Open liweintu opened 1 year ago
To add qibotn as a backend, there're two options.
.benchmarks/libraries/qibo.py
, extend the Qibo class to support qibotn
as a new backend..benchmarks/libraries/
, create a new file qibotn.py
to add qibo-supported TNet backends.IMHO, if we want to package both state vector and TNet support into qibo, option 1 is favoured.
@AleCandido Which one do you recommend?
Given that qibotn
is somehow falling under the hat of Qibo, I would see option 1 as the favored alternative as well, exactly for the same reason you are arguing.
However, in this context I'm not an authoritative source, I believe we should ask for the opinion of @stavros11 (who mostly developed the benchmarks and a large part of Qibo itself), and of course @scarrazza.
Given that
qibotn
is somehow falling under the hat of Qibo, I would see option 1 as the favored alternative as well, exactly for the same reason you are arguing.However, in this context I'm not an authoritative source, I believe we should ask for the opinion of @stavros11 (who mostly developed the benchmarks and a large part of Qibo itself), and of course @scarrazza.
Sure, fair enough.
Hi @stavros11 , @scarrazza , do you have a suggestion about which option we should pick to incorporate TNet backends into the benchmarking? Thanks.
Hi @stavros11 , @scarrazza , do you have a suggestion about which option we should pick to incorporate TNet backends into the benchmarking? Thanks.
I think the idea of qibotn is that it will be an additional qibo backend (see for example qiboteam/qibotn#7), so I would also prefer option 1. Option 2 may be faster to implement now but integrating qibotn with qibo shouldn't be very hard either and doing so will also expose the new backend to all qibo users which could be useful for other applications (other than benchmarks).
I would suggest to give a try to implement the backend in the qibotn repository, for example following the qibojit numba backend or qibolab as template, but I can also provide more details and help with how to do this. This will be the first non-state vector simulator, so it will be a good exercise to see if our backend abstractions are sufficient to support it.
Agreed, then we'll go ahead to follow qibojit
backend as the template. Thanks for the suggestion @stavros11 .
We add the support for the benchmarking of tensor network backend in simulation. First quimb, then cuQuantum.