feat: modify kwai template

[LeonDiao0427]

PR type

• Bug Fix

PR information

Fix the KeyeVLTemplate implementation with the following changes:

• Add RoPE position_ids computation via model.get_rope_index_slowfast()
• Remove incorrect do_resize=False in image preprocessing

[gemini-code-assist]

Code Review

This pull request refactors the Kwai template to support slow-fast multimodal inputs and modularizes visual embedding logic. It also integrates RoPE index calculation into the post-encoding phase. Review feedback indicates that the early return in _post_encode may incorrectly bypass necessary embedding and position ID logic during inference. A suggestion was also made to use a more robust method for calculating dimension products to avoid floating-point conversions.

[gemini-code-assist]

The early return if not self.is_training prevents the computation of inputs_embeds and the new position_ids (via get_rope_index_slowfast) during inference when using the transformers backend. For multimodal models like KeyeVL, replacing token embeddings with visual embeddings and using the correct RoPE indices is essential for correct inference results. Since this PR aims to fix the RoPE implementation, this logic should also be applied during inference.

[gemini-code-assist]

Using np.prod on a slice of a tuple returns a float in many NumPy versions. While it is cast to int here, using math.prod (available in Python 3.8+) is generally more idiomatic for calculating the product of dimensions in a shape tuple and avoids floating-point conversions.

Suggested change

	media_position_ids = torch.arange(numel, device=device) % int(np.prod(thw_tuple[1:]))
	media_position_ids = torch.arange(numel, device=device) % int(torch.prod(torch.tensor(thw_tuple[1:])))

[Jintao-Huang]

thanks!

please pass lint test

pip install pre-commit
pre-commit run --all-files