Package Details: nccl-cuda12.9 2.30.7-2

Git Clone URL: https://aur.archlinux.org/nccl-cuda12.9.git (read-only, click to copy)
Package Base: nccl-cuda12.9
Description: Library for NVIDIA multi-GPU and multi-node collective communication primitives (Maxwell/Pascal/Volta support)
Upstream URL: https://developer.nvidia.com/nccl/
Licenses: BSD-3-Clause
Conflicts: nccl
Provides: nccl
Submitter: piernov
Maintainer: piernov
Last Packager: piernov
Votes: 0
Popularity: 0.000000
First Submitted: 2025-10-18 16:36 (UTC)
Last Updated: 2026-07-11 10:45 (UTC)

Required by (48)

Sources (2)

Latest Comments

gdevenyi commented on 2026-07-11 02:30 (UTC) (edited on 2026-07-11 02:30 (UTC) by gdevenyi) 

==> ERROR: One or more files did not pass the validity check!
error: failed to download sources for 'nccl-cuda12.9-2.30.7-1':
error: packages failed to build: nccl-cuda12.9-2.30.7-1

piernov commented on 2026-07-05 11:11 (UTC)

@yznnyz since the specific compiler version is a requirement of CUDA itself, the NVCC_CCBIN environement variable is defined in /etc/profile.d/cuda.sh, provided by the cuda package. Therefore, it should be part of your environment and does not need to be defined in this PKGBUILD.

yznnyz commented on 2026-07-05 02:18 (UTC)

The build failed because NVCC_CCBIN was not defined, so this line effectively passed an empty CXX to make:

export CXXFLAGS+=" -ffat-lto-objects"
make CXX="$NVCC_CCBIN" CUDA_HOME=/opt/cuda PREFIX=/usr src.build

Changing it to explicitly use GCC 14 fixed the problem for me:

export CC=/usr/bin/gcc-14
export CXX=/usr/bin/g++-14
export NVCC_CCBIN=/usr/bin/g++-14

export CXXFLAGS+=" -ffat-lto-objects"

make \
  CC=/usr/bin/gcc-14 \
  CXX=/usr/bin/g++-14 \
  NVCC_CCBIN=/usr/bin/g++-14 \
  CUDA_HOME=/opt/cuda \
  PREFIX=/usr \
  src.build

After this change, nccl-cuda12.9 builds correctly.

fxyqmngelonzwoxp commented on 2026-05-14 05:55 (UTC)

When I try to use P2P via NCCL on Maxwell cards I get a failure. Running NCCL with trace logs I believe these lines are related.

NCCL INFO misc/utils.cc:297 memory stack hunk malloc(65536) init.cc:557 NCCL WARN Cuda failure 'operation not supported'

Line 557 in init.cc is this CUDACHECK(cudaMemPoolCreate(&comm->memPool, &props));

I tried to look into this and found this comment https://github.com/pytorch/pytorch/issues/157517#issuecomment-3716213481 Basically even if nvidia supports maxwell cards in cuda 12.9 they seem to have dropped support for them in nccl and cudnn earlier.

I can't find out what version NCCL dropped support for maxwell or if it even if they officially list supported gpus for each nccl version rather than just throw unsupported errors for functions on old hardware.

wyf9661 commented on 2026-05-11 00:47 (UTC)

build failed

make[1]: I.: No such file or directory
nvcc warning : Support for offline compilation for architectures prior to '<compute/sm/lto>_75' will be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
-gencode=arch=compute_52,code=sm_52: No such file or directory
/bin/sh: line 1: I.: command not found

cyberpunkrocker commented on 2025-12-21 15:39 (UTC)

Fails to build in a clean chroot:

Compiling       src/device/onerank.cu
nvcc warning : Support for offline compilation for architectures prior to '<compute/sm/lto>_75' will be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
nvcc warning : Support for offline compilation for architectures prior to '<compute/sm/lto>_75' will be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
nvlink error   : Undefined reference to 'ncclDevFuncTable' in '/build/nccl-cuda12.9/src/nccl/build/obj/device/common.cu.o' (target: sm_52)
make[2]: *** [Makefile:119: /build/nccl-cuda12.9/src/nccl/build/obj/device/device_glue.o] Error 255
make[2]: Leaving directory '/build/nccl-cuda12.9/src/nccl/src/device'
make[1]: *** [Makefile:74: /build/nccl-cuda12.9/src/nccl/build/obj/device/manifest] Error 2
make[1]: Leaving directory '/build/nccl-cuda12.9/src/nccl/src'
make: *** [Makefile:25: src.build] Error 2
==> ERROR: A failure occurred in build().
    Aborting...
==> ERROR: Build failed, check /home/madmax/chroot/madmax/build

Supplying cuda-12.9-12.9.1-1, nvidia-580xx-utils-580.119.02-2 and opencl-nvidia-580xx-580.119.02-2 separately to the chroot.