Flash Attention is almost mandatory for LLM training and inference because of the speedup it brings. Usually pip install flash-attn is enough, but for specific hardware (e.g., A800) or bleeding-edge features you sometimes have to build from source.

After a few rounds of frustration, here’s the end-to-end process I used to build flash-attn from source on an A800 server, plus the traps to avoid.

Environment

Hardware/Software

• Server: 64-core CPU, 512G RAM
• GPU: NVIDIA A800 80G (PCIe)
• CUDA: 12.2
• PyTorch: 2.7.1+cu118

Official README

Requirements:

• CUDA toolkit or ROCm toolkit
• PyTorch 2.2 and above.
• packaging Python package (pip install packaging)
• ninja Python package (pip install ninja) *
• Linux. Might work for Windows starting v2.3.2 (positive reports), but Windows builds need more testing.

* Make sure ninja is installed and works (ninja --version should exit 0). If not, reinstall (pip uninstall -y ninja && pip install ninja). Without ninja, builds can take ~2h single-threaded; with ninja, 3–5 minutes on a 64-core box.

To install:

1	pip install flash-attn --no-build-isolation

Or build from source:

1	python setup.py install

If your machine has less than 96GB RAM and lots of CPU cores, limit parallel jobs with MAX_JOBS.

A small rant: “3–5 minutes on a 64-core machine” must be on some dream hardware…

Core Build Steps

Upgrade core build tools

Keep pip, wheel, and setuptools current to dodge weird slowdowns.

1

pip install pip wheel setuptools -U

Old setuptools can be painfully slow—ask me how I know.

Install and verify ninja

The README stresses ninja. Without it, the build falls back to single-threaded and drags on forever.

1. Install

   1	pip install ninja

2. Verify

   The official check is ninja --version followed by echo $? for exit code 0. I prefer a visual check: start the build, open another terminal, run htop, and see dozens of cicc/nvcc processes pegging the CPUs. If not, reinstall ninja with pip uninstall -y ninja && pip install ninja.

   

Set critical env vars

These variables decide whether the build succeeds and how fast.

1. FLASH_ATTN_CUDA_ARCHS: limit architectures

   Tell the compiler which compute capability to target. A100 and A800 are compute capability 8.0, so set it to 80. Without it, the build will generate code for many architectures—wasting time.

   • Lookup: NVIDIA CUDA GPUs

   1	export FLASH_ATTN_CUDA_ARCHS="80"

2. MAX_JOBS: parallel build workers

   Controls how many CPU cores ninja uses. More cores = faster but also more RAM.

   • My run: On 64 cores / 512GB RAM, I set MAX_JOBS=64.
   • Result: 64 cores maxed, peak RAM >300GB, build took ~10–15 minutes. When RAM usage in htop drops, the build is almost done.
   • Tip: If RAM is tighter, start with half the cores, e.g., MAX_JOBS=32.

Install

Combine everything and build:

1

FLASH_ATTN_CUDA_ARCHS="80" MAX_JOBS=64 pip install flash-attn --no-build-isolation

--no-build-isolation is critical so the build process can see the env vars you exported in the current shell.

Troubleshooting

• g++: fatal error: Killed signal terminated program cc1plus

  • Cause: OOM (too many parallel jobs). Lower MAX_JOBS (e.g., 64 → 32 → 16).

• GCC too new

  • At publish time, the build requires gcc-11/g++-11.

  • Fix (Ubuntu 22.04 example):

    1 2 3 4 5

    apt install gcc-11 g++-11 which gcc-11 which g++-11 export CC=/usr/bin/gcc-11 export CXX=/usr/bin/g++-11

• Build takes forever

  • Likely ninja isn’t actually used and the build is single-threaded. Verify/reinstall ninja.

• CUDA-related build errors

  • Usually FLASH_ATTN_CUDA_ARCHS is wrong or PyTorch/CUDA toolkit versions don’t match. Double-check GPU model vs. compute capability, and compare nvidia-smi with nvcc --version.

Alternatives

If source builds are too painful, try:

1. Prebuilt pip wheels: pip install flash-attn sometimes works out of the box, but versions may lag and hardware support isn’t guaranteed.
2. conda: conda install flash-attn -c conda-forge might provide a prebuilt package.

Thanks

Flash Attention on GitHub

Issues in the official repo

Goal: use Visual Studio Code (aka vscode) to develop on a remote server over SSH with a local-like experience. Notebook support is especially smooth with the VS Code notebook extensions.

For consistency, I’ll use these terms:

• Visual Studio Code → vscode
• Remote server (usually Linux) → remote server
• Local machine (Windows/macOS/Linux) → host
• Remote-SSH extension → remote extension
• Jupyter Notebook → notebook
• Extensions → plugins

VS Code’s success owes a lot to open source and a vibrant extension ecosystem. Remote development is powered by the Remote-SSH plugin.

Remote dev isn’t limited to SSH:

• Windows host to local WSL
• Local Docker containers
• GitHub Codespaces

First step: download VS Code and install the remote extension. After installation, a remote icon appears in the lower-left corner.

If the extension fails to install, update VS Code or reinstall.

Configure Server Info

Click the lower-left remote icon and choose the appropriate target. For SSH, pick Connect to Host. Two ways to configure:

Method 1: ssh command

Enter a full ssh command directly. Replace username with your account, host with IP or domain, port with the SSH port. For a jump host, use -J with jump user/host/port.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

# Default port 22
ssh username@host:port
# Examples:
# ssh yuesir@127.0.0.1:22
# ssh yuesir@vccv.cc:22

# With jump host
ssh username@host:port -J username@host:port
# Example:
# Jump host A: user yuesir, host 192.168.0.1, port 22
# Target B: user yueqingyou, host 192.168.0.2, port 2222
# ssh yueqingyou@192.168.0.2:2222 -J yuesir@192.168.0.1:22

Method 2: config file

After connecting once, VS Code writes the config into your SSH config. Choose Connect to Host -> Configure SSH Hosts, usually ~/.ssh/config (e.g., /Users/yuesir/.ssh/config).

Two templates (with/without jump). Host is just a nickname; HostName is the actual host/IP.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

# No jump host
Host MyServer
  HostName <server-host>
  Port <port>
  User <username>
  
# With jump host
Host TargetServer
  HostName 192.168.0.2
  Port 2222
  User yueqingyou
  ProxyJump yuesir@192.168.0.1:22

If it hangs on “Downloading VS Code Server”

This happens when the remote can’t reach the VS Code server URL (offline or blocked). Workaround: download the server locally and upload manually. References:

• Offline install VS Code Server
• Offline install VS Code Server + SSH keys
• VS Code Server download stuck

Configure SSH Key Login

Typing passwords repeatedly is painful, especially with flaky VPNs. Use key-based auth—it’s also what VS Code recommends:

We recommend using key-based authentication (otherwise you may be prompted multiple times).

Check for existing keys on the host

1
2
3
4
5

# Check for .ssh directory
ls -la ~ | grep .ssh
# If present, check for key files
ls -la ~/.ssh | grep id_rsa
ls -la ~/.ssh | grep id_ed25519

If you already have keys, skip creation.

Quick crypto note: RSA vs Ed25519. Summary:

Command	ssh-keygen -t rsa -b 4096	ssh-keygen -t ed25519
Type	RSA	Ed25519
Performance	Slower (still fine)	Faster
Security	High	Higher
Compatibility	Better with legacy systems	Very old systems may not support

Default to ssh-keygen -t ed25519.

Create a key pair on the host

On Windows open cmd/PowerShell; on macOS/Linux use Terminal. Example on Windows 10:

Prefer Ed25519, fall back to RSA:

1
2
3
4
5

# Ed25519
ssh-keygen -t ed25519

# RSA
ssh-keygen -t rsa -b 4096

Press Enter to accept defaults and empty passphrase. Keys are stored in ~/.ssh (e.g., C:/Users/<user>/.ssh or /home/<user>/.ssh). You’ll see id_ed25519.pub (public) and id_ed25519 (private). Keep the private key secret.

Add the public key to the remote server

Cloud providers often let you paste/upload a key when provisioning (see screenshot below).

If you just have SSH access, use the command-line or manual edit:

1
2

# From host
ssh-copy-id -i ~/.ssh/id_ed25519.pub username@host:port

Or manually append the .pub content to ~/.ssh/authorized_keys on the server (create the file if missing), then set permissions:

1
2

chmod 700 ~/.ssh
chmod 600 ~/.ssh/authorized_keys

Restart the SSH service if needed to apply key auth.

Recommended Plugins

Beyond Remote-SSH, these make Python/Notebook life easier:

• ms-python.python (Python tooling, debugger, linting)
• ms-toolsai.jupyter (Notebook support)
• ms-toolsai.jupyter-renderers
• ms-toolsai.jupyter-keymap
• ms-toolsai.jupyter-cell-tags
• ms-toolsai.vscode-jupyter-powertoys
• ms-toolsai.datawrangler (data wrangling)
• ms-toolsai.data-wrangler-vscode
• ms-toolsai.data-wrangler-native-notebook
• ms-toolsai.data-wrangler-lsp
• ms-toolsai.data-wrangler-insiders
• ms-toolsai.data-wrangler-debug
• ms-toolsai.data-wrangler-darwin-x64
• ms-toolsai.data-wrangler-darwin-arm64
• ms-toolsai.data-wrangler-windows-x64
• ms-toolsai.data-wrangler-windows-arm64

The last few are platform-specific binaries; VS Code will fetch the right one automatically.

Final Thoughts

Once SSH and keys are set, connecting is just a click. With the Python and Jupyter extensions, you get a near-local dev experience while running code on the remote server. If you edit authorized_keys with vim, save and exit with:

1
2
3

:wq
# or
ZZ

Close VS Code and reconnect—if it no longer prompts for a password, key auth is working. The rest focuses on language-specific setup and notebooks.

Python

Developing remotely still requires installing language plugins on the remote side. VS Code will prompt you; just ensure you install to the server, not the host.

I recommend managing environments with Conda:

For Notebook + R/Python integration, see the next section.

Notebook

I’ve written several notebook setup posts; note some target the older 6.x series (current releases are 7.x).

If you’re new to notebooks, see the install guide and official docs.

If you’d like a detailed walkthrough, drop a comment (I’ll try not to bail ~~).

R + Notebook

R isn’t limited to RStudio / RStudio Server. Paired with Notebook, visualization can be even smoother.

Python + Notebook

Some common issues are recorded here, though VS Code’s Notebook UI already solves most of them.

Thanks

VS Code Remote Development

Remote development over SSH

Deep learning projects rely on a web of dependencies: torch, transformers, cuda, and many more. Version mismatches often lead to obscure compatibility issues, making it hard to faithfully recreate an environment on a different server.

I already had a stable deep learning Conda environment on one server and wanted to migrate it to another. The easiest path was conda-pack, which can bundle a Conda environment into a single archive that works out of the box on another Linux host.

Export

To move the original environment off the source server, export it first.

1. List all environments

   Use conda env list to see every Conda environment on the machine. Make sure the target isn’t the default base. If you need to move base, clone it to a new name with conda clone first.

   1	conda env list

2. Activate the environment to migrate

   Suppose the environment name is torch:

   1	conda activate torch

3. Install conda-pack

   Install inside the activated environment:

   1	conda install conda-pack -c conda-forge

4. Pack the environment

   Run conda pack -n <env> to generate a tarball. Replace <env> with your environment name, e.g. torch.

   1	conda pack -n torch

   The command creates torch.tar.gz in the current directory.

Packing time scales with environment size. My torch env is large, so the job took a while—thankfully there’s a progress bar. The environment consumed ~20GB and the packed file landed just under 10GB.

Import

After packing and downloading the archive, upload it to the target server and unpack it.

Upload the file to the target

1. Download locally, then upload

   Keeps a local backup: download torch.tar.gz to your laptop, then upload to the target. Use whatever SSH/SFTP tool you like.

2. Use rsync directly

   Transfer straight between servers:

   1	rsync -avz torch.tar.gz user@target_server:/path/to/destination

Restore the environment

1. Ensure no environment with the same name exists

   1	conda env list

   If a duplicate exists, remove it first:

   1 2 3

   conda env remove -p <env_path> # or conda env remove -n <env_name>

2. Prepare the destination directory

   Check where Conda keeps environments, e.g. /mnt/nfs/yuesir/miniconda3/envs. Create a folder matching the environment name:

   1 2	mkdir /mnt/nfs/yuesir/miniconda3/envs/torch chmod 777 /mnt/nfs/yuesir/miniconda3/envs/torch

3. Extract the archive

   1	tar -xzvf torch.tar.gz -C /mnt/nfs/yuesir/miniconda3/envs/torch

4. Activate the new environment

   1	conda activate torch

Verify

After importing, run a few checks.

