# Local Setup Guide (Laptop)

This model is part of the DevaFlow project (custom D3PM, not native `transformers.AutoModel` format).

## 1) Environment

```bash
python3.11 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt
```

## 2) Quick Inference

```python
from inference_api import predict
print(predict("dharmo rakṣati rakṣitaḥ"))
```

## 3) Transformer-Style Use

```python
import torch
from config import CONFIG
from inference import load_model, _build_tokenizers

cfg = CONFIG
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, cfg = load_model("best_model.pt", cfg, device)
src_tok, tgt_tok = _build_tokenizers(cfg)

text = "yadā mano nivarteta viṣayebhyaḥ svabhāvataḥ"
input_ids = torch.tensor([src_tok.encode(text)], dtype=torch.long, device=device)
out = model.generate(
    input_ids,
    num_steps=cfg["inference"]["num_steps"],
    temperature=cfg["inference"]["temperature"],
    top_k=cfg["inference"]["top_k"],
    repetition_penalty=cfg["inference"]["repetition_penalty"],
    diversity_penalty=cfg["inference"]["diversity_penalty"],
)
ids = [x for x in out[0].tolist() if x > 4]
print(tgt_tok.decode(ids).strip())
```

## 4) Full Project Execution

For training, UI, Tasks 1–5, ablation workflow, and HF deployment, use the full project repository and run:

- `python train.py`
- `python inference.py`
- `python app.py`
- `python analysis/run_analysis.py --task <1|2|3|4|5|all>`

Task 4 note:
- `--phase generate_configs` first
- train ablation checkpoints
- then `--phase analyze`