MPS (Apple SIlicon) Support + fix: ensure HighlightedText receives a valid payload (avoid IndexError)

[ZimengXiong]

MPS (Apple Silicon) support and fixed bugs re multinomial sampling across demos and the Gradio app

• Shared device/dtype selection

  • Added select_device() + backend→dtype map:
    • CUDA → torch.bfloat16
    • MPS → torch.float16
    • CPU → torch.float32

• Guarded multinomial sampling

  • Introduced a wrapped torch.multinomial
  • Logit sanitization:
    • Replace NaN/Inf with 0
    • Clamp negatives to 0
    • Renormalize along the last dimension to sum to 1

Accepts tensors of arbitrary rank, flattens to 2D for sampling to satisfy PyTorch’s 1–2D constraint, Reshapes sampled indices back to the original shape

Fixes a runtime crash in the Gradio UI caused by HighlightedText expecting list-of-tuples but getting a string

• Add highlight_message() helper to wrap error/status strings as a HighlightedText-compatible list of (text, color) tuples.
• Replace bare string yields in dream_generate_with_visualization error paths with highlight_message(error_message) so gr.HighlightedText no longer throws IndexError.
• Make UI reset functions and callbacks use an empty list for vis_output_display (instead of "") so the HighlightedText component always receives a consistent type.
• Return [] for skip response in bot_response_generator to maintain consistent component behavior.

[jiacheng-ye]

Thanks, Zimeng, nice work!