Use the same instructions as in 
  ../transformer_translation_python/HowToTrain.README


python transformer_translation_triton.py
python transformer_translation_triton.py --train train.en train.de


RTX 4090 optimizations:

    Block sizes tuned for SM89 architecture (BLOCK_M=64, BLOCK_N=64)
    Memory-efficient attention avoids materializing full attention matrix
    Fused operations reduce global memory traffic