[Versioning] Require torch 2.11 or later and deprecate torch<2.13

[theap06]

Description

Raises the torch floor from 2.1 to 2.11 and deprecates torch 2.11 and 2.12.

• pyproject.toml requires torch>=2.11.0a0. The tensordict range (>=0.14.2,<0.15.0) is unchanged, so pip resolves tensordict 0.14.3 on torch 2.11 and 2.12.

• Importing torchrl on torch<2.13 raises a DeprecationWarning that names v0.17 as the release that requires torch>=2.13 (_warn_if_torch_deprecated in torchrl/_utils.py, tested in test/test_utils.py). This follows the two-release deprecation window.

• README and docs/source/index.rst state the new floor and the deprecation.

• test_native_auto_reset_compile_step_and_maybe_reset now requires torch 2.12: on 2.11, Dynamo cannot trace the self.__dict__.setdefault of the EnvBase key caches.

• olddeps runs the declared minimum: torch 2.11.0 and torchvision 0.26.0 from the cu126 index on osdc-cuda:cuda12.6.3-cudnn-devel-ubuntu24.04. Moving it off torch 2.1 needed the following fixes:

  • install hoptorch, because the job installs torchrl with --no-deps;
  • drop the setuptools<82 pin, which torch 2.1 needed;
  • install tensordict<0.15 on release runs;
  • link mujoco-py from the checkout and install libglew-dev;
  • add an autouse fixture that resets the interaction type between tests, because the serial olddeps run let one leak into test_exploration.

• CLAUDE.md 7c describes olddeps as the minimum supported versions.

• I have raised an issue to propose this change (required for new features and bug fixes)

Types of changes

• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation (update in the documentation)

Testing

Local, macOS arm64, Python 3.12:

• torch 2.11.0 + tensordict 0.14.3:
  • test_utils, test_auto_reset, test_exploration: 334 passed.
  • test_ppo, test_sac, test_iql, test_specs, test_distributions: 5937 passed.
  • The deprecation warning fires on import, and the auto-reset compile test skips.
• torch 2.13.0 + tensordict 0.14.3: no deprecation warning, the auto-reset compile test runs and passes, and test_import_warns_on_deprecated_torch passes.
• uv pip compile (x86_64 manylinux, Python 3.12): tensordict 0.14.3 on torch 2.11 and 2.12, 0.15.1 on 2.13. The vla extra resolves lerobot 0.6.1 on torch 2.11.

Known limitations

• torchrl._torchrl is built against one torch ABI. A wheel built against a newer torch loads without its C++ extension on torch 2.11, so PrioritizedReplayBuffer raises SumSegmentTreeFp32 is not available ([BUG] SumSegmentTreeFp32 not available #3508). Isaac Lab 3.0 users who need prioritized replay have to build torchrl from source until wheels are built per torch version.
• The unittests-isaaclab job still uses the isaac-lab:2.3.0 image (Isaac Sim 5.1, torch 2.7, below this floor; torchrl is installed with --no-deps). Testing Isaac Lab 3.0 needs a follow-up job.
• The CI tensordict pin (.github/ci_commit_pins/tensordict.txt, from before tensordict required torch 2.13) stays. Unpinning and moving to tensordict 0.15 belongs with the torch 2.13 floor in v0.17.
• Version gates below 2.11 (implement_for("torch", ...) and TORCH_VERSION checks) are now dead code. Removing them is a follow-up.

Checklist

• I have read the CONTRIBUTION guide (required)
• My change requires a change to the documentation.
• I have updated the tests accordingly (required for a bug fix or a new feature).
• I have updated the documentation accordingly.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/4595

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 Unclassified Failure

As of commit 5069dd0 with merge base e7a05a7 ():

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

• Unit-tests on Linux / tests-olddeps (3.10, 12.6, remainder) / linux-job (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  test/rb/test_storages.py::TestStorageExport::test_device_storage[cuda-ListStorage-True]

This comment was automatically generated by Dr. CI and updates every 15 minutes.