Nixpkgs: Can't build Pytorch

[GuillaumeDesforges]

Describe the bug
Pytorch fails to build

Executing pythonImportsCheckPhase
Check whether the following modules can be imported: torch
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "<string>", line 1, in <lambda>
  File "/nix/store/2dcsn57cgaxs92ha5swihrab0g3l2h6g-python3-3.7.7/lib/python3.7/importlib/__init__.py", line 127, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 1006, in _gcd_import
  File "<frozen importlib._bootstrap>", line 983, in _find_and_load
  File "<frozen importlib._bootstrap>", line 967, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 677, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 728, in exec_module
  File "<frozen importlib._bootstrap>", line 219, in _call_with_frames_removed
  File "/nix/store/i6fkycsn1q8vk4190hslrvh256wilw5i-python3.7-pytorch-1.4.1/lib/python3.7/site-packages/torch/__init__.py", line 81, in <module>
    from torch._C import *
ImportError: /nix/store/i6fkycsn1q8vk4190hslrvh256wilw5i-python3.7-pytorch-1.4.1/lib/python3.7/site-packages/torch/lib/libtorch_python.so: undefined symbol: _ZN5torch4cuda4nccl6detail16throw_nccl_errorE12ncclResult_t
builder for '/nix/store/lia86jj07nygz23ca3s1gxahx62ipls2-python3.7-pytorch-1.4.1.drv' failed with exit code 1
cannot build derivation '/nix/store/ia0jgpq5w2myqpp8zha2xj7h3f5cj5nn-python3-3.7.7-env.drv': 1 dependencies couldn't be built
cannot build derivation '/nix/store/hxp00cp2lvhhw677d1v8zbp5iv3cqpy1-home-manager-path.drv': 1 dependencies couldn't be built
cannot build derivation '/nix/store/gcpmscdc0d5h3fg5m6ik411q6hhsa0mz-home-manager-generation.drv': 1 dependencies couldn't be built
error: build of '/nix/store/gcpmscdc0d5h3fg5m6ik411q6hhsa0mz-home-manager-generation.drv' failed

To Reproduce
Steps to reproduce the behavior:

1. Set .config/nixpkgs/config.nix to

{
    allowUnfree = true;
    cudaSupport = true;
}

1. Switch to nixpkgs version 20.09pre227577.135073a87b7
2. nix-build '<nixpkgs>' -A python3Packages.pytorch

Expected behavior
Pytorch should build without failure.

Notify maintainers

@teh @thoughtpolice @tscholak

Metadata

$ nix-shell -p nix-info --run "nix-info -m"
 - system: `"x86_64-linux"`
 - host os: `Linux 5.4.35, NixOS, 20.09pre227577.135073a87b7 (Nightingale)`
 - multi-user?: `yes`
 - sandbox: `yes`
 - version: `nix-env (Nix) 2.3.5`
 - channels(theo): `"home-manager"`
 - channels(root): `"nixos-20.09pre227577.135073a87b7"`
 - channels(arsleust): `"home-manager"`
 - nixpkgs: `/nix/var/nix/profiles/per-user/root/channels/nixos`

• NVIDIA SMI working
  NVIDIA-SMI 440.82 Driver Version: 440.82 CUDA Version: 10.2

Maintainer information:

# a list of nixpkgs attributes affected by the problem
attribute:
# a list of nixos modules affected by the problem
module:

[evanjs]

I wonder if this is the sort of issue I was seeing during my many nixpkgs-review runs for #77714
Attaching the error I encountered just in case.
python37Packages.pytorchWithCuda.log

There were also some failures in a nixpkgs-review log that @jonringer posted, but I wasn't sure what to make of them, or if they were more builds that were failing when attempting builds with 128 cores as mentioned here.

[GuillaumeDesforges]

Your log (python37Packages.pytorchWithCuda.log) gives exactly the same error for the same symbol...

So if you now say:

Looks like pytorch and batchgenerators build fine on master with Pillow 7.1.2.

I'll try to rerun a build now

[evanjs]

Your log (python37Packages.pytorchWithCuda.log) gives exactly the same error for the same symbol...

So if you now say:

Looks like pytorch and batchgenerators build fine on master with Pillow 7.1.2.

I'll try to rerun a build now

I need to stop posting so late.
I did not realize the same symbol was there.

And yes, everything seemed to build fine on hydra.
Though, I suppose there’s a chance this might be an issue with the build environment, though I can’t imagine what that might be.

[GuillaumeDesforges]

Same error on 20.09pre228204.467ce5a9f45

To my knowledge hydra builds pytorch without CUDA, @tbenst was working on a Hydra for CUDA related libs I believe. Seeing as cuda and nccl are part of the name of the symbol _ZN5torch4cuda4nccl6detail16throw_nccl_errorE12ncclResult_t, I think this is an issue related to CUDA.

[GuillaumeDesforges]

Ok I have seen this: https://github.com/pytorch/pytorch/issues/32638

Corrected by https://github.com/pytorch/pytorch/commit/58cffbff91844217b9f4546e50ea0ab95544cb6a

Which is part of milestone 1.4.1 : https://github.com/pytorch/pytorch/milestone/14
but 1.5.0 seems on the way and the issue still not closed.

Anyway, I will investigate upstream and PR update accordingly

[jonringer]

since https://github.com/NixOS/nixpkgs/pull/89802 was merged, and I'm able to build this on master:

[16:46:18] jon@nixos ~/projects/nixpkgs (master)
$ nix-build -A python3Packages.pytorch
/nix/store/nr2p2w6mrs8rqnj43ngzw3jq5gixsjai-python3.7-pytorch-1.5.0

I'm going to considered this issue resolved.