Let me be a bit more specific. There will be three modes that such
integration is possible
- Allow TVM function to declare a function that can be called imperatively
- Allow MXNet symbolic graph to be compiled to a standalone TVM module(no
mxnet runtime dependency) that can be deployed to the device you like with
support for OpenCL/Metal/ javascript etc.
- This is called AOT mode and is usually good for deployment, this is
what we will be PR next week
- Allow mix TVM compiled function with Gluon operator, and possibly shift
some of mxnet runtime to work with TVM runtime to support OpenCL/Metal
directly in MXNet. This should be the longer term and needed if we want to
do training
Hope this clarify your question
Tianqi
On Thu, Aug 17, 2017 at 5:58 PM, Dominic Divakaruni <
[email protected]> wrote:
> Congratulations!!
>
> Can you share what you are thinking with regards to how you propose to
> integrate TVM into MXNet?
>
> On Thu, Aug 17, 2017 at 12:52 PM Minjie Wang <[email protected]> wrote:
>
> > Great news! This is a huge step towards a highly efficient deep learning
> > system that is portable on different hardware. Thanks Tianqi and the
> > efforts of all the contributors.
> >
> > On Thu, Aug 17, 2017 at 3:41 PM, Tianqi Chen <[email protected]>
> > wrote:
> >
> > > Hi Guys:
> > > I am super excited to announce DMLC/TVM our deep learning
> compilation
> > > stack. There will be followups on mxnet to add the official support
> soon.
> > > To check what it is, see the announcement
> > >
> > > http://tvmlang.org/2017/08/17/tvm-release-announcement.html
> > >
> > >
> > > Tianqi
> > >
> >
> >
> >
> > --
> > Minjie Wang
> > *New York University | Computer Science*
> > 715 Broadway, New York, NY, 10009
> >
> --
>
>
> Dominic Divakaruni
> 206.475.9200 Cell
>
