Bump actions/setup-python from 6.1.0 to 6.3.0

Bumps [actions/setup-python](https://github.com/actions/setup-python) from 6.1.0 to 6.3.0. - [Release notes](https://github.com/actions/setup-python/releases) - [Commits](https://github.com/actions/setup-python/compare/83679a892e2d95755f2dac6acb0bfd1e9ac5d548...ece7cb06caefa5fff74198d8649806c4678c61a1) --- updated-dependencies: - dependency-name: actions/setup-python dependency-version: 6.3.0 dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com>
2026-06-30 12:18:49 +00:00
2441 changed files with 45629 additions and 229205 deletions
@@ -1,103 +0,0 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Audit vLLM compiled libraries for PyTorch stable ABI compliance."""
 import fnmatch
 import sys
 from pathlib import Path
 from torch_abi_audit import inspect_package
 from torch_abi_audit.report import ExtensionReport, PackageReport
 # Temporary allowlist of extensions not yet on the stable ABI.
 # Shrink and remove over time.
 ALLOWED_UNSTABLE_LIBRARIES: tuple[str, ...] = (
    "_flashkda_C.abi3.so",
    "vllm_flash_attn/_vllm_fa2_C.abi3.so",
    "vllm_flash_attn/_vllm_fa3_C.abi3.so",
    "third_party/deep_gemm/_C*.so",
 )
 def _relative_path(lib: ExtensionReport, package_root: Path) -> str:
    try:
        return lib.path.relative_to(package_root).as_posix()
    except ValueError:
        return lib.path.name
 def _is_torch_unstable(lib: ExtensionReport) -> bool:
    return lib.error is None and lib.torch.uses_torch and not lib.torch.stable
 def _matches_allowlist(rel_path: str, patterns: tuple[str, ...]) -> bool:
    return any(fnmatch.fnmatch(rel_path, pattern) for pattern in patterns)
 def _iter_libs(report: PackageReport) -> tuple[ExtensionReport, ...]:
    return (*report.extensions, *report.bundled_libs)
 def _collect_unstable(report: PackageReport) -> list[str]:
    return sorted(
        _relative_path(lib, report.root)
        for lib in _iter_libs(report)
        if _is_torch_unstable(lib)
    )
 def _find_stale_allowlist_entries(
    report: PackageReport, patterns: tuple[str, ...]
 ) -> list[str]:
    """Allowlist patterns that match a built library which is no longer unstable."""
    stale: list[str] = []
    for pattern in patterns:
        for lib in _iter_libs(report):
            if lib.error is not None:
                continue
            if not fnmatch.fnmatch(_relative_path(lib, report.root), pattern):
                continue
            if not _is_torch_unstable(lib):
                stale.append(pattern)
                break
    return stale
 def check_torch_abi(
    package: str = "vllm",
    patterns: tuple[str, ...] = ALLOWED_UNSTABLE_LIBRARIES,
 ) -> int:
    report = inspect_package(package)
    if report.error:
        print(f"error: failed to inspect {package!r}: {report.error}", file=sys.stderr)
        return 2
    unstable = _collect_unstable(report)
    unexpected = [
        rel_path for rel_path in unstable if not _matches_allowlist(rel_path, patterns)
    ]
    stale = _find_stale_allowlist_entries(report, patterns)
    if unexpected or stale:
        if unexpected:
            print(
                "Not allowed: torch-unstable libraries outside "
                f"ALLOWED_UNSTABLE_LIBRARIES: {', '.join(unexpected)}",
                file=sys.stderr,
            )
        if stale:
            print(
                "Not allowed: stale ALLOWED_UNSTABLE_LIBRARIES entries: "
                f"{', '.join(stale)}",
                file=sys.stderr,
            )
        return 1
    print("Torch stable ABI check passed.")
    return 0
 if __name__ == "__main__":
    print(">>> Auditing vLLM extension modules for PyTorch stable ABI compliance")
    sys.exit(check_torch_abi())
@@ -14,7 +14,6 @@ run_all_patterns:
  - "setup.py"
  - "csrc/"
  - "cmake/"
  - ".buildkite/check-torch-abi.py"
 run_all_exclude_patterns:
  - "docker/Dockerfile."
  - "csrc/cpu/"
@@ -3,7 +3,6 @@ job_dirs:
  - ".buildkite/intel_jobs"
 run_all_patterns:
  - ".buildkite/ci_config_intel.yaml"
  - ".buildkite/scripts/hardware_ci/run-intel-test.sh"
  - "docker/Dockerfile"
  - "docker/Dockerfile.xpu"
  - "CMakeLists.txt"
@@ -8,7 +8,6 @@ run_all_patterns:
  - "docker/docker-bake-rocm.hcl"
  - ".buildkite/hardware_tests/amd.yaml"
  - ".buildkite/scripts/ci-bake-rocm.sh"
  - ".buildkite/scripts/rocm/"
  - ".buildkite/scripts/hardware_ci/run-amd-test.py"
  - ".buildkite/scripts/hardware_ci/run-amd-test.sh"
  - "CMakeLists.txt"
@@ -1,30 +1,5 @@
 group: Hardware - AMD Build
 # ROCm image flow:
 # 1. Refresh the long-lived ROCm base image only when Dockerfile.rocm_base changes.
 # 2. Build ci_base from either the stable base or the freshly refreshed base.
 # 3. Build the per-commit ROCm CI image and smoke-test it before GPU jobs run.
 steps:
  - label: "AMD: :docker: refresh ROCm base"
    key: refresh-rocm-base-amd
    depends_on: []
    device: amd_cpu
    no_plugin: true
    commands:
      - bash .buildkite/scripts/rocm/refresh-base-image.sh
    env:
      DOCKER_BUILDKIT: "1"
      BUILDKIT_PROGRESS: "tty"
      TERM: "xterm-256color"
    retry:
      automatic:
        - exit_status: 1  # Transient Docker/BuildKit failure
          limit: 1
        - exit_status: -1  # Agent was lost
          limit: 1
        - exit_status: -10  # Agent was lost
          limit: 1
  # Ensure ci_base is up-to-date before building the test image.
  # Compares a content hash of ci_base-affecting files against the remote
  # image label. If hashes match the build is skipped (< 30 s); if they
@@ -32,24 +7,19 @@ steps:
  - label: "AMD: :docker: ensure ci_base"
    key: ensure-ci-base-amd
    soft_fail: false
-    depends_on:
+    depends_on: []
      - refresh-rocm-base-amd
    device: amd_cpu
    no_plugin: true
    commands:
-      - bash .buildkite/scripts/rocm/build-ci-base.sh
+      - bash .buildkite/scripts/ci-bake-rocm.sh ci-base-rocm-ci-with-deps
    env:
      DOCKER_BUILDKIT: "1"
      BUILDKIT_PROGRESS: "tty"
      TERM: "xterm-256color"
      VLLM_BAKE_FILE: "docker/docker-bake-rocm.hcl"
      PYTORCH_ROCM_ARCH: "gfx90a;gfx942;gfx950"
      REMOTE_VLLM: "1"
      VLLM_BRANCH: "$BUILDKITE_COMMIT"
    retry:
      automatic:
        - exit_status: 1  # Transient Docker/BuildKit failure
          limit: 1
        - exit_status: -1  # Agent was lost
          limit: 1
        - exit_status: -10  # Agent was lost
@@ -63,12 +33,35 @@ steps:
    device: amd_cpu
    no_plugin: true
    commands:
-      - bash .buildkite/scripts/rocm/build-test-image.sh
+      - |
-      - bash .buildkite/scripts/rocm/smoke-test-image.sh
+        if [[ "${ROCM_CI_ARTIFACT_ONLY:-0}" == "1" ]]; then
          echo "ROCM_CI_ARTIFACT_ONLY=1; building ROCm wheel artifact only"
          IMAGE_TAG="" bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-artifacts
        else
          bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-wheel
        fi
      - |
        docker run --rm --network=none --entrypoint /bin/bash "rocm/vllm-ci:${BUILDKITE_COMMIT}" -ec '
          if [ ! -d /vllm-workspace ]; then echo Missing directory: /vllm-workspace >&2; exit 1; fi
          if [ ! -d /vllm-workspace/tests ]; then echo Missing directory: /vllm-workspace/tests >&2; exit 1; fi
          if [ ! -d /vllm-workspace/src/vllm ]; then echo Missing directory: /vllm-workspace/src/vllm >&2; exit 1; fi
          if [ ! -x /vllm-workspace/src/vllm/vllm-rs ]; then echo Missing executable: /vllm-workspace/src/vllm/vllm-rs >&2; exit 1; fi
          command -v python3
          command -v uv
          command -v pytest
          if ! command -v amd-smi >/dev/null 2>&1 && ! command -v rocminfo >/dev/null 2>&1; then
            echo No ROCm CLI found in image >&2
            exit 1
          fi
          python3 - <<PY
        import torch, vllm
        print(torch.__version__)
        print(vllm.__version__)
        PY
          echo AMD image smoke OK
        '
    env:
      DOCKER_BUILDKIT: "1"
      BUILDKIT_PROGRESS: "tty"
      TERM: "xterm-256color"
      VLLM_BAKE_FILE: "docker/docker-bake-rocm.hcl"
      PYTORCH_ROCM_ARCH: "gfx90a;gfx942;gfx950"
      IMAGE_TAG: "rocm/vllm-ci:$BUILDKITE_COMMIT"
@@ -76,8 +69,6 @@ steps:
      VLLM_BRANCH: "$BUILDKITE_COMMIT"
    retry:
      automatic:
        - exit_status: 1  # Transient Docker/BuildKit failure
          limit: 1
        - exit_status: -1  # Agent was lost
          limit: 1
        - exit_status: -10  # Agent was lost
@@ -17,22 +17,16 @@ steps:
  - tests/kernels/test_awq_int4_to_int8.py
  - tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
  - tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
  - tests/kernels/mamba/test_cpu_short_conv.py
  - tests/kernels/mamba/test_causal_conv1d.py
  - tests/kernels/mamba/test_mamba_ssm.py
  commands:
    - |
      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
      pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
      pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
      pytest -x -v -s tests/kernels/moe/test_cpu_quant_fused_moe.py
      pytest -x -v -s tests/kernels/mamba/test_cpu_short_conv.py
      pytest -x -v -s tests/kernels/test_onednn.py
      pytest -x -v -s tests/kernels/test_awq_int4_to_int8.py
      pytest -x -v -s tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
-      pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
+      pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py"
      pytest -x -v -s tests/kernels/mamba/test_causal_conv1d.py
      pytest -x -v -s tests/kernels/mamba/test_mamba_ssm.py"
 # Note: SDE can't be downloaded from CI host because of AWS WAF
 # - label: CPU-Compatibility Tests
@@ -145,22 +139,9 @@ steps:
  commands:
    - |
      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
-      pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py --ignore=tests/models/multimodal/generation/test_qwen2_5_vl.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+      pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
  parallelism: 4
 - label: CPU-Qwen2.5-VL Multimodal Tests
  depends_on: []
  device: intel_cpu
  no_plugin: true
  source_file_dependencies:
  # - vllm/
  - vllm/model_executor/layers/rotary_embedding
  - tests/models/multimodal/generation/
  commands:
    - |
      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 40m "
      VLLM_CI_ENV=0 pytest -x -v -s tests/models/multimodal/generation/test_qwen2_5_vl.py"
 - label: "Arm CPU Test"
  depends_on: []
  soft_fail: false
@@ -18,7 +18,7 @@ steps:
  - label: "XPU example Test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 50
+    timeout_in_minutes: 30
    optional: true
    device: intel_gpu
    agent_tags:
@@ -39,13 +39,13 @@ steps:
  - label: "XPU V1 test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 70
+    timeout_in_minutes: 30
    optional: true
    device: intel_gpu
    agent_tags:
      label: production
      gpu: 1+
-      mem: 24+
+      mem: 16+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -60,7 +60,7 @@ steps:
  - label: "XPU server test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 45
+    timeout_in_minutes: 30
    optional: true
    device: intel_gpu
    agent_tags:
@@ -79,18 +79,12 @@ setup_buildx_builder() {
    docker buildx ls | grep -E '^\*|^NAME' || docker buildx ls
 }
 annotate_image_tags() {
    .buildkite/scripts/annotate-image-build.sh \
        "${IMAGE_TAG:-}" "${IMAGE_TAG_LATEST:-}"
 }
 check_and_skip_if_image_exists() {
    if [[ -n "${IMAGE_TAG:-}" ]]; then
        echo "--- :mag: Checking if image exists"
        if docker manifest inspect "${IMAGE_TAG}" >/dev/null 2>&1; then
            echo "Image already exists: ${IMAGE_TAG}"
            echo "Skipping build"
            annotate_image_tags
            exit 0
        fi
        echo "Image not found, proceeding with build"
@@ -177,18 +171,6 @@ BRANCH=$4
 IMAGE_TAG=$5
 IMAGE_TAG_LATEST=${6:-} # only used for main branch, optional
 # When TORCH_NIGHTLY=1, build the base CI image against PyTorch nightly so the
 # entire existing pipeline runs on nightly torch (CUDA/GPU lane only). Delegate
 # to the dedicated nightly build (PYTORCH_NIGHTLY=1, CUDA 13.0) and tag it at the
 # normal IMAGE_TAG that every test step already pulls -- no separate image tag,
 # no duplicate "vLLM Against PyTorch Nightly" pipeline section.
 if [[ "${TORCH_NIGHTLY:-0}" == "1" ]]; then
    echo "--- :warning: TORCH_NIGHTLY=1 -- building base image on PyTorch nightly"
    SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
    exec "${SCRIPT_DIR}/image_build_torch_nightly.sh" \
        "${REGISTRY}" "${REPO}" "${BUILDKITE_COMMIT}" "${BRANCH}" "${IMAGE_TAG}"
 fi
 # build config
 TARGET="test-ci"
 VLLM_BAKE_FILE_PATH="${VLLM_BAKE_FILE_PATH:-docker/docker-bake.hcl}"
@@ -272,5 +254,3 @@ echo "--- :docker: Building ${TARGET}"
 docker --debug buildx bake -f "${VLLM_BAKE_FILE_PATH}" -f "${CI_HCL_PATH}" --progress plain "${TARGET}"
 echo "--- :white_check_mark: Build complete"
 annotate_image_tags
@@ -9,31 +9,30 @@ fi
 REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 IMAGE="$REGISTRY/$REPO:$BUILDKITE_COMMIT-arm64"
 # authenticate with AWS ECR
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
 # skip build if image already exists
-if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64) ]]; then
  echo "Image found"
 else
  echo "Image not found, proceeding with build..."
-  # build for arm64 GPU targets: Grace/GH200 (sm_90),
+else
-  # Blackwell/Thor (sm_100/sm_103/sm_110), and DGX Spark/GB10
+  echo "Image found"
-  # (sm_121, family-covered by 12.0 under CUDA 13)
+  exit 0
  docker build --file docker/Dockerfile \
    --platform linux/arm64 \
    --build-arg max_jobs=16 \
    --build-arg nvcc_threads=4 \
    --build-arg torch_cuda_arch_list="9.0 10.0 11.0 12.0" \
    --build-arg USE_SCCACHE=1 \
    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
    --tag "$IMAGE" \
    --target test \
    --progress plain .
  # push
  docker push "$IMAGE"
 fi
-.buildkite/scripts/annotate-image-build.sh "$IMAGE"
+# build for arm64 GPU targets: Grace/GH200 (sm_90) and DGX Spark/GB10
 # (sm_121, family-covered by 12.0 under CUDA 13)
 docker build --file docker/Dockerfile \
  --platform linux/arm64 \
  --build-arg max_jobs=16 \
  --build-arg nvcc_threads=4 \
  --build-arg torch_cuda_arch_list="9.0 12.0" \
  --build-arg USE_SCCACHE=1 \
  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64 \
  --target test \
  --progress plain .
 # push
 docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64
@@ -9,26 +9,26 @@ fi
 REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 IMAGE="$REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu"
 # authenticate with AWS ECR
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
 # skip build if image already exists
-if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
  echo "Image found"
 else
  echo "Image not found, proceeding with build..."
-  # build
+else
-  docker build --file docker/Dockerfile.cpu \
+  echo "Image found"
-    --build-arg max_jobs=16 \
+  exit 0
    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
    --build-arg VLLM_CPU_X86=true \
    --tag "$IMAGE" \
    --target vllm-test \
    --progress plain .
  # push
  docker push "$IMAGE"
 fi
-.buildkite/scripts/annotate-image-build.sh "$IMAGE"
+# build
 docker build --file docker/Dockerfile.cpu \
  --build-arg max_jobs=16 \
  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
  --build-arg VLLM_CPU_X86=true \
  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
  --target vllm-test \
  --progress plain .
 # push
 docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu
@@ -9,25 +9,25 @@ fi
 REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 IMAGE="$REGISTRY/$REPO:$BUILDKITE_COMMIT-arm64-cpu"
 # authenticate with AWS ECR
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
 # skip build if image already exists
-if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu) ]]; then
  echo "Image found"
 else
  echo "Image not found, proceeding with build..."
-  # build
+else
-  docker build --file docker/Dockerfile.cpu \
+  echo "Image found"
-    --build-arg max_jobs=16 \
+  exit 0
    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
    --tag "$IMAGE" \
    --target vllm-test \
    --progress plain .
  # push
  docker push "$IMAGE"
 fi
-.buildkite/scripts/annotate-image-build.sh "$IMAGE"
+# build
 docker build --file docker/Dockerfile.cpu \
  --build-arg max_jobs=16 \
  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu \
  --target vllm-test \
  --progress plain .
 # push
 docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu
@@ -9,26 +9,26 @@ fi
 REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 IMAGE="$REGISTRY/$REPO:$BUILDKITE_COMMIT-hpu"
 # authenticate with AWS ECR
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
 # skip build if image already exists
-if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu) ]]; then
  echo "Image found"
 else
  echo "Image not found, proceeding with build..."
-  # build
+else
-  docker build \
+  echo "Image found"
-    --file tests/pytorch_ci_hud_benchmark/Dockerfile.hpu \
+  exit 0
    --build-arg max_jobs=16 \
    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
    --tag "$IMAGE" \
    --progress plain \
    https://github.com/vllm-project/vllm-gaudi.git
  # push
  docker push "$IMAGE"
 fi
-.buildkite/scripts/annotate-image-build.sh "$IMAGE"
+# build
 docker build \
  --file tests/pytorch_ci_hud_benchmark/Dockerfile.hpu \
  --build-arg max_jobs=16 \
  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu \
  --progress plain \
  https://github.com/vllm-project/vllm-gaudi.git
 # push
 docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu
@@ -40,7 +40,6 @@ docker buildx ls
 echo "--- :mag: Checking if image already exists"
 if docker manifest inspect "$IMAGE_TAG" >/dev/null 2>&1; then
  echo "Image found: $IMAGE_TAG — skipping build"
  .buildkite/scripts/annotate-image-build.sh "$IMAGE_TAG"
  exit 0
 fi
 echo "Image not found, proceeding with build..."
