The invention relates to the technical field of computers, and provides a distributed AI
large model warehouse downloading acceleration method, which comprises the following steps that: a newly added node completes synchronization of networking and a file mapping relation table between the newly added node and nodes of the whole network based on a gossip protocol; each node constructs a Hash ring through a consistent Hash
algorithm based on the node identifier, determines a target Key according to the deployment identifier, and
clockwise selects a next node of the target Key as a main node in the Hash ring; when a local
intranet node receives a downloading request, firstly searching a target file in a local file
list; if the target file is not found locally, initiating an
intranet back-to-source request to a target node recorded in the file mapping relation table to obtain the target file; and if the
intranet is not hit, a
public network back-to-source request is initiated to the main node to obtain the target file, a decentralized networking
mechanism based on the gossip protocol and a file mapping relation sharing mechanism are adopted, so that repeated downloading among the nodes is effectively avoided, bandwidth waste is reduced, and
public network bandwidth occupation is remarkably reduced.