gliner2-multi-v1

GLiNER2 ONNX export

gliner2-multi-v1 ONNX

• model.onnx (FP32 export)
• model_fp16.onnx (FP16 weights converted from FP32)
• model_int8.onnx (dynamic INT8 quantization via onnxruntime)
• tokenizer files copied verbatim from the HF model
• a small config.json describing runtime constraints

The export script follows the same design:

• ONNX graph includes only encoder + span head tensor computation
• Schema logic, label mapping, and decoding stay outside the graph
• Inputs: input_ids, attention_mask (optionally token_type_ids)
• Output: span_logits
• Export with torch.onnx.export (opset 19) and dynamic batch/sequence axes
• Convert FP32 weights to FP16 with convert_float_to_float16
• Quantize with onnxruntime.quantization.quantize_dynamic(QInt8)

Usage

Enter the dev shell (adds Python + ONNX deps):

nix develop

Install the Python dependencies with Pipenv:

cd onnx
pipenv install
cd ..

Export (run from the onnx directory so Pipenv finds the Pipfile):

pipenv run python export.py \
  --model-id fastino/gliner2-multi-v1 \
  --output-dir gliner2-multi-v1

Validation is enabled by default and compares the exported ONNX output to the PyTorch output for a dummy batch. It also runs a small extraction-method check (entities, classification, JSON) using identical decoding logic. To skip validation or to load the quantized model, use:

pipenv run python export.py --no-validate
pipenv run python export.py --no-validate-extraction
pipenv run python export.py --validate-quantized
pipenv run python export.py --no-fp16

The output directory will include:

• model.onnx
• model_fp16.onnx
• model_int8.onnx
• tokenizer.json
• tokenizer_config.json
• special_tokens_map.json
• added_tokens.json
• spm.model
• config.json

Notes

• The export uses a fixed max_seq_len (default 512) and expects inputs padded or truncated to that length. This matches the published bundle's runtime config.
• The span_logits label axis is aligned to token positions in the input sequence. Use label marker token positions ([E], [C], [R], [L]) to map logits back to schema labels. Label mapping and decoding are intentionally handled outside the graph.