Siyuan (Ryans) Zhuang

This affects: https://github.com/ray-project/ray/pull/22553

> A small nit: many of our previous comments followed the 80-character limit. Is it possible to stretch them automatically as well? Currently seems not supported. I can do it...

I can update this PR after #763 and #759 get merged to minimize the trouble for re-formatting.

accidentally request again... so I convert it to the draft

对，的确是floor，不是round。。。然而不知道为啥我写题解时却发现行为和round一样。。。惊出了我一身冷汗。。。这样我就可以放心地改回去了

@0rzx 这里还有一个比较科学的解释： 在题目的生成源代码里面运行： ```python print((np.mean(targets, axis=0).astype(np.uint8) - np.mean(targets, axis=0)).mean()) print(images.mean() - images[choices].mean()) ``` 会得到 ``` -0.47768702651515166 2.2311801805161053 ``` 虽然由于取整，带来了一个大约是 0.5 的负的 bias，但是由于样本选择问题，导致平均像素值低了2.2311801805161053，后者对于50000个样本而言是一个巨大的bias。如果此时intercept设为0，那么显然 Lasso 不得不将这个 bias 加权到其它图片上面，因而不能得到解，甚至导致正确答案的权重严重偏低。

Because ray-autoscaler is using it. For GCP and AWS, their CLIs are already installed.

hmmm, it is mostly used by ray autoscaler for monitoring

Oh, I just follows the original CLI behavior, where enter in the empty line exits the console. This indeed sounds not user friendly. I'll improve it. Also the suggestion from...

Great job! Can you resolve the conflict?