Multi-GPU parallelization of the iCATS clustering procedure

Optimize the iCATS clustering procedure for execution on multiple GPUs to improve its parallel scalability.

Background

The iCATS framework uses interaction-aware clustering to estimate important query-key regions for sparse attention in diffusion-transformer video generation. Although the method reduces the computational cost of forming interaction-based clustering objectives, the paper states that its clustering procedure is not yet optimized for multi-GPU parallel execution. Developing such an implementation is identified as future work, leaving the scalability of this component across multiple GPUs unresolved.