Run autoresearch experiments on an 8 GB laptop GPU under WSL2
At a glance
A developer ran autoresearch on an RTX 4070 Laptop GPU and shared the experiment history, resulting configuration, and a fix for compilation startup overhead. This workflow is listed under autonomous-experiments, model-training, consumer-gpu.
- Agents / tools
- autoresearch, NVIDIA RTX 4070 Laptop GPU, WSL2
Building a tool for these workflows? See weekly rates.
What they did
The developer used an ASUS ROG NUC with an RTX 4070 Laptop GPU, 8 GB of VRAM, and Windows 11 with WSL2. The autoresearch loop changes training code, runs a five-minute experiment, evaluates validation bits per byte, and retains or discards the candidate. The linked fork contains configuration tables and a progress history.
How it went
The author reports 282 experiments with 38 retained changes, reducing validation bits per byte from 2.770532 to 2.495532.
Worth knowing
Compilation startup could consume the experiment budget; the author reports excluding the first ten steps from timing. Several settings that helped on other GPUs performed worse on this hardware.