1. List environments

   1	conda env list

   You should see the migrated environment in the list.

2. Activate it

   1	conda activate <env_name>

3. Inspect installed packages

   1	conda list

   Confirm versions and dependencies look right.

Extra

For deep learning workloads (e.g., PyTorch) that depend on specific driver versions, verify CUDA driver compatibility between the two servers. CUDA drivers are backward compatible—if they don’t match, either upgrade the target’s driver (root required) or downgrade PyTorch (risky because of dependency complexity). In tricky cases, a clean install might be safer.

For choosing CUDA/PyTorch combos, see my earlier post:

Thanks

Using conda pack for environment migration

I used to use Hexo framework. But with the increasing number of articles, custom projects are more and more, coupled with the NPM plug-ins are in constant conflict, have not dared to arbitrarily install new plug-ins, the difficulty of maintenance has risen greatly. Former Hexo blog project file has been a “shit”.

As the saying goes, “break and then stand up”, taking advantage of this period of time is relatively free, I directly switch to the Hugo framework. My first and deepest impression of Hugo is that it is very fast, more than twice as fast as Hexo, of course, I also found some of the places that are not adapted to: the original blog webp image compression and watermarking features are automatically generated by the plug-in, this is the use of Hugo’s highly customizable and rich image processing api restore the blog’s image structure.

Hugo Image Customization

Hugo has a Render Hooks function, you can replace some of the components to generate styles, which contains images resources. Here is the official introduction, if you are interested, you can go to learn more about it.

When rendering Markdown to HTML, render hooks override the conversion. Each render hook is a template, with one template for each supported element type :: Rendering hooks are used to render HTML to Markdown.

• Code blocks
• Code blocks
• Headings
• Images
• Links

For images, we need to create or modify the content oflayouts/_default/_markup/render-image.htmlfile for customization. If this is the first time you know about this concept, please first create this file in the appropriate path in your blog root directory, note that it is not in the theme theme directory.

Next I will give a reference article and code for each treatment, I will give detailed comments at the end of the article all the code, each subsection of the code is for reference only, similar to pseudo-code, can not be used directly.

Webp Compression

Show Results

The memory usage of a static blog is dominated by images, which can be more than 100kb, making the already fast blog even worse (the blog is hosted on Cloudflare, which is free, and the speed in China is impressive). The common png, jpg and other image files can be compressed into webp format to significantly reduce the footprint.

The image above, for example, takes up 39.35 kb before compression and 8.87 kb after compression.

Principle Analysis

Refer to Hugo’s method of automatically converting images to webp, which mainly utilizes theimage.Resizemethod, official docs means that you can specify the specific aspect and format.

Reference article added to recognize image EXIF orientation to prevent image orientation change after format conversion.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

{{ with $rawImage.Exif }}
  {{ if eq .Tags.Orientation 1 }}
      {{ $image = $image.Resize (print $image.Width "x" $image.Height " webp") }}
  {{ else if eq .Tags.Orientation 3 }}
      {{ $image = $image.Resize (print $image.Width "x" $image.Height " webp r180") }}
  {{ else if eq .Tags.Orientation 6 }}
      {{ $image = $image.Resize (print $image.Height "x" $image.Width " webp r270") }}
  {{ else if eq .Tags.Orientation 8 }}
      {{ $image = $image.Resize (print $image.Height "x" $image.Width " webp r90") }}
  {{ end }}
{{ end }}

Add Watermarks

Method Source

I basically didn’t find what I needed on the Chinese Internet, and finally found a three-year-old article in the blog of the author of the PaperMod theme (which is the theme used in this blog): Watermarking images with HUGO.

The main idea is to use the image.Filter and image.Overlay methods to overlay a watermarked image onto each image.

1
2

{{- $logo := $logo.Resize (printf "%.0fx jpg" $size) }}
{{- $image := $image.Filter (images.Overlay $logo (sub $image.Width $logo.Width) (sub $image.Height $logo.Height) ) }}

This requires a simple structure for the watermark image, and it is recommended to use a website logo or blog URL to create the watermark overlay. My watermark is made through this tool.

Code Show

The original author will cover the watermark to the lower right corner of the picture, I modified the centered transparent tilt placed, not too much impact on the reading experience, but also to better protect their image copyright.

1
2
3
4
5

{{ $size := math.Round (mul $image.Height 0.80) }}
{{ $size := cond (ge $size 250) $size 250.0 }}
{{ $logo := $logo.Filter (images.Opacity 0.5)  }}
{{ $logo := $logo.Resize (printf "%.0fx png r30" $size) }}
{{ $image = $image.Filter (images.Overlay $logo (div (sub $image.Width $logo.Width) 2) (div (sub $image.Height $logo.Height) 2)) }}

You need to calculate the length and width of the original image to determine the logo position, then use theimage.Opacitymethod to set the transparency, and finally center it.

Using Fancybox

My blog images usually contain a lot of textual information, and it would be very difficult to view the original image without zooming in. It so happens that fancybox is a js library that solves this kind of problem. The official website description is as follows:

Fancybox is the ultimate JavaScript lightbox alternative that sets the standard for premium user experience in multimedia display. Supports images, videos, maps, inline content, iframes and any other HTML content.

Reference Tutorial

The functionality is very powerful and will be shown here only for the image gallerygallery function. This step happens to be referenced in a previous tutorial I wrote for the Hexo Cactus theme:

Code Show

Pay special attention to the first need to import the relevant css file and js file, and then modify the img upper layer of the a tag.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

<a data-fancybox="gallery" data-src="{{ $image.RelPermalink }}" data-caption="{{ .Text }}">
    <picture>
        <source type="image/webp" srcset="{{ $image.RelPermalink }}">
        <img
            width="{{ $rawImage.Width }}px"
            height="{{ $rawImage.Height }}px"
            style="width: 100%; height: 100%; max-height: 36rem;"
            loading="lazy"
            src="{{ $rawImage.RelPermalink }}"
            alt="{{ .Text }}"
            {{ with .Title }} title="{{ . }}" {{ end }}
        />
    </picture>
</a>

Add the picture tag is to adapt old devices that can’t parse webp files to display the original image and improve code compatibility. The width and height parameters are added to minimize the impact of CLS.

Summary

There is no guarantee that the code can be used 100% directly, a better way is to learn the principle step by step and modify it with your own theme features.

Code Summary

I’ll try to comment the following code in detail, just remove the comments when you use it. If there are other questions, you can leave a message in the comments section.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63

{{/* Get the resource image of the current page matching */}}
{{ $image := .Page.Resources.GetMatch .Destination }}
{{/* Get logo images from resources, transparent background images work best */}}
{{/* The image needs to be placed in the assets/images directory in the root directory */}}
{{ $logo := (resources.Get "images/logo.png") }}
{{/* Check if the media type of the image is not gif or svg */}}
{{ if not (in (slice "gif" "svg") $image.MediaType.SubType) }}
    {{/* Preserve references to the original image */}}
    {{ $rawImage := $image }}
    {{/* Calculate 80% of the image height and make sure the minimum value is 250 */}}
    {{ $size := math.Round (mul $image.Height 0.80) }}
    {{ $size := cond (ge $size 250) $size 250.0 }}
    {{/* Adjust the transparency of the logo to 50% */}}
    {{ $logo := $logo.Filter (images.Opacity 0.5)  }}
    {{/* Resize the logo to the calculated $size in height, keeping the aspect ratio and rounding the corners with a radius of 30 */}}
    {{ $logo := $logo.Resize (printf "%.0fx png r30" $size) }}
    {{/* If the original image has EXIF information */}}
    {{ with $rawImage.Exif }}
        {{/* Resize and rotate the image according to its orientation */}}
        {{ if eq .Tags.Orientation 1 }}
            {{ $image = $image.Resize (print $image.Width "x" $image.Height " webp") }}
        {{ else if eq .Tags.Orientation 3 }}
            {{ $image = $image.Resize (print $image.Width "x" $image.Height " webp r180") }}
        {{ else if eq .Tags.Orientation 6 }}
            {{ $image = $image.Resize (print $image.Height "x" $image.Width " webp r270") }}
        {{ else if eq .Tags.Orientation 8 }}
            {{ $image = $image.Resize (print $image.Height "x" $image.Width " webp r90") }}
        {{ end }}
    {{ else }}
        {{/* If there is no EXIF information, then only resize, no rotation */}}
        {{ $image = $image.Resize (print $image.Width "x" $image.Height " webp") }}
    {{ end }}
    {{/* Add a logo watermark to the image, positioned in the center of the image */}}
    {{ $image = $image.Filter (images.Overlay $logo (div (sub $image.Width $logo.Width) 2) (div (sub $image.Height $logo.Height) 2)) }}
    {{/* Create a link to a data attribute that meets fancybox's requirements, showing the processed image */}}
    <a data-fancybox="gallery" data-src="{{ $image.RelPermalink }}" data-caption="{{ .Text }}">
        <picture>
            <source type="image/webp" srcset="{{ $image.RelPermalink }}">
            <img
                width="{{ $rawImage.Width }}px"
                height="{{ $rawImage.Height }}px"
                style="width: 100%; height: 100%; max-height: 36rem;"
                loading="lazy"
                src="{{ $rawImage.RelPermalink }}"
                alt="{{ .Text }}"
                {{ with .Title }} title="{{ . }}" {{ end }}
            />
        </picture>
    </a>
{{ else }}
    {{/* No webp compression for gif or svg images */}}
    <a data-fancybox="gallery" data-src="{{ $image.RelPermalink }}" data-caption="{{ .Text }}">
        <img
            width="{{ $image.Width }}px"
            height="{{ $image.Height }}px"
            style="width: 100%; height: 100%; max-height: 36rem;"
            loading="lazy"
            src="{{ $image.RelPermalink }}"
            alt="{{ .Text }}"
            {{ with .Title }} title="{{ . }}" {{ end }}
        />
    </a>
{{ end }}

Thanks

Hugo: Image render hooks

Hugo图片自动转webp的方法

Resource: Hugging Face Doc

FP16

If your model doesn’t work well with mixed precision, for example if it wasn’t pretrained in mixed precision, you may encounter overflow or underflow issues which can cause NaN loss. For these cases, you should use full fp32 precision by explicitly disabling the default fp16 mode.

ZERO

stage 1 Optimizer state

stage 2 Optimizer state + gradient

stage 3 optimizer state + gradient + model parameters (weights)

The inference stage zero-1 zero-2 does nothing, only zero-3 can be set (no optimizer is needed for the inference stage and no gradient is generated). Also if Transformers<4.28, you need to set synced_gpus=True when generating.

Using multiple GPUs with ZeRO-3 for generation requires synchronizing the GPUs by setting synced_gpus=True in the generate() method. Otherwise, if one GPU is finished generating before another one, the whole system hangs because the remaining GPUs haven’t received the weight shard from the GPU that finished first.

For Transformers>=4.28, if synced_gpus is automatically set to True if multiple GPUs are detected during generation.

If offload is configured, you need to choose an optimizer that is adapted to both CPU and GPU, and the commonly used Adam cannot be used.

DeepSpeed and Transformers optimizer and scheduler can be mixed and matched as long as you don’t enable offload_optimizer. When offload_optimizer is enabled, you could use a non-DeepSpeed optimizer (except for LAMB) as long as it has both a CPU and GPU implementation.

Multi-GPU Optimization

Resource: Hugging Face Doc

It’s mostly just a few latitudes of parallelism: Data + Pipeline + Tensor.

Confusion

FP16

Only speeds up the training process, not necessarily reduces the memory footprint (maybe even by a factor of 1.5).

While mixed precision training results in faster computations, it can also lead to more GPU memory being utilized, especially for small batch sizes. This is because the model is now present on the GPU in both 16-bit and 32-bit precision (1.5x the original model on the GPU). From: Hugging Face Doc

Note: when using mixed precision with a small model and a large batch size, there will be some memory savings but with a large model and a small batch size, the memory use will be larger.

