Clarify transformer-style usage for custom D3PM runtime

README.md

@@ -39,7 +39,7 @@ print(predict("dharmo rakṣati rakṣitaḥ")["output"])
 
 ## Transformer-Style Usage (Recommended)
 
-Use this model as a reusable generation object:
+Use this model like a transformer pipeline pattern: load once, call `generate(text)` many times.
 
 ```python
 import torch
@@ -68,9 +68,16 @@ def generate(text: str):
 print(generate("yadā mano nivarteta viṣayebhyaḥ svabhāvataḥ"))
 ```
 
+### Minimal 3-Step Pattern
+
+1. `load_model(...)` once at app startup  
+2. `encode -> model.generate(...) -> decode` for each request  
+3. Reuse loaded model/tokenizers for all requests
+
 ## About `transformers` Compatibility
 
 - This repo does not expose `config.json` + `model.safetensors` in `transformers` format.
+- This is not a PEFT/LoRA adapter repository.
 - If you want full `AutoModel`/`pipeline` compatibility, you must create a wrapper architecture and export weights into HF Transformers conventions.
 - For production today, use:
   - `inference_api.py` for Python apps