program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3505.3.2"}, {"coremlc-version", "3505.4.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<fp32, [32]> bos_emb, tensor<fp32, [2, 1, 512, 16, 64]> cache0, tensor<fp32, [2, 1, 512, 16, 64]> cache1, tensor<fp32, [2, 1, 512, 16, 64]> cache2, tensor<fp32, [2, 1, 512, 16, 64]> cache3, tensor<fp32, [2, 1, 512, 16, 64]> cache4, tensor<fp32, [2, 1, 512, 16, 64]> cache5, tensor<fp32, [1]> position0, tensor<fp32, [1]> position1, tensor<fp32, [1]> position2, tensor<fp32, [1]> position3, tensor<fp32, [1]> position4, tensor<fp32, [1]> position5, tensor<fp32, [1, 1, 32]> sequence) {
            tensor<fp32, [1024, 32]> input_linear_weight = const()[name = tensor<string, []>("input_linear_weight"), val = tensor<fp32, [1024, 32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp32, [1024]> norm0_1_bias = const()[name = tensor<string, []>("norm0_1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(131200)))];
            tensor<fp32, [1024]> norm0_1_weight = const()[name = tensor<string, []>("norm0_1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135360)))];
            tensor<fp32, [3072, 1024]> attn0_in_proj_weight = const()[name = tensor<string, []>("attn0_in_proj_weight"), val = tensor<fp32, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(139520)))];
            tensor<fp32, [1024, 1024]> attn0_out_proj_weight = const()[name = tensor<string, []>("attn0_out_proj_weight"), val = tensor<fp32, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(12722496)))];
            tensor<fp32, [1024]> norm0_2_bias = const()[name = tensor<string, []>("norm0_2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16916864)))];
            tensor<fp32, [1024]> norm0_2_weight = const()[name = tensor<string, []>("norm0_2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16921024)))];
            tensor<fp32, [4096, 1024]> linear0_1_weight = const()[name = tensor<string, []>("linear0_1_weight"), val = tensor<fp32, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16925184)))];
            tensor<fp32, [1024, 4096]> linear0_2_weight = const()[name = tensor<string, []>("linear0_2_weight"), val = tensor<fp32, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33702464)))];
            tensor<fp32, [1024]> norm1_1_bias = const()[name = tensor<string, []>("norm1_1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50479744)))];
            tensor<fp32, [1024]> norm1_1_weight = const()[name = tensor<string, []>("norm1_1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50483904)))];
            tensor<fp32, [3072, 1024]> attn1_in_proj_weight = const()[name = tensor<string, []>("attn1_in_proj_weight"), val = tensor<fp32, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(50488064)))];
            tensor<fp32, [1024, 1024]> attn1_out_proj_weight = const()[name = tensor<string, []>("attn1_out_proj_weight"), val = tensor<fp32, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63071040)))];
            tensor<fp32, [1024]> norm1_2_bias = const()[name = tensor<string, []>("norm1_2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67265408)))];
            tensor<fp32, [1024]> norm1_2_weight = const()[name = tensor<string, []>("norm1_2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67269568)))];
            tensor<fp32, [4096, 1024]> linear1_1_weight = const()[name = tensor<string, []>("linear1_1_weight"), val = tensor<fp32, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67273728)))];
            tensor<fp32, [1024, 4096]> linear1_2_weight = const()[name = tensor<string, []>("linear1_2_weight"), val = tensor<fp32, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(84051008)))];
            tensor<fp32, [1024]> norm2_1_bias = const()[name = tensor<string, []>("norm2_1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100828288)))];
            tensor<fp32, [1024]> norm2_1_weight = const()[name = tensor<string, []>("norm2_1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100832448)))];
            tensor<fp32, [3072, 1024]> attn2_in_proj_weight = const()[name = tensor<string, []>("attn2_in_proj_weight"), val = tensor<fp32, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(100836608)))];
            tensor<fp32, [1024, 1024]> attn2_out_proj_weight = const()[name = tensor<string, []>("attn2_out_proj_weight"), val = tensor<fp32, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113419584)))];
            tensor<fp32, [1024]> norm2_2_bias = const()[name = tensor<string, []>("norm2_2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117613952)))];
            tensor<fp32, [1024]> norm2_2_weight = const()[name = tensor<string, []>("norm2_2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117618112)))];
            tensor<fp32, [4096, 1024]> linear2_1_weight = const()[name = tensor<string, []>("linear2_1_weight"), val = tensor<fp32, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117622272)))];
            tensor<fp32, [1024, 4096]> linear2_2_weight = const()[name = tensor<string, []>("linear2_2_weight"), val = tensor<fp32, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134399552)))];
            tensor<fp32, [1024]> norm3_1_bias = const()[name = tensor<string, []>("norm3_1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151176832)))];
            tensor<fp32, [1024]> norm3_1_weight = const()[name = tensor<string, []>("norm3_1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151180992)))];
            tensor<fp32, [3072, 1024]> attn3_in_proj_weight = const()[name = tensor<string, []>("attn3_in_proj_weight"), val = tensor<fp32, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(151185152)))];
            tensor<fp32, [1024, 1024]> attn3_out_proj_weight = const()[name = tensor<string, []>("attn3_out_proj_weight"), val = tensor<fp32, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(163768128)))];
            tensor<fp32, [1024]> norm3_2_bias = const()[name = tensor<string, []>("norm3_2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167962496)))];
            tensor<fp32, [1024]> norm3_2_weight = const()[name = tensor<string, []>("norm3_2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167966656)))];
            tensor<fp32, [4096, 1024]> linear3_1_weight = const()[name = tensor<string, []>("linear3_1_weight"), val = tensor<fp32, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(167970816)))];
            tensor<fp32, [1024, 4096]> linear3_2_weight = const()[name = tensor<string, []>("linear3_2_weight"), val = tensor<fp32, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184748096)))];
            tensor<fp32, [1024]> norm4_1_bias = const()[name = tensor<string, []>("norm4_1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201525376)))];
            tensor<fp32, [1024]> norm4_1_weight = const()[name = tensor<string, []>("norm4_1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201529536)))];
            tensor<fp32, [3072, 1024]> attn4_in_proj_weight = const()[name = tensor<string, []>("attn4_in_proj_weight"), val = tensor<fp32, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(201533696)))];
            tensor<fp32, [1024, 1024]> attn4_out_proj_weight = const()[name = tensor<string, []>("attn4_out_proj_weight"), val = tensor<fp32, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214116672)))];
            tensor<fp32, [1024]> norm4_2_bias = const()[name = tensor<string, []>("norm4_2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218311040)))];
            tensor<fp32, [1024]> norm4_2_weight = const()[name = tensor<string, []>("norm4_2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218315200)))];
            tensor<fp32, [4096, 1024]> linear4_1_weight = const()[name = tensor<string, []>("linear4_1_weight"), val = tensor<fp32, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218319360)))];
            tensor<fp32, [1024, 4096]> linear4_2_weight = const()[name = tensor<string, []>("linear4_2_weight"), val = tensor<fp32, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235096640)))];
            tensor<fp32, [1024]> norm5_1_bias = const()[name = tensor<string, []>("norm5_1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251873920)))];
            tensor<fp32, [1024]> norm5_1_weight = const()[name = tensor<string, []>("norm5_1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251878080)))];
            tensor<fp32, [3072, 1024]> attn5_in_proj_weight = const()[name = tensor<string, []>("attn5_in_proj_weight"), val = tensor<fp32, [3072, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251882240)))];
            tensor<fp32, [1024, 1024]> attn5_out_proj_weight = const()[name = tensor<string, []>("attn5_out_proj_weight"), val = tensor<fp32, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(264465216)))];
            tensor<fp32, [1024]> norm5_2_bias = const()[name = tensor<string, []>("norm5_2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268659584)))];
            tensor<fp32, [1024]> norm5_2_weight = const()[name = tensor<string, []>("norm5_2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268663744)))];
            tensor<fp32, [4096, 1024]> linear5_1_weight = const()[name = tensor<string, []>("linear5_1_weight"), val = tensor<fp32, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268667904)))];
            tensor<fp32, [1024, 4096]> linear5_2_weight = const()[name = tensor<string, []>("linear5_2_weight"), val = tensor<fp32, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(285445184)))];
            tensor<fp32, [1024]> out_norm_bias = const()[name = tensor<string, []>("out_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302222464)))];
            tensor<fp32, [1024]> out_norm_weight = const()[name = tensor<string, []>("out_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302226624)))];
            tensor<fp32, [1]> out_eos_bias = const()[name = tensor<string, []>("out_eos_bias"), val = tensor<fp32, [1]>([-0x1.42p-2])];
            tensor<fp32, [1, 1024]> out_eos_weight = const()[name = tensor<string, []>("out_eos_weight"), val = tensor<fp32, [1, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302230784)))];
            tensor<bool, [1, 1, 32]> var_54 = not_equal(x = sequence, y = sequence)[name = tensor<string, []>("op_54")];
            tensor<int32, [2]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [2]>([0, 1])];
            tensor<fp32, [1, 1, 32]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb)[name = tensor<string, []>("expand_dims_0")];
            tensor<fp32, [1, 1, 32]> input_1 = select(a = expand_dims_0, b = sequence, cond = var_54)[name = tensor<string, []>("input_1")];
            tensor<fp32, [1024]> linear_0_bias_0 = const()[name = tensor<string, []>("linear_0_bias_0"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302234944)))];
            tensor<fp32, [1, 1, 1024]> input_3 = linear(bias = linear_0_bias_0, weight = input_linear_weight, x = input_1)[name = tensor<string, []>("linear_0")];
            tensor<fp32, []> var_60 = const()[name = tensor<string, []>("op_60"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> x_1_axes_0 = const()[name = tensor<string, []>("x_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> x_1 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias, epsilon = var_60, gamma = norm0_1_weight, x = input_3)[name = tensor<string, []>("x_1")];
            tensor<fp32, [3072]> linear_1_bias_0 = const()[name = tensor<string, []>("linear_1_bias_0"), val = tensor<fp32, [3072]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302239104)))];
            tensor<fp32, [1, 1, 3072]> var_92 = linear(bias = linear_1_bias_0, weight = attn0_in_proj_weight, x = x_1)[name = tensor<string, []>("linear_1")];
            tensor<int32, [5]> var_96 = const()[name = tensor<string, []>("op_96"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<fp32, [1, 1, 3, 16, 64]> qkv_1 = reshape(shape = var_96, x = var_92)[name = tensor<string, []>("qkv_1")];
            tensor<int32, [5]> q_1_begin_0 = const()[name = tensor<string, []>("q_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_1_end_0 = const()[name = tensor<string, []>("q_1_end_0"), val = tensor<int32, [5]>([1, 1, 1, 16, 64])];
            tensor<bool, [5]> q_1_end_mask_0 = const()[name = tensor<string, []>("q_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_1_squeeze_mask_0 = const()[name = tensor<string, []>("q_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> q_1 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1)[name = tensor<string, []>("q_1")];
            tensor<int32, [5]> k_1_begin_0 = const()[name = tensor<string, []>("k_1_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_1_end_0 = const()[name = tensor<string, []>("k_1_end_0"), val = tensor<int32, [5]>([1, 1, 2, 16, 64])];
            tensor<bool, [5]> k_1_end_mask_0 = const()[name = tensor<string, []>("k_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_1_squeeze_mask_0 = const()[name = tensor<string, []>("k_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> k_1 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1)[name = tensor<string, []>("k_1")];
            tensor<int32, [5]> v_1_begin_0 = const()[name = tensor<string, []>("v_1_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_1_end_0 = const()[name = tensor<string, []>("v_1_end_0"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<bool, [5]> v_1_end_mask_0 = const()[name = tensor<string, []>("v_1_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_1_squeeze_mask_0 = const()[name = tensor<string, []>("v_1_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> v_1 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1)[name = tensor<string, []>("v_1")];
            tensor<fp32, [32]> freqs_1 = const()[name = tensor<string, []>("freqs_1"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302251456)))];
            tensor<int32, [4]> var_200 = const()[name = tensor<string, []>("op_200"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp32, [1, 1, 1, 1]> ts_5 = reshape(shape = var_200, x = position0)[name = tensor<string, []>("ts_5")];
            tensor<int32, [5]> var_204 = const()[name = tensor<string, []>("op_204"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> q_complex_1 = reshape(shape = var_204, x = q_1)[name = tensor<string, []>("q_complex_1")];
            tensor<int32, [5]> var_208 = const()[name = tensor<string, []>("op_208"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> k_complex_1 = reshape(shape = var_208, x = k_1)[name = tensor<string, []>("k_complex_1")];
            tensor<int32, [5]> var_212_begin_0 = const()[name = tensor<string, []>("op_212_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_212_end_0 = const()[name = tensor<string, []>("op_212_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_212_end_mask_0 = const()[name = tensor<string, []>("op_212_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_212_squeeze_mask_0 = const()[name = tensor<string, []>("op_212_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_212 = slice_by_index(begin = var_212_begin_0, end = var_212_end_0, end_mask = var_212_end_mask_0, squeeze_mask = var_212_squeeze_mask_0, x = q_complex_1)[name = tensor<string, []>("op_212")];
            tensor<int32, [5]> var_220_begin_0 = const()[name = tensor<string, []>("op_220_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_220_end_0 = const()[name = tensor<string, []>("op_220_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_220_end_mask_0 = const()[name = tensor<string, []>("op_220_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_220_squeeze_mask_0 = const()[name = tensor<string, []>("op_220_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_220 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, squeeze_mask = var_220_squeeze_mask_0, x = q_complex_1)[name = tensor<string, []>("op_220")];
            tensor<int32, [5]> var_228_begin_0 = const()[name = tensor<string, []>("op_228_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_228_end_0 = const()[name = tensor<string, []>("op_228_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_228_end_mask_0 = const()[name = tensor<string, []>("op_228_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_228_squeeze_mask_0 = const()[name = tensor<string, []>("op_228_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_228 = slice_by_index(begin = var_228_begin_0, end = var_228_end_0, end_mask = var_228_end_mask_0, squeeze_mask = var_228_squeeze_mask_0, x = k_complex_1)[name = tensor<string, []>("op_228")];
            tensor<int32, [5]> var_236_begin_0 = const()[name = tensor<string, []>("op_236_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_236_end_0 = const()[name = tensor<string, []>("op_236_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_236_end_mask_0 = const()[name = tensor<string, []>("op_236_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_236_squeeze_mask_0 = const()[name = tensor<string, []>("op_236_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_236 = slice_by_index(begin = var_236_begin_0, end = var_236_end_0, end_mask = var_236_end_mask_0, squeeze_mask = var_236_squeeze_mask_0, x = k_complex_1)[name = tensor<string, []>("op_236")];
            tensor<fp32, [1, 1, 1, 32]> var_242 = mul(x = freqs_1, y = ts_5)[name = tensor<string, []>("op_242")];
            tensor<fp32, [1, 1, 1, 32]> rotr_1 = cos(x = var_242)[name = tensor<string, []>("rotr_1")];
            tensor<fp32, [1, 1, 1, 32]> roti_1 = sin(x = var_242)[name = tensor<string, []>("roti_1")];
            tensor<fp32, [1, 1, 16, 32]> var_246 = mul(x = var_212, y = rotr_1)[name = tensor<string, []>("op_246")];
            tensor<fp32, [1, 1, 16, 32]> var_247 = mul(x = var_220, y = roti_1)[name = tensor<string, []>("op_247")];
            tensor<fp32, [1, 1, 16, 32]> qor_1 = sub(x = var_246, y = var_247)[name = tensor<string, []>("qor_1")];
            tensor<fp32, [1, 1, 16, 32]> var_250 = mul(x = var_212, y = roti_1)[name = tensor<string, []>("op_250")];
            tensor<fp32, [1, 1, 16, 32]> var_251 = mul(x = var_220, y = rotr_1)[name = tensor<string, []>("op_251")];
            tensor<fp32, [1, 1, 16, 32]> qoi_1 = add(x = var_250, y = var_251)[name = tensor<string, []>("qoi_1")];
            tensor<fp32, [1, 1, 16, 32]> var_254 = mul(x = var_228, y = rotr_1)[name = tensor<string, []>("op_254")];
            tensor<fp32, [1, 1, 16, 32]> var_255 = mul(x = var_236, y = roti_1)[name = tensor<string, []>("op_255")];
            tensor<fp32, [1, 1, 16, 32]> kor_1 = sub(x = var_254, y = var_255)[name = tensor<string, []>("kor_1")];
            tensor<fp32, [1, 1, 16, 32]> var_258 = mul(x = var_228, y = roti_1)[name = tensor<string, []>("op_258")];
            tensor<fp32, [1, 1, 16, 32]> var_259 = mul(x = var_236, y = rotr_1)[name = tensor<string, []>("op_259")];
            tensor<fp32, [1, 1, 16, 32]> koi_1 = add(x = var_258, y = var_259)[name = tensor<string, []>("koi_1")];
            tensor<int32, []> qo_1_axis_0 = const()[name = tensor<string, []>("qo_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor<string, []>("qo_1")];
            tensor<int32, []> ko_1_axis_0 = const()[name = tensor<string, []>("ko_1_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor<string, []>("ko_1")];
            tensor<int32, [4]> var_288 = const()[name = tensor<string, []>("op_288"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> q_3 = reshape(shape = var_288, x = qo_1)[name = tensor<string, []>("q_3")];
            tensor<int32, [4]> var_290 = const()[name = tensor<string, []>("op_290"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> k_3 = reshape(shape = var_290, x = ko_1)[name = tensor<string, []>("k_3")];
            tensor<fp32, []> _inversed_312_y_0 = const()[name = tensor<string, []>("_inversed_312_y_0"), val = tensor<fp32, []>(0x1p-9)];
            tensor<fp32, [1, 1, 1, 1]> _inversed_312 = mul(x = ts_5, y = _inversed_312_y_0)[name = tensor<string, []>("_inversed_312")];
            tensor<fp32, [1, 1, 1, 1]> var_313 = floor(x = _inversed_312)[name = tensor<string, []>("op_313")];
            tensor<fp32, []> var_314 = const()[name = tensor<string, []>("op_314"), val = tensor<fp32, []>(0x1p+9)];
            tensor<fp32, [1, 1, 1, 1]> var_315 = mul(x = var_313, y = var_314)[name = tensor<string, []>("op_315")];
