Run autoresearch on your gaming PC (Windows + RTX GPUs fork)

[jsegov]

Hi everyone!

I loved the autoresearch concept but wanted to make it accessible for local Windows environments. I’ve put together a fork optimized specifically for native Windows and desktop RTX GPUs.

Key changes in this fork:

• Native Windows Support: No WSL or unofficial Triton-on-Windows stacks required.
• Consumer GPU Focus: Tailored for Ampere, Ada, and Blackwell cards with 10GB+ VRAM.
• Simplified Runtime: Uses PyTorch SDPA and eager execution instead of FA3/torch.compile for better stability on consumer hardware.
• Easy Setup: Fully integrated with uv and uses tinystories_gpt4_clean for practical local training.

Check it out here: jsegov/autoresearch-win-rtx

I’ve tested this on an RTX 3080 10GB, I'd love to hear how it performs on your 30/40/50-series cards!

[aniruddhaadak80]

This is a fantastic initiative. Lowering the barrier to entry so anyone with a gaming PC can spin up their own AI researcher is exactly what this project needs to go completely mainstream.

Getting all the dependencies (especially the Triton/FlashAttention kernels) to compile nicely natively on Windows without C++ compiler headaches is notoriously tricky. Your fork is going to save Windows users hours of debugging frustration. Great work sharing this with the community.