python313Packages.torch-bin Build Failue

[Luke2]

Hello, I am using flake.nix to set up my project environment. But I failed to build the python313Packages.torch-bin.

Here is the output.

error: Cannot build '/nix/store/d79qgcmzih32xnh1y5lvcqfw2p5wzkj6-python3.13-torch-2.10.0.drv'.
       Reason: builder failed with exit code 1.
       Output paths:
         /nix/store/ph94rl5rb5cb563g9ffj5x5vwz542qq0-python3.13-torch-2.10.0
         /nix/store/yk1x8jp8kdfa6i6a9jddyzxrzhkrjqg8-python3.13-torch-2.10.0-dist
       Last 25 log lines:
       > Running phase: updateAutotoolsGnuConfigScriptsPhase
       > Running phase: configurePhase
       > no configure script, doing nothing
       > Running phase: buildPhase
       > no Makefile or custom buildPhase, doing nothing
       > Running phase: pythonRuntimeDepsCheckHook
       > Executing pythonRuntimeDepsCheck
       > Checking runtime dependencies for torch-2.10.0-cp313-cp313-linux_x86_64.whl
       >   - fsspec not installed
       >   - cuda-bindings==12.9.4 not satisfied by version 12.8.0
       >   - nvidia-cuda-nvrtc-cu12 not installed
       >   - nvidia-cuda-runtime-cu12 not installed
       >   - nvidia-cuda-cupti-cu12 not installed
       >   - nvidia-cudnn-cu12 not installed
       >   - nvidia-cublas-cu12 not installed
       >   - nvidia-cufft-cu12 not installed
       >   - nvidia-curand-cu12 not installed
       >   - nvidia-cusolver-cu12 not installed
       >   - nvidia-cusparse-cu12 not installed
       >   - nvidia-cusparselt-cu12 not installed
       >   - nvidia-nccl-cu12 not installed
       >   - nvidia-nvshmem-cu12 not installed
       >   - nvidia-nvtx-cu12 not installed
       >   - nvidia-nvjitlink-cu12 not installed
       >   - nvidia-cufile-cu12 not installed
       For full logs, run:
         nix log /nix/store/d79qgcmzih32xnh1y5lvcqfw2p5wzkj6-python3.13-torch-2.10.0.drv

I am using nixos-unstable channel. And not only my flake.nix can’t build but also the nix-shell -p python313Packages.torch-bin can’t build. Is this my fault?

[Luke2]

Should I just override the python313Packages.torch-bin in nixpkgs unstable-channel or even disable the runtime dependencies checker?

[magicquark]

@Luke2

Is this my fault?

No. The package requires exactly version 12.9.4 of the cuda-bindings, but the version in nixpkgs unstable is 12.8.0, so the dependency check fails.

@Luke2

disable the runtime dependencies checker

You can do this, you can also be more precise and only disable the check for a specific dependency instead of all of them.

That said, while this is commonly done, it is not guaranteed to work as the pinning or restriction of a package version might be needed, e.g. if the API has changed making the code incompatible.

An alternative way, would be to bump the version of the cuda-bindings yourself, and use that to override the version of cuda-bindings used in torch-bin.

[Luke2]

Thanks for reply me, when will they fix this package?

[waffle8946]

If there’s not already a PR, there is no timeline (nixpkgs is a vounteer project). People who want to see something fixed must step up.

[Luke2]

I see. I now report it to NixOS/nixpkgs an wating for solving. Thanks for replying me.

[sirati]

if you want to you can try to use my WIP project that deals with torch/cuda

example/default.nix

Properly dealing with ML / CUDA / Python

I have now impl quite a lot of the above goal (yes i use LLMs, but I do read the code afterwards, still rn this is WIP so this is only on a best-effort bases) i am now able to build from bin or source a compatible set of torch, flash-attn, mamba-ssm, and causal-conv1d, while being able to specify the python, cuda, torch, and a special cuda-packages version for pascal gpus. (comment says otherwise but cuda 13.0 is supported technically, just missing cuda-packages v13 to be provided) for now …

edit: if you do this please explicitly pin to the latest commit of the feature-py-version-resolve branch, because as long as I am modifying the infrastructure, it will cause you a complete rebuild of all python packages, whenever I push a change and you run nix flake update which is unnecessary unless you require a newer package I add support for that is broken when added via extraDeps from nixpkgs

edit: also be careful with flash-attn there is no wheel yet for torch 2.10, so you will have to build it from source and it OOMs my 96Gb ram and swap machine in 5 seconds if i dont restrict max jobs and cores

[Luke2]

I found that the package python313Packages.torch can be built and its torch.cuda.is_available() return True. So I just need to use python313Packages.torch instead of python313Packages.torch-bin. Thanks.