Hello Vladimir,

The Apache MXNet team has accepted your talk. We are very excited to have you 
present on this event. Your talk has been scheduled at 14:16 PM PST on Dec 14th 
for 15 minutes.

The agenda/time is still evolving. We will inform you if there are any changes.
Here is the updated website -  https://s.apache.org/MXNet2020

Action Required:

  *   We request you to send your recorded talk (15 minutes) to us by Dec 6th 
midnight PST.
  *   Please respond by end of tomorrow to confirm your presentation and 
attendance.

Question Answers:
We plan to have a slack channel in parallel to your recording where you can 
respond to questions while your recording plays. We will add the slack channel 
names to the respective talks in a week.

Please reach out if you have any questions. We will send out the webcast 
details by end of Wednesday.

Thank you
Vartika





From: Vladimir Cherepanov <[email protected]>
Date: Friday, November 20, 2020 at 11:23 AM
To: apachemxnetday <[email protected]>
Subject: Talk proposal - MXNet AMP (Automatic Mixed Precision)

MXNet AMP (Automatic Mixed Precision)

Automatic Mixed Precision allows some parts of a model to run in fp16 mode, 
which improves performance without loosing numerical precision. In MXNet 2.0 it 
is now a core feature. MXNet AMP is implemented by classifying all ops with 
respect to their ability to be executed in fp16 mode, and casting their inputs 
accordingly. The casting is currently implemented by monkey-patching Python op 
functions. We'll present some benchmarking / performance gain examples. AMP 
benefits are very model dependent. Some real challenges - significantly 
increased memory consumption in the case of RNNs - a fix is understood, but is 
still to be implemented. Some perceived - vanishing loss scale - can happen 
during automatic loss scaling - is an indication of numerical instability, 
rather than an AMP-specific problem - AMP just exposes it. Future development 
includes, among other things, layout management (implemented in the container, 
but not yet upstreamed). There's also a strong early feedback to add API to be 
able to turn AMP off - a natural way appears to be to make it controlled by a 
Python context manager.

Vladimir

Reply via email to