Your own GGUF
Hello, I have a few questions that I can't seem to find a clear answer to. Does it make sense to make your own GGUF? I noticed that when I compile llamacpp (vulkan or rocm), the processing and generation is a bit better, does it work similarly with doing GGUF yourself? If I use Vulkan, is it worth doing GGUF using llama-quantize vulkan version (not rocm version)? To what extent does it make sense to place certain model elements at higher precision (conversation, document analysis)? I use gemma 4 31B the most.
评论
?
参与讨论