New CUDA 10.1 update 1 install missing cusolver lsqvqr device functions

I am trying to find least square solutions to large, sparse, tall matrices on a GPU.
According to the documentation, cuSOLVER’s cusolverSpXcsrlsqvqr functions should do exactly that, with both Host (CPU) versions and non-host (GPU) versions.
unfortunately, my installation of the CUDA toolkit is missing the GPU version of these functions. I cannot run any programs relying on them, and I have confirmed that they are missing from libcusolver.so by running

$ nm -D /usr/local/cuda/lib64/libcusolver.so | grep “lsqvqr”

This reports the Host version of the function for all four datatypes, but the normal (GPU) versions do not appear.
Searching for other families of functions in this way (e.g. “lsvqr”) shows the expected GPU functions as well as the Host version of the functions.

I am running up-to-date CentOS 7 (64 bit) with CUDA toolkit version 10.1 update 1, installed per the instructions for the offline RPM, NVIDIA driver version 430.14, and CUDA runtime version 10.2.

Where is that function defined? Please identify the exact link to the online documentation.

If by X you mean cusolverSpcsrlsqvqr()

then those functions (the S and D variants) are only defined for host code paths, and that is documented:

https://docs.nvidia.com/cuda/cusolver/index.html#cusolver-lt-t-gt-csrlsqvqr

“Remark 2: only CPU (Host) path is provided.”

Note that using X as a replacement for is confusing, because the cusolver library also defines functions with X to mean something specific (i.e. specific functions)

Yes, that is the function family that I am talking about. I’m sorry I missed that very small note at the bottom of the documentation, that could probably be clearer.

Are there plans for the GPU paths to be provided in the near future?
It seems pointless to have only a CPU path in the CUDA library, there’s already plenty of well-optimized libraries for sparse matrix least squares solutions that run on the CPU.

I’m not able to make statements about the future on these forums. If you’d like to see a change in CUDA functionality, the recommended approach is to file a bug, using the word “enhancement” in the description. THe instructions for bug filing are linked to a sticky post at the top of the CUDA programming forum