| program(1.3) |
| [buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.2"}})] |
| { |
| func main<ios18>(tensor<fp16, [1, 80, 3000]> mel) { |
| int32 var_10 = const()[name = string("op_10"), val = int32(-2)]; |
| tensor<int32, [3]> var_13 = const()[name = string("op_13"), val = tensor<int32, [3]>([0, 2, 1])]; |
| string var_188_pad_type_0 = const()[name = string("op_188_pad_type_0"), val = string("custom")]; |
| tensor<int32, [2]> var_188_pad_0 = const()[name = string("op_188_pad_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [1]> var_188_strides_0 = const()[name = string("op_188_strides_0"), val = tensor<int32, [1]>([1])]; |
| tensor<int32, [1]> var_188_dilations_0 = const()[name = string("op_188_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 var_188_groups_0 = const()[name = string("op_188_groups_0"), val = int32(1)]; |
| tensor<fp16, [384, 80, 3]> self_encoder_conv1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 80, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64))), lut = tensor<fp16, [12, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(69248))))[name = string("self_encoder_conv1_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_conv1_bias_to_fp16 = const()[name = string("self_encoder_conv1_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(70848)))]; |
| tensor<fp16, [1, 384, 3000]> var_188_cast_fp16 = conv(bias = self_encoder_conv1_bias_to_fp16, dilations = var_188_dilations_0, groups = var_188_groups_0, pad = var_188_pad_0, pad_type = var_188_pad_type_0, strides = var_188_strides_0, weight = self_encoder_conv1_weight_to_fp16_palettized, x = mel)[name = string("op_188_cast_fp16")]; |
| string input_1_mode_0 = const()[name = string("input_1_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 384, 3000]> input_1_cast_fp16 = gelu(mode = input_1_mode_0, x = var_188_cast_fp16)[name = string("input_1_cast_fp16")]; |
| string var_190_pad_type_0 = const()[name = string("op_190_pad_type_0"), val = string("custom")]; |
| tensor<int32, [2]> var_190_pad_0 = const()[name = string("op_190_pad_0"), val = tensor<int32, [2]>([1, 1])]; |
| tensor<int32, [1]> var_190_strides_0 = const()[name = string("op_190_strides_0"), val = tensor<int32, [1]>([2])]; |
| tensor<int32, [1]> var_190_dilations_0 = const()[name = string("op_190_dilations_0"), val = tensor<int32, [1]>([1])]; |
| int32 var_190_groups_0 = const()[name = string("op_190_groups_0"), val = int32(1)]; |
| tensor<fp16, [384, 384, 3]> self_encoder_conv2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(71680))), lut = tensor<fp16, [12, 1, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403520))))[name = string("self_encoder_conv2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_conv2_bias_to_fp16 = const()[name = string("self_encoder_conv2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405120)))]; |
| tensor<fp16, [1, 384, 1500]> var_190_cast_fp16 = conv(bias = self_encoder_conv2_bias_to_fp16, dilations = var_190_dilations_0, groups = var_190_groups_0, pad = var_190_pad_0, pad_type = var_190_pad_type_0, strides = var_190_strides_0, weight = self_encoder_conv2_weight_to_fp16_palettized, x = input_1_cast_fp16)[name = string("op_190_cast_fp16")]; |
| string inputs_embeds_1_mode_0 = const()[name = string("inputs_embeds_1_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 384, 1500]> inputs_embeds_1_cast_fp16 = gelu(mode = inputs_embeds_1_mode_0, x = var_190_cast_fp16)[name = string("inputs_embeds_1_cast_fp16")]; |
| tensor<fp16, [1500, 384]> var_12_to_fp16 = const()[name = string("op_12_to_fp16"), val = tensor<fp16, [1500, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405952)))]; |
| tensor<fp16, [1, 1500, 384]> inputs_embeds_3_cast_fp16 = transpose(perm = var_13, x = inputs_embeds_1_cast_fp16)[name = string("transpose_103")]; |
| tensor<fp16, [1, 1500, 384]> input_5_cast_fp16 = add(x = inputs_embeds_3_cast_fp16, y = var_12_to_fp16)[name = string("input_5_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_1_axes_0 = const()[name = string("hidden_states_1_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_0_self_attn_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_0_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1558016)))]; |
| tensor<fp16, [384]> self_encoder_layers_0_self_attn_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_0_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1558848)))]; |
| fp16 var_29_to_fp16 = const()[name = string("op_29_to_fp16"), val = fp16(0x1.5p-17)]; |
| tensor<fp16, [1, 1500, 384]> hidden_states_1_cast_fp16 = layer_norm(axes = hidden_states_1_axes_0, beta = self_encoder_layers_0_self_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_0_self_attn_layer_norm_weight_to_fp16, x = input_5_cast_fp16)[name = string("hidden_states_1_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_0_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1559680))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1670336))))[name = string("self_encoder_layers_0_self_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_0_self_attn_q_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_0_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1671936)))]; |
| tensor<fp16, [1, 1500, 384]> linear_0_cast_fp16 = linear(bias = self_encoder_layers_0_self_attn_q_proj_bias_to_fp16, weight = self_encoder_layers_0_self_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_1_cast_fp16)[name = string("linear_0_cast_fp16")]; |
| fp16 var_198_to_fp16 = const()[name = string("op_198_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1500, 384]> var_199_cast_fp16 = mul(x = linear_0_cast_fp16, y = var_198_to_fp16)[name = string("op_199_cast_fp16")]; |
| tensor<int32, [4]> var_200 = const()[name = string("op_200"), val = tensor<int32, [4]>([1, 1500, -1, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_201_cast_fp16 = reshape(shape = var_200, x = var_199_cast_fp16)[name = string("op_201_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_0_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1672768))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1783424))))[name = string("self_encoder_layers_0_self_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1785024)))]; |
| tensor<fp16, [1, 1500, 384]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_encoder_layers_0_self_attn_k_proj_weight_to_fp16_palettized, x = hidden_states_1_cast_fp16)[name = string("linear_1_cast_fp16")]; |
| tensor<int32, [4]> var_205 = const()[name = string("op_205"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_206_cast_fp16 = reshape(shape = var_205, x = linear_1_cast_fp16)[name = string("op_206_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_0_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1785856))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1896512))))[name = string("self_encoder_layers_0_self_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_0_self_attn_v_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_0_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1898112)))]; |
| tensor<fp16, [1, 1500, 384]> linear_2_cast_fp16 = linear(bias = self_encoder_layers_0_self_attn_v_proj_bias_to_fp16, weight = self_encoder_layers_0_self_attn_v_proj_weight_to_fp16_palettized, x = hidden_states_1_cast_fp16)[name = string("linear_2_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_210_cast_fp16 = reshape(shape = var_205, x = linear_2_cast_fp16)[name = string("op_210_cast_fp16")]; |
| tensor<int32, [4]> var_211_perm_0 = const()[name = string("op_211_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| bool matmul_0_transpose_y_0 = const()[name = string("matmul_0_transpose_y_0"), val = bool(true)]; |
| bool matmul_0_transpose_x_0 = const()[name = string("matmul_0_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_40_perm_0 = const()[name = string("transpose_40_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_41_perm_0 = const()[name = string("transpose_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_41 = transpose(perm = transpose_41_perm_0, x = var_206_cast_fp16)[name = string("transpose_101")]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_40 = transpose(perm = transpose_40_perm_0, x = var_201_cast_fp16)[name = string("transpose_102")]; |
| tensor<fp16, [1, 6, 1500, 1500]> matmul_0_cast_fp16 = matmul(transpose_x = matmul_0_transpose_x_0, transpose_y = matmul_0_transpose_y_0, x = transpose_40, y = transpose_41)[name = string("matmul_0_cast_fp16")]; |
| int32 softmax_0_axis_0 = const()[name = string("softmax_0_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1500, 1500]> softmax_0_cast_fp16 = softmax(axis = softmax_0_axis_0, x = matmul_0_cast_fp16)[name = string("softmax_0_cast_fp16")]; |
| bool attn_output_1_transpose_x_0 = const()[name = string("attn_output_1_transpose_x_0"), val = bool(false)]; |
| bool attn_output_1_transpose_y_0 = const()[name = string("attn_output_1_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_211_cast_fp16 = transpose(perm = var_211_perm_0, x = var_210_cast_fp16)[name = string("transpose_100")]; |
| tensor<fp16, [1, 6, 1500, 64]> attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = softmax_0_cast_fp16, y = var_211_cast_fp16)[name = string("attn_output_1_cast_fp16")]; |
| tensor<int32, [4]> var_214_perm_0 = const()[name = string("op_214_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_216 = const()[name = string("op_216"), val = tensor<int32, [3]>([1, 1500, -1])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_214_cast_fp16 = transpose(perm = var_214_perm_0, x = attn_output_1_cast_fp16)[name = string("transpose_99")]; |
| tensor<fp16, [1, 1500, 384]> var_217_cast_fp16 = reshape(shape = var_216, x = var_214_cast_fp16)[name = string("op_217_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_0_self_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1898944))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2009600))))[name = string("self_encoder_layers_0_self_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_0_self_attn_out_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_0_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2011200)))]; |
| tensor<fp16, [1, 1500, 384]> linear_3_cast_fp16 = linear(bias = self_encoder_layers_0_self_attn_out_proj_bias_to_fp16, weight = self_encoder_layers_0_self_attn_out_proj_weight_to_fp16_palettized, x = var_217_cast_fp16)[name = string("linear_3_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_13_cast_fp16 = add(x = input_5_cast_fp16, y = linear_3_cast_fp16)[name = string("input_13_cast_fp16")]; |
| tensor<int32, [1]> input_15_axes_0 = const()[name = string("input_15_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_0_final_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_0_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2012032)))]; |
| tensor<fp16, [384]> self_encoder_layers_0_final_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_0_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2012864)))]; |
| tensor<fp16, [1, 1500, 384]> input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = self_encoder_layers_0_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_0_final_layer_norm_weight_to_fp16, x = input_13_cast_fp16)[name = string("input_15_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_encoder_layers_0_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2013696))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2456128))))[name = string("self_encoder_layers_0_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_encoder_layers_0_fc1_bias_to_fp16 = const()[name = string("self_encoder_layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2462336)))]; |
| tensor<fp16, [1, 1500, 1536]> linear_4_cast_fp16 = linear(bias = self_encoder_layers_0_fc1_bias_to_fp16, weight = self_encoder_layers_0_fc1_weight_to_fp16_palettized, x = input_15_cast_fp16)[name = string("linear_4_cast_fp16")]; |
