llama.cpp on a jailbroken PS5: run any GGUF model natively on the console, starting with Qwen 3.8
diff --git a/common/log.cpp b/common/log.cpp
index 25f12684e..51b3d28fd 100644
--- a/common/log.cpp
+++ b/common/log.cpp
@@ -204,6 +204,10 @@ private:
size_t head;
size_t tail;
+#if defined(__PROSPERO__)
+ size_t dropped = 0; // entries refused since the writer last caught up
+#endif
+
bool print_entry(const common_log_entry & e) const {
if (e.is_end) return true;
@@ -239,8 +243,19 @@ public:
void add(enum ggml_log_level level, const char * fmt, va_list args) {
std::unique_lock<std::mutex> lock(mtx);
+#if defined(__PROSPERO__)
+ // PS5LM: never wait for the writer. It prints every line to stdout and
+ // to --log-file, and when that stalls a full queue froze the server
+ // thread mid-print_timings (/slots and /metrics hung, /health stayed
+ // 200). Drop the line and count it; the writer reports the count.
+ if (running && is_full()) {
+ dropped++;
+ return;
+ }
+#else
// block if the queue is full
cv_full.wait(lock, [this]() { return !running || !is_full(); });
+#endif
if (!running) {
// discard messages while the worker thread is paused
@@ -306,8 +321,16 @@ public:
std::unique_lock<std::mutex> lock(mtx);
+#if defined(__PROSPERO__)
+ // PS5LM: drop instead of waiting, as in add()
+ if (running && is_full()) {
+ dropped++;
+ return;
+ }
+#else
// block if the queue is full
cv_full.wait(lock, [this]() { return !running || !is_full(); });
+#endif
if (!running) {
// discard messages while the worker thread is paused
@@ -358,6 +381,23 @@ public:
head = next_head;
cv_full.notify_all();
+#if defined(__PROSPERO__)
+ // PS5LM: one line for everything add() dropped, once there is
+ // room for it. The loop prints it on the next pass.
+ if (dropped > 0 && !is_full()) {
+ auto & entry = queue[tail];
+ snprintf(entry.msg.data(), entry.msg.size(), "%zu log lines dropped, the log writer fell behind\n", dropped);
+ entry.is_end = false;
+ entry.level = GGML_LOG_LEVEL_WARN;
+ entry.prefix = prefix;
+ entry.jsonl = common_log_jsonl;
+ entry.is_json = false;
+ entry.timestamp = timestamps ? t_us() - t_start : 0;
+ tail = (tail + 1) % queue.size();
+ dropped = 0;
+ }
+#endif
+
if (stop) {
break;
}