llama.cpp

76689995 - Merge branch 'master' into gpu-sampling

Merge branch 'master' into gpu-sampling Let's keep `master's` cumsum implementation for it's likely better AMD perf and add back pure-CUB-implementation in follow-up commit