量子位官网T2·Infra
58 pts
Shishi Tech released a multimodal model serving product that combines residual-based adaptive caching, SGLang cache management, modality-aware token compression and a Context Engine to speed up video generation. Under its own test setup it reports a 2.48x end-to-end speedup versus a dense baseline while preserving output quality, trading cache reuse against error recalibration rather than simply cutting steps.