast_fp16 = rsqrt(epsilon = var_4775_epsilon_0, x = var_4774_cast_fp16)[name = string("op_4775_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_115_cast_fp16 = mul(x = y_135_cast_fp16, y = var_4775_cast_fp16)[name = string("out_115_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88149760)))];
            tensor<fp16, [1, 630, 768]> var_4777_cast_fp16 = mul(x = out_115_cast_fp16, y = t_encoder_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("op_4777_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_177_cast_fp16 = add(x = x_163_cast_fp16, y = var_4777_cast_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4780_cast_fp16 = abs(x = x_177_cast_fp16)[name = string("op_4780_cast_fp16")];
            bool var_4783 = const()[name = string("op_4783"), val = bool(true)];
            tensor<int32, [1]> var_4784_axes_0 = const()[name = string("op_4784_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4784_cast_fp16 = reduce_max(axes = var_4784_axes_0, keep_dims = var_4783, x = var_4780_cast_fp16)[name = string("op_4784_cast_fp16")];
            fp16 var_4785_to_fp16 = const()[name = string("op_4785_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_141_cast_fp16 = maximum(x = var_4784_cast_fp16, y = var_4785_to_fp16)[name = string("s_141_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_137_cast_fp16 = real_div(x = x_177_cast_fp16, y = s_141_cast_fp16)[name = string("y_137_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4788_cast_fp16 = mul(x = y_137_cast_fp16, y = y_137_cast_fp16)[name = string("op_4788_cast_fp16")];
            tensor<int32, [1]> var_4793_axes_0 = const()[name = string("op_4793_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4793_keep_dims_0 = const()[name = string("op_4793_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4793_cast_fp16 = reduce_mean(axes = var_4793_axes_0, keep_dims = var_4793_keep_dims_0, x = var_4788_cast_fp16)[name = string("op_4793_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4794_cast_fp16 = mul(x = s_141_cast_fp16, y = s_141_cast_fp16)[name = string("op_4794_cast_fp16")];
            fp32 var_4795_epsilon_0 = const()[name = string("op_4795_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4795_cast_fp16 = inverse(epsilon = var_4795_epsilon_0, x = var_4794_cast_fp16)[name = string("op_4795_cast_fp16")];
            fp16 var_4796_to_fp16 = const()[name = string("op_4796_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4797_cast_fp16 = mul(x = var_4795_cast_fp16, y = var_4796_to_fp16)[name = string("op_4797_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4799_cast_fp16 = add(x = var_4793_cast_fp16, y = var_4797_cast_fp16)[name = string("op_4799_cast_fp16")];
            fp32 var_4800_epsilon_0 = const()[name = string("op_4800_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4800_cast_fp16 = rsqrt(epsilon = var_4800_epsilon_0, x = var_4799_cast_fp16)[name = string("op_4800_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_117_cast_fp16 = mul(x = y_137_cast_fp16, y = var_4800_cast_fp16)[name = string("out_117_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88151360)))];
            tensor<fp16, [1, 630, 768]> input_97_cast_fp16 = mul(x = out_117_cast_fp16, y = t_encoder_layers_9_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(88152960))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90512320))))[name = string("t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_68_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_68_cast_fp16")];
            string var_4807_mode_0 = const()[name = string("op_4807_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_4807_cast_fp16 = gelu(mode = var_4807_mode_0, x = linear_68_cast_fp16)[name = string("op_4807_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(90518528))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92877888))))[name = string("t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_69_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_97_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_101_cast_fp16 = mul(x = var_4807_cast_fp16, y = linear_69_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92884096))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95243456))))[name = string("t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_9_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4815_cast_fp16 = abs(x = linear_70_cast_fp16)[name = string("op_4815_cast_fp16")];
            bool var_4818 = const()[name = string("op_4818"), val = bool(true)];
            tensor<int32, [1]> var_4819_axes_0 = const()[name = string("op_4819_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4819_cast_fp16 = reduce_max(axes = var_4819_axes_0, keep_dims = var_4818, x = var_4815_cast_fp16)[name = string("op_4819_cast_fp16")];
            fp16 var_4820_to_fp16 = const()[name = string("op_4820_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_143_cast_fp16 = maximum(x = var_4819_cast_fp16, y = var_4820_to_fp16)[name = string("s_143_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_139_cast_fp16 = real_div(x = linear_70_cast_fp16, y = s_143_cast_fp16)[name = string("y_139_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4823_cast_fp16 = mul(x = y_139_cast_fp16, y = y_139_cast_fp16)[name = string("op_4823_cast_fp16")];
            tensor<int32, [1]> var_4828_axes_0 = const()[name = string("op_4828_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4828_keep_dims_0 = const()[name = string("op_4828_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4828_cast_fp16 = reduce_mean(axes = var_4828_axes_0, keep_dims = var_4828_keep_dims_0, x = var_4823_cast_fp16)[name = string("op_4828_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4829_cast_fp16 = mul(x = s_143_cast_fp16, y = s_143_cast_fp16)[name = string("op_4829_cast_fp16")];
            fp32 var_4830_epsilon_0 = const()[name = string("op_4830_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4830_cast_fp16 = inverse(epsilon = var_4830_epsilon_0, x = var_4829_cast_fp16)[name = string("op_4830_cast_fp16")];
            fp16 var_4831_to_fp16 = const()[name = string("op_4831_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4832_cast_fp16 = mul(x = var_4830_cast_fp16, y = var_4831_to_fp16)[name = string("op_4832_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4834_cast_fp16 = add(x = var_4828_cast_fp16, y = var_4832_cast_fp16)[name = string("op_4834_cast_fp16")];
            fp32 var_4835_epsilon_0 = const()[name = string("op_4835_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4835_cast_fp16 = rsqrt(epsilon = var_4835_epsilon_0, x = var_4834_cast_fp16)[name = string("op_4835_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_119_cast_fp16 = mul(x = y_139_cast_fp16, y = var_4835_cast_fp16)[name = string("out_119_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95245056)))];
            tensor<fp16, [1, 630, 768]> var_4837_cast_fp16 = mul(x = out_119_cast_fp16, y = t_encoder_layers_9_post_feedforward_layernorm_weight_to_fp16)[name = string("op_4837_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_181_cast_fp16 = add(x = x_177_cast_fp16, y = var_4837_cast_fp16)[name = string("x_181_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4840_cast_fp16 = abs(x = x_181_cast_fp16)[name = string("op_4840_cast_fp16")];
            bool var_4843 = const()[name = string("op_4843"), val = bool(true)];
            tensor<int32, [1]> var_4844_axes_0 = const()[name = string("op_4844_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_4844_cast_fp16 = reduce_max(axes = var_4844_axes_0, keep_dims = var_4843, x = var_4840_cast_fp16)[name = string("op_4844_cast_fp16")];
            fp16 var_4845_to_fp16 = const()[name = string("op_4845_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_145_cast_fp16 = maximum(x = var_4844_cast_fp16, y = var_4845_to_fp16)[name = string("s_145_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_141_cast_fp16 = real_div(x = x_181_cast_fp16, y = s_145_cast_fp16)[name = string("y_141_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_4848_cast_fp16 = mul(x = y_141_cast_fp16, y = y_141_cast_fp16)[name = string("op_4848_cast_fp16")];
            tensor<int32, [1]> var_4853_axes_0 = const()[name = string("op_4853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4853_keep_dims_0 = const()[name = string("op_4853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_4853_cast_fp16 = reduce_mean(axes = var_4853_axes_0, keep_dims = var_4853_keep_dims_0, x = var_4848_cast_fp16)[name = string("op_4853_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4854_cast_fp16 = mul(x = s_145_cast_fp16, y = s_145_cast_fp16)[name = string("op_4854_cast_fp16")];
            fp32 var_4855_epsilon_0 = const()[name = string("op_4855_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_4855_cast_fp16 = inverse(epsilon = var_4855_epsilon_0, x = var_4854_cast_fp16)[name = string("op_4855_cast_fp16")];
            fp16 var_4856_to_fp16 = const()[name = string("op_4856_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_4857_cast_fp16 = mul(x = var_4855_cast_fp16, y = var_4856_to_fp16)[name = string("op_4857_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_4859_cast_fp16 = add(x = var_4853_cast_fp16, y = var_4857_cast_fp16)[name = string("op_4859_cast_fp16")];
            fp32 var_4860_epsilon_0 = const()[name = string("op_4860_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_4860_cast_fp16 = rsqrt(epsilon = var_4860_epsilon_0, x = var_4859_cast_fp16)[name = string("op_4860_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_121_cast_fp16 = mul(x = y_141_cast_fp16, y = var_4860_cast_fp16)[name = string("out_121_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95246656)))];
            tensor<fp16, [1, 630, 768]> input_103_cast_fp16 = mul(x = out_121_cast_fp16, y = t_encoder_layers_10_input_layernorm_weight_to_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95248256))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95838144))))[name = string("t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_71_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_4870 = const()[name = string("op_4870"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_183_cast_fp16 = reshape(shape = var_4870, x = linear_71_cast_fp16)[name = string("x_183_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4872_cast_fp16 = abs(x = x_183_cast_fp16)[name = string("op_4872_cast_fp16")];
            bool var_4875 = const()[name = string("op_4875"), val = bool(true)];
            tensor<int32, [1]> var_4876_axes_0 = const()[name = string("op_4876_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4876_cast_fp16 = reduce_max(axes = var_4876_axes_0, keep_dims = var_4875, x = var_4872_cast_fp16)[name = string("op_4876_cast_fp16")];
            fp16 var_4877_to_fp16 = const()[name = string("op_4877_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_147_cast_fp16 = maximum(x = var_4876_cast_fp16, y = var_4877_to_fp16)[name = string("s_147_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_143_cast_fp16 = real_div(x = x_183_cast_fp16, y = s_147_cast_fp16)[name = string("y_143_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4880_cast_fp16 = mul(x = y_143_cast_fp16, y = y_143_cast_fp16)[name = string("op_4880_cast_fp16")];
            tensor<int32, [1]> var_4885_axes_0 = const()[name = string("op_4885_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4885_keep_dims_0 = const()[name = string("op_4885_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4885_cast_fp16 = reduce_mean(axes = var_4885_axes_0, keep_dims = var_4885_keep_dims_0, x = var_4880_cast_fp16)[name = string("op_4885_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4886_cast_fp16 = mul(x = s_147_cast_fp16, y = s_147_cast_fp16)[name = string("op_4886_cast_fp16")];
            fp32 var_4887_epsilon_0 = const()[name = string("op_4887_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4887_cast_fp16 = inverse(epsilon = var_4887_epsilon_0, x = var_4886_cast_fp16)[name = string("op_4887_cast_fp16")];
            fp16 var_4888_to_fp16 = const()[name = string("op_4888_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4889_cast_fp16 = mul(x = var_4887_cast_fp16, y = var_4888_to_fp16)[name = string("op_4889_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4891_cast_fp16 = add(x = var_4885_cast_fp16, y = var_4889_cast_fp16)[name = string("op_4891_cast_fp16")];
            fp32 var_4892_epsilon_0 = const()[name = string("op_4892_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4892_cast_fp16 = rsqrt(epsilon = var_4892_epsilon_0, x = var_4891_cast_fp16)[name = string("op_4892_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_123_cast_fp16 = mul(x = y_143_cast_fp16, y = var_4892_cast_fp16)[name = string("out_123_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95839744)))];
            tensor<fp16, [1, 630, 12, 64]> x_189_cast_fp16 = mul(x = out_123_cast_fp16, y = t_encoder_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_189_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(95839936))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96429824))))[name = string("t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_4902 = const()[name = string("op_4902"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_185_cast_fp16 = reshape(shape = var_4902, x = linear_72_cast_fp16)[name = string("x_185_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4904_cast_fp16 = abs(x = x_185_cast_fp16)[name = string("op_4904_cast_fp16")];
            bool var_4907 = const()[name = string("op_4907"), val = bool(true)];
            tensor<int32, [1]> var_4908_axes_0 = const()[name = string("op_4908_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4908_cast_fp16 = reduce_max(axes = var_4908_axes_0, keep_dims = var_4907, x = var_4904_cast_fp16)[name = string("op_4908_cast_fp16")];
            fp16 var_4909_to_fp16 = const()[name = string("op_4909_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_149_cast_fp16 = maximum(x = var_4908_cast_fp16, y = var_4909_to_fp16)[name = string("s_149_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_145_cast_fp16 = real_div(x = x_185_cast_fp16, y = s_149_cast_fp16)[name = string("y_145_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4912_cast_fp16 = mul(x = y_145_cast_fp16, y = y_145_cast_fp16)[name = string("op_4912_cast_fp16")];
            tensor<int32, [1]> var_4917_axes_0 = const()[name = string("op_4917_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4917_keep_dims_0 = const()[name = string("op_4917_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4917_cast_fp16 = reduce_mean(axes = var_4917_axes_0, keep_dims = var_4917_keep_dims_0, x = var_4912_cast_fp16)[name = string("op_4917_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4918_cast_fp16 = mul(x = s_149_cast_fp16, y = s_149_cast_fp16)[name = string("op_4918_cast_fp16")];
            fp32 var_4919_epsilon_0 = const()[name = string("op_4919_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4919_cast_fp16 = inverse(epsilon = var_4919_epsilon_0, x = var_4918_cast_fp16)[name = string("op_4919_cast_fp16")];
            fp16 var_4920_to_fp16 = const()[name = string("op_4920_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = string("op_4921_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4923_cast_fp16 = add(x = var_4917_cast_fp16, y = var_4921_cast_fp16)[name = string("op_4923_cast_fp16")];
            fp32 var_4924_epsilon_0 = const()[name = string("op_4924_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4924_cast_fp16 = rsqrt(epsilon = var_4924_epsilon_0, x = var_4923_cast_fp16)[name = string("op_4924_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_125_cast_fp16 = mul(x = y_145_cast_fp16, y = var_4924_cast_fp16)[name = string("out_125_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96431424)))];
            tensor<fp16, [1, 630, 12, 64]> x_191_cast_fp16 = mul(x = out_125_cast_fp16, y = t_encoder_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_191_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96431616))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97021504))))[name = string("t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_73_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<int32, [4]> var_4934 = const()[name = string("op_4934"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_187_cast_fp16 = reshape(shape = var_4934, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4936_cast_fp16 = abs(x = x_187_cast_fp16)[name = string("op_4936_cast_fp16")];
            bool var_4939 = const()[name = string("op_4939"), val = bool(true)];
            tensor<int32, [1]> var_4940_axes_0 = const()[name = string("op_4940_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_4940_cast_fp16 = reduce_max(axes = var_4940_axes_0, keep_dims = var_4939, x = var_4936_cast_fp16)[name = string("op_4940_cast_fp16")];
            fp16 var_4941_to_fp16 = const()[name = string("op_4941_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_151_cast_fp16 = maximum(x = var_4940_cast_fp16, y = var_4941_to_fp16)[name = string("s_151_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_147_cast_fp16 = real_div(x = x_187_cast_fp16, y = s_151_cast_fp16)[name = string("y_147_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_4944_cast_fp16 = mul(x = y_147_cast_fp16, y = y_147_cast_fp16)[name = string("op_4944_cast_fp16")];
            tensor<int32, [1]> var_4949_axes_0 = const()[name = string("op_4949_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4949_keep_dims_0 = const()[name = string("op_4949_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_4949_cast_fp16 = reduce_mean(axes = var_4949_axes_0, keep_dims = var_4949_keep_dims_0, x = var_4944_cast_fp16)[name = string("op_4949_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4950_cast_fp16 = mul(x = s_151_cast_fp16, y = s_151_cast_fp16)[name = string("op_4950_cast_fp16")];
            fp32 var_4951_epsilon_0 = const()[name = string("op_4951_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_4951_cast_fp16 = inverse(epsilon = var_4951_epsilon_0, x = var_4950_cast_fp16)[name = string("op_4951_cast_fp16")];
            fp16 var_4952_to_fp16 = const()[name = string("op_4952_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_4953_cast_fp16 = mul(x = var_4951_cast_fp16, y = var_4952_to_fp16)[name = string("op_4953_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_4955_cast_fp16 = add(x = var_4949_cast_fp16, y = var_4953_cast_fp16)[name = string("op_4955_cast_fp16")];
            fp32 var_4956_epsilon_0 = const()[name = string("op_4956_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_4956_cast_fp16 = rsqrt(epsilon = var_4956_epsilon_0, x = var_4955_cast_fp16)[name = string("op_4956_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_41_cast_fp16 = mul(x = y_147_cast_fp16, y = var_4956_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<int32, [4]> z_81_begin_0 = const()[name = string("z_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_81_end_0 = const()[name = string("z_81_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_81_end_mask_0 = const()[name = string("z_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_81_cast_fp16 = slice_by_index(begin = z_81_begin_0, end = z_81_end_0, end_mask = z_81_end_mask_0, x = x_189_cast_fp16)[name = string("z_81_cast_fp16")];
            tensor<int32, [4]> var_4967_begin_0 = const()[name = string("op_4967_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4967_end_0 = const()[name = string("op_4967_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4967_end_mask_0 = const()[name = string("op_4967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4967_cast_fp16 = slice_by_index(begin = var_4967_begin_0, end = var_4967_end_0, end_mask = var_4967_end_mask_0, x = z_81_cast_fp16)[name = string("op_4967_cast_fp16")];
            fp16 const_40_promoted_to_fp16 = const()[name = string("const_40_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4968_cast_fp16 = mul(x = var_4967_cast_fp16, y = const_40_promoted_to_fp16)[name = string("op_4968_cast_fp16")];
            tensor<int32, [4]> var_4973_begin_0 = const()[name = string("op_4973_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4973_end_0 = const()[name = string("op_4973_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4973_end_mask_0 = const()[name = string("op_4973_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, x = z_81_cast_fp16)[name = string("op_4973_cast_fp16")];
            int32 var_4975 = const()[name = string("op_4975"), val = int32(-1)];
            bool rot_81_interleave_0 = const()[name = string("rot_81_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_81_cast_fp16 = concat(axis = var_4975, interleave = rot_81_interleave_0, values = (var_4968_cast_fp16, var_4973_cast_fp16))[name = string("rot_81_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4977_cast_fp16 = mul(x = z_81_cast_fp16, y = c_1_cast_fp16)[name = string("op_4977_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4978_cast_fp16 = mul(x = rot_81_cast_fp16, y = s_9_cast_fp16)[name = string("op_4978_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_4980_cast_fp16 = add(x = var_4977_cast_fp16, y = var_4978_cast_fp16)[name = string("op_4980_cast_fp16")];
            tensor<int32, [4]> z_83_begin_0 = const()[name = string("z_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_83_end_0 = const()[name = string("z_83_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_83_end_mask_0 = const()[name = string("z_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_83_cast_fp16 = slice_by_index(begin = z_83_begin_0, end = z_83_end_0, end_mask = z_83_end_mask_0, x = x_189_cast_fp16)[name = string("z_83_cast_fp16")];
            tensor<int32, [4]> var_4990_begin_0 = const()[name = string("op_4990_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_4990_end_0 = const()[name = string("op_4990_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_4990_end_mask_0 = const()[name = string("op_4990_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_4990_cast_fp16 = slice_by_index(begin = var_4990_begin_0, end = var_4990_end_0, end_mask = var_4990_end_mask_0, x = z_83_cast_fp16)[name = string("op_4990_cast_fp16")];
            fp16 const_41_promoted_to_fp16 = const()[name = string("const_41_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_4991_cast_fp16 = mul(x = var_4990_cast_fp16, y = const_41_promoted_to_fp16)[name = string("op_4991_cast_fp16")];
            tensor<int32, [4]> var_4996_begin_0 = const()[name = string("op_4996_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4996_end_0 = const()[name = string("op_4996_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_4996_end_mask_0 = const()[name = string("op_4996_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_4996_cast_fp16 = slice_by_index(begin = var_4996_begin_0, end = var_4996_end_0, end_mask = var_4996_end_mask_0, x = z_83_cast_fp16)[name = string("op_4996_cast_fp16")];
            int32 var_4998 = const()[name = string("op_4998"), val = int32(-1)];
            bool rot_83_interleave_0 = const()[name = string("rot_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_83_cast_fp16 = concat(axis = var_4998, interleave = rot_83_interleave_0, values = (var_4991_cast_fp16, var_4996_cast_fp16))[name = string("rot_83_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5000_cast_fp16 = mul(x = z_83_cast_fp16, y = c_cast_fp16)[name = string("op_5000_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5001_cast_fp16 = mul(x = rot_83_cast_fp16, y = s_11_cast_fp16)[name = string("op_5001_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5003_cast_fp16 = add(x = var_5000_cast_fp16, y = var_5001_cast_fp16)[name = string("op_5003_cast_fp16")];
            int32 var_5005 = const()[name = string("op_5005"), val = int32(-1)];
            bool var_5006_interleave_0 = const()[name = string("op_5006_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_5006_cast_fp16 = concat(axis = var_5005, interleave = var_5006_interleave_0, values = (var_4980_cast_fp16, var_5003_cast_fp16))[name = string("op_5006_cast_fp16")];
            tensor<int32, [4]> z_85_begin_0 = const()[name = string("z_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_85_end_0 = const()[name = string("z_85_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_85_end_mask_0 = const()[name = string("z_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_85_cast_fp16 = slice_by_index(begin = z_85_begin_0, end = z_85_end_0, end_mask = z_85_end_mask_0, x = x_191_cast_fp16)[name = string("z_85_cast_fp16")];
            tensor<int32, [4]> var_5019_begin_0 = const()[name = string("op_5019_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5019_end_0 = const()[name = string("op_5019_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5019_end_mask_0 = const()[name = string("op_5019_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5019_cast_fp16 = slice_by_index(begin = var_5019_begin_0, end = var_5019_end_0, end_mask = var_5019_end_mask_0, x = z_85_cast_fp16)[name = string("op_5019_cast_fp16")];
            fp16 const_42_promoted_to_fp16 = const()[name = string("const_42_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5020_cast_fp16 = mul(x = var_5019_cast_fp16, y = const_42_promoted_to_fp16)[name = string("op_5020_cast_fp16")];
            tensor<int32, [4]> var_5025_begin_0 = const()[name = string("op_5025_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5025_end_0 = const()[name = string("op_5025_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5025_end_mask_0 = const()[name = string("op_5025_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5025_cast_fp16 = slice_by_index(begin = var_5025_begin_0, end = var_5025_end_0, end_mask = var_5025_end_mask_0, x = z_85_cast_fp16)[name = string("op_5025_cast_fp16")];
            int32 var_5027 = const()[name = string("op_5027"), val = int32(-1)];
            bool rot_85_interleave_0 = const()[name = string("rot_85_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_85_cast_fp16 = concat(axis = var_5027, interleave = rot_85_interleave_0, values = (var_5020_cast_fp16, var_5025_cast_fp16))[name = string("rot_85_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5029_cast_fp16 = mul(x = z_85_cast_fp16, y = c_1_cast_fp16)[name = string("op_5029_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5030_cast_fp16 = mul(x = rot_85_cast_fp16, y = s_9_cast_fp16)[name = string("op_5030_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5032_cast_fp16 = add(x = var_5029_cast_fp16, y = var_5030_cast_fp16)[name = string("op_5032_cast_fp16")];
            tensor<int32, [4]> z_87_begin_0 = const()[name = string("z_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_87_end_0 = const()[name = string("z_87_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_87_end_mask_0 = const()[name = string("z_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_87_cast_fp16 = slice_by_index(begin = z_87_begin_0, end = z_87_end_0, end_mask = z_87_end_mask_0, x = x_191_cast_fp16)[name = string("z_87_cast_fp16")];
