* llama4 support for linearized experts * clean up fsdp2 sharding to prevent hang * add yaml config * cleanup example [skip ci]
Wing Lian