Fix: use return_dict=True + **encoded for transformers >=4.47 compatibility

Force rebuild to pick up traceback diagnostic

Return full traceback in response for diagnosis

Try float32 to see actual failure mode vs bfloat16 AttributeError

Minimal ChatInterface, return not yield, no Blocks wrapper — isolate inference

show_error=True in Blocks, robust history parsing for Gradio 5

Add show_error=True and full traceback capture to expose hidden exception

Simplify: drop threading/streamer, use synchronous generate + torch.no_grad

Better exception capture to diagnose generation failure

Switch float16->bfloat16: float16 unsupported for CPU matmul

Switch to float16 (~6GB vs 12GB), capture thread exceptions

Fix: add streamer timeout, reduce default tokens to 150, honest CPU timing

Fix Gradio 5 API: remove retry/undo/clear_btn, add type=messages to Chatbot

Fix: pin gradio==5.29.1 (Python 3.13 compatible, no audioop)

Fix: upgrade to Gradio 5 for Python 3.13 compatibility

Deploy Feynman Explainer Gradio app (app.py)