Improve modern GPU compatibility and document RTX 5090 validation

[XiaoBinGan]

Summary

This PR improves modern GPU usability and documents a successful community validation run on an NVIDIA GeForce RTX 5090.

Changes included

• add automatic device selection in config/config.py

  • uses CUDA when available
  • falls back to CPU otherwise

• add lightweight runtime diagnostics in scripts/train_transformer.py

  • PyTorch version
  • CUDA version
  • configured device
  • GPU name
  • GPU capability
  • total VRAM
  • step time
  • throughput
  • elapsed time between eval intervals
  • peak VRAM allocated / reserved
  • checkpoint metadata for device / PyTorch / CUDA version

• improve checkpoint loading compatibility in scripts/generate_text.py

  • supports newer PyTorch versions where torch.load behavior changed

• update README.md

  • add community-validated RTX 5090 note
  • document a successful official README flow run for the 13M configuration
  • record tested environment and observed training behavior

Validation performed

Tested on:

• GPU: NVIDIA GeForce RTX 5090
• PyTorch: 2.11.0+cu128
• CUDA: 12.8

Validated successfully:

• official dataset download flow
• official preprocessing flow
• README-recommended 13M training flow
• checkpoint saving
• checkpoint loading
• text generation

Notes

• The RTX 5090 validation in this PR confirms the official 13M workflow.
• Larger configurations are still hardware/config dependent and are not claimed as fully benchmarked here.

[XiaoBinGan]

Tested locally on an RTX 5090 with PyTorch 2.11.0+cu128 and CUDA 12.8.

I followed the repo's official dataset download + preprocessing flow, switched to the README-recommended 13M configuration for validation, completed training successfully, saved a checkpoint, and verified checkpoint loading + text generation.