| string input_19_mode_0 = const()[name = string("input_19_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1500, 1536]> input_19_cast_fp16 = gelu(mode = input_19_mode_0, x = linear_4_cast_fp16)[name = string("input_19_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_encoder_layers_0_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2465472))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2907904))))[name = string("self_encoder_layers_0_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_0_fc2_bias_to_fp16 = const()[name = string("self_encoder_layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2909504)))]; |
| tensor<fp16, [1, 1500, 384]> linear_5_cast_fp16 = linear(bias = self_encoder_layers_0_fc2_bias_to_fp16, weight = self_encoder_layers_0_fc2_weight_to_fp16_palettized, x = input_19_cast_fp16)[name = string("linear_5_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_25_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = string("input_25_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_7_axes_0 = const()[name = string("hidden_states_7_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_1_self_attn_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_1_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2910336)))]; |
| tensor<fp16, [384]> self_encoder_layers_1_self_attn_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_1_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2911168)))]; |
| tensor<fp16, [1, 1500, 384]> hidden_states_7_cast_fp16 = layer_norm(axes = hidden_states_7_axes_0, beta = self_encoder_layers_1_self_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_1_self_attn_layer_norm_weight_to_fp16, x = input_25_cast_fp16)[name = string("hidden_states_7_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2912000))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3022656))))[name = string("self_encoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3024256)))]; |
| tensor<fp16, [1, 1500, 384]> linear_6_cast_fp16 = linear(bias = self_encoder_layers_1_self_attn_q_proj_bias_to_fp16, weight = self_encoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_7_cast_fp16)[name = string("linear_6_cast_fp16")]; |
| fp16 var_230_to_fp16 = const()[name = string("op_230_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1500, 384]> var_231_cast_fp16 = mul(x = linear_6_cast_fp16, y = var_230_to_fp16)[name = string("op_231_cast_fp16")]; |
| tensor<int32, [4]> var_232 = const()[name = string("op_232"), val = tensor<int32, [4]>([1, 1500, -1, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_233_cast_fp16 = reshape(shape = var_232, x = var_231_cast_fp16)[name = string("op_233_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3025088))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3135744))))[name = string("self_encoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1500, 384]> linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_encoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = hidden_states_7_cast_fp16)[name = string("linear_7_cast_fp16")]; |
| tensor<int32, [4]> var_237 = const()[name = string("op_237"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_238_cast_fp16 = reshape(shape = var_237, x = linear_7_cast_fp16)[name = string("op_238_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3137344))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3248000))))[name = string("self_encoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3249600)))]; |
| tensor<fp16, [1, 1500, 384]> linear_8_cast_fp16 = linear(bias = self_encoder_layers_1_self_attn_v_proj_bias_to_fp16, weight = self_encoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = hidden_states_7_cast_fp16)[name = string("linear_8_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_242_cast_fp16 = reshape(shape = var_237, x = linear_8_cast_fp16)[name = string("op_242_cast_fp16")]; |
| tensor<int32, [4]> var_243_perm_0 = const()[name = string("op_243_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| bool matmul_1_transpose_y_0 = const()[name = string("matmul_1_transpose_y_0"), val = bool(true)]; |
| bool matmul_1_transpose_x_0 = const()[name = string("matmul_1_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_42_perm_0 = const()[name = string("transpose_42_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_43_perm_0 = const()[name = string("transpose_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_43 = transpose(perm = transpose_43_perm_0, x = var_238_cast_fp16)[name = string("transpose_97")]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_42 = transpose(perm = transpose_42_perm_0, x = var_233_cast_fp16)[name = string("transpose_98")]; |
| tensor<fp16, [1, 6, 1500, 1500]> matmul_1_cast_fp16 = matmul(transpose_x = matmul_1_transpose_x_0, transpose_y = matmul_1_transpose_y_0, x = transpose_42, y = transpose_43)[name = string("matmul_1_cast_fp16")]; |
| int32 softmax_1_axis_0 = const()[name = string("softmax_1_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1500, 1500]> softmax_1_cast_fp16 = softmax(axis = softmax_1_axis_0, x = matmul_1_cast_fp16)[name = string("softmax_1_cast_fp16")]; |
| bool attn_output_5_transpose_x_0 = const()[name = string("attn_output_5_transpose_x_0"), val = bool(false)]; |
| bool attn_output_5_transpose_y_0 = const()[name = string("attn_output_5_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_243_cast_fp16 = transpose(perm = var_243_perm_0, x = var_242_cast_fp16)[name = string("transpose_96")]; |
| tensor<fp16, [1, 6, 1500, 64]> attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = softmax_1_cast_fp16, y = var_243_cast_fp16)[name = string("attn_output_5_cast_fp16")]; |
| tensor<int32, [4]> var_246_perm_0 = const()[name = string("op_246_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_248 = const()[name = string("op_248"), val = tensor<int32, [3]>([1, 1500, -1])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_246_cast_fp16 = transpose(perm = var_246_perm_0, x = attn_output_5_cast_fp16)[name = string("transpose_95")]; |
| tensor<fp16, [1, 1500, 384]> var_249_cast_fp16 = reshape(shape = var_248, x = var_246_cast_fp16)[name = string("op_249_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_1_self_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3250432))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3361088))))[name = string("self_encoder_layers_1_self_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3362688)))]; |
| tensor<fp16, [1, 1500, 384]> linear_9_cast_fp16 = linear(bias = self_encoder_layers_1_self_attn_out_proj_bias_to_fp16, weight = self_encoder_layers_1_self_attn_out_proj_weight_to_fp16_palettized, x = var_249_cast_fp16)[name = string("linear_9_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = string("input_31_cast_fp16")]; |
| tensor<int32, [1]> input_33_axes_0 = const()[name = string("input_33_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_1_final_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_1_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3363520)))]; |
| tensor<fp16, [384]> self_encoder_layers_1_final_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_1_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3364352)))]; |
| tensor<fp16, [1, 1500, 384]> input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = self_encoder_layers_1_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_1_final_layer_norm_weight_to_fp16, x = input_31_cast_fp16)[name = string("input_33_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_encoder_layers_1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3365184))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3807616))))[name = string("self_encoder_layers_1_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_encoder_layers_1_fc1_bias_to_fp16 = const()[name = string("self_encoder_layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3813824)))]; |
| tensor<fp16, [1, 1500, 1536]> linear_10_cast_fp16 = linear(bias = self_encoder_layers_1_fc1_bias_to_fp16, weight = self_encoder_layers_1_fc1_weight_to_fp16_palettized, x = input_33_cast_fp16)[name = string("linear_10_cast_fp16")]; |
| string input_37_mode_0 = const()[name = string("input_37_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1500, 1536]> input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_10_cast_fp16)[name = string("input_37_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_encoder_layers_1_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(3816960))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4259392))))[name = string("self_encoder_layers_1_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_1_fc2_bias_to_fp16 = const()[name = string("self_encoder_layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4260992)))]; |
| tensor<fp16, [1, 1500, 384]> linear_11_cast_fp16 = linear(bias = self_encoder_layers_1_fc2_bias_to_fp16, weight = self_encoder_layers_1_fc2_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = string("linear_11_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_43_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = string("input_43_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_13_axes_0 = const()[name = string("hidden_states_13_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_2_self_attn_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_2_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4261824)))]; |
| tensor<fp16, [384]> self_encoder_layers_2_self_attn_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_2_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4262656)))]; |
| tensor<fp16, [1, 1500, 384]> hidden_states_13_cast_fp16 = layer_norm(axes = hidden_states_13_axes_0, beta = self_encoder_layers_2_self_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_2_self_attn_layer_norm_weight_to_fp16, x = input_43_cast_fp16)[name = string("hidden_states_13_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4263488))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4374144))))[name = string("self_encoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4375744)))]; |
| tensor<fp16, [1, 1500, 384]> linear_12_cast_fp16 = linear(bias = self_encoder_layers_2_self_attn_q_proj_bias_to_fp16, weight = self_encoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_13_cast_fp16)[name = string("linear_12_cast_fp16")]; |
| fp16 var_262_to_fp16 = const()[name = string("op_262_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1500, 384]> var_263_cast_fp16 = mul(x = linear_12_cast_fp16, y = var_262_to_fp16)[name = string("op_263_cast_fp16")]; |
| tensor<int32, [4]> var_264 = const()[name = string("op_264"), val = tensor<int32, [4]>([1, 1500, -1, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_265_cast_fp16 = reshape(shape = var_264, x = var_263_cast_fp16)[name = string("op_265_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4376576))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4487232))))[name = string("self_encoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1500, 384]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_encoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = hidden_states_13_cast_fp16)[name = string("linear_13_cast_fp16")]; |
| tensor<int32, [4]> var_269 = const()[name = string("op_269"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_270_cast_fp16 = reshape(shape = var_269, x = linear_13_cast_fp16)[name = string("op_270_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4488832))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4599488))))[name = string("self_encoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601088)))]; |
| tensor<fp16, [1, 1500, 384]> linear_14_cast_fp16 = linear(bias = self_encoder_layers_2_self_attn_v_proj_bias_to_fp16, weight = self_encoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = hidden_states_13_cast_fp16)[name = string("linear_14_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_274_cast_fp16 = reshape(shape = var_269, x = linear_14_cast_fp16)[name = string("op_274_cast_fp16")]; |