            tensor<int32, [4]> var_5042_begin_0 = const()[name = string("op_5042_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5042_end_0 = const()[name = string("op_5042_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5042_end_mask_0 = const()[name = string("op_5042_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5042_cast_fp16 = slice_by_index(begin = var_5042_begin_0, end = var_5042_end_0, end_mask = var_5042_end_mask_0, x = z_87_cast_fp16)[name = string("op_5042_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5043_cast_fp16 = mul(x = var_5042_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_5043_cast_fp16")];
            tensor<int32, [4]> var_5048_begin_0 = const()[name = string("op_5048_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5048_end_0 = const()[name = string("op_5048_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5048_end_mask_0 = const()[name = string("op_5048_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5048_cast_fp16 = slice_by_index(begin = var_5048_begin_0, end = var_5048_end_0, end_mask = var_5048_end_mask_0, x = z_87_cast_fp16)[name = string("op_5048_cast_fp16")];
            int32 var_5050 = const()[name = string("op_5050"), val = int32(-1)];
            bool rot_87_interleave_0 = const()[name = string("rot_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_87_cast_fp16 = concat(axis = var_5050, interleave = rot_87_interleave_0, values = (var_5043_cast_fp16, var_5048_cast_fp16))[name = string("rot_87_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5052_cast_fp16 = mul(x = z_87_cast_fp16, y = c_cast_fp16)[name = string("op_5052_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5053_cast_fp16 = mul(x = rot_87_cast_fp16, y = s_11_cast_fp16)[name = string("op_5053_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5055_cast_fp16 = add(x = var_5052_cast_fp16, y = var_5053_cast_fp16)[name = string("op_5055_cast_fp16")];
            int32 var_5057 = const()[name = string("op_5057"), val = int32(-1)];
            bool var_5058_interleave_0 = const()[name = string("op_5058_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_5058_cast_fp16 = concat(axis = var_5057, interleave = var_5058_interleave_0, values = (var_5032_cast_fp16, var_5055_cast_fp16))[name = string("op_5058_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5068_transpose_x_0 = const()[name = string("op_5068_transpose_x_0"), val = bool(false)];
            bool var_5068_transpose_y_0 = const()[name = string("op_5068_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_85 = transpose(perm = transpose_85_perm_0, x = var_5058_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 12, 630, 64]> transpose_84 = transpose(perm = transpose_84_perm_0, x = var_5006_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 12, 630, 630]> var_5068_cast_fp16 = matmul(transpose_x = var_5068_transpose_x_0, transpose_y = var_5068_transpose_y_0, x = transpose_84, y = transpose_85)[name = string("op_5068_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_5070_cast_fp16 = add(x = var_5068_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5070_cast_fp16")];
            int32 var_5071 = const()[name = string("op_5071"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_147_cast_fp16 = softmax(axis = var_5071, x = var_5070_cast_fp16)[name = string("w_147_cast_fp16")];
            bool var_5074_transpose_x_0 = const()[name = string("op_5074_transpose_x_0"), val = bool(false)];
            bool var_5074_transpose_y_0 = const()[name = string("op_5074_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = v_41_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 12, 630, 64]> var_5074_cast_fp16 = matmul(transpose_x = var_5074_transpose_x_0, transpose_y = var_5074_transpose_y_0, x = w_147_cast_fp16, y = v_43_cast_fp16)[name = string("op_5074_cast_fp16")];
            tensor<int32, [4]> var_5077_perm_0 = const()[name = string("op_5077_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5081 = const()[name = string("op_5081"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_5077_cast_fp16 = transpose(perm = var_5077_perm_0, x = var_5074_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 630, 768]> input_105_cast_fp16 = reshape(shape = var_5081, x = var_5077_cast_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97023104))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97612992))))[name = string("t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5086_cast_fp16 = abs(x = linear_74_cast_fp16)[name = string("op_5086_cast_fp16")];
            bool var_5089 = const()[name = string("op_5089"), val = bool(true)];
            tensor<int32, [1]> var_5090_axes_0 = const()[name = string("op_5090_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5090_cast_fp16 = reduce_max(axes = var_5090_axes_0, keep_dims = var_5089, x = var_5086_cast_fp16)[name = string("op_5090_cast_fp16")];
            fp16 var_5091_to_fp16 = const()[name = string("op_5091_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_153_cast_fp16 = maximum(x = var_5090_cast_fp16, y = var_5091_to_fp16)[name = string("s_153_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_149_cast_fp16 = real_div(x = linear_74_cast_fp16, y = s_153_cast_fp16)[name = string("y_149_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5094_cast_fp16 = mul(x = y_149_cast_fp16, y = y_149_cast_fp16)[name = string("op_5094_cast_fp16")];
            tensor<int32, [1]> var_5099_axes_0 = const()[name = string("op_5099_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5099_keep_dims_0 = const()[name = string("op_5099_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5099_cast_fp16 = reduce_mean(axes = var_5099_axes_0, keep_dims = var_5099_keep_dims_0, x = var_5094_cast_fp16)[name = string("op_5099_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5100_cast_fp16 = mul(x = s_153_cast_fp16, y = s_153_cast_fp16)[name = string("op_5100_cast_fp16")];
            fp32 var_5101_epsilon_0 = const()[name = string("op_5101_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5101_cast_fp16 = inverse(epsilon = var_5101_epsilon_0, x = var_5100_cast_fp16)[name = string("op_5101_cast_fp16")];
            fp16 var_5102_to_fp16 = const()[name = string("op_5102_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5103_cast_fp16 = mul(x = var_5101_cast_fp16, y = var_5102_to_fp16)[name = string("op_5103_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5105_cast_fp16 = add(x = var_5099_cast_fp16, y = var_5103_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, 630, 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, 630, 768]> out_127_cast_fp16 = mul(x = y_149_cast_fp16, y = var_5106_cast_fp16)[name = string("out_127_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97614592)))];
            tensor<fp16, [1, 630, 768]> var_5108_cast_fp16 = mul(x = out_127_cast_fp16, y = t_encoder_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("op_5108_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_195_cast_fp16 = add(x = x_181_cast_fp16, y = var_5108_cast_fp16)[name = string("x_195_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5111_cast_fp16 = abs(x = x_195_cast_fp16)[name = string("op_5111_cast_fp16")];
            bool var_5114 = const()[name = string("op_5114"), val = bool(true)];
            tensor<int32, [1]> var_5115_axes_0 = const()[name = string("op_5115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5115_cast_fp16 = reduce_max(axes = var_5115_axes_0, keep_dims = var_5114, x = var_5111_cast_fp16)[name = string("op_5115_cast_fp16")];
            fp16 var_5116_to_fp16 = const()[name = string("op_5116_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_155_cast_fp16 = maximum(x = var_5115_cast_fp16, y = var_5116_to_fp16)[name = string("s_155_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_151_cast_fp16 = real_div(x = x_195_cast_fp16, y = s_155_cast_fp16)[name = string("y_151_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5119_cast_fp16 = mul(x = y_151_cast_fp16, y = y_151_cast_fp16)[name = string("op_5119_cast_fp16")];
            tensor<int32, [1]> var_5124_axes_0 = const()[name = string("op_5124_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5124_keep_dims_0 = const()[name = string("op_5124_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5124_cast_fp16 = reduce_mean(axes = var_5124_axes_0, keep_dims = var_5124_keep_dims_0, x = var_5119_cast_fp16)[name = string("op_5124_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5125_cast_fp16 = mul(x = s_155_cast_fp16, y = s_155_cast_fp16)[name = string("op_5125_cast_fp16")];
            fp32 var_5126_epsilon_0 = const()[name = string("op_5126_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5126_cast_fp16 = inverse(epsilon = var_5126_epsilon_0, x = var_5125_cast_fp16)[name = string("op_5126_cast_fp16")];
            fp16 var_5127_to_fp16 = const()[name = string("op_5127_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5128_cast_fp16 = mul(x = var_5126_cast_fp16, y = var_5127_to_fp16)[name = string("op_5128_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5130_cast_fp16 = add(x = var_5124_cast_fp16, y = var_5128_cast_fp16)[name = string("op_5130_cast_fp16")];
            fp32 var_5131_epsilon_0 = const()[name = string("op_5131_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5131_cast_fp16 = rsqrt(epsilon = var_5131_epsilon_0, x = var_5130_cast_fp16)[name = string("op_5131_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_129_cast_fp16 = mul(x = y_151_cast_fp16, y = var_5131_cast_fp16)[name = string("out_129_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97616192)))];
            tensor<fp16, [1, 630, 768]> input_107_cast_fp16 = mul(x = out_129_cast_fp16, y = t_encoder_layers_10_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(97617792))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99977152))))[name = string("t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_75_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_75_cast_fp16")];
            string var_5138_mode_0 = const()[name = string("op_5138_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_5138_cast_fp16 = gelu(mode = var_5138_mode_0, x = linear_75_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(99983360))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102342720))))[name = string("t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_76_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_107_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_111_cast_fp16 = mul(x = var_5138_cast_fp16, y = linear_76_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102348928))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104708288))))[name = string("t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_10_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5146_cast_fp16 = abs(x = linear_77_cast_fp16)[name = string("op_5146_cast_fp16")];
            bool var_5149 = const()[name = string("op_5149"), val = bool(true)];
            tensor<int32, [1]> var_5150_axes_0 = const()[name = string("op_5150_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5150_cast_fp16 = reduce_max(axes = var_5150_axes_0, keep_dims = var_5149, x = var_5146_cast_fp16)[name = string("op_5150_cast_fp16")];
            fp16 var_5151_to_fp16 = const()[name = string("op_5151_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_157_cast_fp16 = maximum(x = var_5150_cast_fp16, y = var_5151_to_fp16)[name = string("s_157_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_153_cast_fp16 = real_div(x = linear_77_cast_fp16, y = s_157_cast_fp16)[name = string("y_153_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5154_cast_fp16 = mul(x = y_153_cast_fp16, y = y_153_cast_fp16)[name = string("op_5154_cast_fp16")];
            tensor<int32, [1]> var_5159_axes_0 = const()[name = string("op_5159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5159_keep_dims_0 = const()[name = string("op_5159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5159_cast_fp16 = reduce_mean(axes = var_5159_axes_0, keep_dims = var_5159_keep_dims_0, x = var_5154_cast_fp16)[name = string("op_5159_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5160_cast_fp16 = mul(x = s_157_cast_fp16, y = s_157_cast_fp16)[name = string("op_5160_cast_fp16")];
            fp32 var_5161_epsilon_0 = const()[name = string("op_5161_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5161_cast_fp16 = inverse(epsilon = var_5161_epsilon_0, x = var_5160_cast_fp16)[name = string("op_5161_cast_fp16")];
            fp16 var_5162_to_fp16 = const()[name = string("op_5162_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5163_cast_fp16 = mul(x = var_5161_cast_fp16, y = var_5162_to_fp16)[name = string("op_5163_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5165_cast_fp16 = add(x = var_5159_cast_fp16, y = var_5163_cast_fp16)[name = string("op_5165_cast_fp16")];
            fp32 var_5166_epsilon_0 = const()[name = string("op_5166_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5166_cast_fp16 = rsqrt(epsilon = var_5166_epsilon_0, x = var_5165_cast_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_131_cast_fp16 = mul(x = y_153_cast_fp16, y = var_5166_cast_fp16)[name = string("out_131_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104709888)))];
            tensor<fp16, [1, 630, 768]> var_5168_cast_fp16 = mul(x = out_131_cast_fp16, y = t_encoder_layers_10_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5168_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_199_cast_fp16 = add(x = x_195_cast_fp16, y = var_5168_cast_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5171_cast_fp16 = abs(x = x_199_cast_fp16)[name = string("op_5171_cast_fp16")];
            bool var_5174 = const()[name = string("op_5174"), val = bool(true)];
            tensor<int32, [1]> var_5175_axes_0 = const()[name = string("op_5175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5175_cast_fp16 = reduce_max(axes = var_5175_axes_0, keep_dims = var_5174, x = var_5171_cast_fp16)[name = string("op_5175_cast_fp16")];
            fp16 var_5176_to_fp16 = const()[name = string("op_5176_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_159_cast_fp16 = maximum(x = var_5175_cast_fp16, y = var_5176_to_fp16)[name = string("s_159_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_155_cast_fp16 = real_div(x = x_199_cast_fp16, y = s_159_cast_fp16)[name = string("y_155_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5179_cast_fp16 = mul(x = y_155_cast_fp16, y = y_155_cast_fp16)[name = string("op_5179_cast_fp16")];
            tensor<int32, [1]> var_5184_axes_0 = const()[name = string("op_5184_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5184_keep_dims_0 = const()[name = string("op_5184_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5184_cast_fp16 = reduce_mean(axes = var_5184_axes_0, keep_dims = var_5184_keep_dims_0, x = var_5179_cast_fp16)[name = string("op_5184_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5185_cast_fp16 = mul(x = s_159_cast_fp16, y = s_159_cast_fp16)[name = string("op_5185_cast_fp16")];
            fp32 var_5186_epsilon_0 = const()[name = string("op_5186_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5186_cast_fp16 = inverse(epsilon = var_5186_epsilon_0, x = var_5185_cast_fp16)[name = string("op_5186_cast_fp16")];
            fp16 var_5187_to_fp16 = const()[name = string("op_5187_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5188_cast_fp16 = mul(x = var_5186_cast_fp16, y = var_5187_to_fp16)[name = string("op_5188_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5190_cast_fp16 = add(x = var_5184_cast_fp16, y = var_5188_cast_fp16)[name = string("op_5190_cast_fp16")];
            fp32 var_5191_epsilon_0 = const()[name = string("op_5191_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5191_cast_fp16 = rsqrt(epsilon = var_5191_epsilon_0, x = var_5190_cast_fp16)[name = string("op_5191_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_133_cast_fp16 = mul(x = y_155_cast_fp16, y = var_5191_cast_fp16)[name = string("out_133_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104711488)))];
            tensor<fp16, [1, 630, 768]> input_113_cast_fp16 = mul(x = out_133_cast_fp16, y = t_encoder_layers_11_input_layernorm_weight_to_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104713088))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105302976))))[name = string("t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_5201 = const()[name = string("op_5201"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_201_cast_fp16 = reshape(shape = var_5201, x = linear_78_cast_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5203_cast_fp16 = abs(x = x_201_cast_fp16)[name = string("op_5203_cast_fp16")];
            bool var_5206 = const()[name = string("op_5206"), val = bool(true)];
            tensor<int32, [1]> var_5207_axes_0 = const()[name = string("op_5207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5207_cast_fp16 = reduce_max(axes = var_5207_axes_0, keep_dims = var_5206, x = var_5203_cast_fp16)[name = string("op_5207_cast_fp16")];
            fp16 var_5208_to_fp16 = const()[name = string("op_5208_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_161_cast_fp16 = maximum(x = var_5207_cast_fp16, y = var_5208_to_fp16)[name = string("s_161_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_157_cast_fp16 = real_div(x = x_201_cast_fp16, y = s_161_cast_fp16)[name = string("y_157_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5211_cast_fp16 = mul(x = y_157_cast_fp16, y = y_157_cast_fp16)[name = string("op_5211_cast_fp16")];
            tensor<int32, [1]> var_5216_axes_0 = const()[name = string("op_5216_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5216_keep_dims_0 = const()[name = string("op_5216_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5216_cast_fp16 = reduce_mean(axes = var_5216_axes_0, keep_dims = var_5216_keep_dims_0, x = var_5211_cast_fp16)[name = string("op_5216_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5217_cast_fp16 = mul(x = s_161_cast_fp16, y = s_161_cast_fp16)[name = string("op_5217_cast_fp16")];
            fp32 var_5218_epsilon_0 = const()[name = string("op_5218_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5218_cast_fp16 = inverse(epsilon = var_5218_epsilon_0, x = var_5217_cast_fp16)[name = string("op_5218_cast_fp16")];
            fp16 var_5219_to_fp16 = const()[name = string("op_5219_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5220_cast_fp16 = mul(x = var_5218_cast_fp16, y = var_5219_to_fp16)[name = string("op_5220_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5222_cast_fp16 = add(x = var_5216_cast_fp16, y = var_5220_cast_fp16)[name = string("op_5222_cast_fp16")];
            fp32 var_5223_epsilon_0 = const()[name = string("op_5223_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5223_cast_fp16 = rsqrt(epsilon = var_5223_epsilon_0, x = var_5222_cast_fp16)[name = string("op_5223_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_135_cast_fp16 = mul(x = y_157_cast_fp16, y = var_5223_cast_fp16)[name = string("out_135_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20121856)))];
            tensor<fp16, [1, 630, 12, 64]> x_207_cast_fp16 = mul(x = out_135_cast_fp16, y = t_encoder_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_207_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105304576))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105894464))))[name = string("t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_79_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_5233 = const()[name = string("op_5233"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_203_cast_fp16 = reshape(shape = var_5233, x = linear_79_cast_fp16)[name = string("x_203_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5235_cast_fp16 = abs(x = x_203_cast_fp16)[name = string("op_5235_cast_fp16")];
            bool var_5238 = const()[name = string("op_5238"), val = bool(true)];
            tensor<int32, [1]> var_5239_axes_0 = const()[name = string("op_5239_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5239_cast_fp16 = reduce_max(axes = var_5239_axes_0, keep_dims = var_5238, x = var_5235_cast_fp16)[name = string("op_5239_cast_fp16")];
            fp16 var_5240_to_fp16 = const()[name = string("op_5240_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_163_cast_fp16 = maximum(x = var_5239_cast_fp16, y = var_5240_to_fp16)[name = string("s_163_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_159_cast_fp16 = real_div(x = x_203_cast_fp16, y = s_163_cast_fp16)[name = string("y_159_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5243_cast_fp16 = mul(x = y_159_cast_fp16, y = y_159_cast_fp16)[name = string("op_5243_cast_fp16")];
            tensor<int32, [1]> var_5248_axes_0 = const()[name = string("op_5248_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5248_keep_dims_0 = const()[name = string("op_5248_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5248_cast_fp16 = reduce_mean(axes = var_5248_axes_0, keep_dims = var_5248_keep_dims_0, x = var_5243_cast_fp16)[name = string("op_5248_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5249_cast_fp16 = mul(x = s_163_cast_fp16, y = s_163_cast_fp16)[name = string("op_5249_cast_fp16")];
            fp32 var_5250_epsilon_0 = const()[name = string("op_5250_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5250_cast_fp16 = inverse(epsilon = var_5250_epsilon_0, x = var_5249_cast_fp16)[name = string("op_5250_cast_fp16")];
            fp16 var_5251_to_fp16 = const()[name = string("op_5251_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5252_cast_fp16 = mul(x = var_5250_cast_fp16, y = var_5251_to_fp16)[name = string("op_5252_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5254_cast_fp16 = add(x = var_5248_cast_fp16, y = var_5252_cast_fp16)[name = string("op_5254_cast_fp16")];
            fp32 var_5255_epsilon_0 = const()[name = string("op_5255_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5255_cast_fp16 = rsqrt(epsilon = var_5255_epsilon_0, x = var_5254_cast_fp16)[name = string("op_5255_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_137_cast_fp16 = mul(x = y_159_cast_fp16, y = var_5255_cast_fp16)[name = string("out_137_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20713536)))];
            tensor<fp16, [1, 630, 12, 64]> x_209_cast_fp16 = mul(x = out_137_cast_fp16, y = t_encoder_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_209_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(105896064))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106485952))))[name = string("t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<int32, [4]> var_5265 = const()[name = string("op_5265"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_205_cast_fp16 = reshape(shape = var_5265, x = linear_80_cast_fp16)[name = string("x_205_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5267_cast_fp16 = abs(x = x_205_cast_fp16)[name = string("op_5267_cast_fp16")];
            bool var_5270 = const()[name = string("op_5270"), val = bool(true)];
            tensor<int32, [1]> var_5271_axes_0 = const()[name = string("op_5271_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5271_cast_fp16 = reduce_max(axes = var_5271_axes_0, keep_dims = var_5270, x = var_5267_cast_fp16)[name = string("op_5271_cast_fp16")];
            fp16 var_5272_to_fp16 = const()[name = string("op_5272_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_165_cast_fp16 = maximum(x = var_5271_cast_fp16, y = var_5272_to_fp16)[name = string("s_165_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_161_cast_fp16 = real_div(x = x_205_cast_fp16, y = s_165_cast_fp16)[name = string("y_161_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5275_cast_fp16 = mul(x = y_161_cast_fp16, y = y_161_cast_fp16)[name = string("op_5275_cast_fp16")];
            tensor<int32, [1]> var_5280_axes_0 = const()[name = string("op_5280_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5280_keep_dims_0 = const()[name = string("op_5280_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5280_cast_fp16 = reduce_mean(axes = var_5280_axes_0, keep_dims = var_5280_keep_dims_0, x = var_5275_cast_fp16)[name = string("op_5280_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5281_cast_fp16 = mul(x = s_165_cast_fp16, y = s_165_cast_fp16)[name = string("op_5281_cast_fp16")];
            fp32 var_5282_epsilon_0 = const()[name = string("op_5282_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5282_cast_fp16 = inverse(epsilon = var_5282_epsilon_0, x = var_5281_cast_fp16)[name = string("op_5282_cast_fp16")];
            fp16 var_5283_to_fp16 = const()[name = string("op_5283_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5284_cast_fp16 = mul(x = var_5282_cast_fp16, y = var_5283_to_fp16)[name = string("op_5284_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5286_cast_fp16 = add(x = var_5280_cast_fp16, y = var_5284_cast_fp16)[name = string("op_5286_cast_fp16")];
            fp32 var_5287_epsilon_0 = const()[name = string("op_5287_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5287_cast_fp16 = rsqrt(epsilon = var_5287_epsilon_0, x = var_5286_cast_fp16)[name = string("op_5287_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_45_cast_fp16 = mul(x = y_161_cast_fp16, y = var_5287_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<int32, [4]> z_89_begin_0 = const()[name = string("z_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_89_end_0 = const()[name = string("z_89_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_89_end_mask_0 = const()[name = string("z_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_89_cast_fp16 = slice_by_index(begin = z_89_begin_0, end = z_89_end_0, end_mask = z_89_end_mask_0, x = x_207_cast_fp16)[name = string("z_89_cast_fp16")];
            tensor<int32, [4]> var_5298_begin_0 = const()[name = string("op_5298_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5298_end_0 = const()[name = string("op_5298_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5298_end_mask_0 = const()[name = string("op_5298_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5298_cast_fp16 = slice_by_index(begin = var_5298_begin_0, end = var_5298_end_0, end_mask = var_5298_end_mask_0, x = z_89_cast_fp16)[name = string("op_5298_cast_fp16")];
            fp16 const_44_promoted_to_fp16 = const()[name = string("const_44_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5299_cast_fp16 = mul(x = var_5298_cast_fp16, y = const_44_promoted_to_fp16)[name = string("op_5299_cast_fp16")];
            tensor<int32, [4]> var_5304_begin_0 = const()[name = string("op_5304_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5304_end_0 = const()[name = string("op_5304_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5304_end_mask_0 = const()[name = string("op_5304_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5304_cast_fp16 = slice_by_index(begin = var_5304_begin_0, end = var_5304_end_0, end_mask = var_5304_end_mask_0, x = z_89_cast_fp16)[name = string("op_5304_cast_fp16")];
            int32 var_5306 = const()[name = string("op_5306"), val = int32(-1)];
