TensorFlow in nvidia-docker: failed call to cuInit: CUDA_ERROR_UNKNOWN

Question

I have been working on getting an application that relies on TensorFlow to work as a docker container with nvidia-docker . I have compiled my application on top of the tensorflow/tensorflow:latest-gpu-py3 image. I run my docker container with the following command:

sudo nvidia-docker run -d -p 9090:9090 -v /src/weights:/weights myname/myrepo:mylabel

When looking at the logs through portainer I see the following:

2017-05-16 03:41:47.715682: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-16 03:41:47.715896: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-16 03:41:47.715948: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-05-16 03:41:47.715978: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-05-16 03:41:47.716002: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
2017-05-16 03:41:47.718076: E tensorflow/stream_executor/cuda/cuda_driver.cc:405] failed call to cuInit: CUDA_ERROR_UNKNOWN
2017-05-16 03:41:47.718177: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:158] retrieving CUDA diagnostic information for host: 1e22bdaf82f1
2017-05-16 03:41:47.718216: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:165] hostname: 1e22bdaf82f1
2017-05-16 03:41:47.718298: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:189] libcuda reported version is: 367.57.0
2017-05-16 03:41:47.718398: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:369] driver version file contents: """NVRM version: NVIDIA UNIX x86_64 Kernel Module  367.57  Mon Oct  3 20:37:01 PDT 2016
GCC version:  gcc version 4.8.4 (Ubuntu 4.8.4-2ubuntu1~14.04.3) 
"""
2017-05-16 03:41:47.718455: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:193] kernel reported version is: 367.57.0
2017-05-16 03:41:47.718484: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:300] kernel version seems to match DSO: 367.57.0

The container does seem to start properly, and my application does appear to be running. When I send requests to it for predictions the predictions are returned correctly - however at the slow speed I would expect when running inference on the CPU, so I think it's pretty clear that the GPU is not being used for some reason. I've also tried running nvidia-smi from within that same container to make sure it is seeing my GPU and these are the results for that:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 367.57                 Driver Version: 367.57                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GRID K1             Off  | 0000:00:07.0     Off |                  N/A |
| N/A   28C    P8     7W /  31W |     25MiB /  4036MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

I'm certainly no expert in this - but it does appear that the GPU is visible from inside the container. Any ideas on how to get this working with TensorFlow?

Answer 1

I run tensorflow on my ubuntu16.04 desktop.

I run code with GPU works well days before. But today I cannot find gpu device with below code

import tensorflow as tf from tensorflow.python.client import device_lib as _device_lib with tf.Session() as sess: local_device_protos = _device_lib.list_local_devices() print(local_device_protos) [print(x.name) for x in local_device_protos]

And I realize the below issue , when I run tf.Session()

cuda_driver.cc:406] failed call to cuInit: CUDA_ERROR_UNKNOWN

I check my Nvidia driver in the system details, and nvcc -V , nvida-smi to check driver ,cuda and cudnn. Everything seems well.

Then I went to Additional Drivers to check driver detail, there I find there are many versions of the NVIDIA driver and the latest version selected. But when I first install the driver there is only one.

So I select a old version, and apply the change.

Then I run the tf.Session() the issue is also here. I think I should reboot my computer, after I rebooted it, this issue gone.

sess = tf.Session() 2018-07-01 12:02:41.336648: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA 2018-07-01 12:02:41.464166: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:898] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero 2018-07-01 12:02:41.464482: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1356] Found device 0 with properties: name: GeForce GTX 1070 major: 6 minor: 1 memoryClockRate(GHz): 1.8225 pciBusID: 0000:01:00.0 totalMemory: 7.93GiB freeMemory: 7.27GiB 2018-07-01 12:02:41.464494: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1435] Adding visible gpu devices: 0 2018-07-01 12:02:42.308689: I tensorflow/core/common_runtime/gpu/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix: 2018-07-01 12:02:42.308721: I tensorflow/core/common_runtime/gpu/gpu_device.cc:929] 0 2018-07-01 12:02:42.308729: I tensorflow/core/common_runtime/gpu/gpu_device.cc:942] 0: N 2018-07-01 12:02:42.309686: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 7022 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability:

Answer 2

Maybe the problem is related to JIT caching files permissions, created by GPU. On linux, by default, cache files were created at ~/.nv/ComputeCache. Setting another directory for JIT cache solves the problem. Just do

export CUDA_CACHE_PATH=/tmp/nvidia

before running something on GPU.

Answer 3

I tried installing nvidia-modrpobe, but still the same error. Then a simple system reboot worked for me

Answer 4

In my case this command fails:

docker run --gpus all --runtime=nvidia -it --rm tensorflow/tensorflow:latest-gpu \                                                                                                                                                     
   python -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"

Adding --privileged solves the problem:

docker run --gpus all --runtime=nvidia --privileged -it --rm tensorflow/tensorflow:latest-gpu \                                                                                                                                                     
   python -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"