[enhancement]: Support pytorch 2.5.0+ for using ROCm in linux kernel 6.8.0+

[chuzhe-as-a-dev]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

No response

What should this feature add?

Allow torch version 2.5.0+.

Alternatives

Alternative would be downgrading linux kernel to an older version. This is obviously a temporary "fix", and eventually people will be moving to new versions...

Additional Content

Currently, torch is pinned "<2.5.0", and therefore can only work with ROCm 6.1 (see this).

However, with a recent linux kernel update, older ROCm version does not compile correctly, see this. While this issue is reported with ROCm 6.2.0, I have encountered the exact same failure and compile error log with ROCm 6.1.2, and I believe this is general to all previous ROCm versions.

A fix is added in ROCm 6.2.1, which is only supported by pytorch 2.5.0+.

Please consider bumping up the pytorch version, thanks!

[Vargol]

Please note pytorch 2.5.x is a mess on MacOS, using more memory and is a lot slower.
pytorch/pytorch#139389
For the report

For example a standard SDXL run via Diffusers using 2.4.1, the python binary reports using 9.5Gb and runs at 5.7 seconds per iteration, under 2.5.1 or nightly it reports 14.9 GB and runs at 8.5 s/i

Also 2.5.0 breaks GGUF on macOS,, but I believe that fix got cherry picked into 2.5.1

Any move to pytorch 2.5 will need to take that into account.

[chuzhe-as-a-dev]

@Vargol Thanks for the reply! Really looking forward to PyTorch stabilizing. Hope it happens soon.