llama.cpp and ik_llama.cpp FP4 Support

·2026.04.26 00:42

Key point

Different FP4 inference support has been added to llama.cpp and ik_llama.cpp.

Nvidia-style NVFP4 support has been merged into llama.cpp. The related CUDA kernels went into mmq.cuh, mmvq.cu, convert.cu, and others.

ik_llama.cpp already supports MXFP4, with CPU-side implementations for AVX2, NEON, Zen4, and GPU-side implementation for CUDA.

The two formats are not the same.

Since both are in the 4-bit floating point family, VRAM savings are expected once model support follows.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.