[4031427256] Upstream Reporter: Lalatendu Mohanty
Upstream issue status: Open
Upstream description:

Notebooks embeds `[[tool.uv.index]]` blocks with CUDA-version-specific PyTorch URLs (`https://download.pytorch.org/whl/cu128`) in pyproject.toml, and uses local version suffixes (`torch==2.9.1+cu128`) to select the right wheel. Both patterns break on AIPCC.

Base-containers CUDA images should set `UV_EXTRA_INDEX_URL` (or equivalent) pointing at the appropriate PyTorch index for the CUDA version. This lets consumers use plain `torch==2.9.1` in pyproject.toml ? the correct accelerator wheel is selected by the build environment, not the dependency file.

*What to do:*

• Set `UV_EXTRA_INDEX_URL` (or equivalent) in CUDA base images to the PyTorch CUDA index for that version
• Same for ROCm base images pointing at the PyTorch ROCm index (when added)
• Consumers should then use plain version pins (no `+cu128` suffix, no `[[tool.uv.index]]` blocks)

*Affected consumer files:*

• `jupyter/pytorch/ubi9-python-3.12/pyproject.toml`
• `jupyter/trustyai/ubi9-python-3.12/pyproject.toml`
• `jupyter/rocm/pytorch/ubi9-python-3.12/pyproject.toml`
• `runtimes/pytorch/ubi9-python-3.12/pyproject.toml`
• `runtimes/rocm-pytorch/ubi9-python-3.12/pyproject.toml`

Parent: #77