TaoLv commented on a change in pull request #17138: Interleaved MHA for CPU path URL: https://github.com/apache/incubator-mxnet/pull/17138#discussion_r360753605
########## File path: src/operator/contrib/transformer.cc ########## @@ -122,6 +122,531 @@ static bool InterleavedMatMulEncDecValAttShape(const NodeAttrs& attrs, return true; } +void strided_batch_sgemm(bool transA, bool transB, Review comment: I don't expect this function will be widely used except these mha operators. Also linalg* functions need high level abstraction but I'm not sure how to wrap the pointers a/b/c here into Tensor while still keeping the m/n/k and strides arguments, and how to implement the gpu part. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
