Telco Troubleshooting Agentic Challenge
Hi everyone,
am trying to run the Track A starter code locally i can start the Agent Tool Server with the training split using:
$env:DATA_SPLIT="train"
py -3.11 server.py
and the tool server seems to work because the notebook successfully calls /tools however when the agent starts the first scenario, it fails at the LLM call with:
openai.APIConnectionError: Connection error
In my notebook, the runner is configured with:
model_url="http://localhost:8000/v1"
model_name="qwen/qwen3.5-35b-a3b"
AGENT_API_KEY="dummy"
my understanding is that server.py is only the tool/data server on port 7860, and the agent also needs a separate OpenAI-compatible Qwen3.5-35B-A3B model server running on port 8000 i dont have an OpenRouter/API key, and i dont see one mentioned in the competition instructions.
Could the organizers or other participants clarify the expected setup?
Specifically:
1. Are participants expected to run Qwen3.5-35B-A3B locally with vLLM/SGLang?
2. Is there a hosted model endpoint or API key provided by the competition?
3. If not, what is the recommended way to run the model for Phase 1 testing?
4. Should the dummy AGENT_API_KEY work only for a local OpenAI-compatible server?
Thanks!
I used the model api.
I added my own AGENT_API_KEY that you can get from API Keys | Settings | OpenRouter
and added this line to resolve a proxy issue I faced
# Clear proxy env vars that can intercept outgoing HTTPS and strip auth headers
for _proxy_key in ["http_proxy", "https_proxy", "HTTP_PROXY", "HTTPS_PROXY"]:
os.environ.pop(_proxy_key, None)
I m not sure if this is what you need. But I hope it can help you to start processing.
yeah that made it clear thank you
1) Are participants expected to run Qwen3.5-35B-A3B locally with vLLM/SGLang? Yes, given the configuration (model_url="http://localhost:8000/v1"), the competition expects you to serve the model yourself using an OpenAI-compatible server. The recommended tools are vLLM or SGLang because they efficiently handle large models like Qwen3.5-35B-A3B. 2) Is there a hosted model endpoint or API key provided by the competition? No. The competition does not provide a hosted endpoint or an API key. The AGENT_API_KEY="dummy" is just a placeholder to satisfy the OpenAI client library. You are expected to run the model locally or on your own cloud GPU instance. 3) If not, what is the recommended way to run the model for Phase 1 testing? Local deployment or using a commercial cloud service are both recommended. 4) Should the dummy AGENT_API_KEY work only for a local OpenAI-compatible server? Yes. The dummy key works because the local OpenAI-compatible server (vLLM/SGLang) does not validate the API key by default. It simply ignores the Authorization header. If you were connecting to a hosted service like OpenRouter or a commercial endpoint, you would need a real key. For local development, dummy is perfectly fine
Due to hardware limitations, I cannot run Qwen3.5-35B-A3B locally on my machine. Would it be acceptable to use a third-party OpenAI-compatible API provider (e.g., OpenRouter) or is the expectation strictly to self-host the model using vLLM/SGLang on a local machine or a user-managed cloud GPU instance?
For sure, you can use a third party service.
Just to be 100% sure @AntonioDeDomenico , Is using OpenRouter to serve the Qwen3.5-35B-A3B model allowed in all phases of the competition and also during final evaluation/code review?
I'm asking for clarification as there's a part of the competition rule page that states "You may only use tools available to everyone i.e. no paid services or free trials that require a credit card." under Reproducibility of submitted code.
Although, it could just be that it was copied as part of the general competitions rule template but doesn't exactly apply to this competition. Need to be sure to avoid being disqualified later.
Hi, no problem in using openrouter.