b8361

The llama.cpp project, the popular open-source C++ implementation for running Llama and other AI models, has released a critical fix in commit b8361 that resolves system crashes on AMD APU and integrated GPU systems. The issue stemmed from how the software handled the hipMemAdviseSetCoarseGrain function call—a performance optimization hint that's not applicable to unified memory architecture (UMA) systems found in AMD APUs like the Strix Halo (gfx1151). Previously, the code treated any error from this function as fatal, causing immediate crashes on affected systems.

The fix, contributed with assistance from Claude Sonnet 4.6, changes the error handling to treat hipMemAdviseSetCoarseGrain as an optional performance hint rather than a required operation. The updated code now calls the function without error checking and clears any resulting errors with hipGetLastError(), preventing crashes while maintaining functionality. This addresses a specific compatibility problem where AMD APUs on Windows were affected by a ROCm runtime bug limiting hipMallocManaged to approximately 64GB regardless of available system RAM.

Beyond the immediate fix, the update includes enhanced debugging capabilities with GGML_LOG_DEBUG pre-allocation logging to help diagnose memory issues on APU systems. The team also added storage of totalGlobalMem in device information for better system reporting. This fix is particularly important for users running AI models on consumer-grade AMD systems with integrated graphics, expanding the hardware compatibility of the llama.cpp ecosystem while upstream fixes for the underlying ROCm runtime bug continue development.