[JenD86]

Why

Steering, residual-norm calibration and extraction hooked model.model.layers directly. A PEFT-wrapped model (e.g. one carrying an Afterlife LoRA adapter) doesn't have that path: peft_model.model is the *ForCausalLM, so its layers are at peft_model.model.model.layers, and the first steered turn (or create) raises AttributeError. Found while running Modulo and Afterlife together in habitat.

What

• New modulo/layers.py:
  • unwrap() goes through PEFT via get_base_model().
  • decoder_layers() tries model.layers, the multimodal language_model layouts, OPT, transformer.h and gpt_neox.layers, in that order. If none match, it falls back to the ModuleList as long as num_hidden_layers. If that fails too, it raises a clear error.
  • num_layers() reads the config, including multimodal text_config.
• agent.py, extraction.py and cli.py now use the helpers instead of model.model.layers / config.num_hidden_layers.
• Docs (README, AGENTS.md, docstrings) now say plain or PEFT-wrapped. The troubleshooting entry is rewritten.

Plain models are unaffected: model.layers is tried first.

Testing

• New tests/test_layers.py covers:
  • the same layers being found with and without PEFT
  • a hook on the PEFT-wrapped model's layers changing the logits
  • the other layouts, the fallback and the failure case
• Full suite: 167 passed, 1 skipped (CPU).
• On an RTX 5090: Qwen3-8B with an Afterlife-trained SFT adapter loaded through PEFT, steered with curiosity +4 during the same turn. It answered with the adapter's trained fact, with steering active.

🤖 Generated with Claude Code