            tensor<fp32, [1, 1, 1, 1]> write_indices_float_3 = sub(x = ts_5, y = var_315)[name = tensor<string, []>("write_indices_float_3")];
            tensor<string, []> var_322_dtype_0 = const()[name = tensor<string, []>("op_322_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [4]> write_indices_1_reps_0 = const()[name = tensor<string, []>("write_indices_1_reps_0"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<int32, [1, 1, 1, 1]> var_322 = cast(dtype = var_322_dtype_0, x = write_indices_float_3)[name = tensor<string, []>("cast_113")];
            tensor<int32, [1, 1, 16, 64]> write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_322)[name = tensor<string, []>("write_indices_1")];
            tensor<int32, [5]> var_330_begin_0 = const()[name = tensor<string, []>("op_330_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_330_end_0 = const()[name = tensor<string, []>("op_330_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> var_330_end_mask_0 = const()[name = tensor<string, []>("op_330_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_330_squeeze_mask_0 = const()[name = tensor<string, []>("op_330_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_330 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, squeeze_mask = var_330_squeeze_mask_0, x = cache0)[name = tensor<string, []>("op_330")];
            tensor<int32, []> var_332_axis_0 = const()[name = tensor<string, []>("op_332_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_332_mode_0 = const()[name = tensor<string, []>("op_332_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_332_validate_indices_0 = const()[name = tensor<string, []>("op_332_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_332 = scatter_along_axis(axis = var_332_axis_0, data = var_330, indices = write_indices_1, mode = var_332_mode_0, updates = k_3, validate_indices = var_332_validate_indices_0)[name = tensor<string, []>("op_332")];
            tensor<int32, [5]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_3 = const()[name = tensor<string, []>("concat_3"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_12 = const()[name = tensor<string, []>("shape_12"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_0 = const()[name = tensor<string, []>("reduce_prod_0"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_0_start_0 = const()[name = tensor<string, []>("range_1d_0_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_0_step_0 = const()[name = tensor<string, []>("range_1d_0_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor<string, []>("range_1d_0")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor<string, []>("reshape_0")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor<string, []>("slice_by_index_0")];
            tensor<int32, [1]> reshape_1_shape_0 = const()[name = tensor<string, []>("reshape_1_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor<string, []>("reshape_1")];
            tensor<int32, [1]> reshape_2_shape_0 = const()[name = tensor<string, []>("reshape_2_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_2 = reshape(shape = reshape_2_shape_0, x = var_332)[name = tensor<string, []>("reshape_2")];
            tensor<int32, [1]> reshape_3_shape_0 = const()[name = tensor<string, []>("reshape_3_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_3 = reshape(shape = reshape_3_shape_0, x = cache0)[name = tensor<string, []>("reshape_3")];
            tensor<string, []> scatter_0_mode_0 = const()[name = tensor<string, []>("scatter_0_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_0_axis_0 = const()[name = tensor<string, []>("scatter_0_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_0_validate_indices_0 = const()[name = tensor<string, []>("scatter_0_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_0 = scatter(axis = scatter_0_axis_0, data = reshape_3, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2, validate_indices = scatter_0_validate_indices_0)[name = tensor<string, []>("scatter_0")];
            tensor<fp32, [2, 1, 512, 16, 64]> reshape_4 = reshape(shape = shape_12, x = scatter_0)[name = tensor<string, []>("reshape_4")];
            tensor<int32, [5]> var_340_begin_0 = const()[name = tensor<string, []>("op_340_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_340_end_0 = const()[name = tensor<string, []>("op_340_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> var_340_end_mask_0 = const()[name = tensor<string, []>("op_340_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_340_squeeze_mask_0 = const()[name = tensor<string, []>("op_340_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_340 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, squeeze_mask = var_340_squeeze_mask_0, x = reshape_4)[name = tensor<string, []>("op_340")];
            tensor<int32, []> var_342_axis_0 = const()[name = tensor<string, []>("op_342_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_342_mode_0 = const()[name = tensor<string, []>("op_342_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_342_validate_indices_0 = const()[name = tensor<string, []>("op_342_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_342 = scatter_along_axis(axis = var_342_axis_0, data = var_340, indices = write_indices_1, mode = var_342_mode_0, updates = v_1, validate_indices = var_342_validate_indices_0)[name = tensor<string, []>("op_342")];
            tensor<int32, [5]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_13 = const()[name = tensor<string, []>("shape_13"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_1 = const()[name = tensor<string, []>("reduce_prod_1"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_1_start_0 = const()[name = tensor<string, []>("range_1d_1_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_1_step_0 = const()[name = tensor<string, []>("range_1d_1_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor<string, []>("range_1d_1")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor<string, []>("reshape_5")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor<string, []>("slice_by_index_1")];
            tensor<int32, [1]> reshape_6_shape_0 = const()[name = tensor<string, []>("reshape_6_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor<string, []>("reshape_6")];
            tensor<int32, [1]> reshape_7_shape_0 = const()[name = tensor<string, []>("reshape_7_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_7 = reshape(shape = reshape_7_shape_0, x = var_342)[name = tensor<string, []>("reshape_7")];
            tensor<int32, [1]> reshape_8_shape_0 = const()[name = tensor<string, []>("reshape_8_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_8 = reshape(shape = reshape_8_shape_0, x = reshape_4)[name = tensor<string, []>("reshape_8")];
            tensor<string, []> scatter_1_mode_0 = const()[name = tensor<string, []>("scatter_1_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_1_axis_0 = const()[name = tensor<string, []>("scatter_1_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_1_validate_indices_0 = const()[name = tensor<string, []>("scatter_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_1 = scatter(axis = scatter_1_axis_0, data = reshape_8, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7, validate_indices = scatter_1_validate_indices_0)[name = tensor<string, []>("scatter_1")];
            tensor<fp32, [2, 1, 512, 16, 64]> new_cache_1_internal_tensor_assign_2 = reshape(shape = shape_13, x = scatter_1)[name = tensor<string, []>("reshape_9")];
            tensor<int32, [5]> keys_1_begin_0 = const()[name = tensor<string, []>("keys_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> keys_1_end_0 = const()[name = tensor<string, []>("keys_1_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> keys_1_end_mask_0 = const()[name = tensor<string, []>("keys_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> keys_1_squeeze_mask_0 = const()[name = tensor<string, []>("keys_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> keys_1 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = new_cache_1_internal_tensor_assign_2)[name = tensor<string, []>("keys_1")];
            tensor<int32, [5]> values_1_begin_0 = const()[name = tensor<string, []>("values_1_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> values_1_end_0 = const()[name = tensor<string, []>("values_1_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> values_1_end_mask_0 = const()[name = tensor<string, []>("values_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> values_1_squeeze_mask_0 = const()[name = tensor<string, []>("values_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> values_1 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = new_cache_1_internal_tensor_assign_2)[name = tensor<string, []>("values_1")];
            tensor<bool, [1, 512, 16, 64]> var_354 = not_equal(x = keys_1, y = keys_1)[name = tensor<string, []>("op_354")];
            tensor<fp32, [1, 512, 16, 64]> var_360 = const()[name = tensor<string, []>("op_360"), val = tensor<fp32, [1, 512, 16, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302251648)))];
            tensor<fp32, [1, 512, 16, 64]> keys_3 = select(a = var_360, b = keys_1, cond = var_354)[name = tensor<string, []>("keys_3")];
            tensor<bool, [1, 512, 16, 64]> var_362 = not_equal(x = values_1, y = values_1)[name = tensor<string, []>("op_362")];
            tensor<fp32, [1, 512, 16, 64]> values_3 = select(a = var_360, b = values_1, cond = var_362)[name = tensor<string, []>("values_3")];
            tensor<int32, [4]> var_386 = const()[name = tensor<string, []>("op_386"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_399 = const()[name = tensor<string, []>("op_399"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<fp32, [1, 1, 1]> var_400 = reshape(shape = var_399, x = position0)[name = tensor<string, []>("op_400")];
            tensor<fp32, []> var_417 = const()[name = tensor<string, []>("op_417"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1, 1, 1]> valid_len_1 = add(x = var_400, y = var_417)[name = tensor<string, []>("valid_len_1")];
            tensor<fp32, [1, 1, 512]> k_positions_1_promoted = const()[name = tensor<string, []>("k_positions_1_promoted"), val = tensor<fp32, [1, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304348864)))];
            tensor<bool, [1, 1, 512]> valid_mask_1 = less(x = k_positions_1_promoted, y = valid_len_1)[name = tensor<string, []>("valid_mask_1")];
            tensor<bool, [1, 1, 512]> causal_mask_1 = less_equal(x = k_positions_1_promoted, y = var_400)[name = tensor<string, []>("causal_mask_1")];
            tensor<bool, [1, 1, 512]> attn_mask_1 = logical_and(x = valid_mask_1, y = causal_mask_1)[name = tensor<string, []>("attn_mask_1")];
            tensor<int32, [1]> attn_mask_3_axes_0 = const()[name = tensor<string, []>("attn_mask_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 1, 512]> attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor<string, []>("attn_mask_3")];
            tensor<fp32, [1]> var_429 = const()[name = tensor<string, []>("op_429"), val = tensor<fp32, [1]>([0x1.fffe5cp-4])];
            tensor<bool, []> var_435_transpose_x_0 = const()[name = tensor<string, []>("op_435_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_435_transpose_y_0 = const()[name = tensor<string, []>("op_435_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = tensor<string, []>("transpose_18_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_19_perm_0 = const()[name = tensor<string, []>("transpose_19_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp32, [1, 16, 64, 512]> transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3)[name = tensor<string, []>("transpose_51")];
            tensor<fp32, [1, 16, 1, 64]> transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3)[name = tensor<string, []>("transpose_52")];
            tensor<fp32, [1, 16, 1, 512]> var_435 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor<string, []>("op_435")];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_1 = mul(x = var_435, y = var_429)[name = tensor<string, []>("attn_weights_1")];
            tensor<bool, [1, 1, 1, 512]> var_437 = logical_not(x = attn_mask_3)[name = tensor<string, []>("op_437")];
            tensor<fp32, []> var_438 = const()[name = tensor<string, []>("op_438"), val = tensor<fp32, []>(-0x1.ff933cp+127)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_3 = select(a = var_438, b = attn_weights_1, cond = var_437)[name = tensor<string, []>("attn_weights_3")];
            tensor<int32, []> var_440 = const()[name = tensor<string, []>("op_440"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_5 = softmax(axis = var_440, x = attn_weights_3)[name = tensor<string, []>("attn_weights_5")];
            tensor<bool, []> attn_output_1_transpose_x_0 = const()[name = tensor<string, []>("attn_output_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_1_transpose_y_0 = const()[name = tensor<string, []>("attn_output_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 16, 512, 64]> values_5 = transpose(perm = var_386, x = values_3)[name = tensor<string, []>("transpose_53")];
            tensor<fp32, [1, 16, 1, 64]> attn_output_1 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5, y = values_5)[name = tensor<string, []>("attn_output_1")];
            tensor<int32, [4]> var_448 = const()[name = tensor<string, []>("op_448"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_451 = const()[name = tensor<string, []>("op_451"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp32, [1, 1, 16, 64]> var_449 = transpose(perm = var_448, x = attn_output_1)[name = tensor<string, []>("transpose_50")];
            tensor<fp32, [1, 1, 1024]> input_5 = reshape(shape = var_451, x = var_449)[name = tensor<string, []>("input_5")];
            tensor<fp32, [1, 1, 1024]> attn_out_1 = linear(bias = linear_0_bias_0, weight = attn0_out_proj_weight, x = input_5)[name = tensor<string, []>("linear_2")];
            tensor<fp32, []> var_457 = const()[name = tensor<string, []>("op_457"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1]> var_458 = add(x = position0, y = var_457)[name = tensor<string, []>("op_458")];
            tensor<fp32, [1, 1, 1024]> input_7 = add(x = input_3, y = attn_out_1)[name = tensor<string, []>("input_7")];
            tensor<fp32, []> var_462 = const()[name = tensor<string, []>("op_462"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> input_9_axes_0 = const()[name = tensor<string, []>("input_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> input_9 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias, epsilon = var_462, gamma = norm0_2_weight, x = input_7)[name = tensor<string, []>("input_9")];
            tensor<fp32, [4096]> linear_3_bias_0 = const()[name = tensor<string, []>("linear_3_bias_0"), val = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304350976)))];
            tensor<fp32, [1, 1, 4096]> var_470 = linear(bias = linear_3_bias_0, weight = linear0_1_weight, x = input_9)[name = tensor<string, []>("linear_3")];
            tensor<string, []> input_11_mode_0 = const()[name = tensor<string, []>("input_11_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp32, [1, 1, 4096]> input_11 = gelu(mode = input_11_mode_0, x = var_470)[name = tensor<string, []>("input_11")];
            tensor<fp32, [1, 1, 1024]> ffn_out_1 = linear(bias = linear_0_bias_0, weight = linear0_2_weight, x = input_11)[name = tensor<string, []>("linear_4")];
            tensor<fp32, [1, 1, 1024]> input_13 = add(x = input_7, y = ffn_out_1)[name = tensor<string, []>("input_13")];
            tensor<fp32, []> var_479 = const()[name = tensor<string, []>("op_479"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> x_3_axes_0 = const()[name = tensor<string, []>("x_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> x_3 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias, epsilon = var_479, gamma = norm1_1_weight, x = input_13)[name = tensor<string, []>("x_3")];
            tensor<fp32, [1, 1, 3072]> var_511 = linear(bias = linear_1_bias_0, weight = attn1_in_proj_weight, x = x_3)[name = tensor<string, []>("linear_5")];
            tensor<int32, [5]> var_515 = const()[name = tensor<string, []>("op_515"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<fp32, [1, 1, 3, 16, 64]> qkv_3 = reshape(shape = var_515, x = var_511)[name = tensor<string, []>("qkv_3")];
            tensor<int32, [5]> q_7_begin_0 = const()[name = tensor<string, []>("q_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_7_end_0 = const()[name = tensor<string, []>("q_7_end_0"), val = tensor<int32, [5]>([1, 1, 1, 16, 64])];
            tensor<bool, [5]> q_7_end_mask_0 = const()[name = tensor<string, []>("q_7_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_7_squeeze_mask_0 = const()[name = tensor<string, []>("q_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> q_7 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3)[name = tensor<string, []>("q_7")];
            tensor<int32, [5]> k_5_begin_0 = const()[name = tensor<string, []>("k_5_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_5_end_0 = const()[name = tensor<string, []>("k_5_end_0"), val = tensor<int32, [5]>([1, 1, 2, 16, 64])];
            tensor<bool, [5]> k_5_end_mask_0 = const()[name = tensor<string, []>("k_5_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_5_squeeze_mask_0 = const()[name = tensor<string, []>("k_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> k_5 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3)[name = tensor<string, []>("k_5")];
            tensor<int32, [5]> v_3_begin_0 = const()[name = tensor<string, []>("v_3_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_3_end_0 = const()[name = tensor<string, []>("v_3_end_0"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<bool, [5]> v_3_end_mask_0 = const()[name = tensor<string, []>("v_3_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_3_squeeze_mask_0 = const()[name = tensor<string, []>("v_3_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> v_3 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3)[name = tensor<string, []>("v_3")];
            tensor<fp32, [32]> freqs_3 = const()[name = tensor<string, []>("freqs_3"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304367424)))];
            tensor<int32, [4]> var_619 = const()[name = tensor<string, []>("op_619"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp32, [1, 1, 1, 1]> ts_11 = reshape(shape = var_619, x = position1)[name = tensor<string, []>("ts_11")];
            tensor<int32, [5]> var_623 = const()[name = tensor<string, []>("op_623"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> q_complex_3 = reshape(shape = var_623, x = q_7)[name = tensor<string, []>("q_complex_3")];
            tensor<int32, [5]> var_627 = const()[name = tensor<string, []>("op_627"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> k_complex_3 = reshape(shape = var_627, x = k_5)[name = tensor<string, []>("k_complex_3")];
            tensor<int32, [5]> var_631_begin_0 = const()[name = tensor<string, []>("op_631_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_631_end_0 = const()[name = tensor<string, []>("op_631_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_631_end_mask_0 = const()[name = tensor<string, []>("op_631_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_631_squeeze_mask_0 = const()[name = tensor<string, []>("op_631_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_631 = slice_by_index(begin = var_631_begin_0, end = var_631_end_0, end_mask = var_631_end_mask_0, squeeze_mask = var_631_squeeze_mask_0, x = q_complex_3)[name = tensor<string, []>("op_631")];
            tensor<int32, [5]> var_639_begin_0 = const()[name = tensor<string, []>("op_639_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_639_end_0 = const()[name = tensor<string, []>("op_639_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_639_end_mask_0 = const()[name = tensor<string, []>("op_639_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_639_squeeze_mask_0 = const()[name = tensor<string, []>("op_639_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_639 = slice_by_index(begin = var_639_begin_0, end = var_639_end_0, end_mask = var_639_end_mask_0, squeeze_mask = var_639_squeeze_mask_0, x = q_complex_3)[name = tensor<string, []>("op_639")];
