Distribute and run LLMs with a single file.
C++ cross-platformggufllama-cpplocal-ailocal-inferencelocal-llmopen-source-aisingle-file-executablespeech-to-text
diff --git a/tools/server/server-stream.cpp b/tools/server/server-stream.cpp
--- a/llama.cpp/tools/server/server-stream.cpp
+++ b/llama.cpp/tools/server/server-stream.cpp
@@ -7,6 +7,7 @@
#include <memory>
#include <utility>
#include <shared_mutex>
+#include <csignal>
enum class stream_read_status {
OK,
@@ -354,6 +355,25 @@ void stream_session_manager::stop_gc() {
}
void stream_session_manager::gc_loop() {
+ // Block every signal on this thread so pthread_cond_timedwait() can't be
+ // interrupted with EINTR (libcxx rethrows EINTR as std::system_error, and
+ // nothing here catches it, so the process std::terminate()s). Without this,
+ // Ctrl+C on `llama-server` aborts with "condition_variable timed_wait
+ // failed: Interrupted system call" instead of shutting down: SIGINT lands
+ // on this thread while it sits in the 60s GC wait. Keep synchronous fault
+ // signals deliverable so real crashes still surface. Same treatment as the
+ // log worker, the task queue and the model-stopping thread.
+ sigset_t block_mask;
+ sigfillset(&block_mask);
+ sigdelset(&block_mask, SIGSEGV);
+ sigdelset(&block_mask, SIGBUS);
+ sigdelset(&block_mask, SIGILL);
+ sigdelset(&block_mask, SIGFPE);
+ sigdelset(&block_mask, SIGABRT);
+ sigdelset(&block_mask, SIGSYS);
+ sigdelset(&block_mask, SIGTRAP);
+ pthread_sigmask(SIG_BLOCK, &block_mask, nullptr);
+
while (true) {
{
std::unique_lock<std::mutex> lock(gc_wake_mu);