Wave Forcing from SJTU-DENG Lab: 8 GPUs can't fix a missing sgl-dlab package

Wave Forcing targets a specific bottleneck in streaming video generation: naive diffusion models force each new chunk to wait until the previous chunk is fully denoised, which blocks pipelining across GPUs. The method redesigns chunk-level causal attention so a new chunk can condition on a mixed-noise frontier, not just a clean prior chunk. That, in turn, enables Wave Parallelism, where each stage of a chunk's processing runs on separate GPU replicas. The dedicated WaveRT runtime handles the cross-GPU exchange of K/V state, prefetching communication under longer attention computations. The repo includes multiple exchange-mode implementations (sync, overlap, one-sided, staggered, relay, paged).

The gap between the design and the distributed code is the central issue. README's quick start treats an 8-GPU 5+2 configuration as the minimum, and config.py constrains parallelism so that wp_size=1 collapses to a degenerate rf_step=0 case, effectively removing the pipelining the method is meant to provide. Yet the published dependency chain does not support even that setup: uv sync installs no sglang, and the registry lookup for WanRollingForcingPipeline fails. The only reference to RollingForcing outside config.py is absent, indicating the class lives in a non-public fork. A separate comparison script also hits an API mismatch: maybe_init_distributed_environment_and_model_parallel() no longer accepts cfg_degree but expects enable_cfg_parallel. Fixing that mismatch still leaves the missing pipeline as the blocker.

The license and weights are public and permissive for inference, but reproducibility remains out of reach from the published README and pyproject.toml. The author frames this as a defect in the tool's public dependency scope rather than a limitation of the hardware environment, meaning teams with ample GPU capacity still cannot verify the reported acceleration without access to the authors' internal package.