Inlay

I don't see the cost of memory coming down until we replace Multi-head attention with linear complexity methods such as those in the linear recurrence class (gated delta nets, etc.). We really need more work in that space.