Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

Merge dev (keep $(NVCC_STD) and -ftz=false)

JustVuggdb0720c511c6
diff --cc c/Makefile
index 129c46fc,8b2d3344..26b97fa5
--- a/c/Makefile
+++ b/c/Makefile
@@@ -222,9 -227,13 +234,13 @@@ ifneq ($(IS_WIN),
# '/Wextra'"), which made `make cuda-dll` unbuildable as shipped. -W3 is
# the MSVC warning level (dash form, NOT /W3: MSYS make would mangle the
# slash-form into a filesystem path).
- NVCCFLAGS ?= -O3 -std=$(NVCC_STD) $(CUDA_GENCODE) -Xcompiler=-W3
-NVCCFLAGS ?= -O3 -std=c++17 -ftz=false $(CUDA_GENCODE) -Xcompiler=-W3
++NVCCFLAGS ?= -O3 -std=$(NVCC_STD) -ftz=false $(CUDA_GENCODE) -Xcompiler=-W3
else
- NVCCFLAGS ?= -O3 -std=$(NVCC_STD) $(CUDA_GENCODE) -Xcompiler=-Wall,-Wextra
+ # -ftz=false is nvcc's default, pinned explicitly: the fmt=8 kernels are
+ # cross-tier parity instruments and a flushed scale*subnormal contribution
+ # silently diverges from the CPU reference (backend_cuda.cu also #errors on
+ # fast-math builds, but -ftz alone defines no macro to guard on).
-NVCCFLAGS ?= -O3 -std=c++17 -ftz=false $(CUDA_GENCODE) -Xcompiler=-Wall,-Wextra
++NVCCFLAGS ?= -O3 -std=$(NVCC_STD) -ftz=false $(CUDA_GENCODE) -Xcompiler=-Wall,-Wextra
# glibc >= 2.41 declares the C23 math additions (rsqrt/rsqrtf, sinpi/cospi, ...)
# with __THROW -> noexcept(true), clashing with CUDA's crt/math_functions.h
# under C++17 (noexcept is part of the function type since C++17). Force-include