tqchen commented on pull request #5779:
URL: https://github.com/apache/incubator-tvm/pull/5779#issuecomment-643646043


   I don't think we disagree in many cases. 
   
   Let me try to rephrase @t-vi's  the dtype part of the arguments:
   
   - F0: NLP and quantization worklaods needs mixed precisions(fp16, i8, i32).
   - F1: In order to support NLP workloads well, we will need to able to handle 
other data types well other than fp32.
   
   I think we all agrees to F0 and F1.  Now one of @masahi 's argument is 
   
   - K0: Most users tries to use fp32 by default, it would be nice to keep 
things to be compatible to that interface when adding dtype support.
   
   As we can see that K0 do not necessarily conflict with F0 and F1. As an 
outsider to the discussion, it is a bit harder for me to express my 
thought(without looking at the code). It would be easier if we can list 
interface candidates during the discussion, and discuss their pros and cons. 
There is certainly a tradeoff we need to make, in terms of ease of use, level 
of compatibility etc.
   
   I need a bit more time to dig in order to understand the name argument. But 
at a high level (this is my opinion) I think it makes sense to inheritate names 
from the source(if they are available) while allow users to override them. The 
main reason why certain name/shapes are requirement in frontend is that many 
cases these information are incomplete. Again having interface candidates will 
be helpful here.
   
   The main reason for a discussion is not necessarily argue for which side is 
right, but actually to clarify the intent, and reach concensus(sometimes both 
sides actually already agrees to). 
   
   So that in the future others can look into it. Also in cases like this 
discussions are also important to find out the best ways for integeration (e.g. 
not about whether or not shall we do dtype handling, but how to best do them).
   
   Discussions also serves the good purpose on learning. It would be great for 
everything to share the  understanding  of the "status quo" and contentious 
points. Sometimes the problem of the disagreement is not what we need to do to 
support these, but how to best resolve the situation. 
   
   We can certainly create followup discussions in the forum.
   
   
   


----------------------------------------------------------------
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]


Reply via email to