program(1.0) [buildInfo = dict, tensor>({{"coremlc-component-MIL", "3304.5.2"}, {"coremlc-version", "3304.6.2"}})] { func main(tensor cache_length, tensor decoder_key_padding_mask, tensor encoder_output_embeds, tensor input_ids, tensor key_cache, tensor kv_cache_update_mask, tensor value_cache) { tensor var_24_axis_0 = const()[name = tensor("op_24_axis_0"), val = tensor(0)]; tensor var_24_batch_dims_0 = const()[name = tensor("op_24_batch_dims_0"), val = tensor(0)]; tensor embed_tokens_weight_to_fp16 = const()[name = tensor("embed_tokens_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; tensor var_24_cast_fp16 = gather(axis = var_24_axis_0, batch_dims = var_24_batch_dims_0, indices = input_ids, x = embed_tokens_weight_to_fp16)[name = tensor("op_24_cast_fp16")]; tensor var_28_axis_0 = const()[name = tensor("op_28_axis_0"), val = tensor(0)]; tensor var_28_batch_dims_0 = const()[name = tensor("op_28_batch_dims_0"), val = tensor(0)]; tensor embed_positions_weight_to_fp16 = const()[name = tensor("embed_positions_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(132777088)))]; tensor var_28_cast_fp16 = gather(axis = var_28_axis_0, batch_dims = var_28_batch_dims_0, indices = cache_length, x = embed_positions_weight_to_fp16)[name = tensor("op_28_cast_fp16")]; tensor hidden_states_1_cast_fp16 = add(x = var_24_cast_fp16, y = var_28_cast_fp16)[name = tensor("hidden_states_1_cast_fp16")]; tensor var_42_axes_0 = const()[name = tensor("op_42_axes_0"), val = tensor([2])]; tensor var_42_cast_fp16 = expand_dims(axes = var_42_axes_0, x = hidden_states_1_cast_fp16)[name = tensor("op_42_cast_fp16")]; tensor inputs_1_axes_0 = const()[name = tensor("inputs_1_axes_0"), val = tensor([3])]; tensor inputs_1_cast_fp16 = expand_dims(axes = inputs_1_axes_0, x = var_42_cast_fp16)[name = tensor("inputs_1_cast_fp16")]; tensor tile_0 = const()[name = tensor("tile_0"), val = tensor([1280, 1280, 1280, 1280])]; tensor var_47_axis_0 = const()[name = tensor("op_47_axis_0"), val = tensor(1)]; tensor var_47_cast_fp16_0, tensor var_47_cast_fp16_1, tensor var_47_cast_fp16_2, tensor var_47_cast_fp16_3 = split(axis = var_47_axis_0, split_sizes = tile_0, x = key_cache)[name = tensor("op_47_cast_fp16")]; tensor tile_1 = const()[name = tensor("tile_1"), val = tensor([1280, 1280, 1280, 1280])]; tensor var_54_axis_0 = const()[name = tensor("op_54_axis_0"), val = tensor(1)]; tensor var_54_cast_fp16_0, tensor var_54_cast_fp16_1, tensor var_54_cast_fp16_2, tensor var_54_cast_fp16_3 = split(axis = var_54_axis_0, split_sizes = tile_1, x = value_cache)[name = tensor("op_54_cast_fp16")]; tensor var_64 = const()[name = tensor("op_64"), val = tensor(3)]; tensor out_1_axes_0 = const()[name = tensor("out_1_axes_0"), val = tensor([1])]; tensor var_90_to_fp16 = const()[name = tensor("op_90_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_1_cast_fp16 = layer_norm(axes = out_1_axes_0, epsilon = var_90_to_fp16, x = inputs_1_cast_fp16)[name = tensor("out_1_cast_fp16")]; tensor obj_1_mean_0_to_fp16 = const()[name = tensor("obj_1_mean_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133924032)))]; tensor obj_1_variance_0_to_fp16 = const()[name = tensor("obj_1_variance_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133926656)))]; tensor obj_1_gamma_0_to_fp16 = const()[name = tensor("obj_1_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133929280)))]; tensor obj_1_beta_0_to_fp16 = const()[name = tensor("obj_1_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133931904)))]; tensor obj_1_epsilon_0_to_fp16 = const()[name = tensor("obj_1_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_1_cast_fp16 = batch_norm(beta = obj_1_beta_0_to_fp16, epsilon = obj_1_epsilon_0_to_fp16, gamma = obj_1_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_1_cast_fp16)[name = tensor("obj_1_cast_fp16")]; tensor pretrained_out_1_pad_type_0 = const()[name = tensor("pretrained_out_1_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_1_strides_0 = const()[name = tensor("pretrained_out_1_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_1_pad_0 = const()[name = tensor("pretrained_out_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_1_dilations_0 = const()[name = tensor("pretrained_out_1_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_1_groups_0 = const()[name = tensor("pretrained_out_1_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133934528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134753792))), name = tensor("layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134753920)))]; tensor pretrained_out_1_cast_fp16 = conv(bias = layers_0_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_1_dilations_0, groups = pretrained_out_1_groups_0, pad = pretrained_out_1_pad_0, pad_type = pretrained_out_1_pad_type_0, strides = pretrained_out_1_strides_0, weight = layers_0_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_1_cast_fp16")]; tensor input_1_pad_type_0 = const()[name = tensor("input_1_pad_type_0"), val = tensor("valid")]; tensor input_1_strides_0 = const()[name = tensor("input_1_strides_0"), val = tensor([1, 1])]; tensor input_1_pad_0 = const()[name = tensor("input_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_1_dilations_0 = const()[name = tensor("input_1_dilations_0"), val = tensor([1, 1])]; tensor input_1_groups_0 = const()[name = tensor("input_1_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134756544)))]; tensor input_1_cast_fp16 = conv(dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = layers_0_self_attn_q_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_1_cast_fp16")]; tensor lora_out_1_pad_type_0 = const()[name = tensor("lora_out_1_pad_type_0"), val = tensor("valid")]; tensor lora_out_1_strides_0 = const()[name = tensor("lora_out_1_strides_0"), val = tensor([1, 1])]; tensor lora_out_1_pad_0 = const()[name = tensor("lora_out_1_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_1_dilations_0 = const()[name = tensor("lora_out_1_dilations_0"), val = tensor([1, 1])]; tensor lora_out_1_groups_0 = const()[name = tensor("lora_out_1_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134797568)))]; tensor lora_out_1_cast_fp16 = conv(dilations = lora_out_1_dilations_0, groups = lora_out_1_groups_0, pad = lora_out_1_pad_0, pad_type = lora_out_1_pad_type_0, strides = lora_out_1_strides_0, weight = layers_0_self_attn_q_proj_loraB_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("lora_out_1_cast_fp16")]; tensor query_1_cast_fp16 = add(x = pretrained_out_1_cast_fp16, y = lora_out_1_cast_fp16)[name = tensor("query_1_cast_fp16")]; tensor pretrained_out_3_pad_type_0 = const()[name = tensor("pretrained_out_3_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_3_strides_0 = const()[name = tensor("pretrained_out_3_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_3_pad_0 = const()[name = tensor("pretrained_out_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_3_dilations_0 = const()[name = tensor("pretrained_out_3_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_3_groups_0 = const()[name = tensor("pretrained_out_3_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134838592))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135657856))), name = tensor("layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_3_cast_fp16 = conv(dilations = pretrained_out_3_dilations_0, groups = pretrained_out_3_groups_0, pad = pretrained_out_3_pad_0, pad_type = pretrained_out_3_pad_type_0, strides = pretrained_out_3_strides_0, weight = layers_0_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_3_cast_fp16")]; tensor input_3_pad_type_0 = const()[name = tensor("input_3_pad_type_0"), val = tensor("valid")]; tensor input_3_strides_0 = const()[name = tensor("input_3_strides_0"), val = tensor([1, 1])]; tensor input_3_pad_0 = const()[name = tensor("input_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_3_dilations_0 = const()[name = tensor("input_3_dilations_0"), val = tensor([1, 1])]; tensor input_3_groups_0 = const()[name = tensor("input_3_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135657984)))]; tensor input_3_cast_fp16 = conv(dilations = input_3_dilations_0, groups = input_3_groups_0, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = input_3_strides_0, weight = layers_0_self_attn_k_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_3_cast_fp16")]; tensor lora_out_3_pad_type_0 = const()[name = tensor("lora_out_3_pad_type_0"), val = tensor("valid")]; tensor lora_out_3_strides_0 = const()[name = tensor("lora_out_3_strides_0"), val = tensor([1, 1])]; tensor lora_out_3_pad_0 = const()[name = tensor("lora_out_3_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_3_dilations_0 = const()[name = tensor("lora_out_3_dilations_0"), val = tensor([1, 1])]; tensor lora_out_3_groups_0 = const()[name = tensor("lora_out_3_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135699008)))]; tensor lora_out_3_cast_fp16 = conv(dilations = lora_out_3_dilations_0, groups = lora_out_3_groups_0, pad = lora_out_3_pad_0, pad_type = lora_out_3_pad_type_0, strides = lora_out_3_strides_0, weight = layers_0_self_attn_k_proj_loraB_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("lora_out_3_cast_fp16")]; tensor current_key_1_cast_fp16 = add(x = pretrained_out_3_cast_fp16, y = lora_out_3_cast_fp16)[name = tensor("current_key_1_cast_fp16")]; tensor pretrained_out_5_pad_type_0 = const()[name = tensor("pretrained_out_5_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_5_strides_0 = const()[name = tensor("pretrained_out_5_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_5_pad_0 = const()[name = tensor("pretrained_out_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_5_dilations_0 = const()[name = tensor("pretrained_out_5_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_5_groups_0 = const()[name = tensor("pretrained_out_5_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135740032))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136559296))), name = tensor("layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136559424)))]; tensor pretrained_out_5_cast_fp16 = conv(bias = layers_0_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_5_dilations_0, groups = pretrained_out_5_groups_0, pad = pretrained_out_5_pad_0, pad_type = pretrained_out_5_pad_type_0, strides = pretrained_out_5_strides_0, weight = layers_0_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_1_cast_fp16)[name = tensor("pretrained_out_5_cast_fp16")]; tensor input_5_pad_type_0 = const()[name = tensor("input_5_pad_type_0"), val = tensor("valid")]; tensor input_5_strides_0 = const()[name = tensor("input_5_strides_0"), val = tensor([1, 1])]; tensor input_5_pad_0 = const()[name = tensor("input_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_5_dilations_0 = const()[name = tensor("input_5_dilations_0"), val = tensor([1, 1])]; tensor input_5_groups_0 = const()[name = tensor("input_5_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136562048)))]; tensor input_5_cast_fp16 = conv(dilations = input_5_dilations_0, groups = input_5_groups_0, pad = input_5_pad_0, pad_type = input_5_pad_type_0, strides = input_5_strides_0, weight = layers_0_self_attn_v_proj_loraA_weight_to_fp16, x = obj_1_cast_fp16)[name = tensor("input_5_cast_fp16")]; tensor lora_out_5_pad_type_0 = const()[name = tensor("lora_out_5_pad_type_0"), val = tensor("valid")]; tensor lora_out_5_strides_0 = const()[name = tensor("lora_out_5_strides_0"), val = tensor([1, 1])]; tensor lora_out_5_pad_0 = const()[name = tensor("lora_out_5_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_5_dilations_0 = const()[name = tensor("lora_out_5_dilations_0"), val = tensor([1, 1])]; tensor lora_out_5_groups_0 = const()[name = tensor("lora_out_5_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136603072)))]; tensor lora_out_5_cast_fp16 = conv(dilations = lora_out_5_dilations_0, groups = lora_out_5_groups_0, pad = lora_out_5_pad_0, pad_type = lora_out_5_pad_type_0, strides = lora_out_5_strides_0, weight = layers_0_self_attn_v_proj_loraB_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("lora_out_5_cast_fp16")]; tensor current_value_1_cast_fp16 = add(x = pretrained_out_5_cast_fp16, y = lora_out_5_cast_fp16)[name = tensor("current_value_1_cast_fp16")]; tensor var_173_axes_0 = const()[name = tensor("op_173_axes_0"), val = tensor([1])]; tensor var_173_cast_fp16 = expand_dims(axes = var_173_axes_0, x = kv_cache_update_mask)[name = tensor("op_173_cast_fp16")]; tensor var_174_axes_0 = const()[name = tensor("op_174_axes_0"), val = tensor([2])]; tensor var_174_cast_fp16 = expand_dims(axes = var_174_axes_0, x = var_173_cast_fp16)[name = tensor("op_174_cast_fp16")]; tensor var_176_cast_fp16 = mul(x = current_key_1_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_176_cast_fp16")]; tensor var_65_to_fp16 = const()[name = tensor("op_65_to_fp16"), val = tensor(0x1p+0)]; tensor var_177_cast_fp16 = sub(x = var_65_to_fp16, y = var_174_cast_fp16)[name = tensor("op_177_cast_fp16")]; tensor var_178_cast_fp16 = mul(x = var_47_cast_fp16_0, y = var_177_cast_fp16)[name = tensor("op_178_cast_fp16")]; tensor key_1_cast_fp16 = add(x = var_176_cast_fp16, y = var_178_cast_fp16)[name = tensor("key_1_cast_fp16")]; tensor var_180_cast_fp16 = mul(x = current_value_1_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_180_cast_fp16")]; tensor var_182_cast_fp16 = mul(x = var_54_cast_fp16_0, y = var_177_cast_fp16)[name = tensor("op_182_cast_fp16")]; tensor value_1_cast_fp16 = add(x = var_180_cast_fp16, y = var_182_cast_fp16)[name = tensor("value_1_cast_fp16")]; tensor var_185 = const()[name = tensor("op_185"), val = tensor([1, 20, 64, -1])]; tensor mh_q_1_cast_fp16 = reshape(shape = var_185, x = query_1_cast_fp16)[name = tensor("mh_q_1_cast_fp16")]; tensor var_187_to_fp16 = const()[name = tensor("op_187_to_fp16"), val = tensor(0x1p-3)]; tensor var_188_cast_fp16 = mul(x = mh_q_1_cast_fp16, y = var_187_to_fp16)[name = tensor("op_188_cast_fp16")]; tensor var_189 = const()[name = tensor("op_189"), val = tensor([1, 20, 64, -1])]; tensor var_190_cast_fp16 = reshape(shape = var_189, x = key_1_cast_fp16)[name = tensor("op_190_cast_fp16")]; tensor mh_w_1_transpose_x_0 = const()[name = tensor("mh_w_1_transpose_x_0"), val = tensor(true)]; tensor mh_w_1_transpose_y_0 = const()[name = tensor("mh_w_1_transpose_y_0"), val = tensor(false)]; tensor mh_w_1_cast_fp16 = matmul(transpose_x = mh_w_1_transpose_x_0, transpose_y = mh_w_1_transpose_y_0, x = var_188_cast_fp16, y = var_190_cast_fp16)[name = tensor("mh_w_1_cast_fp16")]; tensor var_194_axes_0 = const()[name = tensor("op_194_axes_0"), val = tensor([1])]; tensor var_194_cast_fp16 = expand_dims(axes = var_194_axes_0, x = decoder_key_padding_mask)[name = tensor("op_194_cast_fp16")]; tensor var_195_axes_0 = const()[name = tensor("op_195_axes_0"), val = tensor([2])]; tensor var_195_cast_fp16 = expand_dims(axes = var_195_axes_0, x = var_194_cast_fp16)[name = tensor("op_195_cast_fp16")]; tensor mh_w_3_cast_fp16 = add(x = mh_w_1_cast_fp16, y = var_195_cast_fp16)[name = tensor("mh_w_3_cast_fp16")]; tensor var_198_cast_fp16 = softmax(axis = var_64, x = mh_w_3_cast_fp16)[name = tensor("op_198_cast_fp16")]; tensor var_199 = const()[name = tensor("op_199"), val = tensor([1, 20, 64, -1])]; tensor var_200_cast_fp16 = reshape(shape = var_199, x = value_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; tensor attn_1_transpose_x_0 = const()[name = tensor("attn_1_transpose_x_0"), val = tensor(false)]; tensor attn_1_transpose_y_0 = const()[name = tensor("attn_1_transpose_y_0"), val = tensor(true)]; tensor attn_1_cast_fp16 = matmul(transpose_x = attn_1_transpose_x_0, transpose_y = attn_1_transpose_y_0, x = var_200_cast_fp16, y = var_198_cast_fp16)[name = tensor("attn_1_cast_fp16")]; tensor var_203 = const()[name = tensor("op_203"), val = tensor([1, 1280, 1, -1])]; tensor input_7_cast_fp16 = reshape(shape = var_203, x = attn_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; tensor pretrained_out_7_pad_type_0 = const()[name = tensor("pretrained_out_7_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_7_strides_0 = const()[name = tensor("pretrained_out_7_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_7_pad_0 = const()[name = tensor("pretrained_out_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_7_dilations_0 = const()[name = tensor("pretrained_out_7_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_7_groups_0 = const()[name = tensor("pretrained_out_7_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(136644096))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137463360))), name = tensor("layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137463488)))]; tensor pretrained_out_7_cast_fp16 = conv(bias = layers_0_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_7_dilations_0, groups = pretrained_out_7_groups_0, pad = pretrained_out_7_pad_0, pad_type = pretrained_out_7_pad_type_0, strides = pretrained_out_7_strides_0, weight = layers_0_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_7_cast_fp16)[name = tensor("pretrained_out_7_cast_fp16")]; tensor input_9_pad_type_0 = const()[name = tensor("input_9_pad_type_0"), val = tensor("valid")]; tensor input_9_strides_0 = const()[name = tensor("input_9_strides_0"), val = tensor([1, 1])]; tensor input_9_pad_0 = const()[name = tensor("input_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_9_dilations_0 = const()[name = tensor("input_9_dilations_0"), val = tensor([1, 1])]; tensor input_9_groups_0 = const()[name = tensor("input_9_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137466112)))]; tensor input_9_cast_fp16 = conv(dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = layers_0_self_attn_o_proj_loraA_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; tensor lora_out_7_pad_type_0 = const()[name = tensor("lora_out_7_pad_type_0"), val = tensor("valid")]; tensor lora_out_7_strides_0 = const()[name = tensor("lora_out_7_strides_0"), val = tensor([1, 1])]; tensor lora_out_7_pad_0 = const()[name = tensor("lora_out_7_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_7_dilations_0 = const()[name = tensor("lora_out_7_dilations_0"), val = tensor([1, 1])]; tensor lora_out_7_groups_0 = const()[name = tensor("lora_out_7_groups_0"), val = tensor(1)]; tensor layers_0_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137507136)))]; tensor lora_out_7_cast_fp16 = conv(dilations = lora_out_7_dilations_0, groups = lora_out_7_groups_0, pad = lora_out_7_pad_0, pad_type = lora_out_7_pad_type_0, strides = lora_out_7_strides_0, weight = layers_0_self_attn_o_proj_loraB_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("lora_out_7_cast_fp16")]; tensor obj_7_cast_fp16 = add(x = pretrained_out_7_cast_fp16, y = lora_out_7_cast_fp16)[name = tensor("obj_7_cast_fp16")]; tensor inputs_3_cast_fp16 = add(x = inputs_1_cast_fp16, y = obj_7_cast_fp16)[name = tensor("inputs_3_cast_fp16")]; tensor out_3_axes_0 = const()[name = tensor("out_3_axes_0"), val = tensor([1])]; tensor var_241_to_fp16 = const()[name = tensor("op_241_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_3_cast_fp16 = layer_norm(axes = out_3_axes_0, epsilon = var_241_to_fp16, x = inputs_3_cast_fp16)[name = tensor("out_3_cast_fp16")]; tensor obj_9_gamma_0_to_fp16 = const()[name = tensor("obj_9_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137548160)))]; tensor obj_9_beta_0_to_fp16 = const()[name = tensor("obj_9_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137550784)))]; tensor obj_9_epsilon_0_to_fp16 = const()[name = tensor("obj_9_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_9_cast_fp16 = batch_norm(beta = obj_9_beta_0_to_fp16, epsilon = obj_9_epsilon_0_to_fp16, gamma = obj_9_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_3_cast_fp16)[name = tensor("obj_9_cast_fp16")]; tensor pretrained_out_9_pad_type_0 = const()[name = tensor("pretrained_out_9_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_9_strides_0 = const()[name = tensor("pretrained_out_9_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_9_pad_0 = const()[name = tensor("pretrained_out_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_9_dilations_0 = const()[name = tensor("pretrained_out_9_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_9_groups_0 = const()[name = tensor("pretrained_out_9_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(137553408))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138372672))), name = tensor("layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138372800)))]; tensor pretrained_out_9_cast_fp16 = conv(bias = layers_0_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_9_dilations_0, groups = pretrained_out_9_groups_0, pad = pretrained_out_9_pad_0, pad_type = pretrained_out_9_pad_type_0, strides = pretrained_out_9_strides_0, weight = layers_0_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_9_cast_fp16)[name = tensor("pretrained_out_9_cast_fp16")]; tensor input_11_pad_type_0 = const()[name = tensor("input_11_pad_type_0"), val = tensor("valid")]; tensor input_11_strides_0 = const()[name = tensor("input_11_strides_0"), val = tensor([1, 1])]; tensor input_11_pad_0 = const()[name = tensor("input_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_11_dilations_0 = const()[name = tensor("input_11_dilations_0"), val = tensor([1, 1])]; tensor input_11_groups_0 = const()[name = tensor("input_11_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138375424)))]; tensor input_11_cast_fp16 = conv(dilations = input_11_dilations_0, groups = input_11_groups_0, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = input_11_strides_0, weight = layers_0_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; tensor lora_out_9_pad_type_0 = const()[name = tensor("lora_out_9_pad_type_0"), val = tensor("valid")]; tensor lora_out_9_strides_0 = const()[name = tensor("lora_out_9_strides_0"), val = tensor([1, 1])]; tensor lora_out_9_pad_0 = const()[name = tensor("lora_out_9_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_9_dilations_0 = const()[name = tensor("lora_out_9_dilations_0"), val = tensor([1, 1])]; tensor lora_out_9_groups_0 = const()[name = tensor("lora_out_9_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138416448)))]; tensor lora_out_9_cast_fp16 = conv(dilations = lora_out_9_dilations_0, groups = lora_out_9_groups_0, pad = lora_out_9_pad_0, pad_type = lora_out_9_pad_type_0, strides = lora_out_9_strides_0, weight = layers_0_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("lora_out_9_cast_fp16")]; tensor query_3_cast_fp16 = add(x = pretrained_out_9_cast_fp16, y = lora_out_9_cast_fp16)[name = tensor("query_3_cast_fp16")]; tensor pretrained_out_11_pad_type_0 = const()[name = tensor("pretrained_out_11_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_11_strides_0 = const()[name = tensor("pretrained_out_11_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_11_pad_0 = const()[name = tensor("pretrained_out_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_11_dilations_0 = const()[name = tensor("pretrained_out_11_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_11_groups_0 = const()[name = tensor("pretrained_out_11_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138457472))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139276736))), name = tensor("layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_11_cast_fp16 = conv(dilations = pretrained_out_11_dilations_0, groups = pretrained_out_11_groups_0, pad = pretrained_out_11_pad_0, pad_type = pretrained_out_11_pad_type_0, strides = pretrained_out_11_strides_0, weight = layers_0_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_11_cast_fp16")]; tensor input_13_pad_type_0 = const()[name = tensor("input_13_pad_type_0"), val = tensor("valid")]; tensor input_13_strides_0 = const()[name = tensor("input_13_strides_0"), val = tensor([1, 1])]; tensor input_13_pad_0 = const()[name = tensor("input_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_13_dilations_0 = const()[name = tensor("input_13_dilations_0"), val = tensor([1, 1])]; tensor input_13_groups_0 = const()[name = tensor("input_13_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139276864)))]; tensor input_13_cast_fp16 = conv(dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = layers_0_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_13_cast_fp16")]; tensor lora_out_11_pad_type_0 = const()[name = tensor("lora_out_11_pad_type_0"), val = tensor("valid")]; tensor lora_out_11_strides_0 = const()[name = tensor("lora_out_11_strides_0"), val = tensor([1, 1])]; tensor lora_out_11_pad_0 = const()[name = tensor("lora_out_11_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_11_dilations_0 = const()[name = tensor("lora_out_11_dilations_0"), val = tensor([1, 1])]; tensor lora_out_11_groups_0 = const()[name = tensor("lora_out_11_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139317888)))]; tensor lora_out_11_cast_fp16 = conv(dilations = lora_out_11_dilations_0, groups = lora_out_11_groups_0, pad = lora_out_11_pad_0, pad_type = lora_out_11_pad_type_0, strides = lora_out_11_strides_0, weight = layers_0_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("lora_out_11_cast_fp16")]; tensor key_3_cast_fp16 = add(x = pretrained_out_11_cast_fp16, y = lora_out_11_cast_fp16)[name = tensor("key_3_cast_fp16")]; tensor pretrained_out_13_pad_type_0 = const()[name = tensor("pretrained_out_13_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_13_strides_0 = const()[name = tensor("pretrained_out_13_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_13_pad_0 = const()[name = tensor("pretrained_out_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_13_dilations_0 = const()[name = tensor("pretrained_out_13_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_13_groups_0 = const()[name = tensor("pretrained_out_13_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(139358912))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140178176))), name = tensor("layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140178304)))]; tensor pretrained_out_13_cast_fp16 = conv(bias = layers_0_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_13_dilations_0, groups = pretrained_out_13_groups_0, pad = pretrained_out_13_pad_0, pad_type = pretrained_out_13_pad_type_0, strides = pretrained_out_13_strides_0, weight = layers_0_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_13_cast_fp16")]; tensor input_15_pad_type_0 = const()[name = tensor("input_15_pad_type_0"), val = tensor("valid")]; tensor input_15_strides_0 = const()[name = tensor("input_15_strides_0"), val = tensor([1, 1])]; tensor input_15_pad_0 = const()[name = tensor("input_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_15_dilations_0 = const()[name = tensor("input_15_dilations_0"), val = tensor([1, 1])]; tensor input_15_groups_0 = const()[name = tensor("input_15_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140180928)))]; tensor input_15_cast_fp16 = conv(dilations = input_15_dilations_0, groups = input_15_groups_0, pad = input_15_pad_0, pad_type = input_15_pad_type_0, strides = input_15_strides_0, weight = layers_0_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_15_cast_fp16")]; tensor lora_out_13_pad_type_0 = const()[name = tensor("lora_out_13_pad_type_0"), val = tensor("valid")]; tensor lora_out_13_strides_0 = const()[name = tensor("lora_out_13_strides_0"), val = tensor([1, 1])]; tensor lora_out_13_pad_0 = const()[name = tensor("lora_out_13_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_13_dilations_0 = const()[name = tensor("lora_out_13_dilations_0"), val = tensor([1, 1])]; tensor lora_out_13_groups_0 = const()[name = tensor("lora_out_13_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140221952)))]; tensor lora_out_13_cast_fp16 = conv(dilations = lora_out_13_dilations_0, groups = lora_out_13_groups_0, pad = lora_out_13_pad_0, pad_type = lora_out_13_pad_type_0, strides = lora_out_13_strides_0, weight = layers_0_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("lora_out_13_cast_fp16")]; tensor value_3_cast_fp16 = add(x = pretrained_out_13_cast_fp16, y = lora_out_13_cast_fp16)[name = tensor("value_3_cast_fp16")]; tensor var_324 = const()[name = tensor("op_324"), val = tensor([1, 20, 64, -1])]; tensor mh_q_3_cast_fp16 = reshape(shape = var_324, x = query_3_cast_fp16)[name = tensor("mh_q_3_cast_fp16")]; tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(0x1p-3)]; tensor var_327_cast_fp16 = mul(x = mh_q_3_cast_fp16, y = var_326_to_fp16)[name = tensor("op_327_cast_fp16")]; tensor var_328 = const()[name = tensor("op_328"), val = tensor([1, 20, 64, -1])]; tensor var_329_cast_fp16 = reshape(shape = var_328, x = key_3_cast_fp16)[name = tensor("op_329_cast_fp16")]; tensor mh_w_5_transpose_x_0 = const()[name = tensor("mh_w_5_transpose_x_0"), val = tensor(true)]; tensor mh_w_5_transpose_y_0 = const()[name = tensor("mh_w_5_transpose_y_0"), val = tensor(false)]; tensor mh_w_5_cast_fp16 = matmul(transpose_x = mh_w_5_transpose_x_0, transpose_y = mh_w_5_transpose_y_0, x = var_327_cast_fp16, y = var_329_cast_fp16)[name = tensor("mh_w_5_cast_fp16")]; tensor obj_13_cast_fp16 = softmax(axis = var_64, x = mh_w_5_cast_fp16)[name = tensor("obj_13_cast_fp16")]; tensor var_333 = const()[name = tensor("op_333"), val = tensor([1, 20, 64, -1])]; tensor var_334_cast_fp16 = reshape(shape = var_333, x = value_3_cast_fp16)[name = tensor("op_334_cast_fp16")]; tensor attn_3_transpose_x_0 = const()[name = tensor("attn_3_transpose_x_0"), val = tensor(false)]; tensor attn_3_transpose_y_0 = const()[name = tensor("attn_3_transpose_y_0"), val = tensor(true)]; tensor attn_3_cast_fp16 = matmul(transpose_x = attn_3_transpose_x_0, transpose_y = attn_3_transpose_y_0, x = var_334_cast_fp16, y = obj_13_cast_fp16)[name = tensor("attn_3_cast_fp16")]; tensor var_337 = const()[name = tensor("op_337"), val = tensor([1, 1280, 1, -1])]; tensor input_17_cast_fp16 = reshape(shape = var_337, x = attn_3_cast_fp16)[name = tensor("input_17_cast_fp16")]; tensor pretrained_out_15_pad_type_0 = const()[name = tensor("pretrained_out_15_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_15_strides_0 = const()[name = tensor("pretrained_out_15_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_15_pad_0 = const()[name = tensor("pretrained_out_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_15_dilations_0 = const()[name = tensor("pretrained_out_15_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_15_groups_0 = const()[name = tensor("pretrained_out_15_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(140262976))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141082240))), name = tensor("layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141082368)))]; tensor pretrained_out_15_cast_fp16 = conv(bias = layers_0_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_15_dilations_0, groups = pretrained_out_15_groups_0, pad = pretrained_out_15_pad_0, pad_type = pretrained_out_15_pad_type_0, strides = pretrained_out_15_strides_0, weight = layers_0_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_17_cast_fp16)[name = tensor("pretrained_out_15_cast_fp16")]; tensor input_19_pad_type_0 = const()[name = tensor("input_19_pad_type_0"), val = tensor("valid")]; tensor input_19_strides_0 = const()[name = tensor("input_19_strides_0"), val = tensor([1, 1])]; tensor input_19_pad_0 = const()[name = tensor("input_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_19_dilations_0 = const()[name = tensor("input_19_dilations_0"), val = tensor([1, 1])]; tensor input_19_groups_0 = const()[name = tensor("input_19_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141084992)))]; tensor input_19_cast_fp16 = conv(dilations = input_19_dilations_0, groups = input_19_groups_0, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = input_19_strides_0, weight = layers_0_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; tensor lora_out_15_pad_type_0 = const()[name = tensor("lora_out_15_pad_type_0"), val = tensor("valid")]; tensor lora_out_15_strides_0 = const()[name = tensor("lora_out_15_strides_0"), val = tensor([1, 1])]; tensor lora_out_15_pad_0 = const()[name = tensor("lora_out_15_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_15_dilations_0 = const()[name = tensor("lora_out_15_dilations_0"), val = tensor([1, 1])]; tensor lora_out_15_groups_0 = const()[name = tensor("lora_out_15_groups_0"), val = tensor(1)]; tensor layers_0_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_0_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141126016)))]; tensor lora_out_15_cast_fp16 = conv(dilations = lora_out_15_dilations_0, groups = lora_out_15_groups_0, pad = lora_out_15_pad_0, pad_type = lora_out_15_pad_type_0, strides = lora_out_15_strides_0, weight = layers_0_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("lora_out_15_cast_fp16")]; tensor obj_11_cast_fp16 = add(x = pretrained_out_15_cast_fp16, y = lora_out_15_cast_fp16)[name = tensor("obj_11_cast_fp16")]; tensor inputs_5_cast_fp16 = add(x = inputs_3_cast_fp16, y = obj_11_cast_fp16)[name = tensor("inputs_5_cast_fp16")]; tensor out_5_axes_0 = const()[name = tensor("out_5_axes_0"), val = tensor([1])]; tensor var_371_to_fp16 = const()[name = tensor("op_371_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_5_cast_fp16 = layer_norm(axes = out_5_axes_0, epsilon = var_371_to_fp16, x = inputs_5_cast_fp16)[name = tensor("out_5_cast_fp16")]; tensor input_21_gamma_0_to_fp16 = const()[name = tensor("input_21_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141167040)))]; tensor input_21_beta_0_to_fp16 = const()[name = tensor("input_21_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141169664)))]; tensor input_21_epsilon_0_to_fp16 = const()[name = tensor("input_21_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor input_21_cast_fp16 = batch_norm(beta = input_21_beta_0_to_fp16, epsilon = input_21_epsilon_0_to_fp16, gamma = input_21_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_5_cast_fp16)[name = tensor("input_21_cast_fp16")]; tensor pretrained_out_17_pad_type_0 = const()[name = tensor("pretrained_out_17_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_17_strides_0 = const()[name = tensor("pretrained_out_17_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_17_pad_0 = const()[name = tensor("pretrained_out_17_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_17_dilations_0 = const()[name = tensor("pretrained_out_17_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_17_groups_0 = const()[name = tensor("pretrained_out_17_groups_0"), val = tensor(1)]; tensor layers_0_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141172288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144449152))), name = tensor("layers_0_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_0_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144449280)))]; tensor pretrained_out_17_cast_fp16 = conv(bias = layers_0_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_17_dilations_0, groups = pretrained_out_17_groups_0, pad = pretrained_out_17_pad_0, pad_type = pretrained_out_17_pad_type_0, strides = pretrained_out_17_strides_0, weight = layers_0_fc1_pretrained_weight_to_fp16_palettized, x = input_21_cast_fp16)[name = tensor("pretrained_out_17_cast_fp16")]; tensor input_23_pad_type_0 = const()[name = tensor("input_23_pad_type_0"), val = tensor("valid")]; tensor input_23_strides_0 = const()[name = tensor("input_23_strides_0"), val = tensor([1, 1])]; tensor input_23_pad_0 = const()[name = tensor("input_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_23_dilations_0 = const()[name = tensor("input_23_dilations_0"), val = tensor([1, 1])]; tensor input_23_groups_0 = const()[name = tensor("input_23_groups_0"), val = tensor(1)]; tensor layers_0_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_0_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144459584)))]; tensor input_23_cast_fp16 = conv(dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = layers_0_fc1_loraA_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; tensor lora_out_17_pad_type_0 = const()[name = tensor("lora_out_17_pad_type_0"), val = tensor("valid")]; tensor lora_out_17_strides_0 = const()[name = tensor("lora_out_17_strides_0"), val = tensor([1, 1])]; tensor lora_out_17_pad_0 = const()[name = tensor("lora_out_17_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_17_dilations_0 = const()[name = tensor("lora_out_17_dilations_0"), val = tensor([1, 1])]; tensor lora_out_17_groups_0 = const()[name = tensor("lora_out_17_groups_0"), val = tensor(1)]; tensor layers_0_fc1_loraB_weight_to_fp16 = const()[name = tensor("layers_0_fc1_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144500608)))]; tensor lora_out_17_cast_fp16 = conv(dilations = lora_out_17_dilations_0, groups = lora_out_17_groups_0, pad = lora_out_17_pad_0, pad_type = lora_out_17_pad_type_0, strides = lora_out_17_strides_0, weight = layers_0_fc1_loraB_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("lora_out_17_cast_fp16")]; tensor input_25_cast_fp16 = add(x = pretrained_out_17_cast_fp16, y = lora_out_17_cast_fp16)[name = tensor("input_25_cast_fp16")]; tensor input_27_mode_0 = const()[name = tensor("input_27_mode_0"), val = tensor("EXACT")]; tensor input_27_cast_fp16 = gelu(mode = input_27_mode_0, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; tensor pretrained_out_19_pad_type_0 = const()[name = tensor("pretrained_out_19_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_19_strides_0 = const()[name = tensor("pretrained_out_19_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_19_pad_0 = const()[name = tensor("pretrained_out_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_19_dilations_0 = const()[name = tensor("pretrained_out_19_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_19_groups_0 = const()[name = tensor("pretrained_out_19_groups_0"), val = tensor(1)]; tensor layers_0_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(144664512))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147941376))), name = tensor("layers_0_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_0_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_0_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147941504)))]; tensor pretrained_out_19_cast_fp16 = conv(bias = layers_0_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_19_dilations_0, groups = pretrained_out_19_groups_0, pad = pretrained_out_19_pad_0, pad_type = pretrained_out_19_pad_type_0, strides = pretrained_out_19_strides_0, weight = layers_0_fc2_pretrained_weight_to_fp16_palettized, x = input_27_cast_fp16)[name = tensor("pretrained_out_19_cast_fp16")]; tensor input_29_pad_type_0 = const()[name = tensor("input_29_pad_type_0"), val = tensor("valid")]; tensor input_29_strides_0 = const()[name = tensor("input_29_strides_0"), val = tensor([1, 1])]; tensor input_29_pad_0 = const()[name = tensor("input_29_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_29_dilations_0 = const()[name = tensor("input_29_dilations_0"), val = tensor([1, 1])]; tensor input_29_groups_0 = const()[name = tensor("input_29_groups_0"), val = tensor(1)]; tensor layers_0_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_0_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147944128)))]; tensor input_29_cast_fp16 = conv(dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = layers_0_fc2_loraA_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; tensor lora_out_19_pad_type_0 = const()[name = tensor("lora_out_19_pad_type_0"), val = tensor("valid")]; tensor lora_out_19_strides_0 = const()[name = tensor("lora_out_19_strides_0"), val = tensor([1, 1])]; tensor lora_out_19_pad_0 = const()[name = tensor("lora_out_19_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_19_dilations_0 = const()[name = tensor("lora_out_19_dilations_0"), val = tensor([1, 1])]; tensor lora_out_19_groups_0 = const()[name = tensor("lora_out_19_groups_0"), val = tensor(1)]; tensor layers_0_fc2_loraB_weight_to_fp16 = const()[name = tensor("layers_0_fc2_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148108032)))]; tensor lora_out_19_cast_fp16 = conv(dilations = lora_out_19_dilations_0, groups = lora_out_19_groups_0, pad = lora_out_19_pad_0, pad_type = lora_out_19_pad_type_0, strides = lora_out_19_strides_0, weight = layers_0_fc2_loraB_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("lora_out_19_cast_fp16")]; tensor hidden_states_3_cast_fp16 = add(x = pretrained_out_19_cast_fp16, y = lora_out_19_cast_fp16)[name = tensor("hidden_states_3_cast_fp16")]; tensor inputs_7_cast_fp16 = add(x = inputs_5_cast_fp16, y = hidden_states_3_cast_fp16)[name = tensor("inputs_7_cast_fp16")]; tensor var_438 = const()[name = tensor("op_438"), val = tensor(3)]; tensor out_7_axes_0 = const()[name = tensor("out_7_axes_0"), val = tensor([1])]; tensor var_464_to_fp16 = const()[name = tensor("op_464_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_7_cast_fp16 = layer_norm(axes = out_7_axes_0, epsilon = var_464_to_fp16, x = inputs_7_cast_fp16)[name = tensor("out_7_cast_fp16")]; tensor obj_15_gamma_0_to_fp16 = const()[name = tensor("obj_15_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148149056)))]; tensor obj_15_beta_0_to_fp16 = const()[name = tensor("obj_15_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148151680)))]; tensor obj_15_epsilon_0_to_fp16 = const()[name = tensor("obj_15_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_15_cast_fp16 = batch_norm(beta = obj_15_beta_0_to_fp16, epsilon = obj_15_epsilon_0_to_fp16, gamma = obj_15_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_7_cast_fp16)[name = tensor("obj_15_cast_fp16")]; tensor pretrained_out_21_pad_type_0 = const()[name = tensor("pretrained_out_21_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_21_strides_0 = const()[name = tensor("pretrained_out_21_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_21_pad_0 = const()[name = tensor("pretrained_out_21_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_21_dilations_0 = const()[name = tensor("pretrained_out_21_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_21_groups_0 = const()[name = tensor("pretrained_out_21_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148154304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148973568))), name = tensor("layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148973696)))]; tensor pretrained_out_21_cast_fp16 = conv(bias = layers_1_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_21_dilations_0, groups = pretrained_out_21_groups_0, pad = pretrained_out_21_pad_0, pad_type = pretrained_out_21_pad_type_0, strides = pretrained_out_21_strides_0, weight = layers_1_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("pretrained_out_21_cast_fp16")]; tensor input_31_pad_type_0 = const()[name = tensor("input_31_pad_type_0"), val = tensor("valid")]; tensor input_31_strides_0 = const()[name = tensor("input_31_strides_0"), val = tensor([1, 1])]; tensor input_31_pad_0 = const()[name = tensor("input_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_31_dilations_0 = const()[name = tensor("input_31_dilations_0"), val = tensor([1, 1])]; tensor input_31_groups_0 = const()[name = tensor("input_31_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(148976320)))]; tensor input_31_cast_fp16 = conv(dilations = input_31_dilations_0, groups = input_31_groups_0, pad = input_31_pad_0, pad_type = input_31_pad_type_0, strides = input_31_strides_0, weight = layers_1_self_attn_q_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("input_31_cast_fp16")]; tensor lora_out_21_pad_type_0 = const()[name = tensor("lora_out_21_pad_type_0"), val = tensor("valid")]; tensor lora_out_21_strides_0 = const()[name = tensor("lora_out_21_strides_0"), val = tensor([1, 1])]; tensor lora_out_21_pad_0 = const()[name = tensor("lora_out_21_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_21_dilations_0 = const()[name = tensor("lora_out_21_dilations_0"), val = tensor([1, 1])]; tensor lora_out_21_groups_0 = const()[name = tensor("lora_out_21_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149017344)))]; tensor lora_out_21_cast_fp16 = conv(dilations = lora_out_21_dilations_0, groups = lora_out_21_groups_0, pad = lora_out_21_pad_0, pad_type = lora_out_21_pad_type_0, strides = lora_out_21_strides_0, weight = layers_1_self_attn_q_proj_loraB_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("lora_out_21_cast_fp16")]; tensor query_5_cast_fp16 = add(x = pretrained_out_21_cast_fp16, y = lora_out_21_cast_fp16)[name = tensor("query_5_cast_fp16")]; tensor pretrained_out_23_pad_type_0 = const()[name = tensor("pretrained_out_23_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_23_strides_0 = const()[name = tensor("pretrained_out_23_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_23_pad_0 = const()[name = tensor("pretrained_out_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_23_dilations_0 = const()[name = tensor("pretrained_out_23_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_23_groups_0 = const()[name = tensor("pretrained_out_23_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149058368))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149877632))), name = tensor("layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_23_cast_fp16 = conv(dilations = pretrained_out_23_dilations_0, groups = pretrained_out_23_groups_0, pad = pretrained_out_23_pad_0, pad_type = pretrained_out_23_pad_type_0, strides = pretrained_out_23_strides_0, weight = layers_1_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("pretrained_out_23_cast_fp16")]; tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1, 1])]; tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1, 1])]; tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149877760)))]; tensor input_33_cast_fp16 = conv(dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = layers_1_self_attn_k_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("input_33_cast_fp16")]; tensor lora_out_23_pad_type_0 = const()[name = tensor("lora_out_23_pad_type_0"), val = tensor("valid")]; tensor lora_out_23_strides_0 = const()[name = tensor("lora_out_23_strides_0"), val = tensor([1, 1])]; tensor lora_out_23_pad_0 = const()[name = tensor("lora_out_23_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_23_dilations_0 = const()[name = tensor("lora_out_23_dilations_0"), val = tensor([1, 1])]; tensor lora_out_23_groups_0 = const()[name = tensor("lora_out_23_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149918784)))]; tensor lora_out_23_cast_fp16 = conv(dilations = lora_out_23_dilations_0, groups = lora_out_23_groups_0, pad = lora_out_23_pad_0, pad_type = lora_out_23_pad_type_0, strides = lora_out_23_strides_0, weight = layers_1_self_attn_k_proj_loraB_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("lora_out_23_cast_fp16")]; tensor current_key_3_cast_fp16 = add(x = pretrained_out_23_cast_fp16, y = lora_out_23_cast_fp16)[name = tensor("current_key_3_cast_fp16")]; tensor pretrained_out_25_pad_type_0 = const()[name = tensor("pretrained_out_25_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_25_strides_0 = const()[name = tensor("pretrained_out_25_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_25_pad_0 = const()[name = tensor("pretrained_out_25_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_25_dilations_0 = const()[name = tensor("pretrained_out_25_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_25_groups_0 = const()[name = tensor("pretrained_out_25_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(149959808))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150779072))), name = tensor("layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150779200)))]; tensor pretrained_out_25_cast_fp16 = conv(bias = layers_1_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_25_dilations_0, groups = pretrained_out_25_groups_0, pad = pretrained_out_25_pad_0, pad_type = pretrained_out_25_pad_type_0, strides = pretrained_out_25_strides_0, weight = layers_1_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_15_cast_fp16)[name = tensor("pretrained_out_25_cast_fp16")]; tensor input_35_pad_type_0 = const()[name = tensor("input_35_pad_type_0"), val = tensor("valid")]; tensor input_35_strides_0 = const()[name = tensor("input_35_strides_0"), val = tensor([1, 1])]; tensor input_35_pad_0 = const()[name = tensor("input_35_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_35_dilations_0 = const()[name = tensor("input_35_dilations_0"), val = tensor([1, 1])]; tensor input_35_groups_0 = const()[name = tensor("input_35_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150781824)))]; tensor input_35_cast_fp16 = conv(dilations = input_35_dilations_0, groups = input_35_groups_0, pad = input_35_pad_0, pad_type = input_35_pad_type_0, strides = input_35_strides_0, weight = layers_1_self_attn_v_proj_loraA_weight_to_fp16, x = obj_15_cast_fp16)[name = tensor("input_35_cast_fp16")]; tensor lora_out_25_pad_type_0 = const()[name = tensor("lora_out_25_pad_type_0"), val = tensor("valid")]; tensor lora_out_25_strides_0 = const()[name = tensor("lora_out_25_strides_0"), val = tensor([1, 1])]; tensor lora_out_25_pad_0 = const()[name = tensor("lora_out_25_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_25_dilations_0 = const()[name = tensor("lora_out_25_dilations_0"), val = tensor([1, 1])]; tensor lora_out_25_groups_0 = const()[name = tensor("lora_out_25_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150822848)))]; tensor lora_out_25_cast_fp16 = conv(dilations = lora_out_25_dilations_0, groups = lora_out_25_groups_0, pad = lora_out_25_pad_0, pad_type = lora_out_25_pad_type_0, strides = lora_out_25_strides_0, weight = layers_1_self_attn_v_proj_loraB_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("lora_out_25_cast_fp16")]; tensor current_value_3_cast_fp16 = add(x = pretrained_out_25_cast_fp16, y = lora_out_25_cast_fp16)[name = tensor("current_value_3_cast_fp16")]; tensor var_550_cast_fp16 = mul(x = current_key_3_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_550_cast_fp16")]; tensor var_552_cast_fp16 = mul(x = var_47_cast_fp16_1, y = var_177_cast_fp16)[name = tensor("op_552_cast_fp16")]; tensor key_5_cast_fp16 = add(x = var_550_cast_fp16, y = var_552_cast_fp16)[name = tensor("key_5_cast_fp16")]; tensor var_554_cast_fp16 = mul(x = current_value_3_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_554_cast_fp16")]; tensor var_556_cast_fp16 = mul(x = var_54_cast_fp16_1, y = var_177_cast_fp16)[name = tensor("op_556_cast_fp16")]; tensor value_5_cast_fp16 = add(x = var_554_cast_fp16, y = var_556_cast_fp16)[name = tensor("value_5_cast_fp16")]; tensor var_559 = const()[name = tensor("op_559"), val = tensor([1, 20, 64, -1])]; tensor mh_q_5_cast_fp16 = reshape(shape = var_559, x = query_5_cast_fp16)[name = tensor("mh_q_5_cast_fp16")]; tensor var_561_to_fp16 = const()[name = tensor("op_561_to_fp16"), val = tensor(0x1p-3)]; tensor var_562_cast_fp16 = mul(x = mh_q_5_cast_fp16, y = var_561_to_fp16)[name = tensor("op_562_cast_fp16")]; tensor var_563 = const()[name = tensor("op_563"), val = tensor([1, 20, 64, -1])]; tensor var_564_cast_fp16 = reshape(shape = var_563, x = key_5_cast_fp16)[name = tensor("op_564_cast_fp16")]; tensor mh_w_7_transpose_x_0 = const()[name = tensor("mh_w_7_transpose_x_0"), val = tensor(true)]; tensor mh_w_7_transpose_y_0 = const()[name = tensor("mh_w_7_transpose_y_0"), val = tensor(false)]; tensor mh_w_7_cast_fp16 = matmul(transpose_x = mh_w_7_transpose_x_0, transpose_y = mh_w_7_transpose_y_0, x = var_562_cast_fp16, y = var_564_cast_fp16)[name = tensor("mh_w_7_cast_fp16")]; tensor mh_w_9_cast_fp16 = add(x = mh_w_7_cast_fp16, y = var_195_cast_fp16)[name = tensor("mh_w_9_cast_fp16")]; tensor var_572_cast_fp16 = softmax(axis = var_438, x = mh_w_9_cast_fp16)[name = tensor("op_572_cast_fp16")]; tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 20, 64, -1])]; tensor var_574_cast_fp16 = reshape(shape = var_573, x = value_5_cast_fp16)[name = tensor("op_574_cast_fp16")]; tensor attn_5_transpose_x_0 = const()[name = tensor("attn_5_transpose_x_0"), val = tensor(false)]; tensor attn_5_transpose_y_0 = const()[name = tensor("attn_5_transpose_y_0"), val = tensor(true)]; tensor attn_5_cast_fp16 = matmul(transpose_x = attn_5_transpose_x_0, transpose_y = attn_5_transpose_y_0, x = var_574_cast_fp16, y = var_572_cast_fp16)[name = tensor("attn_5_cast_fp16")]; tensor var_577 = const()[name = tensor("op_577"), val = tensor([1, 1280, 1, -1])]; tensor input_37_cast_fp16 = reshape(shape = var_577, x = attn_5_cast_fp16)[name = tensor("input_37_cast_fp16")]; tensor pretrained_out_27_pad_type_0 = const()[name = tensor("pretrained_out_27_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_27_strides_0 = const()[name = tensor("pretrained_out_27_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_27_pad_0 = const()[name = tensor("pretrained_out_27_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_27_dilations_0 = const()[name = tensor("pretrained_out_27_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_27_groups_0 = const()[name = tensor("pretrained_out_27_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(150863872))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151683136))), name = tensor("layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151683264)))]; tensor pretrained_out_27_cast_fp16 = conv(bias = layers_1_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_27_dilations_0, groups = pretrained_out_27_groups_0, pad = pretrained_out_27_pad_0, pad_type = pretrained_out_27_pad_type_0, strides = pretrained_out_27_strides_0, weight = layers_1_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_37_cast_fp16)[name = tensor("pretrained_out_27_cast_fp16")]; tensor input_39_pad_type_0 = const()[name = tensor("input_39_pad_type_0"), val = tensor("valid")]; tensor input_39_strides_0 = const()[name = tensor("input_39_strides_0"), val = tensor([1, 1])]; tensor input_39_pad_0 = const()[name = tensor("input_39_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_39_dilations_0 = const()[name = tensor("input_39_dilations_0"), val = tensor([1, 1])]; tensor input_39_groups_0 = const()[name = tensor("input_39_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151685888)))]; tensor input_39_cast_fp16 = conv(dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = layers_1_self_attn_o_proj_loraA_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; tensor lora_out_27_pad_type_0 = const()[name = tensor("lora_out_27_pad_type_0"), val = tensor("valid")]; tensor lora_out_27_strides_0 = const()[name = tensor("lora_out_27_strides_0"), val = tensor([1, 1])]; tensor lora_out_27_pad_0 = const()[name = tensor("lora_out_27_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_27_dilations_0 = const()[name = tensor("lora_out_27_dilations_0"), val = tensor([1, 1])]; tensor lora_out_27_groups_0 = const()[name = tensor("lora_out_27_groups_0"), val = tensor(1)]; tensor layers_1_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151726912)))]; tensor lora_out_27_cast_fp16 = conv(dilations = lora_out_27_dilations_0, groups = lora_out_27_groups_0, pad = lora_out_27_pad_0, pad_type = lora_out_27_pad_type_0, strides = lora_out_27_strides_0, weight = layers_1_self_attn_o_proj_loraB_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("lora_out_27_cast_fp16")]; tensor obj_21_cast_fp16 = add(x = pretrained_out_27_cast_fp16, y = lora_out_27_cast_fp16)[name = tensor("obj_21_cast_fp16")]; tensor inputs_9_cast_fp16 = add(x = inputs_7_cast_fp16, y = obj_21_cast_fp16)[name = tensor("inputs_9_cast_fp16")]; tensor out_9_axes_0 = const()[name = tensor("out_9_axes_0"), val = tensor([1])]; tensor var_615_to_fp16 = const()[name = tensor("op_615_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_9_cast_fp16 = layer_norm(axes = out_9_axes_0, epsilon = var_615_to_fp16, x = inputs_9_cast_fp16)[name = tensor("out_9_cast_fp16")]; tensor obj_23_gamma_0_to_fp16 = const()[name = tensor("obj_23_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151767936)))]; tensor obj_23_beta_0_to_fp16 = const()[name = tensor("obj_23_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151770560)))]; tensor obj_23_epsilon_0_to_fp16 = const()[name = tensor("obj_23_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_23_cast_fp16 = batch_norm(beta = obj_23_beta_0_to_fp16, epsilon = obj_23_epsilon_0_to_fp16, gamma = obj_23_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_9_cast_fp16)[name = tensor("obj_23_cast_fp16")]; tensor pretrained_out_29_pad_type_0 = const()[name = tensor("pretrained_out_29_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_29_strides_0 = const()[name = tensor("pretrained_out_29_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_29_pad_0 = const()[name = tensor("pretrained_out_29_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_29_dilations_0 = const()[name = tensor("pretrained_out_29_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_29_groups_0 = const()[name = tensor("pretrained_out_29_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151773184))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152592448))), name = tensor("layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152592576)))]; tensor pretrained_out_29_cast_fp16 = conv(bias = layers_1_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_29_dilations_0, groups = pretrained_out_29_groups_0, pad = pretrained_out_29_pad_0, pad_type = pretrained_out_29_pad_type_0, strides = pretrained_out_29_strides_0, weight = layers_1_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_23_cast_fp16)[name = tensor("pretrained_out_29_cast_fp16")]; tensor input_41_pad_type_0 = const()[name = tensor("input_41_pad_type_0"), val = tensor("valid")]; tensor input_41_strides_0 = const()[name = tensor("input_41_strides_0"), val = tensor([1, 1])]; tensor input_41_pad_0 = const()[name = tensor("input_41_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_41_dilations_0 = const()[name = tensor("input_41_dilations_0"), val = tensor([1, 1])]; tensor input_41_groups_0 = const()[name = tensor("input_41_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152595200)))]; tensor input_41_cast_fp16 = conv(dilations = input_41_dilations_0, groups = input_41_groups_0, pad = input_41_pad_0, pad_type = input_41_pad_type_0, strides = input_41_strides_0, weight = layers_1_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_23_cast_fp16)[name = tensor("input_41_cast_fp16")]; tensor lora_out_29_pad_type_0 = const()[name = tensor("lora_out_29_pad_type_0"), val = tensor("valid")]; tensor lora_out_29_strides_0 = const()[name = tensor("lora_out_29_strides_0"), val = tensor([1, 1])]; tensor lora_out_29_pad_0 = const()[name = tensor("lora_out_29_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_29_dilations_0 = const()[name = tensor("lora_out_29_dilations_0"), val = tensor([1, 1])]; tensor lora_out_29_groups_0 = const()[name = tensor("lora_out_29_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152636224)))]; tensor lora_out_29_cast_fp16 = conv(dilations = lora_out_29_dilations_0, groups = lora_out_29_groups_0, pad = lora_out_29_pad_0, pad_type = lora_out_29_pad_type_0, strides = lora_out_29_strides_0, weight = layers_1_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("lora_out_29_cast_fp16")]; tensor query_7_cast_fp16 = add(x = pretrained_out_29_cast_fp16, y = lora_out_29_cast_fp16)[name = tensor("query_7_cast_fp16")]; tensor pretrained_out_31_pad_type_0 = const()[name = tensor("pretrained_out_31_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_31_strides_0 = const()[name = tensor("pretrained_out_31_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_31_pad_0 = const()[name = tensor("pretrained_out_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_31_dilations_0 = const()[name = tensor("pretrained_out_31_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_31_groups_0 = const()[name = tensor("pretrained_out_31_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152677248))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153496512))), name = tensor("layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_31_cast_fp16 = conv(dilations = pretrained_out_31_dilations_0, groups = pretrained_out_31_groups_0, pad = pretrained_out_31_pad_0, pad_type = pretrained_out_31_pad_type_0, strides = pretrained_out_31_strides_0, weight = layers_1_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_31_cast_fp16")]; tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1, 1])]; tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1, 1])]; tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153496640)))]; tensor input_43_cast_fp16 = conv(dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = layers_1_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_43_cast_fp16")]; tensor lora_out_31_pad_type_0 = const()[name = tensor("lora_out_31_pad_type_0"), val = tensor("valid")]; tensor lora_out_31_strides_0 = const()[name = tensor("lora_out_31_strides_0"), val = tensor([1, 1])]; tensor lora_out_31_pad_0 = const()[name = tensor("lora_out_31_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_31_dilations_0 = const()[name = tensor("lora_out_31_dilations_0"), val = tensor([1, 1])]; tensor lora_out_31_groups_0 = const()[name = tensor("lora_out_31_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153537664)))]; tensor lora_out_31_cast_fp16 = conv(dilations = lora_out_31_dilations_0, groups = lora_out_31_groups_0, pad = lora_out_31_pad_0, pad_type = lora_out_31_pad_type_0, strides = lora_out_31_strides_0, weight = layers_1_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("lora_out_31_cast_fp16")]; tensor key_7_cast_fp16 = add(x = pretrained_out_31_cast_fp16, y = lora_out_31_cast_fp16)[name = tensor("key_7_cast_fp16")]; tensor pretrained_out_33_pad_type_0 = const()[name = tensor("pretrained_out_33_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_33_strides_0 = const()[name = tensor("pretrained_out_33_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_33_pad_0 = const()[name = tensor("pretrained_out_33_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_33_dilations_0 = const()[name = tensor("pretrained_out_33_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_33_groups_0 = const()[name = tensor("pretrained_out_33_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(153578688))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154397952))), name = tensor("layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154398080)))]; tensor pretrained_out_33_cast_fp16 = conv(bias = layers_1_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_33_dilations_0, groups = pretrained_out_33_groups_0, pad = pretrained_out_33_pad_0, pad_type = pretrained_out_33_pad_type_0, strides = pretrained_out_33_strides_0, weight = layers_1_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_33_cast_fp16")]; tensor input_45_pad_type_0 = const()[name = tensor("input_45_pad_type_0"), val = tensor("valid")]; tensor input_45_strides_0 = const()[name = tensor("input_45_strides_0"), val = tensor([1, 1])]; tensor input_45_pad_0 = const()[name = tensor("input_45_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_45_dilations_0 = const()[name = tensor("input_45_dilations_0"), val = tensor([1, 1])]; tensor input_45_groups_0 = const()[name = tensor("input_45_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154400704)))]; tensor input_45_cast_fp16 = conv(dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = layers_1_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_45_cast_fp16")]; tensor lora_out_33_pad_type_0 = const()[name = tensor("lora_out_33_pad_type_0"), val = tensor("valid")]; tensor lora_out_33_strides_0 = const()[name = tensor("lora_out_33_strides_0"), val = tensor([1, 1])]; tensor lora_out_33_pad_0 = const()[name = tensor("lora_out_33_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_33_dilations_0 = const()[name = tensor("lora_out_33_dilations_0"), val = tensor([1, 1])]; tensor lora_out_33_groups_0 = const()[name = tensor("lora_out_33_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154441728)))]; tensor lora_out_33_cast_fp16 = conv(dilations = lora_out_33_dilations_0, groups = lora_out_33_groups_0, pad = lora_out_33_pad_0, pad_type = lora_out_33_pad_type_0, strides = lora_out_33_strides_0, weight = layers_1_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("lora_out_33_cast_fp16")]; tensor value_7_cast_fp16 = add(x = pretrained_out_33_cast_fp16, y = lora_out_33_cast_fp16)[name = tensor("value_7_cast_fp16")]; tensor var_698 = const()[name = tensor("op_698"), val = tensor([1, 20, 64, -1])]; tensor mh_q_7_cast_fp16 = reshape(shape = var_698, x = query_7_cast_fp16)[name = tensor("mh_q_7_cast_fp16")]; tensor var_700_to_fp16 = const()[name = tensor("op_700_to_fp16"), val = tensor(0x1p-3)]; tensor var_701_cast_fp16 = mul(x = mh_q_7_cast_fp16, y = var_700_to_fp16)[name = tensor("op_701_cast_fp16")]; tensor var_702 = const()[name = tensor("op_702"), val = tensor([1, 20, 64, -1])]; tensor var_703_cast_fp16 = reshape(shape = var_702, x = key_7_cast_fp16)[name = tensor("op_703_cast_fp16")]; tensor mh_w_11_transpose_x_0 = const()[name = tensor("mh_w_11_transpose_x_0"), val = tensor(true)]; tensor mh_w_11_transpose_y_0 = const()[name = tensor("mh_w_11_transpose_y_0"), val = tensor(false)]; tensor mh_w_11_cast_fp16 = matmul(transpose_x = mh_w_11_transpose_x_0, transpose_y = mh_w_11_transpose_y_0, x = var_701_cast_fp16, y = var_703_cast_fp16)[name = tensor("mh_w_11_cast_fp16")]; tensor obj_27_cast_fp16 = softmax(axis = var_438, x = mh_w_11_cast_fp16)[name = tensor("obj_27_cast_fp16")]; tensor var_707 = const()[name = tensor("op_707"), val = tensor([1, 20, 64, -1])]; tensor var_708_cast_fp16 = reshape(shape = var_707, x = value_7_cast_fp16)[name = tensor("op_708_cast_fp16")]; tensor attn_7_transpose_x_0 = const()[name = tensor("attn_7_transpose_x_0"), val = tensor(false)]; tensor attn_7_transpose_y_0 = const()[name = tensor("attn_7_transpose_y_0"), val = tensor(true)]; tensor attn_7_cast_fp16 = matmul(transpose_x = attn_7_transpose_x_0, transpose_y = attn_7_transpose_y_0, x = var_708_cast_fp16, y = obj_27_cast_fp16)[name = tensor("attn_7_cast_fp16")]; tensor var_711 = const()[name = tensor("op_711"), val = tensor([1, 1280, 1, -1])]; tensor input_47_cast_fp16 = reshape(shape = var_711, x = attn_7_cast_fp16)[name = tensor("input_47_cast_fp16")]; tensor pretrained_out_35_pad_type_0 = const()[name = tensor("pretrained_out_35_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_35_strides_0 = const()[name = tensor("pretrained_out_35_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_35_pad_0 = const()[name = tensor("pretrained_out_35_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_35_dilations_0 = const()[name = tensor("pretrained_out_35_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_35_groups_0 = const()[name = tensor("pretrained_out_35_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154482752))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155302016))), name = tensor("layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155302144)))]; tensor pretrained_out_35_cast_fp16 = conv(bias = layers_1_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_35_dilations_0, groups = pretrained_out_35_groups_0, pad = pretrained_out_35_pad_0, pad_type = pretrained_out_35_pad_type_0, strides = pretrained_out_35_strides_0, weight = layers_1_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_47_cast_fp16)[name = tensor("pretrained_out_35_cast_fp16")]; tensor input_49_pad_type_0 = const()[name = tensor("input_49_pad_type_0"), val = tensor("valid")]; tensor input_49_strides_0 = const()[name = tensor("input_49_strides_0"), val = tensor([1, 1])]; tensor input_49_pad_0 = const()[name = tensor("input_49_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_49_dilations_0 = const()[name = tensor("input_49_dilations_0"), val = tensor([1, 1])]; tensor input_49_groups_0 = const()[name = tensor("input_49_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155304768)))]; tensor input_49_cast_fp16 = conv(dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = layers_1_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; tensor lora_out_35_pad_type_0 = const()[name = tensor("lora_out_35_pad_type_0"), val = tensor("valid")]; tensor lora_out_35_strides_0 = const()[name = tensor("lora_out_35_strides_0"), val = tensor([1, 1])]; tensor lora_out_35_pad_0 = const()[name = tensor("lora_out_35_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_35_dilations_0 = const()[name = tensor("lora_out_35_dilations_0"), val = tensor([1, 1])]; tensor lora_out_35_groups_0 = const()[name = tensor("lora_out_35_groups_0"), val = tensor(1)]; tensor layers_1_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_1_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155345792)))]; tensor lora_out_35_cast_fp16 = conv(dilations = lora_out_35_dilations_0, groups = lora_out_35_groups_0, pad = lora_out_35_pad_0, pad_type = lora_out_35_pad_type_0, strides = lora_out_35_strides_0, weight = layers_1_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("lora_out_35_cast_fp16")]; tensor obj_25_cast_fp16 = add(x = pretrained_out_35_cast_fp16, y = lora_out_35_cast_fp16)[name = tensor("obj_25_cast_fp16")]; tensor inputs_11_cast_fp16 = add(x = inputs_9_cast_fp16, y = obj_25_cast_fp16)[name = tensor("inputs_11_cast_fp16")]; tensor out_11_axes_0 = const()[name = tensor("out_11_axes_0"), val = tensor([1])]; tensor var_745_to_fp16 = const()[name = tensor("op_745_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_11_cast_fp16 = layer_norm(axes = out_11_axes_0, epsilon = var_745_to_fp16, x = inputs_11_cast_fp16)[name = tensor("out_11_cast_fp16")]; tensor input_51_gamma_0_to_fp16 = const()[name = tensor("input_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155386816)))]; tensor input_51_beta_0_to_fp16 = const()[name = tensor("input_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155389440)))]; tensor input_51_epsilon_0_to_fp16 = const()[name = tensor("input_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor input_51_cast_fp16 = batch_norm(beta = input_51_beta_0_to_fp16, epsilon = input_51_epsilon_0_to_fp16, gamma = input_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_11_cast_fp16)[name = tensor("input_51_cast_fp16")]; tensor pretrained_out_37_pad_type_0 = const()[name = tensor("pretrained_out_37_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_37_strides_0 = const()[name = tensor("pretrained_out_37_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_37_pad_0 = const()[name = tensor("pretrained_out_37_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_37_dilations_0 = const()[name = tensor("pretrained_out_37_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_37_groups_0 = const()[name = tensor("pretrained_out_37_groups_0"), val = tensor(1)]; tensor layers_1_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155392064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158668928))), name = tensor("layers_1_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_1_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158669056)))]; tensor pretrained_out_37_cast_fp16 = conv(bias = layers_1_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_37_dilations_0, groups = pretrained_out_37_groups_0, pad = pretrained_out_37_pad_0, pad_type = pretrained_out_37_pad_type_0, strides = pretrained_out_37_strides_0, weight = layers_1_fc1_pretrained_weight_to_fp16_palettized, x = input_51_cast_fp16)[name = tensor("pretrained_out_37_cast_fp16")]; tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1, 1])]; tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1, 1])]; tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; tensor layers_1_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_1_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158679360)))]; tensor input_53_cast_fp16 = conv(dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = layers_1_fc1_loraA_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; tensor lora_out_37_pad_type_0 = const()[name = tensor("lora_out_37_pad_type_0"), val = tensor("valid")]; tensor lora_out_37_strides_0 = const()[name = tensor("lora_out_37_strides_0"), val = tensor([1, 1])]; tensor lora_out_37_pad_0 = const()[name = tensor("lora_out_37_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_37_dilations_0 = const()[name = tensor("lora_out_37_dilations_0"), val = tensor([1, 1])]; tensor lora_out_37_groups_0 = const()[name = tensor("lora_out_37_groups_0"), val = tensor(1)]; tensor layers_1_fc1_loraB_weight_to_fp16 = const()[name = tensor("layers_1_fc1_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158720384)))]; tensor lora_out_37_cast_fp16 = conv(dilations = lora_out_37_dilations_0, groups = lora_out_37_groups_0, pad = lora_out_37_pad_0, pad_type = lora_out_37_pad_type_0, strides = lora_out_37_strides_0, weight = layers_1_fc1_loraB_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("lora_out_37_cast_fp16")]; tensor input_55_cast_fp16 = add(x = pretrained_out_37_cast_fp16, y = lora_out_37_cast_fp16)[name = tensor("input_55_cast_fp16")]; tensor input_57_mode_0 = const()[name = tensor("input_57_mode_0"), val = tensor("EXACT")]; tensor input_57_cast_fp16 = gelu(mode = input_57_mode_0, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; tensor pretrained_out_39_pad_type_0 = const()[name = tensor("pretrained_out_39_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_39_strides_0 = const()[name = tensor("pretrained_out_39_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_39_pad_0 = const()[name = tensor("pretrained_out_39_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_39_dilations_0 = const()[name = tensor("pretrained_out_39_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_39_groups_0 = const()[name = tensor("pretrained_out_39_groups_0"), val = tensor(1)]; tensor layers_1_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158884288))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162161152))), name = tensor("layers_1_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_1_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_1_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162161280)))]; tensor pretrained_out_39_cast_fp16 = conv(bias = layers_1_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_39_dilations_0, groups = pretrained_out_39_groups_0, pad = pretrained_out_39_pad_0, pad_type = pretrained_out_39_pad_type_0, strides = pretrained_out_39_strides_0, weight = layers_1_fc2_pretrained_weight_to_fp16_palettized, x = input_57_cast_fp16)[name = tensor("pretrained_out_39_cast_fp16")]; tensor input_59_pad_type_0 = const()[name = tensor("input_59_pad_type_0"), val = tensor("valid")]; tensor input_59_strides_0 = const()[name = tensor("input_59_strides_0"), val = tensor([1, 1])]; tensor input_59_pad_0 = const()[name = tensor("input_59_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_59_dilations_0 = const()[name = tensor("input_59_dilations_0"), val = tensor([1, 1])]; tensor input_59_groups_0 = const()[name = tensor("input_59_groups_0"), val = tensor(1)]; tensor layers_1_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_1_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162163904)))]; tensor input_59_cast_fp16 = conv(dilations = input_59_dilations_0, groups = input_59_groups_0, pad = input_59_pad_0, pad_type = input_59_pad_type_0, strides = input_59_strides_0, weight = layers_1_fc2_loraA_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; tensor lora_out_39_pad_type_0 = const()[name = tensor("lora_out_39_pad_type_0"), val = tensor("valid")]; tensor lora_out_39_strides_0 = const()[name = tensor("lora_out_39_strides_0"), val = tensor([1, 1])]; tensor lora_out_39_pad_0 = const()[name = tensor("lora_out_39_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_39_dilations_0 = const()[name = tensor("lora_out_39_dilations_0"), val = tensor([1, 1])]; tensor lora_out_39_groups_0 = const()[name = tensor("lora_out_39_groups_0"), val = tensor(1)]; tensor layers_1_fc2_loraB_weight_to_fp16 = const()[name = tensor("layers_1_fc2_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162327808)))]; tensor lora_out_39_cast_fp16 = conv(dilations = lora_out_39_dilations_0, groups = lora_out_39_groups_0, pad = lora_out_39_pad_0, pad_type = lora_out_39_pad_type_0, strides = lora_out_39_strides_0, weight = layers_1_fc2_loraB_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("lora_out_39_cast_fp16")]; tensor hidden_states_5_cast_fp16 = add(x = pretrained_out_39_cast_fp16, y = lora_out_39_cast_fp16)[name = tensor("hidden_states_5_cast_fp16")]; tensor inputs_13_cast_fp16 = add(x = inputs_11_cast_fp16, y = hidden_states_5_cast_fp16)[name = tensor("inputs_13_cast_fp16")]; tensor var_812 = const()[name = tensor("op_812"), val = tensor(3)]; tensor out_13_axes_0 = const()[name = tensor("out_13_axes_0"), val = tensor([1])]; tensor var_838_to_fp16 = const()[name = tensor("op_838_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_13_cast_fp16 = layer_norm(axes = out_13_axes_0, epsilon = var_838_to_fp16, x = inputs_13_cast_fp16)[name = tensor("out_13_cast_fp16")]; tensor obj_29_gamma_0_to_fp16 = const()[name = tensor("obj_29_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162368832)))]; tensor obj_29_beta_0_to_fp16 = const()[name = tensor("obj_29_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162371456)))]; tensor obj_29_epsilon_0_to_fp16 = const()[name = tensor("obj_29_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_29_cast_fp16 = batch_norm(beta = obj_29_beta_0_to_fp16, epsilon = obj_29_epsilon_0_to_fp16, gamma = obj_29_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_13_cast_fp16)[name = tensor("obj_29_cast_fp16")]; tensor pretrained_out_41_pad_type_0 = const()[name = tensor("pretrained_out_41_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_41_strides_0 = const()[name = tensor("pretrained_out_41_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_41_pad_0 = const()[name = tensor("pretrained_out_41_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_41_dilations_0 = const()[name = tensor("pretrained_out_41_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_41_groups_0 = const()[name = tensor("pretrained_out_41_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(162374080))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163193344))), name = tensor("layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163193472)))]; tensor pretrained_out_41_cast_fp16 = conv(bias = layers_2_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_41_dilations_0, groups = pretrained_out_41_groups_0, pad = pretrained_out_41_pad_0, pad_type = pretrained_out_41_pad_type_0, strides = pretrained_out_41_strides_0, weight = layers_2_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_41_cast_fp16")]; tensor input_61_pad_type_0 = const()[name = tensor("input_61_pad_type_0"), val = tensor("valid")]; tensor input_61_strides_0 = const()[name = tensor("input_61_strides_0"), val = tensor([1, 1])]; tensor input_61_pad_0 = const()[name = tensor("input_61_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_61_dilations_0 = const()[name = tensor("input_61_dilations_0"), val = tensor([1, 1])]; tensor input_61_groups_0 = const()[name = tensor("input_61_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163196096)))]; tensor input_61_cast_fp16 = conv(dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = layers_2_self_attn_q_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_61_cast_fp16")]; tensor lora_out_41_pad_type_0 = const()[name = tensor("lora_out_41_pad_type_0"), val = tensor("valid")]; tensor lora_out_41_strides_0 = const()[name = tensor("lora_out_41_strides_0"), val = tensor([1, 1])]; tensor lora_out_41_pad_0 = const()[name = tensor("lora_out_41_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_41_dilations_0 = const()[name = tensor("lora_out_41_dilations_0"), val = tensor([1, 1])]; tensor lora_out_41_groups_0 = const()[name = tensor("lora_out_41_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163237120)))]; tensor lora_out_41_cast_fp16 = conv(dilations = lora_out_41_dilations_0, groups = lora_out_41_groups_0, pad = lora_out_41_pad_0, pad_type = lora_out_41_pad_type_0, strides = lora_out_41_strides_0, weight = layers_2_self_attn_q_proj_loraB_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("lora_out_41_cast_fp16")]; tensor query_9_cast_fp16 = add(x = pretrained_out_41_cast_fp16, y = lora_out_41_cast_fp16)[name = tensor("query_9_cast_fp16")]; tensor pretrained_out_43_pad_type_0 = const()[name = tensor("pretrained_out_43_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_43_strides_0 = const()[name = tensor("pretrained_out_43_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_43_pad_0 = const()[name = tensor("pretrained_out_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_43_dilations_0 = const()[name = tensor("pretrained_out_43_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_43_groups_0 = const()[name = tensor("pretrained_out_43_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163278144))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164097408))), name = tensor("layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_43_cast_fp16 = conv(dilations = pretrained_out_43_dilations_0, groups = pretrained_out_43_groups_0, pad = pretrained_out_43_pad_0, pad_type = pretrained_out_43_pad_type_0, strides = pretrained_out_43_strides_0, weight = layers_2_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_43_cast_fp16")]; tensor input_63_pad_type_0 = const()[name = tensor("input_63_pad_type_0"), val = tensor("valid")]; tensor input_63_strides_0 = const()[name = tensor("input_63_strides_0"), val = tensor([1, 1])]; tensor input_63_pad_0 = const()[name = tensor("input_63_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_63_dilations_0 = const()[name = tensor("input_63_dilations_0"), val = tensor([1, 1])]; tensor input_63_groups_0 = const()[name = tensor("input_63_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164097536)))]; tensor input_63_cast_fp16 = conv(dilations = input_63_dilations_0, groups = input_63_groups_0, pad = input_63_pad_0, pad_type = input_63_pad_type_0, strides = input_63_strides_0, weight = layers_2_self_attn_k_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_63_cast_fp16")]; tensor lora_out_43_pad_type_0 = const()[name = tensor("lora_out_43_pad_type_0"), val = tensor("valid")]; tensor lora_out_43_strides_0 = const()[name = tensor("lora_out_43_strides_0"), val = tensor([1, 1])]; tensor lora_out_43_pad_0 = const()[name = tensor("lora_out_43_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_43_dilations_0 = const()[name = tensor("lora_out_43_dilations_0"), val = tensor([1, 1])]; tensor lora_out_43_groups_0 = const()[name = tensor("lora_out_43_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164138560)))]; tensor lora_out_43_cast_fp16 = conv(dilations = lora_out_43_dilations_0, groups = lora_out_43_groups_0, pad = lora_out_43_pad_0, pad_type = lora_out_43_pad_type_0, strides = lora_out_43_strides_0, weight = layers_2_self_attn_k_proj_loraB_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("lora_out_43_cast_fp16")]; tensor current_key_5_cast_fp16 = add(x = pretrained_out_43_cast_fp16, y = lora_out_43_cast_fp16)[name = tensor("current_key_5_cast_fp16")]; tensor pretrained_out_45_pad_type_0 = const()[name = tensor("pretrained_out_45_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_45_strides_0 = const()[name = tensor("pretrained_out_45_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_45_pad_0 = const()[name = tensor("pretrained_out_45_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_45_dilations_0 = const()[name = tensor("pretrained_out_45_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_45_groups_0 = const()[name = tensor("pretrained_out_45_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164179584))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164998848))), name = tensor("layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(164998976)))]; tensor pretrained_out_45_cast_fp16 = conv(bias = layers_2_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_45_dilations_0, groups = pretrained_out_45_groups_0, pad = pretrained_out_45_pad_0, pad_type = pretrained_out_45_pad_type_0, strides = pretrained_out_45_strides_0, weight = layers_2_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_29_cast_fp16)[name = tensor("pretrained_out_45_cast_fp16")]; tensor input_65_pad_type_0 = const()[name = tensor("input_65_pad_type_0"), val = tensor("valid")]; tensor input_65_strides_0 = const()[name = tensor("input_65_strides_0"), val = tensor([1, 1])]; tensor input_65_pad_0 = const()[name = tensor("input_65_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_65_dilations_0 = const()[name = tensor("input_65_dilations_0"), val = tensor([1, 1])]; tensor input_65_groups_0 = const()[name = tensor("input_65_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165001600)))]; tensor input_65_cast_fp16 = conv(dilations = input_65_dilations_0, groups = input_65_groups_0, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = input_65_strides_0, weight = layers_2_self_attn_v_proj_loraA_weight_to_fp16, x = obj_29_cast_fp16)[name = tensor("input_65_cast_fp16")]; tensor lora_out_45_pad_type_0 = const()[name = tensor("lora_out_45_pad_type_0"), val = tensor("valid")]; tensor lora_out_45_strides_0 = const()[name = tensor("lora_out_45_strides_0"), val = tensor([1, 1])]; tensor lora_out_45_pad_0 = const()[name = tensor("lora_out_45_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_45_dilations_0 = const()[name = tensor("lora_out_45_dilations_0"), val = tensor([1, 1])]; tensor lora_out_45_groups_0 = const()[name = tensor("lora_out_45_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165042624)))]; tensor lora_out_45_cast_fp16 = conv(dilations = lora_out_45_dilations_0, groups = lora_out_45_groups_0, pad = lora_out_45_pad_0, pad_type = lora_out_45_pad_type_0, strides = lora_out_45_strides_0, weight = layers_2_self_attn_v_proj_loraB_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("lora_out_45_cast_fp16")]; tensor current_value_5_cast_fp16 = add(x = pretrained_out_45_cast_fp16, y = lora_out_45_cast_fp16)[name = tensor("current_value_5_cast_fp16")]; tensor var_924_cast_fp16 = mul(x = current_key_5_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_924_cast_fp16")]; tensor var_926_cast_fp16 = mul(x = var_47_cast_fp16_2, y = var_177_cast_fp16)[name = tensor("op_926_cast_fp16")]; tensor key_9_cast_fp16 = add(x = var_924_cast_fp16, y = var_926_cast_fp16)[name = tensor("key_9_cast_fp16")]; tensor var_928_cast_fp16 = mul(x = current_value_5_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_928_cast_fp16")]; tensor var_930_cast_fp16 = mul(x = var_54_cast_fp16_2, y = var_177_cast_fp16)[name = tensor("op_930_cast_fp16")]; tensor value_9_cast_fp16 = add(x = var_928_cast_fp16, y = var_930_cast_fp16)[name = tensor("value_9_cast_fp16")]; tensor var_933 = const()[name = tensor("op_933"), val = tensor([1, 20, 64, -1])]; tensor mh_q_9_cast_fp16 = reshape(shape = var_933, x = query_9_cast_fp16)[name = tensor("mh_q_9_cast_fp16")]; tensor var_935_to_fp16 = const()[name = tensor("op_935_to_fp16"), val = tensor(0x1p-3)]; tensor var_936_cast_fp16 = mul(x = mh_q_9_cast_fp16, y = var_935_to_fp16)[name = tensor("op_936_cast_fp16")]; tensor var_937 = const()[name = tensor("op_937"), val = tensor([1, 20, 64, -1])]; tensor var_938_cast_fp16 = reshape(shape = var_937, x = key_9_cast_fp16)[name = tensor("op_938_cast_fp16")]; tensor mh_w_13_transpose_x_0 = const()[name = tensor("mh_w_13_transpose_x_0"), val = tensor(true)]; tensor mh_w_13_transpose_y_0 = const()[name = tensor("mh_w_13_transpose_y_0"), val = tensor(false)]; tensor mh_w_13_cast_fp16 = matmul(transpose_x = mh_w_13_transpose_x_0, transpose_y = mh_w_13_transpose_y_0, x = var_936_cast_fp16, y = var_938_cast_fp16)[name = tensor("mh_w_13_cast_fp16")]; tensor mh_w_15_cast_fp16 = add(x = mh_w_13_cast_fp16, y = var_195_cast_fp16)[name = tensor("mh_w_15_cast_fp16")]; tensor var_946_cast_fp16 = softmax(axis = var_812, x = mh_w_15_cast_fp16)[name = tensor("op_946_cast_fp16")]; tensor var_947 = const()[name = tensor("op_947"), val = tensor([1, 20, 64, -1])]; tensor var_948_cast_fp16 = reshape(shape = var_947, x = value_9_cast_fp16)[name = tensor("op_948_cast_fp16")]; tensor attn_9_transpose_x_0 = const()[name = tensor("attn_9_transpose_x_0"), val = tensor(false)]; tensor attn_9_transpose_y_0 = const()[name = tensor("attn_9_transpose_y_0"), val = tensor(true)]; tensor attn_9_cast_fp16 = matmul(transpose_x = attn_9_transpose_x_0, transpose_y = attn_9_transpose_y_0, x = var_948_cast_fp16, y = var_946_cast_fp16)[name = tensor("attn_9_cast_fp16")]; tensor var_951 = const()[name = tensor("op_951"), val = tensor([1, 1280, 1, -1])]; tensor input_67_cast_fp16 = reshape(shape = var_951, x = attn_9_cast_fp16)[name = tensor("input_67_cast_fp16")]; tensor pretrained_out_47_pad_type_0 = const()[name = tensor("pretrained_out_47_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_47_strides_0 = const()[name = tensor("pretrained_out_47_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_47_pad_0 = const()[name = tensor("pretrained_out_47_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_47_dilations_0 = const()[name = tensor("pretrained_out_47_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_47_groups_0 = const()[name = tensor("pretrained_out_47_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165083648))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165902912))), name = tensor("layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165903040)))]; tensor pretrained_out_47_cast_fp16 = conv(bias = layers_2_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_47_dilations_0, groups = pretrained_out_47_groups_0, pad = pretrained_out_47_pad_0, pad_type = pretrained_out_47_pad_type_0, strides = pretrained_out_47_strides_0, weight = layers_2_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_67_cast_fp16)[name = tensor("pretrained_out_47_cast_fp16")]; tensor input_69_pad_type_0 = const()[name = tensor("input_69_pad_type_0"), val = tensor("valid")]; tensor input_69_strides_0 = const()[name = tensor("input_69_strides_0"), val = tensor([1, 1])]; tensor input_69_pad_0 = const()[name = tensor("input_69_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_69_dilations_0 = const()[name = tensor("input_69_dilations_0"), val = tensor([1, 1])]; tensor input_69_groups_0 = const()[name = tensor("input_69_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165905664)))]; tensor input_69_cast_fp16 = conv(dilations = input_69_dilations_0, groups = input_69_groups_0, pad = input_69_pad_0, pad_type = input_69_pad_type_0, strides = input_69_strides_0, weight = layers_2_self_attn_o_proj_loraA_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; tensor lora_out_47_pad_type_0 = const()[name = tensor("lora_out_47_pad_type_0"), val = tensor("valid")]; tensor lora_out_47_strides_0 = const()[name = tensor("lora_out_47_strides_0"), val = tensor([1, 1])]; tensor lora_out_47_pad_0 = const()[name = tensor("lora_out_47_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_47_dilations_0 = const()[name = tensor("lora_out_47_dilations_0"), val = tensor([1, 1])]; tensor lora_out_47_groups_0 = const()[name = tensor("lora_out_47_groups_0"), val = tensor(1)]; tensor layers_2_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165946688)))]; tensor lora_out_47_cast_fp16 = conv(dilations = lora_out_47_dilations_0, groups = lora_out_47_groups_0, pad = lora_out_47_pad_0, pad_type = lora_out_47_pad_type_0, strides = lora_out_47_strides_0, weight = layers_2_self_attn_o_proj_loraB_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("lora_out_47_cast_fp16")]; tensor obj_35_cast_fp16 = add(x = pretrained_out_47_cast_fp16, y = lora_out_47_cast_fp16)[name = tensor("obj_35_cast_fp16")]; tensor inputs_15_cast_fp16 = add(x = inputs_13_cast_fp16, y = obj_35_cast_fp16)[name = tensor("inputs_15_cast_fp16")]; tensor out_15_axes_0 = const()[name = tensor("out_15_axes_0"), val = tensor([1])]; tensor var_989_to_fp16 = const()[name = tensor("op_989_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_15_cast_fp16 = layer_norm(axes = out_15_axes_0, epsilon = var_989_to_fp16, x = inputs_15_cast_fp16)[name = tensor("out_15_cast_fp16")]; tensor obj_37_gamma_0_to_fp16 = const()[name = tensor("obj_37_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165987712)))]; tensor obj_37_beta_0_to_fp16 = const()[name = tensor("obj_37_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165990336)))]; tensor obj_37_epsilon_0_to_fp16 = const()[name = tensor("obj_37_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_37_cast_fp16 = batch_norm(beta = obj_37_beta_0_to_fp16, epsilon = obj_37_epsilon_0_to_fp16, gamma = obj_37_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_15_cast_fp16)[name = tensor("obj_37_cast_fp16")]; tensor pretrained_out_49_pad_type_0 = const()[name = tensor("pretrained_out_49_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_49_strides_0 = const()[name = tensor("pretrained_out_49_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_49_pad_0 = const()[name = tensor("pretrained_out_49_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_49_dilations_0 = const()[name = tensor("pretrained_out_49_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_49_groups_0 = const()[name = tensor("pretrained_out_49_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(165992960))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166812224))), name = tensor("layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166812352)))]; tensor pretrained_out_49_cast_fp16 = conv(bias = layers_2_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_49_dilations_0, groups = pretrained_out_49_groups_0, pad = pretrained_out_49_pad_0, pad_type = pretrained_out_49_pad_type_0, strides = pretrained_out_49_strides_0, weight = layers_2_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_37_cast_fp16)[name = tensor("pretrained_out_49_cast_fp16")]; tensor input_71_pad_type_0 = const()[name = tensor("input_71_pad_type_0"), val = tensor("valid")]; tensor input_71_strides_0 = const()[name = tensor("input_71_strides_0"), val = tensor([1, 1])]; tensor input_71_pad_0 = const()[name = tensor("input_71_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_71_dilations_0 = const()[name = tensor("input_71_dilations_0"), val = tensor([1, 1])]; tensor input_71_groups_0 = const()[name = tensor("input_71_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166814976)))]; tensor input_71_cast_fp16 = conv(dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = layers_2_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_37_cast_fp16)[name = tensor("input_71_cast_fp16")]; tensor lora_out_49_pad_type_0 = const()[name = tensor("lora_out_49_pad_type_0"), val = tensor("valid")]; tensor lora_out_49_strides_0 = const()[name = tensor("lora_out_49_strides_0"), val = tensor([1, 1])]; tensor lora_out_49_pad_0 = const()[name = tensor("lora_out_49_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_49_dilations_0 = const()[name = tensor("lora_out_49_dilations_0"), val = tensor([1, 1])]; tensor lora_out_49_groups_0 = const()[name = tensor("lora_out_49_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166856000)))]; tensor lora_out_49_cast_fp16 = conv(dilations = lora_out_49_dilations_0, groups = lora_out_49_groups_0, pad = lora_out_49_pad_0, pad_type = lora_out_49_pad_type_0, strides = lora_out_49_strides_0, weight = layers_2_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("lora_out_49_cast_fp16")]; tensor query_11_cast_fp16 = add(x = pretrained_out_49_cast_fp16, y = lora_out_49_cast_fp16)[name = tensor("query_11_cast_fp16")]; tensor pretrained_out_51_pad_type_0 = const()[name = tensor("pretrained_out_51_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_51_strides_0 = const()[name = tensor("pretrained_out_51_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_51_pad_0 = const()[name = tensor("pretrained_out_51_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_51_dilations_0 = const()[name = tensor("pretrained_out_51_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_51_groups_0 = const()[name = tensor("pretrained_out_51_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166897024))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167716288))), name = tensor("layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_51_cast_fp16 = conv(dilations = pretrained_out_51_dilations_0, groups = pretrained_out_51_groups_0, pad = pretrained_out_51_pad_0, pad_type = pretrained_out_51_pad_type_0, strides = pretrained_out_51_strides_0, weight = layers_2_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_51_cast_fp16")]; tensor input_73_pad_type_0 = const()[name = tensor("input_73_pad_type_0"), val = tensor("valid")]; tensor input_73_strides_0 = const()[name = tensor("input_73_strides_0"), val = tensor([1, 1])]; tensor input_73_pad_0 = const()[name = tensor("input_73_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_73_dilations_0 = const()[name = tensor("input_73_dilations_0"), val = tensor([1, 1])]; tensor input_73_groups_0 = const()[name = tensor("input_73_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167716416)))]; tensor input_73_cast_fp16 = conv(dilations = input_73_dilations_0, groups = input_73_groups_0, pad = input_73_pad_0, pad_type = input_73_pad_type_0, strides = input_73_strides_0, weight = layers_2_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_73_cast_fp16")]; tensor lora_out_51_pad_type_0 = const()[name = tensor("lora_out_51_pad_type_0"), val = tensor("valid")]; tensor lora_out_51_strides_0 = const()[name = tensor("lora_out_51_strides_0"), val = tensor([1, 1])]; tensor lora_out_51_pad_0 = const()[name = tensor("lora_out_51_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_51_dilations_0 = const()[name = tensor("lora_out_51_dilations_0"), val = tensor([1, 1])]; tensor lora_out_51_groups_0 = const()[name = tensor("lora_out_51_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167757440)))]; tensor lora_out_51_cast_fp16 = conv(dilations = lora_out_51_dilations_0, groups = lora_out_51_groups_0, pad = lora_out_51_pad_0, pad_type = lora_out_51_pad_type_0, strides = lora_out_51_strides_0, weight = layers_2_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("lora_out_51_cast_fp16")]; tensor key_11_cast_fp16 = add(x = pretrained_out_51_cast_fp16, y = lora_out_51_cast_fp16)[name = tensor("key_11_cast_fp16")]; tensor pretrained_out_53_pad_type_0 = const()[name = tensor("pretrained_out_53_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_53_strides_0 = const()[name = tensor("pretrained_out_53_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_53_pad_0 = const()[name = tensor("pretrained_out_53_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_53_dilations_0 = const()[name = tensor("pretrained_out_53_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_53_groups_0 = const()[name = tensor("pretrained_out_53_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(167798464))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168617728))), name = tensor("layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168617856)))]; tensor pretrained_out_53_cast_fp16 = conv(bias = layers_2_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_53_dilations_0, groups = pretrained_out_53_groups_0, pad = pretrained_out_53_pad_0, pad_type = pretrained_out_53_pad_type_0, strides = pretrained_out_53_strides_0, weight = layers_2_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_53_cast_fp16")]; tensor input_75_pad_type_0 = const()[name = tensor("input_75_pad_type_0"), val = tensor("valid")]; tensor input_75_strides_0 = const()[name = tensor("input_75_strides_0"), val = tensor([1, 1])]; tensor input_75_pad_0 = const()[name = tensor("input_75_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_75_dilations_0 = const()[name = tensor("input_75_dilations_0"), val = tensor([1, 1])]; tensor input_75_groups_0 = const()[name = tensor("input_75_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168620480)))]; tensor input_75_cast_fp16 = conv(dilations = input_75_dilations_0, groups = input_75_groups_0, pad = input_75_pad_0, pad_type = input_75_pad_type_0, strides = input_75_strides_0, weight = layers_2_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_75_cast_fp16")]; tensor lora_out_53_pad_type_0 = const()[name = tensor("lora_out_53_pad_type_0"), val = tensor("valid")]; tensor lora_out_53_strides_0 = const()[name = tensor("lora_out_53_strides_0"), val = tensor([1, 1])]; tensor lora_out_53_pad_0 = const()[name = tensor("lora_out_53_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_53_dilations_0 = const()[name = tensor("lora_out_53_dilations_0"), val = tensor([1, 1])]; tensor lora_out_53_groups_0 = const()[name = tensor("lora_out_53_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168661504)))]; tensor lora_out_53_cast_fp16 = conv(dilations = lora_out_53_dilations_0, groups = lora_out_53_groups_0, pad = lora_out_53_pad_0, pad_type = lora_out_53_pad_type_0, strides = lora_out_53_strides_0, weight = layers_2_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("lora_out_53_cast_fp16")]; tensor value_11_cast_fp16 = add(x = pretrained_out_53_cast_fp16, y = lora_out_53_cast_fp16)[name = tensor("value_11_cast_fp16")]; tensor var_1072 = const()[name = tensor("op_1072"), val = tensor([1, 20, 64, -1])]; tensor mh_q_11_cast_fp16 = reshape(shape = var_1072, x = query_11_cast_fp16)[name = tensor("mh_q_11_cast_fp16")]; tensor var_1074_to_fp16 = const()[name = tensor("op_1074_to_fp16"), val = tensor(0x1p-3)]; tensor var_1075_cast_fp16 = mul(x = mh_q_11_cast_fp16, y = var_1074_to_fp16)[name = tensor("op_1075_cast_fp16")]; tensor var_1076 = const()[name = tensor("op_1076"), val = tensor([1, 20, 64, -1])]; tensor var_1077_cast_fp16 = reshape(shape = var_1076, x = key_11_cast_fp16)[name = tensor("op_1077_cast_fp16")]; tensor mh_w_17_transpose_x_0 = const()[name = tensor("mh_w_17_transpose_x_0"), val = tensor(true)]; tensor mh_w_17_transpose_y_0 = const()[name = tensor("mh_w_17_transpose_y_0"), val = tensor(false)]; tensor mh_w_17_cast_fp16 = matmul(transpose_x = mh_w_17_transpose_x_0, transpose_y = mh_w_17_transpose_y_0, x = var_1075_cast_fp16, y = var_1077_cast_fp16)[name = tensor("mh_w_17_cast_fp16")]; tensor obj_41_cast_fp16 = softmax(axis = var_812, x = mh_w_17_cast_fp16)[name = tensor("obj_41_cast_fp16")]; tensor var_1081 = const()[name = tensor("op_1081"), val = tensor([1, 20, 64, -1])]; tensor var_1082_cast_fp16 = reshape(shape = var_1081, x = value_11_cast_fp16)[name = tensor("op_1082_cast_fp16")]; tensor attn_11_transpose_x_0 = const()[name = tensor("attn_11_transpose_x_0"), val = tensor(false)]; tensor attn_11_transpose_y_0 = const()[name = tensor("attn_11_transpose_y_0"), val = tensor(true)]; tensor attn_11_cast_fp16 = matmul(transpose_x = attn_11_transpose_x_0, transpose_y = attn_11_transpose_y_0, x = var_1082_cast_fp16, y = obj_41_cast_fp16)[name = tensor("attn_11_cast_fp16")]; tensor var_1085 = const()[name = tensor("op_1085"), val = tensor([1, 1280, 1, -1])]; tensor input_77_cast_fp16 = reshape(shape = var_1085, x = attn_11_cast_fp16)[name = tensor("input_77_cast_fp16")]; tensor pretrained_out_55_pad_type_0 = const()[name = tensor("pretrained_out_55_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_55_strides_0 = const()[name = tensor("pretrained_out_55_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_55_pad_0 = const()[name = tensor("pretrained_out_55_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_55_dilations_0 = const()[name = tensor("pretrained_out_55_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_55_groups_0 = const()[name = tensor("pretrained_out_55_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168702528))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169521792))), name = tensor("layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169521920)))]; tensor pretrained_out_55_cast_fp16 = conv(bias = layers_2_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_55_dilations_0, groups = pretrained_out_55_groups_0, pad = pretrained_out_55_pad_0, pad_type = pretrained_out_55_pad_type_0, strides = pretrained_out_55_strides_0, weight = layers_2_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_77_cast_fp16)[name = tensor("pretrained_out_55_cast_fp16")]; tensor input_79_pad_type_0 = const()[name = tensor("input_79_pad_type_0"), val = tensor("valid")]; tensor input_79_strides_0 = const()[name = tensor("input_79_strides_0"), val = tensor([1, 1])]; tensor input_79_pad_0 = const()[name = tensor("input_79_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_79_dilations_0 = const()[name = tensor("input_79_dilations_0"), val = tensor([1, 1])]; tensor input_79_groups_0 = const()[name = tensor("input_79_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169524544)))]; tensor input_79_cast_fp16 = conv(dilations = input_79_dilations_0, groups = input_79_groups_0, pad = input_79_pad_0, pad_type = input_79_pad_type_0, strides = input_79_strides_0, weight = layers_2_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; tensor lora_out_55_pad_type_0 = const()[name = tensor("lora_out_55_pad_type_0"), val = tensor("valid")]; tensor lora_out_55_strides_0 = const()[name = tensor("lora_out_55_strides_0"), val = tensor([1, 1])]; tensor lora_out_55_pad_0 = const()[name = tensor("lora_out_55_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_55_dilations_0 = const()[name = tensor("lora_out_55_dilations_0"), val = tensor([1, 1])]; tensor lora_out_55_groups_0 = const()[name = tensor("lora_out_55_groups_0"), val = tensor(1)]; tensor layers_2_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_2_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169565568)))]; tensor lora_out_55_cast_fp16 = conv(dilations = lora_out_55_dilations_0, groups = lora_out_55_groups_0, pad = lora_out_55_pad_0, pad_type = lora_out_55_pad_type_0, strides = lora_out_55_strides_0, weight = layers_2_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("lora_out_55_cast_fp16")]; tensor obj_39_cast_fp16 = add(x = pretrained_out_55_cast_fp16, y = lora_out_55_cast_fp16)[name = tensor("obj_39_cast_fp16")]; tensor inputs_17_cast_fp16 = add(x = inputs_15_cast_fp16, y = obj_39_cast_fp16)[name = tensor("inputs_17_cast_fp16")]; tensor out_17_axes_0 = const()[name = tensor("out_17_axes_0"), val = tensor([1])]; tensor var_1122_to_fp16 = const()[name = tensor("op_1122_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_17_cast_fp16 = layer_norm(axes = out_17_axes_0, epsilon = var_1122_to_fp16, x = inputs_17_cast_fp16)[name = tensor("out_17_cast_fp16")]; tensor input_81_gamma_0_to_fp16 = const()[name = tensor("input_81_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169606592)))]; tensor input_81_beta_0_to_fp16 = const()[name = tensor("input_81_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169609216)))]; tensor input_81_epsilon_0_to_fp16 = const()[name = tensor("input_81_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor input_81_cast_fp16 = batch_norm(beta = input_81_beta_0_to_fp16, epsilon = input_81_epsilon_0_to_fp16, gamma = input_81_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_17_cast_fp16)[name = tensor("input_81_cast_fp16")]; tensor pretrained_out_57_pad_type_0 = const()[name = tensor("pretrained_out_57_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_57_strides_0 = const()[name = tensor("pretrained_out_57_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_57_pad_0 = const()[name = tensor("pretrained_out_57_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_57_dilations_0 = const()[name = tensor("pretrained_out_57_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_57_groups_0 = const()[name = tensor("pretrained_out_57_groups_0"), val = tensor(1)]; tensor layers_2_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(169611840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172888704))), name = tensor("layers_2_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_2_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172888832)))]; tensor pretrained_out_57_cast_fp16 = conv(bias = layers_2_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_57_dilations_0, groups = pretrained_out_57_groups_0, pad = pretrained_out_57_pad_0, pad_type = pretrained_out_57_pad_type_0, strides = pretrained_out_57_strides_0, weight = layers_2_fc1_pretrained_weight_to_fp16_palettized, x = input_81_cast_fp16)[name = tensor("pretrained_out_57_cast_fp16")]; tensor input_83_pad_type_0 = const()[name = tensor("input_83_pad_type_0"), val = tensor("valid")]; tensor input_83_strides_0 = const()[name = tensor("input_83_strides_0"), val = tensor([1, 1])]; tensor input_83_pad_0 = const()[name = tensor("input_83_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_83_dilations_0 = const()[name = tensor("input_83_dilations_0"), val = tensor([1, 1])]; tensor input_83_groups_0 = const()[name = tensor("input_83_groups_0"), val = tensor(1)]; tensor layers_2_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_2_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172899136)))]; tensor input_83_cast_fp16 = conv(dilations = input_83_dilations_0, groups = input_83_groups_0, pad = input_83_pad_0, pad_type = input_83_pad_type_0, strides = input_83_strides_0, weight = layers_2_fc1_loraA_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; tensor lora_out_57_pad_type_0 = const()[name = tensor("lora_out_57_pad_type_0"), val = tensor("valid")]; tensor lora_out_57_strides_0 = const()[name = tensor("lora_out_57_strides_0"), val = tensor([1, 1])]; tensor lora_out_57_pad_0 = const()[name = tensor("lora_out_57_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_57_dilations_0 = const()[name = tensor("lora_out_57_dilations_0"), val = tensor([1, 1])]; tensor lora_out_57_groups_0 = const()[name = tensor("lora_out_57_groups_0"), val = tensor(1)]; tensor layers_2_fc1_loraB_weight_to_fp16 = const()[name = tensor("layers_2_fc1_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172940160)))]; tensor lora_out_57_cast_fp16 = conv(dilations = lora_out_57_dilations_0, groups = lora_out_57_groups_0, pad = lora_out_57_pad_0, pad_type = lora_out_57_pad_type_0, strides = lora_out_57_strides_0, weight = layers_2_fc1_loraB_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("lora_out_57_cast_fp16")]; tensor input_85_cast_fp16 = add(x = pretrained_out_57_cast_fp16, y = lora_out_57_cast_fp16)[name = tensor("input_85_cast_fp16")]; tensor input_87_mode_0 = const()[name = tensor("input_87_mode_0"), val = tensor("EXACT")]; tensor input_87_cast_fp16 = gelu(mode = input_87_mode_0, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; tensor pretrained_out_59_pad_type_0 = const()[name = tensor("pretrained_out_59_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_59_strides_0 = const()[name = tensor("pretrained_out_59_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_59_pad_0 = const()[name = tensor("pretrained_out_59_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_59_dilations_0 = const()[name = tensor("pretrained_out_59_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_59_groups_0 = const()[name = tensor("pretrained_out_59_groups_0"), val = tensor(1)]; tensor layers_2_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(173104064))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176380928))), name = tensor("layers_2_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_2_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_2_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176381056)))]; tensor pretrained_out_59_cast_fp16 = conv(bias = layers_2_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_59_dilations_0, groups = pretrained_out_59_groups_0, pad = pretrained_out_59_pad_0, pad_type = pretrained_out_59_pad_type_0, strides = pretrained_out_59_strides_0, weight = layers_2_fc2_pretrained_weight_to_fp16_palettized, x = input_87_cast_fp16)[name = tensor("pretrained_out_59_cast_fp16")]; tensor input_89_pad_type_0 = const()[name = tensor("input_89_pad_type_0"), val = tensor("valid")]; tensor input_89_strides_0 = const()[name = tensor("input_89_strides_0"), val = tensor([1, 1])]; tensor input_89_pad_0 = const()[name = tensor("input_89_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_89_dilations_0 = const()[name = tensor("input_89_dilations_0"), val = tensor([1, 1])]; tensor input_89_groups_0 = const()[name = tensor("input_89_groups_0"), val = tensor(1)]; tensor layers_2_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_2_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176383680)))]; tensor input_89_cast_fp16 = conv(dilations = input_89_dilations_0, groups = input_89_groups_0, pad = input_89_pad_0, pad_type = input_89_pad_type_0, strides = input_89_strides_0, weight = layers_2_fc2_loraA_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; tensor lora_out_59_pad_type_0 = const()[name = tensor("lora_out_59_pad_type_0"), val = tensor("valid")]; tensor lora_out_59_strides_0 = const()[name = tensor("lora_out_59_strides_0"), val = tensor([1, 1])]; tensor lora_out_59_pad_0 = const()[name = tensor("lora_out_59_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_59_dilations_0 = const()[name = tensor("lora_out_59_dilations_0"), val = tensor([1, 1])]; tensor lora_out_59_groups_0 = const()[name = tensor("lora_out_59_groups_0"), val = tensor(1)]; tensor layers_2_fc2_loraB_weight_to_fp16 = const()[name = tensor("layers_2_fc2_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176547584)))]; tensor lora_out_59_cast_fp16 = conv(dilations = lora_out_59_dilations_0, groups = lora_out_59_groups_0, pad = lora_out_59_pad_0, pad_type = lora_out_59_pad_type_0, strides = lora_out_59_strides_0, weight = layers_2_fc2_loraB_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("lora_out_59_cast_fp16")]; tensor hidden_states_7_cast_fp16 = add(x = pretrained_out_59_cast_fp16, y = lora_out_59_cast_fp16)[name = tensor("hidden_states_7_cast_fp16")]; tensor inputs_19_cast_fp16 = add(x = inputs_17_cast_fp16, y = hidden_states_7_cast_fp16)[name = tensor("inputs_19_cast_fp16")]; tensor var_1190 = const()[name = tensor("op_1190"), val = tensor(3)]; tensor out_19_axes_0 = const()[name = tensor("out_19_axes_0"), val = tensor([1])]; tensor var_1216_to_fp16 = const()[name = tensor("op_1216_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_19_cast_fp16 = layer_norm(axes = out_19_axes_0, epsilon = var_1216_to_fp16, x = inputs_19_cast_fp16)[name = tensor("out_19_cast_fp16")]; tensor obj_43_gamma_0_to_fp16 = const()[name = tensor("obj_43_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176588608)))]; tensor obj_43_beta_0_to_fp16 = const()[name = tensor("obj_43_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176591232)))]; tensor obj_43_epsilon_0_to_fp16 = const()[name = tensor("obj_43_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_43_cast_fp16 = batch_norm(beta = obj_43_beta_0_to_fp16, epsilon = obj_43_epsilon_0_to_fp16, gamma = obj_43_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_19_cast_fp16)[name = tensor("obj_43_cast_fp16")]; tensor pretrained_out_61_pad_type_0 = const()[name = tensor("pretrained_out_61_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_61_strides_0 = const()[name = tensor("pretrained_out_61_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_61_pad_0 = const()[name = tensor("pretrained_out_61_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_61_dilations_0 = const()[name = tensor("pretrained_out_61_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_61_groups_0 = const()[name = tensor("pretrained_out_61_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176593856))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177413120))), name = tensor("layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_self_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177413248)))]; tensor pretrained_out_61_cast_fp16 = conv(bias = layers_3_self_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_61_dilations_0, groups = pretrained_out_61_groups_0, pad = pretrained_out_61_pad_0, pad_type = pretrained_out_61_pad_type_0, strides = pretrained_out_61_strides_0, weight = layers_3_self_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("pretrained_out_61_cast_fp16")]; tensor input_91_pad_type_0 = const()[name = tensor("input_91_pad_type_0"), val = tensor("valid")]; tensor input_91_strides_0 = const()[name = tensor("input_91_strides_0"), val = tensor([1, 1])]; tensor input_91_pad_0 = const()[name = tensor("input_91_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_91_dilations_0 = const()[name = tensor("input_91_dilations_0"), val = tensor([1, 1])]; tensor input_91_groups_0 = const()[name = tensor("input_91_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177415872)))]; tensor input_91_cast_fp16 = conv(dilations = input_91_dilations_0, groups = input_91_groups_0, pad = input_91_pad_0, pad_type = input_91_pad_type_0, strides = input_91_strides_0, weight = layers_3_self_attn_q_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("input_91_cast_fp16")]; tensor lora_out_61_pad_type_0 = const()[name = tensor("lora_out_61_pad_type_0"), val = tensor("valid")]; tensor lora_out_61_strides_0 = const()[name = tensor("lora_out_61_strides_0"), val = tensor([1, 1])]; tensor lora_out_61_pad_0 = const()[name = tensor("lora_out_61_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_61_dilations_0 = const()[name = tensor("lora_out_61_dilations_0"), val = tensor([1, 1])]; tensor lora_out_61_groups_0 = const()[name = tensor("lora_out_61_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177456896)))]; tensor lora_out_61_cast_fp16 = conv(dilations = lora_out_61_dilations_0, groups = lora_out_61_groups_0, pad = lora_out_61_pad_0, pad_type = lora_out_61_pad_type_0, strides = lora_out_61_strides_0, weight = layers_3_self_attn_q_proj_loraB_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("lora_out_61_cast_fp16")]; tensor query_13_cast_fp16 = add(x = pretrained_out_61_cast_fp16, y = lora_out_61_cast_fp16)[name = tensor("query_13_cast_fp16")]; tensor pretrained_out_63_pad_type_0 = const()[name = tensor("pretrained_out_63_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_63_strides_0 = const()[name = tensor("pretrained_out_63_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_63_pad_0 = const()[name = tensor("pretrained_out_63_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_63_dilations_0 = const()[name = tensor("pretrained_out_63_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_63_groups_0 = const()[name = tensor("pretrained_out_63_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177497920))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178317184))), name = tensor("layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_63_cast_fp16 = conv(dilations = pretrained_out_63_dilations_0, groups = pretrained_out_63_groups_0, pad = pretrained_out_63_pad_0, pad_type = pretrained_out_63_pad_type_0, strides = pretrained_out_63_strides_0, weight = layers_3_self_attn_k_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("pretrained_out_63_cast_fp16")]; tensor input_93_pad_type_0 = const()[name = tensor("input_93_pad_type_0"), val = tensor("valid")]; tensor input_93_strides_0 = const()[name = tensor("input_93_strides_0"), val = tensor([1, 1])]; tensor input_93_pad_0 = const()[name = tensor("input_93_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_93_dilations_0 = const()[name = tensor("input_93_dilations_0"), val = tensor([1, 1])]; tensor input_93_groups_0 = const()[name = tensor("input_93_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178317312)))]; tensor input_93_cast_fp16 = conv(dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = layers_3_self_attn_k_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("input_93_cast_fp16")]; tensor lora_out_63_pad_type_0 = const()[name = tensor("lora_out_63_pad_type_0"), val = tensor("valid")]; tensor lora_out_63_strides_0 = const()[name = tensor("lora_out_63_strides_0"), val = tensor([1, 1])]; tensor lora_out_63_pad_0 = const()[name = tensor("lora_out_63_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_63_dilations_0 = const()[name = tensor("lora_out_63_dilations_0"), val = tensor([1, 1])]; tensor lora_out_63_groups_0 = const()[name = tensor("lora_out_63_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178358336)))]; tensor lora_out_63_cast_fp16 = conv(dilations = lora_out_63_dilations_0, groups = lora_out_63_groups_0, pad = lora_out_63_pad_0, pad_type = lora_out_63_pad_type_0, strides = lora_out_63_strides_0, weight = layers_3_self_attn_k_proj_loraB_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("lora_out_63_cast_fp16")]; tensor current_key_cast_fp16 = add(x = pretrained_out_63_cast_fp16, y = lora_out_63_cast_fp16)[name = tensor("current_key_cast_fp16")]; tensor pretrained_out_65_pad_type_0 = const()[name = tensor("pretrained_out_65_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_65_strides_0 = const()[name = tensor("pretrained_out_65_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_65_pad_0 = const()[name = tensor("pretrained_out_65_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_65_dilations_0 = const()[name = tensor("pretrained_out_65_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_65_groups_0 = const()[name = tensor("pretrained_out_65_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(178399360))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179218624))), name = tensor("layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_self_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179218752)))]; tensor pretrained_out_65_cast_fp16 = conv(bias = layers_3_self_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_65_dilations_0, groups = pretrained_out_65_groups_0, pad = pretrained_out_65_pad_0, pad_type = pretrained_out_65_pad_type_0, strides = pretrained_out_65_strides_0, weight = layers_3_self_attn_v_proj_pretrained_weight_to_fp16_palettized, x = obj_43_cast_fp16)[name = tensor("pretrained_out_65_cast_fp16")]; tensor input_95_pad_type_0 = const()[name = tensor("input_95_pad_type_0"), val = tensor("valid")]; tensor input_95_strides_0 = const()[name = tensor("input_95_strides_0"), val = tensor([1, 1])]; tensor input_95_pad_0 = const()[name = tensor("input_95_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_95_dilations_0 = const()[name = tensor("input_95_dilations_0"), val = tensor([1, 1])]; tensor input_95_groups_0 = const()[name = tensor("input_95_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179221376)))]; tensor input_95_cast_fp16 = conv(dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = layers_3_self_attn_v_proj_loraA_weight_to_fp16, x = obj_43_cast_fp16)[name = tensor("input_95_cast_fp16")]; tensor lora_out_65_pad_type_0 = const()[name = tensor("lora_out_65_pad_type_0"), val = tensor("valid")]; tensor lora_out_65_strides_0 = const()[name = tensor("lora_out_65_strides_0"), val = tensor([1, 1])]; tensor lora_out_65_pad_0 = const()[name = tensor("lora_out_65_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_65_dilations_0 = const()[name = tensor("lora_out_65_dilations_0"), val = tensor([1, 1])]; tensor lora_out_65_groups_0 = const()[name = tensor("lora_out_65_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179262400)))]; tensor lora_out_65_cast_fp16 = conv(dilations = lora_out_65_dilations_0, groups = lora_out_65_groups_0, pad = lora_out_65_pad_0, pad_type = lora_out_65_pad_type_0, strides = lora_out_65_strides_0, weight = layers_3_self_attn_v_proj_loraB_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("lora_out_65_cast_fp16")]; tensor current_value_cast_fp16 = add(x = pretrained_out_65_cast_fp16, y = lora_out_65_cast_fp16)[name = tensor("current_value_cast_fp16")]; tensor var_1302_cast_fp16 = mul(x = current_key_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_1302_cast_fp16")]; tensor var_1304_cast_fp16 = mul(x = var_47_cast_fp16_3, y = var_177_cast_fp16)[name = tensor("op_1304_cast_fp16")]; tensor key_13_cast_fp16 = add(x = var_1302_cast_fp16, y = var_1304_cast_fp16)[name = tensor("key_13_cast_fp16")]; tensor var_1306_cast_fp16 = mul(x = current_value_cast_fp16, y = var_174_cast_fp16)[name = tensor("op_1306_cast_fp16")]; tensor var_1308_cast_fp16 = mul(x = var_54_cast_fp16_3, y = var_177_cast_fp16)[name = tensor("op_1308_cast_fp16")]; tensor value_13_cast_fp16 = add(x = var_1306_cast_fp16, y = var_1308_cast_fp16)[name = tensor("value_13_cast_fp16")]; tensor var_1311 = const()[name = tensor("op_1311"), val = tensor([1, 20, 64, -1])]; tensor mh_q_13_cast_fp16 = reshape(shape = var_1311, x = query_13_cast_fp16)[name = tensor("mh_q_13_cast_fp16")]; tensor var_1313_to_fp16 = const()[name = tensor("op_1313_to_fp16"), val = tensor(0x1p-3)]; tensor var_1314_cast_fp16 = mul(x = mh_q_13_cast_fp16, y = var_1313_to_fp16)[name = tensor("op_1314_cast_fp16")]; tensor var_1315 = const()[name = tensor("op_1315"), val = tensor([1, 20, 64, -1])]; tensor var_1316_cast_fp16 = reshape(shape = var_1315, x = key_13_cast_fp16)[name = tensor("op_1316_cast_fp16")]; tensor mh_w_19_transpose_x_0 = const()[name = tensor("mh_w_19_transpose_x_0"), val = tensor(true)]; tensor mh_w_19_transpose_y_0 = const()[name = tensor("mh_w_19_transpose_y_0"), val = tensor(false)]; tensor mh_w_19_cast_fp16 = matmul(transpose_x = mh_w_19_transpose_x_0, transpose_y = mh_w_19_transpose_y_0, x = var_1314_cast_fp16, y = var_1316_cast_fp16)[name = tensor("mh_w_19_cast_fp16")]; tensor mh_w_21_cast_fp16 = add(x = mh_w_19_cast_fp16, y = var_195_cast_fp16)[name = tensor("mh_w_21_cast_fp16")]; tensor var_1324_cast_fp16 = softmax(axis = var_1190, x = mh_w_21_cast_fp16)[name = tensor("op_1324_cast_fp16")]; tensor var_1325 = const()[name = tensor("op_1325"), val = tensor([1, 20, 64, -1])]; tensor var_1326_cast_fp16 = reshape(shape = var_1325, x = value_13_cast_fp16)[name = tensor("op_1326_cast_fp16")]; tensor attn_13_transpose_x_0 = const()[name = tensor("attn_13_transpose_x_0"), val = tensor(false)]; tensor attn_13_transpose_y_0 = const()[name = tensor("attn_13_transpose_y_0"), val = tensor(true)]; tensor attn_13_cast_fp16 = matmul(transpose_x = attn_13_transpose_x_0, transpose_y = attn_13_transpose_y_0, x = var_1326_cast_fp16, y = var_1324_cast_fp16)[name = tensor("attn_13_cast_fp16")]; tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([1, 1280, 1, -1])]; tensor input_97_cast_fp16 = reshape(shape = var_1329, x = attn_13_cast_fp16)[name = tensor("input_97_cast_fp16")]; tensor pretrained_out_67_pad_type_0 = const()[name = tensor("pretrained_out_67_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_67_strides_0 = const()[name = tensor("pretrained_out_67_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_67_pad_0 = const()[name = tensor("pretrained_out_67_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_67_dilations_0 = const()[name = tensor("pretrained_out_67_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_67_groups_0 = const()[name = tensor("pretrained_out_67_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179303424))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180122688))), name = tensor("layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_self_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180122816)))]; tensor pretrained_out_67_cast_fp16 = conv(bias = layers_3_self_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_67_dilations_0, groups = pretrained_out_67_groups_0, pad = pretrained_out_67_pad_0, pad_type = pretrained_out_67_pad_type_0, strides = pretrained_out_67_strides_0, weight = layers_3_self_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_97_cast_fp16)[name = tensor("pretrained_out_67_cast_fp16")]; tensor input_99_pad_type_0 = const()[name = tensor("input_99_pad_type_0"), val = tensor("valid")]; tensor input_99_strides_0 = const()[name = tensor("input_99_strides_0"), val = tensor([1, 1])]; tensor input_99_pad_0 = const()[name = tensor("input_99_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_99_dilations_0 = const()[name = tensor("input_99_dilations_0"), val = tensor([1, 1])]; tensor input_99_groups_0 = const()[name = tensor("input_99_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180125440)))]; tensor input_99_cast_fp16 = conv(dilations = input_99_dilations_0, groups = input_99_groups_0, pad = input_99_pad_0, pad_type = input_99_pad_type_0, strides = input_99_strides_0, weight = layers_3_self_attn_o_proj_loraA_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; tensor lora_out_67_pad_type_0 = const()[name = tensor("lora_out_67_pad_type_0"), val = tensor("valid")]; tensor lora_out_67_strides_0 = const()[name = tensor("lora_out_67_strides_0"), val = tensor([1, 1])]; tensor lora_out_67_pad_0 = const()[name = tensor("lora_out_67_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_67_dilations_0 = const()[name = tensor("lora_out_67_dilations_0"), val = tensor([1, 1])]; tensor lora_out_67_groups_0 = const()[name = tensor("lora_out_67_groups_0"), val = tensor(1)]; tensor layers_3_self_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_self_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180166464)))]; tensor lora_out_67_cast_fp16 = conv(dilations = lora_out_67_dilations_0, groups = lora_out_67_groups_0, pad = lora_out_67_pad_0, pad_type = lora_out_67_pad_type_0, strides = lora_out_67_strides_0, weight = layers_3_self_attn_o_proj_loraB_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("lora_out_67_cast_fp16")]; tensor obj_49_cast_fp16 = add(x = pretrained_out_67_cast_fp16, y = lora_out_67_cast_fp16)[name = tensor("obj_49_cast_fp16")]; tensor inputs_21_cast_fp16 = add(x = inputs_19_cast_fp16, y = obj_49_cast_fp16)[name = tensor("inputs_21_cast_fp16")]; tensor out_21_axes_0 = const()[name = tensor("out_21_axes_0"), val = tensor([1])]; tensor var_1367_to_fp16 = const()[name = tensor("op_1367_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_21_cast_fp16 = layer_norm(axes = out_21_axes_0, epsilon = var_1367_to_fp16, x = inputs_21_cast_fp16)[name = tensor("out_21_cast_fp16")]; tensor obj_51_gamma_0_to_fp16 = const()[name = tensor("obj_51_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180207488)))]; tensor obj_51_beta_0_to_fp16 = const()[name = tensor("obj_51_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180210112)))]; tensor obj_51_epsilon_0_to_fp16 = const()[name = tensor("obj_51_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor obj_51_cast_fp16 = batch_norm(beta = obj_51_beta_0_to_fp16, epsilon = obj_51_epsilon_0_to_fp16, gamma = obj_51_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_21_cast_fp16)[name = tensor("obj_51_cast_fp16")]; tensor pretrained_out_69_pad_type_0 = const()[name = tensor("pretrained_out_69_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_69_strides_0 = const()[name = tensor("pretrained_out_69_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_69_pad_0 = const()[name = tensor("pretrained_out_69_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_69_dilations_0 = const()[name = tensor("pretrained_out_69_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_69_groups_0 = const()[name = tensor("pretrained_out_69_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180212736))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181032000))), name = tensor("layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181032128)))]; tensor pretrained_out_69_cast_fp16 = conv(bias = layers_3_encoder_attn_q_proj_pretrained_bias_to_fp16, dilations = pretrained_out_69_dilations_0, groups = pretrained_out_69_groups_0, pad = pretrained_out_69_pad_0, pad_type = pretrained_out_69_pad_type_0, strides = pretrained_out_69_strides_0, weight = layers_3_encoder_attn_q_proj_pretrained_weight_to_fp16_palettized, x = obj_51_cast_fp16)[name = tensor("pretrained_out_69_cast_fp16")]; tensor input_101_pad_type_0 = const()[name = tensor("input_101_pad_type_0"), val = tensor("valid")]; tensor input_101_strides_0 = const()[name = tensor("input_101_strides_0"), val = tensor([1, 1])]; tensor input_101_pad_0 = const()[name = tensor("input_101_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_101_dilations_0 = const()[name = tensor("input_101_dilations_0"), val = tensor([1, 1])]; tensor input_101_groups_0 = const()[name = tensor("input_101_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_q_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181034752)))]; tensor input_101_cast_fp16 = conv(dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = layers_3_encoder_attn_q_proj_loraA_weight_to_fp16, x = obj_51_cast_fp16)[name = tensor("input_101_cast_fp16")]; tensor lora_out_69_pad_type_0 = const()[name = tensor("lora_out_69_pad_type_0"), val = tensor("valid")]; tensor lora_out_69_strides_0 = const()[name = tensor("lora_out_69_strides_0"), val = tensor([1, 1])]; tensor lora_out_69_pad_0 = const()[name = tensor("lora_out_69_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_69_dilations_0 = const()[name = tensor("lora_out_69_dilations_0"), val = tensor([1, 1])]; tensor lora_out_69_groups_0 = const()[name = tensor("lora_out_69_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_q_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_q_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181075776)))]; tensor lora_out_69_cast_fp16 = conv(dilations = lora_out_69_dilations_0, groups = lora_out_69_groups_0, pad = lora_out_69_pad_0, pad_type = lora_out_69_pad_type_0, strides = lora_out_69_strides_0, weight = layers_3_encoder_attn_q_proj_loraB_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("lora_out_69_cast_fp16")]; tensor query_cast_fp16 = add(x = pretrained_out_69_cast_fp16, y = lora_out_69_cast_fp16)[name = tensor("query_cast_fp16")]; tensor pretrained_out_71_pad_type_0 = const()[name = tensor("pretrained_out_71_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_71_strides_0 = const()[name = tensor("pretrained_out_71_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_71_pad_0 = const()[name = tensor("pretrained_out_71_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_71_dilations_0 = const()[name = tensor("pretrained_out_71_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_71_groups_0 = const()[name = tensor("pretrained_out_71_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181116800))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181936064))), name = tensor("layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor pretrained_out_71_cast_fp16 = conv(dilations = pretrained_out_71_dilations_0, groups = pretrained_out_71_groups_0, pad = pretrained_out_71_pad_0, pad_type = pretrained_out_71_pad_type_0, strides = pretrained_out_71_strides_0, weight = layers_3_encoder_attn_k_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_71_cast_fp16")]; tensor input_103_pad_type_0 = const()[name = tensor("input_103_pad_type_0"), val = tensor("valid")]; tensor input_103_strides_0 = const()[name = tensor("input_103_strides_0"), val = tensor([1, 1])]; tensor input_103_pad_0 = const()[name = tensor("input_103_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_103_dilations_0 = const()[name = tensor("input_103_dilations_0"), val = tensor([1, 1])]; tensor input_103_groups_0 = const()[name = tensor("input_103_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_k_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181936192)))]; tensor input_103_cast_fp16 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = layers_3_encoder_attn_k_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_103_cast_fp16")]; tensor lora_out_71_pad_type_0 = const()[name = tensor("lora_out_71_pad_type_0"), val = tensor("valid")]; tensor lora_out_71_strides_0 = const()[name = tensor("lora_out_71_strides_0"), val = tensor([1, 1])]; tensor lora_out_71_pad_0 = const()[name = tensor("lora_out_71_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_71_dilations_0 = const()[name = tensor("lora_out_71_dilations_0"), val = tensor([1, 1])]; tensor lora_out_71_groups_0 = const()[name = tensor("lora_out_71_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_k_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_k_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(181977216)))]; tensor lora_out_71_cast_fp16 = conv(dilations = lora_out_71_dilations_0, groups = lora_out_71_groups_0, pad = lora_out_71_pad_0, pad_type = lora_out_71_pad_type_0, strides = lora_out_71_strides_0, weight = layers_3_encoder_attn_k_proj_loraB_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("lora_out_71_cast_fp16")]; tensor key_cast_fp16 = add(x = pretrained_out_71_cast_fp16, y = lora_out_71_cast_fp16)[name = tensor("key_cast_fp16")]; tensor pretrained_out_73_pad_type_0 = const()[name = tensor("pretrained_out_73_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_73_strides_0 = const()[name = tensor("pretrained_out_73_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_73_pad_0 = const()[name = tensor("pretrained_out_73_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_73_dilations_0 = const()[name = tensor("pretrained_out_73_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_73_groups_0 = const()[name = tensor("pretrained_out_73_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182018240))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182837504))), name = tensor("layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182837632)))]; tensor pretrained_out_73_cast_fp16 = conv(bias = layers_3_encoder_attn_v_proj_pretrained_bias_to_fp16, dilations = pretrained_out_73_dilations_0, groups = pretrained_out_73_groups_0, pad = pretrained_out_73_pad_0, pad_type = pretrained_out_73_pad_type_0, strides = pretrained_out_73_strides_0, weight = layers_3_encoder_attn_v_proj_pretrained_weight_to_fp16_palettized, x = encoder_output_embeds)[name = tensor("pretrained_out_73_cast_fp16")]; tensor input_105_pad_type_0 = const()[name = tensor("input_105_pad_type_0"), val = tensor("valid")]; tensor input_105_strides_0 = const()[name = tensor("input_105_strides_0"), val = tensor([1, 1])]; tensor input_105_pad_0 = const()[name = tensor("input_105_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_105_dilations_0 = const()[name = tensor("input_105_dilations_0"), val = tensor([1, 1])]; tensor input_105_groups_0 = const()[name = tensor("input_105_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_v_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182840256)))]; tensor input_105_cast_fp16 = conv(dilations = input_105_dilations_0, groups = input_105_groups_0, pad = input_105_pad_0, pad_type = input_105_pad_type_0, strides = input_105_strides_0, weight = layers_3_encoder_attn_v_proj_loraA_weight_to_fp16, x = encoder_output_embeds)[name = tensor("input_105_cast_fp16")]; tensor lora_out_73_pad_type_0 = const()[name = tensor("lora_out_73_pad_type_0"), val = tensor("valid")]; tensor lora_out_73_strides_0 = const()[name = tensor("lora_out_73_strides_0"), val = tensor([1, 1])]; tensor lora_out_73_pad_0 = const()[name = tensor("lora_out_73_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_73_dilations_0 = const()[name = tensor("lora_out_73_dilations_0"), val = tensor([1, 1])]; tensor lora_out_73_groups_0 = const()[name = tensor("lora_out_73_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_v_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_v_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182881280)))]; tensor lora_out_73_cast_fp16 = conv(dilations = lora_out_73_dilations_0, groups = lora_out_73_groups_0, pad = lora_out_73_pad_0, pad_type = lora_out_73_pad_type_0, strides = lora_out_73_strides_0, weight = layers_3_encoder_attn_v_proj_loraB_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("lora_out_73_cast_fp16")]; tensor value_cast_fp16 = add(x = pretrained_out_73_cast_fp16, y = lora_out_73_cast_fp16)[name = tensor("value_cast_fp16")]; tensor var_1450 = const()[name = tensor("op_1450"), val = tensor([1, 20, 64, -1])]; tensor mh_q_cast_fp16 = reshape(shape = var_1450, x = query_cast_fp16)[name = tensor("mh_q_cast_fp16")]; tensor var_1452_to_fp16 = const()[name = tensor("op_1452_to_fp16"), val = tensor(0x1p-3)]; tensor var_1453_cast_fp16 = mul(x = mh_q_cast_fp16, y = var_1452_to_fp16)[name = tensor("op_1453_cast_fp16")]; tensor var_1454 = const()[name = tensor("op_1454"), val = tensor([1, 20, 64, -1])]; tensor var_1455_cast_fp16 = reshape(shape = var_1454, x = key_cast_fp16)[name = tensor("op_1455_cast_fp16")]; tensor mh_w_transpose_x_0 = const()[name = tensor("mh_w_transpose_x_0"), val = tensor(true)]; tensor mh_w_transpose_y_0 = const()[name = tensor("mh_w_transpose_y_0"), val = tensor(false)]; tensor mh_w_cast_fp16 = matmul(transpose_x = mh_w_transpose_x_0, transpose_y = mh_w_transpose_y_0, x = var_1453_cast_fp16, y = var_1455_cast_fp16)[name = tensor("mh_w_cast_fp16")]; tensor obj_55_cast_fp16 = softmax(axis = var_1190, x = mh_w_cast_fp16)[name = tensor("obj_55_cast_fp16")]; tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1, 20, 64, -1])]; tensor var_1460_cast_fp16 = reshape(shape = var_1459, x = value_cast_fp16)[name = tensor("op_1460_cast_fp16")]; tensor attn_transpose_x_0 = const()[name = tensor("attn_transpose_x_0"), val = tensor(false)]; tensor attn_transpose_y_0 = const()[name = tensor("attn_transpose_y_0"), val = tensor(true)]; tensor attn_cast_fp16 = matmul(transpose_x = attn_transpose_x_0, transpose_y = attn_transpose_y_0, x = var_1460_cast_fp16, y = obj_55_cast_fp16)[name = tensor("attn_cast_fp16")]; tensor var_1463 = const()[name = tensor("op_1463"), val = tensor([1, 1280, 1, -1])]; tensor input_107_cast_fp16 = reshape(shape = var_1463, x = attn_cast_fp16)[name = tensor("input_107_cast_fp16")]; tensor pretrained_out_75_pad_type_0 = const()[name = tensor("pretrained_out_75_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_75_strides_0 = const()[name = tensor("pretrained_out_75_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_75_pad_0 = const()[name = tensor("pretrained_out_75_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_75_dilations_0 = const()[name = tensor("pretrained_out_75_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_75_groups_0 = const()[name = tensor("pretrained_out_75_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(182922304))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183741568))), name = tensor("layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 