| tensor<int32, [4]> var_275_perm_0 = const()[name = string("op_275_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| bool matmul_2_transpose_y_0 = const()[name = string("matmul_2_transpose_y_0"), val = bool(true)]; |
| bool matmul_2_transpose_x_0 = const()[name = string("matmul_2_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_44_perm_0 = const()[name = string("transpose_44_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_45_perm_0 = const()[name = string("transpose_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_45 = transpose(perm = transpose_45_perm_0, x = var_270_cast_fp16)[name = string("transpose_93")]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_44 = transpose(perm = transpose_44_perm_0, x = var_265_cast_fp16)[name = string("transpose_94")]; |
| tensor<fp16, [1, 6, 1500, 1500]> matmul_2_cast_fp16 = matmul(transpose_x = matmul_2_transpose_x_0, transpose_y = matmul_2_transpose_y_0, x = transpose_44, y = transpose_45)[name = string("matmul_2_cast_fp16")]; |
| int32 softmax_2_axis_0 = const()[name = string("softmax_2_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1500, 1500]> softmax_2_cast_fp16 = softmax(axis = softmax_2_axis_0, x = matmul_2_cast_fp16)[name = string("softmax_2_cast_fp16")]; |
| bool attn_output_9_transpose_x_0 = const()[name = string("attn_output_9_transpose_x_0"), val = bool(false)]; |
| bool attn_output_9_transpose_y_0 = const()[name = string("attn_output_9_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_275_cast_fp16 = transpose(perm = var_275_perm_0, x = var_274_cast_fp16)[name = string("transpose_92")]; |
| tensor<fp16, [1, 6, 1500, 64]> attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = softmax_2_cast_fp16, y = var_275_cast_fp16)[name = string("attn_output_9_cast_fp16")]; |
| tensor<int32, [4]> var_278_perm_0 = const()[name = string("op_278_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_280 = const()[name = string("op_280"), val = tensor<int32, [3]>([1, 1500, -1])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_278_cast_fp16 = transpose(perm = var_278_perm_0, x = attn_output_9_cast_fp16)[name = string("transpose_91")]; |
| tensor<fp16, [1, 1500, 384]> var_281_cast_fp16 = reshape(shape = var_280, x = var_278_cast_fp16)[name = string("op_281_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_2_self_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4601920))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4712576))))[name = string("self_encoder_layers_2_self_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4714176)))]; |
| tensor<fp16, [1, 1500, 384]> linear_15_cast_fp16 = linear(bias = self_encoder_layers_2_self_attn_out_proj_bias_to_fp16, weight = self_encoder_layers_2_self_attn_out_proj_weight_to_fp16_palettized, x = var_281_cast_fp16)[name = string("linear_15_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = string("input_49_cast_fp16")]; |
| tensor<int32, [1]> input_51_axes_0 = const()[name = string("input_51_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_2_final_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_2_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4715008)))]; |
| tensor<fp16, [384]> self_encoder_layers_2_final_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_2_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4715840)))]; |
| tensor<fp16, [1, 1500, 384]> input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = self_encoder_layers_2_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_2_final_layer_norm_weight_to_fp16, x = input_49_cast_fp16)[name = string("input_51_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_encoder_layers_2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(4716672))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5159104))))[name = string("self_encoder_layers_2_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_encoder_layers_2_fc1_bias_to_fp16 = const()[name = string("self_encoder_layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5165312)))]; |
| tensor<fp16, [1, 1500, 1536]> linear_16_cast_fp16 = linear(bias = self_encoder_layers_2_fc1_bias_to_fp16, weight = self_encoder_layers_2_fc1_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = string("linear_16_cast_fp16")]; |
| string input_55_mode_0 = const()[name = string("input_55_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1500, 1536]> input_55_cast_fp16 = gelu(mode = input_55_mode_0, x = linear_16_cast_fp16)[name = string("input_55_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_encoder_layers_2_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5168448))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5610880))))[name = string("self_encoder_layers_2_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_2_fc2_bias_to_fp16 = const()[name = string("self_encoder_layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5612480)))]; |
| tensor<fp16, [1, 1500, 384]> linear_17_cast_fp16 = linear(bias = self_encoder_layers_2_fc2_bias_to_fp16, weight = self_encoder_layers_2_fc2_weight_to_fp16_palettized, x = input_55_cast_fp16)[name = string("linear_17_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_61_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = string("input_61_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_19_axes_0 = const()[name = string("hidden_states_19_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_3_self_attn_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_3_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5613312)))]; |
| tensor<fp16, [384]> self_encoder_layers_3_self_attn_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_3_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5614144)))]; |
| tensor<fp16, [1, 1500, 384]> hidden_states_19_cast_fp16 = layer_norm(axes = hidden_states_19_axes_0, beta = self_encoder_layers_3_self_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_3_self_attn_layer_norm_weight_to_fp16, x = input_61_cast_fp16)[name = string("hidden_states_19_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5614976))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5725632))))[name = string("self_encoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5727232)))]; |
| tensor<fp16, [1, 1500, 384]> linear_18_cast_fp16 = linear(bias = self_encoder_layers_3_self_attn_q_proj_bias_to_fp16, weight = self_encoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_19_cast_fp16)[name = string("linear_18_cast_fp16")]; |
| fp16 var_294_to_fp16 = const()[name = string("op_294_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1500, 384]> var_295_cast_fp16 = mul(x = linear_18_cast_fp16, y = var_294_to_fp16)[name = string("op_295_cast_fp16")]; |
| tensor<int32, [4]> var_296 = const()[name = string("op_296"), val = tensor<int32, [4]>([1, 1500, -1, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_297_cast_fp16 = reshape(shape = var_296, x = var_295_cast_fp16)[name = string("op_297_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5728064))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5838720))))[name = string("self_encoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1500, 384]> linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_encoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = hidden_states_19_cast_fp16)[name = string("linear_19_cast_fp16")]; |
| tensor<int32, [4]> var_301 = const()[name = string("op_301"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_302_cast_fp16 = reshape(shape = var_301, x = linear_19_cast_fp16)[name = string("op_302_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5840320))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5950976))))[name = string("self_encoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5952576)))]; |
| tensor<fp16, [1, 1500, 384]> linear_20_cast_fp16 = linear(bias = self_encoder_layers_3_self_attn_v_proj_bias_to_fp16, weight = self_encoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = hidden_states_19_cast_fp16)[name = string("linear_20_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_306_cast_fp16 = reshape(shape = var_301, x = linear_20_cast_fp16)[name = string("op_306_cast_fp16")]; |
| tensor<int32, [4]> var_307_perm_0 = const()[name = string("op_307_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| bool matmul_3_transpose_y_0 = const()[name = string("matmul_3_transpose_y_0"), val = bool(true)]; |
| bool matmul_3_transpose_x_0 = const()[name = string("matmul_3_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_46_perm_0 = const()[name = string("transpose_46_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_47_perm_0 = const()[name = string("transpose_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_47 = transpose(perm = transpose_47_perm_0, x = var_302_cast_fp16)[name = string("transpose_89")]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_46 = transpose(perm = transpose_46_perm_0, x = var_297_cast_fp16)[name = string("transpose_90")]; |
| tensor<fp16, [1, 6, 1500, 1500]> matmul_3_cast_fp16 = matmul(transpose_x = matmul_3_transpose_x_0, transpose_y = matmul_3_transpose_y_0, x = transpose_46, y = transpose_47)[name = string("matmul_3_cast_fp16")]; |
| int32 softmax_3_axis_0 = const()[name = string("softmax_3_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1500, 1500]> softmax_3_cast_fp16 = softmax(axis = softmax_3_axis_0, x = matmul_3_cast_fp16)[name = string("softmax_3_cast_fp16")]; |
| bool attn_output_13_transpose_x_0 = const()[name = string("attn_output_13_transpose_x_0"), val = bool(false)]; |
| bool attn_output_13_transpose_y_0 = const()[name = string("attn_output_13_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_307_cast_fp16 = transpose(perm = var_307_perm_0, x = var_306_cast_fp16)[name = string("transpose_88")]; |
| tensor<fp16, [1, 6, 1500, 64]> attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = softmax_3_cast_fp16, y = var_307_cast_fp16)[name = string("attn_output_13_cast_fp16")]; |
| tensor<int32, [4]> var_310_perm_0 = const()[name = string("op_310_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_312 = const()[name = string("op_312"), val = tensor<int32, [3]>([1, 1500, -1])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_310_cast_fp16 = transpose(perm = var_310_perm_0, x = attn_output_13_cast_fp16)[name = string("transpose_87")]; |
| tensor<fp16, [1, 1500, 384]> var_313_cast_fp16 = reshape(shape = var_312, x = var_310_cast_fp16)[name = string("op_313_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_encoder_layers_3_self_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5953408))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6064064))))[name = string("self_encoder_layers_3_self_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = string("self_encoder_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6065664)))]; |
| tensor<fp16, [1, 1500, 384]> linear_21_cast_fp16 = linear(bias = self_encoder_layers_3_self_attn_out_proj_bias_to_fp16, weight = self_encoder_layers_3_self_attn_out_proj_weight_to_fp16_palettized, x = var_313_cast_fp16)[name = string("linear_21_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = string("input_67_cast_fp16")]; |
| tensor<int32, [1]> input_69_axes_0 = const()[name = string("input_69_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layers_3_final_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layers_3_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6066496)))]; |
| tensor<fp16, [384]> self_encoder_layers_3_final_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layers_3_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6067328)))]; |
| tensor<fp16, [1, 1500, 384]> input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = self_encoder_layers_3_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layers_3_final_layer_norm_weight_to_fp16, x = input_67_cast_fp16)[name = string("input_69_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_encoder_layers_3_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6068160))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6510592))))[name = string("self_encoder_layers_3_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_encoder_layers_3_fc1_bias_to_fp16 = const()[name = string("self_encoder_layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6516800)))]; |
