Tag
User @binsquares reports that GPU acceleration on smolvm achieves nearly 90% of host performance when running llama.cpp via the Vulkan backend.