The Pulse
PyTorch 2.15 Makes CUDA 13.2 Its Default GPU Build
PyTorch 2.15 ships CUDA 13.2 as its default GPU wheel, with CUDA 13.4 wheels also available and experimental Rubin support. The release removes CUDA 13.0 builds and drops binary support for Maxwell, Pascal and Volta GPUs as CUDA 12.6 builds
AI.info Team · Written with AI · fact-checked ·
PyTorch 2.15 makes CUDA 13.2 its default GPU build, with CUDA Toolkit 13.2.2 specified for the default wheel. The PyTorch Foundation announced the release on October 28. CUDA 13.4 wheels ship alongside it for Linux x86_64, Linux aarch64 and Windows, with experimental support for NVIDIA Rubin GPUs in the 13.4 builds.
CUDA 13.2 remains the default
Users who need a different build can still choose the CUDA 13.4 wheels; the release blog labels those wheels generally available, while describing Rubin support within them as experimental. PyTorch’s release compatibility table lists CUDA 13.2 as stable and CUDA 13.4 as experimental for version 2.15, alongside cuDNN 9.26.0.51 and NCCL 2.30.7 for both builds.
CUDA 13.2 was already identified as the Stable/Default PyPI version for PyTorch 2.14, while CUDA 13.0 was listed as stable. For 2.15, the release plan keeps CUDA 13.2 as the default and removes CUDA 13.0 builds from the matrix. The release compatibility table records CUDA 13.2 as stable and CUDA 13.4 as experimental.
CUDA 13.2.2 also fixes a compiler bug
The toolkit version matters beyond the number attached to the wheel. In a September notice to developers, Andrey Talman, a PyTorch contributor, described the change to nightly builds:
“Starting the week of Monday, Sept 28 2026, we are removing CUDA 13.0 from the PyTorch nightly build matrix and CI. Nightly CUDA builds will be 13.2 (stable / default) and 13.4 (prototype).”
The notice says CUDA 13.2.2 fixes a compiler issue that could produce stale or corrupted register values in certain kernels with two or more nested levels of thread divergence. It says kernels with only one divergence level were unaffected, and that CUDA 13.0 predates the fix.
PyTorch’s shift away from CUDA 13.0 therefore follows a specific compatibility and correctness rationale, not just a version refresh. The team’s developer notice told users of nightly builds pinned to the cu130 package index to update those references to cu132; default nightly users would receive the newer build automatically.
Older NVIDIA GPUs lose the newer wheels
CUDA 13.x does not support NVIDIA’s pre-Turing architectures. PyTorch’s release planning document says dropping CUDA 12.6 builds in 2.15 removes binary support for Maxwell, Pascal and Volta GPUs, including cards that relied on CUDA 12.x. Turing and newer architectures remain supported by CUDA 13.2 and 13.4, according to the project’s CUDA support plan.
For owners of affected cards, the release does not make existing installations stop working, but PyTorch 2.15’s new CUDA wheels are not a path forward for those GPUs. The project’s earlier warning advised users of Maxwell, Pascal and Volta to remain on PyTorch 2.14 with CUDA 12.6 or an earlier release. That is the concrete trade-off in 2.15: CUDA 13.2 as the default and CUDA 13.4 as an alternate build, while dropping support for hardware that CUDA 13.x cannot target.
2.15 brings a larger release
CUDA’s default sits among a broader set of updates. PyTorch says 2.15 includes 3,117 commits from 583 contributors since version 2.14. The release also adds ROCm 10.0 wheels alongside ROCm 7.14, expands Python support through 3.15, and makes a rewritten NCCL backend the default for NVIDIA process groups.
For teams upgrading, the immediate checks are practical: confirm the GPU architecture is Turing or newer, review any pinned CUDA wheel index, and test dependent extensions against the new toolkit. Those decisions determine whether a project can adopt the new default directly, needs the CUDA 13.4 build, or must keep an older PyTorch release for its hardware.