| tensor<fp16, [1, 1500, 1536]> linear_22_cast_fp16 = linear(bias = self_encoder_layers_3_fc1_bias_to_fp16, weight = self_encoder_layers_3_fc1_weight_to_fp16_palettized, x = input_69_cast_fp16)[name = string("linear_22_cast_fp16")]; |
| string input_73_mode_0 = const()[name = string("input_73_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1500, 1536]> input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_22_cast_fp16)[name = string("input_73_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_encoder_layers_3_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6519936))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6962368))))[name = string("self_encoder_layers_3_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_encoder_layers_3_fc2_bias_to_fp16 = const()[name = string("self_encoder_layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6963968)))]; |
| tensor<fp16, [1, 1500, 384]> linear_23_cast_fp16 = linear(bias = self_encoder_layers_3_fc2_bias_to_fp16, weight = self_encoder_layers_3_fc2_weight_to_fp16_palettized, x = input_73_cast_fp16)[name = string("linear_23_cast_fp16")]; |
| tensor<fp16, [1, 1500, 384]> input_79_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = string("input_79_cast_fp16")]; |
| tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_encoder_layer_norm_weight_to_fp16 = const()[name = string("self_encoder_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6964800)))]; |
| tensor<fp16, [384]> self_encoder_layer_norm_bias_to_fp16 = const()[name = string("self_encoder_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6965632)))]; |
| tensor<fp16, [1, 1500, 384]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = self_encoder_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_encoder_layer_norm_weight_to_fp16, x = input_79_cast_fp16)[name = string("input_91_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_0_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6966464))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7077120))))[name = string("self_decoder_layers_0_encoder_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1500, 384]> linear_26_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_decoder_layers_0_encoder_attn_k_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_26_cast_fp16")]; |
| tensor<int32, [4]> var_341 = const()[name = string("op_341"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_342_cast_fp16 = reshape(shape = var_341, x = linear_26_cast_fp16)[name = string("op_342_cast_fp16")]; |
| tensor<int32, [4]> var_343_perm_0 = const()[name = string("op_343_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<fp16, [384, 384]> self_decoder_layers_0_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7078720))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7189376))))[name = string("self_decoder_layers_0_encoder_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_0_encoder_attn_v_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_0_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7190976)))]; |
| tensor<fp16, [1, 1500, 384]> linear_27_cast_fp16 = linear(bias = self_decoder_layers_0_encoder_attn_v_proj_bias_to_fp16, weight = self_decoder_layers_0_encoder_attn_v_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_27_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_346_cast_fp16 = reshape(shape = var_341, x = linear_27_cast_fp16)[name = string("op_346_cast_fp16")]; |
| tensor<int32, [4]> var_347_perm_0 = const()[name = string("op_347_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool key_11_interleave_0 = const()[name = string("key_11_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_343_cast_fp16 = transpose(perm = var_343_perm_0, x = var_342_cast_fp16)[name = string("transpose_86")]; |
| tensor<fp16, [1, 6, 1500, 64]> key_11_cast_fp16 = concat(axis = var_10, interleave = key_11_interleave_0, values = var_343_cast_fp16)[name = string("key_11_cast_fp16")]; |
| bool value_11_interleave_0 = const()[name = string("value_11_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_347_cast_fp16 = transpose(perm = var_347_perm_0, x = var_346_cast_fp16)[name = string("transpose_85")]; |
| tensor<fp16, [1, 6, 1500, 64]> value_11_cast_fp16 = concat(axis = var_10, interleave = value_11_interleave_0, values = var_347_cast_fp16)[name = string("value_11_cast_fp16")]; |
| bool matmul_5_transpose_y_0 = const()[name = string("matmul_5_transpose_y_0"), val = bool(true)]; |
| bool matmul_5_transpose_x_0 = const()[name = string("matmul_5_transpose_x_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> var_338_to_fp16 = const()[name = string("op_338_to_fp16"), val = tensor<fp16, [1, 6, 1, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7191808)))]; |
| tensor<fp16, [1, 6, 1, 1500]> matmul_5_cast_fp16 = matmul(transpose_x = matmul_5_transpose_x_0, transpose_y = matmul_5_transpose_y_0, x = var_338_to_fp16, y = key_11_cast_fp16)[name = string("matmul_5_cast_fp16")]; |
| int32 softmax_5_axis_0 = const()[name = string("softmax_5_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1, 1500]> softmax_5_cast_fp16 = softmax(axis = softmax_5_axis_0, x = matmul_5_cast_fp16)[name = string("softmax_5_cast_fp16")]; |
| bool attn_output_21_transpose_x_0 = const()[name = string("attn_output_21_transpose_x_0"), val = bool(false)]; |
| bool attn_output_21_transpose_y_0 = const()[name = string("attn_output_21_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = softmax_5_cast_fp16, y = value_11_cast_fp16)[name = string("attn_output_21_cast_fp16")]; |
| tensor<int32, [4]> var_354_perm_0 = const()[name = string("op_354_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_356 = const()[name = string("op_356"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp16, [1, 1, 6, 64]> var_354_cast_fp16 = transpose(perm = var_354_perm_0, x = attn_output_21_cast_fp16)[name = string("transpose_84")]; |
| tensor<fp16, [1, 1, 384]> var_357_cast_fp16 = reshape(shape = var_356, x = var_354_cast_fp16)[name = string("op_357_cast_fp16")]; |
| tensor<fp16, [384, 384]> input_97_weight_0_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7192640))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7303296))))[name = string("input_97_weight_0_to_fp16_palettized")]; |
| tensor<fp16, [384]> input_97_bias_0_to_fp16 = const()[name = string("input_97_bias_0_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7304896)))]; |
| tensor<fp16, [1, 1, 384]> input_97_cast_fp16 = linear(bias = input_97_bias_0_to_fp16, weight = input_97_weight_0_to_fp16_palettized, x = var_357_cast_fp16)[name = string("input_97_cast_fp16")]; |
| tensor<int32, [1]> input_99_axes_0 = const()[name = string("input_99_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_0_final_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_0_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7305728)))]; |
| tensor<fp16, [384]> self_decoder_layers_0_final_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_0_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7306560)))]; |
| tensor<fp16, [1, 1, 384]> input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = self_decoder_layers_0_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_0_final_layer_norm_weight_to_fp16, x = input_97_cast_fp16)[name = string("input_99_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_decoder_layers_0_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7307392))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7749824))))[name = string("self_decoder_layers_0_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_decoder_layers_0_fc1_bias_to_fp16 = const()[name = string("self_decoder_layers_0_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7756032)))]; |
| tensor<fp16, [1, 1, 1536]> linear_29_cast_fp16 = linear(bias = self_decoder_layers_0_fc1_bias_to_fp16, weight = self_decoder_layers_0_fc1_weight_to_fp16_palettized, x = input_99_cast_fp16)[name = string("linear_29_cast_fp16")]; |
| string input_103_mode_0 = const()[name = string("input_103_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1, 1536]> input_103_cast_fp16 = gelu(mode = input_103_mode_0, x = linear_29_cast_fp16)[name = string("input_103_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_decoder_layers_0_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(7759168))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8201600))))[name = string("self_decoder_layers_0_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_0_fc2_bias_to_fp16 = const()[name = string("self_decoder_layers_0_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8203200)))]; |
| tensor<fp16, [1, 1, 384]> linear_30_cast_fp16 = linear(bias = self_decoder_layers_0_fc2_bias_to_fp16, weight = self_decoder_layers_0_fc2_weight_to_fp16_palettized, x = input_103_cast_fp16)[name = string("linear_30_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_109_cast_fp16 = add(x = input_97_cast_fp16, y = linear_30_cast_fp16)[name = string("input_109_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_35_axes_0 = const()[name = string("hidden_states_35_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_1_self_attn_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_1_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8204032)))]; |
| tensor<fp16, [384]> self_decoder_layers_1_self_attn_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_1_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8204864)))]; |
| tensor<fp16, [1, 1, 384]> hidden_states_35_cast_fp16 = layer_norm(axes = hidden_states_35_axes_0, beta = self_decoder_layers_1_self_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_1_self_attn_layer_norm_weight_to_fp16, x = input_109_cast_fp16)[name = string("hidden_states_35_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8205696))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8316352))))[name = string("self_decoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_1_self_attn_q_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_1_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8317952)))]; |
| tensor<fp16, [1, 1, 384]> linear_31_cast_fp16 = linear(bias = self_decoder_layers_1_self_attn_q_proj_bias_to_fp16, weight = self_decoder_layers_1_self_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_35_cast_fp16)[name = string("linear_31_cast_fp16")]; |
| fp16 var_370_to_fp16 = const()[name = string("op_370_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1, 384]> var_371_cast_fp16 = mul(x = linear_31_cast_fp16, y = var_370_to_fp16)[name = string("op_371_cast_fp16")]; |
| tensor<int32, [4]> var_372 = const()[name = string("op_372"), val = tensor<int32, [4]>([1, 1, -1, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_373_cast_fp16 = reshape(shape = var_372, x = var_371_cast_fp16)[name = string("op_373_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8318784))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8429440))))[name = string("self_decoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1, 384]> linear_32_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_decoder_layers_1_self_attn_k_proj_weight_to_fp16_palettized, x = hidden_states_35_cast_fp16)[name = string("linear_32_cast_fp16")]; |
| tensor<int32, [4]> var_377 = const()[name = string("op_377"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_378_cast_fp16 = reshape(shape = var_377, x = linear_32_cast_fp16)[name = string("op_378_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8431040))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8541696))))[name = string("self_decoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_1_self_attn_v_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_1_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8543296)))]; |
