mxnet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Lin Yuan <apefor...@gmail.com>
Subject [PROPOSAL] Large tensor support in MXNet
Date Mon, 03 Dec 2018 06:20:33 GMT
Dear Community,

As some of you may have already encountered, MXNet has a limitation in
supporting tensors of size greater than ~4.3 billion elements (2^32). The
root cause is because in MXNet backend 32-bit integer type is used as the
default integer data type for both computation and storage in many places.

Uplifting this limitation, however, is not simply replacing all 32-bit
integers by a larger data type (64-bit integer) as I detailed them in the
design proposal at
https://cwiki.apache.org/confluence/display/MXNET/Large+Tensor+Support. It
requires a systematic approach to address this problem in MXNet backend as
well as all APIs in different language bindings.

I will appreciate your suggestions in solving this problem systematically
and elegantly, as well as your help to support in different language
bindings other than Python.

Please add your comment in the design proposal or create tickets in the
JIRA epic:  https://issues.apache.org/jira/browse/MXNET-1184

Best Regards,

Lin

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message