[LTXPipeline] Update latents dtype to match VAE dtype

[james-p-xu]

What does this PR do?

The LTX Video HuggingFace Documentation states:

Note: The recommended dtype is for the transformer component. The VAE and text encoders can be either torch.float32, torch.bfloat16 or torch.float16 but the recommended dtype is torch.bfloat16 as used in the original repository.

This implies that the VAE dtype can be different than the rest of the pipeline. e.g. the VAE could be in fp32 while the transformer is in fp16.

However, when running with fp32 VAE and fp16 transformer, I hit the following error.

RuntimeError: Input type (c10::Half) and bias type (float) should be the same

I believe the proper fix should be to ensure that the latents tensor (input) is the same dtype as the VAE (bias?). We can ensure this by casting the latents tensor to the proper dtype before the VAE decoder is run.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[james-p-xu]

cc: @a-r-r-o-w, not able to request you as a reviewer. Happy to chat about any changes / other ideas you might have here.

[a-r-r-o-w]

Thanks, makes sense!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.