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
It appears to be a race between the bf_getbuffer handler on JAX arrays and the allocator. The JAX code is just zeroing the buffer view object right at the start of exporting it, so I doubt it's a JAX bug.
My guess: perhaps there's a race in memoryview or NumPy, where the consumer of the exported buffer doesn't have adequate synchronization?
Description
In this TSAN run https://github.com/jax-ml/jax/actions/runs/13134558166/job/36646914845?pr=26300 we see the following race, which I've seen a few times.
It appears to be a race between the bf_getbuffer handler on JAX arrays and the allocator. The JAX code is just zeroing the buffer view object right at the start of exporting it, so I doubt it's a JAX bug.
My guess: perhaps there's a race in memoryview or NumPy, where the consumer of the exported buffer doesn't have adequate synchronization?
python/cpython#127716 implies that memoryview isn't thread safe yet.
System info (python version, jaxlib version, accelerator, etc.)
Python 3.13t
The text was updated successfully, but these errors were encountered: