Josh Bleecher Snyder

Actually, I have a follow-up. :) One of the factors is listed as the size of the IDs. But IDs are strings, which means that there's also a 2 word...

> I don't like how it complicates graph instantiation Seconded. But OTOH this is exactly the sort of scenario in which generics shine.

I’m skirting the edge of what I can do now, at 500gb of RAM. So for me, it’d be pretty useful; I could use uint32s and scale up. But I...

new api looks great. closing this again. :)

This does not reproduce using the upstream llama3 tokenizer.model and tiktoken.

> Zero3 should handle frozen modules. I think the trouble is that range freezing relies on having shape information available, and once deepspeed has wrapped the model, that shape information...

I think there are going to end up being some interesting design challenges here, for which we are going to have to develop principles as we go. This came up...

>> "any currently-online node with tag T". > What does this represent? A specific service? Yes, exactly. (Or slightly more precisely: A specific service in a specific environment, like prod...

I don't have a need for intersection of tags, but that sounds not unreasonable to me. I do definitely still want this feature, though. :)

I plan to keep running this for a little while longer, gathering data, but I thought I would share it in case anyone else wants to play with it. (I...