| tensor<fp16, [1, 1, 384]> linear_33_cast_fp16 = linear(bias = self_decoder_layers_1_self_attn_v_proj_bias_to_fp16, weight = self_decoder_layers_1_self_attn_v_proj_weight_to_fp16_palettized, x = hidden_states_35_cast_fp16)[name = string("linear_33_cast_fp16")]; |
| tensor<fp16, [1, 1, 6, 64]> var_382_cast_fp16 = reshape(shape = var_377, x = linear_33_cast_fp16)[name = string("op_382_cast_fp16")]; |
| tensor<int32, [4]> var_383_perm_0 = const()[name = string("op_383_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool key_13_interleave_0 = const()[name = string("key_13_interleave_0"), val = bool(false)]; |
| int32 const_22 = const()[name = string("const_22"), val = int32(1)]; |
| tensor<fp16, [1, 1, 6, 64]> key_13_cast_fp16 = concat(axis = const_22, interleave = key_13_interleave_0, values = var_378_cast_fp16)[name = string("key_13_cast_fp16")]; |
| bool value_13_interleave_0 = const()[name = string("value_13_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> var_383_cast_fp16 = transpose(perm = var_383_perm_0, x = var_382_cast_fp16)[name = string("transpose_83")]; |
| tensor<fp16, [1, 6, 1, 64]> value_13_cast_fp16 = concat(axis = var_10, interleave = value_13_interleave_0, values = var_383_cast_fp16)[name = string("value_13_cast_fp16")]; |
| bool matmul_6_transpose_y_0 = const()[name = string("matmul_6_transpose_y_0"), val = bool(true)]; |
| bool matmul_6_transpose_x_0 = const()[name = string("matmul_6_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_48_perm_0 = const()[name = string("transpose_48_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_49_perm_0 = const()[name = string("transpose_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_49 = transpose(perm = transpose_49_perm_0, x = key_13_cast_fp16)[name = string("transpose_81")]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_48 = transpose(perm = transpose_48_perm_0, x = var_373_cast_fp16)[name = string("transpose_82")]; |
| tensor<fp16, [1, 6, 1, 1]> matmul_6_cast_fp16 = matmul(transpose_x = matmul_6_transpose_x_0, transpose_y = matmul_6_transpose_y_0, x = transpose_48, y = transpose_49)[name = string("matmul_6_cast_fp16")]; |
| int32 softmax_6_axis_0 = const()[name = string("softmax_6_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1, 1]> softmax_6_cast_fp16 = softmax(axis = softmax_6_axis_0, x = matmul_6_cast_fp16)[name = string("softmax_6_cast_fp16")]; |
| bool attn_output_25_transpose_x_0 = const()[name = string("attn_output_25_transpose_x_0"), val = bool(false)]; |
| bool attn_output_25_transpose_y_0 = const()[name = string("attn_output_25_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = softmax_6_cast_fp16, y = value_13_cast_fp16)[name = string("attn_output_25_cast_fp16")]; |
| tensor<int32, [4]> var_390_perm_0 = const()[name = string("op_390_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_392 = const()[name = string("op_392"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp16, [1, 1, 6, 64]> var_390_cast_fp16 = transpose(perm = var_390_perm_0, x = attn_output_25_cast_fp16)[name = string("transpose_80")]; |
| tensor<fp16, [1, 1, 384]> var_393_cast_fp16 = reshape(shape = var_392, x = var_390_cast_fp16)[name = string("op_393_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_self_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8544128))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8654784))))[name = string("self_decoder_layers_1_self_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_1_self_attn_out_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_1_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8656384)))]; |
| tensor<fp16, [1, 1, 384]> linear_34_cast_fp16 = linear(bias = self_decoder_layers_1_self_attn_out_proj_bias_to_fp16, weight = self_decoder_layers_1_self_attn_out_proj_weight_to_fp16_palettized, x = var_393_cast_fp16)[name = string("linear_34_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_34_cast_fp16)[name = string("input_115_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_39_axes_0 = const()[name = string("hidden_states_39_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_1_encoder_attn_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_1_encoder_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8657216)))]; |
| tensor<fp16, [384]> self_decoder_layers_1_encoder_attn_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_1_encoder_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8658048)))]; |
| tensor<fp16, [1, 1, 384]> hidden_states_39_cast_fp16 = layer_norm(axes = hidden_states_39_axes_0, beta = self_decoder_layers_1_encoder_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_1_encoder_attn_layer_norm_weight_to_fp16, x = input_115_cast_fp16)[name = string("hidden_states_39_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8658880))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8769536))))[name = string("self_decoder_layers_1_encoder_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_1_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_1_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8771136)))]; |
| tensor<fp16, [1, 1, 384]> linear_35_cast_fp16 = linear(bias = self_decoder_layers_1_encoder_attn_q_proj_bias_to_fp16, weight = self_decoder_layers_1_encoder_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_39_cast_fp16)[name = string("linear_35_cast_fp16")]; |
| fp16 var_401_to_fp16 = const()[name = string("op_401_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1, 384]> var_402_cast_fp16 = mul(x = linear_35_cast_fp16, y = var_401_to_fp16)[name = string("op_402_cast_fp16")]; |
| tensor<int32, [4]> var_403 = const()[name = string("op_403"), val = tensor<int32, [4]>([1, 1, -1, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_404_cast_fp16 = reshape(shape = var_403, x = var_402_cast_fp16)[name = string("op_404_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8771968))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8882624))))[name = string("self_decoder_layers_1_encoder_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1500, 384]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_decoder_layers_1_encoder_attn_k_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_36_cast_fp16")]; |
| tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_409_cast_fp16 = reshape(shape = var_408, x = linear_36_cast_fp16)[name = string("op_409_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8884224))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8994880))))[name = string("self_decoder_layers_1_encoder_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_1_encoder_attn_v_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_1_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8996480)))]; |
| tensor<fp16, [1, 1500, 384]> linear_37_cast_fp16 = linear(bias = self_decoder_layers_1_encoder_attn_v_proj_bias_to_fp16, weight = self_decoder_layers_1_encoder_attn_v_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_37_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_413_cast_fp16 = reshape(shape = var_408, x = linear_37_cast_fp16)[name = string("op_413_cast_fp16")]; |
| tensor<int32, [4]> var_414_perm_0 = const()[name = string("op_414_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool key_15_interleave_0 = const()[name = string("key_15_interleave_0"), val = bool(false)]; |
| int32 const_23 = const()[name = string("const_23"), val = int32(1)]; |
| tensor<fp16, [1, 1500, 6, 64]> key_15_cast_fp16 = concat(axis = const_23, interleave = key_15_interleave_0, values = var_409_cast_fp16)[name = string("key_15_cast_fp16")]; |
| bool value_15_interleave_0 = const()[name = string("value_15_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_414_cast_fp16 = transpose(perm = var_414_perm_0, x = var_413_cast_fp16)[name = string("transpose_79")]; |
| tensor<fp16, [1, 6, 1500, 64]> value_15_cast_fp16 = concat(axis = var_10, interleave = value_15_interleave_0, values = var_414_cast_fp16)[name = string("value_15_cast_fp16")]; |
| bool matmul_7_transpose_y_0 = const()[name = string("matmul_7_transpose_y_0"), val = bool(true)]; |
| bool matmul_7_transpose_x_0 = const()[name = string("matmul_7_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_50_perm_0 = const()[name = string("transpose_50_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_51_perm_0 = const()[name = string("transpose_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_51 = transpose(perm = transpose_51_perm_0, x = key_15_cast_fp16)[name = string("transpose_77")]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_50 = transpose(perm = transpose_50_perm_0, x = var_404_cast_fp16)[name = string("transpose_78")]; |
| tensor<fp16, [1, 6, 1, 1500]> matmul_7_cast_fp16 = matmul(transpose_x = matmul_7_transpose_x_0, transpose_y = matmul_7_transpose_y_0, x = transpose_50, y = transpose_51)[name = string("matmul_7_cast_fp16")]; |
| int32 softmax_7_axis_0 = const()[name = string("softmax_7_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1, 1500]> softmax_7_cast_fp16 = softmax(axis = softmax_7_axis_0, x = matmul_7_cast_fp16)[name = string("softmax_7_cast_fp16")]; |
| bool attn_output_29_transpose_x_0 = const()[name = string("attn_output_29_transpose_x_0"), val = bool(false)]; |
| bool attn_output_29_transpose_y_0 = const()[name = string("attn_output_29_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = softmax_7_cast_fp16, y = value_15_cast_fp16)[name = string("attn_output_29_cast_fp16")]; |
| tensor<int32, [4]> var_421_perm_0 = const()[name = string("op_421_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_423 = const()[name = string("op_423"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp16, [1, 1, 6, 64]> var_421_cast_fp16 = transpose(perm = var_421_perm_0, x = attn_output_29_cast_fp16)[name = string("transpose_76")]; |
| tensor<fp16, [1, 1, 384]> var_424_cast_fp16 = reshape(shape = var_423, x = var_421_cast_fp16)[name = string("op_424_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_1_encoder_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8997312))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9107968))))[name = string("self_decoder_layers_1_encoder_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_1_encoder_attn_out_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_1_encoder_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9109568)))]; |
| tensor<fp16, [1, 1, 384]> linear_38_cast_fp16 = linear(bias = self_decoder_layers_1_encoder_attn_out_proj_bias_to_fp16, weight = self_decoder_layers_1_encoder_attn_out_proj_weight_to_fp16_palettized, x = var_424_cast_fp16)[name = string("linear_38_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_38_cast_fp16)[name = string("input_121_cast_fp16")]; |
| tensor<int32, [1]> input_123_axes_0 = const()[name = string("input_123_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_1_final_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_1_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9110400)))]; |
| tensor<fp16, [384]> self_decoder_layers_1_final_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_1_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9111232)))]; |
| tensor<fp16, [1, 1, 384]> input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = self_decoder_layers_1_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_1_final_layer_norm_weight_to_fp16, x = input_121_cast_fp16)[name = string("input_123_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_decoder_layers_1_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9112064))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9554496))))[name = string("self_decoder_layers_1_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_decoder_layers_1_fc1_bias_to_fp16 = const()[name = string("self_decoder_layers_1_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9560704)))]; |
| tensor<fp16, [1, 1, 1536]> linear_39_cast_fp16 = linear(bias = self_decoder_layers_1_fc1_bias_to_fp16, weight = self_decoder_layers_1_fc1_weight_to_fp16_palettized, x = input_123_cast_fp16)[name = string("linear_39_cast_fp16")]; |
