program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}})]
{
    func main<ios17>(tensor<fp32, [1, 128, 15]> cache_s0_l0_conv1, tensor<fp32, [1, 128, 15]> cache_s0_l0_conv2, tensor<fp32, [64, 1, 128]> cache_s0_l0_key, tensor<fp32, [1, 1, 64, 96]> cache_s0_l0_nonlin_attn, tensor<fp32, [64, 1, 48]> cache_s0_l0_val1, tensor<fp32, [64, 1, 48]> cache_s0_l0_val2, tensor<fp32, [1, 128, 15]> cache_s1_l0_conv1, tensor<fp32, [1, 128, 15]> cache_s1_l0_conv2, tensor<fp32, [32, 1, 128]> cache_s1_l0_key, tensor<fp32, [1, 1, 32, 96]> cache_s1_l0_nonlin_attn, tensor<fp32, [32, 1, 48]> cache_s1_l0_val1, tensor<fp32, [32, 1, 48]> cache_s1_l0_val2, tensor<fp32, [1, 128, 7]> cache_s2_l0_conv1, tensor<fp32, [1, 128, 7]> cache_s2_l0_conv2, tensor<fp32, [16, 1, 128]> cache_s2_l0_key, tensor<fp32, [1, 1, 16, 96]> cache_s2_l0_nonlin_attn, tensor<fp32, [16, 1, 48]> cache_s2_l0_val1, tensor<fp32, [16, 1, 48]> cache_s2_l0_val2, tensor<fp32, [1, 128, 7]> cache_s3_l0_conv1, tensor<fp32, [1, 128, 7]> cache_s3_l0_conv2, tensor<fp32, [8, 1, 256]> cache_s3_l0_key, tensor<fp32, [1, 1, 8, 96]> cache_s3_l0_nonlin_attn, tensor<fp32, [8, 1, 96]> cache_s3_l0_val1, tensor<fp32, [8, 1, 96]> cache_s3_l0_val2, tensor<fp32, [1, 128, 7]> cache_s4_l0_conv1, tensor<fp32, [1, 128, 7]> cache_s4_l0_conv2, tensor<fp32, [16, 1, 128]> cache_s4_l0_key, tensor<fp32, [1, 1, 16, 96]> cache_s4_l0_nonlin_attn, tensor<fp32, [16, 1, 48]> cache_s4_l0_val1, tensor<fp32, [16, 1, 48]> cache_s4_l0_val2, tensor<fp32, [1, 128, 15]> cache_s5_l0_conv1, tensor<fp32, [1, 128, 15]> cache_s5_l0_conv2, tensor<fp32, [32, 1, 128]> cache_s5_l0_key, tensor<fp32, [1, 1, 32, 96]> cache_s5_l0_nonlin_attn, tensor<fp32, [32, 1, 48]> cache_s5_l0_val1, tensor<fp32, [32, 1, 48]> cache_s5_l0_val2, tensor<fp32, [1, 128, 3, 19]> cached_embed_left_pad, tensor<int32, [1]> processed_lens, tensor<fp32, [1, 45, 80]> x) {
            tensor<int32, []> var_46 = const()[name = tensor<string, []>("op_46"), val = tensor<int32, []>(0)];
            tensor<int32, []> var_48 = const()[name = tensor<string, []>("op_48"), val = tensor<int32, []>(1)];
            tensor<int32, []> var_49 = const()[name = tensor<string, []>("op_49"), val = tensor<int32, []>(2)];
            tensor<int32, []> var_59 = const()[name = tensor<string, []>("op_59"), val = tensor<int32, []>(-1)];
            tensor<int32, [1]> input_3_axes_0 = const()[name = tensor<string, []>("input_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<string, []> x_to_fp16_dtype_0 = const()[name = tensor<string, []>("x_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 45, 80]> x_to_fp16 = cast(dtype = x_to_fp16_dtype_0, x = x)[name = tensor<string, []>("cast_38")];
            tensor<fp16, [1, 1, 45, 80]> input_3_cast_fp16 = expand_dims(axes = input_3_axes_0, x = x_to_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [4]>([0, 0, 1, 1])];
            tensor<int32, [2]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [8, 1, 3, 3]> encoder_embed_conv_0_weight_to_fp16 = const()[name = tensor<string, []>("encoder_embed_conv_0_weight_to_fp16"), val = tensor<fp16, [8, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [8]> encoder_embed_conv_0_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_conv_0_bias_to_fp16"), val = tensor<fp16, [8]>([0x1.4f4p-1, 0x1.25cp-2, 0x1.81cp-3, 0x1.2a8p-3, 0x1.ff4p-4, -0x1.05p-1, 0x1.dd8p-5, 0x1.08cp-4])];
            tensor<fp16, [1, 8, 43, 80]> x_25_cast_fp16 = conv(bias = encoder_embed_conv_0_bias_to_fp16, dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = encoder_embed_conv_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<fp16, []> var_1320_to_fp16 = const()[name = tensor<string, []>("op_1320_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 8, 43, 80]> y_2_cast_fp16 = sub(x = x_25_cast_fp16, y = var_1320_to_fp16)[name = tensor<string, []>("y_2_cast_fp16")];
            tensor<fp16, []> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 8, 43, 80]> max_value_2_cast_fp16 = maximum(x = const_2_to_fp16, y = y_2_cast_fp16)[name = tensor<string, []>("max_value_2_cast_fp16")];
            tensor<fp16, [1, 8, 43, 80]> var_1323_cast_fp16 = sub(x = const_2_to_fp16, y = y_2_cast_fp16)[name = tensor<string, []>("op_1323_cast_fp16")];
            tensor<fp16, [1, 8, 43, 80]> diff_2_cast_fp16 = abs(x = var_1323_cast_fp16)[name = tensor<string, []>("diff_2_cast_fp16")];
            tensor<fp16, []> const_3_promoted_to_fp16 = const()[name = tensor<string, []>("const_3_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 8, 43, 80]> var_1325_cast_fp16 = mul(x = diff_2_cast_fp16, y = const_3_promoted_to_fp16)[name = tensor<string, []>("op_1325_cast_fp16")];
            tensor<fp16, [1, 8, 43, 80]> var_1326_cast_fp16 = exp(x = var_1325_cast_fp16)[name = tensor<string, []>("op_1326_cast_fp16")];
            tensor<fp32, []> var_1327_epsilon_0 = const()[name = tensor<string, []>("op_1327_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 8, 43, 80]> var_1327_cast_fp16 = log(epsilon = var_1327_epsilon_0, x = var_1326_cast_fp16)[name = tensor<string, []>("op_1327_cast_fp16")];
            tensor<fp16, [1, 8, 43, 80]> var_1328_cast_fp16 = add(x = max_value_2_cast_fp16, y = var_1327_cast_fp16)[name = tensor<string, []>("op_1328_cast_fp16")];
            tensor<fp16, []> var_1329_to_fp16 = const()[name = tensor<string, []>("op_1329_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 8, 43, 80]> var_1330_cast_fp16 = mul(x = x_25_cast_fp16, y = var_1329_to_fp16)[name = tensor<string, []>("op_1330_cast_fp16")];
            tensor<fp16, [1, 8, 43, 80]> var_1331_cast_fp16 = sub(x = var_1328_cast_fp16, y = var_1330_cast_fp16)[name = tensor<string, []>("op_1331_cast_fp16")];
            tensor<fp16, []> var_1332_to_fp16 = const()[name = tensor<string, []>("op_1332_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 8, 43, 80]> input_4_cast_fp16 = sub(x = var_1331_cast_fp16, y = var_1332_to_fp16)[name = tensor<string, []>("input_4_cast_fp16")];
            tensor<string, []> x_6_pad_type_0 = const()[name = tensor<string, []>("x_6_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_6_strides_0 = const()[name = tensor<string, []>("x_6_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [4]> x_6_pad_0 = const()[name = tensor<string, []>("x_6_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_6_dilations_0 = const()[name = tensor<string, []>("x_6_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_6_groups_0 = const()[name = tensor<string, []>("x_6_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [32, 8, 3, 3]> encoder_embed_conv_4_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [2304]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2688))), name = tensor<string, []>("encoder_embed_conv_4_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([32, 8, 3, 3])];
            tensor<fp16, [32]> encoder_embed_conv_4_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_conv_4_bias_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3264)))];
            tensor<fp16, [1, 32, 21, 39]> x_6_cast_fp16 = conv(bias = encoder_embed_conv_4_bias_to_fp16, dilations = x_6_dilations_0, groups = x_6_groups_0, pad = x_6_pad_0, pad_type = x_6_pad_type_0, strides = x_6_strides_0, weight = encoder_embed_conv_4_weight_to_fp16_palettized, x = input_4_cast_fp16)[name = tensor<string, []>("x_6_cast_fp16")];
            tensor<fp16, []> var_1343_to_fp16 = const()[name = tensor<string, []>("op_1343_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 32, 21, 39]> y_4_cast_fp16 = sub(x = x_6_cast_fp16, y = var_1343_to_fp16)[name = tensor<string, []>("y_4_cast_fp16")];
            tensor<fp16, []> const_4_to_fp16 = const()[name = tensor<string, []>("const_4_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 32, 21, 39]> max_value_4_cast_fp16 = maximum(x = const_4_to_fp16, y = y_4_cast_fp16)[name = tensor<string, []>("max_value_4_cast_fp16")];
            tensor<fp16, [1, 32, 21, 39]> var_1346_cast_fp16 = sub(x = const_4_to_fp16, y = y_4_cast_fp16)[name = tensor<string, []>("op_1346_cast_fp16")];
            tensor<fp16, [1, 32, 21, 39]> diff_4_cast_fp16 = abs(x = var_1346_cast_fp16)[name = tensor<string, []>("diff_4_cast_fp16")];
            tensor<fp16, []> const_5_promoted_to_fp16 = const()[name = tensor<string, []>("const_5_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 32, 21, 39]> var_1348_cast_fp16 = mul(x = diff_4_cast_fp16, y = const_5_promoted_to_fp16)[name = tensor<string, []>("op_1348_cast_fp16")];
            tensor<fp16, [1, 32, 21, 39]> var_1349_cast_fp16 = exp(x = var_1348_cast_fp16)[name = tensor<string, []>("op_1349_cast_fp16")];
            tensor<fp32, []> var_1350_epsilon_0 = const()[name = tensor<string, []>("op_1350_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 32, 21, 39]> var_1350_cast_fp16 = log(epsilon = var_1350_epsilon_0, x = var_1349_cast_fp16)[name = tensor<string, []>("op_1350_cast_fp16")];
            tensor<fp16, [1, 32, 21, 39]> var_1351_cast_fp16 = add(x = max_value_4_cast_fp16, y = var_1350_cast_fp16)[name = tensor<string, []>("op_1351_cast_fp16")];
            tensor<fp16, []> var_1352_to_fp16 = const()[name = tensor<string, []>("op_1352_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 32, 21, 39]> var_1353_cast_fp16 = mul(x = x_6_cast_fp16, y = var_1352_to_fp16)[name = tensor<string, []>("op_1353_cast_fp16")];
            tensor<fp16, [1, 32, 21, 39]> var_1354_cast_fp16 = sub(x = var_1351_cast_fp16, y = var_1353_cast_fp16)[name = tensor<string, []>("op_1354_cast_fp16")];
            tensor<fp16, []> var_1355_to_fp16 = const()[name = tensor<string, []>("op_1355_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 32, 21, 39]> input_5_cast_fp16 = sub(x = var_1354_cast_fp16, y = var_1355_to_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
            tensor<string, []> x_10_pad_type_0 = const()[name = tensor<string, []>("x_10_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_10_strides_0 = const()[name = tensor<string, []>("x_10_strides_0"), val = tensor<int32, [2]>([1, 2])];
            tensor<int32, [4]> x_10_pad_0 = const()[name = tensor<string, []>("x_10_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_10_dilations_0 = const()[name = tensor<string, []>("x_10_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_10_groups_0 = const()[name = tensor<string, []>("x_10_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [128, 32, 3, 3]> encoder_embed_conv_7_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [36864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3392))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40320))), name = tensor<string, []>("encoder_embed_conv_7_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([128, 32, 3, 3])];
            tensor<fp16, [128]> encoder_embed_conv_7_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_conv_7_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40896)))];
            tensor<fp16, [1, 128, 19, 19]> x_10_cast_fp16 = conv(bias = encoder_embed_conv_7_bias_to_fp16, dilations = x_10_dilations_0, groups = x_10_groups_0, pad = x_10_pad_0, pad_type = x_10_pad_type_0, strides = x_10_strides_0, weight = encoder_embed_conv_7_weight_to_fp16_palettized, x = input_5_cast_fp16)[name = tensor<string, []>("x_10_cast_fp16")];
            tensor<fp16, []> var_1366_to_fp16 = const()[name = tensor<string, []>("op_1366_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [1, 128, 19, 19]> y_5_cast_fp16 = sub(x = x_10_cast_fp16, y = var_1366_to_fp16)[name = tensor<string, []>("y_5_cast_fp16")];
            tensor<fp16, []> const_6_to_fp16 = const()[name = tensor<string, []>("const_6_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 128, 19, 19]> max_value_3_cast_fp16 = maximum(x = const_6_to_fp16, y = y_5_cast_fp16)[name = tensor<string, []>("max_value_3_cast_fp16")];
            tensor<fp16, [1, 128, 19, 19]> var_1369_cast_fp16 = sub(x = const_6_to_fp16, y = y_5_cast_fp16)[name = tensor<string, []>("op_1369_cast_fp16")];
            tensor<fp16, [1, 128, 19, 19]> diff_3_cast_fp16 = abs(x = var_1369_cast_fp16)[name = tensor<string, []>("diff_3_cast_fp16")];
            tensor<fp16, []> const_7_promoted_to_fp16 = const()[name = tensor<string, []>("const_7_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 128, 19, 19]> var_1371_cast_fp16 = mul(x = diff_3_cast_fp16, y = const_7_promoted_to_fp16)[name = tensor<string, []>("op_1371_cast_fp16")];
            tensor<fp16, [1, 128, 19, 19]> var_1372_cast_fp16 = exp(x = var_1371_cast_fp16)[name = tensor<string, []>("op_1372_cast_fp16")];
            tensor<fp32, []> var_1373_epsilon_0 = const()[name = tensor<string, []>("op_1373_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 128, 19, 19]> var_1373_cast_fp16 = log(epsilon = var_1373_epsilon_0, x = var_1372_cast_fp16)[name = tensor<string, []>("op_1373_cast_fp16")];
            tensor<fp16, [1, 128, 19, 19]> var_1374_cast_fp16 = add(x = max_value_3_cast_fp16, y = var_1373_cast_fp16)[name = tensor<string, []>("op_1374_cast_fp16")];
            tensor<fp16, []> var_1375_to_fp16 = const()[name = tensor<string, []>("op_1375_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 128, 19, 19]> var_1376_cast_fp16 = mul(x = x_10_cast_fp16, y = var_1375_to_fp16)[name = tensor<string, []>("op_1376_cast_fp16")];
            tensor<fp16, [1, 128, 19, 19]> var_1377_cast_fp16 = sub(x = var_1374_cast_fp16, y = var_1376_cast_fp16)[name = tensor<string, []>("op_1377_cast_fp16")];
            tensor<fp16, []> var_1378_to_fp16 = const()[name = tensor<string, []>("op_1378_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [1, 128, 19, 19]> x0_7_cast_fp16 = sub(x = var_1377_cast_fp16, y = var_1378_to_fp16)[name = tensor<string, []>("x0_7_cast_fp16")];
            tensor<int32, [4]> var_1388_begin_0 = const()[name = tensor<string, []>("op_1388_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1388_end_0 = const()[name = tensor<string, []>("op_1388_end_0"), val = tensor<int32, [4]>([1, 128, 16, 19])];
            tensor<bool, [4]> var_1388_end_mask_0 = const()[name = tensor<string, []>("op_1388_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 128, 16, 19]> var_1388_cast_fp16 = slice_by_index(begin = var_1388_begin_0, end = var_1388_end_0, end_mask = var_1388_end_mask_0, x = x0_7_cast_fp16)[name = tensor<string, []>("op_1388_cast_fp16")];
            tensor<bool, []> x0_5_interleave_0 = const()[name = tensor<string, []>("x0_5_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cached_embed_left_pad_to_fp16_dtype_0 = const()[name = tensor<string, []>("cached_embed_left_pad_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 3, 19]> cached_embed_left_pad_to_fp16 = cast(dtype = cached_embed_left_pad_to_fp16_dtype_0, x = cached_embed_left_pad)[name = tensor<string, []>("cast_37")];
            tensor<fp16, [1, 128, 22, 19]> x0_5_cast_fp16 = concat(axis = var_49, interleave = x0_5_interleave_0, values = (cached_embed_left_pad_to_fp16, x0_7_cast_fp16))[name = tensor<string, []>("x0_5_cast_fp16")];
            tensor<int32, [4]> var_1397_begin_0 = const()[name = tensor<string, []>("op_1397_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1397_end_0 = const()[name = tensor<string, []>("op_1397_end_0"), val = tensor<int32, [4]>([1, 128, 19, 19])];
            tensor<bool, [4]> var_1397_end_mask_0 = const()[name = tensor<string, []>("op_1397_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 128, 3, 19]> new_cached_embed_left_pad = slice_by_index(begin = var_1397_begin_0, end = var_1397_end_0, end_mask = var_1397_end_mask_0, x = x0_5_cast_fp16)[name = tensor<string, []>("op_1397_cast_fp16")];
            tensor<string, []> input0_3_pad_type_0 = const()[name = tensor<string, []>("input0_3_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [4]> input0_3_pad_0 = const()[name = tensor<string, []>("input0_3_pad_0"), val = tensor<int32, [4]>([0, 0, 3, 3])];
            tensor<int32, []> input0_3_groups_0 = const()[name = tensor<string, []>("input0_3_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [2]> input0_3_strides_0 = const()[name = tensor<string, []>("input0_3_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input0_3_dilations_0 = const()[name = tensor<string, []>("input0_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [128, 1, 7, 7]> encoder_embed_convnext_depthwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6272]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41216))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(47552))), name = tensor<string, []>("encoder_embed_convnext_depthwise_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([128, 1, 7, 7])];
            tensor<fp16, [128]> encoder_embed_convnext_depthwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_convnext_depthwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48128)))];
            tensor<fp16, [1, 128, 16, 19]> input0_3_cast_fp16 = conv(bias = encoder_embed_convnext_depthwise_conv_bias_to_fp16, dilations = input0_3_dilations_0, groups = input0_3_groups_0, pad = input0_3_pad_0, pad_type = input0_3_pad_type_0, strides = input0_3_strides_0, weight = encoder_embed_convnext_depthwise_conv_weight_to_fp16_palettized, x = x0_5_cast_fp16)[name = tensor<string, []>("input0_3_cast_fp16")];
            tensor<string, []> x_9_pad_type_0 = const()[name = tensor<string, []>("x_9_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_9_strides_0 = const()[name = tensor<string, []>("x_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_9_dilations_0 = const()[name = tensor<string, []>("x_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_9_groups_0 = const()[name = tensor<string, []>("x_9_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [384, 128, 1, 1]> encoder_embed_convnext_pointwise_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [49152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48448))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97664))), name = tensor<string, []>("encoder_embed_convnext_pointwise_conv1_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([384, 128, 1, 1])];
            tensor<fp16, [384]> encoder_embed_convnext_pointwise_conv1_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_convnext_pointwise_conv1_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98240)))];
            tensor<fp16, [1, 384, 16, 19]> x_9_cast_fp16 = conv(bias = encoder_embed_convnext_pointwise_conv1_bias_to_fp16, dilations = x_9_dilations_0, groups = x_9_groups_0, pad = x_9_pad_0, pad_type = x_9_pad_type_0, strides = x_9_strides_0, weight = encoder_embed_convnext_pointwise_conv1_weight_to_fp16_palettized, x = input0_3_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
            tensor<fp16, []> var_1423_to_fp16 = const()[name = tensor<string, []>("op_1423_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [1, 384, 16, 19]> y_1_cast_fp16 = sub(x = x_9_cast_fp16, y = var_1423_to_fp16)[name = tensor<string, []>("y_1_cast_fp16")];
            tensor<fp16, []> const_9_to_fp16 = const()[name = tensor<string, []>("const_9_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 384, 16, 19]> max_value_1_cast_fp16 = maximum(x = const_9_to_fp16, y = y_1_cast_fp16)[name = tensor<string, []>("max_value_1_cast_fp16")];
            tensor<fp16, [1, 384, 16, 19]> var_1426_cast_fp16 = sub(x = const_9_to_fp16, y = y_1_cast_fp16)[name = tensor<string, []>("op_1426_cast_fp16")];
            tensor<fp16, [1, 384, 16, 19]> diff_1_cast_fp16 = abs(x = var_1426_cast_fp16)[name = tensor<string, []>("diff_1_cast_fp16")];
            tensor<fp16, []> const_10_promoted_to_fp16 = const()[name = tensor<string, []>("const_10_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 384, 16, 19]> var_1428_cast_fp16 = mul(x = diff_1_cast_fp16, y = const_10_promoted_to_fp16)[name = tensor<string, []>("op_1428_cast_fp16")];
            tensor<fp16, [1, 384, 16, 19]> var_1429_cast_fp16 = exp(x = var_1428_cast_fp16)[name = tensor<string, []>("op_1429_cast_fp16")];
            tensor<fp32, []> var_1430_epsilon_0 = const()[name = tensor<string, []>("op_1430_epsilon_0"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp16, [1, 384, 16, 19]> var_1430_cast_fp16 = log(epsilon = var_1430_epsilon_0, x = var_1429_cast_fp16)[name = tensor<string, []>("op_1430_cast_fp16")];
            tensor<fp16, [1, 384, 16, 19]> var_1431_cast_fp16 = add(x = max_value_1_cast_fp16, y = var_1430_cast_fp16)[name = tensor<string, []>("op_1431_cast_fp16")];
            tensor<fp16, []> var_1432_to_fp16 = const()[name = tensor<string, []>("op_1432_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [1, 384, 16, 19]> var_1433_cast_fp16 = mul(x = x_9_cast_fp16, y = var_1432_to_fp16)[name = tensor<string, []>("op_1433_cast_fp16")];
            tensor<fp16, [1, 384, 16, 19]> var_1434_cast_fp16 = sub(x = var_1431_cast_fp16, y = var_1433_cast_fp16)[name = tensor<string, []>("op_1434_cast_fp16")];
            tensor<fp16, []> var_1435_to_fp16 = const()[name = tensor<string, []>("op_1435_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [1, 384, 16, 19]> input_7_cast_fp16 = sub(x = var_1434_cast_fp16, y = var_1435_to_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<string, []> x_13_pad_type_0 = const()[name = tensor<string, []>("x_13_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, [2]> x_13_strides_0 = const()[name = tensor<string, []>("x_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_13_pad_0 = const()[name = tensor<string, []>("x_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_13_dilations_0 = const()[name = tensor<string, []>("x_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, []> x_13_groups_0 = const()[name = tensor<string, []>("x_13_groups_0"), val = tensor<int32, []>(1)];
            tensor<fp16, [128, 384, 1, 1]> encoder_embed_convnext_pointwise_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [49152]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99072))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148288))), name = tensor<string, []>("encoder_embed_convnext_pointwise_conv2_weight_to_fp16_palettized"), shape = tensor<uint32, [4]>([128, 384, 1, 1])];
            tensor<fp16, [128]> encoder_embed_convnext_pointwise_conv2_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_convnext_pointwise_conv2_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(148864)))];
            tensor<fp16, [1, 128, 16, 19]> x_13_cast_fp16 = conv(bias = encoder_embed_convnext_pointwise_conv2_bias_to_fp16, dilations = x_13_dilations_0, groups = x_13_groups_0, pad = x_13_pad_0, pad_type = x_13_pad_type_0, strides = x_13_strides_0, weight = encoder_embed_convnext_pointwise_conv2_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor<string, []>("x_13_cast_fp16")];
            tensor<fp16, [1, 128, 16, 19]> x1_1_cast_fp16 = add(x = var_1388_cast_fp16, y = x_13_cast_fp16)[name = tensor<string, []>("x1_1_cast_fp16")];
            tensor<int32, [4]> var_1459_perm_0 = const()[name = tensor<string, []>("op_1459_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1462 = const()[name = tensor<string, []>("op_1462"), val = tensor<int32, [3]>([1, 16, 2432])];
            tensor<fp16, [1, 16, 128, 19]> var_1459_cast_fp16 = transpose(perm = var_1459_perm_0, x = x1_1_cast_fp16)[name = tensor<string, []>("transpose_116")];
            tensor<fp16, [1, 16, 2432]> input1_1_cast_fp16 = reshape(shape = var_1462, x = var_1459_cast_fp16)[name = tensor<string, []>("input1_1_cast_fp16")];
            tensor<fp16, [128, 2432]> encoder_embed_out_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [311296]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149184))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460544))), name = tensor<string, []>("encoder_embed_out_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 2432])];
            tensor<fp16, [128]> encoder_embed_out_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_out_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461120)))];
            tensor<fp16, [1, 16, 128]> linear_0_cast_fp16 = linear(bias = encoder_embed_out_bias_to_fp16, weight = encoder_embed_out_weight_to_fp16_palettized, x = input1_1_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<fp16, [128]> encoder_embed_out_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_embed_out_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461440)))];
            tensor<fp16, [1, 16, 128]> var_1474_cast_fp16 = sub(x = linear_0_cast_fp16, y = encoder_embed_out_norm_bias_to_fp16)[name = tensor<string, []>("op_1474_cast_fp16")];
            tensor<fp16, []> var_1468_promoted_to_fp16 = const()[name = tensor<string, []>("op_1468_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [1, 16, 128]> var_1475_cast_fp16 = pow(x = var_1474_cast_fp16, y = var_1468_promoted_to_fp16)[name = tensor<string, []>("op_1475_cast_fp16")];
            tensor<int32, [1]> var_1477_axes_0 = const()[name = tensor<string, []>("op_1477_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, []> var_1477_keep_dims_0 = const()[name = tensor<string, []>("op_1477_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 16, 1]> var_1477_cast_fp16 = reduce_mean(axes = var_1477_axes_0, keep_dims = var_1477_keep_dims_0, x = var_1475_cast_fp16)[name = tensor<string, []>("op_1477_cast_fp16")];
            tensor<fp16, []> var_1467_to_fp16 = const()[name = tensor<string, []>("op_1467_to_fp16"), val = tensor<fp16, []>(-0x1p-1)];
            tensor<fp16, [1, 16, 1]> var_1478_cast_fp16 = pow(x = var_1477_cast_fp16, y = var_1467_to_fp16)[name = tensor<string, []>("op_1478_cast_fp16")];
            tensor<fp16, []> var_1479_to_fp16 = const()[name = tensor<string, []>("op_1479_to_fp16"), val = tensor<fp16, []>(0x1.c94p-3)];
            tensor<fp16, [1, 16, 1]> scales_2_cast_fp16 = mul(x = var_1478_cast_fp16, y = var_1479_to_fp16)[name = tensor<string, []>("scales_2_cast_fp16")];
            tensor<fp16, [1, 16, 128]> var_1481_cast_fp16 = mul(x = linear_0_cast_fp16, y = scales_2_cast_fp16)[name = tensor<string, []>("op_1481_cast_fp16")];
            tensor<bool, [1, 16]> src_key_padding_mask_1 = const()[name = tensor<string, []>("src_key_padding_mask_1"), val = tensor<bool, [1, 16]>([[false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false]])];
            tensor<int32, [1, 64]> expand_dims_0 = const()[name = tensor<string, []>("expand_dims_0"), val = tensor<int32, [1, 64]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63]])];
            tensor<int32, [1]> var_1492_axes_0 = const()[name = tensor<string, []>("op_1492_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1]> var_1492 = expand_dims(axes = var_1492_axes_0, x = processed_lens)[name = tensor<string, []>("op_1492")];
            tensor<bool, [1, 64]> var_1493 = less_equal(x = var_1492, y = expand_dims_0)[name = tensor<string, []>("op_1493")];
            tensor<int32, [1]> var_1494 = const()[name = tensor<string, []>("op_1494"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 64]> processed_mask0_1 = reverse(axes = var_1494, x = var_1493)[name = tensor<string, []>("processed_mask0_1")];
            tensor<string, []> processed_lens_promoted_to_fp16_dtype_0 = const()[name = tensor<string, []>("processed_lens_promoted_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1]> x_lens0_1_to_fp16 = const()[name = tensor<string, []>("x_lens0_1_to_fp16"), val = tensor<fp16, [1]>([0x1p+4])];
            tensor<fp16, [1]> processed_lens_to_fp16 = cast(dtype = processed_lens_promoted_to_fp16_dtype_0, x = processed_lens)[name = tensor<string, []>("cast_36")];
            tensor<fp16, [1]> new_processed_lens = add(x = processed_lens_to_fp16, y = x_lens0_1_to_fp16)[name = tensor<string, []>("op_1496_cast_fp16")];
            tensor<bool, []> src_key_padding_mask0_1_interleave_0 = const()[name = tensor<string, []>("src_key_padding_mask0_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<bool, [1, 80]> src_key_padding_mask0_1 = concat(axis = var_48, interleave = src_key_padding_mask0_1_interleave_0, values = (processed_mask0_1, src_key_padding_mask_1))[name = tensor<string, []>("src_key_padding_mask0_1")];
            tensor<int32, [3]> var_1499 = const()[name = tensor<string, []>("op_1499"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<fp16, [272, 128]> encoder_encoders_0_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [34816]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461760))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(496640))), name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([272, 128])];
            tensor<fp16, [272]> encoder_encoders_0_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [272]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497216)))];
            tensor<fp16, [16, 1, 128]> x2_1_cast_fp16 = transpose(perm = var_1499, x = var_1481_cast_fp16)[name = tensor<string, []>("transpose_115")];
            tensor<fp16, [16, 1, 272]> linear_1_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_self_attn_weights_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized, x = x2_1_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<int32, [3]> q_1_begin_0 = const()[name = tensor<string, []>("q_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> q_1_end_0 = const()[name = tensor<string, []>("q_1_end_0"), val = tensor<int32, [3]>([16, 1, 128])];
            tensor<bool, [3]> q_1_end_mask_0 = const()[name = tensor<string, []>("q_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [16, 1, 128]> q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, x = linear_1_cast_fp16)[name = tensor<string, []>("q_1_cast_fp16")];
            tensor<int32, [3]> k_1_begin_0 = const()[name = tensor<string, []>("k_1_begin_0"), val = tensor<int32, [3]>([0, 0, 128])];
            tensor<int32, [3]> k_1_end_0 = const()[name = tensor<string, []>("k_1_end_0"), val = tensor<int32, [3]>([16, 1, 256])];
            tensor<bool, [3]> k_1_end_mask_0 = const()[name = tensor<string, []>("k_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [16, 1, 128]> k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, x = linear_1_cast_fp16)[name = tensor<string, []>("k_1_cast_fp16")];
            tensor<int32, [3]> p_1_begin_0 = const()[name = tensor<string, []>("p_1_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> p_1_end_0 = const()[name = tensor<string, []>("p_1_end_0"), val = tensor<int32, [3]>([16, 1, 272])];
            tensor<bool, [3]> p_1_end_mask_0 = const()[name = tensor<string, []>("p_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [16, 1, 16]> p_1_cast_fp16 = slice_by_index(begin = p_1_begin_0, end = p_1_end_0, end_mask = p_1_end_mask_0, x = linear_1_cast_fp16)[name = tensor<string, []>("p_1_cast_fp16")];
            tensor<bool, []> k0_1_interleave_0 = const()[name = tensor<string, []>("k0_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s0_l0_key_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s0_l0_key_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [64, 1, 128]> cache_s0_l0_key_to_fp16 = cast(dtype = cache_s0_l0_key_to_fp16_dtype_0, x = cache_s0_l0_key)[name = tensor<string, []>("cast_35")];
            tensor<fp16, [80, 1, 128]> k0_1_cast_fp16 = concat(axis = var_46, interleave = k0_1_interleave_0, values = (cache_s0_l0_key_to_fp16, k_1_cast_fp16))[name = tensor<string, []>("k0_1_cast_fp16")];
            tensor<int32, [3]> var_1553_begin_0 = const()[name = tensor<string, []>("op_1553_begin_0"), val = tensor<int32, [3]>([16, 0, 0])];
            tensor<int32, [3]> var_1553_end_0 = const()[name = tensor<string, []>("op_1553_end_0"), val = tensor<int32, [3]>([80, 1, 128])];
            tensor<bool, [3]> var_1553_end_mask_0 = const()[name = tensor<string, []>("op_1553_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [64, 1, 128]> new_cache_s0_l0_key = slice_by_index(begin = var_1553_begin_0, end = var_1553_end_0, end_mask = var_1553_end_mask_0, x = k0_1_cast_fp16)[name = tensor<string, []>("op_1553_cast_fp16")];
            tensor<int32, [4]> var_1559 = const()[name = tensor<string, []>("op_1559"), val = tensor<int32, [4]>([16, 1, 4, 32])];
            tensor<fp16, [16, 1, 4, 32]> q0_1_cast_fp16 = reshape(shape = var_1559, x = q_1_cast_fp16)[name = tensor<string, []>("q0_1_cast_fp16")];
            tensor<int32, [4]> var_1561 = const()[name = tensor<string, []>("op_1561"), val = tensor<int32, [4]>([16, 1, 4, 4])];
            tensor<fp16, [16, 1, 4, 4]> p0_1_cast_fp16 = reshape(shape = var_1561, x = p_1_cast_fp16)[name = tensor<string, []>("p0_1_cast_fp16")];
            tensor<int32, [4]> var_1563 = const()[name = tensor<string, []>("op_1563"), val = tensor<int32, [4]>([80, 1, 4, 32])];
            tensor<fp16, [80, 1, 4, 32]> k1_1_cast_fp16 = reshape(shape = var_1563, x = k0_1_cast_fp16)[name = tensor<string, []>("k1_1_cast_fp16")];
            tensor<int32, [4]> var_1567 = const()[name = tensor<string, []>("op_1567"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> attn_scores_1_transpose_x_0 = const()[name = tensor<string, []>("attn_scores_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_scores_1_transpose_y_0 = const()[name = tensor<string, []>("attn_scores_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = tensor<string, []>("transpose_24_perm_0"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<int32, [4]> transpose_25_perm_0 = const()[name = tensor<string, []>("transpose_25_perm_0"), val = tensor<int32, [4]>([2, 1, -1, -4])];
            tensor<fp16, [4, 1, 32, 80]> transpose_25 = transpose(perm = transpose_25_perm_0, x = k1_1_cast_fp16)[name = tensor<string, []>("transpose_113")];
            tensor<fp16, [4, 1, 16, 32]> transpose_24 = transpose(perm = transpose_24_perm_0, x = q0_1_cast_fp16)[name = tensor<string, []>("transpose_114")];
            tensor<fp16, [4, 1, 16, 80]> attn_scores_1_cast_fp16 = matmul(transpose_x = attn_scores_1_transpose_x_0, transpose_y = attn_scores_1_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor<string, []>("attn_scores_1_cast_fp16")];
            tensor<bool, []> pos_scores_1_transpose_x_0 = const()[name = tensor<string, []>("pos_scores_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_scores_1_transpose_y_0 = const()[name = tensor<string, []>("pos_scores_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 4, 95]> pos_emb_3_to_fp16 = const()[name = tensor<string, []>("pos_emb_3_to_fp16"), val = tensor<fp16, [4, 1, 4, 95]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(497856)))];
            tensor<fp16, [4, 1, 16, 4]> p1_1_cast_fp16 = transpose(perm = var_1567, x = p0_1_cast_fp16)[name = tensor<string, []>("transpose_112")];
            tensor<fp16, [4, 1, 16, 95]> pos_scores_1_cast_fp16 = matmul(transpose_x = pos_scores_1_transpose_x_0, transpose_y = pos_scores_1_transpose_y_0, x = p1_1_cast_fp16, y = pos_emb_3_to_fp16)[name = tensor<string, []>("pos_scores_1_cast_fp16")];
            tensor<int32, [2]> var_1610 = const()[name = tensor<string, []>("op_1610"), val = tensor<int32, [2]>([-1, 95])];
            tensor<fp16, [64, 95]> pos_scores0_1_cast_fp16 = reshape(shape = var_1610, x = pos_scores_1_cast_fp16)[name = tensor<string, []>("pos_scores0_1_cast_fp16")];
            tensor<bool, []> pos_scores1_1_validate_indices_0 = const()[name = tensor<string, []>("pos_scores1_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<uint16, [64, 80]> indexes_1_to_uint16 = const()[name = tensor<string, []>("indexes_1_to_uint16"), val = tensor<uint16, [64, 80]>([[15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94], [14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93], [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92], [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91], [11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90], [10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89], [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88], [8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87], [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79], [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94], [14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93], [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92], [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91], [11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90], [10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89], [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88], [8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87], [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79], [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94], [14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93], [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92], [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91], [11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90], [10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89], [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88], [8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87], [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79], [15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94], [14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93], [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92], [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91], [11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90], [10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89], [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88], [8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87], [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79]])];
            tensor<fp16, [64, 80]> pos_scores1_1_cast_fp16_cast_uint16 = gather_along_axis(axis = var_48, indices = indexes_1_to_uint16, validate_indices = pos_scores1_1_validate_indices_0, x = pos_scores0_1_cast_fp16)[name = tensor<string, []>("pos_scores1_1_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_1613 = const()[name = tensor<string, []>("op_1613"), val = tensor<int32, [4]>([4, 1, 16, 80])];
            tensor<fp16, [4, 1, 16, 80]> pos_scores2_1_cast_fp16 = reshape(shape = var_1613, x = pos_scores1_1_cast_fp16_cast_uint16)[name = tensor<string, []>("pos_scores2_1_cast_fp16")];
            tensor<fp16, [4, 1, 16, 80]> attn_scores0_1_cast_fp16 = add(x = attn_scores_1_cast_fp16, y = pos_scores2_1_cast_fp16)[name = tensor<string, []>("attn_scores0_1_cast_fp16")];
            tensor<int32, [1]> var_1616_axes_0 = const()[name = tensor<string, []>("op_1616_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 80]> var_1616 = expand_dims(axes = var_1616_axes_0, x = src_key_padding_mask0_1)[name = tensor<string, []>("op_1616")];
            tensor<fp16, []> cast_7_to_fp16 = const()[name = tensor<string, []>("cast_7_to_fp16"), val = tensor<fp16, []>(-0x1.f4p+9)];
            tensor<fp16, [4, 1, 16, 80]> attn_scores1_1_cast_fp16 = select(a = cast_7_to_fp16, b = attn_scores0_1_cast_fp16, cond = var_1616)[name = tensor<string, []>("attn_scores1_1_cast_fp16")];
            tensor<fp16, [4, 1, 16, 80]> attn_weights_1_cast_fp16 = softmax(axis = var_59, x = attn_scores1_1_cast_fp16)[name = tensor<string, []>("attn_weights_1_cast_fp16")];
            tensor<fp16, [144, 128]> encoder_encoders_0_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500992))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(519488))), name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([144, 128])];
            tensor<fp16, [144]> encoder_encoders_0_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520064)))];
            tensor<fp16, [16, 1, 144]> linear_3_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_feed_forward1_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized, x = x2_1_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, []> var_1631_to_fp16 = const()[name = tensor<string, []>("op_1631_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [16, 1, 144]> x_offset_2_cast_fp16 = sub(x = linear_3_cast_fp16, y = var_1631_to_fp16)[name = tensor<string, []>("x_offset_2_cast_fp16")];
            tensor<fp16, [16, 1, 144]> var_1633_cast_fp16 = exp(x = x_offset_2_cast_fp16)[name = tensor<string, []>("op_1633_cast_fp16")];
            tensor<fp16, []> var_1634_to_fp16 = const()[name = tensor<string, []>("op_1634_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [16, 1, 144]> var_1635_cast_fp16 = add(x = var_1633_cast_fp16, y = var_1634_to_fp16)[name = tensor<string, []>("op_1635_cast_fp16")];
            tensor<fp32, []> var_1636_epsilon_0 = const()[name = tensor<string, []>("op_1636_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [16, 1, 144]> var_1636_cast_fp16 = log(epsilon = var_1636_epsilon_0, x = var_1635_cast_fp16)[name = tensor<string, []>("op_1636_cast_fp16")];
            tensor<fp16, []> var_1619_to_fp16 = const()[name = tensor<string, []>("op_1619_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [16, 1, 144]> var_1638_cast_fp16 = equal(x = var_1636_cast_fp16, y = var_1619_to_fp16)[name = tensor<string, []>("op_1638_cast_fp16")];
            tensor<fp16, [16, 1, 144]> log_sum0_2_cast_fp16 = select(a = x_offset_2_cast_fp16, b = var_1636_cast_fp16, cond = var_1638_cast_fp16)[name = tensor<string, []>("log_sum0_2_cast_fp16")];
            tensor<fp16, []> var_1640_to_fp16 = const()[name = tensor<string, []>("op_1640_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [16, 1, 144]> var_1641_cast_fp16 = mul(x = linear_3_cast_fp16, y = var_1640_to_fp16)[name = tensor<string, []>("op_1641_cast_fp16")];
            tensor<fp16, [16, 1, 144]> var_1642_cast_fp16 = sub(x = log_sum0_2_cast_fp16, y = var_1641_cast_fp16)[name = tensor<string, []>("op_1642_cast_fp16")];
            tensor<fp16, []> var_1643_to_fp16 = const()[name = tensor<string, []>("op_1643_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [16, 1, 144]> x_15_cast_fp16 = sub(x = var_1642_cast_fp16, y = var_1643_to_fp16)[name = tensor<string, []>("x_15_cast_fp16")];
            tensor<fp16, [128, 144]> encoder_encoders_0_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(520448))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538944))), name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 144])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539520)))];
            tensor<fp16, [16, 1, 128]> linear_4_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_feed_forward1_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized, x = x_15_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<fp16, [16, 1, 128]> input2_1_cast_fp16 = add(x = x2_1_cast_fp16, y = linear_4_cast_fp16)[name = tensor<string, []>("input2_1_cast_fp16")];
            tensor<int32, [4]> attn_weights0_1_begin_0 = const()[name = tensor<string, []>("attn_weights0_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_weights0_1_end_0 = const()[name = tensor<string, []>("attn_weights0_1_end_0"), val = tensor<int32, [4]>([1, 1, 16, 80])];
            tensor<bool, [4]> attn_weights0_1_end_mask_0 = const()[name = tensor<string, []>("attn_weights0_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1, 16, 80]> attn_weights0_1_cast_fp16 = slice_by_index(begin = attn_weights0_1_begin_0, end = attn_weights0_1_end_0, end_mask = attn_weights0_1_end_mask_0, x = attn_weights_1_cast_fp16)[name = tensor<string, []>("attn_weights0_1_cast_fp16")];
            tensor<fp16, [288, 128]> encoder_encoders_0_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [36864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(539840))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(576768))), name = tensor<string, []>("encoder_encoders_0_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([288, 128])];
            tensor<fp16, [288]> encoder_encoders_0_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577344)))];
            tensor<fp16, [16, 1, 288]> linear_5_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_nonlin_attention_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized, x = input2_1_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<int32, [3]> var_1657_split_sizes_0 = const()[name = tensor<string, []>("op_1657_split_sizes_0"), val = tensor<int32, [3]>([96, 96, 96])];
            tensor<int32, []> var_1657_axis_0 = const()[name = tensor<string, []>("op_1657_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [16, 1, 96]> var_1657_cast_fp16_0, tensor<fp16, [16, 1, 96]> var_1657_cast_fp16_1, tensor<fp16, [16, 1, 96]> var_1657_cast_fp16_2 = split(axis = var_1657_axis_0, split_sizes = var_1657_split_sizes_0, x = linear_5_cast_fp16)[name = tensor<string, []>("op_1657_cast_fp16")];
            tensor<fp16, [16, 1, 96]> var_1660_cast_fp16 = tanh(x = var_1657_cast_fp16_0)[name = tensor<string, []>("op_1660_cast_fp16")];
            tensor<int32, [1]> var_1661_axes_0 = const()[name = tensor<string, []>("op_1661_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [16, 1, 96, 1]> var_1661_cast_fp16 = expand_dims(axes = var_1661_axes_0, x = var_1660_cast_fp16)[name = tensor<string, []>("op_1661_cast_fp16")];
            tensor<int32, [3]> var_1662 = const()[name = tensor<string, []>("op_1662"), val = tensor<int32, [3]>([16, 1, 96])];
            tensor<fp16, [16, 1, 96]> s_1_cast_fp16 = reshape(shape = var_1662, x = var_1661_cast_fp16)[name = tensor<string, []>("s_1_cast_fp16")];
            tensor<fp16, [16, 1, 96]> x5_1_cast_fp16 = mul(x = var_1657_cast_fp16_1, y = s_1_cast_fp16)[name = tensor<string, []>("x5_1_cast_fp16")];
            tensor<int32, [4]> var_1676 = const()[name = tensor<string, []>("op_1676"), val = tensor<int32, [4]>([16, 1, 1, -1])];
            tensor<fp16, [16, 1, 1, 96]> var_1677_cast_fp16 = reshape(shape = var_1676, x = x5_1_cast_fp16)[name = tensor<string, []>("op_1677_cast_fp16")];
            tensor<int32, [4]> var_1678 = const()[name = tensor<string, []>("op_1678"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<bool, []> x_pad_1_interleave_0 = const()[name = tensor<string, []>("x_pad_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s0_l0_nonlin_attn_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s0_l0_nonlin_attn_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 16, 96]> x6_1_cast_fp16 = transpose(perm = var_1678, x = var_1677_cast_fp16)[name = tensor<string, []>("transpose_111")];
            tensor<fp16, [1, 1, 64, 96]> cache_s0_l0_nonlin_attn_to_fp16 = cast(dtype = cache_s0_l0_nonlin_attn_to_fp16_dtype_0, x = cache_s0_l0_nonlin_attn)[name = tensor<string, []>("cast_34")];
            tensor<fp16, [1, 1, 80, 96]> x_pad_1_cast_fp16 = concat(axis = var_49, interleave = x_pad_1_interleave_0, values = (cache_s0_l0_nonlin_attn_to_fp16, x6_1_cast_fp16))[name = tensor<string, []>("x_pad_1_cast_fp16")];
            tensor<int32, [4]> var_1684_begin_0 = const()[name = tensor<string, []>("op_1684_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1684_end_0 = const()[name = tensor<string, []>("op_1684_end_0"), val = tensor<int32, [4]>([1, 1, 80, 96])];
            tensor<bool, [4]> var_1684_end_mask_0 = const()[name = tensor<string, []>("op_1684_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 64, 96]> new_cache_s0_l0_nonlin_attn = slice_by_index(begin = var_1684_begin_0, end = var_1684_end_0, end_mask = var_1684_end_mask_0, x = x_pad_1_cast_fp16)[name = tensor<string, []>("op_1684_cast_fp16")];
            tensor<bool, []> x7_1_transpose_x_0 = const()[name = tensor<string, []>("x7_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x7_1_transpose_y_0 = const()[name = tensor<string, []>("x7_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 16, 96]> x7_1_cast_fp16 = matmul(transpose_x = x7_1_transpose_x_0, transpose_y = x7_1_transpose_y_0, x = attn_weights0_1_cast_fp16, y = x_pad_1_cast_fp16)[name = tensor<string, []>("x7_1_cast_fp16")];
            tensor<int32, [4]> var_1687 = const()[name = tensor<string, []>("op_1687"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_1689 = const()[name = tensor<string, []>("op_1689"), val = tensor<int32, [3]>([16, 1, -1])];
            tensor<fp16, [16, 1, 1, 96]> var_1688_cast_fp16 = transpose(perm = var_1687, x = x7_1_cast_fp16)[name = tensor<string, []>("transpose_110")];
            tensor<fp16, [16, 1, 96]> x8_1_cast_fp16 = reshape(shape = var_1689, x = var_1688_cast_fp16)[name = tensor<string, []>("x8_1_cast_fp16")];
            tensor<fp16, [16, 1, 96]> input4_1_cast_fp16 = mul(x = x8_1_cast_fp16, y = var_1657_cast_fp16_2)[name = tensor<string, []>("input4_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_0_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577984))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590336))), name = tensor<string, []>("encoder_encoders_0_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590912)))];
