Versions

NCCL4Py exposes top-level helpers to inspect the installed NCCL stack: nccl4py itself plus the native libraries libnccl.so and libnccl_ep.so.

import nccl
nccl.show_versions()      # human-readable block to stdout
v = nccl.get_version()    # programmatic snapshot

show_versions

nccl.show_versions() → None

Print a summary of the installed NCCL stack to stdout.

For each component, reports the release version, the CUDA toolkit it was built with, and (for native libraries) the path of the loaded .so.

nccl.get_version() → VersionInfo

Return a structured snapshot of NCCL stack versions.

Returns:: VersionInfo with nccl4py + libnccl + libnccl_ep versions, CUDA build variants, and loaded .so paths.

class nccl.VersionInfo(nccl4py: Version, nccl: LibraryInfo | None, nccl_ep: LibraryInfo | None)

Bases: object

Aggregate version snapshot of the NCCL stack.

nccl: LibraryInfo | None: Version/CUDA-variant/path of the libnccl.so nccl4py is using, or None when it cannot be loaded.

nccl_ep: LibraryInfo | None: Version/CUDA-variant/path of the libnccl_ep.so nccl4py is using, or None when it cannot be loaded.

class nccl.LibraryInfo(version: Version, cuda_variant: Version | None, path: Path | None)

Bases: object

Version, CUDA build variant, and loaded-path info for a shared library.

cuda_variant: Version | None: CUDA toolkit major.minor the library was built with (e.g. 12.9), or None if it could not be read from the library.

path: Path | None: Path of the loaded libnccl.so, or None if it cannot be determined.