After seeing Langchain's CUA workflow, I got inspired to try to automate applying to jobs by using this cua agent, scrapybara, and a local llm (so its much more affordable).
Took me a while to find a LLM that works with my older graphics card, but I eventually found Qwen/Qwen2.5-VL-3B-Instruct --dtype=half. I barely had enough vram to run this.
Following my own blended implementation of these 2 guides, local multi-agent systems and CUA agents, I set aside 5hrs to build a cua agent to help save time in the job application process.
The current tooling is quite rigid. Langgraph's cua_agent() only works with open_ai. I tried hard to fit my locally running qwen model into scrapybara's client.act(). Through all the model setup, I finally got it accepted!
INFO 04-07 14:26:28 [model_runner.py:1146] Model loading took 7.1557 GiB and 225.087367 seconds
I only have 7.5GiB available for use
ValueError: No available memory for the cache blocks. Try increasing `gpu_memory_utilization` when initializing the engine.
Hahahah! Time to try a different model :D
This has been quite a fun little challenge! Times up. Back to applying regularly.
pip install
vllm serve Qwen/Qwen2.5-VL-3B-Instruct --dtype=half
python src/main.py