jmamou

"""
Demonstrates the actual output damage caused by the STANDALONE vocab-mismatch
bug during inference — without requiring model weights or a GPU.

The STANDALONE speculative decoding loop assembles output token-by-token:
  - Accepted draft tokens  → decoded by TARGET tokenizer  (CORRUPTED if vocab differs)
  - Rejected positions      → resampled by target          (correct)

This script simulates that assembly:
  1. "Target output" is produced by encoding the ground-truth continuation