            tensor<fp16, [16, 1, 128]> linear_6_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_nonlin_attention_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized, x = input4_1_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<fp16, [16, 1, 128]> x9_1_cast_fp16 = add(x = input2_1_cast_fp16, y = linear_6_cast_fp16)[name = tensor<string, []>("x9_1_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_0_layers_0_self_attn1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(591232))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(597440))), name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_0_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(598016)))];
            tensor<fp16, [16, 1, 48]> linear_7_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_self_attn1_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_self_attn1_in_proj_weight_to_fp16_palettized, x = x9_1_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<bool, []> x10_1_interleave_0 = const()[name = tensor<string, []>("x10_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s0_l0_val1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s0_l0_val1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [64, 1, 48]> cache_s0_l0_val1_to_fp16 = cast(dtype = cache_s0_l0_val1_to_fp16_dtype_0, x = cache_s0_l0_val1)[name = tensor<string, []>("cast_33")];
            tensor<fp16, [80, 1, 48]> x10_1_cast_fp16 = concat(axis = var_46, interleave = x10_1_interleave_0, values = (cache_s0_l0_val1_to_fp16, linear_7_cast_fp16))[name = tensor<string, []>("x10_1_cast_fp16")];
            tensor<int32, [3]> var_1714_begin_0 = const()[name = tensor<string, []>("op_1714_begin_0"), val = tensor<int32, [3]>([16, 0, 0])];
            tensor<int32, [3]> var_1714_end_0 = const()[name = tensor<string, []>("op_1714_end_0"), val = tensor<int32, [3]>([80, 1, 48])];
            tensor<bool, [3]> var_1714_end_mask_0 = const()[name = tensor<string, []>("op_1714_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [64, 1, 48]> new_cache_s0_l0_val1 = slice_by_index(begin = var_1714_begin_0, end = var_1714_end_0, end_mask = var_1714_end_mask_0, x = x10_1_cast_fp16)[name = tensor<string, []>("op_1714_cast_fp16")];
            tensor<int32, [4]> var_1715 = const()[name = tensor<string, []>("op_1715"), val = tensor<int32, [4]>([80, 1, 4, -1])];
            tensor<fp16, [80, 1, 4, 12]> var_1716_cast_fp16 = reshape(shape = var_1715, x = x10_1_cast_fp16)[name = tensor<string, []>("op_1716_cast_fp16")];
            tensor<int32, [4]> var_1717 = const()[name = tensor<string, []>("op_1717"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x12_1_transpose_x_0 = const()[name = tensor<string, []>("x12_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x12_1_transpose_y_0 = const()[name = tensor<string, []>("x12_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 80, 12]> x11_1_cast_fp16 = transpose(perm = var_1717, x = var_1716_cast_fp16)[name = tensor<string, []>("transpose_109")];
            tensor<fp16, [4, 1, 16, 12]> x12_1_cast_fp16 = matmul(transpose_x = x12_1_transpose_x_0, transpose_y = x12_1_transpose_y_0, x = attn_weights_1_cast_fp16, y = x11_1_cast_fp16)[name = tensor<string, []>("x12_1_cast_fp16")];
            tensor<int32, [4]> var_1722 = const()[name = tensor<string, []>("op_1722"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_1727 = const()[name = tensor<string, []>("op_1727"), val = tensor<int32, [3]>([16, 1, 48])];
            tensor<fp16, [16, 1, 4, 12]> var_1723_cast_fp16 = transpose(perm = var_1722, x = x12_1_cast_fp16)[name = tensor<string, []>("transpose_108")];
            tensor<fp16, [16, 1, 48]> input5_1_cast_fp16 = reshape(shape = var_1727, x = var_1723_cast_fp16)[name = tensor<string, []>("input5_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_0_layers_0_self_attn1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(598208))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604416))), name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(604992)))];
            tensor<fp16, [16, 1, 128]> linear_8_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_self_attn1_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_self_attn1_out_proj_weight_to_fp16_palettized, x = input5_1_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<fp16, [16, 1, 128]> input6_1_cast_fp16 = add(x = x9_1_cast_fp16, y = linear_8_cast_fp16)[name = tensor<string, []>("input6_1_cast_fp16")];
            tensor<int32, [2]> src_key_padding_mask1_1_begin_0 = const()[name = tensor<string, []>("src_key_padding_mask1_1_begin_0"), val = tensor<int32, [2]>([0, 64])];
            tensor<int32, [2]> src_key_padding_mask1_1_end_0 = const()[name = tensor<string, []>("src_key_padding_mask1_1_end_0"), val = tensor<int32, [2]>([1, 80])];
            tensor<bool, [2]> src_key_padding_mask1_1_end_mask_0 = const()[name = tensor<string, []>("src_key_padding_mask1_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<bool, [1, 16]> src_key_padding_mask1_1 = slice_by_index(begin = src_key_padding_mask1_1_begin_0, end = src_key_padding_mask1_1_end_0, end_mask = src_key_padding_mask1_1_end_mask_0, x = src_key_padding_mask0_1)[name = tensor<string, []>("src_key_padding_mask1_1")];
            tensor<fp16, [256, 128]> encoder_encoders_0_layers_0_conv_module1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(605312))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(638144))), name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_0_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(638720)))];
            tensor<fp16, [16, 1, 256]> linear_9_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_conv_module1_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_conv_module1_in_proj_weight_to_fp16_palettized, x = input6_1_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<int32, [2]> var_1738_split_sizes_0 = const()[name = tensor<string, []>("op_1738_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_1738_axis_0 = const()[name = tensor<string, []>("op_1738_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [16, 1, 128]> var_1738_cast_fp16_0, tensor<fp16, [16, 1, 128]> var_1738_cast_fp16_1 = split(axis = var_1738_axis_0, split_sizes = var_1738_split_sizes_0, x = linear_9_cast_fp16)[name = tensor<string, []>("op_1738_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1740_cast_fp16 = sigmoid(x = var_1738_cast_fp16_1)[name = tensor<string, []>("op_1740_cast_fp16")];
            tensor<fp16, [16, 1, 128]> x14_1_cast_fp16 = mul(x = var_1738_cast_fp16_0, y = var_1740_cast_fp16)[name = tensor<string, []>("x14_1_cast_fp16")];
            tensor<int32, [3]> var_1742 = const()[name = tensor<string, []>("op_1742"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_1744_axes_0 = const()[name = tensor<string, []>("op_1744_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 16]> var_1744 = expand_dims(axes = var_1744_axes_0, x = src_key_padding_mask1_1)[name = tensor<string, []>("op_1744")];
            tensor<int32, [3]> var_1745_reps_0 = const()[name = tensor<string, []>("op_1745_reps_0"), val = tensor<int32, [3]>([1, 128, 1])];
            tensor<bool, [1, 128, 16]> var_1745 = tile(reps = var_1745_reps_0, x = var_1744)[name = tensor<string, []>("op_1745")];
            tensor<fp16, []> var_62_to_fp16 = const()[name = tensor<string, []>("op_62_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
            tensor<fp16, [1, 128, 16]> x15_1_cast_fp16 = transpose(perm = var_1742, x = x14_1_cast_fp16)[name = tensor<string, []>("transpose_107")];
            tensor<fp16, [1, 128, 16]> x16_1_cast_fp16 = select(a = var_62_to_fp16, b = x15_1_cast_fp16, cond = var_1745)[name = tensor<string, []>("x16_1_cast_fp16")];
            tensor<bool, []> x17_1_interleave_0 = const()[name = tensor<string, []>("x17_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s0_l0_conv1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s0_l0_conv1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 15]> cache_s0_l0_conv1_to_fp16 = cast(dtype = cache_s0_l0_conv1_to_fp16_dtype_0, x = cache_s0_l0_conv1)[name = tensor<string, []>("cast_32")];
            tensor<fp16, [1, 128, 31]> x17_1_cast_fp16 = concat(axis = var_49, interleave = x17_1_interleave_0, values = (cache_s0_l0_conv1_to_fp16, x16_1_cast_fp16))[name = tensor<string, []>("x17_1_cast_fp16")];
            tensor<int32, [3]> var_1752_begin_0 = const()[name = tensor<string, []>("op_1752_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> var_1752_end_0 = const()[name = tensor<string, []>("op_1752_end_0"), val = tensor<int32, [3]>([1, 128, 31])];
            tensor<bool, [3]> var_1752_end_mask_0 = const()[name = tensor<string, []>("op_1752_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 15]> new_cache_s0_l0_conv1 = slice_by_index(begin = var_1752_begin_0, end = var_1752_end_0, end_mask = var_1752_end_mask_0, x = x17_1_cast_fp16)[name = tensor<string, []>("op_1752_cast_fp16")];
            tensor<string, []> x_causal_2_pad_type_0 = const()[name = tensor<string, []>("x_causal_2_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_2_groups_0 = const()[name = tensor<string, []>("x_causal_2_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_2_strides_0 = const()[name = tensor<string, []>("x_causal_2_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_2_pad_0 = const()[name = tensor<string, []>("x_causal_2_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_2_dilations_0 = const()[name = tensor<string, []>("x_causal_2_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 16]> encoder_encoders_0_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(639296)))];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(643456)))];
            tensor<fp16, [1, 128, 16]> x_causal_2_cast_fp16 = conv(bias = encoder_encoders_0_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_2_dilations_0, groups = x_causal_2_groups_0, pad = x_causal_2_pad_0, pad_type = x_causal_2_pad_type_0, strides = x_causal_2_strides_0, weight = encoder_encoders_0_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16, x = x17_1_cast_fp16)[name = tensor<string, []>("x_causal_2_cast_fp16")];
            tensor<int32, [3]> input8_1_begin_0 = const()[name = tensor<string, []>("input8_1_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> input8_1_end_0 = const()[name = tensor<string, []>("input8_1_end_0"), val = tensor<int32, [3]>([1, 128, 31])];
            tensor<bool, [3]> input8_1_end_mask_0 = const()[name = tensor<string, []>("input8_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 16]> input8_1_cast_fp16 = slice_by_index(begin = input8_1_begin_0, end = input8_1_end_0, end_mask = input8_1_end_mask_0, x = x17_1_cast_fp16)[name = tensor<string, []>("input8_1_cast_fp16")];
            tensor<string, []> x_chunk_5_pad_type_0 = const()[name = tensor<string, []>("x_chunk_5_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_5_pad_0 = const()[name = tensor<string, []>("x_chunk_5_pad_0"), val = tensor<int32, [2]>([15, 15])];
            tensor<int32, []> x_chunk_5_groups_0 = const()[name = tensor<string, []>("x_chunk_5_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_5_strides_0 = const()[name = tensor<string, []>("x_chunk_5_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_5_dilations_0 = const()[name = tensor<string, []>("x_chunk_5_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 31]> encoder_encoders_0_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3968]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(643776))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(647808))), name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 1, 31])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648384)))];
            tensor<fp16, [1, 128, 16]> x_chunk_5_cast_fp16 = conv(bias = encoder_encoders_0_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_5_dilations_0, groups = x_chunk_5_groups_0, pad = x_chunk_5_pad_0, pad_type = x_chunk_5_pad_type_0, strides = x_chunk_5_strides_0, weight = encoder_encoders_0_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized, x = input8_1_cast_fp16)[name = tensor<string, []>("x_chunk_5_cast_fp16")];
            tensor<fp16, [128, 16]> chunk_scale_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale_1_to_fp16"), val = tensor<fp16, [128, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(648704)))];
            tensor<fp16, [1, 128, 16]> x_chunk_3_cast_fp16 = mul(x = x_chunk_5_cast_fp16, y = chunk_scale_1_to_fp16)[name = tensor<string, []>("x_chunk_3_cast_fp16")];
            tensor<fp16, [1, 128, 16]> x18_1_cast_fp16 = add(x = x_chunk_3_cast_fp16, y = x_causal_2_cast_fp16)[name = tensor<string, []>("x18_1_cast_fp16")];
            tensor<int32, [3]> var_1792 = const()[name = tensor<string, []>("op_1792"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_1801_to_fp16 = const()[name = tensor<string, []>("op_1801_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [16, 1, 128]> x19_1_cast_fp16 = transpose(perm = var_1792, x = x18_1_cast_fp16)[name = tensor<string, []>("transpose_106")];
            tensor<fp16, [16, 1, 128]> x_offset_4_cast_fp16 = sub(x = x19_1_cast_fp16, y = var_1801_to_fp16)[name = tensor<string, []>("x_offset_4_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1803_cast_fp16 = exp(x = x_offset_4_cast_fp16)[name = tensor<string, []>("op_1803_cast_fp16")];
            tensor<fp16, []> var_1804_to_fp16 = const()[name = tensor<string, []>("op_1804_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [16, 1, 128]> var_1805_cast_fp16 = add(x = var_1803_cast_fp16, y = var_1804_to_fp16)[name = tensor<string, []>("op_1805_cast_fp16")];
            tensor<fp32, []> var_1806_epsilon_0 = const()[name = tensor<string, []>("op_1806_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [16, 1, 128]> var_1806_cast_fp16 = log(epsilon = var_1806_epsilon_0, x = var_1805_cast_fp16)[name = tensor<string, []>("op_1806_cast_fp16")];
            tensor<fp16, []> var_1798_to_fp16 = const()[name = tensor<string, []>("op_1798_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [16, 1, 128]> var_1808_cast_fp16 = equal(x = var_1806_cast_fp16, y = var_1798_to_fp16)[name = tensor<string, []>("op_1808_cast_fp16")];
            tensor<fp16, [16, 1, 128]> log_sum0_4_cast_fp16 = select(a = x_offset_4_cast_fp16, b = var_1806_cast_fp16, cond = var_1808_cast_fp16)[name = tensor<string, []>("log_sum0_4_cast_fp16")];
            tensor<fp16, []> var_1810_to_fp16 = const()[name = tensor<string, []>("op_1810_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [16, 1, 128]> var_1811_cast_fp16 = mul(x = x19_1_cast_fp16, y = var_1810_to_fp16)[name = tensor<string, []>("op_1811_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1812_cast_fp16 = sub(x = log_sum0_4_cast_fp16, y = var_1811_cast_fp16)[name = tensor<string, []>("op_1812_cast_fp16")];
            tensor<fp16, []> var_1813_to_fp16 = const()[name = tensor<string, []>("op_1813_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [16, 1, 128]> x0_9_cast_fp16 = sub(x = var_1812_cast_fp16, y = var_1813_to_fp16)[name = tensor<string, []>("x0_9_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_0_layers_0_conv_module1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(652864))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669312))), name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(669888)))];
            tensor<fp16, [16, 1, 128]> linear_10_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_conv_module1_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_conv_module1_out_proj_weight_to_fp16_palettized, x = x0_9_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<fp16, [16, 1, 128]> input9_1_cast_fp16 = add(x = input6_1_cast_fp16, y = linear_10_cast_fp16)[name = tensor<string, []>("input9_1_cast_fp16")];
            tensor<fp16, [192, 128]> encoder_encoders_0_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(670208))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(694848))), name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([192, 128])];
            tensor<fp16, [192]> encoder_encoders_0_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695424)))];
            tensor<fp16, [16, 1, 192]> linear_11_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_feed_forward2_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized, x = input9_1_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, []> var_1829_to_fp16 = const()[name = tensor<string, []>("op_1829_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [16, 1, 192]> x_offset_6_cast_fp16 = sub(x = linear_11_cast_fp16, y = var_1829_to_fp16)[name = tensor<string, []>("x_offset_6_cast_fp16")];
            tensor<fp16, [16, 1, 192]> var_1831_cast_fp16 = exp(x = x_offset_6_cast_fp16)[name = tensor<string, []>("op_1831_cast_fp16")];
            tensor<fp16, []> var_1832_to_fp16 = const()[name = tensor<string, []>("op_1832_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [16, 1, 192]> var_1833_cast_fp16 = add(x = var_1831_cast_fp16, y = var_1832_to_fp16)[name = tensor<string, []>("op_1833_cast_fp16")];
            tensor<fp32, []> var_1834_epsilon_0 = const()[name = tensor<string, []>("op_1834_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [16, 1, 192]> var_1834_cast_fp16 = log(epsilon = var_1834_epsilon_0, x = var_1833_cast_fp16)[name = tensor<string, []>("op_1834_cast_fp16")];
            tensor<fp16, []> var_1817_to_fp16 = const()[name = tensor<string, []>("op_1817_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [16, 1, 192]> var_1836_cast_fp16 = equal(x = var_1834_cast_fp16, y = var_1817_to_fp16)[name = tensor<string, []>("op_1836_cast_fp16")];
            tensor<fp16, [16, 1, 192]> log_sum0_6_cast_fp16 = select(a = x_offset_6_cast_fp16, b = var_1834_cast_fp16, cond = var_1836_cast_fp16)[name = tensor<string, []>("log_sum0_6_cast_fp16")];
            tensor<fp16, []> var_1838_to_fp16 = const()[name = tensor<string, []>("op_1838_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [16, 1, 192]> var_1839_cast_fp16 = mul(x = linear_11_cast_fp16, y = var_1838_to_fp16)[name = tensor<string, []>("op_1839_cast_fp16")];
            tensor<fp16, [16, 1, 192]> var_1840_cast_fp16 = sub(x = log_sum0_6_cast_fp16, y = var_1839_cast_fp16)[name = tensor<string, []>("op_1840_cast_fp16")];
            tensor<fp16, []> var_1841_to_fp16 = const()[name = tensor<string, []>("op_1841_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [16, 1, 192]> x_17_cast_fp16 = sub(x = var_1840_cast_fp16, y = var_1841_to_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<fp16, [128, 192]> encoder_encoders_0_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(695872))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(720512))), name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 192])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721088)))];
            tensor<fp16, [16, 1, 128]> linear_12_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_feed_forward2_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized, x = x_17_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<fp16, [16, 1, 128]> src_3_cast_fp16 = add(x = input9_1_cast_fp16, y = linear_12_cast_fp16)[name = tensor<string, []>("src_3_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1847_cast_fp16 = sub(x = src_3_cast_fp16, y = x2_1_cast_fp16)[name = tensor<string, []>("op_1847_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721408)))];
            tensor<fp16, [16, 1, 128]> var_1848_cast_fp16 = mul(x = var_1847_cast_fp16, y = encoder_encoders_0_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_1848_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1849_cast_fp16 = add(x = x2_1_cast_fp16, y = var_1848_cast_fp16)[name = tensor<string, []>("op_1849_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_0_layers_0_self_attn2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(721728))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(727936))), name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_0_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728512)))];
            tensor<fp16, [16, 1, 48]> linear_13_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_self_attn2_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_self_attn2_in_proj_weight_to_fp16_palettized, x = var_1849_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<bool, []> x20_1_interleave_0 = const()[name = tensor<string, []>("x20_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s0_l0_val2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s0_l0_val2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [64, 1, 48]> cache_s0_l0_val2_to_fp16 = cast(dtype = cache_s0_l0_val2_to_fp16_dtype_0, x = cache_s0_l0_val2)[name = tensor<string, []>("cast_31")];
            tensor<fp16, [80, 1, 48]> x20_1_cast_fp16 = concat(axis = var_46, interleave = x20_1_interleave_0, values = (cache_s0_l0_val2_to_fp16, linear_13_cast_fp16))[name = tensor<string, []>("x20_1_cast_fp16")];
            tensor<int32, [3]> var_1868_begin_0 = const()[name = tensor<string, []>("op_1868_begin_0"), val = tensor<int32, [3]>([16, 0, 0])];
            tensor<int32, [3]> var_1868_end_0 = const()[name = tensor<string, []>("op_1868_end_0"), val = tensor<int32, [3]>([80, 1, 48])];
            tensor<bool, [3]> var_1868_end_mask_0 = const()[name = tensor<string, []>("op_1868_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [64, 1, 48]> new_cache_s0_l0_val2 = slice_by_index(begin = var_1868_begin_0, end = var_1868_end_0, end_mask = var_1868_end_mask_0, x = x20_1_cast_fp16)[name = tensor<string, []>("op_1868_cast_fp16")];
            tensor<int32, [4]> var_1869 = const()[name = tensor<string, []>("op_1869"), val = tensor<int32, [4]>([80, 1, 4, -1])];
            tensor<fp16, [80, 1, 4, 12]> var_1870_cast_fp16 = reshape(shape = var_1869, x = x20_1_cast_fp16)[name = tensor<string, []>("op_1870_cast_fp16")];
            tensor<int32, [4]> var_1871 = const()[name = tensor<string, []>("op_1871"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x22_1_transpose_x_0 = const()[name = tensor<string, []>("x22_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x22_1_transpose_y_0 = const()[name = tensor<string, []>("x22_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 80, 12]> x21_1_cast_fp16 = transpose(perm = var_1871, x = var_1870_cast_fp16)[name = tensor<string, []>("transpose_105")];
            tensor<fp16, [4, 1, 16, 12]> x22_1_cast_fp16 = matmul(transpose_x = x22_1_transpose_x_0, transpose_y = x22_1_transpose_y_0, x = attn_weights_1_cast_fp16, y = x21_1_cast_fp16)[name = tensor<string, []>("x22_1_cast_fp16")];
            tensor<int32, [4]> var_1876 = const()[name = tensor<string, []>("op_1876"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_1881 = const()[name = tensor<string, []>("op_1881"), val = tensor<int32, [3]>([16, 1, 48])];
            tensor<fp16, [16, 1, 4, 12]> var_1877_cast_fp16 = transpose(perm = var_1876, x = x22_1_cast_fp16)[name = tensor<string, []>("transpose_104")];
            tensor<fp16, [16, 1, 48]> input10_1_cast_fp16 = reshape(shape = var_1881, x = var_1877_cast_fp16)[name = tensor<string, []>("input10_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_0_layers_0_self_attn2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(728704))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(734912))), name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(735488)))];
            tensor<fp16, [16, 1, 128]> linear_14_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_self_attn2_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_self_attn2_out_proj_weight_to_fp16_palettized, x = input10_1_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<fp16, [16, 1, 128]> input11_1_cast_fp16 = add(x = var_1849_cast_fp16, y = linear_14_cast_fp16)[name = tensor<string, []>("input11_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_0_layers_0_conv_module2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(735808))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(768640))), name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_0_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(769216)))];
            tensor<fp16, [16, 1, 256]> linear_15_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_conv_module2_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_conv_module2_in_proj_weight_to_fp16_palettized, x = input11_1_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<int32, [2]> var_1892_split_sizes_0 = const()[name = tensor<string, []>("op_1892_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_1892_axis_0 = const()[name = tensor<string, []>("op_1892_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [16, 1, 128]> var_1892_cast_fp16_0, tensor<fp16, [16, 1, 128]> var_1892_cast_fp16_1 = split(axis = var_1892_axis_0, split_sizes = var_1892_split_sizes_0, x = linear_15_cast_fp16)[name = tensor<string, []>("op_1892_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1894_cast_fp16 = sigmoid(x = var_1892_cast_fp16_1)[name = tensor<string, []>("op_1894_cast_fp16")];
            tensor<fp16, [16, 1, 128]> x24_1_cast_fp16 = mul(x = var_1892_cast_fp16_0, y = var_1894_cast_fp16)[name = tensor<string, []>("x24_1_cast_fp16")];
            tensor<int32, [3]> var_1896 = const()[name = tensor<string, []>("op_1896"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 16]> x25_1_cast_fp16 = transpose(perm = var_1896, x = x24_1_cast_fp16)[name = tensor<string, []>("transpose_103")];
            tensor<fp16, [1, 128, 16]> x26_1_cast_fp16 = select(a = var_62_to_fp16, b = x25_1_cast_fp16, cond = var_1745)[name = tensor<string, []>("x26_1_cast_fp16")];
            tensor<bool, []> x27_1_interleave_0 = const()[name = tensor<string, []>("x27_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s0_l0_conv2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s0_l0_conv2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 15]> cache_s0_l0_conv2_to_fp16 = cast(dtype = cache_s0_l0_conv2_to_fp16_dtype_0, x = cache_s0_l0_conv2)[name = tensor<string, []>("cast_30")];
            tensor<fp16, [1, 128, 31]> x27_1_cast_fp16 = concat(axis = var_49, interleave = x27_1_interleave_0, values = (cache_s0_l0_conv2_to_fp16, x26_1_cast_fp16))[name = tensor<string, []>("x27_1_cast_fp16")];
            tensor<int32, [3]> var_1906_begin_0 = const()[name = tensor<string, []>("op_1906_begin_0"), val = tensor<int32, [3]>([0, 0, 16])];
            tensor<int32, [3]> var_1906_end_0 = const()[name = tensor<string, []>("op_1906_end_0"), val = tensor<int32, [3]>([1, 128, 31])];
            tensor<bool, [3]> var_1906_end_mask_0 = const()[name = tensor<string, []>("op_1906_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 15]> new_cache_s0_l0_conv2 = slice_by_index(begin = var_1906_begin_0, end = var_1906_end_0, end_mask = var_1906_end_mask_0, x = x27_1_cast_fp16)[name = tensor<string, []>("op_1906_cast_fp16")];
            tensor<string, []> x_causal_4_pad_type_0 = const()[name = tensor<string, []>("x_causal_4_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_4_groups_0 = const()[name = tensor<string, []>("x_causal_4_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_4_strides_0 = const()[name = tensor<string, []>("x_causal_4_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_4_pad_0 = const()[name = tensor<string, []>("x_causal_4_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_4_dilations_0 = const()[name = tensor<string, []>("x_causal_4_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 16]> encoder_encoders_0_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(769792)))];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(773952)))];
            tensor<fp16, [1, 128, 16]> x_causal_4_cast_fp16 = conv(bias = encoder_encoders_0_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_4_dilations_0, groups = x_causal_4_groups_0, pad = x_causal_4_pad_0, pad_type = x_causal_4_pad_type_0, strides = x_causal_4_strides_0, weight = encoder_encoders_0_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16, x = x27_1_cast_fp16)[name = tensor<string, []>("x_causal_4_cast_fp16")];
            tensor<int32, [3]> input13_1_begin_0 = const()[name = tensor<string, []>("input13_1_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> input13_1_end_0 = const()[name = tensor<string, []>("input13_1_end_0"), val = tensor<int32, [3]>([1, 128, 31])];
            tensor<bool, [3]> input13_1_end_mask_0 = const()[name = tensor<string, []>("input13_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 16]> input13_1_cast_fp16 = slice_by_index(begin = input13_1_begin_0, end = input13_1_end_0, end_mask = input13_1_end_mask_0, x = x27_1_cast_fp16)[name = tensor<string, []>("input13_1_cast_fp16")];
            tensor<string, []> x_chunk_7_pad_type_0 = const()[name = tensor<string, []>("x_chunk_7_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_7_pad_0 = const()[name = tensor<string, []>("x_chunk_7_pad_0"), val = tensor<int32, [2]>([15, 15])];
            tensor<int32, []> x_chunk_7_groups_0 = const()[name = tensor<string, []>("x_chunk_7_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_7_strides_0 = const()[name = tensor<string, []>("x_chunk_7_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_7_dilations_0 = const()[name = tensor<string, []>("x_chunk_7_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 31]> encoder_encoders_0_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3968]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(774272))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778304))), name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 1, 31])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(778880)))];
            tensor<fp16, [1, 128, 16]> x_chunk_7_cast_fp16 = conv(bias = encoder_encoders_0_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_7_dilations_0, groups = x_chunk_7_groups_0, pad = x_chunk_7_pad_0, pad_type = x_chunk_7_pad_type_0, strides = x_chunk_7_strides_0, weight = encoder_encoders_0_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized, x = input13_1_cast_fp16)[name = tensor<string, []>("x_chunk_7_cast_fp16")];
            tensor<fp16, [128, 16]> chunk_scale0_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale0_1_to_fp16"), val = tensor<fp16, [128, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(779200)))];
            tensor<fp16, [1, 128, 16]> x_chunk0_1_cast_fp16 = mul(x = x_chunk_7_cast_fp16, y = chunk_scale0_1_to_fp16)[name = tensor<string, []>("x_chunk0_1_cast_fp16")];
            tensor<fp16, [1, 128, 16]> x28_1_cast_fp16 = add(x = x_chunk0_1_cast_fp16, y = x_causal_4_cast_fp16)[name = tensor<string, []>("x28_1_cast_fp16")];
            tensor<int32, [3]> var_1946 = const()[name = tensor<string, []>("op_1946"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_1955_to_fp16 = const()[name = tensor<string, []>("op_1955_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [16, 1, 128]> x29_1_cast_fp16 = transpose(perm = var_1946, x = x28_1_cast_fp16)[name = tensor<string, []>("transpose_102")];
            tensor<fp16, [16, 1, 128]> x_offset_8_cast_fp16 = sub(x = x29_1_cast_fp16, y = var_1955_to_fp16)[name = tensor<string, []>("x_offset_8_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1957_cast_fp16 = exp(x = x_offset_8_cast_fp16)[name = tensor<string, []>("op_1957_cast_fp16")];
            tensor<fp16, []> var_1958_to_fp16 = const()[name = tensor<string, []>("op_1958_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [16, 1, 128]> var_1959_cast_fp16 = add(x = var_1957_cast_fp16, y = var_1958_to_fp16)[name = tensor<string, []>("op_1959_cast_fp16")];
            tensor<fp32, []> var_1960_epsilon_0 = const()[name = tensor<string, []>("op_1960_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [16, 1, 128]> var_1960_cast_fp16 = log(epsilon = var_1960_epsilon_0, x = var_1959_cast_fp16)[name = tensor<string, []>("op_1960_cast_fp16")];
            tensor<fp16, []> var_1952_to_fp16 = const()[name = tensor<string, []>("op_1952_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [16, 1, 128]> var_1962_cast_fp16 = equal(x = var_1960_cast_fp16, y = var_1952_to_fp16)[name = tensor<string, []>("op_1962_cast_fp16")];
            tensor<fp16, [16, 1, 128]> log_sum0_8_cast_fp16 = select(a = x_offset_8_cast_fp16, b = var_1960_cast_fp16, cond = var_1962_cast_fp16)[name = tensor<string, []>("log_sum0_8_cast_fp16")];
            tensor<fp16, []> var_1964_to_fp16 = const()[name = tensor<string, []>("op_1964_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [16, 1, 128]> var_1965_cast_fp16 = mul(x = x29_1_cast_fp16, y = var_1964_to_fp16)[name = tensor<string, []>("op_1965_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_1966_cast_fp16 = sub(x = log_sum0_8_cast_fp16, y = var_1965_cast_fp16)[name = tensor<string, []>("op_1966_cast_fp16")];
            tensor<fp16, []> var_1967_to_fp16 = const()[name = tensor<string, []>("op_1967_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [16, 1, 128]> x0_11_cast_fp16 = sub(x = var_1966_cast_fp16, y = var_1967_to_fp16)[name = tensor<string, []>("x0_11_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_0_layers_0_conv_module2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(783360))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(799808))), name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800384)))];
            tensor<fp16, [16, 1, 128]> linear_16_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_conv_module2_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_conv_module2_out_proj_weight_to_fp16_palettized, x = x0_11_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<fp16, [16, 1, 128]> input14_1_cast_fp16 = add(x = input11_1_cast_fp16, y = linear_16_cast_fp16)[name = tensor<string, []>("input14_1_cast_fp16")];
            tensor<fp16, [240, 128]> encoder_encoders_0_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(800704))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(831488))), name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([240, 128])];
            tensor<fp16, [240]> encoder_encoders_0_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832064)))];
            tensor<fp16, [16, 1, 240]> linear_17_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_feed_forward3_in_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized, x = input14_1_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, []> var_1983_to_fp16 = const()[name = tensor<string, []>("op_1983_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [16, 1, 240]> x_offset_10_cast_fp16 = sub(x = linear_17_cast_fp16, y = var_1983_to_fp16)[name = tensor<string, []>("x_offset_10_cast_fp16")];
            tensor<fp16, [16, 1, 240]> var_1985_cast_fp16 = exp(x = x_offset_10_cast_fp16)[name = tensor<string, []>("op_1985_cast_fp16")];
            tensor<fp16, []> var_1986_to_fp16 = const()[name = tensor<string, []>("op_1986_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [16, 1, 240]> var_1987_cast_fp16 = add(x = var_1985_cast_fp16, y = var_1986_to_fp16)[name = tensor<string, []>("op_1987_cast_fp16")];
            tensor<fp32, []> var_1988_epsilon_0 = const()[name = tensor<string, []>("op_1988_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [16, 1, 240]> var_1988_cast_fp16 = log(epsilon = var_1988_epsilon_0, x = var_1987_cast_fp16)[name = tensor<string, []>("op_1988_cast_fp16")];
            tensor<fp16, []> var_1971_to_fp16 = const()[name = tensor<string, []>("op_1971_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [16, 1, 240]> var_1990_cast_fp16 = equal(x = var_1988_cast_fp16, y = var_1971_to_fp16)[name = tensor<string, []>("op_1990_cast_fp16")];
            tensor<fp16, [16, 1, 240]> log_sum0_10_cast_fp16 = select(a = x_offset_10_cast_fp16, b = var_1988_cast_fp16, cond = var_1990_cast_fp16)[name = tensor<string, []>("log_sum0_10_cast_fp16")];
            tensor<fp16, []> var_1992_to_fp16 = const()[name = tensor<string, []>("op_1992_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [16, 1, 240]> var_1993_cast_fp16 = mul(x = linear_17_cast_fp16, y = var_1992_to_fp16)[name = tensor<string, []>("op_1993_cast_fp16")];
            tensor<fp16, [16, 1, 240]> var_1994_cast_fp16 = sub(x = log_sum0_10_cast_fp16, y = var_1993_cast_fp16)[name = tensor<string, []>("op_1994_cast_fp16")];
            tensor<fp16, []> var_1995_to_fp16 = const()[name = tensor<string, []>("op_1995_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [16, 1, 240]> x_21_cast_fp16 = sub(x = var_1994_cast_fp16, y = var_1995_to_fp16)[name = tensor<string, []>("x_21_cast_fp16")];
            tensor<fp16, [128, 240]> encoder_encoders_0_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832640))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(863424))), name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 240])];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(864000)))];
            tensor<fp16, [16, 1, 128]> linear_18_cast_fp16 = linear(bias = encoder_encoders_0_layers_0_feed_forward3_out_proj_bias_to_fp16, weight = encoder_encoders_0_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized, x = x_21_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, [16, 1, 128]> x30_1_cast_fp16 = add(x = input14_1_cast_fp16, y = linear_18_cast_fp16)[name = tensor<string, []>("x30_1_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(864320)))];
            tensor<fp16, [16, 1, 128]> var_2006_cast_fp16 = sub(x = x30_1_cast_fp16, y = encoder_encoders_0_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("op_2006_cast_fp16")];
            tensor<fp16, []> var_2000_promoted_to_fp16 = const()[name = tensor<string, []>("op_2000_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [16, 1, 128]> var_2007_cast_fp16 = pow(x = var_2006_cast_fp16, y = var_2000_promoted_to_fp16)[name = tensor<string, []>("op_2007_cast_fp16")];
            tensor<int32, [1]> var_2009_axes_0 = const()[name = tensor<string, []>("op_2009_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, []> var_2009_keep_dims_0 = const()[name = tensor<string, []>("op_2009_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [16, 1, 1]> var_2009_cast_fp16 = reduce_mean(axes = var_2009_axes_0, keep_dims = var_2009_keep_dims_0, x = var_2007_cast_fp16)[name = tensor<string, []>("op_2009_cast_fp16")];
            tensor<fp16, []> var_1999_to_fp16 = const()[name = tensor<string, []>("op_1999_to_fp16"), val = tensor<fp16, []>(-0x1p-1)];
            tensor<fp16, [16, 1, 1]> var_2010_cast_fp16 = pow(x = var_2009_cast_fp16, y = var_1999_to_fp16)[name = tensor<string, []>("op_2010_cast_fp16")];
            tensor<fp16, []> var_2011_to_fp16 = const()[name = tensor<string, []>("op_2011_to_fp16"), val = tensor<fp16, []>(0x1.1acp-2)];
            tensor<fp16, [16, 1, 1]> scales_4_cast_fp16 = mul(x = var_2010_cast_fp16, y = var_2011_to_fp16)[name = tensor<string, []>("scales_4_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_2013_cast_fp16 = mul(x = x30_1_cast_fp16, y = scales_4_cast_fp16)[name = tensor<string, []>("op_2013_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_2016_cast_fp16 = sub(x = var_2013_cast_fp16, y = x2_1_cast_fp16)[name = tensor<string, []>("op_2016_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_0_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_0_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(864640)))];
            tensor<fp16, [16, 1, 128]> var_2017_cast_fp16 = mul(x = var_2016_cast_fp16, y = encoder_encoders_0_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_2017_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_2018_cast_fp16 = add(x = x2_1_cast_fp16, y = var_2017_cast_fp16)[name = tensor<string, []>("op_2018_cast_fp16")];
            tensor<int32, [2]> key_padding_mask0_1_begin_0 = const()[name = tensor<string, []>("key_padding_mask0_1_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> key_padding_mask0_1_end_0 = const()[name = tensor<string, []>("key_padding_mask0_1_end_0"), val = tensor<int32, [2]>([1, 80])];
            tensor<bool, [2]> key_padding_mask0_1_end_mask_0 = const()[name = tensor<string, []>("key_padding_mask0_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<int32, [2]> key_padding_mask0_1_stride_0 = const()[name = tensor<string, []>("key_padding_mask0_1_stride_0"), val = tensor<int32, [2]>([1, 2])];
            tensor<bool, [1, 40]> key_padding_mask0_1 = slice_by_index(begin = key_padding_mask0_1_begin_0, end = key_padding_mask0_1_end_0, end_mask = key_padding_mask0_1_end_mask_0, stride = key_padding_mask0_1_stride_0, x = src_key_padding_mask0_1)[name = tensor<string, []>("key_padding_mask0_1")];
            tensor<int32, [4]> var_2039 = const()[name = tensor<string, []>("op_2039"), val = tensor<int32, [4]>([8, 2, 1, 128])];
            tensor<fp16, [8, 2, 1, 128]> src0_9_cast_fp16 = reshape(shape = var_2039, x = var_2018_cast_fp16)[name = tensor<string, []>("src0_9_cast_fp16")];
            tensor<fp16, [2, 1, 1]> weights0_2_to_fp16 = const()[name = tensor<string, []>("weights0_2_to_fp16"), val = tensor<fp16, [2, 1, 1]>([[[0x1p-1]], [[0x1p-1]]])];
            tensor<fp16, [8, 2, 1, 128]> var_2044_cast_fp16 = mul(x = src0_9_cast_fp16, y = weights0_2_to_fp16)[name = tensor<string, []>("op_2044_cast_fp16")];
            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> x_23_keep_dims_0 = const()[name = tensor<string, []>("x_23_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [8, 1, 128]> x_23_cast_fp16 = reduce_sum(axes = x_23_axes_0, keep_dims = x_23_keep_dims_0, x = var_2044_cast_fp16)[name = tensor<string, []>("x_23_cast_fp16")];
            tensor<fp16, [272, 128]> encoder_encoders_1_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [34816]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(864960))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(899840))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([272, 128])];
            tensor<fp16, [272]> encoder_encoders_1_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [272]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(900416)))];
            tensor<fp16, [8, 1, 272]> linear_19_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized, x = x_23_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<int32, [3]> q2_1_begin_0 = const()[name = tensor<string, []>("q2_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> q2_1_end_0 = const()[name = tensor<string, []>("q2_1_end_0"), val = tensor<int32, [3]>([8, 1, 128])];
            tensor<bool, [3]> q2_1_end_mask_0 = const()[name = tensor<string, []>("q2_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 1, 128]> q2_1_cast_fp16 = slice_by_index(begin = q2_1_begin_0, end = q2_1_end_0, end_mask = q2_1_end_mask_0, x = linear_19_cast_fp16)[name = tensor<string, []>("q2_1_cast_fp16")];
            tensor<int32, [3]> k3_1_begin_0 = const()[name = tensor<string, []>("k3_1_begin_0"), val = tensor<int32, [3]>([0, 0, 128])];
            tensor<int32, [3]> k3_1_end_0 = const()[name = tensor<string, []>("k3_1_end_0"), val = tensor<int32, [3]>([8, 1, 256])];
            tensor<bool, [3]> k3_1_end_mask_0 = const()[name = tensor<string, []>("k3_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 1, 128]> k3_1_cast_fp16 = slice_by_index(begin = k3_1_begin_0, end = k3_1_end_0, end_mask = k3_1_end_mask_0, x = linear_19_cast_fp16)[name = tensor<string, []>("k3_1_cast_fp16")];
            tensor<int32, [3]> p2_1_begin_0 = const()[name = tensor<string, []>("p2_1_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> p2_1_end_0 = const()[name = tensor<string, []>("p2_1_end_0"), val = tensor<int32, [3]>([8, 1, 272])];
            tensor<bool, [3]> p2_1_end_mask_0 = const()[name = tensor<string, []>("p2_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 1, 16]> p2_1_cast_fp16 = slice_by_index(begin = p2_1_begin_0, end = p2_1_end_0, end_mask = p2_1_end_mask_0, x = linear_19_cast_fp16)[name = tensor<string, []>("p2_1_cast_fp16")];
            tensor<bool, []> k4_1_interleave_0 = const()[name = tensor<string, []>("k4_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s1_l0_key_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s1_l0_key_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [32, 1, 128]> cache_s1_l0_key_to_fp16 = cast(dtype = cache_s1_l0_key_to_fp16_dtype_0, x = cache_s1_l0_key)[name = tensor<string, []>("cast_29")];
            tensor<fp16, [40, 1, 128]> k4_1_cast_fp16 = concat(axis = var_46, interleave = k4_1_interleave_0, values = (cache_s1_l0_key_to_fp16, k3_1_cast_fp16))[name = tensor<string, []>("k4_1_cast_fp16")];
            tensor<int32, [3]> var_2097_begin_0 = const()[name = tensor<string, []>("op_2097_begin_0"), val = tensor<int32, [3]>([8, 0, 0])];
            tensor<int32, [3]> var_2097_end_0 = const()[name = tensor<string, []>("op_2097_end_0"), val = tensor<int32, [3]>([40, 1, 128])];
            tensor<bool, [3]> var_2097_end_mask_0 = const()[name = tensor<string, []>("op_2097_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [32, 1, 128]> new_cache_s1_l0_key = slice_by_index(begin = var_2097_begin_0, end = var_2097_end_0, end_mask = var_2097_end_mask_0, x = k4_1_cast_fp16)[name = tensor<string, []>("op_2097_cast_fp16")];
            tensor<int32, [4]> var_2103 = const()[name = tensor<string, []>("op_2103"), val = tensor<int32, [4]>([8, 1, 4, 32])];
            tensor<fp16, [8, 1, 4, 32]> q3_1_cast_fp16 = reshape(shape = var_2103, x = q2_1_cast_fp16)[name = tensor<string, []>("q3_1_cast_fp16")];
            tensor<int32, [4]> var_2105 = const()[name = tensor<string, []>("op_2105"), val = tensor<int32, [4]>([8, 1, 4, 4])];
            tensor<fp16, [8, 1, 4, 4]> p3_1_cast_fp16 = reshape(shape = var_2105, x = p2_1_cast_fp16)[name = tensor<string, []>("p3_1_cast_fp16")];
            tensor<int32, [4]> var_2107 = const()[name = tensor<string, []>("op_2107"), val = tensor<int32, [4]>([40, 1, 4, 32])];
            tensor<fp16, [40, 1, 4, 32]> k5_1_cast_fp16 = reshape(shape = var_2107, x = k4_1_cast_fp16)[name = tensor<string, []>("k5_1_cast_fp16")];
            tensor<int32, [4]> var_2111 = const()[name = tensor<string, []>("op_2111"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> attn_scores2_1_transpose_x_0 = const()[name = tensor<string, []>("attn_scores2_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_scores2_1_transpose_y_0 = const()[name = tensor<string, []>("attn_scores2_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = tensor<string, []>("transpose_26_perm_0"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<int32, [4]> transpose_27_perm_0 = const()[name = tensor<string, []>("transpose_27_perm_0"), val = tensor<int32, [4]>([2, 1, -1, -4])];
            tensor<fp16, [4, 1, 32, 40]> transpose_27 = transpose(perm = transpose_27_perm_0, x = k5_1_cast_fp16)[name = tensor<string, []>("transpose_100")];
            tensor<fp16, [4, 1, 8, 32]> transpose_26 = transpose(perm = transpose_26_perm_0, x = q3_1_cast_fp16)[name = tensor<string, []>("transpose_101")];
            tensor<fp16, [4, 1, 8, 40]> attn_scores2_1_cast_fp16 = matmul(transpose_x = attn_scores2_1_transpose_x_0, transpose_y = attn_scores2_1_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor<string, []>("attn_scores2_1_cast_fp16")];
            tensor<bool, []> pos_scores3_1_transpose_x_0 = const()[name = tensor<string, []>("pos_scores3_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_scores3_1_transpose_y_0 = const()[name = tensor<string, []>("pos_scores3_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 4, 47]> pos_emb0_1_to_fp16 = const()[name = tensor<string, []>("pos_emb0_1_to_fp16"), val = tensor<fp16, [4, 1, 4, 47]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(901056)))];
            tensor<fp16, [4, 1, 8, 4]> p4_1_cast_fp16 = transpose(perm = var_2111, x = p3_1_cast_fp16)[name = tensor<string, []>("transpose_99")];
            tensor<fp16, [4, 1, 8, 47]> pos_scores3_1_cast_fp16 = matmul(transpose_x = pos_scores3_1_transpose_x_0, transpose_y = pos_scores3_1_transpose_y_0, x = p4_1_cast_fp16, y = pos_emb0_1_to_fp16)[name = tensor<string, []>("pos_scores3_1_cast_fp16")];
            tensor<int32, [2]> var_2154 = const()[name = tensor<string, []>("op_2154"), val = tensor<int32, [2]>([-1, 47])];
            tensor<fp16, [32, 47]> pos_scores4_1_cast_fp16 = reshape(shape = var_2154, x = pos_scores3_1_cast_fp16)[name = tensor<string, []>("pos_scores4_1_cast_fp16")];
            tensor<bool, []> pos_scores5_1_validate_indices_0 = const()[name = tensor<string, []>("pos_scores5_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<uint16, [32, 40]> indexes0_1_to_uint16 = const()[name = tensor<string, []>("indexes0_1_to_uint16"), val = tensor<uint16, [32, 40]>([[7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39], [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39], [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39], [7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46], [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45], [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44], [4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39]])];
