2 ms·
Both still compute FP16 gradients. It comes back to the original point that Mac is just going to be too slow even on 8B model.
by lostmsu 2y ago
Both still compute FP16 gradients. It comes back to the original point that Mac is just going to be too slow even on 8B model.