Sparsh: Breaking the Communication Bottleneck in Sequence Parallel Video Diffusion Inference with Predictive Sparse Communication

This artifact provides the Sparsh implementation for predictive sparse communication in distributed video diffusion inference. It includes the necessary source code for SparshWan2.2 and SparshWan2.1, covering the modified sequence-parallel attention path, KV prediction logic, adaptive online controller, and evaluation scripts. Model checkpoints are intentionally excluded and should be downloaded separately following the project documentation. The artifact supports reproducing the main communication reduction results, including predictive-step Q-only communication and full adaptive async prediction runs. Users should first install the dependencies listed in each subdirectory, prepare the corresponding Wan model checkpoints, and then follow the commands in SparshWan2.2/ARTIFACT_OVERVIEW.md. That document contains the detailed reproduction workflow, expected log outputs.