            tensor<fp16, [32, 40]> pos_scores5_1_cast_fp16_cast_uint16 = gather_along_axis(axis = var_48, indices = indexes0_1_to_uint16, validate_indices = pos_scores5_1_validate_indices_0, x = pos_scores4_1_cast_fp16)[name = tensor<string, []>("pos_scores5_1_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_2157 = const()[name = tensor<string, []>("op_2157"), val = tensor<int32, [4]>([4, 1, 8, 40])];
            tensor<fp16, [4, 1, 8, 40]> pos_scores6_1_cast_fp16 = reshape(shape = var_2157, x = pos_scores5_1_cast_fp16_cast_uint16)[name = tensor<string, []>("pos_scores6_1_cast_fp16")];
            tensor<fp16, [4, 1, 8, 40]> attn_scores3_1_cast_fp16 = add(x = attn_scores2_1_cast_fp16, y = pos_scores6_1_cast_fp16)[name = tensor<string, []>("attn_scores3_1_cast_fp16")];
            tensor<int32, [1]> var_2160_axes_0 = const()[name = tensor<string, []>("op_2160_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 40]> var_2160 = expand_dims(axes = var_2160_axes_0, x = key_padding_mask0_1)[name = tensor<string, []>("op_2160")];
            tensor<fp16, []> cast_27_to_fp16 = const()[name = tensor<string, []>("cast_27_to_fp16"), val = tensor<fp16, []>(-0x1.f4p+9)];
            tensor<fp16, [4, 1, 8, 40]> attn_scores4_1_cast_fp16 = select(a = cast_27_to_fp16, b = attn_scores3_1_cast_fp16, cond = var_2160)[name = tensor<string, []>("attn_scores4_1_cast_fp16")];
            tensor<fp16, [4, 1, 8, 40]> attn_weights1_1_cast_fp16 = softmax(axis = var_59, x = attn_scores4_1_cast_fp16)[name = tensor<string, []>("attn_weights1_1_cast_fp16")];
            tensor<fp16, [144, 128]> encoder_encoders_1_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(902656))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(921152))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([144, 128])];
            tensor<fp16, [144]> encoder_encoders_1_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(921728)))];
            tensor<fp16, [8, 1, 144]> linear_21_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized, x = x_23_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<fp16, []> var_2175_to_fp16 = const()[name = tensor<string, []>("op_2175_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [8, 1, 144]> x_offset_12_cast_fp16 = sub(x = linear_21_cast_fp16, y = var_2175_to_fp16)[name = tensor<string, []>("x_offset_12_cast_fp16")];
            tensor<fp16, [8, 1, 144]> var_2177_cast_fp16 = exp(x = x_offset_12_cast_fp16)[name = tensor<string, []>("op_2177_cast_fp16")];
            tensor<fp16, []> var_2178_to_fp16 = const()[name = tensor<string, []>("op_2178_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 144]> var_2179_cast_fp16 = add(x = var_2177_cast_fp16, y = var_2178_to_fp16)[name = tensor<string, []>("op_2179_cast_fp16")];
            tensor<fp32, []> var_2180_epsilon_0 = const()[name = tensor<string, []>("op_2180_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 144]> var_2180_cast_fp16 = log(epsilon = var_2180_epsilon_0, x = var_2179_cast_fp16)[name = tensor<string, []>("op_2180_cast_fp16")];
            tensor<fp16, []> var_2163_to_fp16 = const()[name = tensor<string, []>("op_2163_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 144]> var_2182_cast_fp16 = equal(x = var_2180_cast_fp16, y = var_2163_to_fp16)[name = tensor<string, []>("op_2182_cast_fp16")];
            tensor<fp16, [8, 1, 144]> log_sum0_12_cast_fp16 = select(a = x_offset_12_cast_fp16, b = var_2180_cast_fp16, cond = var_2182_cast_fp16)[name = tensor<string, []>("log_sum0_12_cast_fp16")];
            tensor<fp16, []> var_2184_to_fp16 = const()[name = tensor<string, []>("op_2184_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 144]> var_2185_cast_fp16 = mul(x = linear_21_cast_fp16, y = var_2184_to_fp16)[name = tensor<string, []>("op_2185_cast_fp16")];
            tensor<fp16, [8, 1, 144]> var_2186_cast_fp16 = sub(x = log_sum0_12_cast_fp16, y = var_2185_cast_fp16)[name = tensor<string, []>("op_2186_cast_fp16")];
            tensor<fp16, []> var_2187_to_fp16 = const()[name = tensor<string, []>("op_2187_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [8, 1, 144]> x_29_cast_fp16 = sub(x = var_2186_cast_fp16, y = var_2187_to_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
            tensor<fp16, [128, 144]> encoder_encoders_1_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(922112))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(940608))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 144])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(941184)))];
            tensor<fp16, [8, 1, 128]> linear_22_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized, x = x_29_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input15_1_cast_fp16 = add(x = x_23_cast_fp16, y = linear_22_cast_fp16)[name = tensor<string, []>("input15_1_cast_fp16")];
            tensor<int32, [4]> attn_weights2_1_begin_0 = const()[name = tensor<string, []>("attn_weights2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_weights2_1_end_0 = const()[name = tensor<string, []>("attn_weights2_1_end_0"), val = tensor<int32, [4]>([1, 1, 8, 40])];
            tensor<bool, [4]> attn_weights2_1_end_mask_0 = const()[name = tensor<string, []>("attn_weights2_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1, 8, 40]> attn_weights2_1_cast_fp16 = slice_by_index(begin = attn_weights2_1_begin_0, end = attn_weights2_1_end_0, end_mask = attn_weights2_1_end_mask_0, x = attn_weights1_1_cast_fp16)[name = tensor<string, []>("attn_weights2_1_cast_fp16")];
            tensor<fp16, [288, 128]> encoder_encoders_1_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [36864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(941504))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(978432))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([288, 128])];
            tensor<fp16, [288]> encoder_encoders_1_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(979008)))];
            tensor<fp16, [8, 1, 288]> linear_23_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized, x = input15_1_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<int32, [3]> var_2201_split_sizes_0 = const()[name = tensor<string, []>("op_2201_split_sizes_0"), val = tensor<int32, [3]>([96, 96, 96])];
            tensor<int32, []> var_2201_axis_0 = const()[name = tensor<string, []>("op_2201_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [8, 1, 96]> var_2201_cast_fp16_0, tensor<fp16, [8, 1, 96]> var_2201_cast_fp16_1, tensor<fp16, [8, 1, 96]> var_2201_cast_fp16_2 = split(axis = var_2201_axis_0, split_sizes = var_2201_split_sizes_0, x = linear_23_cast_fp16)[name = tensor<string, []>("op_2201_cast_fp16")];
            tensor<fp16, [8, 1, 96]> var_2204_cast_fp16 = tanh(x = var_2201_cast_fp16_0)[name = tensor<string, []>("op_2204_cast_fp16")];
            tensor<int32, [1]> var_2205_axes_0 = const()[name = tensor<string, []>("op_2205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [8, 1, 96, 1]> var_2205_cast_fp16 = expand_dims(axes = var_2205_axes_0, x = var_2204_cast_fp16)[name = tensor<string, []>("op_2205_cast_fp16")];
            tensor<int32, [3]> var_2206 = const()[name = tensor<string, []>("op_2206"), val = tensor<int32, [3]>([8, 1, 96])];
            tensor<fp16, [8, 1, 96]> s0_1_cast_fp16 = reshape(shape = var_2206, x = var_2205_cast_fp16)[name = tensor<string, []>("s0_1_cast_fp16")];
            tensor<fp16, [8, 1, 96]> x32_1_cast_fp16 = mul(x = var_2201_cast_fp16_1, y = s0_1_cast_fp16)[name = tensor<string, []>("x32_1_cast_fp16")];
            tensor<int32, [4]> var_2220 = const()[name = tensor<string, []>("op_2220"), val = tensor<int32, [4]>([8, 1, 1, -1])];
            tensor<fp16, [8, 1, 1, 96]> var_2221_cast_fp16 = reshape(shape = var_2220, x = x32_1_cast_fp16)[name = tensor<string, []>("op_2221_cast_fp16")];
            tensor<int32, [4]> var_2222 = const()[name = tensor<string, []>("op_2222"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<bool, []> x_pad0_1_interleave_0 = const()[name = tensor<string, []>("x_pad0_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s1_l0_nonlin_attn_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s1_l0_nonlin_attn_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 8, 96]> x33_1_cast_fp16 = transpose(perm = var_2222, x = var_2221_cast_fp16)[name = tensor<string, []>("transpose_98")];
            tensor<fp16, [1, 1, 32, 96]> cache_s1_l0_nonlin_attn_to_fp16 = cast(dtype = cache_s1_l0_nonlin_attn_to_fp16_dtype_0, x = cache_s1_l0_nonlin_attn)[name = tensor<string, []>("cast_28")];
            tensor<fp16, [1, 1, 40, 96]> x_pad0_1_cast_fp16 = concat(axis = var_49, interleave = x_pad0_1_interleave_0, values = (cache_s1_l0_nonlin_attn_to_fp16, x33_1_cast_fp16))[name = tensor<string, []>("x_pad0_1_cast_fp16")];
            tensor<int32, [4]> var_2228_begin_0 = const()[name = tensor<string, []>("op_2228_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_2228_end_0 = const()[name = tensor<string, []>("op_2228_end_0"), val = tensor<int32, [4]>([1, 1, 40, 96])];
            tensor<bool, [4]> var_2228_end_mask_0 = const()[name = tensor<string, []>("op_2228_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 32, 96]> new_cache_s1_l0_nonlin_attn = slice_by_index(begin = var_2228_begin_0, end = var_2228_end_0, end_mask = var_2228_end_mask_0, x = x_pad0_1_cast_fp16)[name = tensor<string, []>("op_2228_cast_fp16")];
            tensor<bool, []> x34_1_transpose_x_0 = const()[name = tensor<string, []>("x34_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x34_1_transpose_y_0 = const()[name = tensor<string, []>("x34_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 8, 96]> x34_1_cast_fp16 = matmul(transpose_x = x34_1_transpose_x_0, transpose_y = x34_1_transpose_y_0, x = attn_weights2_1_cast_fp16, y = x_pad0_1_cast_fp16)[name = tensor<string, []>("x34_1_cast_fp16")];
            tensor<int32, [4]> var_2231 = const()[name = tensor<string, []>("op_2231"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_2233 = const()[name = tensor<string, []>("op_2233"), val = tensor<int32, [3]>([8, 1, -1])];
            tensor<fp16, [8, 1, 1, 96]> var_2232_cast_fp16 = transpose(perm = var_2231, x = x34_1_cast_fp16)[name = tensor<string, []>("transpose_97")];
            tensor<fp16, [8, 1, 96]> x35_1_cast_fp16 = reshape(shape = var_2233, x = var_2232_cast_fp16)[name = tensor<string, []>("x35_1_cast_fp16")];
            tensor<fp16, [8, 1, 96]> input17_1_cast_fp16 = mul(x = x35_1_cast_fp16, y = var_2201_cast_fp16_2)[name = tensor<string, []>("input17_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_1_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(979648))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(992000))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(992576)))];
            tensor<fp16, [8, 1, 128]> linear_24_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized, x = input17_1_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x36_1_cast_fp16 = add(x = input15_1_cast_fp16, y = linear_24_cast_fp16)[name = tensor<string, []>("x36_1_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_1_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(992896))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(999104))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_1_encoder_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(999680)))];
            tensor<fp16, [8, 1, 48]> linear_25_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_self_attn1_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized, x = x36_1_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<bool, []> x37_1_interleave_0 = const()[name = tensor<string, []>("x37_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s1_l0_val1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s1_l0_val1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [32, 1, 48]> cache_s1_l0_val1_to_fp16 = cast(dtype = cache_s1_l0_val1_to_fp16_dtype_0, x = cache_s1_l0_val1)[name = tensor<string, []>("cast_27")];
            tensor<fp16, [40, 1, 48]> x37_1_cast_fp16 = concat(axis = var_46, interleave = x37_1_interleave_0, values = (cache_s1_l0_val1_to_fp16, linear_25_cast_fp16))[name = tensor<string, []>("x37_1_cast_fp16")];
            tensor<int32, [3]> var_2258_begin_0 = const()[name = tensor<string, []>("op_2258_begin_0"), val = tensor<int32, [3]>([8, 0, 0])];
            tensor<int32, [3]> var_2258_end_0 = const()[name = tensor<string, []>("op_2258_end_0"), val = tensor<int32, [3]>([40, 1, 48])];
            tensor<bool, [3]> var_2258_end_mask_0 = const()[name = tensor<string, []>("op_2258_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [32, 1, 48]> new_cache_s1_l0_val1 = slice_by_index(begin = var_2258_begin_0, end = var_2258_end_0, end_mask = var_2258_end_mask_0, x = x37_1_cast_fp16)[name = tensor<string, []>("op_2258_cast_fp16")];
            tensor<int32, [4]> var_2259 = const()[name = tensor<string, []>("op_2259"), val = tensor<int32, [4]>([40, 1, 4, -1])];
            tensor<fp16, [40, 1, 4, 12]> var_2260_cast_fp16 = reshape(shape = var_2259, x = x37_1_cast_fp16)[name = tensor<string, []>("op_2260_cast_fp16")];
            tensor<int32, [4]> var_2261 = const()[name = tensor<string, []>("op_2261"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x39_1_transpose_x_0 = const()[name = tensor<string, []>("x39_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x39_1_transpose_y_0 = const()[name = tensor<string, []>("x39_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 40, 12]> x38_1_cast_fp16 = transpose(perm = var_2261, x = var_2260_cast_fp16)[name = tensor<string, []>("transpose_96")];
            tensor<fp16, [4, 1, 8, 12]> x39_1_cast_fp16 = matmul(transpose_x = x39_1_transpose_x_0, transpose_y = x39_1_transpose_y_0, x = attn_weights1_1_cast_fp16, y = x38_1_cast_fp16)[name = tensor<string, []>("x39_1_cast_fp16")];
            tensor<int32, [4]> var_2266 = const()[name = tensor<string, []>("op_2266"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_2271 = const()[name = tensor<string, []>("op_2271"), val = tensor<int32, [3]>([8, 1, 48])];
            tensor<fp16, [8, 1, 4, 12]> var_2267_cast_fp16 = transpose(perm = var_2266, x = x39_1_cast_fp16)[name = tensor<string, []>("transpose_95")];
            tensor<fp16, [8, 1, 48]> input18_1_cast_fp16 = reshape(shape = var_2271, x = var_2267_cast_fp16)[name = tensor<string, []>("input18_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_1_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(999872))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1006080))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1006656)))];
            tensor<fp16, [8, 1, 128]> linear_26_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_self_attn1_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized, x = input18_1_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input19_1_cast_fp16 = add(x = x36_1_cast_fp16, y = linear_26_cast_fp16)[name = tensor<string, []>("input19_1_cast_fp16")];
            tensor<int32, [2]> src_key_padding_mask3_1_begin_0 = const()[name = tensor<string, []>("src_key_padding_mask3_1_begin_0"), val = tensor<int32, [2]>([0, 32])];
            tensor<int32, [2]> src_key_padding_mask3_1_end_0 = const()[name = tensor<string, []>("src_key_padding_mask3_1_end_0"), val = tensor<int32, [2]>([1, 40])];
            tensor<bool, [2]> src_key_padding_mask3_1_end_mask_0 = const()[name = tensor<string, []>("src_key_padding_mask3_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<bool, [1, 8]> src_key_padding_mask3_1 = slice_by_index(begin = src_key_padding_mask3_1_begin_0, end = src_key_padding_mask3_1_end_0, end_mask = src_key_padding_mask3_1_end_mask_0, x = key_padding_mask0_1)[name = tensor<string, []>("src_key_padding_mask3_1")];
            tensor<fp16, [256, 128]> encoder_encoders_1_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1006976))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1039808))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_1_encoder_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1040384)))];
            tensor<fp16, [8, 1, 256]> linear_27_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_conv_module1_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized, x = input19_1_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<int32, [2]> var_2282_split_sizes_0 = const()[name = tensor<string, []>("op_2282_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_2282_axis_0 = const()[name = tensor<string, []>("op_2282_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [8, 1, 128]> var_2282_cast_fp16_0, tensor<fp16, [8, 1, 128]> var_2282_cast_fp16_1 = split(axis = var_2282_axis_0, split_sizes = var_2282_split_sizes_0, x = linear_27_cast_fp16)[name = tensor<string, []>("op_2282_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2284_cast_fp16 = sigmoid(x = var_2282_cast_fp16_1)[name = tensor<string, []>("op_2284_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x41_1_cast_fp16 = mul(x = var_2282_cast_fp16_0, y = var_2284_cast_fp16)[name = tensor<string, []>("x41_1_cast_fp16")];
            tensor<int32, [3]> var_2286 = const()[name = tensor<string, []>("op_2286"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_2288_axes_0 = const()[name = tensor<string, []>("op_2288_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 8]> var_2288 = expand_dims(axes = var_2288_axes_0, x = src_key_padding_mask3_1)[name = tensor<string, []>("op_2288")];
            tensor<int32, [3]> var_2289_reps_0 = const()[name = tensor<string, []>("op_2289_reps_0"), val = tensor<int32, [3]>([1, 128, 1])];
            tensor<bool, [1, 128, 8]> var_2289 = tile(reps = var_2289_reps_0, x = var_2288)[name = tensor<string, []>("op_2289")];
            tensor<fp16, [1, 128, 8]> x42_1_cast_fp16 = transpose(perm = var_2286, x = x41_1_cast_fp16)[name = tensor<string, []>("transpose_94")];
            tensor<fp16, [1, 128, 8]> x43_1_cast_fp16 = select(a = var_62_to_fp16, b = x42_1_cast_fp16, cond = var_2289)[name = tensor<string, []>("x43_1_cast_fp16")];
            tensor<bool, []> x44_1_interleave_0 = const()[name = tensor<string, []>("x44_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s1_l0_conv1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s1_l0_conv1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 15]> cache_s1_l0_conv1_to_fp16 = cast(dtype = cache_s1_l0_conv1_to_fp16_dtype_0, x = cache_s1_l0_conv1)[name = tensor<string, []>("cast_26")];
            tensor<fp16, [1, 128, 23]> x44_1_cast_fp16 = concat(axis = var_49, interleave = x44_1_interleave_0, values = (cache_s1_l0_conv1_to_fp16, x43_1_cast_fp16))[name = tensor<string, []>("x44_1_cast_fp16")];
            tensor<int32, [3]> var_2296_begin_0 = const()[name = tensor<string, []>("op_2296_begin_0"), val = tensor<int32, [3]>([0, 0, 8])];
            tensor<int32, [3]> var_2296_end_0 = const()[name = tensor<string, []>("op_2296_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> var_2296_end_mask_0 = const()[name = tensor<string, []>("op_2296_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 15]> new_cache_s1_l0_conv1 = slice_by_index(begin = var_2296_begin_0, end = var_2296_end_0, end_mask = var_2296_end_mask_0, x = x44_1_cast_fp16)[name = tensor<string, []>("op_2296_cast_fp16")];
            tensor<string, []> x_causal_6_pad_type_0 = const()[name = tensor<string, []>("x_causal_6_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_6_groups_0 = const()[name = tensor<string, []>("x_causal_6_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_6_strides_0 = const()[name = tensor<string, []>("x_causal_6_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_6_pad_0 = const()[name = tensor<string, []>("x_causal_6_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_6_dilations_0 = const()[name = tensor<string, []>("x_causal_6_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 16]> encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1040960)))];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1045120)))];
            tensor<fp16, [1, 128, 8]> x_causal_6_cast_fp16 = conv(bias = encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_6_dilations_0, groups = x_causal_6_groups_0, pad = x_causal_6_pad_0, pad_type = x_causal_6_pad_type_0, strides = x_causal_6_strides_0, weight = encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16, x = x44_1_cast_fp16)[name = tensor<string, []>("x_causal_6_cast_fp16")];
            tensor<int32, [3]> input21_1_begin_0 = const()[name = tensor<string, []>("input21_1_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> input21_1_end_0 = const()[name = tensor<string, []>("input21_1_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> input21_1_end_mask_0 = const()[name = tensor<string, []>("input21_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 8]> input21_1_cast_fp16 = slice_by_index(begin = input21_1_begin_0, end = input21_1_end_0, end_mask = input21_1_end_mask_0, x = x44_1_cast_fp16)[name = tensor<string, []>("input21_1_cast_fp16")];
            tensor<string, []> x_chunk_9_pad_type_0 = const()[name = tensor<string, []>("x_chunk_9_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_9_pad_0 = const()[name = tensor<string, []>("x_chunk_9_pad_0"), val = tensor<int32, [2]>([15, 15])];
            tensor<int32, []> x_chunk_9_groups_0 = const()[name = tensor<string, []>("x_chunk_9_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_9_strides_0 = const()[name = tensor<string, []>("x_chunk_9_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_9_dilations_0 = const()[name = tensor<string, []>("x_chunk_9_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 31]> encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3968]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1045440))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1049472))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 1, 31])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050048)))];
            tensor<fp16, [1, 128, 8]> x_chunk_9_cast_fp16 = conv(bias = encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_9_dilations_0, groups = x_chunk_9_groups_0, pad = x_chunk_9_pad_0, pad_type = x_chunk_9_pad_type_0, strides = x_chunk_9_strides_0, weight = encoder_encoders_1_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized, x = input21_1_cast_fp16)[name = tensor<string, []>("x_chunk_9_cast_fp16")];
            tensor<fp16, [128, 8]> chunk_scale1_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale1_1_to_fp16"), val = tensor<fp16, [128, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1050368)))];
            tensor<fp16, [1, 128, 8]> x_chunk1_1_cast_fp16 = mul(x = x_chunk_9_cast_fp16, y = chunk_scale1_1_to_fp16)[name = tensor<string, []>("x_chunk1_1_cast_fp16")];
            tensor<fp16, [1, 128, 8]> x45_1_cast_fp16 = add(x = x_chunk1_1_cast_fp16, y = x_causal_6_cast_fp16)[name = tensor<string, []>("x45_1_cast_fp16")];
            tensor<int32, [3]> var_2336 = const()[name = tensor<string, []>("op_2336"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_2345_to_fp16 = const()[name = tensor<string, []>("op_2345_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> x46_1_cast_fp16 = transpose(perm = var_2336, x = x45_1_cast_fp16)[name = tensor<string, []>("transpose_93")];
            tensor<fp16, [8, 1, 128]> x_offset_14_cast_fp16 = sub(x = x46_1_cast_fp16, y = var_2345_to_fp16)[name = tensor<string, []>("x_offset_14_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2347_cast_fp16 = exp(x = x_offset_14_cast_fp16)[name = tensor<string, []>("op_2347_cast_fp16")];
            tensor<fp16, []> var_2348_to_fp16 = const()[name = tensor<string, []>("op_2348_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> var_2349_cast_fp16 = add(x = var_2347_cast_fp16, y = var_2348_to_fp16)[name = tensor<string, []>("op_2349_cast_fp16")];
            tensor<fp32, []> var_2350_epsilon_0 = const()[name = tensor<string, []>("op_2350_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 128]> var_2350_cast_fp16 = log(epsilon = var_2350_epsilon_0, x = var_2349_cast_fp16)[name = tensor<string, []>("op_2350_cast_fp16")];
            tensor<fp16, []> var_2342_to_fp16 = const()[name = tensor<string, []>("op_2342_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 128]> var_2352_cast_fp16 = equal(x = var_2350_cast_fp16, y = var_2342_to_fp16)[name = tensor<string, []>("op_2352_cast_fp16")];
            tensor<fp16, [8, 1, 128]> log_sum0_14_cast_fp16 = select(a = x_offset_14_cast_fp16, b = var_2350_cast_fp16, cond = var_2352_cast_fp16)[name = tensor<string, []>("log_sum0_14_cast_fp16")];
            tensor<fp16, []> var_2354_to_fp16 = const()[name = tensor<string, []>("op_2354_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 128]> var_2355_cast_fp16 = mul(x = x46_1_cast_fp16, y = var_2354_to_fp16)[name = tensor<string, []>("op_2355_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2356_cast_fp16 = sub(x = log_sum0_14_cast_fp16, y = var_2355_cast_fp16)[name = tensor<string, []>("op_2356_cast_fp16")];
            tensor<fp16, []> var_2357_to_fp16 = const()[name = tensor<string, []>("op_2357_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [8, 1, 128]> x0_13_cast_fp16 = sub(x = var_2356_cast_fp16, y = var_2357_to_fp16)[name = tensor<string, []>("x0_13_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_1_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1052480))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1068928))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1069504)))];
            tensor<fp16, [8, 1, 128]> linear_28_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_conv_module1_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized, x = x0_13_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input22_1_cast_fp16 = add(x = input19_1_cast_fp16, y = linear_28_cast_fp16)[name = tensor<string, []>("input22_1_cast_fp16")];
            tensor<fp16, [192, 128]> encoder_encoders_1_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1069824))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1094464))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([192, 128])];
            tensor<fp16, [192]> encoder_encoders_1_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1095040)))];
            tensor<fp16, [8, 1, 192]> linear_29_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized, x = input22_1_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<fp16, []> var_2373_to_fp16 = const()[name = tensor<string, []>("op_2373_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [8, 1, 192]> x_offset_16_cast_fp16 = sub(x = linear_29_cast_fp16, y = var_2373_to_fp16)[name = tensor<string, []>("x_offset_16_cast_fp16")];
            tensor<fp16, [8, 1, 192]> var_2375_cast_fp16 = exp(x = x_offset_16_cast_fp16)[name = tensor<string, []>("op_2375_cast_fp16")];
            tensor<fp16, []> var_2376_to_fp16 = const()[name = tensor<string, []>("op_2376_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 192]> var_2377_cast_fp16 = add(x = var_2375_cast_fp16, y = var_2376_to_fp16)[name = tensor<string, []>("op_2377_cast_fp16")];
            tensor<fp32, []> var_2378_epsilon_0 = const()[name = tensor<string, []>("op_2378_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 192]> var_2378_cast_fp16 = log(epsilon = var_2378_epsilon_0, x = var_2377_cast_fp16)[name = tensor<string, []>("op_2378_cast_fp16")];
            tensor<fp16, []> var_2361_to_fp16 = const()[name = tensor<string, []>("op_2361_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 192]> var_2380_cast_fp16 = equal(x = var_2378_cast_fp16, y = var_2361_to_fp16)[name = tensor<string, []>("op_2380_cast_fp16")];
            tensor<fp16, [8, 1, 192]> log_sum0_16_cast_fp16 = select(a = x_offset_16_cast_fp16, b = var_2378_cast_fp16, cond = var_2380_cast_fp16)[name = tensor<string, []>("log_sum0_16_cast_fp16")];
            tensor<fp16, []> var_2382_to_fp16 = const()[name = tensor<string, []>("op_2382_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 192]> var_2383_cast_fp16 = mul(x = linear_29_cast_fp16, y = var_2382_to_fp16)[name = tensor<string, []>("op_2383_cast_fp16")];
            tensor<fp16, [8, 1, 192]> var_2384_cast_fp16 = sub(x = log_sum0_16_cast_fp16, y = var_2383_cast_fp16)[name = tensor<string, []>("op_2384_cast_fp16")];
            tensor<fp16, []> var_2385_to_fp16 = const()[name = tensor<string, []>("op_2385_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [8, 1, 192]> x_31_cast_fp16 = sub(x = var_2384_cast_fp16, y = var_2385_to_fp16)[name = tensor<string, []>("x_31_cast_fp16")];
            tensor<fp16, [128, 192]> encoder_encoders_1_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1095488))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1120128))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 192])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1120704)))];
            tensor<fp16, [8, 1, 128]> linear_30_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized, x = x_31_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<fp16, [8, 1, 128]> src1_1_cast_fp16 = add(x = input22_1_cast_fp16, y = linear_30_cast_fp16)[name = tensor<string, []>("src1_1_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2391_cast_fp16 = sub(x = src1_1_cast_fp16, y = x_23_cast_fp16)[name = tensor<string, []>("op_2391_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1121024)))];
            tensor<fp16, [8, 1, 128]> var_2392_cast_fp16 = mul(x = var_2391_cast_fp16, y = encoder_encoders_1_encoder_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_2392_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2393_cast_fp16 = add(x = x_23_cast_fp16, y = var_2392_cast_fp16)[name = tensor<string, []>("op_2393_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_1_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1121344))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1127552))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_1_encoder_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1128128)))];
            tensor<fp16, [8, 1, 48]> linear_31_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_self_attn2_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized, x = var_2393_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<bool, []> x47_1_interleave_0 = const()[name = tensor<string, []>("x47_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s1_l0_val2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s1_l0_val2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [32, 1, 48]> cache_s1_l0_val2_to_fp16 = cast(dtype = cache_s1_l0_val2_to_fp16_dtype_0, x = cache_s1_l0_val2)[name = tensor<string, []>("cast_25")];
            tensor<fp16, [40, 1, 48]> x47_1_cast_fp16 = concat(axis = var_46, interleave = x47_1_interleave_0, values = (cache_s1_l0_val2_to_fp16, linear_31_cast_fp16))[name = tensor<string, []>("x47_1_cast_fp16")];
            tensor<int32, [3]> var_2412_begin_0 = const()[name = tensor<string, []>("op_2412_begin_0"), val = tensor<int32, [3]>([8, 0, 0])];
            tensor<int32, [3]> var_2412_end_0 = const()[name = tensor<string, []>("op_2412_end_0"), val = tensor<int32, [3]>([40, 1, 48])];
            tensor<bool, [3]> var_2412_end_mask_0 = const()[name = tensor<string, []>("op_2412_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [32, 1, 48]> new_cache_s1_l0_val2 = slice_by_index(begin = var_2412_begin_0, end = var_2412_end_0, end_mask = var_2412_end_mask_0, x = x47_1_cast_fp16)[name = tensor<string, []>("op_2412_cast_fp16")];
            tensor<int32, [4]> var_2413 = const()[name = tensor<string, []>("op_2413"), val = tensor<int32, [4]>([40, 1, 4, -1])];
            tensor<fp16, [40, 1, 4, 12]> var_2414_cast_fp16 = reshape(shape = var_2413, x = x47_1_cast_fp16)[name = tensor<string, []>("op_2414_cast_fp16")];
            tensor<int32, [4]> var_2415 = const()[name = tensor<string, []>("op_2415"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x49_1_transpose_x_0 = const()[name = tensor<string, []>("x49_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x49_1_transpose_y_0 = const()[name = tensor<string, []>("x49_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 40, 12]> x48_1_cast_fp16 = transpose(perm = var_2415, x = var_2414_cast_fp16)[name = tensor<string, []>("transpose_92")];
            tensor<fp16, [4, 1, 8, 12]> x49_1_cast_fp16 = matmul(transpose_x = x49_1_transpose_x_0, transpose_y = x49_1_transpose_y_0, x = attn_weights1_1_cast_fp16, y = x48_1_cast_fp16)[name = tensor<string, []>("x49_1_cast_fp16")];
            tensor<int32, [4]> var_2420 = const()[name = tensor<string, []>("op_2420"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_2425 = const()[name = tensor<string, []>("op_2425"), val = tensor<int32, [3]>([8, 1, 48])];
            tensor<fp16, [8, 1, 4, 12]> var_2421_cast_fp16 = transpose(perm = var_2420, x = x49_1_cast_fp16)[name = tensor<string, []>("transpose_91")];
            tensor<fp16, [8, 1, 48]> input23_1_cast_fp16 = reshape(shape = var_2425, x = var_2421_cast_fp16)[name = tensor<string, []>("input23_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_1_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1128320))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1134528))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1135104)))];
            tensor<fp16, [8, 1, 128]> linear_32_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_self_attn2_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized, x = input23_1_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input24_1_cast_fp16 = add(x = var_2393_cast_fp16, y = linear_32_cast_fp16)[name = tensor<string, []>("input24_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_1_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1135424))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168256))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_1_encoder_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1168832)))];
            tensor<fp16, [8, 1, 256]> linear_33_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_conv_module2_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized, x = input24_1_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<int32, [2]> var_2436_split_sizes_0 = const()[name = tensor<string, []>("op_2436_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_2436_axis_0 = const()[name = tensor<string, []>("op_2436_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [8, 1, 128]> var_2436_cast_fp16_0, tensor<fp16, [8, 1, 128]> var_2436_cast_fp16_1 = split(axis = var_2436_axis_0, split_sizes = var_2436_split_sizes_0, x = linear_33_cast_fp16)[name = tensor<string, []>("op_2436_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2438_cast_fp16 = sigmoid(x = var_2436_cast_fp16_1)[name = tensor<string, []>("op_2438_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x51_1_cast_fp16 = mul(x = var_2436_cast_fp16_0, y = var_2438_cast_fp16)[name = tensor<string, []>("x51_1_cast_fp16")];
            tensor<int32, [3]> var_2440 = const()[name = tensor<string, []>("op_2440"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 8]> x52_1_cast_fp16 = transpose(perm = var_2440, x = x51_1_cast_fp16)[name = tensor<string, []>("transpose_90")];
            tensor<fp16, [1, 128, 8]> x53_1_cast_fp16 = select(a = var_62_to_fp16, b = x52_1_cast_fp16, cond = var_2289)[name = tensor<string, []>("x53_1_cast_fp16")];
            tensor<bool, []> x54_1_interleave_0 = const()[name = tensor<string, []>("x54_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s1_l0_conv2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s1_l0_conv2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 15]> cache_s1_l0_conv2_to_fp16 = cast(dtype = cache_s1_l0_conv2_to_fp16_dtype_0, x = cache_s1_l0_conv2)[name = tensor<string, []>("cast_24")];
            tensor<fp16, [1, 128, 23]> x54_1_cast_fp16 = concat(axis = var_49, interleave = x54_1_interleave_0, values = (cache_s1_l0_conv2_to_fp16, x53_1_cast_fp16))[name = tensor<string, []>("x54_1_cast_fp16")];
            tensor<int32, [3]> var_2450_begin_0 = const()[name = tensor<string, []>("op_2450_begin_0"), val = tensor<int32, [3]>([0, 0, 8])];
            tensor<int32, [3]> var_2450_end_0 = const()[name = tensor<string, []>("op_2450_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> var_2450_end_mask_0 = const()[name = tensor<string, []>("op_2450_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 15]> new_cache_s1_l0_conv2 = slice_by_index(begin = var_2450_begin_0, end = var_2450_end_0, end_mask = var_2450_end_mask_0, x = x54_1_cast_fp16)[name = tensor<string, []>("op_2450_cast_fp16")];
            tensor<string, []> x_causal_8_pad_type_0 = const()[name = tensor<string, []>("x_causal_8_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_8_groups_0 = const()[name = tensor<string, []>("x_causal_8_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_8_strides_0 = const()[name = tensor<string, []>("x_causal_8_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_8_pad_0 = const()[name = tensor<string, []>("x_causal_8_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_8_dilations_0 = const()[name = tensor<string, []>("x_causal_8_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 16]> encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1169408)))];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1173568)))];
            tensor<fp16, [1, 128, 8]> x_causal_8_cast_fp16 = conv(bias = encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_8_dilations_0, groups = x_causal_8_groups_0, pad = x_causal_8_pad_0, pad_type = x_causal_8_pad_type_0, strides = x_causal_8_strides_0, weight = encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16, x = x54_1_cast_fp16)[name = tensor<string, []>("x_causal_8_cast_fp16")];
            tensor<int32, [3]> input26_1_begin_0 = const()[name = tensor<string, []>("input26_1_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> input26_1_end_0 = const()[name = tensor<string, []>("input26_1_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> input26_1_end_mask_0 = const()[name = tensor<string, []>("input26_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 8]> input26_1_cast_fp16 = slice_by_index(begin = input26_1_begin_0, end = input26_1_end_0, end_mask = input26_1_end_mask_0, x = x54_1_cast_fp16)[name = tensor<string, []>("input26_1_cast_fp16")];
            tensor<string, []> x_chunk_11_pad_type_0 = const()[name = tensor<string, []>("x_chunk_11_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_11_pad_0 = const()[name = tensor<string, []>("x_chunk_11_pad_0"), val = tensor<int32, [2]>([15, 15])];
            tensor<int32, []> x_chunk_11_groups_0 = const()[name = tensor<string, []>("x_chunk_11_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_11_strides_0 = const()[name = tensor<string, []>("x_chunk_11_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_11_dilations_0 = const()[name = tensor<string, []>("x_chunk_11_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 31]> encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3968]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1173888))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1177920))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 1, 31])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1178496)))];
            tensor<fp16, [1, 128, 8]> x_chunk_11_cast_fp16 = conv(bias = encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_11_dilations_0, groups = x_chunk_11_groups_0, pad = x_chunk_11_pad_0, pad_type = x_chunk_11_pad_type_0, strides = x_chunk_11_strides_0, weight = encoder_encoders_1_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized, x = input26_1_cast_fp16)[name = tensor<string, []>("x_chunk_11_cast_fp16")];
            tensor<fp16, [128, 8]> chunk_scale2_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale2_1_to_fp16"), val = tensor<fp16, [128, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1178816)))];
            tensor<fp16, [1, 128, 8]> x_chunk2_1_cast_fp16 = mul(x = x_chunk_11_cast_fp16, y = chunk_scale2_1_to_fp16)[name = tensor<string, []>("x_chunk2_1_cast_fp16")];
            tensor<fp16, [1, 128, 8]> x55_1_cast_fp16 = add(x = x_chunk2_1_cast_fp16, y = x_causal_8_cast_fp16)[name = tensor<string, []>("x55_1_cast_fp16")];
            tensor<int32, [3]> var_2490 = const()[name = tensor<string, []>("op_2490"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_2499_to_fp16 = const()[name = tensor<string, []>("op_2499_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> x56_1_cast_fp16 = transpose(perm = var_2490, x = x55_1_cast_fp16)[name = tensor<string, []>("transpose_89")];
            tensor<fp16, [8, 1, 128]> x_offset_18_cast_fp16 = sub(x = x56_1_cast_fp16, y = var_2499_to_fp16)[name = tensor<string, []>("x_offset_18_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2501_cast_fp16 = exp(x = x_offset_18_cast_fp16)[name = tensor<string, []>("op_2501_cast_fp16")];
            tensor<fp16, []> var_2502_to_fp16 = const()[name = tensor<string, []>("op_2502_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> var_2503_cast_fp16 = add(x = var_2501_cast_fp16, y = var_2502_to_fp16)[name = tensor<string, []>("op_2503_cast_fp16")];
            tensor<fp32, []> var_2504_epsilon_0 = const()[name = tensor<string, []>("op_2504_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 128]> var_2504_cast_fp16 = log(epsilon = var_2504_epsilon_0, x = var_2503_cast_fp16)[name = tensor<string, []>("op_2504_cast_fp16")];
            tensor<fp16, []> var_2496_to_fp16 = const()[name = tensor<string, []>("op_2496_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 128]> var_2506_cast_fp16 = equal(x = var_2504_cast_fp16, y = var_2496_to_fp16)[name = tensor<string, []>("op_2506_cast_fp16")];
            tensor<fp16, [8, 1, 128]> log_sum0_18_cast_fp16 = select(a = x_offset_18_cast_fp16, b = var_2504_cast_fp16, cond = var_2506_cast_fp16)[name = tensor<string, []>("log_sum0_18_cast_fp16")];
            tensor<fp16, []> var_2508_to_fp16 = const()[name = tensor<string, []>("op_2508_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 128]> var_2509_cast_fp16 = mul(x = x56_1_cast_fp16, y = var_2508_to_fp16)[name = tensor<string, []>("op_2509_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2510_cast_fp16 = sub(x = log_sum0_18_cast_fp16, y = var_2509_cast_fp16)[name = tensor<string, []>("op_2510_cast_fp16")];
            tensor<fp16, []> var_2511_to_fp16 = const()[name = tensor<string, []>("op_2511_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [8, 1, 128]> x0_15_cast_fp16 = sub(x = var_2510_cast_fp16, y = var_2511_to_fp16)[name = tensor<string, []>("x0_15_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_1_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1180928))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1197376))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1197952)))];
            tensor<fp16, [8, 1, 128]> linear_34_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_conv_module2_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized, x = x0_15_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input27_1_cast_fp16 = add(x = input24_1_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("input27_1_cast_fp16")];
            tensor<fp16, [240, 128]> encoder_encoders_1_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1198272))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1229056))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([240, 128])];
            tensor<fp16, [240]> encoder_encoders_1_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1229632)))];
            tensor<fp16, [8, 1, 240]> linear_35_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized, x = input27_1_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<fp16, []> var_2527_to_fp16 = const()[name = tensor<string, []>("op_2527_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [8, 1, 240]> x_offset_20_cast_fp16 = sub(x = linear_35_cast_fp16, y = var_2527_to_fp16)[name = tensor<string, []>("x_offset_20_cast_fp16")];
            tensor<fp16, [8, 1, 240]> var_2529_cast_fp16 = exp(x = x_offset_20_cast_fp16)[name = tensor<string, []>("op_2529_cast_fp16")];
            tensor<fp16, []> var_2530_to_fp16 = const()[name = tensor<string, []>("op_2530_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 240]> var_2531_cast_fp16 = add(x = var_2529_cast_fp16, y = var_2530_to_fp16)[name = tensor<string, []>("op_2531_cast_fp16")];
            tensor<fp32, []> var_2532_epsilon_0 = const()[name = tensor<string, []>("op_2532_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 240]> var_2532_cast_fp16 = log(epsilon = var_2532_epsilon_0, x = var_2531_cast_fp16)[name = tensor<string, []>("op_2532_cast_fp16")];
            tensor<fp16, []> var_2515_to_fp16 = const()[name = tensor<string, []>("op_2515_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 240]> var_2534_cast_fp16 = equal(x = var_2532_cast_fp16, y = var_2515_to_fp16)[name = tensor<string, []>("op_2534_cast_fp16")];
            tensor<fp16, [8, 1, 240]> log_sum0_20_cast_fp16 = select(a = x_offset_20_cast_fp16, b = var_2532_cast_fp16, cond = var_2534_cast_fp16)[name = tensor<string, []>("log_sum0_20_cast_fp16")];
            tensor<fp16, []> var_2536_to_fp16 = const()[name = tensor<string, []>("op_2536_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 240]> var_2537_cast_fp16 = mul(x = linear_35_cast_fp16, y = var_2536_to_fp16)[name = tensor<string, []>("op_2537_cast_fp16")];
            tensor<fp16, [8, 1, 240]> var_2538_cast_fp16 = sub(x = log_sum0_20_cast_fp16, y = var_2537_cast_fp16)[name = tensor<string, []>("op_2538_cast_fp16")];
            tensor<fp16, []> var_2539_to_fp16 = const()[name = tensor<string, []>("op_2539_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [8, 1, 240]> x_35_cast_fp16 = sub(x = var_2538_cast_fp16, y = var_2539_to_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<fp16, [128, 240]> encoder_encoders_1_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1230208))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1260992))), name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 240])];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1261568)))];
            tensor<fp16, [8, 1, 128]> linear_36_cast_fp16 = linear(bias = encoder_encoders_1_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16, weight = encoder_encoders_1_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized, x = x_35_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x57_1_cast_fp16 = add(x = input27_1_cast_fp16, y = linear_36_cast_fp16)[name = tensor<string, []>("x57_1_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1261888)))];
            tensor<fp16, [8, 1, 128]> var_2550_cast_fp16 = sub(x = x57_1_cast_fp16, y = encoder_encoders_1_encoder_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("op_2550_cast_fp16")];
            tensor<fp16, []> var_2544_promoted_to_fp16 = const()[name = tensor<string, []>("op_2544_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [8, 1, 128]> var_2551_cast_fp16 = pow(x = var_2550_cast_fp16, y = var_2544_promoted_to_fp16)[name = tensor<string, []>("op_2551_cast_fp16")];
            tensor<int32, [1]> var_2553_axes_0 = const()[name = tensor<string, []>("op_2553_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, []> var_2553_keep_dims_0 = const()[name = tensor<string, []>("op_2553_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [8, 1, 1]> var_2553_cast_fp16 = reduce_mean(axes = var_2553_axes_0, keep_dims = var_2553_keep_dims_0, x = var_2551_cast_fp16)[name = tensor<string, []>("op_2553_cast_fp16")];
            tensor<fp16, []> var_2543_to_fp16 = const()[name = tensor<string, []>("op_2543_to_fp16"), val = tensor<fp16, []>(-0x1p-1)];
            tensor<fp16, [8, 1, 1]> var_2554_cast_fp16 = pow(x = var_2553_cast_fp16, y = var_2543_to_fp16)[name = tensor<string, []>("op_2554_cast_fp16")];
            tensor<fp16, []> var_2555_to_fp16 = const()[name = tensor<string, []>("op_2555_to_fp16"), val = tensor<fp16, []>(0x1.1a8p-2)];
            tensor<fp16, [8, 1, 1]> scales_6_cast_fp16 = mul(x = var_2554_cast_fp16, y = var_2555_to_fp16)[name = tensor<string, []>("scales_6_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2557_cast_fp16 = mul(x = x57_1_cast_fp16, y = scales_6_cast_fp16)[name = tensor<string, []>("op_2557_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2560_cast_fp16 = sub(x = var_2557_cast_fp16, y = x_23_cast_fp16)[name = tensor<string, []>("op_2560_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_1_encoder_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_encoder_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1262208)))];
            tensor<fp16, [8, 1, 128]> var_2561_cast_fp16 = mul(x = var_2560_cast_fp16, y = encoder_encoders_1_encoder_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_2561_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_2562_cast_fp16 = add(x = x_23_cast_fp16, y = var_2561_cast_fp16)[name = tensor<string, []>("op_2562_cast_fp16")];
