mediumDetection & drivers

CUDA / ROCm version mismatch

The app fails with a CUDA/ROCm version or 'no kernel image' error.

Scope
Application/runtime compatibility, not hardware.
Check recent changes
  • ·New framework version
  • ·New image
  • ·Driver downgrade
Diagnostics

Driver's max CUDA

bash
nvidia-smi | grep 'CUDA Version'

Expect: ≥ the CUDA the framework was built against.

Framework build

bash
python -c 'import torch; print(torch.version.cuda, torch.cuda.get_arch_list())'

Expect: CUDA ≤ driver's; arch list includes your GPU's compute capability.

Likely causes & fixes
  • high likelihoodFramework built for newer CUDA than the driver supports

    Fix · Upgrade the driver or use a matching framework build.

  • medium likelihoodGPU arch not in the build ('no kernel image')

    Fix · Use a build that includes your compute capability / gfx arch.

Validation
  • ✓A tiny GPU tensor op runs without error
Escalation

None — resolve by aligning versions.

Data to collect first
  • ·Driver CUDA capability
  • ·Framework's built CUDA/ROCm version
  • ·GPU compute capability / gfx arch