| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1.2"}, {"coremltools-component-torch", "2.14.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] |
| { |
| func main<ios18>(tensor<fp16, [1, 104, 1024]> embeddings, tensor<fp16, [1, 1, 1, 104]> key_bias) { |
| int32 var_84 = const()[name = string("op_84"), val = int32(-1)]; |
| tensor<int32, [1]> input_1_axes_0 = const()[name = string("input_1_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_0_ln1_weight_to_fp16 = const()[name = string("m_layers_0_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))]; |
| tensor<fp16, [1024]> m_layers_0_ln1_bias_to_fp16 = const()[name = string("m_layers_0_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2176)))]; |
| fp16 var_87_to_fp16 = const()[name = string("op_87_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_1_cast_fp16 = layer_norm(axes = input_1_axes_0, beta = m_layers_0_ln1_bias_to_fp16, epsilon = var_87_to_fp16, gamma = m_layers_0_ln1_weight_to_fp16, x = embeddings)[name = string("input_1_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_0_q_weight_to_fp16 = const()[name = string("m_layers_0_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4288)))]; |
| tensor<fp16, [1024]> m_layers_0_q_bias_to_fp16 = const()[name = string("m_layers_0_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2101504)))]; |
| tensor<fp16, [1, 104, 1024]> linear_0_cast_fp16 = linear(bias = m_layers_0_q_bias_to_fp16, weight = m_layers_0_q_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")]; |
| tensor<int32, [4]> var_110 = const()[name = string("op_110"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_111_cast_fp16 = reshape(shape = var_110, x = linear_0_cast_fp16)[name = string("op_111_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_0_k_weight_to_fp16 = const()[name = string("m_layers_0_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2103616)))]; |
| tensor<fp16, [1024]> m_layers_0_k_bias_to_fp16 = const()[name = string("m_layers_0_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4200832)))]; |
| tensor<fp16, [1, 104, 1024]> linear_1_cast_fp16 = linear(bias = m_layers_0_k_bias_to_fp16, weight = m_layers_0_k_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")]; |
| tensor<int32, [4]> var_116 = const()[name = string("op_116"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_117_cast_fp16 = reshape(shape = var_116, x = linear_1_cast_fp16)[name = string("op_117_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_0_v_weight_to_fp16 = const()[name = string("m_layers_0_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4202944)))]; |
| tensor<fp16, [1024]> m_layers_0_v_bias_to_fp16 = const()[name = string("m_layers_0_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6300160)))]; |
| tensor<fp16, [1, 104, 1024]> linear_2_cast_fp16 = linear(bias = m_layers_0_v_bias_to_fp16, weight = m_layers_0_v_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")]; |
| tensor<int32, [4]> var_122 = const()[name = string("op_122"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_123_cast_fp16 = reshape(shape = var_122, x = linear_2_cast_fp16)[name = string("op_123_cast_fp16")]; |
| tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_125_to_fp16 = const()[name = string("op_125_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_126_cast_fp16 = mul(x = var_111_cast_fp16, y = var_125_to_fp16)[name = string("op_126_cast_fp16")]; |
| bool var_128_transpose_x_0 = const()[name = string("op_128_transpose_x_0"), val = bool(false)]; |
| bool var_128_transpose_y_0 = const()[name = string("op_128_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_72_perm_0 = const()[name = string("transpose_72_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_73_perm_0 = const()[name = string("transpose_73_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_73 = transpose(perm = transpose_73_perm_0, x = var_117_cast_fp16)[name = string("transpose_213")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_72 = transpose(perm = transpose_72_perm_0, x = var_126_cast_fp16)[name = string("transpose_214")]; |
| tensor<fp16, [1, 16, 104, 104]> var_128_cast_fp16 = matmul(transpose_x = var_128_transpose_x_0, transpose_y = var_128_transpose_y_0, x = transpose_72, y = transpose_73)[name = string("op_128_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_129_cast_fp16 = add(x = var_128_cast_fp16, y = key_bias)[name = string("op_129_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_1_cast_fp16 = softmax(axis = var_84, x = var_129_cast_fp16)[name = string("att_1_cast_fp16")]; |
| bool var_131_transpose_x_0 = const()[name = string("op_131_transpose_x_0"), val = bool(false)]; |
| bool var_131_transpose_y_0 = const()[name = string("op_131_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_123_cast_fp16)[name = string("transpose_215")]; |
| tensor<fp16, [1, 16, 104, 64]> var_131_cast_fp16 = matmul(transpose_x = var_131_transpose_x_0, transpose_y = var_131_transpose_y_0, x = att_1_cast_fp16, y = v_1_cast_fp16)[name = string("op_131_cast_fp16")]; |
| tensor<int32, [4]> var_132_perm_0 = const()[name = string("op_132_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_133 = const()[name = string("op_133"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_132_cast_fp16 = transpose(perm = var_132_perm_0, x = var_131_cast_fp16)[name = string("transpose_212")]; |
| tensor<fp16, [1, 104, 1024]> input_3_cast_fp16 = reshape(shape = var_133, x = var_132_cast_fp16)[name = string("input_3_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_0_o_weight_to_fp16 = const()[name = string("m_layers_0_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6302272)))]; |
| tensor<fp16, [1024]> m_layers_0_o_bias_to_fp16 = const()[name = string("m_layers_0_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8399488)))]; |
| tensor<fp16, [1, 104, 1024]> linear_3_cast_fp16 = linear(bias = m_layers_0_o_bias_to_fp16, weight = m_layers_0_o_weight_to_fp16, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_5_cast_fp16 = add(x = embeddings, y = linear_3_cast_fp16)[name = string("input_5_cast_fp16")]; |
| tensor<int32, [1]> input_7_axes_0 = const()[name = string("input_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_0_ln2_weight_to_fp16 = const()[name = string("m_layers_0_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8401600)))]; |
| tensor<fp16, [1024]> m_layers_0_ln2_bias_to_fp16 = const()[name = string("m_layers_0_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8403712)))]; |
| tensor<fp16, [1, 104, 1024]> input_7_cast_fp16 = layer_norm(axes = input_7_axes_0, beta = m_layers_0_ln2_bias_to_fp16, epsilon = var_87_to_fp16, gamma = m_layers_0_ln2_weight_to_fp16, x = input_5_cast_fp16)[name = string("input_7_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_0_fc1_weight_to_fp16 = const()[name = string("m_layers_0_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8405824)))]; |
| tensor<fp16, [4096]> m_layers_0_fc1_bias_to_fp16 = const()[name = string("m_layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16794496)))]; |
| tensor<fp16, [1, 104, 4096]> linear_4_cast_fp16 = linear(bias = m_layers_0_fc1_bias_to_fp16, weight = m_layers_0_fc1_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_4_cast_fp16")]; |
| string input_9_mode_0 = const()[name = string("input_9_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_9_cast_fp16 = gelu(mode = input_9_mode_0, x = linear_4_cast_fp16)[name = string("input_9_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_0_fc2_weight_to_fp16 = const()[name = string("m_layers_0_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16802752)))]; |
| tensor<fp16, [1024]> m_layers_0_fc2_bias_to_fp16 = const()[name = string("m_layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25191424)))]; |
| tensor<fp16, [1, 104, 1024]> linear_5_cast_fp16 = linear(bias = m_layers_0_fc2_bias_to_fp16, weight = m_layers_0_fc2_weight_to_fp16, x = input_9_cast_fp16)[name = string("linear_5_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_3_cast_fp16 = add(x = input_5_cast_fp16, y = linear_5_cast_fp16)[name = string("x_3_cast_fp16")]; |
| int32 var_154 = const()[name = string("op_154"), val = int32(-1)]; |
| tensor<int32, [1]> input_11_axes_0 = const()[name = string("input_11_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_1_ln1_weight_to_fp16 = const()[name = string("m_layers_1_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25193536)))]; |
| tensor<fp16, [1024]> m_layers_1_ln1_bias_to_fp16 = const()[name = string("m_layers_1_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25195648)))]; |
| fp16 var_157_to_fp16 = const()[name = string("op_157_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = m_layers_1_ln1_bias_to_fp16, epsilon = var_157_to_fp16, gamma = m_layers_1_ln1_weight_to_fp16, x = x_3_cast_fp16)[name = string("input_11_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_1_q_weight_to_fp16 = const()[name = string("m_layers_1_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25197760)))]; |
| tensor<fp16, [1024]> m_layers_1_q_bias_to_fp16 = const()[name = string("m_layers_1_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27294976)))]; |
| tensor<fp16, [1, 104, 1024]> linear_6_cast_fp16 = linear(bias = m_layers_1_q_bias_to_fp16, weight = m_layers_1_q_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_6_cast_fp16")]; |
| tensor<int32, [4]> var_180 = const()[name = string("op_180"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_181_cast_fp16 = reshape(shape = var_180, x = linear_6_cast_fp16)[name = string("op_181_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_1_k_weight_to_fp16 = const()[name = string("m_layers_1_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27297088)))]; |
| tensor<fp16, [1024]> m_layers_1_k_bias_to_fp16 = const()[name = string("m_layers_1_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29394304)))]; |
| tensor<fp16, [1, 104, 1024]> linear_7_cast_fp16 = linear(bias = m_layers_1_k_bias_to_fp16, weight = m_layers_1_k_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")]; |
| tensor<int32, [4]> var_186 = const()[name = string("op_186"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_187_cast_fp16 = reshape(shape = var_186, x = linear_7_cast_fp16)[name = string("op_187_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_1_v_weight_to_fp16 = const()[name = string("m_layers_1_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29396416)))]; |
| tensor<fp16, [1024]> m_layers_1_v_bias_to_fp16 = const()[name = string("m_layers_1_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31493632)))]; |
| tensor<fp16, [1, 104, 1024]> linear_8_cast_fp16 = linear(bias = m_layers_1_v_bias_to_fp16, weight = m_layers_1_v_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_8_cast_fp16")]; |
| tensor<int32, [4]> var_192 = const()[name = string("op_192"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_193_cast_fp16 = reshape(shape = var_192, x = linear_8_cast_fp16)[name = string("op_193_cast_fp16")]; |
| tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_195_to_fp16 = const()[name = string("op_195_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_196_cast_fp16 = mul(x = var_181_cast_fp16, y = var_195_to_fp16)[name = string("op_196_cast_fp16")]; |
| bool var_198_transpose_x_0 = const()[name = string("op_198_transpose_x_0"), val = bool(false)]; |
| bool var_198_transpose_y_0 = const()[name = string("op_198_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_74_perm_0 = const()[name = string("transpose_74_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_75_perm_0 = const()[name = string("transpose_75_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_75 = transpose(perm = transpose_75_perm_0, x = var_187_cast_fp16)[name = string("transpose_209")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_74 = transpose(perm = transpose_74_perm_0, x = var_196_cast_fp16)[name = string("transpose_210")]; |
| tensor<fp16, [1, 16, 104, 104]> var_198_cast_fp16 = matmul(transpose_x = var_198_transpose_x_0, transpose_y = var_198_transpose_y_0, x = transpose_74, y = transpose_75)[name = string("op_198_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_199_cast_fp16 = add(x = var_198_cast_fp16, y = key_bias)[name = string("op_199_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_3_cast_fp16 = softmax(axis = var_154, x = var_199_cast_fp16)[name = string("att_3_cast_fp16")]; |
| bool var_201_transpose_x_0 = const()[name = string("op_201_transpose_x_0"), val = bool(false)]; |
| bool var_201_transpose_y_0 = const()[name = string("op_201_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_193_cast_fp16)[name = string("transpose_211")]; |
| tensor<fp16, [1, 16, 104, 64]> var_201_cast_fp16 = matmul(transpose_x = var_201_transpose_x_0, transpose_y = var_201_transpose_y_0, x = att_3_cast_fp16, y = v_3_cast_fp16)[name = string("op_201_cast_fp16")]; |
| tensor<int32, [4]> var_202_perm_0 = const()[name = string("op_202_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_203 = const()[name = string("op_203"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_202_cast_fp16 = transpose(perm = var_202_perm_0, x = var_201_cast_fp16)[name = string("transpose_208")]; |
| tensor<fp16, [1, 104, 1024]> input_13_cast_fp16 = reshape(shape = var_203, x = var_202_cast_fp16)[name = string("input_13_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_1_o_weight_to_fp16 = const()[name = string("m_layers_1_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31495744)))]; |
| tensor<fp16, [1024]> m_layers_1_o_bias_to_fp16 = const()[name = string("m_layers_1_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33592960)))]; |
| tensor<fp16, [1, 104, 1024]> linear_9_cast_fp16 = linear(bias = m_layers_1_o_bias_to_fp16, weight = m_layers_1_o_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_9_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_15_cast_fp16 = add(x = x_3_cast_fp16, y = linear_9_cast_fp16)[name = string("input_15_cast_fp16")]; |
| tensor<int32, [1]> input_17_axes_0 = const()[name = string("input_17_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_1_ln2_weight_to_fp16 = const()[name = string("m_layers_1_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33595072)))]; |
| tensor<fp16, [1024]> m_layers_1_ln2_bias_to_fp16 = const()[name = string("m_layers_1_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33597184)))]; |
| tensor<fp16, [1, 104, 1024]> input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = m_layers_1_ln2_bias_to_fp16, epsilon = var_157_to_fp16, gamma = m_layers_1_ln2_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_17_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_1_fc1_weight_to_fp16 = const()[name = string("m_layers_1_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33599296)))]; |
| tensor<fp16, [4096]> m_layers_1_fc1_bias_to_fp16 = const()[name = string("m_layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41987968)))]; |
| tensor<fp16, [1, 104, 4096]> linear_10_cast_fp16 = linear(bias = m_layers_1_fc1_bias_to_fp16, weight = m_layers_1_fc1_weight_to_fp16, x = input_17_cast_fp16)[name = string("linear_10_cast_fp16")]; |
| string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_10_cast_fp16)[name = string("input_19_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_1_fc2_weight_to_fp16 = const()[name = string("m_layers_1_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41996224)))]; |
| tensor<fp16, [1024]> m_layers_1_fc2_bias_to_fp16 = const()[name = string("m_layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50384896)))]; |
| tensor<fp16, [1, 104, 1024]> linear_11_cast_fp16 = linear(bias = m_layers_1_fc2_bias_to_fp16, weight = m_layers_1_fc2_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_11_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_5_cast_fp16 = add(x = input_15_cast_fp16, y = linear_11_cast_fp16)[name = string("x_5_cast_fp16")]; |
| int32 var_224 = const()[name = string("op_224"), val = int32(-1)]; |
| tensor<int32, [1]> input_21_axes_0 = const()[name = string("input_21_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_2_ln1_weight_to_fp16 = const()[name = string("m_layers_2_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50387008)))]; |
| tensor<fp16, [1024]> m_layers_2_ln1_bias_to_fp16 = const()[name = string("m_layers_2_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50389120)))]; |
| fp16 var_227_to_fp16 = const()[name = string("op_227_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = m_layers_2_ln1_bias_to_fp16, epsilon = var_227_to_fp16, gamma = m_layers_2_ln1_weight_to_fp16, x = x_5_cast_fp16)[name = string("input_21_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_2_q_weight_to_fp16 = const()[name = string("m_layers_2_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50391232)))]; |
| tensor<fp16, [1024]> m_layers_2_q_bias_to_fp16 = const()[name = string("m_layers_2_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52488448)))]; |
| tensor<fp16, [1, 104, 1024]> linear_12_cast_fp16 = linear(bias = m_layers_2_q_bias_to_fp16, weight = m_layers_2_q_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_12_cast_fp16")]; |
| tensor<int32, [4]> var_250 = const()[name = string("op_250"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_251_cast_fp16 = reshape(shape = var_250, x = linear_12_cast_fp16)[name = string("op_251_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_2_k_weight_to_fp16 = const()[name = string("m_layers_2_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(52490560)))]; |
| tensor<fp16, [1024]> m_layers_2_k_bias_to_fp16 = const()[name = string("m_layers_2_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54587776)))]; |
| tensor<fp16, [1, 104, 1024]> linear_13_cast_fp16 = linear(bias = m_layers_2_k_bias_to_fp16, weight = m_layers_2_k_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_13_cast_fp16")]; |
| tensor<int32, [4]> var_256 = const()[name = string("op_256"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_257_cast_fp16 = reshape(shape = var_256, x = linear_13_cast_fp16)[name = string("op_257_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_2_v_weight_to_fp16 = const()[name = string("m_layers_2_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(54589888)))]; |
| tensor<fp16, [1024]> m_layers_2_v_bias_to_fp16 = const()[name = string("m_layers_2_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56687104)))]; |
| tensor<fp16, [1, 104, 1024]> linear_14_cast_fp16 = linear(bias = m_layers_2_v_bias_to_fp16, weight = m_layers_2_v_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")]; |
| tensor<int32, [4]> var_262 = const()[name = string("op_262"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_263_cast_fp16 = reshape(shape = var_262, x = linear_14_cast_fp16)[name = string("op_263_cast_fp16")]; |
| tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_265_to_fp16 = const()[name = string("op_265_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_266_cast_fp16 = mul(x = var_251_cast_fp16, y = var_265_to_fp16)[name = string("op_266_cast_fp16")]; |
| bool var_268_transpose_x_0 = const()[name = string("op_268_transpose_x_0"), val = bool(false)]; |
| bool var_268_transpose_y_0 = const()[name = string("op_268_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_76_perm_0 = const()[name = string("transpose_76_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_77_perm_0 = const()[name = string("transpose_77_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_77 = transpose(perm = transpose_77_perm_0, x = var_257_cast_fp16)[name = string("transpose_205")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_76 = transpose(perm = transpose_76_perm_0, x = var_266_cast_fp16)[name = string("transpose_206")]; |
| tensor<fp16, [1, 16, 104, 104]> var_268_cast_fp16 = matmul(transpose_x = var_268_transpose_x_0, transpose_y = var_268_transpose_y_0, x = transpose_76, y = transpose_77)[name = string("op_268_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_269_cast_fp16 = add(x = var_268_cast_fp16, y = key_bias)[name = string("op_269_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_5_cast_fp16 = softmax(axis = var_224, x = var_269_cast_fp16)[name = string("att_5_cast_fp16")]; |
| bool var_271_transpose_x_0 = const()[name = string("op_271_transpose_x_0"), val = bool(false)]; |
| bool var_271_transpose_y_0 = const()[name = string("op_271_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_263_cast_fp16)[name = string("transpose_207")]; |
| tensor<fp16, [1, 16, 104, 64]> var_271_cast_fp16 = matmul(transpose_x = var_271_transpose_x_0, transpose_y = var_271_transpose_y_0, x = att_5_cast_fp16, y = v_5_cast_fp16)[name = string("op_271_cast_fp16")]; |
| tensor<int32, [4]> var_272_perm_0 = const()[name = string("op_272_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_273 = const()[name = string("op_273"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_272_cast_fp16 = transpose(perm = var_272_perm_0, x = var_271_cast_fp16)[name = string("transpose_204")]; |
| tensor<fp16, [1, 104, 1024]> input_23_cast_fp16 = reshape(shape = var_273, x = var_272_cast_fp16)[name = string("input_23_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_2_o_weight_to_fp16 = const()[name = string("m_layers_2_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(56689216)))]; |
| tensor<fp16, [1024]> m_layers_2_o_bias_to_fp16 = const()[name = string("m_layers_2_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58786432)))]; |
| tensor<fp16, [1, 104, 1024]> linear_15_cast_fp16 = linear(bias = m_layers_2_o_bias_to_fp16, weight = m_layers_2_o_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_15_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_25_cast_fp16 = add(x = x_5_cast_fp16, y = linear_15_cast_fp16)[name = string("input_25_cast_fp16")]; |
| tensor<int32, [1]> input_27_axes_0 = const()[name = string("input_27_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_2_ln2_weight_to_fp16 = const()[name = string("m_layers_2_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58788544)))]; |
| tensor<fp16, [1024]> m_layers_2_ln2_bias_to_fp16 = const()[name = string("m_layers_2_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58790656)))]; |
| tensor<fp16, [1, 104, 1024]> input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = m_layers_2_ln2_bias_to_fp16, epsilon = var_227_to_fp16, gamma = m_layers_2_ln2_weight_to_fp16, x = input_25_cast_fp16)[name = string("input_27_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_2_fc1_weight_to_fp16 = const()[name = string("m_layers_2_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58792768)))]; |
| tensor<fp16, [4096]> m_layers_2_fc1_bias_to_fp16 = const()[name = string("m_layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67181440)))]; |
| tensor<fp16, [1, 104, 4096]> linear_16_cast_fp16 = linear(bias = m_layers_2_fc1_bias_to_fp16, weight = m_layers_2_fc1_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_16_cast_fp16")]; |
| string input_29_mode_0 = const()[name = string("input_29_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_29_cast_fp16 = gelu(mode = input_29_mode_0, x = linear_16_cast_fp16)[name = string("input_29_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_2_fc2_weight_to_fp16 = const()[name = string("m_layers_2_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67189696)))]; |
| tensor<fp16, [1024]> m_layers_2_fc2_bias_to_fp16 = const()[name = string("m_layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75578368)))]; |
| tensor<fp16, [1, 104, 1024]> linear_17_cast_fp16 = linear(bias = m_layers_2_fc2_bias_to_fp16, weight = m_layers_2_fc2_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_17_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_7_cast_fp16 = add(x = input_25_cast_fp16, y = linear_17_cast_fp16)[name = string("x_7_cast_fp16")]; |
| int32 var_294 = const()[name = string("op_294"), val = int32(-1)]; |
| tensor<int32, [1]> input_31_axes_0 = const()[name = string("input_31_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_3_ln1_weight_to_fp16 = const()[name = string("m_layers_3_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75580480)))]; |
| tensor<fp16, [1024]> m_layers_3_ln1_bias_to_fp16 = const()[name = string("m_layers_3_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75582592)))]; |
