ci: pin the official vLLM main commit image

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

docker/Dockerfile.ci

@@ -1,10 +1,11 @@
-# Target vLLM version: v0.30.0 (tag @ ced6857afa0e).
-# Use the matching release image and its coherent CUDA/PyTorch/FlashInfer
-# runtime. Do not reinstall a commit wheel or repair dependencies on top of
-# an older base image at a release boundary. For a future unreleased target,
-# restore an explicit wheel pin and validate its complete runtime stack.
+# Target an unreleased vLLM main commit. The official commit image carries
+# the matching CUDA/PyTorch/FlashInfer stack; do not overlay a nightly wheel
+# on the previous release image. Copilot updates this commit pin, which also
+# selects the image used by CI. The selected commit must have a published
+# nightly image as well as a cu130 wheel.
+ARG VLLM_PRECOMPILED_WHEEL_COMMIT=ac9126e58aa7bbab1856ba6593ba4d5003fea516
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.30.0
+ARG VLLM_BASE_TAG=nightly-${VLLM_PRECOMPILED_WHEEL_COMMIT}
 
 # We use registry-based layer caching for the dependencies,
 # where the tag is determined based on the hash of the contents
@@ -12,6 +13,7 @@ ARG VLLM_BASE_TAG=v0.30.0
 # everything except the actual project dir, since that should not
 # change much across branches.
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG} AS deps
+ARG VLLM_PRECOMPILED_WHEEL_COMMIT
 # Keep APP_DIR outside /workspace so k8s Buildkite workspace mounts cannot
 # shadow the editable install target (#5364). Docker WORKDIR creates /opt if
 # needed.
@@ -42,6 +44,9 @@ ENV SETUPTOOLS_SCM_PRETEND_VERSION=0.0.0
 # Do NOT use a non-editable install here — that copies an empty stub into
 # site-packages and breaks ``import vllm_omni.entrypoints``.
 RUN uv pip install --system -e ".[dev]"
+# Fail the image build if dependency resolution replaced the selected vLLM
+# or its compiled CUDA runtime no longer imports.
+RUN python3 -c "import vllm, torch, flashinfer; import vllm._C_stable_libtorch; assert '${VLLM_PRECOMPILED_WHEEL_COMMIT}'[:8] in vllm.__version__, vllm.__version__"
 
 # The final stage just copies the working dir on top of the
 # dependencies image. This uses --link so that we can skip

tests/buildkite/test_rocm_dockerfile.py

@@ -10,7 +10,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 REPO_ROOT = Path(__file__).resolve().parents[2]
 AMD_TEMPLATE = REPO_ROOT / ".buildkite/amd/test-template-amd-omni.j2"
 AMD_BUILD_SCRIPT = REPO_ROOT / ".buildkite/amd/scripts/build-ci-image.sh"
-CI_DOCKERFILE = REPO_ROOT / "docker/Dockerfile.ci"
+CUDA_RELEASE_DOCKERFILE = REPO_ROOT / "docker/Dockerfile.cuda"
 ROCM_DOCKERFILE = REPO_ROOT / "docker/Dockerfile.rocm"
 ROCM_DOCKERIGNORE = REPO_ROOT / "docker/Dockerfile.rocm.dockerignore"
 
@@ -30,14 +30,17 @@ def _line_index(lines: list[str], prefix: str) -> int:
     return matches[0]
 
 
-def test_rocm_base_tracks_ci_vllm_release() -> None:
-    ci_release = _docker_arg(CI_DOCKERFILE, "VLLM_BASE_TAG")
+def test_rocm_base_tracks_cuda_vllm_release() -> None:
+    # CUDA CI can target a main-branch nightly independently of the stable
+    # release images used by the CUDA and ROCm distribution Dockerfiles.
+    cuda_base = _docker_arg(CUDA_RELEASE_DOCKERFILE, "BASE_IMAGE")
+    _, _, cuda_release = cuda_base.rpartition(":")
     rocm_base = _docker_arg(ROCM_DOCKERFILE, "BASE_IMAGE")
 
     image_ref, separator, image_tag = rocm_base.rpartition(":")
     assert separator, f"expected a tagged ROCm base image, got {rocm_base}"
     assert image_ref.rsplit("/", 1)[-1] == "vllm-openai-rocm", image_ref
-    assert image_tag == ci_release
+    assert image_tag == cuda_release
 
 
 def test_rocm_defaults_to_prebuilt_base_image() -> None:
@@ -103,11 +106,12 @@ def test_rocm_build_context_excludes_git_history() -> None:
     assert ".git" in patterns
 
 
-def test_rocm_source_ref_tracks_ci_vllm_release() -> None:
-    ci_release = _docker_arg(CI_DOCKERFILE, "VLLM_BASE_TAG")
+def test_rocm_source_ref_tracks_cuda_vllm_release() -> None:
+    cuda_base = _docker_arg(CUDA_RELEASE_DOCKERFILE, "BASE_IMAGE")
+    _, _, cuda_release = cuda_base.rpartition(":")
     rocm_source_ref = _docker_arg(ROCM_DOCKERFILE, "VLLM_VERSION_OR_COMMIT_HASH")
 
-    assert rocm_source_ref == ci_release
+    assert rocm_source_ref == cuda_release
 
 
 def test_rocm_dockerfile_contains_vllm_api_canary() -> None: