The application provides a
large model training acceleration method and
system based on
CUDA shared memory and a PCIe
expansion card, hardware configuration information is acquired to identify the PCIe
expansion card and build an expansion storage space, and a unified management heterogeneous memory space is formed in combination with GPU
shared memory; in the training process, according to a preset initialization strategy, model data blocks are loaded to different levels of the heterogeneous memory,
data access frequency is monitored in real time,
data migration instructions are dynamically generated based on
multiple frequency thresholds, and intelligent scheduling of the data blocks among the levels of storage is realized; a transparent address remapping mechanism is used to enable
CUDA kernel functions to seamlessly access the heterogeneous memory space; periodic data snapshots and persistent storage are simultaneously supported, and quick
recovery after interruption of training is ensured; the application breaks through the GPU
display memory capacity limit, improves the
large model training efficiency, reduces the dependence on multi-card hardware and cost, and guarantees
data security and training reliability.