Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
test(cuda): wire test_fp8_cuda into cuda-test and add independent footprint oracle
monotophic9480d88cf3d1
diff --git a/c/Makefile b/c/Makefileindex d646cc9c..39856bc8 100644--- a/c/Makefile+++ b/c/Makefile@@ -839,7 +839,7 @@ rans: $(RANSLIB)$(RANSLIB): tools/rans_ctypes.c rans.h$(CC) $(CFLAGS) -fPIC -shared $< -o $@ $(LDFLAGS)-cuda-test: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/test_backend_cuda.cu tests/test_ragged_attention.cu tests/test_absorb_determinism.cu tests/test_mxfp4_cuda.cu tests/mxfp4_ref.c tests/test_fp8_warp_cuda.cu+cuda-test: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/test_backend_cuda.cu tests/test_ragged_attention.cu tests/test_absorb_determinism.cu tests/test_mxfp4_cuda.cu tests/mxfp4_ref.c tests/test_fp8_warp_cuda.cu tests/test_fp8_cuda.cu@command -v "$(GPUCC)" >/dev/null 2>&1 || { echo "$(GPUCC_NAME) not found" >&2; exit 1; }"$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_backend_cuda.cu -o backend_cuda_test$(EXE) $(ANS_NVCC_LIBS)./backend_cuda_test$(EXE)@@ -864,6 +864,11 @@ cuda-test: backend_cuda.cu backend_cuda.h backend_gpu_compat.h tests/test_backen# one is expected to fail."$(GPUCC)" $(GPUFLAGS) backend_cuda.cu tests/test_absorb_determinism.cu -o absorb_determinism_test$(EXE)./absorb_determinism_test$(EXE)+ # fmt=8 (fp8-e4m3) kernel oracle + public API (LUT gate, dense, expert+ # group, byte accounting). Includes backend_cuda.cu itself, so only the+ # test file is compiled. Also runs BEFORE the MXFP4 test (same reason).+ "$(GPUCC)" $(GPUFLAGS) tests/test_fp8_cuda.cu -o fp8_cuda_test$(EXE) $(ANS_NVCC_LIBS)+ ./fp8_cuda_test$(EXE)# MXFP4 (fmt=7) against the CPU decoder in quant.h. The reference is built# as C on purpose: quant.h uses _Thread_local, which nvcc's C++ front end# rejects, and re-implementing it for the test would defeat the comparison.diff --git a/c/tests/test_fp8_cuda.cu b/c/tests/test_fp8_cuda.cuindex 8547624f..6c98c51e 100644--- a/c/tests/test_fp8_cuda.cu+++ b/c/tests/test_fp8_cuda.cu@@ -283,6 +283,16 @@ int main(void){got,want,bt->weight_bytes,bt->scale_count);return 1;}+ /* Independent oracle: fmt=8 footprint from quant.h's documented+ * layout -- O*I raw e4m3 bytes + ceil(O/128)*ceil(I/128) f32 block+ * scales (hs_n above) -- no struct fields, so the accessor cannot+ * validate itself against the counts upload happened to store. */+ size_t indep=(size_t)I*D+(size_t)hs_n*sizeof(float);+ if(got!=indep||bt->scale_count!=(size_t)hs_n){+ printf("FAIL bytes-check fmt8 independent: got %zu want %zu (%d*%d + %d*4), scale_count %zu (want %d)\n",+ got,indep,I,D,hs_n,bt->scale_count,hs_n);+ return 1;+ }coli_cuda_tensor_free(bt);free(bw);free(bs);