Small models, large batch sizes

• MEMORY SAVINGS: For small models, the model parameters and weights take up relatively little memory. When training with large batch sizes, the memory occupied by activations and gradients becomes the main memory usage part. In this case, converting activations and gradients to FP16 can significantly reduce this portion of memory usage, as FP16 occupies half the memory of FP32. Even when considering the retention of FP32-accurate copies of the model weights, the overall memory usage is still reduced because the memory savings of the activations and gradients outweigh the memory overhead of the additional copies of the weights.
• EFFICIENCY AND SAVINGS: Large batch sizes mean that more data is processed per iteration, which increases the amount of memory required to compute activations and gradients. Since this part now uses FP16, the overall memory requirement is reduced compared to using only FP32.

Large model, small batch size

• Increased memory usage: for large models, the model parameters and weights themselves take up a lot of memory. In mixed-precision training, even if the activations and gradients are stored as FP16, a copy of the FP32-precision weights still needs to be retained to ensure the accuracy of the updates. This means that the additional memory overhead (due to the FP32 copy of the weights) is a relatively large part of the total memory footprint relative to the size of the model itself. When the batch size is small, the activations and gradients take up relatively little memory, so the memory saving effect from FP16 is not enough to offset the additional memory usage added by keeping FP32 weight copies.
• Impact of weighted copies: in this case, the additional memory used by FP32 weighted copies becomes a significant factor because the model itself is large. Even if FP16 is used for activation and gradient, the total memory usage may still increase due to FP32 weighted copies.

Gradient accumulation

Effective batch = gradient accumulation * actual batch.

1

training_args = TrainingArguments(per_device_train_batch_size=1, gradient_accumulation_steps=4, **default_args)

In the above example, your effective batch size becomes 4.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

for epoch in range(2): # Assume 2 epochs of training
    for i, (inputs, labels) in enumerate(dataloader):: inputs = torch.tensor(inputs, dtype=torch.float32): # Suppose train 2 epochs.
        inputs = torch.tensor(inputs, dtype=torch.float32)
        labels = torch.tensor(labels, dtype=torch.long)
        # Forward propagation
        outputs = model(inputs)
        loss = criterion(outputs, labels) / accumulation_steps # Note that this is divided by the number of accumulated steps
        # Backward propagation
        loss.backward()
        # Update model parameters every accumulation_steps
        if (i + 1) % accumulation_steps == 0 or (i + 1) == len(dataloader).
            optimizer.step() # update parameters
            optimizer.zero_grad() # clear the gradient

BP is required each time, but the optimizer updates the parameters only when the cumulative number of gradient steps is reached.

optimizer.step() vs optimizer.zero_grad()

Normal training flow: zero_grad then step Avoid the gradient computed by previous batch to affect the current optimizer state.

Gradient accumulation process: step then zero_grad As the name suggests, we want to let the gradient accumulation take effect and then update the parameters at the end.

Prioritize the per_device_train_batch_size setting that guarantees the GPU limit is reached before considering accumulation.

While it is advised to max out GPU usage as much as possible, a high number of gradient accumulation steps can result in a more pronounced training slowdown. Consider the following example. Let’s say, the per_device_train_batch_size=4 without gradient accumulation hits the GPU’s limit. If you would like to train with batches of size 64, do not set the per_device_train_batch_size to 1 and gradient_accumulation_steps to 64. Instead, keep per_device_train_batch_size=4 and set gradient_accumulation_steps=16. This results in the same effective batch size while making better use of the available GPU resources.

It is not known whether gradient accumulation causes excess memory usage. Or rather I don’t think it does, I think the gradient just changes from one number to another number resulting from multiple accumulations, the data format doesn’t change and the number of bytes taken up doesn’t change. In practice it seems to have an effect, but I can’t find a reason to rationalize it.

Reference

Hugging Face Doc

CUDA driver

Check the version of CUDA already installed on the server:

1

nvcc -V

Showing version 11.4 here, but it doesn’t matter. Next to check the driver version, this relates to the latest CUDA version that we can install in the conda virtual environment.

1
2
3
4

nvidia-smi

### Output
### Driver Version: 535.129.03   CUDA Version: 12.2

Tips:

The CUDA Version shown here is 12.2, which is actually the latest CUDA version that the CUDA driver installed on the server supports to install, which means that the highest CUDA version we can install next is 12.2, and the CUDA version that has already been installed on this machine is 11.4.

Because we are a NON-ROOT user, we can’t change the installed driver version, but we can install the new CUDA version through the Conda virtual environment.

Conda

When configuring python-related environments for non-root users, I extremely recommend using the conda package manager for dependency management (in fact, it’s also recommended for those using R).

I’m installing the miniconda version here (the minimized version), but you can also install anaconda (the full version).

The blog has been written about related content before.

1
2
3

# I installed the environment with python 3.10
conda create -n torch python=3.10
conda activate torch

Pytorch

You can check the Pytorch history release at Pytorch History Release and combine it with your own needs (e.g., [Colossal-AI](https:// colossalai.org/zh-Hans/docs/get_started/installation/) requires PyTorch >= 1.11 and PyTorch <= 2.1). I ended up with pytorch==2.1.0 and CUDA==11.8.

1
2
3
4
5

# CUDA 11.8
# Less than the maximum supported driver version 12.2 but more than the installed CUDA version 11.4
# Try not to jump to a bigger version
# For example, if I have 11.4 installed on my machine, then I would install it with cuda=11.8. Does it seem to be a problem to install 12.1?
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=11.8 -c pytorch -c nvidia

Thinking:

As for why pytorch doesn’t redistribute for some cuda versions, in this issue one of pytorch’s developers explains that pytorch-cuda113 can be used on cuda114 This simply means that there is no need to re-release it, and that the 113 version can be used for the 114 version.

Check availability

1
2
3
4
5
6
7

(torch) [username/pwd]$ python
Python 3.10.13 (main, Sep 11 2023, 13:44:35) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.is_available()
True
>>> exit()

CUDA

At Nvidia-Cuda look for the CUDA version corresponding to the installed Pytorch, in my case 11.8.

1
2
3
4
5

# In nvidia, this channel doesn't even show a progress bar when downloading packages.
# It will only show 0% or 100%, but the download speed is fine, it's not walled off, just wait patiently.
# Officially: conda install nvidia/label/cuda-11.8.0::cuda

conda install cuda -c nvidia/label/cuda-11.8.0

At this point, enter the nvcc -V command to find that the installed version is now 11.8.

Cuda installed using conda

1
2

# This prevents programs that cannot find the CUDA (11.8) installed by conda from using the CUDA (11.4) that comes with the system.
export CUDA_HOME=$CONDA_PREFIX

Extra

gcc clang g++ clang++

1
2

# The server comes with a version of gcc that is too low and doesn't even support g++. Use conda to install it and it works straight away.
conda install -c conda-forge cxx-compiler

xformers

1
2
3
4
5
6
7

# Official repository suggested commands
conda install xformers -c xformers

# Be sure to double-check the package change log before confirming the installation.
# There is a chance that it could replace an installed version of pytorch, so don't install it.
# Try adding -c nvidia -c pytorch and you should be fine.
conda install xformers -c xformers -c nvidia -c pytorch

transformers & datasets & accelerate

1
2

# Installation with conda may not be the latest version
pip install transformers datasets accelerate -U

Reference

Nvidia-Cuda

Pytorch History Release

Pytorch Github Issue

I’ve seen several guides on sniffing the Star Rail gacha link on Android (ADB or desktop assisted). Thanks to the WeChat public account 星穹铁道工坊 for sharing.

For Star Rail sniffing:

The simplest approach is on a PC or iOS. If you only have Android, the ADB method above is simpler for non-tinkerers.

The method here isn’t simple—it’s fairly involved. But hey, tinkering is about persistence.

This guide targets rooted users who want to operate entirely on-device. The same approach works for other apps besides Star Rail.

Prep

Prerequisites:

• Magisk installed (KernelSU unknown because we need Magisk modules)
• LSPosed installed (Xposed compatibility unknown)

Install these apps/modules:

• Http Canary (v3.3.6): packet capture (aka “Little Yellow Bird”)
• AiWanJi Toolbox: fixes the Http Canary cert
• TrustMeAlready & JustTrustMe: partially handle apps that pin certificates
• “CA certificate completion” Magisk module

I’ve bundled them on Lanzou: download (password: ft5q). Steps below explain the setup.

Http Canary

This section shows how to install Http Canary’s certificate into the system store on newer Android versions. You can also read the original (in Chinese) here by user 密西根碎冰蓝. My short version:

• Tip: in Shamiko choose Blacklist mode (default). Don’t hide root from Http Canary even if it already has root.

• Flash the 抓包CA证书补齐_18002.zip module via Magisk, reboot to apply.

• Install and open Http Canary. If prompted to install a cert, tap Skip.

  We rely on the Magisk module to install the cert, not the in-app flow.

• Reopen Http Canary; it will still say the cert is missing. Force stop the app, install AiWanJi Toolbox (grant root via Magisk module mode), and run the fix:

  Path: App-related -> Third-party app assist -> Http Canary cert fix.

  

• In Http Canary open the side menu -> Settings -> HttpCanary root certificate. The subtitle under Install HttpCanary root certificate should say Installed. If not, reinstall and repeat the step above.

  

• Tap Add root certificate to system (Root), then tap Move and grant root. Even if the move fails, the setup still worked for me.

  

• Flash the LSPosed modules JustTrustMe and TrustMeAlready, then tick the target apps (both modules need the app checked).

  

  

That’s all the environment setup.

Star Rail

This part is Star Rail-specific, but the same logic applies to other apps.

Note:

Even with this setup, some apps can’t be sniffed due to enhanced network security (e.g., banking apps). For those you’ll need other techniques (reverse-engineering, etc.). See this thread.

Star Rail has similar protections—if you start capturing before entering the game you’ll see network errors. Instead, log in normally, enter the Warp page, then start Http Canary. Return to the Warp screen, view Details, then History to grab the data.

Back in Http Canary, find the URL starting with api-takumi.mihoyo.com. Open it and long-press the highlighted area to copy.

Wrap-up

Most of this post is about getting Http Canary running. For apps without heavy network hardening, that’s enough. For stricter apps, tailor the approach accordingly.

Recently, I need to recompile an iOS jailbreak plugin named PassBy, and found that it reported errors frequently on my new device, so I recorded the exploration and solution process below.

Since I’m not a professional, and I don’t even know much about iOS development, this is a record of my exploration with the attitude of learning. If there are any mistakes, welcome to criticize and correct in the comment section.

By the way, I would like to advertise my other related article haha:

iOS Jailbreak Plugin PassBy Customization, documenting my attempts to solve PassBy’s non-recognition of Bluetooth wearable devices other than the Apple Watch.

Trying to compile

Firstly

System: macOS Monterey 12.6.5

Chip type: Intel chip, not M series

Xcode version: 14.2

I first tried to compile directly by typing make package, which prompted openURL has been deprecated.

At this point I guessed that the Xcode version was too high to cause the mismatch, and I was going to go to the source code on GitHub first to see if the author had made any updates. Unfortunately, there was nothing I could do but try to modify the code.

According to the error message I first found the relevant c file, searched for the openURL keyword and commented out the relevant code.

Continue to look for, you can find that this is not affect the main function of the code, should be to perform a donation and source code to open the URL of the operation, because I am not familiar with Objective-C language, the simplest way is to directly comment out, but fortunately does not affect the main function of the plug-in. The two URLs I pasted below:

• Donate (Donate): https://paypal.me/giorgioiavicoli
• Source Code: https://github.com/giorgioiavicoli/PassBy

Secondly

Trying to continue compiling, unsurprisingly, a new problem arises.

This time, the problem is even more difficult, as theos is directly prompted by the Xcode compiler that it cannot find the Message framework.

This problem would not be something I could fix with a simple modification. My guess is that the newer versions of Xcode have discarded the framework, which leaves me with two options: either refactor the code (which would be too difficult for me to do), or switch to an older version of Xcode (which is a bit of an unknown quantity until I get my hands on it, and I can’t guarantee that I’ll be able to fix it).

I can only try to change the Xcode version.

Change the Xcode version

First check the current version, you can see that it is version 14.2.

On the Xcode support page you can clearly see the minimum system requirements for each Xcode version as well as the included SDKs and other information.

Next go to Apple’s official download page about all versions of Xcode and just download an older version. I downloaded version 12.5.1.

The next step is to extract the .xip file, change the name of the extracted Xcode.app to a name like Xcode_12.5.app, and finally move it to the /Application directory.

As you can see in the image above, my old version of Xcode doesn’t open directly, but it doesn’t matter for the purposes of this chapter, we don’t need to open the old version of Xcode directly to operate it. Please read on to find out why.

It takes a while to unpack, so be patient.

