ci: drop PyTorch 2.0-2.5 and raise the minimum supported version to 2.6

[bhimrazy]

What does this PR do?

Drops the end-of-life PyTorch minors and raises the floor to 2.6, keeping the latest five supported minors (2.6–2.10). Mirrors Lightning-AI/pytorch-lightning#21851.

• CPU matrix: drop the 2.0–2.5 rows, retarget oldest/macOS/Windows to 2.6.0 — 23 → 17 jobs
• Docker: drop the 2.0.1–2.5.1 CUDA images — 14 → 8 builds
• GPU lanes: oldest image moves to torch 2.6
• Requirements: torch >=2.6.0, torchaudio >=2.6.0, torchvision >=0.21.0, pytorch-lightning >=2.5.1 (first release CI-tested against torch 2.6)

The GPU lanes hardcoded TORCH_VER == "2.0" to gate the oldest requirements, which silently goes stale when the floor moves on the CPU side. Now derived from requirements/base.txt, with a fail-fast check if an image sits below the declared floor.

How the floor check works

TORCH_MIN is read from requirements/base.txt and used both to gate the oldest-requirements branch and to reject a mismatched image:

ERROR: image provides torch 2.4, below the 2.6 floor in requirements/base.txt.
       Bump the images in .lightning/workflows/ together with the floor.

The comparison goes through packaging.version.Version rather than a string compare, so 2.10 > 2.6 orders correctly. It runs before adjust-torch-versions.py, which rewrites the >= pin to an exact version.

Dependency floors raised for the newer images

The GPU and docker images move from Python 3.10 to 3.11, so a few packages need floors that actually publish cp311 wheels — each carries an inline note explaining why:

• onnxruntime >=1.12.0 → >=1.15.0
• scikit-image >=0.19.0 → >=0.20.0
• mecab-ko >=1.0.0 → >=1.0.1

PIP_BREAK_SYSTEM_PACKAGES=1 was added to integrations.yml, since the newer pytorch/pytorch images ship a PEP 668 "externally managed" system Python.

Docker, docs, and a bug fixed along the way

• dockers/ubuntu-cuda/Dockerfile — defaults moved to CUDA 12.4.1 / Python 3.11 / torch 2.6.
• dockers/README.md — the build example pointed at a dockers/base-cuda/Dockerfile path that no longer exists, and at an unsupported CUDA 11.7.1 / torch 2.0 combination.
• .lightning/workflows/unittests.yml referenced an undefined lowercase ${cuda_version} when computing CUDA_VERSION_M_M, silently producing an empty CUDA_VERSION_MM and a broken wheel index URL (.../whl/cu/torch_stable.html). Fixed to ${CUDA_VERSION%.*}.

Follow-ups: docs/support policy, then removing the dead _TORCH_GREATER_EQUAL_2_1..2_5 paths.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 36%. Comparing base (d184220) to head (02307a0).
⚠️ Report is 12 commits behind head on master.
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@          Coverage Diff           @@
##           master   #3449   +/-   ##
======================================
- Coverage      37%     36%   -0%     
======================================
  Files         349     349           
  Lines       19901   19910    +9     
======================================
- Hits         7264    7263    -1     
- Misses      12637   12647   +10     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request