The Metal accorr kernel creates a new MTLCommandQueue on every call, which leaks GPU resources over long surrogate/permutation loops until newCommandQueue() returns None and the run crashes (AttributeError: 'NoneType' object has no attribute 'commandBuffer').
#279 (21be8d5) fixed this on master (cached queue + autorelease pool + buffer release), but the current PyPI release 0.6.0 predates it — so GPU users have to pin the git commit. Could a patch release including #279 be cut?
The Metal
accorrkernel creates a newMTLCommandQueueon every call, which leaks GPU resources over long surrogate/permutation loops untilnewCommandQueue()returnsNoneand the run crashes (AttributeError: 'NoneType' object has no attribute 'commandBuffer').#279 (21be8d5) fixed this on
master(cached queue + autorelease pool + buffer release), but the current PyPI release 0.6.0 predates it — so GPU users have to pin the git commit. Could a patch release including #279 be cut?