@@ -67,5 +66,3 @@ docker buildx build --file docker/Dockerfile \
  --progress plain .
 echo "--- :white_check_mark: Torch nightly image build complete: $IMAGE_TAG"
 .buildkite/scripts/annotate-image-build.sh "$IMAGE_TAG"
@@ -9,26 +9,26 @@ fi
 REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 IMAGE="$REGISTRY/$REPO:$BUILDKITE_COMMIT-xpu"
 # authenticate with AWS ECR
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY" || true
 aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin 936637512419.dkr.ecr.us-east-1.amazonaws.com || true
 # skip build if image already exists
-if docker manifest inspect "$IMAGE" &> /dev/null; then
+if ! docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-xpu &> /dev/null; then
  echo "Image found"
 else
  echo "Image not found, proceeding with build..."
-  # build
+else
-  docker build \
+  echo "Image found"
-    --file docker/Dockerfile.xpu \
+  exit 0
    --build-arg max_jobs=16 \
    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
    --tag "$IMAGE" \
    --progress plain .
  # push
  docker push "$IMAGE"
 fi
-.buildkite/scripts/annotate-image-build.sh "$IMAGE"
+# build
 docker build \
  --file docker/Dockerfile.xpu \
  --build-arg max_jobs=16 \
  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-xpu \
  --progress plain .
 # push
 docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-xpu
@@ -3,7 +3,7 @@ depends_on:
  - image-build-xpu
 steps:
 - label: XPU Sleep Mode
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: intel_gpu
  agent_tags:
    label: production
@@ -1,26 +0,0 @@
 group: Benchmarks
 depends_on: 
  - image-build-xpu
 steps:
 - label: Benchmarks CLI Test
  key: benchmarks-cli-test
  timeout_in_minutes: 40
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
  - vllm/
  - tests/benchmarks/
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      pytest -v -s benchmarks/'
@@ -2,44 +2,6 @@ group: Engine Intel
 depends_on:
  - image-build-xpu
 steps:
 - label: Engine
  key: engine
  timeout_in_minutes: 40
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
  - vllm/compilation/
  - vllm/config/
  - vllm/engine/
  - vllm/entrypoints/logger.py
  - vllm/envs.py
  - vllm/logger.py
  - vllm/logging_utils/
  - vllm/platforms/
  - vllm/sequence.py
  - vllm/triton_utils/
  - vllm/utils/
  - tests/engine
  - tests/test_sequence
  - tests/test_config
  - tests/test_logger
  - tests/test_vllm_port
  - tests/test_jit_monitor.py
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      pytest -v -s engine/test_arg_utils.py test_sequence.py test_logger.py test_vllm_port.py test_jit_monitor.py'
 - label: Engine (1 GPU)
  timeout_in_minutes: 30
  device: intel_gpu
@@ -61,41 +23,3 @@ steps:
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py'
 - label: V1 e2e (2 GPUs)
  timeout_in_minutes: 30
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 2+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
    - vllm/compilation/
    - vllm/config/
    - vllm/distributed/
    - vllm/engine/
    - vllm/envs.py
    - vllm/forward_context.py
    - vllm/inputs/
    - vllm/logger.py
    - vllm/logging_utils/
    - vllm/model_executor/
    - vllm/multimodal/
    - vllm/platforms/
    - vllm/sampling_params.py
    - vllm/transformers_utils/
    - vllm/triton_utils/
    - vllm/utils/
    - vllm/v1/
    - tests/v1/e2e/spec_decode
  commands:
  - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"'
@@ -81,12 +81,10 @@ steps:
      'cd tests &&
      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
      set -o pipefail &&
-      pytest -v -s lora/test_punica_ops.py::test_kernels &&
+      pytest -v -s lora/test_punica_ops.py --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[expand-0-xpu:0-dtype0-3-43264-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype1-1-2049-64-128-16]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-128-1-32]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-256-1-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-256-8-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels[expand-0-xpu:0-dtype0-3-2049-128-8-16]" --deselect="tests/lora/test_punica_ops.py::test_kernels[shrink-0-xpu:0-dtype0-1-2049-128-8-32]" --deselect="tests/lora/test_punica_ops.py::test_kernels[expand-0-xpu:0-dtype1-1-2049-256-128-32]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype0-3-64256-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype1-2-29696-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype1-3-49408-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[shrink-0-xpu:0-dtype0-2-16384-32-4-4]" --deselect="tests/lora/test_punica_ops.py::test_kernels_hidden_size[expand-0-xpu:0-dtype0-2-51328-32-4-4]"'
      pytest -v -s lora/test_punica_ops.py::test_kernels_hidden_size &&
      pytest -v -s lora/test_punica_ops.py::test_add_lora_fused_moe_early_exit'
 - label: LoRA Punica FP8/XPU Ops
-  timeout_in_minutes: 60
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
@@ -3,7 +3,7 @@ depends_on:
  - image-build-xpu
 steps:
 - label: V1 Core + KV + Metrics
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: intel_gpu
  agent_tags:
    label: production
@@ -33,12 +33,12 @@ steps:
      pytest -v -s v1/executor'
 - label: V1 Sample + Logits
-  timeout_in_minutes: 90
+  timeout_in_minutes: 30
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
-    mem: 24+
+    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -72,31 +72,9 @@ steps:
      pytest -v -s v1/test_oracle.py &&
      pytest -v -s v1/test_request.py &&
      pytest -v -s v1/test_outputs.py &&
-      pytest -v -s v1/sample'
+      pytest -v -s v1/sample/test_topk_topp_sampler.py &&
-
+      pytest -v -s v1/sample/test_logprobs.py &&
- label: Basic Models Tests (Initialization)
+      pytest -v -s v1/sample/test_logprobs_e2e.py'
  timeout_in_minutes: 60
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
    - vllm/
    - tests/models/test_initialization.py
    - tests/models/registry.py
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'export VLLM_XPU_FUSED_MOE_USE_REF=1 &&
      cd tests &&
      pytest -v -s models/test_initialization.py::test_can_initialize_large_subset[Eagle3MiniMaxM2ForCausalLM]'
 - label: XPU CPU Offload
  timeout_in_minutes: 60
@@ -125,37 +103,9 @@ steps:
      pytest -v -s v1/kv_offload &&
      pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'
 - label: NixlConnector PD accuracy (4 GPUs)
  timeout_in_minutes: 60
  num_devices: 4
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 4+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
    - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
    - vllm/v1/worker/kv_connector_model_runner_mixin.py
    - tests/v1/kv_connector/nixl_integration/
    - vllm/platforms/xpu.py
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=1 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
      PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'
 - label: Regression
  key: regression
-  timeout_in_minutes: 50
+  timeout_in_minutes: 30
  device: intel_gpu
  agent_tags:
    label: production
@@ -183,13 +133,13 @@ steps:
  commands:
  - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'pip install modelscope\<1.38 &&
+      'pip install modelscope &&
      cd tests &&
      pytest -v -s test_regression.py'
 - label: Metrics, Tracing (2 GPUs)
  key: metrics-tracing-2-gpus
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  num_devices: 2
  device: intel_gpu
  agent_tags:
@@ -225,7 +175,7 @@ steps:
 - label: Async Engine, Inputs, Utils, Worker
  key: async-engine-inputs-utils-worker
-  timeout_in_minutes: 55
+  timeout_in_minutes: 30
  device: intel_gpu
  agent_tags:
    label: production
@@ -262,25 +212,3 @@ steps:
    pytest -v -s detokenizer &&
    pytest -v -s -m "not cpu_test" ./multimodal &&
    pytest -v -s utils_ --ignore=utils_/test_mem_utils.py'
 - label: Fusion Unit Tests
  timeout_in_minutes: 30
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
    - vllm/compilation/
    - tests/compile/passes/test_qk_norm_rope_fusion.py
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      pytest -v -s compile/passes/test_qk_norm_rope_fusion.py'
@@ -1,33 +0,0 @@
 group: Model Executor Intel
 depends_on:
  - image-build-xpu
 steps:
 - label: Model Executor (Intel)
  key: model-executor-intel
  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
  - vllm/engine/arg_utils.py
  - vllm/config/model.py
  - vllm/model_executor
  - tests/model_executor
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'apt-get update && apt-get install -y curl libsodium23 &&
      pip3 install tensorizer==2.10.1 &&
      pip3 install runai-model-streamer[s3,gcs,azure]\>=0.15.7 &&
      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
      export PYTHONFAULTHANDLER=1 &&
      cd tests &&
      pytest -v -s model_executor -m "not slow_test" --ignore="model_executor/layers/test_rocm_unquantized_gemm.py" --deselect="tests/model_executor/model_loader/test_reload.py::test_kv_scale_reload"'
@@ -8,7 +8,7 @@ steps:
  agent_tags:
    label: production
    gpu: 2+
-    mem: 24+
+    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -28,9 +28,7 @@ steps:
      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
      cd tests &&
      pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics" &&
      pytest -v -s v1/e2e/general/test_context_length.py &&
      ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram" &&
      pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0" &&
      pytest -v -s v1/e2e/general/test_min_tokens.py'
 - label: Model Runner V2 Examples (Intel)
@@ -62,55 +60,3 @@ steps:
      python3 basic/offline_inference/generate.py --model facebook/opt-125m &&
      python3 generate/multimodal/vision_language_offline.py --seed 0 &&
      python3 features/automatic_prefix_caching/prefix_caching_offline.py'
 - label: Model Runner V2 Distributed (2 GPUs)
  timeout_in_minutes: 50
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 2+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
    - vllm/v1/worker/gpu/
    - vllm/v1/worker/gpu_worker.py
    - tests/basic_correctness/test_basic_correctness.py
    - tests/v1/distributed/test_async_llm_dp.py
    - tests/v1/distributed/test_eagle_dp.py
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
      cd tests &&
      TARGET_TEST_SUITE=L4 pytest -v -s basic_correctness/test_basic_correctness.py -m "distributed\(num_gpus=2\)" -k "not ray and not True"'
 - label: Model Runner V2 Spec Decode
  timeout_in_minutes: 50
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
  - vllm/v1/worker/gpu/
  - vllm/v1/worker/gpu_worker.py
  - tests/v1/spec_decode/test_max_len.py
  - tests/v1/spec_decode/test_rejection_sampler_utils.py
  - tests/v1/e2e/spec_decode/test_spec_decode.py
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
      cd tests &&
      pytest -v -s v1/spec_decode/test_synthetic_rejection_sampler_utils.py'
@@ -4,12 +4,12 @@ depends_on:
 steps:
 - label: Distributed Model Tests (2 GPUs)
  key: distributed-model-tests-2-gpus
-  timeout_in_minutes: 65
+  timeout_in_minutes: 50
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 2+
-    mem: 16+
+    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -4,32 +4,7 @@ depends_on:
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
  key: multi-modal-models-standard-1-qwen2
-  timeout_in_minutes: 70
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'pip install av &&
      cd tests &&
      pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2" &&
      pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model'
 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
  key: multi-modal-models-standard-2-qwen3-gemma
  timeout_in_minutes: 70
  device: intel_gpu
  agent_tags:
    label: production
@@ -47,17 +22,19 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'cd tests &&
+      'pip install av git+https://github.com/TIGER-AI-Lab/Mantis.git &&
-      pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model'
+      cd tests &&
      pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2" &&
      pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model'
- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
+- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
-  key: multi-modal-models-standard-3-llava-qwen2-vl
+  key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 65
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
-    mem: 24+
+    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -70,12 +47,12 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'cd tests &&
+      'pip install git+https://github.com/TIGER-AI-Lab/Mantis.git &&
-      pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma" &&
+      cd tests &&
-      pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model'
+      pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model'
- label: "Multi-Modal Models (Standard) 4: other + whisper"
+- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
-  key: multi-modal-models-standard-4-other-whisper
+  key: multi-modal-models-standard-3-llava-qwen2-vl
  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
@@ -94,18 +71,43 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'pip install av &&
+      'pip install git+https://github.com/TIGER-AI-Lab/Mantis.git &&
      cd tests &&
      pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma" &&
      pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model'
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
  key: multi-modal-models-standard-4-other-whisper
  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'pip install av git+https://github.com/TIGER-AI-Lab/Mantis.git &&
      cd tests &&
      pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py  --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/processing'
 - label: Multi-Modal Processor # 44min
  key: multi-modal-processor
-  timeout_in_minutes: 60
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
-    mem: 24+
+    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -119,7 +121,7 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'pip install av matplotlib ftfy &&
+      'pip install av matplotlib ftfy git+https://github.com/TIGER-AI-Lab/Mantis.git &&
      pip install open-clip-torch --no-deps &&
      cd tests &&
      pytest -v -s models/multimodal/processing/test_tensor_schema.py
@@ -1,28 +0,0 @@
 group: Quantization
 depends_on:
  - image-build-xpu
 steps:
 - label: Quantization
  key: quantization
  timeout_in_minutes: 30
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  no_plugin: true
  working_dir: "."
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 16+
  source_file_dependencies:
  - csrc/
  - vllm/model_executor/layers/quantization
  - tests/quantization
  commands:
  # - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
  - >-
    bash .buildkite/scripts/hardware_ci/run-intel-test.sh
    'VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s tests/quantization/test_per_token_kv_cache.py --deselect="tests/quantization/test_per_token_kv_cache.py::test_triton_unified_attention_per_token_head_scale[int4-16-128-num_heads0-seq_lens1]"'
@@ -1,29 +0,0 @@
 group: Samplers Intel
 depends_on:
  - image-build-xpu
 steps:
 - label: Samplers Test (FlashInfer)
  key: samplers-test-flashinfer-intel
  timeout_in_minutes: 40
  device: intel_gpu
  agent_tags:
    label: production
    gpu: 1+
    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
    REGISTRY: "public.ecr.aws/q9t5s3a7"
    REPO: "vllm-ci-test-repo"
    VLLM_TEST_DEVICE: "xpu"
  source_file_dependencies:
  - vllm/model_executor/layers
  - vllm/sampling_metadata.py
  - tests/samplers
  - tests/conftest.py
  - vllm/entrypoints/generate/beam_search
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers'
@@ -17,7 +17,7 @@ steps:
  - label: "XPU example Test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 50
+    timeout_in_minutes: 30
    device: intel_gpu
    agent_tags:
      label: production
@@ -42,46 +42,21 @@ steps:
        python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --kv-cache-dtype fp8 &&
        python3 examples/basic/offline_inference/generate.py --model nvidia/Llama-3.1-8B-Instruct-FP8 --block-size 64 --enforce-eager --quantization modelopt --kv-cache-dtype fp8 --attention-backend TRITON_ATTN --max-model-len 4096 &&
        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192 &&
        python3 examples/basic/offline_inference/generate.py --model TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ --block-size 64 --enforce-eager &&
        python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 &&
        python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel &&
        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192 &&
        VLLM_XPU_FUSED_MOE_USE_REF=1 python3 examples/basic/offline_inference/generate.py --model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 --enforce-eager -tp 2 --max-model-len 8192 &&
        python3 examples/basic/offline_inference/generate.py --model INCModel/Qwen3-30B-A3B-Instruct-2507-MXFP4-LLMC --enforce-eager -tp 2 --max-model-len 8192
        '
  - label: "XPU W8A8 FP8 Linear Examples"
    depends_on:
      - image-build-xpu
    timeout_in_minutes: 60
    device: intel_gpu
    agent_tags:
      label: production
      gpu: 1+
      mem: 24+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
      REPO: "vllm-ci-test-repo"
      VLLM_TEST_DEVICE: "xpu"
    source_file_dependencies:
      - vllm/
      - .buildkite/intel_jobs/test-intel.yaml
    commands:
      - >-
        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
        'python3 examples/basic/offline_inference/generate.py --linear-backend xpu --model RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8 --enforce-eager --max-model-len 4096 &&
        python3 examples/basic/offline_inference/generate.py --linear-backend xpu --model neuralmagic/Llama-3.2-1B-Instruct-FP8-dynamic --enforce-eager --max-model-len 4096 &&
        python3 examples/basic/offline_inference/generate.py --linear-backend xpu --model meta-llama/Llama-3.2-1B-Instruct --quantization fp8 --enforce-eager --max-model-len 4096
        '
  - label: "XPU V1 test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 70
+    timeout_in_minutes: 30
    device: intel_gpu
    agent_tags:
      label: production
      gpu: 1+
-      mem: 24+
+      mem: 16+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -93,19 +68,19 @@ steps:
      - >-
        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
        'cd tests &&
        bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
        pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py &&
        pytest -v -s v1/engine --ignore=v1/engine/test_output_processor.py &&
        pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py -k "not test_topk_only and not test_topp_only and not test_topk_and_topp" &&
        pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py &&
        pytest -v -s v1/structured_output &&
        pytest -v -s v1/test_serial_utils.py &&
        pytest -v -s v1/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]" &&
        pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py &&
        pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py'
  - label: "XPU server test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 45
+    timeout_in_minutes: 30
    device: intel_gpu
    agent_tags:
      label: production
@@ -145,52 +120,4 @@ steps:
      - >-
        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
        'cd tests &&
-        pytest -v -s quantization/test_auto_round.py &&
+        pytest -v -s quantization/test_auto_round.py'
        pytest -v -s quantization/test_online.py'
  - label: "XPU GPQA Eval (GPT-OSS)"
    depends_on:
      - image-build-xpu
    timeout_in_minutes: 60
    device: intel_gpu
    agent_tags:
      label: production
      gpu: 1+
      mem: 24+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
      REPO: "vllm-ci-test-repo"
      VLLM_TEST_DEVICE: "xpu"
    source_file_dependencies:
      - vllm/
      - tests/evals/gpt_oss/
      - .buildkite/intel_jobs/test-intel.yaml
    commands:
      - >-
        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
        'pip install "gpt-oss[eval]==0.0.5" &&
        cd tests &&
        pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-xpu.txt'
  - label: "XPU compressed tensors FP8 test"
    depends_on:
      - image-build-xpu
    timeout_in_minutes: 60
    device: intel_gpu
    agent_tags:
      label: production
      gpu: 1+
      mem: 16+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
      REPO: "vllm-ci-test-repo"
      VLLM_TEST_DEVICE: "xpu"
    source_file_dependencies:
      - vllm/
      - tests/quantization/test_compressed_tensors.py
      - .buildkite/intel_jobs/test-intel.yaml
    commands:
      - >-
        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
        'cd tests &&
        pytest -v -s quantization/test_compressed_tensors.py::test_compressed_tensors_fp8'
@@ -1,7 +1,6 @@
 # For hf script, without -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh -m meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 -l 250 -t 8 -f 5
 model_name: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
 rocm_safetensors_load_strategy: lazy
 required_gpu_arch:
  - gfx942
  - gfx950
@@ -72,11 +72,6 @@ def launch_lm_eval(eval_config, tp_size):
    if moe_backend is not None:
        model_args += f"moe_backend={moe_backend},"
    if current_platform.is_rocm():
        rocm_load_strategy = eval_config.get("rocm_safetensors_load_strategy")
        if rocm_load_strategy is not None:
            model_args += f"safetensors_load_strategy={rocm_load_strategy},"
    env_vars = eval_config.get("env_vars", None)
    with scoped_env_vars(env_vars):
