Why Gemma-4 26B MoE works in HuggingFace but breaks in prod inference enginesgithub.com1 points·by maeddesg·há 2 meses·0 comments
VulkanForge – 14 MB Vulkan LLM engine that runs native FP8 models on AMD (Rust)github.com3 points·by maeddesg·há 2 meses·0 comments