runanywhere

/

nv_embedqa_1b_HNPU

like 0

Follow

RunAnywhere, Inc. 8

Adds the v79 (Hexagon v79 / SM8750, e.g. Galaxy S25) bundle alongside v81 (same 5-file layout).

• Bidirectional Llama-3.2 encoder (host-rope graph, 0 in-graph Gather) compiled to a 1.95 GB fp16 v79 context bin (nvembedqa_enc_f16, spill_bytes=0), + llama_embed host-op (embed + llama3 rope factor 32 + bidirectional mask + avg-pool + L2).

Export gate: host-rope embedding vs HF (fp32) cosine = 1.000000 (numerically identical to the reference).
Plane A (qhx_embed on-device, S25/v79): 2048-d, L2=1.00000, encoder ~43 ms. Retrieval — cos(France?, Paris)=0.610, cos(plants?, photosynthesis)=0.493, cos(France?, bread recipe)=0.104 (relevant >> irrelevant).

PR opened unmerged for Plane-B verification before merge.