ernie模型和FasterTransformer冲突？报错symbol runGemmShortApi, version libcublasLt.so.11 not defined in file libcublasLt.so.11 with link time reference

[bdbaigc]

1、python pipeline_service.py 启动服务，一切正常

2023/08/08 23:46:45 start proxy service

2、curl 访问算子
pipeline_service.py直接挂掉，同时std_out打印如下日志：

W0808 23:46:51.424680 206179 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 8.6, Driver API Version: 11.2, Runtime API Version: 11.1
W0808 23:46:51.427105 206179 gpu_resources.cc:91] device: 0, cuDNN Version: 8.1.
python: relocation error: /home/work/paddle_env/cuda_pkgs/cuda-11.1/lib64/libcublas.so: symbol runGemmShortApi, version libcublasLt.so.11 not defined in file libcublasLt.so.11 with link time reference

现象：
load("FasterTransformer", build_dir="/home/work/paddle_env/ppnlp_home/extensions", verbose=True)
这行注释掉程序运行OK，但是这行注释掉会影响其他算子的运行，所以不能注释。

不知道ernie模型和FasterTransformer有啥冲突。

附录：
1）pipeline_service.py 部分代码如下
if name == "main":
service = ModelService(name="model_service")
script_dir = os.path.dirname(os.path.abspath(file))
root_dir = os.path.dirname(script_dir)
load("FasterTransformer", build_dir="/home/work/paddle_env/ppnlp_home/extensions", verbose=True)
config_file = os.path.join(script_dir, "conf/config.yml")
service.prepare_pipeline_config(config_file)
service.run_service()

2）算子初始化代码如下：
import json
from paddle_serving_server.web_service import Op
import numpy as np
from collections import namedtuple
class TextQualityDetectOp(Op):
"""文本去噪模型"""

def init_op(self):
    """初始化
    """
    from paddlenlp.transformers import AutoTokenizer
    self.tokenizer = AutoTokenizer.from_pretrained("ernie-3.0-medium-zh",
                                                    use_fast=True)
    self.fetch_names = [
        "linear_147.tmp_1",
    ]