Support CUDA 12.8, RTX 50xx gpus

[LiuZhiyuan-python]

🚀 The feature, motivation and pitch

I'm learning and deploying deep learning projects on an RTX5070Ti. The nightly version is currently available for pytorch based on cuda 12.8 (about latest pytorch). Can we use pyg with cuda 12.8 and latest pytorch version In the future?

Alternatives

No response

Additional context

No response

[IlliaOvcharenko]

I tried to run current version of PyG available via pip (torch-2.5.0+cu124) with CUDA 12.8 and RTX 5080.

/home/dev/.local/lib/python3.10/site-packages/torch_geometric/typing.py:68: UserWarning: An issue occurred while importing 'pyg-lib'. Disabling its usage. Stacktrace: /home/dev/.local/lib/python3.10/site-packages/libpyg.so: undefined symbol: _ZNK5torch8autograd4Node4nameEv
  warnings.warn(f"An issue occurred while importing 'pyg-lib'. "
/home/dev/.local/lib/python3.10/site-packages/torch_geometric/typing.py:86: UserWarning: An issue occurred while importing 'torch-scatter'. Disabling its usage. Stacktrace: /home/dev/.local/lib/python3.10/site-packages/torch_scatter/_version_cuda.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKSsb
  warnings.warn(f"An issue occurred while importing 'torch-scatter'. "
/home/dev/.local/lib/python3.10/site-packages/torch_geometric/typing.py:97: UserWarning: An issue occurred while importing 'torch-cluster'. Disabling its usage. Stacktrace: /home/dev/.local/lib/python3.10/site-packages/torch_cluster/_version_cuda.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKSsb
  warnings.warn(f"An issue occurred while importing 'torch-cluster'. "
/home/dev/.local/lib/python3.10/site-packages/torch_geometric/typing.py:113: UserWarning: An issue occurred while importing 'torch-spline-conv'. Disabling its usage. Stacktrace: /home/dev/.local/lib/python3.10/site-packages/torch_spline_conv/_version_cuda.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKSsb
  warnings.warn(
/home/dev/.local/lib/python3.10/site-packages/torch_geometric/typing.py:124: UserWarning: An issue occurred while importing 'torch-sparse'. Disabling its usage. Stacktrace: /home/dev/.local/lib/python3.10/site-packages/torch_sparse/_version_cuda.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKSsb
  warnings.warn(f"An issue occurred while importing 'torch-sparse'. "

Traceback (most recent call last):
  ...
  File "/home/dev/codebase/src/metrics.py", line 11, in <module>
    from torch_scatter import scatter_add
  File "/home/dev/.local/lib/python3.10/site-packages/torch_scatter/__init__.py", line 16, in <module>
    torch.ops.load_library(spec.origin)
  File "/home/dev/.local/lib/python3.10/site-packages/torch/_ops.py", line 1392, in load_library
    ctypes.CDLL(path)
  File "/opt/conda/lib/python3.10/ctypes/__init__.py", line 374, in _init_
    self._handle = _dlopen(self._name, mode)
OSError: /home/dev/.local/lib/python3.10/site-packages/torch_scatter/_version_cuda.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKSsb

Is there a way to compile PyG from source on a device, will it solve an issue?

[akihironitta]

Yes, there is a way to build from source. Feel free to refer to https://github.com/pyg-team/pyg-lib#from-master. We will release wheels built with CUDA 12.8 once PyTorch 2.7 is released on 23rd April, 2025.

[akihironitta]

Fro pyg-lib, it's in progress in pyg-team/pyg-lib#442. Wheels for PyTorch 2.7 with CUDA 12.8 should be ready sometime this week.

[akihironitta]

For Linux, nightly pyg-lib wheels with CUDA 12.8 are available. Nightly pyg-lib wheels with CUDA 12.8 for Windows will also be available shortly (https://github.com/pyg-team/pyg-lib/actions/runs/14880351089). You can try them out by following the instrudciton at https://github.com/pyg-team/pyg-lib#form-nightly. Please let us know if you encounter any issues!

https://github.com/pyg-team/pyg-lib#form-nightly

---

For note, if you encounter an error like this one, you'll need to use a newer Ubuntu version (2104 or later) until we can make them manylinux compatible through pyg-team/pyg-lib#447 or pyg-team/pyg-lib#435.

OSError: /lib/x86_64-linux-gnu/libc.so.6: version `GLIBC_2.33' not found (required by /home/aki/.miniconda3/envs/pyg-lib310/lib/python3.10/site-packages/libpyg.so)

[akihironitta]

Keeping this issue open until torch-scatter and torch-sparse adds support for PyTorch 2.7.

[LiuZhiyuan-python]

Thank you for your attention and efforts on this issue! Currently, when I need to from torch_scatter import scatter in my code, the following error occurs: OSError: /home/eric/miniconda3/envs/LeakDetection/bin/../lib/libstdc++.so.6: version `GLIBCXX_3.4.32' not found (required by /home/eric/miniconda3/envs/LeakDetection/lib/python3.10/site-packages/torch_scatter/_version_cuda.so)