NVIDIA GeForce RTX 3080 Ti with CUDA capability sm_86 is incompatible with PyTorch issue occurs when you are “using an outdated version of PyTorch, which is incompatible with sm_86.”
The sm_86 architecture is specific to some of the newer NVIDIA cards, like the RTX 3080Ti, and not all versions of PyTorch and CUDA may support it out of the box.
Steps to fix the Issue:
Step 1: Update PyTorch
Ensure you’re using the latest version of PyTorch, as newer versions are more likely to support the latest hardware. Based on your initial installation method, you can update PyTorch via pip or conda.
# Using pip pip install --upgrade torch torchvision # Using conda conda update pytorch torchvision -c pytorch
Step 2: Check the CUDA Toolkit
Update your CUDA Toolkit to a version that supports sm_86 if it doesn’t already. Usually, the latest version should support the latest architectures.
Step 3: Custom Build
If the mainstream PyTorch build doesn’t support sm_86, you could try building PyTorch from the source with specific flags to enable sm_86 support.
When running the build command, you can specify the architectures you want to support by setting the TORCH_CUDA_ARCH_LIST environment variable or modifying the CMake flags.
TORCH_CUDA_ARCH_LIST="8.6" python setup.py install
Step 4: Check System Environment
Verify that the system environment variables related to CUDA are set correctly. Sometimes, incorrect paths or settings can cause compatibility issues.
Step 5: Driver Update
Make sure your NVIDIA driver is up-to-date. Newer drivers are more likely to support new hardware architectures like sm_86.
Step 6: Check for Updates
Often, issues related to new hardware support are addressed in software updates. Make sure your system, CUDA Toolkit, and PyTorch are fully updated.