| fp16 var_297_to_fp16 = const()[name = string("op_297_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_31_cast_fp16 = layer_norm(axes = input_31_axes_0, beta = m_layers_3_ln1_bias_to_fp16, epsilon = var_297_to_fp16, gamma = m_layers_3_ln1_weight_to_fp16, x = x_7_cast_fp16)[name = string("input_31_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_3_q_weight_to_fp16 = const()[name = string("m_layers_3_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75584704)))]; |
| tensor<fp16, [1024]> m_layers_3_q_bias_to_fp16 = const()[name = string("m_layers_3_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77681920)))]; |
| tensor<fp16, [1, 104, 1024]> linear_18_cast_fp16 = linear(bias = m_layers_3_q_bias_to_fp16, weight = m_layers_3_q_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_18_cast_fp16")]; |
| tensor<int32, [4]> var_320 = const()[name = string("op_320"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_321_cast_fp16 = reshape(shape = var_320, x = linear_18_cast_fp16)[name = string("op_321_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_3_k_weight_to_fp16 = const()[name = string("m_layers_3_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(77684032)))]; |
| tensor<fp16, [1024]> m_layers_3_k_bias_to_fp16 = const()[name = string("m_layers_3_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79781248)))]; |
| tensor<fp16, [1, 104, 1024]> linear_19_cast_fp16 = linear(bias = m_layers_3_k_bias_to_fp16, weight = m_layers_3_k_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_19_cast_fp16")]; |
| tensor<int32, [4]> var_326 = const()[name = string("op_326"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_327_cast_fp16 = reshape(shape = var_326, x = linear_19_cast_fp16)[name = string("op_327_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_3_v_weight_to_fp16 = const()[name = string("m_layers_3_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79783360)))]; |
| tensor<fp16, [1024]> m_layers_3_v_bias_to_fp16 = const()[name = string("m_layers_3_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81880576)))]; |
| tensor<fp16, [1, 104, 1024]> linear_20_cast_fp16 = linear(bias = m_layers_3_v_bias_to_fp16, weight = m_layers_3_v_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_20_cast_fp16")]; |
| tensor<int32, [4]> var_332 = const()[name = string("op_332"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_333_cast_fp16 = reshape(shape = var_332, x = linear_20_cast_fp16)[name = string("op_333_cast_fp16")]; |
| tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_335_to_fp16 = const()[name = string("op_335_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_336_cast_fp16 = mul(x = var_321_cast_fp16, y = var_335_to_fp16)[name = string("op_336_cast_fp16")]; |
| bool var_338_transpose_x_0 = const()[name = string("op_338_transpose_x_0"), val = bool(false)]; |
| bool var_338_transpose_y_0 = const()[name = string("op_338_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_78_perm_0 = const()[name = string("transpose_78_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_79_perm_0 = const()[name = string("transpose_79_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_79 = transpose(perm = transpose_79_perm_0, x = var_327_cast_fp16)[name = string("transpose_201")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_78 = transpose(perm = transpose_78_perm_0, x = var_336_cast_fp16)[name = string("transpose_202")]; |
| tensor<fp16, [1, 16, 104, 104]> var_338_cast_fp16 = matmul(transpose_x = var_338_transpose_x_0, transpose_y = var_338_transpose_y_0, x = transpose_78, y = transpose_79)[name = string("op_338_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_339_cast_fp16 = add(x = var_338_cast_fp16, y = key_bias)[name = string("op_339_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_7_cast_fp16 = softmax(axis = var_294, x = var_339_cast_fp16)[name = string("att_7_cast_fp16")]; |
| bool var_341_transpose_x_0 = const()[name = string("op_341_transpose_x_0"), val = bool(false)]; |
| bool var_341_transpose_y_0 = const()[name = string("op_341_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_333_cast_fp16)[name = string("transpose_203")]; |
| tensor<fp16, [1, 16, 104, 64]> var_341_cast_fp16 = matmul(transpose_x = var_341_transpose_x_0, transpose_y = var_341_transpose_y_0, x = att_7_cast_fp16, y = v_7_cast_fp16)[name = string("op_341_cast_fp16")]; |
| tensor<int32, [4]> var_342_perm_0 = const()[name = string("op_342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_343 = const()[name = string("op_343"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_342_cast_fp16 = transpose(perm = var_342_perm_0, x = var_341_cast_fp16)[name = string("transpose_200")]; |
| tensor<fp16, [1, 104, 1024]> input_33_cast_fp16 = reshape(shape = var_343, x = var_342_cast_fp16)[name = string("input_33_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_3_o_weight_to_fp16 = const()[name = string("m_layers_3_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(81882688)))]; |
| tensor<fp16, [1024]> m_layers_3_o_bias_to_fp16 = const()[name = string("m_layers_3_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83979904)))]; |
| tensor<fp16, [1, 104, 1024]> linear_21_cast_fp16 = linear(bias = m_layers_3_o_bias_to_fp16, weight = m_layers_3_o_weight_to_fp16, x = input_33_cast_fp16)[name = string("linear_21_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_35_cast_fp16 = add(x = x_7_cast_fp16, y = linear_21_cast_fp16)[name = string("input_35_cast_fp16")]; |
| tensor<int32, [1]> input_37_axes_0 = const()[name = string("input_37_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_3_ln2_weight_to_fp16 = const()[name = string("m_layers_3_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83982016)))]; |
| tensor<fp16, [1024]> m_layers_3_ln2_bias_to_fp16 = const()[name = string("m_layers_3_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83984128)))]; |
| tensor<fp16, [1, 104, 1024]> input_37_cast_fp16 = layer_norm(axes = input_37_axes_0, beta = m_layers_3_ln2_bias_to_fp16, epsilon = var_297_to_fp16, gamma = m_layers_3_ln2_weight_to_fp16, x = input_35_cast_fp16)[name = string("input_37_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_3_fc1_weight_to_fp16 = const()[name = string("m_layers_3_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83986240)))]; |
| tensor<fp16, [4096]> m_layers_3_fc1_bias_to_fp16 = const()[name = string("m_layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92374912)))]; |
| tensor<fp16, [1, 104, 4096]> linear_22_cast_fp16 = linear(bias = m_layers_3_fc1_bias_to_fp16, weight = m_layers_3_fc1_weight_to_fp16, x = input_37_cast_fp16)[name = string("linear_22_cast_fp16")]; |
| string input_39_mode_0 = const()[name = string("input_39_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_39_cast_fp16 = gelu(mode = input_39_mode_0, x = linear_22_cast_fp16)[name = string("input_39_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_3_fc2_weight_to_fp16 = const()[name = string("m_layers_3_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92383168)))]; |
| tensor<fp16, [1024]> m_layers_3_fc2_bias_to_fp16 = const()[name = string("m_layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100771840)))]; |
| tensor<fp16, [1, 104, 1024]> linear_23_cast_fp16 = linear(bias = m_layers_3_fc2_bias_to_fp16, weight = m_layers_3_fc2_weight_to_fp16, x = input_39_cast_fp16)[name = string("linear_23_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_9_cast_fp16 = add(x = input_35_cast_fp16, y = linear_23_cast_fp16)[name = string("x_9_cast_fp16")]; |
| int32 var_364 = const()[name = string("op_364"), val = int32(-1)]; |
| tensor<int32, [1]> input_41_axes_0 = const()[name = string("input_41_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_4_ln1_weight_to_fp16 = const()[name = string("m_layers_4_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100773952)))]; |
| tensor<fp16, [1024]> m_layers_4_ln1_bias_to_fp16 = const()[name = string("m_layers_4_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100776064)))]; |
| fp16 var_367_to_fp16 = const()[name = string("op_367_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = m_layers_4_ln1_bias_to_fp16, epsilon = var_367_to_fp16, gamma = m_layers_4_ln1_weight_to_fp16, x = x_9_cast_fp16)[name = string("input_41_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_4_q_weight_to_fp16 = const()[name = string("m_layers_4_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100778176)))]; |
| tensor<fp16, [1024]> m_layers_4_q_bias_to_fp16 = const()[name = string("m_layers_4_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102875392)))]; |
| tensor<fp16, [1, 104, 1024]> linear_24_cast_fp16 = linear(bias = m_layers_4_q_bias_to_fp16, weight = m_layers_4_q_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_24_cast_fp16")]; |
| tensor<int32, [4]> var_390 = const()[name = string("op_390"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_391_cast_fp16 = reshape(shape = var_390, x = linear_24_cast_fp16)[name = string("op_391_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_4_k_weight_to_fp16 = const()[name = string("m_layers_4_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(102877504)))]; |
| tensor<fp16, [1024]> m_layers_4_k_bias_to_fp16 = const()[name = string("m_layers_4_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104974720)))]; |
| tensor<fp16, [1, 104, 1024]> linear_25_cast_fp16 = linear(bias = m_layers_4_k_bias_to_fp16, weight = m_layers_4_k_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_25_cast_fp16")]; |
| tensor<int32, [4]> var_396 = const()[name = string("op_396"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_397_cast_fp16 = reshape(shape = var_396, x = linear_25_cast_fp16)[name = string("op_397_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_4_v_weight_to_fp16 = const()[name = string("m_layers_4_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(104976832)))]; |
| tensor<fp16, [1024]> m_layers_4_v_bias_to_fp16 = const()[name = string("m_layers_4_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107074048)))]; |
| tensor<fp16, [1, 104, 1024]> linear_26_cast_fp16 = linear(bias = m_layers_4_v_bias_to_fp16, weight = m_layers_4_v_weight_to_fp16, x = input_41_cast_fp16)[name = string("linear_26_cast_fp16")]; |
| tensor<int32, [4]> var_402 = const()[name = string("op_402"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_403_cast_fp16 = reshape(shape = var_402, x = linear_26_cast_fp16)[name = string("op_403_cast_fp16")]; |
| tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_405_to_fp16 = const()[name = string("op_405_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_406_cast_fp16 = mul(x = var_391_cast_fp16, y = var_405_to_fp16)[name = string("op_406_cast_fp16")]; |
| bool var_408_transpose_x_0 = const()[name = string("op_408_transpose_x_0"), val = bool(false)]; |
| bool var_408_transpose_y_0 = const()[name = string("op_408_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_80_perm_0 = const()[name = string("transpose_80_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_81_perm_0 = const()[name = string("transpose_81_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_81 = transpose(perm = transpose_81_perm_0, x = var_397_cast_fp16)[name = string("transpose_197")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_80 = transpose(perm = transpose_80_perm_0, x = var_406_cast_fp16)[name = string("transpose_198")]; |
| tensor<fp16, [1, 16, 104, 104]> var_408_cast_fp16 = matmul(transpose_x = var_408_transpose_x_0, transpose_y = var_408_transpose_y_0, x = transpose_80, y = transpose_81)[name = string("op_408_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_409_cast_fp16 = add(x = var_408_cast_fp16, y = key_bias)[name = string("op_409_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_9_cast_fp16 = softmax(axis = var_364, x = var_409_cast_fp16)[name = string("att_9_cast_fp16")]; |
| bool var_411_transpose_x_0 = const()[name = string("op_411_transpose_x_0"), val = bool(false)]; |
| bool var_411_transpose_y_0 = const()[name = string("op_411_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_403_cast_fp16)[name = string("transpose_199")]; |
| tensor<fp16, [1, 16, 104, 64]> var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = att_9_cast_fp16, y = v_9_cast_fp16)[name = string("op_411_cast_fp16")]; |
| tensor<int32, [4]> var_412_perm_0 = const()[name = string("op_412_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_413 = const()[name = string("op_413"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_412_cast_fp16 = transpose(perm = var_412_perm_0, x = var_411_cast_fp16)[name = string("transpose_196")]; |
| tensor<fp16, [1, 104, 1024]> input_43_cast_fp16 = reshape(shape = var_413, x = var_412_cast_fp16)[name = string("input_43_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_4_o_weight_to_fp16 = const()[name = string("m_layers_4_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(107076160)))]; |
| tensor<fp16, [1024]> m_layers_4_o_bias_to_fp16 = const()[name = string("m_layers_4_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109173376)))]; |
| tensor<fp16, [1, 104, 1024]> linear_27_cast_fp16 = linear(bias = m_layers_4_o_bias_to_fp16, weight = m_layers_4_o_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_27_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_45_cast_fp16 = add(x = x_9_cast_fp16, y = linear_27_cast_fp16)[name = string("input_45_cast_fp16")]; |
| tensor<int32, [1]> input_47_axes_0 = const()[name = string("input_47_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_4_ln2_weight_to_fp16 = const()[name = string("m_layers_4_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109175488)))]; |
| tensor<fp16, [1024]> m_layers_4_ln2_bias_to_fp16 = const()[name = string("m_layers_4_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109177600)))]; |
| tensor<fp16, [1, 104, 1024]> input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = m_layers_4_ln2_bias_to_fp16, epsilon = var_367_to_fp16, gamma = m_layers_4_ln2_weight_to_fp16, x = input_45_cast_fp16)[name = string("input_47_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_4_fc1_weight_to_fp16 = const()[name = string("m_layers_4_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109179712)))]; |
| tensor<fp16, [4096]> m_layers_4_fc1_bias_to_fp16 = const()[name = string("m_layers_4_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117568384)))]; |
| tensor<fp16, [1, 104, 4096]> linear_28_cast_fp16 = linear(bias = m_layers_4_fc1_bias_to_fp16, weight = m_layers_4_fc1_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_28_cast_fp16")]; |
| string input_49_mode_0 = const()[name = string("input_49_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_28_cast_fp16)[name = string("input_49_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_4_fc2_weight_to_fp16 = const()[name = string("m_layers_4_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117576640)))]; |
| tensor<fp16, [1024]> m_layers_4_fc2_bias_to_fp16 = const()[name = string("m_layers_4_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125965312)))]; |
| tensor<fp16, [1, 104, 1024]> linear_29_cast_fp16 = linear(bias = m_layers_4_fc2_bias_to_fp16, weight = m_layers_4_fc2_weight_to_fp16, x = input_49_cast_fp16)[name = string("linear_29_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_11_cast_fp16 = add(x = input_45_cast_fp16, y = linear_29_cast_fp16)[name = string("x_11_cast_fp16")]; |
| int32 var_434 = const()[name = string("op_434"), val = int32(-1)]; |
| tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_5_ln1_weight_to_fp16 = const()[name = string("m_layers_5_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125967424)))]; |
| tensor<fp16, [1024]> m_layers_5_ln1_bias_to_fp16 = const()[name = string("m_layers_5_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125969536)))]; |
| fp16 var_437_to_fp16 = const()[name = string("op_437_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = m_layers_5_ln1_bias_to_fp16, epsilon = var_437_to_fp16, gamma = m_layers_5_ln1_weight_to_fp16, x = x_11_cast_fp16)[name = string("input_51_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_5_q_weight_to_fp16 = const()[name = string("m_layers_5_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125971648)))]; |
| tensor<fp16, [1024]> m_layers_5_q_bias_to_fp16 = const()[name = string("m_layers_5_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128068864)))]; |
| tensor<fp16, [1, 104, 1024]> linear_30_cast_fp16 = linear(bias = m_layers_5_q_bias_to_fp16, weight = m_layers_5_q_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_30_cast_fp16")]; |
| tensor<int32, [4]> var_460 = const()[name = string("op_460"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_461_cast_fp16 = reshape(shape = var_460, x = linear_30_cast_fp16)[name = string("op_461_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_5_k_weight_to_fp16 = const()[name = string("m_layers_5_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(128070976)))]; |
| tensor<fp16, [1024]> m_layers_5_k_bias_to_fp16 = const()[name = string("m_layers_5_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130168192)))]; |
| tensor<fp16, [1, 104, 1024]> linear_31_cast_fp16 = linear(bias = m_layers_5_k_bias_to_fp16, weight = m_layers_5_k_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_31_cast_fp16")]; |
| tensor<int32, [4]> var_466 = const()[name = string("op_466"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_467_cast_fp16 = reshape(shape = var_466, x = linear_31_cast_fp16)[name = string("op_467_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_5_v_weight_to_fp16 = const()[name = string("m_layers_5_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130170304)))]; |
| tensor<fp16, [1024]> m_layers_5_v_bias_to_fp16 = const()[name = string("m_layers_5_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132267520)))]; |
| tensor<fp16, [1, 104, 1024]> linear_32_cast_fp16 = linear(bias = m_layers_5_v_bias_to_fp16, weight = m_layers_5_v_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_32_cast_fp16")]; |
| tensor<int32, [4]> var_472 = const()[name = string("op_472"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_473_cast_fp16 = reshape(shape = var_472, x = linear_32_cast_fp16)[name = string("op_473_cast_fp16")]; |
| tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_475_to_fp16 = const()[name = string("op_475_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_476_cast_fp16 = mul(x = var_461_cast_fp16, y = var_475_to_fp16)[name = string("op_476_cast_fp16")]; |
| bool var_478_transpose_x_0 = const()[name = string("op_478_transpose_x_0"), val = bool(false)]; |
| bool var_478_transpose_y_0 = const()[name = string("op_478_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_82_perm_0 = const()[name = string("transpose_82_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_83_perm_0 = const()[name = string("transpose_83_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_83 = transpose(perm = transpose_83_perm_0, x = var_467_cast_fp16)[name = string("transpose_193")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_82 = transpose(perm = transpose_82_perm_0, x = var_476_cast_fp16)[name = string("transpose_194")]; |
| tensor<fp16, [1, 16, 104, 104]> var_478_cast_fp16 = matmul(transpose_x = var_478_transpose_x_0, transpose_y = var_478_transpose_y_0, x = transpose_82, y = transpose_83)[name = string("op_478_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_479_cast_fp16 = add(x = var_478_cast_fp16, y = key_bias)[name = string("op_479_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_11_cast_fp16 = softmax(axis = var_434, x = var_479_cast_fp16)[name = string("att_11_cast_fp16")]; |
| bool var_481_transpose_x_0 = const()[name = string("op_481_transpose_x_0"), val = bool(false)]; |
| bool var_481_transpose_y_0 = const()[name = string("op_481_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_473_cast_fp16)[name = string("transpose_195")]; |
| tensor<fp16, [1, 16, 104, 64]> var_481_cast_fp16 = matmul(transpose_x = var_481_transpose_x_0, transpose_y = var_481_transpose_y_0, x = att_11_cast_fp16, y = v_11_cast_fp16)[name = string("op_481_cast_fp16")]; |
| tensor<int32, [4]> var_482_perm_0 = const()[name = string("op_482_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_483 = const()[name = string("op_483"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_482_cast_fp16 = transpose(perm = var_482_perm_0, x = var_481_cast_fp16)[name = string("transpose_192")]; |
| tensor<fp16, [1, 104, 1024]> input_53_cast_fp16 = reshape(shape = var_483, x = var_482_cast_fp16)[name = string("input_53_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_5_o_weight_to_fp16 = const()[name = string("m_layers_5_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132269632)))]; |
| tensor<fp16, [1024]> m_layers_5_o_bias_to_fp16 = const()[name = string("m_layers_5_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134366848)))]; |
| tensor<fp16, [1, 104, 1024]> linear_33_cast_fp16 = linear(bias = m_layers_5_o_bias_to_fp16, weight = m_layers_5_o_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_33_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_55_cast_fp16 = add(x = x_11_cast_fp16, y = linear_33_cast_fp16)[name = string("input_55_cast_fp16")]; |
| tensor<int32, [1]> input_57_axes_0 = const()[name = string("input_57_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_5_ln2_weight_to_fp16 = const()[name = string("m_layers_5_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134368960)))]; |
| tensor<fp16, [1024]> m_layers_5_ln2_bias_to_fp16 = const()[name = string("m_layers_5_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134371072)))]; |
| tensor<fp16, [1, 104, 1024]> input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = m_layers_5_ln2_bias_to_fp16, epsilon = var_437_to_fp16, gamma = m_layers_5_ln2_weight_to_fp16, x = input_55_cast_fp16)[name = string("input_57_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_5_fc1_weight_to_fp16 = const()[name = string("m_layers_5_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134373184)))]; |
| tensor<fp16, [4096]> m_layers_5_fc1_bias_to_fp16 = const()[name = string("m_layers_5_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142761856)))]; |
| tensor<fp16, [1, 104, 4096]> linear_34_cast_fp16 = linear(bias = m_layers_5_fc1_bias_to_fp16, weight = m_layers_5_fc1_weight_to_fp16, x = input_57_cast_fp16)[name = string("linear_34_cast_fp16")]; |
| string input_59_mode_0 = const()[name = string("input_59_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_34_cast_fp16)[name = string("input_59_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_5_fc2_weight_to_fp16 = const()[name = string("m_layers_5_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142770112)))]; |
| tensor<fp16, [1024]> m_layers_5_fc2_bias_to_fp16 = const()[name = string("m_layers_5_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151158784)))]; |
| tensor<fp16, [1, 104, 1024]> linear_35_cast_fp16 = linear(bias = m_layers_5_fc2_bias_to_fp16, weight = m_layers_5_fc2_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_35_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_13_cast_fp16 = add(x = input_55_cast_fp16, y = linear_35_cast_fp16)[name = string("x_13_cast_fp16")]; |
| int32 var_504 = const()[name = string("op_504"), val = int32(-1)]; |
| tensor<int32, [1]> input_61_axes_0 = const()[name = string("input_61_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_6_ln1_weight_to_fp16 = const()[name = string("m_layers_6_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151160896)))]; |
| tensor<fp16, [1024]> m_layers_6_ln1_bias_to_fp16 = const()[name = string("m_layers_6_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151163008)))]; |
| fp16 var_507_to_fp16 = const()[name = string("op_507_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_61_cast_fp16 = layer_norm(axes = input_61_axes_0, beta = m_layers_6_ln1_bias_to_fp16, epsilon = var_507_to_fp16, gamma = m_layers_6_ln1_weight_to_fp16, x = x_13_cast_fp16)[name = string("input_61_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_6_q_weight_to_fp16 = const()[name = string("m_layers_6_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151165120)))]; |
