Tian, Feng

mkldnn has changed its .so location, which is under lib64 under centos, lib under ubuntu. as you are using prepare_env.sh with gcc compiler, it will go to make path rather...

it's old version, has been updated at 2016. pls refer to latest readme/wiki to do multi-node training.

why it's a useful option for your case? it's just preallocate blobs and same with prefetch of new data layer. did you compare your perf running on old multi-node code...

the latest data reader has considered this. there is no such option needed.

thanks Maciej, yes, this is what we prefer. thanks for the quick update

> We still have to work on the `NASConfig` class, as it currently does not have an option to pass Dynas/NASBase objects to it, but given the timeline I think...

@lvliang-intel conflict needs to be solved

@jgong5 "compression dim" means the 4bits weight compression direction, either along output channel or along input channel. "zero point dim" means the 4bits zero point compression direction, same with "compression...

per offline discussion, we decided to take solution 2, that's IPEX to be directly compatible with HF format, INC will also generate HF compatible format in export_compressed_model().

I was confused on that as well. here is my understanding: (pls correct me if anybody has insight on this) The VNET was for PROMISE12 competition. this competition doesn't share...