Closed PatriceVignola closed 2 years ago
Until we have a DML implementation that doesn't explode in memory usage, we have to emulate this operator on the CPU. We could also simply decide to not implement it, but it has been seen in many models in the wild and breaks colocation graphs.
Until we have a DML implementation that doesn't explode in memory usage, we have to emulate this operator on the CPU. We could also simply decide to not implement it, but it has been seen in many models in the wild and breaks colocation graphs.