| tensor<fp16, [1024]> m_layers_6_q_bias_to_fp16 = const()[name = string("m_layers_6_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153262336)))]; |
| tensor<fp16, [1, 104, 1024]> linear_36_cast_fp16 = linear(bias = m_layers_6_q_bias_to_fp16, weight = m_layers_6_q_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_36_cast_fp16")]; |
| tensor<int32, [4]> var_530 = const()[name = string("op_530"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_531_cast_fp16 = reshape(shape = var_530, x = linear_36_cast_fp16)[name = string("op_531_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_6_k_weight_to_fp16 = const()[name = string("m_layers_6_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(153264448)))]; |
| tensor<fp16, [1024]> m_layers_6_k_bias_to_fp16 = const()[name = string("m_layers_6_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155361664)))]; |
| tensor<fp16, [1, 104, 1024]> linear_37_cast_fp16 = linear(bias = m_layers_6_k_bias_to_fp16, weight = m_layers_6_k_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_37_cast_fp16")]; |
| tensor<int32, [4]> var_536 = const()[name = string("op_536"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_537_cast_fp16 = reshape(shape = var_536, x = linear_37_cast_fp16)[name = string("op_537_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_6_v_weight_to_fp16 = const()[name = string("m_layers_6_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(155363776)))]; |
| tensor<fp16, [1024]> m_layers_6_v_bias_to_fp16 = const()[name = string("m_layers_6_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157460992)))]; |
| tensor<fp16, [1, 104, 1024]> linear_38_cast_fp16 = linear(bias = m_layers_6_v_bias_to_fp16, weight = m_layers_6_v_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_38_cast_fp16")]; |
| tensor<int32, [4]> var_542 = const()[name = string("op_542"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_543_cast_fp16 = reshape(shape = var_542, x = linear_38_cast_fp16)[name = string("op_543_cast_fp16")]; |
| tensor<int32, [4]> v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_545_to_fp16 = const()[name = string("op_545_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_546_cast_fp16 = mul(x = var_531_cast_fp16, y = var_545_to_fp16)[name = string("op_546_cast_fp16")]; |
| bool var_548_transpose_x_0 = const()[name = string("op_548_transpose_x_0"), val = bool(false)]; |
| bool var_548_transpose_y_0 = const()[name = string("op_548_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_84_perm_0 = const()[name = string("transpose_84_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_85_perm_0 = const()[name = string("transpose_85_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_85 = transpose(perm = transpose_85_perm_0, x = var_537_cast_fp16)[name = string("transpose_189")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_84 = transpose(perm = transpose_84_perm_0, x = var_546_cast_fp16)[name = string("transpose_190")]; |
| tensor<fp16, [1, 16, 104, 104]> var_548_cast_fp16 = matmul(transpose_x = var_548_transpose_x_0, transpose_y = var_548_transpose_y_0, x = transpose_84, y = transpose_85)[name = string("op_548_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_549_cast_fp16 = add(x = var_548_cast_fp16, y = key_bias)[name = string("op_549_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_13_cast_fp16 = softmax(axis = var_504, x = var_549_cast_fp16)[name = string("att_13_cast_fp16")]; |
| bool var_551_transpose_x_0 = const()[name = string("op_551_transpose_x_0"), val = bool(false)]; |
| bool var_551_transpose_y_0 = const()[name = string("op_551_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_543_cast_fp16)[name = string("transpose_191")]; |
| tensor<fp16, [1, 16, 104, 64]> var_551_cast_fp16 = matmul(transpose_x = var_551_transpose_x_0, transpose_y = var_551_transpose_y_0, x = att_13_cast_fp16, y = v_13_cast_fp16)[name = string("op_551_cast_fp16")]; |
| tensor<int32, [4]> var_552_perm_0 = const()[name = string("op_552_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_553 = const()[name = string("op_553"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_552_cast_fp16 = transpose(perm = var_552_perm_0, x = var_551_cast_fp16)[name = string("transpose_188")]; |
| tensor<fp16, [1, 104, 1024]> input_63_cast_fp16 = reshape(shape = var_553, x = var_552_cast_fp16)[name = string("input_63_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_6_o_weight_to_fp16 = const()[name = string("m_layers_6_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(157463104)))]; |
| tensor<fp16, [1024]> m_layers_6_o_bias_to_fp16 = const()[name = string("m_layers_6_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159560320)))]; |
| tensor<fp16, [1, 104, 1024]> linear_39_cast_fp16 = linear(bias = m_layers_6_o_bias_to_fp16, weight = m_layers_6_o_weight_to_fp16, x = input_63_cast_fp16)[name = string("linear_39_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_65_cast_fp16 = add(x = x_13_cast_fp16, y = linear_39_cast_fp16)[name = string("input_65_cast_fp16")]; |
| tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_6_ln2_weight_to_fp16 = const()[name = string("m_layers_6_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159562432)))]; |
| tensor<fp16, [1024]> m_layers_6_ln2_bias_to_fp16 = const()[name = string("m_layers_6_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159564544)))]; |
| tensor<fp16, [1, 104, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = m_layers_6_ln2_bias_to_fp16, epsilon = var_507_to_fp16, gamma = m_layers_6_ln2_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_6_fc1_weight_to_fp16 = const()[name = string("m_layers_6_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159566656)))]; |
| tensor<fp16, [4096]> m_layers_6_fc1_bias_to_fp16 = const()[name = string("m_layers_6_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167955328)))]; |
| tensor<fp16, [1, 104, 4096]> linear_40_cast_fp16 = linear(bias = m_layers_6_fc1_bias_to_fp16, weight = m_layers_6_fc1_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_40_cast_fp16")]; |
| string input_69_mode_0 = const()[name = string("input_69_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_69_cast_fp16 = gelu(mode = input_69_mode_0, x = linear_40_cast_fp16)[name = string("input_69_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_6_fc2_weight_to_fp16 = const()[name = string("m_layers_6_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167963584)))]; |
| tensor<fp16, [1024]> m_layers_6_fc2_bias_to_fp16 = const()[name = string("m_layers_6_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176352256)))]; |
| tensor<fp16, [1, 104, 1024]> linear_41_cast_fp16 = linear(bias = m_layers_6_fc2_bias_to_fp16, weight = m_layers_6_fc2_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_41_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_15_cast_fp16 = add(x = input_65_cast_fp16, y = linear_41_cast_fp16)[name = string("x_15_cast_fp16")]; |
| int32 var_574 = const()[name = string("op_574"), val = int32(-1)]; |
| tensor<int32, [1]> input_71_axes_0 = const()[name = string("input_71_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_7_ln1_weight_to_fp16 = const()[name = string("m_layers_7_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176354368)))]; |
| tensor<fp16, [1024]> m_layers_7_ln1_bias_to_fp16 = const()[name = string("m_layers_7_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176356480)))]; |
| fp16 var_577_to_fp16 = const()[name = string("op_577_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = m_layers_7_ln1_bias_to_fp16, epsilon = var_577_to_fp16, gamma = m_layers_7_ln1_weight_to_fp16, x = x_15_cast_fp16)[name = string("input_71_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_7_q_weight_to_fp16 = const()[name = string("m_layers_7_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176358592)))]; |
| tensor<fp16, [1024]> m_layers_7_q_bias_to_fp16 = const()[name = string("m_layers_7_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178455808)))]; |
| tensor<fp16, [1, 104, 1024]> linear_42_cast_fp16 = linear(bias = m_layers_7_q_bias_to_fp16, weight = m_layers_7_q_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_42_cast_fp16")]; |
| tensor<int32, [4]> var_600 = const()[name = string("op_600"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_601_cast_fp16 = reshape(shape = var_600, x = linear_42_cast_fp16)[name = string("op_601_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_7_k_weight_to_fp16 = const()[name = string("m_layers_7_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(178457920)))]; |
| tensor<fp16, [1024]> m_layers_7_k_bias_to_fp16 = const()[name = string("m_layers_7_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180555136)))]; |
| tensor<fp16, [1, 104, 1024]> linear_43_cast_fp16 = linear(bias = m_layers_7_k_bias_to_fp16, weight = m_layers_7_k_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_43_cast_fp16")]; |
| tensor<int32, [4]> var_606 = const()[name = string("op_606"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_607_cast_fp16 = reshape(shape = var_606, x = linear_43_cast_fp16)[name = string("op_607_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_7_v_weight_to_fp16 = const()[name = string("m_layers_7_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(180557248)))]; |
| tensor<fp16, [1024]> m_layers_7_v_bias_to_fp16 = const()[name = string("m_layers_7_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182654464)))]; |
| tensor<fp16, [1, 104, 1024]> linear_44_cast_fp16 = linear(bias = m_layers_7_v_bias_to_fp16, weight = m_layers_7_v_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_44_cast_fp16")]; |
| tensor<int32, [4]> var_612 = const()[name = string("op_612"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_613_cast_fp16 = reshape(shape = var_612, x = linear_44_cast_fp16)[name = string("op_613_cast_fp16")]; |
| tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_615_to_fp16 = const()[name = string("op_615_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_616_cast_fp16 = mul(x = var_601_cast_fp16, y = var_615_to_fp16)[name = string("op_616_cast_fp16")]; |
| bool var_618_transpose_x_0 = const()[name = string("op_618_transpose_x_0"), val = bool(false)]; |
| bool var_618_transpose_y_0 = const()[name = string("op_618_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_86_perm_0 = const()[name = string("transpose_86_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_87_perm_0 = const()[name = string("transpose_87_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_87 = transpose(perm = transpose_87_perm_0, x = var_607_cast_fp16)[name = string("transpose_185")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_86 = transpose(perm = transpose_86_perm_0, x = var_616_cast_fp16)[name = string("transpose_186")]; |
| tensor<fp16, [1, 16, 104, 104]> var_618_cast_fp16 = matmul(transpose_x = var_618_transpose_x_0, transpose_y = var_618_transpose_y_0, x = transpose_86, y = transpose_87)[name = string("op_618_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_619_cast_fp16 = add(x = var_618_cast_fp16, y = key_bias)[name = string("op_619_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_15_cast_fp16 = softmax(axis = var_574, x = var_619_cast_fp16)[name = string("att_15_cast_fp16")]; |
| bool var_621_transpose_x_0 = const()[name = string("op_621_transpose_x_0"), val = bool(false)]; |
| bool var_621_transpose_y_0 = const()[name = string("op_621_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_613_cast_fp16)[name = string("transpose_187")]; |
| tensor<fp16, [1, 16, 104, 64]> var_621_cast_fp16 = matmul(transpose_x = var_621_transpose_x_0, transpose_y = var_621_transpose_y_0, x = att_15_cast_fp16, y = v_15_cast_fp16)[name = string("op_621_cast_fp16")]; |
| tensor<int32, [4]> var_622_perm_0 = const()[name = string("op_622_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_623 = const()[name = string("op_623"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_622_cast_fp16 = transpose(perm = var_622_perm_0, x = var_621_cast_fp16)[name = string("transpose_184")]; |
| tensor<fp16, [1, 104, 1024]> input_73_cast_fp16 = reshape(shape = var_623, x = var_622_cast_fp16)[name = string("input_73_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_7_o_weight_to_fp16 = const()[name = string("m_layers_7_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(182656576)))]; |
| tensor<fp16, [1024]> m_layers_7_o_bias_to_fp16 = const()[name = string("m_layers_7_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184753792)))]; |
| tensor<fp16, [1, 104, 1024]> linear_45_cast_fp16 = linear(bias = m_layers_7_o_bias_to_fp16, weight = m_layers_7_o_weight_to_fp16, x = input_73_cast_fp16)[name = string("linear_45_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_75_cast_fp16 = add(x = x_15_cast_fp16, y = linear_45_cast_fp16)[name = string("input_75_cast_fp16")]; |
| tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_7_ln2_weight_to_fp16 = const()[name = string("m_layers_7_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184755904)))]; |
| tensor<fp16, [1024]> m_layers_7_ln2_bias_to_fp16 = const()[name = string("m_layers_7_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184758016)))]; |
| tensor<fp16, [1, 104, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = m_layers_7_ln2_bias_to_fp16, epsilon = var_577_to_fp16, gamma = m_layers_7_ln2_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_7_fc1_weight_to_fp16 = const()[name = string("m_layers_7_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184760128)))]; |
| tensor<fp16, [4096]> m_layers_7_fc1_bias_to_fp16 = const()[name = string("m_layers_7_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193148800)))]; |
| tensor<fp16, [1, 104, 4096]> linear_46_cast_fp16 = linear(bias = m_layers_7_fc1_bias_to_fp16, weight = m_layers_7_fc1_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_46_cast_fp16")]; |
| string input_79_mode_0 = const()[name = string("input_79_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_79_cast_fp16 = gelu(mode = input_79_mode_0, x = linear_46_cast_fp16)[name = string("input_79_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_7_fc2_weight_to_fp16 = const()[name = string("m_layers_7_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(193157056)))]; |
| tensor<fp16, [1024]> m_layers_7_fc2_bias_to_fp16 = const()[name = string("m_layers_7_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201545728)))]; |
| tensor<fp16, [1, 104, 1024]> linear_47_cast_fp16 = linear(bias = m_layers_7_fc2_bias_to_fp16, weight = m_layers_7_fc2_weight_to_fp16, x = input_79_cast_fp16)[name = string("linear_47_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_17_cast_fp16 = add(x = input_75_cast_fp16, y = linear_47_cast_fp16)[name = string("x_17_cast_fp16")]; |
| int32 var_644 = const()[name = string("op_644"), val = int32(-1)]; |
| tensor<int32, [1]> input_81_axes_0 = const()[name = string("input_81_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_8_ln1_weight_to_fp16 = const()[name = string("m_layers_8_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201547840)))]; |
| tensor<fp16, [1024]> m_layers_8_ln1_bias_to_fp16 = const()[name = string("m_layers_8_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201549952)))]; |
| fp16 var_647_to_fp16 = const()[name = string("op_647_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = m_layers_8_ln1_bias_to_fp16, epsilon = var_647_to_fp16, gamma = m_layers_8_ln1_weight_to_fp16, x = x_17_cast_fp16)[name = string("input_81_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_8_q_weight_to_fp16 = const()[name = string("m_layers_8_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201552064)))]; |
| tensor<fp16, [1024]> m_layers_8_q_bias_to_fp16 = const()[name = string("m_layers_8_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203649280)))]; |
| tensor<fp16, [1, 104, 1024]> linear_48_cast_fp16 = linear(bias = m_layers_8_q_bias_to_fp16, weight = m_layers_8_q_weight_to_fp16, x = input_81_cast_fp16)[name = string("linear_48_cast_fp16")]; |
| tensor<int32, [4]> var_670 = const()[name = string("op_670"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_671_cast_fp16 = reshape(shape = var_670, x = linear_48_cast_fp16)[name = string("op_671_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_8_k_weight_to_fp16 = const()[name = string("m_layers_8_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(203651392)))]; |
| tensor<fp16, [1024]> m_layers_8_k_bias_to_fp16 = const()[name = string("m_layers_8_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205748608)))]; |
| tensor<fp16, [1, 104, 1024]> linear_49_cast_fp16 = linear(bias = m_layers_8_k_bias_to_fp16, weight = m_layers_8_k_weight_to_fp16, x = input_81_cast_fp16)[name = string("linear_49_cast_fp16")]; |
| tensor<int32, [4]> var_676 = const()[name = string("op_676"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_677_cast_fp16 = reshape(shape = var_676, x = linear_49_cast_fp16)[name = string("op_677_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_8_v_weight_to_fp16 = const()[name = string("m_layers_8_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(205750720)))]; |
| tensor<fp16, [1024]> m_layers_8_v_bias_to_fp16 = const()[name = string("m_layers_8_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207847936)))]; |
| tensor<fp16, [1, 104, 1024]> linear_50_cast_fp16 = linear(bias = m_layers_8_v_bias_to_fp16, weight = m_layers_8_v_weight_to_fp16, x = input_81_cast_fp16)[name = string("linear_50_cast_fp16")]; |
| tensor<int32, [4]> var_682 = const()[name = string("op_682"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_683_cast_fp16 = reshape(shape = var_682, x = linear_50_cast_fp16)[name = string("op_683_cast_fp16")]; |
| tensor<int32, [4]> v_17_perm_0 = const()[name = string("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_685_to_fp16 = const()[name = string("op_685_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_686_cast_fp16 = mul(x = var_671_cast_fp16, y = var_685_to_fp16)[name = string("op_686_cast_fp16")]; |
| bool var_688_transpose_x_0 = const()[name = string("op_688_transpose_x_0"), val = bool(false)]; |
| bool var_688_transpose_y_0 = const()[name = string("op_688_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_88_perm_0 = const()[name = string("transpose_88_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_89_perm_0 = const()[name = string("transpose_89_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_89 = transpose(perm = transpose_89_perm_0, x = var_677_cast_fp16)[name = string("transpose_181")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_88 = transpose(perm = transpose_88_perm_0, x = var_686_cast_fp16)[name = string("transpose_182")]; |
| tensor<fp16, [1, 16, 104, 104]> var_688_cast_fp16 = matmul(transpose_x = var_688_transpose_x_0, transpose_y = var_688_transpose_y_0, x = transpose_88, y = transpose_89)[name = string("op_688_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_689_cast_fp16 = add(x = var_688_cast_fp16, y = key_bias)[name = string("op_689_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_17_cast_fp16 = softmax(axis = var_644, x = var_689_cast_fp16)[name = string("att_17_cast_fp16")]; |
| bool var_691_transpose_x_0 = const()[name = string("op_691_transpose_x_0"), val = bool(false)]; |
| bool var_691_transpose_y_0 = const()[name = string("op_691_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_683_cast_fp16)[name = string("transpose_183")]; |
| tensor<fp16, [1, 16, 104, 64]> var_691_cast_fp16 = matmul(transpose_x = var_691_transpose_x_0, transpose_y = var_691_transpose_y_0, x = att_17_cast_fp16, y = v_17_cast_fp16)[name = string("op_691_cast_fp16")]; |
| tensor<int32, [4]> var_692_perm_0 = const()[name = string("op_692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_693 = const()[name = string("op_693"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_692_cast_fp16 = transpose(perm = var_692_perm_0, x = var_691_cast_fp16)[name = string("transpose_180")]; |
| tensor<fp16, [1, 104, 1024]> input_83_cast_fp16 = reshape(shape = var_693, x = var_692_cast_fp16)[name = string("input_83_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_8_o_weight_to_fp16 = const()[name = string("m_layers_8_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(207850048)))]; |
| tensor<fp16, [1024]> m_layers_8_o_bias_to_fp16 = const()[name = string("m_layers_8_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209947264)))]; |
| tensor<fp16, [1, 104, 1024]> linear_51_cast_fp16 = linear(bias = m_layers_8_o_bias_to_fp16, weight = m_layers_8_o_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_51_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_85_cast_fp16 = add(x = x_17_cast_fp16, y = linear_51_cast_fp16)[name = string("input_85_cast_fp16")]; |
| tensor<int32, [1]> input_87_axes_0 = const()[name = string("input_87_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_8_ln2_weight_to_fp16 = const()[name = string("m_layers_8_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209949376)))]; |
| tensor<fp16, [1024]> m_layers_8_ln2_bias_to_fp16 = const()[name = string("m_layers_8_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209951488)))]; |
| tensor<fp16, [1, 104, 1024]> input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = m_layers_8_ln2_bias_to_fp16, epsilon = var_647_to_fp16, gamma = m_layers_8_ln2_weight_to_fp16, x = input_85_cast_fp16)[name = string("input_87_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_8_fc1_weight_to_fp16 = const()[name = string("m_layers_8_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209953600)))]; |
| tensor<fp16, [4096]> m_layers_8_fc1_bias_to_fp16 = const()[name = string("m_layers_8_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218342272)))]; |
| tensor<fp16, [1, 104, 4096]> linear_52_cast_fp16 = linear(bias = m_layers_8_fc1_bias_to_fp16, weight = m_layers_8_fc1_weight_to_fp16, x = input_87_cast_fp16)[name = string("linear_52_cast_fp16")]; |
| string input_89_mode_0 = const()[name = string("input_89_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_89_cast_fp16 = gelu(mode = input_89_mode_0, x = linear_52_cast_fp16)[name = string("input_89_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_8_fc2_weight_to_fp16 = const()[name = string("m_layers_8_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218350528)))]; |
| tensor<fp16, [1024]> m_layers_8_fc2_bias_to_fp16 = const()[name = string("m_layers_8_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226739200)))]; |
| tensor<fp16, [1, 104, 1024]> linear_53_cast_fp16 = linear(bias = m_layers_8_fc2_bias_to_fp16, weight = m_layers_8_fc2_weight_to_fp16, x = input_89_cast_fp16)[name = string("linear_53_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_19_cast_fp16 = add(x = input_85_cast_fp16, y = linear_53_cast_fp16)[name = string("x_19_cast_fp16")]; |
| int32 var_714 = const()[name = string("op_714"), val = int32(-1)]; |
| tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_9_ln1_weight_to_fp16 = const()[name = string("m_layers_9_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226741312)))]; |
| tensor<fp16, [1024]> m_layers_9_ln1_bias_to_fp16 = const()[name = string("m_layers_9_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226743424)))]; |
| fp16 var_717_to_fp16 = const()[name = string("op_717_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = m_layers_9_ln1_bias_to_fp16, epsilon = var_717_to_fp16, gamma = m_layers_9_ln1_weight_to_fp16, x = x_19_cast_fp16)[name = string("input_91_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_9_q_weight_to_fp16 = const()[name = string("m_layers_9_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226745536)))]; |
| tensor<fp16, [1024]> m_layers_9_q_bias_to_fp16 = const()[name = string("m_layers_9_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228842752)))]; |
| tensor<fp16, [1, 104, 1024]> linear_54_cast_fp16 = linear(bias = m_layers_9_q_bias_to_fp16, weight = m_layers_9_q_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_54_cast_fp16")]; |
| tensor<int32, [4]> var_740 = const()[name = string("op_740"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_741_cast_fp16 = reshape(shape = var_740, x = linear_54_cast_fp16)[name = string("op_741_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_9_k_weight_to_fp16 = const()[name = string("m_layers_9_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228844864)))]; |
