trending/colibri
1
0
Fork 0
mirror of https://github.com/JustVugg/colibri.git synced 2026-10-02 02:54:37 +08:00
Code Issues Packages Projects Releases Wiki Activity

docs: add CUDA-backend build recipe for glibc >= 2.41

Browse Source 
Building the CUDA backend on Debian trixie / Fedora 40+ etc. hits a
glibc-2.41 sinpi/cospi exception-spec clash with CUDA <= 12.9. Documents
the micromamba + CUDA 13.3 route that fixes it, and the dead ends (Debian
nvidia-cuda-toolkit conflicts with a .run/dkms driver; the pip nvcc wheel
ships only ptxas).
This commit is contained in:
Peter Groom
2026-07-26 10:08:15 -03:00
parent 81f08a09e5
commit ab6c370483
1 changed files with 33 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/BUILD-cuda-glibc241.md
+33
View File
@@ -0,0 +1,33 @@
# Building the CUDA backend on glibc ≥ 2.41 (Debian trixie and newer)
**Author:** Peter Groom · 2026-07-18
Symptom: `make glm CUDA=1` fails with four errors like
mathcalls.h(79): error: exception specification is incompatible with that of
previous function "cospi" (declared at ... crt/math_functions.h)
Cause: glibc 2.41 added C23 `sinpi/cospi/sinpif/cospif` with exception specs that clash
with CUDA ≤ 12.9's `math_functions.h` declarations. **CUDA 13.x headers fix it.**
Recipe that needs neither root-owned toolkits nor distro driver packages (safe next to a
`.run`-installed or dkms driver — important in LXC setups where userland must match the
host kernel module version):
```bash
# micromamba (single static binary) + NVIDIA's conda channel = complete nvcc
curl -Ls https://micro.mamba.pm/api/micromamba/linux-64/latest | tar -xj -C /opt bin/micromamba
/opt/bin/micromamba create -y -p /opt/cuda13 -c nvidia -c conda-forge cuda-nvcc=13 cuda-cudart-dev=13
ln -s /opt/cuda13/lib /opt/cuda13/lib64 # Makefile expects lib64; conda ships lib
make glm CUDA=1 CUDA_ARCH=sm_86 CUDA_HOME=/opt/cuda13 # your arch here
```
The binary links `libcudart.so.13` dynamically — ship it next to the binary and run with
`LD_LIBRARY_PATH`, or drop it in `/usr/local/lib`. Driver ≥ the runtime's minimum is the
only host requirement (any recent driver runs CUDA 13 binaries).
Dead ends, for the record: Debian's `nvidia-cuda-toolkit` hard-depends on Debian's driver
userland (version-clashes a `.run`/dkms driver — and its `nvidia-installer-cleanup`
postinst will try to delete your hand-installed userland); pip's `nvidia-cuda-nvcc-cu12`
wheel ships only `ptxas`, not the compiler frontend.