feat: add Turing (20~ Series) GPU compatibility for Z Image Turbo

[devgdovg]

Motivation

Make Nunchaku Z-Image-Turbo model able to run with Turing GPUs.

Modifications

• Make SVDQW4A4Linear able to change dtype to torch.float16 when initializing.
• Manually convert state dict dtype to torch.float16 when loading to Turing GPU in order to avoid range overflow.

Checklist

• Code is formatted using Pre-Commit hooks.
• Relevant unit tests are added in the tests directory following the guidance in Contribution Guide.
• Documentation and example scripts in examples are updated if necessary.
• Throughput/latency benchmarks and quality evaluations are included where applicable.
• For reviewers: If you're only helping merge the main branch and haven't contributed code to this PR, please remove yourself as a co-author when merging.
• Please feel free to join our Slack, Discord or WeChat to discuss your PR.

[anr2me]

Nice, now we can use it on a free T4 GPUs at Kaggle/Colab 😁

[xynu617]

可以在我的Tesla T10上运行了，谢谢

[ataquino]

How can I switch to this PR locally to test it out?

[anr2me]

How can I switch to this PR locally to test it out?

After you git pull the nunchaku repo, do:

git fetch origin pull/833/head:pr-833

and then

git merge pr-833 -m "pr-833 or any message"

or (if it can't be merged successfully)

git checkout pr-833

[lmxyy]

could you add a test for this?

[lmxyy]

passing tests. approved