Pure GPU id say use vllm, honestly probably no need to get into SGLang for what your doing, tips for VLLM are honestly just copy other peoples configs, use prebuilt dockers, etc, Vllm has alot of knobs and dials and being on RTX 6000 pros in my experience if your doing it from scratch your gonna need some trial and error.

If your doing offloading (GPUS + CPU) then go with llama.cpp, its also better if you just want pure simplicity, it can obviously do pure GPU too and if VLLM on NVFP4 (nvfp4 is the quant optimzied for blackwell cards like yours, best option 90% of the time) is too much of a hastle (I have some configs on a dual RTX 6000 pro setup for minimax if you can't figure it out) then llama.cpp will make your life easier, no NVFP4 but you get the most used quant style which is GGUF, id always recommend getting ones from unsloth, the UD versions are often better in my experience and they always publish them.

AWQ is mostly for vllm/sglang, I woudnt use it unless you had too and in this case the models I suggested should have NVFP4 and if your offloading then you should use GGUFs anyways, I woudnt touch REAP in general, too much quality loss.