The paper reports that the modified Mamba-2 block is tensor-parallel friendly and reduces synchronization points per block by half. It also describes sequence/context parallelism in which devices process contiguous sequence chunks and pass recurrent states between workers, with communication linear in the number of workers.

unavailable

cf33461905

This is source verification of forward tensor-parallel synchronization points, not an independent multi-device runtime or hardware-performance benchmark. · Backward gradient collectives and the paper's qualitative context-parallel recurrent-state exchange were recorded or discussed as scope context but were not included in the assigned scalar measurement. · The contract's paper Markdown path was unavailable, so the relevant fixed PDF pages were inspected directly. · The independent AI scientific assessor is unavailable. Numeric evidence is preserved, and CiteArk does not fall back to a fixed-threshold verdict.