| tensor<fp16, [1024]> m_layers_9_k_bias_to_fp16 = const()[name = string("m_layers_9_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230942080)))]; |
| tensor<fp16, [1, 104, 1024]> linear_55_cast_fp16 = linear(bias = m_layers_9_k_bias_to_fp16, weight = m_layers_9_k_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_55_cast_fp16")]; |
| tensor<int32, [4]> var_746 = const()[name = string("op_746"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_747_cast_fp16 = reshape(shape = var_746, x = linear_55_cast_fp16)[name = string("op_747_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_9_v_weight_to_fp16 = const()[name = string("m_layers_9_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230944192)))]; |
| tensor<fp16, [1024]> m_layers_9_v_bias_to_fp16 = const()[name = string("m_layers_9_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233041408)))]; |
| tensor<fp16, [1, 104, 1024]> linear_56_cast_fp16 = linear(bias = m_layers_9_v_bias_to_fp16, weight = m_layers_9_v_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_56_cast_fp16")]; |
| tensor<int32, [4]> var_752 = const()[name = string("op_752"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_753_cast_fp16 = reshape(shape = var_752, x = linear_56_cast_fp16)[name = string("op_753_cast_fp16")]; |
| tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_755_to_fp16 = const()[name = string("op_755_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_756_cast_fp16 = mul(x = var_741_cast_fp16, y = var_755_to_fp16)[name = string("op_756_cast_fp16")]; |
| bool var_758_transpose_x_0 = const()[name = string("op_758_transpose_x_0"), val = bool(false)]; |
| bool var_758_transpose_y_0 = const()[name = string("op_758_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_90_perm_0 = const()[name = string("transpose_90_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_91_perm_0 = const()[name = string("transpose_91_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_91 = transpose(perm = transpose_91_perm_0, x = var_747_cast_fp16)[name = string("transpose_177")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_90 = transpose(perm = transpose_90_perm_0, x = var_756_cast_fp16)[name = string("transpose_178")]; |
| tensor<fp16, [1, 16, 104, 104]> var_758_cast_fp16 = matmul(transpose_x = var_758_transpose_x_0, transpose_y = var_758_transpose_y_0, x = transpose_90, y = transpose_91)[name = string("op_758_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_759_cast_fp16 = add(x = var_758_cast_fp16, y = key_bias)[name = string("op_759_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_19_cast_fp16 = softmax(axis = var_714, x = var_759_cast_fp16)[name = string("att_19_cast_fp16")]; |
| bool var_761_transpose_x_0 = const()[name = string("op_761_transpose_x_0"), val = bool(false)]; |
| bool var_761_transpose_y_0 = const()[name = string("op_761_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_753_cast_fp16)[name = string("transpose_179")]; |
| tensor<fp16, [1, 16, 104, 64]> var_761_cast_fp16 = matmul(transpose_x = var_761_transpose_x_0, transpose_y = var_761_transpose_y_0, x = att_19_cast_fp16, y = v_19_cast_fp16)[name = string("op_761_cast_fp16")]; |
| tensor<int32, [4]> var_762_perm_0 = const()[name = string("op_762_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_763 = const()[name = string("op_763"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_762_cast_fp16 = transpose(perm = var_762_perm_0, x = var_761_cast_fp16)[name = string("transpose_176")]; |
| tensor<fp16, [1, 104, 1024]> input_93_cast_fp16 = reshape(shape = var_763, x = var_762_cast_fp16)[name = string("input_93_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_9_o_weight_to_fp16 = const()[name = string("m_layers_9_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(233043520)))]; |
| tensor<fp16, [1024]> m_layers_9_o_bias_to_fp16 = const()[name = string("m_layers_9_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235140736)))]; |
| tensor<fp16, [1, 104, 1024]> linear_57_cast_fp16 = linear(bias = m_layers_9_o_bias_to_fp16, weight = m_layers_9_o_weight_to_fp16, x = input_93_cast_fp16)[name = string("linear_57_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_95_cast_fp16 = add(x = x_19_cast_fp16, y = linear_57_cast_fp16)[name = string("input_95_cast_fp16")]; |
| tensor<int32, [1]> input_97_axes_0 = const()[name = string("input_97_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_9_ln2_weight_to_fp16 = const()[name = string("m_layers_9_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235142848)))]; |
| tensor<fp16, [1024]> m_layers_9_ln2_bias_to_fp16 = const()[name = string("m_layers_9_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235144960)))]; |
| tensor<fp16, [1, 104, 1024]> input_97_cast_fp16 = layer_norm(axes = input_97_axes_0, beta = m_layers_9_ln2_bias_to_fp16, epsilon = var_717_to_fp16, gamma = m_layers_9_ln2_weight_to_fp16, x = input_95_cast_fp16)[name = string("input_97_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_9_fc1_weight_to_fp16 = const()[name = string("m_layers_9_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(235147072)))]; |
| tensor<fp16, [4096]> m_layers_9_fc1_bias_to_fp16 = const()[name = string("m_layers_9_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243535744)))]; |
| tensor<fp16, [1, 104, 4096]> linear_58_cast_fp16 = linear(bias = m_layers_9_fc1_bias_to_fp16, weight = m_layers_9_fc1_weight_to_fp16, x = input_97_cast_fp16)[name = string("linear_58_cast_fp16")]; |
| string input_99_mode_0 = const()[name = string("input_99_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_99_cast_fp16 = gelu(mode = input_99_mode_0, x = linear_58_cast_fp16)[name = string("input_99_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_9_fc2_weight_to_fp16 = const()[name = string("m_layers_9_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243544000)))]; |
| tensor<fp16, [1024]> m_layers_9_fc2_bias_to_fp16 = const()[name = string("m_layers_9_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251932672)))]; |
| tensor<fp16, [1, 104, 1024]> linear_59_cast_fp16 = linear(bias = m_layers_9_fc2_bias_to_fp16, weight = m_layers_9_fc2_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_59_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_21_cast_fp16 = add(x = input_95_cast_fp16, y = linear_59_cast_fp16)[name = string("x_21_cast_fp16")]; |
| int32 var_784 = const()[name = string("op_784"), val = int32(-1)]; |
| tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_10_ln1_weight_to_fp16 = const()[name = string("m_layers_10_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251934784)))]; |
| tensor<fp16, [1024]> m_layers_10_ln1_bias_to_fp16 = const()[name = string("m_layers_10_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251936896)))]; |
| fp16 var_787_to_fp16 = const()[name = string("op_787_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = m_layers_10_ln1_bias_to_fp16, epsilon = var_787_to_fp16, gamma = m_layers_10_ln1_weight_to_fp16, x = x_21_cast_fp16)[name = string("input_101_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_10_q_weight_to_fp16 = const()[name = string("m_layers_10_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251939008)))]; |
| tensor<fp16, [1024]> m_layers_10_q_bias_to_fp16 = const()[name = string("m_layers_10_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254036224)))]; |
| tensor<fp16, [1, 104, 1024]> linear_60_cast_fp16 = linear(bias = m_layers_10_q_bias_to_fp16, weight = m_layers_10_q_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_60_cast_fp16")]; |
| tensor<int32, [4]> var_810 = const()[name = string("op_810"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_811_cast_fp16 = reshape(shape = var_810, x = linear_60_cast_fp16)[name = string("op_811_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_10_k_weight_to_fp16 = const()[name = string("m_layers_10_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(254038336)))]; |
| tensor<fp16, [1024]> m_layers_10_k_bias_to_fp16 = const()[name = string("m_layers_10_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256135552)))]; |
| tensor<fp16, [1, 104, 1024]> linear_61_cast_fp16 = linear(bias = m_layers_10_k_bias_to_fp16, weight = m_layers_10_k_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_61_cast_fp16")]; |
| tensor<int32, [4]> var_816 = const()[name = string("op_816"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_817_cast_fp16 = reshape(shape = var_816, x = linear_61_cast_fp16)[name = string("op_817_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_10_v_weight_to_fp16 = const()[name = string("m_layers_10_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(256137664)))]; |
| tensor<fp16, [1024]> m_layers_10_v_bias_to_fp16 = const()[name = string("m_layers_10_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258234880)))]; |
| tensor<fp16, [1, 104, 1024]> linear_62_cast_fp16 = linear(bias = m_layers_10_v_bias_to_fp16, weight = m_layers_10_v_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_62_cast_fp16")]; |
| tensor<int32, [4]> var_822 = const()[name = string("op_822"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_823_cast_fp16 = reshape(shape = var_822, x = linear_62_cast_fp16)[name = string("op_823_cast_fp16")]; |
| tensor<int32, [4]> v_21_perm_0 = const()[name = string("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_825_to_fp16 = const()[name = string("op_825_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_826_cast_fp16 = mul(x = var_811_cast_fp16, y = var_825_to_fp16)[name = string("op_826_cast_fp16")]; |
| bool var_828_transpose_x_0 = const()[name = string("op_828_transpose_x_0"), val = bool(false)]; |
| bool var_828_transpose_y_0 = const()[name = string("op_828_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_92_perm_0 = const()[name = string("transpose_92_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_93_perm_0 = const()[name = string("transpose_93_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_93 = transpose(perm = transpose_93_perm_0, x = var_817_cast_fp16)[name = string("transpose_173")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_92 = transpose(perm = transpose_92_perm_0, x = var_826_cast_fp16)[name = string("transpose_174")]; |
| tensor<fp16, [1, 16, 104, 104]> var_828_cast_fp16 = matmul(transpose_x = var_828_transpose_x_0, transpose_y = var_828_transpose_y_0, x = transpose_92, y = transpose_93)[name = string("op_828_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_829_cast_fp16 = add(x = var_828_cast_fp16, y = key_bias)[name = string("op_829_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_21_cast_fp16 = softmax(axis = var_784, x = var_829_cast_fp16)[name = string("att_21_cast_fp16")]; |
| bool var_831_transpose_x_0 = const()[name = string("op_831_transpose_x_0"), val = bool(false)]; |
| bool var_831_transpose_y_0 = const()[name = string("op_831_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_823_cast_fp16)[name = string("transpose_175")]; |
| tensor<fp16, [1, 16, 104, 64]> var_831_cast_fp16 = matmul(transpose_x = var_831_transpose_x_0, transpose_y = var_831_transpose_y_0, x = att_21_cast_fp16, y = v_21_cast_fp16)[name = string("op_831_cast_fp16")]; |
| tensor<int32, [4]> var_832_perm_0 = const()[name = string("op_832_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_833 = const()[name = string("op_833"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_832_cast_fp16 = transpose(perm = var_832_perm_0, x = var_831_cast_fp16)[name = string("transpose_172")]; |
| tensor<fp16, [1, 104, 1024]> input_103_cast_fp16 = reshape(shape = var_833, x = var_832_cast_fp16)[name = string("input_103_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_10_o_weight_to_fp16 = const()[name = string("m_layers_10_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(258236992)))]; |
| tensor<fp16, [1024]> m_layers_10_o_bias_to_fp16 = const()[name = string("m_layers_10_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260334208)))]; |
| tensor<fp16, [1, 104, 1024]> linear_63_cast_fp16 = linear(bias = m_layers_10_o_bias_to_fp16, weight = m_layers_10_o_weight_to_fp16, x = input_103_cast_fp16)[name = string("linear_63_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_105_cast_fp16 = add(x = x_21_cast_fp16, y = linear_63_cast_fp16)[name = string("input_105_cast_fp16")]; |
| tensor<int32, [1]> input_107_axes_0 = const()[name = string("input_107_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_10_ln2_weight_to_fp16 = const()[name = string("m_layers_10_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260336320)))]; |
| tensor<fp16, [1024]> m_layers_10_ln2_bias_to_fp16 = const()[name = string("m_layers_10_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260338432)))]; |
| tensor<fp16, [1, 104, 1024]> input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = m_layers_10_ln2_bias_to_fp16, epsilon = var_787_to_fp16, gamma = m_layers_10_ln2_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_10_fc1_weight_to_fp16 = const()[name = string("m_layers_10_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260340544)))]; |
| tensor<fp16, [4096]> m_layers_10_fc1_bias_to_fp16 = const()[name = string("m_layers_10_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268729216)))]; |
| tensor<fp16, [1, 104, 4096]> linear_64_cast_fp16 = linear(bias = m_layers_10_fc1_bias_to_fp16, weight = m_layers_10_fc1_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_64_cast_fp16")]; |
| string input_109_mode_0 = const()[name = string("input_109_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_64_cast_fp16)[name = string("input_109_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_10_fc2_weight_to_fp16 = const()[name = string("m_layers_10_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268737472)))]; |
| tensor<fp16, [1024]> m_layers_10_fc2_bias_to_fp16 = const()[name = string("m_layers_10_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277126144)))]; |
| tensor<fp16, [1, 104, 1024]> linear_65_cast_fp16 = linear(bias = m_layers_10_fc2_bias_to_fp16, weight = m_layers_10_fc2_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_65_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_23_cast_fp16 = add(x = input_105_cast_fp16, y = linear_65_cast_fp16)[name = string("x_23_cast_fp16")]; |
| int32 var_854 = const()[name = string("op_854"), val = int32(-1)]; |
| tensor<int32, [1]> input_111_axes_0 = const()[name = string("input_111_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_11_ln1_weight_to_fp16 = const()[name = string("m_layers_11_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277128256)))]; |
| tensor<fp16, [1024]> m_layers_11_ln1_bias_to_fp16 = const()[name = string("m_layers_11_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277130368)))]; |
| fp16 var_857_to_fp16 = const()[name = string("op_857_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = m_layers_11_ln1_bias_to_fp16, epsilon = var_857_to_fp16, gamma = m_layers_11_ln1_weight_to_fp16, x = x_23_cast_fp16)[name = string("input_111_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_11_q_weight_to_fp16 = const()[name = string("m_layers_11_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(277132480)))]; |
| tensor<fp16, [1024]> m_layers_11_q_bias_to_fp16 = const()[name = string("m_layers_11_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279229696)))]; |
| tensor<fp16, [1, 104, 1024]> linear_66_cast_fp16 = linear(bias = m_layers_11_q_bias_to_fp16, weight = m_layers_11_q_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_66_cast_fp16")]; |
| tensor<int32, [4]> var_880 = const()[name = string("op_880"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_881_cast_fp16 = reshape(shape = var_880, x = linear_66_cast_fp16)[name = string("op_881_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_11_k_weight_to_fp16 = const()[name = string("m_layers_11_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(279231808)))]; |
| tensor<fp16, [1024]> m_layers_11_k_bias_to_fp16 = const()[name = string("m_layers_11_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281329024)))]; |
| tensor<fp16, [1, 104, 1024]> linear_67_cast_fp16 = linear(bias = m_layers_11_k_bias_to_fp16, weight = m_layers_11_k_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_67_cast_fp16")]; |
| tensor<int32, [4]> var_886 = const()[name = string("op_886"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_887_cast_fp16 = reshape(shape = var_886, x = linear_67_cast_fp16)[name = string("op_887_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_11_v_weight_to_fp16 = const()[name = string("m_layers_11_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(281331136)))]; |
| tensor<fp16, [1024]> m_layers_11_v_bias_to_fp16 = const()[name = string("m_layers_11_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283428352)))]; |
| tensor<fp16, [1, 104, 1024]> linear_68_cast_fp16 = linear(bias = m_layers_11_v_bias_to_fp16, weight = m_layers_11_v_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_68_cast_fp16")]; |
| tensor<int32, [4]> var_892 = const()[name = string("op_892"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_893_cast_fp16 = reshape(shape = var_892, x = linear_68_cast_fp16)[name = string("op_893_cast_fp16")]; |
| tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_895_to_fp16 = const()[name = string("op_895_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_896_cast_fp16 = mul(x = var_881_cast_fp16, y = var_895_to_fp16)[name = string("op_896_cast_fp16")]; |
| bool var_898_transpose_x_0 = const()[name = string("op_898_transpose_x_0"), val = bool(false)]; |
| bool var_898_transpose_y_0 = const()[name = string("op_898_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_94_perm_0 = const()[name = string("transpose_94_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_95_perm_0 = const()[name = string("transpose_95_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_95 = transpose(perm = transpose_95_perm_0, x = var_887_cast_fp16)[name = string("transpose_169")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_94 = transpose(perm = transpose_94_perm_0, x = var_896_cast_fp16)[name = string("transpose_170")]; |
| tensor<fp16, [1, 16, 104, 104]> var_898_cast_fp16 = matmul(transpose_x = var_898_transpose_x_0, transpose_y = var_898_transpose_y_0, x = transpose_94, y = transpose_95)[name = string("op_898_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_899_cast_fp16 = add(x = var_898_cast_fp16, y = key_bias)[name = string("op_899_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_23_cast_fp16 = softmax(axis = var_854, x = var_899_cast_fp16)[name = string("att_23_cast_fp16")]; |
| bool var_901_transpose_x_0 = const()[name = string("op_901_transpose_x_0"), val = bool(false)]; |
| bool var_901_transpose_y_0 = const()[name = string("op_901_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_893_cast_fp16)[name = string("transpose_171")]; |
| tensor<fp16, [1, 16, 104, 64]> var_901_cast_fp16 = matmul(transpose_x = var_901_transpose_x_0, transpose_y = var_901_transpose_y_0, x = att_23_cast_fp16, y = v_23_cast_fp16)[name = string("op_901_cast_fp16")]; |
| tensor<int32, [4]> var_902_perm_0 = const()[name = string("op_902_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_903 = const()[name = string("op_903"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_902_cast_fp16 = transpose(perm = var_902_perm_0, x = var_901_cast_fp16)[name = string("transpose_168")]; |
| tensor<fp16, [1, 104, 1024]> input_113_cast_fp16 = reshape(shape = var_903, x = var_902_cast_fp16)[name = string("input_113_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_11_o_weight_to_fp16 = const()[name = string("m_layers_11_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(283430464)))]; |
| tensor<fp16, [1024]> m_layers_11_o_bias_to_fp16 = const()[name = string("m_layers_11_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285527680)))]; |
| tensor<fp16, [1, 104, 1024]> linear_69_cast_fp16 = linear(bias = m_layers_11_o_bias_to_fp16, weight = m_layers_11_o_weight_to_fp16, x = input_113_cast_fp16)[name = string("linear_69_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_115_cast_fp16 = add(x = x_23_cast_fp16, y = linear_69_cast_fp16)[name = string("input_115_cast_fp16")]; |
| tensor<int32, [1]> input_117_axes_0 = const()[name = string("input_117_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_11_ln2_weight_to_fp16 = const()[name = string("m_layers_11_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285529792)))]; |
| tensor<fp16, [1024]> m_layers_11_ln2_bias_to_fp16 = const()[name = string("m_layers_11_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285531904)))]; |
| tensor<fp16, [1, 104, 1024]> input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = m_layers_11_ln2_bias_to_fp16, epsilon = var_857_to_fp16, gamma = m_layers_11_ln2_weight_to_fp16, x = input_115_cast_fp16)[name = string("input_117_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_11_fc1_weight_to_fp16 = const()[name = string("m_layers_11_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285534016)))]; |
| tensor<fp16, [4096]> m_layers_11_fc1_bias_to_fp16 = const()[name = string("m_layers_11_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293922688)))]; |
| tensor<fp16, [1, 104, 4096]> linear_70_cast_fp16 = linear(bias = m_layers_11_fc1_bias_to_fp16, weight = m_layers_11_fc1_weight_to_fp16, x = input_117_cast_fp16)[name = string("linear_70_cast_fp16")]; |
| string input_119_mode_0 = const()[name = string("input_119_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_70_cast_fp16)[name = string("input_119_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_11_fc2_weight_to_fp16 = const()[name = string("m_layers_11_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293930944)))]; |
| tensor<fp16, [1024]> m_layers_11_fc2_bias_to_fp16 = const()[name = string("m_layers_11_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302319616)))]; |
| tensor<fp16, [1, 104, 1024]> linear_71_cast_fp16 = linear(bias = m_layers_11_fc2_bias_to_fp16, weight = m_layers_11_fc2_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_71_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_25_cast_fp16 = add(x = input_115_cast_fp16, y = linear_71_cast_fp16)[name = string("x_25_cast_fp16")]; |
| int32 var_924 = const()[name = string("op_924"), val = int32(-1)]; |
| tensor<int32, [1]> input_121_axes_0 = const()[name = string("input_121_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_12_ln1_weight_to_fp16 = const()[name = string("m_layers_12_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302321728)))]; |
| tensor<fp16, [1024]> m_layers_12_ln1_bias_to_fp16 = const()[name = string("m_layers_12_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302323840)))]; |
| fp16 var_927_to_fp16 = const()[name = string("op_927_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_121_cast_fp16 = layer_norm(axes = input_121_axes_0, beta = m_layers_12_ln1_bias_to_fp16, epsilon = var_927_to_fp16, gamma = m_layers_12_ln1_weight_to_fp16, x = x_25_cast_fp16)[name = string("input_121_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_12_q_weight_to_fp16 = const()[name = string("m_layers_12_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302325952)))]; |
| tensor<fp16, [1024]> m_layers_12_q_bias_to_fp16 = const()[name = string("m_layers_12_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304423168)))]; |
| tensor<fp16, [1, 104, 1024]> linear_72_cast_fp16 = linear(bias = m_layers_12_q_bias_to_fp16, weight = m_layers_12_q_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_72_cast_fp16")]; |
| tensor<int32, [4]> var_950 = const()[name = string("op_950"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_951_cast_fp16 = reshape(shape = var_950, x = linear_72_cast_fp16)[name = string("op_951_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_12_k_weight_to_fp16 = const()[name = string("m_layers_12_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(304425280)))]; |
| tensor<fp16, [1024]> m_layers_12_k_bias_to_fp16 = const()[name = string("m_layers_12_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306522496)))]; |
