Update model card: multi-deployment (torch/onnx) support

README.md

@@ -8,61 +8,52 @@ tags:
 - robotics
 - IB-Robot
 - ascend
+- torch
 - edge-deployment
 ---
 
 # Model Card for FullSubNet (IB-Robot)
 
-FullSubNet speech enhancement model with cumulative Laplace normalization, compiled as stateful Ascend 310P OM artifacts for streaming 4-channel microphone array processing. Packaged for the [IB-Robot](https://gitcode.com/openeuler/IB_Robot) framework.
+FullSubNet speech enhancement model with cumulative Laplace normalization for
+streaming 4-channel microphone array processing, packaged for the
+[IB-Robot](https://atomgit.com/openeuler/IB_Robot) framework with three
+deployments sharing one stream contract.
 
-## Repository Structure
-
-- `inference_manifest.json` — deployment routing (schema v3, stateful stream contract)
-- `assets/adapter.json` — algorithm contract (n_fft=512, hop=256, cumulative_laplace_norm)
-- `artifacts/ascend/fullsubnet/` — two stateful OM modules:
-  - `fullsubnet_cum_stateful_fb_b4_t2_fp16.om` — full-band encoder (batch=4, time_steps=2)
-  - `fullsubnet_cum_stateful_sb_b4_t2_fp16.om` — sub-band decoder (batch=1028, time_steps=2)
-
-## Deployment
+## Deployments
 
-| Target | Backend | Runtime | Hardware |
-|--------|---------|---------|----------|
-| `ascend_310p` | ascend | ACL (cann-8.1.RC1) | Ascend 310P1 |
+| deployment | backend | artifacts | notes |
+|---|---|---|---|
+| `ascend_310p` | Ascend ACL (Ascend310P1) | stateful FB/SB OM pair | board-side streaming (LSTM state via state_links) |
+| `torch_cuda` | PyTorch CUDA | — (ckpt in `assets/`) | host-side streaming executor |
+| `torch_cpu` | PyTorch CPU | — (ckpt in `assets/`) | CPU fallback (watch the 128 ms hop budget) |
 
-**Execution:** stateful 2-stage pipeline `fullsubnet_fb` -> `fullsubnet_sb` with LSTM hidden/cell state banks (double-buffered, runtime-exclusive).
+All deployments share `tensor_model/fullsubnet/enhance`
+(`observation.audio_4ch [-1,4] -> voice.audio_enhanced_4ch [-1,4]`) with a
+stateful stream execution contract (`state_bank_mode: runtime_exclusive`,
+`max_open_streams: 1`).
 
-**Pipeline:**
-1. 4-channel audio (512 samples) -> STFT (512-pt, hop 256) -> magnitude [4,2,257]
-2. Cumulative Laplace normalization
-3. FB OM -> features [4,2,257] + LSTM state
-4. Host-side sub-band feature construction -> [1028,2,32]
-5. SB OM -> compressed complex mask [1028,2,2]
-6. CIRM decompression -> iSTFT/OLA -> enhanced audio [512,4]
+## Weights provenance
 
-**Input:** `observation.audio_4ch` float32 [-1,4]
-**Output:** `voice.audio_enhanced_4ch` float32 [-1,4]
+- `assets/cum_fullsubnet_best_model_218epochs.tar` = Audio-WestlakeU/FullSubNet
+  official v0.2 release checkpoint (sha256
+  `d08d09107eb276b8dc3d2d9fff995f4354a51fa3347125f52f8b9aea7c339f81`)
+- The 310P stateful OM pair was converted from the same checkpoint
+- `assets/cum_fullsubnet_best_model_218epochs.manifest.json` pins the digest and
+  `norm_type: cumulative_laplace_norm` (do not mix with offline-norm checkpoints)
 
-## Source Model
+Note: the `.tar` file is the upstream PyTorch serialization container
+(legacy `torch.save` format), loaded directly by `torch.load` — no extraction
+step.
 
-This bundle's OM artifacts were compiled from the upstream FullSubNet checkpoint:
-
-- **Source checkpoint:** [cum_fullsubnet_best_model_218epochs.tar](https://github.com/Audio-WestlakeU/FullSubNet/releases/download/v0.2/cum_fullsubnet_best_model_218epochs.tar) (218 epochs, 5.6M params)
-- **Source repository:** [github.com/Audio-WestlakeU/FullSubNet](https://github.com/Audio-WestlakeU/FullSubNet)
-
-Download script: `scripts/download_speech_direction_models.sh`. The OM artifacts were compiled from the torch checkpoint via ATC for Ascend 310P1 (cann-8.1.RC1, fp16).
+## Repository Structure
 
-> **Note:** Must use `cumulative_laplace_norm` normalization. Not compatible with `offline_laplace_norm` configurations.
+- `inference_manifest.json` — deployment routing (schema v3, stateful stream contract)
+- `assets/adapter.json` — algorithm contract (STFT 512/256, T=2, look-ahead 2)
+- `assets/cum_fullsubnet_best_model_218epochs.tar` — Torch checkpoint
+- `artifacts/ascend/fullsubnet/*.om` — stateful FB/SB OM modules
 
-## Citation
+## Usage
 
-    @inproceedings{fullsubnet,
-      title     = {FullSubNet: A Full-Band and Sub-Band Fusion Model for Real-Time Speech Enhancement},
-      author    = {Chen, Jun and Du, Zilin and others},
-      booktitle = {ICASSP},
-      year      = {2022}
-    }
-    @software{ib_robot,
-      title   = {IB-Robot: Intelligence Boom Robot},
-      url     = {https://gitcode.com/openeuler/IB_Robot},
-      license = {Apache-2.0}
-    }
+Host: `voice_asr_service.speech_direction` with backend `stateful_torch_cuda` /
+`stateful_torch_cpu`; board: select `ascend_310p` via the unified inference
+runtime.