Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(tests): the qwen36 fill-wait test needs compat.h to build on Windows
JustVugg0df0fa6b1d16
diff --git a/c/Makefile b/c/Makefileindex c0b97fd0..f6fde8c3 100644--- a/c/Makefile+++ b/c/Makefile@@ -1714,7 +1714,7 @@ tests/test_qwen36_tier_invariants$(EXE): tests/test_qwen36_tier_invariants.c tes# until the last enqueued expert is RESIDENT, not merely dequeued -- the engine# frees the RAM int8 copies right after it (#1360 saw the gap one run in# fifteen; here the slow upload hook makes it every run without the fix).-tests/test_qwen36_tier_fill_wait$(EXE): tests/test_qwen36_tier_fill_wait.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h+tests/test_qwen36_tier_fill_wait$(EXE): tests/test_qwen36_tier_fill_wait.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h compat.h$(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)# Same fake backend: the automatic placement (COLI_PLACE unset/"auto") putsdiff --git a/c/tests/test_qwen36_tier_fill_wait.c b/c/tests/test_qwen36_tier_fill_wait.cindex fcb9c631..3c2afdd5 100644--- a/c/tests/test_qwen36_tier_fill_wait.c+++ b/c/tests/test_qwen36_tier_fill_wait.c@@ -24,6 +24,9 @@#include "../qwen36_tier.c"+#include "../compat.h" /* setenv: MinGW has none, and qwen36_tier.c does not+ * pull it in the way an engine .c does */+static int fails;static void check(int ok, const char *what) {if (!ok) { printf(" FAIL: %s\n", what); fails++; }