Sliding Window Memory Layout for Large Time-Series Model Training

The patent segments the time-series data into non-overlapping chunks and processes them individually, avoiding the creation of multiple overlapping window copies. Each chunk is processed once and fed into the foundation model sequentially, reducing memory requirements while maintaining training effectiveness through proper chunk selection and processing strategy.

The patent eliminates unnecessary data copying by using zero-copy operations where the foundation model processes data directly from the source buffer without creating intermediate copies. This is achieved through direct memory access and pointer manipulation, preventing the memory footprint explosion that occurs with traditional overlapping window techniques that copy data multiple times.

The patent implements zero-copy data transfer from CPU to GPU, eliminating redundant data copies. Instead of copying batched time series data multiple times between different memory spaces, the system uses direct memory access and shared memory buffers that allow both CPU and GPU to access the same data without duplication, significantly reducing memory usage while maintaining processing capability.

The patent nests data structures efficiently by organizing time-series data in a hierarchical manner where smaller data chunks are stored within larger buffers, and memory pointers reference these nested structures. This allows the system to process large volumes of data without allocating separate memory spaces for each data level, reducing overall memory requirements.