Vocotype Qwen3-ASR Android Adapter INT4
This repository hosts the Android-usable Qwen3-ASR adapter encoder and the
resident Qualcomm QNN decoder artifacts used by Vocotype.
Files
encoder.adapter.int4.onnx: pre-fused Android ONNX encoder with the Vocotype
Qwen3-ASR speaker adapter.chunk0-rtu.bin .. chunk6-rtu.bin: resident W8A16+u8-attention decoder
step chunks (4 layers each, GQA KV expand folded into the Q-head M
dimension).chunk7-rtu.bin: resident step output head (final norm + full-vocabulary
LM head).prefill-eq-group3000.bin .. prefill-eq-group3003.bin: resident prefill
attention/FFN layer cores (equalized W8A16).prefill-eq-group2099.bin: resident prefill output head.decoder_layer*_input_norm.f32, decoder_layer*_post_attention_norm.f32:
CPU-side RMSNorm weights.decoder_ffn_dc.f32: per-layer FFN device DC-bias correction vectors
(measured on SM8850, subtracted before the FFN residual add).runtime/context-hotword-bias-20260705/: runtime contract for the Android
context hotword bias promoted on 2026-07-05.
Android execution boundary
The production IME path runs the pre-merged Voco adapter encoder followed by
the resident decoder session protocol: prefill over a fixed 128-token bucket,
then greedy decode steps over a 127-position KV ring cache. Decoder MatMul,
attention, FFN, and output-head work runs on the Qualcomm HTP NPU through
cached context binaries; RMS normalization, residual additions, and token
selection remain on CPU. Step decode is ~108 ms/token on SM8850 with the
current W8A16 layout (fixture-verified 26/26 tokens and live-audio
regression-tested on 2026-08-11).
