Linux Py3.12 CPU CI fails to resolve `torch-tensorrt` when testing older pinned PyTorch (2.4.1 / 2.5.1)

[littlebullGit]

Bug description

Summary

The Test PyTorch / pl-cpu workflow fails consistently on Linux + Python 3.12 for configurations pinned to PyTorch 2.4.1 and 2.5.1, due to an unsatisfiable dependency resolution involving torch-tensorrt.

This does not reproduce on macOS/Windows because the torch-tensorrt requirement is currently Linux-only.

Affected CI job(s)

• Test PyTorch / pl-cpu (ubuntu-22.04, lightning, 3.12.7, 2.4.1)
• Test PyTorch / pl-cpu (ubuntu-22.04, lightning, 3.12.7, 2.5.1)

Current behavior / error

CI fails during dependency installation (using uv pip install) with “No solution found” because the available torch-tensorrt wheels from the configured sources require newer torch versions than the job’s pinned torch.

Typical failure excerpt pattern:

• Resolver only sees torch-tensorrt==2.6.x+cu126, 2.7.x+cu12x, 2.8.x+cu12x, 2.9.x+cu12x …
• Those versions depend on torch==2.6.0 or torch>=2.7.0 (etc.)
• But CI pins torch==2.4.1 / torch==2.5.1

What version are you seeing the problem on?

v2.5

Reproduced in studio

No response

How to reproduce the bug

## Repro / relevant configuration
The failing install happens in `.github/workflows/ci-tests-pytorch.yml`:


uv pip install ".[${EXTRA_PREFIX}extra,${EXTRA_PREFIX}test,${EXTRA_PREFIX}strategies]" \
  --upgrade \
  --find-links="${TORCH_URL}" \
  --find-links="https://download.pytorch.org/whl/torch-tensorrt"


The requirement is currently in `requirements/pytorch/test.txt`:


torch-tensorrt; platform_system == "Linux" and python_version >= "3.12"

Error messages and logs

Run uv pip install ".[${EXTRA_PREFIX}extra,${EXTRA_PREFIX}test,${EXTRA_PREFIX}strategies]"
× No solution found when resolving dependencies:
╰─▶ Because only the following versions of torch-tensorrt{sys_platform ==
'linux'} are available:
torch-tensorrt{sys_platform == 'linux'}==2.6.0+cu126
torch-tensorrt{sys_platform == 'linux'}==2.6.1+cu126
torch-tensorrt{sys_platform == 'linux'}==2.7.0+cu126
torch-tensorrt{sys_platform == 'linux'}==2.7.0+cu128
torch-tensorrt{sys_platform == 'linux'}==2.8.0+cu126
torch-tensorrt{sys_platform == 'linux'}==2.8.0+cu128
torch-tensorrt{sys_platform == 'linux'}==2.8.0+cu129
torch-tensorrt{sys_platform == 'linux'}==2.9.0+cu126
torch-tensorrt{sys_platform == 'linux'}==2.9.0+cu128
torch-tensorrt{sys_platform == 'linux'}==2.9.0+cu130
and torch-tensorrt<=2.6.1+cu126 depends on torch==2.6.0, we can conclude
that torch-tensorrt{sys_platform == 'linux'}<2.6.1+cu126 depends on
torch==2.6.0.
And because torch-tensorrt==2.6.1+cu126 depends on torch==2.6.0, we
can conclude that torch-tensorrt{sys_platform == 'linux'}<2.7.0+cu126
depends on torch==2.6.0.
And because torch-tensorrt>=2.7.0+cu126,<=2.7.0+cu128 depends on
torch>=2.7.0,<2.8.0 and torch>=2.7.0,<2.8.0, we can conclude that
torch-tensorrt{sys_platform == 'linux'}<2.8.0+cu126 depends on one of:
torch==2.6.0
torch>=2.7.0,<2.8.0

Environment

Current environment

#- PyTorch Lightning Version (e.g., 2.5.0):
#- PyTorch Version (e.g., 2.5):
#- Python version (e.g., 3.12):
#- OS (e.g., Linux):
#- CUDA/cuDNN version:
#- GPU models and configuration:
#- How you installed Lightning(`conda`, `pip`, source):

More info

Root cause analysis (why the resolver fails)

1. torch-tensorrt is pulled unconditionally for Linux + Py3.12 test installs, including CPU-only CI jobs.
2. CI also points the resolver at the PyTorch torch-tensorrt wheel index (https://download.pytorch.org/whl/torch-tensorrt).
3. That index currently appears to publish only CUDA-tagged torch-tensorrt wheels (e.g. +cu126, +cu128, …) that require torch>=2.6 / >=2.7+.
4. Therefore, for Linux + Py3.12 jobs pinned to torch==2.4.1 or torch==2.5.1, resolution is mathematically unsatisfiable.

Suggestion (keep TRT coverage; avoid breaking CPU jobs)

Update .github/workflows/ci-tests-pytorch.yml:

1. Compute an env var (e.g. INSTALL_TORCH_TRT=1/0) based on the matrix torch version (and OS).
2. If disabled, remove the torch-tensorrt line from requirements/pytorch/test.txt before running uv pip install ....
3. Only add --find-links=https://download.pytorch.org/whl/torch-tensorrt when INSTALL_TORCH_TRT=1.

I can work on it if suggestion is OK.

cc @ethanwharris

[bhimrazy]

It looks like Torch-TensorRT wheels for torch < 2.6.0 may have been dropped.

Related discussion opened by @GdoongMathew: pytorch/TensorRT#3967

We’ll see if the authors provide any clarification there.

[GdoongMathew]

Related to #21412 & #21413

[GdoongMathew]

@bhimrazy , since pytorch/test-infra#7599 and pytorch/TensorRT#3967 is fixed, shall we reverse part of the #21413?