Upload README.md with huggingface_hub

README.md

@@ -1,31 +1,16 @@
----
-license: cc-by-4.0
-base_model: nvidia/canary-1b-flash
-tags: [asr, hexagon, qualcomm, npu, qhexrt, fastconformer, aed]
----
+# canary-1b-flash → Hexagon NPU (v81) — encoder + AED decoder + recipe
 
-# Canary-1b-flash → Qualcomm Hexagon NPU (v81)
+NeMo Canary-1B-flash AED ASR ([nvidia/canary-1b-flash](https://huggingface.co/nvidia/canary-1b-flash)) on the Qualcomm Hexagon NPU (**v81 / SM8850 / soc_model 87**).
 
-NVIDIA **canary-1b-flash** (FastConformer AED ASR) ported to the Qualcomm Hexagon NPU (SM8750 / **v81** /
-`soc_model 87`) and **device-validated on a Samsung S25**.
+## Device validation (Samsung S25, Hexagon v81)
+Both networks self-replicated from the NeMo weights + validated exact offline, compiled to v81 bins:
+- **`v81/canary_conformer.bin`** (32-layer FastConformer encoder, fp32, 1.65 GB) — on-device raw-I/O parity **cos 0.999029** vs offline.
+- **`v81/canary_decoder.bin`** (4-layer MBart-like AED transformer decoder, 146 MB) — teacher-forced argmax == gold.
+- **Full autoregressive device decode** (encoder mem → decoder loop, `canary2` EN-ASR prompt seed) → **" Hey Jarvis." — exact gold match.**
 
-## Device validation (S25, v81)
-- **Encoder** graph (`v81/canary_conformer.bin`, 32-layer FastConformer): fed real inputs → device memory
-  **cosine 0.999029** vs the offline reference.
-- **Decoder** graph (`v81/canary_decoder.bin`, 4-layer transformer AED): teacher-forced device argmax ==
-  gold tokens.
-- **Full autoregressive decode on-device** (encoder memory → decoder loop): →
-  `[2014,2045,1387,1174,2042,1170,2048]` = **" Hey Jarvis." — exact gold match.**
+## What's in this bundle
+- `v81/canary_conformer.bin`, `v81/canary_decoder.bin` — the two QNN context binaries.
+- `recipe/` — `conformer.py` (encoder), `decoder.py` (AED decoder), `oracle.py` (full pipeline + canary2 multitask prompt + greedy decode).
 
-## How it was built
-The HF `parakeet` port is subtly numerically wrong for Canary's conformer (gates only ~0.86). This bundle
-uses a **bit-exact self-replication of the NeMo model** (`recipe/conformer.py` — full 32L encoder cos 1.000003
-vs NeMo; `recipe/decoder.py` — 4L AED reproduces the gold token chain). Both are exported to 2D-I/O ONNX and
-compiled to v81 context binaries. See `recipe/oracle.py` for the end-to-end oracle + the 23-iteration debug log.
-
-## Pipeline
-mel(128) → FastConformer encoder (graph) → memory; canary2 prompt seed
-`[7,4,16,62,62,5,9,11,13]` → host embed (tok_emb + FixedPosEnc + LN) → 4L AED decoder (graph, self+cross-attn) →
-argmax loop → detok. Encoder/decoder run on the NPU; mel/subsample/embed/detok are host-side.
-
-Arch-pinned to v81 (`dsp_arch` + `soc_model 87` baked in). Part of the RunAnywhere Hexagon model set.
+## Pipeline (NPU vs host)
+host log-mel(128) → **conformer.bin [NPU]** → host `canary2` prompt seed → **canary_decoder.bin [NPU]** autoregressive AED loop (self+cross-attn KV cache) → detok. A one-command `qhx_asr` AED host-op (adapting `src/hostop/asr_whisper.cpp` to this KV convention) is the productization follow-up; `oracle.py` is the reference decode. Arch-pinned v81. License: nvidia-open-model-license.