ci: add support for vllm sanity testing on Github

[nv-anants]

Overview:

This PR introduces CI support for running vLLM sanity checks within the GitHub Actions pipeline. It integrates a .github/workflows/container-validation-backends.yml workflow, ensuring that at minimal, the vLLM end to end functionality is automatically verified every time on public PR's.

This PR also includes build optimizations based on the container strategy DEP, which result in reducing the build time on Github for dynamo vllm container from 1 hour to 20 minutes. Build optimizations include:

• Restructure Dockerfile.vllm to leverage the Dynamo base container instead of building Dynamo from source in the vllm container. This enables us to decouple the Dynamo and vLLM build from each other.
• Leverage sccache s3 bucket to speed up NIXL, UCX, Dynamo, and vLLM build.
• Add the stage definition, purpose for each stage defined in the framework Dockerfile
• Updated build.sh to pass in new parameters such as --use_sccache, --sccache-bucket, and --sccache-region

Details:

• Restructure Dockerfile.vllm to follow the proposed format in the container strategy: https://github.com/ai-dynamo/enhancements/tree/tusharma/container-strategy
• Add new Github workflow for testing vLLM end to end on a custom Dynamo GPU runner.

Where should the reviewer start?

Check the Github actions workflow: NVIDIA Github Validation / Build and Test - vllm (push)

• .github/workflows/container-validation-backends.yml
• container/Dockerfile.vllm
• container/build.sh
• container/Dockerfile

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• closes: OPS-711, OPS-702

Summary by CodeRabbit

• New Features

  • Optional build caching to speed container builds.
  • ARM64 build support.
  • Streamlined vLLM images with distinct runtime and local-dev targets for faster setup.

• Refactor

  • Reworked container builds to multi-stage, configurable pipelines for improved performance and flexibility.

• Chores

  • Added and updated CI workflows to build/test containers on GPU runners with concurrency controls and junit reporting.

• Tests

  • Promoted an end-to-end vLLM test by removing the “slow” marker, increasing its execution frequency.