1280, 1, 1])]; tensor layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183741696)))]; tensor pretrained_out_75_cast_fp16 = conv(bias = layers_3_encoder_attn_o_proj_pretrained_bias_to_fp16, dilations = pretrained_out_75_dilations_0, groups = pretrained_out_75_groups_0, pad = pretrained_out_75_pad_0, pad_type = pretrained_out_75_pad_type_0, strides = pretrained_out_75_strides_0, weight = layers_3_encoder_attn_o_proj_pretrained_weight_to_fp16_palettized, x = input_107_cast_fp16)[name = tensor("pretrained_out_75_cast_fp16")]; tensor input_109_pad_type_0 = const()[name = tensor("input_109_pad_type_0"), val = tensor("valid")]; tensor input_109_strides_0 = const()[name = tensor("input_109_strides_0"), val = tensor([1, 1])]; tensor input_109_pad_0 = const()[name = tensor("input_109_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_109_dilations_0 = const()[name = tensor("input_109_dilations_0"), val = tensor([1, 1])]; tensor input_109_groups_0 = const()[name = tensor("input_109_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_o_proj_loraA_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183744320)))]; tensor input_109_cast_fp16 = conv(dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = layers_3_encoder_attn_o_proj_loraA_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; tensor lora_out_75_pad_type_0 = const()[name = tensor("lora_out_75_pad_type_0"), val = tensor("valid")]; tensor lora_out_75_strides_0 = const()[name = tensor("lora_out_75_strides_0"), val = tensor([1, 1])]; tensor lora_out_75_pad_0 = const()[name = tensor("lora_out_75_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_75_dilations_0 = const()[name = tensor("lora_out_75_dilations_0"), val = tensor([1, 1])]; tensor lora_out_75_groups_0 = const()[name = tensor("lora_out_75_groups_0"), val = tensor(1)]; tensor layers_3_encoder_attn_o_proj_loraB_weight_to_fp16 = const()[name = tensor("layers_3_encoder_attn_o_proj_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183785344)))]; tensor lora_out_75_cast_fp16 = conv(dilations = lora_out_75_dilations_0, groups = lora_out_75_groups_0, pad = lora_out_75_pad_0, pad_type = lora_out_75_pad_type_0, strides = lora_out_75_strides_0, weight = layers_3_encoder_attn_o_proj_loraB_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("lora_out_75_cast_fp16")]; tensor obj_53_cast_fp16 = add(x = pretrained_out_75_cast_fp16, y = lora_out_75_cast_fp16)[name = tensor("obj_53_cast_fp16")]; tensor inputs_23_cast_fp16 = add(x = inputs_21_cast_fp16, y = obj_53_cast_fp16)[name = tensor("inputs_23_cast_fp16")]; tensor out_23_axes_0 = const()[name = tensor("out_23_axes_0"), val = tensor([1])]; tensor var_1500_to_fp16 = const()[name = tensor("op_1500_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_23_cast_fp16 = layer_norm(axes = out_23_axes_0, epsilon = var_1500_to_fp16, x = inputs_23_cast_fp16)[name = tensor("out_23_cast_fp16")]; tensor input_111_gamma_0_to_fp16 = const()[name = tensor("input_111_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183826368)))]; tensor input_111_beta_0_to_fp16 = const()[name = tensor("input_111_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183828992)))]; tensor input_111_epsilon_0_to_fp16 = const()[name = tensor("input_111_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor input_111_cast_fp16 = batch_norm(beta = input_111_beta_0_to_fp16, epsilon = input_111_epsilon_0_to_fp16, gamma = input_111_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_23_cast_fp16)[name = tensor("input_111_cast_fp16")]; tensor pretrained_out_77_pad_type_0 = const()[name = tensor("pretrained_out_77_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_77_strides_0 = const()[name = tensor("pretrained_out_77_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_77_pad_0 = const()[name = tensor("pretrained_out_77_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_77_dilations_0 = const()[name = tensor("pretrained_out_77_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_77_groups_0 = const()[name = tensor("pretrained_out_77_groups_0"), val = tensor(1)]; tensor layers_3_fc1_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183831616))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187108480))), name = tensor("layers_3_fc1_pretrained_weight_to_fp16_palettized"), shape = tensor([5120, 1280, 1, 1])]; tensor layers_3_fc1_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_fc1_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187108608)))]; tensor pretrained_out_77_cast_fp16 = conv(bias = layers_3_fc1_pretrained_bias_to_fp16, dilations = pretrained_out_77_dilations_0, groups = pretrained_out_77_groups_0, pad = pretrained_out_77_pad_0, pad_type = pretrained_out_77_pad_type_0, strides = pretrained_out_77_strides_0, weight = layers_3_fc1_pretrained_weight_to_fp16_palettized, x = input_111_cast_fp16)[name = tensor("pretrained_out_77_cast_fp16")]; tensor input_113_pad_type_0 = const()[name = tensor("input_113_pad_type_0"), val = tensor("valid")]; tensor input_113_strides_0 = const()[name = tensor("input_113_strides_0"), val = tensor([1, 1])]; tensor input_113_pad_0 = const()[name = tensor("input_113_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_113_dilations_0 = const()[name = tensor("input_113_dilations_0"), val = tensor([1, 1])]; tensor input_113_groups_0 = const()[name = tensor("input_113_groups_0"), val = tensor(1)]; tensor layers_3_fc1_loraA_weight_to_fp16 = const()[name = tensor("layers_3_fc1_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187118912)))]; tensor input_113_cast_fp16 = conv(dilations = input_113_dilations_0, groups = input_113_groups_0, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = input_113_strides_0, weight = layers_3_fc1_loraA_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; tensor lora_out_77_pad_type_0 = const()[name = tensor("lora_out_77_pad_type_0"), val = tensor("valid")]; tensor lora_out_77_strides_0 = const()[name = tensor("lora_out_77_strides_0"), val = tensor([1, 1])]; tensor lora_out_77_pad_0 = const()[name = tensor("lora_out_77_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_77_dilations_0 = const()[name = tensor("lora_out_77_dilations_0"), val = tensor([1, 1])]; tensor lora_out_77_groups_0 = const()[name = tensor("lora_out_77_groups_0"), val = tensor(1)]; tensor layers_3_fc1_loraB_weight_to_fp16 = const()[name = tensor("layers_3_fc1_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187159936)))]; tensor lora_out_77_cast_fp16 = conv(dilations = lora_out_77_dilations_0, groups = lora_out_77_groups_0, pad = lora_out_77_pad_0, pad_type = lora_out_77_pad_type_0, strides = lora_out_77_strides_0, weight = layers_3_fc1_loraB_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("lora_out_77_cast_fp16")]; tensor input_115_cast_fp16 = add(x = pretrained_out_77_cast_fp16, y = lora_out_77_cast_fp16)[name = tensor("input_115_cast_fp16")]; tensor input_117_mode_0 = const()[name = tensor("input_117_mode_0"), val = tensor("EXACT")]; tensor input_117_cast_fp16 = gelu(mode = input_117_mode_0, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; tensor pretrained_out_pad_type_0 = const()[name = tensor("pretrained_out_pad_type_0"), val = tensor("valid")]; tensor pretrained_out_strides_0 = const()[name = tensor("pretrained_out_strides_0"), val = tensor([1, 1])]; tensor pretrained_out_pad_0 = const()[name = tensor("pretrained_out_pad_0"), val = tensor([0, 0, 0, 0])]; tensor pretrained_out_dilations_0 = const()[name = tensor("pretrained_out_dilations_0"), val = tensor([1, 1])]; tensor pretrained_out_groups_0 = const()[name = tensor("pretrained_out_groups_0"), val = tensor(1)]; tensor layers_3_fc2_pretrained_weight_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(187323840))), lut = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192239104))), name = tensor("layers_3_fc2_pretrained_weight_to_fp16_palettized"), shape = tensor([1280, 5120, 1, 1])]; tensor layers_3_fc2_pretrained_bias_to_fp16 = const()[name = tensor("layers_3_fc2_pretrained_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192239296)))]; tensor pretrained_out_cast_fp16 = conv(bias = layers_3_fc2_pretrained_bias_to_fp16, dilations = pretrained_out_dilations_0, groups = pretrained_out_groups_0, pad = pretrained_out_pad_0, pad_type = pretrained_out_pad_type_0, strides = pretrained_out_strides_0, weight = layers_3_fc2_pretrained_weight_to_fp16_palettized, x = input_117_cast_fp16)[name = tensor("pretrained_out_cast_fp16")]; tensor input_pad_type_0 = const()[name = tensor("input_pad_type_0"), val = tensor("valid")]; tensor input_strides_0 = const()[name = tensor("input_strides_0"), val = tensor([1, 1])]; tensor input_pad_0 = const()[name = tensor("input_pad_0"), val = tensor([0, 0, 0, 0])]; tensor input_dilations_0 = const()[name = tensor("input_dilations_0"), val = tensor([1, 1])]; tensor input_groups_0 = const()[name = tensor("input_groups_0"), val = tensor(1)]; tensor layers_3_fc2_loraA_weight_to_fp16 = const()[name = tensor("layers_3_fc2_loraA_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192241920)))]; tensor input_cast_fp16 = conv(dilations = input_dilations_0, groups = input_groups_0, pad = input_pad_0, pad_type = input_pad_type_0, strides = input_strides_0, weight = layers_3_fc2_loraA_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_cast_fp16")]; tensor lora_out_pad_type_0 = const()[name = tensor("lora_out_pad_type_0"), val = tensor("valid")]; tensor lora_out_strides_0 = const()[name = tensor("lora_out_strides_0"), val = tensor([1, 1])]; tensor lora_out_pad_0 = const()[name = tensor("lora_out_pad_0"), val = tensor([0, 0, 0, 0])]; tensor lora_out_dilations_0 = const()[name = tensor("lora_out_dilations_0"), val = tensor([1, 1])]; tensor lora_out_groups_0 = const()[name = tensor("lora_out_groups_0"), val = tensor(1)]; tensor layers_3_fc2_loraB_weight_to_fp16 = const()[name = tensor("layers_3_fc2_loraB_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192405824)))]; tensor lora_out_cast_fp16 = conv(dilations = lora_out_dilations_0, groups = lora_out_groups_0, pad = lora_out_pad_0, pad_type = lora_out_pad_type_0, strides = lora_out_strides_0, weight = layers_3_fc2_loraB_weight_to_fp16, x = input_cast_fp16)[name = tensor("lora_out_cast_fp16")]; tensor hidden_states_9_cast_fp16 = add(x = pretrained_out_cast_fp16, y = lora_out_cast_fp16)[name = tensor("hidden_states_9_cast_fp16")]; tensor inputs_cast_fp16 = add(x = inputs_23_cast_fp16, y = hidden_states_9_cast_fp16)[name = tensor("inputs_cast_fp16")]; tensor out_axes_0 = const()[name = tensor("out_axes_0"), val = tensor([1])]; tensor var_1575_to_fp16 = const()[name = tensor("op_1575_to_fp16"), val = tensor(0x1.5p-17)]; tensor out_cast_fp16 = layer_norm(axes = out_axes_0, epsilon = var_1575_to_fp16, x = inputs_cast_fp16)[name = tensor("out_cast_fp16")]; tensor hidden_states_gamma_0_to_fp16 = const()[name = tensor("hidden_states_gamma_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192446848)))]; tensor hidden_states_beta_0_to_fp16 = const()[name = tensor("hidden_states_beta_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192449472)))]; tensor hidden_states_epsilon_0_to_fp16 = const()[name = tensor("hidden_states_epsilon_0_to_fp16"), val = tensor(0x1.5p-17)]; tensor hidden_states_cast_fp16 = batch_norm(beta = hidden_states_beta_0_to_fp16, epsilon = hidden_states_epsilon_0_to_fp16, gamma = hidden_states_gamma_0_to_fp16, mean = obj_1_mean_0_to_fp16, variance = obj_1_variance_0_to_fp16, x = out_cast_fp16)[name = tensor("hidden_states_cast_fp16")]; tensor var_1586_axes_0 = const()[name = tensor("op_1586_axes_0"), val = tensor([2])]; tensor var_1586_cast_fp16 = squeeze(axes = var_1586_axes_0, x = hidden_states_cast_fp16)[name = tensor("op_1586_cast_fp16")]; tensor var_1589_perm_0 = const()[name = tensor("op_1589_perm_0"), val = tensor([0, 2, 1])]; tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192452096)))]; tensor var_1589_cast_fp16 = transpose(perm = var_1589_perm_0, x = var_1586_cast_fp16)[name = tensor("transpose_0")]; tensor logits = linear(bias = linear_0_bias_0_to_fp16, weight = embed_tokens_weight_to_fp16, x = var_1589_cast_fp16)[name = tensor("linear_0_cast_fp16")]; tensor var_1593 = const()[name = tensor("op_1593"), val = tensor(1)]; tensor obj_59_interleave_0 = const()[name = tensor("obj_59_interleave_0"), val = tensor(false)]; tensor key_cache_updates = concat(axis = var_1593, interleave = obj_59_interleave_0, values = (current_key_1_cast_fp16, current_key_3_cast_fp16, current_key_5_cast_fp16, current_key_cast_fp16))[name = tensor("obj_59_cast_fp16")]; tensor var_1596 = const()[name = tensor("op_1596"), val = tensor(1)]; tensor obj_61_interleave_0 = const()[name = tensor("obj_61_interleave_0"), val = tensor(false)]; tensor value_cache_updates = concat(axis = var_1596, interleave = obj_61_interleave_0, values = (current_value_1_cast_fp16, current_value_3_cast_fp16, current_value_5_cast_fp16, current_value_cast_fp16))[name = tensor("obj_61_cast_fp16")]; tensor var_1607_begin_0 = const()[name = tensor("op_1607_begin_0"), val = tensor([0, 4, 0, 0])]; tensor var_1607_end_0 = const()[name = tensor("op_1607_end_0"), val = tensor([1, 5, 1, 1500])]; tensor var_1607_end_mask_0 = const()[name = tensor("op_1607_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1607_cast_fp16 = slice_by_index(begin = var_1607_begin_0, end = var_1607_end_0, end_mask = var_1607_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_1607_cast_fp16")]; tensor var_1610_begin_0 = const()[name = tensor("op_1610_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1610_end_0 = const()[name = tensor("op_1610_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_1610_end_mask_0 = const()[name = tensor("op_1610_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_1610_squeeze_mask_0 = const()[name = tensor("op_1610_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_1610_cast_fp16 = slice_by_index(begin = var_1610_begin_0, end = var_1610_end_0, end_mask = var_1610_end_mask_0, squeeze_mask = var_1610_squeeze_mask_0, x = var_1607_cast_fp16)[name = tensor("op_1610_cast_fp16")]; tensor var_1625_begin_0 = const()[name = tensor("op_1625_begin_0"), val = tensor([0, 11, 0, 0])]; tensor var_1625_end_0 = const()[name = tensor("op_1625_end_0"), val = tensor([1, 12, 1, 1500])]; tensor var_1625_end_mask_0 = const()[name = tensor("op_1625_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1625_cast_fp16 = slice_by_index(begin = var_1625_begin_0, end = var_1625_end_0, end_mask = var_1625_end_mask_0, x = obj_41_cast_fp16)[name = tensor("op_1625_cast_fp16")]; tensor var_1628_begin_0 = const()[name = tensor("op_1628_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1628_end_0 = const()[name = tensor("op_1628_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_1628_end_mask_0 = const()[name = tensor("op_1628_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_1628_squeeze_mask_0 = const()[name = tensor("op_1628_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_1628_cast_fp16 = slice_by_index(begin = var_1628_begin_0, end = var_1628_end_0, end_mask = var_1628_end_mask_0, squeeze_mask = var_1628_squeeze_mask_0, x = var_1625_cast_fp16)[name = tensor("op_1628_cast_fp16")]; tensor var_1643_begin_0 = const()[name = tensor("op_1643_begin_0"), val = tensor([0, 3, 0, 0])]; tensor var_1643_end_0 = const()[name = tensor("op_1643_end_0"), val = tensor([1, 4, 1, 1500])]; tensor var_1643_end_mask_0 = const()[name = tensor("op_1643_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1643_cast_fp16 = slice_by_index(begin = var_1643_begin_0, end = var_1643_end_0, end_mask = var_1643_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1643_cast_fp16")]; tensor var_1646_begin_0 = const()[name = tensor("op_1646_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1646_end_0 = const()[name = tensor("op_1646_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_1646_end_mask_0 = const()[name = tensor("op_1646_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_1646_squeeze_mask_0 = const()[name = tensor("op_1646_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_1646_cast_fp16 = slice_by_index(begin = var_1646_begin_0, end = var_1646_end_0, end_mask = var_1646_end_mask_0, squeeze_mask = var_1646_squeeze_mask_0, x = var_1643_cast_fp16)[name = tensor("op_1646_cast_fp16")]; tensor var_1661_begin_0 = const()[name = tensor("op_1661_begin_0"), val = tensor([0, 6, 0, 0])]; tensor var_1661_end_0 = const()[name = tensor("op_1661_end_0"), val = tensor([1, 7, 1, 1500])]; tensor var_1661_end_mask_0 = const()[name = tensor("op_1661_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1661_cast_fp16 = slice_by_index(begin = var_1661_begin_0, end = var_1661_end_0, end_mask = var_1661_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1661_cast_fp16")]; tensor var_1664_begin_0 = const()[name = tensor("op_1664_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1664_end_0 = const()[name = tensor("op_1664_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_1664_end_mask_0 = const()[name = tensor("op_1664_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_1664_squeeze_mask_0 = const()[name = tensor("op_1664_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_1664_cast_fp16 = slice_by_index(begin = var_1664_begin_0, end = var_1664_end_0, end_mask = var_1664_end_mask_0, squeeze_mask = var_1664_squeeze_mask_0, x = var_1661_cast_fp16)[name = tensor("op_1664_cast_fp16")]; tensor var_1679_begin_0 = const()[name = tensor("op_1679_begin_0"), val = tensor([0, 11, 0, 0])]; tensor var_1679_end_0 = const()[name = tensor("op_1679_end_0"), val = tensor([1, 12, 1, 1500])]; tensor var_1679_end_mask_0 = const()[name = tensor("op_1679_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1679_cast_fp16 = slice_by_index(begin = var_1679_begin_0, end = var_1679_end_0, end_mask = var_1679_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1679_cast_fp16")]; tensor var_1682_begin_0 = const()[name = tensor("op_1682_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1682_end_0 = const()[name = tensor("op_1682_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_1682_end_mask_0 = const()[name = tensor("op_1682_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_1682_squeeze_mask_0 = const()[name = tensor("op_1682_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, squeeze_mask = var_1682_squeeze_mask_0, x = var_1679_cast_fp16)[name = tensor("op_1682_cast_fp16")]; tensor var_1697_begin_0 = const()[name = tensor("op_1697_begin_0"), val = tensor([0, 14, 0, 0])]; tensor var_1697_end_0 = const()[name = tensor("op_1697_end_0"), val = tensor([1, 15, 1, 1500])]; tensor var_1697_end_mask_0 = const()[name = tensor("op_1697_end_mask_0"), val = tensor([true, false, true, true])]; tensor var_1697_cast_fp16 = slice_by_index(begin = var_1697_begin_0, end = var_1697_end_0, end_mask = var_1697_end_mask_0, x = obj_55_cast_fp16)[name = tensor("op_1697_cast_fp16")]; tensor var_1700_begin_0 = const()[name = tensor("op_1700_begin_0"), val = tensor([0, 0, 0, 0])]; tensor var_1700_end_0 = const()[name = tensor("op_1700_end_0"), val = tensor([1, 1, 1, 1500])]; tensor var_1700_end_mask_0 = const()[name = tensor("op_1700_end_mask_0"), val = tensor([true, true, false, true])]; tensor var_1700_squeeze_mask_0 = const()[name = tensor("op_1700_squeeze_mask_0"), val = tensor([false, false, true, false])]; tensor var_1700_cast_fp16 = slice_by_index(begin = var_1700_begin_0, end = var_1700_end_0, end_mask = var_1700_end_mask_0, squeeze_mask = var_1700_squeeze_mask_0, x = var_1697_cast_fp16)[name = tensor("op_1700_cast_fp16")]; tensor var_1707 = const()[name = tensor("op_1707"), val = tensor(1)]; tensor var_1708_interleave_0 = const()[name = tensor("op_1708_interleave_0"), val = tensor(false)]; tensor var_1708_cast_fp16 = concat(axis = var_1707, interleave = var_1708_interleave_0, values = (var_1610_cast_fp16, var_1628_cast_fp16, var_1646_cast_fp16, var_1664_cast_fp16, var_1682_cast_fp16, var_1700_cast_fp16))[name = tensor("op_1708_cast_fp16")]; tensor var_1711 = const()[name = tensor("op_1711"), val = tensor(false)]; tensor obj_axes_0 = const()[name = tensor("obj_axes_0"), val = tensor([1])]; tensor alignment_heads_weights = reduce_mean(axes = obj_axes_0, keep_dims = var_1711, x = var_1708_cast_fp16)[name = tensor("obj_cast_fp16")]; } -> (logits, key_cache_updates, value_cache_updates, alignment_heads_weights); }