Module build 

fdl nccl build – compile libnccl from NVIDIA’s NCCL source.

Why this exists: libtorch wheels statically link NCCL into libtorch_cuda.so. On heterogeneous-arch rigs (Pascal sm_61 + Blackwell sm_120, etc.), that static NCCL can corrupt the shared CUmodule loader state during cross-host bootstrap, which surfaces as cudaErrorNoKernelImageForDevice on otherwise-working kernels. Building libnccl as a standalone .so and LD_PRELOAD-ing it gives NCCL its own CUmodule scope and sidesteps the issue.

Output layout (mirrors fdl libtorch build):

libtorch/nccl/builds/v<version>-<archs>/
    include/nccl.h
    lib/libnccl.so          -> libnccl.so.2
    lib/libnccl.so.2        -> libnccl.so.<X.Y.Z>
    lib/libnccl.so.<X.Y.Z>
    lib/libnccl_static.a
    lib/pkgconfig/

Structs

BuildOpts

Functions

run