| string input_127_mode_0 = const()[name = string("input_127_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1, 1536]> input_127_cast_fp16 = gelu(mode = input_127_mode_0, x = linear_39_cast_fp16)[name = string("input_127_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_decoder_layers_1_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(9563840))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10006272))))[name = string("self_decoder_layers_1_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_1_fc2_bias_to_fp16 = const()[name = string("self_decoder_layers_1_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10007872)))]; |
| tensor<fp16, [1, 1, 384]> linear_40_cast_fp16 = linear(bias = self_decoder_layers_1_fc2_bias_to_fp16, weight = self_decoder_layers_1_fc2_weight_to_fp16_palettized, x = input_127_cast_fp16)[name = string("linear_40_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_133_cast_fp16 = add(x = input_121_cast_fp16, y = linear_40_cast_fp16)[name = string("input_133_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_45_axes_0 = const()[name = string("hidden_states_45_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_2_self_attn_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_2_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10008704)))]; |
| tensor<fp16, [384]> self_decoder_layers_2_self_attn_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_2_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10009536)))]; |
| tensor<fp16, [1, 1, 384]> hidden_states_45_cast_fp16 = layer_norm(axes = hidden_states_45_axes_0, beta = self_decoder_layers_2_self_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_2_self_attn_layer_norm_weight_to_fp16, x = input_133_cast_fp16)[name = string("hidden_states_45_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10010368))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10121024))))[name = string("self_decoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_2_self_attn_q_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_2_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10122624)))]; |
| tensor<fp16, [1, 1, 384]> linear_41_cast_fp16 = linear(bias = self_decoder_layers_2_self_attn_q_proj_bias_to_fp16, weight = self_decoder_layers_2_self_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_45_cast_fp16)[name = string("linear_41_cast_fp16")]; |
| fp16 var_437_to_fp16 = const()[name = string("op_437_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1, 384]> var_438_cast_fp16 = mul(x = linear_41_cast_fp16, y = var_437_to_fp16)[name = string("op_438_cast_fp16")]; |
| tensor<int32, [4]> var_439 = const()[name = string("op_439"), val = tensor<int32, [4]>([1, 1, -1, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_440_cast_fp16 = reshape(shape = var_439, x = var_438_cast_fp16)[name = string("op_440_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10123456))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10234112))))[name = string("self_decoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1, 384]> linear_42_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_decoder_layers_2_self_attn_k_proj_weight_to_fp16_palettized, x = hidden_states_45_cast_fp16)[name = string("linear_42_cast_fp16")]; |
| tensor<int32, [4]> var_444 = const()[name = string("op_444"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_445_cast_fp16 = reshape(shape = var_444, x = linear_42_cast_fp16)[name = string("op_445_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10235712))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10346368))))[name = string("self_decoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_2_self_attn_v_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_2_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10347968)))]; |
| tensor<fp16, [1, 1, 384]> linear_43_cast_fp16 = linear(bias = self_decoder_layers_2_self_attn_v_proj_bias_to_fp16, weight = self_decoder_layers_2_self_attn_v_proj_weight_to_fp16_palettized, x = hidden_states_45_cast_fp16)[name = string("linear_43_cast_fp16")]; |
| tensor<fp16, [1, 1, 6, 64]> var_449_cast_fp16 = reshape(shape = var_444, x = linear_43_cast_fp16)[name = string("op_449_cast_fp16")]; |
| tensor<int32, [4]> var_450_perm_0 = const()[name = string("op_450_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool key_17_interleave_0 = const()[name = string("key_17_interleave_0"), val = bool(false)]; |
| int32 const_24 = const()[name = string("const_24"), val = int32(1)]; |
| tensor<fp16, [1, 1, 6, 64]> key_17_cast_fp16 = concat(axis = const_24, interleave = key_17_interleave_0, values = var_445_cast_fp16)[name = string("key_17_cast_fp16")]; |
| bool value_17_interleave_0 = const()[name = string("value_17_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> var_450_cast_fp16 = transpose(perm = var_450_perm_0, x = var_449_cast_fp16)[name = string("transpose_75")]; |
| tensor<fp16, [1, 6, 1, 64]> value_17_cast_fp16 = concat(axis = var_10, interleave = value_17_interleave_0, values = var_450_cast_fp16)[name = string("value_17_cast_fp16")]; |
| bool matmul_8_transpose_y_0 = const()[name = string("matmul_8_transpose_y_0"), val = bool(true)]; |
| bool matmul_8_transpose_x_0 = const()[name = string("matmul_8_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_52_perm_0 = const()[name = string("transpose_52_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_53_perm_0 = const()[name = string("transpose_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_53 = transpose(perm = transpose_53_perm_0, x = key_17_cast_fp16)[name = string("transpose_73")]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_52 = transpose(perm = transpose_52_perm_0, x = var_440_cast_fp16)[name = string("transpose_74")]; |
| tensor<fp16, [1, 6, 1, 1]> matmul_8_cast_fp16 = matmul(transpose_x = matmul_8_transpose_x_0, transpose_y = matmul_8_transpose_y_0, x = transpose_52, y = transpose_53)[name = string("matmul_8_cast_fp16")]; |
| int32 softmax_8_axis_0 = const()[name = string("softmax_8_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1, 1]> softmax_8_cast_fp16 = softmax(axis = softmax_8_axis_0, x = matmul_8_cast_fp16)[name = string("softmax_8_cast_fp16")]; |
| bool attn_output_33_transpose_x_0 = const()[name = string("attn_output_33_transpose_x_0"), val = bool(false)]; |
| bool attn_output_33_transpose_y_0 = const()[name = string("attn_output_33_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = softmax_8_cast_fp16, y = value_17_cast_fp16)[name = string("attn_output_33_cast_fp16")]; |
| tensor<int32, [4]> var_457_perm_0 = const()[name = string("op_457_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_459 = const()[name = string("op_459"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp16, [1, 1, 6, 64]> var_457_cast_fp16 = transpose(perm = var_457_perm_0, x = attn_output_33_cast_fp16)[name = string("transpose_72")]; |
| tensor<fp16, [1, 1, 384]> var_460_cast_fp16 = reshape(shape = var_459, x = var_457_cast_fp16)[name = string("op_460_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_self_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10348800))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10459456))))[name = string("self_decoder_layers_2_self_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_2_self_attn_out_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_2_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10461056)))]; |
| tensor<fp16, [1, 1, 384]> linear_44_cast_fp16 = linear(bias = self_decoder_layers_2_self_attn_out_proj_bias_to_fp16, weight = self_decoder_layers_2_self_attn_out_proj_weight_to_fp16_palettized, x = var_460_cast_fp16)[name = string("linear_44_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_44_cast_fp16)[name = string("input_139_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_49_axes_0 = const()[name = string("hidden_states_49_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_2_encoder_attn_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_2_encoder_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10461888)))]; |
| tensor<fp16, [384]> self_decoder_layers_2_encoder_attn_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_2_encoder_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10462720)))]; |
| tensor<fp16, [1, 1, 384]> hidden_states_49_cast_fp16 = layer_norm(axes = hidden_states_49_axes_0, beta = self_decoder_layers_2_encoder_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_2_encoder_attn_layer_norm_weight_to_fp16, x = input_139_cast_fp16)[name = string("hidden_states_49_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10463552))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10574208))))[name = string("self_decoder_layers_2_encoder_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_2_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_2_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10575808)))]; |
| tensor<fp16, [1, 1, 384]> linear_45_cast_fp16 = linear(bias = self_decoder_layers_2_encoder_attn_q_proj_bias_to_fp16, weight = self_decoder_layers_2_encoder_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_49_cast_fp16)[name = string("linear_45_cast_fp16")]; |
| fp16 var_468_to_fp16 = const()[name = string("op_468_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1, 384]> var_469_cast_fp16 = mul(x = linear_45_cast_fp16, y = var_468_to_fp16)[name = string("op_469_cast_fp16")]; |
| tensor<int32, [4]> var_470 = const()[name = string("op_470"), val = tensor<int32, [4]>([1, 1, -1, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_471_cast_fp16 = reshape(shape = var_470, x = var_469_cast_fp16)[name = string("op_471_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10576640))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10687296))))[name = string("self_decoder_layers_2_encoder_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1500, 384]> linear_46_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_decoder_layers_2_encoder_attn_k_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_46_cast_fp16")]; |
| tensor<int32, [4]> var_475 = const()[name = string("op_475"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_476_cast_fp16 = reshape(shape = var_475, x = linear_46_cast_fp16)[name = string("op_476_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10688896))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10799552))))[name = string("self_decoder_layers_2_encoder_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_2_encoder_attn_v_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_2_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10801152)))]; |
| tensor<fp16, [1, 1500, 384]> linear_47_cast_fp16 = linear(bias = self_decoder_layers_2_encoder_attn_v_proj_bias_to_fp16, weight = self_decoder_layers_2_encoder_attn_v_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_47_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_480_cast_fp16 = reshape(shape = var_475, x = linear_47_cast_fp16)[name = string("op_480_cast_fp16")]; |
| tensor<int32, [4]> var_481_perm_0 = const()[name = string("op_481_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool key_19_interleave_0 = const()[name = string("key_19_interleave_0"), val = bool(false)]; |
| int32 const_25 = const()[name = string("const_25"), val = int32(1)]; |
| tensor<fp16, [1, 1500, 6, 64]> key_19_cast_fp16 = concat(axis = const_25, interleave = key_19_interleave_0, values = var_476_cast_fp16)[name = string("key_19_cast_fp16")]; |
| bool value_19_interleave_0 = const()[name = string("value_19_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_481_cast_fp16 = transpose(perm = var_481_perm_0, x = var_480_cast_fp16)[name = string("transpose_71")]; |
| tensor<fp16, [1, 6, 1500, 64]> value_19_cast_fp16 = concat(axis = var_10, interleave = value_19_interleave_0, values = var_481_cast_fp16)[name = string("value_19_cast_fp16")]; |