        results = lm_eval.simple_evaluate(
@@ -28,6 +28,11 @@
        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
      }
    },
    {
        "dataset_name": "sharegpt",
        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
      }
    },
    {
      "test_name": "serving_llama8B_tp1_random_128_128",
      "server_parameters": {
@@ -3,8 +3,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 #
 # Append a build artifact line to the Buildkite annotation.
-# Usage: annotate-build-artifact.sh <label> <value> <context>
+# Usage: annotate-build-artifact.sh <label> <value>
 set -e
 echo "- **${1}**: \`${2}\`" | \
-  buildkite-agent annotate --append --style 'info' \
+  buildkite-agent annotate --append --style 'info' --context 'release-artifacts'
    --context "${3:?context is required}"
@@ -1,36 +0,0 @@
 #!/bin/bash
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 #
 # Append the Docker image tag(s) an image-build step pushed to a Buildkite
 # annotation, so the built image tags show up on the build page instead of
 # being buried in the job logs.
 #
 # Usage: annotate-image-build.sh <image_tag> [<image_tag> ...]
 set -euo pipefail
 # buildkite-agent only exists on Buildkite agents; no-op elsewhere so the
 # image build scripts stay runnable locally.
 if ! command -v buildkite-agent >/dev/null 2>&1; then
    echo "buildkite-agent not found; skipping image tag annotation"
    exit 0
 fi
 label="${BUILDKITE_LABEL:-Image build}"
 content=""
 for image in "$@"; do
    [[ -n "$image" ]] || continue
    content+="- **${label}**: \`${image}\`"$'\n'
 done
 if [[ -z "$content" ]]; then
    echo "No image tags provided; nothing to annotate"
    exit 0
 fi
 # Best-effort: a flaky annotation must never fail an otherwise successful
 # (and expensive) image build.
 if ! printf '%s' "$content" | \
    buildkite-agent annotate --append --style 'info' --context 'docker-images'; then
    echo "warning: failed to annotate build with image tags"
 fi
@@ -1,35 +0,0 @@
 #!/usr/bin/env bash
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 #
 # Build the macOS arm64 CPU wheel natively on a macOS agent (the `macmini`
 # queue) into artifacts/dist/ for upload-nightly-wheels.sh.
 set -euo pipefail
 # The macmini queue uses persistent checkouts, so refresh tags for setuptools-scm.
 git fetch --tags --force origin
 # The Rust frontend build needs protoc.
 if ! command -v protoc >/dev/null 2>&1; then
  brew install protobuf
 fi
 # upload-nightly-wheels.sh expects exactly one wheel.
 rm -rf artifacts/dist
 mkdir -p artifacts/dist
 export VLLM_TARGET_DEVICE=cpu
 export VLLM_REQUIRE_RUST_FRONTEND=1
 export MACOSX_DEPLOYMENT_TARGET=11.0
 # uv's CPython is universal2; force an arm64-only build and tag so the wheel
 # isn't mislabelled universal2 and installed on Intel Macs where import fails.
 export ARCHFLAGS="-arch arm64"
 export _PYTHON_HOST_PLATFORM="macosx-11.0-arm64"
 export CMAKE_BUILD_PARALLEL_LEVEL="${CMAKE_BUILD_PARALLEL_LEVEL:-4}"
 uv venv --python 3.12
 uv pip install -r requirements/build/cpu.txt --index-strategy unsafe-best-match
 uv build --wheel --no-build-isolation -o artifacts/dist
 ls -l artifacts/dist/*.whl
@@ -29,11 +29,7 @@ if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
        TORCH_INDEX_URL=""
    fi
 else
-    if [ "${TORCH_NIGHTLY:-0}" = "1" ]; then
+    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
        TORCH_INDEX_URL="https://download.pytorch.org/whl/nightly/cu130"
    else
        TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
    fi
 fi
 echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"
@@ -15,9 +15,9 @@ set -euo pipefail
 DEFAULT_REPO_SLUG="vllm-project/vllm"
 DEFAULT_CI_HCL_SOURCE="docker/ci-rocm.hcl"
-DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tools/install_protoc.sh rust-toolchain.toml tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh .buildkite/scripts/rocm/build-ci-base.sh"
+DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh"
 DEFAULT_CI_BASE_DOCKERFILE="docker/Dockerfile.rocm"
-DEFAULT_CI_BASE_DOCKERFILE_STAGES="base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust-toolchain build_nixl build_rocshmem build_deepep mori_base ci_base"
+DEFAULT_CI_BASE_DOCKERFILE_STAGES="base build_rixl build_rocshmem build_deepep mori_base ci_base"
 DEFAULT_CI_BASE_METADATA_VERSION="1"
 IMAGE_EXISTED_BEFORE_BUILD=0
@@ -285,7 +285,7 @@ get_content_arg_names() {
    fi | awk 'NF && !seen[$0]++'
 }
-compute_ci_base_content_hash_once() {
+compute_ci_base_content_hash() {
    local -a content_paths=()
    local -a content_args=()
    local dockerfile="${CI_BASE_DOCKERFILE:-}"
@@ -301,8 +301,7 @@ compute_ci_base_content_hash_once() {
        if [[ -n "${dockerfile}" ]]; then
            printf 'dockerfile:%s\n' "${dockerfile}"
            printf 'resolved-build-args:\n'
-            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}" \
+            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}"
                || return 1
            if [[ -n "${stages}" ]]; then
                printf 'dockerfile-stages:%s\n' "${stages}"
                if [[ -f "${dockerfile}" ]]; then
@@ -315,53 +314,6 @@ compute_ci_base_content_hash_once() {
    } | sha256sum | cut -d' ' -f1
 }
 compute_ci_base_content_hash() {
    local attempts="${CI_BASE_HASH_ATTEMPTS:-3}"
    local delay_secs="${CI_BASE_HASH_RETRY_DELAY:-5}"
    local attempt=0
    local hash=""
    local failed=0
    local -a hashes=()
    if [[ ! "${attempts}" =~ ^[1-9][0-9]*$ ]]; then
        echo "Invalid CI_BASE_HASH_ATTEMPTS: ${attempts}" >&2
        return 1
    fi
    if [[ ! "${delay_secs}" =~ ^[0-9]+$ ]]; then
        echo "Invalid CI_BASE_HASH_RETRY_DELAY: ${delay_secs}" >&2
        return 1
    fi
    for ((attempt = 1; attempt <= attempts; attempt++)); do
        if ! hash=$(compute_ci_base_content_hash_once); then
            echo "ci_base content hash calculation ${attempt}/${attempts} failed" >&2
            failed=1
        else
            hashes+=("${hash}")
            echo "ci_base content hash calculation ${attempt}/${attempts}: ${hash}" >&2
        fi
        if ((attempt < attempts)); then
            sleep "${delay_secs}"
        fi
    done
    if ((failed)) || ((${#hashes[@]} != attempts)); then
        echo "Could not calculate a reliable ci_base content hash" >&2
        return 1
    fi
    for hash in "${hashes[@]:1}"; do
        if [[ "${hash}" != "${hashes[0]}" ]]; then
            echo "ci_base content hash changed between calculations" >&2
            printf '  observed: %s\n' "${hashes[@]}" >&2
            return 1
        fi
    done
    printf '%s\n' "${hashes[0]}"
 }
 extract_dockerfile_arg_default() {
    local dockerfile="$1"
    local arg_name="$2"
@@ -414,11 +366,7 @@ hash_dockerfile_arg_values() {
        printf 'arg:%s=%s\n' "${arg_name}" "${arg_value:-<empty>}"
        if [[ "${arg_name}" == "BASE_IMAGE" && -n "${arg_value}" ]]; then
            digest=$(resolve_image_digest "${arg_value}")
-            if [[ -z "${digest}" ]]; then
+            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest:-unknown}"
                echo "Failed to resolve digest for BASE_IMAGE=${arg_value}" >&2
                return 1
            fi
            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest}"
        fi
    done
 }
@@ -445,16 +393,6 @@ should_upload_wheel_artifacts() {
        || "${TARGET}" == *"artifact"* ]]
 }
 set_buildkite_metadata() {
    local key="$1"
    local value="$2"
    [[ -n "${value}" ]] || return 0
    if command -v buildkite-agent >/dev/null 2>&1; then
        buildkite-agent meta-data set "${key}" "${value}" || true
    fi
 }
 get_remote_image_label() {
    local image_ref="$1"
    local label_key="$2"
@@ -795,8 +733,6 @@ configure_ci_base_image_refs() {
    if is_ci_base_target; then
        IMAGE_TAG="${primary_tag}"
        CI_BASE_IMAGE="${primary_tag}"
        export CI_BASE_IMAGE
        export IMAGE_TAG
        echo "ci_base primary image tag: ${CI_BASE_IMAGE_TAG}"
@@ -814,10 +750,6 @@ configure_ci_base_image_refs() {
            echo "ci_base stable alias will not be pushed for this build"
            echo "Set NIGHTLY=1 on ${CI_BASE_STABLE_BRANCH:-main} to refresh ${stable_tag}"
        fi
        set_buildkite_metadata "rocm-ci-base-image" "${CI_BASE_IMAGE_TAG}"
        set_buildkite_metadata "rocm-ci-base-image-content" "${content_tag}"
        set_buildkite_metadata "rocm-ci-base-image-commit" "${CI_BASE_IMAGE_TAG_COMMIT_REF:-}"
        set_buildkite_metadata "rocm-ci-base-image-stable" "${CI_BASE_IMAGE_TAG_STABLE:-}"
        return 0
    fi
@@ -1159,8 +1091,8 @@ ci_base_metadata_pairs() {
    metadata_pair "vllm.rocm.nic_backend" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIC_BACKEND")"
    metadata_pair "vllm.rocm.ainic_version" "$(resolve_dockerfile_arg_value "${dockerfile}" "AINIC_VERSION")"
    metadata_pair "vllm.rocm.ubuntu_codename" "$(resolve_dockerfile_arg_value "${dockerfile}" "UBUNTU_CODENAME")"
-    metadata_pair "vllm.rocm.nixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_REPO")"
+    metadata_pair "vllm.rocm.rixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_REPO")"
-    metadata_pair "vllm.rocm.nixl_commit" "${NIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_BRANCH")}"
+    metadata_pair "vllm.rocm.rixl_commit" "${RIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_BRANCH")}"
    metadata_pair "vllm.rocm.ucx_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_REPO")"
    metadata_pair "vllm.rocm.ucx_commit" "${UCX_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_BRANCH")}"
    metadata_pair "vllm.rocm.rocshmem_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "ROCSHMEM_REPO")"
@@ -1169,7 +1101,7 @@ ci_base_metadata_pairs() {
    metadata_pair "vllm.rocm.deepep_commit" "${DEEPEP_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_BRANCH")}"
    metadata_pair "vllm.rocm.deepep_nic" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_NIC")"
    metadata_pair "vllm.rocm.deepep_rocm_arch" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_ROCM_ARCH")"
-    metadata_pair "vllm.rocm.nixl_cache_key" "${NIXL_CACHE_KEY:-}"
+    metadata_pair "vllm.rocm.rixl_cache_key" "${RIXL_CACHE_KEY:-}"
    metadata_pair "vllm.rocm.rocshmem_cache_key" "${ROCSHMEM_CACHE_KEY:-}"
    metadata_pair "vllm.rocm.deepep_cache_key" "${DEEPEP_CACHE_KEY:-}"
@@ -1263,24 +1195,12 @@ uses_rocm_csrc_cache() {
    esac
 }
 uses_rocm_rust_cache() {
    case "${TARGET}" in
        rust-rocm-ci|test-rocm-ci|test-rocm-ci-with-wheel|test-rocm-ci-with-artifacts|export-wheel-rocm)
            return 0
            ;;
        *)
            return 1
            ;;
    esac
 }
 compute_rocm_csrc_content_hash() {
    local bake_dir=""
    local dockerfile_rocm=""
    local -a content_paths=(
        "requirements/common.txt"
        "requirements/rocm.txt"
        "pyproject.toml"
        "setup.py"
        "CMakeLists.txt"
        "cmake"
@@ -1324,56 +1244,6 @@ compute_rocm_csrc_content_hash_if_needed() {
    echo "ROCm csrc content cache ref: ${ROCM_CSRC_CONTENT_CACHE_REF}"
 }
 compute_rocm_rust_content_hash() {
    local bake_dir=""
    local dockerfile_rocm=""
    local -a content_paths=(
        "requirements/build/rust.txt"
        "rust/Cargo.lock"
        "rust/Cargo.toml"
        "rust/proto"
        "rust/src"
        "rust-toolchain.toml"
        "tools/build_rust.py"
        "tools/install_protoc.sh"
        "build_rust.sh"
    )
    local -a content_args=()
    bake_dir=$(dirname "${VLLM_BAKE_FILE}")
    dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
    mapfile -t content_args < <(
        get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
    )
    {
        printf 'rust-input-files-hash:%s\n' "$(compute_content_hash "${content_paths[@]}")"
        printf 'dockerfile:%s\n' "${dockerfile_rocm}"
        printf 'resolved-build-args:\n'
        hash_dockerfile_arg_values "${dockerfile_rocm}" "${content_args[@]}"
        printf 'dockerfile-stages:base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build\n'
        if [[ -f "${dockerfile_rocm}" ]]; then
            hash_dockerfile_stages "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build"
        else
            printf 'missing:%s\n' "${dockerfile_rocm}"
        fi
    } | sha256sum | cut -d' ' -f1
 }
 compute_rocm_rust_content_hash_if_needed() {
    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
    if [[ "${ROCM_RUST_CONTENT_CACHE:-1}" == "0" ]] || ! uses_rocm_rust_cache; then
        return 0
    fi
    ROCM_RUST_CONTENT_HASH=$(compute_rocm_rust_content_hash)
    ROCM_RUST_CONTENT_CACHE_REF="${cache_repo}:rust-rocm-input-${ROCM_RUST_CONTENT_HASH}"
    export ROCM_RUST_CONTENT_HASH
    export ROCM_RUST_CONTENT_CACHE_REF
    echo "ROCm Rust content cache ref: ${ROCM_RUST_CONTENT_CACHE_REF}"
 }
 write_hcl_string_list_entries() {
    local indent="$1"
    local value=""
@@ -1431,7 +1301,6 @@ write_rocm_build_arg_override() {
                "${CI_BASE_DOCKERFILE_STAGES:-${DEFAULT_CI_BASE_DOCKERFILE_STAGES}}" \
                "${CI_BASE_CONTENT_ARGS:-}"
            get_content_arg_names "${dockerfile_rocm}" "base csrc-build" "${ROCM_CSRC_CONTENT_ARGS:-}"
            get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
        } | awk 'NF && !seen[$0]++'
    )
@@ -1480,133 +1349,46 @@ validate_cache_export_mode() {
    esac
 }
 validate_content_cache_export_mode() {
    local mode="$1"
    local env_name="$2"
    case "${mode}" in
        missing|always|never)
            ;;
        *)
            echo "Error: ${env_name} must be one of: missing, always, never"
            exit 1
            ;;
    esac
 }
 should_export_content_cache_ref() {
    local cache_ref="$1"
    local cache_name="$2"
    local mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
    case "${mode}" in
        always)
            echo "${cache_name} content cache export mode is always; exporting ${cache_ref}"
            return 0
            ;;
        never)
            echo "${cache_name} content cache export mode is never; not exporting ${cache_ref}"
            return 1
            ;;
        missing|"")
            if docker buildx imagetools inspect "${cache_ref}" >/dev/null 2>&1; then
                echo "${cache_name} content cache exists; not re-exporting ${cache_ref}"
                return 1
            fi
            echo "${cache_name} content cache missing; will export ${cache_ref}"
            return 0
            ;;
        *)
            echo "Error: ROCM_CONTENT_CACHE_EXPORT_MODE must be one of: missing, always, never"
            exit 1
            ;;
    esac
 }
 write_rocm_cache_override() {
    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
    local content_cache_export_mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
    local csrc_cache_to_mode="${ROCM_CSRC_CACHE_TO_MODE:-max}"
    local rust_cache_to_mode="${ROCM_RUST_CACHE_TO_MODE:-max}"
    local rocm_cache_to_mode="${ROCM_FINAL_CACHE_TO_MODE:-min}"
-    local -a csrc_content_cache_from=()
+    local -a content_cache_from=()
    local -a rust_content_cache_from=()
    local -a combined_content_cache_from=()
    local -a csrc_cache_to=()
    local -a rust_cache_to=()
    local -a rocm_cache_to=()
    local -a export_wheel_cache_to=()
    local export_csrc_cache=1
    local export_rust_cache=1
-    if ! uses_rocm_csrc_cache && ! uses_rocm_rust_cache; then
+    if ! uses_rocm_csrc_cache; then
        return 0
    fi
    validate_content_cache_export_mode \
        "${content_cache_export_mode}" \
        "ROCM_CONTENT_CACHE_EXPORT_MODE"
    validate_cache_export_mode "${csrc_cache_to_mode}" "ROCM_CSRC_CACHE_TO_MODE"
    validate_cache_export_mode "${rust_cache_to_mode}" "ROCM_RUST_CACHE_TO_MODE"
    validate_cache_export_mode "${rocm_cache_to_mode}" "ROCM_FINAL_CACHE_TO_MODE"
    echo "ROCm content cache export mode: ${content_cache_export_mode}"
    echo "ROCm csrc cache export mode: ${csrc_cache_to_mode}"
    echo "ROCm Rust cache export mode: ${rust_cache_to_mode}"
    echo "ROCm final image cache export mode: ${rocm_cache_to_mode}"
    if [[ -n "${ROCM_CSRC_CONTENT_CACHE_REF:-}" ]]; then
-        csrc_content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
+        content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
-        if should_export_content_cache_ref "${ROCM_CSRC_CONTENT_CACHE_REF}" "ROCm csrc"; then
+        csrc_cache_to+=(
-            csrc_cache_to+=(
+            "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
-                "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
+        )
            )
        else
            export_csrc_cache=0
        fi
    fi
    if [[ -n "${ROCM_RUST_CONTENT_CACHE_REF:-}" ]]; then
        rust_content_cache_from+=("type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF}")
        if should_export_content_cache_ref "${ROCM_RUST_CONTENT_CACHE_REF}" "ROCm Rust"; then
            rust_cache_to+=(
                "type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF},mode=${rust_cache_to_mode},ignore-error=true"
            )
        else
            export_rust_cache=0
        fi
    fi
    combined_content_cache_from=("${csrc_content_cache_from[@]}" "${rust_content_cache_from[@]}")
