Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
ci(hip): assert backend_cuda.o is position-independent
recvikingfda0071a4aa9
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.ymlindex b2de6e91..e7a48740 100644--- a/.github/workflows/ci.yml+++ b/.github/workflows/ci.yml@@ -90,8 +90,23 @@ jobs:# link of the test binary below covers the single-source claim, and# the engine link is exercised on real ROCm hardware (GPU_BACKENDS.md).docker run --rm -v "$PWD":/src -w /src rocm/dev-ubuntu-24.04:6.2 bash -c '- apt-get update && apt-get install -y --no-install-recommends make gcc g++ rocwmma-dev &&- make -C c gpu-compile HIP=1 HIP_ARCH=gfx1100'+ apt-get update && apt-get install -y --no-install-recommends make gcc g++ rocwmma-dev binutils &&+ make -C c gpu-compile HIP=1 HIP_ARCH=gfx1100 &&+ # The engine link is skipped here (see above), so assert the property+ # that link needs instead: backend_cuda.o must be position-independent.+ # hipcc emits a non-PIC/PIE object unless told otherwise, while gcc+ # links -pie by default on the distros users build on, so absolute+ # 32-bit relocations here mean `make HIP=1 colibri` dies with+ # R_X86_64_32 ... can not be used when making a PIE object+ # That is exactly what happened before -fPIE was added to HIPCCFLAGS:+ # this job stayed green over a HIP build that could not link at all,+ # because gpu-compile links the TEST binary with the hipcc driver+ # (PIC-agnostic) and never runs the gcc engine link. One readelf keeps+ # that regression from silently coming back.+ if readelf -r c/backend_cuda.o | grep -qE "R_X86_64_32S?"; then+ echo "backend_cuda.o has absolute 32-bit relocations: the PIE link of colibri will fail (HIPCCFLAGS needs -fPIE or -fPIC)" >&2+ exit 1+ fi'echo "HIP syntax check passed"windows-cuda-build: