Pulse large model deployment verification method for brain-like computing platform

Abstract

The application discloses a kind of brain-like computing platform-oriented pulse big model deployment verification method, belong to brain-like computing technical field, comprising: determine the target linear layer of the piece to be executed in the model to be deployed;The weight and / or activation of target linear layer are quantized with low bit to convert into integer form;According to operator type, the heterogeneous module division is carried out to the model to be deployed, linear calculation module is distributed to brain-like chip side execution, auxiliary operator module and state management module are distributed to peripheral coprocessing unit execution;The weight matrix of quantized target linear layer is structured and blocked, and the blocked target linear layer is mapped to multi-chip neuron array;According to mapping result, generate neuron connection topology, parameter configuration, pulse transmission rule and running control information, and complete compilation and deployment.The method can improve the deployment realizability of big model on brain-like platform under the premise of ensuring the correctness of model function, reduce storage and communication overhead.

Images