[eigenvalue]

Fair enough, but that still uses a lot more memory during training than what DeepSeek is doing.