Martin Leitner-Ankerl

Thanks for the PR! Note though that the documentation is created automatically from the sourcecode. So you'd have to update the wording in nanobench.h and regenerate the documentation (see src/docs/generate.sh)....

exactly. Also, if you just use `static inline void enableEnumFlags(Name) {}` instead of `friend`, then it's not possible to make the enum type private.

No, currently that's not possible. On startup nanobench internally tunes itself to the accuracy of the `std::chrono::steady_clock`, and that wouldn't be easily possible with a customizable clock. Why do you...

Sorry currently I don't have much time to work on this. I'll keep the ticket open, but don't expect anything soon

I've tried to add it when creating the benchmark, but unfortunately the map doesn't have a default constructor so I've stopped right there. See https://github.com/martinus/map_benchmark#maps-i-couldnt-add

I just want to note, when g++ was built with the `--with-ld`, this overrides all flags, and `-B/usr/local/libexec/mold` doesn't work. That is the case in our environment, and it took...

After rebasing I get a compile error in `spend.cpp` `ChooseSelectionResult`, adding a bunch of `std::move` helps

Hi @dsiroky, thanks for the bug report! This seems to work for the try_emplace methods, but not for operator[]. I'll have a look

> probably you can reduce the n_batch to 256, or smaller to get rid of this OOM error I edited the modelfile and tried different settings for `num_batch`, and `PARAMETER...

All my problems seem to disappear when I limit the amount of VRAM for ollama. My 7900 XT has 20GB, so I have limited it now to 18GB like so...