            bool rot_89_interleave_0 = const()[name = string("rot_89_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_89_cast_fp16 = concat(axis = var_5306, interleave = rot_89_interleave_0, values = (var_5299_cast_fp16, var_5304_cast_fp16))[name = string("rot_89_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5308_cast_fp16 = mul(x = z_89_cast_fp16, y = c_1_cast_fp16)[name = string("op_5308_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5309_cast_fp16 = mul(x = rot_89_cast_fp16, y = s_9_cast_fp16)[name = string("op_5309_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5311_cast_fp16 = add(x = var_5308_cast_fp16, y = var_5309_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<int32, [4]> z_91_begin_0 = const()[name = string("z_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_91_end_0 = const()[name = string("z_91_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_91_end_mask_0 = const()[name = string("z_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_91_cast_fp16 = slice_by_index(begin = z_91_begin_0, end = z_91_end_0, end_mask = z_91_end_mask_0, x = x_207_cast_fp16)[name = string("z_91_cast_fp16")];
            tensor<int32, [4]> var_5321_begin_0 = const()[name = string("op_5321_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5321_end_0 = const()[name = string("op_5321_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5321_end_mask_0 = const()[name = string("op_5321_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5321_cast_fp16 = slice_by_index(begin = var_5321_begin_0, end = var_5321_end_0, end_mask = var_5321_end_mask_0, x = z_91_cast_fp16)[name = string("op_5321_cast_fp16")];
            fp16 const_45_promoted_to_fp16 = const()[name = string("const_45_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5322_cast_fp16 = mul(x = var_5321_cast_fp16, y = const_45_promoted_to_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, 0])];
            tensor<int32, [4]> var_5327_end_0 = const()[name = string("op_5327_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5327_end_mask_0 = const()[name = string("op_5327_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 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 = z_91_cast_fp16)[name = string("op_5327_cast_fp16")];
            int32 var_5329 = const()[name = string("op_5329"), val = int32(-1)];
            bool rot_91_interleave_0 = const()[name = string("rot_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_91_cast_fp16 = concat(axis = var_5329, interleave = rot_91_interleave_0, values = (var_5322_cast_fp16, var_5327_cast_fp16))[name = string("rot_91_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5331_cast_fp16 = mul(x = z_91_cast_fp16, y = c_cast_fp16)[name = string("op_5331_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5332_cast_fp16 = mul(x = rot_91_cast_fp16, y = s_11_cast_fp16)[name = string("op_5332_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5334_cast_fp16 = add(x = var_5331_cast_fp16, y = var_5332_cast_fp16)[name = string("op_5334_cast_fp16")];
            int32 var_5336 = const()[name = string("op_5336"), val = int32(-1)];
            bool var_5337_interleave_0 = const()[name = string("op_5337_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_5337_cast_fp16 = concat(axis = var_5336, interleave = var_5337_interleave_0, values = (var_5311_cast_fp16, var_5334_cast_fp16))[name = string("op_5337_cast_fp16")];
            tensor<int32, [4]> z_93_begin_0 = const()[name = string("z_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_93_end_0 = const()[name = string("z_93_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_93_end_mask_0 = const()[name = string("z_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_93_cast_fp16 = slice_by_index(begin = z_93_begin_0, end = z_93_end_0, end_mask = z_93_end_mask_0, x = x_209_cast_fp16)[name = string("z_93_cast_fp16")];
            tensor<int32, [4]> var_5350_begin_0 = const()[name = string("op_5350_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5350_end_0 = const()[name = string("op_5350_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5350_end_mask_0 = const()[name = string("op_5350_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5350_cast_fp16 = slice_by_index(begin = var_5350_begin_0, end = var_5350_end_0, end_mask = var_5350_end_mask_0, x = z_93_cast_fp16)[name = string("op_5350_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5351_cast_fp16 = mul(x = var_5350_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_5351_cast_fp16")];
            tensor<int32, [4]> var_5356_begin_0 = const()[name = string("op_5356_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5356_end_0 = const()[name = string("op_5356_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5356_end_mask_0 = const()[name = string("op_5356_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5356_cast_fp16 = slice_by_index(begin = var_5356_begin_0, end = var_5356_end_0, end_mask = var_5356_end_mask_0, x = z_93_cast_fp16)[name = string("op_5356_cast_fp16")];
            int32 var_5358 = const()[name = string("op_5358"), val = int32(-1)];
            bool rot_93_interleave_0 = const()[name = string("rot_93_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_93_cast_fp16 = concat(axis = var_5358, interleave = rot_93_interleave_0, values = (var_5351_cast_fp16, var_5356_cast_fp16))[name = string("rot_93_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5360_cast_fp16 = mul(x = z_93_cast_fp16, y = c_1_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5361_cast_fp16 = mul(x = rot_93_cast_fp16, y = s_9_cast_fp16)[name = string("op_5361_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5363_cast_fp16 = add(x = var_5360_cast_fp16, y = var_5361_cast_fp16)[name = string("op_5363_cast_fp16")];
            tensor<int32, [4]> z_95_begin_0 = const()[name = string("z_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_95_end_0 = const()[name = string("z_95_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_95_end_mask_0 = const()[name = string("z_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_95_cast_fp16 = slice_by_index(begin = z_95_begin_0, end = z_95_end_0, end_mask = z_95_end_mask_0, x = x_209_cast_fp16)[name = string("z_95_cast_fp16")];
            tensor<int32, [4]> var_5373_begin_0 = const()[name = string("op_5373_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5373_end_0 = const()[name = string("op_5373_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5373_end_mask_0 = const()[name = string("op_5373_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5373_cast_fp16 = slice_by_index(begin = var_5373_begin_0, end = var_5373_end_0, end_mask = var_5373_end_mask_0, x = z_95_cast_fp16)[name = string("op_5373_cast_fp16")];
            fp16 const_47_promoted_to_fp16 = const()[name = string("const_47_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5374_cast_fp16 = mul(x = var_5373_cast_fp16, y = const_47_promoted_to_fp16)[name = string("op_5374_cast_fp16")];
            tensor<int32, [4]> var_5379_begin_0 = const()[name = string("op_5379_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5379_end_0 = const()[name = string("op_5379_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5379_end_mask_0 = const()[name = string("op_5379_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5379_cast_fp16 = slice_by_index(begin = var_5379_begin_0, end = var_5379_end_0, end_mask = var_5379_end_mask_0, x = z_95_cast_fp16)[name = string("op_5379_cast_fp16")];
            int32 var_5381 = const()[name = string("op_5381"), val = int32(-1)];
            bool rot_95_interleave_0 = const()[name = string("rot_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_95_cast_fp16 = concat(axis = var_5381, interleave = rot_95_interleave_0, values = (var_5374_cast_fp16, var_5379_cast_fp16))[name = string("rot_95_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5383_cast_fp16 = mul(x = z_95_cast_fp16, y = c_cast_fp16)[name = string("op_5383_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5384_cast_fp16 = mul(x = rot_95_cast_fp16, y = s_11_cast_fp16)[name = string("op_5384_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5386_cast_fp16 = add(x = var_5383_cast_fp16, y = var_5384_cast_fp16)[name = string("op_5386_cast_fp16")];
            int32 var_5388 = const()[name = string("op_5388"), val = int32(-1)];
            bool var_5389_interleave_0 = const()[name = string("op_5389_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_5389_cast_fp16 = concat(axis = var_5388, interleave = var_5389_interleave_0, values = (var_5363_cast_fp16, var_5386_cast_fp16))[name = string("op_5389_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5399_transpose_x_0 = const()[name = string("op_5399_transpose_x_0"), val = bool(false)];
            bool var_5399_transpose_y_0 = const()[name = string("op_5399_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_87 = transpose(perm = transpose_87_perm_0, x = var_5389_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 12, 630, 64]> transpose_86 = transpose(perm = transpose_86_perm_0, x = var_5337_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 12, 630, 630]> var_5399_cast_fp16 = matmul(transpose_x = var_5399_transpose_x_0, transpose_y = var_5399_transpose_y_0, x = transpose_86, y = transpose_87)[name = string("op_5399_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_5401_cast_fp16 = add(x = var_5399_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5401_cast_fp16")];
            int32 var_5402 = const()[name = string("op_5402"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_161_cast_fp16 = softmax(axis = var_5402, x = var_5401_cast_fp16)[name = string("w_161_cast_fp16")];
            bool var_5405_transpose_x_0 = const()[name = string("op_5405_transpose_x_0"), val = bool(false)];
            bool var_5405_transpose_y_0 = const()[name = string("op_5405_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = v_45_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 12, 630, 64]> var_5405_cast_fp16 = matmul(transpose_x = var_5405_transpose_x_0, transpose_y = var_5405_transpose_y_0, x = w_161_cast_fp16, y = v_47_cast_fp16)[name = string("op_5405_cast_fp16")];
            tensor<int32, [4]> var_5408_perm_0 = const()[name = string("op_5408_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5412 = const()[name = string("op_5412"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_5408_cast_fp16 = transpose(perm = var_5408_perm_0, x = var_5405_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 630, 768]> input_115_cast_fp16 = reshape(shape = var_5412, x = var_5408_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(106487552))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107077440))))[name = string("t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_81_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5417_cast_fp16 = abs(x = linear_81_cast_fp16)[name = string("op_5417_cast_fp16")];
            bool var_5420 = const()[name = string("op_5420"), val = bool(true)];
            tensor<int32, [1]> var_5421_axes_0 = const()[name = string("op_5421_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5421_cast_fp16 = reduce_max(axes = var_5421_axes_0, keep_dims = var_5420, x = var_5417_cast_fp16)[name = string("op_5421_cast_fp16")];
            fp16 var_5422_to_fp16 = const()[name = string("op_5422_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_167_cast_fp16 = maximum(x = var_5421_cast_fp16, y = var_5422_to_fp16)[name = string("s_167_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_163_cast_fp16 = real_div(x = linear_81_cast_fp16, y = s_167_cast_fp16)[name = string("y_163_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5425_cast_fp16 = mul(x = y_163_cast_fp16, y = y_163_cast_fp16)[name = string("op_5425_cast_fp16")];
            tensor<int32, [1]> var_5430_axes_0 = const()[name = string("op_5430_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5430_keep_dims_0 = const()[name = string("op_5430_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5430_cast_fp16 = reduce_mean(axes = var_5430_axes_0, keep_dims = var_5430_keep_dims_0, x = var_5425_cast_fp16)[name = string("op_5430_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5431_cast_fp16 = mul(x = s_167_cast_fp16, y = s_167_cast_fp16)[name = string("op_5431_cast_fp16")];
            fp32 var_5432_epsilon_0 = const()[name = string("op_5432_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5432_cast_fp16 = inverse(epsilon = var_5432_epsilon_0, x = var_5431_cast_fp16)[name = string("op_5432_cast_fp16")];
            fp16 var_5433_to_fp16 = const()[name = string("op_5433_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = string("op_5434_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5436_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5434_cast_fp16)[name = string("op_5436_cast_fp16")];
            fp32 var_5437_epsilon_0 = const()[name = string("op_5437_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5437_cast_fp16 = rsqrt(epsilon = var_5437_epsilon_0, x = var_5436_cast_fp16)[name = string("op_5437_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_139_cast_fp16 = mul(x = y_163_cast_fp16, y = var_5437_cast_fp16)[name = string("out_139_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107079040)))];
            tensor<fp16, [1, 630, 768]> var_5439_cast_fp16 = mul(x = out_139_cast_fp16, y = t_encoder_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("op_5439_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_213_cast_fp16 = add(x = x_199_cast_fp16, y = var_5439_cast_fp16)[name = string("x_213_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5442_cast_fp16 = abs(x = x_213_cast_fp16)[name = string("op_5442_cast_fp16")];
            bool var_5445 = const()[name = string("op_5445"), val = bool(true)];
            tensor<int32, [1]> var_5446_axes_0 = const()[name = string("op_5446_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5446_cast_fp16 = reduce_max(axes = var_5446_axes_0, keep_dims = var_5445, x = var_5442_cast_fp16)[name = string("op_5446_cast_fp16")];
            fp16 var_5447_to_fp16 = const()[name = string("op_5447_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_169_cast_fp16 = maximum(x = var_5446_cast_fp16, y = var_5447_to_fp16)[name = string("s_169_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_165_cast_fp16 = real_div(x = x_213_cast_fp16, y = s_169_cast_fp16)[name = string("y_165_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5450_cast_fp16 = mul(x = y_165_cast_fp16, y = y_165_cast_fp16)[name = string("op_5450_cast_fp16")];
            tensor<int32, [1]> var_5455_axes_0 = const()[name = string("op_5455_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5455_keep_dims_0 = const()[name = string("op_5455_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5455_cast_fp16 = reduce_mean(axes = var_5455_axes_0, keep_dims = var_5455_keep_dims_0, x = var_5450_cast_fp16)[name = string("op_5455_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5456_cast_fp16 = mul(x = s_169_cast_fp16, y = s_169_cast_fp16)[name = string("op_5456_cast_fp16")];
            fp32 var_5457_epsilon_0 = const()[name = string("op_5457_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5457_cast_fp16 = inverse(epsilon = var_5457_epsilon_0, x = var_5456_cast_fp16)[name = string("op_5457_cast_fp16")];
            fp16 var_5458_to_fp16 = const()[name = string("op_5458_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5459_cast_fp16 = mul(x = var_5457_cast_fp16, y = var_5458_to_fp16)[name = string("op_5459_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5461_cast_fp16 = add(x = var_5455_cast_fp16, y = var_5459_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, 630, 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, 630, 768]> out_141_cast_fp16 = mul(x = y_165_cast_fp16, y = var_5462_cast_fp16)[name = string("out_141_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107080640)))];
            tensor<fp16, [1, 630, 768]> input_117_cast_fp16 = mul(x = out_141_cast_fp16, y = t_encoder_layers_11_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107082240))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109441600))))[name = string("t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_82_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_82_cast_fp16")];
            string var_5469_mode_0 = const()[name = string("op_5469_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_5469_cast_fp16 = gelu(mode = var_5469_mode_0, x = linear_82_cast_fp16)[name = string("op_5469_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109447808))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111807168))))[name = string("t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_83_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_117_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_121_cast_fp16 = mul(x = var_5469_cast_fp16, y = linear_83_cast_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(111813376))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114172736))))[name = string("t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_11_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5477_cast_fp16 = abs(x = linear_84_cast_fp16)[name = string("op_5477_cast_fp16")];
            bool var_5480 = const()[name = string("op_5480"), val = bool(true)];
            tensor<int32, [1]> var_5481_axes_0 = const()[name = string("op_5481_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5481_cast_fp16 = reduce_max(axes = var_5481_axes_0, keep_dims = var_5480, x = var_5477_cast_fp16)[name = string("op_5481_cast_fp16")];
            fp16 var_5482_to_fp16 = const()[name = string("op_5482_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_171_cast_fp16 = maximum(x = var_5481_cast_fp16, y = var_5482_to_fp16)[name = string("s_171_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_167_cast_fp16 = real_div(x = linear_84_cast_fp16, y = s_171_cast_fp16)[name = string("y_167_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5485_cast_fp16 = mul(x = y_167_cast_fp16, y = y_167_cast_fp16)[name = string("op_5485_cast_fp16")];
            tensor<int32, [1]> var_5490_axes_0 = const()[name = string("op_5490_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5490_keep_dims_0 = const()[name = string("op_5490_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5490_cast_fp16 = reduce_mean(axes = var_5490_axes_0, keep_dims = var_5490_keep_dims_0, x = var_5485_cast_fp16)[name = string("op_5490_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5491_cast_fp16 = mul(x = s_171_cast_fp16, y = s_171_cast_fp16)[name = string("op_5491_cast_fp16")];
            fp32 var_5492_epsilon_0 = const()[name = string("op_5492_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5492_cast_fp16 = inverse(epsilon = var_5492_epsilon_0, x = var_5491_cast_fp16)[name = string("op_5492_cast_fp16")];
            fp16 var_5493_to_fp16 = const()[name = string("op_5493_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5494_cast_fp16 = mul(x = var_5492_cast_fp16, y = var_5493_to_fp16)[name = string("op_5494_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5496_cast_fp16 = add(x = var_5490_cast_fp16, y = var_5494_cast_fp16)[name = string("op_5496_cast_fp16")];
            fp32 var_5497_epsilon_0 = const()[name = string("op_5497_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5497_cast_fp16 = rsqrt(epsilon = var_5497_epsilon_0, x = var_5496_cast_fp16)[name = string("op_5497_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_143_cast_fp16 = mul(x = y_167_cast_fp16, y = var_5497_cast_fp16)[name = string("out_143_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114174336)))];
            tensor<fp16, [1, 630, 768]> var_5499_cast_fp16 = mul(x = out_143_cast_fp16, y = t_encoder_layers_11_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5499_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_217_cast_fp16 = add(x = x_213_cast_fp16, y = var_5499_cast_fp16)[name = string("x_217_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5502_cast_fp16 = abs(x = x_217_cast_fp16)[name = string("op_5502_cast_fp16")];
            bool var_5505 = const()[name = string("op_5505"), val = bool(true)];
            tensor<int32, [1]> var_5506_axes_0 = const()[name = string("op_5506_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5506_cast_fp16 = reduce_max(axes = var_5506_axes_0, keep_dims = var_5505, x = var_5502_cast_fp16)[name = string("op_5506_cast_fp16")];
            fp16 var_5507_to_fp16 = const()[name = string("op_5507_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_173_cast_fp16 = maximum(x = var_5506_cast_fp16, y = var_5507_to_fp16)[name = string("s_173_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_169_cast_fp16 = real_div(x = x_217_cast_fp16, y = s_173_cast_fp16)[name = string("y_169_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5510_cast_fp16 = mul(x = y_169_cast_fp16, y = y_169_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<int32, [1]> var_5515_axes_0 = const()[name = string("op_5515_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5515_keep_dims_0 = const()[name = string("op_5515_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5515_cast_fp16 = reduce_mean(axes = var_5515_axes_0, keep_dims = var_5515_keep_dims_0, x = var_5510_cast_fp16)[name = string("op_5515_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5516_cast_fp16 = mul(x = s_173_cast_fp16, y = s_173_cast_fp16)[name = string("op_5516_cast_fp16")];
            fp32 var_5517_epsilon_0 = const()[name = string("op_5517_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5517_cast_fp16 = inverse(epsilon = var_5517_epsilon_0, x = var_5516_cast_fp16)[name = string("op_5517_cast_fp16")];
            fp16 var_5518_to_fp16 = const()[name = string("op_5518_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5519_cast_fp16 = mul(x = var_5517_cast_fp16, y = var_5518_to_fp16)[name = string("op_5519_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5521_cast_fp16 = add(x = var_5515_cast_fp16, y = var_5519_cast_fp16)[name = string("op_5521_cast_fp16")];
            fp32 var_5522_epsilon_0 = const()[name = string("op_5522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5522_cast_fp16 = rsqrt(epsilon = var_5522_epsilon_0, x = var_5521_cast_fp16)[name = string("op_5522_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_145_cast_fp16 = mul(x = y_169_cast_fp16, y = var_5522_cast_fp16)[name = string("out_145_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114175936)))];
            tensor<fp16, [1, 630, 768]> input_123_cast_fp16 = mul(x = out_145_cast_fp16, y = t_encoder_layers_12_input_layernorm_weight_to_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114177536))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114767424))))[name = string("t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_85_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_5532 = const()[name = string("op_5532"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_219_cast_fp16 = reshape(shape = var_5532, x = linear_85_cast_fp16)[name = string("x_219_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5534_cast_fp16 = abs(x = x_219_cast_fp16)[name = string("op_5534_cast_fp16")];
            bool var_5537 = const()[name = string("op_5537"), val = bool(true)];
            tensor<int32, [1]> var_5538_axes_0 = const()[name = string("op_5538_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5538_cast_fp16 = reduce_max(axes = var_5538_axes_0, keep_dims = var_5537, x = var_5534_cast_fp16)[name = string("op_5538_cast_fp16")];
            fp16 var_5539_to_fp16 = const()[name = string("op_5539_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_175_cast_fp16 = maximum(x = var_5538_cast_fp16, y = var_5539_to_fp16)[name = string("s_175_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_171_cast_fp16 = real_div(x = x_219_cast_fp16, y = s_175_cast_fp16)[name = string("y_171_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5542_cast_fp16 = mul(x = y_171_cast_fp16, y = y_171_cast_fp16)[name = string("op_5542_cast_fp16")];
            tensor<int32, [1]> var_5547_axes_0 = const()[name = string("op_5547_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5547_keep_dims_0 = const()[name = string("op_5547_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5547_cast_fp16 = reduce_mean(axes = var_5547_axes_0, keep_dims = var_5547_keep_dims_0, x = var_5542_cast_fp16)[name = string("op_5547_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5548_cast_fp16 = mul(x = s_175_cast_fp16, y = s_175_cast_fp16)[name = string("op_5548_cast_fp16")];
            fp32 var_5549_epsilon_0 = const()[name = string("op_5549_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5549_cast_fp16 = inverse(epsilon = var_5549_epsilon_0, x = var_5548_cast_fp16)[name = string("op_5549_cast_fp16")];
            fp16 var_5550_to_fp16 = const()[name = string("op_5550_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5551_cast_fp16 = mul(x = var_5549_cast_fp16, y = var_5550_to_fp16)[name = string("op_5551_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5553_cast_fp16 = add(x = var_5547_cast_fp16, y = var_5551_cast_fp16)[name = string("op_5553_cast_fp16")];
            fp32 var_5554_epsilon_0 = const()[name = string("op_5554_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5554_cast_fp16 = rsqrt(epsilon = var_5554_epsilon_0, x = var_5553_cast_fp16)[name = string("op_5554_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_147_cast_fp16 = mul(x = y_171_cast_fp16, y = var_5554_cast_fp16)[name = string("out_147_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769024)))];
            tensor<fp16, [1, 630, 12, 64]> x_225_cast_fp16 = mul(x = out_147_cast_fp16, y = t_encoder_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_225_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769216))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115359104))))[name = string("t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_5564 = const()[name = string("op_5564"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_221_cast_fp16 = reshape(shape = var_5564, x = linear_86_cast_fp16)[name = string("x_221_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5566_cast_fp16 = abs(x = x_221_cast_fp16)[name = string("op_5566_cast_fp16")];
            bool var_5569 = const()[name = string("op_5569"), val = bool(true)];
            tensor<int32, [1]> var_5570_axes_0 = const()[name = string("op_5570_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5570_cast_fp16 = reduce_max(axes = var_5570_axes_0, keep_dims = var_5569, x = var_5566_cast_fp16)[name = string("op_5570_cast_fp16")];
            fp16 var_5571_to_fp16 = const()[name = string("op_5571_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_177_cast_fp16 = maximum(x = var_5570_cast_fp16, y = var_5571_to_fp16)[name = string("s_177_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_173_cast_fp16 = real_div(x = x_221_cast_fp16, y = s_177_cast_fp16)[name = string("y_173_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5574_cast_fp16 = mul(x = y_173_cast_fp16, y = y_173_cast_fp16)[name = string("op_5574_cast_fp16")];
