Closed BenjaminRemez closed 1 month ago
@maleadt On my machine, #451 did not resolve the issue. Currently on master, the problem seems worse than before --- if this occurs once, after quitting and restarting Julia, using oneAPI
crashes Julia itself. This is tested on WSL, and the issue persists even if the virtual machine is shut down and restarted. The only remedy is restarting the physical Windows machine - so it appears this crash has an observable effect even outside WSL.
The following throws an
UndefVarError
forOutOfGPUMemoryError
. Note that after this, allocating any otheroneArray
throws again the sameUndefVarError
forOutOfGPUMemoryError
, and subsequently many errors are thrown on runningexit()
in the REPL. Tested on WSL2.This does not occur for larger sizes of the arrays, so I suppose this is about overflowing the GPU with too many submitted tasks before the first ones start to complete - this is fine, but I suppose a graceful failure that does not kill the GPU is the desired behavior. Likewise, without
@btime
the failure occurs only probabilistically.