| tensor<fp16, [1, 104, 1024]> linear_73_cast_fp16 = linear(bias = m_layers_12_k_bias_to_fp16, weight = m_layers_12_k_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_73_cast_fp16")]; |
| tensor<int32, [4]> var_956 = const()[name = string("op_956"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_957_cast_fp16 = reshape(shape = var_956, x = linear_73_cast_fp16)[name = string("op_957_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_12_v_weight_to_fp16 = const()[name = string("m_layers_12_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(306524608)))]; |
| tensor<fp16, [1024]> m_layers_12_v_bias_to_fp16 = const()[name = string("m_layers_12_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308621824)))]; |
| tensor<fp16, [1, 104, 1024]> linear_74_cast_fp16 = linear(bias = m_layers_12_v_bias_to_fp16, weight = m_layers_12_v_weight_to_fp16, x = input_121_cast_fp16)[name = string("linear_74_cast_fp16")]; |
| tensor<int32, [4]> var_962 = const()[name = string("op_962"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_963_cast_fp16 = reshape(shape = var_962, x = linear_74_cast_fp16)[name = string("op_963_cast_fp16")]; |
| tensor<int32, [4]> v_25_perm_0 = const()[name = string("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_965_to_fp16 = const()[name = string("op_965_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_966_cast_fp16 = mul(x = var_951_cast_fp16, y = var_965_to_fp16)[name = string("op_966_cast_fp16")]; |
| bool var_968_transpose_x_0 = const()[name = string("op_968_transpose_x_0"), val = bool(false)]; |
| bool var_968_transpose_y_0 = const()[name = string("op_968_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_96_perm_0 = const()[name = string("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_97_perm_0 = const()[name = string("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_97 = transpose(perm = transpose_97_perm_0, x = var_957_cast_fp16)[name = string("transpose_165")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_966_cast_fp16)[name = string("transpose_166")]; |
| tensor<fp16, [1, 16, 104, 104]> var_968_cast_fp16 = matmul(transpose_x = var_968_transpose_x_0, transpose_y = var_968_transpose_y_0, x = transpose_96, y = transpose_97)[name = string("op_968_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_969_cast_fp16 = add(x = var_968_cast_fp16, y = key_bias)[name = string("op_969_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_25_cast_fp16 = softmax(axis = var_924, x = var_969_cast_fp16)[name = string("att_25_cast_fp16")]; |
| bool var_971_transpose_x_0 = const()[name = string("op_971_transpose_x_0"), val = bool(false)]; |
| bool var_971_transpose_y_0 = const()[name = string("op_971_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_963_cast_fp16)[name = string("transpose_167")]; |
| tensor<fp16, [1, 16, 104, 64]> var_971_cast_fp16 = matmul(transpose_x = var_971_transpose_x_0, transpose_y = var_971_transpose_y_0, x = att_25_cast_fp16, y = v_25_cast_fp16)[name = string("op_971_cast_fp16")]; |
| tensor<int32, [4]> var_972_perm_0 = const()[name = string("op_972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_973 = const()[name = string("op_973"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_972_cast_fp16 = transpose(perm = var_972_perm_0, x = var_971_cast_fp16)[name = string("transpose_164")]; |
| tensor<fp16, [1, 104, 1024]> input_123_cast_fp16 = reshape(shape = var_973, x = var_972_cast_fp16)[name = string("input_123_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_12_o_weight_to_fp16 = const()[name = string("m_layers_12_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(308623936)))]; |
| tensor<fp16, [1024]> m_layers_12_o_bias_to_fp16 = const()[name = string("m_layers_12_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310721152)))]; |
| tensor<fp16, [1, 104, 1024]> linear_75_cast_fp16 = linear(bias = m_layers_12_o_bias_to_fp16, weight = m_layers_12_o_weight_to_fp16, x = input_123_cast_fp16)[name = string("linear_75_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_125_cast_fp16 = add(x = x_25_cast_fp16, y = linear_75_cast_fp16)[name = string("input_125_cast_fp16")]; |
| tensor<int32, [1]> input_127_axes_0 = const()[name = string("input_127_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_12_ln2_weight_to_fp16 = const()[name = string("m_layers_12_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310723264)))]; |
| tensor<fp16, [1024]> m_layers_12_ln2_bias_to_fp16 = const()[name = string("m_layers_12_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310725376)))]; |
| tensor<fp16, [1, 104, 1024]> input_127_cast_fp16 = layer_norm(axes = input_127_axes_0, beta = m_layers_12_ln2_bias_to_fp16, epsilon = var_927_to_fp16, gamma = m_layers_12_ln2_weight_to_fp16, x = input_125_cast_fp16)[name = string("input_127_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_12_fc1_weight_to_fp16 = const()[name = string("m_layers_12_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310727488)))]; |
| tensor<fp16, [4096]> m_layers_12_fc1_bias_to_fp16 = const()[name = string("m_layers_12_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319116160)))]; |
| tensor<fp16, [1, 104, 4096]> linear_76_cast_fp16 = linear(bias = m_layers_12_fc1_bias_to_fp16, weight = m_layers_12_fc1_weight_to_fp16, x = input_127_cast_fp16)[name = string("linear_76_cast_fp16")]; |
| string input_129_mode_0 = const()[name = string("input_129_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_129_cast_fp16 = gelu(mode = input_129_mode_0, x = linear_76_cast_fp16)[name = string("input_129_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_12_fc2_weight_to_fp16 = const()[name = string("m_layers_12_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319124416)))]; |
| tensor<fp16, [1024]> m_layers_12_fc2_bias_to_fp16 = const()[name = string("m_layers_12_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327513088)))]; |
| tensor<fp16, [1, 104, 1024]> linear_77_cast_fp16 = linear(bias = m_layers_12_fc2_bias_to_fp16, weight = m_layers_12_fc2_weight_to_fp16, x = input_129_cast_fp16)[name = string("linear_77_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_27_cast_fp16 = add(x = input_125_cast_fp16, y = linear_77_cast_fp16)[name = string("x_27_cast_fp16")]; |
| int32 var_994 = const()[name = string("op_994"), val = int32(-1)]; |
| tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_13_ln1_weight_to_fp16 = const()[name = string("m_layers_13_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327515200)))]; |
| tensor<fp16, [1024]> m_layers_13_ln1_bias_to_fp16 = const()[name = string("m_layers_13_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327517312)))]; |
| fp16 var_997_to_fp16 = const()[name = string("op_997_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = m_layers_13_ln1_bias_to_fp16, epsilon = var_997_to_fp16, gamma = m_layers_13_ln1_weight_to_fp16, x = x_27_cast_fp16)[name = string("input_131_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_13_q_weight_to_fp16 = const()[name = string("m_layers_13_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327519424)))]; |
| tensor<fp16, [1024]> m_layers_13_q_bias_to_fp16 = const()[name = string("m_layers_13_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329616640)))]; |
| tensor<fp16, [1, 104, 1024]> linear_78_cast_fp16 = linear(bias = m_layers_13_q_bias_to_fp16, weight = m_layers_13_q_weight_to_fp16, x = input_131_cast_fp16)[name = string("linear_78_cast_fp16")]; |
| tensor<int32, [4]> var_1020 = const()[name = string("op_1020"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1021_cast_fp16 = reshape(shape = var_1020, x = linear_78_cast_fp16)[name = string("op_1021_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_13_k_weight_to_fp16 = const()[name = string("m_layers_13_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329618752)))]; |
| tensor<fp16, [1024]> m_layers_13_k_bias_to_fp16 = const()[name = string("m_layers_13_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331715968)))]; |
| tensor<fp16, [1, 104, 1024]> linear_79_cast_fp16 = linear(bias = m_layers_13_k_bias_to_fp16, weight = m_layers_13_k_weight_to_fp16, x = input_131_cast_fp16)[name = string("linear_79_cast_fp16")]; |
| tensor<int32, [4]> var_1026 = const()[name = string("op_1026"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1027_cast_fp16 = reshape(shape = var_1026, x = linear_79_cast_fp16)[name = string("op_1027_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_13_v_weight_to_fp16 = const()[name = string("m_layers_13_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331718080)))]; |
| tensor<fp16, [1024]> m_layers_13_v_bias_to_fp16 = const()[name = string("m_layers_13_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333815296)))]; |
| tensor<fp16, [1, 104, 1024]> linear_80_cast_fp16 = linear(bias = m_layers_13_v_bias_to_fp16, weight = m_layers_13_v_weight_to_fp16, x = input_131_cast_fp16)[name = string("linear_80_cast_fp16")]; |
| tensor<int32, [4]> var_1032 = const()[name = string("op_1032"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1033_cast_fp16 = reshape(shape = var_1032, x = linear_80_cast_fp16)[name = string("op_1033_cast_fp16")]; |
| tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1035_to_fp16 = const()[name = string("op_1035_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1036_cast_fp16 = mul(x = var_1021_cast_fp16, y = var_1035_to_fp16)[name = string("op_1036_cast_fp16")]; |
| bool var_1038_transpose_x_0 = const()[name = string("op_1038_transpose_x_0"), val = bool(false)]; |
| bool var_1038_transpose_y_0 = const()[name = string("op_1038_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_98_perm_0 = const()[name = string("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_99_perm_0 = const()[name = string("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_99 = transpose(perm = transpose_99_perm_0, x = var_1027_cast_fp16)[name = string("transpose_161")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_1036_cast_fp16)[name = string("transpose_162")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1038_cast_fp16 = matmul(transpose_x = var_1038_transpose_x_0, transpose_y = var_1038_transpose_y_0, x = transpose_98, y = transpose_99)[name = string("op_1038_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1039_cast_fp16 = add(x = var_1038_cast_fp16, y = key_bias)[name = string("op_1039_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_27_cast_fp16 = softmax(axis = var_994, x = var_1039_cast_fp16)[name = string("att_27_cast_fp16")]; |
| bool var_1041_transpose_x_0 = const()[name = string("op_1041_transpose_x_0"), val = bool(false)]; |
| bool var_1041_transpose_y_0 = const()[name = string("op_1041_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_1033_cast_fp16)[name = string("transpose_163")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1041_cast_fp16 = matmul(transpose_x = var_1041_transpose_x_0, transpose_y = var_1041_transpose_y_0, x = att_27_cast_fp16, y = v_27_cast_fp16)[name = string("op_1041_cast_fp16")]; |
| tensor<int32, [4]> var_1042_perm_0 = const()[name = string("op_1042_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1043 = const()[name = string("op_1043"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1042_cast_fp16 = transpose(perm = var_1042_perm_0, x = var_1041_cast_fp16)[name = string("transpose_160")]; |
| tensor<fp16, [1, 104, 1024]> input_133_cast_fp16 = reshape(shape = var_1043, x = var_1042_cast_fp16)[name = string("input_133_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_13_o_weight_to_fp16 = const()[name = string("m_layers_13_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333817408)))]; |
| tensor<fp16, [1024]> m_layers_13_o_bias_to_fp16 = const()[name = string("m_layers_13_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335914624)))]; |
| tensor<fp16, [1, 104, 1024]> linear_81_cast_fp16 = linear(bias = m_layers_13_o_bias_to_fp16, weight = m_layers_13_o_weight_to_fp16, x = input_133_cast_fp16)[name = string("linear_81_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_135_cast_fp16 = add(x = x_27_cast_fp16, y = linear_81_cast_fp16)[name = string("input_135_cast_fp16")]; |
| tensor<int32, [1]> input_137_axes_0 = const()[name = string("input_137_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_13_ln2_weight_to_fp16 = const()[name = string("m_layers_13_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335916736)))]; |
| tensor<fp16, [1024]> m_layers_13_ln2_bias_to_fp16 = const()[name = string("m_layers_13_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335918848)))]; |
| tensor<fp16, [1, 104, 1024]> input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = m_layers_13_ln2_bias_to_fp16, epsilon = var_997_to_fp16, gamma = m_layers_13_ln2_weight_to_fp16, x = input_135_cast_fp16)[name = string("input_137_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_13_fc1_weight_to_fp16 = const()[name = string("m_layers_13_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335920960)))]; |
| tensor<fp16, [4096]> m_layers_13_fc1_bias_to_fp16 = const()[name = string("m_layers_13_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344309632)))]; |
| tensor<fp16, [1, 104, 4096]> linear_82_cast_fp16 = linear(bias = m_layers_13_fc1_bias_to_fp16, weight = m_layers_13_fc1_weight_to_fp16, x = input_137_cast_fp16)[name = string("linear_82_cast_fp16")]; |
| string input_139_mode_0 = const()[name = string("input_139_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_139_cast_fp16 = gelu(mode = input_139_mode_0, x = linear_82_cast_fp16)[name = string("input_139_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_13_fc2_weight_to_fp16 = const()[name = string("m_layers_13_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344317888)))]; |
| tensor<fp16, [1024]> m_layers_13_fc2_bias_to_fp16 = const()[name = string("m_layers_13_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352706560)))]; |
| tensor<fp16, [1, 104, 1024]> linear_83_cast_fp16 = linear(bias = m_layers_13_fc2_bias_to_fp16, weight = m_layers_13_fc2_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_83_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_29_cast_fp16 = add(x = input_135_cast_fp16, y = linear_83_cast_fp16)[name = string("x_29_cast_fp16")]; |
| int32 var_1064 = const()[name = string("op_1064"), val = int32(-1)]; |
| tensor<int32, [1]> input_141_axes_0 = const()[name = string("input_141_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_14_ln1_weight_to_fp16 = const()[name = string("m_layers_14_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352708672)))]; |
| tensor<fp16, [1024]> m_layers_14_ln1_bias_to_fp16 = const()[name = string("m_layers_14_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352710784)))]; |
| fp16 var_1067_to_fp16 = const()[name = string("op_1067_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = m_layers_14_ln1_bias_to_fp16, epsilon = var_1067_to_fp16, gamma = m_layers_14_ln1_weight_to_fp16, x = x_29_cast_fp16)[name = string("input_141_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_14_q_weight_to_fp16 = const()[name = string("m_layers_14_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352712896)))]; |
| tensor<fp16, [1024]> m_layers_14_q_bias_to_fp16 = const()[name = string("m_layers_14_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354810112)))]; |
| tensor<fp16, [1, 104, 1024]> linear_84_cast_fp16 = linear(bias = m_layers_14_q_bias_to_fp16, weight = m_layers_14_q_weight_to_fp16, x = input_141_cast_fp16)[name = string("linear_84_cast_fp16")]; |
| tensor<int32, [4]> var_1090 = const()[name = string("op_1090"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1091_cast_fp16 = reshape(shape = var_1090, x = linear_84_cast_fp16)[name = string("op_1091_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_14_k_weight_to_fp16 = const()[name = string("m_layers_14_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(354812224)))]; |
| tensor<fp16, [1024]> m_layers_14_k_bias_to_fp16 = const()[name = string("m_layers_14_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356909440)))]; |
| tensor<fp16, [1, 104, 1024]> linear_85_cast_fp16 = linear(bias = m_layers_14_k_bias_to_fp16, weight = m_layers_14_k_weight_to_fp16, x = input_141_cast_fp16)[name = string("linear_85_cast_fp16")]; |
| tensor<int32, [4]> var_1096 = const()[name = string("op_1096"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1097_cast_fp16 = reshape(shape = var_1096, x = linear_85_cast_fp16)[name = string("op_1097_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_14_v_weight_to_fp16 = const()[name = string("m_layers_14_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(356911552)))]; |
| tensor<fp16, [1024]> m_layers_14_v_bias_to_fp16 = const()[name = string("m_layers_14_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359008768)))]; |
| tensor<fp16, [1, 104, 1024]> linear_86_cast_fp16 = linear(bias = m_layers_14_v_bias_to_fp16, weight = m_layers_14_v_weight_to_fp16, x = input_141_cast_fp16)[name = string("linear_86_cast_fp16")]; |
| tensor<int32, [4]> var_1102 = const()[name = string("op_1102"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1103_cast_fp16 = reshape(shape = var_1102, x = linear_86_cast_fp16)[name = string("op_1103_cast_fp16")]; |
| tensor<int32, [4]> v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1105_to_fp16 = const()[name = string("op_1105_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1106_cast_fp16 = mul(x = var_1091_cast_fp16, y = var_1105_to_fp16)[name = string("op_1106_cast_fp16")]; |
| bool var_1108_transpose_x_0 = const()[name = string("op_1108_transpose_x_0"), val = bool(false)]; |
| bool var_1108_transpose_y_0 = const()[name = string("op_1108_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_100_perm_0 = const()[name = string("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_101_perm_0 = const()[name = string("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_101 = transpose(perm = transpose_101_perm_0, x = var_1097_cast_fp16)[name = string("transpose_157")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_1106_cast_fp16)[name = string("transpose_158")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1108_cast_fp16 = matmul(transpose_x = var_1108_transpose_x_0, transpose_y = var_1108_transpose_y_0, x = transpose_100, y = transpose_101)[name = string("op_1108_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1109_cast_fp16 = add(x = var_1108_cast_fp16, y = key_bias)[name = string("op_1109_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_29_cast_fp16 = softmax(axis = var_1064, x = var_1109_cast_fp16)[name = string("att_29_cast_fp16")]; |
| bool var_1111_transpose_x_0 = const()[name = string("op_1111_transpose_x_0"), val = bool(false)]; |
| bool var_1111_transpose_y_0 = const()[name = string("op_1111_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_1103_cast_fp16)[name = string("transpose_159")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1111_cast_fp16 = matmul(transpose_x = var_1111_transpose_x_0, transpose_y = var_1111_transpose_y_0, x = att_29_cast_fp16, y = v_29_cast_fp16)[name = string("op_1111_cast_fp16")]; |
| tensor<int32, [4]> var_1112_perm_0 = const()[name = string("op_1112_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1113 = const()[name = string("op_1113"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1112_cast_fp16 = transpose(perm = var_1112_perm_0, x = var_1111_cast_fp16)[name = string("transpose_156")]; |
| tensor<fp16, [1, 104, 1024]> input_143_cast_fp16 = reshape(shape = var_1113, x = var_1112_cast_fp16)[name = string("input_143_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_14_o_weight_to_fp16 = const()[name = string("m_layers_14_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(359010880)))]; |
| tensor<fp16, [1024]> m_layers_14_o_bias_to_fp16 = const()[name = string("m_layers_14_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361108096)))]; |
| tensor<fp16, [1, 104, 1024]> linear_87_cast_fp16 = linear(bias = m_layers_14_o_bias_to_fp16, weight = m_layers_14_o_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_87_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_145_cast_fp16 = add(x = x_29_cast_fp16, y = linear_87_cast_fp16)[name = string("input_145_cast_fp16")]; |
| tensor<int32, [1]> input_147_axes_0 = const()[name = string("input_147_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_14_ln2_weight_to_fp16 = const()[name = string("m_layers_14_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361110208)))]; |
| tensor<fp16, [1024]> m_layers_14_ln2_bias_to_fp16 = const()[name = string("m_layers_14_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361112320)))]; |
| tensor<fp16, [1, 104, 1024]> input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = m_layers_14_ln2_bias_to_fp16, epsilon = var_1067_to_fp16, gamma = m_layers_14_ln2_weight_to_fp16, x = input_145_cast_fp16)[name = string("input_147_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_14_fc1_weight_to_fp16 = const()[name = string("m_layers_14_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361114432)))]; |
| tensor<fp16, [4096]> m_layers_14_fc1_bias_to_fp16 = const()[name = string("m_layers_14_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369503104)))]; |
| tensor<fp16, [1, 104, 4096]> linear_88_cast_fp16 = linear(bias = m_layers_14_fc1_bias_to_fp16, weight = m_layers_14_fc1_weight_to_fp16, x = input_147_cast_fp16)[name = string("linear_88_cast_fp16")]; |
| string input_149_mode_0 = const()[name = string("input_149_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_149_cast_fp16 = gelu(mode = input_149_mode_0, x = linear_88_cast_fp16)[name = string("input_149_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_14_fc2_weight_to_fp16 = const()[name = string("m_layers_14_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369511360)))]; |
| tensor<fp16, [1024]> m_layers_14_fc2_bias_to_fp16 = const()[name = string("m_layers_14_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377900032)))]; |
| tensor<fp16, [1, 104, 1024]> linear_89_cast_fp16 = linear(bias = m_layers_14_fc2_bias_to_fp16, weight = m_layers_14_fc2_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_89_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_31_cast_fp16 = add(x = input_145_cast_fp16, y = linear_89_cast_fp16)[name = string("x_31_cast_fp16")]; |
| int32 var_1134 = const()[name = string("op_1134"), val = int32(-1)]; |
| tensor<int32, [1]> input_151_axes_0 = const()[name = string("input_151_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_15_ln1_weight_to_fp16 = const()[name = string("m_layers_15_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377902144)))]; |
| tensor<fp16, [1024]> m_layers_15_ln1_bias_to_fp16 = const()[name = string("m_layers_15_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377904256)))]; |
| fp16 var_1137_to_fp16 = const()[name = string("op_1137_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_151_cast_fp16 = layer_norm(axes = input_151_axes_0, beta = m_layers_15_ln1_bias_to_fp16, epsilon = var_1137_to_fp16, gamma = m_layers_15_ln1_weight_to_fp16, x = x_31_cast_fp16)[name = string("input_151_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_15_q_weight_to_fp16 = const()[name = string("m_layers_15_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377906368)))]; |
| tensor<fp16, [1024]> m_layers_15_q_bias_to_fp16 = const()[name = string("m_layers_15_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380003584)))]; |
| tensor<fp16, [1, 104, 1024]> linear_90_cast_fp16 = linear(bias = m_layers_15_q_bias_to_fp16, weight = m_layers_15_q_weight_to_fp16, x = input_151_cast_fp16)[name = string("linear_90_cast_fp16")]; |
| tensor<int32, [4]> var_1160 = const()[name = string("op_1160"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1161_cast_fp16 = reshape(shape = var_1160, x = linear_90_cast_fp16)[name = string("op_1161_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_15_k_weight_to_fp16 = const()[name = string("m_layers_15_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380005696)))]; |
| tensor<fp16, [1024]> m_layers_15_k_bias_to_fp16 = const()[name = string("m_layers_15_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382102912)))]; |
| tensor<fp16, [1, 104, 1024]> linear_91_cast_fp16 = linear(bias = m_layers_15_k_bias_to_fp16, weight = m_layers_15_k_weight_to_fp16, x = input_151_cast_fp16)[name = string("linear_91_cast_fp16")]; |