            tensor<int32, [1]> var_5579_axes_0 = const()[name = string("op_5579_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5579_keep_dims_0 = const()[name = string("op_5579_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5579_cast_fp16 = reduce_mean(axes = var_5579_axes_0, keep_dims = var_5579_keep_dims_0, x = var_5574_cast_fp16)[name = string("op_5579_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5580_cast_fp16 = mul(x = s_177_cast_fp16, y = s_177_cast_fp16)[name = string("op_5580_cast_fp16")];
            fp32 var_5581_epsilon_0 = const()[name = string("op_5581_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5581_cast_fp16 = inverse(epsilon = var_5581_epsilon_0, x = var_5580_cast_fp16)[name = string("op_5581_cast_fp16")];
            fp16 var_5582_to_fp16 = const()[name = string("op_5582_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5583_cast_fp16 = mul(x = var_5581_cast_fp16, y = var_5582_to_fp16)[name = string("op_5583_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5585_cast_fp16 = add(x = var_5579_cast_fp16, y = var_5583_cast_fp16)[name = string("op_5585_cast_fp16")];
            fp32 var_5586_epsilon_0 = const()[name = string("op_5586_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5586_cast_fp16 = rsqrt(epsilon = var_5586_epsilon_0, x = var_5585_cast_fp16)[name = string("op_5586_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_149_cast_fp16 = mul(x = y_173_cast_fp16, y = var_5586_cast_fp16)[name = string("out_149_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360704)))];
            tensor<fp16, [1, 630, 12, 64]> x_227_cast_fp16 = mul(x = out_149_cast_fp16, y = t_encoder_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_227_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360896))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115950784))))[name = string("t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_87_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<int32, [4]> var_5596 = const()[name = string("op_5596"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_223_cast_fp16 = reshape(shape = var_5596, x = linear_87_cast_fp16)[name = string("x_223_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5598_cast_fp16 = abs(x = x_223_cast_fp16)[name = string("op_5598_cast_fp16")];
            bool var_5601 = const()[name = string("op_5601"), val = bool(true)];
            tensor<int32, [1]> var_5602_axes_0 = const()[name = string("op_5602_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5602_cast_fp16 = reduce_max(axes = var_5602_axes_0, keep_dims = var_5601, x = var_5598_cast_fp16)[name = string("op_5602_cast_fp16")];
            fp16 var_5603_to_fp16 = const()[name = string("op_5603_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_179_cast_fp16 = maximum(x = var_5602_cast_fp16, y = var_5603_to_fp16)[name = string("s_179_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_175_cast_fp16 = real_div(x = x_223_cast_fp16, y = s_179_cast_fp16)[name = string("y_175_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5606_cast_fp16 = mul(x = y_175_cast_fp16, y = y_175_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<int32, [1]> var_5611_axes_0 = const()[name = string("op_5611_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5611_keep_dims_0 = const()[name = string("op_5611_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5611_cast_fp16 = reduce_mean(axes = var_5611_axes_0, keep_dims = var_5611_keep_dims_0, x = var_5606_cast_fp16)[name = string("op_5611_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5612_cast_fp16 = mul(x = s_179_cast_fp16, y = s_179_cast_fp16)[name = string("op_5612_cast_fp16")];
            fp32 var_5613_epsilon_0 = const()[name = string("op_5613_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5613_cast_fp16 = inverse(epsilon = var_5613_epsilon_0, x = var_5612_cast_fp16)[name = string("op_5613_cast_fp16")];
            fp16 var_5614_to_fp16 = const()[name = string("op_5614_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5615_cast_fp16 = mul(x = var_5613_cast_fp16, y = var_5614_to_fp16)[name = string("op_5615_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5617_cast_fp16 = add(x = var_5611_cast_fp16, y = var_5615_cast_fp16)[name = string("op_5617_cast_fp16")];
            fp32 var_5618_epsilon_0 = const()[name = string("op_5618_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5618_cast_fp16 = rsqrt(epsilon = var_5618_epsilon_0, x = var_5617_cast_fp16)[name = string("op_5618_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_49_cast_fp16 = mul(x = y_175_cast_fp16, y = var_5618_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<int32, [4]> z_97_begin_0 = const()[name = string("z_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_97_end_0 = const()[name = string("z_97_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_97_end_mask_0 = const()[name = string("z_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_97_cast_fp16 = slice_by_index(begin = z_97_begin_0, end = z_97_end_0, end_mask = z_97_end_mask_0, x = x_225_cast_fp16)[name = string("z_97_cast_fp16")];
            tensor<int32, [4]> var_5629_begin_0 = const()[name = string("op_5629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5629_end_0 = const()[name = string("op_5629_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5629_end_mask_0 = const()[name = string("op_5629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> 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 = z_97_cast_fp16)[name = string("op_5629_cast_fp16")];
            fp16 const_48_promoted_to_fp16 = const()[name = string("const_48_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5630_cast_fp16 = mul(x = var_5629_cast_fp16, y = const_48_promoted_to_fp16)[name = string("op_5630_cast_fp16")];
            tensor<int32, [4]> var_5635_begin_0 = const()[name = string("op_5635_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5635_end_0 = const()[name = string("op_5635_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5635_end_mask_0 = const()[name = string("op_5635_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5635_cast_fp16 = slice_by_index(begin = var_5635_begin_0, end = var_5635_end_0, end_mask = var_5635_end_mask_0, x = z_97_cast_fp16)[name = string("op_5635_cast_fp16")];
            int32 var_5637 = const()[name = string("op_5637"), val = int32(-1)];
            bool rot_97_interleave_0 = const()[name = string("rot_97_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_97_cast_fp16 = concat(axis = var_5637, interleave = rot_97_interleave_0, values = (var_5630_cast_fp16, var_5635_cast_fp16))[name = string("rot_97_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5639_cast_fp16 = mul(x = z_97_cast_fp16, y = c_1_cast_fp16)[name = string("op_5639_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5640_cast_fp16 = mul(x = rot_97_cast_fp16, y = s_9_cast_fp16)[name = string("op_5640_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5642_cast_fp16 = add(x = var_5639_cast_fp16, y = var_5640_cast_fp16)[name = string("op_5642_cast_fp16")];
            tensor<int32, [4]> z_99_begin_0 = const()[name = string("z_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_99_end_0 = const()[name = string("z_99_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_99_end_mask_0 = const()[name = string("z_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_99_cast_fp16 = slice_by_index(begin = z_99_begin_0, end = z_99_end_0, end_mask = z_99_end_mask_0, x = x_225_cast_fp16)[name = string("z_99_cast_fp16")];
            tensor<int32, [4]> var_5652_begin_0 = const()[name = string("op_5652_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5652_end_0 = const()[name = string("op_5652_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5652_end_mask_0 = const()[name = string("op_5652_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5652_cast_fp16 = slice_by_index(begin = var_5652_begin_0, end = var_5652_end_0, end_mask = var_5652_end_mask_0, x = z_99_cast_fp16)[name = string("op_5652_cast_fp16")];
            fp16 const_49_promoted_to_fp16 = const()[name = string("const_49_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5653_cast_fp16 = mul(x = var_5652_cast_fp16, y = const_49_promoted_to_fp16)[name = string("op_5653_cast_fp16")];
            tensor<int32, [4]> var_5658_begin_0 = const()[name = string("op_5658_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5658_end_0 = const()[name = string("op_5658_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5658_end_mask_0 = const()[name = string("op_5658_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5658_cast_fp16 = slice_by_index(begin = var_5658_begin_0, end = var_5658_end_0, end_mask = var_5658_end_mask_0, x = z_99_cast_fp16)[name = string("op_5658_cast_fp16")];
            int32 var_5660 = const()[name = string("op_5660"), val = int32(-1)];
            bool rot_99_interleave_0 = const()[name = string("rot_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_99_cast_fp16 = concat(axis = var_5660, interleave = rot_99_interleave_0, values = (var_5653_cast_fp16, var_5658_cast_fp16))[name = string("rot_99_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5662_cast_fp16 = mul(x = z_99_cast_fp16, y = c_cast_fp16)[name = string("op_5662_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5663_cast_fp16 = mul(x = rot_99_cast_fp16, y = s_11_cast_fp16)[name = string("op_5663_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5665_cast_fp16 = add(x = var_5662_cast_fp16, y = var_5663_cast_fp16)[name = string("op_5665_cast_fp16")];
            int32 var_5667 = const()[name = string("op_5667"), val = int32(-1)];
            bool var_5668_interleave_0 = const()[name = string("op_5668_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_5668_cast_fp16 = concat(axis = var_5667, interleave = var_5668_interleave_0, values = (var_5642_cast_fp16, var_5665_cast_fp16))[name = string("op_5668_cast_fp16")];
            tensor<int32, [4]> z_101_begin_0 = const()[name = string("z_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_101_end_0 = const()[name = string("z_101_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_101_end_mask_0 = const()[name = string("z_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_101_cast_fp16 = slice_by_index(begin = z_101_begin_0, end = z_101_end_0, end_mask = z_101_end_mask_0, x = x_227_cast_fp16)[name = string("z_101_cast_fp16")];
            tensor<int32, [4]> var_5681_begin_0 = const()[name = string("op_5681_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5681_end_0 = const()[name = string("op_5681_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5681_end_mask_0 = const()[name = string("op_5681_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5681_cast_fp16 = slice_by_index(begin = var_5681_begin_0, end = var_5681_end_0, end_mask = var_5681_end_mask_0, x = z_101_cast_fp16)[name = string("op_5681_cast_fp16")];
            fp16 const_50_promoted_to_fp16 = const()[name = string("const_50_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5682_cast_fp16 = mul(x = var_5681_cast_fp16, y = const_50_promoted_to_fp16)[name = string("op_5682_cast_fp16")];
            tensor<int32, [4]> var_5687_begin_0 = const()[name = string("op_5687_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5687_end_0 = const()[name = string("op_5687_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5687_end_mask_0 = const()[name = string("op_5687_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5687_cast_fp16 = slice_by_index(begin = var_5687_begin_0, end = var_5687_end_0, end_mask = var_5687_end_mask_0, x = z_101_cast_fp16)[name = string("op_5687_cast_fp16")];
            int32 var_5689 = const()[name = string("op_5689"), val = int32(-1)];
            bool rot_101_interleave_0 = const()[name = string("rot_101_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_101_cast_fp16 = concat(axis = var_5689, interleave = rot_101_interleave_0, values = (var_5682_cast_fp16, var_5687_cast_fp16))[name = string("rot_101_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5691_cast_fp16 = mul(x = z_101_cast_fp16, y = c_1_cast_fp16)[name = string("op_5691_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5692_cast_fp16 = mul(x = rot_101_cast_fp16, y = s_9_cast_fp16)[name = string("op_5692_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5694_cast_fp16 = add(x = var_5691_cast_fp16, y = var_5692_cast_fp16)[name = string("op_5694_cast_fp16")];
            tensor<int32, [4]> z_103_begin_0 = const()[name = string("z_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_103_end_0 = const()[name = string("z_103_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_103_end_mask_0 = const()[name = string("z_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_103_cast_fp16 = slice_by_index(begin = z_103_begin_0, end = z_103_end_0, end_mask = z_103_end_mask_0, x = x_227_cast_fp16)[name = string("z_103_cast_fp16")];
            tensor<int32, [4]> var_5704_begin_0 = const()[name = string("op_5704_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5704_end_0 = const()[name = string("op_5704_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5704_end_mask_0 = const()[name = string("op_5704_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5704_cast_fp16 = slice_by_index(begin = var_5704_begin_0, end = var_5704_end_0, end_mask = var_5704_end_mask_0, x = z_103_cast_fp16)[name = string("op_5704_cast_fp16")];
            fp16 const_51_promoted_to_fp16 = const()[name = string("const_51_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5705_cast_fp16 = mul(x = var_5704_cast_fp16, y = const_51_promoted_to_fp16)[name = string("op_5705_cast_fp16")];
            tensor<int32, [4]> var_5710_begin_0 = const()[name = string("op_5710_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5710_end_0 = const()[name = string("op_5710_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5710_end_mask_0 = const()[name = string("op_5710_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5710_cast_fp16 = slice_by_index(begin = var_5710_begin_0, end = var_5710_end_0, end_mask = var_5710_end_mask_0, x = z_103_cast_fp16)[name = string("op_5710_cast_fp16")];
            int32 var_5712 = const()[name = string("op_5712"), val = int32(-1)];
            bool rot_103_interleave_0 = const()[name = string("rot_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_103_cast_fp16 = concat(axis = var_5712, interleave = rot_103_interleave_0, values = (var_5705_cast_fp16, var_5710_cast_fp16))[name = string("rot_103_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5714_cast_fp16 = mul(x = z_103_cast_fp16, y = c_cast_fp16)[name = string("op_5714_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5715_cast_fp16 = mul(x = rot_103_cast_fp16, y = s_11_cast_fp16)[name = string("op_5715_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5717_cast_fp16 = add(x = var_5714_cast_fp16, y = var_5715_cast_fp16)[name = string("op_5717_cast_fp16")];
            int32 var_5719 = const()[name = string("op_5719"), val = int32(-1)];
            bool var_5720_interleave_0 = const()[name = string("op_5720_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_5720_cast_fp16 = concat(axis = var_5719, interleave = var_5720_interleave_0, values = (var_5694_cast_fp16, var_5717_cast_fp16))[name = string("op_5720_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_5730_transpose_x_0 = const()[name = string("op_5730_transpose_x_0"), val = bool(false)];
            bool var_5730_transpose_y_0 = const()[name = string("op_5730_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_89 = transpose(perm = transpose_89_perm_0, x = var_5720_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 12, 630, 64]> transpose_88 = transpose(perm = transpose_88_perm_0, x = var_5668_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 12, 630, 630]> var_5730_cast_fp16 = matmul(transpose_x = var_5730_transpose_x_0, transpose_y = var_5730_transpose_y_0, x = transpose_88, y = transpose_89)[name = string("op_5730_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_5732_cast_fp16 = add(x = var_5730_cast_fp16, y = var_1524_cast_fp16)[name = string("op_5732_cast_fp16")];
            int32 var_5733 = const()[name = string("op_5733"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_175_cast_fp16 = softmax(axis = var_5733, x = var_5732_cast_fp16)[name = string("w_175_cast_fp16")];
            bool var_5736_transpose_x_0 = const()[name = string("op_5736_transpose_x_0"), val = bool(false)];
            bool var_5736_transpose_y_0 = const()[name = string("op_5736_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = v_49_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 12, 630, 64]> var_5736_cast_fp16 = matmul(transpose_x = var_5736_transpose_x_0, transpose_y = var_5736_transpose_y_0, x = w_175_cast_fp16, y = v_51_cast_fp16)[name = string("op_5736_cast_fp16")];
            tensor<int32, [4]> var_5739_perm_0 = const()[name = string("op_5739_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5743 = const()[name = string("op_5743"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_5739_cast_fp16 = transpose(perm = var_5739_perm_0, x = var_5736_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 630, 768]> input_125_cast_fp16 = reshape(shape = var_5743, x = var_5739_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115952384))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116542272))))[name = string("t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5748_cast_fp16 = abs(x = linear_88_cast_fp16)[name = string("op_5748_cast_fp16")];
            bool var_5751 = const()[name = string("op_5751"), val = bool(true)];
            tensor<int32, [1]> var_5752_axes_0 = const()[name = string("op_5752_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5752_cast_fp16 = reduce_max(axes = var_5752_axes_0, keep_dims = var_5751, x = var_5748_cast_fp16)[name = string("op_5752_cast_fp16")];
            fp16 var_5753_to_fp16 = const()[name = string("op_5753_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_181_cast_fp16 = maximum(x = var_5752_cast_fp16, y = var_5753_to_fp16)[name = string("s_181_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_177_cast_fp16 = real_div(x = linear_88_cast_fp16, y = s_181_cast_fp16)[name = string("y_177_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5756_cast_fp16 = mul(x = y_177_cast_fp16, y = y_177_cast_fp16)[name = string("op_5756_cast_fp16")];
            tensor<int32, [1]> var_5761_axes_0 = const()[name = string("op_5761_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5761_keep_dims_0 = const()[name = string("op_5761_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5761_cast_fp16 = reduce_mean(axes = var_5761_axes_0, keep_dims = var_5761_keep_dims_0, x = var_5756_cast_fp16)[name = string("op_5761_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5762_cast_fp16 = mul(x = s_181_cast_fp16, y = s_181_cast_fp16)[name = string("op_5762_cast_fp16")];
            fp32 var_5763_epsilon_0 = const()[name = string("op_5763_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5763_cast_fp16 = inverse(epsilon = var_5763_epsilon_0, x = var_5762_cast_fp16)[name = string("op_5763_cast_fp16")];
            fp16 var_5764_to_fp16 = const()[name = string("op_5764_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5765_cast_fp16 = mul(x = var_5763_cast_fp16, y = var_5764_to_fp16)[name = string("op_5765_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5767_cast_fp16 = add(x = var_5761_cast_fp16, y = var_5765_cast_fp16)[name = string("op_5767_cast_fp16")];
            fp32 var_5768_epsilon_0 = const()[name = string("op_5768_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5768_cast_fp16 = rsqrt(epsilon = var_5768_epsilon_0, x = var_5767_cast_fp16)[name = string("op_5768_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_151_cast_fp16 = mul(x = y_177_cast_fp16, y = var_5768_cast_fp16)[name = string("out_151_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116543872)))];
            tensor<fp16, [1, 630, 768]> var_5770_cast_fp16 = mul(x = out_151_cast_fp16, y = t_encoder_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("op_5770_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_231_cast_fp16 = add(x = x_217_cast_fp16, y = var_5770_cast_fp16)[name = string("x_231_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5773_cast_fp16 = abs(x = x_231_cast_fp16)[name = string("op_5773_cast_fp16")];
            bool var_5776 = const()[name = string("op_5776"), val = bool(true)];
            tensor<int32, [1]> var_5777_axes_0 = const()[name = string("op_5777_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5777_cast_fp16 = reduce_max(axes = var_5777_axes_0, keep_dims = var_5776, x = var_5773_cast_fp16)[name = string("op_5777_cast_fp16")];
            fp16 var_5778_to_fp16 = const()[name = string("op_5778_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_183_cast_fp16 = maximum(x = var_5777_cast_fp16, y = var_5778_to_fp16)[name = string("s_183_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_179_cast_fp16 = real_div(x = x_231_cast_fp16, y = s_183_cast_fp16)[name = string("y_179_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5781_cast_fp16 = mul(x = y_179_cast_fp16, y = y_179_cast_fp16)[name = string("op_5781_cast_fp16")];
            tensor<int32, [1]> var_5786_axes_0 = const()[name = string("op_5786_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5786_keep_dims_0 = const()[name = string("op_5786_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5786_cast_fp16 = reduce_mean(axes = var_5786_axes_0, keep_dims = var_5786_keep_dims_0, x = var_5781_cast_fp16)[name = string("op_5786_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5787_cast_fp16 = mul(x = s_183_cast_fp16, y = s_183_cast_fp16)[name = string("op_5787_cast_fp16")];
            fp32 var_5788_epsilon_0 = const()[name = string("op_5788_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5788_cast_fp16 = inverse(epsilon = var_5788_epsilon_0, x = var_5787_cast_fp16)[name = string("op_5788_cast_fp16")];
            fp16 var_5789_to_fp16 = const()[name = string("op_5789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5790_cast_fp16 = mul(x = var_5788_cast_fp16, y = var_5789_to_fp16)[name = string("op_5790_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5792_cast_fp16 = add(x = var_5786_cast_fp16, y = var_5790_cast_fp16)[name = string("op_5792_cast_fp16")];
            fp32 var_5793_epsilon_0 = const()[name = string("op_5793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5793_cast_fp16 = rsqrt(epsilon = var_5793_epsilon_0, x = var_5792_cast_fp16)[name = string("op_5793_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_153_cast_fp16 = mul(x = y_179_cast_fp16, y = var_5793_cast_fp16)[name = string("out_153_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116545472)))];
            tensor<fp16, [1, 630, 768]> input_127_cast_fp16 = mul(x = out_153_cast_fp16, y = t_encoder_layers_12_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116547072))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118906432))))[name = string("t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_89_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_89_cast_fp16")];
            string var_5800_mode_0 = const()[name = string("op_5800_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_5800_cast_fp16 = gelu(mode = var_5800_mode_0, x = linear_89_cast_fp16)[name = string("op_5800_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(118912640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121272000))))[name = string("t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_90_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_127_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_131_cast_fp16 = mul(x = var_5800_cast_fp16, y = linear_90_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(121278208))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123637568))))[name = string("t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_12_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5808_cast_fp16 = abs(x = linear_91_cast_fp16)[name = string("op_5808_cast_fp16")];
            bool var_5811 = const()[name = string("op_5811"), val = bool(true)];
            tensor<int32, [1]> var_5812_axes_0 = const()[name = string("op_5812_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5812_cast_fp16 = reduce_max(axes = var_5812_axes_0, keep_dims = var_5811, x = var_5808_cast_fp16)[name = string("op_5812_cast_fp16")];
            fp16 var_5813_to_fp16 = const()[name = string("op_5813_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_185_cast_fp16 = maximum(x = var_5812_cast_fp16, y = var_5813_to_fp16)[name = string("s_185_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_181_cast_fp16 = real_div(x = linear_91_cast_fp16, y = s_185_cast_fp16)[name = string("y_181_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5816_cast_fp16 = mul(x = y_181_cast_fp16, y = y_181_cast_fp16)[name = string("op_5816_cast_fp16")];
            tensor<int32, [1]> var_5821_axes_0 = const()[name = string("op_5821_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5821_keep_dims_0 = const()[name = string("op_5821_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5821_cast_fp16 = reduce_mean(axes = var_5821_axes_0, keep_dims = var_5821_keep_dims_0, x = var_5816_cast_fp16)[name = string("op_5821_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5822_cast_fp16 = mul(x = s_185_cast_fp16, y = s_185_cast_fp16)[name = string("op_5822_cast_fp16")];
            fp32 var_5823_epsilon_0 = const()[name = string("op_5823_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5823_cast_fp16 = inverse(epsilon = var_5823_epsilon_0, x = var_5822_cast_fp16)[name = string("op_5823_cast_fp16")];
            fp16 var_5824_to_fp16 = const()[name = string("op_5824_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5825_cast_fp16 = mul(x = var_5823_cast_fp16, y = var_5824_to_fp16)[name = string("op_5825_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5827_cast_fp16 = add(x = var_5821_cast_fp16, y = var_5825_cast_fp16)[name = string("op_5827_cast_fp16")];
            fp32 var_5828_epsilon_0 = const()[name = string("op_5828_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5828_cast_fp16 = rsqrt(epsilon = var_5828_epsilon_0, x = var_5827_cast_fp16)[name = string("op_5828_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_155_cast_fp16 = mul(x = y_181_cast_fp16, y = var_5828_cast_fp16)[name = string("out_155_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123639168)))];
            tensor<fp16, [1, 630, 768]> var_5830_cast_fp16 = mul(x = out_155_cast_fp16, y = t_encoder_layers_12_post_feedforward_layernorm_weight_to_fp16)[name = string("op_5830_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_235_cast_fp16 = add(x = x_231_cast_fp16, y = var_5830_cast_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5833_cast_fp16 = abs(x = x_235_cast_fp16)[name = string("op_5833_cast_fp16")];
            bool var_5836 = const()[name = string("op_5836"), val = bool(true)];
            tensor<int32, [1]> var_5837_axes_0 = const()[name = string("op_5837_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_5837_cast_fp16 = reduce_max(axes = var_5837_axes_0, keep_dims = var_5836, x = var_5833_cast_fp16)[name = string("op_5837_cast_fp16")];
