Which version should I be using now (previously clblas)?

[stl3]

I've been using llama-b3085-bin-win-clblast-x64, but in the latest versions I no longer see clblas, so I assume it has either been swapped with something better. My question is which one do I download now with similar functionality? I used clblas to split between my cpu (7950x3d) and gpu (4070S). Before when I was using clblas I was able to balance a nice usage of vram, sysram, gpu/cpu usage (wanted to avoid offloading everything to gpu, needed vram for other things). Would love some advice thank you.

[dspasyuk]

@stl3 The current version should work fine. Assuming you are using Linux:
git clone https://github.com/ggerganov/llama.cpp.git; cd llama.cpp; sed -i 's/-arch=native/-arch=all/g' Makefile; make clean && LLAMA_CUDA=1 make -j 6;
Use this manual for more details: https://github.com/ggerganov/llama.cpp/blob/master/examples/main/README.md

[stl3]

Ahh I do have linux machines but I'm actually running this on windows llama-b3085-bin-win-clblast-x64. Which current release should I use to replace clblas?

cudart-llama-bin-win-cu11.7.1-x64.zip
cudart-llama-bin-win-cu12.2.0-x64.zip
llama-b3334-bin-macos-arm64.zip
llama-b3334-bin-macos-x64.zip
llama-b3334-bin-ubuntu-x64.zip
llama-b3334-bin-win-avx-x64.zip
llama-b3334-bin-win-avx2-x64.zip
llama-b3334-bin-win-avx512-x64.zip
llama-b3334-bin-win-cuda-cu11.7.1-x64.zip
llama-b3334-bin-win-cuda-cu12.2.0-x64.zip
llama-b3334-bin-win-kompute-x64.zip
llama-b3334-bin-win-llvm-arm64.zip
llama-b3334-bin-win-msvc-arm64.zip
llama-b3334-bin-win-noavx-x64.zip
llama-b3334-bin-win-openblas-x64.zip
llama-b3334-bin-win-rpc-x64.zip
llama-b3334-bin-win-sycl-x64.zip
llama-b3334-bin-win-vulkan-x64.zip

I want to split workload between gpu+vram and cpu+sysram like I did previously with clblas (am using 7950x3d cpu with 64GB sysram and 4070 Super with 12GB vram), but also able to limit threads and ngl like in clblas.

[Mar2ck]

Since you have a modern Nvidia gpu you should go with win-cuda-cu12, the CUDA backend can do everything clblast could and more. Make sure you install CUDA toolkit if you haven't already.

[stl3]

I don't mind trying cuda except it comes with an extra 3GB to download/install toolkit whereas running clblas I could just run it from the ~8MB download (~20MB decompressed) without installing anything extra.

[Mar2ck]

Vulkan should work without installing anything extra, it's much more lightweight but not as optimized as CUDA.