runanywhere

/

nv_rerankqa_1b_HNPU

like 0

Follow

RunAnywhere, Inc. 16

Adds a v79/ bundle (Hexagon v79 / SM8750, Galaxy S25, soc_model 69) alongside the existing arch dirs. Compiled with QAIRT 2.48. Nothing under v75/ or v81/ is touched.

Device validation — SM8750 / v79

5/5 correct ranking over 10 labelled query/passage pairs, matching the HF fp32 oracle's 5/5; max |Δ| 0.013 and mean |Δ| 0.0059 on the raw logits.

Validated on a real Galaxy S25 via QHexRT Plane A (qhx_asr / qhx_embed / qhx_generate). Full record, build recipe and gotchas: forge/recipes/nvidia/*/results/device_v79.json.

Gate B (raw-I/O parity) has not been run for this model on this arch; the accuracy evidence above is end-to-end plus the offline export gates.