Created by: AoZhang

• 环境 1）PaddlePaddle版本：gpu版，1.6.1/1.6.2 2）CPU：Intel(R) Xeon(R) CPU E5-2620 v2 3）GPU：NVIDIA K40m 4）系统环境：CentOS 6.3 ，Python版本 3.6.5

• 问题描述 预测阶段，config设置了 disable_gpu() 和 enable_mkldnn()，但是 create_paddle_predictor 时依然使用了 GPU。如果测试对应的 GPU 卡正在被占用且显存不足，则会报如下的错误：

>>> predictor = create_paddle_predictor(config.to_native_config())
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
paddle.fluid.core_avx.EnforceNotMet:

--------------------------------------------
C++ Call Stacks (More useful to developers):
--------------------------------------------
0   std::string paddle::platform::GetTraceBackString<char const*>(char const*&&, char const*, int)
1   paddle::platform::EnforceNotMet::EnforceNotMet(std::__exception_ptr::exception_ptr, char const*, int)
2   paddle::platform::GpuMemoryUsage(unsigned long*, unsigned long*)
3   paddle::AnalysisConfig::fraction_of_gpu_memory_for_pool() const
4   paddle::AnalysisConfig::ToNativeConfig() const

----------------------
Error Message Summary:
----------------------
PaddleCheckError: cudaMemGetInfo failed in paddle::platform::GetMemoryUsage, error code : 2, Please see detail in http
s://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430
e0038: out of memory at [/paddle/paddle/fluid/platform/gpu_info.cc:188]

• 复现代码 操作系统环境变量中 CUDA_VISIBLE_DEVICES=0，且正在被使用

from paddle.fluid.core_avx import AnalysisConfig, create_paddle_predictor
config = AnalysisConfig("lstm/model", "lstm/params")
config.disable_gpu()
config.enable_mkldnn()
predictor = create_paddle_predictor(config.to_native_config())

模型： lstm.tar.gz