    # Docker Hub cache exports are best-effort. A cache-only target failure can
    # otherwise cancel the sibling image target before its manifest is pushed.
    if [[ -n "${BUILDKITE_COMMIT:-}" ]]; then
-        if [[ ${export_csrc_cache} -eq 1 ]]; then
+        csrc_cache_to+=(
-            csrc_cache_to+=(
+            "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
-                "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
+        )
            )
        fi
        if [[ ${export_rust_cache} -eq 1 ]]; then
            rust_cache_to+=(
                "type=registry,ref=${cache_repo}:rust-rocm-${BUILDKITE_COMMIT},mode=${rust_cache_to_mode},ignore-error=true"
            )
        fi
        rocm_cache_to+=(
            "type=registry,ref=${cache_repo}:rocm-${BUILDKITE_COMMIT},mode=${rocm_cache_to_mode},ignore-error=true"
        )
    fi
    if [[ -n "${ROCM_CACHE_BRANCH_TAG:-}" ]]; then
-        if [[ ${export_csrc_cache} -eq 1 ]]; then
+        csrc_cache_to+=(
-            csrc_cache_to+=(
+            "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
-                "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
+        )
            )
        fi
        if [[ ${export_rust_cache} -eq 1 ]]; then
            rust_cache_to+=(
                "type=registry,ref=${cache_repo}:rust-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rust_cache_to_mode},ignore-error=true"
            )
        fi
        rocm_cache_to+=(
            "type=registry,ref=${cache_repo}:rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rocm_cache_to_mode},ignore-error=true"
        )
@@ -1624,7 +1406,7 @@ target "csrc-rocm-ci" {
  cache-from = concat(
    get_cache_from_rocm_csrc(),
 EOF
-        write_hcl_string_list "    " "${csrc_content_cache_from[@]}"
+        write_hcl_string_list "    " "${content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1632,23 +1414,11 @@ EOF
        cat <<EOF
 }
 target "rust-rocm-ci" {
  cache-from = concat(
    get_cache_from_rocm_rust(),
 EOF
        write_hcl_string_list "    " "${rust_content_cache_from[@]}"
        cat <<EOF
  )
 EOF
        write_hcl_string_list_attr "  " "cache-to" "${rust_cache_to[@]}"
        cat <<EOF
 }
 target "test-rocm-ci" {
  cache-from = concat(
    get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
+        write_hcl_string_list "    " "${content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1660,7 +1430,7 @@ target "export-wheel-rocm" {
  cache-from = concat(
    get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
+        write_hcl_string_list "    " "${content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1686,7 +1456,7 @@ extract_dependency_pins() {
        return 0
    fi
-    for var in NIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
+    for var in RIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
        if [[ -n "${!var:-}" ]]; then
            echo "Using provided ${var}: ${!var}"
            continue
@@ -1706,30 +1476,30 @@ extract_dependency_pins() {
 compute_dependency_cache_keys() {
    local bake_dir=""
    local dockerfile_rocm=""
-    local nixl_branch=""
+    local rixl_branch=""
    local ucx_branch=""
    local rocshmem_branch=""
    local deepep_branch=""
-    local nixl_material=""
+    local rixl_material=""
    local rocshmem_material=""
    local deepep_material=""
    bake_dir=$(dirname "${VLLM_BAKE_FILE}")
    dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
-    nixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "NIXL_BRANCH")
+    rixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "RIXL_BRANCH")
    ucx_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "UCX_BRANCH")
    rocshmem_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "ROCSHMEM_BRANCH")
    deepep_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "DEEPEP_BRANCH")
-    if [[ -n "${nixl_branch}" && -n "${ucx_branch}" ]]; then
+    if [[ -n "${rixl_branch}" && -n "${ucx_branch}" ]]; then
-        nixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_nixl")
+        rixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_rixl")
-        NIXL_CACHE_KEY=$(
+        RIXL_CACHE_KEY=$(
            compose_dependency_cache_key \
-                "${nixl_branch}-ucx-${ucx_branch}" \
+                "${rixl_branch}-ucx-${ucx_branch}" \
-                "${nixl_material}"
+                "${rixl_material}"
        )
-        export NIXL_CACHE_KEY
+        export RIXL_CACHE_KEY
-        echo "NIXL dependency cache key: ${NIXL_CACHE_KEY}"
+        echo "RIXL dependency cache key: ${RIXL_CACHE_KEY}"
    fi
    if [[ -n "${rocshmem_branch}" ]]; then
@@ -1780,11 +1550,11 @@ dependency_cache_ref_for_target() {
    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
    case "${target}" in
-        nixl-rocm-ci)
+        rixl-rocm-ci)
-            if [[ -n "${NIXL_CACHE_KEY:-}" ]]; then
+            if [[ -n "${RIXL_CACHE_KEY:-}" ]]; then
-                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_CACHE_KEY}"
+                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_CACHE_KEY}"
-            elif [[ -n "${NIXL_BRANCH:-}" ]]; then
+            elif [[ -n "${RIXL_BRANCH:-}" ]]; then
-                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
+                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
            fi
            ;;
        rocshmem-rocm-ci)
@@ -1815,7 +1585,7 @@ add_dependency_cache_target() {
 resolve_ci_base_dependency_targets() {
    local mode="${ROCM_DEP_CACHE_EXPORT_MODE:-missing}"
-    local nixl_ref=""
+    local rixl_ref=""
    local rocshmem_ref=""
    local deepep_ref=""
@@ -1824,7 +1594,7 @@ resolve_ci_base_dependency_targets() {
    case "${mode}" in
        always)
            echo "ROCM_DEP_CACHE_EXPORT_MODE=always; exporting all dependency caches serially"
-            for target in nixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
+            for target in rixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
                if [[ -n "$(dependency_cache_ref_for_target "${target}")" ]]; then
                    add_dependency_cache_target "${target}"
                fi
@@ -1844,13 +1614,13 @@ resolve_ci_base_dependency_targets() {
            ;;
    esac
-    if [[ "${mode}" != "always" && -n "${NIXL_CACHE_KEY:-}" ]]; then
+    if [[ "${mode}" != "always" && -n "${RIXL_CACHE_KEY:-}" ]]; then
-        nixl_ref=$(dependency_cache_ref_for_target "nixl-rocm-ci")
+        rixl_ref=$(dependency_cache_ref_for_target "rixl-rocm-ci")
-        if dependency_cache_ref_exists "${nixl_ref}"; then
+        if dependency_cache_ref_exists "${rixl_ref}"; then
-            echo "NIXL dependency cache exists: ${nixl_ref}"
+            echo "RIXL dependency cache exists: ${rixl_ref}"
        else
-            echo "NIXL dependency cache missing; will seed: ${nixl_ref}"
+            echo "RIXL dependency cache missing; will seed: ${rixl_ref}"
-            add_dependency_cache_target "nixl-rocm-ci"
+            add_dependency_cache_target "rixl-rocm-ci"
        fi
    fi
@@ -1950,8 +1720,8 @@ confirm_remote_image_push() {
        fi
        if [[ -z "${remote_revision}" \
-              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 ]] \
+              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 \
-            && image_tag_is_commit_scoped; then
+              && image_tag_is_commit_scoped ]]; then
            echo "Remote image exists under a commit-scoped tag; accepting push despite missing revision label."
            return 0
        fi
@@ -2009,10 +1779,7 @@ seed_dependency_caches_if_needed() {
        echo "--- :docker: Seeding ${target}"
        echo "Expected cache ref: ${cache_ref}"
-        docker buildx bake \
+        docker buildx bake "${BAKE_FILES[@]}" --progress plain "${target}"
            "${BAKE_FILES[@]}" \
            --progress "${BUILDKIT_PROGRESS:-plain}" \
            "${target}"
        verify_dependency_cache_ref "${cache_ref}"
    done
 }
@@ -2040,10 +1807,7 @@ run_bake() {
    local build_rc=0
    echo "--- :docker: Building ${TARGET}"
-    docker buildx bake \
+    docker buildx bake "${BAKE_FILES[@]}" --progress plain "${BAKE_TARGETS[@]}" || build_rc=$?
        "${BAKE_FILES[@]}" \
        --progress "${BUILDKIT_PROGRESS:-plain}" \
        "${BAKE_TARGETS[@]}" || build_rc=$?
    if [[ ${build_rc} -eq 0 ]]; then
        echo "--- :white_check_mark: Build complete"
@@ -2081,57 +1845,36 @@ upload_wheel_artifacts_if_present() {
    local wheel_dir="./wheel-export"
    local artifact_dir="artifacts/vllm-rocm-install"
    local archive_name="vllm-rocm-install.tar.gz"
    local metadata_dir="${wheel_dir}/.vllm-ci-artifact"
    local native_base_image=""
    local whl=""
    local whl_name=""
    local -a wheels=()
    if ! should_upload_wheel_artifacts; then
        return 0
    fi
-    if [[ -d "${wheel_dir}" ]]; then
+    if [[ ! -d "${wheel_dir}" ]] || ! ls "${wheel_dir}"/*.whl >/dev/null 2>&1; then
-        mapfile -t wheels < <(find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print)
+        echo "No ROCm wheel artifacts found in ${wheel_dir}"
-    fi
+        return 0
    if [[ ${#wheels[@]} -ne 1 ]]; then
        echo "Expected exactly one ROCm wheel in ${wheel_dir}; found ${#wheels[@]}" >&2
        return 1
    fi
    whl="${wheels[0]}"
    whl_name=$(basename "${whl}")
    native_base_image="${CI_BASE_IMAGE_TAG_COMMIT_REF:-${CI_BASE_IMAGE:-}}"
    if [[ -z "${native_base_image}" ]]; then
        echo "Native ROCm artifact requires a ci_base image reference" >&2
        return 1
    fi
    echo "--- :package: Uploading ROCm vLLM install artifact"
-    rm -rf "${artifact_dir}" "${metadata_dir}"
+    mkdir -p "${artifact_dir}"
    mkdir -p "${artifact_dir}" "${metadata_dir}"
    printf '%s\n' "${BUILDKITE_COMMIT:-local}" > "${metadata_dir}/commit.txt"
    printf '%s\n' "${native_base_image}" > "${metadata_dir}/native-base-image.txt"
    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${metadata_dir}/ci-base-image.txt"
    printf '%s\n' "${IMAGE_TAG:-}" > "${metadata_dir}/fallback-image.txt"
    printf '%s\n' "${whl_name}" > "${metadata_dir}/wheel-filename.txt"
    tar -C "${wheel_dir}" -czf "${artifact_dir}/${archive_name}" .
    (
        cd "${artifact_dir}"
        sha256sum "${archive_name}" > "${archive_name}.sha256"
    )
    echo "Created ${archive_name}: $(du -sh "${artifact_dir}/${archive_name}" | cut -f1)"
-    cp "${metadata_dir}"/*.txt "${artifact_dir}/"
+    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${artifact_dir}/ci-base-image.txt"
-    cp "${whl}" "${artifact_dir}/${whl_name}"
+    printf '%s\n' "${IMAGE_TAG:-}" > "${artifact_dir}/fallback-image.txt"
-    echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"
+
    for whl in "${wheel_dir}"/*.whl; do
        [[ -f "${whl}" ]] || continue
        whl_name=$(basename "${whl}")
        cp "${whl}" "${artifact_dir}/${whl_name}"
        echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"
    done
    if command -v buildkite-agent >/dev/null 2>&1; then
-        buildkite-agent artifact upload "${artifact_dir}/*" || return 1
+        buildkite-agent artifact upload "${artifact_dir}/*"
        echo "ROCm vLLM install artifacts uploaded to ${artifact_dir}/"
    elif [[ "${BUILDKITE:-false}" == "true" ]]; then
        echo "buildkite-agent not found; cannot upload required ROCm artifacts" >&2
        return 1
    else
        echo "Not in Buildkite, skipping artifact upload"
    fi
@@ -2155,7 +1898,6 @@ main() {
    compute_dependency_cache_keys
    write_ci_base_label_override
    compute_rocm_csrc_content_hash_if_needed
    compute_rocm_rust_content_hash_if_needed
    write_rocm_cache_override
    resolve_ci_base_dependency_targets
    print_bake_config
@@ -2163,11 +1905,6 @@ main() {
        echo "BAKE_PRINT_ONLY=1 set; skipping build"
        return 0
    fi
    if should_upload_wheel_artifacts; then
        # wheel-export is an output directory, not a BuildKit cache. Starting
        # clean prevents a failed/retried export from packaging a stale wheel.
        rm -rf ./wheel-export
    fi
    seed_dependency_caches_if_needed
    run_bake
    upload_wheel_artifacts_if_present
@@ -45,10 +45,8 @@ $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --curr
 echo "Uploading indices to $S3_COMMIT_PREFIX"
 aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
-# copy to /nightly/ only when enabled for a main branch build that is not a PR
+# copy to /nightly/ only if it is on the main branch and not a PR
-if [[ "${UPDATE_NIGHTLY_INDEX:-1}" == "1" && \
+if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
      "$BUILDKITE_BRANCH" == "main" && \
      "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
    echo "Uploading indices to overwrite /nightly/"
    aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/nightly/"
 fi
@@ -69,7 +67,7 @@ pure_version="${version%%+*}"
 echo "Pure version (without variant): $pure_version"
 # re-generate and copy to /<pure_version>/ only if it does not have "dev" in the version
-if [[ "${UPDATE_VERSION_INDEX:-1}" == "1" && "$version" != *"dev"* ]]; then
+if [[ "$version" != *"dev"* ]]; then
    echo "Re-generating indices for /$pure_version/"
    rm -rf "${INDICES_OUTPUT_DIR:?}"
    mkdir -p "$INDICES_OUTPUT_DIR"
@@ -1,7 +1,7 @@
 #!/bin/bash
-# This script runs ROCm tests either directly in a native CI pod or inside the
+# This script runs tests inside the corresponding ROCm docker container.
-# corresponding Docker container. Multi-node tests continue to use Docker.
+# It handles both single-node and multi-node test configurations.
 #
 # Multi-node detection: Instead of matching on fragile group names, we detect
 # multi-node jobs structurally by looking for the bracket command syntax
@@ -28,34 +28,6 @@
 ###############################################################################
 set -o pipefail
 : "${BUILDKIT_PROGRESS:=plain}"
 : "${TERM:=xterm-256color}"
 : "${FORCE_COLOR:=1}"
 : "${CLICOLOR_FORCE:=1}"
 : "${PY_COLORS:=1}"
 : "${ROCM_DOCKER_TTY:=1}"
 : "${PYTHONFAULTHANDLER:=1}"
 : "${PYTEST_TIMEOUT:=2400}"
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --color"* ]]; then
  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--color=yes"
 fi
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations="* ]]; then
  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations=25"
 fi
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations-min="* ]]; then
  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations-min=1.0"
 fi
 # Dump stacks after 25 minutes, then stop an individual test after 40 minutes.
 if [[ " ${PYTEST_ADDOPTS:-} " != *" faulthandler_timeout="* ]]; then
  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }-o faulthandler_timeout=1500"
 fi
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --timeout-method="* &&
  " ${PYTEST_ADDOPTS:-} " != *" --timeout-method "* ]]; then
  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--timeout-method=thread"
