-
Notifications
You must be signed in to change notification settings - Fork 811
HAS_GPU should not be required for builds #278
Comments
Hi Kaixhin, We recently changed HAS_GPU to be automatically set to whether there was a functional GPU in the system rather than just looking for the presence of the CUDA libraries (which people install for debugging purposes on non-GPU systems). You should be able to override this with a manual setting as follows: make HAS_GPU=true Can you try that? Thanks! |
I've changed this line from File "/usr/local/bin/neon", line 49, in <module>
from neon.backends import gen_backend
...
OSError: CUDA driver library not found Any ideas on how to debug? An issue that might be relevant is NVIDIA/nvidia-docker#37. |
Thanks for the pointer. Following up on that, it seems likely that you are running into the same issue as NVIDIA/nvidia-docker#172 which might be addressed by adding /usr/lib/x86_64-linux-gnu to the LD_LIBRARY_PATH. Can you confirm? |
I've added this location to the front of LD_LIBRARY_PATH before ENV LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}
# Clone neon repo and move into it
RUN cd /root && git clone https://github.com/NervanaSystems/neon.git && cd neon && \
# Make (no multithreading to save resources/prevent concurrency errors)
make sysinstall HAS_GPU=true
# Set ~/neon as working directory
WORKDIR /root/neon |
It seems that although the latest builds for
kaixhin/cuda-neon
are succeeding, they aren't actually building CUDA support. The issue isHAS_GPU
in the Makefile, which fails to pick up a GPU during the build. Would it be possible to check that the toolkit works instead of looking for an actual GPU?The text was updated successfully, but these errors were encountered: