SIGN IN SIGN UP

ci: fix nightly wheel build broken by wheel_base caching (#2110)

## What
Fix the nightly wheel build broken by #1971, and simplify the nightly's
CUDA version handling.

## Why?
PR #1971 added the PR CI's cached wheel_base mechanism to
build-wheel-nightly-matrix.yaml. The nightly builds everything from
scratch on every run and has no pre-built wheel_base image to pull, so
ci-demo's attempt to build/pull
nixl-wheel-base-cu${CUDA_VARIANT}-manylinux_2_28:CI_MANAGED fails
immediately. Additionally, PR #1971 added CUDA_VERSION and CUDA_VARIANT
as explicit JJB parameters and cron arguments, which are redundant since
the CUDA version is already encoded in BASE_TAG (e.g.
13.2.1-devel-ubi8).

## How?

Remove from build-wheel-nightly-matrix.yaml:
  - registry_path at the job level
  - The runs_on_dockers wheel_base entry
  - CI_IMAGE_TAG and WHEEL_BASE_IMAGE_NAME from env
  - --cuda-version and --wheel-base-image from the Build Wheel step

Keep the cuObject additions from #1971 that the nightly needs:
  - MANYLINUX_IMAGE / MANYLINUX_IMAGE_TAG in env
- --manylinux-image, --manylinux-image-tag, --torch-versions in Build
Wheel step
- TORCH_VERSIONS auto-selection based on CUDA major version derived from
BASE_TAG

Simplify proj-jjb.yaml:
- Remove CUDA_VERSION and CUDA_VARIANT JJB params and cron arguments —
BASE_TAG already carries this information
- TORCH_VERSIONS default is now empty (auto-derived) rather than
hardcoded

Extend build-container.sh:
- When building Dockerfile.manylinux, auto-derive CUDA_VERSION from
BASE_IMAGE_TAG if not set explicitly, so callers don't need to pass it
separately

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **New Features**
* Nightly wheel builds now support simplified CUDA 12 and CUDA 13
scheduling.
* CUDA and PyTorch versions can be derived automatically from the
selected base image.
* Builds now create wheels directly from the manylinux Dockerfile
without requiring a prebuilt wheel image.

* **Improvements**
* Removed redundant CUDA configuration options from nightly build
settings.
  * Added support for selecting the base image directly.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Signed-off-by: Adit Ranadive <aranadive@nvidia.com>
A
Adit Ranadive committed
ed15b1ac2cd00ffa3ea454ede9fde1e9db516814
Parent: 6d855fc
Committed by GitHub <noreply@github.com> on 8/18/2026, 12:59:30 PM