CN-121981166-A - Data processing method and switch

Abstract

The application discloses a data processing method and a switch, the method comprises the steps that a first switch receives at least one word metadata sent by at least one data processor and a bitmap corresponding to each word metadata, the bitmap is used for indicating at least one target expert processor corresponding to the word metadata in a plurality of expert processors, the first switch determines the target expert processor corresponding to the word metadata in the plurality of expert processors according to a forwarding rule and the bitmap corresponding to the word metadata, the forwarding rule is determined according to intelligent reasoning capability information related to the processing efficiency of the expert processors, the intelligent reasoning capability information of the expert processors comprises any one or more of processing resource information, network topology information and communication tree structure information of the expert processors, and the first switch sends the word metadata to the target expert processor corresponding to the word metadata.

Inventors

• LI YANG
• WANG YONGGONG

Assignees

• 联想(北京)有限公司

Dates

Publication Date

20260505

Application Date

20260122

Claims (10)

1. 1. A method of data processing, the method comprising: the method comprises the steps that a first switch receives at least one word metadata sent by at least one data processor and a bitmap corresponding to each word metadata, wherein the bitmap is used for indicating at least one target expert processor corresponding to the word metadata in a plurality of expert processors; the first switch determines a target expert processor corresponding to the word metadata in a plurality of expert processors according to a forwarding rule and a bitmap corresponding to the word metadata, wherein the forwarding rule is determined according to intelligent reasoning capability information of the plurality of expert processors, and the intelligent reasoning capability information of the expert processors comprises any one or more of processing resource information, network topology information and communication tree structure information of the expert processors; And the first switch sends the word metadata to a target expert processor corresponding to the word metadata.
2. 2. The method of claim 1, the bitmap comprising a plurality of bits, the forwarding rule comprising correspondence of the plurality of bits and the plurality of expert processors; The first switch determines a target expert processor corresponding to the word metadata in a plurality of expert processors according to a forwarding rule and a bitmap corresponding to the word metadata, and the first switch comprises: the first switch analyzes the numerical value of each bit in the bitmap corresponding to the word metadata to obtain at least one target bit with a preset target value in the bitmap; and the first switch determines an expert processor corresponding to the target bit as a target expert processor corresponding to the word metadata based on the corresponding relation.
3. 3. The method of claim 2, each of the bits corresponding to an expert network; determining the forwarding rule according to intelligent reasoning capability information of the plurality of expert processors, including: And determining the corresponding relation between the plurality of bits and the plurality of expert processors in the forwarding rule according to the processing resource information for operating the expert network in the expert processor and the resource requirement of the expert network corresponding to the bits.
4. 4. The method of claim 2, determining the forwarding rule from intelligent reasoning capability information of the plurality of expert processors, comprising: Determining communication overhead of the first switch for transmitting the word metadata to the expert processor according to the network topology information of the expert processor; And determining the corresponding relation between the plurality of bits in the forwarding rule and the plurality of expert processors according to the communication overhead of sending the word metadata to the expert processors.
5. 5. The method of claim 2, determining the forwarding rule from intelligent reasoning capability information of the plurality of expert processors, comprising: determining communication overhead of the expert processor feedback processing result according to the communication tree structure information of the expert processor; and determining the corresponding relation between the plurality of bits in the forwarding rule and the plurality of expert processors according to the communication cost of the expert processor feedback processing result.
6. 6. The method of claim 1, the forwarding rule comprising a communication connection relationship of a plurality of interfaces of the expert processor and the first switch; the first switch sends the word metadata to a target expert processor corresponding to the word metadata, and the target expert processor comprises: the first switch determines a target interface connected with a target expert processor corresponding to the word metadata according to the forwarding rule; And the first switch broadcasts the word metadata to a target expert processor corresponding to the word metadata through the target interface.
7. 7. The method of claim 1, further comprising: A second switch receives the word metadata fed back by the target expert processor and expert processing results corresponding to the word metadata, wherein the expert processing results corresponding to the word metadata are obtained by processing the word metadata by the target expert processor, and the second switch and the first switch are the same or different; and the second switch feeds back the fusion processing result corresponding to the word metadata to a data processor sending the word metadata, wherein the fusion processing result corresponding to the word metadata is obtained by fusion of a plurality of expert processing results corresponding to the same word metadata.
8. 8. The method of claim 1, the first switch sending the word metadata to a target expert processor to which the word metadata corresponds, comprising: the first switch sequentially sends a plurality of word metadata to a target expert processor corresponding to the word metadata in sequence; The sending sequence of the plurality of word metadata is determined based on at least one of the priority, the data quantity and the data type corresponding to the word metadata, and the priority, the data quantity and the data type corresponding to the word metadata are sent to the first switch by the data processor.
9. 9. The method of claim 1, the first switch receiving at least one word metadata sent by at least one data processor, and a bitmap corresponding to each of the word metadata, comprising: the first switch receives at least one word metadata output by at least one data processor and a bitmap corresponding to each word metadata through a remote direct memory access network card of a server node to which the at least one data processor belongs.
10. 10. A switch comprising a switch processor and at least one interface; The at least one interface is used for receiving at least one word metadata sent by at least one data processor and a bitmap corresponding to each word metadata, wherein the bitmap is used for indicating at least one target expert processor corresponding to the word metadata in a plurality of expert processors; The switching processor is configured to: Determining target expert processors corresponding to the word metadata in a plurality of expert processors according to forwarding rules and bitmaps corresponding to the word metadata, wherein the forwarding rules are determined according to intelligent reasoning capability information of the expert processors, and the intelligent reasoning capability information of the expert processors comprises any one or more of processing resource information, network topology information and communication tree structure information of the expert processors; And sending the word metadata to a target expert processor corresponding to the word metadata.

