Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
tests: guard -x none on the HIP build
huppiflupp654ff83fd071
diff --git a/c/Makefile b/c/Makefileindex 2a755b48..3ca63b71 100644--- a/c/Makefile+++ b/c/Makefile@@ -984,7 +984,12 @@ cuda-test: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/test_backen# below no longer pulls the runtime, and CI could not catch it because CI# never runs cuda-test. -Xclang appears nowhere else in CFLAGS.$(CC) $(filter-out -Xclang -fopenmp,$(CFLAGS)) -fno-openmp -fPIC -c tests/mxfp4_ref.c -o tests/mxfp4_ref.o- "$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_mxfp4_cuda.cu -x none tests/mxfp4_ref.o -o mxfp4_cuda_test$(EXE) $(GPU_TEST_LIBS)+ # -x none only under HIP: hip-test is `$(MAKE) cuda-test HIP=1`, so this+ # line also runs under nvcc for CUDA users, and nvcc accepts only c, c+++ # and cu for -x -- it fails the invocation on anything else. CI never+ # executes cuda-test, so that breakage would surface only on the next+ # CUDA box to run the suite.+ "$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_mxfp4_cuda.cu $(if $(filter 1,$(HIP)),-x none) tests/mxfp4_ref.o -o mxfp4_cuda_test$(EXE) $(GPU_TEST_LIBS)./mxfp4_cuda_test$(EXE)# Allocator footprint (#687): what a cudaMalloc really takes off the card,# which is what the expert tier has to be charged rather than the logical