            fp16 var_5838_to_fp16 = const()[name = string("op_5838_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_187_cast_fp16 = maximum(x = var_5837_cast_fp16, y = var_5838_to_fp16)[name = string("s_187_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_183_cast_fp16 = real_div(x = x_235_cast_fp16, y = s_187_cast_fp16)[name = string("y_183_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_5841_cast_fp16 = mul(x = y_183_cast_fp16, y = y_183_cast_fp16)[name = string("op_5841_cast_fp16")];
            tensor<int32, [1]> var_5846_axes_0 = const()[name = string("op_5846_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5846_keep_dims_0 = const()[name = string("op_5846_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_5846_cast_fp16 = reduce_mean(axes = var_5846_axes_0, keep_dims = var_5846_keep_dims_0, x = var_5841_cast_fp16)[name = string("op_5846_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5847_cast_fp16 = mul(x = s_187_cast_fp16, y = s_187_cast_fp16)[name = string("op_5847_cast_fp16")];
            fp32 var_5848_epsilon_0 = const()[name = string("op_5848_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_5848_cast_fp16 = inverse(epsilon = var_5848_epsilon_0, x = var_5847_cast_fp16)[name = string("op_5848_cast_fp16")];
            fp16 var_5849_to_fp16 = const()[name = string("op_5849_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_5850_cast_fp16 = mul(x = var_5848_cast_fp16, y = var_5849_to_fp16)[name = string("op_5850_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_5852_cast_fp16 = add(x = var_5846_cast_fp16, y = var_5850_cast_fp16)[name = string("op_5852_cast_fp16")];
            fp32 var_5853_epsilon_0 = const()[name = string("op_5853_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_5853_cast_fp16 = rsqrt(epsilon = var_5853_epsilon_0, x = var_5852_cast_fp16)[name = string("op_5853_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_157_cast_fp16 = mul(x = y_183_cast_fp16, y = var_5853_cast_fp16)[name = string("out_157_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123640768)))];
            tensor<fp16, [1, 630, 768]> input_133_cast_fp16 = mul(x = out_157_cast_fp16, y = t_encoder_layers_13_input_layernorm_weight_to_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(123642368))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124232256))))[name = string("t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5863 = const()[name = string("op_5863"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_237_cast_fp16 = reshape(shape = var_5863, x = linear_92_cast_fp16)[name = string("x_237_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5865_cast_fp16 = abs(x = x_237_cast_fp16)[name = string("op_5865_cast_fp16")];
            bool var_5868 = const()[name = string("op_5868"), val = bool(true)];
            tensor<int32, [1]> var_5869_axes_0 = const()[name = string("op_5869_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5869_cast_fp16 = reduce_max(axes = var_5869_axes_0, keep_dims = var_5868, x = var_5865_cast_fp16)[name = string("op_5869_cast_fp16")];
            fp16 var_5870_to_fp16 = const()[name = string("op_5870_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_189_cast_fp16 = maximum(x = var_5869_cast_fp16, y = var_5870_to_fp16)[name = string("s_189_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_185_cast_fp16 = real_div(x = x_237_cast_fp16, y = s_189_cast_fp16)[name = string("y_185_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5873_cast_fp16 = mul(x = y_185_cast_fp16, y = y_185_cast_fp16)[name = string("op_5873_cast_fp16")];
            tensor<int32, [1]> var_5878_axes_0 = const()[name = string("op_5878_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5878_keep_dims_0 = const()[name = string("op_5878_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5878_cast_fp16 = reduce_mean(axes = var_5878_axes_0, keep_dims = var_5878_keep_dims_0, x = var_5873_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5879_cast_fp16 = mul(x = s_189_cast_fp16, y = s_189_cast_fp16)[name = string("op_5879_cast_fp16")];
            fp32 var_5880_epsilon_0 = const()[name = string("op_5880_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5880_cast_fp16 = inverse(epsilon = var_5880_epsilon_0, x = var_5879_cast_fp16)[name = string("op_5880_cast_fp16")];
            fp16 var_5881_to_fp16 = const()[name = string("op_5881_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5882_cast_fp16 = mul(x = var_5880_cast_fp16, y = var_5881_to_fp16)[name = string("op_5882_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5884_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5882_cast_fp16)[name = string("op_5884_cast_fp16")];
            fp32 var_5885_epsilon_0 = const()[name = string("op_5885_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5885_cast_fp16 = rsqrt(epsilon = var_5885_epsilon_0, x = var_5884_cast_fp16)[name = string("op_5885_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_159_cast_fp16 = mul(x = y_185_cast_fp16, y = var_5885_cast_fp16)[name = string("out_159_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(76910080)))];
            tensor<fp16, [1, 630, 12, 64]> x_243_cast_fp16 = mul(x = out_159_cast_fp16, y = t_encoder_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_243_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124233856))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124823744))))[name = string("t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_93_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5895 = const()[name = string("op_5895"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_239_cast_fp16 = reshape(shape = var_5895, x = linear_93_cast_fp16)[name = string("x_239_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5897_cast_fp16 = abs(x = x_239_cast_fp16)[name = string("op_5897_cast_fp16")];
            bool var_5900 = const()[name = string("op_5900"), val = bool(true)];
            tensor<int32, [1]> var_5901_axes_0 = const()[name = string("op_5901_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5901_cast_fp16 = reduce_max(axes = var_5901_axes_0, keep_dims = var_5900, x = var_5897_cast_fp16)[name = string("op_5901_cast_fp16")];
            fp16 var_5902_to_fp16 = const()[name = string("op_5902_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_191_cast_fp16 = maximum(x = var_5901_cast_fp16, y = var_5902_to_fp16)[name = string("s_191_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_187_cast_fp16 = real_div(x = x_239_cast_fp16, y = s_191_cast_fp16)[name = string("y_187_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5905_cast_fp16 = mul(x = y_187_cast_fp16, y = y_187_cast_fp16)[name = string("op_5905_cast_fp16")];
            tensor<int32, [1]> var_5910_axes_0 = const()[name = string("op_5910_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5910_keep_dims_0 = const()[name = string("op_5910_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5910_cast_fp16 = reduce_mean(axes = var_5910_axes_0, keep_dims = var_5910_keep_dims_0, x = var_5905_cast_fp16)[name = string("op_5910_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5911_cast_fp16 = mul(x = s_191_cast_fp16, y = s_191_cast_fp16)[name = string("op_5911_cast_fp16")];
            fp32 var_5912_epsilon_0 = const()[name = string("op_5912_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5912_cast_fp16 = inverse(epsilon = var_5912_epsilon_0, x = var_5911_cast_fp16)[name = string("op_5912_cast_fp16")];
            fp16 var_5913_to_fp16 = const()[name = string("op_5913_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5914_cast_fp16 = mul(x = var_5912_cast_fp16, y = var_5913_to_fp16)[name = string("op_5914_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5916_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5914_cast_fp16)[name = string("op_5916_cast_fp16")];
            fp32 var_5917_epsilon_0 = const()[name = string("op_5917_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5917_cast_fp16 = rsqrt(epsilon = var_5917_epsilon_0, x = var_5916_cast_fp16)[name = string("op_5917_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_161_cast_fp16 = mul(x = y_187_cast_fp16, y = var_5917_cast_fp16)[name = string("out_161_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77501760)))];
            tensor<fp16, [1, 630, 12, 64]> x_245_cast_fp16 = mul(x = out_161_cast_fp16, y = t_encoder_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_245_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(124825344))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125415232))))[name = string("t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<int32, [4]> var_5927 = const()[name = string("op_5927"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_241_cast_fp16 = reshape(shape = var_5927, x = linear_94_cast_fp16)[name = string("x_241_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5929_cast_fp16 = abs(x = x_241_cast_fp16)[name = string("op_5929_cast_fp16")];
            bool var_5932 = const()[name = string("op_5932"), val = bool(true)];
            tensor<int32, [1]> var_5933_axes_0 = const()[name = string("op_5933_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_5933_cast_fp16 = reduce_max(axes = var_5933_axes_0, keep_dims = var_5932, x = var_5929_cast_fp16)[name = string("op_5933_cast_fp16")];
            fp16 var_5934_to_fp16 = const()[name = string("op_5934_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_193_cast_fp16 = maximum(x = var_5933_cast_fp16, y = var_5934_to_fp16)[name = string("s_193_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_189_cast_fp16 = real_div(x = x_241_cast_fp16, y = s_193_cast_fp16)[name = string("y_189_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_5937_cast_fp16 = mul(x = y_189_cast_fp16, y = y_189_cast_fp16)[name = string("op_5937_cast_fp16")];
            tensor<int32, [1]> var_5942_axes_0 = const()[name = string("op_5942_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5942_keep_dims_0 = const()[name = string("op_5942_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_5942_cast_fp16 = reduce_mean(axes = var_5942_axes_0, keep_dims = var_5942_keep_dims_0, x = var_5937_cast_fp16)[name = string("op_5942_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5943_cast_fp16 = mul(x = s_193_cast_fp16, y = s_193_cast_fp16)[name = string("op_5943_cast_fp16")];
            fp32 var_5944_epsilon_0 = const()[name = string("op_5944_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_5944_cast_fp16 = inverse(epsilon = var_5944_epsilon_0, x = var_5943_cast_fp16)[name = string("op_5944_cast_fp16")];
            fp16 var_5945_to_fp16 = const()[name = string("op_5945_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_5946_cast_fp16 = mul(x = var_5944_cast_fp16, y = var_5945_to_fp16)[name = string("op_5946_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_5948_cast_fp16 = add(x = var_5942_cast_fp16, y = var_5946_cast_fp16)[name = string("op_5948_cast_fp16")];
            fp32 var_5949_epsilon_0 = const()[name = string("op_5949_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_5949_cast_fp16 = rsqrt(epsilon = var_5949_epsilon_0, x = var_5948_cast_fp16)[name = string("op_5949_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_53_cast_fp16 = mul(x = y_189_cast_fp16, y = var_5949_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<int32, [4]> z_105_begin_0 = const()[name = string("z_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_105_end_0 = const()[name = string("z_105_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_105_end_mask_0 = const()[name = string("z_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_105_cast_fp16 = slice_by_index(begin = z_105_begin_0, end = z_105_end_0, end_mask = z_105_end_mask_0, x = x_243_cast_fp16)[name = string("z_105_cast_fp16")];
            tensor<int32, [4]> var_5960_begin_0 = const()[name = string("op_5960_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5960_end_0 = const()[name = string("op_5960_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5960_end_mask_0 = const()[name = string("op_5960_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5960_cast_fp16 = slice_by_index(begin = var_5960_begin_0, end = var_5960_end_0, end_mask = var_5960_end_mask_0, x = z_105_cast_fp16)[name = string("op_5960_cast_fp16")];
            fp16 const_52_promoted_to_fp16 = const()[name = string("const_52_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5961_cast_fp16 = mul(x = var_5960_cast_fp16, y = const_52_promoted_to_fp16)[name = string("op_5961_cast_fp16")];
            tensor<int32, [4]> var_5966_begin_0 = const()[name = string("op_5966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5966_end_0 = const()[name = string("op_5966_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5966_end_mask_0 = const()[name = string("op_5966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5966_cast_fp16 = slice_by_index(begin = var_5966_begin_0, end = var_5966_end_0, end_mask = var_5966_end_mask_0, x = z_105_cast_fp16)[name = string("op_5966_cast_fp16")];
            int32 var_5968 = const()[name = string("op_5968"), val = int32(-1)];
            bool rot_105_interleave_0 = const()[name = string("rot_105_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_105_cast_fp16 = concat(axis = var_5968, interleave = rot_105_interleave_0, values = (var_5961_cast_fp16, var_5966_cast_fp16))[name = string("rot_105_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5970_cast_fp16 = mul(x = z_105_cast_fp16, y = c_1_cast_fp16)[name = string("op_5970_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5971_cast_fp16 = mul(x = rot_105_cast_fp16, y = s_9_cast_fp16)[name = string("op_5971_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5973_cast_fp16 = add(x = var_5970_cast_fp16, y = var_5971_cast_fp16)[name = string("op_5973_cast_fp16")];
            tensor<int32, [4]> z_107_begin_0 = const()[name = string("z_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_107_end_0 = const()[name = string("z_107_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_107_end_mask_0 = const()[name = string("z_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_107_cast_fp16 = slice_by_index(begin = z_107_begin_0, end = z_107_end_0, end_mask = z_107_end_mask_0, x = x_243_cast_fp16)[name = string("z_107_cast_fp16")];
            tensor<int32, [4]> var_5983_begin_0 = const()[name = string("op_5983_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_5983_end_0 = const()[name = string("op_5983_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_5983_end_mask_0 = const()[name = string("op_5983_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_5983_cast_fp16 = slice_by_index(begin = var_5983_begin_0, end = var_5983_end_0, end_mask = var_5983_end_mask_0, x = z_107_cast_fp16)[name = string("op_5983_cast_fp16")];
            fp16 const_53_promoted_to_fp16 = const()[name = string("const_53_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_5984_cast_fp16 = mul(x = var_5983_cast_fp16, y = const_53_promoted_to_fp16)[name = string("op_5984_cast_fp16")];
            tensor<int32, [4]> var_5989_begin_0 = const()[name = string("op_5989_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5989_end_0 = const()[name = string("op_5989_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_5989_end_mask_0 = const()[name = string("op_5989_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_5989_cast_fp16 = slice_by_index(begin = var_5989_begin_0, end = var_5989_end_0, end_mask = var_5989_end_mask_0, x = z_107_cast_fp16)[name = string("op_5989_cast_fp16")];
            int32 var_5991 = const()[name = string("op_5991"), val = int32(-1)];
            bool rot_107_interleave_0 = const()[name = string("rot_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_107_cast_fp16 = concat(axis = var_5991, interleave = rot_107_interleave_0, values = (var_5984_cast_fp16, var_5989_cast_fp16))[name = string("rot_107_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5993_cast_fp16 = mul(x = z_107_cast_fp16, y = c_cast_fp16)[name = string("op_5993_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5994_cast_fp16 = mul(x = rot_107_cast_fp16, y = s_11_cast_fp16)[name = string("op_5994_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_5996_cast_fp16 = add(x = var_5993_cast_fp16, y = var_5994_cast_fp16)[name = string("op_5996_cast_fp16")];
            int32 var_5998 = const()[name = string("op_5998"), val = int32(-1)];
            bool var_5999_interleave_0 = const()[name = string("op_5999_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_5999_cast_fp16 = concat(axis = var_5998, interleave = var_5999_interleave_0, values = (var_5973_cast_fp16, var_5996_cast_fp16))[name = string("op_5999_cast_fp16")];
            tensor<int32, [4]> z_109_begin_0 = const()[name = string("z_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_109_end_0 = const()[name = string("z_109_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_109_end_mask_0 = const()[name = string("z_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_109_cast_fp16 = slice_by_index(begin = z_109_begin_0, end = z_109_end_0, end_mask = z_109_end_mask_0, x = x_245_cast_fp16)[name = string("z_109_cast_fp16")];
            tensor<int32, [4]> var_6012_begin_0 = const()[name = string("op_6012_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6012_end_0 = const()[name = string("op_6012_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6012_end_mask_0 = const()[name = string("op_6012_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6012_cast_fp16 = slice_by_index(begin = var_6012_begin_0, end = var_6012_end_0, end_mask = var_6012_end_mask_0, x = z_109_cast_fp16)[name = string("op_6012_cast_fp16")];
            fp16 const_54_promoted_to_fp16 = const()[name = string("const_54_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6013_cast_fp16 = mul(x = var_6012_cast_fp16, y = const_54_promoted_to_fp16)[name = string("op_6013_cast_fp16")];
            tensor<int32, [4]> var_6018_begin_0 = const()[name = string("op_6018_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6018_end_0 = const()[name = string("op_6018_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6018_end_mask_0 = const()[name = string("op_6018_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6018_cast_fp16 = slice_by_index(begin = var_6018_begin_0, end = var_6018_end_0, end_mask = var_6018_end_mask_0, x = z_109_cast_fp16)[name = string("op_6018_cast_fp16")];
            int32 var_6020 = const()[name = string("op_6020"), val = int32(-1)];
            bool rot_109_interleave_0 = const()[name = string("rot_109_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_109_cast_fp16 = concat(axis = var_6020, interleave = rot_109_interleave_0, values = (var_6013_cast_fp16, var_6018_cast_fp16))[name = string("rot_109_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6022_cast_fp16 = mul(x = z_109_cast_fp16, y = c_1_cast_fp16)[name = string("op_6022_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6023_cast_fp16 = mul(x = rot_109_cast_fp16, y = s_9_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6025_cast_fp16 = add(x = var_6022_cast_fp16, y = var_6023_cast_fp16)[name = string("op_6025_cast_fp16")];
            tensor<int32, [4]> z_111_begin_0 = const()[name = string("z_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_111_end_0 = const()[name = string("z_111_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_111_end_mask_0 = const()[name = string("z_111_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_111_cast_fp16 = slice_by_index(begin = z_111_begin_0, end = z_111_end_0, end_mask = z_111_end_mask_0, x = x_245_cast_fp16)[name = string("z_111_cast_fp16")];
            tensor<int32, [4]> var_6035_begin_0 = const()[name = string("op_6035_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6035_end_0 = const()[name = string("op_6035_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6035_end_mask_0 = const()[name = string("op_6035_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6035_cast_fp16 = slice_by_index(begin = var_6035_begin_0, end = var_6035_end_0, end_mask = var_6035_end_mask_0, x = z_111_cast_fp16)[name = string("op_6035_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6036_cast_fp16 = mul(x = var_6035_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_6036_cast_fp16")];
            tensor<int32, [4]> var_6041_begin_0 = const()[name = string("op_6041_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6041_end_0 = const()[name = string("op_6041_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6041_end_mask_0 = const()[name = string("op_6041_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6041_cast_fp16 = slice_by_index(begin = var_6041_begin_0, end = var_6041_end_0, end_mask = var_6041_end_mask_0, x = z_111_cast_fp16)[name = string("op_6041_cast_fp16")];
            int32 var_6043 = const()[name = string("op_6043"), val = int32(-1)];
            bool rot_111_interleave_0 = const()[name = string("rot_111_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_111_cast_fp16 = concat(axis = var_6043, interleave = rot_111_interleave_0, values = (var_6036_cast_fp16, var_6041_cast_fp16))[name = string("rot_111_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6045_cast_fp16 = mul(x = z_111_cast_fp16, y = c_cast_fp16)[name = string("op_6045_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6046_cast_fp16 = mul(x = rot_111_cast_fp16, y = s_11_cast_fp16)[name = string("op_6046_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6048_cast_fp16 = add(x = var_6045_cast_fp16, y = var_6046_cast_fp16)[name = string("op_6048_cast_fp16")];
            int32 var_6050 = const()[name = string("op_6050"), val = int32(-1)];
            bool var_6051_interleave_0 = const()[name = string("op_6051_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_6051_cast_fp16 = concat(axis = var_6050, interleave = var_6051_interleave_0, values = (var_6025_cast_fp16, var_6048_cast_fp16))[name = string("op_6051_cast_fp16")];
            tensor<int32, [4]> v_55_perm_0 = const()[name = string("v_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6061_transpose_x_0 = const()[name = string("op_6061_transpose_x_0"), val = bool(false)];
            bool var_6061_transpose_y_0 = const()[name = string("op_6061_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_91 = transpose(perm = transpose_91_perm_0, x = var_6051_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 12, 630, 64]> transpose_90 = transpose(perm = transpose_90_perm_0, x = var_5999_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 12, 630, 630]> var_6061_cast_fp16 = matmul(transpose_x = var_6061_transpose_x_0, transpose_y = var_6061_transpose_y_0, x = transpose_90, y = transpose_91)[name = string("op_6061_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_6063_cast_fp16 = add(x = var_6061_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6063_cast_fp16")];
            int32 var_6064 = const()[name = string("op_6064"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_189_cast_fp16 = softmax(axis = var_6064, x = var_6063_cast_fp16)[name = string("w_189_cast_fp16")];
            bool var_6067_transpose_x_0 = const()[name = string("op_6067_transpose_x_0"), val = bool(false)];
            bool var_6067_transpose_y_0 = const()[name = string("op_6067_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_55_cast_fp16 = transpose(perm = v_55_perm_0, x = v_53_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 12, 630, 64]> var_6067_cast_fp16 = matmul(transpose_x = var_6067_transpose_x_0, transpose_y = var_6067_transpose_y_0, x = w_189_cast_fp16, y = v_55_cast_fp16)[name = string("op_6067_cast_fp16")];
            tensor<int32, [4]> var_6070_perm_0 = const()[name = string("op_6070_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6074 = const()[name = string("op_6074"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_6070_cast_fp16 = transpose(perm = var_6070_perm_0, x = var_6067_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 630, 768]> input_135_cast_fp16 = reshape(shape = var_6074, x = var_6070_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125416832))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126006720))))[name = string("t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_95_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6079_cast_fp16 = abs(x = linear_95_cast_fp16)[name = string("op_6079_cast_fp16")];
            bool var_6082 = const()[name = string("op_6082"), val = bool(true)];
            tensor<int32, [1]> var_6083_axes_0 = const()[name = string("op_6083_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6083_cast_fp16 = reduce_max(axes = var_6083_axes_0, keep_dims = var_6082, x = var_6079_cast_fp16)[name = string("op_6083_cast_fp16")];
            fp16 var_6084_to_fp16 = const()[name = string("op_6084_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_195_cast_fp16 = maximum(x = var_6083_cast_fp16, y = var_6084_to_fp16)[name = string("s_195_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_191_cast_fp16 = real_div(x = linear_95_cast_fp16, y = s_195_cast_fp16)[name = string("y_191_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6087_cast_fp16 = mul(x = y_191_cast_fp16, y = y_191_cast_fp16)[name = string("op_6087_cast_fp16")];
            tensor<int32, [1]> var_6092_axes_0 = const()[name = string("op_6092_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6092_keep_dims_0 = const()[name = string("op_6092_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6092_cast_fp16 = reduce_mean(axes = var_6092_axes_0, keep_dims = var_6092_keep_dims_0, x = var_6087_cast_fp16)[name = string("op_6092_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6093_cast_fp16 = mul(x = s_195_cast_fp16, y = s_195_cast_fp16)[name = string("op_6093_cast_fp16")];
            fp32 var_6094_epsilon_0 = const()[name = string("op_6094_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6094_cast_fp16 = inverse(epsilon = var_6094_epsilon_0, x = var_6093_cast_fp16)[name = string("op_6094_cast_fp16")];
            fp16 var_6095_to_fp16 = const()[name = string("op_6095_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6096_cast_fp16 = mul(x = var_6094_cast_fp16, y = var_6095_to_fp16)[name = string("op_6096_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6098_cast_fp16 = add(x = var_6092_cast_fp16, y = var_6096_cast_fp16)[name = string("op_6098_cast_fp16")];
            fp32 var_6099_epsilon_0 = const()[name = string("op_6099_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6099_cast_fp16 = rsqrt(epsilon = var_6099_epsilon_0, x = var_6098_cast_fp16)[name = string("op_6099_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_163_cast_fp16 = mul(x = y_191_cast_fp16, y = var_6099_cast_fp16)[name = string("out_163_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126008320)))];
            tensor<fp16, [1, 630, 768]> var_6101_cast_fp16 = mul(x = out_163_cast_fp16, y = t_encoder_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("op_6101_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_249_cast_fp16 = add(x = x_235_cast_fp16, y = var_6101_cast_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6104_cast_fp16 = abs(x = x_249_cast_fp16)[name = string("op_6104_cast_fp16")];
