llama.cpp b10549 Released with Tensor Splitting for LFM2 and LFM2MOE Models
The llama.cpp project has released version b10549, which introduces tensor splitting support for LFM2 and LFM2MOE models. This pull request was assisted by deepseek-v4-flash. This update allows users to run Liquid AI's LFM2 and LFM2MOE models across multiple GPUs, reducing latency and enabling the execution of larger model configurations that exceed a single GPU's memory. Tensor splitting (TP) in llama.cpp distributes transformer layers across multiple GPUs to parallelize workloads, though performance remains bottlenecked by GPU interconnect speeds. The release also provides pre-built binaries for various platforms, though macOS KleidiAI and openEuler support are currently disabled.
## BACKGROUND
llama.cpp is a popular open-source LLM inference engine written in C/C++ designed for efficient local execution. Liquid Foundation Models (LFMs), developed by Liquid AI, are designed to run efficiently on edge devices like phones and laptops under tight memory constraints. Tensor splitting (or tensor parallelism) is a technique used to divide a model's tensors across multiple hardware accelerators to fit larger models into memory and speed up computation.