[question] 请问 KExpertsCPU 的weights_cpu dtype为什么是torch.float32, 而不是torch.bfloat16等其他类型？

1.请问 KExpertsCPU 的weights_cpu dtype为什么是torch.float32, 而不是torch.bfloat16等其他类型？
2.另外KExpertsCPU.forward() 内部weights: weights = weights.contiguous().to(torch.float32).cpu(), 也是使用的torch.float32, 请问这个是为什么？
请帮忙解答，谢谢。

Start by locating KExpertsCPU and reading KExpertsCPU.forward(), focusing on weights_cpu and the conversion to torch.float32 on CPU. Document the rationale for using torch.float32 instead of torch.bfloat16 or other dtypes, addressing both questions in the issue.

Written by the indexing model from the issue text.