 fi
 export BUILDKIT_PROGRESS TERM FORCE_COLOR CLICOLOR_FORCE PY_COLORS PYTEST_ADDOPTS PYTEST_TIMEOUT ROCM_DOCKER_TTY
 export PYTHONFAULTHANDLER
 # Export Python path for commands that run directly on the host. Containerized
 # tests set this to /vllm-workspace below so spawned Python processes do not
 # depend on their current working directory.
@@ -70,28 +42,6 @@ report_docker_usage() {
  docker system df || true
 }
 clear_ci_orchestration_env() {
  unset -v \
    VLLM_TEST_GROUP_NAME \
    VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE \
    VLLM_CI_ARTIFACT_STEP \
    VLLM_TEST_CACHE \
    VLLM_CI_EXECUTION_MODE \
    VLLM_CI_WORKSPACE \
    VLLM_CI_REQUIRE_WORKSPACE_MOUNT \
    VLLM_TEST_COMMANDS \
    VLLM_CI_BRANCH \
    VLLM_CI_BASE_IMAGE \
    VLLM_CI_FALLBACK_IMAGE \
    VLLM_CI_DOCKER_DISABLED \
    VLLM_CI_ARTIFACT_GLOB \
    VLLM_CI_ARTIFACT_CHECKSUM_GLOB \
    VLLM_CI_EXPECTED_GPU_COUNT \
    VLLM_CI_USE_ARTIFACTS \
    VLLM_CI_RESULTS_ROOT \
    VLLM_ALLOW_DEPRECATED_BEAM_SEARCH
 }
 cleanup_network() {
  local max_nodes=${NUM_NODES:-2}
  for node in $(seq 0 $((max_nodes - 1))); do
@@ -184,11 +134,7 @@ prepare_artifact_image() {
  fi
  cp "${wheel_dir}"/*.whl "${context_dir}/wheels/" || return 1
-  tar -C "${wheel_dir}" \
+  tar -C "${wheel_dir}" --exclude='*.whl' -cf - . \
    --exclude='*.whl' \
    --exclude='.vllm-ci-artifact' \
    --exclude='./.vllm-ci-artifact' \
    -cf - . \
    | tar -C "${workspace_dir}" -xf - || return 1
  cat > "${context_dir}/Dockerfile" <<'EOF'
 ARG BASE_IMAGE
@@ -203,7 +149,6 @@ EOF
  echo "--- Building local ROCm test image"
  docker build \
    --pull=false \
    --progress "${BUILDKIT_PROGRESS}" \
    --build-arg "BASE_IMAGE=${base_image}" \
    -t "${artifact_image}" \
    "${context_dir}" || return 1
@@ -211,276 +156,6 @@ EOF
  return 0
 }
 is_native_runtime() {
  [[ "${AMD_CI_RUNTIME:-}" == "native" || "${NATIVE_CI:-}" == "true" ]]
 }
 validate_native_workspace() {
  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
  local workspace_real=""
  local checkout_real=""
  local workspace_mount=""
  mkdir -p "${workspace_dir}" || return 1
  workspace_real=$(readlink -m "${workspace_dir}") || return 1
  if [[ -n "${BUILDKITE_BUILD_CHECKOUT_PATH:-}" ]]; then
    checkout_real=$(readlink -m "${BUILDKITE_BUILD_CHECKOUT_PATH}") || return 1
    if [[ "${checkout_real}" == "${workspace_real}" \
      || "${checkout_real}" == "${workspace_real}/"* \
      || "${workspace_real}" == "${checkout_real}/"* ]]; then
      echo "Refusing to replace ${workspace_real}; it overlaps the Buildkite checkout ${checkout_real}" >&2
      return 1
    fi
  fi
  if [[ "${VLLM_CI_REQUIRE_WORKSPACE_MOUNT:-1}" == "1" ]]; then
    if ! command -v findmnt >/dev/null 2>&1; then
      echo "findmnt is required to verify the native workspace mount" >&2
      return 1
    fi
    workspace_mount=$(findmnt -n -T "${workspace_real}" -o TARGET 2>/dev/null || true)
    if [[ "$(readlink -m "${workspace_mount:-/}")" != "${workspace_real}" ]]; then
      echo "Native CI requires a dedicated volume mounted at ${workspace_real}" >&2
      return 1
    fi
  fi
 }
 prepare_native_workspace() {
  if [[ "${VLLM_CI_USE_ARTIFACTS:-0}" != "1" ]]; then
    echo "Native CI requires VLLM_CI_USE_ARTIFACTS=1"
    return 1
  fi
  if ! command -v buildkite-agent >/dev/null 2>&1; then
    echo "buildkite-agent not found; cannot download ROCm wheel artifact"
    return 1
  fi
  validate_native_workspace || return 1
  local artifact_glob="${VLLM_CI_ARTIFACT_GLOB:-artifacts/vllm-rocm-install/vllm-rocm-install.tar.gz}"
  local artifact_checksum_glob="${VLLM_CI_ARTIFACT_CHECKSUM_GLOB:-${artifact_glob}.sha256}"
  local artifact_step="${VLLM_CI_ARTIFACT_STEP:-image-build-amd}"
  local archive=""
  local checksum=""
  local download_dir=""
  local metadata_dir=""
  local recorded_base=""
  local recorded_commit=""
  local recorded_wheel=""
  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
  local wheel_dir=""
  local attempt=0
  local attempt_dir=""
  local -a archives=()
  local -a checksums=()
  local -a wheels=()
  artifact_work_dir=$(mktemp -d -t vllm-rocm-artifact.XXXXXX) || return 1
  wheel_dir="${artifact_work_dir}/wheels"
  mkdir -p "${wheel_dir}" || return 1
  echo "--- Downloading ROCm wheel artifact from ${artifact_step} (native in-pod)"
  for attempt in 1 2 3; do
    attempt_dir="${artifact_work_dir}/download-${attempt}"
    rm -rf "${attempt_dir}" || return 1
    mkdir -p "${attempt_dir}" || return 1
    if buildkite-agent artifact download \
      "${artifact_glob}" "${attempt_dir}" --step "${artifact_step}" \
      && buildkite-agent artifact download \
        "${artifact_checksum_glob}" "${attempt_dir}" --step "${artifact_step}"; then
      download_dir="${attempt_dir}"
      break
    fi
    echo "Artifact download attempt ${attempt}/3 failed"
    if [[ "${attempt}" -lt 3 ]]; then
      sleep $((attempt * 2))
    fi
  done
  if [[ -z "${download_dir}" ]]; then
    echo "Failed to download ${artifact_glob} and ${artifact_checksum_glob} from ${artifact_step}"
    return 1
  fi
  mapfile -t archives < <(
    find "${download_dir}" -name "vllm-rocm-install.tar.gz" -type f -print
  )
  mapfile -t checksums < <(
    find "${download_dir}" -name "vllm-rocm-install.tar.gz.sha256" -type f -print
  )
  if [[ ${#archives[@]} -ne 1 || ${#checksums[@]} -ne 1 ]]; then
    echo "Expected exactly one ROCm archive and checksum; found ${#archives[@]} archive(s) and ${#checksums[@]} checksum(s)" >&2
    return 1
  fi
  archive="${archives[0]}"
  checksum="${checksums[0]}"
  if [[ "$(dirname "${archive}")" != "$(dirname "${checksum}")" ]]; then
    echo "ROCm archive and checksum were downloaded to different directories" >&2
    return 1
  fi
  (
    cd "$(dirname "${archive}")"
    sha256sum -c "$(basename "${checksum}")"
  ) || return 1
  tar --no-same-owner -xzf "${archive}" -C "${wheel_dir}" || return 1
  mapfile -t wheels < <(
    find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print
  )
  if [[ ${#wheels[@]} -ne 1 ]]; then
    echo "ROCm artifact must contain exactly one top-level wheel; found ${#wheels[@]}" >&2
    return 1
  fi
  metadata_dir="${wheel_dir}/.vllm-ci-artifact"
  for metadata_file in commit.txt native-base-image.txt wheel-filename.txt; do
    if [[ ! -s "${metadata_dir}/${metadata_file}" ]]; then
      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
      return 1
    fi
  done
  for metadata_file in ci-base-image.txt fallback-image.txt; do
    if [[ ! -f "${metadata_dir}/${metadata_file}" ]]; then
      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
      return 1
    fi
  done
  recorded_commit=$(tr -d '\r\n' < "${metadata_dir}/commit.txt")
  recorded_base=$(tr -d '\r\n' < "${metadata_dir}/native-base-image.txt")
  recorded_wheel=$(tr -d '\r\n' < "${metadata_dir}/wheel-filename.txt")
  if [[ -z "${BUILDKITE_COMMIT:-}" || "${recorded_commit}" != "${BUILDKITE_COMMIT}" ]]; then
    echo "ROCm artifact commit ${recorded_commit} does not match ${BUILDKITE_COMMIT:-unset}" >&2
    return 1
  fi
  if [[ -z "${VLLM_CI_BASE_IMAGE:-}" || "${recorded_base}" != "${VLLM_CI_BASE_IMAGE}" ]]; then
    echo "ROCm artifact base ${recorded_base} does not match ${VLLM_CI_BASE_IMAGE:-unset}" >&2
    return 1
  fi
  if [[ "${recorded_wheel}" != "$(basename "${wheels[0]}")" ]]; then
    echo "ROCm artifact wheel manifest ${recorded_wheel} does not match $(basename "${wheels[0]}")" >&2
    return 1
  fi
  for required_dir in tests .buildkite requirements; do
    if [[ ! -d "${wheel_dir}/${required_dir}" ]]; then
      echo "ROCm wheel artifact did not contain ${required_dir}/" >&2
      return 1
    fi
  done
  echo "--- Installing ROCm wheel into pod environment"
  python3 -m pip install --no-deps --force-reinstall "${wheels[0]}" || return 1
  echo "--- Preparing ${workspace_dir} from artifact"
  find "${workspace_dir}" -mindepth 1 -maxdepth 1 -exec rm -rf -- {} + || return 1
  tar -C "${wheel_dir}" \
    --exclude='*.whl' \
    --exclude='.vllm-ci-artifact' \
    --exclude='./.vllm-ci-artifact' \
    -cf - . | tar --no-same-owner -C "${workspace_dir}" -xf - || return 1
  if [[ ! -d "${workspace_dir}/tests" ]]; then
    echo "Failed to stage the native test workspace" >&2
    return 1
  fi
  return 0
 }
 initialize_native_environment() {
  local job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
  local job_id_suffix=""
  local native_root=""
  local hf_fstype=""
  local hf_mount=""
  if [[ "$(id -u)" -ne 0 ]]; then
    echo "Native ROCm CI currently requires the ci_base container to run as root" >&2
    return 1
  fi
  job_id="${job_id//[^A-Za-z0-9_.-]/_}"
  job_id_suffix="${job_id##*-}"
  job_id_suffix="${job_id_suffix:0:12}"
  native_root="/tmp/vllm-native-${job_id}"
  TMPDIR="/tmp/vllm-${job_id_suffix}/tmp"
  VLLM_RPC_BASE_PATH="/tmp"
  TORCHINDUCTOR_CACHE_DIR="${native_root}/cache/torchinductor"
  TRITON_CACHE_DIR="${native_root}/cache/triton"
  VLLM_CACHE_ROOT="${native_root}/cache/vllm"
  XDG_CACHE_HOME="${native_root}/cache/xdg"
  : "${HF_HOME:=/home/buildkite-agent/huggingface}"
  # datasets uses POSIX locks that are unsupported by the shared HF NFS cache.
  # Keep processed datasets job-local while retaining the persistent Hub cache.
  HF_DATASETS_CACHE="${native_root}/cache/huggingface/datasets"
  : "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
  : "${HF_HUB_ETAG_TIMEOUT:=60}"
  export TMPDIR VLLM_RPC_BASE_PATH
  export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
  export HF_HOME HF_DATASETS_CACHE HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
  export PYTORCH_ROCM_ARCH=""
  mkdir -p "${TMPDIR}" \
    "${TORCHINDUCTOR_CACHE_DIR}" \
    "${TRITON_CACHE_DIR}" \
    "${VLLM_CACHE_ROOT}" \
    "${XDG_CACHE_HOME}" \
    "${HF_HOME}" \
    "${HF_DATASETS_CACHE}" || return 1
  echo "Native compile caches: VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT} TORCHINDUCTOR_CACHE_DIR=${TORCHINDUCTOR_CACHE_DIR}"
  if [[ "${VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE:-0}" == "1" ]]; then
    if ! command -v findmnt >/dev/null 2>&1; then
      echo "findmnt is required to verify the native Hugging Face cache mount" >&2
      return 1
    fi
    hf_mount=$(findmnt -n -T "${HF_HOME}" -o TARGET 2>/dev/null || true)
    if [[ -z "${hf_mount}" || "${hf_mount}" == "/" ]]; then
      echo "Native CI requires a persistent volume mounted at or above ${HF_HOME}" >&2
      return 1
    fi
  fi
  if command -v findmnt >/dev/null 2>&1; then
    hf_fstype=$(findmnt -n -T "${HF_HOME}" -o FSTYPE 2>/dev/null || true)
  fi
  if [[ "${hf_fstype}" == nfs || "${hf_fstype}" == nfs4 ]]; then
    # Keep hf-xet state local and avoid vectored writes on shared NFS.
    export HF_XET_CACHE="${native_root}/cache/hf-xet"
    export HF_XET_HIGH_PERFORMANCE=0
    export HF_XET_RECONSTRUCTION_USE_VECTORED_WRITE=0
    mkdir -p "${HF_XET_CACHE}" || return 1
    echo "Configured hf-xet for shared ${hf_fstype} cache at ${HF_HOME}"
  fi
 }
 run_native_preflight() {
  local expected_gpus="${VLLM_CI_EXPECTED_GPU_COUNT:-1}"
  if [[ ! "${expected_gpus}" =~ ^[0-9]+$ ]]; then
    echo "Invalid VLLM_CI_EXPECTED_GPU_COUNT=${expected_gpus}" >&2
    return 1
  fi
  python3 -c "import encodings, importlib.metadata as im, importlib.util as iu; [im.version(d) for d in ('transformers', 'torch', 'ray', 'sympy', 'markupsafe', 'vllm')]; missing=[m for m in ('torch.utils.model_zoo', 'transformers.models.nomic_bert', 'ray.dag', 'sympy.physics', 'markupsafe._speedups') if iu.find_spec(m) is None]; assert not missing, missing" || return 1
  if [[ "${expected_gpus}" == "0" ]]; then
    echo "Native CPU-only AMD job: skipping ROCm device validation"
    return 0
  fi
  echo "--- ROCm info"
  rocminfo || return 1
  VLLM_CI_EXPECTED_GPU_COUNT="${expected_gpus}" python3 - <<'PY'
 import os
 import torch
 expected = int(os.environ["VLLM_CI_EXPECTED_GPU_COUNT"])
 assert torch.version.hip, "PyTorch is not a ROCm build"
 assert torch.cuda.is_available(), "ROCm GPU is not available to PyTorch"
 actual = torch.cuda.device_count()
 assert actual == expected, f"Expected {expected} ROCm GPU(s), found {actual}"
 PY
 }
 is_multi_node() {
  local cmds="$1"
  # Primary signal: NUM_NODES environment variable set by the pipeline
@@ -663,58 +338,7 @@ re_quote_pytest_markers() {
 # Main
 ###############################################################################
-if is_native_runtime; then
+# --- GPU initialization ---
  echo "--- Native in-pod ROCm CI (AMD_CI_RUNTIME=${AMD_CI_RUNTIME:-unset}, NATIVE_CI=${NATIVE_CI:-unset})"
  artifact_work_dir=""
  cleanup_native_workspace() {
    if [[ -n "${artifact_work_dir}" ]]; then
      rm -rf "${artifact_work_dir}"
    fi
  }
  trap cleanup_native_workspace EXIT
  if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
    commands="${VLLM_TEST_COMMANDS}"
    commands_source="env"
  else
    commands="$*"
    commands_source="argv"
    if [[ -z "$commands" ]]; then
      echo "Error: No test commands provided for native CI." >&2
      exit 1
    fi
  fi
  if [[ "$commands_source" == "argv" ]]; then
    commands=$(re_quote_pytest_markers "$commands")
  fi
  if is_multi_node "$commands"; then
    echo "Native CI does not support multi-node jobs yet."
    exit 1
  fi
  if ! initialize_native_environment; then
    echo "Failed to initialize the native test environment"
    exit 1
  fi
  if ! prepare_native_workspace; then
    echo "Failed to prepare native test workspace"
    exit 1
  fi
  export PYTHONPATH="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
  echo "Native test commands: $commands"
  run_native_preflight || exit 1
  # Keep AMD CI orchestration variables out of vLLM's runtime environment.
  clear_ci_orchestration_env
  /bin/bash -o pipefail -c "${commands}"
  handle_pytest_exit "$?"
 fi
 # --- GPU initialization for legacy Docker execution ---
 echo "--- ROCm info"
 rocminfo
@@ -816,27 +440,23 @@ fi
 echo "Final commands: $commands"
-standalone_merge_base_env=()
+# The ROCm test image often ships /vllm-workspace without .git (artifact tarball unpack).
-if [[ "$commands" == *python_only_compile.sh* ]]; then
+# tests/standalone_tests/python_only_compile.sh uses merge-base(HEAD, origin/main) for
-  # The ROCm test image often ships /vllm-workspace without .git. Resolve the
+# wheels.vllm.ai; compute on the agent (full git checkout) and pass into the container.
-  # wheels.vllm.ai commit from the agent checkout for this test only.
+vllm_standalone_merge_base=""
-  vllm_standalone_merge_base=""
+checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
-  checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
+if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
-  if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
+  checkout="."
    checkout="."
  fi
  # Pass safe.directory per-command because Buildkite uses mixed user IDs.
  if git -c "safe.directory=${checkout}" -C "${checkout}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
    vllm_standalone_merge_base="$(
      git -c "safe.directory=${checkout}" -C "${checkout}" merge-base HEAD origin/main 2>/dev/null || true
    )"
  fi
  if [[ -z "${vllm_standalone_merge_base}" ]]; then
    vllm_standalone_merge_base="${BUILDKITE_COMMIT:-}"
  fi
  echo "INFO: passing CI_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
  standalone_merge_base_env=(-e "CI_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}")
 fi
 if git -C "${checkout}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
  vllm_standalone_merge_base="$(
    git -C "${checkout}" merge-base HEAD origin/main 2>/dev/null || true
  )"
 fi
 if [[ -z "${vllm_standalone_merge_base}" ]]; then
  vllm_standalone_merge_base="${BUILDKITE_COMMIT:-}"
 fi
 echo "INFO: passing VLLM_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
 MYPYTHONPATH="/vllm-workspace"
@@ -867,7 +487,6 @@ else
 fi
 # --- Route: multi-node vs single-node ---
 clear_ci_orchestration_env
 if is_multi_node "$commands"; then
  echo "--- Multi-node job detected"
  export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
@@ -914,37 +533,14 @@ if is_multi_node "$commands"; then
 else
  echo "--- Single-node job"
  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
  docker_run_terminal_args=(-i)
  if [[ "${ROCM_DOCKER_TTY}" == "1" ]]; then
    docker_run_terminal_args+=(-t)
    echo "Docker interactive stdin: enabled; TTY allocation: enabled"
  else
    echo "Docker interactive stdin: enabled; TTY allocation: disabled"
  fi
  ulimit_core_hard=$(ulimit -H -c)
  if [[ "$ulimit_core_hard" == "unlimited" ]]; then
    # docker run can't pass "unlimited" to --ulimit
    ulimit_core_hard="-1"
  fi
   # Disable core dumps in the ROCm test container unless the ROCm debug agent is enabled
  coredump_flags="--ulimit core=0:$ulimit_core_hard"
  if [[ "$commands" == *"ROCm debug agent enabled"* ]]; then
    # Works around https://github.com/rocm/rocm-systems/issues/6206
    coredump_flags='-e HSA_COREDUMP_PATTERN="/tmp/gpucore.%p"'
  else
    echo "ROCm debug agent not enabled, coredumps are disabled in the test container."
  fi
  docker run \
    "${docker_run_terminal_args[@]}" \
    --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
    $RDMA_FLAGS \
    --network=host \
    --shm-size=16gb \
    --group-add "$render_gid" \
    --rm \
    $coredump_flags \
    -e HF_TOKEN \
    -e "HF_HUB_DOWNLOAD_TIMEOUT=${HF_HUB_DOWNLOAD_TIMEOUT}" \
    -e "HF_HUB_ETAG_TIMEOUT=${HF_HUB_ETAG_TIMEOUT}" \
@@ -952,13 +548,6 @@ else
    -e AWS_SECRET_ACCESS_KEY \
    -e BUILDKITE_PARALLEL_JOB \
    -e BUILDKITE_PARALLEL_JOB_COUNT \
    -e TERM \
    -e FORCE_COLOR \
    -e CLICOLOR_FORCE \
    -e PY_COLORS \
    -e PYTHONFAULTHANDLER \
    -e PYTEST_ADDOPTS \
    -e PYTEST_TIMEOUT \
    -v "${HF_CACHE}:${HF_MOUNT}" \
    -e "HF_HOME=${HF_MOUNT}" \
    -e "PYTHONPATH=${MYPYTHONPATH}" \
@@ -968,7 +557,7 @@ else
    -e "VLLM_CACHE_ROOT=${CONTAINER_CACHE_ROOT}/vllm" \
    -e "XDG_CACHE_HOME=${CONTAINER_CACHE_ROOT}/xdg" \
    -e "PYTORCH_ROCM_ARCH=" \
-    "${standalone_merge_base_env[@]}" \
+    -e "VLLM_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}" \
    --name "${container_name}" \
    "${image_name}" \
    /bin/bash -c "${CONTAINER_PREFLIGHT} && ${commands}"
@@ -1,11 +1,10 @@
 #!/bin/bash
 set -euox pipefail
-export VLLM_CPU_KVCACHE_SPACE=1
+export VLLM_CPU_KVCACHE_SPACE=1 
 export VLLM_CPU_CI_ENV=1
-# Skip torch.compile via vLLM's --enforce-eager flag (passed below) instead of
+# Reduce sub-processes for acceleration
-# TORCH_COMPILE_DISABLE=1, which torch 2.12 no longer treats as a silent no-op
+export TORCH_COMPILE_DISABLE=1 
 # when callers specify fullgraph=True.
 export VLLM_ENABLE_V1_MULTIPROCESSING=0
 SDE_ARCHIVE="sde-external-10.7.0-2026-02-18-lin.tar.xz"
@@ -50,15 +49,15 @@ wait_for_pid_and_check_log() {
 }
 # Test Sky Lake (AVX512F)
-./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_0.log 2>&1 &
+./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_0.log 2>&1 &
 PID_TEST_0=$!
 # Test Cascade Lake (AVX512F + VNNI)
-./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_1.log 2>&1 &
+./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_1.log 2>&1 &
 PID_TEST_1=$!
