Thorsten Kurth

I was thinking that cuda::toolkit should add cuda::cuda_driver but maybe it does not anymore since the driver lib is not always required. But aluminum needs it for cudaStreamWrite32 and 3...

Actually the matter is more complicated: when I am trying to build lbann while being inside a docker container, it works fine. When I am using a Dockerfile issuing the...

Ok, I found it: I need to link libcuda.so from stubs into some directory where the linker can find it. I was expecting it to barf on missing lcuda but...

Hello @Weigaa, thanks for your insights. Before trying to dig deeper into this issue, did you make sure that your drive is GDS compatible? Each gds installation comes with a...

Hello @Weigaa , A few things to mention here: 1) it seems that your installation does not support GDS. When you installed CUDA on the system, did you also install...

Hello Weigaa, sorry for the late reply. I am trying to figure that out with the GDS developers. So good news is, P100 is supported. Therefore, the reason why you...

Hello @Weigaa, so this is a GDS issue: inside DALI, I am printing the CUFILE error I receive. It says: Error in thread 0: [/opt/dali/dali/util/std_cufile.cc:73] CUFile import failed: ./Inputnumpy.npy. GPUDirect...

Please make sure that you point it to an empty directory on the NVME drive. I recommend to create some dummy files there or let gdsio write some files there...

Hello @Weigaa, I apologize for the silence but I was/am on vacation till July 15. What I read from the table is that GDS compat mode is faster than GDS...

Hello @Weigaa, What you did sounds correct. In more recent GDS versions, there is a simpler method to enforce compat mode but your solution works. I think this is an...