            tensor<int32, [1]> var_2571_axes_0 = const()[name = tensor<string, []>("op_2571_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [8, 1, 1, 128]> var_2571_cast_fp16 = expand_dims(axes = var_2571_axes_0, x = var_2562_cast_fp16)[name = tensor<string, []>("op_2571_cast_fp16")];
            tensor<int32, [4]> src_5_reps_0 = const()[name = tensor<string, []>("src_5_reps_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [8, 2, 1, 128]> src_5_cast_fp16 = tile(reps = src_5_reps_0, x = var_2571_cast_fp16)[name = tensor<string, []>("src_5_cast_fp16")];
            tensor<int32, [3]> var_2577 = const()[name = tensor<string, []>("op_2577"), val = tensor<int32, [3]>([16, 1, 128])];
            tensor<fp16, [16, 1, 128]> var_2578_cast_fp16 = reshape(shape = var_2577, x = src_5_cast_fp16)[name = tensor<string, []>("op_2578_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_2585_cast_fp16 = sub(x = var_2578_cast_fp16, y = var_2018_cast_fp16)[name = tensor<string, []>("op_2585_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_1_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_1_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1262528)))];
            tensor<fp16, [16, 1, 128]> var_2586_cast_fp16 = mul(x = var_2585_cast_fp16, y = encoder_encoders_1_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_2586_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_2587_cast_fp16 = add(x = var_2018_cast_fp16, y = var_2586_cast_fp16)[name = tensor<string, []>("op_2587_cast_fp16")];
            tensor<int32, [2]> key_padding_mask1_1_begin_0 = const()[name = tensor<string, []>("key_padding_mask1_1_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> key_padding_mask1_1_end_0 = const()[name = tensor<string, []>("key_padding_mask1_1_end_0"), val = tensor<int32, [2]>([1, 80])];
            tensor<bool, [2]> key_padding_mask1_1_end_mask_0 = const()[name = tensor<string, []>("key_padding_mask1_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<int32, [2]> key_padding_mask1_1_stride_0 = const()[name = tensor<string, []>("key_padding_mask1_1_stride_0"), val = tensor<int32, [2]>([1, 4])];
            tensor<bool, [1, 20]> key_padding_mask1_1 = slice_by_index(begin = key_padding_mask1_1_begin_0, end = key_padding_mask1_1_end_0, end_mask = key_padding_mask1_1_end_mask_0, stride = key_padding_mask1_1_stride_0, x = src_key_padding_mask0_1)[name = tensor<string, []>("key_padding_mask1_1")];
            tensor<int32, [4]> var_2609 = const()[name = tensor<string, []>("op_2609"), val = tensor<int32, [4]>([4, 4, 1, 128])];
            tensor<fp16, [4, 4, 1, 128]> src0_11_cast_fp16 = reshape(shape = var_2609, x = var_2587_cast_fp16)[name = tensor<string, []>("src0_11_cast_fp16")];
            tensor<fp16, [4, 1, 1]> weights0_4_to_fp16 = const()[name = tensor<string, []>("weights0_4_to_fp16"), val = tensor<fp16, [4, 1, 1]>([[[0x1.96p-2]], [[0x1.45cp-4]], [[0x1.8b8p-2]], [[0x1.19cp-3]]])];
            tensor<fp16, [4, 4, 1, 128]> var_2614_cast_fp16 = mul(x = src0_11_cast_fp16, y = weights0_4_to_fp16)[name = tensor<string, []>("op_2614_cast_fp16")];
            tensor<int32, [1]> x_37_axes_0 = const()[name = tensor<string, []>("x_37_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> x_37_keep_dims_0 = const()[name = tensor<string, []>("x_37_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 128]> x_37_cast_fp16 = reduce_sum(axes = x_37_axes_0, keep_dims = x_37_keep_dims_0, x = var_2614_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<fp16, [272, 128]> encoder_encoders_2_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [34816]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1262848))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1297728))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([272, 128])];
            tensor<fp16, [272]> encoder_encoders_2_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [272]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1298304)))];
            tensor<fp16, [4, 1, 272]> linear_37_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized, x = x_37_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<int32, [3]> q5_1_begin_0 = const()[name = tensor<string, []>("q5_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> q5_1_end_0 = const()[name = tensor<string, []>("q5_1_end_0"), val = tensor<int32, [3]>([4, 1, 128])];
            tensor<bool, [3]> q5_1_end_mask_0 = const()[name = tensor<string, []>("q5_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [4, 1, 128]> q5_1_cast_fp16 = slice_by_index(begin = q5_1_begin_0, end = q5_1_end_0, end_mask = q5_1_end_mask_0, x = linear_37_cast_fp16)[name = tensor<string, []>("q5_1_cast_fp16")];
            tensor<int32, [3]> k7_1_begin_0 = const()[name = tensor<string, []>("k7_1_begin_0"), val = tensor<int32, [3]>([0, 0, 128])];
            tensor<int32, [3]> k7_1_end_0 = const()[name = tensor<string, []>("k7_1_end_0"), val = tensor<int32, [3]>([4, 1, 256])];
            tensor<bool, [3]> k7_1_end_mask_0 = const()[name = tensor<string, []>("k7_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [4, 1, 128]> k7_1_cast_fp16 = slice_by_index(begin = k7_1_begin_0, end = k7_1_end_0, end_mask = k7_1_end_mask_0, x = linear_37_cast_fp16)[name = tensor<string, []>("k7_1_cast_fp16")];
            tensor<int32, [3]> p5_1_begin_0 = const()[name = tensor<string, []>("p5_1_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> p5_1_end_0 = const()[name = tensor<string, []>("p5_1_end_0"), val = tensor<int32, [3]>([4, 1, 272])];
            tensor<bool, [3]> p5_1_end_mask_0 = const()[name = tensor<string, []>("p5_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [4, 1, 16]> p5_1_cast_fp16 = slice_by_index(begin = p5_1_begin_0, end = p5_1_end_0, end_mask = p5_1_end_mask_0, x = linear_37_cast_fp16)[name = tensor<string, []>("p5_1_cast_fp16")];
            tensor<bool, []> k8_1_interleave_0 = const()[name = tensor<string, []>("k8_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s2_l0_key_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s2_l0_key_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [16, 1, 128]> cache_s2_l0_key_to_fp16 = cast(dtype = cache_s2_l0_key_to_fp16_dtype_0, x = cache_s2_l0_key)[name = tensor<string, []>("cast_23")];
            tensor<fp16, [20, 1, 128]> k8_1_cast_fp16 = concat(axis = var_46, interleave = k8_1_interleave_0, values = (cache_s2_l0_key_to_fp16, k7_1_cast_fp16))[name = tensor<string, []>("k8_1_cast_fp16")];
            tensor<int32, [3]> var_2667_begin_0 = const()[name = tensor<string, []>("op_2667_begin_0"), val = tensor<int32, [3]>([4, 0, 0])];
            tensor<int32, [3]> var_2667_end_0 = const()[name = tensor<string, []>("op_2667_end_0"), val = tensor<int32, [3]>([20, 1, 128])];
            tensor<bool, [3]> var_2667_end_mask_0 = const()[name = tensor<string, []>("op_2667_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [16, 1, 128]> new_cache_s2_l0_key = slice_by_index(begin = var_2667_begin_0, end = var_2667_end_0, end_mask = var_2667_end_mask_0, x = k8_1_cast_fp16)[name = tensor<string, []>("op_2667_cast_fp16")];
            tensor<int32, [4]> var_2673 = const()[name = tensor<string, []>("op_2673"), val = tensor<int32, [4]>([4, 1, 4, 32])];
            tensor<fp16, [4, 1, 4, 32]> q6_1_cast_fp16 = reshape(shape = var_2673, x = q5_1_cast_fp16)[name = tensor<string, []>("q6_1_cast_fp16")];
            tensor<int32, [4]> var_2675 = const()[name = tensor<string, []>("op_2675"), val = tensor<int32, [4]>([4, 1, 4, 4])];
            tensor<fp16, [4, 1, 4, 4]> p6_1_cast_fp16 = reshape(shape = var_2675, x = p5_1_cast_fp16)[name = tensor<string, []>("p6_1_cast_fp16")];
            tensor<int32, [4]> var_2677 = const()[name = tensor<string, []>("op_2677"), val = tensor<int32, [4]>([20, 1, 4, 32])];
            tensor<fp16, [20, 1, 4, 32]> k9_1_cast_fp16 = reshape(shape = var_2677, x = k8_1_cast_fp16)[name = tensor<string, []>("k9_1_cast_fp16")];
            tensor<int32, [4]> var_2681 = const()[name = tensor<string, []>("op_2681"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> attn_scores5_1_transpose_x_0 = const()[name = tensor<string, []>("attn_scores5_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_scores5_1_transpose_y_0 = const()[name = tensor<string, []>("attn_scores5_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = tensor<string, []>("transpose_28_perm_0"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = tensor<string, []>("transpose_29_perm_0"), val = tensor<int32, [4]>([2, 1, -1, -4])];
            tensor<fp16, [4, 1, 32, 20]> transpose_29 = transpose(perm = transpose_29_perm_0, x = k9_1_cast_fp16)[name = tensor<string, []>("transpose_87")];
            tensor<fp16, [4, 1, 4, 32]> transpose_28 = transpose(perm = transpose_28_perm_0, x = q6_1_cast_fp16)[name = tensor<string, []>("transpose_88")];
            tensor<fp16, [4, 1, 4, 20]> attn_scores5_1_cast_fp16 = matmul(transpose_x = attn_scores5_1_transpose_x_0, transpose_y = attn_scores5_1_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor<string, []>("attn_scores5_1_cast_fp16")];
            tensor<bool, []> pos_scores7_1_transpose_x_0 = const()[name = tensor<string, []>("pos_scores7_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_scores7_1_transpose_y_0 = const()[name = tensor<string, []>("pos_scores7_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 4, 23]> pos_emb1_1_to_fp16 = const()[name = tensor<string, []>("pos_emb1_1_to_fp16"), val = tensor<fp16, [4, 1, 4, 23]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1298944)))];
            tensor<fp16, [4, 1, 4, 4]> p7_1_cast_fp16 = transpose(perm = var_2681, x = p6_1_cast_fp16)[name = tensor<string, []>("transpose_86")];
            tensor<fp16, [4, 1, 4, 23]> pos_scores7_1_cast_fp16 = matmul(transpose_x = pos_scores7_1_transpose_x_0, transpose_y = pos_scores7_1_transpose_y_0, x = p7_1_cast_fp16, y = pos_emb1_1_to_fp16)[name = tensor<string, []>("pos_scores7_1_cast_fp16")];
            tensor<int32, [2]> var_2724 = const()[name = tensor<string, []>("op_2724"), val = tensor<int32, [2]>([-1, 23])];
            tensor<fp16, [16, 23]> pos_scores8_1_cast_fp16 = reshape(shape = var_2724, x = pos_scores7_1_cast_fp16)[name = tensor<string, []>("pos_scores8_1_cast_fp16")];
            tensor<bool, []> pos_scores9_1_validate_indices_0 = const()[name = tensor<string, []>("pos_scores9_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<uint16, [16, 20]> indexes1_1_to_uint16 = const()[name = tensor<string, []>("indexes1_1_to_uint16"), val = tensor<uint16, [16, 20]>([[3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19], [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22], [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]])];
            tensor<fp16, [16, 20]> pos_scores9_1_cast_fp16_cast_uint16 = gather_along_axis(axis = var_48, indices = indexes1_1_to_uint16, validate_indices = pos_scores9_1_validate_indices_0, x = pos_scores8_1_cast_fp16)[name = tensor<string, []>("pos_scores9_1_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_2727 = const()[name = tensor<string, []>("op_2727"), val = tensor<int32, [4]>([4, 1, 4, 20])];
            tensor<fp16, [4, 1, 4, 20]> pos_scores10_1_cast_fp16 = reshape(shape = var_2727, x = pos_scores9_1_cast_fp16_cast_uint16)[name = tensor<string, []>("pos_scores10_1_cast_fp16")];
            tensor<fp16, [4, 1, 4, 20]> attn_scores6_1_cast_fp16 = add(x = attn_scores5_1_cast_fp16, y = pos_scores10_1_cast_fp16)[name = tensor<string, []>("attn_scores6_1_cast_fp16")];
            tensor<int32, [1]> var_2730_axes_0 = const()[name = tensor<string, []>("op_2730_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 20]> var_2730 = expand_dims(axes = var_2730_axes_0, x = key_padding_mask1_1)[name = tensor<string, []>("op_2730")];
            tensor<fp16, []> cast_48_to_fp16 = const()[name = tensor<string, []>("cast_48_to_fp16"), val = tensor<fp16, []>(-0x1.f4p+9)];
            tensor<fp16, [4, 1, 4, 20]> attn_scores7_1_cast_fp16 = select(a = cast_48_to_fp16, b = attn_scores6_1_cast_fp16, cond = var_2730)[name = tensor<string, []>("attn_scores7_1_cast_fp16")];
            tensor<fp16, [4, 1, 4, 20]> attn_weights3_1_cast_fp16 = softmax(axis = var_59, x = attn_scores7_1_cast_fp16)[name = tensor<string, []>("attn_weights3_1_cast_fp16")];
            tensor<fp16, [144, 128]> encoder_encoders_2_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1299776))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1318272))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([144, 128])];
            tensor<fp16, [144]> encoder_encoders_2_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1318848)))];
            tensor<fp16, [4, 1, 144]> linear_39_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized, x = x_37_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, []> var_2745_to_fp16 = const()[name = tensor<string, []>("op_2745_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [4, 1, 144]> x_offset_22_cast_fp16 = sub(x = linear_39_cast_fp16, y = var_2745_to_fp16)[name = tensor<string, []>("x_offset_22_cast_fp16")];
            tensor<fp16, [4, 1, 144]> var_2747_cast_fp16 = exp(x = x_offset_22_cast_fp16)[name = tensor<string, []>("op_2747_cast_fp16")];
            tensor<fp16, []> var_2748_to_fp16 = const()[name = tensor<string, []>("op_2748_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 144]> var_2749_cast_fp16 = add(x = var_2747_cast_fp16, y = var_2748_to_fp16)[name = tensor<string, []>("op_2749_cast_fp16")];
            tensor<fp32, []> var_2750_epsilon_0 = const()[name = tensor<string, []>("op_2750_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 144]> var_2750_cast_fp16 = log(epsilon = var_2750_epsilon_0, x = var_2749_cast_fp16)[name = tensor<string, []>("op_2750_cast_fp16")];
            tensor<fp16, []> var_2733_to_fp16 = const()[name = tensor<string, []>("op_2733_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 144]> var_2752_cast_fp16 = equal(x = var_2750_cast_fp16, y = var_2733_to_fp16)[name = tensor<string, []>("op_2752_cast_fp16")];
            tensor<fp16, [4, 1, 144]> log_sum0_22_cast_fp16 = select(a = x_offset_22_cast_fp16, b = var_2750_cast_fp16, cond = var_2752_cast_fp16)[name = tensor<string, []>("log_sum0_22_cast_fp16")];
            tensor<fp16, []> var_2754_to_fp16 = const()[name = tensor<string, []>("op_2754_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 144]> var_2755_cast_fp16 = mul(x = linear_39_cast_fp16, y = var_2754_to_fp16)[name = tensor<string, []>("op_2755_cast_fp16")];
            tensor<fp16, [4, 1, 144]> var_2756_cast_fp16 = sub(x = log_sum0_22_cast_fp16, y = var_2755_cast_fp16)[name = tensor<string, []>("op_2756_cast_fp16")];
            tensor<fp16, []> var_2757_to_fp16 = const()[name = tensor<string, []>("op_2757_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [4, 1, 144]> x_43_cast_fp16 = sub(x = var_2756_cast_fp16, y = var_2757_to_fp16)[name = tensor<string, []>("x_43_cast_fp16")];
            tensor<fp16, [128, 144]> encoder_encoders_2_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1319232))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1337728))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 144])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1338304)))];
            tensor<fp16, [4, 1, 128]> linear_40_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized, x = x_43_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input28_1_cast_fp16 = add(x = x_37_cast_fp16, y = linear_40_cast_fp16)[name = tensor<string, []>("input28_1_cast_fp16")];
            tensor<int32, [4]> attn_weights4_1_begin_0 = const()[name = tensor<string, []>("attn_weights4_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_weights4_1_end_0 = const()[name = tensor<string, []>("attn_weights4_1_end_0"), val = tensor<int32, [4]>([1, 1, 4, 20])];
            tensor<bool, [4]> attn_weights4_1_end_mask_0 = const()[name = tensor<string, []>("attn_weights4_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1, 4, 20]> attn_weights4_1_cast_fp16 = slice_by_index(begin = attn_weights4_1_begin_0, end = attn_weights4_1_end_0, end_mask = attn_weights4_1_end_mask_0, x = attn_weights3_1_cast_fp16)[name = tensor<string, []>("attn_weights4_1_cast_fp16")];
            tensor<fp16, [288, 128]> encoder_encoders_2_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [36864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1338624))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1375552))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([288, 128])];
            tensor<fp16, [288]> encoder_encoders_2_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1376128)))];
            tensor<fp16, [4, 1, 288]> linear_41_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized, x = input28_1_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<int32, [3]> var_2771_split_sizes_0 = const()[name = tensor<string, []>("op_2771_split_sizes_0"), val = tensor<int32, [3]>([96, 96, 96])];
            tensor<int32, []> var_2771_axis_0 = const()[name = tensor<string, []>("op_2771_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [4, 1, 96]> var_2771_cast_fp16_0, tensor<fp16, [4, 1, 96]> var_2771_cast_fp16_1, tensor<fp16, [4, 1, 96]> var_2771_cast_fp16_2 = split(axis = var_2771_axis_0, split_sizes = var_2771_split_sizes_0, x = linear_41_cast_fp16)[name = tensor<string, []>("op_2771_cast_fp16")];
            tensor<fp16, [4, 1, 96]> var_2774_cast_fp16 = tanh(x = var_2771_cast_fp16_0)[name = tensor<string, []>("op_2774_cast_fp16")];
            tensor<int32, [1]> var_2775_axes_0 = const()[name = tensor<string, []>("op_2775_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 1, 96, 1]> var_2775_cast_fp16 = expand_dims(axes = var_2775_axes_0, x = var_2774_cast_fp16)[name = tensor<string, []>("op_2775_cast_fp16")];
            tensor<int32, [3]> var_2776 = const()[name = tensor<string, []>("op_2776"), val = tensor<int32, [3]>([4, 1, 96])];
            tensor<fp16, [4, 1, 96]> s1_1_cast_fp16 = reshape(shape = var_2776, x = var_2775_cast_fp16)[name = tensor<string, []>("s1_1_cast_fp16")];
            tensor<fp16, [4, 1, 96]> x59_1_cast_fp16 = mul(x = var_2771_cast_fp16_1, y = s1_1_cast_fp16)[name = tensor<string, []>("x59_1_cast_fp16")];
            tensor<int32, [4]> var_2790 = const()[name = tensor<string, []>("op_2790"), val = tensor<int32, [4]>([4, 1, 1, -1])];
            tensor<fp16, [4, 1, 1, 96]> var_2791_cast_fp16 = reshape(shape = var_2790, x = x59_1_cast_fp16)[name = tensor<string, []>("op_2791_cast_fp16")];
            tensor<int32, [4]> var_2792 = const()[name = tensor<string, []>("op_2792"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<bool, []> x_pad1_1_interleave_0 = const()[name = tensor<string, []>("x_pad1_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s2_l0_nonlin_attn_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s2_l0_nonlin_attn_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 96]> x60_1_cast_fp16 = transpose(perm = var_2792, x = var_2791_cast_fp16)[name = tensor<string, []>("transpose_85")];
            tensor<fp16, [1, 1, 16, 96]> cache_s2_l0_nonlin_attn_to_fp16 = cast(dtype = cache_s2_l0_nonlin_attn_to_fp16_dtype_0, x = cache_s2_l0_nonlin_attn)[name = tensor<string, []>("cast_22")];
            tensor<fp16, [1, 1, 20, 96]> x_pad1_1_cast_fp16 = concat(axis = var_49, interleave = x_pad1_1_interleave_0, values = (cache_s2_l0_nonlin_attn_to_fp16, x60_1_cast_fp16))[name = tensor<string, []>("x_pad1_1_cast_fp16")];
            tensor<int32, [4]> var_2798_begin_0 = const()[name = tensor<string, []>("op_2798_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_2798_end_0 = const()[name = tensor<string, []>("op_2798_end_0"), val = tensor<int32, [4]>([1, 1, 20, 96])];
            tensor<bool, [4]> var_2798_end_mask_0 = const()[name = tensor<string, []>("op_2798_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 16, 96]> new_cache_s2_l0_nonlin_attn = slice_by_index(begin = var_2798_begin_0, end = var_2798_end_0, end_mask = var_2798_end_mask_0, x = x_pad1_1_cast_fp16)[name = tensor<string, []>("op_2798_cast_fp16")];
            tensor<bool, []> x61_1_transpose_x_0 = const()[name = tensor<string, []>("x61_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x61_1_transpose_y_0 = const()[name = tensor<string, []>("x61_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 4, 96]> x61_1_cast_fp16 = matmul(transpose_x = x61_1_transpose_x_0, transpose_y = x61_1_transpose_y_0, x = attn_weights4_1_cast_fp16, y = x_pad1_1_cast_fp16)[name = tensor<string, []>("x61_1_cast_fp16")];
            tensor<int32, [4]> var_2801 = const()[name = tensor<string, []>("op_2801"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_2803 = const()[name = tensor<string, []>("op_2803"), val = tensor<int32, [3]>([4, 1, -1])];
            tensor<fp16, [4, 1, 1, 96]> var_2802_cast_fp16 = transpose(perm = var_2801, x = x61_1_cast_fp16)[name = tensor<string, []>("transpose_84")];
            tensor<fp16, [4, 1, 96]> x62_1_cast_fp16 = reshape(shape = var_2803, x = var_2802_cast_fp16)[name = tensor<string, []>("x62_1_cast_fp16")];
            tensor<fp16, [4, 1, 96]> input30_1_cast_fp16 = mul(x = x62_1_cast_fp16, y = var_2771_cast_fp16_2)[name = tensor<string, []>("input30_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_2_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1376768))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1389120))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1389696)))];
            tensor<fp16, [4, 1, 128]> linear_42_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized, x = input30_1_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x63_1_cast_fp16 = add(x = input28_1_cast_fp16, y = linear_42_cast_fp16)[name = tensor<string, []>("x63_1_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_2_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1390016))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1396224))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_2_encoder_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1396800)))];
            tensor<fp16, [4, 1, 48]> linear_43_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_self_attn1_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized, x = x63_1_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<bool, []> x64_1_interleave_0 = const()[name = tensor<string, []>("x64_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s2_l0_val1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s2_l0_val1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [16, 1, 48]> cache_s2_l0_val1_to_fp16 = cast(dtype = cache_s2_l0_val1_to_fp16_dtype_0, x = cache_s2_l0_val1)[name = tensor<string, []>("cast_21")];
            tensor<fp16, [20, 1, 48]> x64_1_cast_fp16 = concat(axis = var_46, interleave = x64_1_interleave_0, values = (cache_s2_l0_val1_to_fp16, linear_43_cast_fp16))[name = tensor<string, []>("x64_1_cast_fp16")];
            tensor<int32, [3]> var_2828_begin_0 = const()[name = tensor<string, []>("op_2828_begin_0"), val = tensor<int32, [3]>([4, 0, 0])];
            tensor<int32, [3]> var_2828_end_0 = const()[name = tensor<string, []>("op_2828_end_0"), val = tensor<int32, [3]>([20, 1, 48])];
            tensor<bool, [3]> var_2828_end_mask_0 = const()[name = tensor<string, []>("op_2828_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [16, 1, 48]> new_cache_s2_l0_val1 = slice_by_index(begin = var_2828_begin_0, end = var_2828_end_0, end_mask = var_2828_end_mask_0, x = x64_1_cast_fp16)[name = tensor<string, []>("op_2828_cast_fp16")];
            tensor<int32, [4]> var_2829 = const()[name = tensor<string, []>("op_2829"), val = tensor<int32, [4]>([20, 1, 4, -1])];
            tensor<fp16, [20, 1, 4, 12]> var_2830_cast_fp16 = reshape(shape = var_2829, x = x64_1_cast_fp16)[name = tensor<string, []>("op_2830_cast_fp16")];
            tensor<int32, [4]> var_2831 = const()[name = tensor<string, []>("op_2831"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x66_1_transpose_x_0 = const()[name = tensor<string, []>("x66_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x66_1_transpose_y_0 = const()[name = tensor<string, []>("x66_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 20, 12]> x65_1_cast_fp16 = transpose(perm = var_2831, x = var_2830_cast_fp16)[name = tensor<string, []>("transpose_83")];
            tensor<fp16, [4, 1, 4, 12]> x66_1_cast_fp16 = matmul(transpose_x = x66_1_transpose_x_0, transpose_y = x66_1_transpose_y_0, x = attn_weights3_1_cast_fp16, y = x65_1_cast_fp16)[name = tensor<string, []>("x66_1_cast_fp16")];
            tensor<int32, [4]> var_2836 = const()[name = tensor<string, []>("op_2836"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_2841 = const()[name = tensor<string, []>("op_2841"), val = tensor<int32, [3]>([4, 1, 48])];
            tensor<fp16, [4, 1, 4, 12]> var_2837_cast_fp16 = transpose(perm = var_2836, x = x66_1_cast_fp16)[name = tensor<string, []>("transpose_82")];
            tensor<fp16, [4, 1, 48]> input31_1_cast_fp16 = reshape(shape = var_2841, x = var_2837_cast_fp16)[name = tensor<string, []>("input31_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_2_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1396992))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1403200))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1403776)))];
            tensor<fp16, [4, 1, 128]> linear_44_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_self_attn1_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized, x = input31_1_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input32_1_cast_fp16 = add(x = x63_1_cast_fp16, y = linear_44_cast_fp16)[name = tensor<string, []>("input32_1_cast_fp16")];
            tensor<int32, [2]> src_key_padding_mask5_1_begin_0 = const()[name = tensor<string, []>("src_key_padding_mask5_1_begin_0"), val = tensor<int32, [2]>([0, 16])];
            tensor<int32, [2]> src_key_padding_mask5_1_end_0 = const()[name = tensor<string, []>("src_key_padding_mask5_1_end_0"), val = tensor<int32, [2]>([1, 20])];
            tensor<bool, [2]> src_key_padding_mask5_1_end_mask_0 = const()[name = tensor<string, []>("src_key_padding_mask5_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<bool, [1, 4]> src_key_padding_mask5_1 = slice_by_index(begin = src_key_padding_mask5_1_begin_0, end = src_key_padding_mask5_1_end_0, end_mask = src_key_padding_mask5_1_end_mask_0, x = key_padding_mask1_1)[name = tensor<string, []>("src_key_padding_mask5_1")];
            tensor<fp16, [256, 128]> encoder_encoders_2_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1404096))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1436928))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_2_encoder_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1437504)))];
            tensor<fp16, [4, 1, 256]> linear_45_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_conv_module1_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized, x = input32_1_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<int32, [2]> var_2852_split_sizes_0 = const()[name = tensor<string, []>("op_2852_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_2852_axis_0 = const()[name = tensor<string, []>("op_2852_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [4, 1, 128]> var_2852_cast_fp16_0, tensor<fp16, [4, 1, 128]> var_2852_cast_fp16_1 = split(axis = var_2852_axis_0, split_sizes = var_2852_split_sizes_0, x = linear_45_cast_fp16)[name = tensor<string, []>("op_2852_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_2854_cast_fp16 = sigmoid(x = var_2852_cast_fp16_1)[name = tensor<string, []>("op_2854_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x68_1_cast_fp16 = mul(x = var_2852_cast_fp16_0, y = var_2854_cast_fp16)[name = tensor<string, []>("x68_1_cast_fp16")];
            tensor<int32, [3]> var_2856 = const()[name = tensor<string, []>("op_2856"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_2858_axes_0 = const()[name = tensor<string, []>("op_2858_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 4]> var_2858 = expand_dims(axes = var_2858_axes_0, x = src_key_padding_mask5_1)[name = tensor<string, []>("op_2858")];
            tensor<int32, [3]> var_2859_reps_0 = const()[name = tensor<string, []>("op_2859_reps_0"), val = tensor<int32, [3]>([1, 128, 1])];
            tensor<bool, [1, 128, 4]> var_2859 = tile(reps = var_2859_reps_0, x = var_2858)[name = tensor<string, []>("op_2859")];
            tensor<fp16, [1, 128, 4]> x69_1_cast_fp16 = transpose(perm = var_2856, x = x68_1_cast_fp16)[name = tensor<string, []>("transpose_81")];
            tensor<fp16, [1, 128, 4]> x70_1_cast_fp16 = select(a = var_62_to_fp16, b = x69_1_cast_fp16, cond = var_2859)[name = tensor<string, []>("x70_1_cast_fp16")];
            tensor<bool, []> x71_1_interleave_0 = const()[name = tensor<string, []>("x71_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s2_l0_conv1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s2_l0_conv1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 7]> cache_s2_l0_conv1_to_fp16 = cast(dtype = cache_s2_l0_conv1_to_fp16_dtype_0, x = cache_s2_l0_conv1)[name = tensor<string, []>("cast_20")];
            tensor<fp16, [1, 128, 11]> x71_1_cast_fp16 = concat(axis = var_49, interleave = x71_1_interleave_0, values = (cache_s2_l0_conv1_to_fp16, x70_1_cast_fp16))[name = tensor<string, []>("x71_1_cast_fp16")];
            tensor<int32, [3]> var_2866_begin_0 = const()[name = tensor<string, []>("op_2866_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_2866_end_0 = const()[name = tensor<string, []>("op_2866_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> var_2866_end_mask_0 = const()[name = tensor<string, []>("op_2866_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 7]> new_cache_s2_l0_conv1 = slice_by_index(begin = var_2866_begin_0, end = var_2866_end_0, end_mask = var_2866_end_mask_0, x = x71_1_cast_fp16)[name = tensor<string, []>("op_2866_cast_fp16")];
            tensor<string, []> x_causal_10_pad_type_0 = const()[name = tensor<string, []>("x_causal_10_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_10_groups_0 = const()[name = tensor<string, []>("x_causal_10_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_10_strides_0 = const()[name = tensor<string, []>("x_causal_10_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_10_pad_0 = const()[name = tensor<string, []>("x_causal_10_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_10_dilations_0 = const()[name = tensor<string, []>("x_causal_10_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 8]> encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1438080)))];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1440192)))];
            tensor<fp16, [1, 128, 4]> x_causal_10_cast_fp16 = conv(bias = encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_10_dilations_0, groups = x_causal_10_groups_0, pad = x_causal_10_pad_0, pad_type = x_causal_10_pad_type_0, strides = x_causal_10_strides_0, weight = encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16, x = x71_1_cast_fp16)[name = tensor<string, []>("x_causal_10_cast_fp16")];
            tensor<int32, [3]> input34_1_begin_0 = const()[name = tensor<string, []>("input34_1_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> input34_1_end_0 = const()[name = tensor<string, []>("input34_1_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> input34_1_end_mask_0 = const()[name = tensor<string, []>("input34_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 4]> input34_1_cast_fp16 = slice_by_index(begin = input34_1_begin_0, end = input34_1_end_0, end_mask = input34_1_end_mask_0, x = x71_1_cast_fp16)[name = tensor<string, []>("input34_1_cast_fp16")];
            tensor<string, []> x_chunk_13_pad_type_0 = const()[name = tensor<string, []>("x_chunk_13_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_13_pad_0 = const()[name = tensor<string, []>("x_chunk_13_pad_0"), val = tensor<int32, [2]>([7, 7])];
            tensor<int32, []> x_chunk_13_groups_0 = const()[name = tensor<string, []>("x_chunk_13_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_13_strides_0 = const()[name = tensor<string, []>("x_chunk_13_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_13_dilations_0 = const()[name = tensor<string, []>("x_chunk_13_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 15]> encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1440512)))];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1444416)))];
            tensor<fp16, [1, 128, 4]> x_chunk_13_cast_fp16 = conv(bias = encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_13_dilations_0, groups = x_chunk_13_groups_0, pad = x_chunk_13_pad_0, pad_type = x_chunk_13_pad_type_0, strides = x_chunk_13_strides_0, weight = encoder_encoders_2_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16, x = input34_1_cast_fp16)[name = tensor<string, []>("x_chunk_13_cast_fp16")];
            tensor<fp16, [128, 4]> chunk_scale3_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale3_1_to_fp16"), val = tensor<fp16, [128, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1444736)))];
            tensor<fp16, [1, 128, 4]> x_chunk3_1_cast_fp16 = mul(x = x_chunk_13_cast_fp16, y = chunk_scale3_1_to_fp16)[name = tensor<string, []>("x_chunk3_1_cast_fp16")];
            tensor<fp16, [1, 128, 4]> x72_1_cast_fp16 = add(x = x_chunk3_1_cast_fp16, y = x_causal_10_cast_fp16)[name = tensor<string, []>("x72_1_cast_fp16")];
            tensor<int32, [3]> var_2906 = const()[name = tensor<string, []>("op_2906"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_2915_to_fp16 = const()[name = tensor<string, []>("op_2915_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> x73_1_cast_fp16 = transpose(perm = var_2906, x = x72_1_cast_fp16)[name = tensor<string, []>("transpose_80")];
            tensor<fp16, [4, 1, 128]> x_offset_24_cast_fp16 = sub(x = x73_1_cast_fp16, y = var_2915_to_fp16)[name = tensor<string, []>("x_offset_24_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_2917_cast_fp16 = exp(x = x_offset_24_cast_fp16)[name = tensor<string, []>("op_2917_cast_fp16")];
            tensor<fp16, []> var_2918_to_fp16 = const()[name = tensor<string, []>("op_2918_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> var_2919_cast_fp16 = add(x = var_2917_cast_fp16, y = var_2918_to_fp16)[name = tensor<string, []>("op_2919_cast_fp16")];
            tensor<fp32, []> var_2920_epsilon_0 = const()[name = tensor<string, []>("op_2920_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 128]> var_2920_cast_fp16 = log(epsilon = var_2920_epsilon_0, x = var_2919_cast_fp16)[name = tensor<string, []>("op_2920_cast_fp16")];
            tensor<fp16, []> var_2912_to_fp16 = const()[name = tensor<string, []>("op_2912_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 128]> var_2922_cast_fp16 = equal(x = var_2920_cast_fp16, y = var_2912_to_fp16)[name = tensor<string, []>("op_2922_cast_fp16")];
            tensor<fp16, [4, 1, 128]> log_sum0_24_cast_fp16 = select(a = x_offset_24_cast_fp16, b = var_2920_cast_fp16, cond = var_2922_cast_fp16)[name = tensor<string, []>("log_sum0_24_cast_fp16")];
            tensor<fp16, []> var_2924_to_fp16 = const()[name = tensor<string, []>("op_2924_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 128]> var_2925_cast_fp16 = mul(x = x73_1_cast_fp16, y = var_2924_to_fp16)[name = tensor<string, []>("op_2925_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_2926_cast_fp16 = sub(x = log_sum0_24_cast_fp16, y = var_2925_cast_fp16)[name = tensor<string, []>("op_2926_cast_fp16")];
            tensor<fp16, []> var_2927_to_fp16 = const()[name = tensor<string, []>("op_2927_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [4, 1, 128]> x0_17_cast_fp16 = sub(x = var_2926_cast_fp16, y = var_2927_to_fp16)[name = tensor<string, []>("x0_17_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_2_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1445824))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1462272))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1462848)))];
            tensor<fp16, [4, 1, 128]> linear_46_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_conv_module1_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized, x = x0_17_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input35_1_cast_fp16 = add(x = input32_1_cast_fp16, y = linear_46_cast_fp16)[name = tensor<string, []>("input35_1_cast_fp16")];
            tensor<fp16, [192, 128]> encoder_encoders_2_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1463168))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1487808))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([192, 128])];
            tensor<fp16, [192]> encoder_encoders_2_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1488384)))];
            tensor<fp16, [4, 1, 192]> linear_47_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized, x = input35_1_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, []> var_2943_to_fp16 = const()[name = tensor<string, []>("op_2943_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [4, 1, 192]> x_offset_26_cast_fp16 = sub(x = linear_47_cast_fp16, y = var_2943_to_fp16)[name = tensor<string, []>("x_offset_26_cast_fp16")];
            tensor<fp16, [4, 1, 192]> var_2945_cast_fp16 = exp(x = x_offset_26_cast_fp16)[name = tensor<string, []>("op_2945_cast_fp16")];
            tensor<fp16, []> var_2946_to_fp16 = const()[name = tensor<string, []>("op_2946_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 192]> var_2947_cast_fp16 = add(x = var_2945_cast_fp16, y = var_2946_to_fp16)[name = tensor<string, []>("op_2947_cast_fp16")];
            tensor<fp32, []> var_2948_epsilon_0 = const()[name = tensor<string, []>("op_2948_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 192]> var_2948_cast_fp16 = log(epsilon = var_2948_epsilon_0, x = var_2947_cast_fp16)[name = tensor<string, []>("op_2948_cast_fp16")];
            tensor<fp16, []> var_2931_to_fp16 = const()[name = tensor<string, []>("op_2931_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 192]> var_2950_cast_fp16 = equal(x = var_2948_cast_fp16, y = var_2931_to_fp16)[name = tensor<string, []>("op_2950_cast_fp16")];
            tensor<fp16, [4, 1, 192]> log_sum0_26_cast_fp16 = select(a = x_offset_26_cast_fp16, b = var_2948_cast_fp16, cond = var_2950_cast_fp16)[name = tensor<string, []>("log_sum0_26_cast_fp16")];
            tensor<fp16, []> var_2952_to_fp16 = const()[name = tensor<string, []>("op_2952_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 192]> var_2953_cast_fp16 = mul(x = linear_47_cast_fp16, y = var_2952_to_fp16)[name = tensor<string, []>("op_2953_cast_fp16")];
            tensor<fp16, [4, 1, 192]> var_2954_cast_fp16 = sub(x = log_sum0_26_cast_fp16, y = var_2953_cast_fp16)[name = tensor<string, []>("op_2954_cast_fp16")];
            tensor<fp16, []> var_2955_to_fp16 = const()[name = tensor<string, []>("op_2955_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [4, 1, 192]> x_45_cast_fp16 = sub(x = var_2954_cast_fp16, y = var_2955_to_fp16)[name = tensor<string, []>("x_45_cast_fp16")];
            tensor<fp16, [128, 192]> encoder_encoders_2_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1488832))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1513472))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 192])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1514048)))];
            tensor<fp16, [4, 1, 128]> linear_48_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized, x = x_45_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<fp16, [4, 1, 128]> src4_1_cast_fp16 = add(x = input35_1_cast_fp16, y = linear_48_cast_fp16)[name = tensor<string, []>("src4_1_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_2961_cast_fp16 = sub(x = src4_1_cast_fp16, y = x_37_cast_fp16)[name = tensor<string, []>("op_2961_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1514368)))];
            tensor<fp16, [4, 1, 128]> var_2962_cast_fp16 = mul(x = var_2961_cast_fp16, y = encoder_encoders_2_encoder_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_2962_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_2963_cast_fp16 = add(x = x_37_cast_fp16, y = var_2962_cast_fp16)[name = tensor<string, []>("op_2963_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_2_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1514688))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1520896))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_2_encoder_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1521472)))];
            tensor<fp16, [4, 1, 48]> linear_49_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_self_attn2_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized, x = var_2963_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<bool, []> x74_1_interleave_0 = const()[name = tensor<string, []>("x74_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s2_l0_val2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s2_l0_val2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [16, 1, 48]> cache_s2_l0_val2_to_fp16 = cast(dtype = cache_s2_l0_val2_to_fp16_dtype_0, x = cache_s2_l0_val2)[name = tensor<string, []>("cast_19")];
            tensor<fp16, [20, 1, 48]> x74_1_cast_fp16 = concat(axis = var_46, interleave = x74_1_interleave_0, values = (cache_s2_l0_val2_to_fp16, linear_49_cast_fp16))[name = tensor<string, []>("x74_1_cast_fp16")];
            tensor<int32, [3]> var_2982_begin_0 = const()[name = tensor<string, []>("op_2982_begin_0"), val = tensor<int32, [3]>([4, 0, 0])];
            tensor<int32, [3]> var_2982_end_0 = const()[name = tensor<string, []>("op_2982_end_0"), val = tensor<int32, [3]>([20, 1, 48])];
            tensor<bool, [3]> var_2982_end_mask_0 = const()[name = tensor<string, []>("op_2982_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [16, 1, 48]> new_cache_s2_l0_val2 = slice_by_index(begin = var_2982_begin_0, end = var_2982_end_0, end_mask = var_2982_end_mask_0, x = x74_1_cast_fp16)[name = tensor<string, []>("op_2982_cast_fp16")];
            tensor<int32, [4]> var_2983 = const()[name = tensor<string, []>("op_2983"), val = tensor<int32, [4]>([20, 1, 4, -1])];
            tensor<fp16, [20, 1, 4, 12]> var_2984_cast_fp16 = reshape(shape = var_2983, x = x74_1_cast_fp16)[name = tensor<string, []>("op_2984_cast_fp16")];
            tensor<int32, [4]> var_2985 = const()[name = tensor<string, []>("op_2985"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x76_1_transpose_x_0 = const()[name = tensor<string, []>("x76_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x76_1_transpose_y_0 = const()[name = tensor<string, []>("x76_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 20, 12]> x75_1_cast_fp16 = transpose(perm = var_2985, x = var_2984_cast_fp16)[name = tensor<string, []>("transpose_79")];
            tensor<fp16, [4, 1, 4, 12]> x76_1_cast_fp16 = matmul(transpose_x = x76_1_transpose_x_0, transpose_y = x76_1_transpose_y_0, x = attn_weights3_1_cast_fp16, y = x75_1_cast_fp16)[name = tensor<string, []>("x76_1_cast_fp16")];
            tensor<int32, [4]> var_2990 = const()[name = tensor<string, []>("op_2990"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_2995 = const()[name = tensor<string, []>("op_2995"), val = tensor<int32, [3]>([4, 1, 48])];
            tensor<fp16, [4, 1, 4, 12]> var_2991_cast_fp16 = transpose(perm = var_2990, x = x76_1_cast_fp16)[name = tensor<string, []>("transpose_78")];
            tensor<fp16, [4, 1, 48]> input36_1_cast_fp16 = reshape(shape = var_2995, x = var_2991_cast_fp16)[name = tensor<string, []>("input36_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_2_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1521664))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1527872))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1528448)))];
            tensor<fp16, [4, 1, 128]> linear_50_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_self_attn2_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized, x = input36_1_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input37_1_cast_fp16 = add(x = var_2963_cast_fp16, y = linear_50_cast_fp16)[name = tensor<string, []>("input37_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_2_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1528768))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1561600))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_2_encoder_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1562176)))];
            tensor<fp16, [4, 1, 256]> linear_51_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_conv_module2_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized, x = input37_1_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")];
            tensor<int32, [2]> var_3006_split_sizes_0 = const()[name = tensor<string, []>("op_3006_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_3006_axis_0 = const()[name = tensor<string, []>("op_3006_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [4, 1, 128]> var_3006_cast_fp16_0, tensor<fp16, [4, 1, 128]> var_3006_cast_fp16_1 = split(axis = var_3006_axis_0, split_sizes = var_3006_split_sizes_0, x = linear_51_cast_fp16)[name = tensor<string, []>("op_3006_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_3008_cast_fp16 = sigmoid(x = var_3006_cast_fp16_1)[name = tensor<string, []>("op_3008_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x78_1_cast_fp16 = mul(x = var_3006_cast_fp16_0, y = var_3008_cast_fp16)[name = tensor<string, []>("x78_1_cast_fp16")];
            tensor<int32, [3]> var_3010 = const()[name = tensor<string, []>("op_3010"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 4]> x79_1_cast_fp16 = transpose(perm = var_3010, x = x78_1_cast_fp16)[name = tensor<string, []>("transpose_77")];
            tensor<fp16, [1, 128, 4]> x80_1_cast_fp16 = select(a = var_62_to_fp16, b = x79_1_cast_fp16, cond = var_2859)[name = tensor<string, []>("x80_1_cast_fp16")];
            tensor<bool, []> x81_1_interleave_0 = const()[name = tensor<string, []>("x81_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s2_l0_conv2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s2_l0_conv2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 7]> cache_s2_l0_conv2_to_fp16 = cast(dtype = cache_s2_l0_conv2_to_fp16_dtype_0, x = cache_s2_l0_conv2)[name = tensor<string, []>("cast_18")];
            tensor<fp16, [1, 128, 11]> x81_1_cast_fp16 = concat(axis = var_49, interleave = x81_1_interleave_0, values = (cache_s2_l0_conv2_to_fp16, x80_1_cast_fp16))[name = tensor<string, []>("x81_1_cast_fp16")];
            tensor<int32, [3]> var_3020_begin_0 = const()[name = tensor<string, []>("op_3020_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_3020_end_0 = const()[name = tensor<string, []>("op_3020_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> var_3020_end_mask_0 = const()[name = tensor<string, []>("op_3020_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 7]> new_cache_s2_l0_conv2 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, x = x81_1_cast_fp16)[name = tensor<string, []>("op_3020_cast_fp16")];
            tensor<string, []> x_causal_12_pad_type_0 = const()[name = tensor<string, []>("x_causal_12_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_12_groups_0 = const()[name = tensor<string, []>("x_causal_12_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_12_strides_0 = const()[name = tensor<string, []>("x_causal_12_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_12_pad_0 = const()[name = tensor<string, []>("x_causal_12_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_12_dilations_0 = const()[name = tensor<string, []>("x_causal_12_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 8]> encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1562752)))];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1564864)))];
            tensor<fp16, [1, 128, 4]> x_causal_12_cast_fp16 = conv(bias = encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_12_dilations_0, groups = x_causal_12_groups_0, pad = x_causal_12_pad_0, pad_type = x_causal_12_pad_type_0, strides = x_causal_12_strides_0, weight = encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16, x = x81_1_cast_fp16)[name = tensor<string, []>("x_causal_12_cast_fp16")];
            tensor<int32, [3]> input39_1_begin_0 = const()[name = tensor<string, []>("input39_1_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> input39_1_end_0 = const()[name = tensor<string, []>("input39_1_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> input39_1_end_mask_0 = const()[name = tensor<string, []>("input39_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 4]> input39_1_cast_fp16 = slice_by_index(begin = input39_1_begin_0, end = input39_1_end_0, end_mask = input39_1_end_mask_0, x = x81_1_cast_fp16)[name = tensor<string, []>("input39_1_cast_fp16")];
            tensor<string, []> x_chunk_15_pad_type_0 = const()[name = tensor<string, []>("x_chunk_15_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_15_pad_0 = const()[name = tensor<string, []>("x_chunk_15_pad_0"), val = tensor<int32, [2]>([7, 7])];
            tensor<int32, []> x_chunk_15_groups_0 = const()[name = tensor<string, []>("x_chunk_15_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_15_strides_0 = const()[name = tensor<string, []>("x_chunk_15_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_15_dilations_0 = const()[name = tensor<string, []>("x_chunk_15_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 15]> encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1565184)))];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1569088)))];
            tensor<fp16, [1, 128, 4]> x_chunk_15_cast_fp16 = conv(bias = encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_15_dilations_0, groups = x_chunk_15_groups_0, pad = x_chunk_15_pad_0, pad_type = x_chunk_15_pad_type_0, strides = x_chunk_15_strides_0, weight = encoder_encoders_2_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16, x = input39_1_cast_fp16)[name = tensor<string, []>("x_chunk_15_cast_fp16")];
            tensor<fp16, [128, 4]> chunk_scale4_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale4_1_to_fp16"), val = tensor<fp16, [128, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1569408)))];
            tensor<fp16, [1, 128, 4]> x_chunk4_1_cast_fp16 = mul(x = x_chunk_15_cast_fp16, y = chunk_scale4_1_to_fp16)[name = tensor<string, []>("x_chunk4_1_cast_fp16")];
            tensor<fp16, [1, 128, 4]> x82_1_cast_fp16 = add(x = x_chunk4_1_cast_fp16, y = x_causal_12_cast_fp16)[name = tensor<string, []>("x82_1_cast_fp16")];
            tensor<int32, [3]> var_3060 = const()[name = tensor<string, []>("op_3060"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_3069_to_fp16 = const()[name = tensor<string, []>("op_3069_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> x83_1_cast_fp16 = transpose(perm = var_3060, x = x82_1_cast_fp16)[name = tensor<string, []>("transpose_76")];
            tensor<fp16, [4, 1, 128]> x_offset_28_cast_fp16 = sub(x = x83_1_cast_fp16, y = var_3069_to_fp16)[name = tensor<string, []>("x_offset_28_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_3071_cast_fp16 = exp(x = x_offset_28_cast_fp16)[name = tensor<string, []>("op_3071_cast_fp16")];
            tensor<fp16, []> var_3072_to_fp16 = const()[name = tensor<string, []>("op_3072_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> var_3073_cast_fp16 = add(x = var_3071_cast_fp16, y = var_3072_to_fp16)[name = tensor<string, []>("op_3073_cast_fp16")];
            tensor<fp32, []> var_3074_epsilon_0 = const()[name = tensor<string, []>("op_3074_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 128]> var_3074_cast_fp16 = log(epsilon = var_3074_epsilon_0, x = var_3073_cast_fp16)[name = tensor<string, []>("op_3074_cast_fp16")];
            tensor<fp16, []> var_3066_to_fp16 = const()[name = tensor<string, []>("op_3066_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 128]> var_3076_cast_fp16 = equal(x = var_3074_cast_fp16, y = var_3066_to_fp16)[name = tensor<string, []>("op_3076_cast_fp16")];
            tensor<fp16, [4, 1, 128]> log_sum0_28_cast_fp16 = select(a = x_offset_28_cast_fp16, b = var_3074_cast_fp16, cond = var_3076_cast_fp16)[name = tensor<string, []>("log_sum0_28_cast_fp16")];
            tensor<fp16, []> var_3078_to_fp16 = const()[name = tensor<string, []>("op_3078_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 128]> var_3079_cast_fp16 = mul(x = x83_1_cast_fp16, y = var_3078_to_fp16)[name = tensor<string, []>("op_3079_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_3080_cast_fp16 = sub(x = log_sum0_28_cast_fp16, y = var_3079_cast_fp16)[name = tensor<string, []>("op_3080_cast_fp16")];
            tensor<fp16, []> var_3081_to_fp16 = const()[name = tensor<string, []>("op_3081_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [4, 1, 128]> x0_19_cast_fp16 = sub(x = var_3080_cast_fp16, y = var_3081_to_fp16)[name = tensor<string, []>("x0_19_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_2_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1570496))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1586944))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1587520)))];
            tensor<fp16, [4, 1, 128]> linear_52_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_conv_module2_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized, x = x0_19_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input40_1_cast_fp16 = add(x = input37_1_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("input40_1_cast_fp16")];
            tensor<fp16, [240, 128]> encoder_encoders_2_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1587840))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1618624))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([240, 128])];
            tensor<fp16, [240]> encoder_encoders_2_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1619200)))];
            tensor<fp16, [4, 1, 240]> linear_53_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized, x = input40_1_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<fp16, []> var_3097_to_fp16 = const()[name = tensor<string, []>("op_3097_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [4, 1, 240]> x_offset_30_cast_fp16 = sub(x = linear_53_cast_fp16, y = var_3097_to_fp16)[name = tensor<string, []>("x_offset_30_cast_fp16")];
            tensor<fp16, [4, 1, 240]> var_3099_cast_fp16 = exp(x = x_offset_30_cast_fp16)[name = tensor<string, []>("op_3099_cast_fp16")];
            tensor<fp16, []> var_3100_to_fp16 = const()[name = tensor<string, []>("op_3100_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 240]> var_3101_cast_fp16 = add(x = var_3099_cast_fp16, y = var_3100_to_fp16)[name = tensor<string, []>("op_3101_cast_fp16")];
            tensor<fp32, []> var_3102_epsilon_0 = const()[name = tensor<string, []>("op_3102_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 240]> var_3102_cast_fp16 = log(epsilon = var_3102_epsilon_0, x = var_3101_cast_fp16)[name = tensor<string, []>("op_3102_cast_fp16")];
            tensor<fp16, []> var_3085_to_fp16 = const()[name = tensor<string, []>("op_3085_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 240]> var_3104_cast_fp16 = equal(x = var_3102_cast_fp16, y = var_3085_to_fp16)[name = tensor<string, []>("op_3104_cast_fp16")];
            tensor<fp16, [4, 1, 240]> log_sum0_30_cast_fp16 = select(a = x_offset_30_cast_fp16, b = var_3102_cast_fp16, cond = var_3104_cast_fp16)[name = tensor<string, []>("log_sum0_30_cast_fp16")];
            tensor<fp16, []> var_3106_to_fp16 = const()[name = tensor<string, []>("op_3106_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 240]> var_3107_cast_fp16 = mul(x = linear_53_cast_fp16, y = var_3106_to_fp16)[name = tensor<string, []>("op_3107_cast_fp16")];
            tensor<fp16, [4, 1, 240]> var_3108_cast_fp16 = sub(x = log_sum0_30_cast_fp16, y = var_3107_cast_fp16)[name = tensor<string, []>("op_3108_cast_fp16")];
            tensor<fp16, []> var_3109_to_fp16 = const()[name = tensor<string, []>("op_3109_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [4, 1, 240]> x_49_cast_fp16 = sub(x = var_3108_cast_fp16, y = var_3109_to_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
            tensor<fp16, [128, 240]> encoder_encoders_2_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1619776))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1650560))), name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 240])];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1651136)))];
            tensor<fp16, [4, 1, 128]> linear_54_cast_fp16 = linear(bias = encoder_encoders_2_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16, weight = encoder_encoders_2_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized, x = x_49_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x84_1_cast_fp16 = add(x = input40_1_cast_fp16, y = linear_54_cast_fp16)[name = tensor<string, []>("x84_1_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1651456)))];
            tensor<fp16, [4, 1, 128]> var_3120_cast_fp16 = sub(x = x84_1_cast_fp16, y = encoder_encoders_2_encoder_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("op_3120_cast_fp16")];
            tensor<fp16, []> var_3114_promoted_to_fp16 = const()[name = tensor<string, []>("op_3114_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [4, 1, 128]> var_3121_cast_fp16 = pow(x = var_3120_cast_fp16, y = var_3114_promoted_to_fp16)[name = tensor<string, []>("op_3121_cast_fp16")];
            tensor<int32, [1]> var_3123_axes_0 = const()[name = tensor<string, []>("op_3123_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, []> var_3123_keep_dims_0 = const()[name = tensor<string, []>("op_3123_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1]> var_3123_cast_fp16 = reduce_mean(axes = var_3123_axes_0, keep_dims = var_3123_keep_dims_0, x = var_3121_cast_fp16)[name = tensor<string, []>("op_3123_cast_fp16")];
            tensor<fp16, []> var_3113_to_fp16 = const()[name = tensor<string, []>("op_3113_to_fp16"), val = tensor<fp16, []>(-0x1p-1)];
            tensor<fp16, [4, 1, 1]> var_3124_cast_fp16 = pow(x = var_3123_cast_fp16, y = var_3113_to_fp16)[name = tensor<string, []>("op_3124_cast_fp16")];
            tensor<fp16, []> var_3125_to_fp16 = const()[name = tensor<string, []>("op_3125_to_fp16"), val = tensor<fp16, []>(0x1.f28p-3)];
            tensor<fp16, [4, 1, 1]> scales_8_cast_fp16 = mul(x = var_3124_cast_fp16, y = var_3125_to_fp16)[name = tensor<string, []>("scales_8_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_3127_cast_fp16 = mul(x = x84_1_cast_fp16, y = scales_8_cast_fp16)[name = tensor<string, []>("op_3127_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_3130_cast_fp16 = sub(x = var_3127_cast_fp16, y = x_37_cast_fp16)[name = tensor<string, []>("op_3130_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_2_encoder_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_encoder_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1651776)))];
            tensor<fp16, [4, 1, 128]> var_3131_cast_fp16 = mul(x = var_3130_cast_fp16, y = encoder_encoders_2_encoder_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_3131_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_3132_cast_fp16 = add(x = x_37_cast_fp16, y = var_3131_cast_fp16)[name = tensor<string, []>("op_3132_cast_fp16")];
            tensor<int32, [1]> var_3142_axes_0 = const()[name = tensor<string, []>("op_3142_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 1, 128]> var_3142_cast_fp16 = expand_dims(axes = var_3142_axes_0, x = var_3132_cast_fp16)[name = tensor<string, []>("op_3142_cast_fp16")];
            tensor<int32, [4]> src_7_reps_0 = const()[name = tensor<string, []>("src_7_reps_0"), val = tensor<int32, [4]>([1, 4, 1, 1])];
            tensor<fp16, [4, 4, 1, 128]> src_7_cast_fp16 = tile(reps = src_7_reps_0, x = var_3142_cast_fp16)[name = tensor<string, []>("src_7_cast_fp16")];
            tensor<int32, [3]> var_3148 = const()[name = tensor<string, []>("op_3148"), val = tensor<int32, [3]>([16, 1, 128])];
            tensor<fp16, [16, 1, 128]> var_3149_cast_fp16 = reshape(shape = var_3148, x = src_7_cast_fp16)[name = tensor<string, []>("op_3149_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_3156_cast_fp16 = sub(x = var_3149_cast_fp16, y = var_2587_cast_fp16)[name = tensor<string, []>("op_3156_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_2_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_2_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1652096)))];
            tensor<fp16, [16, 1, 128]> var_3157_cast_fp16 = mul(x = var_3156_cast_fp16, y = encoder_encoders_2_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_3157_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_3158_cast_fp16 = add(x = var_2587_cast_fp16, y = var_3157_cast_fp16)[name = tensor<string, []>("op_3158_cast_fp16")];
            tensor<int32, [2]> key_padding_mask2_1_begin_0 = const()[name = tensor<string, []>("key_padding_mask2_1_begin_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [2]> key_padding_mask2_1_end_0 = const()[name = tensor<string, []>("key_padding_mask2_1_end_0"), val = tensor<int32, [2]>([1, 80])];
            tensor<bool, [2]> key_padding_mask2_1_end_mask_0 = const()[name = tensor<string, []>("key_padding_mask2_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<int32, [2]> key_padding_mask2_1_stride_0 = const()[name = tensor<string, []>("key_padding_mask2_1_stride_0"), val = tensor<int32, [2]>([1, 8])];
            tensor<bool, [1, 10]> key_padding_mask2_1 = slice_by_index(begin = key_padding_mask2_1_begin_0, end = key_padding_mask2_1_end_0, end_mask = key_padding_mask2_1_end_mask_0, stride = key_padding_mask2_1_stride_0, x = src_key_padding_mask0_1)[name = tensor<string, []>("key_padding_mask2_1")];
            tensor<int32, [4]> var_3180 = const()[name = tensor<string, []>("op_3180"), val = tensor<int32, [4]>([2, 8, 1, 128])];
            tensor<fp16, [2, 8, 1, 128]> src0_13_cast_fp16 = reshape(shape = var_3180, x = var_3158_cast_fp16)[name = tensor<string, []>("src0_13_cast_fp16")];
            tensor<fp16, [8, 1, 1]> weights0_6_to_fp16 = const()[name = tensor<string, []>("weights0_6_to_fp16"), val = tensor<fp16, [8, 1, 1]>([[[0x1.e28p-4]], [[0x1.cbcp-5]], [[0x1.3b4p-6]], [[0x1.0a4p-2]], [[0x1.66p-9]], [[0x1.99cp-9]], [[0x1.308p-10]], [[0x1.148p-1]]])];
            tensor<fp16, [2, 8, 1, 128]> var_3185_cast_fp16 = mul(x = src0_13_cast_fp16, y = weights0_6_to_fp16)[name = tensor<string, []>("op_3185_cast_fp16")];
            tensor<int32, [1]> x_51_axes_0 = const()[name = tensor<string, []>("x_51_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> x_51_keep_dims_0 = const()[name = tensor<string, []>("x_51_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [2, 1, 128]> x_51_cast_fp16 = reduce_sum(axes = x_51_axes_0, keep_dims = x_51_keep_dims_0, x = var_3185_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<fp16, [544, 128]> encoder_encoders_3_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [69632]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1652416))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1722112))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([544, 128])];
            tensor<fp16, [544]> encoder_encoders_3_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [544]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1722688)))];
            tensor<fp16, [2, 1, 544]> linear_55_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized, x = x_51_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<int32, [3]> q8_1_begin_0 = const()[name = tensor<string, []>("q8_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> q8_1_end_0 = const()[name = tensor<string, []>("q8_1_end_0"), val = tensor<int32, [3]>([2, 1, 256])];
            tensor<bool, [3]> q8_1_end_mask_0 = const()[name = tensor<string, []>("q8_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [2, 1, 256]> q8_1_cast_fp16 = slice_by_index(begin = q8_1_begin_0, end = q8_1_end_0, end_mask = q8_1_end_mask_0, x = linear_55_cast_fp16)[name = tensor<string, []>("q8_1_cast_fp16")];
            tensor<int32, [3]> k11_1_begin_0 = const()[name = tensor<string, []>("k11_1_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> k11_1_end_0 = const()[name = tensor<string, []>("k11_1_end_0"), val = tensor<int32, [3]>([2, 1, 512])];
            tensor<bool, [3]> k11_1_end_mask_0 = const()[name = tensor<string, []>("k11_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [2, 1, 256]> k11_1_cast_fp16 = slice_by_index(begin = k11_1_begin_0, end = k11_1_end_0, end_mask = k11_1_end_mask_0, x = linear_55_cast_fp16)[name = tensor<string, []>("k11_1_cast_fp16")];
            tensor<int32, [3]> p8_1_begin_0 = const()[name = tensor<string, []>("p8_1_begin_0"), val = tensor<int32, [3]>([0, 0, 512])];
            tensor<int32, [3]> p8_1_end_0 = const()[name = tensor<string, []>("p8_1_end_0"), val = tensor<int32, [3]>([2, 1, 544])];
            tensor<bool, [3]> p8_1_end_mask_0 = const()[name = tensor<string, []>("p8_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [2, 1, 32]> p8_1_cast_fp16 = slice_by_index(begin = p8_1_begin_0, end = p8_1_end_0, end_mask = p8_1_end_mask_0, x = linear_55_cast_fp16)[name = tensor<string, []>("p8_1_cast_fp16")];
            tensor<bool, []> k12_1_interleave_0 = const()[name = tensor<string, []>("k12_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s3_l0_key_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s3_l0_key_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [8, 1, 256]> cache_s3_l0_key_to_fp16 = cast(dtype = cache_s3_l0_key_to_fp16_dtype_0, x = cache_s3_l0_key)[name = tensor<string, []>("cast_17")];
            tensor<fp16, [10, 1, 256]> k12_1_cast_fp16 = concat(axis = var_46, interleave = k12_1_interleave_0, values = (cache_s3_l0_key_to_fp16, k11_1_cast_fp16))[name = tensor<string, []>("k12_1_cast_fp16")];
            tensor<int32, [3]> var_3238_begin_0 = const()[name = tensor<string, []>("op_3238_begin_0"), val = tensor<int32, [3]>([2, 0, 0])];
            tensor<int32, [3]> var_3238_end_0 = const()[name = tensor<string, []>("op_3238_end_0"), val = tensor<int32, [3]>([10, 1, 256])];
            tensor<bool, [3]> var_3238_end_mask_0 = const()[name = tensor<string, []>("op_3238_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 1, 256]> new_cache_s3_l0_key = slice_by_index(begin = var_3238_begin_0, end = var_3238_end_0, end_mask = var_3238_end_mask_0, x = k12_1_cast_fp16)[name = tensor<string, []>("op_3238_cast_fp16")];
            tensor<int32, [4]> var_3244 = const()[name = tensor<string, []>("op_3244"), val = tensor<int32, [4]>([2, 1, 8, 32])];
            tensor<fp16, [2, 1, 8, 32]> q9_1_cast_fp16 = reshape(shape = var_3244, x = q8_1_cast_fp16)[name = tensor<string, []>("q9_1_cast_fp16")];
            tensor<int32, [4]> var_3246 = const()[name = tensor<string, []>("op_3246"), val = tensor<int32, [4]>([2, 1, 8, 4])];
            tensor<fp16, [2, 1, 8, 4]> p9_1_cast_fp16 = reshape(shape = var_3246, x = p8_1_cast_fp16)[name = tensor<string, []>("p9_1_cast_fp16")];
            tensor<int32, [4]> var_3248 = const()[name = tensor<string, []>("op_3248"), val = tensor<int32, [4]>([10, 1, 8, 32])];
            tensor<fp16, [10, 1, 8, 32]> k13_1_cast_fp16 = reshape(shape = var_3248, x = k12_1_cast_fp16)[name = tensor<string, []>("k13_1_cast_fp16")];
            tensor<int32, [4]> var_3252 = const()[name = tensor<string, []>("op_3252"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> attn_scores8_1_transpose_x_0 = const()[name = tensor<string, []>("attn_scores8_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_scores8_1_transpose_y_0 = const()[name = tensor<string, []>("attn_scores8_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_30_perm_0 = const()[name = tensor<string, []>("transpose_30_perm_0"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<int32, [4]> transpose_31_perm_0 = const()[name = tensor<string, []>("transpose_31_perm_0"), val = tensor<int32, [4]>([2, 1, -1, -4])];
            tensor<fp16, [8, 1, 32, 10]> transpose_31 = transpose(perm = transpose_31_perm_0, x = k13_1_cast_fp16)[name = tensor<string, []>("transpose_74")];
            tensor<fp16, [8, 1, 2, 32]> transpose_30 = transpose(perm = transpose_30_perm_0, x = q9_1_cast_fp16)[name = tensor<string, []>("transpose_75")];
            tensor<fp16, [8, 1, 2, 10]> attn_scores8_1_cast_fp16 = matmul(transpose_x = attn_scores8_1_transpose_x_0, transpose_y = attn_scores8_1_transpose_y_0, x = transpose_30, y = transpose_31)[name = tensor<string, []>("attn_scores8_1_cast_fp16")];
            tensor<bool, []> pos_scores11_1_transpose_x_0 = const()[name = tensor<string, []>("pos_scores11_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_scores11_1_transpose_y_0 = const()[name = tensor<string, []>("pos_scores11_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [8, 1, 4, 11]> pos_emb2_1_to_fp16 = const()[name = tensor<string, []>("pos_emb2_1_to_fp16"), val = tensor<fp16, [8, 1, 4, 11]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1723840)))];
            tensor<fp16, [8, 1, 2, 4]> p10_1_cast_fp16 = transpose(perm = var_3252, x = p9_1_cast_fp16)[name = tensor<string, []>("transpose_73")];
            tensor<fp16, [8, 1, 2, 11]> pos_scores11_1_cast_fp16 = matmul(transpose_x = pos_scores11_1_transpose_x_0, transpose_y = pos_scores11_1_transpose_y_0, x = p10_1_cast_fp16, y = pos_emb2_1_to_fp16)[name = tensor<string, []>("pos_scores11_1_cast_fp16")];
            tensor<int32, [2]> var_3295 = const()[name = tensor<string, []>("op_3295"), val = tensor<int32, [2]>([-1, 11])];
            tensor<fp16, [16, 11]> pos_scores12_1_cast_fp16 = reshape(shape = var_3295, x = pos_scores11_1_cast_fp16)[name = tensor<string, []>("pos_scores12_1_cast_fp16")];
            tensor<bool, []> pos_scores13_1_validate_indices_0 = const()[name = tensor<string, []>("pos_scores13_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<uint16, [16, 10]> indexes2_1_to_uint16 = const()[name = tensor<string, []>("indexes2_1_to_uint16"), val = tensor<uint16, [16, 10]>([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])];
            tensor<fp16, [16, 10]> pos_scores13_1_cast_fp16_cast_uint16 = gather_along_axis(axis = var_48, indices = indexes2_1_to_uint16, validate_indices = pos_scores13_1_validate_indices_0, x = pos_scores12_1_cast_fp16)[name = tensor<string, []>("pos_scores13_1_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_3298 = const()[name = tensor<string, []>("op_3298"), val = tensor<int32, [4]>([8, 1, 2, 10])];
            tensor<fp16, [8, 1, 2, 10]> pos_scores14_1_cast_fp16 = reshape(shape = var_3298, x = pos_scores13_1_cast_fp16_cast_uint16)[name = tensor<string, []>("pos_scores14_1_cast_fp16")];
            tensor<fp16, [8, 1, 2, 10]> attn_scores9_1_cast_fp16 = add(x = attn_scores8_1_cast_fp16, y = pos_scores14_1_cast_fp16)[name = tensor<string, []>("attn_scores9_1_cast_fp16")];
            tensor<int32, [1]> var_3301_axes_0 = const()[name = tensor<string, []>("op_3301_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 10]> var_3301 = expand_dims(axes = var_3301_axes_0, x = key_padding_mask2_1)[name = tensor<string, []>("op_3301")];
            tensor<fp16, []> cast_69_to_fp16 = const()[name = tensor<string, []>("cast_69_to_fp16"), val = tensor<fp16, []>(-0x1.f4p+9)];
            tensor<fp16, [8, 1, 2, 10]> attn_scores10_1_cast_fp16 = select(a = cast_69_to_fp16, b = attn_scores9_1_cast_fp16, cond = var_3301)[name = tensor<string, []>("attn_scores10_1_cast_fp16")];
            tensor<fp16, [8, 1, 2, 10]> attn_weights5_1_cast_fp16 = softmax(axis = var_59, x = attn_scores10_1_cast_fp16)[name = tensor<string, []>("attn_weights5_1_cast_fp16")];
            tensor<fp16, [144, 128]> encoder_encoders_3_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1724608))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1743104))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([144, 128])];
            tensor<fp16, [144]> encoder_encoders_3_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1743680)))];
            tensor<fp16, [2, 1, 144]> linear_57_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized, x = x_51_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
            tensor<fp16, []> var_3316_to_fp16 = const()[name = tensor<string, []>("op_3316_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [2, 1, 144]> x_offset_32_cast_fp16 = sub(x = linear_57_cast_fp16, y = var_3316_to_fp16)[name = tensor<string, []>("x_offset_32_cast_fp16")];
            tensor<fp16, [2, 1, 144]> var_3318_cast_fp16 = exp(x = x_offset_32_cast_fp16)[name = tensor<string, []>("op_3318_cast_fp16")];
            tensor<fp16, []> var_3319_to_fp16 = const()[name = tensor<string, []>("op_3319_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [2, 1, 144]> var_3320_cast_fp16 = add(x = var_3318_cast_fp16, y = var_3319_to_fp16)[name = tensor<string, []>("op_3320_cast_fp16")];
            tensor<fp32, []> var_3321_epsilon_0 = const()[name = tensor<string, []>("op_3321_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [2, 1, 144]> var_3321_cast_fp16 = log(epsilon = var_3321_epsilon_0, x = var_3320_cast_fp16)[name = tensor<string, []>("op_3321_cast_fp16")];
            tensor<fp16, []> var_3304_to_fp16 = const()[name = tensor<string, []>("op_3304_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [2, 1, 144]> var_3323_cast_fp16 = equal(x = var_3321_cast_fp16, y = var_3304_to_fp16)[name = tensor<string, []>("op_3323_cast_fp16")];
            tensor<fp16, [2, 1, 144]> log_sum0_32_cast_fp16 = select(a = x_offset_32_cast_fp16, b = var_3321_cast_fp16, cond = var_3323_cast_fp16)[name = tensor<string, []>("log_sum0_32_cast_fp16")];
            tensor<fp16, []> var_3325_to_fp16 = const()[name = tensor<string, []>("op_3325_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [2, 1, 144]> var_3326_cast_fp16 = mul(x = linear_57_cast_fp16, y = var_3325_to_fp16)[name = tensor<string, []>("op_3326_cast_fp16")];
            tensor<fp16, [2, 1, 144]> var_3327_cast_fp16 = sub(x = log_sum0_32_cast_fp16, y = var_3326_cast_fp16)[name = tensor<string, []>("op_3327_cast_fp16")];
            tensor<fp16, []> var_3328_to_fp16 = const()[name = tensor<string, []>("op_3328_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [2, 1, 144]> x_57_cast_fp16 = sub(x = var_3327_cast_fp16, y = var_3328_to_fp16)[name = tensor<string, []>("x_57_cast_fp16")];
            tensor<fp16, [128, 144]> encoder_encoders_3_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1744064))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1762560))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 144])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1763136)))];
            tensor<fp16, [2, 1, 128]> linear_58_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized, x = x_57_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
            tensor<fp16, [2, 1, 128]> input41_1_cast_fp16 = add(x = x_51_cast_fp16, y = linear_58_cast_fp16)[name = tensor<string, []>("input41_1_cast_fp16")];
            tensor<int32, [4]> attn_weights6_1_begin_0 = const()[name = tensor<string, []>("attn_weights6_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_weights6_1_end_0 = const()[name = tensor<string, []>("attn_weights6_1_end_0"), val = tensor<int32, [4]>([1, 1, 2, 10])];
            tensor<bool, [4]> attn_weights6_1_end_mask_0 = const()[name = tensor<string, []>("attn_weights6_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1, 2, 10]> attn_weights6_1_cast_fp16 = slice_by_index(begin = attn_weights6_1_begin_0, end = attn_weights6_1_end_0, end_mask = attn_weights6_1_end_mask_0, x = attn_weights5_1_cast_fp16)[name = tensor<string, []>("attn_weights6_1_cast_fp16")];
            tensor<fp16, [288, 128]> encoder_encoders_3_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [36864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1763456))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1800384))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([288, 128])];
            tensor<fp16, [288]> encoder_encoders_3_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1800960)))];
            tensor<fp16, [2, 1, 288]> linear_59_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized, x = input41_1_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
            tensor<int32, [3]> var_3342_split_sizes_0 = const()[name = tensor<string, []>("op_3342_split_sizes_0"), val = tensor<int32, [3]>([96, 96, 96])];
            tensor<int32, []> var_3342_axis_0 = const()[name = tensor<string, []>("op_3342_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [2, 1, 96]> var_3342_cast_fp16_0, tensor<fp16, [2, 1, 96]> var_3342_cast_fp16_1, tensor<fp16, [2, 1, 96]> var_3342_cast_fp16_2 = split(axis = var_3342_axis_0, split_sizes = var_3342_split_sizes_0, x = linear_59_cast_fp16)[name = tensor<string, []>("op_3342_cast_fp16")];
            tensor<fp16, [2, 1, 96]> var_3345_cast_fp16 = tanh(x = var_3342_cast_fp16_0)[name = tensor<string, []>("op_3345_cast_fp16")];
            tensor<int32, [1]> var_3346_axes_0 = const()[name = tensor<string, []>("op_3346_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [2, 1, 96, 1]> var_3346_cast_fp16 = expand_dims(axes = var_3346_axes_0, x = var_3345_cast_fp16)[name = tensor<string, []>("op_3346_cast_fp16")];
            tensor<int32, [3]> var_3347 = const()[name = tensor<string, []>("op_3347"), val = tensor<int32, [3]>([2, 1, 96])];
            tensor<fp16, [2, 1, 96]> s2_1_cast_fp16 = reshape(shape = var_3347, x = var_3346_cast_fp16)[name = tensor<string, []>("s2_1_cast_fp16")];
            tensor<fp16, [2, 1, 96]> x86_1_cast_fp16 = mul(x = var_3342_cast_fp16_1, y = s2_1_cast_fp16)[name = tensor<string, []>("x86_1_cast_fp16")];
            tensor<int32, [4]> var_3361 = const()[name = tensor<string, []>("op_3361"), val = tensor<int32, [4]>([2, 1, 1, -1])];
            tensor<fp16, [2, 1, 1, 96]> var_3362_cast_fp16 = reshape(shape = var_3361, x = x86_1_cast_fp16)[name = tensor<string, []>("op_3362_cast_fp16")];
            tensor<int32, [4]> var_3363 = const()[name = tensor<string, []>("op_3363"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<bool, []> x_pad2_1_interleave_0 = const()[name = tensor<string, []>("x_pad2_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s3_l0_nonlin_attn_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s3_l0_nonlin_attn_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 2, 96]> x87_1_cast_fp16 = transpose(perm = var_3363, x = var_3362_cast_fp16)[name = tensor<string, []>("transpose_72")];
            tensor<fp16, [1, 1, 8, 96]> cache_s3_l0_nonlin_attn_to_fp16 = cast(dtype = cache_s3_l0_nonlin_attn_to_fp16_dtype_0, x = cache_s3_l0_nonlin_attn)[name = tensor<string, []>("cast_16")];
            tensor<fp16, [1, 1, 10, 96]> x_pad2_1_cast_fp16 = concat(axis = var_49, interleave = x_pad2_1_interleave_0, values = (cache_s3_l0_nonlin_attn_to_fp16, x87_1_cast_fp16))[name = tensor<string, []>("x_pad2_1_cast_fp16")];
            tensor<int32, [4]> var_3369_begin_0 = const()[name = tensor<string, []>("op_3369_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_3369_end_0 = const()[name = tensor<string, []>("op_3369_end_0"), val = tensor<int32, [4]>([1, 1, 10, 96])];
            tensor<bool, [4]> var_3369_end_mask_0 = const()[name = tensor<string, []>("op_3369_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 8, 96]> new_cache_s3_l0_nonlin_attn = slice_by_index(begin = var_3369_begin_0, end = var_3369_end_0, end_mask = var_3369_end_mask_0, x = x_pad2_1_cast_fp16)[name = tensor<string, []>("op_3369_cast_fp16")];
            tensor<bool, []> x88_1_transpose_x_0 = const()[name = tensor<string, []>("x88_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x88_1_transpose_y_0 = const()[name = tensor<string, []>("x88_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 2, 96]> x88_1_cast_fp16 = matmul(transpose_x = x88_1_transpose_x_0, transpose_y = x88_1_transpose_y_0, x = attn_weights6_1_cast_fp16, y = x_pad2_1_cast_fp16)[name = tensor<string, []>("x88_1_cast_fp16")];
            tensor<int32, [4]> var_3372 = const()[name = tensor<string, []>("op_3372"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_3374 = const()[name = tensor<string, []>("op_3374"), val = tensor<int32, [3]>([2, 1, -1])];
            tensor<fp16, [2, 1, 1, 96]> var_3373_cast_fp16 = transpose(perm = var_3372, x = x88_1_cast_fp16)[name = tensor<string, []>("transpose_71")];
            tensor<fp16, [2, 1, 96]> x89_1_cast_fp16 = reshape(shape = var_3374, x = var_3373_cast_fp16)[name = tensor<string, []>("x89_1_cast_fp16")];
            tensor<fp16, [2, 1, 96]> input43_1_cast_fp16 = mul(x = x89_1_cast_fp16, y = var_3342_cast_fp16_2)[name = tensor<string, []>("input43_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_3_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1801600))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1813952))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1814528)))];
            tensor<fp16, [2, 1, 128]> linear_60_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized, x = input43_1_cast_fp16)[name = tensor<string, []>("linear_60_cast_fp16")];
            tensor<fp16, [2, 1, 128]> x90_1_cast_fp16 = add(x = input41_1_cast_fp16, y = linear_60_cast_fp16)[name = tensor<string, []>("x90_1_cast_fp16")];
            tensor<fp16, [96, 128]> encoder_encoders_3_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1814848))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1827200))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([96, 128])];
            tensor<fp16, [96]> encoder_encoders_3_encoder_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [96]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1827776)))];
            tensor<fp16, [2, 1, 96]> linear_61_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_self_attn1_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized, x = x90_1_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
            tensor<bool, []> x91_1_interleave_0 = const()[name = tensor<string, []>("x91_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s3_l0_val1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s3_l0_val1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [8, 1, 96]> cache_s3_l0_val1_to_fp16 = cast(dtype = cache_s3_l0_val1_to_fp16_dtype_0, x = cache_s3_l0_val1)[name = tensor<string, []>("cast_15")];
            tensor<fp16, [10, 1, 96]> x91_1_cast_fp16 = concat(axis = var_46, interleave = x91_1_interleave_0, values = (cache_s3_l0_val1_to_fp16, linear_61_cast_fp16))[name = tensor<string, []>("x91_1_cast_fp16")];
            tensor<int32, [3]> var_3399_begin_0 = const()[name = tensor<string, []>("op_3399_begin_0"), val = tensor<int32, [3]>([2, 0, 0])];
            tensor<int32, [3]> var_3399_end_0 = const()[name = tensor<string, []>("op_3399_end_0"), val = tensor<int32, [3]>([10, 1, 96])];
            tensor<bool, [3]> var_3399_end_mask_0 = const()[name = tensor<string, []>("op_3399_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 1, 96]> new_cache_s3_l0_val1 = slice_by_index(begin = var_3399_begin_0, end = var_3399_end_0, end_mask = var_3399_end_mask_0, x = x91_1_cast_fp16)[name = tensor<string, []>("op_3399_cast_fp16")];
            tensor<int32, [4]> var_3400 = const()[name = tensor<string, []>("op_3400"), val = tensor<int32, [4]>([10, 1, 8, -1])];
            tensor<fp16, [10, 1, 8, 12]> var_3401_cast_fp16 = reshape(shape = var_3400, x = x91_1_cast_fp16)[name = tensor<string, []>("op_3401_cast_fp16")];
            tensor<int32, [4]> var_3402 = const()[name = tensor<string, []>("op_3402"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x93_1_transpose_x_0 = const()[name = tensor<string, []>("x93_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x93_1_transpose_y_0 = const()[name = tensor<string, []>("x93_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [8, 1, 10, 12]> x92_1_cast_fp16 = transpose(perm = var_3402, x = var_3401_cast_fp16)[name = tensor<string, []>("transpose_70")];
            tensor<fp16, [8, 1, 2, 12]> x93_1_cast_fp16 = matmul(transpose_x = x93_1_transpose_x_0, transpose_y = x93_1_transpose_y_0, x = attn_weights5_1_cast_fp16, y = x92_1_cast_fp16)[name = tensor<string, []>("x93_1_cast_fp16")];
            tensor<int32, [4]> var_3407 = const()[name = tensor<string, []>("op_3407"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_3412 = const()[name = tensor<string, []>("op_3412"), val = tensor<int32, [3]>([2, 1, 96])];
            tensor<fp16, [2, 1, 8, 12]> var_3408_cast_fp16 = transpose(perm = var_3407, x = x93_1_cast_fp16)[name = tensor<string, []>("transpose_69")];
            tensor<fp16, [2, 1, 96]> input44_1_cast_fp16 = reshape(shape = var_3412, x = var_3408_cast_fp16)[name = tensor<string, []>("input44_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_3_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1828032))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1840384))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1840960)))];
            tensor<fp16, [2, 1, 128]> linear_62_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_self_attn1_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized, x = input44_1_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
            tensor<fp16, [2, 1, 128]> input45_1_cast_fp16 = add(x = x90_1_cast_fp16, y = linear_62_cast_fp16)[name = tensor<string, []>("input45_1_cast_fp16")];
            tensor<int32, [2]> src_key_padding_mask7_1_begin_0 = const()[name = tensor<string, []>("src_key_padding_mask7_1_begin_0"), val = tensor<int32, [2]>([0, 8])];
            tensor<int32, [2]> src_key_padding_mask7_1_end_0 = const()[name = tensor<string, []>("src_key_padding_mask7_1_end_0"), val = tensor<int32, [2]>([1, 10])];
            tensor<bool, [2]> src_key_padding_mask7_1_end_mask_0 = const()[name = tensor<string, []>("src_key_padding_mask7_1_end_mask_0"), val = tensor<bool, [2]>([true, true])];
            tensor<bool, [1, 2]> src_key_padding_mask7_1 = slice_by_index(begin = src_key_padding_mask7_1_begin_0, end = src_key_padding_mask7_1_end_0, end_mask = src_key_padding_mask7_1_end_mask_0, x = key_padding_mask2_1)[name = tensor<string, []>("src_key_padding_mask7_1")];
            tensor<fp16, [256, 128]> encoder_encoders_3_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1841280))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1874112))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_3_encoder_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1874688)))];
            tensor<fp16, [2, 1, 256]> linear_63_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_conv_module1_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized, x = input45_1_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
            tensor<int32, [2]> var_3423_split_sizes_0 = const()[name = tensor<string, []>("op_3423_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_3423_axis_0 = const()[name = tensor<string, []>("op_3423_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [2, 1, 128]> var_3423_cast_fp16_0, tensor<fp16, [2, 1, 128]> var_3423_cast_fp16_1 = split(axis = var_3423_axis_0, split_sizes = var_3423_split_sizes_0, x = linear_63_cast_fp16)[name = tensor<string, []>("op_3423_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3425_cast_fp16 = sigmoid(x = var_3423_cast_fp16_1)[name = tensor<string, []>("op_3425_cast_fp16")];
            tensor<fp16, [2, 1, 128]> x95_1_cast_fp16 = mul(x = var_3423_cast_fp16_0, y = var_3425_cast_fp16)[name = tensor<string, []>("x95_1_cast_fp16")];
            tensor<int32, [3]> var_3427 = const()[name = tensor<string, []>("op_3427"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<int32, [1]> var_3429_axes_0 = const()[name = tensor<string, []>("op_3429_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 2]> var_3429 = expand_dims(axes = var_3429_axes_0, x = src_key_padding_mask7_1)[name = tensor<string, []>("op_3429")];
            tensor<int32, [3]> var_3430_reps_0 = const()[name = tensor<string, []>("op_3430_reps_0"), val = tensor<int32, [3]>([1, 128, 1])];
            tensor<bool, [1, 128, 2]> var_3430 = tile(reps = var_3430_reps_0, x = var_3429)[name = tensor<string, []>("op_3430")];
            tensor<fp16, [1, 128, 2]> x96_1_cast_fp16 = transpose(perm = var_3427, x = x95_1_cast_fp16)[name = tensor<string, []>("transpose_68")];
            tensor<fp16, [1, 128, 2]> x97_1_cast_fp16 = select(a = var_62_to_fp16, b = x96_1_cast_fp16, cond = var_3430)[name = tensor<string, []>("x97_1_cast_fp16")];
            tensor<bool, []> x98_1_interleave_0 = const()[name = tensor<string, []>("x98_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s3_l0_conv1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s3_l0_conv1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 7]> cache_s3_l0_conv1_to_fp16 = cast(dtype = cache_s3_l0_conv1_to_fp16_dtype_0, x = cache_s3_l0_conv1)[name = tensor<string, []>("cast_14")];
            tensor<fp16, [1, 128, 9]> x98_1_cast_fp16 = concat(axis = var_49, interleave = x98_1_interleave_0, values = (cache_s3_l0_conv1_to_fp16, x97_1_cast_fp16))[name = tensor<string, []>("x98_1_cast_fp16")];
            tensor<int32, [3]> var_3437_begin_0 = const()[name = tensor<string, []>("op_3437_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
            tensor<int32, [3]> var_3437_end_0 = const()[name = tensor<string, []>("op_3437_end_0"), val = tensor<int32, [3]>([1, 128, 9])];
            tensor<bool, [3]> var_3437_end_mask_0 = const()[name = tensor<string, []>("op_3437_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 7]> new_cache_s3_l0_conv1 = slice_by_index(begin = var_3437_begin_0, end = var_3437_end_0, end_mask = var_3437_end_mask_0, x = x98_1_cast_fp16)[name = tensor<string, []>("op_3437_cast_fp16")];
            tensor<string, []> x_causal_14_pad_type_0 = const()[name = tensor<string, []>("x_causal_14_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_14_groups_0 = const()[name = tensor<string, []>("x_causal_14_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_14_strides_0 = const()[name = tensor<string, []>("x_causal_14_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_14_pad_0 = const()[name = tensor<string, []>("x_causal_14_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_14_dilations_0 = const()[name = tensor<string, []>("x_causal_14_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 8]> encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1875264)))];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1877376)))];
            tensor<fp16, [1, 128, 2]> x_causal_14_cast_fp16 = conv(bias = encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_14_dilations_0, groups = x_causal_14_groups_0, pad = x_causal_14_pad_0, pad_type = x_causal_14_pad_type_0, strides = x_causal_14_strides_0, weight = encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16, x = x98_1_cast_fp16)[name = tensor<string, []>("x_causal_14_cast_fp16")];
            tensor<int32, [3]> input47_1_begin_0 = const()[name = tensor<string, []>("input47_1_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> input47_1_end_0 = const()[name = tensor<string, []>("input47_1_end_0"), val = tensor<int32, [3]>([1, 128, 9])];
            tensor<bool, [3]> input47_1_end_mask_0 = const()[name = tensor<string, []>("input47_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 2]> input47_1_cast_fp16 = slice_by_index(begin = input47_1_begin_0, end = input47_1_end_0, end_mask = input47_1_end_mask_0, x = x98_1_cast_fp16)[name = tensor<string, []>("input47_1_cast_fp16")];
            tensor<string, []> x_chunk_17_pad_type_0 = const()[name = tensor<string, []>("x_chunk_17_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_17_pad_0 = const()[name = tensor<string, []>("x_chunk_17_pad_0"), val = tensor<int32, [2]>([7, 7])];
            tensor<int32, []> x_chunk_17_groups_0 = const()[name = tensor<string, []>("x_chunk_17_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_17_strides_0 = const()[name = tensor<string, []>("x_chunk_17_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_17_dilations_0 = const()[name = tensor<string, []>("x_chunk_17_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 15]> encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1877696)))];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1881600)))];
            tensor<fp16, [1, 128, 2]> x_chunk_17_cast_fp16 = conv(bias = encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_17_dilations_0, groups = x_chunk_17_groups_0, pad = x_chunk_17_pad_0, pad_type = x_chunk_17_pad_type_0, strides = x_chunk_17_strides_0, weight = encoder_encoders_3_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16, x = input47_1_cast_fp16)[name = tensor<string, []>("x_chunk_17_cast_fp16")];
            tensor<fp16, [128, 2]> chunk_scale5_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale5_1_to_fp16"), val = tensor<fp16, [128, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1881920)))];
            tensor<fp16, [1, 128, 2]> x_chunk5_1_cast_fp16 = mul(x = x_chunk_17_cast_fp16, y = chunk_scale5_1_to_fp16)[name = tensor<string, []>("x_chunk5_1_cast_fp16")];
            tensor<fp16, [1, 128, 2]> x99_1_cast_fp16 = add(x = x_chunk5_1_cast_fp16, y = x_causal_14_cast_fp16)[name = tensor<string, []>("x99_1_cast_fp16")];
            tensor<int32, [3]> var_3477 = const()[name = tensor<string, []>("op_3477"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_3486_to_fp16 = const()[name = tensor<string, []>("op_3486_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [2, 1, 128]> x100_1_cast_fp16 = transpose(perm = var_3477, x = x99_1_cast_fp16)[name = tensor<string, []>("transpose_67")];
            tensor<fp16, [2, 1, 128]> x_offset_34_cast_fp16 = sub(x = x100_1_cast_fp16, y = var_3486_to_fp16)[name = tensor<string, []>("x_offset_34_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3488_cast_fp16 = exp(x = x_offset_34_cast_fp16)[name = tensor<string, []>("op_3488_cast_fp16")];
            tensor<fp16, []> var_3489_to_fp16 = const()[name = tensor<string, []>("op_3489_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [2, 1, 128]> var_3490_cast_fp16 = add(x = var_3488_cast_fp16, y = var_3489_to_fp16)[name = tensor<string, []>("op_3490_cast_fp16")];
            tensor<fp32, []> var_3491_epsilon_0 = const()[name = tensor<string, []>("op_3491_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [2, 1, 128]> var_3491_cast_fp16 = log(epsilon = var_3491_epsilon_0, x = var_3490_cast_fp16)[name = tensor<string, []>("op_3491_cast_fp16")];
            tensor<fp16, []> var_3483_to_fp16 = const()[name = tensor<string, []>("op_3483_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [2, 1, 128]> var_3493_cast_fp16 = equal(x = var_3491_cast_fp16, y = var_3483_to_fp16)[name = tensor<string, []>("op_3493_cast_fp16")];
            tensor<fp16, [2, 1, 128]> log_sum0_34_cast_fp16 = select(a = x_offset_34_cast_fp16, b = var_3491_cast_fp16, cond = var_3493_cast_fp16)[name = tensor<string, []>("log_sum0_34_cast_fp16")];
            tensor<fp16, []> var_3495_to_fp16 = const()[name = tensor<string, []>("op_3495_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [2, 1, 128]> var_3496_cast_fp16 = mul(x = x100_1_cast_fp16, y = var_3495_to_fp16)[name = tensor<string, []>("op_3496_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3497_cast_fp16 = sub(x = log_sum0_34_cast_fp16, y = var_3496_cast_fp16)[name = tensor<string, []>("op_3497_cast_fp16")];
            tensor<fp16, []> var_3498_to_fp16 = const()[name = tensor<string, []>("op_3498_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [2, 1, 128]> x0_21_cast_fp16 = sub(x = var_3497_cast_fp16, y = var_3498_to_fp16)[name = tensor<string, []>("x0_21_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_3_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1882496))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1898944))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1899520)))];
            tensor<fp16, [2, 1, 128]> linear_64_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_conv_module1_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized, x = x0_21_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
            tensor<fp16, [2, 1, 128]> input48_1_cast_fp16 = add(x = input45_1_cast_fp16, y = linear_64_cast_fp16)[name = tensor<string, []>("input48_1_cast_fp16")];
            tensor<fp16, [192, 128]> encoder_encoders_3_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1899840))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1924480))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([192, 128])];
            tensor<fp16, [192]> encoder_encoders_3_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1925056)))];
            tensor<fp16, [2, 1, 192]> linear_65_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized, x = input48_1_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
            tensor<fp16, []> var_3514_to_fp16 = const()[name = tensor<string, []>("op_3514_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [2, 1, 192]> x_offset_36_cast_fp16 = sub(x = linear_65_cast_fp16, y = var_3514_to_fp16)[name = tensor<string, []>("x_offset_36_cast_fp16")];
            tensor<fp16, [2, 1, 192]> var_3516_cast_fp16 = exp(x = x_offset_36_cast_fp16)[name = tensor<string, []>("op_3516_cast_fp16")];
            tensor<fp16, []> var_3517_to_fp16 = const()[name = tensor<string, []>("op_3517_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [2, 1, 192]> var_3518_cast_fp16 = add(x = var_3516_cast_fp16, y = var_3517_to_fp16)[name = tensor<string, []>("op_3518_cast_fp16")];
