Helios distilled dev

[Fatemanx]

Summary

This PR adds native Helios-Distilled integration to LightX2V following the existing runner / network / scheduler / text encoder / vae layering, instead of wrapping the upstream HeliosPyramidPipeline as a black-box pipeline bridge.

The public entry point is narrowed to model_cls=helios_distilled only. This avoids advertising generic Helios/Base support that is not actually implemented.

Main Changes

• Add native Helios modules under lightx2v/models/.../helios/:
  • HeliosModel and HeliosTransformer3DModel
  • HeliosDistilledScheduler and vendored HeliosDMDScheduler
  • HeliosTextEncoder (UMT5 path)
  • HeliosVAE
  • HeliosRunner for native t2v/i2v
• Add config and CLI support for model_cls=helios_distilled
• Parse local HF-style Helios directories from set_config.py
• Restrict support to distilled checkpoints and reject unsupported base metadata explicitly
• Add minimal runnable configs and scripts for Helios-Distilled t2v/i2v

Validation

• Static validation:
  • python -m py_compile passed for the changed runtime files
• Runtime validation:
  • --model_cls helios is now rejected explicitly
  • --model_cls helios_distilled runs successfully for local i2v
• End-to-end sample used for verification:
  • input image: assets/inputs/imgs/girl.png
  • prompt: the girl is dancing
  • output: 97 frames, 640x384, 24 fps

Known Limitations

• This PR supports Helios-Distilled only; it does not support Helios base checkpoints.
• This PR focuses on t2v/i2v CLI and pipeline integration. It does not add Gradio-side Helios model assembly.

[gemini-code-assist]

Code Review

This pull request introduces native integration for the Helios-Distilled model, supporting both Text-to-Video (T2V) and Image-to-Video (I2V) generation. It adds the necessary model components, including the transformer, text encoder, VAE, scheduler, and runner, along with configurations and shell scripts. The review identified a potential device mismatch in the VAE decoding process when CPU offloading is enabled, as well as incorrect --model_cls arguments in the provided run scripts that would cause validation failures.

[gemini-code-assist]

In decode, latents_mean and latents_std are moved to latents.device and latents.dtype. However, current_latents is constructed on self.model.device and self.model.dtype. If latents are on the CPU (e.g., due to CPU offloading) while self.model is on the GPU, this will cause a device mismatch runtime error (RuntimeError: Expected all tensors to be on the same device...). They should be moved to self.model.device and self.model.dtype instead, matching the implementation in prepare_image_latents.

Suggested change

	latents_mean = self.latents_mean.to(device=latents.device, dtype=latents.dtype)
	latents_std = self.latents_std.to(device=latents.device, dtype=latents.dtype)
	current_latents = latents.to(self.model.device, dtype=self.model.dtype) / latents_std + latents_mean
	latents_mean = self.latents_mean.to(device=self.model.device, dtype=self.model.dtype)
	latents_std = self.latents_std.to(device=self.model.device, dtype=self.model.dtype)
	current_latents = latents.to(self.model.device, dtype=self.model.dtype) / latents_std + latents_mean

[gemini-code-assist]

The script specifies --model_cls helios, but "helios" is not in the SUPPORTED_MODEL_CLASSES list in lightx2v/infer.py. Running this script will result in an invalid --model_cls error. It should be updated to --model_cls helios_distilled.

Suggested change

	--model_cls helios \
	--model_cls helios_distilled \

[gemini-code-assist]

The script specifies --model_cls helios, but "helios" is not in the SUPPORTED_MODEL_CLASSES list in lightx2v/infer.py. Running this script will result in an invalid --model_cls error. It should be updated to --model_cls helios_distilled.

Suggested change

	--model_cls helios \
	--model_cls helios_distilled \