feat: video models and seeding for inference

[begumcig]

Description

This PR introduces several updates to the inference handlers:

DiffusersHandler Enhancements:

• Added support for video models.
• Added support for return_dict=False, which skips wrapping outputs in a pipeline output object and instead returns a tuple directly.

Seeding Support Across Handlers:

• All inference handlers now support optional seeding.

For DiffusersHandler, seeding is crucial since it directly affects the generation process.
For other models, inference is usually deterministic, but we added seeding for reproducibility in cases where randomness is involved.

Users can explicitly set or unset the seed. By default, no seed is applied, leaving seeding control to the user.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

Unit tests were added in tests/test_handler.py to validate the new functionality:

• Verified that seeding and unseeding work as expected for both standard inference handlers and the DiffusersHandler.

• Added tests for image and video model outputs using lightweight/tiny models to ensure correctness and efficiency.

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

Now you can configure the seed like:

input_text = "photo of a cute prune."
pipe = FluxPipeline.from_pretrained("katuni4ka/tiny-random-flux")
pruna_model = PrunaModel(pipe)
pruna_model.inference_handler.configure_seed("per_sample",global_seed=42)
output_image = pruna_model.run_inference(input_text)

[davidberenstein1957]

some small comments, but feel free to mege after resolving,

[davidberenstein1957]

Is this intended to overwrite? Perhaps checking if it is present and overwriting only if it is not is better?

[begumcig]

Yes! In the previous implementation we already automatically transformed everything to Tensors using torchvision.transforms.toTensor (line 87), so this is me just refactoring it using the more "official" way

[davidberenstein1957]

sassy 😂

[davidberenstein1957]

why were we doin this again? this is only a problem when using them in tests like this, but this is not the behavior we expect users to deal with, correct?

[begumcig]

Exactly, we are doing all this here because the pipe is not a PrunaModel. Indeed when we use PrunaModel and Evaluation Agent we don't have to deal with this at all. This is the exact feature I am testing here :D

[simlang]

only have minor comments. super clean 💅

[simlang]

i would just cast it to f32, tbh.
This could actually change the value of the input, because the precision is different between bf16 and f16

[begumcig]

Actually moving this entire type casting part to inference handler! And yes will use f32 tysm