You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Address known thread-safety issues in cuda.core/look into call_once pattern #1389
I did an audit before for this type of issues but will have another sweep soon. Similar code still exists around _decide_nvjitlink_or_driver (and _lazy_init in _linker.pyx), but only a minor issue.
I.e. this doesn't use cdef ... _inited so is no additional free-threading issue. The function may get called twice but looks idempotent except for issuing a warning which is a nuisance.
That said: I'll look into establishing a call-once pattern which wouldn't just fix the nuisance but hopefully also make this clean (or at least obviously so).
changed the title [-]Address known thread-safety issues in `cuda.core`[/-][+]Address known thread-safety issues in `cuda.core`/look into `call_once` pattern[/+]on Aug 28, 2026
Activity
leofang commentedon Dec 16, 2025
Example:
cuda-python/cuda_core/cuda/core/experimental/_launcher.pyx
Lines 20 to 35 in 027ba10
leofang commentedon Dec 17, 2025
Sebastian's comment here is relevant: #1390 (comment)
seberg commentedon Aug 28, 2026
I did an audit before for this type of issues but will have another sweep soon. Similar code still exists around
_decide_nvjitlink_or_driver(and_lazy_initin_linker.pyx), but only a minor issue.I.e. this doesn't use
cdef ... _initedso is no additional free-threading issue. The function may get called twice but looks idempotent except for issuing a warning which is a nuisance.That said: I'll look into establishing a
call-oncepattern which wouldn't just fix the nuisance but hopefully also make this clean (or at least obviously so).[-]Address known thread-safety issues in `cuda.core`[/-][+]Address known thread-safety issues in `cuda.core`/look into `call_once` pattern[/+]