gliner2.5-multi-v1-onnx

GLiNER2.5 multi-v1 (ONNX)

ONNX export of fastino/gliner2.5-multi-v1 — the boundary architecture (BoundaryExtractor), which predicts where entities start and end instead of scoring a fixed-width span grid.

Exported for SecuPi's unstructured AI service (SPAI), which runs it through ONNX Runtime from Java. Published so that service can fetch it without a PyTorch dependency; anyone else is welcome to use it under the same licence as the base model.

Files

There is no span_rep.onnx or count_embed.onnx: the boundary architecture has no span grid and no count embedding. classifier.onnx keeps the same contract as the GLiNER2 exports (hidden_state [n,768] -> logit [n,1]).

How to run it

The prompt is unchanged from GLiNER2:

( [P] entities ( [E] label … ) ) [SEP_TEXT] <lowercased text>

1. Run encoder.onnx over that.

2. query_states = hidden states at the [E] positions → [1, n_labels, 768]. token_states = hidden state at the first sub-token of each word → [1, n_words, 768] (token_pooling: "first" — the head sees words, not sub-tokens).

3. Run onnx/boundary.onnx:

   input	shape
   token_states	[1, n_words, 768]
   text_mask	[1, n_words] bool
   query_states	[1, n_labels, 768]
   query_mask	[1, n_labels] bool

   output	shape
   pair_logits	[1, n_labels, 192]
   candidate_indices	[1, n_labels, 192, 2]
   candidate_valid	[1, n_labels, 192] bool
   null_logits	[1, n_labels]

4. A candidate is an entity when candidate_valid and sigmoid(pair_logit) >= threshold. Its indices are word boundaries with an exclusive end: (1, 3) means words[1:3].

5. The pool intentionally contains overlapping spans of the same label. Keep the highest-scoring and drop anything overlapping it, per label, to match the reference implementation.

A runnable version of the above, which diffs its output against native PyTorch, ships with the export tooling as verify_gliner25_export.py.

Export notes

Two source-level obstacles are worked around at export time:

• torch.sort(..., stable=True) lowers to aten::sort's out-variant, which has no ONNX symbolic. Stability is preserved rather than dropped where it is load-bearing — the candidate dedup sorts integer keys because they contain duplicates — by folding the position into the key.
• torch.eye(n, dtype=torch.bool) becomes EyeLike, for which ONNX Runtime has no boolean kernel; an arange comparison produces the same mask.

Verified against native PyTorch: identical sets of valid candidates (0 missing, 0 extra) with max score difference 2.1e-05, across word counts from 9 to 63 and 2 to 6 labels; and identical end-to-end entity output on the tooling's sample documents.

Provenance

Exported from fastino/gliner2.5-multi-v1 at revision aaecfe45db1d828c963717054ccb868e8ad1f1d5.

Licence

Apache 2.0, inherited from the base model. This repository redistributes the same weights in a different serialisation; all credit for the model belongs to Fastino.