Fabrizio Milo

Also in Linux it takes a lot of time. Any clue why? ```bash Top 10 plugins slowing vim's startup ==================================== 1 31.292 vim-conda 2 25.968 nerdtree 3 17.585 vim-textobj-entire 4...

Any directions on where we should check for the compatibility?

tdqm is not in conda but is in pip `pip install tqdm`

I just noticed in the startup logs that `/health` is enabled

exactly, the endpoint is there but is not documented. I changed the title of the issue.

@galenzhao where should the code that you posted go? is a patch to the api library ?

I was just looking in this direction to bring back the size of the model towards a consumer-level size. Seems deepspeed's team has done some work on that front although...

I can figure out the details, it really depends on what hardware we have available, if cloud / bare metal or k8s.

Not sure, which benchmarks are we running, and on what hardware?

oh, I see, you mean the accuracy of the model is not the same when the weights are merged. Yes this is still the case with this PR.