Igor Pissolati

@grantbey It's finished, but since the maintainers showed no interest whatsoever in merging it, I didn't resolve the merge conflicts. If you can't do that by yourself, you should just...

@grantbey I rebased the PR to the last master commit :+1:

@goerch Here is a simple test: https://github.com/ggerganov/llama.cpp/pull/1931/commits/6f7dabab441566078446ef868e573cd309fe62be

@goerch Yeah, that's fine. You just have to be aware about: https://github.com/ggerganov/llama.cpp/blob/6f7dabab441566078446ef868e573cd309fe62be/llama-util.h#L554-L555 The trie algorithm used in the PR is a port from the huggingface repository, as written in the...

The important part is the `split` method https://github.com/ggerganov/llama.cpp/blob/6f7dabab441566078446ef868e573cd309fe62be/llama-util.h#L575-L577 If using the ported version isn't an option, it would be necessary to reimplement it using the other trie algorithm. >IANAL, but...

> I'll recommend to either implement a trie from scratch, or use a linear search algorithm - we are not tokenizing billions of tokens, so not sure what we gain...

Looks like this PR was superseded by https://github.com/ggerganov/llama.cpp/pull/3538, from what I could see it looks great. Thanks for your attention @goerch! I don't think I have anything more to contribute.

@kolinfluence You can find the instructions for training here: https://github.com/ggerganov/llama.cpp/blob/master/examples/train-text-from-scratch/README.md But this is the command that I used: ```sh train-text-from-scratch --vocab-model "chronos-13b.ggmlv3.q4_0.bin" --ctx 64 --embd 768 --head 12 --layer 6...

It looks good, but as I commented at the source, in my point of view, this is not advisable because it would increase the number of requests received by Bancho,...

I'm too lazy to make a list of friends, if you do it would help :v This problem that you mentioned in commit can be because you isn't setting the...