llama.cpp Discussions

[MillionthOdin16]

I created this as a place to discuss llama.cpp specific info that doesn't directly require action for this repo, but is still related :)

Performance wise, there's some cool stuff being investigated related to generation speed as the context grows. It looks like there's a significant drop in tokens/S, which was introduced in a last couple weeks. If they can find the cause, gen should be much faster :D

ggml-org/llama.cpp#603

[MillionthOdin16]

Just FYI, llama CPP received a big bug fix in the last hour that significantly improves token per second performance at large context lengths 🔥

[abetlen]

Pushed

[sergedc]

What version of llama.cpp is this python binding using?
I see a file called libllama.so (or libllama.dll in windows)
Is there a way to update it?

On my machine, the model load is 3 second using the llama.cpp but 40seconds using this python binding.
I know llama.cpp improved loading speed significantly recently.

[abetlen]

It depends on the version / where you install from but if you check the vendor/llama.cpp folder and select the commit / version that'll be the version installed.

[hongbopeng]

any way to bring "-n N, --n_predict N: number of tokens to predict" to python interface? without it, it only retrieves a fixed max number of words not the full content. it is defaulted at 128, but should be set at -1 for getting full content.