            tensor<int32, [5]> var_647_begin_0 = const()[name = tensor<string, []>("op_647_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_647_end_0 = const()[name = tensor<string, []>("op_647_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_647_end_mask_0 = const()[name = tensor<string, []>("op_647_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_647_squeeze_mask_0 = const()[name = tensor<string, []>("op_647_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_647 = slice_by_index(begin = var_647_begin_0, end = var_647_end_0, end_mask = var_647_end_mask_0, squeeze_mask = var_647_squeeze_mask_0, x = k_complex_3)[name = tensor<string, []>("op_647")];
            tensor<int32, [5]> var_655_begin_0 = const()[name = tensor<string, []>("op_655_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_655_end_0 = const()[name = tensor<string, []>("op_655_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_655_end_mask_0 = const()[name = tensor<string, []>("op_655_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_655_squeeze_mask_0 = const()[name = tensor<string, []>("op_655_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_655 = slice_by_index(begin = var_655_begin_0, end = var_655_end_0, end_mask = var_655_end_mask_0, squeeze_mask = var_655_squeeze_mask_0, x = k_complex_3)[name = tensor<string, []>("op_655")];
            tensor<fp32, [1, 1, 1, 32]> var_661 = mul(x = freqs_3, y = ts_11)[name = tensor<string, []>("op_661")];
            tensor<fp32, [1, 1, 1, 32]> rotr_3 = cos(x = var_661)[name = tensor<string, []>("rotr_3")];
            tensor<fp32, [1, 1, 1, 32]> roti_3 = sin(x = var_661)[name = tensor<string, []>("roti_3")];
            tensor<fp32, [1, 1, 16, 32]> var_665 = mul(x = var_631, y = rotr_3)[name = tensor<string, []>("op_665")];
            tensor<fp32, [1, 1, 16, 32]> var_666 = mul(x = var_639, y = roti_3)[name = tensor<string, []>("op_666")];
            tensor<fp32, [1, 1, 16, 32]> qor_5 = sub(x = var_665, y = var_666)[name = tensor<string, []>("qor_5")];
            tensor<fp32, [1, 1, 16, 32]> var_669 = mul(x = var_631, y = roti_3)[name = tensor<string, []>("op_669")];
            tensor<fp32, [1, 1, 16, 32]> var_670 = mul(x = var_639, y = rotr_3)[name = tensor<string, []>("op_670")];
            tensor<fp32, [1, 1, 16, 32]> qoi_5 = add(x = var_669, y = var_670)[name = tensor<string, []>("qoi_5")];
            tensor<fp32, [1, 1, 16, 32]> var_673 = mul(x = var_647, y = rotr_3)[name = tensor<string, []>("op_673")];
            tensor<fp32, [1, 1, 16, 32]> var_674 = mul(x = var_655, y = roti_3)[name = tensor<string, []>("op_674")];
            tensor<fp32, [1, 1, 16, 32]> kor_5 = sub(x = var_673, y = var_674)[name = tensor<string, []>("kor_5")];
            tensor<fp32, [1, 1, 16, 32]> var_677 = mul(x = var_647, y = roti_3)[name = tensor<string, []>("op_677")];
            tensor<fp32, [1, 1, 16, 32]> var_678 = mul(x = var_655, y = rotr_3)[name = tensor<string, []>("op_678")];
            tensor<fp32, [1, 1, 16, 32]> koi_5 = add(x = var_677, y = var_678)[name = tensor<string, []>("koi_5")];
            tensor<int32, []> qo_3_axis_0 = const()[name = tensor<string, []>("qo_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> qo_3 = stack(axis = qo_3_axis_0, values = (qor_5, qoi_5))[name = tensor<string, []>("qo_3")];
            tensor<int32, []> ko_3_axis_0 = const()[name = tensor<string, []>("ko_3_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> ko_3 = stack(axis = ko_3_axis_0, values = (kor_5, koi_5))[name = tensor<string, []>("ko_3")];
            tensor<int32, [4]> var_707 = const()[name = tensor<string, []>("op_707"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> q_9 = reshape(shape = var_707, x = qo_3)[name = tensor<string, []>("q_9")];
            tensor<int32, [4]> var_709 = const()[name = tensor<string, []>("op_709"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> k_7 = reshape(shape = var_709, x = ko_3)[name = tensor<string, []>("k_7")];
            tensor<fp32, []> _inversed_731_y_0 = const()[name = tensor<string, []>("_inversed_731_y_0"), val = tensor<fp32, []>(0x1p-9)];
            tensor<fp32, [1, 1, 1, 1]> _inversed_731 = mul(x = ts_11, y = _inversed_731_y_0)[name = tensor<string, []>("_inversed_731")];
            tensor<fp32, [1, 1, 1, 1]> var_732 = floor(x = _inversed_731)[name = tensor<string, []>("op_732")];
            tensor<fp32, []> var_733 = const()[name = tensor<string, []>("op_733"), val = tensor<fp32, []>(0x1p+9)];
            tensor<fp32, [1, 1, 1, 1]> var_734 = mul(x = var_732, y = var_733)[name = tensor<string, []>("op_734")];
            tensor<fp32, [1, 1, 1, 1]> write_indices_float_7 = sub(x = ts_11, y = var_734)[name = tensor<string, []>("write_indices_float_7")];
            tensor<string, []> var_741_dtype_0 = const()[name = tensor<string, []>("op_741_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [4]> write_indices_3_reps_0 = const()[name = tensor<string, []>("write_indices_3_reps_0"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<int32, [1, 1, 1, 1]> var_741 = cast(dtype = var_741_dtype_0, x = write_indices_float_7)[name = tensor<string, []>("cast_112")];
            tensor<int32, [1, 1, 16, 64]> write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_741)[name = tensor<string, []>("write_indices_3")];
            tensor<int32, [5]> var_749_begin_0 = const()[name = tensor<string, []>("op_749_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_749_end_0 = const()[name = tensor<string, []>("op_749_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> var_749_end_mask_0 = const()[name = tensor<string, []>("op_749_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_749_squeeze_mask_0 = const()[name = tensor<string, []>("op_749_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_749 = slice_by_index(begin = var_749_begin_0, end = var_749_end_0, end_mask = var_749_end_mask_0, squeeze_mask = var_749_squeeze_mask_0, x = cache1)[name = tensor<string, []>("op_749")];
            tensor<int32, []> var_751_axis_0 = const()[name = tensor<string, []>("op_751_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_751_mode_0 = const()[name = tensor<string, []>("op_751_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_751_validate_indices_0 = const()[name = tensor<string, []>("op_751_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_751 = scatter_along_axis(axis = var_751_axis_0, data = var_749, indices = write_indices_3, mode = var_751_mode_0, updates = k_7, validate_indices = var_751_validate_indices_0)[name = tensor<string, []>("op_751")];
            tensor<int32, [5]> concat_9 = const()[name = tensor<string, []>("concat_9"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_10 = const()[name = tensor<string, []>("concat_10"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_14 = const()[name = tensor<string, []>("shape_14"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_2 = const()[name = tensor<string, []>("reduce_prod_2"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_2_start_0 = const()[name = tensor<string, []>("range_1d_2_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_2_step_0 = const()[name = tensor<string, []>("range_1d_2_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor<string, []>("range_1d_2")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor<string, []>("reshape_10")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor<string, []>("slice_by_index_2")];
            tensor<int32, [1]> reshape_11_shape_0 = const()[name = tensor<string, []>("reshape_11_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor<string, []>("reshape_11")];
            tensor<int32, [1]> reshape_12_shape_0 = const()[name = tensor<string, []>("reshape_12_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_12 = reshape(shape = reshape_12_shape_0, x = var_751)[name = tensor<string, []>("reshape_12")];
            tensor<int32, [1]> reshape_13_shape_0 = const()[name = tensor<string, []>("reshape_13_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_13 = reshape(shape = reshape_13_shape_0, x = cache1)[name = tensor<string, []>("reshape_13")];
            tensor<string, []> scatter_2_mode_0 = const()[name = tensor<string, []>("scatter_2_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_2_axis_0 = const()[name = tensor<string, []>("scatter_2_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_2_validate_indices_0 = const()[name = tensor<string, []>("scatter_2_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_2 = scatter(axis = scatter_2_axis_0, data = reshape_13, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12, validate_indices = scatter_2_validate_indices_0)[name = tensor<string, []>("scatter_2")];
            tensor<fp32, [2, 1, 512, 16, 64]> reshape_14 = reshape(shape = shape_14, x = scatter_2)[name = tensor<string, []>("reshape_14")];
            tensor<int32, [5]> var_759_begin_0 = const()[name = tensor<string, []>("op_759_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_759_end_0 = const()[name = tensor<string, []>("op_759_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> var_759_end_mask_0 = const()[name = tensor<string, []>("op_759_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_759_squeeze_mask_0 = const()[name = tensor<string, []>("op_759_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_759 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, squeeze_mask = var_759_squeeze_mask_0, x = reshape_14)[name = tensor<string, []>("op_759")];
            tensor<int32, []> var_761_axis_0 = const()[name = tensor<string, []>("op_761_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_761_mode_0 = const()[name = tensor<string, []>("op_761_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_761_validate_indices_0 = const()[name = tensor<string, []>("op_761_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_761 = scatter_along_axis(axis = var_761_axis_0, data = var_759, indices = write_indices_3, mode = var_761_mode_0, updates = v_3, validate_indices = var_761_validate_indices_0)[name = tensor<string, []>("op_761")];
            tensor<int32, [5]> concat_11 = const()[name = tensor<string, []>("concat_11"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_12 = const()[name = tensor<string, []>("concat_12"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_15 = const()[name = tensor<string, []>("shape_15"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_3 = const()[name = tensor<string, []>("reduce_prod_3"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_3_start_0 = const()[name = tensor<string, []>("range_1d_3_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_3_step_0 = const()[name = tensor<string, []>("range_1d_3_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor<string, []>("range_1d_3")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor<string, []>("reshape_15")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor<string, []>("slice_by_index_3")];
            tensor<int32, [1]> reshape_16_shape_0 = const()[name = tensor<string, []>("reshape_16_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor<string, []>("reshape_16")];
            tensor<int32, [1]> reshape_17_shape_0 = const()[name = tensor<string, []>("reshape_17_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_17 = reshape(shape = reshape_17_shape_0, x = var_761)[name = tensor<string, []>("reshape_17")];
            tensor<int32, [1]> reshape_18_shape_0 = const()[name = tensor<string, []>("reshape_18_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_18 = reshape(shape = reshape_18_shape_0, x = reshape_14)[name = tensor<string, []>("reshape_18")];
            tensor<string, []> scatter_3_mode_0 = const()[name = tensor<string, []>("scatter_3_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_3_axis_0 = const()[name = tensor<string, []>("scatter_3_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_3_validate_indices_0 = const()[name = tensor<string, []>("scatter_3_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_3 = scatter(axis = scatter_3_axis_0, data = reshape_18, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17, validate_indices = scatter_3_validate_indices_0)[name = tensor<string, []>("scatter_3")];
            tensor<fp32, [2, 1, 512, 16, 64]> new_cache_3_internal_tensor_assign_2 = reshape(shape = shape_15, x = scatter_3)[name = tensor<string, []>("reshape_19")];
            tensor<int32, [5]> keys_7_begin_0 = const()[name = tensor<string, []>("keys_7_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> keys_7_end_0 = const()[name = tensor<string, []>("keys_7_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> keys_7_end_mask_0 = const()[name = tensor<string, []>("keys_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> keys_7_squeeze_mask_0 = const()[name = tensor<string, []>("keys_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> keys_7 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = new_cache_3_internal_tensor_assign_2)[name = tensor<string, []>("keys_7")];
            tensor<int32, [5]> values_7_begin_0 = const()[name = tensor<string, []>("values_7_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> values_7_end_0 = const()[name = tensor<string, []>("values_7_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> values_7_end_mask_0 = const()[name = tensor<string, []>("values_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> values_7_squeeze_mask_0 = const()[name = tensor<string, []>("values_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> values_7 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = new_cache_3_internal_tensor_assign_2)[name = tensor<string, []>("values_7")];
            tensor<bool, [1, 512, 16, 64]> var_773 = not_equal(x = keys_7, y = keys_7)[name = tensor<string, []>("op_773")];
            tensor<fp32, [1, 512, 16, 64]> keys_9 = select(a = var_360, b = keys_7, cond = var_773)[name = tensor<string, []>("keys_9")];
            tensor<bool, [1, 512, 16, 64]> var_781 = not_equal(x = values_7, y = values_7)[name = tensor<string, []>("op_781")];
            tensor<fp32, [1, 512, 16, 64]> values_9 = select(a = var_360, b = values_7, cond = var_781)[name = tensor<string, []>("values_9")];
            tensor<int32, [4]> var_805 = const()[name = tensor<string, []>("op_805"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_818 = const()[name = tensor<string, []>("op_818"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<fp32, [1, 1, 1]> var_819 = reshape(shape = var_818, x = position1)[name = tensor<string, []>("op_819")];
            tensor<fp32, []> var_836 = const()[name = tensor<string, []>("op_836"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1, 1, 1]> valid_len_3 = add(x = var_819, y = var_836)[name = tensor<string, []>("valid_len_3")];
            tensor<bool, [1, 1, 512]> valid_mask_3 = less(x = k_positions_1_promoted, y = valid_len_3)[name = tensor<string, []>("valid_mask_3")];
            tensor<bool, [1, 1, 512]> causal_mask_3 = less_equal(x = k_positions_1_promoted, y = var_819)[name = tensor<string, []>("causal_mask_3")];
            tensor<bool, [1, 1, 512]> attn_mask_5 = logical_and(x = valid_mask_3, y = causal_mask_3)[name = tensor<string, []>("attn_mask_5")];
            tensor<int32, [1]> attn_mask_7_axes_0 = const()[name = tensor<string, []>("attn_mask_7_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 1, 512]> attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor<string, []>("attn_mask_7")];
            tensor<fp32, [1]> var_848 = const()[name = tensor<string, []>("op_848"), val = tensor<fp32, [1]>([0x1.fffe5cp-4])];
            tensor<bool, []> var_854_transpose_x_0 = const()[name = tensor<string, []>("op_854_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_854_transpose_y_0 = const()[name = tensor<string, []>("op_854_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = tensor<string, []>("transpose_20_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_21_perm_0 = const()[name = tensor<string, []>("transpose_21_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp32, [1, 16, 64, 512]> transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9)[name = tensor<string, []>("transpose_47")];
            tensor<fp32, [1, 16, 1, 64]> transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9)[name = tensor<string, []>("transpose_48")];
            tensor<fp32, [1, 16, 1, 512]> var_854 = matmul(transpose_x = var_854_transpose_x_0, transpose_y = var_854_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor<string, []>("op_854")];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_7 = mul(x = var_854, y = var_848)[name = tensor<string, []>("attn_weights_7")];
            tensor<bool, [1, 1, 1, 512]> var_856 = logical_not(x = attn_mask_7)[name = tensor<string, []>("op_856")];
            tensor<fp32, []> var_857 = const()[name = tensor<string, []>("op_857"), val = tensor<fp32, []>(-0x1.ff933cp+127)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_9 = select(a = var_857, b = attn_weights_7, cond = var_856)[name = tensor<string, []>("attn_weights_9")];
            tensor<int32, []> var_859 = const()[name = tensor<string, []>("op_859"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_11 = softmax(axis = var_859, x = attn_weights_9)[name = tensor<string, []>("attn_weights_11")];
            tensor<bool, []> attn_output_3_transpose_x_0 = const()[name = tensor<string, []>("attn_output_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_3_transpose_y_0 = const()[name = tensor<string, []>("attn_output_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 16, 512, 64]> values_11 = transpose(perm = var_805, x = values_9)[name = tensor<string, []>("transpose_49")];
            tensor<fp32, [1, 16, 1, 64]> attn_output_3 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11, y = values_11)[name = tensor<string, []>("attn_output_3")];
            tensor<int32, [4]> var_867 = const()[name = tensor<string, []>("op_867"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_870 = const()[name = tensor<string, []>("op_870"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp32, [1, 1, 16, 64]> var_868 = transpose(perm = var_867, x = attn_output_3)[name = tensor<string, []>("transpose_46")];
            tensor<fp32, [1, 1, 1024]> input_15 = reshape(shape = var_870, x = var_868)[name = tensor<string, []>("input_15")];
            tensor<fp32, [1, 1, 1024]> attn_out_3 = linear(bias = linear_0_bias_0, weight = attn1_out_proj_weight, x = input_15)[name = tensor<string, []>("linear_6")];
            tensor<fp32, []> var_876 = const()[name = tensor<string, []>("op_876"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1]> var_877 = add(x = position1, y = var_876)[name = tensor<string, []>("op_877")];
            tensor<fp32, [1, 1, 1024]> input_17 = add(x = input_13, y = attn_out_3)[name = tensor<string, []>("input_17")];
            tensor<fp32, []> var_881 = const()[name = tensor<string, []>("op_881"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> input_19_axes_0 = const()[name = tensor<string, []>("input_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> input_19 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias, epsilon = var_881, gamma = norm1_2_weight, x = input_17)[name = tensor<string, []>("input_19")];
            tensor<fp32, [1, 1, 4096]> var_889 = linear(bias = linear_3_bias_0, weight = linear1_1_weight, x = input_19)[name = tensor<string, []>("linear_7")];
            tensor<string, []> input_21_mode_0 = const()[name = tensor<string, []>("input_21_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp32, [1, 1, 4096]> input_21 = gelu(mode = input_21_mode_0, x = var_889)[name = tensor<string, []>("input_21")];
            tensor<fp32, [1, 1, 1024]> ffn_out_3 = linear(bias = linear_0_bias_0, weight = linear1_2_weight, x = input_21)[name = tensor<string, []>("linear_8")];
            tensor<fp32, [1, 1, 1024]> input_23 = add(x = input_17, y = ffn_out_3)[name = tensor<string, []>("input_23")];
            tensor<fp32, []> var_898 = const()[name = tensor<string, []>("op_898"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> x_5_axes_0 = const()[name = tensor<string, []>("x_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> x_5 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias, epsilon = var_898, gamma = norm2_1_weight, x = input_23)[name = tensor<string, []>("x_5")];
