session_run hangs on GPU (libtensorflow-gpu) about hello_tf_c_api

[Neargye]

Could you please try this function?

hello_tf_c_api/src/tf_utils.hpp

Line 117 in 93a048e

TF_SessionOptions* CreateSessionOptions(double gpu_memory_fraction, TF_Status* status = nullptr);

[apivovarov]

I tried to replace

auto options = TF_NewSessionOptions();

with

auto options = tf_utils::CreateSessionOptions(0.75, status);

in src/session_run.cpp

I also tried other values - 0.95, 0.85, 0.65 and 0.55.

I did not help. ./session_run hangs.

Another guy from our company also faced the issue on his aws ec2 p3.2xlarge box (sm_70, Tesla V100-SXM2-16GB).

Should I try to build libtensorflow.so from sources on p3.2xlarge?

I also tried conda tensorflow-gpu-1.15 on my p3.2xlarge box - it works fine (regular Python API)

[apivovarov]

I built libtensorflow.so from sources (with CUDA 10.0.130, cudnn 7.6.5.32 and compute capability 70). ./session_run works now!

Download libtensorflow.so

[Neargye]

Thank you for sharing the solution and the assembled library!