surgery_metadata.json

{
  "surgery": "VibeVoice-1.5B -> VibeVoice-4B",
  "original_llm": "Qwen2.5-1.5B (Qwen2)",
  "new_llm": "Qwen3.5-4B-Base text decoder (Qwen3_5TextModel)",
  "old_hidden_size": 1536,
  "new_hidden_size": 2560,
  "old_vocab_size": 151936,
  "new_vocab_size": 248320,
  "strategy": "Zero-Init Padding (ZIP)",
  "approach": "state_dict (no transformers 5.x needed for surgery)",
  "notes": [
    "Acoustic/Semantic tokenizers are unchanged.",
    "Diffusion head expanded from 1536 to 2560 with ZIP.",
    "Connectors expanded from 1536 to 2560 with ZIP.",
    "adaLN modulation layers re-zeroed to preserve identity behavior.",
    "lm_head tied to new Qwen3.5 embed_tokens.",
    "New capacity dims (1536:2560) contribute ZERO at init time.",
    "Fine-tuning is REQUIRED to make new capacity useful.",
    "Loading the saved model requires transformers 5.x + VibeVoice repo import."
  ]
}