Result of nixpkgs-review pr 179912 run on x86_64-linux 1

• cudaPackages.cuda-samples is known broken based on the comment in this PR
• mathematica-cuda is an existing failure due to missing the installer download

That leaves ethminer-cuda and cupy, which appear to be broken by this 11.7 upgrade. @dguibert could you add the appropriate overrides to keep these two packages on 11.6 (or whatever the appropriate versions are)?

AFAICT cupy error looks like

cupy_backends/cuda/libs/cutensor.cpp: In function ‘uint64_t __pyx_f_13cupy_backends_4cuda_4libs_8cutensor_contractionGetWorkspaceSize(__pyx_obj_13cupy_backends_4cuda_4libs_8cutensor_Handle*, __pyx_obj_13cupy_backen
ds_4cuda_4libs_8cutensor_ContractionDescriptor*, __pyx_obj_13cupy_backends_4cuda_4libs_8cutensor_ContractionFind*, int, int)’:
cupy_backends/cuda/libs/cutensor.cpp:5717:17: error: ‘CUTENSOR_VERSION’ was not declared in this scope; did you mean ‘CUTENSOR_OP_SIN’?
 5717 |   __pyx_t_1 = ((CUTENSOR_VERSION < 0x2904) != 0);
      |                 ^~~~~~~~~~~~~~~~
      |                 CUTENSOR_OP_SIN
cupy_backends/cuda/libs/cutensor.cpp: In function ‘uint64_t __pyx_f_13cupy_backends_4cuda_4libs_8cutensor_reductionGetWorkspaceSize(__pyx_obj_13cupy_backends_4cuda_4libs_8cutensor_Handle*, intptr_t, __pyx_obj_13cup
y_backends_4cuda_4libs_8cutensor_TensorDescriptor*, intptr_t, intptr_t, __pyx_obj_13cupy_backends_4cuda_4libs_8cutensor_TensorDescriptor*, intptr_t, intptr_t, __pyx_obj_13cupy_backends_4cuda_4libs_8cutensor_TensorD
escriptor*, intptr_t, int, int, int)’:
cupy_backends/cuda/libs/cutensor.cpp:6584:17: error: ‘CUTENSOR_VERSION’ was not declared in this scope; did you mean ‘CUTENSOR_OP_SIN’?
 6584 |   __pyx_t_1 = ((CUTENSOR_VERSION < 0x2904) != 0);
      |                 ^~~~~~~~~~~~~~~~
      |                 CUTENSOR_OP_SIN

cupy has been broken on master for a while

ah gotcha, looks like the same is true for ethminer... in that case I guess we're good to merge

https://gist.github.com/GrahamcOfBorg/45ac7f5bc9e02a74cb1e4264f365417f

Seems this PR broke eval on master.

11.7.0 would need to be added here, assuming it's compatible. (cc @aidalgol)

@zowoq uh oh, sorry about that! I'll revert. Looks like we'll need to rebase onto latest master and try again

thanks for the heads up!

11.7.0 would need to be added here, assuming it's compatible. (cc @aidalgol)

@samuela
That version is not. The list of supported versions in that file is exactly as listed on the nvidia download page. There is a newer version of TensorRT that supports CUDA 11.7, but it requires cuDNN 8.4, which is not yet in nixpkgs.

Hmm interesting... so we should be marking TensorRT as broken in that case? at least we can't break evaluation haha

long term solution of course is to package cuDNN 8.4...

Hmm interesting... so we should be marking TensorRT as broken in that case? at least we can't break evaluation haha

That is already done automatically in the TensorRT derivations (see here).

Mm, I see what you mean. It appears that merging this did break eval however. I haven't dug into all the details just yet, but the logs suggest it has something to do with TensorRT...

It appears that the case where the CUDA version is not in tensorRTDefaultVersion is not handled, which is why adding CUDA 11.7 broke eval. I'm not sure how best to handle this case. Perhaps define a default version for CUDA 11.7, and let the logic in generic.nix mark it as broken, because 11.7 is not in the list of supported CUDA versions, but it will at least evaluate.

@zowoq What was the exact command that was failing? I just want to make sure that I'm testing correctly

See OfBorg's README for the command it runs.

long term solution of course is to package cuDNN 8.4...

introduced by a43d296 within this PR.

Mm, I see what you mean. It appears that merging this did break eval however. I haven't dug into all the details just yet, but the logs suggest it has something to do with TensorRT...

Adding a line: "11.7" = "8.4.0"; to tensorRTDefaultVersion should be enough.

This pull request has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/how-to-install-a-specific-version-of-cuda-and-cudnn/21725/4