ltx-kernels build fails on rolling-release toolchains (CUDA toolkit mismatch + -Werror on glibc/Python POSIX macro clash)

[zmelumian972]

ltx-kernels fails to build on rolling-release toolchains (CachyOS/Arch)

Environment

• OS: CachyOS (Arch-based), GCC 16, glibc with _POSIX_C_SOURCE defaulting to 202405L
• Python: 3.12.13, built with pyconfig.h pinning _POSIX_C_SOURCE to 200809L
• uv sync --group kernels on a clean checkout

Issue 1: wrong CUDA toolkit picked up when CUDA_HOME/CUDA_PATH is preset

packages/ltx-kernels/setup.py only auto-selects the pip-installed nvidia-cuda-nvcc
toolchain when CUDA_HOME/CUDA_PATH are unset. Distros that export CUDA_PATH
globally (e.g. CachyOS points it at /opt/cuda) cause the build to silently use the
system CUDA toolkit instead, which can be a different minor version than the pinned
nvidia-cuda-nvcc==13.2.* / nvidia-cuda-cccl==13.2.* dependency group. Result:

/usr/include/cccl/cuda/std/__cccl/cuda_toolkit.h:41:8: error: #error
"CUDA compiler and CUDA toolkit headers are incompatible, please check your include paths"

Workaround: explicitly point CUDA_HOME/CUDA_PATH at the venv's bundled toolkit
before syncing:

CUDA_HOME=$(pwd)/.venv/lib/python3.12/site-packages/nvidia/cu13 \
CUDA_PATH=$(pwd)/.venv/lib/python3.12/site-packages/nvidia/cu13 \
uv sync --group kernels

Issue 2: -Werror fatal on a harmless glibc/Python POSIX macro mismatch

all2all.cpp includes ATen headers (which pull in glibc's <features.h>, defining
_POSIX_C_SOURCE 202405L) before pybind11's Python.h (which defines
_POSIX_C_SOURCE 200809L). On newer glibc these values differ, producing:

pyconfig.h:1877:9: error: '_POSIX_C_SOURCE' redefined [-Werror]
pyconfig.h:1898:9: error: '_XOPEN_SOURCE' redefined [-Werror]
cc1plus: all warnings being treated as errors

This diagnostic has no named -W group (no [-Werror=...] tag), so it can't be
selectively silenced under -Werror via -Wno-error=<name>. all2all_args in
setup.py hard-codes -Werror, making this fatal on any toolchain where glibc and
CPython disagree on the POSIX feature-test level.

Workaround applied locally: drop -Werror from all2all_args in
packages/ltx-kernels/setup.py (keeping -Wall -Wextra).

[ManuelFCastillo]

Thanks for the clear write-up, it made both of these quick to track down. I've opened #313 with fixes at the cause rather than the workarounds:

• setup.py: a bare CUDA_PATH (as Arch/CachyOS export globally) no longer vetoes the pip nvidia-cuda-nvcc toolkit. An explicit CUDA_HOME is still honoured.
• all2all.cpp: <torch/python.h> now leads the include list so <Python.h> precedes glibc's <features.h>. torch's wrapper push/undef/pops _POSIX_C_SOURCE / _XOPEN_SOURCE around <Python.h>; pybind11's wrapper doesn't, and it was landing first. -Werror stays.

I reproduced the exact diagnostics with a minimal include-order matrix on GCC 16.2 / glibc 2.43 (details in the PR), but I don't have a CachyOS host with an NVIDIA card. If you get a chance, a plain uv sync --group kernels on that branch, with no CUDA_HOME / CUDA_PATH overrides and -Werror still in place, would confirm it end to end:

git fetch https://github.com/ManuelFCastillo/LTX-2 fix/kernels-rolling-release-build && git checkout FETCH_HEAD