A method and system for parallel training of a large-scale neural network of a perception model structure
By combining semantically enhanced computation graphs and hardware topology profiles, illegal paths are eliminated. A multi-dimensional cost model and a shortest path optimization algorithm with real-time weight correction are used to solve the long-tail effect problem in heterogeneous cluster computing in existing technologies, thereby improving the throughput of model training.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NO 15 INST OF CHINA ELECTRONICS TECH GRP
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-17
AI Technical Summary
Existing automation methods lack a deep understanding of the logical semantics of the model when dealing with highly integrated operators and non-Transformer structures, which leads to the inability to fully utilize the hardware potential of heterogeneous clusters and results in a long tail effect in computation.
By acquiring the original model code, it is transformed into a semantically enhanced computational graph. Combined with hardware topology profiling, constraint rule library, and fault tolerance contingency plan library, logically illegal paths and inefficient paths are eliminated. The shortest path optimization algorithm with multidimensional cost model and real-time weight correction is used to extract the best parallel strategy quadruple for training.
Ensuring the integrity of computational logic and avoiding communication spikes significantly improves model training throughput in heterogeneous architectures.
Smart Images
Figure CN122414248A_ABST