4 results about "Collective communication" patented technology

A processing system dynamically and selectively quantizes or compresses variable-length input data for a collective operation to fit within a predetermined limit, executes the collective operation on the compressed data, and converts the data back to variable length results by dequantizing or decompressing the results of the collective operation.

This invention discloses a GPU ensemble communication method, system, device, medium, and product. The method first maps ensemble communication interfaces from different GPU device manufacturers to a unified standard API interface based on GPU device manufacturer information, generating an interface mapping relationship. When responding to a ensemble communication request initiated by a user calling the API interface, the method translates the ensemble communication request into a corresponding manufacturer's ensemble communication interface request based on the GPU type used by the user and the interface mapping relationship, thereby invoking the corresponding manufacturer's ensemble communication interface to perform ensemble communication operations. Using this invention can reduce the development and maintenance complexity of multi-GPU ensemble communication systems and improve the overall performance and efficiency of distributed training of AI models.

Disclosed in the embodiments of the present application is a collective communication method. The method is applied to a collective communication system, wherein the collective communication system comprises a management node and N computing nodes, the N computing nodes being used for concurrently processing a first communication task, and N being an integer greater than 1. The method comprises: a management node receiving first information, wherein the first information indicates that at least one first link has failed, and each first link is a communication link between N computing nodes that is used for processing the first communication task; the management node receiving N pieces of second information from the N computing nodes, wherein each piece of second information is used for indicating whether original data of the first communication task that is locally stored in a computing node has been changed; and if the original data in the N computing nodes has not been changed, the management node sending third information to the N computing nodes, so as to instruct the N computing nodes to re-process the first communication task on the basis of the original data. In this way, the time required for task recovery after a communication fault occurs can be reduced.

A technique for efficiently performing integrated matrix multiplication to compute an output matrix based on two input matrices is described. Data corresponding to a first input tile of the first input matrix is ​​retrieved and stored in a local buffer. For each output tile of the output matrix, a non-blocking remote call is initiated to retrieve data corresponding to the next tile of the first input matrix into the local buffer, and concurrently with the processing of this remote call, the output tile is iteratively computed using the data from the first input matrix stored in the local buffer and a one-dimensional sequence of input tiles from the second input matrix. The output matrix is ​​generated based on the iteratively computed output tiles.