 # Test Cooper Lake (AVX512F + VNNI + BF16)
-./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_2.log 2>&1 &
+./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_2.log 2>&1 &
 PID_TEST_2=$!
 wait_for_pid_and_check_log $PID_TEST_0 test_0.log
@@ -38,11 +38,7 @@ function cpu_tests() {
    pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
    pytest -x -v -s tests/kernels/core/test_cpu_activation.py
    pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
-    pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
+    pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py"
    pytest -x -v -s tests/kernels/moe/test_cpu_int4_moe.py
    pytest -x -v -s tests/kernels/mamba/test_cpu_short_conv.py
    pytest -x -v -s tests/kernels/mamba/test_causal_conv1d.py
    pytest -x -v -s tests/kernels/mamba/test_mamba_ssm.py"
  # skip tests requiring model downloads if HF_TOKEN is not set
  # due to rate-limits
@@ -66,6 +62,7 @@ function cpu_tests() {
    set -e
    pytest -x -v -s tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs"
  # basic online serving
  docker exec cpu-test bash -c '
    set -e
@@ -99,4 +96,3 @@ function cpu_tests() {
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
 timeout 2h bash -c cpu_tests
@@ -21,7 +21,6 @@ case "${test_suite}" in
    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --kv-cache-dtype fp8
    python3 examples/basic/offline_inference/generate.py --model nvidia/Llama-3.1-8B-Instruct-FP8 --block-size 64 --enforce-eager --quantization modelopt --kv-cache-dtype fp8 --attention-backend TRITON_ATTN --max-model-len 4096
    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192
    python3 examples/basic/offline_inference/generate.py --model TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ --block-size 64 --enforce-eager 
    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2
    python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192
@@ -35,7 +34,6 @@ case "${test_suite}" in
    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
    pytest -v -s v1/structured_output
    pytest -v -s v1/test_serial_utils.py
    pytest -v -s v1/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]"
    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py
    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py
    ;;
@@ -360,7 +360,7 @@ export HF_TOKEN ZE_AFFINITY_MASK
    --ipc=host \
    --privileged \
    -v /dev/dri/by-path:/dev/dri/by-path \
-    -v "/data/huggingface:/root/.cache/huggingface" \
+    -v "${HOME}/.cache/huggingface:/root/.cache/huggingface" \
    --entrypoint='' \
    -e HF_TOKEN \
    -e ZE_AFFINITY_MASK \
@@ -369,7 +369,7 @@ export HF_TOKEN ZE_AFFINITY_MASK
    -e CMDS \
    --name "${container_name}" \
    "${IMAGE}" \
-    bash -c 'set -e; echo "ZE_AFFINITY_MASK is ${ZE_AFFINITY_MASK:-}"; eval "$CMDS"' \
+    bash -c 'set -e; source /opt/intel/oneapi/setvars.sh --force; source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force; echo "ZE_AFFINITY_MASK is ${ZE_AFFINITY_MASK:-}"; eval "$CMDS"' \
    >/dev/null
 } 9>/tmp/docker-pull.lock
@@ -85,7 +85,7 @@ RUN pip config set global.index-url http://cache-service-vllm.nginx-pypi-cache.s
 # Install for pytest to make the docker build cache layer always valid
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install pytest>=6.0  'modelscope<1.38'
+    pip install pytest>=6.0  modelscope
 WORKDIR /workspace/vllm
@@ -130,22 +130,6 @@ docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${ROCM_BASE_CACHE_KEY}-rocm
 docker push vllm/vllm-openai-rocm:latest-base
 docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
 # ---- XPU ----
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-xpu
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-xpu vllm/vllm-openai-xpu:latest-x86_64
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${COMMIT}-x86_64-xpu vllm/vllm-openai-xpu:v${RELEASE_VERSION}-x86_64
 docker push vllm/vllm-openai-xpu:latest-x86_64
 docker push vllm/vllm-openai-xpu:v${RELEASE_VERSION}-x86_64
 docker manifest rm vllm/vllm-openai-xpu:latest || true
 docker manifest rm vllm/vllm-openai-xpu:v${RELEASE_VERSION} || true
 docker manifest create vllm/vllm-openai-xpu:latest vllm/vllm-openai-xpu:latest-x86_64 --amend
 docker manifest create vllm/vllm-openai-xpu:v${RELEASE_VERSION} vllm/vllm-openai-xpu:v${RELEASE_VERSION}-x86_64 --amend
 docker manifest push vllm/vllm-openai-xpu:latest
 docker manifest push vllm/vllm-openai-xpu:v${RELEASE_VERSION}
 # ---- CPU ----
 # CPU images are behind separate block steps and may not have been built.
 # All-or-nothing: inspect both arches first, then either publish everything
@@ -1,32 +0,0 @@
 #!/usr/bin/env bash
 # Build the ROCm ci_base image, optionally from a freshly rebuilt ROCm base.
 set -euo pipefail
 metadata_get() {
    local key="$1"
    if command -v buildkite-agent >/dev/null 2>&1; then
        buildkite-agent meta-data get "${key}" 2>/dev/null || true
    fi
 }
 main() {
    local base_refreshed=""
    base_refreshed="$(metadata_get rocm-base-refresh)"
    if [[ "${base_refreshed}" == "1" ]]; then
        export BASE_IMAGE
        export CI_BASE_PUSH_STABLE_TAG
        BASE_IMAGE="$(metadata_get rocm-base-image)"
        CI_BASE_PUSH_STABLE_TAG="$(metadata_get rocm-base-push-stable-tag)"
        CI_BASE_PUSH_STABLE_TAG="${CI_BASE_PUSH_STABLE_TAG:-0}"
        echo "Using refreshed ROCm base image for ci_base: ${BASE_IMAGE}"
        echo "Push stable ci_base tag: ${CI_BASE_PUSH_STABLE_TAG}"
    fi
    bash .buildkite/scripts/ci-bake-rocm.sh ci-base-rocm-ci-with-deps
 }
 main "$@"
@@ -1,68 +0,0 @@
 #!/usr/bin/env bash
 # Build the ROCm CI test image or wheel artifact.
 #
 # When Dockerfile.rocm_base changes, always build the full image so downstream
 # ROCm tests can validate the freshly rebuilt base -> ci_base -> ci image chain.
 set -euo pipefail
 metadata_get() {
    local key="$1"
    if command -v buildkite-agent >/dev/null 2>&1; then
        buildkite-agent meta-data get "${key}" 2>/dev/null || true
    fi
 }
 use_ci_base_if_present() {
    local ci_base_image=""
    ci_base_image="$(metadata_get rocm-ci-base-image)"
    if [[ -z "${ci_base_image}" ]]; then
        return 1
    fi
    export CI_BASE_IMAGE="${ci_base_image}"
    echo "Using ROCm ci_base image selected by the preceding build step: ${CI_BASE_IMAGE}"
 }
 use_refreshed_base_if_present() {
    local base_refreshed=""
    base_refreshed="$(metadata_get rocm-base-refresh)"
    if [[ "${base_refreshed}" != "1" ]]; then
        return 1
    fi
    export BASE_IMAGE
    export IMAGE_TAG_LATEST
    BASE_IMAGE="$(metadata_get rocm-base-image)"
    IMAGE_TAG_LATEST="$(metadata_get rocm-ci-image-descriptive)"
    echo "Using refreshed ROCm base image for test image: ${BASE_IMAGE}"
    if [[ -n "${IMAGE_TAG_LATEST}" ]]; then
        echo "Also tagging full ROCm CI image as: ${IMAGE_TAG_LATEST}"
    fi
    return 0
 }
 main() {
    local base_refreshed=0
    use_ci_base_if_present || true
    if use_refreshed_base_if_present; then
        base_refreshed=1
    fi
    if [[ "${ROCM_CI_ARTIFACT_ONLY:-0}" == "1" && "${base_refreshed}" != "1" ]]; then
        echo "ROCM_CI_ARTIFACT_ONLY=1; building ROCm wheel artifact only"
        IMAGE_TAG="" bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-artifacts
        return
    fi
    bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-wheel
 }
 main "$@"
@@ -1,513 +0,0 @@
 #!/usr/bin/env bash
 # Build and publish a fresh ROCm base image when Dockerfile.rocm_base changes.
 #
 # Normal AMD CI builds should not pay for this path. The script no-ops unless
 # docker/Dockerfile.rocm_base changed relative to the branch base, the previous
 # main commit, or ROCM_BASE_REFRESH_FORCE=1 is set.
 set -euo pipefail
 DOCKERFILE="${ROCM_BASE_DOCKERFILE:-docker/Dockerfile.rocm_base}"
 BASE_REPO="${ROCM_BASE_IMAGE_REPO:-rocm/vllm-dev}"
 CI_IMAGE_REPO="${ROCM_CI_IMAGE_REPO:-rocm/vllm-ci}"
 BUILDER_NAME="${ROCM_BASE_BUILDER_NAME:-vllm-rocm-base-builder}"
 DEFAULT_ROCM_BASE_METADATA_VERSION="1"
 DEFAULT_ROCM_BASE_CONTENT_FILES="${DOCKERFILE}"
 DEFAULT_ROCM_BASE_CONTENT_ARGS="BASE_IMAGE TRITON_BRANCH TRITON_REPO PYTORCH_BRANCH PYTORCH_REPO PYTORCH_VISION_BRANCH PYTORCH_VISION_REPO PYTORCH_AUDIO_BRANCH PYTORCH_AUDIO_REPO FA_BRANCH FA_REPO AITER_BRANCH AITER_REPO MORI_BRANCH MORI_REPO PYTORCH_ROCM_ARCH PYTHON_VERSION USE_SCCACHE"
 metadata_set() {
    local key="$1"
    local value="$2"
    [[ -n "${value}" ]] || return 0
    if command -v buildkite-agent >/dev/null 2>&1; then
        buildkite-agent meta-data set "${key}" "${value}" || true
    fi
 }
 compute_content_hash() {
    local path=""
    local file=""
    for path in "$@"; do
        if [[ -d "${path}" ]]; then
            while IFS= read -r -d '' file; do
                printf 'file:%s\n' "${file}"
                sha256sum "${file}"
            done < <(find "${path}" -type f -print0 | sort -z)
        elif [[ -f "${path}" ]]; then
            printf 'file:%s\n' "${path}"
            sha256sum "${path}"
        else
            printf 'missing:%s\n' "${path}"
        fi
    done | sha256sum | cut -d' ' -f1
 }
 clean_docker_tag() {
    local input="$1"
    echo "${input}" | sed 's/[^a-zA-Z0-9._-]/_/g' | cut -c1-128
 }
 tag_component() {
    local input="$1"
    local max_chars="${2:-24}"
    clean_docker_tag "${input:-unknown}" | cut -c1-"${max_chars}"
 }
 extract_arg_default() {
    local arg_name="$1"
    sed -n -E "s/^[[:space:]]*ARG[[:space:]]+${arg_name}=\"?([^\"[:space:]]+)\"?.*/\\1/p" \
        "${DOCKERFILE}" | head -1
 }
 resolve_image_digest() {
    local image_ref="$1"
    docker buildx imagetools inspect "${image_ref}" 2>/dev/null \
        | sed -n -E 's/^Digest:[[:space:]]+//p' \
        | head -1 || true
 }
 resolve_rocm_base_arg_value() {
    local arg_name="$1"
    local use_sccache="$2"
    case "${arg_name}" in
        USE_SCCACHE)
            printf '%s\n' "${use_sccache}"
            ;;
        *)
            extract_arg_default "${arg_name}"
            ;;
    esac
 }
 hash_rocm_base_arg_values() {
    local use_sccache="$1"
    local base_image_digest="$2"
    local arg_name=""
    local arg_value=""
    shift 2 || true
    for arg_name in "$@"; do
        [[ -n "${arg_name}" ]] || continue
        arg_value=$(resolve_rocm_base_arg_value "${arg_name}" "${use_sccache}")
        printf 'arg:%s=%s\n' "${arg_name}" "${arg_value:-<empty>}"
        if [[ "${arg_name}" == "BASE_IMAGE" && -n "${arg_value}" ]]; then
            printf 'arg:%s.digest=%s\n' "${arg_name}" "${base_image_digest:-unknown}"
        fi
    done
 }
 rocm_version_from_base_image() {
    local base_image="$1"
    local version=""
    version="$(sed -n -E 's/.*:([0-9]+\.[0-9]+(\.[0-9]+)?)-.*/\1/p' <<<"${base_image}")"
    tag_component "${version:-${base_image}}" 16
 }
 git_diff_changed_base() {
    local range="$1"
    [[ -n "$(git diff --name-only "${range}" -- "${DOCKERFILE}" 2>/dev/null)" ]]
 }
 short_git_ref() {
    local ref="$1"
    git rev-parse --short "${ref}" 2>/dev/null || printf '%s\n' "${ref}"
 }
 extract_arg_default_from_ref() {
    local ref="$1"
    local arg_name="$2"
    local content=""
    content="$(git show "${ref}:${DOCKERFILE}" 2>/dev/null || true)"
    sed -n -E "s/^[[:space:]]*ARG[[:space:]]+${arg_name}=\"?([^\"[:space:]]+)\"?.*/\\1/p" \
        <<<"${content}" | head -1
 }
 log_arg_default_changes() {
    local old_ref="$1"
    local new_ref="$2"
    local content_args="${ROCM_BASE_CONTENT_ARGS:-${DEFAULT_ROCM_BASE_CONTENT_ARGS}}"
    local arg_name=""
    local old_value=""
    local new_value=""
    local changed=0
    echo "Changed ROCm base ARG defaults:"
    for arg_name in ${content_args}; do
        old_value="$(extract_arg_default_from_ref "${old_ref}" "${arg_name}")"
        new_value="$(extract_arg_default_from_ref "${new_ref}" "${arg_name}")"
        if [[ "${old_value}" != "${new_value}" ]]; then
            echo "  - ${arg_name}: ${old_value:-<unset>} -> ${new_value:-<unset>}"
            changed=1
        fi
    done
    if [[ "${changed}" == "0" ]]; then
        echo "  - none detected; Dockerfile instructions changed outside tracked ARG defaults"
    fi
 }
 log_arg_line_diff() {
    local range="$1"
    local arg_diff=""
    arg_diff="$(
        git diff --unified=0 "${range}" -- "${DOCKERFILE}" 2>/dev/null \
            | awk '/^[+-][[:space:]]*ARG[[:space:]]/ && $0 !~ /^(---|\+\+\+)/ { print "  " $0 }' \
            || true
    )"
    if [[ -n "${arg_diff}" ]]; then
        echo "Changed Dockerfile ARG lines:"
        printf '%s\n' "${arg_diff}"
    fi
 }
 log_rocm_base_change_check() {
    local context="$1"
    local range="$2"
    local old_ref="$3"
    local old_short=""
    local head_short=""
    old_short="$(short_git_ref "${old_ref}")"
    head_short="$(short_git_ref HEAD)"
    echo "--- :mag: ROCm base refresh check"
    echo "Context: ${context}"
    echo "Dockerfile: ${DOCKERFILE}"
    echo "Base revision: ${old_short}"
    echo "Head revision: ${head_short}"
    echo "Git diff range: ${range}"
 }
 log_rocm_base_rebuild_reason() {
    local context="$1"
    local range="$2"
    local old_ref="$3"
    local changed_files=""
    log_rocm_base_change_check "${context}" "${range}" "${old_ref}"
    changed_files="$(git diff --name-only "${range}" -- "${DOCKERFILE}" 2>/dev/null || true)"
    echo "Changed files:"
    if [[ -n "${changed_files}" ]]; then
        sed 's/^/  - /' <<<"${changed_files}"
    else
        echo "  - ${DOCKERFILE}"
    fi
    log_arg_default_changes "${old_ref}" HEAD
    log_arg_line_diff "${range}"
    echo "Decision: rebuilding ROCm base image because ${DOCKERFILE} changed."
 }
 rocm_base_changed_in_range() {
    local context="$1"
    local range="$2"
    local old_ref="$3"
    if git_diff_changed_base "${range}"; then
        log_rocm_base_rebuild_reason "${context}" "${range}" "${old_ref}"
        return 0
    fi
    log_rocm_base_change_check "${context}" "${range}" "${old_ref}"
    echo "Decision: ROCm base refresh not required; ${DOCKERFILE} is unchanged."
