Back to Cookbook
CUDA Stack Sanity Check for Scientific Python
Make GPU availability a checklist, not a mystery.
Resolve CUDA driver/runtime mismatches and GPU framework install pitfalls by checking driver versions, supported CUDA toolkits, and whether binaries bundle CUDA runtimes (PyTorch/TensorFlow).
CommunitySubmitted by CommunityWork12 min
INGREDIENTS
🔍Web
PROMPT
You are OpenClaw. Ask for nvidia-smi, OS, framework (PyTorch/TensorFlow/JAX), install method, and the exact error text. Then walk through driver/runtime compatibility, propose the correct install commands for CPU vs GPU builds, and provide a minimal verification script. Include container-specific advice if they are using Docker/Singularity.
Pain point
GPU frameworks install but fail at runtime with driver/runtime mismatch errors, missing libcuda, or no GPU
detected.
Repro/diagnostic steps
- Collect: `nvidia-smi` output (driver + reported CUDA capability).
- Identify framework build: PyTorch/TensorFlow CUDA build tag, and the install command used.
- Run a minimal GPU probe script (framework-specific).
Root causes (common)
- NVIDIA driver too old for the CUDA runtime/toolkit used.
- Confusion between "driver supports CUDA X" and "toolkit installed is CUDA X".
- Multiple CUDA toolkits on PATH/LD_LIBRARY_PATH.
- Container run without GPU runtime flags.
Fix workflow
- Ensure driver meets compatibility requirements per NVIDIA guidance.
- Install the framework build that matches the supported CUDA runtime for your system.
- Prefer prebuilt wheels/containers that bundle CUDA runtime when appropriate.
- Validate with a smoke test (list devices; run one tensor op on GPU).
Expected result
- Framework reports GPU devices and executes a minimal operation on GPU without version errors.
References
- https://docs.nvidia.com/deploy/cuda-compatibility/
- https://pytorch.org/get-started/locally/
- https://www.tensorflow.org/install/pip
- https://stackoverflow.com/questions/3253257/cuda-driver-version-is-insufficient-for-cuda-runtime-version
Tags:#gpu#cuda#python#ml#hpc