PyTorch version parsing fails for source builds due to non-numeric suffixes #112
Description
Describe the bug
The PyTorch version checking code does not correctly handle versions built from source; I am one of those affected. For example, pip shows an installation of the release build of PyTorch 2.5.1 as 2.5.0a0+gita8d6afb. However, the code attempts to convert 0a0 to an integer, which is obviously invalid due to the a character, producing the following error:
2025-02-16 05:07:44 | INFO | configs.config | Found GPU AMD Radeon RX 5500 XT
2025-02-16 05:07:44 | INFO | configs.config | Half-precision floating-point: True, device: cuda:0
Traceback (most recent call last):
File "/home/rdna1_rocm-5.4_container/Software/RVC-Fumiama/web.py", line 15, in <module>
from infer.modules.uvr5.modules import uvr
File "/home/rdna1_rocm-5.4_container/Software/RVC-Fumiama/infer/modules/uvr5/modules.py", line 14, in <module>
config = Config()
^^^^^^^^
File "/home/rdna1_rocm-5.4_container/Software/RVC-Fumiama/configs/config.py", line 30, in __call__
cls._instances[cls] = super(Singleton, cls).__call__(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/rdna1_rocm-5.4_container/Software/RVC-Fumiama/configs/config.py", line 55, in __init__
self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
^^^^^^^^^^^^^^^^^^^^
File "/home/rdna1_rocm-5.4_container/Software/RVC-Fumiama/configs/config.py", line 228, in device_config
if tuple(map(int, torch.__version__.split("+")[0].split("."))) >= (2, 6, 0):
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: invalid literal for int() with base 10: '0a0'
To Reproduce
Create the PyTorch wheel locally, install it with pip, and run the following command:
python web.py
Expected behavior
The code should further initialize and load any required components without this error appearing.
Screenshots
Not applicable.
Desktop (please complete the following information):
- OS and version: Debian 12 Docker
- Python version: 3.11.2
- Commit/Tag with the issue: dev
Additional context
This current method is ineffective because development versions of PyTorch include non-numeric suffixes that break the int() conversion. The fix is rather simple: use the packaging.version module for a more robust solution. This prevents issues for users who build PyTorch or install nightly builds. Here is a proposed fix:
from packaging import version
if version.parse(torch.__version__) >= version.parse("2.6.0"):This approach ensures compatibility with all PyTorch versions regardless of their origin.