            tensor<fp32, [1, 1, 3072]> var_930 = linear(bias = linear_1_bias_0, weight = attn2_in_proj_weight, x = x_5)[name = tensor<string, []>("linear_9")];
            tensor<int32, [5]> var_934 = const()[name = tensor<string, []>("op_934"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<fp32, [1, 1, 3, 16, 64]> qkv_5 = reshape(shape = var_934, x = var_930)[name = tensor<string, []>("qkv_5")];
            tensor<int32, [5]> q_13_begin_0 = const()[name = tensor<string, []>("q_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_13_end_0 = const()[name = tensor<string, []>("q_13_end_0"), val = tensor<int32, [5]>([1, 1, 1, 16, 64])];
            tensor<bool, [5]> q_13_end_mask_0 = const()[name = tensor<string, []>("q_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_13_squeeze_mask_0 = const()[name = tensor<string, []>("q_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> q_13 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5)[name = tensor<string, []>("q_13")];
            tensor<int32, [5]> k_9_begin_0 = const()[name = tensor<string, []>("k_9_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_9_end_0 = const()[name = tensor<string, []>("k_9_end_0"), val = tensor<int32, [5]>([1, 1, 2, 16, 64])];
            tensor<bool, [5]> k_9_end_mask_0 = const()[name = tensor<string, []>("k_9_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_9_squeeze_mask_0 = const()[name = tensor<string, []>("k_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> k_9 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5)[name = tensor<string, []>("k_9")];
            tensor<int32, [5]> v_5_begin_0 = const()[name = tensor<string, []>("v_5_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_5_end_0 = const()[name = tensor<string, []>("v_5_end_0"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<bool, [5]> v_5_end_mask_0 = const()[name = tensor<string, []>("v_5_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_5_squeeze_mask_0 = const()[name = tensor<string, []>("v_5_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> v_5 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5)[name = tensor<string, []>("v_5")];
            tensor<fp32, [32]> freqs_5 = const()[name = tensor<string, []>("freqs_5"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304367616)))];
            tensor<int32, [4]> var_1038 = const()[name = tensor<string, []>("op_1038"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp32, [1, 1, 1, 1]> ts_17 = reshape(shape = var_1038, x = position2)[name = tensor<string, []>("ts_17")];
            tensor<int32, [5]> var_1042 = const()[name = tensor<string, []>("op_1042"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> q_complex_5 = reshape(shape = var_1042, x = q_13)[name = tensor<string, []>("q_complex_5")];
            tensor<int32, [5]> var_1046 = const()[name = tensor<string, []>("op_1046"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> k_complex_5 = reshape(shape = var_1046, x = k_9)[name = tensor<string, []>("k_complex_5")];
            tensor<int32, [5]> var_1050_begin_0 = const()[name = tensor<string, []>("op_1050_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1050_end_0 = const()[name = tensor<string, []>("op_1050_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_1050_end_mask_0 = const()[name = tensor<string, []>("op_1050_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1050_squeeze_mask_0 = const()[name = tensor<string, []>("op_1050_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1050 = slice_by_index(begin = var_1050_begin_0, end = var_1050_end_0, end_mask = var_1050_end_mask_0, squeeze_mask = var_1050_squeeze_mask_0, x = q_complex_5)[name = tensor<string, []>("op_1050")];
            tensor<int32, [5]> var_1058_begin_0 = const()[name = tensor<string, []>("op_1058_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_1058_end_0 = const()[name = tensor<string, []>("op_1058_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_1058_end_mask_0 = const()[name = tensor<string, []>("op_1058_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1058_squeeze_mask_0 = const()[name = tensor<string, []>("op_1058_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1058 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, squeeze_mask = var_1058_squeeze_mask_0, x = q_complex_5)[name = tensor<string, []>("op_1058")];
            tensor<int32, [5]> var_1066_begin_0 = const()[name = tensor<string, []>("op_1066_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1066_end_0 = const()[name = tensor<string, []>("op_1066_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_1066_end_mask_0 = const()[name = tensor<string, []>("op_1066_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1066_squeeze_mask_0 = const()[name = tensor<string, []>("op_1066_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1066 = slice_by_index(begin = var_1066_begin_0, end = var_1066_end_0, end_mask = var_1066_end_mask_0, squeeze_mask = var_1066_squeeze_mask_0, x = k_complex_5)[name = tensor<string, []>("op_1066")];
            tensor<int32, [5]> var_1074_begin_0 = const()[name = tensor<string, []>("op_1074_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_1074_end_0 = const()[name = tensor<string, []>("op_1074_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_1074_end_mask_0 = const()[name = tensor<string, []>("op_1074_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1074_squeeze_mask_0 = const()[name = tensor<string, []>("op_1074_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1074 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = k_complex_5)[name = tensor<string, []>("op_1074")];
            tensor<fp32, [1, 1, 1, 32]> var_1080 = mul(x = freqs_5, y = ts_17)[name = tensor<string, []>("op_1080")];
            tensor<fp32, [1, 1, 1, 32]> rotr_5 = cos(x = var_1080)[name = tensor<string, []>("rotr_5")];
            tensor<fp32, [1, 1, 1, 32]> roti_5 = sin(x = var_1080)[name = tensor<string, []>("roti_5")];
            tensor<fp32, [1, 1, 16, 32]> var_1084 = mul(x = var_1050, y = rotr_5)[name = tensor<string, []>("op_1084")];
            tensor<fp32, [1, 1, 16, 32]> var_1085 = mul(x = var_1058, y = roti_5)[name = tensor<string, []>("op_1085")];
            tensor<fp32, [1, 1, 16, 32]> qor_9 = sub(x = var_1084, y = var_1085)[name = tensor<string, []>("qor_9")];
            tensor<fp32, [1, 1, 16, 32]> var_1088 = mul(x = var_1050, y = roti_5)[name = tensor<string, []>("op_1088")];
            tensor<fp32, [1, 1, 16, 32]> var_1089 = mul(x = var_1058, y = rotr_5)[name = tensor<string, []>("op_1089")];
            tensor<fp32, [1, 1, 16, 32]> qoi_9 = add(x = var_1088, y = var_1089)[name = tensor<string, []>("qoi_9")];
            tensor<fp32, [1, 1, 16, 32]> var_1092 = mul(x = var_1066, y = rotr_5)[name = tensor<string, []>("op_1092")];
            tensor<fp32, [1, 1, 16, 32]> var_1093 = mul(x = var_1074, y = roti_5)[name = tensor<string, []>("op_1093")];
            tensor<fp32, [1, 1, 16, 32]> kor_9 = sub(x = var_1092, y = var_1093)[name = tensor<string, []>("kor_9")];
            tensor<fp32, [1, 1, 16, 32]> var_1096 = mul(x = var_1066, y = roti_5)[name = tensor<string, []>("op_1096")];
            tensor<fp32, [1, 1, 16, 32]> var_1097 = mul(x = var_1074, y = rotr_5)[name = tensor<string, []>("op_1097")];
            tensor<fp32, [1, 1, 16, 32]> koi_9 = add(x = var_1096, y = var_1097)[name = tensor<string, []>("koi_9")];
            tensor<int32, []> qo_5_axis_0 = const()[name = tensor<string, []>("qo_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> qo_5 = stack(axis = qo_5_axis_0, values = (qor_9, qoi_9))[name = tensor<string, []>("qo_5")];
            tensor<int32, []> ko_5_axis_0 = const()[name = tensor<string, []>("ko_5_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> ko_5 = stack(axis = ko_5_axis_0, values = (kor_9, koi_9))[name = tensor<string, []>("ko_5")];
            tensor<int32, [4]> var_1126 = const()[name = tensor<string, []>("op_1126"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> q_15 = reshape(shape = var_1126, x = qo_5)[name = tensor<string, []>("q_15")];
            tensor<int32, [4]> var_1128 = const()[name = tensor<string, []>("op_1128"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> k_11 = reshape(shape = var_1128, x = ko_5)[name = tensor<string, []>("k_11")];
            tensor<fp32, []> _inversed_1150_y_0 = const()[name = tensor<string, []>("_inversed_1150_y_0"), val = tensor<fp32, []>(0x1p-9)];
            tensor<fp32, [1, 1, 1, 1]> _inversed_1150 = mul(x = ts_17, y = _inversed_1150_y_0)[name = tensor<string, []>("_inversed_1150")];
            tensor<fp32, [1, 1, 1, 1]> var_1151 = floor(x = _inversed_1150)[name = tensor<string, []>("op_1151")];
            tensor<fp32, []> var_1152 = const()[name = tensor<string, []>("op_1152"), val = tensor<fp32, []>(0x1p+9)];
            tensor<fp32, [1, 1, 1, 1]> var_1153 = mul(x = var_1151, y = var_1152)[name = tensor<string, []>("op_1153")];
            tensor<fp32, [1, 1, 1, 1]> write_indices_float_11 = sub(x = ts_17, y = var_1153)[name = tensor<string, []>("write_indices_float_11")];
            tensor<string, []> var_1160_dtype_0 = const()[name = tensor<string, []>("op_1160_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [4]> write_indices_5_reps_0 = const()[name = tensor<string, []>("write_indices_5_reps_0"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<int32, [1, 1, 1, 1]> var_1160 = cast(dtype = var_1160_dtype_0, x = write_indices_float_11)[name = tensor<string, []>("cast_111")];
            tensor<int32, [1, 1, 16, 64]> write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1160)[name = tensor<string, []>("write_indices_5")];
            tensor<int32, [5]> var_1168_begin_0 = const()[name = tensor<string, []>("op_1168_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1168_end_0 = const()[name = tensor<string, []>("op_1168_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> var_1168_end_mask_0 = const()[name = tensor<string, []>("op_1168_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1168_squeeze_mask_0 = const()[name = tensor<string, []>("op_1168_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_1168 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, squeeze_mask = var_1168_squeeze_mask_0, x = cache2)[name = tensor<string, []>("op_1168")];
            tensor<int32, []> var_1170_axis_0 = const()[name = tensor<string, []>("op_1170_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_1170_mode_0 = const()[name = tensor<string, []>("op_1170_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_1170_validate_indices_0 = const()[name = tensor<string, []>("op_1170_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_1170 = scatter_along_axis(axis = var_1170_axis_0, data = var_1168, indices = write_indices_5, mode = var_1170_mode_0, updates = k_11, validate_indices = var_1170_validate_indices_0)[name = tensor<string, []>("op_1170")];
            tensor<int32, [5]> concat_16 = const()[name = tensor<string, []>("concat_16"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_17 = const()[name = tensor<string, []>("concat_17"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_16 = const()[name = tensor<string, []>("shape_16"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_4 = const()[name = tensor<string, []>("reduce_prod_4"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_4_start_0 = const()[name = tensor<string, []>("range_1d_4_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_4_step_0 = const()[name = tensor<string, []>("range_1d_4_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor<string, []>("range_1d_4")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor<string, []>("reshape_20")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor<string, []>("slice_by_index_4")];
            tensor<int32, [1]> reshape_21_shape_0 = const()[name = tensor<string, []>("reshape_21_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor<string, []>("reshape_21")];
            tensor<int32, [1]> reshape_22_shape_0 = const()[name = tensor<string, []>("reshape_22_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_22 = reshape(shape = reshape_22_shape_0, x = var_1170)[name = tensor<string, []>("reshape_22")];
            tensor<int32, [1]> reshape_23_shape_0 = const()[name = tensor<string, []>("reshape_23_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_23 = reshape(shape = reshape_23_shape_0, x = cache2)[name = tensor<string, []>("reshape_23")];
            tensor<string, []> scatter_4_mode_0 = const()[name = tensor<string, []>("scatter_4_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_4_axis_0 = const()[name = tensor<string, []>("scatter_4_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_4_validate_indices_0 = const()[name = tensor<string, []>("scatter_4_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_4 = scatter(axis = scatter_4_axis_0, data = reshape_23, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22, validate_indices = scatter_4_validate_indices_0)[name = tensor<string, []>("scatter_4")];
            tensor<fp32, [2, 1, 512, 16, 64]> reshape_24 = reshape(shape = shape_16, x = scatter_4)[name = tensor<string, []>("reshape_24")];
            tensor<int32, [5]> var_1178_begin_0 = const()[name = tensor<string, []>("op_1178_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1178_end_0 = const()[name = tensor<string, []>("op_1178_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> var_1178_end_mask_0 = const()[name = tensor<string, []>("op_1178_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1178_squeeze_mask_0 = const()[name = tensor<string, []>("op_1178_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_1178 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, squeeze_mask = var_1178_squeeze_mask_0, x = reshape_24)[name = tensor<string, []>("op_1178")];
            tensor<int32, []> var_1180_axis_0 = const()[name = tensor<string, []>("op_1180_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_1180_mode_0 = const()[name = tensor<string, []>("op_1180_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_1180_validate_indices_0 = const()[name = tensor<string, []>("op_1180_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_1180 = scatter_along_axis(axis = var_1180_axis_0, data = var_1178, indices = write_indices_5, mode = var_1180_mode_0, updates = v_5, validate_indices = var_1180_validate_indices_0)[name = tensor<string, []>("op_1180")];
            tensor<int32, [5]> concat_18 = const()[name = tensor<string, []>("concat_18"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_19 = const()[name = tensor<string, []>("concat_19"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_17 = const()[name = tensor<string, []>("shape_17"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_5 = const()[name = tensor<string, []>("reduce_prod_5"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_5_start_0 = const()[name = tensor<string, []>("range_1d_5_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_5_step_0 = const()[name = tensor<string, []>("range_1d_5_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor<string, []>("range_1d_5")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor<string, []>("reshape_25")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor<string, []>("slice_by_index_5")];
            tensor<int32, [1]> reshape_26_shape_0 = const()[name = tensor<string, []>("reshape_26_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor<string, []>("reshape_26")];
            tensor<int32, [1]> reshape_27_shape_0 = const()[name = tensor<string, []>("reshape_27_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_27 = reshape(shape = reshape_27_shape_0, x = var_1180)[name = tensor<string, []>("reshape_27")];
            tensor<int32, [1]> reshape_28_shape_0 = const()[name = tensor<string, []>("reshape_28_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_28 = reshape(shape = reshape_28_shape_0, x = reshape_24)[name = tensor<string, []>("reshape_28")];
            tensor<string, []> scatter_5_mode_0 = const()[name = tensor<string, []>("scatter_5_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_5_axis_0 = const()[name = tensor<string, []>("scatter_5_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_5_validate_indices_0 = const()[name = tensor<string, []>("scatter_5_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_5 = scatter(axis = scatter_5_axis_0, data = reshape_28, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27, validate_indices = scatter_5_validate_indices_0)[name = tensor<string, []>("scatter_5")];
            tensor<fp32, [2, 1, 512, 16, 64]> new_cache_5_internal_tensor_assign_2 = reshape(shape = shape_17, x = scatter_5)[name = tensor<string, []>("reshape_29")];
            tensor<int32, [5]> keys_13_begin_0 = const()[name = tensor<string, []>("keys_13_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> keys_13_end_0 = const()[name = tensor<string, []>("keys_13_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> keys_13_end_mask_0 = const()[name = tensor<string, []>("keys_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> keys_13_squeeze_mask_0 = const()[name = tensor<string, []>("keys_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> keys_13 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = new_cache_5_internal_tensor_assign_2)[name = tensor<string, []>("keys_13")];
            tensor<int32, [5]> values_13_begin_0 = const()[name = tensor<string, []>("values_13_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> values_13_end_0 = const()[name = tensor<string, []>("values_13_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> values_13_end_mask_0 = const()[name = tensor<string, []>("values_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> values_13_squeeze_mask_0 = const()[name = tensor<string, []>("values_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> values_13 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = new_cache_5_internal_tensor_assign_2)[name = tensor<string, []>("values_13")];
            tensor<bool, [1, 512, 16, 64]> var_1192 = not_equal(x = keys_13, y = keys_13)[name = tensor<string, []>("op_1192")];
            tensor<fp32, [1, 512, 16, 64]> keys_15 = select(a = var_360, b = keys_13, cond = var_1192)[name = tensor<string, []>("keys_15")];
            tensor<bool, [1, 512, 16, 64]> var_1200 = not_equal(x = values_13, y = values_13)[name = tensor<string, []>("op_1200")];
            tensor<fp32, [1, 512, 16, 64]> values_15 = select(a = var_360, b = values_13, cond = var_1200)[name = tensor<string, []>("values_15")];
            tensor<int32, [4]> var_1224 = const()[name = tensor<string, []>("op_1224"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1237 = const()[name = tensor<string, []>("op_1237"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<fp32, [1, 1, 1]> var_1238 = reshape(shape = var_1237, x = position2)[name = tensor<string, []>("op_1238")];
            tensor<fp32, []> var_1255 = const()[name = tensor<string, []>("op_1255"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1, 1, 1]> valid_len_5 = add(x = var_1238, y = var_1255)[name = tensor<string, []>("valid_len_5")];
            tensor<bool, [1, 1, 512]> valid_mask_5 = less(x = k_positions_1_promoted, y = valid_len_5)[name = tensor<string, []>("valid_mask_5")];
            tensor<bool, [1, 1, 512]> causal_mask_5 = less_equal(x = k_positions_1_promoted, y = var_1238)[name = tensor<string, []>("causal_mask_5")];
            tensor<bool, [1, 1, 512]> attn_mask_9 = logical_and(x = valid_mask_5, y = causal_mask_5)[name = tensor<string, []>("attn_mask_9")];
            tensor<int32, [1]> attn_mask_11_axes_0 = const()[name = tensor<string, []>("attn_mask_11_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 1, 512]> attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor<string, []>("attn_mask_11")];
            tensor<fp32, [1]> var_1267 = const()[name = tensor<string, []>("op_1267"), val = tensor<fp32, [1]>([0x1.fffe5cp-4])];