| bool matmul_9_transpose_y_0 = const()[name = string("matmul_9_transpose_y_0"), val = bool(true)]; |
| bool matmul_9_transpose_x_0 = const()[name = string("matmul_9_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_54_perm_0 = const()[name = string("transpose_54_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_55_perm_0 = const()[name = string("transpose_55_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_55 = transpose(perm = transpose_55_perm_0, x = key_19_cast_fp16)[name = string("transpose_69")]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_54 = transpose(perm = transpose_54_perm_0, x = var_471_cast_fp16)[name = string("transpose_70")]; |
| tensor<fp16, [1, 6, 1, 1500]> matmul_9_cast_fp16 = matmul(transpose_x = matmul_9_transpose_x_0, transpose_y = matmul_9_transpose_y_0, x = transpose_54, y = transpose_55)[name = string("matmul_9_cast_fp16")]; |
| int32 softmax_9_axis_0 = const()[name = string("softmax_9_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1, 1500]> softmax_9_cast_fp16 = softmax(axis = softmax_9_axis_0, x = matmul_9_cast_fp16)[name = string("softmax_9_cast_fp16")]; |
| bool attn_output_37_transpose_x_0 = const()[name = string("attn_output_37_transpose_x_0"), val = bool(false)]; |
| bool attn_output_37_transpose_y_0 = const()[name = string("attn_output_37_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = softmax_9_cast_fp16, y = value_19_cast_fp16)[name = string("attn_output_37_cast_fp16")]; |
| tensor<int32, [4]> var_488_perm_0 = const()[name = string("op_488_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_490 = const()[name = string("op_490"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp16, [1, 1, 6, 64]> var_488_cast_fp16 = transpose(perm = var_488_perm_0, x = attn_output_37_cast_fp16)[name = string("transpose_68")]; |
| tensor<fp16, [1, 1, 384]> var_491_cast_fp16 = reshape(shape = var_490, x = var_488_cast_fp16)[name = string("op_491_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_2_encoder_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10801984))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10912640))))[name = string("self_decoder_layers_2_encoder_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_2_encoder_attn_out_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_2_encoder_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10914240)))]; |
| tensor<fp16, [1, 1, 384]> linear_48_cast_fp16 = linear(bias = self_decoder_layers_2_encoder_attn_out_proj_bias_to_fp16, weight = self_decoder_layers_2_encoder_attn_out_proj_weight_to_fp16_palettized, x = var_491_cast_fp16)[name = string("linear_48_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_48_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, [384]> self_decoder_layers_2_final_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_2_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10915072)))]; |
| tensor<fp16, [384]> self_decoder_layers_2_final_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_2_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10915904)))]; |
| tensor<fp16, [1, 1, 384]> input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = self_decoder_layers_2_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_2_final_layer_norm_weight_to_fp16, x = input_145_cast_fp16)[name = string("input_147_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_decoder_layers_2_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(10916736))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11359168))))[name = string("self_decoder_layers_2_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_decoder_layers_2_fc1_bias_to_fp16 = const()[name = string("self_decoder_layers_2_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11365376)))]; |
| tensor<fp16, [1, 1, 1536]> linear_49_cast_fp16 = linear(bias = self_decoder_layers_2_fc1_bias_to_fp16, weight = self_decoder_layers_2_fc1_weight_to_fp16_palettized, x = input_147_cast_fp16)[name = string("linear_49_cast_fp16")]; |
| string input_151_mode_0 = const()[name = string("input_151_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1, 1536]> input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_49_cast_fp16)[name = string("input_151_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_decoder_layers_2_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11368512))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11810944))))[name = string("self_decoder_layers_2_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_2_fc2_bias_to_fp16 = const()[name = string("self_decoder_layers_2_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11812544)))]; |
| tensor<fp16, [1, 1, 384]> linear_50_cast_fp16 = linear(bias = self_decoder_layers_2_fc2_bias_to_fp16, weight = self_decoder_layers_2_fc2_weight_to_fp16_palettized, x = input_151_cast_fp16)[name = string("linear_50_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_157_cast_fp16 = add(x = input_145_cast_fp16, y = linear_50_cast_fp16)[name = string("input_157_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_55_axes_0 = const()[name = string("hidden_states_55_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_3_self_attn_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_3_self_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11813376)))]; |
| tensor<fp16, [384]> self_decoder_layers_3_self_attn_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_3_self_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11814208)))]; |
| tensor<fp16, [1, 1, 384]> hidden_states_55_cast_fp16 = layer_norm(axes = hidden_states_55_axes_0, beta = self_decoder_layers_3_self_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_3_self_attn_layer_norm_weight_to_fp16, x = input_157_cast_fp16)[name = string("hidden_states_55_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11815040))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11925696))))[name = string("self_decoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_3_self_attn_q_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_3_self_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11927296)))]; |
| tensor<fp16, [1, 1, 384]> linear_51_cast_fp16 = linear(bias = self_decoder_layers_3_self_attn_q_proj_bias_to_fp16, weight = self_decoder_layers_3_self_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_55_cast_fp16)[name = string("linear_51_cast_fp16")]; |
| fp16 var_504_to_fp16 = const()[name = string("op_504_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1, 384]> var_505_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_504_to_fp16)[name = string("op_505_cast_fp16")]; |
| tensor<int32, [4]> var_506 = const()[name = string("op_506"), val = tensor<int32, [4]>([1, 1, -1, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_507_cast_fp16 = reshape(shape = var_506, x = var_505_cast_fp16)[name = string("op_507_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(11928128))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12038784))))[name = string("self_decoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1, 384]> linear_52_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_decoder_layers_3_self_attn_k_proj_weight_to_fp16_palettized, x = hidden_states_55_cast_fp16)[name = string("linear_52_cast_fp16")]; |
| tensor<int32, [4]> var_511 = const()[name = string("op_511"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_512_cast_fp16 = reshape(shape = var_511, x = linear_52_cast_fp16)[name = string("op_512_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12040384))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12151040))))[name = string("self_decoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_3_self_attn_v_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_3_self_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12152640)))]; |
| tensor<fp16, [1, 1, 384]> linear_53_cast_fp16 = linear(bias = self_decoder_layers_3_self_attn_v_proj_bias_to_fp16, weight = self_decoder_layers_3_self_attn_v_proj_weight_to_fp16_palettized, x = hidden_states_55_cast_fp16)[name = string("linear_53_cast_fp16")]; |
| tensor<fp16, [1, 1, 6, 64]> var_516_cast_fp16 = reshape(shape = var_511, x = linear_53_cast_fp16)[name = string("op_516_cast_fp16")]; |
| tensor<int32, [4]> var_517_perm_0 = const()[name = string("op_517_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool key_21_interleave_0 = const()[name = string("key_21_interleave_0"), val = bool(false)]; |
| int32 const_26 = const()[name = string("const_26"), val = int32(1)]; |
| tensor<fp16, [1, 1, 6, 64]> key_21_cast_fp16 = concat(axis = const_26, interleave = key_21_interleave_0, values = var_512_cast_fp16)[name = string("key_21_cast_fp16")]; |
| bool value_21_interleave_0 = const()[name = string("value_21_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> var_517_cast_fp16 = transpose(perm = var_517_perm_0, x = var_516_cast_fp16)[name = string("transpose_67")]; |
| tensor<fp16, [1, 6, 1, 64]> value_21_cast_fp16 = concat(axis = var_10, interleave = value_21_interleave_0, values = var_517_cast_fp16)[name = string("value_21_cast_fp16")]; |
| bool matmul_10_transpose_y_0 = const()[name = string("matmul_10_transpose_y_0"), val = bool(true)]; |
| bool matmul_10_transpose_x_0 = const()[name = string("matmul_10_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_56_perm_0 = const()[name = string("transpose_56_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_57_perm_0 = const()[name = string("transpose_57_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_57 = transpose(perm = transpose_57_perm_0, x = key_21_cast_fp16)[name = string("transpose_65")]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_56 = transpose(perm = transpose_56_perm_0, x = var_507_cast_fp16)[name = string("transpose_66")]; |
| tensor<fp16, [1, 6, 1, 1]> matmul_10_cast_fp16 = matmul(transpose_x = matmul_10_transpose_x_0, transpose_y = matmul_10_transpose_y_0, x = transpose_56, y = transpose_57)[name = string("matmul_10_cast_fp16")]; |
| int32 softmax_10_axis_0 = const()[name = string("softmax_10_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1, 1]> softmax_10_cast_fp16 = softmax(axis = softmax_10_axis_0, x = matmul_10_cast_fp16)[name = string("softmax_10_cast_fp16")]; |
| bool attn_output_41_transpose_x_0 = const()[name = string("attn_output_41_transpose_x_0"), val = bool(false)]; |
| bool attn_output_41_transpose_y_0 = const()[name = string("attn_output_41_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = softmax_10_cast_fp16, y = value_21_cast_fp16)[name = string("attn_output_41_cast_fp16")]; |
| tensor<int32, [4]> var_524_perm_0 = const()[name = string("op_524_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_526 = const()[name = string("op_526"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp16, [1, 1, 6, 64]> var_524_cast_fp16 = transpose(perm = var_524_perm_0, x = attn_output_41_cast_fp16)[name = string("transpose_64")]; |
| tensor<fp16, [1, 1, 384]> var_527_cast_fp16 = reshape(shape = var_526, x = var_524_cast_fp16)[name = string("op_527_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_self_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12153472))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12264128))))[name = string("self_decoder_layers_3_self_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_3_self_attn_out_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_3_self_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12265728)))]; |
| tensor<fp16, [1, 1, 384]> linear_54_cast_fp16 = linear(bias = self_decoder_layers_3_self_attn_out_proj_bias_to_fp16, weight = self_decoder_layers_3_self_attn_out_proj_weight_to_fp16_palettized, x = var_527_cast_fp16)[name = string("linear_54_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_54_cast_fp16)[name = string("input_163_cast_fp16")]; |
| tensor<int32, [1]> hidden_states_59_axes_0 = const()[name = string("hidden_states_59_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_3_encoder_attn_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_3_encoder_attn_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12266560)))]; |