            bool var_6107 = const()[name = string("op_6107"), val = bool(true)];
            tensor<int32, [1]> var_6108_axes_0 = const()[name = string("op_6108_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6108_cast_fp16 = reduce_max(axes = var_6108_axes_0, keep_dims = var_6107, x = var_6104_cast_fp16)[name = string("op_6108_cast_fp16")];
            fp16 var_6109_to_fp16 = const()[name = string("op_6109_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_197_cast_fp16 = maximum(x = var_6108_cast_fp16, y = var_6109_to_fp16)[name = string("s_197_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_193_cast_fp16 = real_div(x = x_249_cast_fp16, y = s_197_cast_fp16)[name = string("y_193_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6112_cast_fp16 = mul(x = y_193_cast_fp16, y = y_193_cast_fp16)[name = string("op_6112_cast_fp16")];
            tensor<int32, [1]> var_6117_axes_0 = const()[name = string("op_6117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6117_keep_dims_0 = const()[name = string("op_6117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6117_cast_fp16 = reduce_mean(axes = var_6117_axes_0, keep_dims = var_6117_keep_dims_0, x = var_6112_cast_fp16)[name = string("op_6117_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6118_cast_fp16 = mul(x = s_197_cast_fp16, y = s_197_cast_fp16)[name = string("op_6118_cast_fp16")];
            fp32 var_6119_epsilon_0 = const()[name = string("op_6119_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6119_cast_fp16 = inverse(epsilon = var_6119_epsilon_0, x = var_6118_cast_fp16)[name = string("op_6119_cast_fp16")];
            fp16 var_6120_to_fp16 = const()[name = string("op_6120_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6121_cast_fp16 = mul(x = var_6119_cast_fp16, y = var_6120_to_fp16)[name = string("op_6121_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6123_cast_fp16 = add(x = var_6117_cast_fp16, y = var_6121_cast_fp16)[name = string("op_6123_cast_fp16")];
            fp32 var_6124_epsilon_0 = const()[name = string("op_6124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6124_cast_fp16 = rsqrt(epsilon = var_6124_epsilon_0, x = var_6123_cast_fp16)[name = string("op_6124_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_165_cast_fp16 = mul(x = y_193_cast_fp16, y = var_6124_cast_fp16)[name = string("out_165_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126009920)))];
            tensor<fp16, [1, 630, 768]> input_137_cast_fp16 = mul(x = out_165_cast_fp16, y = t_encoder_layers_13_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_137_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(126011520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128370880))))[name = string("t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_96_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_96_cast_fp16")];
            string var_6131_mode_0 = const()[name = string("op_6131_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_6131_cast_fp16 = gelu(mode = var_6131_mode_0, x = linear_96_cast_fp16)[name = string("op_6131_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128377088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130736448))))[name = string("t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_97_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_137_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_141_cast_fp16 = mul(x = var_6131_cast_fp16, y = linear_97_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130742656))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133102016))))[name = string("t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_13_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6139_cast_fp16 = abs(x = linear_98_cast_fp16)[name = string("op_6139_cast_fp16")];
            bool var_6142 = const()[name = string("op_6142"), val = bool(true)];
            tensor<int32, [1]> var_6143_axes_0 = const()[name = string("op_6143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6143_cast_fp16 = reduce_max(axes = var_6143_axes_0, keep_dims = var_6142, x = var_6139_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 var_6144_to_fp16 = const()[name = string("op_6144_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_199_cast_fp16 = maximum(x = var_6143_cast_fp16, y = var_6144_to_fp16)[name = string("s_199_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_195_cast_fp16 = real_div(x = linear_98_cast_fp16, y = s_199_cast_fp16)[name = string("y_195_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6147_cast_fp16 = mul(x = y_195_cast_fp16, y = y_195_cast_fp16)[name = string("op_6147_cast_fp16")];
            tensor<int32, [1]> var_6152_axes_0 = const()[name = string("op_6152_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6152_keep_dims_0 = const()[name = string("op_6152_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6152_cast_fp16 = reduce_mean(axes = var_6152_axes_0, keep_dims = var_6152_keep_dims_0, x = var_6147_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6153_cast_fp16 = mul(x = s_199_cast_fp16, y = s_199_cast_fp16)[name = string("op_6153_cast_fp16")];
            fp32 var_6154_epsilon_0 = const()[name = string("op_6154_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6154_cast_fp16 = inverse(epsilon = var_6154_epsilon_0, x = var_6153_cast_fp16)[name = string("op_6154_cast_fp16")];
            fp16 var_6155_to_fp16 = const()[name = string("op_6155_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6156_cast_fp16 = mul(x = var_6154_cast_fp16, y = var_6155_to_fp16)[name = string("op_6156_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6158_cast_fp16 = add(x = var_6152_cast_fp16, y = var_6156_cast_fp16)[name = string("op_6158_cast_fp16")];
            fp32 var_6159_epsilon_0 = const()[name = string("op_6159_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6159_cast_fp16 = rsqrt(epsilon = var_6159_epsilon_0, x = var_6158_cast_fp16)[name = string("op_6159_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_167_cast_fp16 = mul(x = y_195_cast_fp16, y = var_6159_cast_fp16)[name = string("out_167_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133103616)))];
            tensor<fp16, [1, 630, 768]> var_6161_cast_fp16 = mul(x = out_167_cast_fp16, y = t_encoder_layers_13_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6161_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_253_cast_fp16 = add(x = x_249_cast_fp16, y = var_6161_cast_fp16)[name = string("x_253_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6164_cast_fp16 = abs(x = x_253_cast_fp16)[name = string("op_6164_cast_fp16")];
            bool var_6167 = const()[name = string("op_6167"), val = bool(true)];
            tensor<int32, [1]> var_6168_axes_0 = const()[name = string("op_6168_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6168_cast_fp16 = reduce_max(axes = var_6168_axes_0, keep_dims = var_6167, x = var_6164_cast_fp16)[name = string("op_6168_cast_fp16")];
            fp16 var_6169_to_fp16 = const()[name = string("op_6169_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_201_cast_fp16 = maximum(x = var_6168_cast_fp16, y = var_6169_to_fp16)[name = string("s_201_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_197_cast_fp16 = real_div(x = x_253_cast_fp16, y = s_201_cast_fp16)[name = string("y_197_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6172_cast_fp16 = mul(x = y_197_cast_fp16, y = y_197_cast_fp16)[name = string("op_6172_cast_fp16")];
            tensor<int32, [1]> var_6177_axes_0 = const()[name = string("op_6177_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6177_keep_dims_0 = const()[name = string("op_6177_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6177_cast_fp16 = reduce_mean(axes = var_6177_axes_0, keep_dims = var_6177_keep_dims_0, x = var_6172_cast_fp16)[name = string("op_6177_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6178_cast_fp16 = mul(x = s_201_cast_fp16, y = s_201_cast_fp16)[name = string("op_6178_cast_fp16")];
            fp32 var_6179_epsilon_0 = const()[name = string("op_6179_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6179_cast_fp16 = inverse(epsilon = var_6179_epsilon_0, x = var_6178_cast_fp16)[name = string("op_6179_cast_fp16")];
            fp16 var_6180_to_fp16 = const()[name = string("op_6180_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6181_cast_fp16 = mul(x = var_6179_cast_fp16, y = var_6180_to_fp16)[name = string("op_6181_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6183_cast_fp16 = add(x = var_6177_cast_fp16, y = var_6181_cast_fp16)[name = string("op_6183_cast_fp16")];
            fp32 var_6184_epsilon_0 = const()[name = string("op_6184_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6184_cast_fp16 = rsqrt(epsilon = var_6184_epsilon_0, x = var_6183_cast_fp16)[name = string("op_6184_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_169_cast_fp16 = mul(x = y_197_cast_fp16, y = var_6184_cast_fp16)[name = string("out_169_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133105216)))];
            tensor<fp16, [1, 630, 768]> input_143_cast_fp16 = mul(x = out_169_cast_fp16, y = t_encoder_layers_14_input_layernorm_weight_to_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133106816))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133696704))))[name = string("t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_99_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_6194 = const()[name = string("op_6194"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_255_cast_fp16 = reshape(shape = var_6194, x = linear_99_cast_fp16)[name = string("x_255_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6196_cast_fp16 = abs(x = x_255_cast_fp16)[name = string("op_6196_cast_fp16")];
            bool var_6199 = const()[name = string("op_6199"), val = bool(true)];
            tensor<int32, [1]> var_6200_axes_0 = const()[name = string("op_6200_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_6200_cast_fp16 = reduce_max(axes = var_6200_axes_0, keep_dims = var_6199, x = var_6196_cast_fp16)[name = string("op_6200_cast_fp16")];
            fp16 var_6201_to_fp16 = const()[name = string("op_6201_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_203_cast_fp16 = maximum(x = var_6200_cast_fp16, y = var_6201_to_fp16)[name = string("s_203_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_199_cast_fp16 = real_div(x = x_255_cast_fp16, y = s_203_cast_fp16)[name = string("y_199_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6204_cast_fp16 = mul(x = y_199_cast_fp16, y = y_199_cast_fp16)[name = string("op_6204_cast_fp16")];
            tensor<int32, [1]> var_6209_axes_0 = const()[name = string("op_6209_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6209_keep_dims_0 = const()[name = string("op_6209_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_6209_cast_fp16 = reduce_mean(axes = var_6209_axes_0, keep_dims = var_6209_keep_dims_0, x = var_6204_cast_fp16)[name = string("op_6209_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6210_cast_fp16 = mul(x = s_203_cast_fp16, y = s_203_cast_fp16)[name = string("op_6210_cast_fp16")];
            fp32 var_6211_epsilon_0 = const()[name = string("op_6211_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_6211_cast_fp16 = inverse(epsilon = var_6211_epsilon_0, x = var_6210_cast_fp16)[name = string("op_6211_cast_fp16")];
            fp16 var_6212_to_fp16 = const()[name = string("op_6212_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_6213_cast_fp16 = mul(x = var_6211_cast_fp16, y = var_6212_to_fp16)[name = string("op_6213_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6215_cast_fp16 = add(x = var_6209_cast_fp16, y = var_6213_cast_fp16)[name = string("op_6215_cast_fp16")];
            fp32 var_6216_epsilon_0 = const()[name = string("op_6216_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_6216_cast_fp16 = rsqrt(epsilon = var_6216_epsilon_0, x = var_6215_cast_fp16)[name = string("op_6216_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_171_cast_fp16 = mul(x = y_199_cast_fp16, y = var_6216_cast_fp16)[name = string("out_171_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133698304)))];
            tensor<fp16, [1, 630, 12, 64]> x_261_cast_fp16 = mul(x = out_171_cast_fp16, y = t_encoder_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_261_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(133698496))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134288384))))[name = string("t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_100_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_6226 = const()[name = string("op_6226"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_257_cast_fp16 = reshape(shape = var_6226, x = linear_100_cast_fp16)[name = string("x_257_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6228_cast_fp16 = abs(x = x_257_cast_fp16)[name = string("op_6228_cast_fp16")];
            bool var_6231 = const()[name = string("op_6231"), val = bool(true)];
            tensor<int32, [1]> var_6232_axes_0 = const()[name = string("op_6232_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_6232_cast_fp16 = reduce_max(axes = var_6232_axes_0, keep_dims = var_6231, x = var_6228_cast_fp16)[name = string("op_6232_cast_fp16")];
            fp16 var_6233_to_fp16 = const()[name = string("op_6233_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_205_cast_fp16 = maximum(x = var_6232_cast_fp16, y = var_6233_to_fp16)[name = string("s_205_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_201_cast_fp16 = real_div(x = x_257_cast_fp16, y = s_205_cast_fp16)[name = string("y_201_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6236_cast_fp16 = mul(x = y_201_cast_fp16, y = y_201_cast_fp16)[name = string("op_6236_cast_fp16")];
            tensor<int32, [1]> var_6241_axes_0 = const()[name = string("op_6241_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6241_keep_dims_0 = const()[name = string("op_6241_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_6241_cast_fp16 = reduce_mean(axes = var_6241_axes_0, keep_dims = var_6241_keep_dims_0, x = var_6236_cast_fp16)[name = string("op_6241_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6242_cast_fp16 = mul(x = s_205_cast_fp16, y = s_205_cast_fp16)[name = string("op_6242_cast_fp16")];
            fp32 var_6243_epsilon_0 = const()[name = string("op_6243_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_6243_cast_fp16 = inverse(epsilon = var_6243_epsilon_0, x = var_6242_cast_fp16)[name = string("op_6243_cast_fp16")];
            fp16 var_6244_to_fp16 = const()[name = string("op_6244_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_6245_cast_fp16 = mul(x = var_6243_cast_fp16, y = var_6244_to_fp16)[name = string("op_6245_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6247_cast_fp16 = add(x = var_6241_cast_fp16, y = var_6245_cast_fp16)[name = string("op_6247_cast_fp16")];
            fp32 var_6248_epsilon_0 = const()[name = string("op_6248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_6248_cast_fp16 = rsqrt(epsilon = var_6248_epsilon_0, x = var_6247_cast_fp16)[name = string("op_6248_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_173_cast_fp16 = mul(x = y_201_cast_fp16, y = var_6248_cast_fp16)[name = string("out_173_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(86966592)))];
            tensor<fp16, [1, 630, 12, 64]> x_263_cast_fp16 = mul(x = out_173_cast_fp16, y = t_encoder_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_263_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134289984))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134879872))))[name = string("t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_101_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<int32, [4]> var_6258 = const()[name = string("op_6258"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_259_cast_fp16 = reshape(shape = var_6258, x = linear_101_cast_fp16)[name = string("x_259_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6260_cast_fp16 = abs(x = x_259_cast_fp16)[name = string("op_6260_cast_fp16")];
            bool var_6263 = const()[name = string("op_6263"), val = bool(true)];
            tensor<int32, [1]> var_6264_axes_0 = const()[name = string("op_6264_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_6264_cast_fp16 = reduce_max(axes = var_6264_axes_0, keep_dims = var_6263, x = var_6260_cast_fp16)[name = string("op_6264_cast_fp16")];
            fp16 var_6265_to_fp16 = const()[name = string("op_6265_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_207_cast_fp16 = maximum(x = var_6264_cast_fp16, y = var_6265_to_fp16)[name = string("s_207_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_203_cast_fp16 = real_div(x = x_259_cast_fp16, y = s_207_cast_fp16)[name = string("y_203_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6268_cast_fp16 = mul(x = y_203_cast_fp16, y = y_203_cast_fp16)[name = string("op_6268_cast_fp16")];
            tensor<int32, [1]> var_6273_axes_0 = const()[name = string("op_6273_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6273_keep_dims_0 = const()[name = string("op_6273_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_6273_cast_fp16 = reduce_mean(axes = var_6273_axes_0, keep_dims = var_6273_keep_dims_0, x = var_6268_cast_fp16)[name = string("op_6273_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6274_cast_fp16 = mul(x = s_207_cast_fp16, y = s_207_cast_fp16)[name = string("op_6274_cast_fp16")];
            fp32 var_6275_epsilon_0 = const()[name = string("op_6275_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_6275_cast_fp16 = inverse(epsilon = var_6275_epsilon_0, x = var_6274_cast_fp16)[name = string("op_6275_cast_fp16")];
            fp16 var_6276_to_fp16 = const()[name = string("op_6276_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_6277_cast_fp16 = mul(x = var_6275_cast_fp16, y = var_6276_to_fp16)[name = string("op_6277_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6279_cast_fp16 = add(x = var_6273_cast_fp16, y = var_6277_cast_fp16)[name = string("op_6279_cast_fp16")];
            fp32 var_6280_epsilon_0 = const()[name = string("op_6280_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_6280_cast_fp16 = rsqrt(epsilon = var_6280_epsilon_0, x = var_6279_cast_fp16)[name = string("op_6280_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_57_cast_fp16 = mul(x = y_203_cast_fp16, y = var_6280_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<int32, [4]> z_113_begin_0 = const()[name = string("z_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_113_end_0 = const()[name = string("z_113_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_113_end_mask_0 = const()[name = string("z_113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_113_cast_fp16 = slice_by_index(begin = z_113_begin_0, end = z_113_end_0, end_mask = z_113_end_mask_0, x = x_261_cast_fp16)[name = string("z_113_cast_fp16")];
            tensor<int32, [4]> var_6291_begin_0 = const()[name = string("op_6291_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6291_end_0 = const()[name = string("op_6291_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6291_end_mask_0 = const()[name = string("op_6291_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6291_cast_fp16 = slice_by_index(begin = var_6291_begin_0, end = var_6291_end_0, end_mask = var_6291_end_mask_0, x = z_113_cast_fp16)[name = string("op_6291_cast_fp16")];
            fp16 const_56_promoted_to_fp16 = const()[name = string("const_56_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6292_cast_fp16 = mul(x = var_6291_cast_fp16, y = const_56_promoted_to_fp16)[name = string("op_6292_cast_fp16")];
            tensor<int32, [4]> var_6297_begin_0 = const()[name = string("op_6297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6297_end_0 = const()[name = string("op_6297_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6297_end_mask_0 = const()[name = string("op_6297_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6297_cast_fp16 = slice_by_index(begin = var_6297_begin_0, end = var_6297_end_0, end_mask = var_6297_end_mask_0, x = z_113_cast_fp16)[name = string("op_6297_cast_fp16")];
            int32 var_6299 = const()[name = string("op_6299"), val = int32(-1)];
            bool rot_113_interleave_0 = const()[name = string("rot_113_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_113_cast_fp16 = concat(axis = var_6299, interleave = rot_113_interleave_0, values = (var_6292_cast_fp16, var_6297_cast_fp16))[name = string("rot_113_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6301_cast_fp16 = mul(x = z_113_cast_fp16, y = c_1_cast_fp16)[name = string("op_6301_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6302_cast_fp16 = mul(x = rot_113_cast_fp16, y = s_9_cast_fp16)[name = string("op_6302_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6304_cast_fp16 = add(x = var_6301_cast_fp16, y = var_6302_cast_fp16)[name = string("op_6304_cast_fp16")];
            tensor<int32, [4]> z_115_begin_0 = const()[name = string("z_115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_115_end_0 = const()[name = string("z_115_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_115_end_mask_0 = const()[name = string("z_115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_115_cast_fp16 = slice_by_index(begin = z_115_begin_0, end = z_115_end_0, end_mask = z_115_end_mask_0, x = x_261_cast_fp16)[name = string("z_115_cast_fp16")];
            tensor<int32, [4]> var_6314_begin_0 = const()[name = string("op_6314_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6314_end_0 = const()[name = string("op_6314_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6314_end_mask_0 = const()[name = string("op_6314_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> 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 = z_115_cast_fp16)[name = string("op_6314_cast_fp16")];
            fp16 const_57_promoted_to_fp16 = const()[name = string("const_57_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6315_cast_fp16 = mul(x = var_6314_cast_fp16, y = const_57_promoted_to_fp16)[name = string("op_6315_cast_fp16")];
            tensor<int32, [4]> var_6320_begin_0 = const()[name = string("op_6320_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6320_end_0 = const()[name = string("op_6320_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6320_end_mask_0 = const()[name = string("op_6320_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6320_cast_fp16 = slice_by_index(begin = var_6320_begin_0, end = var_6320_end_0, end_mask = var_6320_end_mask_0, x = z_115_cast_fp16)[name = string("op_6320_cast_fp16")];
            int32 var_6322 = const()[name = string("op_6322"), val = int32(-1)];
            bool rot_115_interleave_0 = const()[name = string("rot_115_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_115_cast_fp16 = concat(axis = var_6322, interleave = rot_115_interleave_0, values = (var_6315_cast_fp16, var_6320_cast_fp16))[name = string("rot_115_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6324_cast_fp16 = mul(x = z_115_cast_fp16, y = c_cast_fp16)[name = string("op_6324_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6325_cast_fp16 = mul(x = rot_115_cast_fp16, y = s_11_cast_fp16)[name = string("op_6325_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6327_cast_fp16 = add(x = var_6324_cast_fp16, y = var_6325_cast_fp16)[name = string("op_6327_cast_fp16")];
            int32 var_6329 = const()[name = string("op_6329"), val = int32(-1)];
            bool var_6330_interleave_0 = const()[name = string("op_6330_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_6330_cast_fp16 = concat(axis = var_6329, interleave = var_6330_interleave_0, values = (var_6304_cast_fp16, var_6327_cast_fp16))[name = string("op_6330_cast_fp16")];
            tensor<int32, [4]> z_117_begin_0 = const()[name = string("z_117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_117_end_0 = const()[name = string("z_117_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_117_end_mask_0 = const()[name = string("z_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_117_cast_fp16 = slice_by_index(begin = z_117_begin_0, end = z_117_end_0, end_mask = z_117_end_mask_0, x = x_263_cast_fp16)[name = string("z_117_cast_fp16")];
            tensor<int32, [4]> var_6343_begin_0 = const()[name = string("op_6343_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6343_end_0 = const()[name = string("op_6343_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6343_end_mask_0 = const()[name = string("op_6343_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6343_cast_fp16 = slice_by_index(begin = var_6343_begin_0, end = var_6343_end_0, end_mask = var_6343_end_mask_0, x = z_117_cast_fp16)[name = string("op_6343_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6344_cast_fp16 = mul(x = var_6343_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_6344_cast_fp16")];
            tensor<int32, [4]> var_6349_begin_0 = const()[name = string("op_6349_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6349_end_0 = const()[name = string("op_6349_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6349_end_mask_0 = const()[name = string("op_6349_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6349_cast_fp16 = slice_by_index(begin = var_6349_begin_0, end = var_6349_end_0, end_mask = var_6349_end_mask_0, x = z_117_cast_fp16)[name = string("op_6349_cast_fp16")];
            int32 var_6351 = const()[name = string("op_6351"), val = int32(-1)];
            bool rot_117_interleave_0 = const()[name = string("rot_117_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_117_cast_fp16 = concat(axis = var_6351, interleave = rot_117_interleave_0, values = (var_6344_cast_fp16, var_6349_cast_fp16))[name = string("rot_117_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6353_cast_fp16 = mul(x = z_117_cast_fp16, y = c_1_cast_fp16)[name = string("op_6353_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6354_cast_fp16 = mul(x = rot_117_cast_fp16, y = s_9_cast_fp16)[name = string("op_6354_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6356_cast_fp16 = add(x = var_6353_cast_fp16, y = var_6354_cast_fp16)[name = string("op_6356_cast_fp16")];
            tensor<int32, [4]> z_119_begin_0 = const()[name = string("z_119_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_119_end_0 = const()[name = string("z_119_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_119_end_mask_0 = const()[name = string("z_119_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_119_cast_fp16 = slice_by_index(begin = z_119_begin_0, end = z_119_end_0, end_mask = z_119_end_mask_0, x = x_263_cast_fp16)[name = string("z_119_cast_fp16")];
            tensor<int32, [4]> var_6366_begin_0 = const()[name = string("op_6366_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6366_end_0 = const()[name = string("op_6366_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6366_end_mask_0 = const()[name = string("op_6366_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6366_cast_fp16 = slice_by_index(begin = var_6366_begin_0, end = var_6366_end_0, end_mask = var_6366_end_mask_0, x = z_119_cast_fp16)[name = string("op_6366_cast_fp16")];
