Examples: Add text compression example.

[stduhpf]

This PR adds an example text compression scheme using a language model. This compression scheme is not optimal, but it's not too far from it.

Performance:

Testing it on the source file against classical compression schemes.

Size (bytes)	Name
1480	compress.cpp.qwen2.5-coder-1.5b-q6-k.bin
1487	compress.cpp.llama3-8b-q4-k-m.bin
1557	compress.cpp.starcoder2-3b-q8.bin
3872	compress.cpp.gz
3878	compress.cpp.bz2
3908	compress.cpp.xz
3983	compress.cpp.7z
3999	compress.cpp.zip

Usage:

Compression

./compress --mode compress -m path/to/your/model.gguf -f path/to/the/text/file.txt -o output.bin

Decompression

./compress --mode expand -m path/to/your/model.gguf -f output.bin -o output.txt

Drawbacks

It's very slow compared to traditionnal compression schemes.
It needs the exact same setup for compression and decompression. ( just changing the number of offloaded gpu layers can change the behavior enough to introduce errors )

How it works

TODO (I'm bad at explainig things, but please read the code)

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[ngxson]

Is this the same method as: https://arxiv.org/pdf/2306.04050 ?

[stduhpf]

Is this the same method as: https://arxiv.org/pdf/2306.04050 ?

Interesting, thanks for sharing. At first glance, this does look similar to what I'm doing. At least the part about the ranks is the same.
The main difference is in the compression format. I'm using a bespoke algorithm here, but maybe Arithmetic Coding (like in the paper) would be better.
Actually, this would be pretty much equivalent to arithmetic coding if we assume the token probabilities are decreasing exponentially with rank (wich is not the case in reality, making this less efficient).

[matteoserva]

Just for reference, I found an interesting implementation of arithmetic coding using llama_cpp_python:

https://github.com/AlexBuz/llama-zip