RTX 5070 (sm_120) – AUTOMATIC1111 WebUI fails on GPU, CPU works fine. Looking for a working build / dev branch.

[ArienM02]

I’m having trouble getting AUTOMATIC1111 Stable Diffusion WebUI to run on my GPU.

CPU mode works fine and can generate images, but any attempt to use CUDA on my RTX 5070 fails, including model loading.

I’m trying to figure out whether I’m using the wrong WebUI branch, or there is a DEV / experimental build that actually works on RTX 50-series (sm_120), or this is a current limitation of Windows PyTorch builds.

What works

• WebUI launches normally
• CPU-only inference works and can generate images
• Models load correctly only in CPU mode

What does NOT work

• Any CUDA operation fails with:
  CUDA error: no kernel image is available for execution on the device
• Models fail to load on GPU (fails during torch tensor initialization)

This happens even with:

• PyTorch nightly (cu128)
• xformers disabled
• SDPA attention
• clean venv
• CUDA Toolkit installed

I’ve confirmed this is not just an A1111 issue — even a basic test like:
torch.randn(16, device="cuda")
fails with the same error.

About the “DEV mode” WebUI

I tried using a WebUI fork/branch that advertises DEV mode / RTX 50-series support, but:

• It requires logging into GitHub to download missing components
• The files it tries to fetch no longer exist or are unavailable
• So that build is unusable in its current state

If there is a working DEV branch, a fork that already supports sm_120, or a known-good commit / release

I’d really appreciate a link.

System Specs

GPU: NVIDIA RTX 5070 (12GB, sm_120)
CPU: AMD Ryzen 9 7900X (12-core)
OS: Windows 11
Driver: NVIDIA Studio Driver 591.74
Driver CUDA Capability: 13.1 (per nvidia-smi)
CUDA Toolkit: 12.8 installed
Python: 3.10.6
PyTorch tested: Stable cu128 and Nightly cu128 (torch 2.11 dev)

What I’m trying to find out

Is there a fully working WebUI version or fork that supports RTX 50-series on Windows?
Is this a known limitation of Windows PyTorch binaries for sm_120?
Are people with RTX 50-series currently using: a specific fork, WSL2/Linux, or DirectML instead?

If I’m missing any important system info, please let me know and I’ll add it.

Thanks in advance — I’m happy to test things and report back.