This delivery tracks the official qox/mba-simplifier snapshot:
The model card describes a three-layer MBA simplification pipeline with an
optional neural-guided rule classifier. The live revision was checked on
2026-08-23: 4671ebd370b8c4dfaa10b05515ce053a8960c946.
The neural mba_classifier.safetensors is 8,903,168 bytes with SHA-256
cb8c272eca983366cda13b1681a64afb23f049fcafebf2f0fb26cd85c67fae4e.
Download the files listed in model-lock.json into /tmp/model-agent/mbas,
then run:
python inference.py --model-path /tmp/model-agent/mbas --device npu --output validation/runtime.jsonThe script reconstructs the official rule-classifier Transformer, strictly
loads the checkpoint, tokenizes one bounded MBA expression, and runs one
forward on npu:0. On the verified Ascend910, strict loading succeeded for
2,224,474 parameters, output [1,90] was finite, and the forward took about
0.202241 seconds. CANN reported that
aten::_transformer_encoder_layer_fwd fell back to CPU; this is recorded as a
runtime compatibility boundary. No full simplification accuracy gate is
claimed.


