[CI][ROCm] Add source-aware filtering for AMD jobs

[haic0]

Summary

• Add per-job source_file_dependencies to every AMD ready, merge, and nightly job.
• Compose job-specific model/test/runtime paths with an amd_ci_shared safety key for Buildkite, ROCm image, dependency, and shared runtime changes.
• Keep malformed, empty, missing, or unknown AMD dependency declarations fail-open while leaving CUDA/NPU validation semantics unchanged.
• Preserve full scheduled main nightly runs while filtering PR-label and post-merge suites.

Stacking and GitHub limitation

This helper PR is stacked on vllm-project#7311 and depends on it. Its base is the exact parent branch, haic0:ci/rocm-amd-uploader-refactor-replacement, so the diff contains only AMD source-filtering changes.

GitHub rejected an upstream stacked PR because a PR in vllm-project/vllm-omni cannot use the fork-owned parent branch as its base (GraphQL: The baseRefName is invalid). Therefore this helper PR lives in haic0/vllm-omni. After vllm-project#7311 merges, rebase this branch onto upstream main and open the follow-up PR in vllm-project/vllm-omni against main.

Dependency policy

Every AMD leaf job combines amd_ci_shared with a focused registry key covering the test target, model family, and relevant runtime paths. amd_ci_shared retains the full selected AMD suite for .buildkite/**, ROCm Docker/image inputs, requirements and packaging files, shared test fixtures, and broad platform/config/runtime files where impact cannot be safely inferred.

Missing declarations already keep jobs. AMD now also treats empty, malformed, unknown, or unavailable dependency mappings as fail-open, logging the reason and retaining the job. Scheduled main nightly builds disable source filtering and keep all nightly jobs.

Why this is not a duplicate

Before implementation, I inspected vllm-project#7311 and its comments, searched open PRs for source_file_dependencies AMD ROCm and AMD source filtering, and searched recent merged PRs in the same area. No open or merged PR adds AMD per-job source filtering. vllm-project#6597 introduced the shared filtering machinery for CUDA/NPU; vllm-project#7311 adds the AMD native uploader lifecycle that this follow-up builds on.

Validation

• .venv/bin/python -m pytest --confcutdir=tests/buildkite tests/buildkite/test_upload_pipeline.py tests/buildkite/test_amd_suite_selection.py tests/buildkite/test_amd_pipeline.py tests/buildkite/test_skip_ci.py -q — 133 passed.
• PATH="$PWD/.venv/bin:$HOME/.cargo/bin:$PATH" uvx pre-commit run --files <changed files> — all applicable hooks passed, including Ruff, mypy, YAML, and Buildkite validation.
• Filtered AMD ready selection rendered through minijinja-cli and parsed as valid YAML; the Qwen3-TTS change retained the mapped TTS job (plus the conservative broad CPU lane), omitted Wan22, and leaked no uploader-only dependency fields.
• git diff --check passed.
• No GPU/model evaluation was run or required because this changes CI selection only, not model output.

AI assistance

AI assistance was used for repository investigation, dependency-map drafting, implementation, tests, and validation. The human submitter must review and understand every changed line before merge.

Follow-up

After vllm-project#7311 merges, rebase this branch onto upstream main, open an upstream PR against main, and rerun the focused Buildkite validation against the merged uploader lifecycle.