Alternatively, you can download and update Xcode in the usual way. With a multi-threaded download tool, you can basically run at full broadband speed and avoid App Store download failures. The disadvantage is that decompression is more time-consuming.

Switch Development

Note that the subtitle of this section is not “Switching Xcode Versions”, because for us, all we need is the scripting tools that come with older versions of Xcode in the development directory, such as gcc and so on.

Next just use this command to switch the xcode-select directory:

1

sudo xcode-select -s /Applications/{modify here to a specific filename}/Contents/Developer/

Just change it to your old version of the file name, and enter the password without showing it.

At this point, enter the command again to verify:

1
2
3
4

xcodebuild -version
xcode-select -p
gcc -v
g++ -v

The output of all the above commands can be seen in the words of the name of the old version of Xcode file you chose, something like Xcode_12.5.

Thirdly

Trying to compile again, the problem is finally solved.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

yuesir at iMac in /Volumes/Data/Xcode/PassBy
$ make package
> Making all for tweak PassBy…
make[2]: Nothing to be done for 'internal-library-compile'.
> Making all in passbyprefs…
> Making all for bundle PassByPrefs…
==> Copying resource directories into the bundle wrapper…
==> Compiling PassByRootListController.m (arm64)…
==> Compiling PassByRootListController.m (arm64e)…
==> Linking bundle PassByPrefs (arm64)…
==> Generating debug symbols for PassByPrefs…
==> Stripping PassByPrefs (arm64)…
==> Linking bundle PassByPrefs (arm64e)…
==> Generating debug symbols for PassByPrefs…
==> Stripping PassByPrefs (arm64e)…
==> Merging bundle PassByPrefs…
==> Signing PassByPrefs…
> Making all in passbyflipswitch…
> Making all for bundle PassByFlipswitch…
==> Copying resource directories into the bundle wrapper…
make[3]: Nothing to be done for 'internal-bundle-compile'.
> Making stage for tweak PassBy…
> Making stage in passbyprefs…
> Making stage for bundle PassByPrefs…
> Making stage in passbyflipswitch…
> Making stage for bundle PassByFlipswitch…
dm.pl: building package `com.giorgioiavicoli.passby:iphoneos-arm' in `./packages/com.giorgioiavicoli.passby_1.9.5_iphoneos-arm.deb'

In fact, I tried to restore the above commented out code to normal, but I didn’t realize that it still reported the same error!

I suspected that it might be a problem with the version of Command Line Tools installed. I didn’t try it because I didn’t know what would happen if I installed a lower version of Tools.

Switch Back

1

sudo xcode-select -s /Applications/Xcode.app/Contents/Developer/

Simply verify it:

Script

Here’s another script shared by others, but I didn’t try it myself, so I’m posting the link here as a thank you.

Link to the article: https://www.cnblogs.com/zndxall/p/12463744.html

Thanks

Apple Xcode

Xcode installs multiple versions and switches versions automatically

My other post describes using Conda installing R kernel Notebook on a Linux remote server, but I’ve found that conda doesn’t handle various dependency conflicts very well, often resulting in installations that don’t I found that conda doesn’t handle dependency conflicts well, which often leads to unsuccessful installations.

Later, I also used brew to install it on my macOS system, and found that it was easier and faster, so I summarized the steps.

System: macOS Venture 13.1 (22C65)

Product: Intel Chip MacBook

Requirement

Install Homebrew.

Homebrew is the most used package manager on macOS, and this is its official installation tutorial. But usually, we just need to enter the following command while making sure the network is accessible:

1

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

R

It is so easy to install r package with brew:

1

brew install r

Jupyter

Almost the same as before:

1

brew install jupyterlab

R-Kernel

Use r to install IRkernel:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

# Go to the r console first
R
# Install devtools
install.packages('IRkernel')
# At this point you might be asked to choose a mirror, just choose the nearest one
# Register the kernel with R
IRkernel::installspec()
# Install the RStudio shortcut.
# If RStudio is installed, you can choose to install it, but not if you're using a notebook.
jupyter labextension install @techrah/text-shortcuts

Launch Notebook

1

jupyter notebook

Click Create or New button and find that you can create r-kernel notebooks now.

Landscaping

This part is not necessary and is for record only.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17

# Configure themes
pip install jupyterthemes
# A set of theme settings that I personally like and have gotten used to
jt -t monokai -f fira -fs 13 -cellw 90% -ofs 11 -dfs 11 -T -N
# Restore the default theme
jt -r
# Configure plugins
pip install jupyter_contrib_nbextensions
jupyter contrib nbextension install --user
# Install the plugin manager This step may fail See update 1 later for details
pip install jupyter_nbextensions_configurator
# Suggested configuration items
# Hinterland code completion
# Collapsible headings
# Notify Notify mechanism to run time-consuming tasks and notify when they're done.
# zenmode Hide active status bar, easy to notice the code.
# ExecuteTime Show execution time

Postscript

I didn’t run into any snags with this installation, so if you run into any other issues, feel free to leave them in the comments section.

Update 1

When I repeated this tutorial on 2023.03.20, I found that the Nbextensions tag would not appear. Upon checking, this was caused by the plugin not being updated to adapt a new version of the dependency package nbclassic.

See the GitHub issue for details.

Solution: Downgrade the dependency package. We may update it later.

1

pip install "nbclassic<0.5"

Update 2

2024.03.15 Update: The notebook was officially upgraded to version 7.x a couple months back. I haven’t had to do any 3rd party manipulation with the new version. It can be changed directly inside the settings with a high degree of customization.

What is Oh My Zsh

• Oh My Zsh is a community-driven command line tool, and as its homepage says, Oh My Zsh is a way of life. It is based on the zsh command line and offers theme configuration, plugin mechanism, and already built-in convenience operations. It gives us a whole new way to use the command line.
• Oh My Zsh is an extended toolset based on the zsh command line, providing a rich set of extended features.

Official Website

Oh My Zsh official website

If you are interested in more, you can go to the official website for details.

Preparation

• Install zsh to replace bash, etc. The default shell for macOS systems is currently zsh, other Linux systems may need to install zsh first.

  Ubuntu：

  1 2 3 4

  # Install zsh apt install zsh # set zsh as the default shell (if you don't set it, start zsh with the direct zsh command) chsh -s /bin/zsh

  Centos：

  1 2	yum install zsh chsh -s /bin/zsh

  Same for the others.

• Install git, same as above using apt or yum package manager.

• Finally you need curl or wget, install the same as above.

Install

curl installation

1

sh -c "$(curl -fsSL https://raw.github.com/ohmyzsh/ohmyzsh/master/tools/install.sh)"

wget installation

1

sh -c "$(wget https://raw.github.com/ohmyzsh/ohmyzsh/master/tools/install.sh -O -)"

There is no difference between the two installs.

Themes

I usually use the steeef theme that comes with it, which is pretty clean and easy to use.

If you want to change the theme, just edit the ~/.zshrc file and replace ZSH_THEME with the theme you like after it.

The .zshrc file.

The .zshrc file does the same thing as the .bashrc file. The default shell for most unix systems used to be bash, and the .bashrc file is the configuration file for bash.

The relevant explanation in the online help file, as viewed with the man bash command, is as follows. .bashrc - The individual per-interactive-shell startup file.

This file mainly stores personalized settings, such as command aliases, paths, etc.

Each time you modify .bashrc, use source ~/.bashrc (or . ~/.bashrc) to immediately load the modified settings and make them take effect. Typically .bashrc is called explicitly in the .bash_profile file.

When you log in to linux and start bash, it will first read the ~/.bash_profile file, so ~/.bashrc will be executed and your personalized settings will take effect.

Plugins

The management of omz plugins is very simple, there are two plugins directories, where user is your username:

• /Users/user/.oh-my-zsh/plugins

  The official plugins directory. This directory is already preloaded with many useful plugins, which only need to be activated manually.

• /Users/user/.oh-my-zsh/custom/plugins

  Third-party plugins directory, shortcut command: $ZSH_CUSTOM/plugins.

To install the plugins, just download the plugins to the third-party plugins directory, and then add the name of the corresponding plugins to the plugins variable in the ~/.zshrc configuration file.

Download Plugin

zsh-syntax-highlighting

1
2

# Command syntax highlighting
git clone https://github.com/zsh-users/zsh-syntax-highlighting.git ${ZSH_CUSTOM:-~/.oh-my-zsh/custom}/plugins/zsh-syntax-highlighting

zsh-autosuggestions

1
2

# Command auto-completion
git clone https://github.com/zsh-users/zsh-autosuggestions ${ZSH_CUSTOM:-~/.oh-my-zsh/custom}/plugins/zsh-autosuggestions

If you feel that Tab key is not convenient for completing, you can also customize the shortcut keys for completing. For example, to set comma completion, just add the following line to ~/.zshrc file.

1	bindkey ',' autosuggest-accept

Enable Plugin

In the .zshrc file set.

1

plugins=(git zsh-syntax-highlighting zsh-autosuggestions)

Enable other plugins in the same way.

Configuring to take effect

One way to do this is to exit the terminal and reopen it, another, more common way is to use the source command I described earlier.

1

source ~/.zshrc

Thanks

Oh My Zsh

安装oh my zsh插件

If you don’t like the “bloat” of the official docker client, like me, and just want a separate and pure docker service without a GUI, or just want to try new things and tinker with new methods, then you are welcome to read this article.

Since I am using macOS, there is only one exploration record under Mac. If there are any omissions and errors, please feel free to criticize and point them out in the comment section.

Computer environment for this article.

System：macOS Ventura 13.0.1 (22A400)

CPU：Intel Core i5（not sure if Apple M series processors are common）

Preparation

• Install Homebrew.

  Homebrew is the most used package manager under macOS platform, and this is its official installation tutorial. But usually, we just need to enter the following command while making sure the network is open.

  1	/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Install packages

• Use Homebrew to install docker.

  Installing the most basic docker service is also very simple and can be done with one command.

  Remember not to use the brew install docker --cask command, this will install docker graphically on macOS!!!

  1	brew install docker

• Installing Colima with Homebrew.

  Colima takes its name from Container on Lima. Lima is a virtual machine tool that enables automatic file sharing, port forwarding, and containerd.

  Colima actually starts a virtual machine named colima through Lima. It helps us to simplify the usage process, but of course it is less customizable. If you are interested, you can of course use Lima to do the process manually yourself, which is more customizable, but more complex.

  The installation is also very simple and is done with one command.

  1	brew install colima

  This should take much longer to download, which is normal. As you can see below, Colima has a lot of dependencies.

  1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38

  $ brew deps colima bdw-gc ca-certificates capstone gettext glib gmp gnutls guile jpeg-turbo libevent libidn2 libnghttp2 libpng libslirp libssh libtasn1 libtool libunistring libusb lima lz4 lzo m4 ncurses nettle openssl@1.1 p11-kit pcre2 pixman pkg-config qemu readline snappy unbound vde xz zstd

Start Colima

The official installation documentation describes that creating a docker runtime environment requires only the simplest command.

1

colima start

Perhaps this is also the reason why the author has optimized the process specifically for docker, which is why I recommend using colima, it’s simple enough.

Boot-up

2023.12.20 Update:

Based on feedback from comments, colima already supports boot-up.

The method is also simple, since we installed colima using brew, we can just take over directly using brew service.

1	brew services start colima

If you are using other ways to install, make good use of search engines. Or follow the tip from comments by adding the -foreground parameter to set it up (untested).

1	colima start –foreground

Although Colima still doesn’t provide such a feature so far, it is in development.

Alternatively, for simple single-container docker use, we can create our own shell scripts and set the startup entries to achieve the same effect.

• To create the shell script.

  1 2	touch colima.sh chmod +x colima.sh

• Modify the contents of the script.

  1	vim colima.sh

  1 2	# Just write the start command to colima.sh colima start

• Add to the login entry.

  

• Set the default open method to your terminal program.

  

Enjoy

It’s that simple, and the docker service is up and running. Type docker ps -a and simply test to see if it runs successfully. If you still have errors, keep reading on.

Troubleshooting

The diagram indicates that the docker service is not running, which is a problem with the colima configuration. You can try to reinstall Colima or re-run colima start to check if it is resolved.

Additional

• You can further install docker-compose to use it.

  1	brew install docker-compose

• If you want to run multiple containers, you can make configuration changes to the containers created by Colima.

  Colima’s default resource configuration is 2 CPU / 2 G memory, which may not be enough in some cases. For example, if you start three mysql instances with docker-compose, it will happen that at least one of them will not start properly.