            fp16 const_59_promoted_to_fp16 = const()[name = string("const_59_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6367_cast_fp16 = mul(x = var_6366_cast_fp16, y = const_59_promoted_to_fp16)[name = string("op_6367_cast_fp16")];
            tensor<int32, [4]> var_6372_begin_0 = const()[name = string("op_6372_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6372_end_0 = const()[name = string("op_6372_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6372_end_mask_0 = const()[name = string("op_6372_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6372_cast_fp16 = slice_by_index(begin = var_6372_begin_0, end = var_6372_end_0, end_mask = var_6372_end_mask_0, x = z_119_cast_fp16)[name = string("op_6372_cast_fp16")];
            int32 var_6374 = const()[name = string("op_6374"), val = int32(-1)];
            bool rot_119_interleave_0 = const()[name = string("rot_119_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_119_cast_fp16 = concat(axis = var_6374, interleave = rot_119_interleave_0, values = (var_6367_cast_fp16, var_6372_cast_fp16))[name = string("rot_119_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6376_cast_fp16 = mul(x = z_119_cast_fp16, y = c_cast_fp16)[name = string("op_6376_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6377_cast_fp16 = mul(x = rot_119_cast_fp16, y = s_11_cast_fp16)[name = string("op_6377_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6379_cast_fp16 = add(x = var_6376_cast_fp16, y = var_6377_cast_fp16)[name = string("op_6379_cast_fp16")];
            int32 var_6381 = const()[name = string("op_6381"), val = int32(-1)];
            bool var_6382_interleave_0 = const()[name = string("op_6382_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_6382_cast_fp16 = concat(axis = var_6381, interleave = var_6382_interleave_0, values = (var_6356_cast_fp16, var_6379_cast_fp16))[name = string("op_6382_cast_fp16")];
            tensor<int32, [4]> v_59_perm_0 = const()[name = string("v_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6392_transpose_x_0 = const()[name = string("op_6392_transpose_x_0"), val = bool(false)];
            bool var_6392_transpose_y_0 = const()[name = string("op_6392_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_93 = transpose(perm = transpose_93_perm_0, x = var_6382_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 12, 630, 64]> transpose_92 = transpose(perm = transpose_92_perm_0, x = var_6330_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 12, 630, 630]> var_6392_cast_fp16 = matmul(transpose_x = var_6392_transpose_x_0, transpose_y = var_6392_transpose_y_0, x = transpose_92, y = transpose_93)[name = string("op_6392_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_6394_cast_fp16 = add(x = var_6392_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6394_cast_fp16")];
            int32 var_6395 = const()[name = string("op_6395"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_203_cast_fp16 = softmax(axis = var_6395, x = var_6394_cast_fp16)[name = string("w_203_cast_fp16")];
            bool var_6398_transpose_x_0 = const()[name = string("op_6398_transpose_x_0"), val = bool(false)];
            bool var_6398_transpose_y_0 = const()[name = string("op_6398_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_59_cast_fp16 = transpose(perm = v_59_perm_0, x = v_57_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 12, 630, 64]> var_6398_cast_fp16 = matmul(transpose_x = var_6398_transpose_x_0, transpose_y = var_6398_transpose_y_0, x = w_203_cast_fp16, y = v_59_cast_fp16)[name = string("op_6398_cast_fp16")];
            tensor<int32, [4]> var_6401_perm_0 = const()[name = string("op_6401_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6405 = const()[name = string("op_6405"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_6401_cast_fp16 = transpose(perm = var_6401_perm_0, x = var_6398_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 630, 768]> input_145_cast_fp16 = reshape(shape = var_6405, x = var_6401_cast_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134881472))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135471360))))[name = string("t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_102_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6410_cast_fp16 = abs(x = linear_102_cast_fp16)[name = string("op_6410_cast_fp16")];
            bool var_6413 = const()[name = string("op_6413"), val = bool(true)];
            tensor<int32, [1]> var_6414_axes_0 = const()[name = string("op_6414_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6414_cast_fp16 = reduce_max(axes = var_6414_axes_0, keep_dims = var_6413, x = var_6410_cast_fp16)[name = string("op_6414_cast_fp16")];
            fp16 var_6415_to_fp16 = const()[name = string("op_6415_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_209_cast_fp16 = maximum(x = var_6414_cast_fp16, y = var_6415_to_fp16)[name = string("s_209_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_205_cast_fp16 = real_div(x = linear_102_cast_fp16, y = s_209_cast_fp16)[name = string("y_205_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6418_cast_fp16 = mul(x = y_205_cast_fp16, y = y_205_cast_fp16)[name = string("op_6418_cast_fp16")];
            tensor<int32, [1]> var_6423_axes_0 = const()[name = string("op_6423_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6423_keep_dims_0 = const()[name = string("op_6423_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6423_cast_fp16 = reduce_mean(axes = var_6423_axes_0, keep_dims = var_6423_keep_dims_0, x = var_6418_cast_fp16)[name = string("op_6423_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6424_cast_fp16 = mul(x = s_209_cast_fp16, y = s_209_cast_fp16)[name = string("op_6424_cast_fp16")];
            fp32 var_6425_epsilon_0 = const()[name = string("op_6425_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6425_cast_fp16 = inverse(epsilon = var_6425_epsilon_0, x = var_6424_cast_fp16)[name = string("op_6425_cast_fp16")];
            fp16 var_6426_to_fp16 = const()[name = string("op_6426_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6427_cast_fp16 = mul(x = var_6425_cast_fp16, y = var_6426_to_fp16)[name = string("op_6427_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6429_cast_fp16 = add(x = var_6423_cast_fp16, y = var_6427_cast_fp16)[name = string("op_6429_cast_fp16")];
            fp32 var_6430_epsilon_0 = const()[name = string("op_6430_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6430_cast_fp16 = rsqrt(epsilon = var_6430_epsilon_0, x = var_6429_cast_fp16)[name = string("op_6430_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_175_cast_fp16 = mul(x = y_205_cast_fp16, y = var_6430_cast_fp16)[name = string("out_175_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135472960)))];
            tensor<fp16, [1, 630, 768]> var_6432_cast_fp16 = mul(x = out_175_cast_fp16, y = t_encoder_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("op_6432_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_267_cast_fp16 = add(x = x_253_cast_fp16, y = var_6432_cast_fp16)[name = string("x_267_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6435_cast_fp16 = abs(x = x_267_cast_fp16)[name = string("op_6435_cast_fp16")];
            bool var_6438 = const()[name = string("op_6438"), val = bool(true)];
            tensor<int32, [1]> var_6439_axes_0 = const()[name = string("op_6439_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6439_cast_fp16 = reduce_max(axes = var_6439_axes_0, keep_dims = var_6438, x = var_6435_cast_fp16)[name = string("op_6439_cast_fp16")];
            fp16 var_6440_to_fp16 = const()[name = string("op_6440_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_211_cast_fp16 = maximum(x = var_6439_cast_fp16, y = var_6440_to_fp16)[name = string("s_211_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_207_cast_fp16 = real_div(x = x_267_cast_fp16, y = s_211_cast_fp16)[name = string("y_207_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6443_cast_fp16 = mul(x = y_207_cast_fp16, y = y_207_cast_fp16)[name = string("op_6443_cast_fp16")];
            tensor<int32, [1]> var_6448_axes_0 = const()[name = string("op_6448_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6448_keep_dims_0 = const()[name = string("op_6448_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6448_cast_fp16 = reduce_mean(axes = var_6448_axes_0, keep_dims = var_6448_keep_dims_0, x = var_6443_cast_fp16)[name = string("op_6448_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6449_cast_fp16 = mul(x = s_211_cast_fp16, y = s_211_cast_fp16)[name = string("op_6449_cast_fp16")];
            fp32 var_6450_epsilon_0 = const()[name = string("op_6450_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6450_cast_fp16 = inverse(epsilon = var_6450_epsilon_0, x = var_6449_cast_fp16)[name = string("op_6450_cast_fp16")];
            fp16 var_6451_to_fp16 = const()[name = string("op_6451_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6452_cast_fp16 = mul(x = var_6450_cast_fp16, y = var_6451_to_fp16)[name = string("op_6452_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6454_cast_fp16 = add(x = var_6448_cast_fp16, y = var_6452_cast_fp16)[name = string("op_6454_cast_fp16")];
            fp32 var_6455_epsilon_0 = const()[name = string("op_6455_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6455_cast_fp16 = rsqrt(epsilon = var_6455_epsilon_0, x = var_6454_cast_fp16)[name = string("op_6455_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_177_cast_fp16 = mul(x = y_207_cast_fp16, y = var_6455_cast_fp16)[name = string("out_177_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135474560)))];
            tensor<fp16, [1, 630, 768]> input_147_cast_fp16 = mul(x = out_177_cast_fp16, y = t_encoder_layers_14_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(135476160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137835520))))[name = string("t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_103_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_103_cast_fp16")];
            string var_6462_mode_0 = const()[name = string("op_6462_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_6462_cast_fp16 = gelu(mode = var_6462_mode_0, x = linear_103_cast_fp16)[name = string("op_6462_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(137841728))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140201088))))[name = string("t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_104_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_147_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_151_cast_fp16 = mul(x = var_6462_cast_fp16, y = linear_104_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(140207296))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142566656))))[name = string("t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_14_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6470_cast_fp16 = abs(x = linear_105_cast_fp16)[name = string("op_6470_cast_fp16")];
            bool var_6473 = const()[name = string("op_6473"), val = bool(true)];
            tensor<int32, [1]> var_6474_axes_0 = const()[name = string("op_6474_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6474_cast_fp16 = reduce_max(axes = var_6474_axes_0, keep_dims = var_6473, x = var_6470_cast_fp16)[name = string("op_6474_cast_fp16")];
            fp16 var_6475_to_fp16 = const()[name = string("op_6475_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_213_cast_fp16 = maximum(x = var_6474_cast_fp16, y = var_6475_to_fp16)[name = string("s_213_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_209_cast_fp16 = real_div(x = linear_105_cast_fp16, y = s_213_cast_fp16)[name = string("y_209_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6478_cast_fp16 = mul(x = y_209_cast_fp16, y = y_209_cast_fp16)[name = string("op_6478_cast_fp16")];
            tensor<int32, [1]> var_6483_axes_0 = const()[name = string("op_6483_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6483_keep_dims_0 = const()[name = string("op_6483_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6483_cast_fp16 = reduce_mean(axes = var_6483_axes_0, keep_dims = var_6483_keep_dims_0, x = var_6478_cast_fp16)[name = string("op_6483_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6484_cast_fp16 = mul(x = s_213_cast_fp16, y = s_213_cast_fp16)[name = string("op_6484_cast_fp16")];
            fp32 var_6485_epsilon_0 = const()[name = string("op_6485_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6485_cast_fp16 = inverse(epsilon = var_6485_epsilon_0, x = var_6484_cast_fp16)[name = string("op_6485_cast_fp16")];
            fp16 var_6486_to_fp16 = const()[name = string("op_6486_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6487_cast_fp16 = mul(x = var_6485_cast_fp16, y = var_6486_to_fp16)[name = string("op_6487_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6489_cast_fp16 = add(x = var_6483_cast_fp16, y = var_6487_cast_fp16)[name = string("op_6489_cast_fp16")];
            fp32 var_6490_epsilon_0 = const()[name = string("op_6490_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6490_cast_fp16 = rsqrt(epsilon = var_6490_epsilon_0, x = var_6489_cast_fp16)[name = string("op_6490_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_179_cast_fp16 = mul(x = y_209_cast_fp16, y = var_6490_cast_fp16)[name = string("out_179_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142568256)))];
            tensor<fp16, [1, 630, 768]> var_6492_cast_fp16 = mul(x = out_179_cast_fp16, y = t_encoder_layers_14_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6492_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_271_cast_fp16 = add(x = x_267_cast_fp16, y = var_6492_cast_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6495_cast_fp16 = abs(x = x_271_cast_fp16)[name = string("op_6495_cast_fp16")];
            bool var_6498 = const()[name = string("op_6498"), val = bool(true)];
            tensor<int32, [1]> var_6499_axes_0 = const()[name = string("op_6499_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6499_cast_fp16 = reduce_max(axes = var_6499_axes_0, keep_dims = var_6498, x = var_6495_cast_fp16)[name = string("op_6499_cast_fp16")];
            fp16 var_6500_to_fp16 = const()[name = string("op_6500_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_215_cast_fp16 = maximum(x = var_6499_cast_fp16, y = var_6500_to_fp16)[name = string("s_215_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_211_cast_fp16 = real_div(x = x_271_cast_fp16, y = s_215_cast_fp16)[name = string("y_211_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6503_cast_fp16 = mul(x = y_211_cast_fp16, y = y_211_cast_fp16)[name = string("op_6503_cast_fp16")];
            tensor<int32, [1]> var_6508_axes_0 = const()[name = string("op_6508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6508_keep_dims_0 = const()[name = string("op_6508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6508_cast_fp16 = reduce_mean(axes = var_6508_axes_0, keep_dims = var_6508_keep_dims_0, x = var_6503_cast_fp16)[name = string("op_6508_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6509_cast_fp16 = mul(x = s_215_cast_fp16, y = s_215_cast_fp16)[name = string("op_6509_cast_fp16")];
            fp32 var_6510_epsilon_0 = const()[name = string("op_6510_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6510_cast_fp16 = inverse(epsilon = var_6510_epsilon_0, x = var_6509_cast_fp16)[name = string("op_6510_cast_fp16")];
            fp16 var_6511_to_fp16 = const()[name = string("op_6511_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6512_cast_fp16 = mul(x = var_6510_cast_fp16, y = var_6511_to_fp16)[name = string("op_6512_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6514_cast_fp16 = add(x = var_6508_cast_fp16, y = var_6512_cast_fp16)[name = string("op_6514_cast_fp16")];
            fp32 var_6515_epsilon_0 = const()[name = string("op_6515_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6515_cast_fp16 = rsqrt(epsilon = var_6515_epsilon_0, x = var_6514_cast_fp16)[name = string("op_6515_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_181_cast_fp16 = mul(x = y_211_cast_fp16, y = var_6515_cast_fp16)[name = string("out_181_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142569856)))];
            tensor<fp16, [1, 630, 768]> input_153_cast_fp16 = mul(x = out_181_cast_fp16, y = t_encoder_layers_15_input_layernorm_weight_to_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142571456))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143161344))))[name = string("t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_106_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_q_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_6525 = const()[name = string("op_6525"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_273_cast_fp16 = reshape(shape = var_6525, x = linear_106_cast_fp16)[name = string("x_273_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6527_cast_fp16 = abs(x = x_273_cast_fp16)[name = string("op_6527_cast_fp16")];
            bool var_6530 = const()[name = string("op_6530"), val = bool(true)];
            tensor<int32, [1]> var_6531_axes_0 = const()[name = string("op_6531_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_6531_cast_fp16 = reduce_max(axes = var_6531_axes_0, keep_dims = var_6530, x = var_6527_cast_fp16)[name = string("op_6531_cast_fp16")];
            fp16 var_6532_to_fp16 = const()[name = string("op_6532_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_217_cast_fp16 = maximum(x = var_6531_cast_fp16, y = var_6532_to_fp16)[name = string("s_217_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_213_cast_fp16 = real_div(x = x_273_cast_fp16, y = s_217_cast_fp16)[name = string("y_213_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6535_cast_fp16 = mul(x = y_213_cast_fp16, y = y_213_cast_fp16)[name = string("op_6535_cast_fp16")];
            tensor<int32, [1]> var_6540_axes_0 = const()[name = string("op_6540_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6540_keep_dims_0 = const()[name = string("op_6540_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_6540_cast_fp16 = reduce_mean(axes = var_6540_axes_0, keep_dims = var_6540_keep_dims_0, x = var_6535_cast_fp16)[name = string("op_6540_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6541_cast_fp16 = mul(x = s_217_cast_fp16, y = s_217_cast_fp16)[name = string("op_6541_cast_fp16")];
            fp32 var_6542_epsilon_0 = const()[name = string("op_6542_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_6542_cast_fp16 = inverse(epsilon = var_6542_epsilon_0, x = var_6541_cast_fp16)[name = string("op_6542_cast_fp16")];
            fp16 var_6543_to_fp16 = const()[name = string("op_6543_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_6544_cast_fp16 = mul(x = var_6542_cast_fp16, y = var_6543_to_fp16)[name = string("op_6544_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6546_cast_fp16 = add(x = var_6540_cast_fp16, y = var_6544_cast_fp16)[name = string("op_6546_cast_fp16")];
            fp32 var_6547_epsilon_0 = const()[name = string("op_6547_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_6547_cast_fp16 = rsqrt(epsilon = var_6547_epsilon_0, x = var_6546_cast_fp16)[name = string("op_6547_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_183_cast_fp16 = mul(x = y_213_cast_fp16, y = var_6547_cast_fp16)[name = string("out_183_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(114769024)))];
            tensor<fp16, [1, 630, 12, 64]> x_279_cast_fp16 = mul(x = out_183_cast_fp16, y = t_encoder_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_279_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143162944))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143752832))))[name = string("t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_107_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_k_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_6557 = const()[name = string("op_6557"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_275_cast_fp16 = reshape(shape = var_6557, x = linear_107_cast_fp16)[name = string("x_275_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6559_cast_fp16 = abs(x = x_275_cast_fp16)[name = string("op_6559_cast_fp16")];
            bool var_6562 = const()[name = string("op_6562"), val = bool(true)];
            tensor<int32, [1]> var_6563_axes_0 = const()[name = string("op_6563_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_6563_cast_fp16 = reduce_max(axes = var_6563_axes_0, keep_dims = var_6562, x = var_6559_cast_fp16)[name = string("op_6563_cast_fp16")];
            fp16 var_6564_to_fp16 = const()[name = string("op_6564_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_219_cast_fp16 = maximum(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = string("s_219_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_215_cast_fp16 = real_div(x = x_275_cast_fp16, y = s_219_cast_fp16)[name = string("y_215_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6567_cast_fp16 = mul(x = y_215_cast_fp16, y = y_215_cast_fp16)[name = string("op_6567_cast_fp16")];
            tensor<int32, [1]> var_6572_axes_0 = const()[name = string("op_6572_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6572_keep_dims_0 = const()[name = string("op_6572_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_6572_cast_fp16 = reduce_mean(axes = var_6572_axes_0, keep_dims = var_6572_keep_dims_0, x = var_6567_cast_fp16)[name = string("op_6572_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6573_cast_fp16 = mul(x = s_219_cast_fp16, y = s_219_cast_fp16)[name = string("op_6573_cast_fp16")];
            fp32 var_6574_epsilon_0 = const()[name = string("op_6574_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_6574_cast_fp16 = inverse(epsilon = var_6574_epsilon_0, x = var_6573_cast_fp16)[name = string("op_6574_cast_fp16")];
            fp16 var_6575_to_fp16 = const()[name = string("op_6575_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_6576_cast_fp16 = mul(x = var_6574_cast_fp16, y = var_6575_to_fp16)[name = string("op_6576_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6578_cast_fp16 = add(x = var_6572_cast_fp16, y = var_6576_cast_fp16)[name = string("op_6578_cast_fp16")];
            fp32 var_6579_epsilon_0 = const()[name = string("op_6579_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_6579_cast_fp16 = rsqrt(epsilon = var_6579_epsilon_0, x = var_6578_cast_fp16)[name = string("op_6579_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> out_185_cast_fp16 = mul(x = y_215_cast_fp16, y = var_6579_cast_fp16)[name = string("out_185_cast_fp16")];
            tensor<fp16, [64]> t_encoder_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115360704)))];
            tensor<fp16, [1, 630, 12, 64]> x_281_cast_fp16 = mul(x = out_185_cast_fp16, y = t_encoder_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_281_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(143754432))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144344320))))[name = string("t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_v_proj_linear_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<int32, [4]> var_6589 = const()[name = string("op_6589"), val = tensor<int32, [4]>([1, -1, 12, 64])];
            tensor<fp16, [1, 630, 12, 64]> x_277_cast_fp16 = reshape(shape = var_6589, x = linear_108_cast_fp16)[name = string("x_277_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6591_cast_fp16 = abs(x = x_277_cast_fp16)[name = string("op_6591_cast_fp16")];
            bool var_6594 = const()[name = string("op_6594"), val = bool(true)];
            tensor<int32, [1]> var_6595_axes_0 = const()[name = string("op_6595_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 12, 1]> var_6595_cast_fp16 = reduce_max(axes = var_6595_axes_0, keep_dims = var_6594, x = var_6591_cast_fp16)[name = string("op_6595_cast_fp16")];
            fp16 var_6596_to_fp16 = const()[name = string("op_6596_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 12, 1]> s_221_cast_fp16 = maximum(x = var_6595_cast_fp16, y = var_6596_to_fp16)[name = string("s_221_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> y_217_cast_fp16 = real_div(x = x_277_cast_fp16, y = s_221_cast_fp16)[name = string("y_217_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> var_6599_cast_fp16 = mul(x = y_217_cast_fp16, y = y_217_cast_fp16)[name = string("op_6599_cast_fp16")];
            tensor<int32, [1]> var_6604_axes_0 = const()[name = string("op_6604_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6604_keep_dims_0 = const()[name = string("op_6604_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 12, 1]> var_6604_cast_fp16 = reduce_mean(axes = var_6604_axes_0, keep_dims = var_6604_keep_dims_0, x = var_6599_cast_fp16)[name = string("op_6604_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6605_cast_fp16 = mul(x = s_221_cast_fp16, y = s_221_cast_fp16)[name = string("op_6605_cast_fp16")];
            fp32 var_6606_epsilon_0 = const()[name = string("op_6606_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 12, 1]> var_6606_cast_fp16 = inverse(epsilon = var_6606_epsilon_0, x = var_6605_cast_fp16)[name = string("op_6606_cast_fp16")];
            fp16 var_6607_to_fp16 = const()[name = string("op_6607_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 12, 1]> var_6608_cast_fp16 = mul(x = var_6606_cast_fp16, y = var_6607_to_fp16)[name = string("op_6608_cast_fp16")];
            tensor<fp16, [1, 630, 12, 1]> var_6610_cast_fp16 = add(x = var_6604_cast_fp16, y = var_6608_cast_fp16)[name = string("op_6610_cast_fp16")];
            fp32 var_6611_epsilon_0 = const()[name = string("op_6611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 12, 1]> var_6611_cast_fp16 = rsqrt(epsilon = var_6611_epsilon_0, x = var_6610_cast_fp16)[name = string("op_6611_cast_fp16")];
            tensor<fp16, [1, 630, 12, 64]> v_61_cast_fp16 = mul(x = y_217_cast_fp16, y = var_6611_cast_fp16)[name = string("v_61_cast_fp16")];
            tensor<int32, [4]> z_121_begin_0 = const()[name = string("z_121_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_121_end_0 = const()[name = string("z_121_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_121_end_mask_0 = const()[name = string("z_121_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_121_cast_fp16 = slice_by_index(begin = z_121_begin_0, end = z_121_end_0, end_mask = z_121_end_mask_0, x = x_279_cast_fp16)[name = string("z_121_cast_fp16")];
            tensor<int32, [4]> var_6622_begin_0 = const()[name = string("op_6622_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6622_end_0 = const()[name = string("op_6622_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6622_end_mask_0 = const()[name = string("op_6622_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6622_cast_fp16 = slice_by_index(begin = var_6622_begin_0, end = var_6622_end_0, end_mask = var_6622_end_mask_0, x = z_121_cast_fp16)[name = string("op_6622_cast_fp16")];
            fp16 const_60_promoted_to_fp16 = const()[name = string("const_60_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6623_cast_fp16 = mul(x = var_6622_cast_fp16, y = const_60_promoted_to_fp16)[name = string("op_6623_cast_fp16")];
