Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
test(cuda): group tier regression rules with related tests
ZacharyZcR09f1403a017b
diff --git a/c/Makefile b/c/Makefileindex e36d1cc1..108de162 100644--- a/c/Makefile+++ b/c/Makefile@@ -1855,6 +1855,10 @@ tests/test_qwen36_tier_shutdown$(EXE): tests/test_qwen36_tier_shutdown.c tests/qtests/test_qwen36_tier_invariants$(EXE): tests/test_qwen36_tier_invariants.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h compat.h$(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)+# Async collection errors must drain all devices and reject partial output.+tests/test_qwen36_tier_take_error$(EXE): tests/test_qwen36_tier_take_error.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h compat.h+ $(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)+# Same fake backend, with uploads that take time: qt_fill_wait must not return# until the last enqueued expert is RESIDENT, not merely dequeued -- the engine# frees the RAM int8 copies right after it (#1360 saw the gap one run in@@ -2180,7 +2184,3 @@ tests/test_kv_prefix$(EXE): tests/test_kv_prefix.c kv_prefix.htests/test_kimi_request_state$(EXE): tests/test_kimi_request_state.c kimi_k3.c kv_prefix.h serve_codec.h st.h json.h tok.h tok_unicode.h tok_unicode_o200k.h compat.h quant.h idot.h omp_tune.h route_trace.h$(CC) $(NOCUDA_CFLAGS) tests/test_kimi_request_state.c $(VK_OBJ) -o tests/test_kimi_request_state$(EXE) $(NOCUDA_LDFLAGS)--# Async collection errors must drain all devices and reject partial output.-tests/test_qwen36_tier_take_error$(EXE): tests/test_qwen36_tier_take_error.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h compat.h- $(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)