Using cuda-quantum with PyCharm

[poojarao8]

@zivchen9993 is attempting to use the cuda-quantum docker container as the interpreter for PyCharm and are running into errors. Pasting the error messages here:

Also, I ran the code of the intro with the cudaq.set_target('nvidia') addition and I get two different results depending on where I run from:

If I run on the PyCharm with the cuda-quantum docker image as the interpreter, I receive this error upon running the cudaq.set_target('nvidia') line:
RuntimeError: [custatevec] %CUDA driver version is insufficient for CUDA runtime version in CuStateVecCircuitSimulator (line 334)

If I run on directly on the docker container and paste the same code, the code manages to run the cudaq.set_target('nvidia') line, but running:
@cudaq.kernel
def kernel(qubit_count: int):...           
generates this error:
OSError: could not get source code

@zivchen9993 is using WSL 2 with Ubuntu 22.04 and a Windows 64 bit machine and is pulling the nightly image.

Any guidance on next debugging steps or possible causes for the error would be great. Thanks!

[bmhowe23]

Does the nvidia-smi command work inside the container? If not, is Nvidia Container Toolkit installed on the host (in WSL), and is the container started with --gpus all?

[poojarao8]

Hey @zivchen9993, has anything changed on this?

[zivchen9993]

Hi Pooja, Sorry but not yet...

[bmhowe23]

Hi @zivchen9993 - I suspect the underlying cause of your two issues are different.

For the first issue (PyCharm with the cuda-quantum docker image as the interpreter), I suspect PyCharm is not starting the container using the appropriate GPU resources. Could you please run these Python commands in that environment and verify that nvidia-smi is working in that environment?

import subprocess
subprocess.run(['nvidia-smi'])

For the second issue (pasting code into an interactive Python session leads to OSError: could not get source code), that looks like a limitation of our kernel mode when using an interactive interpreter. I believe your options are to either:

1. Setup your PyCharm environment to use Python notebooks (i.e. .ipynb files instead of regular .py files), or
2. Avoid the @cudaq.kernel decorator syntax and use cudaq.make_kernel() instead (as shown here)

(Full disclosure - I am not very familiar with PyCharm, so I don't know how feasible my .ipynb suggestion really is.)

[zivchen9993]

Hi @bmhowe23, Thanks a lot for the reply!

I've checked and regarding the first issue, you are right, the PyCharm does have a problem with it and it seems that I'm not the only one frustrated by it (and it doesn't seem like there's a solution).
do you know of any nice IDE that won't have this kind of problem?

You were right about the second issue as well, I managed to run directly from the docker container and saw that my GPU was utilized during the process.

Thanks a lot and have a great weekend!

[bmhowe23]

do you know of any nice IDE that won't have this kind of problem?

Many of us use VS Code using these instructions. You can install many VS Code extensions to customize your Python experience. Hopefully this helps!