Supported Cuda Matrix
| CUDA | Default driver | Status |
|---|---|---|
| 12.5 | 545.29 | Full |
| 12.6 | 555.42 | Full |
| 13.0 (preview) | 555.42 | Best effort |
Required Headers / Libs
cuda_runtime.hcublas_v2.hcuda.h(NVML bridge)cupti.h(only when--enable-cupti)
Compile-Time Flags
- For TensorRT integration: TRT 10.x.
- For FlashAttention: ≥ 2.5 SDK.
- For Transformer Engine: TE 1.13+.
ABI Compatibility
PyTorch wheels distributed with CUDA 12.x require that the host driver support the corresponding PTX JIT path. We surface a warning if cudaDriverGetVersion returns older than the wheel's expected minimum.