            tensor<bool, []> var_1273_transpose_x_0 = const()[name = tensor<string, []>("op_1273_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1273_transpose_y_0 = const()[name = tensor<string, []>("op_1273_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = tensor<string, []>("transpose_22_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_23_perm_0 = const()[name = tensor<string, []>("transpose_23_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp32, [1, 16, 64, 512]> transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15)[name = tensor<string, []>("transpose_43")];
            tensor<fp32, [1, 16, 1, 64]> transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15)[name = tensor<string, []>("transpose_44")];
            tensor<fp32, [1, 16, 1, 512]> var_1273 = matmul(transpose_x = var_1273_transpose_x_0, transpose_y = var_1273_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor<string, []>("op_1273")];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_13 = mul(x = var_1273, y = var_1267)[name = tensor<string, []>("attn_weights_13")];
            tensor<bool, [1, 1, 1, 512]> var_1275 = logical_not(x = attn_mask_11)[name = tensor<string, []>("op_1275")];
            tensor<fp32, []> var_1276 = const()[name = tensor<string, []>("op_1276"), val = tensor<fp32, []>(-0x1.ff933cp+127)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_15 = select(a = var_1276, b = attn_weights_13, cond = var_1275)[name = tensor<string, []>("attn_weights_15")];
            tensor<int32, []> var_1278 = const()[name = tensor<string, []>("op_1278"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_17 = softmax(axis = var_1278, x = attn_weights_15)[name = tensor<string, []>("attn_weights_17")];
            tensor<bool, []> attn_output_5_transpose_x_0 = const()[name = tensor<string, []>("attn_output_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_5_transpose_y_0 = const()[name = tensor<string, []>("attn_output_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 16, 512, 64]> values_17 = transpose(perm = var_1224, x = values_15)[name = tensor<string, []>("transpose_45")];
            tensor<fp32, [1, 16, 1, 64]> attn_output_5 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17, y = values_17)[name = tensor<string, []>("attn_output_5")];
            tensor<int32, [4]> var_1286 = const()[name = tensor<string, []>("op_1286"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1289 = const()[name = tensor<string, []>("op_1289"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp32, [1, 1, 16, 64]> var_1287 = transpose(perm = var_1286, x = attn_output_5)[name = tensor<string, []>("transpose_42")];
            tensor<fp32, [1, 1, 1024]> input_25 = reshape(shape = var_1289, x = var_1287)[name = tensor<string, []>("input_25")];
            tensor<fp32, [1, 1, 1024]> attn_out_5 = linear(bias = linear_0_bias_0, weight = attn2_out_proj_weight, x = input_25)[name = tensor<string, []>("linear_10")];
            tensor<fp32, []> var_1295 = const()[name = tensor<string, []>("op_1295"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1]> var_1296 = add(x = position2, y = var_1295)[name = tensor<string, []>("op_1296")];
            tensor<fp32, [1, 1, 1024]> input_27 = add(x = input_23, y = attn_out_5)[name = tensor<string, []>("input_27")];
            tensor<fp32, []> var_1300 = const()[name = tensor<string, []>("op_1300"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> input_29_axes_0 = const()[name = tensor<string, []>("input_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> input_29 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias, epsilon = var_1300, gamma = norm2_2_weight, x = input_27)[name = tensor<string, []>("input_29")];
            tensor<fp32, [1, 1, 4096]> var_1308 = linear(bias = linear_3_bias_0, weight = linear2_1_weight, x = input_29)[name = tensor<string, []>("linear_11")];
            tensor<string, []> input_31_mode_0 = const()[name = tensor<string, []>("input_31_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp32, [1, 1, 4096]> input_31 = gelu(mode = input_31_mode_0, x = var_1308)[name = tensor<string, []>("input_31")];
            tensor<fp32, [1, 1, 1024]> ffn_out_5 = linear(bias = linear_0_bias_0, weight = linear2_2_weight, x = input_31)[name = tensor<string, []>("linear_12")];
            tensor<fp32, [1, 1, 1024]> input_33 = add(x = input_27, y = ffn_out_5)[name = tensor<string, []>("input_33")];
            tensor<fp32, []> var_1317 = const()[name = tensor<string, []>("op_1317"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> x_7_axes_0 = const()[name = tensor<string, []>("x_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> x_7 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias, epsilon = var_1317, gamma = norm3_1_weight, x = input_33)[name = tensor<string, []>("x_7")];
            tensor<fp32, [1, 1, 3072]> var_1349 = linear(bias = linear_1_bias_0, weight = attn3_in_proj_weight, x = x_7)[name = tensor<string, []>("linear_13")];
            tensor<int32, [5]> var_1353 = const()[name = tensor<string, []>("op_1353"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<fp32, [1, 1, 3, 16, 64]> qkv_7 = reshape(shape = var_1353, x = var_1349)[name = tensor<string, []>("qkv_7")];
            tensor<int32, [5]> q_19_begin_0 = const()[name = tensor<string, []>("q_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_19_end_0 = const()[name = tensor<string, []>("q_19_end_0"), val = tensor<int32, [5]>([1, 1, 1, 16, 64])];
            tensor<bool, [5]> q_19_end_mask_0 = const()[name = tensor<string, []>("q_19_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_19_squeeze_mask_0 = const()[name = tensor<string, []>("q_19_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> q_19 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7)[name = tensor<string, []>("q_19")];
            tensor<int32, [5]> k_13_begin_0 = const()[name = tensor<string, []>("k_13_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_13_end_0 = const()[name = tensor<string, []>("k_13_end_0"), val = tensor<int32, [5]>([1, 1, 2, 16, 64])];
            tensor<bool, [5]> k_13_end_mask_0 = const()[name = tensor<string, []>("k_13_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_13_squeeze_mask_0 = const()[name = tensor<string, []>("k_13_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> k_13 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7)[name = tensor<string, []>("k_13")];
            tensor<int32, [5]> v_7_begin_0 = const()[name = tensor<string, []>("v_7_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_7_end_0 = const()[name = tensor<string, []>("v_7_end_0"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<bool, [5]> v_7_end_mask_0 = const()[name = tensor<string, []>("v_7_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_7_squeeze_mask_0 = const()[name = tensor<string, []>("v_7_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> v_7 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7)[name = tensor<string, []>("v_7")];
            tensor<fp32, [32]> freqs_7 = const()[name = tensor<string, []>("freqs_7"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304367808)))];
            tensor<int32, [4]> var_1457 = const()[name = tensor<string, []>("op_1457"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp32, [1, 1, 1, 1]> ts_23 = reshape(shape = var_1457, x = position3)[name = tensor<string, []>("ts_23")];
            tensor<int32, [5]> var_1461 = const()[name = tensor<string, []>("op_1461"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> q_complex_7 = reshape(shape = var_1461, x = q_19)[name = tensor<string, []>("q_complex_7")];
            tensor<int32, [5]> var_1465 = const()[name = tensor<string, []>("op_1465"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> k_complex_7 = reshape(shape = var_1465, x = k_13)[name = tensor<string, []>("k_complex_7")];
            tensor<int32, [5]> var_1469_begin_0 = const()[name = tensor<string, []>("op_1469_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1469_end_0 = const()[name = tensor<string, []>("op_1469_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_1469_end_mask_0 = const()[name = tensor<string, []>("op_1469_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1469_squeeze_mask_0 = const()[name = tensor<string, []>("op_1469_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1469 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, squeeze_mask = var_1469_squeeze_mask_0, x = q_complex_7)[name = tensor<string, []>("op_1469")];
            tensor<int32, [5]> var_1477_begin_0 = const()[name = tensor<string, []>("op_1477_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_1477_end_0 = const()[name = tensor<string, []>("op_1477_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_1477_end_mask_0 = const()[name = tensor<string, []>("op_1477_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1477_squeeze_mask_0 = const()[name = tensor<string, []>("op_1477_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1477 = slice_by_index(begin = var_1477_begin_0, end = var_1477_end_0, end_mask = var_1477_end_mask_0, squeeze_mask = var_1477_squeeze_mask_0, x = q_complex_7)[name = tensor<string, []>("op_1477")];
            tensor<int32, [5]> var_1485_begin_0 = const()[name = tensor<string, []>("op_1485_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1485_end_0 = const()[name = tensor<string, []>("op_1485_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_1485_end_mask_0 = const()[name = tensor<string, []>("op_1485_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1485_squeeze_mask_0 = const()[name = tensor<string, []>("op_1485_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1485 = slice_by_index(begin = var_1485_begin_0, end = var_1485_end_0, end_mask = var_1485_end_mask_0, squeeze_mask = var_1485_squeeze_mask_0, x = k_complex_7)[name = tensor<string, []>("op_1485")];
            tensor<int32, [5]> var_1493_begin_0 = const()[name = tensor<string, []>("op_1493_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_1493_end_0 = const()[name = tensor<string, []>("op_1493_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_1493_end_mask_0 = const()[name = tensor<string, []>("op_1493_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1493_squeeze_mask_0 = const()[name = tensor<string, []>("op_1493_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1493 = slice_by_index(begin = var_1493_begin_0, end = var_1493_end_0, end_mask = var_1493_end_mask_0, squeeze_mask = var_1493_squeeze_mask_0, x = k_complex_7)[name = tensor<string, []>("op_1493")];
            tensor<fp32, [1, 1, 1, 32]> var_1499 = mul(x = freqs_7, y = ts_23)[name = tensor<string, []>("op_1499")];
            tensor<fp32, [1, 1, 1, 32]> rotr_7 = cos(x = var_1499)[name = tensor<string, []>("rotr_7")];
            tensor<fp32, [1, 1, 1, 32]> roti_7 = sin(x = var_1499)[name = tensor<string, []>("roti_7")];
            tensor<fp32, [1, 1, 16, 32]> var_1503 = mul(x = var_1469, y = rotr_7)[name = tensor<string, []>("op_1503")];
            tensor<fp32, [1, 1, 16, 32]> var_1504 = mul(x = var_1477, y = roti_7)[name = tensor<string, []>("op_1504")];
            tensor<fp32, [1, 1, 16, 32]> qor_13 = sub(x = var_1503, y = var_1504)[name = tensor<string, []>("qor_13")];
            tensor<fp32, [1, 1, 16, 32]> var_1507 = mul(x = var_1469, y = roti_7)[name = tensor<string, []>("op_1507")];
            tensor<fp32, [1, 1, 16, 32]> var_1508 = mul(x = var_1477, y = rotr_7)[name = tensor<string, []>("op_1508")];
            tensor<fp32, [1, 1, 16, 32]> qoi_13 = add(x = var_1507, y = var_1508)[name = tensor<string, []>("qoi_13")];
            tensor<fp32, [1, 1, 16, 32]> var_1511 = mul(x = var_1485, y = rotr_7)[name = tensor<string, []>("op_1511")];
            tensor<fp32, [1, 1, 16, 32]> var_1512 = mul(x = var_1493, y = roti_7)[name = tensor<string, []>("op_1512")];
            tensor<fp32, [1, 1, 16, 32]> kor_13 = sub(x = var_1511, y = var_1512)[name = tensor<string, []>("kor_13")];
            tensor<fp32, [1, 1, 16, 32]> var_1515 = mul(x = var_1485, y = roti_7)[name = tensor<string, []>("op_1515")];
            tensor<fp32, [1, 1, 16, 32]> var_1516 = mul(x = var_1493, y = rotr_7)[name = tensor<string, []>("op_1516")];
            tensor<fp32, [1, 1, 16, 32]> koi_13 = add(x = var_1515, y = var_1516)[name = tensor<string, []>("koi_13")];
            tensor<int32, []> qo_7_axis_0 = const()[name = tensor<string, []>("qo_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> qo_7 = stack(axis = qo_7_axis_0, values = (qor_13, qoi_13))[name = tensor<string, []>("qo_7")];
            tensor<int32, []> ko_7_axis_0 = const()[name = tensor<string, []>("ko_7_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> ko_7 = stack(axis = ko_7_axis_0, values = (kor_13, koi_13))[name = tensor<string, []>("ko_7")];
            tensor<int32, [4]> var_1545 = const()[name = tensor<string, []>("op_1545"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> q_21 = reshape(shape = var_1545, x = qo_7)[name = tensor<string, []>("q_21")];
            tensor<int32, [4]> var_1547 = const()[name = tensor<string, []>("op_1547"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> k_15 = reshape(shape = var_1547, x = ko_7)[name = tensor<string, []>("k_15")];
            tensor<fp32, []> _inversed_1569_y_0 = const()[name = tensor<string, []>("_inversed_1569_y_0"), val = tensor<fp32, []>(0x1p-9)];
            tensor<fp32, [1, 1, 1, 1]> _inversed_1569 = mul(x = ts_23, y = _inversed_1569_y_0)[name = tensor<string, []>("_inversed_1569")];
            tensor<fp32, [1, 1, 1, 1]> var_1570 = floor(x = _inversed_1569)[name = tensor<string, []>("op_1570")];
            tensor<fp32, []> var_1571 = const()[name = tensor<string, []>("op_1571"), val = tensor<fp32, []>(0x1p+9)];
            tensor<fp32, [1, 1, 1, 1]> var_1572 = mul(x = var_1570, y = var_1571)[name = tensor<string, []>("op_1572")];
            tensor<fp32, [1, 1, 1, 1]> write_indices_float_15 = sub(x = ts_23, y = var_1572)[name = tensor<string, []>("write_indices_float_15")];
            tensor<string, []> var_1579_dtype_0 = const()[name = tensor<string, []>("op_1579_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [4]> write_indices_7_reps_0 = const()[name = tensor<string, []>("write_indices_7_reps_0"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<int32, [1, 1, 1, 1]> var_1579 = cast(dtype = var_1579_dtype_0, x = write_indices_float_15)[name = tensor<string, []>("cast_110")];
            tensor<int32, [1, 1, 16, 64]> write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1579)[name = tensor<string, []>("write_indices_7")];
            tensor<int32, [5]> var_1587_begin_0 = const()[name = tensor<string, []>("op_1587_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1587_end_0 = const()[name = tensor<string, []>("op_1587_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> var_1587_end_mask_0 = const()[name = tensor<string, []>("op_1587_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1587_squeeze_mask_0 = const()[name = tensor<string, []>("op_1587_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_1587 = slice_by_index(begin = var_1587_begin_0, end = var_1587_end_0, end_mask = var_1587_end_mask_0, squeeze_mask = var_1587_squeeze_mask_0, x = cache3)[name = tensor<string, []>("op_1587")];
            tensor<int32, []> var_1589_axis_0 = const()[name = tensor<string, []>("op_1589_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_1589_mode_0 = const()[name = tensor<string, []>("op_1589_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_1589_validate_indices_0 = const()[name = tensor<string, []>("op_1589_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_1589 = scatter_along_axis(axis = var_1589_axis_0, data = var_1587, indices = write_indices_7, mode = var_1589_mode_0, updates = k_15, validate_indices = var_1589_validate_indices_0)[name = tensor<string, []>("op_1589")];
            tensor<int32, [5]> concat_23 = const()[name = tensor<string, []>("concat_23"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_24 = const()[name = tensor<string, []>("concat_24"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_18 = const()[name = tensor<string, []>("shape_18"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_6 = const()[name = tensor<string, []>("reduce_prod_6"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_6_start_0 = const()[name = tensor<string, []>("range_1d_6_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_6_step_0 = const()[name = tensor<string, []>("range_1d_6_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor<string, []>("range_1d_6")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor<string, []>("reshape_30")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor<string, []>("slice_by_index_6")];
            tensor<int32, [1]> reshape_31_shape_0 = const()[name = tensor<string, []>("reshape_31_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor<string, []>("reshape_31")];
            tensor<int32, [1]> reshape_32_shape_0 = const()[name = tensor<string, []>("reshape_32_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_32 = reshape(shape = reshape_32_shape_0, x = var_1589)[name = tensor<string, []>("reshape_32")];
            tensor<int32, [1]> reshape_33_shape_0 = const()[name = tensor<string, []>("reshape_33_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_33 = reshape(shape = reshape_33_shape_0, x = cache3)[name = tensor<string, []>("reshape_33")];
            tensor<string, []> scatter_6_mode_0 = const()[name = tensor<string, []>("scatter_6_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_6_axis_0 = const()[name = tensor<string, []>("scatter_6_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_6_validate_indices_0 = const()[name = tensor<string, []>("scatter_6_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_6 = scatter(axis = scatter_6_axis_0, data = reshape_33, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32, validate_indices = scatter_6_validate_indices_0)[name = tensor<string, []>("scatter_6")];
            tensor<fp32, [2, 1, 512, 16, 64]> reshape_34 = reshape(shape = shape_18, x = scatter_6)[name = tensor<string, []>("reshape_34")];
            tensor<int32, [5]> var_1597_begin_0 = const()[name = tensor<string, []>("op_1597_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1597_end_0 = const()[name = tensor<string, []>("op_1597_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> var_1597_end_mask_0 = const()[name = tensor<string, []>("op_1597_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_1597_squeeze_mask_0 = const()[name = tensor<string, []>("op_1597_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_1597 = slice_by_index(begin = var_1597_begin_0, end = var_1597_end_0, end_mask = var_1597_end_mask_0, squeeze_mask = var_1597_squeeze_mask_0, x = reshape_34)[name = tensor<string, []>("op_1597")];
            tensor<int32, []> var_1599_axis_0 = const()[name = tensor<string, []>("op_1599_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_1599_mode_0 = const()[name = tensor<string, []>("op_1599_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_1599_validate_indices_0 = const()[name = tensor<string, []>("op_1599_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_1599 = scatter_along_axis(axis = var_1599_axis_0, data = var_1597, indices = write_indices_7, mode = var_1599_mode_0, updates = v_7, validate_indices = var_1599_validate_indices_0)[name = tensor<string, []>("op_1599")];
            tensor<int32, [5]> concat_25 = const()[name = tensor<string, []>("concat_25"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_26 = const()[name = tensor<string, []>("concat_26"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_19 = const()[name = tensor<string, []>("shape_19"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_7 = const()[name = tensor<string, []>("reduce_prod_7"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_7_start_0 = const()[name = tensor<string, []>("range_1d_7_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_7_step_0 = const()[name = tensor<string, []>("range_1d_7_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor<string, []>("range_1d_7")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor<string, []>("reshape_35")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor<string, []>("slice_by_index_7")];
            tensor<int32, [1]> reshape_36_shape_0 = const()[name = tensor<string, []>("reshape_36_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor<string, []>("reshape_36")];
