llama.cpp on a jailbroken PS5: run any GGUF model natively on the console, starting with Qwen 3.8
Driver patches
Three patches to PS5_Mesa's RADV, from PS5CEMU-HAR
(patches/mesa, GPL-3.0), applied on top of the Mesa revision PS5_Vulkan pins (1ba9874):
0007: a destroyed buffer's memory waits until the GPU is done with it, so freed memory still in use cannot fault the GPU and stop the console. Always on.0008: submissions handed to a thread of the driver's own (RADV_PS5_SUBMIT_THREAD=1).0009: how long a wait for the GPU spins and sleeps (RADV_PS5_WAIT_SPIN_US,RADV_PS5_WAIT_SLEEP_US).
0008 and 0009 change nothing unless asked. On the console neither made generation faster
(2026-10-16, Qwen3.5 0.8B and 9B), so the app does not ask. Put the switches in
/data/PS5LM/env.txt to try them. To build RADV with the patches:
cd .deps/src/PS5_Mesa && git am ../../../patches/mesa/*.patch
(cd ../PS5_Vulkan && PS5_MESA_REVISION=$(git -C ../PS5_Mesa rev-parse HEAD) tools/build-radv.sh release)