Remove gfx1103 from gfx110X-all or switch back to publishing gfx110X-dgpu

[amd-aakash]

We are hitting build errors that look like machines are running out of memory after updating rocm-libraries in the gfx110X-all target family on Windows. See ROCm/rocm-libraries#2346.

For now, we plan on reverting the change that included gfx1103 in releases via the gfx110X-all target family, by either excluding it from that family or switching back to the gfx110X-dgpu family.

[erman-gurses]

Synced with Laura (@HereThereBeDragon) about it. Pointed out this is the commit caused the problem:

https://github.com/ROCm/TheRock/commit/8e1bd06ded145b75b7c8d62e7d41f117bf3f3c41#diff-bcf7382c20ca29…

Laura said it is only for Windows so I will revert from gfx110X-all to gfx110X-dgpu only for Windows to make it green.

[ScottTodd]

Laura said it is only for Windows so I will revert from gfx110X-all to gfx110X-dgpu only for Windows to make it green.

I would prefer we keep Linux and Windows GPU families in sync. Is a clean revert of 8e1bd06 possible?

Unwinding this is going to need communication with users who were expecting gfx1103 releases like on #1939.

[erman-gurses]

I agreed since it might cause further issues If they do not sync.
I am double checking if we can clean revert for that

[erman-gurses]

Apart from changes from dgpu to all. I see the changes at ‎ROADMAP.md file

[erman-gurses]

@ScottTodd , If you say OK to revert those above too. I will revert it.

[stellaraccident]

Clean reverts and keeping things in sync is preferred. In any PR, be sure to note that this is a temporary change to restore disrupted service and will be rolled forward deliberately in followons.

[erman-gurses]

Automatic revert did not work - getting message below:
Sorry, this pull request couldn’t be reverted automatically. It may have already been reverted, or the content may have changed since it was merged.
Will revert it manually and post a PR

[erman-gurses]

After clean revert, we have the error below:

"Building PyTorch wheels for gfx110X-dgpu"
Traceback (most recent call last):
  File "C:\home\runner\_work\TheRock\TheRock\external-builds\pytorch\build_prod_wheels.py", line 116, in <module>
    from packaging.version import Version, parse
ModuleNotFoundError: No module named 'packaging'
Traceback (most recent call last):
  File "C:\home\runner\_work\TheRock\TheRock\build_tools\github_actions\write_torch_versions.py", line 116, in <module>
    main(sys.argv[1:])
  File "C:\home\runner\_work\TheRock\TheRock\build_tools\github_actions\write_torch_versions.py", line 109, in main
    raise FileNotFoundError(f"Dist dir '{args.dist_dir}' does not exist")
FileNotFoundError: Dist dir 'C:\home\runner\_work\TheRock\TheRock\output\packages\dist' does not exist
Error: Process completed with exit code 1.

https://github.com/ROCm/TheRock/actions/runs/19082353866/job/54514052862

[ScottTodd]

I just hit that ModuleNotFoundError: No module named 'packaging' myself locally. Must be missing from requirements somewhere.

[erman-gurses]

Checking it.