  In this case, you can change the default configuration with the -c and -m start parameters, or with colima start --edit.

Thanks

Colima

Lima

Docker

Due to the author’s frequent use of conda to configure the R kernel jupyter notebook on Linux remote servers, some steps are forgotten over time (the main problem is that the same pit will be stepped on differently every time).

After summarizing the relevant tutorials and personal experience on the internet, a guide is provided for everyone and future me to refer to.

Guide

Install Conda

To start with an introduction to Conda, Anaconda and some of the connections between them.

Conda is a package manager that supports binary packages, so you don’t need to compile from source when installing, and you can also install native libraries. However, some of the Python libraries in pip are not found in conda.

• Anaconda is a Python scientific computing distribution managed with Conda.
• Conda Forge is another Python distribution that is more package rich than Anaconda. But one pitfall here is that Conda Forge and Anaconda are not fully compatible, so if you have a project that uses packages from both Anaconda and Conda Forge, there is a chance that it will hang.

Usually the servers I use have Anaconda installed beforehand,

Here is the official Anaconda Installing Anaconda on Linux document with detailed step-by-step instructions and command explanations.

2022.10.17 Update Linux - Anaconda installation process under Centos7.

Due to the special domestic environment, it is still recommended to download Anaconda installation scripts directly from domestic sources such as Tsinghua University Mirror Source, usually choosing the x86_64 version of.

Chrome can be downloaded by right clicking and copying the file link address and then using wget in centos:

1
2
3
4
5
6

# Download script
wget https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive/Anaconda3-2022.05-Linux-x86_64.sh
# If prompted without wget use yum install
yum install wget
# Execute the installation script Change the script name yourself
bash Anaconda3-2022.05-Linux-x86_64.sh

Here are some options I encountered myself, for the record. Please follow the script prompts as you see fit and enter.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

# Follow the prompt and type ENTER (enter)
Welcome to Anaconda3 2022.05

In order to continue the installation process, please review the license
agreement.
Please, press ENTER to continue
>> ENTER

# Follow the prompts and enter yes
Do you accept the license terms?
[no] >>>
Please answer 'yes' or 'no':'
>>> yes

# Follow the prompts and type ENTER (enter)
# Anaconda3 will now be installed into this location:
/root/anaconda3

  - Press ENTER to confirm the location
  - Press CTRL-C to abort the installation
  - Or specify a different location below

[/root/anaconda3] >>> ENTER

# conda initialization default no recommend entering yes
# If you choose no, you need to configure the environment variables yourself (Baidu keyword anaconda environment variables)
installation finished.
Do you wish the installer to initialize Anaconda3 by running conda init?
Do you want the installer to initialize Anaconda3 by running conda init?
[no] >>> yes

# Here are some tips, briefly translated
# For conda to take effect you need to close and reopen the current shell (reconnect to ssh)
# If you don't need to activate the conda base environment at startup enter conda config --set auto_activate_base false
==> For changes to take effect, close and re-open your current shell.

If you'd prefer that conda's base environment not be activated on startup,
   set the auto_activate_base parameter to false:

conda config --set auto_activate_base false

Anaconda is now installed, and by the way it is recommended to add domestic sources:

1
2
3
4
5
6
7
8

# Only for Chinese
conda config --add channels https://mirrors.ustc.edu.cn/anaconda/pkgs/main/
conda config --add channels https://mirrors.ustc.edu.cn/anaconda/pkgs/free/
conda config --add channels https://mirrors.ustc.edu.cn/anaconda/cloud/conda-forge/
conda config --add channels https://mirrors.ustc.edu.cn/anaconda/cloud/msys2/
conda config --add channels https://mirrors.ustc.edu.cn/anaconda/cloud/bioconda/
conda config --add channels https://mirrors.ustc.edu.cn/anaconda/cloud/menpo/
conda config --set show_channel_urls yes

Creating Environment

I am creating a new conda environment directly to avoid dependency conflicts with the least effort.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

# View existing conda environments
conda env list
# Create a new environment by name
# your_new_name_to_conda Replace with the name of the environment you want
conda create -n your_new_name_to_conda
# Activate the new environment
conda activate your_new_name_to_conda
# Check the installed dependencies
# The new environment is, of course, free of anything
conda list
# Quit the new environment
conda deactivate

Install R

Note here that you cannot use conda install r directly to install R.

If you enter the above command directly, you will be installing version 3.6 (I think) of R. This version is too old and many R packages are not available. As of 2022.10.10, the latest version of R’s conda is 4.2.1.

The following will explain how to install the latest version of R (4.2.1).

1
2

# conda install r-base via conda-forge
conda install -c conda-forge r-base

conda-forge is also described at the beginning of the article, I feel it can be understood as a distribution inside Linux, maybe?

Before confirming the installation (i.e. before entering y), you can make sure beforehand if the version of r-base is version 4 or higher (currently 4.2.1).

Install R kernel and Jupyter

1
2
3
4

### Specify to install using the default source
# r-irkernel is the r-kernel for jupyter notebook
# jupyter includes jupyter notebook
conda install -c defaults r-irkernel jupyter

Note:

Do not use conda-forge to install r-irkernel and jupyter here.

As mentioned before, Conda Forge and Anaconda are not fully compatible. I personally tried to install it using the conda install -c conda-forge r-irkernel jupyter command and jupyter notebook did not start properly. In this case just uninstall and reinstall.

1 2 3 4

# Uninstall conda uninstall r-irkernel jupyter # Install conda install -c defaults r-irkernel jupyter

Start Notebook

If everything is fine in the previous section, you will be able to start Jupyter Notebook properly here and create R’s notebook.

If it doesn’t start properly, please continue to read [Exception Handling] below.

1
2

# Start Jupyter Notebook
jupyter notebook

Exception Handling

Module Missing

If you are prompted with ModuleNotFoundError: No module named ‘zmq.backend.cffi._cffi’ when starting a notebook, the following is a similar situation.

The solution can be found in an issue under the official repository by uninstalling the module in question using pip and reinstalling it:

1
2
3

# Use pip instead of conda to manage pyzmq packages
pip uninstall pyzmq
pip install pyzmq

Pro-tested to work.

If you have similar ModuleNotFoundError problems, you can try the above method and use pip instead of conda to manage those python packages, uninstall and reinstall them first.

The main thing is to find that exact package name, you can try to search keywords in pypi.

Port Possession

Since it is a server shared by multiple people in the lab, the default port 8888 of the notebook is also occupied. Here is the solution.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

ipython
# Access the python console
Python 3.7.13 (default, Mar 29 2022, 02:18:16) 
Type 'copyright', 'credits' or 'license' for more information
IPython 7.31.1 -- An enhanced Interactive Python. type '? for help.

In [1]: from notebook.auth import passwd

In [2]: passwd()
Enter password: 
Verify password: 
Out [2]: 'argon2:$argon2id$v=19$m=10240,t=10,p=xxxxxxxxx' # Copy the content between '' needed later

In [2]: exit

jupyter notebook --generate-config # This will prompt you where the generated file is
vim ~/.jupyter/jupyter_notebook_config.py # Edit the file prompted in the previous step here

# Add or modify the following
c.NotebookApp.allow_remote_access = True
c.NotebookApp.ip = '*'
c.NotebookApp.password = u'argon2:$argon2id$v=19$m=10240,t=10,p=8$Vc5lLVsXKZLGftp6xrHfEfOkQvIU5YiOhdw' # Paste the previously copied key
c.NotebookApp.open_browser = False
c.NotebookApp.port = 61111 # Set the port and be careful not to duplicate it with others
c.NotebookApp.notebook_dir = '/home/xxx/' # not necessary set jupyter default directory

# Once set, you should be able to access it normally
jupyter notebook

Beautification enhancements

This part is not essential and is for record only.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17

# Configure themes
pip install jupyterthemes
# A set of theme settings that I personally like and have gotten used to
jt -t monokai -f fira -fs 13 -cellw 90% -ofs 11 -dfs 11 -T -N
# Restore the default theme
jt -r
# Configure plugins
pip install jupyter_contrib_nbextensions
jupyter contrib nbextension install --user
pip install jupyter_nbextensions_configurator
# Suggested configuration items
# Hinterland code completion
# Collapsible headings Collapse headings
# Notify notification mechanism to run time-consuming tasks and notify when they are completed
# Codefolding collapsing code
# zenmode Hide the active status bar for easy attention to the code
# ExecuteTime shows the running time

Thanks

All the users who enthusiastically shared their tutorial experience.

This content may be outdated—please cross-check with the official repo.

Heads-up:

As of 2023.10.1, new users can pull the Debian image docker pull whyour/qinglong:debian.

It’s more compatible with some Python deps: e.g., ddddocr installs on the Debian tag but not on the default image.

Preface

After sharing the Qinglong setup and JD beans task, I found another fun project: BiliTools.

Based on the original docs, here’s my configuration tutorial plus personal notes. Take what you need.

Note:

vercel.app was recently blocked in China; you’ll need a proxy to read the original docs.

Prerequisites

A running Qinglong server (local or cloud).

Setup guide: click here.

The project supports many environments (local, Qinglong, Docker, serverless, etc.). This post only covers Qinglong.

Tutorial

Install dependency

In Qinglong Panel -> Dependency Management -> NodeJs -> New dependency, enter @catlair/bilitools and install.

Pull the script

In Qinglong Panel -> Cron Jobs -> New job, set Name (anything), Command (ql raw https://gitee.com/Kudouran/BiliTools/raw/main/tools/bilitools_npm.js), Schedule (0 0 0 * * *), then create and run once manually.

About the cron rule:

This uses a Java-style cron expression. 0 0 0 * * * means once daily at midnight. Adjust with an online tool if needed.

Update npm package

Auto-update looks like this:

1

pnpm remove -g @catlair/bilitools && pnpm add -g @catlair/bilitools

Without pnpm remove -g @catlair/bilitools, old versions linger on disk.

Get cookies

Desktop browser incognito (recommended)

Example with Firefox/Chrome/Edge. Final cookie resembles (line breaks added for readability):

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

_uuid=D2282D0F-257B-845A-BDF5-C770ED288F4001440infoc; buvid3=BF17608E-FB87-4F49-A922-56FD2E284D6F18534infoc;
fingerprint=5502cd4fe9637738de04bd9c3d1bdbc5;
buvid_fp=BF17608E-FB87-4F49-A922-56FD2E284D6F18534infoc;
SESSDATA=21607773%2C1631089673%2C71a42%2A31; bili_jct=dd92c55a6d67041ce2f3fb1650889ea8;
DedeUserID=521268093; DedeUserID__ckMd5=47d541f04b605da9;
sid=ivie73r8; fingerprint3=792b32adfecbe31a4aca53ab7be1ad76;
fingerprint_s=bb6736758e7344a295c2ed6070cc642e;
buvid_fp_plain=BF17608E-FB87-4F49-A922-56FD2E284D6F18534infoc;
CURRENT_FNVAL=80; blackside_state=1; rpdid=|(kmJYYJ)lkR0J'uYu)llkJYJ; _dfcaptcha=a46d7562a42065d43a88c053e283e876;
LIVE_BUVID=AUTO8016188357987702; bsource=search_baidu; PVID=2

To avoid cookie refresh/invalidations (e.g., if you frequently browse Bilibili, the session might rotate), don’t copy cookies from your everyday session. Instead, open a private/incognito window, log in, and capture a fresh cookie. Close the window directly—don’t log out or the cookie dies immediately.

Incognito shortcuts (examples):

Chrome

Open an incognito window, visit Bilibili, log in. Press F12 (or right-click -> Inspect), switch to Network, refresh, click a request (often the first nav request), then copy the cookie.

When copying in Chromium browsers, choose Copy (not “Copy value”).

Firefox

Open a private window, visit Bilibili, log in. Press F12 (or right-click -> Inspect), go to Network, refresh.

Mobile

Not recommended—too many pitfalls.

Config file (important)

Configuration is the core of the project. Only a correct config will run properly—read carefully.

