[CI] Route mmlu and GB300 MMMU-Pro evals through sgl-eval (#34477)

2026-08-12 19:33:21 -07:00
parent cfdfd31826
commit 50cc1aa241
25 changed files with 182 additions and 503 deletions
@@ -119,6 +119,8 @@ jobs:
          UV_SYSTEM_PYTHON: "1"
        run: |
          uv pip install -e "python[dev]" --index-strategy unsafe-best-match --prerelease allow
+          source scripts/ci/utils/sgl_eval_ref.sh
+          uv pip install "$SGL_EVAL_SPEC" --index-strategy unsafe-best-match

      # Hosted runners are ephemeral, so models are re-fetched every run and the
      # Hub occasionally returns 429s. Persist the HF cache in GitHub's cache