paragontasx/ferro
Native OpenCL inference engine for Intel N100-class SoCs: EmbeddingGemma-2, LiquidAI d1-omni and LFM2.5 chat in one C++ binary, on the iGPU. Measured ~2x llama.cpp's own Vulkan backend on the same N100 (2.1-2.3x); OpenAI-compatible server, CPU fallback, bit-exact batch invariance.
GitHub repository with 9 stars and 0 forks.
Language: C++
Topics: c-plus-plus, decision-models, embeddings, gguf, inference, intel-n100, llama-cpp, llm, opencl, self-hosted