Lazy single-user template (edit as needed)

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135

[
  // Lines starting with // are comments
  {
    // Browser User-Agent; grab from http://service.spiritsoft.cn/ua.html
    userAgent: '',
    // General API delay range (seconds)
    apiDelay: [2, 6],
    // Bilibili cookie (see above)
    cookie: "",
    // Features to run
    function: {
      silver2Coin: true,
      addCoins: true,
      liveSignTask: true,
      shareAndWatch: true,
      mangaTask: true,
      supGroupSign: false,
      useCouponBp: true,
      getVipPrivilege: true,
      giveGift: false,
      matchGame: false,
      batchUnfollow: true,
      liveLottery: false,
      liveRedPack: true,
      liveIntimacy: false,
      bigPoint: false,
      judgement: false,
      activityLottery: false,
    },
    // Live red packet
    redPack: {
      // 1: activity link (may be outdated); 2: scan; other = try all
      source: 0,
      // Activity link
      uri: 'https://api.live.bilibili.com/xlive/fuxi-interface/AugRedPacket2022Controller/redPocketPlaying',
      // Interval between rounds (seconds)
      intervalActive: 60,
      // Mid-break: [count, rest minutes (<1 ends)]
      restTime: [-1, -1],
      // Suspected risk cooldown: [count, rest minutes (<1 ends)]
      riskTime: [-1, -1],
      // Concurrent rooms
      linkRoomNum: 1,
      // Bullet comments sent while waiting [fixed] or [min,max]
      dmNum: [10],
      // Handle followed users while waiting (read/move)
      moveUpInWait: true,
      /** Move followed UPs into group */
      moveTag: 'rp关注',
      /** Follow reply handling */
      actFollowMsg: 'read',
      // Total participation limit (-1 = none)
      totalNum: -1,
      // Stop after N consecutive losses (<1 = no limit)
      noWinNum: 10,
      // Stop after N suspected risk events (<1 = no limit)
      riskNum: 5,
    },
    // Auto unfollow
    unFollow: {
      // Interval per unfollow (seconds)
      delay: 3,
      // Mid-break [count, rest minutes] (positive only)
      restTime: [20, -1],
      // Total limit (-1 = none)
      totalNum: -1,
      // Tags to unfollow
      tags: ['天选时刻', 'rp关注'],
    },
    // Notifications
    message: {
      br: '\n',
      // Qinglong-friendly example: WeCom webhook key
      'QYWX_KEY': "Fill your key here",
    },
    // Coin settings
    coin: {
      targetLevel: 6,
      stayCoins: 0,
      targetCoins: 5,
      customizeUp: [],
      todayCoins: 0,
      upperAccMatch: false,
    },
    // Use B-coins
    couponBalance: {
      /** Default charge target (self by default) */
      mid: 0,
      /** Preset days; empty = daily */
      presetTime: [10, 20],
      /** Mode: charge/battery */
      use: '充电',
    },
    // Live gifts
    gift: {
      // Custom gift targets (randomly chosen)
      mids: [],
      // Other fields omitted
    },
    // Intimacy
    intimacy: {
      liveSendMessage: true,
      liveLike: true,
      limitFeed: 1500,
      liveHeart: false,
      whiteList: [],
      blackList: [],
    },
    // Manga
    manga: {
      sign: true,
      buy: false,
      mc: [],
      name: [],
      love: true,
      buyInterval: 2,
      buyWeek: [],
    },
    // Jury
    jury: {
      mode: 1,
      vote: [0, 0, 1],
      once: true,
      opinionMin: 3,
      waitTime: 20,
    },
    // Big points
    bigPoint: {
      isRetry: true,
      isWatch: true,
      epids: [],
      watchDelay: 40,
    },
  },
]

If you use the template, edit these four places:

1. User-Agent required — copy the red text from this site.

2. Cookie required — see above.

3. Coin settings — adjust as needed (e.g., disable once you reach Lv6).

   1 2 3 4 5 6 7 8

   coin: { targetLevel: 6, stayCoins: 0, targetCoins: 5, customizeUp: [], todayCoins: 0, upperAccMatch: false, }

4. Notification — recommended to fill.

   1 2 3 4 5

   message: { br: '\n', 'QYWX_KEY': "Fill your key here", // Other providers are also supported (GOBOT, SCKEY, QQ, etc.) }

Add environment variable

• Config uses json5 (supports comments). Validate with an online json5 validator before use.

• Paste the validated config into the Gzip tool, choose Gzip compress, then copy the output.

  

• In Qinglong Panel -> Environment variables -> New variable, set Name BILITOOLS_CONFIG, Value = the Gzip output, then save.

  

Run manually

Back on the Qinglong tasks page you should see three BiliBili jobs:

Names can be anything—watch the commands:

• First: update dependency
• Third: pull repo
• Second: main program (visible after running the pull once)

When ready, run the main job manually.

Thanks

BiliTools

If iOS keeps killing your apps, try older versions. They often use fewer resources, feel smoother, burn less battery, and show fewer ads. Some apps were free before turning paid (e.g., Caiyun Weather Pro, BaiMiao). Even WiFi Master Key still works unfiltered—so grab it!

Tip: On desktop, use Ctrl+F (Command+F on macOS) to search quickly.

Changelog

2022/09/25

• JD 10.2.0 is the lowest version where sending images on iOS 16 works; 10.2.4 has no Double-11 icon
• Xiaohongshu 7.4 supports new emoji

2022/09/12

• Amap below 9.05.2 loses search

2022/09/03

• WeCom 4.0 supports live meetings
• QQ 8.5.0 supports group classroom
• IdleFish 7.4.10 adds elder mode, no launch ads, phone-login only
• Taobao 10.13.0: first dark-mode release after 10.12.20, no 618 icon; set region to Global to reduce ads
• Alipay 10.2.76 adds elder mode
• NetEase Cloud Music 4.1.2 supports new ncm format download/playback
• Tieba internal 1.0.8 delisted; still downloadable if purchased; no ads; no search
• BaiMiao 2.1.6 no ads

2022/08/31

• Douyin Lite 16.8.0 supports playback speed
• Douyin 16.5.0 supports playback speed

2022/08/28

• Douban 6.x.x removed; minimum usable 7.0.0
• Taobao versions before 10.12.20 all show update popups; 10.12.20 is the first dark-mode version

Perks

Halide Mark II

• RAW camera support
• Switched from paid download (¥30) to subscription (¥20/mo) on 2020-10-22. Download 1.17.3 (850784840) to use all features free.

FiLMiC Pro

• Powerful video camera
• Switched from paid (¥99) to subscription (¥35/week) on 2022-08-25. Download 6.19.7 (851609129) to use basic features free.
• Jailbroken users can unlock Log etc. via iOSGods iAP.

Pixelmator Photo

• Excellent photo editor
• Switched from paid (¥30) to subscription (¥33/mo) on 2022-08-18. Download 2.0.12 (850784840) to use all features free.

Version Recommendations

• iPhone X launched 2017-11-03; versions after that support full screen.
• iOS 13.0 Beta 1 (2019-06-04) introduced dark mode.
• iOS 14.5 Beta 1 (2021-02-02) guarantees support for iOS 14.5+.
• iOS 15.0 Beta 1 (2021-06-18) guarantees support for iOS 15.0+.
• ◆ marks my personal daily version (iPhone 11 Pro Max, iOS 14.2).
• Be cautious updating iOS: many legacy apps can’t be installed on 14.5+ (package incomplete errors).
• Check version histories and release dates on Qimai.
• If a version stops working, leave a comment.

Below are the version lists grouped by category. Version notes largely keep the original wording to avoid losing nuance.

Social

• QQ 6.2.2 (very fast launch, rarely killed; bubble display glitch) 6.5.9 6.6.5 (lowest version with working cloud sync) 6.7.1 (no side-eye emoji; requires overlay install) 7.1.0 (lowest direct-login version) 7.2.5 7.2.8 (full screen) 7.3.2 (Apple Watch) 7.3.5 7.5.5 7.7.8 8.0.6 8.1.0 (first simple mode; space videos won’t load) 8.1.5 (iOS 15.5) ◆8.1.8◆ (minimum for overlay-install dark mode) 8.2.6 (iOS 14.5) 8.2.8 (iOS 15.3) 8.4.1 (minimum for downloading night themes) 8.4.2 (iOS 15.3) 8.4.10 (bull-emoji pack) 8.5.0 (group classroom) 8.7.5 (Caiwang emoji, iOS 16) 8.8.38 (QQ Channels) 8.8.50
• TIM 1.1.5 ◆2.3.0◆ (structured messages, online docs, full screen) 2.5.2 2.5.6 3.0.0 3.1.0 (earlier versions require overlay install; iOS 16 cannot overlay)
• WeChat 6.5.19 6.6.1 6.7.0 (iOS 15.6 crashes) 6.7.4 (iOS 14.5; last with China region; see https://qianling.pw/wechat-blank) 7.0.5 (first floating window; auto-play Moments videos) 7.0.9 7.0.12 (first dark mode; no ads in Official Accounts; iOS 15.3) 7.0.13 (first “poke” + Channels) ◆7.0.14◆ (last with mini-program floating window; supports iPad+iphone login; face unlock broken; 836563902) 7.0.15 (face unlock broken) 7.0.18 (Yue Sheng Shi minimum) 7.0.20 (iOS 14 photo permissions) 7.0.21 (Yue Kang code minimum) 8.0.3 (Apple Watch; 999-sticker limit) 8.0.8 (change notification sound) 8.0.9 (auto desktop login) 8.0.14 (folded groups) 8.0.22 (iOS 16)
• WeCom 3.1.8 (lowest login) 4.0 (live meeting)
• Weibo 6.5.1 (no ads) 7.3.1 (last before homepage live icon; no ads in comments) 8.5.3 8.12.3 (iOS 14.7) 10.11.0 11.0.0 (lowest without outdated-version warning on profile) 11.3.0 (iPad minimum without bugs)
• Weibo Intl 3.6.8 3.8.4 ◆4.1.5◆ (dark mode) 4.2.0 (last without homepage recommendations)
• vvebo 3.1.4 (last before paid IAP)
• DingTalk 4.6.10 ◆5.1.10◆ (logs) 6.0.0 (iOS 15.3) 6.0.26 (iOS 14.7)
• Qzone 8.6.1
• Baidu Tieba 6.8.5 6.9.6 (in-bar search) 7.1.0 (no launch ads) 7.3.0 (VIP can remove ads) 8.9.0 (full screen) 9.2.0 (clean, few ads) 9.3.0 9.3.1 (iOS 14.5) ◆9.6.0◆ (minimum for Safari invoke; pair with Flex for ad blocking) 9.7.1 9.7.11 9.7.5 9.8.4 (iOS 14.7) 9.9.9 (cannot edit images) 10.3.11 (dark mode) 11.0.0 (use Flex to block ads) 11.3.2 12.0.5
• Tieba internal 1.0.8 (delisted; downloadable if purchased; no ads; no search)
• Zhihu ◆6.0.0◆ (search works; no launch ads; iOS 15.3) 6.5.0 6.7.0 (last without launch ads) 6.14.0 (dark mode) 6.65.0 6.62.0 (iOS 14 widgets) 6.67.0
• Douban ◆7.0.0◆ 7.15.0
• Hupu ◆7.3.2◆
• Coolapk 3.1 (last without launch ads; messages broken) 4.5.1 (iOS 15.3) ◆4.9◆ (847456398) 4.9.1 (last blue theme; 848225916)
• WeFeng 5.5.1 (last without ads but network issues) ◆6.4◆ (840909450; use Flex to block ads) 6.6.5
• NGA 5.3.0 (no launch/banner/interstitial ads) 6.0.8 (iOS 14.5/16) 6.3.2
• PiPiXia 1.3.2 1.3.9 (no ads)
• Twitter 7.37.2 (translation) 7.60 (dark mode) 8.44 (no Fleets) 8.47.1 (iOS 14 photo permissions)
• Instagram 124.0 (dark mode) 165.0 (iOS 14 photo permissions)
• Nicegram 8.3.2 (bypasses forced update)