            tensor<int32, [4]> var_6628_begin_0 = const()[name = string("op_6628_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6628_end_0 = const()[name = string("op_6628_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6628_end_mask_0 = const()[name = string("op_6628_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6628_cast_fp16 = slice_by_index(begin = var_6628_begin_0, end = var_6628_end_0, end_mask = var_6628_end_mask_0, x = z_121_cast_fp16)[name = string("op_6628_cast_fp16")];
            int32 var_6630 = const()[name = string("op_6630"), val = int32(-1)];
            bool rot_121_interleave_0 = const()[name = string("rot_121_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_121_cast_fp16 = concat(axis = var_6630, interleave = rot_121_interleave_0, values = (var_6623_cast_fp16, var_6628_cast_fp16))[name = string("rot_121_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6632_cast_fp16 = mul(x = z_121_cast_fp16, y = c_1_cast_fp16)[name = string("op_6632_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6633_cast_fp16 = mul(x = rot_121_cast_fp16, y = s_9_cast_fp16)[name = string("op_6633_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6635_cast_fp16 = add(x = var_6632_cast_fp16, y = var_6633_cast_fp16)[name = string("op_6635_cast_fp16")];
            tensor<int32, [4]> z_123_begin_0 = const()[name = string("z_123_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_123_end_0 = const()[name = string("z_123_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_123_end_mask_0 = const()[name = string("z_123_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_123_cast_fp16 = slice_by_index(begin = z_123_begin_0, end = z_123_end_0, end_mask = z_123_end_mask_0, x = x_279_cast_fp16)[name = string("z_123_cast_fp16")];
            tensor<int32, [4]> var_6645_begin_0 = const()[name = string("op_6645_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6645_end_0 = const()[name = string("op_6645_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6645_end_mask_0 = const()[name = string("op_6645_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6645_cast_fp16 = slice_by_index(begin = var_6645_begin_0, end = var_6645_end_0, end_mask = var_6645_end_mask_0, x = z_123_cast_fp16)[name = string("op_6645_cast_fp16")];
            fp16 const_61_promoted_to_fp16 = const()[name = string("const_61_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6646_cast_fp16 = mul(x = var_6645_cast_fp16, y = const_61_promoted_to_fp16)[name = string("op_6646_cast_fp16")];
            tensor<int32, [4]> var_6651_begin_0 = const()[name = string("op_6651_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6651_end_0 = const()[name = string("op_6651_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6651_end_mask_0 = const()[name = string("op_6651_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6651_cast_fp16 = slice_by_index(begin = var_6651_begin_0, end = var_6651_end_0, end_mask = var_6651_end_mask_0, x = z_123_cast_fp16)[name = string("op_6651_cast_fp16")];
            int32 var_6653 = const()[name = string("op_6653"), val = int32(-1)];
            bool rot_123_interleave_0 = const()[name = string("rot_123_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_123_cast_fp16 = concat(axis = var_6653, interleave = rot_123_interleave_0, values = (var_6646_cast_fp16, var_6651_cast_fp16))[name = string("rot_123_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6655_cast_fp16 = mul(x = z_123_cast_fp16, y = c_cast_fp16)[name = string("op_6655_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6656_cast_fp16 = mul(x = rot_123_cast_fp16, y = s_11_cast_fp16)[name = string("op_6656_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6658_cast_fp16 = add(x = var_6655_cast_fp16, y = var_6656_cast_fp16)[name = string("op_6658_cast_fp16")];
            int32 var_6660 = const()[name = string("op_6660"), val = int32(-1)];
            bool var_6661_interleave_0 = const()[name = string("op_6661_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_6661_cast_fp16 = concat(axis = var_6660, interleave = var_6661_interleave_0, values = (var_6635_cast_fp16, var_6658_cast_fp16))[name = string("op_6661_cast_fp16")];
            tensor<int32, [4]> z_125_begin_0 = const()[name = string("z_125_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> z_125_end_0 = const()[name = string("z_125_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> z_125_end_mask_0 = const()[name = string("z_125_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 32]> z_125_cast_fp16 = slice_by_index(begin = z_125_begin_0, end = z_125_end_0, end_mask = z_125_end_mask_0, x = x_281_cast_fp16)[name = string("z_125_cast_fp16")];
            tensor<int32, [4]> var_6674_begin_0 = const()[name = string("op_6674_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6674_end_0 = const()[name = string("op_6674_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6674_end_mask_0 = const()[name = string("op_6674_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6674_cast_fp16 = slice_by_index(begin = var_6674_begin_0, end = var_6674_end_0, end_mask = var_6674_end_mask_0, x = z_125_cast_fp16)[name = string("op_6674_cast_fp16")];
            fp16 const_62_promoted_to_fp16 = const()[name = string("const_62_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6675_cast_fp16 = mul(x = var_6674_cast_fp16, y = const_62_promoted_to_fp16)[name = string("op_6675_cast_fp16")];
            tensor<int32, [4]> var_6680_begin_0 = const()[name = string("op_6680_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6680_end_0 = const()[name = string("op_6680_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6680_end_mask_0 = const()[name = string("op_6680_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6680_cast_fp16 = slice_by_index(begin = var_6680_begin_0, end = var_6680_end_0, end_mask = var_6680_end_mask_0, x = z_125_cast_fp16)[name = string("op_6680_cast_fp16")];
            int32 var_6682 = const()[name = string("op_6682"), val = int32(-1)];
            bool rot_125_interleave_0 = const()[name = string("rot_125_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_125_cast_fp16 = concat(axis = var_6682, interleave = rot_125_interleave_0, values = (var_6675_cast_fp16, var_6680_cast_fp16))[name = string("rot_125_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6684_cast_fp16 = mul(x = z_125_cast_fp16, y = c_1_cast_fp16)[name = string("op_6684_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6685_cast_fp16 = mul(x = rot_125_cast_fp16, y = s_9_cast_fp16)[name = string("op_6685_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6687_cast_fp16 = add(x = var_6684_cast_fp16, y = var_6685_cast_fp16)[name = string("op_6687_cast_fp16")];
            tensor<int32, [4]> z_begin_0 = const()[name = string("z_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> z_end_0 = const()[name = string("z_end_0"), val = tensor<int32, [4]>([1, 630, 12, 1])];
            tensor<bool, [4]> z_end_mask_0 = const()[name = string("z_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 32]> z_cast_fp16 = slice_by_index(begin = z_begin_0, end = z_end_0, end_mask = z_end_mask_0, x = x_281_cast_fp16)[name = string("z_cast_fp16")];
            tensor<int32, [4]> var_6697_begin_0 = const()[name = string("op_6697_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 16])];
            tensor<int32, [4]> var_6697_end_0 = const()[name = string("op_6697_end_0"), val = tensor<int32, [4]>([1, 630, 12, 32])];
            tensor<bool, [4]> var_6697_end_mask_0 = const()[name = string("op_6697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 630, 12, 16]> var_6697_cast_fp16 = slice_by_index(begin = var_6697_begin_0, end = var_6697_end_0, end_mask = var_6697_end_mask_0, x = z_cast_fp16)[name = string("op_6697_cast_fp16")];
            fp16 const_63_promoted_to_fp16 = const()[name = string("const_63_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 630, 12, 16]> var_6698_cast_fp16 = mul(x = var_6697_cast_fp16, y = const_63_promoted_to_fp16)[name = string("op_6698_cast_fp16")];
            tensor<int32, [4]> var_6703_begin_0 = const()[name = string("op_6703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6703_end_0 = const()[name = string("op_6703_end_0"), val = tensor<int32, [4]>([1, 630, 12, 16])];
            tensor<bool, [4]> var_6703_end_mask_0 = const()[name = string("op_6703_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 630, 12, 16]> var_6703_cast_fp16 = slice_by_index(begin = var_6703_begin_0, end = var_6703_end_0, end_mask = var_6703_end_mask_0, x = z_cast_fp16)[name = string("op_6703_cast_fp16")];
            int32 var_6705 = const()[name = string("op_6705"), val = int32(-1)];
            bool rot_interleave_0 = const()[name = string("rot_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 32]> rot_cast_fp16 = concat(axis = var_6705, interleave = rot_interleave_0, values = (var_6698_cast_fp16, var_6703_cast_fp16))[name = string("rot_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6707_cast_fp16 = mul(x = z_cast_fp16, y = c_cast_fp16)[name = string("op_6707_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6708_cast_fp16 = mul(x = rot_cast_fp16, y = s_11_cast_fp16)[name = string("op_6708_cast_fp16")];
            tensor<fp16, [1, 630, 12, 32]> var_6710_cast_fp16 = add(x = var_6707_cast_fp16, y = var_6708_cast_fp16)[name = string("op_6710_cast_fp16")];
            int32 var_6712 = const()[name = string("op_6712"), val = int32(-1)];
            bool var_6713_interleave_0 = const()[name = string("op_6713_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 630, 12, 64]> var_6713_cast_fp16 = concat(axis = var_6712, interleave = var_6713_interleave_0, values = (var_6687_cast_fp16, var_6710_cast_fp16))[name = string("op_6713_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            bool var_6723_transpose_x_0 = const()[name = string("op_6723_transpose_x_0"), val = bool(false)];
            bool var_6723_transpose_y_0 = const()[name = string("op_6723_transpose_y_0"), val = bool(false)];
            tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp16, [1, 12, 64, 630]> transpose_95 = transpose(perm = transpose_95_perm_0, x = var_6713_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 12, 630, 64]> transpose_94 = transpose(perm = transpose_94_perm_0, x = var_6661_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 12, 630, 630]> var_6723_cast_fp16 = matmul(transpose_x = var_6723_transpose_x_0, transpose_y = var_6723_transpose_y_0, x = transpose_94, y = transpose_95)[name = string("op_6723_cast_fp16")];
            tensor<fp16, [1, 12, 630, 630]> var_6725_cast_fp16 = add(x = var_6723_cast_fp16, y = var_1524_cast_fp16)[name = string("op_6725_cast_fp16")];
            int32 var_6726 = const()[name = string("op_6726"), val = int32(-1)];
            tensor<fp16, [1, 12, 630, 630]> w_217_cast_fp16 = softmax(axis = var_6726, x = var_6725_cast_fp16)[name = string("w_217_cast_fp16")];
            bool var_6729_transpose_x_0 = const()[name = string("op_6729_transpose_x_0"), val = bool(false)];
            bool var_6729_transpose_y_0 = const()[name = string("op_6729_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 12, 630, 64]> v_cast_fp16 = transpose(perm = v_perm_0, x = v_61_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 12, 630, 64]> var_6729_cast_fp16 = matmul(transpose_x = var_6729_transpose_x_0, transpose_y = var_6729_transpose_y_0, x = w_217_cast_fp16, y = v_cast_fp16)[name = string("op_6729_cast_fp16")];
            tensor<int32, [4]> var_6732_perm_0 = const()[name = string("op_6732_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6736 = const()[name = string("op_6736"), val = tensor<int32, [3]>([1, -1, 768])];
            tensor<fp16, [1, 630, 12, 64]> var_6732_cast_fp16 = transpose(perm = var_6732_perm_0, x = var_6729_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 630, 768]> input_155_cast_fp16 = reshape(shape = var_6736, x = var_6732_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [768, 768]> t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144345920))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144935808))))[name = string("t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_109_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_self_attn_o_proj_linear_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6741_cast_fp16 = abs(x = linear_109_cast_fp16)[name = string("op_6741_cast_fp16")];
            bool var_6744 = const()[name = string("op_6744"), val = bool(true)];
            tensor<int32, [1]> var_6745_axes_0 = const()[name = string("op_6745_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6745_cast_fp16 = reduce_max(axes = var_6745_axes_0, keep_dims = var_6744, x = var_6741_cast_fp16)[name = string("op_6745_cast_fp16")];
            fp16 var_6746_to_fp16 = const()[name = string("op_6746_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_223_cast_fp16 = maximum(x = var_6745_cast_fp16, y = var_6746_to_fp16)[name = string("s_223_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_219_cast_fp16 = real_div(x = linear_109_cast_fp16, y = s_223_cast_fp16)[name = string("y_219_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6749_cast_fp16 = mul(x = y_219_cast_fp16, y = y_219_cast_fp16)[name = string("op_6749_cast_fp16")];
            tensor<int32, [1]> var_6754_axes_0 = const()[name = string("op_6754_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6754_keep_dims_0 = const()[name = string("op_6754_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6754_cast_fp16 = reduce_mean(axes = var_6754_axes_0, keep_dims = var_6754_keep_dims_0, x = var_6749_cast_fp16)[name = string("op_6754_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6755_cast_fp16 = mul(x = s_223_cast_fp16, y = s_223_cast_fp16)[name = string("op_6755_cast_fp16")];
            fp32 var_6756_epsilon_0 = const()[name = string("op_6756_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6756_cast_fp16 = inverse(epsilon = var_6756_epsilon_0, x = var_6755_cast_fp16)[name = string("op_6756_cast_fp16")];
            fp16 var_6757_to_fp16 = const()[name = string("op_6757_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6758_cast_fp16 = mul(x = var_6756_cast_fp16, y = var_6757_to_fp16)[name = string("op_6758_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6760_cast_fp16 = add(x = var_6754_cast_fp16, y = var_6758_cast_fp16)[name = string("op_6760_cast_fp16")];
            fp32 var_6761_epsilon_0 = const()[name = string("op_6761_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6761_cast_fp16 = rsqrt(epsilon = var_6761_epsilon_0, x = var_6760_cast_fp16)[name = string("op_6761_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_187_cast_fp16 = mul(x = y_219_cast_fp16, y = var_6761_cast_fp16)[name = string("out_187_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144937408)))];
            tensor<fp16, [1, 630, 768]> var_6763_cast_fp16 = mul(x = out_187_cast_fp16, y = t_encoder_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("op_6763_cast_fp16")];
            tensor<fp16, [1, 630, 768]> x_285_cast_fp16 = add(x = x_271_cast_fp16, y = var_6763_cast_fp16)[name = string("x_285_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6766_cast_fp16 = abs(x = x_285_cast_fp16)[name = string("op_6766_cast_fp16")];
            bool var_6769 = const()[name = string("op_6769"), val = bool(true)];
            tensor<int32, [1]> var_6770_axes_0 = const()[name = string("op_6770_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6770_cast_fp16 = reduce_max(axes = var_6770_axes_0, keep_dims = var_6769, x = var_6766_cast_fp16)[name = string("op_6770_cast_fp16")];
            fp16 var_6771_to_fp16 = const()[name = string("op_6771_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_225_cast_fp16 = maximum(x = var_6770_cast_fp16, y = var_6771_to_fp16)[name = string("s_225_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_221_cast_fp16 = real_div(x = x_285_cast_fp16, y = s_225_cast_fp16)[name = string("y_221_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6774_cast_fp16 = mul(x = y_221_cast_fp16, y = y_221_cast_fp16)[name = string("op_6774_cast_fp16")];
            tensor<int32, [1]> var_6779_axes_0 = const()[name = string("op_6779_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6779_keep_dims_0 = const()[name = string("op_6779_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6779_cast_fp16 = reduce_mean(axes = var_6779_axes_0, keep_dims = var_6779_keep_dims_0, x = var_6774_cast_fp16)[name = string("op_6779_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6780_cast_fp16 = mul(x = s_225_cast_fp16, y = s_225_cast_fp16)[name = string("op_6780_cast_fp16")];
            fp32 var_6781_epsilon_0 = const()[name = string("op_6781_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6781_cast_fp16 = inverse(epsilon = var_6781_epsilon_0, x = var_6780_cast_fp16)[name = string("op_6781_cast_fp16")];
            fp16 var_6782_to_fp16 = const()[name = string("op_6782_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6783_cast_fp16 = mul(x = var_6781_cast_fp16, y = var_6782_to_fp16)[name = string("op_6783_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6785_cast_fp16 = add(x = var_6779_cast_fp16, y = var_6783_cast_fp16)[name = string("op_6785_cast_fp16")];
            fp32 var_6786_epsilon_0 = const()[name = string("op_6786_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6786_cast_fp16 = rsqrt(epsilon = var_6786_epsilon_0, x = var_6785_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_189_cast_fp16 = mul(x = y_221_cast_fp16, y = var_6786_cast_fp16)[name = string("out_189_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144939008)))];
            tensor<fp16, [1, 630, 768]> input_157_cast_fp16 = mul(x = out_189_cast_fp16, y = t_encoder_layers_15_pre_feedforward_layernorm_weight_to_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(144940608))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147299968))))[name = string("t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_110_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_gate_proj_linear_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_110_cast_fp16")];
            string var_6793_mode_0 = const()[name = string("op_6793_mode_0"), val = string("TANH_APPROXIMATION")];
            tensor<fp16, [1, 630, 3072]> var_6793_cast_fp16 = gelu(mode = var_6793_mode_0, x = linear_110_cast_fp16)[name = string("op_6793_cast_fp16")];
            tensor<fp16, [3072, 768]> t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(147306176))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149665536))))[name = string("t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 3072]> linear_111_cast_fp16 = linear(bias = linear_5_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_up_proj_linear_weight_to_fp16_quantized, x = input_157_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 630, 3072]> input_161_cast_fp16 = mul(x = var_6793_cast_fp16, y = linear_111_cast_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [768, 3072]> t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [768, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(149671744))), scale = tensor<fp16, [768, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152031104))))[name = string("t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized")];
            tensor<fp16, [1, 630, 768]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = t_encoder_layers_15_mlp_down_proj_linear_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6801_cast_fp16 = abs(x = linear_112_cast_fp16)[name = string("op_6801_cast_fp16")];
            bool var_6804 = const()[name = string("op_6804"), val = bool(true)];
            tensor<int32, [1]> var_6805_axes_0 = const()[name = string("op_6805_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [1, 630, 1]> var_6805_cast_fp16 = reduce_max(axes = var_6805_axes_0, keep_dims = var_6804, x = var_6801_cast_fp16)[name = string("op_6805_cast_fp16")];
            fp16 var_6806_to_fp16 = const()[name = string("op_6806_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [1, 630, 1]> s_227_cast_fp16 = maximum(x = var_6805_cast_fp16, y = var_6806_to_fp16)[name = string("s_227_cast_fp16")];
            tensor<fp16, [1, 630, 768]> y_223_cast_fp16 = real_div(x = linear_112_cast_fp16, y = s_227_cast_fp16)[name = string("y_223_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6809_cast_fp16 = mul(x = y_223_cast_fp16, y = y_223_cast_fp16)[name = string("op_6809_cast_fp16")];
            tensor<int32, [1]> var_6814_axes_0 = const()[name = string("op_6814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6814_keep_dims_0 = const()[name = string("op_6814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 630, 1]> var_6814_cast_fp16 = reduce_mean(axes = var_6814_axes_0, keep_dims = var_6814_keep_dims_0, x = var_6809_cast_fp16)[name = string("op_6814_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6815_cast_fp16 = mul(x = s_227_cast_fp16, y = s_227_cast_fp16)[name = string("op_6815_cast_fp16")];
            fp32 var_6816_epsilon_0 = const()[name = string("op_6816_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [1, 630, 1]> var_6816_cast_fp16 = inverse(epsilon = var_6816_epsilon_0, x = var_6815_cast_fp16)[name = string("op_6816_cast_fp16")];
            fp16 var_6817_to_fp16 = const()[name = string("op_6817_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 630, 1]> var_6818_cast_fp16 = mul(x = var_6816_cast_fp16, y = var_6817_to_fp16)[name = string("op_6818_cast_fp16")];
            tensor<fp16, [1, 630, 1]> var_6820_cast_fp16 = add(x = var_6814_cast_fp16, y = var_6818_cast_fp16)[name = string("op_6820_cast_fp16")];
            fp32 var_6821_epsilon_0 = const()[name = string("op_6821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 630, 1]> var_6821_cast_fp16 = rsqrt(epsilon = var_6821_epsilon_0, x = var_6820_cast_fp16)[name = string("op_6821_cast_fp16")];
            tensor<fp16, [1, 630, 768]> out_cast_fp16 = mul(x = y_223_cast_fp16, y = var_6821_cast_fp16)[name = string("out_cast_fp16")];
            tensor<fp16, [768]> t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16 = const()[name = string("t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16"), val = tensor<fp16, [768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152032704)))];
            tensor<fp16, [1, 630, 768]> var_6823_cast_fp16 = mul(x = out_cast_fp16, y = t_encoder_layers_15_post_feedforward_layernorm_weight_to_fp16)[name = string("op_6823_cast_fp16")];
            tensor<fp16, [1, 630, 768]> h_cast_fp16 = add(x = x_285_cast_fp16, y = var_6823_cast_fp16)[name = string("h_cast_fp16")];
            tensor<fp16, [1, 630, 768]> var_6828_cast_fp16 = mul(x = h_cast_fp16, y = var_1413_cast_fp16)[name = string("op_6828_cast_fp16")];
            tensor<int32, [3]> var_6831_begin_0 = const()[name = string("op_6831_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_6831_end_0 = const()[name = string("op_6831_end_0"), val = tensor<int32, [3]>([1, 630, 768])];
            tensor<bool, [3]> var_6831_end_mask_0 = const()[name = string("op_6831_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> var_6831_squeeze_mask_0 = const()[name = string("op_6831_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<fp16, [630, 768]> var_6831_cast_fp16 = slice_by_index(begin = var_6831_begin_0, end = var_6831_end_0, end_mask = var_6831_end_mask_0, squeeze_mask = var_6831_squeeze_mask_0, x = var_6828_cast_fp16)[name = string("op_6831_cast_fp16")];
            bool x_transpose_x_0 = const()[name = string("x_transpose_x_0"), val = bool(false)];
            bool x_transpose_y_0 = const()[name = string("x_transpose_y_0"), val = bool(false)];
            tensor<fp16, [70, 768]> x_cast_fp16 = matmul(transpose_x = x_transpose_x_0, transpose_y = x_transpose_y_0, x = pool, y = var_6831_cast_fp16)[name = string("x_cast_fp16")];
            tensor<fp16, [70, 768]> var_6833_cast_fp16 = abs(x = x_cast_fp16)[name = string("op_6833_cast_fp16")];
            bool var_6836 = const()[name = string("op_6836"), val = bool(true)];
            tensor<int32, [1]> var_6837_axes_0 = const()[name = string("op_6837_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [70, 1]> var_6837_cast_fp16 = reduce_max(axes = var_6837_axes_0, keep_dims = var_6836, x = var_6833_cast_fp16)[name = string("op_6837_cast_fp16")];
            fp16 var_6838_to_fp16 = const()[name = string("op_6838_to_fp16"), val = fp16(0x1.a38p-14)];
            tensor<fp16, [70, 1]> s_cast_fp16 = maximum(x = var_6837_cast_fp16, y = var_6838_to_fp16)[name = string("s_cast_fp16")];
            tensor<fp16, [70, 768]> y_cast_fp16 = real_div(x = x_cast_fp16, y = s_cast_fp16)[name = string("y_cast_fp16")];
            tensor<fp16, [70, 768]> var_6841_cast_fp16 = mul(x = y_cast_fp16, y = y_cast_fp16)[name = string("op_6841_cast_fp16")];
            tensor<int32, [1]> var_6846_axes_0 = const()[name = string("op_6846_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6846_keep_dims_0 = const()[name = string("op_6846_keep_dims_0"), val = bool(true)];
            tensor<fp16, [70, 1]> var_6846_cast_fp16 = reduce_mean(axes = var_6846_axes_0, keep_dims = var_6846_keep_dims_0, x = var_6841_cast_fp16)[name = string("op_6846_cast_fp16")];
            tensor<fp16, [70, 1]> var_6847_cast_fp16 = mul(x = s_cast_fp16, y = s_cast_fp16)[name = string("op_6847_cast_fp16")];
            fp32 var_6848_epsilon_0 = const()[name = string("op_6848_epsilon_0"), val = fp32(0x1.a36e2ep-14)];
            tensor<fp16, [70, 1]> var_6848_cast_fp16 = inverse(epsilon = var_6848_epsilon_0, x = var_6847_cast_fp16)[name = string("op_6848_cast_fp16")];
            fp16 var_6849_to_fp16 = const()[name = string("op_6849_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [70, 1]> var_6850_cast_fp16 = mul(x = var_6848_cast_fp16, y = var_6849_to_fp16)[name = string("op_6850_cast_fp16")];
            tensor<fp16, [70, 1]> var_6852_cast_fp16 = add(x = var_6846_cast_fp16, y = var_6850_cast_fp16)[name = string("op_6852_cast_fp16")];
            fp32 var_6853_epsilon_0 = const()[name = string("op_6853_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [70, 1]> var_6853_cast_fp16 = rsqrt(epsilon = var_6853_epsilon_0, x = var_6852_cast_fp16)[name = string("op_6853_cast_fp16")];
            tensor<fp16, [70, 768]> input_cast_fp16 = mul(x = y_cast_fp16, y = var_6853_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [512, 768]> e_embedding_projection_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [512, 768]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152034304))), scale = tensor<fp16, [512, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152427584))))[name = string("e_embedding_projection_weight_to_fp16_quantized")];
            tensor<fp16, [512]> linear_113_bias_0_to_fp16 = const()[name = string("linear_113_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(152428672)))];
            tensor<fp16, [70, 512]> linear_113_cast_fp16 = linear(bias = linear_113_bias_0_to_fp16, weight = e_embedding_projection_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [1]> var_6859_axes_0 = const()[name = string("op_6859_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 70, 512]> image_tokens = expand_dims(axes = var_6859_axes_0, x = linear_113_cast_fp16)[name = string("op_6859_cast_fp16")];
        } -> (image_tokens);
}