# Rebuilding the ONNX files

Needs ~8 GB RAM plus swap (the fp32→fp16 pass peaks around 6 GB), Python 3.11,
and about 4 GB of disk. Runs on CPU; a Mac is fine.

```bash
pip install torch "gliner2[local]" onnx onnxruntime onnxscript onnxconverter-common transformers
python export_onnx.py dynamo          # -> onnx/model.onnx + model.onnx.data, config.json, tokenizer  (~1 min)
python quantize_q4.py                 # -> onnx/model_q4.onnx  (only the q4 step matters; ignore its fp16 error)
python convert_fp16.py model.onnx model_fp16.onnx
python convert_fp16.py model_q4.onnx model_q4f16.onnx MatMulNBits
python verify_parity.py               # fp32 vs the Python library on 3 prompts (expects ~1e-7)
python verify_variants.py             # fp16/q4/q4f16 vs fp32 on cases.json
```

Then rename the external data files to Transformers.js's convention and patch
the `location` field inside each graph:

```python
import onnx, os
from onnx.external_data_helper import _get_all_tensors
for name in ["model", "model_fp16", "model_q4", "model_q4f16"]:
    m = onnx.load(f"onnx/{name}.onnx", load_external_data=False)
    for t in _get_all_tensors(m):
        for kv in t.external_data:
            if kv.key == "location": kv.value = f"{name}.onnx_data"
    onnx.save(m, f"onnx/{name}.onnx"); os.rename(f"onnx/{name}.onnx.data", f"onnx/{name}.onnx_data")
```

`node_test.mjs` runs the whole open-jev path in Node against `reference.json`
(Python token ids and probabilities for `cases.json`).

Paths inside the scripts are absolute (`/home/claude/g2/...`); adjust `OUT` / `D` at the top of each.