Correction. Not "by per-tensor quantization", but "by training-time quantization".