gliner2.5-multi-v1-onnx

ONNX export of fastino/gliner2.5-multi-v1 (GLiNER 2.5 BoundaryExtractor) for onnxruntime-web / WebGPU. Revision 4: same fused encoder as revision 3, plus [R] role-state gather and a separate relation scorer graph.

One encoder pass. JointIE beam search stays in JavaScript.

Host packing and decode: Pastel-Org/gliner2.5-onnx-webgpu (live: gliner25-onnx-webgpu.pages.dev).

Files

File	Role
onnx/model.onnx	Encoder + entity pair path + classifier + cached states
onnx/heads.onnx	SparseRelationScorer only (no encoder). Directional 2H + biaffine.

model.onnx inputs

Name	Shape	Dtype
input_ids	[B, T]	int64
attention_mask	[B, T]	int64
text_word_indices	[B, L]	int64
text_word_mask	[B, L]	float32
query_marker_indices	[B, Q]	int64
query_marker_mask	[B, Q]	float32
cls_marker_indices	[B, K]	int64
cls_marker_mask	[B, K]	float32
rel_marker_indices	[B, R]	int64
rel_marker_mask	[B, R]	float32

Unused heads: pass length-1 indices with mask 0.

model.onnx outputs

Name	Shape
start_logits / end_logits	[B, Q, L+1]
pair_indices / pair_logits / pair_valid	C=192
cls_logits	[B, K]
text_states	[B, L, H]
query_states	[B, Q, H]
rel_role_states	[B, R, H]

Host concatenates each [R] head + [R] tail pair into a 2H relation state.

heads.onnx

Inputs: text_states, rel_states [B, Rel, 2H], head_start/end, tail_start/end, rel_index, pair_mask (all pair tensors [B, P]). Output: rel_logits [B, P]. sigmoid(rel_logits) then JS beam (width 16). Schema constraints stay on the host.

Still not in ONNX

Neural RecordHead instance formation. Repeated JSON objects on the demo are host assignment (i-th mention of each field). Constrained implies/excludes for classification is JS when used.

Credits

• Base checkpoints: Fastino, Apache-2.0.
• ONNX export + JS host: Pastel-Cloud OÜ.