Dali: Installation error with TF plugin 0.15 and g++-4.8

[ben0it8]

Based on the image below, the prebuilt plugin should be available for this compiler version but still, I can't install it.

More info:
Ubuntu 18.04
Python 3.7
TF 2.0

[jantonguirao]

Hi @ben0it8,

This is a known problem with the combination of Python 3.7 and Tensorflow 2.0 (which we just started supportingi recently).
Good news is that we already fixed it. It will be part of our next release, but for now you can just use our nightly builds instead:

pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly/cuda/10.0 nvidia-dali-nightly
pip install --extra-index-url https://developer.download.nvidia.com/compute/redist/nightly/cuda/10.0 nvidia-dali-tf-plugin-nightly

[ben0it8]

Amazing, thanks for the quick reply. I quickly tried this but I get the following RuntimeError when importing Dali:

[JanuszL]

@ben0it8 - could you show the output from nvidia-smi command (GPU model, driver version)?
Has the resale version of DALI works for you (non-nightly)?

[ben0it8]

@JanuszL no, the non-nightly tf-plugin does not install. I've tried TF1.15 too -- same error.

Edit:I use this image https://gist.github.com/ben0it8/28ccae6296c63902789883a8814c8616

[jantonguirao]

@ben0it8 I just tried the Dockerfile from your link and it works.

Are you running nvidia-docker (instead of docker)? You need that to have the CUDA runtime environment available (needed by DALI and tensorflow-gpu).

Regular docker (can't import):

$ docker run --rm test /bin/bash -c "python -c 'import nvidia.dali as dali; print(dali.__version__)'"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/srv/venv/lib/python3.7/site-packages/nvidia/dali/__init__.py", line 17, in <module>
    from . import ops
  File "/srv/venv/lib/python3.7/site-packages/nvidia/dali/ops.py", line 21, in <module>
    from nvidia.dali import backend as b
  File "/srv/venv/lib/python3.7/site-packages/nvidia/dali/backend.py", line 28, in <module>
    Init(OpSpec("CPUAllocator"), OpSpec("PinnedCPUAllocator"), OpSpec("GPUAllocator"))
RuntimeError: CUDA runtime API error cudaErrorInsufficientDriver (35):
CUDA driver version is insufficient for CUDA runtime version

nvidia-docker:

$ nvidia-docker run --rm test /bin/bash -c "python -c 'import nvidia.dali as dali; print(dali.__version__)'"
0.16.0dev.20191030

[ben0it8]

I was using docker run ...--runtime=nvidia, with nvidia-docker it works. thanks!