| tensor<int32, [4]> var_1166 = const()[name = string("op_1166"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1167_cast_fp16 = reshape(shape = var_1166, x = linear_91_cast_fp16)[name = string("op_1167_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_15_v_weight_to_fp16 = const()[name = string("m_layers_15_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382105024)))]; |
| tensor<fp16, [1024]> m_layers_15_v_bias_to_fp16 = const()[name = string("m_layers_15_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384202240)))]; |
| tensor<fp16, [1, 104, 1024]> linear_92_cast_fp16 = linear(bias = m_layers_15_v_bias_to_fp16, weight = m_layers_15_v_weight_to_fp16, x = input_151_cast_fp16)[name = string("linear_92_cast_fp16")]; |
| tensor<int32, [4]> var_1172 = const()[name = string("op_1172"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1173_cast_fp16 = reshape(shape = var_1172, x = linear_92_cast_fp16)[name = string("op_1173_cast_fp16")]; |
| tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1175_to_fp16 = const()[name = string("op_1175_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1176_cast_fp16 = mul(x = var_1161_cast_fp16, y = var_1175_to_fp16)[name = string("op_1176_cast_fp16")]; |
| bool var_1178_transpose_x_0 = const()[name = string("op_1178_transpose_x_0"), val = bool(false)]; |
| bool var_1178_transpose_y_0 = const()[name = string("op_1178_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_102_perm_0 = const()[name = string("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_103_perm_0 = const()[name = string("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_103 = transpose(perm = transpose_103_perm_0, x = var_1167_cast_fp16)[name = string("transpose_153")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1176_cast_fp16)[name = string("transpose_154")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1178_cast_fp16 = matmul(transpose_x = var_1178_transpose_x_0, transpose_y = var_1178_transpose_y_0, x = transpose_102, y = transpose_103)[name = string("op_1178_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1179_cast_fp16 = add(x = var_1178_cast_fp16, y = key_bias)[name = string("op_1179_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_31_cast_fp16 = softmax(axis = var_1134, x = var_1179_cast_fp16)[name = string("att_31_cast_fp16")]; |
| bool var_1181_transpose_x_0 = const()[name = string("op_1181_transpose_x_0"), val = bool(false)]; |
| bool var_1181_transpose_y_0 = const()[name = string("op_1181_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_1173_cast_fp16)[name = string("transpose_155")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1181_cast_fp16 = matmul(transpose_x = var_1181_transpose_x_0, transpose_y = var_1181_transpose_y_0, x = att_31_cast_fp16, y = v_31_cast_fp16)[name = string("op_1181_cast_fp16")]; |
| tensor<int32, [4]> var_1182_perm_0 = const()[name = string("op_1182_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1183 = const()[name = string("op_1183"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1182_cast_fp16 = transpose(perm = var_1182_perm_0, x = var_1181_cast_fp16)[name = string("transpose_152")]; |
| tensor<fp16, [1, 104, 1024]> input_153_cast_fp16 = reshape(shape = var_1183, x = var_1182_cast_fp16)[name = string("input_153_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_15_o_weight_to_fp16 = const()[name = string("m_layers_15_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(384204352)))]; |
| tensor<fp16, [1024]> m_layers_15_o_bias_to_fp16 = const()[name = string("m_layers_15_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386301568)))]; |
| tensor<fp16, [1, 104, 1024]> linear_93_cast_fp16 = linear(bias = m_layers_15_o_bias_to_fp16, weight = m_layers_15_o_weight_to_fp16, x = input_153_cast_fp16)[name = string("linear_93_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_155_cast_fp16 = add(x = x_31_cast_fp16, y = linear_93_cast_fp16)[name = string("input_155_cast_fp16")]; |
| tensor<int32, [1]> input_157_axes_0 = const()[name = string("input_157_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_15_ln2_weight_to_fp16 = const()[name = string("m_layers_15_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386303680)))]; |
| tensor<fp16, [1024]> m_layers_15_ln2_bias_to_fp16 = const()[name = string("m_layers_15_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386305792)))]; |
| tensor<fp16, [1, 104, 1024]> input_157_cast_fp16 = layer_norm(axes = input_157_axes_0, beta = m_layers_15_ln2_bias_to_fp16, epsilon = var_1137_to_fp16, gamma = m_layers_15_ln2_weight_to_fp16, x = input_155_cast_fp16)[name = string("input_157_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_15_fc1_weight_to_fp16 = const()[name = string("m_layers_15_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386307904)))]; |
| tensor<fp16, [4096]> m_layers_15_fc1_bias_to_fp16 = const()[name = string("m_layers_15_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394696576)))]; |
| tensor<fp16, [1, 104, 4096]> linear_94_cast_fp16 = linear(bias = m_layers_15_fc1_bias_to_fp16, weight = m_layers_15_fc1_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_94_cast_fp16")]; |
| string input_159_mode_0 = const()[name = string("input_159_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_159_cast_fp16 = gelu(mode = input_159_mode_0, x = linear_94_cast_fp16)[name = string("input_159_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_15_fc2_weight_to_fp16 = const()[name = string("m_layers_15_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394704832)))]; |
| tensor<fp16, [1024]> m_layers_15_fc2_bias_to_fp16 = const()[name = string("m_layers_15_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403093504)))]; |
| tensor<fp16, [1, 104, 1024]> linear_95_cast_fp16 = linear(bias = m_layers_15_fc2_bias_to_fp16, weight = m_layers_15_fc2_weight_to_fp16, x = input_159_cast_fp16)[name = string("linear_95_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_33_cast_fp16 = add(x = input_155_cast_fp16, y = linear_95_cast_fp16)[name = string("x_33_cast_fp16")]; |
| int32 var_1204 = const()[name = string("op_1204"), val = int32(-1)]; |
| tensor<int32, [1]> input_161_axes_0 = const()[name = string("input_161_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_16_ln1_weight_to_fp16 = const()[name = string("m_layers_16_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403095616)))]; |
| tensor<fp16, [1024]> m_layers_16_ln1_bias_to_fp16 = const()[name = string("m_layers_16_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403097728)))]; |
| fp16 var_1207_to_fp16 = const()[name = string("op_1207_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = m_layers_16_ln1_bias_to_fp16, epsilon = var_1207_to_fp16, gamma = m_layers_16_ln1_weight_to_fp16, x = x_33_cast_fp16)[name = string("input_161_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_16_q_weight_to_fp16 = const()[name = string("m_layers_16_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403099840)))]; |
| tensor<fp16, [1024]> m_layers_16_q_bias_to_fp16 = const()[name = string("m_layers_16_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405197056)))]; |
| tensor<fp16, [1, 104, 1024]> linear_96_cast_fp16 = linear(bias = m_layers_16_q_bias_to_fp16, weight = m_layers_16_q_weight_to_fp16, x = input_161_cast_fp16)[name = string("linear_96_cast_fp16")]; |
| tensor<int32, [4]> var_1230 = const()[name = string("op_1230"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1231_cast_fp16 = reshape(shape = var_1230, x = linear_96_cast_fp16)[name = string("op_1231_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_16_k_weight_to_fp16 = const()[name = string("m_layers_16_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405199168)))]; |
| tensor<fp16, [1024]> m_layers_16_k_bias_to_fp16 = const()[name = string("m_layers_16_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407296384)))]; |
| tensor<fp16, [1, 104, 1024]> linear_97_cast_fp16 = linear(bias = m_layers_16_k_bias_to_fp16, weight = m_layers_16_k_weight_to_fp16, x = input_161_cast_fp16)[name = string("linear_97_cast_fp16")]; |
| tensor<int32, [4]> var_1236 = const()[name = string("op_1236"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1237_cast_fp16 = reshape(shape = var_1236, x = linear_97_cast_fp16)[name = string("op_1237_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_16_v_weight_to_fp16 = const()[name = string("m_layers_16_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407298496)))]; |
| tensor<fp16, [1024]> m_layers_16_v_bias_to_fp16 = const()[name = string("m_layers_16_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409395712)))]; |
| tensor<fp16, [1, 104, 1024]> linear_98_cast_fp16 = linear(bias = m_layers_16_v_bias_to_fp16, weight = m_layers_16_v_weight_to_fp16, x = input_161_cast_fp16)[name = string("linear_98_cast_fp16")]; |
| tensor<int32, [4]> var_1242 = const()[name = string("op_1242"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1243_cast_fp16 = reshape(shape = var_1242, x = linear_98_cast_fp16)[name = string("op_1243_cast_fp16")]; |
| tensor<int32, [4]> v_33_perm_0 = const()[name = string("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1245_to_fp16 = const()[name = string("op_1245_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1246_cast_fp16 = mul(x = var_1231_cast_fp16, y = var_1245_to_fp16)[name = string("op_1246_cast_fp16")]; |
| bool var_1248_transpose_x_0 = const()[name = string("op_1248_transpose_x_0"), val = bool(false)]; |
| bool var_1248_transpose_y_0 = const()[name = string("op_1248_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_104_perm_0 = const()[name = string("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_105_perm_0 = const()[name = string("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_105 = transpose(perm = transpose_105_perm_0, x = var_1237_cast_fp16)[name = string("transpose_149")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1246_cast_fp16)[name = string("transpose_150")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1248_cast_fp16 = matmul(transpose_x = var_1248_transpose_x_0, transpose_y = var_1248_transpose_y_0, x = transpose_104, y = transpose_105)[name = string("op_1248_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1249_cast_fp16 = add(x = var_1248_cast_fp16, y = key_bias)[name = string("op_1249_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_33_cast_fp16 = softmax(axis = var_1204, x = var_1249_cast_fp16)[name = string("att_33_cast_fp16")]; |
| bool var_1251_transpose_x_0 = const()[name = string("op_1251_transpose_x_0"), val = bool(false)]; |
| bool var_1251_transpose_y_0 = const()[name = string("op_1251_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_1243_cast_fp16)[name = string("transpose_151")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1251_cast_fp16 = matmul(transpose_x = var_1251_transpose_x_0, transpose_y = var_1251_transpose_y_0, x = att_33_cast_fp16, y = v_33_cast_fp16)[name = string("op_1251_cast_fp16")]; |
| tensor<int32, [4]> var_1252_perm_0 = const()[name = string("op_1252_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1253 = const()[name = string("op_1253"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1252_cast_fp16 = transpose(perm = var_1252_perm_0, x = var_1251_cast_fp16)[name = string("transpose_148")]; |
| tensor<fp16, [1, 104, 1024]> input_163_cast_fp16 = reshape(shape = var_1253, x = var_1252_cast_fp16)[name = string("input_163_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_16_o_weight_to_fp16 = const()[name = string("m_layers_16_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(409397824)))]; |
| tensor<fp16, [1024]> m_layers_16_o_bias_to_fp16 = const()[name = string("m_layers_16_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411495040)))]; |
| tensor<fp16, [1, 104, 1024]> linear_99_cast_fp16 = linear(bias = m_layers_16_o_bias_to_fp16, weight = m_layers_16_o_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_99_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_165_cast_fp16 = add(x = x_33_cast_fp16, y = linear_99_cast_fp16)[name = string("input_165_cast_fp16")]; |
| tensor<int32, [1]> input_167_axes_0 = const()[name = string("input_167_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_16_ln2_weight_to_fp16 = const()[name = string("m_layers_16_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411497152)))]; |
| tensor<fp16, [1024]> m_layers_16_ln2_bias_to_fp16 = const()[name = string("m_layers_16_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411499264)))]; |
| tensor<fp16, [1, 104, 1024]> input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = m_layers_16_ln2_bias_to_fp16, epsilon = var_1207_to_fp16, gamma = m_layers_16_ln2_weight_to_fp16, x = input_165_cast_fp16)[name = string("input_167_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_16_fc1_weight_to_fp16 = const()[name = string("m_layers_16_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411501376)))]; |
| tensor<fp16, [4096]> m_layers_16_fc1_bias_to_fp16 = const()[name = string("m_layers_16_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419890048)))]; |
| tensor<fp16, [1, 104, 4096]> linear_100_cast_fp16 = linear(bias = m_layers_16_fc1_bias_to_fp16, weight = m_layers_16_fc1_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_100_cast_fp16")]; |
| string input_169_mode_0 = const()[name = string("input_169_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_100_cast_fp16)[name = string("input_169_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_16_fc2_weight_to_fp16 = const()[name = string("m_layers_16_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419898304)))]; |
| tensor<fp16, [1024]> m_layers_16_fc2_bias_to_fp16 = const()[name = string("m_layers_16_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428286976)))]; |
| tensor<fp16, [1, 104, 1024]> linear_101_cast_fp16 = linear(bias = m_layers_16_fc2_bias_to_fp16, weight = m_layers_16_fc2_weight_to_fp16, x = input_169_cast_fp16)[name = string("linear_101_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_35_cast_fp16 = add(x = input_165_cast_fp16, y = linear_101_cast_fp16)[name = string("x_35_cast_fp16")]; |
| int32 var_1274 = const()[name = string("op_1274"), val = int32(-1)]; |
| tensor<int32, [1]> input_171_axes_0 = const()[name = string("input_171_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_17_ln1_weight_to_fp16 = const()[name = string("m_layers_17_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428289088)))]; |
| tensor<fp16, [1024]> m_layers_17_ln1_bias_to_fp16 = const()[name = string("m_layers_17_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428291200)))]; |
| fp16 var_1277_to_fp16 = const()[name = string("op_1277_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = m_layers_17_ln1_bias_to_fp16, epsilon = var_1277_to_fp16, gamma = m_layers_17_ln1_weight_to_fp16, x = x_35_cast_fp16)[name = string("input_171_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_17_q_weight_to_fp16 = const()[name = string("m_layers_17_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(428293312)))]; |
| tensor<fp16, [1024]> m_layers_17_q_bias_to_fp16 = const()[name = string("m_layers_17_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430390528)))]; |
| tensor<fp16, [1, 104, 1024]> linear_102_cast_fp16 = linear(bias = m_layers_17_q_bias_to_fp16, weight = m_layers_17_q_weight_to_fp16, x = input_171_cast_fp16)[name = string("linear_102_cast_fp16")]; |
| tensor<int32, [4]> var_1300 = const()[name = string("op_1300"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1301_cast_fp16 = reshape(shape = var_1300, x = linear_102_cast_fp16)[name = string("op_1301_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_17_k_weight_to_fp16 = const()[name = string("m_layers_17_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(430392640)))]; |
| tensor<fp16, [1024]> m_layers_17_k_bias_to_fp16 = const()[name = string("m_layers_17_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432489856)))]; |
| tensor<fp16, [1, 104, 1024]> linear_103_cast_fp16 = linear(bias = m_layers_17_k_bias_to_fp16, weight = m_layers_17_k_weight_to_fp16, x = input_171_cast_fp16)[name = string("linear_103_cast_fp16")]; |
| tensor<int32, [4]> var_1306 = const()[name = string("op_1306"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1307_cast_fp16 = reshape(shape = var_1306, x = linear_103_cast_fp16)[name = string("op_1307_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_17_v_weight_to_fp16 = const()[name = string("m_layers_17_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432491968)))]; |
| tensor<fp16, [1024]> m_layers_17_v_bias_to_fp16 = const()[name = string("m_layers_17_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434589184)))]; |
| tensor<fp16, [1, 104, 1024]> linear_104_cast_fp16 = linear(bias = m_layers_17_v_bias_to_fp16, weight = m_layers_17_v_weight_to_fp16, x = input_171_cast_fp16)[name = string("linear_104_cast_fp16")]; |
| tensor<int32, [4]> var_1312 = const()[name = string("op_1312"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1313_cast_fp16 = reshape(shape = var_1312, x = linear_104_cast_fp16)[name = string("op_1313_cast_fp16")]; |
| tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1315_to_fp16 = const()[name = string("op_1315_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1316_cast_fp16 = mul(x = var_1301_cast_fp16, y = var_1315_to_fp16)[name = string("op_1316_cast_fp16")]; |
| bool var_1318_transpose_x_0 = const()[name = string("op_1318_transpose_x_0"), val = bool(false)]; |
| bool var_1318_transpose_y_0 = const()[name = string("op_1318_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_106_perm_0 = const()[name = string("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_107_perm_0 = const()[name = string("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_107 = transpose(perm = transpose_107_perm_0, x = var_1307_cast_fp16)[name = string("transpose_145")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1316_cast_fp16)[name = string("transpose_146")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1318_cast_fp16 = matmul(transpose_x = var_1318_transpose_x_0, transpose_y = var_1318_transpose_y_0, x = transpose_106, y = transpose_107)[name = string("op_1318_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1319_cast_fp16 = add(x = var_1318_cast_fp16, y = key_bias)[name = string("op_1319_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_35_cast_fp16 = softmax(axis = var_1274, x = var_1319_cast_fp16)[name = string("att_35_cast_fp16")]; |
| bool var_1321_transpose_x_0 = const()[name = string("op_1321_transpose_x_0"), val = bool(false)]; |
| bool var_1321_transpose_y_0 = const()[name = string("op_1321_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_1313_cast_fp16)[name = string("transpose_147")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1321_cast_fp16 = matmul(transpose_x = var_1321_transpose_x_0, transpose_y = var_1321_transpose_y_0, x = att_35_cast_fp16, y = v_35_cast_fp16)[name = string("op_1321_cast_fp16")]; |
| tensor<int32, [4]> var_1322_perm_0 = const()[name = string("op_1322_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1323 = const()[name = string("op_1323"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1322_cast_fp16 = transpose(perm = var_1322_perm_0, x = var_1321_cast_fp16)[name = string("transpose_144")]; |
| tensor<fp16, [1, 104, 1024]> input_173_cast_fp16 = reshape(shape = var_1323, x = var_1322_cast_fp16)[name = string("input_173_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_17_o_weight_to_fp16 = const()[name = string("m_layers_17_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434591296)))]; |
| tensor<fp16, [1024]> m_layers_17_o_bias_to_fp16 = const()[name = string("m_layers_17_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436688512)))]; |
| tensor<fp16, [1, 104, 1024]> linear_105_cast_fp16 = linear(bias = m_layers_17_o_bias_to_fp16, weight = m_layers_17_o_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_105_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_175_cast_fp16 = add(x = x_35_cast_fp16, y = linear_105_cast_fp16)[name = string("input_175_cast_fp16")]; |
| tensor<int32, [1]> input_177_axes_0 = const()[name = string("input_177_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_17_ln2_weight_to_fp16 = const()[name = string("m_layers_17_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436690624)))]; |
| tensor<fp16, [1024]> m_layers_17_ln2_bias_to_fp16 = const()[name = string("m_layers_17_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436692736)))]; |
| tensor<fp16, [1, 104, 1024]> input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = m_layers_17_ln2_bias_to_fp16, epsilon = var_1277_to_fp16, gamma = m_layers_17_ln2_weight_to_fp16, x = input_175_cast_fp16)[name = string("input_177_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_17_fc1_weight_to_fp16 = const()[name = string("m_layers_17_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436694848)))]; |
| tensor<fp16, [4096]> m_layers_17_fc1_bias_to_fp16 = const()[name = string("m_layers_17_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445083520)))]; |
| tensor<fp16, [1, 104, 4096]> linear_106_cast_fp16 = linear(bias = m_layers_17_fc1_bias_to_fp16, weight = m_layers_17_fc1_weight_to_fp16, x = input_177_cast_fp16)[name = string("linear_106_cast_fp16")]; |
| string input_179_mode_0 = const()[name = string("input_179_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_106_cast_fp16)[name = string("input_179_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_17_fc2_weight_to_fp16 = const()[name = string("m_layers_17_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445091776)))]; |
| tensor<fp16, [1024]> m_layers_17_fc2_bias_to_fp16 = const()[name = string("m_layers_17_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453480448)))]; |
| tensor<fp16, [1, 104, 1024]> linear_107_cast_fp16 = linear(bias = m_layers_17_fc2_bias_to_fp16, weight = m_layers_17_fc2_weight_to_fp16, x = input_179_cast_fp16)[name = string("linear_107_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_37_cast_fp16 = add(x = input_175_cast_fp16, y = linear_107_cast_fp16)[name = string("x_37_cast_fp16")]; |
| int32 var_1344 = const()[name = string("op_1344"), val = int32(-1)]; |
| tensor<int32, [1]> input_181_axes_0 = const()[name = string("input_181_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_18_ln1_weight_to_fp16 = const()[name = string("m_layers_18_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453482560)))]; |
| tensor<fp16, [1024]> m_layers_18_ln1_bias_to_fp16 = const()[name = string("m_layers_18_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453484672)))]; |
| fp16 var_1347_to_fp16 = const()[name = string("op_1347_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_181_cast_fp16 = layer_norm(axes = input_181_axes_0, beta = m_layers_18_ln1_bias_to_fp16, epsilon = var_1347_to_fp16, gamma = m_layers_18_ln1_weight_to_fp16, x = x_37_cast_fp16)[name = string("input_181_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_18_q_weight_to_fp16 = const()[name = string("m_layers_18_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453486784)))]; |
| tensor<fp16, [1024]> m_layers_18_q_bias_to_fp16 = const()[name = string("m_layers_18_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455584000)))]; |
| tensor<fp16, [1, 104, 1024]> linear_108_cast_fp16 = linear(bias = m_layers_18_q_bias_to_fp16, weight = m_layers_18_q_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_108_cast_fp16")]; |
| tensor<int32, [4]> var_1370 = const()[name = string("op_1370"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1371_cast_fp16 = reshape(shape = var_1370, x = linear_108_cast_fp16)[name = string("op_1371_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_18_k_weight_to_fp16 = const()[name = string("m_layers_18_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455586112)))]; |
| tensor<fp16, [1024]> m_layers_18_k_bias_to_fp16 = const()[name = string("m_layers_18_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457683328)))]; |
| tensor<fp16, [1, 104, 1024]> linear_109_cast_fp16 = linear(bias = m_layers_18_k_bias_to_fp16, weight = m_layers_18_k_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_109_cast_fp16")]; |
