Pin DiffEqGPU PyTorch to CUDA 12.6 - #1789

What changed and why

Pin DiffEqGPU's Python environment to Python 3.12 and the official PyTorch 2.13.0 CUDA 12.6 Linux x86-64 wheel, including its SHA-256. The unconstrained PyPI dependency selected torch==2.13.0+cu130; that build contains no compute-capability 7.0 kernels for the V100 runners and explicitly recommended reinstalling from PyTorch's CUDA 12.6 index.

This is a dependency-only change. The Julia CUDA 12.9 preferences are handled separately by #1733. The singleton benchmark-source fixes merged through #1715, with the CRN parameter-count follow-up in #1788.

Failing before

The V100 validation run reported:

Found GPU0 Tesla V100-PCIE-32GB which is of compute capability (CC) 7.0.
Your installed torch==2.13.0+cu130 does not include kernels for this GPU.
For CUDA 12.6 use pip install torch==2.13.0 --index-url https://download.pytorch.org/whl/cu126

The final dependency assertions fail against current master:

Test Summary:                  | Fail  Total  Time
DiffEqGPU PyTorch V100 runtime |    2      2  1.9s
ERROR: Some tests did not pass: 0 passed, 2 failed, 0 errored, 0 broken.

Passing after

A clean CondaPkg/Pixi environment resolved and installed successfully. Importing the package through the benchmark's PythonCall environment produced:

TORCH_VERSION=2.13.0+cu126
TORCH_CUDA=12.6

Repository checks:

$ GROUP=Core julia +1.11.9 --project=. -e 'using Pkg; Pkg.test()'
Test Summary: | Pass  Total   Time
Core          |   45     45  34.6s
Testing SciMLBenchmarks tests passed

$ GROUP=QA julia +1.11.9 --project=. -e 'using Pkg; Pkg.test()'
Test Summary: | Pass  Total     Time
QA            |   21     21  1m44.7s
Testing SciMLBenchmarks tests passed

Runic 1.x checked test/core.jl, typos checked both changed files, and git diff --check passed.

Not verified locally

This host has no NVIDIA GPU, so actual V100 execution remains unverified. The V100 run also requires the Julia CUDA 12.9 preferences from #1733 and the CRN follow-up from #1788. No docs build was run because this changes a benchmark dependency declaration and a private regression test, not public API or docstrings.

PyTorch was already a benchmark dependency, so this pin adds no dependency or license.

Failing V100 validation: https://github.com/SciML/SciMLBenchmarks.jl/actions/runs/33298026014

🤖 Generated with Claude Code

https://chatgpt.com/codex/tasks/01a03a17-ad6f-7131-82fc-d0fd57ea6512