[CI] Add CUDA 13 nightly containers

[csahithi]

Purpose

• This PR adds CUDA 13 nightly containers to dockerhub along with automatic 14 day cleanup (CUDA 13 wheels already exist in release-pipeline.yaml).
• The cleanup script is modified to accept a prefix as a parameter using which cleanup is performed.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[gemini-code-assist]

Code Review

This pull request introduces support for building and publishing nightly Docker containers for CUDA 13. This is achieved by adding new steps to the Buildkite release pipeline for building CUDA 13 images for x86 and arm64, creating a multi-arch manifest, and publishing them to DockerHub. The cleanup-nightly-builds.sh script has also been updated to accept a tag prefix, making it reusable for cleaning up different sets of nightly images.

The changes are logical and follow the existing structure of the pipeline. However, I have identified two high-severity issues in the pipeline configuration file:

1. An invalid CUDA compute capability is specified in the torch_cuda_arch_list for the arm64 build.
2. There is significant code duplication in the pipeline steps, which harms maintainability. I've suggested using YAML anchors to refactor this.

Please see the detailed comments for suggestions on how to address these points.

[gemini-code-assist]

The torch_cuda_arch_list contains 12.0, which is not a valid CUDA compute capability. The latest defined architecture is 9.0 for Hopper. While 10.0+PTX can be for forward-compatibility with upcoming architectures like Blackwell, 12.0 is likely a mistake. At best, it will be ignored by the build system, but it should be removed to avoid confusion and potential issues.

      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cuda13-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."

[wangshangsam]

12.0 is for RTX PRO 6000 Blackwell, no? There's also 12.1 for DGX Spark.

Speaking of which, @csahithi I think it probably make sense to add 12.1 here. I remember I had a late night discussion with @suhara and he was having troubles building an image on DGX Spark. Therefore, probably make sense to include 12.1 here so that our Nemotron folks will be able to just use the nightly built container directly on DGX Spark.

[gemini-code-assist]

This new step to publish nightly CUDA 13.0 images is almost a complete duplicate of the existing step for CUDA 12.9 images. This level of duplication makes the pipeline configuration difficult to maintain and prone to errors, as changes need to be manually synchronized across multiple blocks.

To improve maintainability, I strongly recommend refactoring this using YAML anchors and aliases. You can define a template for the common parts of the job and then reuse it for each CUDA version, only overriding the specific parts like dependencies and tag prefixes.

Here is a conceptual example:

.publish_nightly_template: &publish_nightly_template
  if: build.env("NIGHTLY") == "1"
  agents:
    queue: cpu_queue_postmerge
  plugins:
    - docker-login#v3.0.0:
        username: vllmbot
        password-env: DOCKERHUB_TOKEN
  # ... other common properties

- label: "Build and publish nightly CUDA 12.9 ..."
  <<: *publish_nightly_template
  depends_on:
    - create-multi-arch-manifest
  commands:
    # ... commands with version-specific tags
    - "bash .buildkite/scripts/cleanup-nightly-builds.sh nightly-"

- label: "Build and publish nightly CUDA 13.0 ..."
  <<: *publish_nightly_template
  depends_on:
    - create-multi-arch-manifest-cuda13
  commands:
    # ... commands with version-specific tags
    - "bash .buildkite/scripts/cleanup-nightly-builds.sh cuda13-nightly-"

Even the commands list could be further parameterized using variables to reduce duplication even more. Adopting this pattern will make the pipeline much cleaner and easier to manage.

[wangshangsam]

This is a fair comment and I think it'd be a good idea to refactor them, if @csahithi wanna take on this challenge.

[wangshangsam]

Nicely done, @csahithi ! If we can merge this before Feb 13, maybe we could use a nightly image directly in our MLPerf submission :)

The pipeline should be tested by someone who has access to vLLM's buildkite and ecr account, since this is very infra-setup-dependent. @mgoin do you know whom would be able to help test this out?

[wangshangsam]

Suggested change

- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cuda13-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."

- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cuda13-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."

FLASHINFER_AOT_COMPILE no longer exists I think.

[wangshangsam]

12.0 is for RTX PRO 6000 Blackwell, no? There's also 12.1 for DGX Spark.

Speaking of which, @csahithi I think it probably make sense to add 12.1 here. I remember I had a late night discussion with @suhara and he was having troubles building an image on DGX Spark. Therefore, probably make sense to include 12.1 here so that our Nemotron folks will be able to just use the nightly built container directly on DGX Spark.

[wangshangsam]

Suggested change

- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cuda13-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."

- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cuda13-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."