Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
build(cuda): NVCC_ALLOW_UNSUPPORTED, so a newer host compiler is reachable
Unknown-Findoutaf2e05263445
diff --git a/c/Makefile b/c/Makefileindex 80c151cf..00fbfa88 100644--- a/c/Makefile+++ b/c/Makefile@@ -296,6 +296,23 @@ ifneq ($(LINUX),)NVCCFLAGS += -Xcompiler=-include,$(CURDIR)/glibc_c23_math_compat.hendifendif+# CUDA 13.1 host_config.h accepts only Visual Studio 2019-2022 and refuses+# anything newer outright, so on a box with VS Build Tools 18 (MSVC 19.50)+# EVERY nvcc target fails at the first #include with C1189 -- cuda-test,+# cuda-dll, all of it. nvcc offers -allow-unsupported-compiler to override the+# check, but there was no way to get one flag in: NVCCFLAGS is the only handle+# and replacing it wholesale discards $(CUDA_GENCODE), the -ccbin set above and+# the -Xcompiler warning form that Windows specifically needs -- the same+# argument that made NVCC_STD a variable rather than a rewrite.+#+# Defaults to 0: nvcc's own wording is "may cause compilation failure or+# incorrect run time execution", so an unsupported host compiler is the+# builder's decision to make explicitly, not a default this Makefile makes for+# them. Any result produced with it set should say so.+NVCC_ALLOW_UNSUPPORTED ?= 0+ifeq ($(NVCC_ALLOW_UNSUPPORTED),1)+NVCCFLAGS += -allow-unsupported-compiler+endif# HIP=1 builds the SAME backend for AMD GPUs via ROCm: backend_cuda.cu is# compiled unchanged through backend_gpu_compat.h (one source, two vendors,# like compat.h does for Windows). HIP_ARCH=native targets the GPU in this