ggerganov/llama.cpp Released b11752 with MoE Router Test Fixes
Llama.cpp release b11752 introduces a fix to make Mixture of Experts (MoE) router selection independent of backend noise, ensuring consistent outputs across different hardware backends. Additionally, it updates the device table to correctly distinguish between Apple Silicon host CPUs and their corresponding Metal GPU backends. This update improves the reliability and accuracy of regression testing for advanced architectures like DeepSeek-V2 and Beam by eliminating false-positive test failures caused by minor hardware floating-point differences. Ensuring robust testing is critical for maintaining high performance and cross-platform consistency in open-source LLM inference tools. The fix populates the selection bias with a step-2 ramp, giving every expert its own score range and preventing minor backend ties in top-k selections. Furthermore, device description collisions on Apple SoCs were resolved by appending explicit device names to the test tracking tables.
## BACKGROUND
Mixture of Experts (MoE) is a neural network architecture where only a subset of sub-networks, called experts, are activated for each token through a routing mechanism, enhancing model capacity efficiently. Llama.cpp is a popular framework designed to run large language models locally on diverse consumer hardware, relying heavily on automated tests to verify mathematical equivalence across various CPU and GPU backends.