llama.cpp on a jailbroken PS5: run any GGUF model natively on the console, starting with Qwen 3.8
patches/0002: thread affinity and priority on the PS5
Mert Cobanov51c52d8acb63
diff --git a/patches/0002-ps5-thread-affinity-and-priority.patch b/patches/0002-ps5-thread-affinity-and-priority.patchnew file mode 100644index 0000000..d307727--- /dev/null+++ b/patches/0002-ps5-thread-affinity-and-priority.patch@@ -0,0 +1,99 @@+diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c+index 8620c6c..dd1bbed 100644+--- a/ggml/src/ggml-cpu/ggml-cpu.c++++ b/ggml/src/ggml-cpu/ggml-cpu.c+@@ -2695,6 +2695,51 @@ static bool ggml_thread_apply_priority(int32_t prio) {+ return true;+ }+++#elif defined(__PROSPERO__)++++// PS5: Sony's thread calls stand in for the Linux ones. A payload may only run++// on the CPUs in its affinity mask (0xea00 on firmware 13.60), shared with the++// console's own services, so a mask outside it fails and priority is only ever++// lowered. Sony priorities run 256 (first) to 767 (last); threads start at 700.++int scePthreadSetaffinity(pthread_t thread, uint64_t mask);++int scePthreadSetprio(pthread_t thread, int prio);++++static bool ggml_thread_apply_affinity(const bool * mask) {++ uint64_t bits = 0;++ for (int i = 0; i < 64 && i < GGML_MAX_N_THREADS; i++) {++ if (mask[i]) {++ bits |= 1ull << i;++ }++ }++++ const int err = scePthreadSetaffinity(pthread_self(), bits);++ if (err != 0) {++ fprintf(stderr, "warn: failed to set affinity mask 0x%llx : 0x%08x\n", (unsigned long long) bits, (unsigned) err);++ return false;++ }++++ return true;++}++++static bool ggml_thread_apply_priority(int32_t prio) {++ if (prio == GGML_SCHED_PRIO_NORMAL) {++ return true;++ }++ if (prio != GGML_SCHED_PRIO_LOW) {++ // Spinning workers above the system's own threads starve them.++ fprintf(stderr, "warn: thread priority %d is not raised on the PS5\n", prio);++ return false;++ }++++ const int err = scePthreadSetprio(pthread_self(), 720);++ if (err != 0) {++ fprintf(stderr, "warn: failed to lower thread priority : 0x%08x\n", (unsigned) err);++ return false;++ }++++ return true;++}+++ #else // unsupported platforms++ static bool ggml_thread_apply_affinity(const bool * mask) {+@@ -3296,9 +3341,11 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int+ if (threadpool->pause) {+ // Update main thread prio and affinity to match the threadpool settings+ ggml_thread_apply_priority(threadpool->prio);++#if !defined(__PROSPERO__) // PS5: see ggml_threadpool_new_impl+ if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) {+ ggml_thread_apply_affinity(threadpool->workers[0].cpumask);+ }++#endif++ // resume does cond broadcast+ ggml_threadpool_resume_locked(threadpool);+@@ -3375,9 +3422,14 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(+ if (!threadpool->pause) {+ // Update main thread prio and affinity at the start, otherwise we'll do it in resume+ ggml_thread_apply_priority(threadpool->prio);++#if !defined(__PROSPERO__)++ // PS5: the calling thread keeps the whole mask. Threads it starts later++ // inherit its affinity, so pinning it to one CPU would pin them too++ // (PS5SX2 hit this with a ticker thread, commit f9abb4a).+ if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) {+ ggml_thread_apply_affinity(threadpool->workers[0].cpumask);+ }++#endif+ }+ #endif // GGML_USE_OPENMP++diff --git a/common/common.cpp b/common/common.cpp+index 7e89648..f0693fe 100644+--- a/common/common.cpp++++ b/common/common.cpp+@@ -275,6 +275,11 @@ bool set_process_priority(enum ggml_sched_priority prio) {+ if (prio == GGML_SCHED_PRIO_NORMAL) {+ return true;+ }++#if defined(__PROSPERO__)++ // PS5: what setpriority does to a payload is unknown. ggml lowers its own++ // threads' Sony priority for --prio -1 and refuses to raise them.++ return prio == GGML_SCHED_PRIO_LOW;++#endif++ int p = 0;+ switch (prio) {