プリセット

RK3576 Kev RKLLM

This preset deploys the tested Kev RKLLM path on RK3576 test platform. The engine export capacity is 8192 tokens, while the deployment guard and metadata limit requests to 2048 tokens because the 8K configuration caused global OOM on that device. The package records the runtime, engine, head, tokenizer, and metadata as separate pinned artifacts.

The 500-token reproducibility report for both RKLLM devices is archived at assets/evidence/context500-20261011/. It records 120 original and 120 reverse fixture cases per device; the fixture uses generated context and is not a general long-request accuracy claim.

Deploy the Kev Workbench

The container selects DECISION_BACKEND=rkllm and MODEL_ID=kev. Supply a complete model directory containing model.rkllm, kev_head.npz, tokenizer/tokenizer.json, and the pinned runtime at runtime/librkllmrt.so. The image contains the C helper only when built with the pinned RKLLM 1.3.1 headers.

リモート

Enter SSH credentials. Keep at least 3GB free for the RKLLM bootstrap; the measured fresh artifact peak is about 1.94GB with a 1GB reserve. The optional MODEL_SERVICES_JSON maps another model family to an already running HTTP service; the local Kev model remains the only model loaded by this container. VOICE_SERVICE_URL connects an existing compatible voice service and does not install ASR.

入力が必要な項目（SenseCraft Solution アプリで入力）

Device IP or hostname*
SSH username*
SSH password*
Remote model profiles JSON (optional)
Map another model id to an existing HTTP model service; leave blank to use local Kev. RK3576 deployment context is limited to 2048 tokens.
Existing local voice service URL (optional)

トラブルシューティング

Symptom	Fix
The model is not ready	Check that `model.rkllm`, `kev_head.npz`, `tokenizer/tokenizer.json`, and `runtime/librkllmrt.so` are present in the supplied model directory, then deploy again.
The health response has another backend or model ID	Check the deployment fields and confirm this preset uses `DECISION_BACKEND=rkllm` and `MODEL_ID=kev`.

デプロイ後

Open http://<device-ip>:18768. Verification requires /api/health to report model_ready: true, backend: "rkllm", and model_id: "kev".

Open the Kev Workbench

Submit a request with candidate commands after the health check reports a ready Kev model.

RK3576 second validation

The second RK3576 test platform validation used the same pinned image and verified five HTTP 200 golden cases with 5/5 expected selections. The 2069-word boundary request encoded to 2088 tokens and was rejected with HTTP 422 before NPU execution. This run used the image-built helper and had no external helper, RKNN runtime, or GOMP mount. Raw evidence is assets/evidence/rk3576-release-0.3.1-20261011/second-validation/.

トラブルシューティング

Symptom	Fix
The workbench page does not open	Confirm the health endpoint is ready at port 18768 before opening the page.
A request is rejected with HTTP 422	Check the request size against the preset limit; the documented boundary case is rejected before NPU execution.

プリセット

RK3588 Kev RKLLM

This preset deploys the tested 8K Kev artifact contract on RK3588 test platform. The verified deployment context is 8192 tokens; the result applies to the tested device and does not generalize to other boards with different free memory.

The same 500-token reproducibility report covers this preset at assets/evidence/context500-20261011/, with raw JSONL and machine-generated summaries under rk3588/.

Deploy the Kev Workbench

Provide the pinned engine, FP32 Q/K head, tokenizer, metadata, and RKLLM runtime directory for the target. The service does not download model weights at startup.

トラブルシューティング

Symptom	Fix
The model is not ready	Check the supplied engine, head, tokenizer, metadata, and RKLLM runtime paths, then inspect `/api/health` again.
A request is rejected before inference	Check the request against the verified 8192-token deployment context and the configured `MAX_OPTIONS`.

リモート

Optional remote model profiles and an existing voice service can be configured through the deployment fields.

入力が必要な項目（SenseCraft Solution アプリで入力）

Device IP or hostname*
SSH username*
SSH password*
Remote model profiles JSON (optional)
Existing local voice service URL (optional)

デプロイ後

Open http://<device-ip>:18768 and confirm the model-ready health response before sending a decision request.

Open the Kev Workbench

Submit a request with candidate commands after the health check reports a ready Kev model.

トラブルシューティング

Symptom	Fix
Bootstrap fails before the service is healthy	Recheck the target-filtered cache and the image archive checksum, then rerun the documented bootstrap.
Voice requests fail	This validation covers text decisions only; leave `VOICE_SERVICE_URL` empty or provide an existing compatible `/asr/stream` service.

For the optional existing voice integration, the browser recorder requires HTTPS or localhost and microphone permission. Plain remote HTTP pages cannot use browser recording; text decisions do not depend on voice. VOICE_SERVICE_URL connects an existing compatible /asr/stream service and does not install or validate ASR. The deployed field supplies the server default; the current UI also requires the voice drawer URL when configuring that optional path.

