Agent deployment fails because downloaded local model is reported as missing (beta.21)

[mandresve]

Description

This appears to be related to the model discovery issues previously reported.

After downloading a local RKLLM model, it is possible to select it while creating a new agent. The deployment wizard correctly lists the downloaded model, suggesting that the model registry is aware of it.

However, the final deployment step fails because the backend reports that the model cannot be found.

Steps to reproduce

1. Download a local model (e.g. qwen2.5-3b-rkllm-w8a8.rkllm).
2. Verify that it appears under Downloaded Models.
3. Open Agents.
4. Create a new agent.
5. Select the downloaded local model.
6. Complete the remaining configuration steps.
7. Click Deploy Agent.

Observed behavior

The deployment fails on the final step with the following error:

model 'qwen2.5-3b-rkllm' was not found on the controller, on any online worker, or among configured cloud providers. Download it first or pick a model that is already in the cluster.

This happens even though:

• The model is successfully downloaded.
• It appears in the Downloaded Models dialog.
• It was available for selection during the deployment wizard.

As a result, no agent is created.

Expected behavior

If a model is listed under Downloaded Models and can be selected during agent creation, the deployment should succeed without reporting that the model does not exist.

Additional information

This seems to indicate an inconsistency between the different components of the application:

• The Models dialog recognizes the downloaded model.
• The Deploy Agent wizard can enumerate and select the model.
• The final deployment backend cannot resolve the same model and behaves as if it were never installed.

This may be another synchronization or model registration issue between the Desktop UI and the backend services.

[jaylfc]

Fixed in beta.22. The real cause: an RKLLM model downloaded via the generic path landed on disk without being registered with rkllama, so the file looked present but the deploy-time resolver reported it missing. Downloading an rkllama-backed model now registers it through rkllama's own pull, so it is both selectable and deployable end to end.

[mandresve]

Update: this issue is still reproducible in beta.22.

The behavior remains unchanged.

Current status

• Downloaded models are correctly detected and appear in the model selector.
• The model can be selected during the Deploy Agent workflow.
• The deployment wizard completes until the final step.
• Creating the agent still fails, preventing deployment from completing successfully.

This indicates that the previous issue (models not being detected) has been resolved, but the final agent creation step is still failing with the following error:

model 'qwen2.5-3b-rkllm' was not found on the controller, on any online worker, or among configured cloud providers. Download it first or pick a model that is already in the cluster.

Expected behavior

After selecting a valid downloaded model and completing the deployment wizard, the new agent should be created successfully and appear in the Agents list.

Environment

• taOS Desktop: beta.22
• Platform: Orange Pi 5 Pro (RK3588)
• OS: Debian 12 (Bookworm) ARM64

[jaylfc]

Same root cause as #1599 — the model's backend being down, not the model being missing. v1.0.0-beta.24 makes the deploy failure actionable (it now names the backend you need to start rather than reporting the model as missing). Update, re-run Install NPU backend in Setup so rkllama is running, then redeploy — the model should resolve. Reply here if it still fails and the new message will tell us where.