/unload_model requires undocumented `model_name` key; /unload_models (plural) is a no-op

[Hi5808]

Summary

Two separate bugs in the model-unload endpoints:

1. POST /unload_model requires a JSON body with key model_name. Every other write endpoint in the same API (/api/generate, /api/chat, /load_model) uses model as the key. Calling /unload_model with {"model": "..."} (the natural, consistent choice given the rest of the API) returns a generic 400 with no indication of the actual expected field name — you have to read server.py to discover it wants model_name.

2. POST /unload_models (plural — intended to unload everything) doesn't actually unload anything. It returns 200 OK with {"message": "Models successfully unloaded!"} regardless, but no models are actually unloaded.

Environment

• rkllama version: 0.0.75
• Board: Radxa ROCK 5B+ (RK3588), Armbian 26.8.1

Bug 1: wrong/undocumented key name

Steps to reproduce

# Natural guess based on the rest of the API (model / load_model use "model")
curl -s -X POST http://localhost:8080/unload_model -H "Content-Type: application/json" -d '{"model":"<loaded-model-name>"}'

Actual

{"error":"Please enter the name of the model to be unloaded."}

Expected

Either the endpoint should accept model for consistency with the rest of the API, or — at minimum — the error message should name the actually-required key (model_name) instead of a generic "please enter the name" with no field-name hint.

Root cause (from source)

def unload_model_route():
    data = request.get_json(force=True)
    model_name = data.get('model_name', None)
    if model_name is None:
        return jsonify({"error": "Please enter the name of the model to be unloaded."}), 400

Bug 2: /unload_models (plural) doesn't call stop_all — silent no-op

Steps to reproduce

curl -s -X POST http://localhost:8080/unload_models

Actual

Returns 200 OK, {"message": "Models successfully unloaded!"}, but any previously-loaded model remains loaded and resident in memory (verified via free -h and GET /models / server process RSS before and after — no memory was released).

Root cause (from source)

@app.route('/unload_models', methods=['POST'])
def unload_models_route():
    variables.worker_manager_rkllm.stop_all
    return jsonify({"message": "Models successfully unloaded!"}), 200

variables.worker_manager_rkllm.stop_all is a bare attribute/method reference — it's never actually called (missing ()). Should be variables.worker_manager_rkllm.stop_all().

Expected behavior

• /unload_model should accept model (matching every other endpoint) or clearly document/error on the actual required key.
• /unload_models should actually invoke stop_all() and unload every loaded model, matching its documented/implied behavior and its success response.

Additional context

Full writeup with the underlying benchmark data (this is called out as Finding 5): https://claude.ai/code/artifact/09b27c30-c837-4204-8941-1d107ee75c9d

[danielferr85]

For unload models Ollama uses /api/generate without prompt. You can check that with OpenWebUI + RKLlama and unload a model from there. It works.

The endpoints unload_model and unload_models maybe old or discontinued because were created at the beginning for early stages in rkllama (specially for client part), not ollama standard. Neverthrless, we will fix those endpoints