Skip to content

[Windows] RTX 5070 Ti (Blackwell) - cuInit error 100 fix guide #13695

Description

@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
  • PyTorch: 2.11.0+cu130

Problem

After installing ComfyUI on an RTX 5070 Ti laptop, CUDA was not available:

  • torch.cuda.is_available() returned False
  • cuInit() returned error code 100
  • nvidia-smi worked fine, but no CUDA program could access the GPU

Root Causes and Fixes

  1. Windows GPU Preference: Hybrid graphics laptop routes Python to Intel iGPU by default. Fix: registry key GpuPreference=2 for python.exe
  2. PATH DLL conflict: CUDA Toolkit cudart64_12.dll conflicts with PyTorch cu130 cudart64_13.dll. Fix: prioritize torch/lib in PATH
  3. System-level CUDA_VISIBLE_DEVICES=-1: Must force-override to 0
  4. Blackwell architecture: Needs TORCH_CUDA_ARCH_LIST=12.0

Full fix details and code available. RTX 5070 Ti/5080/5090 laptop users will hit this same issue.

Happy to contribute a PR with updated troubleshooting docs if helpful.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions