PyTorch v2.14.0-rc3 fixes CUDA 13.4 build with cuDNN 9.25
PyTorch's v2.14.0-rc3 patches CUDA 13.4 build failures with cuDNN 9.25
PyTorch v2.14.0-rc3 addresses critical build failures in CUDA 13.4rc1 by bumping cuDNN from 9.24.0.43 to 9.25.0.15. The issue stemmed from missing fused attention engines for compute capability 10.7 in the older cuDNN version, causing RuntimeError: cuDNN Frontend errors during scaled_dot_product_attention operations on Rubin hardware.
The fix was validated on Vera Rubin NVL72 nodes (sm_107, aarch64) using CUDA 13.4 nightly wheels (torch-2.14.0.dev20260809+cu134). Testing showed all backend probes (cudnn, flash, mem_eff, math, default) passing after upgrading cuDNN, compared to failures with cuDNN 9.24.0.43. The patch was authored with assistance from Claude Opus 5 (1M context) and approved by NVIDIA's internal team.
- PyTorch v2.14.0-rc3 fixes CUDA 13.4rc1 build failures by upgrading cuDNN to 9.25.0.15
- Resolves fused attention engine errors on Rubin hardware (compute capability 10.7)
- Tested on Vera Rubin NVL72 nodes with CUDA 13.4 nightly wheels, all backend probes now pass
Why It Matters
Critical fix for AI researchers and developers using CUDA 13.4 with PyTorch, enabling stable operations on next-gen NVIDIA hardware.