Fix: Remove hardcoded CUDA autocast in Kandinsky 5 to fix import warning (#12814)
* Fix: Remove hardcoded CUDA autocast in Kandinsky 5 to fix import warning
* Apply style fixes
* Fix: Remove import-time autocast in Kandinsky to prevent warnings
- Removed @torch.autocast decorator from Kandinsky classes.
- Implemented manual F.linear casting to ensure numerical parity with FP32.
- Verified bit-exact output matches main branch.
Co-authored-by: hlky <hlky@hlky.ac>
* Used _keep_in_fp32_modules to align with standards
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: hlky <hlky@hlky.ac>