Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
test(qwen36): bound the shutdown test with a thread, not SIGALRM
JustVuggfa8142d7d2b8
diff --git a/c/tests/test_qwen36_tier_shutdown.c b/c/tests/test_qwen36_tier_shutdown.cindex fad38885..af8faf7b 100644--- a/c/tests/test_qwen36_tier_shutdown.c+++ b/c/tests/test_qwen36_tier_shutdown.c@@ -12,7 +12,7 @@* This test reproduces exactly that sequence with the fake CUDA backend* (tests/qwen36_fake_cuda.h, no GPU needed): budget one expert in, issue it* without taking it, queue an LFRU swap directly (what qt_lfru_tick_locked- * does), then call qt_shutdown() under an alarm(10) so a hang fails loudly+ * does), then call qt_shutdown() under a 10 s watchdog so a hang fails loudly* instead of wedging CI. */#include <stdio.h>#include <stdlib.h>@@ -34,10 +34,22 @@ static int issue_ok(int device, int count, const float *x) {(void)device; (void)count; (void)x; return 1;}-static void on_alarm(int sig) {- (void)sig;- (void)!write(2,"FAIL: qt_shutdown hung\n",23);- _exit(1);+/* A hang must fail, not sit forever in CI. alarm()/SIGALRM would do it on+ * POSIX, but MinGW has neither, and the Windows job is the one that proved+ * it. A detached thread that sleeps and then checks a flag does the same+ * job everywhere the test already builds (it uses pthread and nanosleep). */+static volatile int shutdown_done = 0;+static void *hang_watchdog(void *arg) {+ (void)arg;+ for (int i = 0; i < 100 && !shutdown_done; i++) { /* 100 x 100 ms = 10 s */+ struct timespec ts = {0, 100000000};+ nanosleep(&ts, NULL);+ }+ if (!shutdown_done) {+ (void)!write(2, "FAIL: qt_shutdown hung\n", 23);+ _exit(1);+ }+ return NULL;}/* Poll a G.mx-guarded predicate without a busy spin. */@@ -112,10 +124,11 @@ int main(void) {}check(parked, "the uploader should have dequeued the swap within 1s");- signal(SIGALRM, on_alarm);- alarm(10);+ pthread_t watchdog;+ if (pthread_create(&watchdog, NULL, hang_watchdog, NULL) == 0)+ pthread_detach(watchdog);qt_shutdown();- alarm(0);+ shutdown_done = 1;check(!G.on, "shutdown_returns_while_a_group_is_open");/* The abandoned swap must leave the victim exactly as the open group left* it, and must not have driven the incoming expert's upload after