Large language model (LLM) training is time-consuming and computationally expensive. Checkpointing periodically saves training states to enable recovery from the most recent checkpoint after failures. In-memory checkpointing stores training states in CPU memory, replicates them to peer nodes to enhance data fault tolerance, and can effectively schedule checkpoint transfers within communication idle windows under symmetric parallelism. However, asymmetric parallelism strategies are widely adopted because they can adapt more flexibly to model architectures and hardware resources and improve training efficiency, but they also introduce three core challenges: parameters exhibit mixed and fine-grained redundancy distribution patterns across devices, making redundancy identification difficult; intra-node and inter-node communication channels are independent, and each process has only a local view, rendering idle window detection ineffective; and existing systems lack the ability to construct globally optimal redundancy groups, thus failing to fully exploit redundancy and idle resources under asymmetric strategies. To this end, this study proposes Mosaic, an in-memory checkpointing system for asymmetric parallelism. Mosaic adopts a two-phase analysis-runtime architecture. In the analysis phase, the proposed system generates an optimal transfer plan through fine-grained parameter redundancy analysis and elimination, communication traffic modeling and idle window detection, and globally optimal redundancy group selection. In the runtime phase, checkpoint traffic is interleaved into idle windows for execution. Experimental results show that, compared with baseline systems, Mosaic reduces additional training overhead by 42.0% to 50.4% under asymmetric parallelism configurations.