デプロイ後

RK3588 0.3.1 release verification

The RKLLM 0.3.1 ef82 image is published at Hugging Face revision db0545065d4da3598d009407fc3bd0c7632ca199, under the existing prefix releases/edge-workbench-0.3.0. On the RK3588 test platform stock deployment, bootstrap exited with code 0. Thirteen target-filtered model/shared files matched the SHA-verified cache and were reused; one new image archive was downloaded from https://hf-mirror.com. This is cache reuse plus one download, not a fresh fourteen-file download.

The image tag was edge-command-workbench:rockchip-rkllm-0.3.1-review. The archive SHA-256 is 251cb7fef080f489bf311b37c5bddbba1cd72e887e71f10ce12fc25a1fe64f77; the image config SHA-256 is d7a81c4f8ef3eaa2c21560b2d37939c6088357ec0c1d41c3089ff5fbc3836de9. The stock service was verified at http://<device-ip>:18768: health reported model_ready=true, backend=rkllm, model_id=kev, max_sequence_tokens=8192, and max_options=16. /api/models reported ready Kev (rkllm, 8192) and Laya (rknn, 90); the same short decision request returned HTTP 200 for both model_id=kev and model_id=laya. The Laya profile used the existing service at http://127.0.0.1:18771.

The helper was verified inside the image at /opt/kev/libkev_runtime.so, SHA-256 022e0102f9dd4c8839ba7acbc68f03d8811bf020cde84d6cf1343829cd66f919; no external helper mount was used. Raw selected request, response, timing, bootstrap, compose and container metadata are in assets/evidence/rk3588-release-0.3.1-20261011/. The prior a556/d870 fourteen-file bootstrap record remains historical and is not replaced. Voice and ASR were not part of this device validation.

プリセット

Historical RK3576 Laya/RKNN (0.1.0)

Deploy the local command workbench to one RK3576 Linux device over SSH. This preset is for RK3576 only; Jetson and RK3588 are not included in this package.

Deploy the Workbench

The deployment prepares the pinned application archive, RKNN model, and tokenizer, then starts the one-container web workbench on port 18768. The target must already have Docker and the RKNN runtime library.

リモート

The deployment uses /home/<username>/edge-command-workbench on the target. It skips Docker installation and Docker image pulling. The image archive is loaded only when the exact local image tag is absent. Before deployment, keep at least 4GB free, install Docker and /usr/lib/librknnrt.so, and expose the tested RK3576 NPU DRM nodes /dev/dri/card1 and /dev/dri/renderD129.

Enter the target IP or hostname, SSH username, and SSH password. An existing compatible local voice service URL is optional. Leave it blank when voice is not needed, then click Deploy.

配線

Power the RK3576 and connect it to the same network as the computer running SenseCraft Solution.
Connect the device's network interface and keep the physical connections stable during deployment.

入力が必要な項目（SenseCraft Solution アプリで入力）

Device IP or hostname*
SSH username*
SSH password*
Existing local voice service URL (optional)
Only set this when an existing compatible local voice service is already running. This solution does not install ASR.

トラブルシューティング

Symptom	Fix
Image archive or tokenizer checksum is missing	Complete the release manifest checksums after building the archive; the bootstrap script stops before starting Docker.
Docker image tag is missing after load	Confirm the archive was built for RK3576 arm64 and contains `edge-command-workbench:rk3576-0.1.0`.
RKNN library error	Check `/usr/lib/librknnrt.so` on the target and keep the read-only compose mount available.
Port 18768 is busy	Stop the service using the port, then retry after checking its ownership.
Dashboard does not open	Check the target IP and confirm `docker compose ps` reports the workbench container as healthy.

デプロイ後

Open the verification page at http://<device-ip>:18768. The page is served by the deployed application. A browser can use HTTP when opened from localhost or the local network; browser microphone access for an optional voice service requires HTTPS or localhost.

Optional voice service

Set VOICE_SERVICE_URL only when a compatible local voice service already exists on the target. This solution does not automatically install a voice service.

Open the Workbench

Open the deployed web workbench and submit a request with candidate commands. A healthy page and a JSON decision response verify that the container, model mount, tokenizer mount, and RKNN backend are reachable.

トラブルシューティング

Symptom	Fix
Health check is not ready	Wait for the model to load, then inspect `docker logs edge-command-workbench`.
Voice button is unavailable	Leave voice disabled or provide an existing compatible `VOICE_SERVICE_URL`; this preset does not install ASR.

デプロイ後

HTTP check

Open http://<device-ip>:18768/api/health and confirm the service reports healthy. The main decision endpoint is POST /api/decide on port 18768.

Edge Command Workbench — デプロイガイド

RK3576 Kev RKLLM

Deploy the Kev Workbench

Open the Kev Workbench

RK3576 second validation

RK3588 Kev RKLLM

Deploy the Kev Workbench

Open the Kev Workbench

RK3588 0.3.1 release verification

Historical RK3576 Laya/RKNN (0.1.0)

Deploy the Workbench

Optional voice service

Open the Workbench

HTTP check