    return 1
 }
 rocm_base_changed() {
    local base_branch="${BUILDKITE_PULL_REQUEST_BASE_BRANCH:-main}"
    local base_ref="refs/remotes/origin/${base_branch}"
    local merge_base=""
    if [[ "${ROCM_BASE_REFRESH_SKIP:-0}" == "1" ]]; then
        echo "ROCM_BASE_REFRESH_SKIP=1 set; skipping ROCm base refresh"
        return 1
    fi
    if [[ "${ROCM_BASE_REFRESH_FORCE:-0}" == "1" ]]; then
        echo "ROCM_BASE_REFRESH_FORCE=1 set; refreshing ROCm base image"
        return 0
    fi
    if ! git rev-parse --is-inside-work-tree >/dev/null 2>&1; then
        echo "Not in a git checkout; skipping ROCm base refresh unless forced"
        return 1
    fi
    if [[ "${BUILDKITE_PULL_REQUEST:-false}" != "false" ]]; then
        git fetch --no-tags --depth=200 origin \
            "+refs/heads/${base_branch}:${base_ref}" >/dev/null 2>&1 || true
        merge_base=$(git merge-base HEAD "${base_ref}" 2>/dev/null || true)
        if [[ -z "${merge_base}" ]]; then
            echo "Unable to determine merge base with PR base ${base_ref}; skipping ROCm base refresh unless forced"
            return 1
        fi
        if rocm_base_changed_in_range \
            "pull request build against ${base_ref}" \
            "${merge_base}...HEAD" \
            "${merge_base}"; then
            return 0
        fi
    elif [[ "${BUILDKITE_BRANCH:-}" == "${ROCM_BASE_STABLE_BRANCH:-main}" ]] \
        && git rev-parse --verify HEAD~1 >/dev/null 2>&1; then
        if rocm_base_changed_in_range \
            "stable branch build; comparing against previous ${ROCM_BASE_STABLE_BRANCH:-main} commit" \
            "HEAD~1..HEAD" \
            "HEAD~1"; then
            return 0
        fi
    else
        git fetch --no-tags --depth=200 origin \
            "+refs/heads/${base_branch}:${base_ref}" >/dev/null 2>&1 || true
        merge_base=$(git merge-base HEAD "${base_ref}" 2>/dev/null || true)
        if [[ -z "${merge_base}" ]]; then
            echo "Unable to determine merge base with branch base ${base_ref}; skipping ROCm base refresh unless forced"
            return 1
        fi
        if rocm_base_changed_in_range \
            "branch build against ${base_ref}" \
            "${merge_base}...HEAD" \
            "${merge_base}"; then
            return 0
        fi
    fi
    return 1
 }
 should_push_stable_tag() {
    if [[ "${BUILDKITE_PULL_REQUEST:-false}" != "false" ]]; then
        return 1
    fi
    if [[ "${ROCM_BASE_PUSH_STABLE_TAG:-}" == "1" ]]; then
        return 0
    fi
    if [[ "${ROCM_BASE_PUSH_STABLE_TAG:-}" == "0" ]]; then
        return 1
    fi
    [[ "${BUILDKITE_PULL_REQUEST:-false}" == "false" \
        && "${BUILDKITE_BRANCH:-}" == "${ROCM_BASE_STABLE_BRANCH:-main}" ]]
 }
 setup_builder() {
    echo "--- :buildkite: Setting up buildx builder for ROCm base"
    if docker buildx inspect "${BUILDER_NAME}" >/dev/null 2>&1; then
        docker buildx use "${BUILDER_NAME}"
    else
        docker buildx create --name "${BUILDER_NAME}" --driver docker-container --use
    fi
    docker buildx inspect --bootstrap
 }
 compute_base_content_hash() {
    local use_sccache="$1"
    local base_image_digest="$2"
    local content_files="${ROCM_BASE_CONTENT_FILES:-${DEFAULT_ROCM_BASE_CONTENT_FILES}}"
    local content_args="${ROCM_BASE_CONTENT_ARGS:-${DEFAULT_ROCM_BASE_CONTENT_ARGS}}"
    local -a content_paths=()
    local -a content_arg_names=()
    read -r -a content_paths <<< "${content_files}"
    read -r -a content_arg_names <<< "${content_args}"
    {
        printf 'content-files-hash:%s\n' "$(compute_content_hash "${content_paths[@]}")"
        printf 'dockerfile:%s\n' "${DOCKERFILE}"
        printf 'resolved-build-args:\n'
        hash_rocm_base_arg_values \
            "${use_sccache}" "${base_image_digest}" "${content_arg_names[@]}"
    } | sha256sum | cut -d' ' -f1
 }
 build_base_image() {
    local use_sccache="${ROCM_BASE_USE_SCCACHE:-${USE_SCCACHE:-0}}"
    local base_hash=""
    local build_date=""
    local build_suffix=""
    local base_image_arg=""
    local base_image_digest=""
    local rocm_version=""
    local triton_arg=""
    local pytorch_arg=""
    local pytorch_vision_arg=""
    local pytorch_audio_arg=""
    local fa_arg=""
    local aiter_arg=""
    local mori_arg=""
    local python_version_arg=""
    local pytorch_rocm_arch_arg=""
    local pytorch_branch=""
    local aiter_branch=""
    local dependency_summary=""
    local descriptor=""
    local ci_descriptor=""
    local descriptive_tag=""
    local stable_tag="${BASE_REPO}:base"
    local ci_descriptive_tag=""
    local content_files="${ROCM_BASE_CONTENT_FILES:-${DEFAULT_ROCM_BASE_CONTENT_FILES}}"
    local content_args="${ROCM_BASE_CONTENT_ARGS:-${DEFAULT_ROCM_BASE_CONTENT_ARGS}}"
    local content_files_hash=""
    local metadata_version="${ROCM_BASE_METADATA_VERSION:-${DEFAULT_ROCM_BASE_METADATA_VERSION}}"
    local -a tags=()
    local -a no_cache_args=()
    local -a sccache_args=()
    local -a content_paths=()
    if [[ ! -f "${DOCKERFILE}" ]]; then
        echo "Error: ROCm base Dockerfile not found: ${DOCKERFILE}" >&2
        exit 1
    fi
    build_date="${ROCM_BASE_TAG_DATE:-$(date -u +%Y%m%d)}"
    if [[ -n "${BUILDKITE_BUILD_NUMBER:-}" ]]; then
        build_suffix="_bk_${BUILDKITE_BUILD_NUMBER}"
    fi
    base_image_arg="$(extract_arg_default BASE_IMAGE)"
    base_image_digest="$(resolve_image_digest "${base_image_arg}")"
    read -r -a content_paths <<< "${content_files}"
    content_files_hash="$(compute_content_hash "${content_paths[@]}")"
    base_hash=$(compute_base_content_hash "${use_sccache}" "${base_image_digest}")
    rocm_version="$(rocm_version_from_base_image "${base_image_arg}")"
    triton_arg="$(extract_arg_default TRITON_BRANCH)"
    pytorch_arg="$(extract_arg_default PYTORCH_BRANCH)"
    pytorch_vision_arg="$(extract_arg_default PYTORCH_VISION_BRANCH)"
    pytorch_audio_arg="$(extract_arg_default PYTORCH_AUDIO_BRANCH)"
    fa_arg="$(extract_arg_default FA_BRANCH)"
    aiter_arg="$(extract_arg_default AITER_BRANCH)"
    mori_arg="$(extract_arg_default MORI_BRANCH)"
    python_version_arg="$(extract_arg_default PYTHON_VERSION)"
    pytorch_rocm_arch_arg="$(extract_arg_default PYTORCH_ROCM_ARCH)"
    pytorch_branch="$(tag_component "${pytorch_arg}" 16)"
    aiter_branch="$(tag_component "${aiter_arg}" 24)"
    dependency_summary="base=${base_image_arg},rocm=${rocm_version},python=${python_version_arg},pytorch=${pytorch_arg},torchvision=${pytorch_vision_arg},torchaudio=${pytorch_audio_arg},triton=${triton_arg},flash-attn=${fa_arg},aiter=${aiter_arg},mori=${mori_arg},pytorch-rocm-arch=${pytorch_rocm_arch_arg}"
    descriptor="$(clean_docker_tag "base_custom_aiter_${aiter_branch}_torch_${pytorch_branch}_${build_date}${build_suffix}")"
    ci_descriptor="$(clean_docker_tag "ci_custom_aiter_${aiter_branch}_torch_${pytorch_branch}_${build_date}${build_suffix}")"
    descriptive_tag="${BASE_REPO}:${descriptor}"
    ci_descriptive_tag="${CI_IMAGE_REPO}:${ci_descriptor}"
    tags=(-t "${descriptive_tag}")
    if should_push_stable_tag; then
        tags+=(-t "${stable_tag}")
        metadata_set "rocm-base-push-stable-tag" "1"
    else
        metadata_set "rocm-base-push-stable-tag" "0"
    fi
    if [[ "${ROCM_BASE_NO_CACHE:-1}" == "1" ]]; then
        no_cache_args=(--no-cache)
    fi
    for env_name in \
        SCCACHE_DOWNLOAD_URL \
        SCCACHE_ENDPOINT \
        SCCACHE_BUCKET_NAME \
        SCCACHE_REGION_NAME \
        SCCACHE_S3_NO_CREDENTIALS; do
        if [[ -n "${!env_name:-}" ]]; then
            sccache_args+=(--build-arg "${env_name}=${!env_name}")
        fi
    done
    echo "--- :docker: Building ROCm base image"
    echo "Dockerfile: ${DOCKERFILE}"
    echo "Descriptive tag: ${descriptive_tag}"
    echo "Stable tag: ${stable_tag} ($(should_push_stable_tag && echo enabled || echo disabled))"
    echo "Content hash: ${base_hash}"
    echo "Dependency summary: ${dependency_summary}"
    echo "USE_SCCACHE: ${use_sccache}"
    docker buildx build \
        "${no_cache_args[@]}" \
        --pull \
        --progress "${BUILDKIT_PROGRESS:-plain}" \
        --file "${DOCKERFILE}" \
        --build-arg "USE_SCCACHE=${use_sccache}" \
        "${sccache_args[@]}" \
        --label "org.opencontainers.image.source=https://github.com/vllm-project/vllm" \
        --label "org.opencontainers.image.vendor=vLLM" \
        --label "org.opencontainers.image.title=vLLM ROCm base" \
        --label "org.opencontainers.image.revision=${BUILDKITE_COMMIT:-}" \
        --label "vllm.rocm_base.metadata_version=${metadata_version}" \
        --label "vllm.rocm_base.content_hash=${base_hash}" \
        --label "vllm.rocm_base.content_files_hash=${content_files_hash}" \
        --label "vllm.rocm_base.dockerfile=${DOCKERFILE}" \
        --label "vllm.rocm_base.image.descriptive=${descriptive_tag}" \
        --label "vllm.rocm_base.image.stable=${stable_tag}" \
        --label "vllm.rocm_base.git_commit=${BUILDKITE_COMMIT:-}" \
        --label "vllm.rocm_base.stable_branch=${ROCM_BASE_STABLE_BRANCH:-main}" \
        --label "vllm.rocm_base.descriptor=${descriptor}" \
        --label "vllm.rocm_base.dependency_summary=${dependency_summary}" \
        --label "vllm.rocm_base.base_image=${base_image_arg}" \
        --label "vllm.rocm_base.base_image_digest=${base_image_digest}" \
        --label "vllm.rocm_base.dependency.rocm=${rocm_version}" \
        --label "vllm.rocm_base.dependency.python=${python_version_arg}" \
        --label "vllm.rocm_base.dependency.pytorch=${pytorch_arg}" \
        --label "vllm.rocm_base.dependency.torchvision=${pytorch_vision_arg}" \
        --label "vllm.rocm_base.dependency.torchaudio=${pytorch_audio_arg}" \
        --label "vllm.rocm_base.dependency.triton=${triton_arg}" \
        --label "vllm.rocm_base.dependency.flash_attention=${fa_arg}" \
        --label "vllm.rocm_base.dependency.aiter=${aiter_arg}" \
        --label "vllm.rocm_base.dependency.mori=${mori_arg}" \
        --label "vllm.rocm_base.pytorch_rocm_arch=${pytorch_rocm_arch_arg}" \
        "${tags[@]}" \
        --push \
        .
    docker buildx imagetools inspect "${descriptive_tag}" >/dev/null
    metadata_set "rocm-base-refresh" "1"
    metadata_set "rocm-base-image" "${descriptive_tag}"
    metadata_set "rocm-base-image-descriptive" "${descriptive_tag}"
    metadata_set "rocm-base-image-stable" "${stable_tag}"
    metadata_set "rocm-base-image-ci-descriptive" "${ci_descriptive_tag}"
    metadata_set "rocm-base-metadata-version" "${metadata_version}"
    metadata_set "rocm-base-content-hash" "${base_hash}"
    metadata_set "rocm-base-content-files-hash" "${content_files_hash}"
    metadata_set "rocm-base-content-files" "${content_files}"
    metadata_set "rocm-base-content-args" "${content_args}"
    metadata_set "rocm-base-base-image-digest" "${base_image_digest}"
    metadata_set "rocm-base-dockerfile" "${DOCKERFILE}"
    metadata_set "rocm-base-descriptor" "${descriptor}"
    metadata_set "rocm-base-dependency-summary" "${dependency_summary}"
    metadata_set "rocm-base-dependency-rocm" "${rocm_version}"
    metadata_set "rocm-base-dependency-python" "${python_version_arg}"
    metadata_set "rocm-base-dependency-pytorch" "${pytorch_arg}"
    metadata_set "rocm-base-dependency-torchvision" "${pytorch_vision_arg}"
    metadata_set "rocm-base-dependency-torchaudio" "${pytorch_audio_arg}"
    metadata_set "rocm-base-dependency-triton" "${triton_arg}"
    metadata_set "rocm-base-dependency-flash-attention" "${fa_arg}"
    metadata_set "rocm-base-dependency-aiter" "${aiter_arg}"
    metadata_set "rocm-base-dependency-mori" "${mori_arg}"
    metadata_set "rocm-base-pytorch-rocm-arch" "${pytorch_rocm_arch_arg}"
    metadata_set "rocm-ci-image-descriptive" "${ci_descriptive_tag}"
    echo "--- :white_check_mark: ROCm base image published"
    echo "Use BASE_IMAGE=${descriptive_tag} for downstream ROCm CI builds"
 }
 main() {
    metadata_set "rocm-base-refresh" "0"
    if ! rocm_base_changed; then
        echo "ROCm base Dockerfile did not change; skipping base image refresh"
        return 0
    fi
    setup_builder
    build_base_image
 }
 main "$@"
@@ -1,32 +0,0 @@
 #!/usr/bin/env bash
 # Fast structural smoke test for the full ROCm CI image.
 set -euo pipefail
 image_ref="${VLLM_CI_SMOKE_IMAGE:-rocm/vllm-ci:${BUILDKITE_COMMIT:?BUILDKITE_COMMIT is required}}"
 docker run --rm --network=none --entrypoint /bin/bash "${image_ref}" -ec '
  if [ ! -d /vllm-workspace ]; then echo Missing directory: /vllm-workspace >&2; exit 1; fi
  if [ ! -d /vllm-workspace/tests ]; then echo Missing directory: /vllm-workspace/tests >&2; exit 1; fi
  if [ ! -d /vllm-workspace/src/vllm ]; then echo Missing directory: /vllm-workspace/src/vllm >&2; exit 1; fi
  if [ ! -x /vllm-workspace/src/vllm/vllm-rs ]; then echo Missing executable: /vllm-workspace/src/vllm/vllm-rs >&2; exit 1; fi
  command -v python3
  command -v uv
  command -v pytest
  if ! command -v amd-smi >/dev/null 2>&1 && ! command -v rocminfo >/dev/null 2>&1; then
    echo No ROCm CLI found in image >&2
    exit 1
  fi
  python3 - <<PY
 import torch
 import vllm
 print(torch.__version__)
 print(vllm.__version__)
 PY
  echo AMD image smoke OK
 '
@@ -109,9 +109,7 @@ run_nodes() {
        if [ "$node" -ne 0 ]; then
            docker exec -d "node$node" /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
        else
-            # Allocate a TTY (-t -i) for the foreground head node so its output
+            docker exec "node$node" /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
            # keeps ANSI color in the Buildkite log (see run-amd-test.sh).
            docker exec -t -i "node$node" /bin/bash -c "cd $WORKING_DIR ; ${COMMANDS[$node]}"
        fi
    done
 }
@@ -8,12 +8,7 @@ if [[ "$MODE" != "style-clippy" && "$MODE" != "test" ]]; then
  exit 2
 fi
-if ROOT_DIR="$(git rev-parse --show-toplevel 2>/dev/null)"; then
+ROOT_DIR="$(git rev-parse --show-toplevel)"
  :
 else
  SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
  ROOT_DIR="$(cd -- "${SCRIPT_DIR}/../.." && pwd -P)"
 fi
 cd "$ROOT_DIR"
 export CARGO_TERM_COLOR="${CARGO_TERM_COLOR:-always}"
@@ -21,20 +16,16 @@ export CARGO_HOME="${CARGO_HOME:-$HOME/.cargo}"
 export RUSTUP_HOME="${RUSTUP_HOME:-$HOME/.rustup}"
 export PATH="$CARGO_HOME/bin:$PATH"
 PROTOC_VERSION="${PROTOC_VERSION:-31.1}"
 CARGO_BINSTALL_VERSION="${CARGO_BINSTALL_VERSION:-1.20.1}"
 UV_VERSION="${UV_VERSION:-0.11.28}"
 PYO3_PYTHON_VERSION="${PYO3_PYTHON_VERSION:-3.12}"
 CARGO_SORT_VERSION_REQ="${CARGO_SORT_VERSION_REQ:-2}"
 CARGO_DENY_VERSION_REQ="${CARGO_DENY_VERSION_REQ:-0.20}"
 CARGO_NEXTEST_VERSION_REQ="${CARGO_NEXTEST_VERSION_REQ:-0.9}"
 log_section() {
  echo "--- $*"
 }
 install_protoc() {
  if command -v protoc >/dev/null 2>&1; then
    return
  fi
  local version="${PROTOC_VERSION:-31.1}"
  local arch
  case "$(uname -m)" in
    x86_64)
@@ -49,17 +40,16 @@ install_protoc() {
      ;;
  esac
-  local url="https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-${arch}.zip"
+  local url="https://github.com/protocolbuffers/protobuf/releases/download/v${version}/protoc-${version}-linux-${arch}.zip"
  local tmp_dir
  tmp_dir="$(mktemp -d)"
-  log_section "Installing protoc ${PROTOC_VERSION}"
+  log_section "Installing protoc ${version}"
  curl -L --proto '=https' --tlsv1.2 -sSf "$url" -o "$tmp_dir/protoc.zip"
  mkdir -p "$CARGO_HOME/bin"
  unzip -q "$tmp_dir/protoc.zip" bin/protoc 'include/*' -d "$CARGO_HOME"
  chmod +x "$CARGO_HOME/bin/protoc"
  rm -rf "$tmp_dir"
  protoc --version
 }
 rust_toolchain() {
@@ -80,48 +70,66 @@ install_rust_toolchain() {
 }
 install_cargo_binstall() {
-  log_section "Installing cargo-binstall ${CARGO_BINSTALL_VERSION}"
+  if command -v cargo-binstall >/dev/null 2>&1; then
    return
  fi
  log_section "Installing cargo-binstall"
  curl -L --proto '=https' --tlsv1.2 -sSf \
-    "https://raw.githubusercontent.com/cargo-bins/cargo-binstall/v${CARGO_BINSTALL_VERSION}/install-from-binstall-release.sh" \
+    https://raw.githubusercontent.com/cargo-bins/cargo-binstall/main/install-from-binstall-release.sh \
-    | env BINSTALL_VERSION="$CARGO_BINSTALL_VERSION" bash
+    | bash
  cargo-binstall -V
 }
 install_cargo_sort() {
-  log_section "Installing cargo-sort ${CARGO_SORT_VERSION_REQ}"
+  if command -v cargo-sort >/dev/null 2>&1; then
-  cargo binstall --no-confirm --force "cargo-sort@${CARGO_SORT_VERSION_REQ}"
+    return
  fi
  log_section "Installing cargo-sort"
  install_cargo_binstall
  cargo binstall --no-confirm cargo-sort
 }
 install_cargo_deny() {
-  log_section "Installing cargo-deny ${CARGO_DENY_VERSION_REQ}"
+  if command -v cargo-deny >/dev/null 2>&1; then
-  cargo binstall --no-confirm --force "cargo-deny@${CARGO_DENY_VERSION_REQ}"
+    return
  fi
  log_section "Installing cargo-deny"
  install_cargo_binstall
  cargo binstall --no-confirm cargo-deny
 }
 install_cargo_nextest() {
-  log_section "Installing cargo-nextest ${CARGO_NEXTEST_VERSION_REQ}"
+  if command -v cargo-nextest >/dev/null 2>&1; then
-  cargo binstall \
+    return
-    --no-confirm \
+  fi
-    --force \
+
-    --secure \
+  log_section "Installing cargo-nextest"
-    "cargo-nextest@${CARGO_NEXTEST_VERSION_REQ}"
+  install_cargo_binstall
  cargo binstall --no-confirm --secure cargo-nextest
 }
 install_uv() {
-  log_section "Installing uv ${UV_VERSION}"
+  if command -v uv >/dev/null 2>&1; then
-  curl -L --proto '=https' --tlsv1.2 -sSf \
+    return
-    "https://github.com/astral-sh/uv/releases/download/${UV_VERSION}/uv-installer.sh" \
+  fi
  log_section "Installing uv"
  curl -LsSf --proto '=https' --tlsv1.2 https://astral.sh/uv/install.sh \
    | env UV_INSTALL_DIR="$CARGO_HOME/bin" sh
  uv --version
 }
 setup_pyo3_python() {
-  log_section "Installing Python ${PYO3_PYTHON_VERSION} for PyO3 tests"
+  local python_version="${PYO3_PYTHON_VERSION:-3.12}"
-  uv python install "$PYO3_PYTHON_VERSION"
+
  log_section "Installing Python ${python_version} for PyO3 tests"
  uv python install "$python_version"
  PYO3_PYTHON="$(uv python find \
    --managed-python \
    --no-project \
    --resolve-links \
-    "$PYO3_PYTHON_VERSION")"
+    "$python_version")"
  export PYO3_PYTHON
  local python_libdir
@@ -143,7 +151,6 @@ PY
 }
 run_style_clippy() {
  install_cargo_binstall
  install_cargo_sort
  install_cargo_deny
@@ -156,8 +163,8 @@ run_style_clippy() {
  log_section "Checking Rust dependency bans"
  cargo deny \
    --manifest-path rust/Cargo.toml \
    --config rust/deny.toml \
    check \
    --config rust/deny.toml \
    bans
  log_section "Running clippy"
@@ -174,7 +181,6 @@ run_style_clippy() {
 run_tests() {
  install_uv
  setup_pyo3_python
  install_cargo_binstall
  install_cargo_nextest
  log_section "Running cargo nextest"
@@ -18,10 +18,6 @@ wait_for_server() {
 MODEL="Qwen/Qwen3-30B-A3B-FP8"
 BACK="allgather_reducescatter"
 if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
  # Disable MOE padding for ROCm since it is causing eplb to fail.
  export VLLM_ROCM_MOE_PADDING=0
 fi
 cleanup() {
  if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
@@ -23,7 +23,6 @@ NC='\033[0m' # No Color
 # Default configuration
 PIPELINE="ci"
 DRY_RUN=true
 TORCH_NIGHTLY=false
 usage() {
    cat <<EOF
@@ -35,14 +34,12 @@ Sets RUN_ALL=1 and NIGHTLY=1 environment variables.
 SAFETY: Dry-run by default. Use --execute to actually trigger a build.
