Resolving core dump issues when running llama.cpp on Ryzen iGPU

Although running llama.cpp on Ryzen iGPU with Vuklan is slow, but it is free. And with a competent model such as Gemma 4, I use it to plan coding changes overnight so that I could apply them with Gemini CLI the next day.

Recently llama.cpp would core dump due to "DeviceLostError" when the context size approaching ~35k. After some trial-and-error and googling, for now the only way to resolve it is to extend the GPU driver lock timeout with "amdgpu.lockup_timeout=1000000,1000000" kernel command line parameters.