Is this for new documentation, or an update to existing docs?
Update
Describe the incorrect/future/missing documentation
The cuda.core 1.1.1 release note says the on-disk program cache directory is created with owner-only permissions (0o700), that those permissions are re-asserted on each use, and that this prevents other local users from reading or injecting cached device code regardless of the process umask. Only the tmp/ staging directory is created 0o700. The cache root, entries/ and the shard directories inherit the umask, and no permissions are re-asserted, so a cache created under a permissive umask stays world-writable after a write.
import os, stat, tempfile
from pathlib import Path
from cuda.core.utils import FileStreamProgramCache
os.umask(0o000)
root = Path(tempfile.mkdtemp()) / "pc"
FileStreamProgramCache(path=root)["k"] = b"hello"
for p in (root, root / "entries", root / "tmp"):
print(p.name, oct(stat.S_IMODE(os.stat(p).st_mode)))pc 0o777
entries 0o777
tmp 0o700
Expected: the note describes what shipped. Cache entry files are 0o600 and tmp/ is 0o700, so cached device code is not readable by other users, but the cache root and entries/ follow the umask and a pre-existing shared cache directory is used as-is.
cuda_core/cuda/core/utils/_program_cache/_file_stream.py also records the trade-off this leaves (a deliberately shared writable entries/ lets a co-owner replace a cached file). Whether the note should state that as well is your call.
If this is a correction, please provide a link to the incorrect documentation. If this is a new documentation request, please link to where you have looked.
https://nvidia.github.io/cuda-python/cuda-core/latest/release/1.1.1-notes.html
Is this for new documentation, or an update to existing docs?
Update
Describe the incorrect/future/missing documentation
The
cuda.core1.1.1 release note says the on-disk program cache directory is created with owner-only permissions (0o700), that those permissions are re-asserted on each use, and that this prevents other local users from reading or injecting cached device code regardless of the processumask. Only thetmp/staging directory is created0o700. The cache root,entries/and the shard directories inherit the umask, and no permissions are re-asserted, so a cache created under a permissive umask stays world-writable after a write.Expected: the note describes what shipped. Cache entry files are
0o600andtmp/is0o700, so cached device code is not readable by other users, but the cache root andentries/follow the umask and a pre-existing shared cache directory is used as-is.cuda_core/cuda/core/utils/_program_cache/_file_stream.pyalso records the trade-off this leaves (a deliberately shared writableentries/lets a co-owner replace a cached file). Whether the note should state that as well is your call.If this is a correction, please provide a link to the incorrect documentation. If this is a new documentation request, please link to where you have looked.
https://nvidia.github.io/cuda-python/cuda-core/latest/release/1.1.1-notes.html