[M-Colley]

This lifts the torch<2.11 / torchvision<0.26 caps from #38 and the <3.13 Python cap. FMPose3D gives the same results on torch 2.15 as on torch 2.10, and no code changes were needed.

Why: FMPose3D currently cannot be installed next to packages that need torch 2.11 or newer, or on Python 3.13 and 3.14. One example is MyoSuite's mjlab backend, which needs torch>=2.13; we would like to retarget FMPose3D poses onto its musculoskeletal models.

The CUDA 13 driver concern from #38

From torch 2.11 on, PyPI's default Linux wheels bring a CUDA 13 runtime, which needs an R580+ driver. On an older driver the fallback to the CPU is not silent:

• torch turns the driver error into a warning: "CUDA initialization: The NVIDIA driver on your system is too old (found version …). Please update your GPU driver … Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver."
• The warning comes from c10/cuda/CUDAFunctions.cpp (device_count) and is unchanged from v2.11.0 to v2.15.0.
• FMPose3D then runs on the CPU.

The README's PyTorch/CUDA note still described the old torch<2.5 pin. It now explains this and shows how to install a CUDA 12 build first, with --index-url https://download.pytorch.org/whl/cu126; that index has torch up to 2.14.1.

Changes

• pyproject.toml: torch>=2.4.1 and torchvision>=0.19.1, with no upper bounds. requires-python = ">=3.10", plus classifiers for Python 3.11 to 3.14.
• README.md: the PyTorch/CUDA note above, and "Python 3.10 or newer".
• .github/workflows/build.yaml:
  • The floor job installed torch 2.4.0, which does not satisfy torch>=2.4.1. So pip install '.[dev]' replaced it with the newest allowed torch, and the floor was never tested. It now installs 2.4.1.
  • torchvision is installed with torch from the CPU index, so the two builds match.
  • A new job runs the newest torch (2.15.0) on Python 3.14. The CPU index has Linux wheels for torch 2.15.0 and torchvision 0.29.0 there.

Testing

All runs used Windows and CPU builds.

	torch 2.10.0, Py 3.12	torch 2.15.0, Py 3.12	torch 2.15.0, Py 3.13	torch 2.15.0, Py 3.14
pytest -m "not functional and not network"	116 passed, 7 skipped, 1 failed*	same	same	same
pytest -m functional (weights from HF and Google Drive)	18 passed, 6 skipped	18 passed, 6 skipped	18 passed, 6 skipped	18 passed, 6 skipped

* test_loss_decreases_over_steps, which draws unseeded noise. It failed in 2 of 10 runs on torch 2.10 and 1 of 10 on torch 2.15. The 6 skips are the animal pipeline, which needs DeepLabCut (not installed).

Seeded outputs, torch 2.15 against torch 2.10:

• 3D lifting of fixed 2D keypoints (8 frames, seed=0): at most 3.9e-7 m apart on Python 3.12, 3.13 and 3.14. The poses reach up to 1.5 m.
• End to end on demo/images/running.png (YOLO, HRNet and lifting, seed=0): identical 2D keypoints and scores, and 3D poses at most 1.8e-7 m apart.

Not tested:

• CUDA builds on a GPU;
• Linux and macOS;
• the animal pipeline;
• torch 2.11 to 2.14 individually;
• the Human3.6M evaluation (I don't have the dataset).

Seen while testing, not changed here

• Flaky test: test_loss_decreases_over_steps fails now and then on any torch version. Seeding it would make it deterministic.
• Unmarked network test: tests/test_demo_human.py::test_2d_pose_estimation downloads the HRNet and YOLO weights from Google Drive, and pip-installs gdown at runtime, but is not marked network. In a venv without pip, such as one made by uv, it fails with ModuleNotFoundError: gdown.
• CI never fails on tests: the pytest step has continue-on-error: true, so test failures don't fail the job.

I can open separate PRs for any of these.

🤖 Generated with Claude Code