PyTorch 2.15 Support #9241

PyTorch 2.15 was released on 2026-10-28 (release notes, blog). MONAI dev declares torch>=2.8.0 and CI tests 2.8.0 / 2.9.1 / 2.10.0 / 2.11.0. This issue tracks adding 2.15.0 to the supported and tested set.

Blocker: PyTorch 2.15 publishes no cp310 wheels

2.15 raises the minimum Python version for binaries to 3.11 (pytorch/pytorch#198233, pytorch/pytorch#199398). The 2.15.0 wheels on PyPI are cp311 / cp312 / cp313 / cp314 / cp315 only, so pip install torch==2.15.0 cannot resolve on Python 3.10. MONAI still declares requires-python = ">=3.10" and the min-dep matrix pairs Python 3.10 with the highest PyTorch version, so those cells cannot move to 2.15 as they are.

Either keep the Python 3.10 cells on 2.14.1 and add a 3.11+ cell for 2.15, or drop Python 3.10 support in the same change. The matrix comments ("Test PyTorch versions under Ubuntu with lowest Python version supported") need updating either way.

Version pairing

• torchvision 0.29.2 is the matching release (requires torch>=2.14). TORCHVISION_VER1 (0.23.0, used with the lowest torch) is unaffected.
• The default PyPI wheel is CUDA 13.2 (torch 2.15.0+cu132); CUDA 13.4 builds are also published.
• Other pins that mention a torch version: Dockerfile (nvcr.io/nvidia/pytorch:25.12-py3), codeql-analysis.yml (PYTORCH_VER: 2.8.0), cron-ngc-bundle.yml and setupapp.yml (2.8.0), WIN_FULLDEP_VER (2.11.0 / torchvision 0.26.0).

Release-note items that touch MONAI code paths

1. Rewritten NCCL backend is now the default ([c10d] Make nccl2 the default NCCL backend pytorch/pytorch#192281, #192895, #193237, #193241). Communicators are created eagerly in init_process_group/new_group, each group is bound to one device (device_id argument), isinstance(pg, ProcessGroupNCCL) is now False, and work.get_future() on an async all_gather returns the per-rank outputs instead of one stacked tensor. nccl-legacy and TORCH_DIST_USE_NCCL2=0 disappear in the next release. Relevant to monai.utils.dist, the *_dist tests and FL workflows.
2. FSDP2 gradient dtype change ([FSDP2] Separate sharded and unsharded gradient dtypes pytorch/pytorch#194434): with MixedPrecisionPolicy(reduce_dtype=None) gradients are now accumulated and reduced in the parameter's grad_dtype (fp32 by default) instead of param_dtype. Affects users wrapping MONAI networks in fully_shard.
3. torch.load(weights_only=True) no longer allowlists nested tensors (Remove NJT (nested jagged tensor) from weights_only safe globals pytorch/pytorch#194380). MONAI checkpoints are unaffected; worth a line in the bundle/checkpoint docs.
4. torch.ops.* calls now use PyTorch's own exception and warning translation (Fix C++ warning propagation for torch.ops and custom ops pytorch/pytorch#193451): pybind11 exception translation is lost and TORCH_WARN from an op becomes a Python warning, which raises under -W error. Relevant to monai/_extensions and the compiled test job.
5. Smaller ones: passing a set of parameters to an optimizer now emits a FutureWarning (Reject a bare set of params, consistent with set inside a param group pytorch/pytorch#198710), and torch.distributed.checkpoint output written by 2.15 cannot be read by 2.14 ([S615805][S694227][PyTorch] Safely preserve tensor strides (#194251) pytorch/pytorch#194251).

Local verification on 2.15.0

Environment: Python 3.13.18, torch 2.15.0+cu132, torchvision 0.29.2+cu132, MONAI dev (1.6.0rc1+91.g609a1108) installed from source with .[all,testing], RTX 5090 (driver 580.178.04). A second environment with the same dependency resolution but torch 2.14.1 was used as the baseline.

No failure is specific to PyTorch 2.15. Every failing test was re-run serially in both environments and fails identically on 2.14.1:

The first six rows are the unbounded-dependency-pin problem tracked in #9069, not a 2.15 problem.

All 17 failing modules were re-run serially on an idle GPU (nothing else holding VRAM) and produced the same results, so none of the above is a memory-pressure artefact or a parallel-run collision. A full GPU quick pass (17,910 tests) on the idle GPU leaves 35 failing tests, all inside the modules listed above.

Smoke checks that pass on 2.15.0: torch.jit.script and torch.jit.trace of a UNet, torch.compile(backend="eager"), torch.load(weights_only=True) on a MONAI-style checkpoint, torch.distributed.checkpoint import, torch.distributed gloo all_reduce, and torch.cuda.nccl availability. (torch.jit.* still works but emits its existing FutureWarning; that warning is present in 2.14 as well.)

Tasks

• Add 2.15.0 to PYTORCH_VER* and the min-dep matrix, and decide the Python 3.10 handling
• Update the Dockerfile, codeql-analysis.yml, cron-ngc-bundle.yml and setupapp.yml pins
• Check monai.utils.dist against the new default NCCL backend, including FL cold start
• Decide whether the requires-python / torch>= floors move with this release
• Keep the dependency-drift failures (see Unit-test results depend on the Python version: unbounded upper pins let 3.12 resolve dependencies CI never sees #9069) out of the 2.15 work