            tensor<int32, [1]> reshape_37_shape_0 = const()[name = tensor<string, []>("reshape_37_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_37 = reshape(shape = reshape_37_shape_0, x = var_1599)[name = tensor<string, []>("reshape_37")];
            tensor<int32, [1]> reshape_38_shape_0 = const()[name = tensor<string, []>("reshape_38_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_38 = reshape(shape = reshape_38_shape_0, x = reshape_34)[name = tensor<string, []>("reshape_38")];
            tensor<string, []> scatter_7_mode_0 = const()[name = tensor<string, []>("scatter_7_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_7_axis_0 = const()[name = tensor<string, []>("scatter_7_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_7_validate_indices_0 = const()[name = tensor<string, []>("scatter_7_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_7 = scatter(axis = scatter_7_axis_0, data = reshape_38, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37, validate_indices = scatter_7_validate_indices_0)[name = tensor<string, []>("scatter_7")];
            tensor<fp32, [2, 1, 512, 16, 64]> new_cache_7_internal_tensor_assign_2 = reshape(shape = shape_19, x = scatter_7)[name = tensor<string, []>("reshape_39")];
            tensor<int32, [5]> keys_19_begin_0 = const()[name = tensor<string, []>("keys_19_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> keys_19_end_0 = const()[name = tensor<string, []>("keys_19_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> keys_19_end_mask_0 = const()[name = tensor<string, []>("keys_19_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> keys_19_squeeze_mask_0 = const()[name = tensor<string, []>("keys_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> keys_19 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = new_cache_7_internal_tensor_assign_2)[name = tensor<string, []>("keys_19")];
            tensor<int32, [5]> values_19_begin_0 = const()[name = tensor<string, []>("values_19_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> values_19_end_0 = const()[name = tensor<string, []>("values_19_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> values_19_end_mask_0 = const()[name = tensor<string, []>("values_19_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> values_19_squeeze_mask_0 = const()[name = tensor<string, []>("values_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> values_19 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = new_cache_7_internal_tensor_assign_2)[name = tensor<string, []>("values_19")];
            tensor<bool, [1, 512, 16, 64]> var_1611 = not_equal(x = keys_19, y = keys_19)[name = tensor<string, []>("op_1611")];
            tensor<fp32, [1, 512, 16, 64]> keys_21 = select(a = var_360, b = keys_19, cond = var_1611)[name = tensor<string, []>("keys_21")];
            tensor<bool, [1, 512, 16, 64]> var_1619 = not_equal(x = values_19, y = values_19)[name = tensor<string, []>("op_1619")];
            tensor<fp32, [1, 512, 16, 64]> values_21 = select(a = var_360, b = values_19, cond = var_1619)[name = tensor<string, []>("values_21")];
            tensor<int32, [4]> var_1643 = const()[name = tensor<string, []>("op_1643"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1656 = const()[name = tensor<string, []>("op_1656"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<fp32, [1, 1, 1]> var_1657 = reshape(shape = var_1656, x = position3)[name = tensor<string, []>("op_1657")];
            tensor<fp32, []> var_1674 = const()[name = tensor<string, []>("op_1674"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1, 1, 1]> valid_len_7 = add(x = var_1657, y = var_1674)[name = tensor<string, []>("valid_len_7")];
            tensor<bool, [1, 1, 512]> valid_mask_7 = less(x = k_positions_1_promoted, y = valid_len_7)[name = tensor<string, []>("valid_mask_7")];
            tensor<bool, [1, 1, 512]> causal_mask_7 = less_equal(x = k_positions_1_promoted, y = var_1657)[name = tensor<string, []>("causal_mask_7")];
            tensor<bool, [1, 1, 512]> attn_mask_13 = logical_and(x = valid_mask_7, y = causal_mask_7)[name = tensor<string, []>("attn_mask_13")];
            tensor<int32, [1]> attn_mask_15_axes_0 = const()[name = tensor<string, []>("attn_mask_15_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 1, 512]> attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor<string, []>("attn_mask_15")];
            tensor<fp32, [1]> var_1686 = const()[name = tensor<string, []>("op_1686"), val = tensor<fp32, [1]>([0x1.fffe5cp-4])];
            tensor<bool, []> var_1692_transpose_x_0 = const()[name = tensor<string, []>("op_1692_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1692_transpose_y_0 = const()[name = tensor<string, []>("op_1692_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]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_25_perm_0 = const()[name = tensor<string, []>("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp32, [1, 16, 64, 512]> transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21)[name = tensor<string, []>("transpose_39")];
            tensor<fp32, [1, 16, 1, 64]> transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21)[name = tensor<string, []>("transpose_40")];
            tensor<fp32, [1, 16, 1, 512]> var_1692 = matmul(transpose_x = var_1692_transpose_x_0, transpose_y = var_1692_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor<string, []>("op_1692")];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_19 = mul(x = var_1692, y = var_1686)[name = tensor<string, []>("attn_weights_19")];
            tensor<bool, [1, 1, 1, 512]> var_1694 = logical_not(x = attn_mask_15)[name = tensor<string, []>("op_1694")];
            tensor<fp32, []> var_1695 = const()[name = tensor<string, []>("op_1695"), val = tensor<fp32, []>(-0x1.ff933cp+127)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_21 = select(a = var_1695, b = attn_weights_19, cond = var_1694)[name = tensor<string, []>("attn_weights_21")];
            tensor<int32, []> var_1697 = const()[name = tensor<string, []>("op_1697"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_23 = softmax(axis = var_1697, x = attn_weights_21)[name = tensor<string, []>("attn_weights_23")];
            tensor<bool, []> attn_output_7_transpose_x_0 = const()[name = tensor<string, []>("attn_output_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_7_transpose_y_0 = const()[name = tensor<string, []>("attn_output_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 16, 512, 64]> values_23 = transpose(perm = var_1643, x = values_21)[name = tensor<string, []>("transpose_41")];
            tensor<fp32, [1, 16, 1, 64]> attn_output_7 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23, y = values_23)[name = tensor<string, []>("attn_output_7")];
            tensor<int32, [4]> var_1705 = const()[name = tensor<string, []>("op_1705"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1708 = const()[name = tensor<string, []>("op_1708"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp32, [1, 1, 16, 64]> var_1706 = transpose(perm = var_1705, x = attn_output_7)[name = tensor<string, []>("transpose_38")];
            tensor<fp32, [1, 1, 1024]> input_35 = reshape(shape = var_1708, x = var_1706)[name = tensor<string, []>("input_35")];
            tensor<fp32, [1, 1, 1024]> attn_out_7 = linear(bias = linear_0_bias_0, weight = attn3_out_proj_weight, x = input_35)[name = tensor<string, []>("linear_14")];
            tensor<fp32, []> var_1714 = const()[name = tensor<string, []>("op_1714"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1]> var_1715 = add(x = position3, y = var_1714)[name = tensor<string, []>("op_1715")];
            tensor<fp32, [1, 1, 1024]> input_37 = add(x = input_33, y = attn_out_7)[name = tensor<string, []>("input_37")];
            tensor<fp32, []> var_1719 = const()[name = tensor<string, []>("op_1719"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> input_39_axes_0 = const()[name = tensor<string, []>("input_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> input_39 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias, epsilon = var_1719, gamma = norm3_2_weight, x = input_37)[name = tensor<string, []>("input_39")];
            tensor<fp32, [1, 1, 4096]> var_1727 = linear(bias = linear_3_bias_0, weight = linear3_1_weight, x = input_39)[name = tensor<string, []>("linear_15")];
            tensor<string, []> input_41_mode_0 = const()[name = tensor<string, []>("input_41_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp32, [1, 1, 4096]> input_41 = gelu(mode = input_41_mode_0, x = var_1727)[name = tensor<string, []>("input_41")];
            tensor<fp32, [1, 1, 1024]> ffn_out_7 = linear(bias = linear_0_bias_0, weight = linear3_2_weight, x = input_41)[name = tensor<string, []>("linear_16")];
            tensor<fp32, [1, 1, 1024]> input_43 = add(x = input_37, y = ffn_out_7)[name = tensor<string, []>("input_43")];
            tensor<fp32, []> var_1736 = const()[name = tensor<string, []>("op_1736"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> x_9_axes_0 = const()[name = tensor<string, []>("x_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> x_9 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias, epsilon = var_1736, gamma = norm4_1_weight, x = input_43)[name = tensor<string, []>("x_9")];
            tensor<fp32, [1, 1, 3072]> var_1768 = linear(bias = linear_1_bias_0, weight = attn4_in_proj_weight, x = x_9)[name = tensor<string, []>("linear_17")];
            tensor<int32, [5]> var_1772 = const()[name = tensor<string, []>("op_1772"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<fp32, [1, 1, 3, 16, 64]> qkv_9 = reshape(shape = var_1772, x = var_1768)[name = tensor<string, []>("qkv_9")];
            tensor<int32, [5]> q_25_begin_0 = const()[name = tensor<string, []>("q_25_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_25_end_0 = const()[name = tensor<string, []>("q_25_end_0"), val = tensor<int32, [5]>([1, 1, 1, 16, 64])];
            tensor<bool, [5]> q_25_end_mask_0 = const()[name = tensor<string, []>("q_25_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_25_squeeze_mask_0 = const()[name = tensor<string, []>("q_25_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> q_25 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9)[name = tensor<string, []>("q_25")];
            tensor<int32, [5]> k_17_begin_0 = const()[name = tensor<string, []>("k_17_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_17_end_0 = const()[name = tensor<string, []>("k_17_end_0"), val = tensor<int32, [5]>([1, 1, 2, 16, 64])];
            tensor<bool, [5]> k_17_end_mask_0 = const()[name = tensor<string, []>("k_17_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_17_squeeze_mask_0 = const()[name = tensor<string, []>("k_17_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> k_17 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9)[name = tensor<string, []>("k_17")];
            tensor<int32, [5]> v_9_begin_0 = const()[name = tensor<string, []>("v_9_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_9_end_0 = const()[name = tensor<string, []>("v_9_end_0"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<bool, [5]> v_9_end_mask_0 = const()[name = tensor<string, []>("v_9_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_9_squeeze_mask_0 = const()[name = tensor<string, []>("v_9_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> v_9 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9)[name = tensor<string, []>("v_9")];
            tensor<fp32, [32]> freqs_9 = const()[name = tensor<string, []>("freqs_9"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304368000)))];
            tensor<int32, [4]> var_1876 = const()[name = tensor<string, []>("op_1876"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp32, [1, 1, 1, 1]> ts_29 = reshape(shape = var_1876, x = position4)[name = tensor<string, []>("ts_29")];
            tensor<int32, [5]> var_1880 = const()[name = tensor<string, []>("op_1880"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> q_complex_9 = reshape(shape = var_1880, x = q_25)[name = tensor<string, []>("q_complex_9")];
            tensor<int32, [5]> var_1884 = const()[name = tensor<string, []>("op_1884"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> k_complex_9 = reshape(shape = var_1884, x = k_17)[name = tensor<string, []>("k_complex_9")];
            tensor<int32, [5]> var_1888_begin_0 = const()[name = tensor<string, []>("op_1888_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1888_end_0 = const()[name = tensor<string, []>("op_1888_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_1888_end_mask_0 = const()[name = tensor<string, []>("op_1888_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1888_squeeze_mask_0 = const()[name = tensor<string, []>("op_1888_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1888 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = q_complex_9)[name = tensor<string, []>("op_1888")];
            tensor<int32, [5]> var_1896_begin_0 = const()[name = tensor<string, []>("op_1896_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_1896_end_0 = const()[name = tensor<string, []>("op_1896_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_1896_end_mask_0 = const()[name = tensor<string, []>("op_1896_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1896_squeeze_mask_0 = const()[name = tensor<string, []>("op_1896_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1896 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, squeeze_mask = var_1896_squeeze_mask_0, x = q_complex_9)[name = tensor<string, []>("op_1896")];
            tensor<int32, [5]> var_1904_begin_0 = const()[name = tensor<string, []>("op_1904_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_1904_end_0 = const()[name = tensor<string, []>("op_1904_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_1904_end_mask_0 = const()[name = tensor<string, []>("op_1904_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1904_squeeze_mask_0 = const()[name = tensor<string, []>("op_1904_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1904 = slice_by_index(begin = var_1904_begin_0, end = var_1904_end_0, end_mask = var_1904_end_mask_0, squeeze_mask = var_1904_squeeze_mask_0, x = k_complex_9)[name = tensor<string, []>("op_1904")];
            tensor<int32, [5]> var_1912_begin_0 = const()[name = tensor<string, []>("op_1912_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_1912_end_0 = const()[name = tensor<string, []>("op_1912_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_1912_end_mask_0 = const()[name = tensor<string, []>("op_1912_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_1912_squeeze_mask_0 = const()[name = tensor<string, []>("op_1912_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_1912 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, squeeze_mask = var_1912_squeeze_mask_0, x = k_complex_9)[name = tensor<string, []>("op_1912")];
            tensor<fp32, [1, 1, 1, 32]> var_1918 = mul(x = freqs_9, y = ts_29)[name = tensor<string, []>("op_1918")];
            tensor<fp32, [1, 1, 1, 32]> rotr_9 = cos(x = var_1918)[name = tensor<string, []>("rotr_9")];
            tensor<fp32, [1, 1, 1, 32]> roti_9 = sin(x = var_1918)[name = tensor<string, []>("roti_9")];
            tensor<fp32, [1, 1, 16, 32]> var_1922 = mul(x = var_1888, y = rotr_9)[name = tensor<string, []>("op_1922")];
            tensor<fp32, [1, 1, 16, 32]> var_1923 = mul(x = var_1896, y = roti_9)[name = tensor<string, []>("op_1923")];
            tensor<fp32, [1, 1, 16, 32]> qor_17 = sub(x = var_1922, y = var_1923)[name = tensor<string, []>("qor_17")];
            tensor<fp32, [1, 1, 16, 32]> var_1926 = mul(x = var_1888, y = roti_9)[name = tensor<string, []>("op_1926")];
            tensor<fp32, [1, 1, 16, 32]> var_1927 = mul(x = var_1896, y = rotr_9)[name = tensor<string, []>("op_1927")];
            tensor<fp32, [1, 1, 16, 32]> qoi_17 = add(x = var_1926, y = var_1927)[name = tensor<string, []>("qoi_17")];
            tensor<fp32, [1, 1, 16, 32]> var_1930 = mul(x = var_1904, y = rotr_9)[name = tensor<string, []>("op_1930")];
            tensor<fp32, [1, 1, 16, 32]> var_1931 = mul(x = var_1912, y = roti_9)[name = tensor<string, []>("op_1931")];
            tensor<fp32, [1, 1, 16, 32]> kor_17 = sub(x = var_1930, y = var_1931)[name = tensor<string, []>("kor_17")];
            tensor<fp32, [1, 1, 16, 32]> var_1934 = mul(x = var_1904, y = roti_9)[name = tensor<string, []>("op_1934")];
            tensor<fp32, [1, 1, 16, 32]> var_1935 = mul(x = var_1912, y = rotr_9)[name = tensor<string, []>("op_1935")];
            tensor<fp32, [1, 1, 16, 32]> koi_17 = add(x = var_1934, y = var_1935)[name = tensor<string, []>("koi_17")];
            tensor<int32, []> qo_9_axis_0 = const()[name = tensor<string, []>("qo_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> qo_9 = stack(axis = qo_9_axis_0, values = (qor_17, qoi_17))[name = tensor<string, []>("qo_9")];
            tensor<int32, []> ko_9_axis_0 = const()[name = tensor<string, []>("ko_9_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> ko_9 = stack(axis = ko_9_axis_0, values = (kor_17, koi_17))[name = tensor<string, []>("ko_9")];
            tensor<int32, [4]> var_1964 = const()[name = tensor<string, []>("op_1964"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> q_27 = reshape(shape = var_1964, x = qo_9)[name = tensor<string, []>("q_27")];
            tensor<int32, [4]> var_1966 = const()[name = tensor<string, []>("op_1966"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> k_19 = reshape(shape = var_1966, x = ko_9)[name = tensor<string, []>("k_19")];
            tensor<fp32, []> _inversed_1988_y_0 = const()[name = tensor<string, []>("_inversed_1988_y_0"), val = tensor<fp32, []>(0x1p-9)];
            tensor<fp32, [1, 1, 1, 1]> _inversed_1988 = mul(x = ts_29, y = _inversed_1988_y_0)[name = tensor<string, []>("_inversed_1988")];
            tensor<fp32, [1, 1, 1, 1]> var_1989 = floor(x = _inversed_1988)[name = tensor<string, []>("op_1989")];
            tensor<fp32, []> var_1990 = const()[name = tensor<string, []>("op_1990"), val = tensor<fp32, []>(0x1p+9)];
            tensor<fp32, [1, 1, 1, 1]> var_1991 = mul(x = var_1989, y = var_1990)[name = tensor<string, []>("op_1991")];
            tensor<fp32, [1, 1, 1, 1]> write_indices_float_19 = sub(x = ts_29, y = var_1991)[name = tensor<string, []>("write_indices_float_19")];
            tensor<string, []> var_1998_dtype_0 = const()[name = tensor<string, []>("op_1998_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [4]> write_indices_9_reps_0 = const()[name = tensor<string, []>("write_indices_9_reps_0"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<int32, [1, 1, 1, 1]> var_1998 = cast(dtype = var_1998_dtype_0, x = write_indices_float_19)[name = tensor<string, []>("cast_109")];
            tensor<int32, [1, 1, 16, 64]> write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_1998)[name = tensor<string, []>("write_indices_9")];
            tensor<int32, [5]> var_2006_begin_0 = const()[name = tensor<string, []>("op_2006_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2006_end_0 = const()[name = tensor<string, []>("op_2006_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> var_2006_end_mask_0 = const()[name = tensor<string, []>("op_2006_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2006_squeeze_mask_0 = const()[name = tensor<string, []>("op_2006_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_2006 = slice_by_index(begin = var_2006_begin_0, end = var_2006_end_0, end_mask = var_2006_end_mask_0, squeeze_mask = var_2006_squeeze_mask_0, x = cache4)[name = tensor<string, []>("op_2006")];
            tensor<int32, []> var_2008_axis_0 = const()[name = tensor<string, []>("op_2008_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_2008_mode_0 = const()[name = tensor<string, []>("op_2008_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_2008_validate_indices_0 = const()[name = tensor<string, []>("op_2008_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_2008 = scatter_along_axis(axis = var_2008_axis_0, data = var_2006, indices = write_indices_9, mode = var_2008_mode_0, updates = k_19, validate_indices = var_2008_validate_indices_0)[name = tensor<string, []>("op_2008")];
