Daniel Socek

> do we need --allow_tf32? No, I removed in the PR amendment

> @dsocek , Is this still good to review/merge or does it require a sync, make style, fast_tests check? @emascarenhas this PR is was good shape, but I need to...

@emascarenhas turns out guidance is not working properly even in the original diffusers. This PR will need to be slightly reworked, but good news is I managed find a fix....

# Updated PR (ready for review) cc: @emascarenhas @libinta @regisss - Fixed guidance on G2 - Improved training logic to avoid OOMs when doing validation - Added support for training...

@regisss my equivalent PR fix is now merged in main diffusers branch (https://github.com/huggingface/diffusers/pull/9010). May be good to also update optimum-habana with the fix too. Could you please review this PR...

> @dsocek Thanks for working on this on both HF and OH. I've tested this pr and seems working. I added some CI tests in attached patch. I've tested it...

Note: Quality of generated video is really poor with `classifier_free_guidance` enabled. I override it to `False` and quality looks good.

@libinta Updated as follows: * Migrated SVD to stable_diffusion folder * Added reference for source file in original `diffusers` * Removed default image to `None`

@regisss Thanks for review, yes for now HPU graph capture disabled. Fixed all other indicated issues in the amendment.

@libinta @regisss Replaced CPU-based `randn_tensor` call with HPU-based (tested and output video looks good)