Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge pull request #1390 from crichalchemist/qwen36-tier-fill-wait
Vincenzo Fornaro2dac79e09342
diff --cc c/Makefileindex f21c59ba,4072a61c..c0b97fd0--- a/c/Makefile+++ b/c/Makefile@@@ -1707,9 -1680,16 +1707,16 @@@ tests/test_qwen36_tier_shutdown$(EXE):# issue blocks stay inside, disjoint and at their device's slot under random# routing on 1-3 devices. Under test-asan the third doubles as a fuzz for the# #1339 class.-tests/test_qwen36_tier_invariants$(EXE): tests/test_qwen36_tier_invariants.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h+tests/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 $@ $(LDFLAGS)+ # Same fake backend, with uploads that take time: qt_fill_wait must not return+ # until the last enqueued expert is RESIDENT, not merely dequeued -- the engine+ # frees the RAM int8 copies right after it (#1360 saw the gap one run in+ # fifteen; here the slow upload hook makes it every run without the fix).+ tests/test_qwen36_tier_fill_wait$(EXE): tests/test_qwen36_tier_fill_wait.c tests/qwen36_fake_cuda.h qwen36_tier.c qwen36_tier.h backend_cuda.h+ $(CC) $(CFLAGS) -DCOLI_CUDA $< -o $@ $(LDFLAGS)+# Same fake backend: the automatic placement (COLI_PLACE unset/"auto") puts# the offered dense trunk into VRAM by bytes-saved-per-byte, prices displaced# experts by heat, keeps the trunk on the CPU when the marginal experts are