It's been mentioned that this model is MLA capable, but it seems like the default vLLM params don't use MLA. Seeing ~0.91MB KV Footprint per token right now. Are you getting MLA to work?
The thing is the CPSC is not killing the demand for the magnets. There are a billion places you can still purchase them online. In fact you can purchase buckyballs in their original packaging at better prices than before from Chinese companies who are simply ripping off the IP of a dead company. The war on magnets faces the same problem as the war on drugs. And similarly, the solution is education, not prohibition.
