PyTorch 2.10 with CUDA 12.6
For PyTorch 2.10 with CUDA 12.6, pytorch release is 2.10; cuda build is CUDA 12.6; python versions is >=3.10, <=(3.14, 3.14t experimental); cudnn version is 9.10.2.21; gpu architectures compiled in, linux wheel is Maxwell(5.0), Pascal(6.0), Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0), recorded from its source on 2026-09-29; source re-checked 2026-10-15.
- Pairing
- PyTorch 2.10 with CUDA 12.6 our reading
- PyTorch release
- 2.10 verified
- CUDA build
- CUDA 12.6 verified
- Python versions
- >=3.10, <=(3.14, 3.14t experimental) verified
- cuDNN version
- 9.10.2.21 verified
- GPU architectures compiled in, Linux wheel
- Maxwell(5.0), Pascal(6.0), Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0) verified
- Build script: base architecture list
- TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0" verified
- Build script: this CUDA build's architecture line
- TORCH_CUDA_ARCH_LIST="5.0;6.0;${TORCH_CUDA_ARCH_LIST}" verified
- GPU architectures compiled in, Windows wheel
- 5.0;6.0;6.1;7.0;7.5;8.0;8.6;9.0 verified
- Minimum driver (minor version compatibility)
- >= 525 verified
- Toolkit driver, Linux x86_64
- >=560.28.03 verified
- Toolkit driver, Windows x86_64
- >=560.76 verified
- Wheel index URL
- https://download.pytorch.org/whl/cu126 our reading
Values marked our reading are our classification of what the source says — the source does not print them in those words. The quote below is the evidence for each one; judge it yourself.
What the source says
CUDA build and cuDNN version
The source prints these values inside the passage below but does not name them there. The passage shows where each one appears; what ties them to this record is the position on the page, not anything the passage says.
| 2.10 | >=3.10, <=(3.14, 3.14t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21) (NCCL 2.27.5), CUDA 12.8 (CUDNN 9.10.2.21) (NCCL 2.27.5) | CUDA 13.0 (CUDNN 9.15.1.9) (NCCL 2.28.9) | ROCm 7.1 |
— raw.githubusercontent.com, retrieved 2026-09-29
Where each value comes from
This source states these in separate places, so each value is shown with the passage that states it.
GPU architectures compiled in, Linux wheel
| 12.6.3 | Maxwell(5.0), Pascal(6.0), Volta(7.0), Turing(7.5), Ampere(8.0, 8.6), Hopper(9.0) | |
Build script: base architecture list
TORCH_CUDA_ARCH_LIST="7.0;7.5;8.0;8.6;9.0"
Build script: this CUDA build's architecture line
12.6) TORCH_CUDA_ARCH_LIST="5.0;6.0;${TORCH_CUDA_ARCH_LIST}" ;; # Only 12.6 includes Legacy Maxwell/Pascal that will be removed in future releases
GPU architectures compiled in, Windows wheel
IF "%BUILD_VISION%" == "" ( set TORCH_CUDA_ARCH_LIST=5.0;6.0;6.1;7.0;7.5;8.0;8.6;9.0
Minimum driver (minor version compatibility)
12.x | >= 525 | < 580
Toolkit driver, Linux x86_64 and Toolkit driver, Windows x86_64
CUDA 12.6 GA | >=560.28.03 | >=560.76
— all from raw.githubusercontent.com, retrieved 2026-09-29
Sources
- raw.githubusercontent.comhttps://raw.githubusercontent.com/pytorch/pytorch/main/RELEASE.md
- raw.githubusercontent.comhttps://raw.githubusercontent.com/pytorch/pytorch/release/2.10/.ci/manywheel/build_cuda.sh
- raw.githubusercontent.comhttps://raw.githubusercontent.com/pytorch/pytorch/release/2.10/.ci/pytorch/windows/cuda126.bat
- raw.githubusercontent.comhttps://raw.githubusercontent.com/pytorch/pytorch/release/2.11/RELEASE.md
- docs.nvidia.comhttps://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html