Bach Le

First, this is amazing work. This makes me wonder whether the entire sampling API should be pulled into something like `llama_samplers` instead. External samplers can evolve independently of the core...

> We may want to ignore `-z start-stop-visibility=hidden` for the sake of compatibility with programs that follow the recommendation in this man page. Other than that, I don't think there's...

Thanks. My knowledge of linker is limited but is it a matter of: 1. Add the visibility to context: https://github.com/rui314/mold/blob/d432e987a019ba213a21cfed89b01ba9041e1a2c/elf/mold.h#L1627 2. Parse cmd args and save the visibility instead of...

That's a local link though. But bgfx doesn't even have file system access (and I like it that way), how will you deal with include file?

I kinda need this too. This is only possible in macro right?

https://github.com/open-telemetry/opentelemetry-specification/blob/main/specification/trace/semantic_conventions/span-general.md#general-thread-attributes On the topic of `thread.id` attribute, `erlang:system_info(scheduler_id)` is suggested. While this is kind of correct, I'm not sure how useful it is compared to say: `self()`. But formatting `self()`...

On the same topic `code.function` also requires some string/binary construction from `?FUNCTION_NAME` and `?FUNCTION_ARITY`. It's probably negligible but it's not nothing, even in the non-recording case. There are 2 ways...

> > Doesn't seem that the `eos_token` is working with either of the convert scripts in this PR > > @m18coppola the instruct models use two different EOS tokens: the...

@teleprint-me Yep, you just have to stop on eot_id instead which is: `128009`. You can use the tokenization tool to test: https://github.com/ggerganov/llama.cpp/blob/master/examples/tokenize/tokenize.cpp ``, `` , ``, `` are all mapped...

@teleprint-me Are you saying that it's a happy coincidence that the current llama.cpp implementation happens to tokenize correctly or there exists character sequences out there that will be tokenized incorrectly?