From: "Srivastava, Rohit Kumar" <[email protected]> Date: Friday, November 20, 2020 at 6:38 PM To: apachemxnetday <[email protected]> Subject: MXNet Presentation Topic
Presentation Topic: Large Tensor Support Abstract: When working with very large amounts of data sometimes int32 limits are not enough, for e.g. when working with graph neural network(GNN) or recommendation systems. Either the attributes of the data can be large or the sheer volume to be loaded into the machine becomes too high. In all such cases MXNet needed support for handling data that has over 2^31-1 units or attributes. For this reason the indexing of tensors and arrays inside MXNet needed to be updated to int64 with an upper limit at 2^63-1 for signed integers. This change had to be done carefully since updating every data type can cause significant slowdowns in operations, 3rd party BLAS and LAPCAK libraries had to be built with int64 support as well and address performance regressions caused by non-optimal implementation of certain ops. Presenters: Rohit Srivastava and Zhaoqi Zhu -Rohit