Shopping

• Taobao 7.12.20 (iOS 15) 8.0.0 (iOS 15.6) 8.6.10 8.8.0 (iOS 14.3) 9.1.1 (iOS 14.3 crash) 9.2.0 (no visual search) 9.17.0 (iOS 14 widget) 9.20.0 (find similar) 10.12.20 (dark mode; earlier versions show update popups) ◆10.13.0◆ (no 618 icon; set region to Global to cut ads) 10.14.10 (850964824)
• Taobao Lite 4.16.4 (overseas)
• Diantao 1.7.6
• Alibaba 7.20.0 (iOS 14.7) ◆9.4.6◆
• JD ◆9.5.0◆ (minimum for payment) 10.0.6 (fix iPhone 12 cart misalignment) 10.2.0 (iOS 16 image send) 10.2.4 (no Double-11 icon)
• Pinduoduo 4.23.0 (iOS 15.6) 4.56.0 4.76.0 (iOS 14.5) 5.2.0 5.12.0 ◆5.30.0◆ (wallet) 5.31.1 (iOS 14 widget)
• Weidian 5.4.7
• Smzdm 9.8.16
• Manmanbuy ◆3.2.40◆
• Suning 7.7.3.8
• Shihuo 6.54.0 (iOS 14.7)
• IdleFish 6.2.5 (minimum usable) 6.2.9 (iOS 14.7) 6.3.7 6.4.80 6.6.60 ◆6.9.80◆ (minimum for bidding; transfers broken) 7.1.10 (iOS 15; overlay install) 7.4.10 (elder mode; phone-login only; iOS 13.7)
• Zhuanzhuan 5.6.1
• Xiaohongshu 6.63 (use Flex to block ads) ◆6.86◆ (iOS 14.7) 7.28 (browse history minimum) 7.40 (new emoji)
• Meituan 11.4.401 (Meituan Youxuan unusable) 11.6.406 11.10.203 (iOS 15)
• Taopiaopiao 10.4.1 (iOS 14 widget)
• Ctrip 8.22 (overlay install)
• Qianniu 8.13.0 9.3.0

Finance

• Alipay 10.1.35 10.2.8 (minimum where bank card list works) ◆10.2.10◆ (iOS 14 widget; mini-programs OK) 10.2.15 10.2.38 (fast cold start) 10.2.53 (can remove Life channel; remove all homepage cards) 10.2.59 (can disable homepage recommendations) 10.2.76 (elder mode)
• UnionPay ◆8.0.7◆ (minimum for transfers)
• Shark Ledger Pro 2.0 (no community bloat)

Entertainment

• Bilibili 6.1.3 6.61.0 (dynamic avatar minimum; need i4Tools download)
• Bilibili HD 2.3 (last with toggleable startup animation)
• Bilibili Concept 2.13.2 2.15.4 2.4 6.48.0 (global mini-window)
• bilibili-弹幕動畫戲劇線上看 2.7.0 (TW region; no game center, mall, live, startup ads; vertical videos fail; 837359696) ◆3.0.3◆ (dark-mode hot switch; live; long-press speed; 841150268) 3.0.4
• Douyin 15.0.0 (840791014; requires overlay install) 16.5.0 (playback speed)
• Douyin Lite 15.0.0 16.8.0 (playback speed)
• TikTok 1.2.2 (last unblocked region; cannot register/login; TW region 823491616)
• YouTube 15.28.2 16.05.9 (CC subtitle button)
• Kuaishou 6.11.8 9.8.50
• iQIYI 5.7.1 (no pre-roll ads) 7.0.1 (login without SMS) 8.12.5 (download 3 videos at once) 9.7.0 (Cinema) 9.8.0 (last version supporting web QR login on secondary device)
• Tencent Video 7.8.0 (iOS 14.7) 8.2.48
• Youku 6.12.0 (no ads) 7.5.5 (iOS 14.7)
• Douyu 4.920
• Huya 6.5.1
• Yunting ◆5.8.0◆ (no ads) 6.0.1 (no launch ads)
• NetEase Cloud Music 2.7.1 3.1.1 (no lossless) 3.7.5 4.0.1 (iOS 14.5) 4.1.2 (last without video; supports new ncm format) ◆4.2.2◆ (first full screen; adds launch ads) 4.3.1 (minimum for Bilibili SIM free traffic) 4.3.2 4.3.5 (old UI last) 5.5.0 (iOS 15.6) 5.6.0 (classic player last) 5.6.3 (iOS 14.5) 5.7.0 (iOS 15.3, iPhone 13) 5.8.1 (overlay install) 6.4.7 (first dark mode) 6.5.2 7.0.0 (iOS 15) 7.2.0 (Listen Together first) 7.2.10 (lyric pinyin) 7.3.01 (top-right spinning disc) 7.3.20 (iOS 14 widget) 8.0.10 (custom bottom tabs; PiP video)
• QQ Music 8.8.5 8.9.9 (web login after uninstalling QQ) 9.2.8 (no live/launch ads) 9.3 (iOS 14.7) 9.7 10.12 (no update popup)
• Bodi Music 1.2.6 (free; import external playlists)
• MOO Music 1.6.5 (simple UI) 2.3.2 (iOS 15.1; earlier versions had clear-history bug)
• Kugou Music 3.9.4 4.0.7 7.0.6 8.8.6 9.7.6
• Kuwo Music 4.9.2
• Xiami 6.1.8
• Quanmin K Ge 3.7.6
• Changba 8.2
• Ximalaya ◆6.6.15◆ 7.0.8
• QingTing FM 6.2.6 (no audio ads)

Reading

• Toutiao 3.1.1 6.0.3 (no ads; iOS 11) 6.2.9 (first with chat emoji) 6.6.0 (can block ads on jailbreak) 7.0.5 7.0.9 7.9.2 (dark mode) 7.9.6 (long-press speed)
• Tencent News 5.1.6
• The Paper 8.1.7 (widget; no hot list)
• WeRead 4.3.0 (dark mode) 5.0.4 (widget) 5.4.0 (Apple Pencil)
• Qidian 4.7.0 (first landscape; block updates with Flex) 5.0.1
• Zhuishushenqi 2.23.1 (no coins; chase-mode) 2.25.1 (switch sources) 2.24.24 (switch sources)
• iReader Pro 6.3.0
• Shuqi 3.9.10 (no horizontal ads)
• BiQuGe 5.5 (no ads/popups)
• Xiangse Guige 2.33.0
• NetEase Comics 4.2.0 (use Flex3 membership patch)

Education

• Notability 10.6 (last free version)
• Chaoxing Study 4.5.6
• Baidu Chinese ◆2.0◆ (minimum usable; 820243885) 2.1.0
• Tencent Translate 3.1 (basic only)
• 50-On Origins ◆1.3.5◆ (last before paywall)
• Sports World Campus 3.1.2 (can reject updates minimum)
• Perfect Campus 5.3.1 (Alipay recharge minimum)
• Driving Test Coach 7.6.4 (iOS 14.5)
• Jiaxiao Yidian Tong 6.1.0 (images display; minimum) 6.1.7 (no update popups; no video lessons; overlay to refresh question bank) 8.1.0 (no paid popups on wrong answers) 9.1.0 (can refresh bank) 9.3.1 10.2.0 (last without launch ads)

Tools

• Baidu 12.13.0.10
• Amap 9.05.2 (minimum working search) 10.55.0 (last with old voice) 10.80.0 (swipe-back)
• Baidu Maps 8.6.5 (no ads) 10.24.5 (full screen) 15.2.5 (dark mode)
• Hello Bike 5.74 (only version without jailbreak detection; iOS 14.5) 6.1.5 (UnionPay) 6.2.0 (no bar on logo)
• Netease Dashen 2.5.0 2.8.0 (no launch ads; QR login works)
• Baidu Netdisk 6.6.0 (upload video without VIP) 6.7.3 (rename extensions/open unknown shares) 6.9.4 7.0.0 8.13.1 (no launch ads) 10.1.70 (messages normal)
• Tianyi Cloud 4.5.0 (last free video upload; daily sign-in adds space)
• Weipan 3.3.2 3.4.3
• Youdao Note 5.7.1 (few ads) 5.9.8 (note + collaboration coexist)
• Evernote 9.3.9
• WPS Office 9.0.0 (no ads) 10.14.0 (iPhone 13 Pro/iOS 15 search & replace without crash)
• Outlook 4.2210.3 (no meeting room booking, etc.)
• JiSpeed 5.16.0 (iOS 15)
• Meiyan Camera 9.4.00 (minimum usable)
• Meitu Xiuxiu 6.0.5 7.0.1
• FiLMiC Pro ◆6.19.7◆ (last pre-subscription; compatible with IAP tweak; 851609129)
• Halide Mark II ◆1.17.3◆ (last paid version; 837845559)
• Sogou Input 2.2.1 (no search) 4.0.0 4.4.0 5.0.1 (clipboard) 7.2.6 (iOS 14.7) 10.22
• Baidu Input 6.0.2 (super skins; no clipboard) 7.1.1 (no voice input)
• iFly Input 6.0 8.0.2000
• UC Browser 10.5.5 (no headlines; cache video) 10.7.11 (cache video) 11.3.1 (cache video, download files)
• QQ Browser 6.1.1 (cache video) 6.3 (can disable headlines) 8.2.1 (first free-traffic version)
• Quark Browser 2.2.9 2.3.1.913 (iOS 15.6)
• Alook 8.2 9.6 11.5
• Speedtest China by Ookla 4.2.10 (839306602)
• DAMA (image privacy blur) 1.1.27 (bug unlocks premium; 850202162)
• Pixelmator Photo 2.0.12 (last before subscription; 850784840)
• Pin 3.2.2 (last with full Labs) 3.15.1
• Keep 5.18.0 7.5.3
• Mi Fit 4.0.3 (minimum login)
• Strava 23.0.0
• Widgy Widgets 1.7 (iOS 14.5)
• Autohome 9.10.6 (vehicle detail pages fail to load)
• Didi 5.1 (first bike version) 5.2.20 (iOS 14.5)
• Caiyun Weather 6.0.4 (dark mode) 6.0.5 (iOS 14 widget) 6.0.7 6.1.1
• Caiyun Weather Pro 5.0.10 (last before paid)
• BaiMiao 2.1.6 (no membership limits, no ads)
• Documents by Readdle (CN) 5.4 (audio/video sniffing, m3u8 download)
• TV Assistant 5.5.0 (audio/video download, file export)

Grab Old Versions

Mobile (TrollStore)

1. Search “新有趣工具箱” in WeChat to install TrollStore online.
2. Download DowngradeApp.
3. Open TrollStore and install.

Desktop

If the “i4Tools Lite” catalog lacks the version you want, capture it with iTunes on desktop. iTunes (v12.6.5.3) download: Apple Support. Packet-capture tool (release post: 52pojie): Lanzou.

Steps:

1. Install iTunes v12.6.5.3 (built-in App Store).
2. Open iTunes and sign in with Apple ID.
3. Click the third-row “♫ Music” tile, click “Edit Menu”, check “Apps”, click “Done”, then click “Apps” to switch view.
4. Download the capture tool. Create a folder for the .exe files, e.g., D:\Program Files\iOS-old-apps (the program creates data files alongside it).
5. Open the capture tool, enter the app name, press Enter to search.
6. Double-click the version you want.
7. Click “Start intercept”.
8. In iTunes, enter the app name in the top-right search box.
9. Click “Store” -> “Search in iTunes Store for …”.
10. Under the correct device tab (iPhone/iPad), click “Get”/“Download”.
11. After download, click “Library” (third row).
12. Right-click the app icon -> “Show in Windows Explorer” to find the .ipa.
13. Back up the package for later use.
14. Install [i4Tools] (“爱思助手”).
15. Open i4Tools -> Apps/Games -> Import Install to sideload the .ipa.

Notes:

• App packages are tied to the Apple ID used to download them. To share, the recipient must log into that Apple ID in App Store before importing via desktop.
• Once an app is added to the iTunes download queue you can stop intercepting and move on to the next.

Block In-App Update Prompts

1. Jailbreak.
2. Install AppSync Unified and Filza File Manager.
3. Open Filza.
4. Go to the app install directory /private/var/containers/Bundle/Application/xxx.
5. Search info.plist and open the largest one.
6. Expand Root and set the version number to the latest release. If the version is x.x.x but the string is x.x.x.x, change the first three segments, leave the last.

Tips & FAQs

Disable Auto-Updates

Go to Settings -> App Store and turn off Automatic Downloads: App Updates.

App is larger than App Store install

For apps installed via desktop i4Tools, go to Settings -> General -> iPhone Storage, tap Offload App then Reinstall App to shrink size. Apps installed with i4Tools Lite or with iTunesMetadata.plist removed cannot be shrunk.

App removed from sale

If the captured version shows “no longer available,” register a new Apple ID and grab once after first logging into iTunes—the pulled version can still be captured.

Downgrade TIM to the first release (1.0.4)

1. Use old iTunes to capture these four: 3.1.1, 2.5.1, 1.1.6, 1.0.4.
2. In i4Tools, overlay-install in order: 3.1.1 → 2.5.1 → 1.1.6 → 1.0.4.
3. Run the app once after each overlay.
4. If a version crashes, reopen a few times.
5. After downgrading to 1.0.4, go to Settings -> General -> iPhone Storage, tap Offload App then Reinstall to reduce size.