            tensor<fp32, []> var_3519_epsilon_0 = const()[name = tensor<string, []>("op_3519_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [2, 1, 192]> var_3519_cast_fp16 = log(epsilon = var_3519_epsilon_0, x = var_3518_cast_fp16)[name = tensor<string, []>("op_3519_cast_fp16")];
            tensor<fp16, []> var_3502_to_fp16 = const()[name = tensor<string, []>("op_3502_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [2, 1, 192]> var_3521_cast_fp16 = equal(x = var_3519_cast_fp16, y = var_3502_to_fp16)[name = tensor<string, []>("op_3521_cast_fp16")];
            tensor<fp16, [2, 1, 192]> log_sum0_36_cast_fp16 = select(a = x_offset_36_cast_fp16, b = var_3519_cast_fp16, cond = var_3521_cast_fp16)[name = tensor<string, []>("log_sum0_36_cast_fp16")];
            tensor<fp16, []> var_3523_to_fp16 = const()[name = tensor<string, []>("op_3523_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [2, 1, 192]> var_3524_cast_fp16 = mul(x = linear_65_cast_fp16, y = var_3523_to_fp16)[name = tensor<string, []>("op_3524_cast_fp16")];
            tensor<fp16, [2, 1, 192]> var_3525_cast_fp16 = sub(x = log_sum0_36_cast_fp16, y = var_3524_cast_fp16)[name = tensor<string, []>("op_3525_cast_fp16")];
            tensor<fp16, []> var_3526_to_fp16 = const()[name = tensor<string, []>("op_3526_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [2, 1, 192]> x_59_cast_fp16 = sub(x = var_3525_cast_fp16, y = var_3526_to_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
            tensor<fp16, [128, 192]> encoder_encoders_3_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1925504))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1950144))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 192])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1950720)))];
            tensor<fp16, [2, 1, 128]> linear_66_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized, x = x_59_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
            tensor<fp16, [2, 1, 128]> src7_1_cast_fp16 = add(x = input48_1_cast_fp16, y = linear_66_cast_fp16)[name = tensor<string, []>("src7_1_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3532_cast_fp16 = sub(x = src7_1_cast_fp16, y = x_51_cast_fp16)[name = tensor<string, []>("op_3532_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1951040)))];
            tensor<fp16, [2, 1, 128]> var_3533_cast_fp16 = mul(x = var_3532_cast_fp16, y = encoder_encoders_3_encoder_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_3533_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3534_cast_fp16 = add(x = x_51_cast_fp16, y = var_3533_cast_fp16)[name = tensor<string, []>("op_3534_cast_fp16")];
            tensor<fp16, [96, 128]> encoder_encoders_3_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1951360))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1963712))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([96, 128])];
            tensor<fp16, [96]> encoder_encoders_3_encoder_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [96]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1964288)))];
            tensor<fp16, [2, 1, 96]> linear_67_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_self_attn2_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized, x = var_3534_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
            tensor<bool, []> x101_1_interleave_0 = const()[name = tensor<string, []>("x101_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s3_l0_val2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s3_l0_val2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [8, 1, 96]> cache_s3_l0_val2_to_fp16 = cast(dtype = cache_s3_l0_val2_to_fp16_dtype_0, x = cache_s3_l0_val2)[name = tensor<string, []>("cast_13")];
            tensor<fp16, [10, 1, 96]> x101_1_cast_fp16 = concat(axis = var_46, interleave = x101_1_interleave_0, values = (cache_s3_l0_val2_to_fp16, linear_67_cast_fp16))[name = tensor<string, []>("x101_1_cast_fp16")];
            tensor<int32, [3]> var_3553_begin_0 = const()[name = tensor<string, []>("op_3553_begin_0"), val = tensor<int32, [3]>([2, 0, 0])];
            tensor<int32, [3]> var_3553_end_0 = const()[name = tensor<string, []>("op_3553_end_0"), val = tensor<int32, [3]>([10, 1, 96])];
            tensor<bool, [3]> var_3553_end_mask_0 = const()[name = tensor<string, []>("op_3553_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 1, 96]> new_cache_s3_l0_val2 = slice_by_index(begin = var_3553_begin_0, end = var_3553_end_0, end_mask = var_3553_end_mask_0, x = x101_1_cast_fp16)[name = tensor<string, []>("op_3553_cast_fp16")];
            tensor<int32, [4]> var_3554 = const()[name = tensor<string, []>("op_3554"), val = tensor<int32, [4]>([10, 1, 8, -1])];
            tensor<fp16, [10, 1, 8, 12]> var_3555_cast_fp16 = reshape(shape = var_3554, x = x101_1_cast_fp16)[name = tensor<string, []>("op_3555_cast_fp16")];
            tensor<int32, [4]> var_3556 = const()[name = tensor<string, []>("op_3556"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x103_1_transpose_x_0 = const()[name = tensor<string, []>("x103_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x103_1_transpose_y_0 = const()[name = tensor<string, []>("x103_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [8, 1, 10, 12]> x102_1_cast_fp16 = transpose(perm = var_3556, x = var_3555_cast_fp16)[name = tensor<string, []>("transpose_66")];
            tensor<fp16, [8, 1, 2, 12]> x103_1_cast_fp16 = matmul(transpose_x = x103_1_transpose_x_0, transpose_y = x103_1_transpose_y_0, x = attn_weights5_1_cast_fp16, y = x102_1_cast_fp16)[name = tensor<string, []>("x103_1_cast_fp16")];
            tensor<int32, [4]> var_3561 = const()[name = tensor<string, []>("op_3561"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_3566 = const()[name = tensor<string, []>("op_3566"), val = tensor<int32, [3]>([2, 1, 96])];
            tensor<fp16, [2, 1, 8, 12]> var_3562_cast_fp16 = transpose(perm = var_3561, x = x103_1_cast_fp16)[name = tensor<string, []>("transpose_65")];
            tensor<fp16, [2, 1, 96]> input49_1_cast_fp16 = reshape(shape = var_3566, x = var_3562_cast_fp16)[name = tensor<string, []>("input49_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_3_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1964544))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1976896))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1977472)))];
            tensor<fp16, [2, 1, 128]> linear_68_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_self_attn2_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized, x = input49_1_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
            tensor<fp16, [2, 1, 128]> input50_1_cast_fp16 = add(x = var_3534_cast_fp16, y = linear_68_cast_fp16)[name = tensor<string, []>("input50_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_3_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1977792))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2010624))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_3_encoder_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2011200)))];
            tensor<fp16, [2, 1, 256]> linear_69_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_conv_module2_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized, x = input50_1_cast_fp16)[name = tensor<string, []>("linear_69_cast_fp16")];
            tensor<int32, [2]> var_3577_split_sizes_0 = const()[name = tensor<string, []>("op_3577_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_3577_axis_0 = const()[name = tensor<string, []>("op_3577_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [2, 1, 128]> var_3577_cast_fp16_0, tensor<fp16, [2, 1, 128]> var_3577_cast_fp16_1 = split(axis = var_3577_axis_0, split_sizes = var_3577_split_sizes_0, x = linear_69_cast_fp16)[name = tensor<string, []>("op_3577_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3579_cast_fp16 = sigmoid(x = var_3577_cast_fp16_1)[name = tensor<string, []>("op_3579_cast_fp16")];
            tensor<fp16, [2, 1, 128]> x105_1_cast_fp16 = mul(x = var_3577_cast_fp16_0, y = var_3579_cast_fp16)[name = tensor<string, []>("x105_1_cast_fp16")];
            tensor<int32, [3]> var_3581 = const()[name = tensor<string, []>("op_3581"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 2]> x106_1_cast_fp16 = transpose(perm = var_3581, x = x105_1_cast_fp16)[name = tensor<string, []>("transpose_64")];
            tensor<fp16, [1, 128, 2]> x107_1_cast_fp16 = select(a = var_62_to_fp16, b = x106_1_cast_fp16, cond = var_3430)[name = tensor<string, []>("x107_1_cast_fp16")];
            tensor<bool, []> x108_1_interleave_0 = const()[name = tensor<string, []>("x108_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s3_l0_conv2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s3_l0_conv2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 7]> cache_s3_l0_conv2_to_fp16 = cast(dtype = cache_s3_l0_conv2_to_fp16_dtype_0, x = cache_s3_l0_conv2)[name = tensor<string, []>("cast_12")];
            tensor<fp16, [1, 128, 9]> x108_1_cast_fp16 = concat(axis = var_49, interleave = x108_1_interleave_0, values = (cache_s3_l0_conv2_to_fp16, x107_1_cast_fp16))[name = tensor<string, []>("x108_1_cast_fp16")];
            tensor<int32, [3]> var_3591_begin_0 = const()[name = tensor<string, []>("op_3591_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
            tensor<int32, [3]> var_3591_end_0 = const()[name = tensor<string, []>("op_3591_end_0"), val = tensor<int32, [3]>([1, 128, 9])];
            tensor<bool, [3]> var_3591_end_mask_0 = const()[name = tensor<string, []>("op_3591_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 7]> new_cache_s3_l0_conv2 = slice_by_index(begin = var_3591_begin_0, end = var_3591_end_0, end_mask = var_3591_end_mask_0, x = x108_1_cast_fp16)[name = tensor<string, []>("op_3591_cast_fp16")];
            tensor<string, []> x_causal_16_pad_type_0 = const()[name = tensor<string, []>("x_causal_16_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_16_groups_0 = const()[name = tensor<string, []>("x_causal_16_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_16_strides_0 = const()[name = tensor<string, []>("x_causal_16_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_16_pad_0 = const()[name = tensor<string, []>("x_causal_16_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_16_dilations_0 = const()[name = tensor<string, []>("x_causal_16_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 8]> encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2011776)))];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2013888)))];
            tensor<fp16, [1, 128, 2]> x_causal_16_cast_fp16 = conv(bias = encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_16_dilations_0, groups = x_causal_16_groups_0, pad = x_causal_16_pad_0, pad_type = x_causal_16_pad_type_0, strides = x_causal_16_strides_0, weight = encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16, x = x108_1_cast_fp16)[name = tensor<string, []>("x_causal_16_cast_fp16")];
            tensor<int32, [3]> input52_1_begin_0 = const()[name = tensor<string, []>("input52_1_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> input52_1_end_0 = const()[name = tensor<string, []>("input52_1_end_0"), val = tensor<int32, [3]>([1, 128, 9])];
            tensor<bool, [3]> input52_1_end_mask_0 = const()[name = tensor<string, []>("input52_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 2]> input52_1_cast_fp16 = slice_by_index(begin = input52_1_begin_0, end = input52_1_end_0, end_mask = input52_1_end_mask_0, x = x108_1_cast_fp16)[name = tensor<string, []>("input52_1_cast_fp16")];
            tensor<string, []> x_chunk_19_pad_type_0 = const()[name = tensor<string, []>("x_chunk_19_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_19_pad_0 = const()[name = tensor<string, []>("x_chunk_19_pad_0"), val = tensor<int32, [2]>([7, 7])];
            tensor<int32, []> x_chunk_19_groups_0 = const()[name = tensor<string, []>("x_chunk_19_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_19_strides_0 = const()[name = tensor<string, []>("x_chunk_19_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_19_dilations_0 = const()[name = tensor<string, []>("x_chunk_19_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 15]> encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2014208)))];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2018112)))];
            tensor<fp16, [1, 128, 2]> x_chunk_19_cast_fp16 = conv(bias = encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_19_dilations_0, groups = x_chunk_19_groups_0, pad = x_chunk_19_pad_0, pad_type = x_chunk_19_pad_type_0, strides = x_chunk_19_strides_0, weight = encoder_encoders_3_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16, x = input52_1_cast_fp16)[name = tensor<string, []>("x_chunk_19_cast_fp16")];
            tensor<fp16, [128, 2]> chunk_scale6_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale6_1_to_fp16"), val = tensor<fp16, [128, 2]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2018432)))];
            tensor<fp16, [1, 128, 2]> x_chunk6_1_cast_fp16 = mul(x = x_chunk_19_cast_fp16, y = chunk_scale6_1_to_fp16)[name = tensor<string, []>("x_chunk6_1_cast_fp16")];
            tensor<fp16, [1, 128, 2]> x109_1_cast_fp16 = add(x = x_chunk6_1_cast_fp16, y = x_causal_16_cast_fp16)[name = tensor<string, []>("x109_1_cast_fp16")];
            tensor<int32, [3]> var_3631 = const()[name = tensor<string, []>("op_3631"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_3640_to_fp16 = const()[name = tensor<string, []>("op_3640_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [2, 1, 128]> x110_1_cast_fp16 = transpose(perm = var_3631, x = x109_1_cast_fp16)[name = tensor<string, []>("transpose_63")];
            tensor<fp16, [2, 1, 128]> x_offset_38_cast_fp16 = sub(x = x110_1_cast_fp16, y = var_3640_to_fp16)[name = tensor<string, []>("x_offset_38_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3642_cast_fp16 = exp(x = x_offset_38_cast_fp16)[name = tensor<string, []>("op_3642_cast_fp16")];
            tensor<fp16, []> var_3643_to_fp16 = const()[name = tensor<string, []>("op_3643_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [2, 1, 128]> var_3644_cast_fp16 = add(x = var_3642_cast_fp16, y = var_3643_to_fp16)[name = tensor<string, []>("op_3644_cast_fp16")];
            tensor<fp32, []> var_3645_epsilon_0 = const()[name = tensor<string, []>("op_3645_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [2, 1, 128]> var_3645_cast_fp16 = log(epsilon = var_3645_epsilon_0, x = var_3644_cast_fp16)[name = tensor<string, []>("op_3645_cast_fp16")];
            tensor<fp16, []> var_3637_to_fp16 = const()[name = tensor<string, []>("op_3637_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [2, 1, 128]> var_3647_cast_fp16 = equal(x = var_3645_cast_fp16, y = var_3637_to_fp16)[name = tensor<string, []>("op_3647_cast_fp16")];
            tensor<fp16, [2, 1, 128]> log_sum0_38_cast_fp16 = select(a = x_offset_38_cast_fp16, b = var_3645_cast_fp16, cond = var_3647_cast_fp16)[name = tensor<string, []>("log_sum0_38_cast_fp16")];
            tensor<fp16, []> var_3649_to_fp16 = const()[name = tensor<string, []>("op_3649_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [2, 1, 128]> var_3650_cast_fp16 = mul(x = x110_1_cast_fp16, y = var_3649_to_fp16)[name = tensor<string, []>("op_3650_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3651_cast_fp16 = sub(x = log_sum0_38_cast_fp16, y = var_3650_cast_fp16)[name = tensor<string, []>("op_3651_cast_fp16")];
            tensor<fp16, []> var_3652_to_fp16 = const()[name = tensor<string, []>("op_3652_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [2, 1, 128]> x0_23_cast_fp16 = sub(x = var_3651_cast_fp16, y = var_3652_to_fp16)[name = tensor<string, []>("x0_23_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_3_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2019008))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2035456))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2036032)))];
            tensor<fp16, [2, 1, 128]> linear_70_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_conv_module2_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized, x = x0_23_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
            tensor<fp16, [2, 1, 128]> input53_1_cast_fp16 = add(x = input50_1_cast_fp16, y = linear_70_cast_fp16)[name = tensor<string, []>("input53_1_cast_fp16")];
            tensor<fp16, [240, 128]> encoder_encoders_3_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2036352))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2067136))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([240, 128])];
            tensor<fp16, [240]> encoder_encoders_3_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2067712)))];
            tensor<fp16, [2, 1, 240]> linear_71_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized, x = input53_1_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
            tensor<fp16, []> var_3668_to_fp16 = const()[name = tensor<string, []>("op_3668_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [2, 1, 240]> x_offset_40_cast_fp16 = sub(x = linear_71_cast_fp16, y = var_3668_to_fp16)[name = tensor<string, []>("x_offset_40_cast_fp16")];
            tensor<fp16, [2, 1, 240]> var_3670_cast_fp16 = exp(x = x_offset_40_cast_fp16)[name = tensor<string, []>("op_3670_cast_fp16")];
            tensor<fp16, []> var_3671_to_fp16 = const()[name = tensor<string, []>("op_3671_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [2, 1, 240]> var_3672_cast_fp16 = add(x = var_3670_cast_fp16, y = var_3671_to_fp16)[name = tensor<string, []>("op_3672_cast_fp16")];
            tensor<fp32, []> var_3673_epsilon_0 = const()[name = tensor<string, []>("op_3673_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [2, 1, 240]> var_3673_cast_fp16 = log(epsilon = var_3673_epsilon_0, x = var_3672_cast_fp16)[name = tensor<string, []>("op_3673_cast_fp16")];
            tensor<fp16, []> var_3656_to_fp16 = const()[name = tensor<string, []>("op_3656_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [2, 1, 240]> var_3675_cast_fp16 = equal(x = var_3673_cast_fp16, y = var_3656_to_fp16)[name = tensor<string, []>("op_3675_cast_fp16")];
            tensor<fp16, [2, 1, 240]> log_sum0_40_cast_fp16 = select(a = x_offset_40_cast_fp16, b = var_3673_cast_fp16, cond = var_3675_cast_fp16)[name = tensor<string, []>("log_sum0_40_cast_fp16")];
            tensor<fp16, []> var_3677_to_fp16 = const()[name = tensor<string, []>("op_3677_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [2, 1, 240]> var_3678_cast_fp16 = mul(x = linear_71_cast_fp16, y = var_3677_to_fp16)[name = tensor<string, []>("op_3678_cast_fp16")];
            tensor<fp16, [2, 1, 240]> var_3679_cast_fp16 = sub(x = log_sum0_40_cast_fp16, y = var_3678_cast_fp16)[name = tensor<string, []>("op_3679_cast_fp16")];
            tensor<fp16, []> var_3680_to_fp16 = const()[name = tensor<string, []>("op_3680_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [2, 1, 240]> x_63_cast_fp16 = sub(x = var_3679_cast_fp16, y = var_3680_to_fp16)[name = tensor<string, []>("x_63_cast_fp16")];
            tensor<fp16, [128, 240]> encoder_encoders_3_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2068288))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2099072))), name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 240])];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2099648)))];
            tensor<fp16, [2, 1, 128]> linear_72_cast_fp16 = linear(bias = encoder_encoders_3_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16, weight = encoder_encoders_3_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized, x = x_63_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
            tensor<fp16, [2, 1, 128]> x111_1_cast_fp16 = add(x = input53_1_cast_fp16, y = linear_72_cast_fp16)[name = tensor<string, []>("x111_1_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2099968)))];
            tensor<fp16, [2, 1, 128]> var_3691_cast_fp16 = sub(x = x111_1_cast_fp16, y = encoder_encoders_3_encoder_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("op_3691_cast_fp16")];
            tensor<fp16, []> var_3685_promoted_to_fp16 = const()[name = tensor<string, []>("op_3685_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [2, 1, 128]> var_3692_cast_fp16 = pow(x = var_3691_cast_fp16, y = var_3685_promoted_to_fp16)[name = tensor<string, []>("op_3692_cast_fp16")];
            tensor<int32, [1]> var_3694_axes_0 = const()[name = tensor<string, []>("op_3694_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, []> var_3694_keep_dims_0 = const()[name = tensor<string, []>("op_3694_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [2, 1, 1]> var_3694_cast_fp16 = reduce_mean(axes = var_3694_axes_0, keep_dims = var_3694_keep_dims_0, x = var_3692_cast_fp16)[name = tensor<string, []>("op_3694_cast_fp16")];
            tensor<fp16, []> var_3684_to_fp16 = const()[name = tensor<string, []>("op_3684_to_fp16"), val = tensor<fp16, []>(-0x1p-1)];
            tensor<fp16, [2, 1, 1]> var_3695_cast_fp16 = pow(x = var_3694_cast_fp16, y = var_3684_to_fp16)[name = tensor<string, []>("op_3695_cast_fp16")];
            tensor<fp16, []> var_3696_to_fp16 = const()[name = tensor<string, []>("op_3696_to_fp16"), val = tensor<fp16, []>(0x1.0f8p-1)];
            tensor<fp16, [2, 1, 1]> scales_10_cast_fp16 = mul(x = var_3695_cast_fp16, y = var_3696_to_fp16)[name = tensor<string, []>("scales_10_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3698_cast_fp16 = mul(x = x111_1_cast_fp16, y = scales_10_cast_fp16)[name = tensor<string, []>("op_3698_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3701_cast_fp16 = sub(x = var_3698_cast_fp16, y = x_51_cast_fp16)[name = tensor<string, []>("op_3701_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_3_encoder_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_encoder_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2100288)))];
            tensor<fp16, [2, 1, 128]> var_3702_cast_fp16 = mul(x = var_3701_cast_fp16, y = encoder_encoders_3_encoder_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_3702_cast_fp16")];
            tensor<fp16, [2, 1, 128]> var_3703_cast_fp16 = add(x = x_51_cast_fp16, y = var_3702_cast_fp16)[name = tensor<string, []>("op_3703_cast_fp16")];
            tensor<int32, [1]> var_3713_axes_0 = const()[name = tensor<string, []>("op_3713_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [2, 1, 1, 128]> var_3713_cast_fp16 = expand_dims(axes = var_3713_axes_0, x = var_3703_cast_fp16)[name = tensor<string, []>("op_3713_cast_fp16")];
            tensor<int32, [4]> src_9_reps_0 = const()[name = tensor<string, []>("src_9_reps_0"), val = tensor<int32, [4]>([1, 8, 1, 1])];
            tensor<fp16, [2, 8, 1, 128]> src_9_cast_fp16 = tile(reps = src_9_reps_0, x = var_3713_cast_fp16)[name = tensor<string, []>("src_9_cast_fp16")];
            tensor<int32, [3]> var_3719 = const()[name = tensor<string, []>("op_3719"), val = tensor<int32, [3]>([16, 1, 128])];
            tensor<fp16, [16, 1, 128]> var_3720_cast_fp16 = reshape(shape = var_3719, x = src_9_cast_fp16)[name = tensor<string, []>("op_3720_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_3727_cast_fp16 = sub(x = var_3720_cast_fp16, y = var_3158_cast_fp16)[name = tensor<string, []>("op_3727_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_3_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_3_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2100608)))];
            tensor<fp16, [16, 1, 128]> var_3728_cast_fp16 = mul(x = var_3727_cast_fp16, y = encoder_encoders_3_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_3728_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_3729_cast_fp16 = add(x = var_3158_cast_fp16, y = var_3728_cast_fp16)[name = tensor<string, []>("op_3729_cast_fp16")];
            tensor<int32, [4]> var_3751 = const()[name = tensor<string, []>("op_3751"), val = tensor<int32, [4]>([4, 4, 1, 128])];
            tensor<fp16, [4, 4, 1, 128]> src0_15_cast_fp16 = reshape(shape = var_3751, x = var_3729_cast_fp16)[name = tensor<string, []>("src0_15_cast_fp16")];
            tensor<fp16, [4, 1, 1]> weights0_8_to_fp16 = const()[name = tensor<string, []>("weights0_8_to_fp16"), val = tensor<fp16, [4, 1, 1]>([[[0x1.fc4p-1]], [[0x1.518p-9]], [[0x1.2ap-9]], [[0x1.2fcp-9]]])];
            tensor<fp16, [4, 4, 1, 128]> var_3756_cast_fp16 = mul(x = src0_15_cast_fp16, y = weights0_8_to_fp16)[name = tensor<string, []>("op_3756_cast_fp16")];
            tensor<int32, [1]> x_65_axes_0 = const()[name = tensor<string, []>("x_65_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> x_65_keep_dims_0 = const()[name = tensor<string, []>("x_65_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 128]> x_65_cast_fp16 = reduce_sum(axes = x_65_axes_0, keep_dims = x_65_keep_dims_0, x = var_3756_cast_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
            tensor<fp16, [272, 128]> encoder_encoders_4_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [34816]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2100928))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2135808))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([272, 128])];
            tensor<fp16, [272]> encoder_encoders_4_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [272]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2136384)))];
            tensor<fp16, [4, 1, 272]> linear_73_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized, x = x_65_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
            tensor<int32, [3]> q11_1_begin_0 = const()[name = tensor<string, []>("q11_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> q11_1_end_0 = const()[name = tensor<string, []>("q11_1_end_0"), val = tensor<int32, [3]>([4, 1, 128])];
            tensor<bool, [3]> q11_1_end_mask_0 = const()[name = tensor<string, []>("q11_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [4, 1, 128]> q11_1_cast_fp16 = slice_by_index(begin = q11_1_begin_0, end = q11_1_end_0, end_mask = q11_1_end_mask_0, x = linear_73_cast_fp16)[name = tensor<string, []>("q11_1_cast_fp16")];
            tensor<int32, [3]> k15_1_begin_0 = const()[name = tensor<string, []>("k15_1_begin_0"), val = tensor<int32, [3]>([0, 0, 128])];
            tensor<int32, [3]> k15_1_end_0 = const()[name = tensor<string, []>("k15_1_end_0"), val = tensor<int32, [3]>([4, 1, 256])];
            tensor<bool, [3]> k15_1_end_mask_0 = const()[name = tensor<string, []>("k15_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [4, 1, 128]> k15_1_cast_fp16 = slice_by_index(begin = k15_1_begin_0, end = k15_1_end_0, end_mask = k15_1_end_mask_0, x = linear_73_cast_fp16)[name = tensor<string, []>("k15_1_cast_fp16")];
            tensor<int32, [3]> p11_1_begin_0 = const()[name = tensor<string, []>("p11_1_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> p11_1_end_0 = const()[name = tensor<string, []>("p11_1_end_0"), val = tensor<int32, [3]>([4, 1, 272])];
            tensor<bool, [3]> p11_1_end_mask_0 = const()[name = tensor<string, []>("p11_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [4, 1, 16]> p11_1_cast_fp16 = slice_by_index(begin = p11_1_begin_0, end = p11_1_end_0, end_mask = p11_1_end_mask_0, x = linear_73_cast_fp16)[name = tensor<string, []>("p11_1_cast_fp16")];
            tensor<bool, []> k16_1_interleave_0 = const()[name = tensor<string, []>("k16_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s4_l0_key_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s4_l0_key_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [16, 1, 128]> cache_s4_l0_key_to_fp16 = cast(dtype = cache_s4_l0_key_to_fp16_dtype_0, x = cache_s4_l0_key)[name = tensor<string, []>("cast_11")];
            tensor<fp16, [20, 1, 128]> k16_1_cast_fp16 = concat(axis = var_46, interleave = k16_1_interleave_0, values = (cache_s4_l0_key_to_fp16, k15_1_cast_fp16))[name = tensor<string, []>("k16_1_cast_fp16")];
            tensor<int32, [3]> var_3809_begin_0 = const()[name = tensor<string, []>("op_3809_begin_0"), val = tensor<int32, [3]>([4, 0, 0])];
            tensor<int32, [3]> var_3809_end_0 = const()[name = tensor<string, []>("op_3809_end_0"), val = tensor<int32, [3]>([20, 1, 128])];
            tensor<bool, [3]> var_3809_end_mask_0 = const()[name = tensor<string, []>("op_3809_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [16, 1, 128]> new_cache_s4_l0_key = slice_by_index(begin = var_3809_begin_0, end = var_3809_end_0, end_mask = var_3809_end_mask_0, x = k16_1_cast_fp16)[name = tensor<string, []>("op_3809_cast_fp16")];
            tensor<int32, [4]> var_3815 = const()[name = tensor<string, []>("op_3815"), val = tensor<int32, [4]>([4, 1, 4, 32])];
            tensor<fp16, [4, 1, 4, 32]> q12_1_cast_fp16 = reshape(shape = var_3815, x = q11_1_cast_fp16)[name = tensor<string, []>("q12_1_cast_fp16")];
            tensor<int32, [4]> var_3817 = const()[name = tensor<string, []>("op_3817"), val = tensor<int32, [4]>([4, 1, 4, 4])];
            tensor<fp16, [4, 1, 4, 4]> p12_1_cast_fp16 = reshape(shape = var_3817, x = p11_1_cast_fp16)[name = tensor<string, []>("p12_1_cast_fp16")];
            tensor<int32, [4]> var_3819 = const()[name = tensor<string, []>("op_3819"), val = tensor<int32, [4]>([20, 1, 4, 32])];
            tensor<fp16, [20, 1, 4, 32]> k17_1_cast_fp16 = reshape(shape = var_3819, x = k16_1_cast_fp16)[name = tensor<string, []>("k17_1_cast_fp16")];
            tensor<int32, [4]> var_3823 = const()[name = tensor<string, []>("op_3823"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> attn_scores11_1_transpose_x_0 = const()[name = tensor<string, []>("attn_scores11_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_scores11_1_transpose_y_0 = const()[name = tensor<string, []>("attn_scores11_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_32_perm_0 = const()[name = tensor<string, []>("transpose_32_perm_0"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<int32, [4]> transpose_33_perm_0 = const()[name = tensor<string, []>("transpose_33_perm_0"), val = tensor<int32, [4]>([2, 1, -1, -4])];
            tensor<fp16, [4, 1, 32, 20]> transpose_33 = transpose(perm = transpose_33_perm_0, x = k17_1_cast_fp16)[name = tensor<string, []>("transpose_61")];
            tensor<fp16, [4, 1, 4, 32]> transpose_32 = transpose(perm = transpose_32_perm_0, x = q12_1_cast_fp16)[name = tensor<string, []>("transpose_62")];
            tensor<fp16, [4, 1, 4, 20]> attn_scores11_1_cast_fp16 = matmul(transpose_x = attn_scores11_1_transpose_x_0, transpose_y = attn_scores11_1_transpose_y_0, x = transpose_32, y = transpose_33)[name = tensor<string, []>("attn_scores11_1_cast_fp16")];
            tensor<bool, []> pos_scores15_1_transpose_x_0 = const()[name = tensor<string, []>("pos_scores15_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_scores15_1_transpose_y_0 = const()[name = tensor<string, []>("pos_scores15_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 4, 23]> pos_emb3_1_to_fp16 = const()[name = tensor<string, []>("pos_emb3_1_to_fp16"), val = tensor<fp16, [4, 1, 4, 23]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2137024)))];
            tensor<fp16, [4, 1, 4, 4]> p13_1_cast_fp16 = transpose(perm = var_3823, x = p12_1_cast_fp16)[name = tensor<string, []>("transpose_60")];
            tensor<fp16, [4, 1, 4, 23]> pos_scores15_1_cast_fp16 = matmul(transpose_x = pos_scores15_1_transpose_x_0, transpose_y = pos_scores15_1_transpose_y_0, x = p13_1_cast_fp16, y = pos_emb3_1_to_fp16)[name = tensor<string, []>("pos_scores15_1_cast_fp16")];
            tensor<int32, [2]> var_3866 = const()[name = tensor<string, []>("op_3866"), val = tensor<int32, [2]>([-1, 23])];
            tensor<fp16, [16, 23]> pos_scores16_1_cast_fp16 = reshape(shape = var_3866, x = pos_scores15_1_cast_fp16)[name = tensor<string, []>("pos_scores16_1_cast_fp16")];
            tensor<bool, []> pos_scores17_1_validate_indices_0 = const()[name = tensor<string, []>("pos_scores17_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [16, 20]> pos_scores17_1_cast_fp16_cast_uint16 = gather_along_axis(axis = var_48, indices = indexes1_1_to_uint16, validate_indices = pos_scores17_1_validate_indices_0, x = pos_scores16_1_cast_fp16)[name = tensor<string, []>("pos_scores17_1_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_3869 = const()[name = tensor<string, []>("op_3869"), val = tensor<int32, [4]>([4, 1, 4, 20])];
            tensor<fp16, [4, 1, 4, 20]> pos_scores18_1_cast_fp16 = reshape(shape = var_3869, x = pos_scores17_1_cast_fp16_cast_uint16)[name = tensor<string, []>("pos_scores18_1_cast_fp16")];
            tensor<fp16, [4, 1, 4, 20]> attn_scores12_1_cast_fp16 = add(x = attn_scores11_1_cast_fp16, y = pos_scores18_1_cast_fp16)[name = tensor<string, []>("attn_scores12_1_cast_fp16")];
            tensor<fp16, []> cast_90_to_fp16 = const()[name = tensor<string, []>("cast_90_to_fp16"), val = tensor<fp16, []>(-0x1.f4p+9)];
            tensor<fp16, [4, 1, 4, 20]> attn_scores13_1_cast_fp16 = select(a = cast_90_to_fp16, b = attn_scores12_1_cast_fp16, cond = var_2730)[name = tensor<string, []>("attn_scores13_1_cast_fp16")];
            tensor<fp16, [4, 1, 4, 20]> attn_weights7_1_cast_fp16 = softmax(axis = var_59, x = attn_scores13_1_cast_fp16)[name = tensor<string, []>("attn_weights7_1_cast_fp16")];
            tensor<fp16, [144, 128]> encoder_encoders_4_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2137856))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2156352))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([144, 128])];
            tensor<fp16, [144]> encoder_encoders_4_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2156928)))];
            tensor<fp16, [4, 1, 144]> linear_75_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized, x = x_65_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
            tensor<fp16, []> var_3887_to_fp16 = const()[name = tensor<string, []>("op_3887_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [4, 1, 144]> x_offset_42_cast_fp16 = sub(x = linear_75_cast_fp16, y = var_3887_to_fp16)[name = tensor<string, []>("x_offset_42_cast_fp16")];
            tensor<fp16, [4, 1, 144]> var_3889_cast_fp16 = exp(x = x_offset_42_cast_fp16)[name = tensor<string, []>("op_3889_cast_fp16")];
            tensor<fp16, []> var_3890_to_fp16 = const()[name = tensor<string, []>("op_3890_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 144]> var_3891_cast_fp16 = add(x = var_3889_cast_fp16, y = var_3890_to_fp16)[name = tensor<string, []>("op_3891_cast_fp16")];
            tensor<fp32, []> var_3892_epsilon_0 = const()[name = tensor<string, []>("op_3892_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 144]> var_3892_cast_fp16 = log(epsilon = var_3892_epsilon_0, x = var_3891_cast_fp16)[name = tensor<string, []>("op_3892_cast_fp16")];
            tensor<fp16, []> var_3875_to_fp16 = const()[name = tensor<string, []>("op_3875_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 144]> var_3894_cast_fp16 = equal(x = var_3892_cast_fp16, y = var_3875_to_fp16)[name = tensor<string, []>("op_3894_cast_fp16")];
            tensor<fp16, [4, 1, 144]> log_sum0_42_cast_fp16 = select(a = x_offset_42_cast_fp16, b = var_3892_cast_fp16, cond = var_3894_cast_fp16)[name = tensor<string, []>("log_sum0_42_cast_fp16")];
            tensor<fp16, []> var_3896_to_fp16 = const()[name = tensor<string, []>("op_3896_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 144]> var_3897_cast_fp16 = mul(x = linear_75_cast_fp16, y = var_3896_to_fp16)[name = tensor<string, []>("op_3897_cast_fp16")];
            tensor<fp16, [4, 1, 144]> var_3898_cast_fp16 = sub(x = log_sum0_42_cast_fp16, y = var_3897_cast_fp16)[name = tensor<string, []>("op_3898_cast_fp16")];
            tensor<fp16, []> var_3899_to_fp16 = const()[name = tensor<string, []>("op_3899_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [4, 1, 144]> x_71_cast_fp16 = sub(x = var_3898_cast_fp16, y = var_3899_to_fp16)[name = tensor<string, []>("x_71_cast_fp16")];
            tensor<fp16, [128, 144]> encoder_encoders_4_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2157312))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2175808))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 144])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2176384)))];
            tensor<fp16, [4, 1, 128]> linear_76_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized, x = x_71_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input54_1_cast_fp16 = add(x = x_65_cast_fp16, y = linear_76_cast_fp16)[name = tensor<string, []>("input54_1_cast_fp16")];
            tensor<int32, [4]> attn_weights8_1_begin_0 = const()[name = tensor<string, []>("attn_weights8_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_weights8_1_end_0 = const()[name = tensor<string, []>("attn_weights8_1_end_0"), val = tensor<int32, [4]>([1, 1, 4, 20])];
            tensor<bool, [4]> attn_weights8_1_end_mask_0 = const()[name = tensor<string, []>("attn_weights8_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1, 4, 20]> attn_weights8_1_cast_fp16 = slice_by_index(begin = attn_weights8_1_begin_0, end = attn_weights8_1_end_0, end_mask = attn_weights8_1_end_mask_0, x = attn_weights7_1_cast_fp16)[name = tensor<string, []>("attn_weights8_1_cast_fp16")];
            tensor<fp16, [288, 128]> encoder_encoders_4_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [36864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2176704))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2213632))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([288, 128])];
            tensor<fp16, [288]> encoder_encoders_4_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2214208)))];
            tensor<fp16, [4, 1, 288]> linear_77_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized, x = input54_1_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
            tensor<int32, [3]> var_3913_split_sizes_0 = const()[name = tensor<string, []>("op_3913_split_sizes_0"), val = tensor<int32, [3]>([96, 96, 96])];
            tensor<int32, []> var_3913_axis_0 = const()[name = tensor<string, []>("op_3913_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [4, 1, 96]> var_3913_cast_fp16_0, tensor<fp16, [4, 1, 96]> var_3913_cast_fp16_1, tensor<fp16, [4, 1, 96]> var_3913_cast_fp16_2 = split(axis = var_3913_axis_0, split_sizes = var_3913_split_sizes_0, x = linear_77_cast_fp16)[name = tensor<string, []>("op_3913_cast_fp16")];
            tensor<fp16, [4, 1, 96]> var_3916_cast_fp16 = tanh(x = var_3913_cast_fp16_0)[name = tensor<string, []>("op_3916_cast_fp16")];
            tensor<int32, [1]> var_3917_axes_0 = const()[name = tensor<string, []>("op_3917_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [4, 1, 96, 1]> var_3917_cast_fp16 = expand_dims(axes = var_3917_axes_0, x = var_3916_cast_fp16)[name = tensor<string, []>("op_3917_cast_fp16")];
            tensor<int32, [3]> var_3918 = const()[name = tensor<string, []>("op_3918"), val = tensor<int32, [3]>([4, 1, 96])];
            tensor<fp16, [4, 1, 96]> s3_1_cast_fp16 = reshape(shape = var_3918, x = var_3917_cast_fp16)[name = tensor<string, []>("s3_1_cast_fp16")];
            tensor<fp16, [4, 1, 96]> x113_1_cast_fp16 = mul(x = var_3913_cast_fp16_1, y = s3_1_cast_fp16)[name = tensor<string, []>("x113_1_cast_fp16")];
            tensor<int32, [4]> var_3932 = const()[name = tensor<string, []>("op_3932"), val = tensor<int32, [4]>([4, 1, 1, -1])];
            tensor<fp16, [4, 1, 1, 96]> var_3933_cast_fp16 = reshape(shape = var_3932, x = x113_1_cast_fp16)[name = tensor<string, []>("op_3933_cast_fp16")];
            tensor<int32, [4]> var_3934 = const()[name = tensor<string, []>("op_3934"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<bool, []> x_pad3_1_interleave_0 = const()[name = tensor<string, []>("x_pad3_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s4_l0_nonlin_attn_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s4_l0_nonlin_attn_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 96]> x114_1_cast_fp16 = transpose(perm = var_3934, x = var_3933_cast_fp16)[name = tensor<string, []>("transpose_59")];
            tensor<fp16, [1, 1, 16, 96]> cache_s4_l0_nonlin_attn_to_fp16 = cast(dtype = cache_s4_l0_nonlin_attn_to_fp16_dtype_0, x = cache_s4_l0_nonlin_attn)[name = tensor<string, []>("cast_10")];
            tensor<fp16, [1, 1, 20, 96]> x_pad3_1_cast_fp16 = concat(axis = var_49, interleave = x_pad3_1_interleave_0, values = (cache_s4_l0_nonlin_attn_to_fp16, x114_1_cast_fp16))[name = tensor<string, []>("x_pad3_1_cast_fp16")];
            tensor<int32, [4]> var_3940_begin_0 = const()[name = tensor<string, []>("op_3940_begin_0"), val = tensor<int32, [4]>([0, 0, 4, 0])];
            tensor<int32, [4]> var_3940_end_0 = const()[name = tensor<string, []>("op_3940_end_0"), val = tensor<int32, [4]>([1, 1, 20, 96])];
            tensor<bool, [4]> var_3940_end_mask_0 = const()[name = tensor<string, []>("op_3940_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 16, 96]> new_cache_s4_l0_nonlin_attn = slice_by_index(begin = var_3940_begin_0, end = var_3940_end_0, end_mask = var_3940_end_mask_0, x = x_pad3_1_cast_fp16)[name = tensor<string, []>("op_3940_cast_fp16")];
            tensor<bool, []> x115_1_transpose_x_0 = const()[name = tensor<string, []>("x115_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x115_1_transpose_y_0 = const()[name = tensor<string, []>("x115_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 4, 96]> x115_1_cast_fp16 = matmul(transpose_x = x115_1_transpose_x_0, transpose_y = x115_1_transpose_y_0, x = attn_weights8_1_cast_fp16, y = x_pad3_1_cast_fp16)[name = tensor<string, []>("x115_1_cast_fp16")];
            tensor<int32, [4]> var_3943 = const()[name = tensor<string, []>("op_3943"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_3945 = const()[name = tensor<string, []>("op_3945"), val = tensor<int32, [3]>([4, 1, -1])];
            tensor<fp16, [4, 1, 1, 96]> var_3944_cast_fp16 = transpose(perm = var_3943, x = x115_1_cast_fp16)[name = tensor<string, []>("transpose_58")];
            tensor<fp16, [4, 1, 96]> x116_1_cast_fp16 = reshape(shape = var_3945, x = var_3944_cast_fp16)[name = tensor<string, []>("x116_1_cast_fp16")];
            tensor<fp16, [4, 1, 96]> input56_1_cast_fp16 = mul(x = x116_1_cast_fp16, y = var_3913_cast_fp16_2)[name = tensor<string, []>("input56_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_4_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2214848))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2227200))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2227776)))];
            tensor<fp16, [4, 1, 128]> linear_78_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized, x = input56_1_cast_fp16)[name = tensor<string, []>("linear_78_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x117_1_cast_fp16 = add(x = input54_1_cast_fp16, y = linear_78_cast_fp16)[name = tensor<string, []>("x117_1_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_4_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2228096))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2234304))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_4_encoder_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2234880)))];
            tensor<fp16, [4, 1, 48]> linear_79_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_self_attn1_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized, x = x117_1_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
            tensor<bool, []> x118_1_interleave_0 = const()[name = tensor<string, []>("x118_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s4_l0_val1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s4_l0_val1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [16, 1, 48]> cache_s4_l0_val1_to_fp16 = cast(dtype = cache_s4_l0_val1_to_fp16_dtype_0, x = cache_s4_l0_val1)[name = tensor<string, []>("cast_9")];
            tensor<fp16, [20, 1, 48]> x118_1_cast_fp16 = concat(axis = var_46, interleave = x118_1_interleave_0, values = (cache_s4_l0_val1_to_fp16, linear_79_cast_fp16))[name = tensor<string, []>("x118_1_cast_fp16")];
            tensor<int32, [3]> var_3970_begin_0 = const()[name = tensor<string, []>("op_3970_begin_0"), val = tensor<int32, [3]>([4, 0, 0])];
            tensor<int32, [3]> var_3970_end_0 = const()[name = tensor<string, []>("op_3970_end_0"), val = tensor<int32, [3]>([20, 1, 48])];
            tensor<bool, [3]> var_3970_end_mask_0 = const()[name = tensor<string, []>("op_3970_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [16, 1, 48]> new_cache_s4_l0_val1 = slice_by_index(begin = var_3970_begin_0, end = var_3970_end_0, end_mask = var_3970_end_mask_0, x = x118_1_cast_fp16)[name = tensor<string, []>("op_3970_cast_fp16")];
            tensor<int32, [4]> var_3971 = const()[name = tensor<string, []>("op_3971"), val = tensor<int32, [4]>([20, 1, 4, -1])];
            tensor<fp16, [20, 1, 4, 12]> var_3972_cast_fp16 = reshape(shape = var_3971, x = x118_1_cast_fp16)[name = tensor<string, []>("op_3972_cast_fp16")];
            tensor<int32, [4]> var_3973 = const()[name = tensor<string, []>("op_3973"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x120_1_transpose_x_0 = const()[name = tensor<string, []>("x120_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x120_1_transpose_y_0 = const()[name = tensor<string, []>("x120_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 20, 12]> x119_1_cast_fp16 = transpose(perm = var_3973, x = var_3972_cast_fp16)[name = tensor<string, []>("transpose_57")];
            tensor<fp16, [4, 1, 4, 12]> x120_1_cast_fp16 = matmul(transpose_x = x120_1_transpose_x_0, transpose_y = x120_1_transpose_y_0, x = attn_weights7_1_cast_fp16, y = x119_1_cast_fp16)[name = tensor<string, []>("x120_1_cast_fp16")];
            tensor<int32, [4]> var_3978 = const()[name = tensor<string, []>("op_3978"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_3983 = const()[name = tensor<string, []>("op_3983"), val = tensor<int32, [3]>([4, 1, 48])];
            tensor<fp16, [4, 1, 4, 12]> var_3979_cast_fp16 = transpose(perm = var_3978, x = x120_1_cast_fp16)[name = tensor<string, []>("transpose_56")];
            tensor<fp16, [4, 1, 48]> input57_1_cast_fp16 = reshape(shape = var_3983, x = var_3979_cast_fp16)[name = tensor<string, []>("input57_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_4_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2235072))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2241280))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2241856)))];
            tensor<fp16, [4, 1, 128]> linear_80_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_self_attn1_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized, x = input57_1_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input58_1_cast_fp16 = add(x = x117_1_cast_fp16, y = linear_80_cast_fp16)[name = tensor<string, []>("input58_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_4_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2242176))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2275008))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_4_encoder_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2275584)))];
            tensor<fp16, [4, 1, 256]> linear_81_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_conv_module1_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized, x = input58_1_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
            tensor<int32, [2]> var_3994_split_sizes_0 = const()[name = tensor<string, []>("op_3994_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_3994_axis_0 = const()[name = tensor<string, []>("op_3994_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [4, 1, 128]> var_3994_cast_fp16_0, tensor<fp16, [4, 1, 128]> var_3994_cast_fp16_1 = split(axis = var_3994_axis_0, split_sizes = var_3994_split_sizes_0, x = linear_81_cast_fp16)[name = tensor<string, []>("op_3994_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_3996_cast_fp16 = sigmoid(x = var_3994_cast_fp16_1)[name = tensor<string, []>("op_3996_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x122_1_cast_fp16 = mul(x = var_3994_cast_fp16_0, y = var_3996_cast_fp16)[name = tensor<string, []>("x122_1_cast_fp16")];
            tensor<int32, [3]> var_3998 = const()[name = tensor<string, []>("op_3998"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 4]> x123_1_cast_fp16 = transpose(perm = var_3998, x = x122_1_cast_fp16)[name = tensor<string, []>("transpose_55")];
            tensor<fp16, [1, 128, 4]> x124_1_cast_fp16 = select(a = var_62_to_fp16, b = x123_1_cast_fp16, cond = var_2859)[name = tensor<string, []>("x124_1_cast_fp16")];
            tensor<bool, []> x125_1_interleave_0 = const()[name = tensor<string, []>("x125_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s4_l0_conv1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s4_l0_conv1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 7]> cache_s4_l0_conv1_to_fp16 = cast(dtype = cache_s4_l0_conv1_to_fp16_dtype_0, x = cache_s4_l0_conv1)[name = tensor<string, []>("cast_8")];
