strands-decider-2B-qwen3.5-v1-2610 ? fused ONNX (fp32 + q4f16)

Fused single-graph build of StrandsAgents/strands-decider-2B-qwen3.5-v1-2610 (revision main), converted with the onnx-community/strands-decider-2B-hobson-v19-ONNX conversion/ pipeline: rank-48 LoRA merged into Qwen3.5-2B-Base (b1485b2f), weights mapped into the OPT text graphs (cosine 0.993 to instruct weights), lm_head dropped, pointer head fused in fp32, empty caches as in-graph zeros.

input_ids [B, L], attention_mask [B, L], answer_pos [B], option_pos [B, K] -> logits [B, K]. Divide by decider.temperature_by_kind[type] in config.json and softmax.

Files

• onnx/model.onnx (+.onnx_data, 7.5 GB): fp32, the faithful build.
• onnx/model_q4f16.onnx (+.onnx_data, 1.1 GB): 4-bit weights, fp32 head.
• onnx/model_q8.onnx (+.onnx_data, 1.7 GB): 8-bit MatMul weights, 4-bit embeds, fp32 head.

Parity (ONNX Runtime CPU vs the upstream torch engine, fp32)

• fp32 fused: worst 0.0021, argmax flips 0 (11 torch-reference cases)
• q4f16 fused: worst 0.2074, argmax flips 1 (tabs/dev flips vs torch, 4-bit cost)
• q8 fused: worst 0.0504, argmax flips 0

Apache-2.0, as the original model and Qwen3.5-2B-Base.