[Windows] RTX 5070 Ti (Blackwell sm_120) - setup and deployment notes

[loongmiaow-pixel]

Environment

• GPU: NVIDIA GeForce RTX 5070 Ti Laptop GPU (Blackwell, compute capability 12.0)
• Driver: 595.79 (CUDA 13.2)
• OS: Windows 11
• Python: 3.14
• diffusers: [latest]

Problem

diffusers on RTX 5070 Ti requires workarounds:

• TORCH_CUDA_ARCH_LIST=12.0 required - Blackwell not auto-detected
• CUDA_VISIBLE_DEVICES=-1 system env var blocks GPU access
• Pipeline inference may fail if CUDA arch not explicitly set

Solution

$env:TORCH_CUDA_ARCH_LIST = "12.0"
$env:CUDA_VISIBLE_DEVICES = "0"
$env:CUDA_MODULE_LOADING = "LAZY"

For hybrid graphics laptops, set Windows GPU Preference:

Set-ItemProperty -Path "HKCU:\Software\Microsoft\DirectX\UserGpuPreferences" -Name "C:\path\to\python.exe" -Value "GpuPreference=2;"

Question

Is there a plan to add Blackwell (sm_120) to the default CUDA architecture detection? RTX 5070/5080/5090 are shipping now and Windows users will need this.