            tensor<int32, [5]> concat_30 = const()[name = tensor<string, []>("concat_30"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_31 = const()[name = tensor<string, []>("concat_31"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_20 = const()[name = tensor<string, []>("shape_20"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_8 = const()[name = tensor<string, []>("reduce_prod_8"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_8_start_0 = const()[name = tensor<string, []>("range_1d_8_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_8_step_0 = const()[name = tensor<string, []>("range_1d_8_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor<string, []>("range_1d_8")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor<string, []>("reshape_40")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor<string, []>("slice_by_index_8")];
            tensor<int32, [1]> reshape_41_shape_0 = const()[name = tensor<string, []>("reshape_41_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor<string, []>("reshape_41")];
            tensor<int32, [1]> reshape_42_shape_0 = const()[name = tensor<string, []>("reshape_42_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_42 = reshape(shape = reshape_42_shape_0, x = var_2008)[name = tensor<string, []>("reshape_42")];
            tensor<int32, [1]> reshape_43_shape_0 = const()[name = tensor<string, []>("reshape_43_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_43 = reshape(shape = reshape_43_shape_0, x = cache4)[name = tensor<string, []>("reshape_43")];
            tensor<string, []> scatter_8_mode_0 = const()[name = tensor<string, []>("scatter_8_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_8_axis_0 = const()[name = tensor<string, []>("scatter_8_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_8_validate_indices_0 = const()[name = tensor<string, []>("scatter_8_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_8 = scatter(axis = scatter_8_axis_0, data = reshape_43, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42, validate_indices = scatter_8_validate_indices_0)[name = tensor<string, []>("scatter_8")];
            tensor<fp32, [2, 1, 512, 16, 64]> reshape_44 = reshape(shape = shape_20, x = scatter_8)[name = tensor<string, []>("reshape_44")];
            tensor<int32, [5]> var_2016_begin_0 = const()[name = tensor<string, []>("op_2016_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2016_end_0 = const()[name = tensor<string, []>("op_2016_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> var_2016_end_mask_0 = const()[name = tensor<string, []>("op_2016_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2016_squeeze_mask_0 = const()[name = tensor<string, []>("op_2016_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_2016 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, squeeze_mask = var_2016_squeeze_mask_0, x = reshape_44)[name = tensor<string, []>("op_2016")];
            tensor<int32, []> var_2018_axis_0 = const()[name = tensor<string, []>("op_2018_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_2018_mode_0 = const()[name = tensor<string, []>("op_2018_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_2018_validate_indices_0 = const()[name = tensor<string, []>("op_2018_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_2018 = scatter_along_axis(axis = var_2018_axis_0, data = var_2016, indices = write_indices_9, mode = var_2018_mode_0, updates = v_9, validate_indices = var_2018_validate_indices_0)[name = tensor<string, []>("op_2018")];
            tensor<int32, [5]> concat_32 = const()[name = tensor<string, []>("concat_32"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_33 = const()[name = tensor<string, []>("concat_33"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_21 = const()[name = tensor<string, []>("shape_21"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_9 = const()[name = tensor<string, []>("reduce_prod_9"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_9_start_0 = const()[name = tensor<string, []>("range_1d_9_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_9_step_0 = const()[name = tensor<string, []>("range_1d_9_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor<string, []>("range_1d_9")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor<string, []>("reshape_45")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor<string, []>("slice_by_index_9")];
            tensor<int32, [1]> reshape_46_shape_0 = const()[name = tensor<string, []>("reshape_46_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor<string, []>("reshape_46")];
            tensor<int32, [1]> reshape_47_shape_0 = const()[name = tensor<string, []>("reshape_47_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_47 = reshape(shape = reshape_47_shape_0, x = var_2018)[name = tensor<string, []>("reshape_47")];
            tensor<int32, [1]> reshape_48_shape_0 = const()[name = tensor<string, []>("reshape_48_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_48 = reshape(shape = reshape_48_shape_0, x = reshape_44)[name = tensor<string, []>("reshape_48")];
            tensor<string, []> scatter_9_mode_0 = const()[name = tensor<string, []>("scatter_9_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_9_axis_0 = const()[name = tensor<string, []>("scatter_9_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_9_validate_indices_0 = const()[name = tensor<string, []>("scatter_9_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_9 = scatter(axis = scatter_9_axis_0, data = reshape_48, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47, validate_indices = scatter_9_validate_indices_0)[name = tensor<string, []>("scatter_9")];
            tensor<fp32, [2, 1, 512, 16, 64]> new_cache_9_internal_tensor_assign_2 = reshape(shape = shape_21, x = scatter_9)[name = tensor<string, []>("reshape_49")];
            tensor<int32, [5]> keys_25_begin_0 = const()[name = tensor<string, []>("keys_25_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> keys_25_end_0 = const()[name = tensor<string, []>("keys_25_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> keys_25_end_mask_0 = const()[name = tensor<string, []>("keys_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> keys_25_squeeze_mask_0 = const()[name = tensor<string, []>("keys_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> keys_25 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = new_cache_9_internal_tensor_assign_2)[name = tensor<string, []>("keys_25")];
            tensor<int32, [5]> values_25_begin_0 = const()[name = tensor<string, []>("values_25_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> values_25_end_0 = const()[name = tensor<string, []>("values_25_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> values_25_end_mask_0 = const()[name = tensor<string, []>("values_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> values_25_squeeze_mask_0 = const()[name = tensor<string, []>("values_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> values_25 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = new_cache_9_internal_tensor_assign_2)[name = tensor<string, []>("values_25")];
            tensor<bool, [1, 512, 16, 64]> var_2030 = not_equal(x = keys_25, y = keys_25)[name = tensor<string, []>("op_2030")];
            tensor<fp32, [1, 512, 16, 64]> keys_27 = select(a = var_360, b = keys_25, cond = var_2030)[name = tensor<string, []>("keys_27")];
            tensor<bool, [1, 512, 16, 64]> var_2038 = not_equal(x = values_25, y = values_25)[name = tensor<string, []>("op_2038")];
            tensor<fp32, [1, 512, 16, 64]> values_27 = select(a = var_360, b = values_25, cond = var_2038)[name = tensor<string, []>("values_27")];
            tensor<int32, [4]> var_2062 = const()[name = tensor<string, []>("op_2062"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2075 = const()[name = tensor<string, []>("op_2075"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<fp32, [1, 1, 1]> var_2076 = reshape(shape = var_2075, x = position4)[name = tensor<string, []>("op_2076")];
            tensor<fp32, []> var_2093 = const()[name = tensor<string, []>("op_2093"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1, 1, 1]> valid_len_9 = add(x = var_2076, y = var_2093)[name = tensor<string, []>("valid_len_9")];
            tensor<bool, [1, 1, 512]> valid_mask_9 = less(x = k_positions_1_promoted, y = valid_len_9)[name = tensor<string, []>("valid_mask_9")];
            tensor<bool, [1, 1, 512]> causal_mask_9 = less_equal(x = k_positions_1_promoted, y = var_2076)[name = tensor<string, []>("causal_mask_9")];
            tensor<bool, [1, 1, 512]> attn_mask_17 = logical_and(x = valid_mask_9, y = causal_mask_9)[name = tensor<string, []>("attn_mask_17")];
            tensor<int32, [1]> attn_mask_19_axes_0 = const()[name = tensor<string, []>("attn_mask_19_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 1, 512]> attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor<string, []>("attn_mask_19")];
            tensor<fp32, [1]> var_2105 = const()[name = tensor<string, []>("op_2105"), val = tensor<fp32, [1]>([0x1.fffe5cp-4])];
            tensor<bool, []> var_2111_transpose_x_0 = const()[name = tensor<string, []>("op_2111_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2111_transpose_y_0 = const()[name = tensor<string, []>("op_2111_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]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_27_perm_0 = const()[name = tensor<string, []>("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp32, [1, 16, 64, 512]> transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27)[name = tensor<string, []>("transpose_35")];
            tensor<fp32, [1, 16, 1, 64]> transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27)[name = tensor<string, []>("transpose_36")];
            tensor<fp32, [1, 16, 1, 512]> var_2111 = matmul(transpose_x = var_2111_transpose_x_0, transpose_y = var_2111_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor<string, []>("op_2111")];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_25 = mul(x = var_2111, y = var_2105)[name = tensor<string, []>("attn_weights_25")];
            tensor<bool, [1, 1, 1, 512]> var_2113 = logical_not(x = attn_mask_19)[name = tensor<string, []>("op_2113")];
            tensor<fp32, []> var_2114 = const()[name = tensor<string, []>("op_2114"), val = tensor<fp32, []>(-0x1.ff933cp+127)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_27 = select(a = var_2114, b = attn_weights_25, cond = var_2113)[name = tensor<string, []>("attn_weights_27")];
            tensor<int32, []> var_2116 = const()[name = tensor<string, []>("op_2116"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_29 = softmax(axis = var_2116, x = attn_weights_27)[name = tensor<string, []>("attn_weights_29")];
            tensor<bool, []> attn_output_9_transpose_x_0 = const()[name = tensor<string, []>("attn_output_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_9_transpose_y_0 = const()[name = tensor<string, []>("attn_output_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 16, 512, 64]> values_29 = transpose(perm = var_2062, x = values_27)[name = tensor<string, []>("transpose_37")];
            tensor<fp32, [1, 16, 1, 64]> attn_output_9 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29, y = values_29)[name = tensor<string, []>("attn_output_9")];
            tensor<int32, [4]> var_2124 = const()[name = tensor<string, []>("op_2124"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2127 = const()[name = tensor<string, []>("op_2127"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp32, [1, 1, 16, 64]> var_2125 = transpose(perm = var_2124, x = attn_output_9)[name = tensor<string, []>("transpose_34")];
            tensor<fp32, [1, 1, 1024]> input_45 = reshape(shape = var_2127, x = var_2125)[name = tensor<string, []>("input_45")];
            tensor<fp32, [1, 1, 1024]> attn_out_9 = linear(bias = linear_0_bias_0, weight = attn4_out_proj_weight, x = input_45)[name = tensor<string, []>("linear_18")];
            tensor<fp32, []> var_2133 = const()[name = tensor<string, []>("op_2133"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1]> var_2134 = add(x = position4, y = var_2133)[name = tensor<string, []>("op_2134")];
            tensor<fp32, [1, 1, 1024]> input_47 = add(x = input_43, y = attn_out_9)[name = tensor<string, []>("input_47")];
            tensor<fp32, []> var_2138 = const()[name = tensor<string, []>("op_2138"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> input_49_axes_0 = const()[name = tensor<string, []>("input_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> input_49 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias, epsilon = var_2138, gamma = norm4_2_weight, x = input_47)[name = tensor<string, []>("input_49")];
            tensor<fp32, [1, 1, 4096]> var_2146 = linear(bias = linear_3_bias_0, weight = linear4_1_weight, x = input_49)[name = tensor<string, []>("linear_19")];
            tensor<string, []> input_51_mode_0 = const()[name = tensor<string, []>("input_51_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp32, [1, 1, 4096]> input_51 = gelu(mode = input_51_mode_0, x = var_2146)[name = tensor<string, []>("input_51")];
            tensor<fp32, [1, 1, 1024]> ffn_out_9 = linear(bias = linear_0_bias_0, weight = linear4_2_weight, x = input_51)[name = tensor<string, []>("linear_20")];
            tensor<fp32, [1, 1, 1024]> input_53 = add(x = input_47, y = ffn_out_9)[name = tensor<string, []>("input_53")];
            tensor<fp32, []> var_2155 = const()[name = tensor<string, []>("op_2155"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> x_axes_0 = const()[name = tensor<string, []>("x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> x = layer_norm(axes = x_axes_0, beta = norm5_1_bias, epsilon = var_2155, gamma = norm5_1_weight, x = input_53)[name = tensor<string, []>("x")];
            tensor<fp32, [1, 1, 3072]> var_2187 = linear(bias = linear_1_bias_0, weight = attn5_in_proj_weight, x = x)[name = tensor<string, []>("linear_21")];
            tensor<int32, [5]> var_2191 = const()[name = tensor<string, []>("op_2191"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<fp32, [1, 1, 3, 16, 64]> qkv = reshape(shape = var_2191, x = var_2187)[name = tensor<string, []>("qkv")];
            tensor<int32, [5]> q_31_begin_0 = const()[name = tensor<string, []>("q_31_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> q_31_end_0 = const()[name = tensor<string, []>("q_31_end_0"), val = tensor<int32, [5]>([1, 1, 1, 16, 64])];
            tensor<bool, [5]> q_31_end_mask_0 = const()[name = tensor<string, []>("q_31_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> q_31_squeeze_mask_0 = const()[name = tensor<string, []>("q_31_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> q_31 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv)[name = tensor<string, []>("q_31")];
            tensor<int32, [5]> k_21_begin_0 = const()[name = tensor<string, []>("k_21_begin_0"), val = tensor<int32, [5]>([0, 0, 1, 0, 0])];
            tensor<int32, [5]> k_21_end_0 = const()[name = tensor<string, []>("k_21_end_0"), val = tensor<int32, [5]>([1, 1, 2, 16, 64])];
            tensor<bool, [5]> k_21_end_mask_0 = const()[name = tensor<string, []>("k_21_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> k_21_squeeze_mask_0 = const()[name = tensor<string, []>("k_21_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> k_21 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv)[name = tensor<string, []>("k_21")];
            tensor<int32, [5]> v_begin_0 = const()[name = tensor<string, []>("v_begin_0"), val = tensor<int32, [5]>([0, 0, 2, 0, 0])];
            tensor<int32, [5]> v_end_0 = const()[name = tensor<string, []>("v_end_0"), val = tensor<int32, [5]>([1, 1, 3, 16, 64])];
            tensor<bool, [5]> v_end_mask_0 = const()[name = tensor<string, []>("v_end_mask_0"), val = tensor<bool, [5]>([true, true, false, true, true])];
            tensor<bool, [5]> v_squeeze_mask_0 = const()[name = tensor<string, []>("v_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, true, false, false])];
            tensor<fp32, [1, 1, 16, 64]> v = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv)[name = tensor<string, []>("v")];
            tensor<fp32, [32]> freqs = const()[name = tensor<string, []>("freqs"), val = tensor<fp32, [32]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304368192)))];
            tensor<int32, [4]> var_2295 = const()[name = tensor<string, []>("op_2295"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<fp32, [1, 1, 1, 1]> ts = reshape(shape = var_2295, x = position5)[name = tensor<string, []>("ts")];
            tensor<int32, [5]> var_2299 = const()[name = tensor<string, []>("op_2299"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> q_complex = reshape(shape = var_2299, x = q_31)[name = tensor<string, []>("q_complex")];
            tensor<int32, [5]> var_2303 = const()[name = tensor<string, []>("op_2303"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<fp32, [1, 1, 16, 32, 2]> k_complex = reshape(shape = var_2303, x = k_21)[name = tensor<string, []>("k_complex")];
            tensor<int32, [5]> var_2307_begin_0 = const()[name = tensor<string, []>("op_2307_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2307_end_0 = const()[name = tensor<string, []>("op_2307_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_2307_end_mask_0 = const()[name = tensor<string, []>("op_2307_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_2307_squeeze_mask_0 = const()[name = tensor<string, []>("op_2307_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_2307 = slice_by_index(begin = var_2307_begin_0, end = var_2307_end_0, end_mask = var_2307_end_mask_0, squeeze_mask = var_2307_squeeze_mask_0, x = q_complex)[name = tensor<string, []>("op_2307")];
            tensor<int32, [5]> var_2315_begin_0 = const()[name = tensor<string, []>("op_2315_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_2315_end_0 = const()[name = tensor<string, []>("op_2315_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_2315_end_mask_0 = const()[name = tensor<string, []>("op_2315_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_2315_squeeze_mask_0 = const()[name = tensor<string, []>("op_2315_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_2315 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, squeeze_mask = var_2315_squeeze_mask_0, x = q_complex)[name = tensor<string, []>("op_2315")];
            tensor<int32, [5]> var_2323_begin_0 = const()[name = tensor<string, []>("op_2323_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2323_end_0 = const()[name = tensor<string, []>("op_2323_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 1])];
            tensor<bool, [5]> var_2323_end_mask_0 = const()[name = tensor<string, []>("op_2323_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_2323_squeeze_mask_0 = const()[name = tensor<string, []>("op_2323_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_2323 = slice_by_index(begin = var_2323_begin_0, end = var_2323_end_0, end_mask = var_2323_end_mask_0, squeeze_mask = var_2323_squeeze_mask_0, x = k_complex)[name = tensor<string, []>("op_2323")];
            tensor<int32, [5]> var_2331_begin_0 = const()[name = tensor<string, []>("op_2331_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 1])];
            tensor<int32, [5]> var_2331_end_0 = const()[name = tensor<string, []>("op_2331_end_0"), val = tensor<int32, [5]>([1, 1, 16, 32, 2])];
            tensor<bool, [5]> var_2331_end_mask_0 = const()[name = tensor<string, []>("op_2331_end_mask_0"), val = tensor<bool, [5]>([true, true, true, true, false])];
            tensor<bool, [5]> var_2331_squeeze_mask_0 = const()[name = tensor<string, []>("op_2331_squeeze_mask_0"), val = tensor<bool, [5]>([false, false, false, false, true])];
            tensor<fp32, [1, 1, 16, 32]> var_2331 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, squeeze_mask = var_2331_squeeze_mask_0, x = k_complex)[name = tensor<string, []>("op_2331")];
            tensor<fp32, [1, 1, 1, 32]> var_2337 = mul(x = freqs, y = ts)[name = tensor<string, []>("op_2337")];
            tensor<fp32, [1, 1, 1, 32]> rotr = cos(x = var_2337)[name = tensor<string, []>("rotr")];
            tensor<fp32, [1, 1, 1, 32]> roti = sin(x = var_2337)[name = tensor<string, []>("roti")];
            tensor<fp32, [1, 1, 16, 32]> var_2341 = mul(x = var_2307, y = rotr)[name = tensor<string, []>("op_2341")];
            tensor<fp32, [1, 1, 16, 32]> var_2342 = mul(x = var_2315, y = roti)[name = tensor<string, []>("op_2342")];
            tensor<fp32, [1, 1, 16, 32]> qor_21 = sub(x = var_2341, y = var_2342)[name = tensor<string, []>("qor_21")];
            tensor<fp32, [1, 1, 16, 32]> var_2345 = mul(x = var_2307, y = roti)[name = tensor<string, []>("op_2345")];
            tensor<fp32, [1, 1, 16, 32]> var_2346 = mul(x = var_2315, y = rotr)[name = tensor<string, []>("op_2346")];
            tensor<fp32, [1, 1, 16, 32]> qoi_21 = add(x = var_2345, y = var_2346)[name = tensor<string, []>("qoi_21")];
            tensor<fp32, [1, 1, 16, 32]> var_2349 = mul(x = var_2323, y = rotr)[name = tensor<string, []>("op_2349")];
            tensor<fp32, [1, 1, 16, 32]> var_2350 = mul(x = var_2331, y = roti)[name = tensor<string, []>("op_2350")];
            tensor<fp32, [1, 1, 16, 32]> kor_21 = sub(x = var_2349, y = var_2350)[name = tensor<string, []>("kor_21")];
            tensor<fp32, [1, 1, 16, 32]> var_2353 = mul(x = var_2323, y = roti)[name = tensor<string, []>("op_2353")];
            tensor<fp32, [1, 1, 16, 32]> var_2354 = mul(x = var_2331, y = rotr)[name = tensor<string, []>("op_2354")];
            tensor<fp32, [1, 1, 16, 32]> koi_21 = add(x = var_2353, y = var_2354)[name = tensor<string, []>("koi_21")];
            tensor<int32, []> qo_axis_0 = const()[name = tensor<string, []>("qo_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> qo = stack(axis = qo_axis_0, values = (qor_21, qoi_21))[name = tensor<string, []>("qo")];
            tensor<int32, []> ko_axis_0 = const()[name = tensor<string, []>("ko_axis_0"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 1, 16, 32, 2]> ko = stack(axis = ko_axis_0, values = (kor_21, koi_21))[name = tensor<string, []>("ko")];
            tensor<int32, [4]> var_2383 = const()[name = tensor<string, []>("op_2383"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> q_33 = reshape(shape = var_2383, x = qo)[name = tensor<string, []>("q_33")];
            tensor<int32, [4]> var_2385 = const()[name = tensor<string, []>("op_2385"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp32, [1, 1, 16, 64]> k = reshape(shape = var_2385, x = ko)[name = tensor<string, []>("k")];
            tensor<fp32, []> _inversed_2407_y_0 = const()[name = tensor<string, []>("_inversed_2407_y_0"), val = tensor<fp32, []>(0x1p-9)];
            tensor<fp32, [1, 1, 1, 1]> _inversed_2407 = mul(x = ts, y = _inversed_2407_y_0)[name = tensor<string, []>("_inversed_2407")];
            tensor<fp32, [1, 1, 1, 1]> var_2408 = floor(x = _inversed_2407)[name = tensor<string, []>("op_2408")];
            tensor<fp32, []> var_2409 = const()[name = tensor<string, []>("op_2409"), val = tensor<fp32, []>(0x1p+9)];
            tensor<fp32, [1, 1, 1, 1]> var_2410 = mul(x = var_2408, y = var_2409)[name = tensor<string, []>("op_2410")];
            tensor<fp32, [1, 1, 1, 1]> write_indices_float = sub(x = ts, y = var_2410)[name = tensor<string, []>("write_indices_float")];
            tensor<string, []> var_2417_dtype_0 = const()[name = tensor<string, []>("op_2417_dtype_0"), val = tensor<string, []>("int32")];
            tensor<int32, [4]> write_indices_reps_0 = const()[name = tensor<string, []>("write_indices_reps_0"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<int32, [1, 1, 1, 1]> var_2417 = cast(dtype = var_2417_dtype_0, x = write_indices_float)[name = tensor<string, []>("cast_108")];
            tensor<int32, [1, 1, 16, 64]> write_indices = tile(reps = write_indices_reps_0, x = var_2417)[name = tensor<string, []>("write_indices")];
            tensor<int32, [5]> var_2425_begin_0 = const()[name = tensor<string, []>("op_2425_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2425_end_0 = const()[name = tensor<string, []>("op_2425_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> var_2425_end_mask_0 = const()[name = tensor<string, []>("op_2425_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2425_squeeze_mask_0 = const()[name = tensor<string, []>("op_2425_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_2425 = slice_by_index(begin = var_2425_begin_0, end = var_2425_end_0, end_mask = var_2425_end_mask_0, squeeze_mask = var_2425_squeeze_mask_0, x = cache5)[name = tensor<string, []>("op_2425")];
            tensor<int32, []> var_2427_axis_0 = const()[name = tensor<string, []>("op_2427_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_2427_mode_0 = const()[name = tensor<string, []>("op_2427_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_2427_validate_indices_0 = const()[name = tensor<string, []>("op_2427_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_2427 = scatter_along_axis(axis = var_2427_axis_0, data = var_2425, indices = write_indices, mode = var_2427_mode_0, updates = k, validate_indices = var_2427_validate_indices_0)[name = tensor<string, []>("op_2427")];
            tensor<int32, [5]> concat_37 = const()[name = tensor<string, []>("concat_37"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_38 = const()[name = tensor<string, []>("concat_38"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_1_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_22 = const()[name = tensor<string, []>("shape_22"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_10 = const()[name = tensor<string, []>("reduce_prod_10"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_10_start_0 = const()[name = tensor<string, []>("range_1d_10_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_10_step_0 = const()[name = tensor<string, []>("range_1d_10_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor<string, []>("range_1d_10")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor<string, []>("reshape_50")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor<string, []>("slice_by_index_10")];
            tensor<int32, [1]> reshape_51_shape_0 = const()[name = tensor<string, []>("reshape_51_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor<string, []>("reshape_51")];
            tensor<int32, [1]> reshape_52_shape_0 = const()[name = tensor<string, []>("reshape_52_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_52 = reshape(shape = reshape_52_shape_0, x = var_2427)[name = tensor<string, []>("reshape_52")];
            tensor<int32, [1]> reshape_53_shape_0 = const()[name = tensor<string, []>("reshape_53_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_53 = reshape(shape = reshape_53_shape_0, x = cache5)[name = tensor<string, []>("reshape_53")];
            tensor<string, []> scatter_10_mode_0 = const()[name = tensor<string, []>("scatter_10_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_10_axis_0 = const()[name = tensor<string, []>("scatter_10_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_10_validate_indices_0 = const()[name = tensor<string, []>("scatter_10_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_10 = scatter(axis = scatter_10_axis_0, data = reshape_53, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52, validate_indices = scatter_10_validate_indices_0)[name = tensor<string, []>("scatter_10")];
            tensor<fp32, [2, 1, 512, 16, 64]> reshape_54 = reshape(shape = shape_22, x = scatter_10)[name = tensor<string, []>("reshape_54")];
            tensor<int32, [5]> var_2435_begin_0 = const()[name = tensor<string, []>("op_2435_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> var_2435_end_0 = const()[name = tensor<string, []>("op_2435_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> var_2435_end_mask_0 = const()[name = tensor<string, []>("op_2435_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> var_2435_squeeze_mask_0 = const()[name = tensor<string, []>("op_2435_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> var_2435 = slice_by_index(begin = var_2435_begin_0, end = var_2435_end_0, end_mask = var_2435_end_mask_0, squeeze_mask = var_2435_squeeze_mask_0, x = reshape_54)[name = tensor<string, []>("op_2435")];
            tensor<int32, []> var_2437_axis_0 = const()[name = tensor<string, []>("op_2437_axis_0"), val = tensor<int32, []>(1)];
            tensor<string, []> var_2437_mode_0 = const()[name = tensor<string, []>("op_2437_mode_0"), val = tensor<string, []>("update")];
            tensor<bool, []> var_2437_validate_indices_0 = const()[name = tensor<string, []>("op_2437_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 512, 16, 64]> var_2437 = scatter_along_axis(axis = var_2437_axis_0, data = var_2435, indices = write_indices, mode = var_2437_mode_0, updates = v, validate_indices = var_2437_validate_indices_0)[name = tensor<string, []>("op_2437")];
            tensor<int32, [5]> concat_39 = const()[name = tensor<string, []>("concat_39"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> concat_40 = const()[name = tensor<string, []>("concat_40"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_2_stride_0"), val = tensor<int32, [5]>([1, 1, 1, 1, 1])];
            tensor<bool, [5]> new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor<string, []>("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<int32, [5]> shape_23 = const()[name = tensor<string, []>("shape_23"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<int32, []> reduce_prod_11 = const()[name = tensor<string, []>("reduce_prod_11"), val = tensor<int32, []>(1048576)];
            tensor<int32, []> range_1d_11_start_0 = const()[name = tensor<string, []>("range_1d_11_start_0"), val = tensor<int32, []>(0)];
            tensor<int32, []> range_1d_11_step_0 = const()[name = tensor<string, []>("range_1d_11_step_0"), val = tensor<int32, []>(1)];
            tensor<int32, [1048576]> range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor<string, []>("range_1d_11")];
            tensor<int32, [2, 1, 512, 16, 64]> reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor<string, []>("reshape_55")];
            tensor<int32, [1, 512, 16, 64]> slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor<string, []>("slice_by_index_11")];
            tensor<int32, [1]> reshape_56_shape_0 = const()[name = tensor<string, []>("reshape_56_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<int32, [524288]> reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor<string, []>("reshape_56")];
            tensor<int32, [1]> reshape_57_shape_0 = const()[name = tensor<string, []>("reshape_57_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [524288]> reshape_57 = reshape(shape = reshape_57_shape_0, x = var_2437)[name = tensor<string, []>("reshape_57")];
            tensor<int32, [1]> reshape_58_shape_0 = const()[name = tensor<string, []>("reshape_58_shape_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1048576]> reshape_58 = reshape(shape = reshape_58_shape_0, x = reshape_54)[name = tensor<string, []>("reshape_58")];
            tensor<string, []> scatter_11_mode_0 = const()[name = tensor<string, []>("scatter_11_mode_0"), val = tensor<string, []>("update")];
            tensor<int32, []> scatter_11_axis_0 = const()[name = tensor<string, []>("scatter_11_axis_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> scatter_11_validate_indices_0 = const()[name = tensor<string, []>("scatter_11_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1048576]> scatter_11 = scatter(axis = scatter_11_axis_0, data = reshape_58, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57, validate_indices = scatter_11_validate_indices_0)[name = tensor<string, []>("scatter_11")];
            tensor<fp32, [2, 1, 512, 16, 64]> new_cache_internal_tensor_assign_2 = reshape(shape = shape_23, x = scatter_11)[name = tensor<string, []>("reshape_59")];
            tensor<int32, [5]> keys_31_begin_0 = const()[name = tensor<string, []>("keys_31_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> keys_31_end_0 = const()[name = tensor<string, []>("keys_31_end_0"), val = tensor<int32, [5]>([1, 1, 512, 16, 64])];
            tensor<bool, [5]> keys_31_end_mask_0 = const()[name = tensor<string, []>("keys_31_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> keys_31_squeeze_mask_0 = const()[name = tensor<string, []>("keys_31_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> keys_31 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = new_cache_internal_tensor_assign_2)[name = tensor<string, []>("keys_31")];
            tensor<int32, [5]> values_31_begin_0 = const()[name = tensor<string, []>("values_31_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> values_31_end_0 = const()[name = tensor<string, []>("values_31_end_0"), val = tensor<int32, [5]>([2, 1, 512, 16, 64])];
            tensor<bool, [5]> values_31_end_mask_0 = const()[name = tensor<string, []>("values_31_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> values_31_squeeze_mask_0 = const()[name = tensor<string, []>("values_31_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp32, [1, 512, 16, 64]> values_31 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = new_cache_internal_tensor_assign_2)[name = tensor<string, []>("values_31")];
            tensor<bool, [1, 512, 16, 64]> var_2449 = not_equal(x = keys_31, y = keys_31)[name = tensor<string, []>("op_2449")];
            tensor<fp32, [1, 512, 16, 64]> keys_33 = select(a = var_360, b = keys_31, cond = var_2449)[name = tensor<string, []>("keys_33")];
            tensor<bool, [1, 512, 16, 64]> var_2457 = not_equal(x = values_31, y = values_31)[name = tensor<string, []>("op_2457")];
            tensor<fp32, [1, 512, 16, 64]> values_33 = select(a = var_360, b = values_31, cond = var_2457)[name = tensor<string, []>("values_33")];
            tensor<int32, [4]> var_2481 = const()[name = tensor<string, []>("op_2481"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2494 = const()[name = tensor<string, []>("op_2494"), val = tensor<int32, [3]>([1, 1, 1])];
            tensor<fp32, [1, 1, 1]> var_2495 = reshape(shape = var_2494, x = position5)[name = tensor<string, []>("op_2495")];
            tensor<fp32, []> var_2512 = const()[name = tensor<string, []>("op_2512"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1, 1, 1]> valid_len = add(x = var_2495, y = var_2512)[name = tensor<string, []>("valid_len")];
            tensor<bool, [1, 1, 512]> valid_mask = less(x = k_positions_1_promoted, y = valid_len)[name = tensor<string, []>("valid_mask")];
            tensor<bool, [1, 1, 512]> causal_mask = less_equal(x = k_positions_1_promoted, y = var_2495)[name = tensor<string, []>("causal_mask")];
            tensor<bool, [1, 1, 512]> attn_mask_21 = logical_and(x = valid_mask, y = causal_mask)[name = tensor<string, []>("attn_mask_21")];
            tensor<int32, [1]> attn_mask_axes_0 = const()[name = tensor<string, []>("attn_mask_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<bool, [1, 1, 1, 512]> attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor<string, []>("attn_mask")];
            tensor<fp32, [1]> var_2524 = const()[name = tensor<string, []>("op_2524"), val = tensor<fp32, [1]>([0x1.fffe5cp-4])];
            tensor<bool, []> var_2530_transpose_x_0 = const()[name = tensor<string, []>("op_2530_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2530_transpose_y_0 = const()[name = tensor<string, []>("op_2530_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]>([0, 2, -3, -1])];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = tensor<string, []>("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
            tensor<fp32, [1, 16, 64, 512]> transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33)[name = tensor<string, []>("transpose_31")];
            tensor<fp32, [1, 16, 1, 64]> transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33)[name = tensor<string, []>("transpose_32")];
            tensor<fp32, [1, 16, 1, 512]> var_2530 = matmul(transpose_x = var_2530_transpose_x_0, transpose_y = var_2530_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor<string, []>("op_2530")];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_31 = mul(x = var_2530, y = var_2524)[name = tensor<string, []>("attn_weights_31")];
            tensor<bool, [1, 1, 1, 512]> var_2532 = logical_not(x = attn_mask)[name = tensor<string, []>("op_2532")];
            tensor<fp32, []> var_2533 = const()[name = tensor<string, []>("op_2533"), val = tensor<fp32, []>(-0x1.ff933cp+127)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights_33 = select(a = var_2533, b = attn_weights_31, cond = var_2532)[name = tensor<string, []>("attn_weights_33")];
            tensor<int32, []> var_2535 = const()[name = tensor<string, []>("op_2535"), val = tensor<int32, []>(-1)];
            tensor<fp32, [1, 16, 1, 512]> attn_weights = softmax(axis = var_2535, x = attn_weights_33)[name = tensor<string, []>("attn_weights")];
            tensor<bool, []> attn_output_transpose_x_0 = const()[name = tensor<string, []>("attn_output_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> attn_output_transpose_y_0 = const()[name = tensor<string, []>("attn_output_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<fp32, [1, 16, 512, 64]> values = transpose(perm = var_2481, x = values_33)[name = tensor<string, []>("transpose_33")];
            tensor<fp32, [1, 16, 1, 64]> attn_output = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights, y = values)[name = tensor<string, []>("attn_output")];
            tensor<int32, [4]> var_2543 = const()[name = tensor<string, []>("op_2543"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2546 = const()[name = tensor<string, []>("op_2546"), val = tensor<int32, [3]>([1, 1, 1024])];
            tensor<fp32, [1, 1, 16, 64]> var_2544 = transpose(perm = var_2543, x = attn_output)[name = tensor<string, []>("transpose_30")];
            tensor<fp32, [1, 1, 1024]> input_55 = reshape(shape = var_2546, x = var_2544)[name = tensor<string, []>("input_55")];
            tensor<fp32, [1, 1, 1024]> attn_out = linear(bias = linear_0_bias_0, weight = attn5_out_proj_weight, x = input_55)[name = tensor<string, []>("linear_22")];
            tensor<fp32, []> var_2552 = const()[name = tensor<string, []>("op_2552"), val = tensor<fp32, []>(0x1p+0)];
            tensor<fp32, [1]> var_2553 = add(x = position5, y = var_2552)[name = tensor<string, []>("op_2553")];
            tensor<fp32, [1, 1, 1024]> input_57 = add(x = input_53, y = attn_out)[name = tensor<string, []>("input_57")];
            tensor<fp32, []> var_2557 = const()[name = tensor<string, []>("op_2557"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> input_59_axes_0 = const()[name = tensor<string, []>("input_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> input_59 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias, epsilon = var_2557, gamma = norm5_2_weight, x = input_57)[name = tensor<string, []>("input_59")];
            tensor<fp32, [1, 1, 4096]> var_2565 = linear(bias = linear_3_bias_0, weight = linear5_1_weight, x = input_59)[name = tensor<string, []>("linear_23")];
            tensor<string, []> input_61_mode_0 = const()[name = tensor<string, []>("input_61_mode_0"), val = tensor<string, []>("EXACT")];
            tensor<fp32, [1, 1, 4096]> input_61 = gelu(mode = input_61_mode_0, x = var_2565)[name = tensor<string, []>("input_61")];
            tensor<fp32, [1, 1, 1024]> ffn_out = linear(bias = linear_0_bias_0, weight = linear5_2_weight, x = input_61)[name = tensor<string, []>("linear_24")];
            tensor<fp32, [1, 1, 1024]> input_63 = add(x = input_57, y = ffn_out)[name = tensor<string, []>("input_63")];
            tensor<fp32, []> var_2574 = const()[name = tensor<string, []>("op_2574"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp32, [1, 1, 1024]> input = layer_norm(axes = input_axes_0, beta = out_norm_bias, epsilon = var_2574, gamma = out_norm_weight, x = input_63)[name = tensor<string, []>("input")];
            tensor<fp32, [1, 1, 1]> var_2582 = linear(bias = out_eos_bias, weight = out_eos_weight, x = input)[name = tensor<string, []>("linear_25")];
        } -> (input, var_2582, new_cache_1_internal_tensor_assign_2, var_458, new_cache_3_internal_tensor_assign_2, var_877, new_cache_5_internal_tensor_assign_2, var_1296, new_cache_7_internal_tensor_assign_2, var_1715, new_cache_9_internal_tensor_assign_2, var_2134, new_cache_internal_tensor_assign_2, var_2553);
}