Problem

When running train.py, there's no indication of which GPU is being used or how much VRAM is available. This makes it harder to:

• Verify the script is using the expected GPU (on multi-GPU systems)
• Debug OOM issues
• Compare results across different hardware

Solution

Print GPU name and VRAM at startup:

gpu_name = torch.cuda.get_device_name(0)
gpu_vram_gb = torch.cuda.get_device_properties(0).total_memory / 1024**3
print(f"GPU: {gpu_name} ({gpu_vram_gb:.1f} GB)")

Example Output

GPU: NVIDIA H100 80GB HBM3 (79.6 GB)
Vocab size: 8,192
Model config: {...}

Benefits

• Easier debugging and verification
• Useful context when sharing results or reporting issues
• Minimal code change, no new dependencies