| tensor<fp16, [384]> self_decoder_layers_3_encoder_attn_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_3_encoder_attn_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12267392)))]; |
| tensor<fp16, [1, 1, 384]> hidden_states_59_cast_fp16 = layer_norm(axes = hidden_states_59_axes_0, beta = self_decoder_layers_3_encoder_attn_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_3_encoder_attn_layer_norm_weight_to_fp16, x = input_163_cast_fp16)[name = string("hidden_states_59_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_encoder_attn_q_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12268224))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12378880))))[name = string("self_decoder_layers_3_encoder_attn_q_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_3_encoder_attn_q_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_3_encoder_attn_q_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12380480)))]; |
| tensor<fp16, [1, 1, 384]> linear_55_cast_fp16 = linear(bias = self_decoder_layers_3_encoder_attn_q_proj_bias_to_fp16, weight = self_decoder_layers_3_encoder_attn_q_proj_weight_to_fp16_palettized, x = hidden_states_59_cast_fp16)[name = string("linear_55_cast_fp16")]; |
| fp16 var_535_to_fp16 = const()[name = string("op_535_to_fp16"), val = fp16(0x1p-3)]; |
| tensor<fp16, [1, 1, 384]> var_536_cast_fp16 = mul(x = linear_55_cast_fp16, y = var_535_to_fp16)[name = string("op_536_cast_fp16")]; |
| tensor<int32, [4]> var_537 = const()[name = string("op_537"), val = tensor<int32, [4]>([1, 1, -1, 64])]; |
| tensor<fp16, [1, 1, 6, 64]> var_538_cast_fp16 = reshape(shape = var_537, x = var_536_cast_fp16)[name = string("op_538_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_encoder_attn_k_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12381312))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12491968))))[name = string("self_decoder_layers_3_encoder_attn_k_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [1, 1500, 384]> linear_56_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = self_decoder_layers_3_encoder_attn_k_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_56_cast_fp16")]; |
| tensor<int32, [4]> var_542 = const()[name = string("op_542"), val = tensor<int32, [4]>([1, -1, 6, 64])]; |
| tensor<fp16, [1, 1500, 6, 64]> var_543_cast_fp16 = reshape(shape = var_542, x = linear_56_cast_fp16)[name = string("op_543_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_encoder_attn_v_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12493568))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12604224))))[name = string("self_decoder_layers_3_encoder_attn_v_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_3_encoder_attn_v_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_3_encoder_attn_v_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12605824)))]; |
| tensor<fp16, [1, 1500, 384]> linear_57_cast_fp16 = linear(bias = self_decoder_layers_3_encoder_attn_v_proj_bias_to_fp16, weight = self_decoder_layers_3_encoder_attn_v_proj_weight_to_fp16_palettized, x = input_91_cast_fp16)[name = string("linear_57_cast_fp16")]; |
| tensor<fp16, [1, 1500, 6, 64]> var_547_cast_fp16 = reshape(shape = var_542, x = linear_57_cast_fp16)[name = string("op_547_cast_fp16")]; |
| tensor<int32, [4]> var_548_perm_0 = const()[name = string("op_548_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| bool key_interleave_0 = const()[name = string("key_interleave_0"), val = bool(false)]; |
| int32 const_27 = const()[name = string("const_27"), val = int32(1)]; |
| tensor<fp16, [1, 1500, 6, 64]> key_cast_fp16 = concat(axis = const_27, interleave = key_interleave_0, values = var_543_cast_fp16)[name = string("key_cast_fp16")]; |
| bool value_interleave_0 = const()[name = string("value_interleave_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1500, 64]> var_548_cast_fp16 = transpose(perm = var_548_perm_0, x = var_547_cast_fp16)[name = string("transpose_63")]; |
| tensor<fp16, [1, 6, 1500, 64]> value_cast_fp16 = concat(axis = var_10, interleave = value_interleave_0, values = var_548_cast_fp16)[name = string("value_cast_fp16")]; |
| bool matmul_11_transpose_y_0 = const()[name = string("matmul_11_transpose_y_0"), val = bool(true)]; |
| bool matmul_11_transpose_x_0 = const()[name = string("matmul_11_transpose_x_0"), val = bool(false)]; |
| tensor<int32, [4]> transpose_58_perm_0 = const()[name = string("transpose_58_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<int32, [4]> transpose_59_perm_0 = const()[name = string("transpose_59_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])]; |
| tensor<fp16, [1, 6, 1500, 64]> transpose_59 = transpose(perm = transpose_59_perm_0, x = key_cast_fp16)[name = string("transpose_61")]; |
| tensor<fp16, [1, 6, 1, 64]> transpose_58 = transpose(perm = transpose_58_perm_0, x = var_538_cast_fp16)[name = string("transpose_62")]; |
| tensor<fp16, [1, 6, 1, 1500]> matmul_11_cast_fp16 = matmul(transpose_x = matmul_11_transpose_x_0, transpose_y = matmul_11_transpose_y_0, x = transpose_58, y = transpose_59)[name = string("matmul_11_cast_fp16")]; |
| int32 softmax_11_axis_0 = const()[name = string("softmax_11_axis_0"), val = int32(-1)]; |
| tensor<fp16, [1, 6, 1, 1500]> softmax_11_cast_fp16 = softmax(axis = softmax_11_axis_0, x = matmul_11_cast_fp16)[name = string("softmax_11_cast_fp16")]; |
| bool attn_output_45_transpose_x_0 = const()[name = string("attn_output_45_transpose_x_0"), val = bool(false)]; |
| bool attn_output_45_transpose_y_0 = const()[name = string("attn_output_45_transpose_y_0"), val = bool(false)]; |
| tensor<fp16, [1, 6, 1, 64]> attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = softmax_11_cast_fp16, y = value_cast_fp16)[name = string("attn_output_45_cast_fp16")]; |
| tensor<int32, [4]> var_555_perm_0 = const()[name = string("op_555_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])]; |
| tensor<int32, [3]> var_557 = const()[name = string("op_557"), val = tensor<int32, [3]>([1, 1, -1])]; |
| tensor<fp16, [1, 1, 6, 64]> var_555_cast_fp16 = transpose(perm = var_555_perm_0, x = attn_output_45_cast_fp16)[name = string("transpose_60")]; |
| tensor<fp16, [1, 1, 384]> var_558_cast_fp16 = reshape(shape = var_557, x = var_555_cast_fp16)[name = string("op_558_cast_fp16")]; |
| tensor<fp16, [384, 384]> self_decoder_layers_3_encoder_attn_out_proj_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12606656))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12717312))))[name = string("self_decoder_layers_3_encoder_attn_out_proj_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_3_encoder_attn_out_proj_bias_to_fp16 = const()[name = string("self_decoder_layers_3_encoder_attn_out_proj_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12718912)))]; |
| tensor<fp16, [1, 1, 384]> linear_58_cast_fp16 = linear(bias = self_decoder_layers_3_encoder_attn_out_proj_bias_to_fp16, weight = self_decoder_layers_3_encoder_attn_out_proj_weight_to_fp16_palettized, x = var_558_cast_fp16)[name = string("linear_58_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_58_cast_fp16)[name = string("input_169_cast_fp16")]; |
| tensor<int32, [1]> input_171_axes_0 = const()[name = string("input_171_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layers_3_final_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layers_3_final_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12719744)))]; |
| tensor<fp16, [384]> self_decoder_layers_3_final_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layers_3_final_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12720576)))]; |
| tensor<fp16, [1, 1, 384]> input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = self_decoder_layers_3_final_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layers_3_final_layer_norm_weight_to_fp16, x = input_169_cast_fp16)[name = string("input_171_cast_fp16")]; |
| tensor<fp16, [1536, 384]> self_decoder_layers_3_fc1_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [1536, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12721408))), lut = tensor<fp16, [48, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13163840))))[name = string("self_decoder_layers_3_fc1_weight_to_fp16_palettized")]; |
| tensor<fp16, [1536]> self_decoder_layers_3_fc1_bias_to_fp16 = const()[name = string("self_decoder_layers_3_fc1_bias_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13170048)))]; |
| tensor<fp16, [1, 1, 1536]> linear_59_cast_fp16 = linear(bias = self_decoder_layers_3_fc1_bias_to_fp16, weight = self_decoder_layers_3_fc1_weight_to_fp16_palettized, x = input_171_cast_fp16)[name = string("linear_59_cast_fp16")]; |
| string input_175_mode_0 = const()[name = string("input_175_mode_0"), val = string("EXACT")]; |
| tensor<fp16, [1, 1, 1536]> input_175_cast_fp16 = gelu(mode = input_175_mode_0, x = linear_59_cast_fp16)[name = string("input_175_cast_fp16")]; |
| tensor<fp16, [384, 1536]> self_decoder_layers_3_fc2_weight_to_fp16_palettized = constexpr_lut_to_dense(indices = tensor<uint6, [384, 1536]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13173184))), lut = tensor<fp16, [12, 1, 64, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13615616))))[name = string("self_decoder_layers_3_fc2_weight_to_fp16_palettized")]; |
| tensor<fp16, [384]> self_decoder_layers_3_fc2_bias_to_fp16 = const()[name = string("self_decoder_layers_3_fc2_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13617216)))]; |
| tensor<fp16, [1, 1, 384]> linear_60_cast_fp16 = linear(bias = self_decoder_layers_3_fc2_bias_to_fp16, weight = self_decoder_layers_3_fc2_weight_to_fp16_palettized, x = input_175_cast_fp16)[name = string("linear_60_cast_fp16")]; |
| tensor<fp16, [1, 1, 384]> input_181_cast_fp16 = add(x = input_169_cast_fp16, y = linear_60_cast_fp16)[name = string("input_181_cast_fp16")]; |
| tensor<int32, [1]> input_axes_0 = const()[name = string("input_axes_0"), val = tensor<int32, [1]>([-1])]; |
| tensor<fp16, [384]> self_decoder_layer_norm_weight_to_fp16 = const()[name = string("self_decoder_layer_norm_weight_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13618048)))]; |
| tensor<fp16, [384]> self_decoder_layer_norm_bias_to_fp16 = const()[name = string("self_decoder_layer_norm_bias_to_fp16"), val = tensor<fp16, [384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13618880)))]; |
| tensor<fp16, [1, 1, 384]> input_cast_fp16 = layer_norm(axes = input_axes_0, beta = self_decoder_layer_norm_bias_to_fp16, epsilon = var_29_to_fp16, gamma = self_decoder_layer_norm_weight_to_fp16, x = input_181_cast_fp16)[name = string("input_cast_fp16")]; |
| tensor<fp16, [99, 384]> self_head_weight_to_fp16 = const()[name = string("self_head_weight_to_fp16"), val = tensor<fp16, [99, 384]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13619712)))]; |
| tensor<fp16, [99]> linear_61_bias_0_to_fp16 = const()[name = string("linear_61_bias_0_to_fp16"), val = tensor<fp16, [99]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13695808)))]; |
| tensor<fp16, [1, 1, 99]> linear_61_cast_fp16 = linear(bias = linear_61_bias_0_to_fp16, weight = self_head_weight_to_fp16, x = input_cast_fp16)[name = string("linear_61_cast_fp16")]; |
| tensor<int32, [3]> var_570_begin_0 = const()[name = string("op_570_begin_0"), val = tensor<int32, [3]>([0, 0, 0])]; |
| tensor<int32, [3]> var_570_end_0 = const()[name = string("op_570_end_0"), val = tensor<int32, [3]>([1, 1, 99])]; |
| tensor<bool, [3]> var_570_end_mask_0 = const()[name = string("op_570_end_mask_0"), val = tensor<bool, [3]>([true, false, true])]; |
| tensor<bool, [3]> var_570_squeeze_mask_0 = const()[name = string("op_570_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])]; |
| tensor<fp16, [1, 99]> logits = slice_by_index(begin = var_570_begin_0, end = var_570_end_0, end_mask = var_570_end_mask_0, squeeze_mask = var_570_squeeze_mask_0, x = linear_61_cast_fp16)[name = string("op_570_cast_fp16")]; |
| } -> (logits); |
| } |
