Accelerating multi-node Large Language Model training with per-layer selective quantization (e.g. FP32 -> FP16) of the transformer architecture.