Ross Wightman

@gabrielilharco that looks li ke a hardware/cluster error not a code error, you can't add the None logic, all but the master are supposed to be None and ALL nodes...

it's the first distributed call now so if there are any issues w/ the interconnect it'll break there first instead of the first step, EDIT: 'any major issues', still can...

you can try commenting it out completely and doing a manual resume to see if resume + training works w/ the same node allocation and no broadcast_object, possible I got...

Also pinging @gmittal and @apsdehal from FLAVA (#218), it looks like this will be a big merge prior to FLAVA being ready. So, some decisions here will impact that PR....

> would appreciate a second look @rwightman k, will take a closer look soon. A higher level question, are se satisfied with the CoCa text gen results so far? I...

@rom1504 this is still most significant change so I don't see the rush but will start picking through, conflicts are small relative to the commitment to maintain and continue getting...

if there's a commitment from @gpucce @iejMac to keep pushing this to a desired performance level once it's merged (and not just let it dangle), we can make a push...

looking better! for fine-tune, how much extra code is it? I feel fine-tune falls closer to a train focused repo than benchmarking?

Has the coca model been tested to work with torchscript (jit=True) in inference mode? Come to think of it has the main CLIP model been tested with the changes? Thinks...

@fffffgggg54 the FeatureDict methods can be used (I've been meaning to do this as an example for swin / maxvit / etc) but the swin ones need some re-org. You...