Contributors

• ⑨RealZJMM怪我丑 (Plex)
• 某年某月某日
• Sunshine
• 1053083281
• 酌酒饮孤单
• 康晓锋
• iuuan
• 偏执又可爱的me
• 那時明月
• 阿立
• 启示录

Repost Notice

View the original

All credit to the original author; this is just a mirrored record.

This is a pure tutorial compiled and tested by me. Any similarity to other guides is coincidental.

If you need related apps, search and download yourself or contact me (comments, etc.).

Let’s dive in.

Prep

If you’ve never registered Kuaishou Speed Edition or never withdrawn cash, first register on a real device and bind WeChat (Settings -> Account & Security -> WeChat) to make withdrawals easier.

If this is a second account and you plan to bind Alipay, also log in on a real device first. Alipay can only be bound during withdrawal, so keep the real-device session logged in. After you capture the packet in a VM (BlackBox or VMOS Pro) and send it to me, I’ll run coins until it’s withdrawable; then you withdraw on the real device via Alipay.

Android guide

BLACKBOX

This method uses BlackBox for packet capture—simpler and faster, and my preferred approach.

Install apps

Install BlackBox64 or BlackBox32, HttpCanary, and the “Yellow Chicken anti-disconnect” module.

If your Kuaishou Speed Edition app is 32-bit, use BlackBox32; if 64-bit, use BlackBox64. If you don’t know, install both and see which one can add the app.

Configure BlackBox

1. Install Kuaishou Speed Edition on the real device. Open BlackBox, tap the bottom-right + to add the app.

   If you can’t find Kuaishou Speed Edition in the picker, switch BlackBox32/64.

2. Tap the three dots top-right -> Settings, enable Xposed Framework, Hide Xposed Framework, and Hide Root.

   

   

3. Open Module Management, tap +, add the Yellow Chicken anti-disconnect module, and check it.

   

4. Restart BlackBox twice (open/close a couple of times).

Target app

In HttpCanary, go to Settings -> Target apps, tap +, choose your virtual app BlackBox (32 or 64—pick the one you’re using).

URL keyword

Back on the main page, tap the search icon, scroll to the bottom, choose URL keyword, and enter salt.

This step is crucial—enter it exactly. If HttpCanary is killed in the background, you must re-enter the keyword.

Start capture

On the main page, tap the bottom-right paper plane icon to start capturing.

If prompted to install a cert, you can skip or install. If you fail to capture later, installing is fine (I captured without installing).

Log into Kuaishou

In BlackBox, open Kuaishou Speed Edition (the one inside BlackBox, not the real device). Make sure it’s the first launch and you’re logged out. Log in with phone + SMS. After login, go to Settings -> Account & Security -> WeChat to bind WeChat, or wait to bind WeChat/Alipay during withdrawal.

If entering the phone/OTP does nothing, kill BlackBox and retry.

If you capture nothing or can’t pass security checks, uninstall the Kuaishou app inside BlackBox, kill related apps, re-add, start HttpCanary, and log in again.

Capture success

If all went well, return to HttpCanary—you should see captured traffic.

Save request

Open any request, go to Request, tap the top-right save icon, choose Save request.

Send the file

Files are saved under HttpCanary->download. Send the latest file (use a file manager). There may be two files: one ending with header and one with raw. Send the header file.

Cleanup

After capturing, you can close all apps. Do not log out of the Kuaishou account, or the data becomes invalid. Just leave the apps alone.

Create a second VM

If you need a second account, do not log out the first. Create another virtual space and repeat the steps in BlackBox.

Swipe left on BlackBox home to switch spaces:

Other steps are the same.

Note:

The second account must use a different phone number and withdrawal method. If you only have one WeChat, capture and send the data first; after I process the coins, bind Alipay during withdrawal.

VMOS PRO

This method uses VMOS Pro. It’s more complex but isolates from the real device and suits multiple accounts.

Install apps

Install HttpCanary (specific version), VMOS Pro (virtual machine with root), and MT Manager (file manager).

HttpCanary must support exporting system certs; I use v9.2.8.1.

Export certificate

Open HttpCanary -> Settings -> SSL certificate settings -> Export HttpCanary root certificate -> System Trusted (.0).

Note the path, typically HttpCanary/cert.

Create VM

Open VMOS Pro and pick the second template—it defaults to Android 7.1.2.

Import files

Inside the VM, tap the dock Add icon and import Kuaishou Speed Edition, MT Manager, and the cert file from above. They’ll auto-install in the VM.

Important: ensure you imported the cert. If VMOSfiletransferstation later appears empty, you forgot the cert!

Get root

In VM settings, enable Superuser. A Superuser app should appear, meaning root access is active.

Grant permissions

Open MT Manager inside the VM. When prompted, allow Superuser (check “remember”).

Move the cert

In MT Manager, left pane: VMOSfiletransferstation; right pane: system. If there’s no system folder, root access failed—redo the previous step.

On the right, go to system->etc->security->cacerts. On the left, open VMOSfiletransferstation to find the cert file. Long-press the cert on the left and Copy—it will appear on the right. Different filenames are normal.

Long-press the cert file on the left and tap Copy. It should appear on the right. Filename differences are normal.

Target app

Back on the real device, open HttpCanary -> Settings -> Target apps, tap +, choose VMOS Pro.

URL keyword

On the main page, tap the search icon, scroll to the bottom, choose URL keyword, and enter salt.

This is critical—enter it correctly.

Start capture

Return to the main page and tap the bottom-right paper plane to start.

If you’re rooted, you can move the cert directly on the real device—no VM needed. The “Install root certificate to capture SSL data” banner is normal for non-root users; ignore it.

Log into Kuaishou

Inside the VM, open Kuaishou Speed Edition for the first time (logged out). Log in with phone + SMS. After login, go to Settings -> Account & Security -> WeChat to bind WeChat, or bind during withdrawal.

If login fails or buttons do nothing, kill all related apps and retry, or try later. Another trick: while logged in, close the VM, wait (even 30 minutes), then start HttpCanary, start the VM, and open Kuaishou—the packet may be captured instantly.

Capture success

If all went well, HttpCanary should show three entries of ~2.8kB each.

Save request

Open any entry, go to Request, tap save top-right, choose Save request.

Send the file

Files live in HttpCanary->download. Send the latest file. You may see two files: one ending with header, one with raw. Send the header file.

iOS guide

My iOS setup:

Jailbroken, iOS 14.2.1

Tools: Crane (jailbreak tweak for multi-instance), Thor (paid sniffer), Stream (free sniffer)

To be continued.

In this post I explained how to add Fancybox to the cactus theme. I later swapped my comment system to Twikoo. While upgrading today I stumbled across: Integrate Twikoo with lightGallery to zoom comment images. Since Fancybox is already installed, I might as well hook Twikoo to Fancybox.

Process

Following the author’s idea, locate the twikoo.init code and modify it. I’ll focus on what changed rather than pasting the original.

Here’s the final code:

Final result

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

twikoo.init({
  // envId: 'abcdef',
  // ...
  onCommentLoaded: function () {
    var commentContents = document.getElementsByClassName('tk-content');
    for (var i = 0; i < commentContents.length; i++) {
      var commentItem = commentContents[i];
      var imgEls = commentItem.getElementsByTagName('img');
      if (imgEls.length > 0) {
        for (var j = 0; j < imgEls.length; j++) {
          var imgEl = imgEls[j];
          // Skip emoji, only wrap normal images
          if (imgEl.className != 'tk-owo-emotion') {
            var aEl = document.createElement('a');
            aEl.setAttribute('data-fancybox', 'gallery');
            aEl.setAttribute('data-src', imgEl.getAttribute('src'));
            aEl.setAttribute('data-caption', 'Comments: '+imgEl.getAttribute('alt'));
            aEl.appendChild(imgEl.cloneNode(false));
            imgEl.parentNode.insertBefore(aEl, imgEl.nextSibling);
            imgEl.remove();
          }
        }
      }
    }
  }
});

Below is how I arrived here—use the idea even if your theme code differs.

Read the original code

For reference, here’s the original snippet:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

twikoo.init({
  // envId: 'abcdef',
  // ...
  onCommentLoaded: function () {
    var commentContents = document.getElementsByClassName('tk-content');
    for (var i = 0; i < commentContents.length; i++) {
      var commentItem = commentContents[i];
      var imgEls = commentItem.getElementsByTagName('img');
      if (imgEls.length > 0) {
        for (var j = 0; j < imgEls.length; j++) {
          var imgEl = imgEls[j];
          var aEl = document.createElement('a');
          aEl.setAttribute('class', 'tk-lg-link');
          aEl.setAttribute('href', imgEl.getAttribute('src'));
          aEl.setAttribute('data-src', imgEl.getAttribute('src'));
          aEl.appendChild(imgEl.cloneNode(false));
          imgEl.parentNode.insertBefore(aEl, imgEl.nextSibling);
          imgEl.remove();
        }
        lightGallery(commentItem, {
          selector: '.tk-lg-link',
          share: false
        });
      }
    }
  }
});

The idea: find images (img) inside the comment area (tk-content), wrap each with the plugin-specific <a> markup, insert it, then remove the original image.

You can swap in other plugins by adjusting the markup:

lightgallery:

<a class="tk-lg-link" data-src="xxx.png" href="xxx.png"><img src="xxx.png" alt="xxx"></a>

fancybox:

<a data-fancybox="gallery" data-src="xxx.png" data-caption="xxx"><img src="xxx.png" alt="xxx"></a>

Change the tags

Copy the structure and switch attributes to Fancybox. For example, lightGallery needs class="tk-lg-link", whereas Fancybox needs data-fancybox="gallery". So replace aEl.setAttribute('class', 'tk-lg-link'); with aEl.setAttribute('data-fancybox', 'gallery');, and so on.

Remove unused code

lightGallery also required:

1
2
3
4

lightGallery(commentItem, {
          selector: '.tk-lg-link',
          share: false
        });

Fancybox doesn’t need this, so delete it. After these changes, try it out—and you’ll hit another issue.

Comment images vs emoji

Emoji are also <img> elements (class="tk-owo-emotion"), so the plugin was wrapping emoji too, which is unnecessary.

Handle emoji

Add a simple check to skip emoji:

1
2
3
4
5
6
7
8
9

if (imgEl.className != 'tk-owo-emotion') {
  var aEl = document.createElement('a');
  aEl.setAttribute('data-fancybox', 'gallery');
  aEl.setAttribute('data-src', imgEl.getAttribute('src'));
  aEl.setAttribute('data-caption', 'Comments: '+imgEl.getAttribute('alt'));
  aEl.appendChild(imgEl.cloneNode(false));
  imgEl.parentNode.insertBefore(aEl, imgEl.nextSibling);
  imgEl.remove();
}

Make sure this condition sits inside the second for-loop.

Result

That’s it—clean galleries without wrapping emoji.

Thanks

iMaeGoo

twikoo

fancybox

I’m at home on vacation and I need to use EasyConnect to connect to the school’s intranet. I occasionally need to connect to the school’s server or access intranet pages, so I plan to download it.

But! EasyConnect is a rogue software, which is often referred to as a cancerous software. I then started looking for ways to deal with it.

Under macOS, EasyConnect primarily behaves as follows.

1. background resident processes EasyMonitor and ECAgentProxy, both running with root privileges.
2. automatically installs root certificates, including the system root certificate and FireFox root certificate, and reinstalls them after deletion.

ps: Resident processes with root privileges means that these processes can read and write anything; installing root certificates means that they can directly perform man-in-the-middle attacks and tamper with data at will.

Method 1

This method is suitable for computers with EasyConnect already installed, with minimal tossing.

1. use sudo su to obtain administrator privileges.

2. Delete (rm)

   /Library/LaunchDaemons/com.sangfor.EasyMonitor.plist.

3. delete /Library/LaunchAgents/com.sangfor.ECAgentProxy.plist.

4. reboot the Mac.

5. Open the keychain and delete the certificate under System Keychain - System - Certificates (keyword sangfor).

Be sure to restart the Mac before deleting the certificate, as the EasyMonitor process in the background will automatically import the certificate after it has been deleted.

This eliminates the problem with EasyConnect. However, so far it causes EasyConnect not to connect, so we need to start EasyConnect in a secure way. Two scripts can be written to achieve start and exit.

startEasyConnect.sh

1
2
3
4

#! /bin/bash

/Applications/EasyConnect.app/Contents/Resources/bin/EasyMonitor > /dev/null 2>&1 &
/Applications/EasyConnect.app/Contents/MacOS/EasyConnect > /dev/null 2>&1 &

killEasyMonitor.sh

1
2
3
4
5