runanywhere

/

canary_1b_flash_HNPU

like 0

Follow

RunAnywhere, Inc. 8

Adds the Hexagon v75 Canary-1B-flash bundle under v75/ (source nvidia/canary-1b-flash, NeMo FastConformer-AED).

Rebuilt for v75 (not v81 bytes):

• canary_conformer.bin (1.6G) — 32-layer FastConformer encoder, self-replicated from the NeMo encoder.* weights (cos 1.0 vs NeMo), 2D I/O sub[91,1024]+pos[181,1024]->mem[91,1024], compiled f32 (fp16 LN-variance overflows) with the >2GB model-lib fallback (blob->.lrodata, -mcmodel=large), O3/soc57.
• canary_decoder.bin (146M) — 4-layer transformer AED decoder, tok_embed[32,1024]+mem->logits[32,5248], f32/soc57.

Reused (arch-independent host blobs, from the v79/v81 bundle): canary_aed_host.bin, canary_frontend.bin (NeMo log-mel + subsample + rel-pos frontend), canary_vocab.hex, the SentencePiece *_tokenizer.model / *_vocab.txt set, and canary-1b-flash.json (manifest, dsp_arch->v75). Host-op canary_aed_transcribe (canary2 seed 7,4,16,62,62,5,9,11,13, n_mels 128, vocab 5248).

Device-validated on the Galaxy S24 (SM8650, serial RFCX200WNWW): qhx_asr on the fixed-length T’=91 clip (hey-jarvis-real, 7.22s) -> TRANSCRIPT “ Hey, Jarvis.” (8 tokens, WER 0.0), encoder 807 ms + decoder 25 ms, total 887 ms, status=Ok. Fixed-shape graph (T’=91), single-length like the v79/v81 dirs.