| Operation | PyTorch 2.4 + CUDA 12.4 | PyTorch 2.6 + CUDA 12.6 | Improvement | |-----------|------------------------|-------------------------|-------------| | MFU (Model FLOPs utilization) | 38.2% | 40.5% | +2.3% | | Kernel launch time (microbench) | 12.4 µs | 8.2 µs | -34% | | cuDNN attention forward (512 seq len) | 0.43 ms | 0.39 ms | -9% |
: While CUDA 13.0 is now the default for pip install torch on PyPI, CUDA 12.6 is the recommended fallback for users with Maxwell (SM 5.0) and Pascal (SM 6.0) GPUs, which are no longer supported by the CUDA 13.x toolkits. pytorch cuda 12.6 news
: These current versions (released in early 2026) maintain CUDA 12.6 as a "legacy" stable build. | Operation | PyTorch 2
import torch print(torch.version.cuda) # Should output "12.6" print(torch.cuda.is_available()) # True pytorch cuda 12.6 news