            tensor<fp16, [1, 128, 11]> x125_1_cast_fp16 = concat(axis = var_49, interleave = x125_1_interleave_0, values = (cache_s4_l0_conv1_to_fp16, x124_1_cast_fp16))[name = tensor<string, []>("x125_1_cast_fp16")];
            tensor<int32, [3]> var_4008_begin_0 = const()[name = tensor<string, []>("op_4008_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_4008_end_0 = const()[name = tensor<string, []>("op_4008_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> var_4008_end_mask_0 = const()[name = tensor<string, []>("op_4008_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 7]> new_cache_s4_l0_conv1 = slice_by_index(begin = var_4008_begin_0, end = var_4008_end_0, end_mask = var_4008_end_mask_0, x = x125_1_cast_fp16)[name = tensor<string, []>("op_4008_cast_fp16")];
            tensor<string, []> x_causal_18_pad_type_0 = const()[name = tensor<string, []>("x_causal_18_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_18_groups_0 = const()[name = tensor<string, []>("x_causal_18_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_18_strides_0 = const()[name = tensor<string, []>("x_causal_18_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_18_pad_0 = const()[name = tensor<string, []>("x_causal_18_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_18_dilations_0 = const()[name = tensor<string, []>("x_causal_18_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 8]> encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2276160)))];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2278272)))];
            tensor<fp16, [1, 128, 4]> x_causal_18_cast_fp16 = conv(bias = encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_18_dilations_0, groups = x_causal_18_groups_0, pad = x_causal_18_pad_0, pad_type = x_causal_18_pad_type_0, strides = x_causal_18_strides_0, weight = encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16, x = x125_1_cast_fp16)[name = tensor<string, []>("x_causal_18_cast_fp16")];
            tensor<int32, [3]> input60_1_begin_0 = const()[name = tensor<string, []>("input60_1_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> input60_1_end_0 = const()[name = tensor<string, []>("input60_1_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> input60_1_end_mask_0 = const()[name = tensor<string, []>("input60_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 4]> input60_1_cast_fp16 = slice_by_index(begin = input60_1_begin_0, end = input60_1_end_0, end_mask = input60_1_end_mask_0, x = x125_1_cast_fp16)[name = tensor<string, []>("input60_1_cast_fp16")];
            tensor<string, []> x_chunk_21_pad_type_0 = const()[name = tensor<string, []>("x_chunk_21_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_21_pad_0 = const()[name = tensor<string, []>("x_chunk_21_pad_0"), val = tensor<int32, [2]>([7, 7])];
            tensor<int32, []> x_chunk_21_groups_0 = const()[name = tensor<string, []>("x_chunk_21_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_21_strides_0 = const()[name = tensor<string, []>("x_chunk_21_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_21_dilations_0 = const()[name = tensor<string, []>("x_chunk_21_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 15]> encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2278592)))];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2282496)))];
            tensor<fp16, [1, 128, 4]> x_chunk_21_cast_fp16 = conv(bias = encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_21_dilations_0, groups = x_chunk_21_groups_0, pad = x_chunk_21_pad_0, pad_type = x_chunk_21_pad_type_0, strides = x_chunk_21_strides_0, weight = encoder_encoders_4_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16, x = input60_1_cast_fp16)[name = tensor<string, []>("x_chunk_21_cast_fp16")];
            tensor<fp16, [128, 4]> chunk_scale7_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale7_1_to_fp16"), val = tensor<fp16, [128, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2282816)))];
            tensor<fp16, [1, 128, 4]> x_chunk7_1_cast_fp16 = mul(x = x_chunk_21_cast_fp16, y = chunk_scale7_1_to_fp16)[name = tensor<string, []>("x_chunk7_1_cast_fp16")];
            tensor<fp16, [1, 128, 4]> x126_1_cast_fp16 = add(x = x_chunk7_1_cast_fp16, y = x_causal_18_cast_fp16)[name = tensor<string, []>("x126_1_cast_fp16")];
            tensor<int32, [3]> var_4048 = const()[name = tensor<string, []>("op_4048"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_4057_to_fp16 = const()[name = tensor<string, []>("op_4057_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> x127_1_cast_fp16 = transpose(perm = var_4048, x = x126_1_cast_fp16)[name = tensor<string, []>("transpose_54")];
            tensor<fp16, [4, 1, 128]> x_offset_44_cast_fp16 = sub(x = x127_1_cast_fp16, y = var_4057_to_fp16)[name = tensor<string, []>("x_offset_44_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4059_cast_fp16 = exp(x = x_offset_44_cast_fp16)[name = tensor<string, []>("op_4059_cast_fp16")];
            tensor<fp16, []> var_4060_to_fp16 = const()[name = tensor<string, []>("op_4060_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> var_4061_cast_fp16 = add(x = var_4059_cast_fp16, y = var_4060_to_fp16)[name = tensor<string, []>("op_4061_cast_fp16")];
            tensor<fp32, []> var_4062_epsilon_0 = const()[name = tensor<string, []>("op_4062_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 128]> var_4062_cast_fp16 = log(epsilon = var_4062_epsilon_0, x = var_4061_cast_fp16)[name = tensor<string, []>("op_4062_cast_fp16")];
            tensor<fp16, []> var_4054_to_fp16 = const()[name = tensor<string, []>("op_4054_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 128]> var_4064_cast_fp16 = equal(x = var_4062_cast_fp16, y = var_4054_to_fp16)[name = tensor<string, []>("op_4064_cast_fp16")];
            tensor<fp16, [4, 1, 128]> log_sum0_44_cast_fp16 = select(a = x_offset_44_cast_fp16, b = var_4062_cast_fp16, cond = var_4064_cast_fp16)[name = tensor<string, []>("log_sum0_44_cast_fp16")];
            tensor<fp16, []> var_4066_to_fp16 = const()[name = tensor<string, []>("op_4066_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 128]> var_4067_cast_fp16 = mul(x = x127_1_cast_fp16, y = var_4066_to_fp16)[name = tensor<string, []>("op_4067_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4068_cast_fp16 = sub(x = log_sum0_44_cast_fp16, y = var_4067_cast_fp16)[name = tensor<string, []>("op_4068_cast_fp16")];
            tensor<fp16, []> var_4069_to_fp16 = const()[name = tensor<string, []>("op_4069_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [4, 1, 128]> x0_25_cast_fp16 = sub(x = var_4068_cast_fp16, y = var_4069_to_fp16)[name = tensor<string, []>("x0_25_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_4_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2283904))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2300352))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2300928)))];
            tensor<fp16, [4, 1, 128]> linear_82_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_conv_module1_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized, x = x0_25_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input61_1_cast_fp16 = add(x = input58_1_cast_fp16, y = linear_82_cast_fp16)[name = tensor<string, []>("input61_1_cast_fp16")];
            tensor<fp16, [192, 128]> encoder_encoders_4_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2301248))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2325888))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([192, 128])];
            tensor<fp16, [192]> encoder_encoders_4_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2326464)))];
            tensor<fp16, [4, 1, 192]> linear_83_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized, x = input61_1_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
            tensor<fp16, []> var_4085_to_fp16 = const()[name = tensor<string, []>("op_4085_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [4, 1, 192]> x_offset_46_cast_fp16 = sub(x = linear_83_cast_fp16, y = var_4085_to_fp16)[name = tensor<string, []>("x_offset_46_cast_fp16")];
            tensor<fp16, [4, 1, 192]> var_4087_cast_fp16 = exp(x = x_offset_46_cast_fp16)[name = tensor<string, []>("op_4087_cast_fp16")];
            tensor<fp16, []> var_4088_to_fp16 = const()[name = tensor<string, []>("op_4088_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 192]> var_4089_cast_fp16 = add(x = var_4087_cast_fp16, y = var_4088_to_fp16)[name = tensor<string, []>("op_4089_cast_fp16")];
            tensor<fp32, []> var_4090_epsilon_0 = const()[name = tensor<string, []>("op_4090_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 192]> var_4090_cast_fp16 = log(epsilon = var_4090_epsilon_0, x = var_4089_cast_fp16)[name = tensor<string, []>("op_4090_cast_fp16")];
            tensor<fp16, []> var_4073_to_fp16 = const()[name = tensor<string, []>("op_4073_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 192]> var_4092_cast_fp16 = equal(x = var_4090_cast_fp16, y = var_4073_to_fp16)[name = tensor<string, []>("op_4092_cast_fp16")];
            tensor<fp16, [4, 1, 192]> log_sum0_46_cast_fp16 = select(a = x_offset_46_cast_fp16, b = var_4090_cast_fp16, cond = var_4092_cast_fp16)[name = tensor<string, []>("log_sum0_46_cast_fp16")];
            tensor<fp16, []> var_4094_to_fp16 = const()[name = tensor<string, []>("op_4094_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 192]> var_4095_cast_fp16 = mul(x = linear_83_cast_fp16, y = var_4094_to_fp16)[name = tensor<string, []>("op_4095_cast_fp16")];
            tensor<fp16, [4, 1, 192]> var_4096_cast_fp16 = sub(x = log_sum0_46_cast_fp16, y = var_4095_cast_fp16)[name = tensor<string, []>("op_4096_cast_fp16")];
            tensor<fp16, []> var_4097_to_fp16 = const()[name = tensor<string, []>("op_4097_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [4, 1, 192]> x_73_cast_fp16 = sub(x = var_4096_cast_fp16, y = var_4097_to_fp16)[name = tensor<string, []>("x_73_cast_fp16")];
            tensor<fp16, [128, 192]> encoder_encoders_4_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2326912))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2351552))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 192])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2352128)))];
            tensor<fp16, [4, 1, 128]> linear_84_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized, x = x_73_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
            tensor<fp16, [4, 1, 128]> src10_1_cast_fp16 = add(x = input61_1_cast_fp16, y = linear_84_cast_fp16)[name = tensor<string, []>("src10_1_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4103_cast_fp16 = sub(x = src10_1_cast_fp16, y = x_65_cast_fp16)[name = tensor<string, []>("op_4103_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2352448)))];
            tensor<fp16, [4, 1, 128]> var_4104_cast_fp16 = mul(x = var_4103_cast_fp16, y = encoder_encoders_4_encoder_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_4104_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4105_cast_fp16 = add(x = x_65_cast_fp16, y = var_4104_cast_fp16)[name = tensor<string, []>("op_4105_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_4_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2352768))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2358976))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_4_encoder_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2359552)))];
            tensor<fp16, [4, 1, 48]> linear_85_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_self_attn2_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized, x = var_4105_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
            tensor<bool, []> x128_1_interleave_0 = const()[name = tensor<string, []>("x128_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s4_l0_val2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s4_l0_val2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [16, 1, 48]> cache_s4_l0_val2_to_fp16 = cast(dtype = cache_s4_l0_val2_to_fp16_dtype_0, x = cache_s4_l0_val2)[name = tensor<string, []>("cast_7")];
            tensor<fp16, [20, 1, 48]> x128_1_cast_fp16 = concat(axis = var_46, interleave = x128_1_interleave_0, values = (cache_s4_l0_val2_to_fp16, linear_85_cast_fp16))[name = tensor<string, []>("x128_1_cast_fp16")];
            tensor<int32, [3]> var_4124_begin_0 = const()[name = tensor<string, []>("op_4124_begin_0"), val = tensor<int32, [3]>([4, 0, 0])];
            tensor<int32, [3]> var_4124_end_0 = const()[name = tensor<string, []>("op_4124_end_0"), val = tensor<int32, [3]>([20, 1, 48])];
            tensor<bool, [3]> var_4124_end_mask_0 = const()[name = tensor<string, []>("op_4124_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [16, 1, 48]> new_cache_s4_l0_val2 = slice_by_index(begin = var_4124_begin_0, end = var_4124_end_0, end_mask = var_4124_end_mask_0, x = x128_1_cast_fp16)[name = tensor<string, []>("op_4124_cast_fp16")];
            tensor<int32, [4]> var_4125 = const()[name = tensor<string, []>("op_4125"), val = tensor<int32, [4]>([20, 1, 4, -1])];
            tensor<fp16, [20, 1, 4, 12]> var_4126_cast_fp16 = reshape(shape = var_4125, x = x128_1_cast_fp16)[name = tensor<string, []>("op_4126_cast_fp16")];
            tensor<int32, [4]> var_4127 = const()[name = tensor<string, []>("op_4127"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x130_1_transpose_x_0 = const()[name = tensor<string, []>("x130_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x130_1_transpose_y_0 = const()[name = tensor<string, []>("x130_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 20, 12]> x129_1_cast_fp16 = transpose(perm = var_4127, x = var_4126_cast_fp16)[name = tensor<string, []>("transpose_53")];
            tensor<fp16, [4, 1, 4, 12]> x130_1_cast_fp16 = matmul(transpose_x = x130_1_transpose_x_0, transpose_y = x130_1_transpose_y_0, x = attn_weights7_1_cast_fp16, y = x129_1_cast_fp16)[name = tensor<string, []>("x130_1_cast_fp16")];
            tensor<int32, [4]> var_4132 = const()[name = tensor<string, []>("op_4132"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_4137 = const()[name = tensor<string, []>("op_4137"), val = tensor<int32, [3]>([4, 1, 48])];
            tensor<fp16, [4, 1, 4, 12]> var_4133_cast_fp16 = transpose(perm = var_4132, x = x130_1_cast_fp16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [4, 1, 48]> input62_1_cast_fp16 = reshape(shape = var_4137, x = var_4133_cast_fp16)[name = tensor<string, []>("input62_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_4_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2359744))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2365952))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2366528)))];
            tensor<fp16, [4, 1, 128]> linear_86_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_self_attn2_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized, x = input62_1_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input63_1_cast_fp16 = add(x = var_4105_cast_fp16, y = linear_86_cast_fp16)[name = tensor<string, []>("input63_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_4_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2366848))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2399680))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_4_encoder_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2400256)))];
            tensor<fp16, [4, 1, 256]> linear_87_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_conv_module2_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized, x = input63_1_cast_fp16)[name = tensor<string, []>("linear_87_cast_fp16")];
            tensor<int32, [2]> var_4148_split_sizes_0 = const()[name = tensor<string, []>("op_4148_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_4148_axis_0 = const()[name = tensor<string, []>("op_4148_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [4, 1, 128]> var_4148_cast_fp16_0, tensor<fp16, [4, 1, 128]> var_4148_cast_fp16_1 = split(axis = var_4148_axis_0, split_sizes = var_4148_split_sizes_0, x = linear_87_cast_fp16)[name = tensor<string, []>("op_4148_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4150_cast_fp16 = sigmoid(x = var_4148_cast_fp16_1)[name = tensor<string, []>("op_4150_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x132_1_cast_fp16 = mul(x = var_4148_cast_fp16_0, y = var_4150_cast_fp16)[name = tensor<string, []>("x132_1_cast_fp16")];
            tensor<int32, [3]> var_4152 = const()[name = tensor<string, []>("op_4152"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 4]> x133_1_cast_fp16 = transpose(perm = var_4152, x = x132_1_cast_fp16)[name = tensor<string, []>("transpose_51")];
            tensor<fp16, [1, 128, 4]> x134_1_cast_fp16 = select(a = var_62_to_fp16, b = x133_1_cast_fp16, cond = var_2859)[name = tensor<string, []>("x134_1_cast_fp16")];
            tensor<bool, []> x135_1_interleave_0 = const()[name = tensor<string, []>("x135_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s4_l0_conv2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s4_l0_conv2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 7]> cache_s4_l0_conv2_to_fp16 = cast(dtype = cache_s4_l0_conv2_to_fp16_dtype_0, x = cache_s4_l0_conv2)[name = tensor<string, []>("cast_6")];
            tensor<fp16, [1, 128, 11]> x135_1_cast_fp16 = concat(axis = var_49, interleave = x135_1_interleave_0, values = (cache_s4_l0_conv2_to_fp16, x134_1_cast_fp16))[name = tensor<string, []>("x135_1_cast_fp16")];
            tensor<int32, [3]> var_4162_begin_0 = const()[name = tensor<string, []>("op_4162_begin_0"), val = tensor<int32, [3]>([0, 0, 4])];
            tensor<int32, [3]> var_4162_end_0 = const()[name = tensor<string, []>("op_4162_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> var_4162_end_mask_0 = const()[name = tensor<string, []>("op_4162_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 7]> new_cache_s4_l0_conv2 = slice_by_index(begin = var_4162_begin_0, end = var_4162_end_0, end_mask = var_4162_end_mask_0, x = x135_1_cast_fp16)[name = tensor<string, []>("op_4162_cast_fp16")];
            tensor<string, []> x_causal_20_pad_type_0 = const()[name = tensor<string, []>("x_causal_20_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_20_groups_0 = const()[name = tensor<string, []>("x_causal_20_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_20_strides_0 = const()[name = tensor<string, []>("x_causal_20_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_20_pad_0 = const()[name = tensor<string, []>("x_causal_20_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_20_dilations_0 = const()[name = tensor<string, []>("x_causal_20_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 8]> encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2400832)))];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2402944)))];
            tensor<fp16, [1, 128, 4]> x_causal_20_cast_fp16 = conv(bias = encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_20_dilations_0, groups = x_causal_20_groups_0, pad = x_causal_20_pad_0, pad_type = x_causal_20_pad_type_0, strides = x_causal_20_strides_0, weight = encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16, x = x135_1_cast_fp16)[name = tensor<string, []>("x_causal_20_cast_fp16")];
            tensor<int32, [3]> input65_1_begin_0 = const()[name = tensor<string, []>("input65_1_begin_0"), val = tensor<int32, [3]>([0, 0, 7])];
            tensor<int32, [3]> input65_1_end_0 = const()[name = tensor<string, []>("input65_1_end_0"), val = tensor<int32, [3]>([1, 128, 11])];
            tensor<bool, [3]> input65_1_end_mask_0 = const()[name = tensor<string, []>("input65_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 4]> input65_1_cast_fp16 = slice_by_index(begin = input65_1_begin_0, end = input65_1_end_0, end_mask = input65_1_end_mask_0, x = x135_1_cast_fp16)[name = tensor<string, []>("input65_1_cast_fp16")];
            tensor<string, []> x_chunk_23_pad_type_0 = const()[name = tensor<string, []>("x_chunk_23_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_23_pad_0 = const()[name = tensor<string, []>("x_chunk_23_pad_0"), val = tensor<int32, [2]>([7, 7])];
            tensor<int32, []> x_chunk_23_groups_0 = const()[name = tensor<string, []>("x_chunk_23_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_23_strides_0 = const()[name = tensor<string, []>("x_chunk_23_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_23_dilations_0 = const()[name = tensor<string, []>("x_chunk_23_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 15]> encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 15]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2403264)))];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2407168)))];
            tensor<fp16, [1, 128, 4]> x_chunk_23_cast_fp16 = conv(bias = encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_23_dilations_0, groups = x_chunk_23_groups_0, pad = x_chunk_23_pad_0, pad_type = x_chunk_23_pad_type_0, strides = x_chunk_23_strides_0, weight = encoder_encoders_4_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16, x = input65_1_cast_fp16)[name = tensor<string, []>("x_chunk_23_cast_fp16")];
            tensor<fp16, [128, 4]> chunk_scale8_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale8_1_to_fp16"), val = tensor<fp16, [128, 4]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2407488)))];
            tensor<fp16, [1, 128, 4]> x_chunk8_1_cast_fp16 = mul(x = x_chunk_23_cast_fp16, y = chunk_scale8_1_to_fp16)[name = tensor<string, []>("x_chunk8_1_cast_fp16")];
            tensor<fp16, [1, 128, 4]> x136_1_cast_fp16 = add(x = x_chunk8_1_cast_fp16, y = x_causal_20_cast_fp16)[name = tensor<string, []>("x136_1_cast_fp16")];
            tensor<int32, [3]> var_4202 = const()[name = tensor<string, []>("op_4202"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_4211_to_fp16 = const()[name = tensor<string, []>("op_4211_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> x137_1_cast_fp16 = transpose(perm = var_4202, x = x136_1_cast_fp16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [4, 1, 128]> x_offset_48_cast_fp16 = sub(x = x137_1_cast_fp16, y = var_4211_to_fp16)[name = tensor<string, []>("x_offset_48_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4213_cast_fp16 = exp(x = x_offset_48_cast_fp16)[name = tensor<string, []>("op_4213_cast_fp16")];
            tensor<fp16, []> var_4214_to_fp16 = const()[name = tensor<string, []>("op_4214_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 128]> var_4215_cast_fp16 = add(x = var_4213_cast_fp16, y = var_4214_to_fp16)[name = tensor<string, []>("op_4215_cast_fp16")];
            tensor<fp32, []> var_4216_epsilon_0 = const()[name = tensor<string, []>("op_4216_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 128]> var_4216_cast_fp16 = log(epsilon = var_4216_epsilon_0, x = var_4215_cast_fp16)[name = tensor<string, []>("op_4216_cast_fp16")];
            tensor<fp16, []> var_4208_to_fp16 = const()[name = tensor<string, []>("op_4208_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 128]> var_4218_cast_fp16 = equal(x = var_4216_cast_fp16, y = var_4208_to_fp16)[name = tensor<string, []>("op_4218_cast_fp16")];
            tensor<fp16, [4, 1, 128]> log_sum0_48_cast_fp16 = select(a = x_offset_48_cast_fp16, b = var_4216_cast_fp16, cond = var_4218_cast_fp16)[name = tensor<string, []>("log_sum0_48_cast_fp16")];
            tensor<fp16, []> var_4220_to_fp16 = const()[name = tensor<string, []>("op_4220_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 128]> var_4221_cast_fp16 = mul(x = x137_1_cast_fp16, y = var_4220_to_fp16)[name = tensor<string, []>("op_4221_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4222_cast_fp16 = sub(x = log_sum0_48_cast_fp16, y = var_4221_cast_fp16)[name = tensor<string, []>("op_4222_cast_fp16")];
            tensor<fp16, []> var_4223_to_fp16 = const()[name = tensor<string, []>("op_4223_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [4, 1, 128]> x0_27_cast_fp16 = sub(x = var_4222_cast_fp16, y = var_4223_to_fp16)[name = tensor<string, []>("x0_27_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_4_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2408576))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2425024))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2425600)))];
            tensor<fp16, [4, 1, 128]> linear_88_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_conv_module2_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized, x = x0_27_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
            tensor<fp16, [4, 1, 128]> input66_1_cast_fp16 = add(x = input63_1_cast_fp16, y = linear_88_cast_fp16)[name = tensor<string, []>("input66_1_cast_fp16")];
            tensor<fp16, [240, 128]> encoder_encoders_4_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2425920))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2456704))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([240, 128])];
            tensor<fp16, [240]> encoder_encoders_4_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2457280)))];
            tensor<fp16, [4, 1, 240]> linear_89_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized, x = input66_1_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
            tensor<fp16, []> var_4239_to_fp16 = const()[name = tensor<string, []>("op_4239_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [4, 1, 240]> x_offset_50_cast_fp16 = sub(x = linear_89_cast_fp16, y = var_4239_to_fp16)[name = tensor<string, []>("x_offset_50_cast_fp16")];
            tensor<fp16, [4, 1, 240]> var_4241_cast_fp16 = exp(x = x_offset_50_cast_fp16)[name = tensor<string, []>("op_4241_cast_fp16")];
            tensor<fp16, []> var_4242_to_fp16 = const()[name = tensor<string, []>("op_4242_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [4, 1, 240]> var_4243_cast_fp16 = add(x = var_4241_cast_fp16, y = var_4242_to_fp16)[name = tensor<string, []>("op_4243_cast_fp16")];
            tensor<fp32, []> var_4244_epsilon_0 = const()[name = tensor<string, []>("op_4244_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [4, 1, 240]> var_4244_cast_fp16 = log(epsilon = var_4244_epsilon_0, x = var_4243_cast_fp16)[name = tensor<string, []>("op_4244_cast_fp16")];
            tensor<fp16, []> var_4227_to_fp16 = const()[name = tensor<string, []>("op_4227_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [4, 1, 240]> var_4246_cast_fp16 = equal(x = var_4244_cast_fp16, y = var_4227_to_fp16)[name = tensor<string, []>("op_4246_cast_fp16")];
            tensor<fp16, [4, 1, 240]> log_sum0_50_cast_fp16 = select(a = x_offset_50_cast_fp16, b = var_4244_cast_fp16, cond = var_4246_cast_fp16)[name = tensor<string, []>("log_sum0_50_cast_fp16")];
            tensor<fp16, []> var_4248_to_fp16 = const()[name = tensor<string, []>("op_4248_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [4, 1, 240]> var_4249_cast_fp16 = mul(x = linear_89_cast_fp16, y = var_4248_to_fp16)[name = tensor<string, []>("op_4249_cast_fp16")];
            tensor<fp16, [4, 1, 240]> var_4250_cast_fp16 = sub(x = log_sum0_50_cast_fp16, y = var_4249_cast_fp16)[name = tensor<string, []>("op_4250_cast_fp16")];
            tensor<fp16, []> var_4251_to_fp16 = const()[name = tensor<string, []>("op_4251_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [4, 1, 240]> x_77_cast_fp16 = sub(x = var_4250_cast_fp16, y = var_4251_to_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
            tensor<fp16, [128, 240]> encoder_encoders_4_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2457856))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2488640))), name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 240])];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2489216)))];
            tensor<fp16, [4, 1, 128]> linear_90_cast_fp16 = linear(bias = encoder_encoders_4_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16, weight = encoder_encoders_4_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized, x = x_77_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
            tensor<fp16, [4, 1, 128]> x138_1_cast_fp16 = add(x = input66_1_cast_fp16, y = linear_90_cast_fp16)[name = tensor<string, []>("x138_1_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2489536)))];
            tensor<fp16, [4, 1, 128]> var_4262_cast_fp16 = sub(x = x138_1_cast_fp16, y = encoder_encoders_4_encoder_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("op_4262_cast_fp16")];
            tensor<fp16, []> var_4256_promoted_to_fp16 = const()[name = tensor<string, []>("op_4256_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [4, 1, 128]> var_4263_cast_fp16 = pow(x = var_4262_cast_fp16, y = var_4256_promoted_to_fp16)[name = tensor<string, []>("op_4263_cast_fp16")];
            tensor<int32, [1]> var_4265_axes_0 = const()[name = tensor<string, []>("op_4265_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, []> var_4265_keep_dims_0 = const()[name = tensor<string, []>("op_4265_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [4, 1, 1]> var_4265_cast_fp16 = reduce_mean(axes = var_4265_axes_0, keep_dims = var_4265_keep_dims_0, x = var_4263_cast_fp16)[name = tensor<string, []>("op_4265_cast_fp16")];
            tensor<fp16, []> var_4255_to_fp16 = const()[name = tensor<string, []>("op_4255_to_fp16"), val = tensor<fp16, []>(-0x1p-1)];
            tensor<fp16, [4, 1, 1]> var_4266_cast_fp16 = pow(x = var_4265_cast_fp16, y = var_4255_to_fp16)[name = tensor<string, []>("op_4266_cast_fp16")];
            tensor<fp16, []> var_4267_to_fp16 = const()[name = tensor<string, []>("op_4267_to_fp16"), val = tensor<fp16, []>(0x1.5ecp-2)];
            tensor<fp16, [4, 1, 1]> scales_12_cast_fp16 = mul(x = var_4266_cast_fp16, y = var_4267_to_fp16)[name = tensor<string, []>("scales_12_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4269_cast_fp16 = mul(x = x138_1_cast_fp16, y = scales_12_cast_fp16)[name = tensor<string, []>("op_4269_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4272_cast_fp16 = sub(x = var_4269_cast_fp16, y = x_65_cast_fp16)[name = tensor<string, []>("op_4272_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_4_encoder_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_encoder_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2489856)))];
            tensor<fp16, [4, 1, 128]> var_4273_cast_fp16 = mul(x = var_4272_cast_fp16, y = encoder_encoders_4_encoder_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_4273_cast_fp16")];
            tensor<fp16, [4, 1, 128]> var_4274_cast_fp16 = add(x = x_65_cast_fp16, y = var_4273_cast_fp16)[name = tensor<string, []>("op_4274_cast_fp16")];
            tensor<int32, [1]> var_4284_axes_0 = const()[name = tensor<string, []>("op_4284_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [4, 1, 1, 128]> var_4284_cast_fp16 = expand_dims(axes = var_4284_axes_0, x = var_4274_cast_fp16)[name = tensor<string, []>("op_4284_cast_fp16")];
            tensor<int32, [4]> src_11_reps_0 = const()[name = tensor<string, []>("src_11_reps_0"), val = tensor<int32, [4]>([1, 4, 1, 1])];
            tensor<fp16, [4, 4, 1, 128]> src_11_cast_fp16 = tile(reps = src_11_reps_0, x = var_4284_cast_fp16)[name = tensor<string, []>("src_11_cast_fp16")];
            tensor<int32, [3]> var_4290 = const()[name = tensor<string, []>("op_4290"), val = tensor<int32, [3]>([16, 1, 128])];
            tensor<fp16, [16, 1, 128]> var_4291_cast_fp16 = reshape(shape = var_4290, x = src_11_cast_fp16)[name = tensor<string, []>("op_4291_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_4298_cast_fp16 = sub(x = var_4291_cast_fp16, y = var_3729_cast_fp16)[name = tensor<string, []>("op_4298_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_4_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_4_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2490176)))];
            tensor<fp16, [16, 1, 128]> var_4299_cast_fp16 = mul(x = var_4298_cast_fp16, y = encoder_encoders_4_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_4299_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_4300_cast_fp16 = add(x = var_3729_cast_fp16, y = var_4299_cast_fp16)[name = tensor<string, []>("op_4300_cast_fp16")];
            tensor<int32, [4]> var_4321 = const()[name = tensor<string, []>("op_4321"), val = tensor<int32, [4]>([8, 2, 1, 128])];
            tensor<fp16, [8, 2, 1, 128]> src0_17_cast_fp16 = reshape(shape = var_4321, x = var_4300_cast_fp16)[name = tensor<string, []>("src0_17_cast_fp16")];
            tensor<fp16, [2, 1, 1]> weights0_10_to_fp16 = const()[name = tensor<string, []>("weights0_10_to_fp16"), val = tensor<fp16, [2, 1, 1]>([[[0x1.b1cp-1]], [[0x1.394p-3]]])];
            tensor<fp16, [8, 2, 1, 128]> var_4326_cast_fp16 = mul(x = src0_17_cast_fp16, y = weights0_10_to_fp16)[name = tensor<string, []>("op_4326_cast_fp16")];
            tensor<int32, [1]> x_79_axes_0 = const()[name = tensor<string, []>("x_79_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> x_79_keep_dims_0 = const()[name = tensor<string, []>("x_79_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [8, 1, 128]> x_79_cast_fp16 = reduce_sum(axes = x_79_axes_0, keep_dims = x_79_keep_dims_0, x = var_4326_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")];
            tensor<fp16, [272, 128]> encoder_encoders_5_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [34816]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2490496))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2525376))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([272, 128])];
            tensor<fp16, [272]> encoder_encoders_5_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16"), val = tensor<fp16, [272]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2525952)))];
            tensor<fp16, [8, 1, 272]> linear_91_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_self_attn_weights_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_self_attn_weights_in_proj_weight_to_fp16_palettized, x = x_79_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
            tensor<int32, [3]> q14_1_begin_0 = const()[name = tensor<string, []>("q14_1_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> q14_1_end_0 = const()[name = tensor<string, []>("q14_1_end_0"), val = tensor<int32, [3]>([8, 1, 128])];
            tensor<bool, [3]> q14_1_end_mask_0 = const()[name = tensor<string, []>("q14_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 1, 128]> q14_1_cast_fp16 = slice_by_index(begin = q14_1_begin_0, end = q14_1_end_0, end_mask = q14_1_end_mask_0, x = linear_91_cast_fp16)[name = tensor<string, []>("q14_1_cast_fp16")];
            tensor<int32, [3]> k19_1_begin_0 = const()[name = tensor<string, []>("k19_1_begin_0"), val = tensor<int32, [3]>([0, 0, 128])];
            tensor<int32, [3]> k19_1_end_0 = const()[name = tensor<string, []>("k19_1_end_0"), val = tensor<int32, [3]>([8, 1, 256])];
            tensor<bool, [3]> k19_1_end_mask_0 = const()[name = tensor<string, []>("k19_1_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<fp16, [8, 1, 128]> k19_1_cast_fp16 = slice_by_index(begin = k19_1_begin_0, end = k19_1_end_0, end_mask = k19_1_end_mask_0, x = linear_91_cast_fp16)[name = tensor<string, []>("k19_1_cast_fp16")];
            tensor<int32, [3]> p14_1_begin_0 = const()[name = tensor<string, []>("p14_1_begin_0"), val = tensor<int32, [3]>([0, 0, 256])];
            tensor<int32, [3]> p14_1_end_0 = const()[name = tensor<string, []>("p14_1_end_0"), val = tensor<int32, [3]>([8, 1, 272])];
            tensor<bool, [3]> p14_1_end_mask_0 = const()[name = tensor<string, []>("p14_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [8, 1, 16]> p14_1_cast_fp16 = slice_by_index(begin = p14_1_begin_0, end = p14_1_end_0, end_mask = p14_1_end_mask_0, x = linear_91_cast_fp16)[name = tensor<string, []>("p14_1_cast_fp16")];
            tensor<bool, []> k20_1_interleave_0 = const()[name = tensor<string, []>("k20_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s5_l0_key_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s5_l0_key_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [32, 1, 128]> cache_s5_l0_key_to_fp16 = cast(dtype = cache_s5_l0_key_to_fp16_dtype_0, x = cache_s5_l0_key)[name = tensor<string, []>("cast_5")];
            tensor<fp16, [40, 1, 128]> k20_1_cast_fp16 = concat(axis = var_46, interleave = k20_1_interleave_0, values = (cache_s5_l0_key_to_fp16, k19_1_cast_fp16))[name = tensor<string, []>("k20_1_cast_fp16")];
            tensor<int32, [3]> var_4379_begin_0 = const()[name = tensor<string, []>("op_4379_begin_0"), val = tensor<int32, [3]>([8, 0, 0])];
            tensor<int32, [3]> var_4379_end_0 = const()[name = tensor<string, []>("op_4379_end_0"), val = tensor<int32, [3]>([40, 1, 128])];
            tensor<bool, [3]> var_4379_end_mask_0 = const()[name = tensor<string, []>("op_4379_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [32, 1, 128]> new_cache_s5_l0_key = slice_by_index(begin = var_4379_begin_0, end = var_4379_end_0, end_mask = var_4379_end_mask_0, x = k20_1_cast_fp16)[name = tensor<string, []>("op_4379_cast_fp16")];
            tensor<int32, [4]> var_4385 = const()[name = tensor<string, []>("op_4385"), val = tensor<int32, [4]>([8, 1, 4, 32])];
            tensor<fp16, [8, 1, 4, 32]> q15_1_cast_fp16 = reshape(shape = var_4385, x = q14_1_cast_fp16)[name = tensor<string, []>("q15_1_cast_fp16")];
            tensor<int32, [4]> var_4387 = const()[name = tensor<string, []>("op_4387"), val = tensor<int32, [4]>([8, 1, 4, 4])];
            tensor<fp16, [8, 1, 4, 4]> p15_1_cast_fp16 = reshape(shape = var_4387, x = p14_1_cast_fp16)[name = tensor<string, []>("p15_1_cast_fp16")];
            tensor<int32, [4]> var_4389 = const()[name = tensor<string, []>("op_4389"), val = tensor<int32, [4]>([40, 1, 4, 32])];
            tensor<fp16, [40, 1, 4, 32]> k21_1_cast_fp16 = reshape(shape = var_4389, x = k20_1_cast_fp16)[name = tensor<string, []>("k21_1_cast_fp16")];
            tensor<int32, [4]> var_4393 = const()[name = tensor<string, []>("op_4393"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> attn_scores14_1_transpose_x_0 = const()[name = tensor<string, []>("attn_scores14_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_scores14_1_transpose_y_0 = const()[name = tensor<string, []>("attn_scores14_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_34_perm_0 = const()[name = tensor<string, []>("transpose_34_perm_0"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<int32, [4]> transpose_35_perm_0 = const()[name = tensor<string, []>("transpose_35_perm_0"), val = tensor<int32, [4]>([2, 1, -1, -4])];
            tensor<fp16, [4, 1, 32, 40]> transpose_35 = transpose(perm = transpose_35_perm_0, x = k21_1_cast_fp16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [4, 1, 8, 32]> transpose_34 = transpose(perm = transpose_34_perm_0, x = q15_1_cast_fp16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [4, 1, 8, 40]> attn_scores14_1_cast_fp16 = matmul(transpose_x = attn_scores14_1_transpose_x_0, transpose_y = attn_scores14_1_transpose_y_0, x = transpose_34, y = transpose_35)[name = tensor<string, []>("attn_scores14_1_cast_fp16")];
            tensor<bool, []> pos_scores19_1_transpose_x_0 = const()[name = tensor<string, []>("pos_scores19_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> pos_scores19_1_transpose_y_0 = const()[name = tensor<string, []>("pos_scores19_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 4, 47]> pos_emb4_1_to_fp16 = const()[name = tensor<string, []>("pos_emb4_1_to_fp16"), val = tensor<fp16, [4, 1, 4, 47]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2526592)))];
            tensor<fp16, [4, 1, 8, 4]> p16_1_cast_fp16 = transpose(perm = var_4393, x = p15_1_cast_fp16)[name = tensor<string, []>("transpose_47")];
            tensor<fp16, [4, 1, 8, 47]> pos_scores19_1_cast_fp16 = matmul(transpose_x = pos_scores19_1_transpose_x_0, transpose_y = pos_scores19_1_transpose_y_0, x = p16_1_cast_fp16, y = pos_emb4_1_to_fp16)[name = tensor<string, []>("pos_scores19_1_cast_fp16")];
            tensor<int32, [2]> var_4436 = const()[name = tensor<string, []>("op_4436"), val = tensor<int32, [2]>([-1, 47])];
            tensor<fp16, [32, 47]> pos_scores20_1_cast_fp16 = reshape(shape = var_4436, x = pos_scores19_1_cast_fp16)[name = tensor<string, []>("pos_scores20_1_cast_fp16")];
            tensor<bool, []> pos_scores21_1_validate_indices_0 = const()[name = tensor<string, []>("pos_scores21_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [32, 40]> pos_scores21_1_cast_fp16_cast_uint16 = gather_along_axis(axis = var_48, indices = indexes0_1_to_uint16, validate_indices = pos_scores21_1_validate_indices_0, x = pos_scores20_1_cast_fp16)[name = tensor<string, []>("pos_scores21_1_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_4439 = const()[name = tensor<string, []>("op_4439"), val = tensor<int32, [4]>([4, 1, 8, 40])];
            tensor<fp16, [4, 1, 8, 40]> pos_scores22_1_cast_fp16 = reshape(shape = var_4439, x = pos_scores21_1_cast_fp16_cast_uint16)[name = tensor<string, []>("pos_scores22_1_cast_fp16")];
            tensor<fp16, [4, 1, 8, 40]> attn_scores15_1_cast_fp16 = add(x = attn_scores14_1_cast_fp16, y = pos_scores22_1_cast_fp16)[name = tensor<string, []>("attn_scores15_1_cast_fp16")];
            tensor<fp16, []> cast_111_to_fp16 = const()[name = tensor<string, []>("cast_111_to_fp16"), val = tensor<fp16, []>(-0x1.f4p+9)];
            tensor<fp16, [4, 1, 8, 40]> attn_scores16_1_cast_fp16 = select(a = cast_111_to_fp16, b = attn_scores15_1_cast_fp16, cond = var_2160)[name = tensor<string, []>("attn_scores16_1_cast_fp16")];
            tensor<fp16, [4, 1, 8, 40]> attn_weights9_1_cast_fp16 = softmax(axis = var_59, x = attn_scores16_1_cast_fp16)[name = tensor<string, []>("attn_weights9_1_cast_fp16")];
            tensor<fp16, [144, 128]> encoder_encoders_5_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2528192))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2546688))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([144, 128])];
            tensor<fp16, [144]> encoder_encoders_5_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2547264)))];
            tensor<fp16, [8, 1, 144]> linear_93_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_feed_forward1_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_feed_forward1_in_proj_weight_to_fp16_palettized, x = x_79_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
            tensor<fp16, []> var_4457_to_fp16 = const()[name = tensor<string, []>("op_4457_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [8, 1, 144]> x_offset_52_cast_fp16 = sub(x = linear_93_cast_fp16, y = var_4457_to_fp16)[name = tensor<string, []>("x_offset_52_cast_fp16")];
            tensor<fp16, [8, 1, 144]> var_4459_cast_fp16 = exp(x = x_offset_52_cast_fp16)[name = tensor<string, []>("op_4459_cast_fp16")];
            tensor<fp16, []> var_4460_to_fp16 = const()[name = tensor<string, []>("op_4460_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 144]> var_4461_cast_fp16 = add(x = var_4459_cast_fp16, y = var_4460_to_fp16)[name = tensor<string, []>("op_4461_cast_fp16")];
            tensor<fp32, []> var_4462_epsilon_0 = const()[name = tensor<string, []>("op_4462_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 144]> var_4462_cast_fp16 = log(epsilon = var_4462_epsilon_0, x = var_4461_cast_fp16)[name = tensor<string, []>("op_4462_cast_fp16")];
            tensor<fp16, []> var_4445_to_fp16 = const()[name = tensor<string, []>("op_4445_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 144]> var_4464_cast_fp16 = equal(x = var_4462_cast_fp16, y = var_4445_to_fp16)[name = tensor<string, []>("op_4464_cast_fp16")];
            tensor<fp16, [8, 1, 144]> log_sum0_52_cast_fp16 = select(a = x_offset_52_cast_fp16, b = var_4462_cast_fp16, cond = var_4464_cast_fp16)[name = tensor<string, []>("log_sum0_52_cast_fp16")];
            tensor<fp16, []> var_4466_to_fp16 = const()[name = tensor<string, []>("op_4466_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 144]> var_4467_cast_fp16 = mul(x = linear_93_cast_fp16, y = var_4466_to_fp16)[name = tensor<string, []>("op_4467_cast_fp16")];
            tensor<fp16, [8, 1, 144]> var_4468_cast_fp16 = sub(x = log_sum0_52_cast_fp16, y = var_4467_cast_fp16)[name = tensor<string, []>("op_4468_cast_fp16")];
            tensor<fp16, []> var_4469_to_fp16 = const()[name = tensor<string, []>("op_4469_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [8, 1, 144]> x_85_cast_fp16 = sub(x = var_4468_cast_fp16, y = var_4469_to_fp16)[name = tensor<string, []>("x_85_cast_fp16")];
            tensor<fp16, [128, 144]> encoder_encoders_5_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [18432]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2547648))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2566144))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 144])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2566720)))];
            tensor<fp16, [8, 1, 128]> linear_94_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_feed_forward1_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_feed_forward1_out_proj_weight_to_fp16_palettized, x = x_85_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input67_1_cast_fp16 = add(x = x_79_cast_fp16, y = linear_94_cast_fp16)[name = tensor<string, []>("input67_1_cast_fp16")];
            tensor<int32, [4]> attn_weights10_1_begin_0 = const()[name = tensor<string, []>("attn_weights10_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> attn_weights10_1_end_0 = const()[name = tensor<string, []>("attn_weights10_1_end_0"), val = tensor<int32, [4]>([1, 1, 8, 40])];
            tensor<bool, [4]> attn_weights10_1_end_mask_0 = const()[name = tensor<string, []>("attn_weights10_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 1, 8, 40]> attn_weights10_1_cast_fp16 = slice_by_index(begin = attn_weights10_1_begin_0, end = attn_weights10_1_end_0, end_mask = attn_weights10_1_end_mask_0, x = attn_weights9_1_cast_fp16)[name = tensor<string, []>("attn_weights10_1_cast_fp16")];
            tensor<fp16, [288, 128]> encoder_encoders_5_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [36864]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2567040))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2603968))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([288, 128])];
            tensor<fp16, [288]> encoder_encoders_5_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16"), val = tensor<fp16, [288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2604544)))];
            tensor<fp16, [8, 1, 288]> linear_95_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_nonlin_attention_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_nonlin_attention_in_proj_weight_to_fp16_palettized, x = input67_1_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
            tensor<int32, [3]> var_4483_split_sizes_0 = const()[name = tensor<string, []>("op_4483_split_sizes_0"), val = tensor<int32, [3]>([96, 96, 96])];
            tensor<int32, []> var_4483_axis_0 = const()[name = tensor<string, []>("op_4483_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [8, 1, 96]> var_4483_cast_fp16_0, tensor<fp16, [8, 1, 96]> var_4483_cast_fp16_1, tensor<fp16, [8, 1, 96]> var_4483_cast_fp16_2 = split(axis = var_4483_axis_0, split_sizes = var_4483_split_sizes_0, x = linear_95_cast_fp16)[name = tensor<string, []>("op_4483_cast_fp16")];
            tensor<fp16, [8, 1, 96]> var_4486_cast_fp16 = tanh(x = var_4483_cast_fp16_0)[name = tensor<string, []>("op_4486_cast_fp16")];
            tensor<int32, [1]> var_4487_axes_0 = const()[name = tensor<string, []>("op_4487_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [8, 1, 96, 1]> var_4487_cast_fp16 = expand_dims(axes = var_4487_axes_0, x = var_4486_cast_fp16)[name = tensor<string, []>("op_4487_cast_fp16")];
            tensor<int32, [3]> var_4488 = const()[name = tensor<string, []>("op_4488"), val = tensor<int32, [3]>([8, 1, 96])];
            tensor<fp16, [8, 1, 96]> s4_1_cast_fp16 = reshape(shape = var_4488, x = var_4487_cast_fp16)[name = tensor<string, []>("s4_1_cast_fp16")];
            tensor<fp16, [8, 1, 96]> x140_1_cast_fp16 = mul(x = var_4483_cast_fp16_1, y = s4_1_cast_fp16)[name = tensor<string, []>("x140_1_cast_fp16")];
            tensor<int32, [4]> var_4502 = const()[name = tensor<string, []>("op_4502"), val = tensor<int32, [4]>([8, 1, 1, -1])];
            tensor<fp16, [8, 1, 1, 96]> var_4503_cast_fp16 = reshape(shape = var_4502, x = x140_1_cast_fp16)[name = tensor<string, []>("op_4503_cast_fp16")];
