nn.Linear(d_model, NUM_AUDIO_TOKENS + 1)   # 1025 classes: 0..1023 audio, 1024 = EOS
eos_id = NUM_AUDIO_TOKENS                  # 1024