| tensor<int32, [4]> var_1376 = const()[name = string("op_1376"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1377_cast_fp16 = reshape(shape = var_1376, x = linear_109_cast_fp16)[name = string("op_1377_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_18_v_weight_to_fp16 = const()[name = string("m_layers_18_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457685440)))]; |
| tensor<fp16, [1024]> m_layers_18_v_bias_to_fp16 = const()[name = string("m_layers_18_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459782656)))]; |
| tensor<fp16, [1, 104, 1024]> linear_110_cast_fp16 = linear(bias = m_layers_18_v_bias_to_fp16, weight = m_layers_18_v_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_110_cast_fp16")]; |
| tensor<int32, [4]> var_1382 = const()[name = string("op_1382"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1383_cast_fp16 = reshape(shape = var_1382, x = linear_110_cast_fp16)[name = string("op_1383_cast_fp16")]; |
| tensor<int32, [4]> v_37_perm_0 = const()[name = string("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1385_to_fp16 = const()[name = string("op_1385_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1386_cast_fp16 = mul(x = var_1371_cast_fp16, y = var_1385_to_fp16)[name = string("op_1386_cast_fp16")]; |
| bool var_1388_transpose_x_0 = const()[name = string("op_1388_transpose_x_0"), val = bool(false)]; |
| bool var_1388_transpose_y_0 = const()[name = string("op_1388_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_108_perm_0 = const()[name = string("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_109_perm_0 = const()[name = string("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_109 = transpose(perm = transpose_109_perm_0, x = var_1377_cast_fp16)[name = string("transpose_141")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1386_cast_fp16)[name = string("transpose_142")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1388_cast_fp16 = matmul(transpose_x = var_1388_transpose_x_0, transpose_y = var_1388_transpose_y_0, x = transpose_108, y = transpose_109)[name = string("op_1388_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1389_cast_fp16 = add(x = var_1388_cast_fp16, y = key_bias)[name = string("op_1389_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_37_cast_fp16 = softmax(axis = var_1344, x = var_1389_cast_fp16)[name = string("att_37_cast_fp16")]; |
| bool var_1391_transpose_x_0 = const()[name = string("op_1391_transpose_x_0"), val = bool(false)]; |
| bool var_1391_transpose_y_0 = const()[name = string("op_1391_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_1383_cast_fp16)[name = string("transpose_143")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1391_cast_fp16 = matmul(transpose_x = var_1391_transpose_x_0, transpose_y = var_1391_transpose_y_0, x = att_37_cast_fp16, y = v_37_cast_fp16)[name = string("op_1391_cast_fp16")]; |
| tensor<int32, [4]> var_1392_perm_0 = const()[name = string("op_1392_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1393 = const()[name = string("op_1393"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1392_cast_fp16 = transpose(perm = var_1392_perm_0, x = var_1391_cast_fp16)[name = string("transpose_140")]; |
| tensor<fp16, [1, 104, 1024]> input_183_cast_fp16 = reshape(shape = var_1393, x = var_1392_cast_fp16)[name = string("input_183_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_18_o_weight_to_fp16 = const()[name = string("m_layers_18_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459784768)))]; |
| tensor<fp16, [1024]> m_layers_18_o_bias_to_fp16 = const()[name = string("m_layers_18_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461881984)))]; |
| tensor<fp16, [1, 104, 1024]> linear_111_cast_fp16 = linear(bias = m_layers_18_o_bias_to_fp16, weight = m_layers_18_o_weight_to_fp16, x = input_183_cast_fp16)[name = string("linear_111_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_185_cast_fp16 = add(x = x_37_cast_fp16, y = linear_111_cast_fp16)[name = string("input_185_cast_fp16")]; |
| tensor<int32, [1]> input_187_axes_0 = const()[name = string("input_187_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_18_ln2_weight_to_fp16 = const()[name = string("m_layers_18_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461884096)))]; |
| tensor<fp16, [1024]> m_layers_18_ln2_bias_to_fp16 = const()[name = string("m_layers_18_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461886208)))]; |
| tensor<fp16, [1, 104, 1024]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = m_layers_18_ln2_bias_to_fp16, epsilon = var_1347_to_fp16, gamma = m_layers_18_ln2_weight_to_fp16, x = input_185_cast_fp16)[name = string("input_187_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_18_fc1_weight_to_fp16 = const()[name = string("m_layers_18_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461888320)))]; |
| tensor<fp16, [4096]> m_layers_18_fc1_bias_to_fp16 = const()[name = string("m_layers_18_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470276992)))]; |
| tensor<fp16, [1, 104, 4096]> linear_112_cast_fp16 = linear(bias = m_layers_18_fc1_bias_to_fp16, weight = m_layers_18_fc1_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_112_cast_fp16")]; |
| string input_189_mode_0 = const()[name = string("input_189_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_189_cast_fp16 = gelu(mode = input_189_mode_0, x = linear_112_cast_fp16)[name = string("input_189_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_18_fc2_weight_to_fp16 = const()[name = string("m_layers_18_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470285248)))]; |
| tensor<fp16, [1024]> m_layers_18_fc2_bias_to_fp16 = const()[name = string("m_layers_18_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478673920)))]; |
| tensor<fp16, [1, 104, 1024]> linear_113_cast_fp16 = linear(bias = m_layers_18_fc2_bias_to_fp16, weight = m_layers_18_fc2_weight_to_fp16, x = input_189_cast_fp16)[name = string("linear_113_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_39_cast_fp16 = add(x = input_185_cast_fp16, y = linear_113_cast_fp16)[name = string("x_39_cast_fp16")]; |
| int32 var_1414 = const()[name = string("op_1414"), val = int32(-1)]; |
| tensor<int32, [1]> input_191_axes_0 = const()[name = string("input_191_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_19_ln1_weight_to_fp16 = const()[name = string("m_layers_19_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478676032)))]; |
| tensor<fp16, [1024]> m_layers_19_ln1_bias_to_fp16 = const()[name = string("m_layers_19_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478678144)))]; |
| fp16 var_1417_to_fp16 = const()[name = string("op_1417_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = m_layers_19_ln1_bias_to_fp16, epsilon = var_1417_to_fp16, gamma = m_layers_19_ln1_weight_to_fp16, x = x_39_cast_fp16)[name = string("input_191_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_19_q_weight_to_fp16 = const()[name = string("m_layers_19_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478680256)))]; |
| tensor<fp16, [1024]> m_layers_19_q_bias_to_fp16 = const()[name = string("m_layers_19_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480777472)))]; |
| tensor<fp16, [1, 104, 1024]> linear_114_cast_fp16 = linear(bias = m_layers_19_q_bias_to_fp16, weight = m_layers_19_q_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_114_cast_fp16")]; |
| tensor<int32, [4]> var_1440 = const()[name = string("op_1440"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1441_cast_fp16 = reshape(shape = var_1440, x = linear_114_cast_fp16)[name = string("op_1441_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_19_k_weight_to_fp16 = const()[name = string("m_layers_19_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(480779584)))]; |
| tensor<fp16, [1024]> m_layers_19_k_bias_to_fp16 = const()[name = string("m_layers_19_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482876800)))]; |
| tensor<fp16, [1, 104, 1024]> linear_115_cast_fp16 = linear(bias = m_layers_19_k_bias_to_fp16, weight = m_layers_19_k_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_115_cast_fp16")]; |
| tensor<int32, [4]> var_1446 = const()[name = string("op_1446"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1447_cast_fp16 = reshape(shape = var_1446, x = linear_115_cast_fp16)[name = string("op_1447_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_19_v_weight_to_fp16 = const()[name = string("m_layers_19_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(482878912)))]; |
| tensor<fp16, [1024]> m_layers_19_v_bias_to_fp16 = const()[name = string("m_layers_19_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484976128)))]; |
| tensor<fp16, [1, 104, 1024]> linear_116_cast_fp16 = linear(bias = m_layers_19_v_bias_to_fp16, weight = m_layers_19_v_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_116_cast_fp16")]; |
| tensor<int32, [4]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1453_cast_fp16 = reshape(shape = var_1452, x = linear_116_cast_fp16)[name = string("op_1453_cast_fp16")]; |
| tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1455_to_fp16 = const()[name = string("op_1455_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1456_cast_fp16 = mul(x = var_1441_cast_fp16, y = var_1455_to_fp16)[name = string("op_1456_cast_fp16")]; |
| bool var_1458_transpose_x_0 = const()[name = string("op_1458_transpose_x_0"), val = bool(false)]; |
| bool var_1458_transpose_y_0 = const()[name = string("op_1458_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_110_perm_0 = const()[name = string("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_111_perm_0 = const()[name = string("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_111 = transpose(perm = transpose_111_perm_0, x = var_1447_cast_fp16)[name = string("transpose_137")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_1456_cast_fp16)[name = string("transpose_138")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1458_cast_fp16 = matmul(transpose_x = var_1458_transpose_x_0, transpose_y = var_1458_transpose_y_0, x = transpose_110, y = transpose_111)[name = string("op_1458_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1459_cast_fp16 = add(x = var_1458_cast_fp16, y = key_bias)[name = string("op_1459_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_39_cast_fp16 = softmax(axis = var_1414, x = var_1459_cast_fp16)[name = string("att_39_cast_fp16")]; |
| bool var_1461_transpose_x_0 = const()[name = string("op_1461_transpose_x_0"), val = bool(false)]; |
| bool var_1461_transpose_y_0 = const()[name = string("op_1461_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_1453_cast_fp16)[name = string("transpose_139")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1461_cast_fp16 = matmul(transpose_x = var_1461_transpose_x_0, transpose_y = var_1461_transpose_y_0, x = att_39_cast_fp16, y = v_39_cast_fp16)[name = string("op_1461_cast_fp16")]; |
| tensor<int32, [4]> var_1462_perm_0 = const()[name = string("op_1462_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1463 = const()[name = string("op_1463"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1462_cast_fp16 = transpose(perm = var_1462_perm_0, x = var_1461_cast_fp16)[name = string("transpose_136")]; |
| tensor<fp16, [1, 104, 1024]> input_193_cast_fp16 = reshape(shape = var_1463, x = var_1462_cast_fp16)[name = string("input_193_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_19_o_weight_to_fp16 = const()[name = string("m_layers_19_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484978240)))]; |
| tensor<fp16, [1024]> m_layers_19_o_bias_to_fp16 = const()[name = string("m_layers_19_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487075456)))]; |
| tensor<fp16, [1, 104, 1024]> linear_117_cast_fp16 = linear(bias = m_layers_19_o_bias_to_fp16, weight = m_layers_19_o_weight_to_fp16, x = input_193_cast_fp16)[name = string("linear_117_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_195_cast_fp16 = add(x = x_39_cast_fp16, y = linear_117_cast_fp16)[name = string("input_195_cast_fp16")]; |
| tensor<int32, [1]> input_197_axes_0 = const()[name = string("input_197_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_19_ln2_weight_to_fp16 = const()[name = string("m_layers_19_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487077568)))]; |
| tensor<fp16, [1024]> m_layers_19_ln2_bias_to_fp16 = const()[name = string("m_layers_19_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487079680)))]; |
| tensor<fp16, [1, 104, 1024]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = m_layers_19_ln2_bias_to_fp16, epsilon = var_1417_to_fp16, gamma = m_layers_19_ln2_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_19_fc1_weight_to_fp16 = const()[name = string("m_layers_19_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487081792)))]; |
| tensor<fp16, [4096]> m_layers_19_fc1_bias_to_fp16 = const()[name = string("m_layers_19_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495470464)))]; |
| tensor<fp16, [1, 104, 4096]> linear_118_cast_fp16 = linear(bias = m_layers_19_fc1_bias_to_fp16, weight = m_layers_19_fc1_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_118_cast_fp16")]; |
| string input_199_mode_0 = const()[name = string("input_199_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_199_cast_fp16 = gelu(mode = input_199_mode_0, x = linear_118_cast_fp16)[name = string("input_199_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_19_fc2_weight_to_fp16 = const()[name = string("m_layers_19_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495478720)))]; |
| tensor<fp16, [1024]> m_layers_19_fc2_bias_to_fp16 = const()[name = string("m_layers_19_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503867392)))]; |
| tensor<fp16, [1, 104, 1024]> linear_119_cast_fp16 = linear(bias = m_layers_19_fc2_bias_to_fp16, weight = m_layers_19_fc2_weight_to_fp16, x = input_199_cast_fp16)[name = string("linear_119_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_41_cast_fp16 = add(x = input_195_cast_fp16, y = linear_119_cast_fp16)[name = string("x_41_cast_fp16")]; |
| int32 var_1484 = const()[name = string("op_1484"), val = int32(-1)]; |
| tensor<int32, [1]> input_201_axes_0 = const()[name = string("input_201_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_20_ln1_weight_to_fp16 = const()[name = string("m_layers_20_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503869504)))]; |
| tensor<fp16, [1024]> m_layers_20_ln1_bias_to_fp16 = const()[name = string("m_layers_20_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503871616)))]; |
| fp16 var_1487_to_fp16 = const()[name = string("op_1487_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = m_layers_20_ln1_bias_to_fp16, epsilon = var_1487_to_fp16, gamma = m_layers_20_ln1_weight_to_fp16, x = x_41_cast_fp16)[name = string("input_201_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_20_q_weight_to_fp16 = const()[name = string("m_layers_20_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503873728)))]; |
| tensor<fp16, [1024]> m_layers_20_q_bias_to_fp16 = const()[name = string("m_layers_20_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505970944)))]; |
| tensor<fp16, [1, 104, 1024]> linear_120_cast_fp16 = linear(bias = m_layers_20_q_bias_to_fp16, weight = m_layers_20_q_weight_to_fp16, x = input_201_cast_fp16)[name = string("linear_120_cast_fp16")]; |
| tensor<int32, [4]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1511_cast_fp16 = reshape(shape = var_1510, x = linear_120_cast_fp16)[name = string("op_1511_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_20_k_weight_to_fp16 = const()[name = string("m_layers_20_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(505973056)))]; |
| tensor<fp16, [1024]> m_layers_20_k_bias_to_fp16 = const()[name = string("m_layers_20_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508070272)))]; |
| tensor<fp16, [1, 104, 1024]> linear_121_cast_fp16 = linear(bias = m_layers_20_k_bias_to_fp16, weight = m_layers_20_k_weight_to_fp16, x = input_201_cast_fp16)[name = string("linear_121_cast_fp16")]; |
| tensor<int32, [4]> var_1516 = const()[name = string("op_1516"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1517_cast_fp16 = reshape(shape = var_1516, x = linear_121_cast_fp16)[name = string("op_1517_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_20_v_weight_to_fp16 = const()[name = string("m_layers_20_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508072384)))]; |
| tensor<fp16, [1024]> m_layers_20_v_bias_to_fp16 = const()[name = string("m_layers_20_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510169600)))]; |
| tensor<fp16, [1, 104, 1024]> linear_122_cast_fp16 = linear(bias = m_layers_20_v_bias_to_fp16, weight = m_layers_20_v_weight_to_fp16, x = input_201_cast_fp16)[name = string("linear_122_cast_fp16")]; |
| tensor<int32, [4]> var_1522 = const()[name = string("op_1522"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1523_cast_fp16 = reshape(shape = var_1522, x = linear_122_cast_fp16)[name = string("op_1523_cast_fp16")]; |
| tensor<int32, [4]> v_41_perm_0 = const()[name = string("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1525_to_fp16 = const()[name = string("op_1525_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1526_cast_fp16 = mul(x = var_1511_cast_fp16, y = var_1525_to_fp16)[name = string("op_1526_cast_fp16")]; |
| bool var_1528_transpose_x_0 = const()[name = string("op_1528_transpose_x_0"), val = bool(false)]; |
| bool var_1528_transpose_y_0 = const()[name = string("op_1528_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_112_perm_0 = const()[name = string("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_113_perm_0 = const()[name = string("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_113 = transpose(perm = transpose_113_perm_0, x = var_1517_cast_fp16)[name = string("transpose_133")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_1526_cast_fp16)[name = string("transpose_134")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1528_cast_fp16 = matmul(transpose_x = var_1528_transpose_x_0, transpose_y = var_1528_transpose_y_0, x = transpose_112, y = transpose_113)[name = string("op_1528_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1529_cast_fp16 = add(x = var_1528_cast_fp16, y = key_bias)[name = string("op_1529_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_41_cast_fp16 = softmax(axis = var_1484, x = var_1529_cast_fp16)[name = string("att_41_cast_fp16")]; |
| bool var_1531_transpose_x_0 = const()[name = string("op_1531_transpose_x_0"), val = bool(false)]; |
| bool var_1531_transpose_y_0 = const()[name = string("op_1531_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_1523_cast_fp16)[name = string("transpose_135")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1531_cast_fp16 = matmul(transpose_x = var_1531_transpose_x_0, transpose_y = var_1531_transpose_y_0, x = att_41_cast_fp16, y = v_41_cast_fp16)[name = string("op_1531_cast_fp16")]; |
| tensor<int32, [4]> var_1532_perm_0 = const()[name = string("op_1532_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1533 = const()[name = string("op_1533"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1532_cast_fp16 = transpose(perm = var_1532_perm_0, x = var_1531_cast_fp16)[name = string("transpose_132")]; |
| tensor<fp16, [1, 104, 1024]> input_203_cast_fp16 = reshape(shape = var_1533, x = var_1532_cast_fp16)[name = string("input_203_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_20_o_weight_to_fp16 = const()[name = string("m_layers_20_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510171712)))]; |
| tensor<fp16, [1024]> m_layers_20_o_bias_to_fp16 = const()[name = string("m_layers_20_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512268928)))]; |
| tensor<fp16, [1, 104, 1024]> linear_123_cast_fp16 = linear(bias = m_layers_20_o_bias_to_fp16, weight = m_layers_20_o_weight_to_fp16, x = input_203_cast_fp16)[name = string("linear_123_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_205_cast_fp16 = add(x = x_41_cast_fp16, y = linear_123_cast_fp16)[name = string("input_205_cast_fp16")]; |
| tensor<int32, [1]> input_207_axes_0 = const()[name = string("input_207_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_20_ln2_weight_to_fp16 = const()[name = string("m_layers_20_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512271040)))]; |
| tensor<fp16, [1024]> m_layers_20_ln2_bias_to_fp16 = const()[name = string("m_layers_20_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512273152)))]; |
| tensor<fp16, [1, 104, 1024]> input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = m_layers_20_ln2_bias_to_fp16, epsilon = var_1487_to_fp16, gamma = m_layers_20_ln2_weight_to_fp16, x = input_205_cast_fp16)[name = string("input_207_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_20_fc1_weight_to_fp16 = const()[name = string("m_layers_20_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512275264)))]; |
| tensor<fp16, [4096]> m_layers_20_fc1_bias_to_fp16 = const()[name = string("m_layers_20_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520663936)))]; |
| tensor<fp16, [1, 104, 4096]> linear_124_cast_fp16 = linear(bias = m_layers_20_fc1_bias_to_fp16, weight = m_layers_20_fc1_weight_to_fp16, x = input_207_cast_fp16)[name = string("linear_124_cast_fp16")]; |
| string input_209_mode_0 = const()[name = string("input_209_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_209_cast_fp16 = gelu(mode = input_209_mode_0, x = linear_124_cast_fp16)[name = string("input_209_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_20_fc2_weight_to_fp16 = const()[name = string("m_layers_20_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520672192)))]; |
| tensor<fp16, [1024]> m_layers_20_fc2_bias_to_fp16 = const()[name = string("m_layers_20_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529060864)))]; |
| tensor<fp16, [1, 104, 1024]> linear_125_cast_fp16 = linear(bias = m_layers_20_fc2_bias_to_fp16, weight = m_layers_20_fc2_weight_to_fp16, x = input_209_cast_fp16)[name = string("linear_125_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_43_cast_fp16 = add(x = input_205_cast_fp16, y = linear_125_cast_fp16)[name = string("x_43_cast_fp16")]; |
| int32 var_1554 = const()[name = string("op_1554"), val = int32(-1)]; |
| tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_21_ln1_weight_to_fp16 = const()[name = string("m_layers_21_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529062976)))]; |
| tensor<fp16, [1024]> m_layers_21_ln1_bias_to_fp16 = const()[name = string("m_layers_21_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529065088)))]; |
| fp16 var_1557_to_fp16 = const()[name = string("op_1557_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = m_layers_21_ln1_bias_to_fp16, epsilon = var_1557_to_fp16, gamma = m_layers_21_ln1_weight_to_fp16, x = x_43_cast_fp16)[name = string("input_211_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_21_q_weight_to_fp16 = const()[name = string("m_layers_21_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(529067200)))]; |
| tensor<fp16, [1024]> m_layers_21_q_bias_to_fp16 = const()[name = string("m_layers_21_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531164416)))]; |
| tensor<fp16, [1, 104, 1024]> linear_126_cast_fp16 = linear(bias = m_layers_21_q_bias_to_fp16, weight = m_layers_21_q_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_126_cast_fp16")]; |
| tensor<int32, [4]> var_1580 = const()[name = string("op_1580"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1581_cast_fp16 = reshape(shape = var_1580, x = linear_126_cast_fp16)[name = string("op_1581_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_21_k_weight_to_fp16 = const()[name = string("m_layers_21_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531166528)))]; |
| tensor<fp16, [1024]> m_layers_21_k_bias_to_fp16 = const()[name = string("m_layers_21_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533263744)))]; |
| tensor<fp16, [1, 104, 1024]> linear_127_cast_fp16 = linear(bias = m_layers_21_k_bias_to_fp16, weight = m_layers_21_k_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_127_cast_fp16")]; |
