autossh exit when working background

 Try to add -N to your arguments.

autossh -M 0 -o "ServerAliveInterval 30" -o "ServerAliveCountMax 3" -f -N -R :5678:127.0.0.1:5678 root@<server_ip>

Through a socks5 proxy

autossh -M 0 -o "ProxyCommand nc -x 127.0.0.1:7891 -X 5 %h %p" -o "ServerAliveInterval 30" -o "ServerAliveCountMax 3" -R :5678:127.0.0.1:5678 root@<server_ip>

Change default user of WSL1&2

Recently, I moved my WSL Ubuntu distribution out of my C drive due to insufficient disk space.

After the "export-import" operation, my default login user for WSL changed to root.

Normally, as a virtual environment, it shouldn't bother me at all. However, my Python packages are installed under a normal user account, which was the default login user.

I tried to find the /etc/wsl.conf file but failed, so I decided to do a small trick to save time.

 Append following code to /root/.bashrc.

PP_NAME=`ps -p $PPID -o comm=`
#echo "PP_NAME: $PP_NAME"
if [ "$PP_NAME" == "init" ]
then
    su <YOUR_USERNAME>
    exit
fi

But it only works when we start wsl from cmd. VSCode's wsl will still start with root user.

Thus, we still need to use the /etc/wsl.conf .

If there is no /etc/wsl.conf, then create it. And fill the file with:

[user]
default=username

If you are using WSL 1, just close all your WSL instances. And restart WSL.

If you are using WSL 2, you need to use 'wsl --shutdown' command to stop your WSL2 VM first, then restart your  WSL.

ChatGPT对于普通人来说，是个灾难。

 

最近ChatGPT的火爆，让所有人，包括我们这些外行都见识到了大模型的厉害。

于是，根据历来传统，想要自主可控的欲望越发的强烈。笔者在自己的老爷机器（i7 4790，32G RAM，gtx1080 8G）上尝试跑了一下GPT2的预训练模型（如果有人感兴趣，可以给大伙弄个简单的流程说明）。

效果嘛：

典型的人工智障。

来看看ChatGPT是咋说的：

---

为啥和ChatGPT差距这么大？

主要原因还是在模型规模上。

为啥不搞个更大的模型啊？

因为笔者不配。。。

我们知道，神经网络或者是深度神经网络的三大要素：模型、数据以及算力。

就GPT来说：

• 模型基本公开。
• 数据就是自然语言。麻烦主要在标注部分，但是一定能解决（毕竟笔者这样的廉价劳动力有的是）。
• 算力，也就是我们的炼丹炉，在大模型上来说就比较难搞了。我们来讨论一下。

目前chatgpt的GPT-3.5还没有具体的信息。所以我们还是以目前公开领域能看到的GPT-3为例。这玩意有1750亿的参数。这是个什么概念呢？

我们用有具体数据的开源模型来估计吧。一个60亿参数的GPT-J模型，其预训练参数的zip包大概60GB。一款3A大作也就这样了。1750亿，扩大了将近30倍！

有人可能会觉得：还不到2TB，现在硬盘也不贵，这有什么？

小伙汁，这就naive啦。这玩意做推理和训练的时候，是要载入内存哒。而且，这么大的模型，要在可接受的时间内计算出结果，用CPU也是不行哒，要用并行计算来加速。一般来说就是GPU了。

仍旧以GPT-J为估算基础，加载这个模型内存要40-50G以上，显存要在13G以上。使用rtx3090 tuning（注意，是tuning，还不是training），需要12个小时。之后的模型参数还会膨胀一些，大概要到80G。

那么GPT-3炼丹大概会是一个什么地狱场景，想必大家心里也应该有数了吧（据说需要几百块A100，训练时间以天计算）。更可怕的是GPT-3现在已经过时了。。。

chatgpt背后的模型据说比GPT-3还要大。这就是我之前文章中认为中小企业无法拥有自己的LLM的主要原因。

分析了一圈。一块A100，便宜的也要6万多。对于一个普通人来说，别说数百块，一块就破产。想要自己跑一个可用的GPT，基本是没什么希望的。

ChatGPT的所作所为大伙最近应该已经看了够多的了。作为一个普通人，要想不被AI卷死，就只能用大企业的服务，老老实实的交AI税。但是别无选择的服务和奴役又有什么区别？

为qbittorrent-nox配置jackett搜索插件(config jackett search plugin for qbittorrent-nox)

1 download jackett plugin(installing the url in webui directly won't work!)

https://raw.githubusercontent.com/qbittorrent/search-plugins/master/nova3/engines/jackett.py

2 move jackett.py to 

~/.local/share/data/qBittorrent/nova3/engines/, or ~/.local/share/qBittorrent/nova3/engines/

3 create a jackett.json file in the same directory with the content below:

{
    "api_key": "YOUR_API_KEY_HERE", 
    "tracker_first": false, 
    "url": "http://127.0.0.1:9117"
}

4 restart your qbittorrent-nox. 

Install docker on ubuntu [Short version]

 Tl;Dr

sudo apt-get remove docker docker-engine docker.io containerd runc

curl -fsSL https://get.docker.com -o get-docker.sh

sudo sh get-docker.sh

 

For ones who have plenty of time to wast :

 https://docs.docker.com/engine/install/ubuntu/

 

Config dockerd to use proxy:

 

sudo mkdir -p /etc/systemd/system/docker.service.d

sudo vi /etc/systemd/system/docker.service.d/http-proxy.conf 

[Service]
Environment="HTTP_PROXY=http://192.168.1.66:7890"
Environment="HTTPS_PROXY=http://192.168.1.66:7890"
Environment="NO_PROXY=localhost,127.0.0.1,192.168.1.66,192.168.1.108"  

sudo systemctl daemon-reload
sudo systemctl restart docker 

sudo systemctl show --property=Environment docker 

Extract shsq

 

As we know the normal Magic number is 0x73717368("hsqs").

But some vendors modified the squashfs format to support LZMA, and changed the magic to 0x71736873("shsq").

But now, squashfs has already supported LZMA. Thus we can modify our binary image and let unsquashfs do its work.

1. Open image with a hex edtor or create your own tool.
2. Change the magic "shsq" to "hsqs" at the beginning of the image.
3. Change compression_id(image_header[20]) from "1-GZIP" to "2-LZMA". 

4. run unsquashfs as usual.

ref:https://dr-emann.github.io/squashfs/ 

Run arm rootfs with QEMU user

Extract or mount your rootfs to a dir. For example: ./rootfs

Install qemu-user-binfmt.

This package will tell kernel run arm ELF with qemu-user.

sudo apt install qemu-user-binfmt

To see what are installed:

ls  /proc/sys/fs/binfmt_misc/

cat /proc/sys/fs/binfmt_misc/qemu-arm

Now, we can find that the interpreter of arm ELF is changed to /usr/bin/qemu-arm-static.

The interpreter could be /usr/bin/qemu-arm depending your installation order of QEMU packages. 

When we call chroot, kernel will use the interpreter to run ELFs. So we need put qemu-arm-static into the chrooted "root".

If your interpreter is /usr/bin/qemu-arm-static:

cp `which qemu-arm-static` ./rootfs/usr/bin/

If your interpreter is /usr/bin/qemu-arm:

cp `which qemu-arm-static` ./rootfs/usr/bin/qemu-arm

 OK, we can chroot now.

chroot ./rootfs/ /bin/ash

Enjoy!