            tensor<int32, [4]> var_4504 = const()[name = tensor<string, []>("op_4504"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<bool, []> x_pad4_1_interleave_0 = const()[name = tensor<string, []>("x_pad4_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s5_l0_nonlin_attn_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s5_l0_nonlin_attn_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 8, 96]> x141_1_cast_fp16 = transpose(perm = var_4504, x = var_4503_cast_fp16)[name = tensor<string, []>("transpose_46")];
            tensor<fp16, [1, 1, 32, 96]> cache_s5_l0_nonlin_attn_to_fp16 = cast(dtype = cache_s5_l0_nonlin_attn_to_fp16_dtype_0, x = cache_s5_l0_nonlin_attn)[name = tensor<string, []>("cast_4")];
            tensor<fp16, [1, 1, 40, 96]> x_pad4_1_cast_fp16 = concat(axis = var_49, interleave = x_pad4_1_interleave_0, values = (cache_s5_l0_nonlin_attn_to_fp16, x141_1_cast_fp16))[name = tensor<string, []>("x_pad4_1_cast_fp16")];
            tensor<int32, [4]> var_4510_begin_0 = const()[name = tensor<string, []>("op_4510_begin_0"), val = tensor<int32, [4]>([0, 0, 8, 0])];
            tensor<int32, [4]> var_4510_end_0 = const()[name = tensor<string, []>("op_4510_end_0"), val = tensor<int32, [4]>([1, 1, 40, 96])];
            tensor<bool, [4]> var_4510_end_mask_0 = const()[name = tensor<string, []>("op_4510_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 32, 96]> new_cache_s5_l0_nonlin_attn = slice_by_index(begin = var_4510_begin_0, end = var_4510_end_0, end_mask = var_4510_end_mask_0, x = x_pad4_1_cast_fp16)[name = tensor<string, []>("op_4510_cast_fp16")];
            tensor<bool, []> x142_1_transpose_x_0 = const()[name = tensor<string, []>("x142_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x142_1_transpose_y_0 = const()[name = tensor<string, []>("x142_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 8, 96]> x142_1_cast_fp16 = matmul(transpose_x = x142_1_transpose_x_0, transpose_y = x142_1_transpose_y_0, x = attn_weights10_1_cast_fp16, y = x_pad4_1_cast_fp16)[name = tensor<string, []>("x142_1_cast_fp16")];
            tensor<int32, [4]> var_4513 = const()[name = tensor<string, []>("op_4513"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_4515 = const()[name = tensor<string, []>("op_4515"), val = tensor<int32, [3]>([8, 1, -1])];
            tensor<fp16, [8, 1, 1, 96]> var_4514_cast_fp16 = transpose(perm = var_4513, x = x142_1_cast_fp16)[name = tensor<string, []>("transpose_45")];
            tensor<fp16, [8, 1, 96]> x143_1_cast_fp16 = reshape(shape = var_4515, x = var_4514_cast_fp16)[name = tensor<string, []>("x143_1_cast_fp16")];
            tensor<fp16, [8, 1, 96]> input69_1_cast_fp16 = mul(x = x143_1_cast_fp16, y = var_4483_cast_fp16_2)[name = tensor<string, []>("input69_1_cast_fp16")];
            tensor<fp16, [128, 96]> encoder_encoders_5_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [12288]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2605184))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2617536))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 96])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2618112)))];
            tensor<fp16, [8, 1, 128]> linear_96_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_nonlin_attention_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_nonlin_attention_out_proj_weight_to_fp16_palettized, x = input69_1_cast_fp16)[name = tensor<string, []>("linear_96_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x144_1_cast_fp16 = add(x = input67_1_cast_fp16, y = linear_96_cast_fp16)[name = tensor<string, []>("x144_1_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_5_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2618432))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2624640))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_5_encoder_layers_0_self_attn1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn1_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2625216)))];
            tensor<fp16, [8, 1, 48]> linear_97_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_self_attn1_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_self_attn1_in_proj_weight_to_fp16_palettized, x = x144_1_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
            tensor<bool, []> x145_1_interleave_0 = const()[name = tensor<string, []>("x145_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s5_l0_val1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s5_l0_val1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [32, 1, 48]> cache_s5_l0_val1_to_fp16 = cast(dtype = cache_s5_l0_val1_to_fp16_dtype_0, x = cache_s5_l0_val1)[name = tensor<string, []>("cast_3")];
            tensor<fp16, [40, 1, 48]> x145_1_cast_fp16 = concat(axis = var_46, interleave = x145_1_interleave_0, values = (cache_s5_l0_val1_to_fp16, linear_97_cast_fp16))[name = tensor<string, []>("x145_1_cast_fp16")];
            tensor<int32, [3]> var_4540_begin_0 = const()[name = tensor<string, []>("op_4540_begin_0"), val = tensor<int32, [3]>([8, 0, 0])];
            tensor<int32, [3]> var_4540_end_0 = const()[name = tensor<string, []>("op_4540_end_0"), val = tensor<int32, [3]>([40, 1, 48])];
            tensor<bool, [3]> var_4540_end_mask_0 = const()[name = tensor<string, []>("op_4540_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [32, 1, 48]> new_cache_s5_l0_val1 = slice_by_index(begin = var_4540_begin_0, end = var_4540_end_0, end_mask = var_4540_end_mask_0, x = x145_1_cast_fp16)[name = tensor<string, []>("op_4540_cast_fp16")];
            tensor<int32, [4]> var_4541 = const()[name = tensor<string, []>("op_4541"), val = tensor<int32, [4]>([40, 1, 4, -1])];
            tensor<fp16, [40, 1, 4, 12]> var_4542_cast_fp16 = reshape(shape = var_4541, x = x145_1_cast_fp16)[name = tensor<string, []>("op_4542_cast_fp16")];
            tensor<int32, [4]> var_4543 = const()[name = tensor<string, []>("op_4543"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x147_1_transpose_x_0 = const()[name = tensor<string, []>("x147_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x147_1_transpose_y_0 = const()[name = tensor<string, []>("x147_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 40, 12]> x146_1_cast_fp16 = transpose(perm = var_4543, x = var_4542_cast_fp16)[name = tensor<string, []>("transpose_44")];
            tensor<fp16, [4, 1, 8, 12]> x147_1_cast_fp16 = matmul(transpose_x = x147_1_transpose_x_0, transpose_y = x147_1_transpose_y_0, x = attn_weights9_1_cast_fp16, y = x146_1_cast_fp16)[name = tensor<string, []>("x147_1_cast_fp16")];
            tensor<int32, [4]> var_4548 = const()[name = tensor<string, []>("op_4548"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_4553 = const()[name = tensor<string, []>("op_4553"), val = tensor<int32, [3]>([8, 1, 48])];
            tensor<fp16, [8, 1, 4, 12]> var_4549_cast_fp16 = transpose(perm = var_4548, x = x147_1_cast_fp16)[name = tensor<string, []>("transpose_43")];
            tensor<fp16, [8, 1, 48]> input70_1_cast_fp16 = reshape(shape = var_4553, x = var_4549_cast_fp16)[name = tensor<string, []>("input70_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_5_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2625408))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2631616))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_self_attn1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2632192)))];
            tensor<fp16, [8, 1, 128]> linear_98_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_self_attn1_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_self_attn1_out_proj_weight_to_fp16_palettized, x = input70_1_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input71_1_cast_fp16 = add(x = x144_1_cast_fp16, y = linear_98_cast_fp16)[name = tensor<string, []>("input71_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_5_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2632512))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2665344))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_5_encoder_layers_0_conv_module1_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2665920)))];
            tensor<fp16, [8, 1, 256]> linear_99_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_conv_module1_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_conv_module1_in_proj_weight_to_fp16_palettized, x = input71_1_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
            tensor<int32, [2]> var_4564_split_sizes_0 = const()[name = tensor<string, []>("op_4564_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_4564_axis_0 = const()[name = tensor<string, []>("op_4564_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [8, 1, 128]> var_4564_cast_fp16_0, tensor<fp16, [8, 1, 128]> var_4564_cast_fp16_1 = split(axis = var_4564_axis_0, split_sizes = var_4564_split_sizes_0, x = linear_99_cast_fp16)[name = tensor<string, []>("op_4564_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4566_cast_fp16 = sigmoid(x = var_4564_cast_fp16_1)[name = tensor<string, []>("op_4566_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x149_1_cast_fp16 = mul(x = var_4564_cast_fp16_0, y = var_4566_cast_fp16)[name = tensor<string, []>("x149_1_cast_fp16")];
            tensor<int32, [3]> var_4568 = const()[name = tensor<string, []>("op_4568"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 8]> x150_1_cast_fp16 = transpose(perm = var_4568, x = x149_1_cast_fp16)[name = tensor<string, []>("transpose_42")];
            tensor<fp16, [1, 128, 8]> x151_1_cast_fp16 = select(a = var_62_to_fp16, b = x150_1_cast_fp16, cond = var_2289)[name = tensor<string, []>("x151_1_cast_fp16")];
            tensor<bool, []> x152_1_interleave_0 = const()[name = tensor<string, []>("x152_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s5_l0_conv1_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s5_l0_conv1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 15]> cache_s5_l0_conv1_to_fp16 = cast(dtype = cache_s5_l0_conv1_to_fp16_dtype_0, x = cache_s5_l0_conv1)[name = tensor<string, []>("cast_2")];
            tensor<fp16, [1, 128, 23]> x152_1_cast_fp16 = concat(axis = var_49, interleave = x152_1_interleave_0, values = (cache_s5_l0_conv1_to_fp16, x151_1_cast_fp16))[name = tensor<string, []>("x152_1_cast_fp16")];
            tensor<int32, [3]> var_4578_begin_0 = const()[name = tensor<string, []>("op_4578_begin_0"), val = tensor<int32, [3]>([0, 0, 8])];
            tensor<int32, [3]> var_4578_end_0 = const()[name = tensor<string, []>("op_4578_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> var_4578_end_mask_0 = const()[name = tensor<string, []>("op_4578_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 15]> new_cache_s5_l0_conv1 = slice_by_index(begin = var_4578_begin_0, end = var_4578_end_0, end_mask = var_4578_end_mask_0, x = x152_1_cast_fp16)[name = tensor<string, []>("op_4578_cast_fp16")];
            tensor<string, []> x_causal_22_pad_type_0 = const()[name = tensor<string, []>("x_causal_22_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_22_groups_0 = const()[name = tensor<string, []>("x_causal_22_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_22_strides_0 = const()[name = tensor<string, []>("x_causal_22_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_22_pad_0 = const()[name = tensor<string, []>("x_causal_22_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_22_dilations_0 = const()[name = tensor<string, []>("x_causal_22_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 16]> encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2666496)))];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2670656)))];
            tensor<fp16, [1, 128, 8]> x_causal_22_cast_fp16 = conv(bias = encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_22_dilations_0, groups = x_causal_22_groups_0, pad = x_causal_22_pad_0, pad_type = x_causal_22_pad_type_0, strides = x_causal_22_strides_0, weight = encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_causal_conv_weight_to_fp16, x = x152_1_cast_fp16)[name = tensor<string, []>("x_causal_22_cast_fp16")];
            tensor<int32, [3]> input73_1_begin_0 = const()[name = tensor<string, []>("input73_1_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> input73_1_end_0 = const()[name = tensor<string, []>("input73_1_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> input73_1_end_mask_0 = const()[name = tensor<string, []>("input73_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 8]> input73_1_cast_fp16 = slice_by_index(begin = input73_1_begin_0, end = input73_1_end_0, end_mask = input73_1_end_mask_0, x = x152_1_cast_fp16)[name = tensor<string, []>("input73_1_cast_fp16")];
            tensor<string, []> x_chunk_25_pad_type_0 = const()[name = tensor<string, []>("x_chunk_25_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_25_pad_0 = const()[name = tensor<string, []>("x_chunk_25_pad_0"), val = tensor<int32, [2]>([15, 15])];
            tensor<int32, []> x_chunk_25_groups_0 = const()[name = tensor<string, []>("x_chunk_25_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_25_strides_0 = const()[name = tensor<string, []>("x_chunk_25_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_25_dilations_0 = const()[name = tensor<string, []>("x_chunk_25_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 31]> encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3968]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2670976))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2675008))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 1, 31])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2675584)))];
            tensor<fp16, [1, 128, 8]> x_chunk_25_cast_fp16 = conv(bias = encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_25_dilations_0, groups = x_chunk_25_groups_0, pad = x_chunk_25_pad_0, pad_type = x_chunk_25_pad_type_0, strides = x_chunk_25_strides_0, weight = encoder_encoders_5_encoder_layers_0_conv_module1_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized, x = input73_1_cast_fp16)[name = tensor<string, []>("x_chunk_25_cast_fp16")];
            tensor<fp16, [128, 8]> chunk_scale9_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale9_1_to_fp16"), val = tensor<fp16, [128, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2675904)))];
            tensor<fp16, [1, 128, 8]> x_chunk9_1_cast_fp16 = mul(x = x_chunk_25_cast_fp16, y = chunk_scale9_1_to_fp16)[name = tensor<string, []>("x_chunk9_1_cast_fp16")];
            tensor<fp16, [1, 128, 8]> x153_1_cast_fp16 = add(x = x_chunk9_1_cast_fp16, y = x_causal_22_cast_fp16)[name = tensor<string, []>("x153_1_cast_fp16")];
            tensor<int32, [3]> var_4618 = const()[name = tensor<string, []>("op_4618"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_4627_to_fp16 = const()[name = tensor<string, []>("op_4627_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> x154_1_cast_fp16 = transpose(perm = var_4618, x = x153_1_cast_fp16)[name = tensor<string, []>("transpose_41")];
            tensor<fp16, [8, 1, 128]> x_offset_54_cast_fp16 = sub(x = x154_1_cast_fp16, y = var_4627_to_fp16)[name = tensor<string, []>("x_offset_54_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4629_cast_fp16 = exp(x = x_offset_54_cast_fp16)[name = tensor<string, []>("op_4629_cast_fp16")];
            tensor<fp16, []> var_4630_to_fp16 = const()[name = tensor<string, []>("op_4630_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> var_4631_cast_fp16 = add(x = var_4629_cast_fp16, y = var_4630_to_fp16)[name = tensor<string, []>("op_4631_cast_fp16")];
            tensor<fp32, []> var_4632_epsilon_0 = const()[name = tensor<string, []>("op_4632_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 128]> var_4632_cast_fp16 = log(epsilon = var_4632_epsilon_0, x = var_4631_cast_fp16)[name = tensor<string, []>("op_4632_cast_fp16")];
            tensor<fp16, []> var_4624_to_fp16 = const()[name = tensor<string, []>("op_4624_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 128]> var_4634_cast_fp16 = equal(x = var_4632_cast_fp16, y = var_4624_to_fp16)[name = tensor<string, []>("op_4634_cast_fp16")];
            tensor<fp16, [8, 1, 128]> log_sum0_54_cast_fp16 = select(a = x_offset_54_cast_fp16, b = var_4632_cast_fp16, cond = var_4634_cast_fp16)[name = tensor<string, []>("log_sum0_54_cast_fp16")];
            tensor<fp16, []> var_4636_to_fp16 = const()[name = tensor<string, []>("op_4636_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 128]> var_4637_cast_fp16 = mul(x = x154_1_cast_fp16, y = var_4636_to_fp16)[name = tensor<string, []>("op_4637_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4638_cast_fp16 = sub(x = log_sum0_54_cast_fp16, y = var_4637_cast_fp16)[name = tensor<string, []>("op_4638_cast_fp16")];
            tensor<fp16, []> var_4639_to_fp16 = const()[name = tensor<string, []>("op_4639_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [8, 1, 128]> x0_29_cast_fp16 = sub(x = var_4638_cast_fp16, y = var_4639_to_fp16)[name = tensor<string, []>("x0_29_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_5_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2678016))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2694464))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_conv_module1_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module1_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2695040)))];
            tensor<fp16, [8, 1, 128]> linear_100_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_conv_module1_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_conv_module1_out_proj_weight_to_fp16_palettized, x = x0_29_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input74_1_cast_fp16 = add(x = input71_1_cast_fp16, y = linear_100_cast_fp16)[name = tensor<string, []>("input74_1_cast_fp16")];
            tensor<fp16, [192, 128]> encoder_encoders_5_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2695360))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2720000))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([192, 128])];
            tensor<fp16, [192]> encoder_encoders_5_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2720576)))];
            tensor<fp16, [8, 1, 192]> linear_101_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_feed_forward2_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_feed_forward2_in_proj_weight_to_fp16_palettized, x = input74_1_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
            tensor<fp16, []> var_4655_to_fp16 = const()[name = tensor<string, []>("op_4655_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [8, 1, 192]> x_offset_56_cast_fp16 = sub(x = linear_101_cast_fp16, y = var_4655_to_fp16)[name = tensor<string, []>("x_offset_56_cast_fp16")];
            tensor<fp16, [8, 1, 192]> var_4657_cast_fp16 = exp(x = x_offset_56_cast_fp16)[name = tensor<string, []>("op_4657_cast_fp16")];
            tensor<fp16, []> var_4658_to_fp16 = const()[name = tensor<string, []>("op_4658_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 192]> var_4659_cast_fp16 = add(x = var_4657_cast_fp16, y = var_4658_to_fp16)[name = tensor<string, []>("op_4659_cast_fp16")];
            tensor<fp32, []> var_4660_epsilon_0 = const()[name = tensor<string, []>("op_4660_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 192]> var_4660_cast_fp16 = log(epsilon = var_4660_epsilon_0, x = var_4659_cast_fp16)[name = tensor<string, []>("op_4660_cast_fp16")];
            tensor<fp16, []> var_4643_to_fp16 = const()[name = tensor<string, []>("op_4643_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 192]> var_4662_cast_fp16 = equal(x = var_4660_cast_fp16, y = var_4643_to_fp16)[name = tensor<string, []>("op_4662_cast_fp16")];
            tensor<fp16, [8, 1, 192]> log_sum0_56_cast_fp16 = select(a = x_offset_56_cast_fp16, b = var_4660_cast_fp16, cond = var_4662_cast_fp16)[name = tensor<string, []>("log_sum0_56_cast_fp16")];
            tensor<fp16, []> var_4664_to_fp16 = const()[name = tensor<string, []>("op_4664_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 192]> var_4665_cast_fp16 = mul(x = linear_101_cast_fp16, y = var_4664_to_fp16)[name = tensor<string, []>("op_4665_cast_fp16")];
            tensor<fp16, [8, 1, 192]> var_4666_cast_fp16 = sub(x = log_sum0_56_cast_fp16, y = var_4665_cast_fp16)[name = tensor<string, []>("op_4666_cast_fp16")];
            tensor<fp16, []> var_4667_to_fp16 = const()[name = tensor<string, []>("op_4667_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [8, 1, 192]> x_87_cast_fp16 = sub(x = var_4666_cast_fp16, y = var_4667_to_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
            tensor<fp16, [128, 192]> encoder_encoders_5_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24576]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2721024))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2745664))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 192])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2746240)))];
            tensor<fp16, [8, 1, 128]> linear_102_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_feed_forward2_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_feed_forward2_out_proj_weight_to_fp16_palettized, x = x_87_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
            tensor<fp16, [8, 1, 128]> src13_1_cast_fp16 = add(x = input74_1_cast_fp16, y = linear_102_cast_fp16)[name = tensor<string, []>("src13_1_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4673_cast_fp16 = sub(x = src13_1_cast_fp16, y = x_79_cast_fp16)[name = tensor<string, []>("op_4673_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_bypass_mid_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_bypass_mid_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2746560)))];
            tensor<fp16, [8, 1, 128]> var_4674_cast_fp16 = mul(x = var_4673_cast_fp16, y = encoder_encoders_5_encoder_layers_0_bypass_mid_bypass_scale_to_fp16)[name = tensor<string, []>("op_4674_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4675_cast_fp16 = add(x = x_79_cast_fp16, y = var_4674_cast_fp16)[name = tensor<string, []>("op_4675_cast_fp16")];
            tensor<fp16, [48, 128]> encoder_encoders_5_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2746880))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2753088))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([48, 128])];
            tensor<fp16, [48]> encoder_encoders_5_encoder_layers_0_self_attn2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn2_in_proj_bias_to_fp16"), val = tensor<fp16, [48]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2753664)))];
            tensor<fp16, [8, 1, 48]> linear_103_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_self_attn2_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_self_attn2_in_proj_weight_to_fp16_palettized, x = var_4675_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
            tensor<bool, []> x155_1_interleave_0 = const()[name = tensor<string, []>("x155_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s5_l0_val2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s5_l0_val2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [32, 1, 48]> cache_s5_l0_val2_to_fp16 = cast(dtype = cache_s5_l0_val2_to_fp16_dtype_0, x = cache_s5_l0_val2)[name = tensor<string, []>("cast_1")];
            tensor<fp16, [40, 1, 48]> x155_1_cast_fp16 = concat(axis = var_46, interleave = x155_1_interleave_0, values = (cache_s5_l0_val2_to_fp16, linear_103_cast_fp16))[name = tensor<string, []>("x155_1_cast_fp16")];
            tensor<int32, [3]> var_4694_begin_0 = const()[name = tensor<string, []>("op_4694_begin_0"), val = tensor<int32, [3]>([8, 0, 0])];
            tensor<int32, [3]> var_4694_end_0 = const()[name = tensor<string, []>("op_4694_end_0"), val = tensor<int32, [3]>([40, 1, 48])];
            tensor<bool, [3]> var_4694_end_mask_0 = const()[name = tensor<string, []>("op_4694_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [32, 1, 48]> new_cache_s5_l0_val2 = slice_by_index(begin = var_4694_begin_0, end = var_4694_end_0, end_mask = var_4694_end_mask_0, x = x155_1_cast_fp16)[name = tensor<string, []>("op_4694_cast_fp16")];
            tensor<int32, [4]> var_4695 = const()[name = tensor<string, []>("op_4695"), val = tensor<int32, [4]>([40, 1, 4, -1])];
            tensor<fp16, [40, 1, 4, 12]> var_4696_cast_fp16 = reshape(shape = var_4695, x = x155_1_cast_fp16)[name = tensor<string, []>("op_4696_cast_fp16")];
            tensor<int32, [4]> var_4697 = const()[name = tensor<string, []>("op_4697"), val = tensor<int32, [4]>([2, 1, -4, -1])];
            tensor<bool, []> x157_1_transpose_x_0 = const()[name = tensor<string, []>("x157_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> x157_1_transpose_y_0 = const()[name = tensor<string, []>("x157_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [4, 1, 40, 12]> x156_1_cast_fp16 = transpose(perm = var_4697, x = var_4696_cast_fp16)[name = tensor<string, []>("transpose_40")];
            tensor<fp16, [4, 1, 8, 12]> x157_1_cast_fp16 = matmul(transpose_x = x157_1_transpose_x_0, transpose_y = x157_1_transpose_y_0, x = attn_weights9_1_cast_fp16, y = x156_1_cast_fp16)[name = tensor<string, []>("x157_1_cast_fp16")];
            tensor<int32, [4]> var_4702 = const()[name = tensor<string, []>("op_4702"), val = tensor<int32, [4]>([2, 1, 0, 3])];
            tensor<int32, [3]> var_4707 = const()[name = tensor<string, []>("op_4707"), val = tensor<int32, [3]>([8, 1, 48])];
            tensor<fp16, [8, 1, 4, 12]> var_4703_cast_fp16 = transpose(perm = var_4702, x = x157_1_cast_fp16)[name = tensor<string, []>("transpose_39")];
            tensor<fp16, [8, 1, 48]> input75_1_cast_fp16 = reshape(shape = var_4707, x = var_4703_cast_fp16)[name = tensor<string, []>("input75_1_cast_fp16")];
            tensor<fp16, [128, 48]> encoder_encoders_5_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [6144]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2753856))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2760064))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 48])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_self_attn2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_self_attn2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2760640)))];
            tensor<fp16, [8, 1, 128]> linear_104_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_self_attn2_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_self_attn2_out_proj_weight_to_fp16_palettized, x = input75_1_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input76_1_cast_fp16 = add(x = var_4675_cast_fp16, y = linear_104_cast_fp16)[name = tensor<string, []>("input76_1_cast_fp16")];
            tensor<fp16, [256, 128]> encoder_encoders_5_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [32768]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2760960))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2793792))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([256, 128])];
            tensor<fp16, [256]> encoder_encoders_5_encoder_layers_0_conv_module2_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_in_proj_bias_to_fp16"), val = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2794368)))];
            tensor<fp16, [8, 1, 256]> linear_105_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_conv_module2_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_conv_module2_in_proj_weight_to_fp16_palettized, x = input76_1_cast_fp16)[name = tensor<string, []>("linear_105_cast_fp16")];
            tensor<int32, [2]> var_4718_split_sizes_0 = const()[name = tensor<string, []>("op_4718_split_sizes_0"), val = tensor<int32, [2]>([128, 128])];
            tensor<int32, []> var_4718_axis_0 = const()[name = tensor<string, []>("op_4718_axis_0"), val = tensor<int32, []>(2)];
            tensor<fp16, [8, 1, 128]> var_4718_cast_fp16_0, tensor<fp16, [8, 1, 128]> var_4718_cast_fp16_1 = split(axis = var_4718_axis_0, split_sizes = var_4718_split_sizes_0, x = linear_105_cast_fp16)[name = tensor<string, []>("op_4718_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4720_cast_fp16 = sigmoid(x = var_4718_cast_fp16_1)[name = tensor<string, []>("op_4720_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x159_1_cast_fp16 = mul(x = var_4718_cast_fp16_0, y = var_4720_cast_fp16)[name = tensor<string, []>("x159_1_cast_fp16")];
            tensor<int32, [3]> var_4722 = const()[name = tensor<string, []>("op_4722"), val = tensor<int32, [3]>([1, 2, 0])];
            tensor<fp16, [1, 128, 8]> x160_1_cast_fp16 = transpose(perm = var_4722, x = x159_1_cast_fp16)[name = tensor<string, []>("transpose_38")];
            tensor<fp16, [1, 128, 8]> x161_1_cast_fp16 = select(a = var_62_to_fp16, b = x160_1_cast_fp16, cond = var_2289)[name = tensor<string, []>("x161_1_cast_fp16")];
            tensor<bool, []> x162_1_interleave_0 = const()[name = tensor<string, []>("x162_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<string, []> cache_s5_l0_conv2_to_fp16_dtype_0 = const()[name = tensor<string, []>("cache_s5_l0_conv2_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 128, 15]> cache_s5_l0_conv2_to_fp16 = cast(dtype = cache_s5_l0_conv2_to_fp16_dtype_0, x = cache_s5_l0_conv2)[name = tensor<string, []>("cast_0")];
            tensor<fp16, [1, 128, 23]> x162_1_cast_fp16 = concat(axis = var_49, interleave = x162_1_interleave_0, values = (cache_s5_l0_conv2_to_fp16, x161_1_cast_fp16))[name = tensor<string, []>("x162_1_cast_fp16")];
            tensor<int32, [3]> var_4732_begin_0 = const()[name = tensor<string, []>("op_4732_begin_0"), val = tensor<int32, [3]>([0, 0, 8])];
            tensor<int32, [3]> var_4732_end_0 = const()[name = tensor<string, []>("op_4732_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> var_4732_end_mask_0 = const()[name = tensor<string, []>("op_4732_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 15]> new_cache_s5_l0_conv2 = slice_by_index(begin = var_4732_begin_0, end = var_4732_end_0, end_mask = var_4732_end_mask_0, x = x162_1_cast_fp16)[name = tensor<string, []>("op_4732_cast_fp16")];
            tensor<string, []> x_causal_1_pad_type_0 = const()[name = tensor<string, []>("x_causal_1_pad_type_0"), val = tensor<string, []>("valid")];
            tensor<int32, []> x_causal_1_groups_0 = const()[name = tensor<string, []>("x_causal_1_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_causal_1_strides_0 = const()[name = tensor<string, []>("x_causal_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [2]> x_causal_1_pad_0 = const()[name = tensor<string, []>("x_causal_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
            tensor<int32, [1]> x_causal_1_dilations_0 = const()[name = tensor<string, []>("x_causal_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 16]> encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16"), val = tensor<fp16, [128, 1, 16]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2794944)))];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2799104)))];
            tensor<fp16, [1, 128, 8]> x_causal_1_cast_fp16 = conv(bias = encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_bias_to_fp16, dilations = x_causal_1_dilations_0, groups = x_causal_1_groups_0, pad = x_causal_1_pad_0, pad_type = x_causal_1_pad_type_0, strides = x_causal_1_strides_0, weight = encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_causal_conv_weight_to_fp16, x = x162_1_cast_fp16)[name = tensor<string, []>("x_causal_1_cast_fp16")];
            tensor<int32, [3]> input78_1_begin_0 = const()[name = tensor<string, []>("input78_1_begin_0"), val = tensor<int32, [3]>([0, 0, 15])];
            tensor<int32, [3]> input78_1_end_0 = const()[name = tensor<string, []>("input78_1_end_0"), val = tensor<int32, [3]>([1, 128, 23])];
            tensor<bool, [3]> input78_1_end_mask_0 = const()[name = tensor<string, []>("input78_1_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
            tensor<fp16, [1, 128, 8]> input78_1_cast_fp16 = slice_by_index(begin = input78_1_begin_0, end = input78_1_end_0, end_mask = input78_1_end_mask_0, x = x162_1_cast_fp16)[name = tensor<string, []>("input78_1_cast_fp16")];
            tensor<string, []> x_chunk_1_pad_type_0 = const()[name = tensor<string, []>("x_chunk_1_pad_type_0"), val = tensor<string, []>("custom")];
            tensor<int32, [2]> x_chunk_1_pad_0 = const()[name = tensor<string, []>("x_chunk_1_pad_0"), val = tensor<int32, [2]>([15, 15])];
            tensor<int32, []> x_chunk_1_groups_0 = const()[name = tensor<string, []>("x_chunk_1_groups_0"), val = tensor<int32, []>(128)];
            tensor<int32, [1]> x_chunk_1_strides_0 = const()[name = tensor<string, []>("x_chunk_1_strides_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1]> x_chunk_1_dilations_0 = const()[name = tensor<string, []>("x_chunk_1_dilations_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [128, 1, 31]> encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [3968]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2799424))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2803456))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized"), shape = tensor<uint32, [3]>([128, 1, 31])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2804032)))];
            tensor<fp16, [1, 128, 8]> x_chunk_1_cast_fp16 = conv(bias = encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_bias_to_fp16, dilations = x_chunk_1_dilations_0, groups = x_chunk_1_groups_0, pad = x_chunk_1_pad_0, pad_type = x_chunk_1_pad_type_0, strides = x_chunk_1_strides_0, weight = encoder_encoders_5_encoder_layers_0_conv_module2_depthwise_conv_chunkwise_conv_weight_to_fp16_palettized, x = input78_1_cast_fp16)[name = tensor<string, []>("x_chunk_1_cast_fp16")];
            tensor<fp16, [128, 8]> chunk_scale10_1_to_fp16 = const()[name = tensor<string, []>("chunk_scale10_1_to_fp16"), val = tensor<fp16, [128, 8]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2804352)))];
            tensor<fp16, [1, 128, 8]> x_chunk10_1_cast_fp16 = mul(x = x_chunk_1_cast_fp16, y = chunk_scale10_1_to_fp16)[name = tensor<string, []>("x_chunk10_1_cast_fp16")];
            tensor<fp16, [1, 128, 8]> x163_1_cast_fp16 = add(x = x_chunk10_1_cast_fp16, y = x_causal_1_cast_fp16)[name = tensor<string, []>("x163_1_cast_fp16")];
            tensor<int32, [3]> var_4772 = const()[name = tensor<string, []>("op_4772"), val = tensor<int32, [3]>([2, 0, 1])];
            tensor<fp16, []> var_4781_to_fp16 = const()[name = tensor<string, []>("op_4781_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> x164_1_cast_fp16 = transpose(perm = var_4772, x = x163_1_cast_fp16)[name = tensor<string, []>("transpose_37")];
            tensor<fp16, [8, 1, 128]> x_offset_58_cast_fp16 = sub(x = x164_1_cast_fp16, y = var_4781_to_fp16)[name = tensor<string, []>("x_offset_58_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4783_cast_fp16 = exp(x = x_offset_58_cast_fp16)[name = tensor<string, []>("op_4783_cast_fp16")];
            tensor<fp16, []> var_4784_to_fp16 = const()[name = tensor<string, []>("op_4784_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 128]> var_4785_cast_fp16 = add(x = var_4783_cast_fp16, y = var_4784_to_fp16)[name = tensor<string, []>("op_4785_cast_fp16")];
            tensor<fp32, []> var_4786_epsilon_0 = const()[name = tensor<string, []>("op_4786_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 128]> var_4786_cast_fp16 = log(epsilon = var_4786_epsilon_0, x = var_4785_cast_fp16)[name = tensor<string, []>("op_4786_cast_fp16")];
            tensor<fp16, []> var_4778_to_fp16 = const()[name = tensor<string, []>("op_4778_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 128]> var_4788_cast_fp16 = equal(x = var_4786_cast_fp16, y = var_4778_to_fp16)[name = tensor<string, []>("op_4788_cast_fp16")];
            tensor<fp16, [8, 1, 128]> log_sum0_58_cast_fp16 = select(a = x_offset_58_cast_fp16, b = var_4786_cast_fp16, cond = var_4788_cast_fp16)[name = tensor<string, []>("log_sum0_58_cast_fp16")];
            tensor<fp16, []> var_4790_to_fp16 = const()[name = tensor<string, []>("op_4790_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 128]> var_4791_cast_fp16 = mul(x = x164_1_cast_fp16, y = var_4790_to_fp16)[name = tensor<string, []>("op_4791_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4792_cast_fp16 = sub(x = log_sum0_58_cast_fp16, y = var_4791_cast_fp16)[name = tensor<string, []>("op_4792_cast_fp16")];
            tensor<fp16, []> var_4793_to_fp16 = const()[name = tensor<string, []>("op_4793_to_fp16"), val = tensor<fp16, []>(0x1.40cp-2)];
            tensor<fp16, [8, 1, 128]> x0_1_cast_fp16 = sub(x = var_4792_cast_fp16, y = var_4793_to_fp16)[name = tensor<string, []>("x0_1_cast_fp16")];
            tensor<fp16, [128, 128]> encoder_encoders_5_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [16384]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2806464))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2822912))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 128])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_conv_module2_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_conv_module2_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2823488)))];
            tensor<fp16, [8, 1, 128]> linear_106_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_conv_module2_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_conv_module2_out_proj_weight_to_fp16_palettized, x = x0_1_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
            tensor<fp16, [8, 1, 128]> input79_1_cast_fp16 = add(x = input76_1_cast_fp16, y = linear_106_cast_fp16)[name = tensor<string, []>("input79_1_cast_fp16")];
            tensor<fp16, [240, 128]> encoder_encoders_5_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2823808))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2854592))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([240, 128])];
            tensor<fp16, [240]> encoder_encoders_5_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16"), val = tensor<fp16, [240]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2855168)))];
            tensor<fp16, [8, 1, 240]> linear_107_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_feed_forward3_in_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_feed_forward3_in_proj_weight_to_fp16_palettized, x = input79_1_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
            tensor<fp16, []> var_4809_to_fp16 = const()[name = tensor<string, []>("op_4809_to_fp16"), val = tensor<fp16, []>(0x1p+2)];
            tensor<fp16, [8, 1, 240]> x_offset_1_cast_fp16 = sub(x = linear_107_cast_fp16, y = var_4809_to_fp16)[name = tensor<string, []>("x_offset_1_cast_fp16")];
            tensor<fp16, [8, 1, 240]> var_4811_cast_fp16 = exp(x = x_offset_1_cast_fp16)[name = tensor<string, []>("op_4811_cast_fp16")];
            tensor<fp16, []> var_4812_to_fp16 = const()[name = tensor<string, []>("op_4812_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<fp16, [8, 1, 240]> var_4813_cast_fp16 = add(x = var_4811_cast_fp16, y = var_4812_to_fp16)[name = tensor<string, []>("op_4813_cast_fp16")];
            tensor<fp32, []> var_4814_epsilon_0 = const()[name = tensor<string, []>("op_4814_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
            tensor<fp16, [8, 1, 240]> var_4814_cast_fp16 = log(epsilon = var_4814_epsilon_0, x = var_4813_cast_fp16)[name = tensor<string, []>("op_4814_cast_fp16")];
            tensor<fp16, []> var_4797_to_fp16 = const()[name = tensor<string, []>("op_4797_to_fp16"), val = tensor<fp16, []>(inf)];
            tensor<bool, [8, 1, 240]> var_4816_cast_fp16 = equal(x = var_4814_cast_fp16, y = var_4797_to_fp16)[name = tensor<string, []>("op_4816_cast_fp16")];
            tensor<fp16, [8, 1, 240]> log_sum0_1_cast_fp16 = select(a = x_offset_1_cast_fp16, b = var_4814_cast_fp16, cond = var_4816_cast_fp16)[name = tensor<string, []>("log_sum0_1_cast_fp16")];
            tensor<fp16, []> var_4818_to_fp16 = const()[name = tensor<string, []>("op_4818_to_fp16"), val = tensor<fp16, []>(0x1.47cp-4)];
            tensor<fp16, [8, 1, 240]> var_4819_cast_fp16 = mul(x = linear_107_cast_fp16, y = var_4818_to_fp16)[name = tensor<string, []>("op_4819_cast_fp16")];
            tensor<fp16, [8, 1, 240]> var_4820_cast_fp16 = sub(x = log_sum0_1_cast_fp16, y = var_4819_cast_fp16)[name = tensor<string, []>("op_4820_cast_fp16")];
            tensor<fp16, []> var_4821_to_fp16 = const()[name = tensor<string, []>("op_4821_to_fp16"), val = tensor<fp16, []>(0x1.1ecp-5)];
            tensor<fp16, [8, 1, 240]> x_1_cast_fp16 = sub(x = var_4820_cast_fp16, y = var_4821_to_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<fp16, [128, 240]> encoder_encoders_5_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [30720]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2855744))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2886528))), name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 240])];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2887104)))];
            tensor<fp16, [8, 1, 128]> linear_108_cast_fp16 = linear(bias = encoder_encoders_5_encoder_layers_0_feed_forward3_out_proj_bias_to_fp16, weight = encoder_encoders_5_encoder_layers_0_feed_forward3_out_proj_weight_to_fp16_palettized, x = x_1_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
            tensor<fp16, [8, 1, 128]> x165_1_cast_fp16 = add(x = input79_1_cast_fp16, y = linear_108_cast_fp16)[name = tensor<string, []>("x165_1_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_norm_bias_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_norm_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2887424)))];
            tensor<fp16, [8, 1, 128]> var_4832_cast_fp16 = sub(x = x165_1_cast_fp16, y = encoder_encoders_5_encoder_layers_0_norm_bias_to_fp16)[name = tensor<string, []>("op_4832_cast_fp16")];
            tensor<fp16, []> var_4826_promoted_to_fp16 = const()[name = tensor<string, []>("op_4826_promoted_to_fp16"), val = tensor<fp16, []>(0x1p+1)];
            tensor<fp16, [8, 1, 128]> var_4833_cast_fp16 = pow(x = var_4832_cast_fp16, y = var_4826_promoted_to_fp16)[name = tensor<string, []>("op_4833_cast_fp16")];
            tensor<int32, [1]> var_4835_axes_0 = const()[name = tensor<string, []>("op_4835_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<bool, []> var_4835_keep_dims_0 = const()[name = tensor<string, []>("op_4835_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp16, [8, 1, 1]> var_4835_cast_fp16 = reduce_mean(axes = var_4835_axes_0, keep_dims = var_4835_keep_dims_0, x = var_4833_cast_fp16)[name = tensor<string, []>("op_4835_cast_fp16")];
            tensor<fp16, []> var_4825_to_fp16 = const()[name = tensor<string, []>("op_4825_to_fp16"), val = tensor<fp16, []>(-0x1p-1)];
            tensor<fp16, [8, 1, 1]> var_4836_cast_fp16 = pow(x = var_4835_cast_fp16, y = var_4825_to_fp16)[name = tensor<string, []>("op_4836_cast_fp16")];
            tensor<fp16, []> var_4837_to_fp16 = const()[name = tensor<string, []>("op_4837_to_fp16"), val = tensor<fp16, []>(0x1.b8p-1)];
            tensor<fp16, [8, 1, 1]> scales_1_cast_fp16 = mul(x = var_4836_cast_fp16, y = var_4837_to_fp16)[name = tensor<string, []>("scales_1_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4839_cast_fp16 = mul(x = x165_1_cast_fp16, y = scales_1_cast_fp16)[name = tensor<string, []>("op_4839_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4842_cast_fp16 = sub(x = var_4839_cast_fp16, y = x_79_cast_fp16)[name = tensor<string, []>("op_4842_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_5_encoder_layers_0_bypass_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_encoder_layers_0_bypass_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2887744)))];
            tensor<fp16, [8, 1, 128]> var_4843_cast_fp16 = mul(x = var_4842_cast_fp16, y = encoder_encoders_5_encoder_layers_0_bypass_bypass_scale_to_fp16)[name = tensor<string, []>("op_4843_cast_fp16")];
            tensor<fp16, [8, 1, 128]> var_4844_cast_fp16 = add(x = x_79_cast_fp16, y = var_4843_cast_fp16)[name = tensor<string, []>("op_4844_cast_fp16")];
            tensor<int32, [1]> var_4853_axes_0 = const()[name = tensor<string, []>("op_4853_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [8, 1, 1, 128]> var_4853_cast_fp16 = expand_dims(axes = var_4853_axes_0, x = var_4844_cast_fp16)[name = tensor<string, []>("op_4853_cast_fp16")];
            tensor<int32, [4]> src_1_reps_0 = const()[name = tensor<string, []>("src_1_reps_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [8, 2, 1, 128]> src_1_cast_fp16 = tile(reps = src_1_reps_0, x = var_4853_cast_fp16)[name = tensor<string, []>("src_1_cast_fp16")];
            tensor<int32, [3]> var_4859 = const()[name = tensor<string, []>("op_4859"), val = tensor<int32, [3]>([16, 1, 128])];
            tensor<fp16, [16, 1, 128]> var_4860_cast_fp16 = reshape(shape = var_4859, x = src_1_cast_fp16)[name = tensor<string, []>("op_4860_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_4867_cast_fp16 = sub(x = var_4860_cast_fp16, y = var_4300_cast_fp16)[name = tensor<string, []>("op_4867_cast_fp16")];
            tensor<fp16, [128]> encoder_encoders_5_out_combiner_bypass_scale_to_fp16 = const()[name = tensor<string, []>("encoder_encoders_5_out_combiner_bypass_scale_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2888064)))];
            tensor<fp16, [16, 1, 128]> var_4868_cast_fp16 = mul(x = var_4867_cast_fp16, y = encoder_encoders_5_out_combiner_bypass_scale_to_fp16)[name = tensor<string, []>("op_4868_cast_fp16")];
            tensor<fp16, [16, 1, 128]> var_4869_cast_fp16 = add(x = var_4300_cast_fp16, y = var_4868_cast_fp16)[name = tensor<string, []>("op_4869_cast_fp16")];
            tensor<bool, []> src15_1_interleave_0 = const()[name = tensor<string, []>("src15_1_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [16, 1, 128]> src15_1_cast_fp16 = concat(axis = var_59, interleave = src15_1_interleave_0, values = var_4869_cast_fp16)[name = tensor<string, []>("src15_1_cast_fp16")];
            tensor<int32, [4]> var_4890 = const()[name = tensor<string, []>("op_4890"), val = tensor<int32, [4]>([8, 2, 1, 128])];
            tensor<fp16, [8, 2, 1, 128]> src0_1_cast_fp16 = reshape(shape = var_4890, x = src15_1_cast_fp16)[name = tensor<string, []>("src0_1_cast_fp16")];
            tensor<fp16, [2, 1, 1]> weights0_1_to_fp16 = const()[name = tensor<string, []>("weights0_1_to_fp16"), val = tensor<fp16, [2, 1, 1]>([[[0x1.c44p-9]], [[0x1.fe4p-1]]])];
            tensor<fp16, [8, 2, 1, 128]> var_4895_cast_fp16 = mul(x = src0_1_cast_fp16, y = weights0_1_to_fp16)[name = tensor<string, []>("op_4895_cast_fp16")];
            tensor<int32, [1]> encoder_out_1_axes_0 = const()[name = tensor<string, []>("encoder_out_1_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, []> encoder_out_1_keep_dims_0 = const()[name = tensor<string, []>("encoder_out_1_keep_dims_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [8, 1, 128]> encoder_out_1_cast_fp16 = reduce_sum(axes = encoder_out_1_axes_0, keep_dims = encoder_out_1_keep_dims_0, x = var_4895_cast_fp16)[name = tensor<string, []>("encoder_out_1_cast_fp16")];
            tensor<int32, [3]> var_4898 = const()[name = tensor<string, []>("op_4898"), val = tensor<int32, [3]>([1, 0, 2])];
            tensor<fp16, [320, 128]> encoder_proj_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [40960]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2888384))), lut = tensor<fp16, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2929408))), name = tensor<string, []>("encoder_proj_weight_to_fp16_palettized"), shape = tensor<uint32, [2]>([320, 128])];
            tensor<fp16, [320]> encoder_proj_bias_to_fp16 = const()[name = tensor<string, []>("encoder_proj_bias_to_fp16"), val = tensor<fp16, [320]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2929984)))];
            tensor<fp16, [1, 8, 128]> input80_1_cast_fp16 = transpose(perm = var_4898, x = encoder_out_1_cast_fp16)[name = tensor<string, []>("transpose_36")];
            tensor<fp16, [1, 8, 320]> encoder_out = linear(bias = encoder_proj_bias_to_fp16, weight = encoder_proj_weight_to_fp16_palettized, x = input80_1_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
        } -> (encoder_out, new_cache_s0_l0_key, new_cache_s0_l0_nonlin_attn, new_cache_s0_l0_val1, new_cache_s0_l0_val2, new_cache_s0_l0_conv1, new_cache_s0_l0_conv2, new_cache_s1_l0_key, new_cache_s1_l0_nonlin_attn, new_cache_s1_l0_val1, new_cache_s1_l0_val2, new_cache_s1_l0_conv1, new_cache_s1_l0_conv2, new_cache_s2_l0_key, new_cache_s2_l0_nonlin_attn, new_cache_s2_l0_val1, new_cache_s2_l0_val2, new_cache_s2_l0_conv1, new_cache_s2_l0_conv2, new_cache_s3_l0_key, new_cache_s3_l0_nonlin_attn, new_cache_s3_l0_val1, new_cache_s3_l0_val2, new_cache_s3_l0_conv1, new_cache_s3_l0_conv2, new_cache_s4_l0_key, new_cache_s4_l0_nonlin_attn, new_cache_s4_l0_val1, new_cache_s4_l0_val2, new_cache_s4_l0_conv1, new_cache_s4_l0_conv2, new_cache_s5_l0_key, new_cache_s5_l0_nonlin_attn, new_cache_s5_l0_val1, new_cache_s5_l0_val2, new_cache_s5_l0_conv1, new_cache_s5_l0_conv2, new_cached_embed_left_pad, new_processed_lens);
}