| tensor<int32, [4]> var_1586 = const()[name = string("op_1586"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1587_cast_fp16 = reshape(shape = var_1586, x = linear_127_cast_fp16)[name = string("op_1587_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_21_v_weight_to_fp16 = const()[name = string("m_layers_21_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533265856)))]; |
| tensor<fp16, [1024]> m_layers_21_v_bias_to_fp16 = const()[name = string("m_layers_21_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535363072)))]; |
| tensor<fp16, [1, 104, 1024]> linear_128_cast_fp16 = linear(bias = m_layers_21_v_bias_to_fp16, weight = m_layers_21_v_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_128_cast_fp16")]; |
| tensor<int32, [4]> var_1592 = const()[name = string("op_1592"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1593_cast_fp16 = reshape(shape = var_1592, x = linear_128_cast_fp16)[name = string("op_1593_cast_fp16")]; |
| tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1595_to_fp16 = const()[name = string("op_1595_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1596_cast_fp16 = mul(x = var_1581_cast_fp16, y = var_1595_to_fp16)[name = string("op_1596_cast_fp16")]; |
| bool var_1598_transpose_x_0 = const()[name = string("op_1598_transpose_x_0"), val = bool(false)]; |
| bool var_1598_transpose_y_0 = const()[name = string("op_1598_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_114_perm_0 = const()[name = string("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_115_perm_0 = const()[name = string("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_115 = transpose(perm = transpose_115_perm_0, x = var_1587_cast_fp16)[name = string("transpose_129")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_1596_cast_fp16)[name = string("transpose_130")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1598_cast_fp16 = matmul(transpose_x = var_1598_transpose_x_0, transpose_y = var_1598_transpose_y_0, x = transpose_114, y = transpose_115)[name = string("op_1598_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1599_cast_fp16 = add(x = var_1598_cast_fp16, y = key_bias)[name = string("op_1599_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_43_cast_fp16 = softmax(axis = var_1554, x = var_1599_cast_fp16)[name = string("att_43_cast_fp16")]; |
| bool var_1601_transpose_x_0 = const()[name = string("op_1601_transpose_x_0"), val = bool(false)]; |
| bool var_1601_transpose_y_0 = const()[name = string("op_1601_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_1593_cast_fp16)[name = string("transpose_131")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1601_cast_fp16 = matmul(transpose_x = var_1601_transpose_x_0, transpose_y = var_1601_transpose_y_0, x = att_43_cast_fp16, y = v_43_cast_fp16)[name = string("op_1601_cast_fp16")]; |
| tensor<int32, [4]> var_1602_perm_0 = const()[name = string("op_1602_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1603 = const()[name = string("op_1603"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1602_cast_fp16 = transpose(perm = var_1602_perm_0, x = var_1601_cast_fp16)[name = string("transpose_128")]; |
| tensor<fp16, [1, 104, 1024]> input_213_cast_fp16 = reshape(shape = var_1603, x = var_1602_cast_fp16)[name = string("input_213_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_21_o_weight_to_fp16 = const()[name = string("m_layers_21_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535365184)))]; |
| tensor<fp16, [1024]> m_layers_21_o_bias_to_fp16 = const()[name = string("m_layers_21_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537462400)))]; |
| tensor<fp16, [1, 104, 1024]> linear_129_cast_fp16 = linear(bias = m_layers_21_o_bias_to_fp16, weight = m_layers_21_o_weight_to_fp16, x = input_213_cast_fp16)[name = string("linear_129_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_215_cast_fp16 = add(x = x_43_cast_fp16, y = linear_129_cast_fp16)[name = string("input_215_cast_fp16")]; |
| tensor<int32, [1]> input_217_axes_0 = const()[name = string("input_217_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_21_ln2_weight_to_fp16 = const()[name = string("m_layers_21_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537464512)))]; |
| tensor<fp16, [1024]> m_layers_21_ln2_bias_to_fp16 = const()[name = string("m_layers_21_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537466624)))]; |
| tensor<fp16, [1, 104, 1024]> input_217_cast_fp16 = layer_norm(axes = input_217_axes_0, beta = m_layers_21_ln2_bias_to_fp16, epsilon = var_1557_to_fp16, gamma = m_layers_21_ln2_weight_to_fp16, x = input_215_cast_fp16)[name = string("input_217_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_21_fc1_weight_to_fp16 = const()[name = string("m_layers_21_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(537468736)))]; |
| tensor<fp16, [4096]> m_layers_21_fc1_bias_to_fp16 = const()[name = string("m_layers_21_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545857408)))]; |
| tensor<fp16, [1, 104, 4096]> linear_130_cast_fp16 = linear(bias = m_layers_21_fc1_bias_to_fp16, weight = m_layers_21_fc1_weight_to_fp16, x = input_217_cast_fp16)[name = string("linear_130_cast_fp16")]; |
| string input_219_mode_0 = const()[name = string("input_219_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_219_cast_fp16 = gelu(mode = input_219_mode_0, x = linear_130_cast_fp16)[name = string("input_219_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_21_fc2_weight_to_fp16 = const()[name = string("m_layers_21_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545865664)))]; |
| tensor<fp16, [1024]> m_layers_21_fc2_bias_to_fp16 = const()[name = string("m_layers_21_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554254336)))]; |
| tensor<fp16, [1, 104, 1024]> linear_131_cast_fp16 = linear(bias = m_layers_21_fc2_bias_to_fp16, weight = m_layers_21_fc2_weight_to_fp16, x = input_219_cast_fp16)[name = string("linear_131_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_45_cast_fp16 = add(x = input_215_cast_fp16, y = linear_131_cast_fp16)[name = string("x_45_cast_fp16")]; |
| int32 var_1624 = const()[name = string("op_1624"), val = int32(-1)]; |
| tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_22_ln1_weight_to_fp16 = const()[name = string("m_layers_22_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554256448)))]; |
| tensor<fp16, [1024]> m_layers_22_ln1_bias_to_fp16 = const()[name = string("m_layers_22_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554258560)))]; |
| fp16 var_1627_to_fp16 = const()[name = string("op_1627_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = m_layers_22_ln1_bias_to_fp16, epsilon = var_1627_to_fp16, gamma = m_layers_22_ln1_weight_to_fp16, x = x_45_cast_fp16)[name = string("input_221_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_22_q_weight_to_fp16 = const()[name = string("m_layers_22_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(554260672)))]; |
| tensor<fp16, [1024]> m_layers_22_q_bias_to_fp16 = const()[name = string("m_layers_22_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556357888)))]; |
| tensor<fp16, [1, 104, 1024]> linear_132_cast_fp16 = linear(bias = m_layers_22_q_bias_to_fp16, weight = m_layers_22_q_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_132_cast_fp16")]; |
| tensor<int32, [4]> var_1650 = const()[name = string("op_1650"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1651_cast_fp16 = reshape(shape = var_1650, x = linear_132_cast_fp16)[name = string("op_1651_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_22_k_weight_to_fp16 = const()[name = string("m_layers_22_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(556360000)))]; |
| tensor<fp16, [1024]> m_layers_22_k_bias_to_fp16 = const()[name = string("m_layers_22_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558457216)))]; |
| tensor<fp16, [1, 104, 1024]> linear_133_cast_fp16 = linear(bias = m_layers_22_k_bias_to_fp16, weight = m_layers_22_k_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_133_cast_fp16")]; |
| tensor<int32, [4]> var_1656 = const()[name = string("op_1656"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1657_cast_fp16 = reshape(shape = var_1656, x = linear_133_cast_fp16)[name = string("op_1657_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_22_v_weight_to_fp16 = const()[name = string("m_layers_22_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(558459328)))]; |
| tensor<fp16, [1024]> m_layers_22_v_bias_to_fp16 = const()[name = string("m_layers_22_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560556544)))]; |
| tensor<fp16, [1, 104, 1024]> linear_134_cast_fp16 = linear(bias = m_layers_22_v_bias_to_fp16, weight = m_layers_22_v_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_134_cast_fp16")]; |
| tensor<int32, [4]> var_1662 = const()[name = string("op_1662"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1663_cast_fp16 = reshape(shape = var_1662, x = linear_134_cast_fp16)[name = string("op_1663_cast_fp16")]; |
| tensor<int32, [4]> v_45_perm_0 = const()[name = string("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1665_to_fp16 = const()[name = string("op_1665_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1666_cast_fp16 = mul(x = var_1651_cast_fp16, y = var_1665_to_fp16)[name = string("op_1666_cast_fp16")]; |
| bool var_1668_transpose_x_0 = const()[name = string("op_1668_transpose_x_0"), val = bool(false)]; |
| bool var_1668_transpose_y_0 = const()[name = string("op_1668_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_116_perm_0 = const()[name = string("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_117_perm_0 = const()[name = string("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_117 = transpose(perm = transpose_117_perm_0, x = var_1657_cast_fp16)[name = string("transpose_125")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_1666_cast_fp16)[name = string("transpose_126")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1668_cast_fp16 = matmul(transpose_x = var_1668_transpose_x_0, transpose_y = var_1668_transpose_y_0, x = transpose_116, y = transpose_117)[name = string("op_1668_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1669_cast_fp16 = add(x = var_1668_cast_fp16, y = key_bias)[name = string("op_1669_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_45_cast_fp16 = softmax(axis = var_1624, x = var_1669_cast_fp16)[name = string("att_45_cast_fp16")]; |
| bool var_1671_transpose_x_0 = const()[name = string("op_1671_transpose_x_0"), val = bool(false)]; |
| bool var_1671_transpose_y_0 = const()[name = string("op_1671_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_1663_cast_fp16)[name = string("transpose_127")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1671_cast_fp16 = matmul(transpose_x = var_1671_transpose_x_0, transpose_y = var_1671_transpose_y_0, x = att_45_cast_fp16, y = v_45_cast_fp16)[name = string("op_1671_cast_fp16")]; |
| tensor<int32, [4]> var_1672_perm_0 = const()[name = string("op_1672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1673 = const()[name = string("op_1673"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1672_cast_fp16 = transpose(perm = var_1672_perm_0, x = var_1671_cast_fp16)[name = string("transpose_124")]; |
| tensor<fp16, [1, 104, 1024]> input_223_cast_fp16 = reshape(shape = var_1673, x = var_1672_cast_fp16)[name = string("input_223_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_22_o_weight_to_fp16 = const()[name = string("m_layers_22_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560558656)))]; |
| tensor<fp16, [1024]> m_layers_22_o_bias_to_fp16 = const()[name = string("m_layers_22_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562655872)))]; |
| tensor<fp16, [1, 104, 1024]> linear_135_cast_fp16 = linear(bias = m_layers_22_o_bias_to_fp16, weight = m_layers_22_o_weight_to_fp16, x = input_223_cast_fp16)[name = string("linear_135_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_225_cast_fp16 = add(x = x_45_cast_fp16, y = linear_135_cast_fp16)[name = string("input_225_cast_fp16")]; |
| tensor<int32, [1]> input_227_axes_0 = const()[name = string("input_227_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_22_ln2_weight_to_fp16 = const()[name = string("m_layers_22_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562657984)))]; |
| tensor<fp16, [1024]> m_layers_22_ln2_bias_to_fp16 = const()[name = string("m_layers_22_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562660096)))]; |
| tensor<fp16, [1, 104, 1024]> input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = m_layers_22_ln2_bias_to_fp16, epsilon = var_1627_to_fp16, gamma = m_layers_22_ln2_weight_to_fp16, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_22_fc1_weight_to_fp16 = const()[name = string("m_layers_22_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562662208)))]; |
| tensor<fp16, [4096]> m_layers_22_fc1_bias_to_fp16 = const()[name = string("m_layers_22_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571050880)))]; |
| tensor<fp16, [1, 104, 4096]> linear_136_cast_fp16 = linear(bias = m_layers_22_fc1_bias_to_fp16, weight = m_layers_22_fc1_weight_to_fp16, x = input_227_cast_fp16)[name = string("linear_136_cast_fp16")]; |
| string input_229_mode_0 = const()[name = string("input_229_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_136_cast_fp16)[name = string("input_229_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_22_fc2_weight_to_fp16 = const()[name = string("m_layers_22_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571059136)))]; |
| tensor<fp16, [1024]> m_layers_22_fc2_bias_to_fp16 = const()[name = string("m_layers_22_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579447808)))]; |
| tensor<fp16, [1, 104, 1024]> linear_137_cast_fp16 = linear(bias = m_layers_22_fc2_bias_to_fp16, weight = m_layers_22_fc2_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_137_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> x_cast_fp16 = add(x = input_225_cast_fp16, y = linear_137_cast_fp16)[name = string("x_cast_fp16")]; |
| int32 var_1694 = const()[name = string("op_1694"), val = int32(-1)]; |
| tensor<int32, [1]> input_231_axes_0 = const()[name = string("input_231_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_23_ln1_weight_to_fp16 = const()[name = string("m_layers_23_ln1_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579449920)))]; |
| tensor<fp16, [1024]> m_layers_23_ln1_bias_to_fp16 = const()[name = string("m_layers_23_ln1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579452032)))]; |
| fp16 var_1697_to_fp16 = const()[name = string("op_1697_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = m_layers_23_ln1_bias_to_fp16, epsilon = var_1697_to_fp16, gamma = m_layers_23_ln1_weight_to_fp16, x = x_cast_fp16)[name = string("input_231_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_23_q_weight_to_fp16 = const()[name = string("m_layers_23_q_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579454144)))]; |
| tensor<fp16, [1024]> m_layers_23_q_bias_to_fp16 = const()[name = string("m_layers_23_q_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581551360)))]; |
| tensor<fp16, [1, 104, 1024]> linear_138_cast_fp16 = linear(bias = m_layers_23_q_bias_to_fp16, weight = m_layers_23_q_weight_to_fp16, x = input_231_cast_fp16)[name = string("linear_138_cast_fp16")]; |
| tensor<int32, [4]> var_1720 = const()[name = string("op_1720"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1721_cast_fp16 = reshape(shape = var_1720, x = linear_138_cast_fp16)[name = string("op_1721_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_23_k_weight_to_fp16 = const()[name = string("m_layers_23_k_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581553472)))]; |
| tensor<fp16, [1024]> m_layers_23_k_bias_to_fp16 = const()[name = string("m_layers_23_k_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583650688)))]; |
| tensor<fp16, [1, 104, 1024]> linear_139_cast_fp16 = linear(bias = m_layers_23_k_bias_to_fp16, weight = m_layers_23_k_weight_to_fp16, x = input_231_cast_fp16)[name = string("linear_139_cast_fp16")]; |
| tensor<int32, [4]> var_1726 = const()[name = string("op_1726"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1727_cast_fp16 = reshape(shape = var_1726, x = linear_139_cast_fp16)[name = string("op_1727_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_23_v_weight_to_fp16 = const()[name = string("m_layers_23_v_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583652800)))]; |
| tensor<fp16, [1024]> m_layers_23_v_bias_to_fp16 = const()[name = string("m_layers_23_v_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585750016)))]; |
| tensor<fp16, [1, 104, 1024]> linear_140_cast_fp16 = linear(bias = m_layers_23_v_bias_to_fp16, weight = m_layers_23_v_weight_to_fp16, x = input_231_cast_fp16)[name = string("linear_140_cast_fp16")]; |
| tensor<int32, [4]> var_1732 = const()[name = string("op_1732"), val = tensor<int32, [4]>([1, 104, 16, -1])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1733_cast_fp16 = reshape(shape = var_1732, x = linear_140_cast_fp16)[name = string("op_1733_cast_fp16")]; |
| tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| fp16 var_1735_to_fp16 = const()[name = string("op_1735_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 104, 16, 64]> var_1736_cast_fp16 = mul(x = var_1721_cast_fp16, y = var_1735_to_fp16)[name = string("op_1736_cast_fp16")]; |
| bool var_1738_transpose_x_0 = const()[name = string("op_1738_transpose_x_0"), val = bool(false)]; |
| bool var_1738_transpose_y_0 = const()[name = string("op_1738_transpose_y_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_118_perm_0 = const()[name = string("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_119_perm_0 = const()[name = string("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])]; |
| tensor<fp16, [1, 16, 64, 104]> transpose_119 = transpose(perm = transpose_119_perm_0, x = var_1727_cast_fp16)[name = string("transpose_121")]; |
| tensor<fp16, [1, 16, 104, 64]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_1736_cast_fp16)[name = string("transpose_122")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1738_cast_fp16 = matmul(transpose_x = var_1738_transpose_x_0, transpose_y = var_1738_transpose_y_0, x = transpose_118, y = transpose_119)[name = string("op_1738_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> var_1739_cast_fp16 = add(x = var_1738_cast_fp16, y = key_bias)[name = string("op_1739_cast_fp16")]; |
| tensor<fp16, [1, 16, 104, 104]> att_cast_fp16 = softmax(axis = var_1694, x = var_1739_cast_fp16)[name = string("att_cast_fp16")]; |
| bool var_1741_transpose_x_0 = const()[name = string("op_1741_transpose_x_0"), val = bool(false)]; |
| bool var_1741_transpose_y_0 = const()[name = string("op_1741_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 16, 104, 64]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_1733_cast_fp16)[name = string("transpose_123")]; |
| tensor<fp16, [1, 16, 104, 64]> var_1741_cast_fp16 = matmul(transpose_x = var_1741_transpose_x_0, transpose_y = var_1741_transpose_y_0, x = att_cast_fp16, y = v_47_cast_fp16)[name = string("op_1741_cast_fp16")]; |
| tensor<int32, [4]> var_1742_perm_0 = const()[name = string("op_1742_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_1743 = const()[name = string("op_1743"), val = tensor<int32, [3]>([1, 104, 1024])]; |
| tensor<fp16, [1, 104, 16, 64]> var_1742_cast_fp16 = transpose(perm = var_1742_perm_0, x = var_1741_cast_fp16)[name = string("transpose_120")]; |
| tensor<fp16, [1, 104, 1024]> input_233_cast_fp16 = reshape(shape = var_1743, x = var_1742_cast_fp16)[name = string("input_233_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_layers_23_o_weight_to_fp16 = const()[name = string("m_layers_23_o_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585752128)))]; |
| tensor<fp16, [1024]> m_layers_23_o_bias_to_fp16 = const()[name = string("m_layers_23_o_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587849344)))]; |
| tensor<fp16, [1, 104, 1024]> linear_141_cast_fp16 = linear(bias = m_layers_23_o_bias_to_fp16, weight = m_layers_23_o_weight_to_fp16, x = input_233_cast_fp16)[name = string("linear_141_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_235_cast_fp16 = add(x = x_cast_fp16, y = linear_141_cast_fp16)[name = string("input_235_cast_fp16")]; |
| tensor<int32, [1]> input_237_axes_0 = const()[name = string("input_237_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_layers_23_ln2_weight_to_fp16 = const()[name = string("m_layers_23_ln2_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587851456)))]; |
| tensor<fp16, [1024]> m_layers_23_ln2_bias_to_fp16 = const()[name = string("m_layers_23_ln2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587853568)))]; |
| tensor<fp16, [1, 104, 1024]> input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = m_layers_23_ln2_bias_to_fp16, epsilon = var_1697_to_fp16, gamma = m_layers_23_ln2_weight_to_fp16, x = input_235_cast_fp16)[name = string("input_237_cast_fp16")]; |
| tensor<fp16, [4096, 1024]> m_layers_23_fc1_weight_to_fp16 = const()[name = string("m_layers_23_fc1_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587855680)))]; |
| tensor<fp16, [4096]> m_layers_23_fc1_bias_to_fp16 = const()[name = string("m_layers_23_fc1_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596244352)))]; |
| tensor<fp16, [1, 104, 4096]> linear_142_cast_fp16 = linear(bias = m_layers_23_fc1_bias_to_fp16, weight = m_layers_23_fc1_weight_to_fp16, x = input_237_cast_fp16)[name = string("linear_142_cast_fp16")]; |
| string input_239_mode_0 = const()[name = string("input_239_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 4096]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_142_cast_fp16)[name = string("input_239_cast_fp16")]; |
| tensor<fp16, [1024, 4096]> m_layers_23_fc2_weight_to_fp16 = const()[name = string("m_layers_23_fc2_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(596252608)))]; |
| tensor<fp16, [1024]> m_layers_23_fc2_bias_to_fp16 = const()[name = string("m_layers_23_fc2_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604641280)))]; |
| tensor<fp16, [1, 104, 1024]> linear_143_cast_fp16 = linear(bias = m_layers_23_fc2_bias_to_fp16, weight = m_layers_23_fc2_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_143_cast_fp16")]; |
| tensor<fp16, [1, 104, 1024]> input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_143_cast_fp16)[name = string("input_241_cast_fp16")]; |
| tensor<int32, [1]> input_243_axes_0 = const()[name = string("input_243_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [1024]> m_ln_post_weight_to_fp16 = const()[name = string("m_ln_post_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604643392)))]; |
| tensor<fp16, [1024]> m_ln_post_bias_to_fp16 = const()[name = string("m_ln_post_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604645504)))]; |
| fp16 var_1762_to_fp16 = const()[name = string("op_1762_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 104, 1024]> input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = m_ln_post_bias_to_fp16, epsilon = var_1762_to_fp16, gamma = m_ln_post_weight_to_fp16, x = input_241_cast_fp16)[name = string("input_243_cast_fp16")]; |
| tensor<fp16, [1024, 1024]> m_proj1_weight_to_fp16 = const()[name = string("m_proj1_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604647616)))]; |
| tensor<fp16, [1024]> m_proj1_bias_to_fp16 = const()[name = string("m_proj1_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606744832)))]; |
| tensor<fp16, [1, 104, 1024]> linear_144_cast_fp16 = linear(bias = m_proj1_bias_to_fp16, weight = m_proj1_weight_to_fp16, x = input_243_cast_fp16)[name = string("linear_144_cast_fp16")]; |
| string input_mode_0 = const()[name = string("input_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 104, 1024]> input_cast_fp16 = gelu(mode = input_mode_0, x = linear_144_cast_fp16)[name = string("input_cast_fp16")]; |
| tensor<fp16, [2048, 1024]> m_proj2_weight_to_fp16 = const()[name = string("m_proj2_weight_to_fp16"), val = tensor<fp16, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(606746944)))]; |
| tensor<fp16, [2048]> m_proj2_bias_to_fp16 = const()[name = string("m_proj2_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610941312)))]; |
| tensor<fp16, [1, 104, 2048]> tokens = linear(bias = m_proj2_bias_to_fp16, weight = m_proj2_weight_to_fp16, x = input_cast_fp16)[name = string("linear_145_cast_fp16")]; |
| } -> (tokens); |
| } |
