Llama 3 BPE tokenization needs improvement

[coder543]

What is the issue?

This PR just merged on llama.cpp, which contained important improvements to how tokenization worked for Llama 3 and other models. An example of the issue is noted here.

Hopefully ollama can update to the latest llama.cpp quickly and make a new release.

OS

Linux

GPU

Nvidia

CPU

AMD

Ollama version

all versions up to this point

[MoonRide303]

You might want to wait for ggml-org/llama.cpp#6965 to be merged, too (should happen soon).

[coder543]

ggml-org/llama.cpp#6965 has been merged now. I'm unclear when things were fixed in ollama, but I just tested with 0.1.35, and I can't reproduce it anymore. Closing.

[dpublic]

The llama.cpp commit link in ollama is dated 4/30 and ggml-org/llama.cpp#6965 was merged to llama.cpp on 5/9.
So, it doesn't look like this merge was included with the last 0.1.37 ollama release.
Does that mean ollama was changed to handle the previous llama.cpp behavior and a future llama.cpp sync in ollama will change behavior?