[FEA]: Add support for pathfinder.find_bitcode_lib("nccl_device")

[xiakun-lu]

Closes #1831

Description

Adds support for cuda.pathfinder.find_bitcode_lib("nccl")

The new bitcode descriptor resolves NCCL device bitcode as libnccl_device.bc under nvidia/nccl/lib, matching the NCCL wheel package layout. The API name is nccl, consistent with the existing pathfinder naming used for NCCL dynamic library discovery.

The bitcode search-order test now covers all entries in SUPPORTED_BITCODE_LIBS, including nccl, while accounting for descriptors that define multiple site-packages candidate directories.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[rparolin]

lgtm, but I defer to @rwgk for the final approval.

[leofang]

Q: is this for #1831?

[rwgk]

Thanks! My main concern is nccl vs nccl_device, it'd be inconsistent as-is.

The other suggestion is minor, but also very easy.

As a sanity check I looked at the CI test logs. It looks good, e.g.:

INFO test_locate_bitcode_lib[nccl]: lib_path='/opt/hostedtoolcache/Python/3.10.20/x64/lib/python3.10/site-packages/nvidia/nccl/lib/libnccl_device.bc'

[rwgk]

Could you please make this nccl_device to match the existing nvshmem_device convention, and move the new entry before nvshmem_device to keep this list in lexicographic order?

[rwgk]

I still see nccl, could you please make it nccl_device for internal consistency?

[xiakun-lu]

In list DESCRIPTOR_CATALOG of dynamic libs, the name of nvshmem is "nvshmem_host" and the name of nccl is "nccl", which reflect the .so filenames libnvshmem_host.so.3 and libnccl.so.2. nccl as a key in _SUPPORTED_BITCODE_LIBS_INFO doesn't match nvshmem, but it's consistent with nccl's name in dynamic lib list DESCRIPTOR_CATALOG. I'd like to keep the key simple for nccl, so I don't have to think about it should end with _device or _host when I call these pathfinder APIs.

[rwgk]

so I don't have to think

It's gonna be surprising/confusing for anyone else, especially because the filename is libnccl_device.bc. Unless there is more than a personal preference (?), nccl_device is clearly the more consistent name here.

[xiakun-lu]

matching filename libnccl_device.bc is a good point.

[xiakun-lu]

done

[rwgk]

It's safer and more readable to avoid the defaults here and below. I tried out the change below. Could you please apply?

@@ -36,14 +36,14 @@ def clear_find_bitcode_lib_cache():
     get_cuda_path_or_home.cache_clear()


-def _make_bitcode_lib_file(dir_path: Path, libname: str = "device") -> str:
+def _make_bitcode_lib_file(dir_path: Path, libname: str) -> str:
     dir_path.mkdir(parents=True, exist_ok=True)
     file_path = dir_path / _bitcode_lib_filename(libname)
     file_path.touch()
     return str(file_path)


-def _bitcode_lib_dir_under(anchor_dir: Path, libname: str = "device") -> Path:
+def _bitcode_lib_dir_under(anchor_dir: Path, libname: str) -> Path:
     return anchor_dir / _bitcode_lib_info(libname)["rel_path"]


@@ -138,7 +138,7 @@ def test_locate_bitcode_lib_search_order(monkeypatch, tmp_path, libname):
 @pytest.mark.usefixtures("clear_find_bitcode_lib_cache")
 def test_find_bitcode_lib_not_found_error_includes_cuda_home_directory_listing(monkeypatch, tmp_path):
     cuda_home = tmp_path / "cuda-home"
-    lib_dir = _bitcode_lib_dir_under(cuda_home)
+    lib_dir = _bitcode_lib_dir_under(cuda_home, "device")
     lib_dir.mkdir(parents=True, exist_ok=True)
     extra_file = lib_dir / "README.txt"
     extra_file.write_text("placeholder", encoding="utf-8")

[xiakun-lu]

applied.

[rwgk]

Thanks!

Will it be helpful if this is released soon? (It's easy, <30 minutes effort.)

[xiakun-lu]

Thanks!

Will it be helpful if this is released soon? (It's easy, <30 minutes effort.)

definitely helpful! would appreciate it.

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.