Description

Data processing method and switch Technical Field The present application relates to the field of data processing technologies, and in particular, to a data processing method and an exchange. Background The mixed expert model is a type of reasoning model commonly used in the technical field of artificial intelligence, and comprises a plurality of mutually independent expert neural networks (expert networks for short) and at least one gating network for selecting the expert networks. When the method is used for reasoning, the gate control network selects at least one target expert network from a plurality of expert networks based on the input data, and then the input data is transmitted to the target expert network for processing. However, in practical application, the gate control network and the plurality of expert networks of the hybrid expert model are respectively operated on different processors, and data interaction between the gate control network and the expert networks can occupy resources of the processors and consume a great deal of time, which is not beneficial to improving performance of the hybrid expert model. Disclosure of Invention Therefore, the application discloses the following technical scheme: the first aspect of the present application provides a data processing method, the method comprising: the method comprises the steps that a first switch receives at least one word metadata sent by at least one data processor and a bitmap corresponding to each word metadata, wherein the bitmap is used for indicating at least one target expert processor corresponding to the word metadata in a plurality of expert processors; the first switch determines a target expert processor corresponding to the word metadata in a plurality of expert processors according to a forwarding rule and a bitmap corresponding to the word metadata, wherein the forwarding rule is determined according to intelligent reasoning capability information of the plurality of expert processors, and the intelligent reasoning capability information of the expert processors comprises any one or more of processing resource information, network topology information and communication tree structure information of the expert processors; And the first switch sends the word metadata to a target expert processor corresponding to the word metadata. Optionally, the bitmap includes a plurality of bits, and the forwarding rule includes correspondence between the plurality of bits and the plurality of expert processors; The first switch determines a target expert processor corresponding to the word metadata in a plurality of expert processors according to a forwarding rule and a bitmap corresponding to the word metadata, and the first switch comprises: the first switch analyzes the numerical value of each bit in the bitmap corresponding to the word metadata to obtain at least one target bit with a preset target value in the bitmap; and the first switch determines an expert processor corresponding to the target bit as a target expert processor corresponding to the word metadata based on the corresponding relation. Optionally, each bit corresponds to an expert network; determining the forwarding rule according to intelligent reasoning capability information of the plurality of expert processors, including: And determining the corresponding relation between the plurality of bits and the plurality of expert processors in the forwarding rule according to the processing resource information for operating the expert network in the expert processor and the resource requirement of the expert network corresponding to the bits. Optionally, determining the forwarding rule according to intelligent reasoning capability information of the plurality of expert processors includes: Determining communication overhead of the first switch for transmitting the word metadata to the expert processor according to the network topology information of the expert processor; And determining the corresponding relation between the plurality of bits in the forwarding rule and the plurality of expert processors according to the communication overhead of sending the word metadata to the expert processors. Optionally, determining the forwarding rule according to intelligent reasoning capability information of the plurality of expert processors includes: determining communication overhead of the expert processor feedback processing result according to the communication tree structure information of the expert processor; and determining the corresponding relation between the plurality of bits in the forwarding rule and the plurality of expert processors according to the communication cost of the expert processor feedback processing result. Optionally, the forwarding rule includes a communication connection relationship between a plurality of the expert processors and a plurality of interfaces of the first switch; the first switch sends the word metadata to a target expert processor