 Options:
-    --execute        Actually trigger the build (default: dry-run)
+    --execute       Actually trigger the build (default: dry-run)
-    --pipeline       Buildkite pipeline slug (default: ${PIPELINE})
+    --pipeline      Buildkite pipeline slug (default: ${PIPELINE})
-    --commit         Override commit SHA (default: current HEAD)
+    --commit        Override commit SHA (default: current HEAD)
-    --branch         Override branch name (default: current branch)
+    --branch        Override branch name (default: current branch)
-    --message        Custom build message (default: auto-generated)
+    --message       Custom build message (default: auto-generated)
-    --torch-nightly  Also build and run the full suite against torch nightly
+    --help          Show this help message
                     (sets TORCH_NIGHTLY=1)
    --help           Show this help message
 Prerequisites:
    - bk CLI installed: brew tap buildkite/buildkite && brew install buildkite/buildkite/bk
@@ -52,7 +49,6 @@ Examples:
    $(basename "$0")                        # Dry-run, show what would happen
    $(basename "$0") --execute              # Actually trigger the build
    $(basename "$0") --pipeline ci-shadow   # Dry-run with different pipeline
    $(basename "$0") --torch-nightly        # Dry-run a full torch-nightly run
 EOF
    exit 1
 }
@@ -100,10 +96,6 @@ while [[ $# -gt 0 ]]; do
            MESSAGE="$2"
            shift 2
            ;;
        --torch-nightly)
            TORCH_NIGHTLY=true
            shift
            ;;
        --help|-h)
            usage
            ;;
@@ -179,17 +171,11 @@ if [[ $(echo "$REMOTE_BRANCHES" | wc -l) -gt 5 ]]; then
 fi
 echo ""
 # Environment variables passed to the build.
 BUILD_ENV=("RUN_ALL=1" "NIGHTLY=1")
 if [[ "$TORCH_NIGHTLY" == true ]]; then
    BUILD_ENV+=("TORCH_NIGHTLY=1")
 fi
 log_info "Pipeline: ${PIPELINE}"
 log_info "Branch: ${BRANCH}"
 log_info "Commit: ${COMMIT}"
 log_info "Message: ${MESSAGE}"
-log_info "Environment: ${BUILD_ENV[*]}"
+log_info "Environment: RUN_ALL=1, NIGHTLY=1"
 echo ""
 # Build the command
@@ -201,10 +187,9 @@ CMD=(bk build create
    --commit "${COMMIT}"
    --branch "${BRANCH}"
    --message "${MESSAGE}"
    --env "RUN_ALL=1"
    --env "NIGHTLY=1"
 )
 for env_var in "${BUILD_ENV[@]}"; do
    CMD+=(--env "${env_var}")
 done
 if [[ "$DRY_RUN" == true ]]; then
    echo "=========================================="
@@ -225,14 +210,8 @@ if [[ "$DRY_RUN" == true ]]; then
    echo "    --commit '$(escape_for_shell "${COMMIT}")' \\"
    echo "    --branch '$(escape_for_shell "${BRANCH}")' \\"
    echo "    --message '$(escape_for_shell "${MESSAGE}")' \\"
-    last_idx=$(( ${#BUILD_ENV[@]} - 1 ))
+    echo "    --env 'RUN_ALL=1' \\"
-    for i in "${!BUILD_ENV[@]}"; do
+    echo "    --env 'NIGHTLY=1'"
        if [[ $i -eq $last_idx ]]; then
            echo "    --env '$(escape_for_shell "${BUILD_ENV[$i]}")'"
        else
            echo "    --env '$(escape_for_shell "${BUILD_ENV[$i]}")' \\"
        fi
    done
    echo ""
    echo "=========================================="
    echo -e "${YELLOW}To actually trigger this build, run:${NC}"
@@ -6,14 +6,8 @@ set -ex
 # manylinux platform tag with auditwheel.
 # Index generation is handled separately by generate-and-upload-nightly-index.sh.
-# auditwheel is Linux-only; macOS wheels already carry a valid tag, so skip the
+# shellcheck source=lib/manylinux.sh
-# manylinux retag for them.
+source .buildkite/scripts/lib/manylinux.sh
 WHEEL_PLATFORM="${VLLM_WHEEL_PLATFORM:-linux}"
 if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
  # shellcheck source=lib/manylinux.sh
  source .buildkite/scripts/lib/manylinux.sh
 fi
 BUCKET="vllm-wheels"
 SUBPATH=$BUILDKITE_COMMIT
@@ -33,10 +27,8 @@ wheel="${wheel_files[0]}"
 # ========= detect manylinux tag and rename ==========
-if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
+wheel="$(apply_manylinux_tag "$wheel")"
-  wheel="$(apply_manylinux_tag "$wheel")"
+echo "Renamed wheel to: $wheel"
  echo "Renamed wheel to: $wheel"
 fi
 # Extract the version from the wheel
 version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
@@ -113,8 +113,8 @@ $PYTHON .buildkite/scripts/generate-nightly-index.py \
 echo "Uploading indices to $S3_COMMIT_PREFIX"
 aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
-# Only scheduled nightly builds should update the moving nightly index.
+# Update rocm/nightly/ if on main branch and not a PR
-if [[ "${NIGHTLY:-0}" == "1" ]]; then
+if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]] || [[ "$NIGHTLY" == "1" ]]; then
    echo "Updating rocm/nightly/ index..."
    aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/rocm/nightly/"
 fi
@@ -147,7 +147,7 @@ echo ""
 echo "Install command (by commit):"
 echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/$ROCM_SUBPATH/"
 echo ""
-if [[ "${NIGHTLY:-0}" == "1" ]]; then
+if [[ "$BUILDKITE_BRANCH" == "main" ]] || [[ "$NIGHTLY" == "1" ]]; then
    echo "Install command (nightly):"
    echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/rocm/nightly/"
 fi
@@ -1,13 +0,0 @@
 #!/bin/bash
 set -euo pipefail
 REGISTRY="public.ecr.aws/q9t5s3a7"
 REPO="vllm-release-repo"
 ARCH_TAG="${BUILDKITE_COMMIT}-$(uname -m)-xpu"
 PLATFORM_TAG="${BUILDKITE_COMMIT}-xpu"
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin ${REGISTRY}
 docker manifest rm ${REGISTRY}/${REPO}:${PLATFORM_TAG} || true
 docker manifest create ${REGISTRY}/${REPO}:${PLATFORM_TAG} ${REGISTRY}/${REPO}:${ARCH_TAG} --amend
 docker manifest push ${REGISTRY}/${REPO}:${PLATFORM_TAG}
@@ -1,21 +0,0 @@
 #!/bin/bash
 set -ex
 ORIG_TAG_NAME="$BUILDKITE_COMMIT"
 REPO="vllm/vllm-openai-xpu"
 echo "Pushing original XPU tag ${ORIG_TAG_NAME}-xpu to nightly tags in ${REPO}"
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64-xpu
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64-xpu ${REPO}:nightly-x86_64
 docker push ${REPO}:nightly-x86_64
 docker manifest rm ${REPO}:nightly || true
 docker manifest rm ${REPO}:nightly-"$BUILDKITE_COMMIT" || true
 docker manifest create ${REPO}:nightly ${REPO}:nightly-x86_64 --amend
 docker manifest create ${REPO}:nightly-"$BUILDKITE_COMMIT" ${REPO}:nightly-x86_64 --amend
 docker manifest push ${REPO}:nightly
 docker manifest push ${REPO}:nightly-"$BUILDKITE_COMMIT"
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: V1 attention (H100-MI300)
  key: v1-attention-h100-mi300
-  timeout_in_minutes: 85
+  timeout_in_minutes: 30
  device: h100
  source_file_dependencies:
    - vllm/config/attention.py
@@ -12,13 +12,11 @@ steps:
    - vllm/v1/attention
    - tests/v1/attention
  commands:
-    - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s v1/attention
  parallelism: 2
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 70
      timeout_in_minutes: 125
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -32,7 +30,7 @@ steps:
 - label: V1 attention (B200)
  key: v1-attention-b200
-  timeout_in_minutes: 80
+  timeout_in_minutes: 30
  device: b200-k8s
  source_file_dependencies:
    - vllm/config/attention.py
@@ -40,5 +38,4 @@ steps:
    - vllm/v1/attention
    - tests/v1/attention
  commands:
-    - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s v1/attention
  parallelism: 2
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Correctness
  key: basic-correctness
-  timeout_in_minutes: 68
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -18,8 +18,7 @@ steps:
  - pytest -v -s basic_correctness/test_cpu_offload.py
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 50
      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Benchmarks CLI Test
  key: benchmarks-cli-test
-  timeout_in_minutes: 45
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -13,9 +13,7 @@ steps:
  - pytest -v -s benchmarks/
  mirror:
    amd:
      dind: false
      device: mi300_1
      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
@@ -25,7 +23,7 @@ steps:
  num_gpus: 2
  optional: true
  working_dir: "/vllm-workspace/"
-  timeout_in_minutes: 20
+  timeout_in_minutes: 10
  source_file_dependencies:
  - benchmarks/attention_benchmarks/
  - vllm/v1/attention/
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Sequence Parallel Correctness Tests (2 GPUs)
  key: sequence-parallel-correctness-tests-2-gpus
-  timeout_in_minutes: 80
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/"
  num_devices: 2
  source_file_dependencies:
@@ -19,7 +19,7 @@ steps:
 - label: Sequence Parallel Correctness Tests (2xH100)
  key: sequence-parallel-correctness-tests-2xh100
-  timeout_in_minutes: 75
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/"
  device: h100
  optional: true
@@ -30,7 +30,7 @@ steps:
 - label: AsyncTP Correctness Tests (2xH100)
  key: asynctp-correctness-tests-2xh100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/"
  device: h100
  optional: true
@@ -41,7 +41,7 @@ steps:
 - label: AsyncTP Correctness Tests (B200)
  key: asynctp-correctness-tests-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  optional: true
@@ -52,7 +52,7 @@ steps:
 - label: Distributed Compile Unit Tests (2xH100)
  key: distributed-compile-unit-tests-2xh100
-  timeout_in_minutes: 45
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/"
  device: h100
  num_devices: 2
@@ -66,7 +66,7 @@ steps:
 - label: Fusion and Compile Unit Tests (2xB200)
  key: fusion-and-compile-unit-tests-2xb200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  source_file_dependencies:
@@ -96,7 +96,7 @@ steps:
 - label: Fusion E2E Quick (H100)
  key: fusion-e2e-quick-h100
-  timeout_in_minutes: 25
+  timeout_in_minutes: 15
  working_dir: "/vllm-workspace/"
  device: h100
  num_devices: 1
@@ -115,7 +115,7 @@ steps:
 - label: Fusion E2E Config Sweep (H100)
  key: fusion-e2e-config-sweep-h100
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/"
  device: h100
  num_devices: 1
@@ -149,7 +149,7 @@ steps:
 - label: Fusion E2E TP2 Quick (H100)
  key: fusion-e2e-tp2-quick-h100
-  timeout_in_minutes: 35
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/"
  device: h100
  num_devices: 2
@@ -167,7 +167,7 @@ steps:
 - label: Fusion E2E TP2 AR-RMS Config Sweep (H100)
  key: fusion-e2e-tp2-ar-rms-config-sweep-h100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  working_dir: "/vllm-workspace/"
  device: h100
  num_devices: 2
@@ -207,7 +207,7 @@ steps:
 - label: Fusion E2E TP2 (B200)
  key: fusion-e2e-tp2-b200
-  timeout_in_minutes: 45
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  num_devices: 2
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Platform Tests
  key: platform-tests
-  timeout_in_minutes: 20
+  timeout_in_minutes: 15
  device: h200_18gb
  source_file_dependencies:
  - vllm/envs.py
@@ -16,21 +16,16 @@ steps:
  commands:
    - pytest -v -s cuda/test_cuda_context.py
    - pytest -v -s cuda/test_platform_no_cuda_init.py
    - pytest -v -s cuda/test_cuda_compatibility_path.py
 - label: Cudagraph
  device: h200_35gb
  key: cudagraph
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  source_file_dependencies:
  - tests/v1/cudagraph
  - vllm/v1/cudagraph_dispatcher.py
  - vllm/config/compilation.py
  - vllm/compilation
  - vllm/v1/worker/encoder_cudagraph.py
  - vllm/v1/worker/encoder_cudagraph_defs.py
  commands:
    - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
    - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
-    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py
+    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py
    - pytest -v -s v1/cudagraph/test_encoder_cudagraph.py
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed NixlConnector PD accuracy (4 GPUs)
  key: distributed-nixlconnector-pd-accuracy-4-gpus
-  timeout_in_minutes: 55
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -15,9 +15,8 @@ steps:
    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 60
+      timeout_in_minutes: 110
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -30,7 +29,7 @@ steps:
 - label: Distributed FlashInfer NixlConnector PD accuracy (4 GPUs)
  key: distributed-flashinfer-nixlconnector-pd-accuracy-4-gpus
-  timeout_in_minutes: 55
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -40,19 +39,6 @@ steps:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - FLASHINFER=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 - label: Push NixlConnector PP prefill PD accuracy (4 GPUs)
  key: push-nixlconnector-pp-prefill-pd-accuracy-4-gpus
  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
    - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
    - tests/v1/kv_connector/nixl_integration/
    - tests/v1/kv_connector/nixl_push_integration/
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - bash v1/kv_connector/nixl_push_integration/config_sweep_accuracy_test.sh
 - label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
  key: dp-ep-distributed-nixlconnector-pd-accuracy-tests-4-gpus
  timeout_in_minutes: 30
@@ -66,9 +52,8 @@ steps:
    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 40
+      timeout_in_minutes: 50
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -81,7 +66,7 @@ steps:
 - label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
  key: crosslayer-kv-layout-distributed-nixlconnector-pd-accuracy-tests-4-gpus
-  timeout_in_minutes: 55
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -92,9 +77,8 @@ steps:
    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 60
+      timeout_in_minutes: 110
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -107,7 +91,7 @@ steps:
 - label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs)
  key: hybrid-ssm-nixlconnector-pd-accuracy-tests-4-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -118,9 +102,8 @@ steps:
    - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 55
+      timeout_in_minutes: 60
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -131,22 +114,6 @@ steps:
        - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
        - HYBRID_SSM=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 - label: NixlConnector PD edge case test (2 GPUs)
  key: nixlconnector-pd-edge-cases-2-gpus
  timeout_in_minutes: 40
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
    - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
    - vllm/v1/core/sched/
    - tests/v1/kv_connector/nixl_integration/
  env:
    PREFILL_GPU_ID: "0"
    DECODE_GPU_ID: "1"
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - bash v1/kv_connector/nixl_integration/run_edge_case_test.sh
 - label: Hybrid SSM NixlConnector PD prefix cache test (2 GPUs)
  key: hybrid-ssm-nixlconnector-pd-prefix-cache-2-gpus
  timeout_in_minutes: 25
@@ -163,7 +130,7 @@ steps:
 - label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
  key: multiconnector-nixl-offloading-pd-accuracy-2-gpus
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -178,7 +145,7 @@ steps:
 - label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
  key: nixlconnector-pd-spec-decode-acceptance-2-gpus
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: a100
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
@@ -191,9 +158,8 @@ steps:
    - bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
  mirror:
    amd:
      dind: false
      device: mi300_2
-      timeout_in_minutes: 45
+      timeout_in_minutes: 60
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -203,11 +169,11 @@ steps:
        - vllm/platforms/rocm.py
      commands:
        - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-        - KV_CACHE_MEMORY_BYTES=8G ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
+        - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
 - label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
  key: multiconnector-nixl-offloading-pd-edge-cases-2-gpus
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -219,25 +185,3 @@ steps:
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
 # P TP 4 - D DPEP 4 test case for DSv4-Flash
 - label: DSv4-Flash Disaggregated DP EP
  key: dsv4-flash-disaggregated
  timeout_in_minutes: 60
  device: h200
  optional: true
  working_dir: "/vllm-workspace/tests"
  num_devices: 8
  env:
    ENABLE_HMA_FLAG: "1"
    DP_EP: "1"
    GPU_MEMORY_UTILIZATION: "0.85"
    PREFILLER_TP_SIZE: "4"
    DECODER_TP_SIZE: "4"
    PREFILL_BLOCK_SIZE: "256"
    DECODE_BLOCK_SIZE: "256"
    MODEL_NAMES: "deepseek-ai/DeepSeek-V4-Flash"
    VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--kv-cache-dtype,fp8"
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Comm Ops
  key: distributed-comm-ops
-  timeout_in_minutes: 25
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -18,7 +18,7 @@ steps:
 - label: Distributed DP Tests (2 GPUs)
  key: distributed-dp-tests-2-gpus
-  timeout_in_minutes: 35
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -39,9 +39,7 @@ steps:
  - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
  mirror:
    amd:
      dind: false
      device: mi300_2
      timeout_in_minutes: 45
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -57,7 +55,7 @@ steps:
 - label: Distributed Compile + RPC Tests (2 GPUs)
  key: distributed-compile-rpc-tests-2-gpus
-  timeout_in_minutes: 65
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -80,7 +78,7 @@ steps:
 - label: Distributed Torchrun + Shutdown Tests (2 GPUs)
  key: distributed-torchrun-shutdown-tests-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -135,7 +133,7 @@ steps:
 - label: Distributed DP Tests (4 GPUs)
  key: distributed-dp-tests-4-gpus
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -156,7 +154,7 @@ steps:
 - label: Distributed Compile + Comm (4 GPUs)
  key: distributed-compile-comm-4-gpus
-  timeout_in_minutes: 70
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -178,7 +176,7 @@ steps:
 - label: Distributed Tests (8xH100)
  key: distributed-tests-8xh100
-  timeout_in_minutes: 20
+  timeout_in_minutes: 10
  device: h100
  num_devices: 8
  working_dir: "/vllm-workspace/tests"
@@ -214,7 +212,7 @@ steps:
 - label: Distributed Tests (2xH100-2xMI300)
  key: distributed-tests-2xh100-2xmi300
-  timeout_in_minutes: 30
+  timeout_in_minutes: 15
  device: h100
  optional: true
  working_dir: "/vllm-workspace/"
@@ -235,7 +233,7 @@ steps:
  num_devices: 2
  commands:
    - pytest -v -s tests/distributed/test_context_parallel.py
-    - pytest -v -s tests/distributed/test_nccl_symm_mem.py
+    - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
    - pytest -v -s tests/v1/distributed/test_dbo.py
    - pytest -v -s tests/distributed/test_mnnvl_alltoall.py
@@ -261,7 +259,7 @@ steps:
 - label: Pipeline + Context Parallelism (4 GPUs)
  key: pipeline-context-parallelism-4-gpus
-  timeout_in_minutes: 55
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -276,7 +274,7 @@ steps:
 - label: RayExecutorV2 (4 GPUs)
  key: rayexecutorv2-4-gpus
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -3,7 +3,7 @@ depends_on:
  - image-build-cpu
 steps:
 - label: Docker Build Metadata
-  timeout_in_minutes: 20
+  timeout_in_minutes: 10
  device: cpu-small
  source_file_dependencies:
    - .buildkite/release-pipeline.yaml
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: DeepSeek V2-Lite Sync EPLB Accuracy (4xH100)
  key: deepseek-v2-lite-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 4
@@ -14,7 +14,7 @@ steps:
 - label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (4xH100)
  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 4
@@ -24,7 +24,7 @@ steps:
 - label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (2xB200)
  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-2xb200
-  timeout_in_minutes: 20
+  timeout_in_minutes: 60
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -34,7 +34,7 @@ steps:
 - label: Qwen3-30B-A3B-FP8 DP4 Async EPLB Accuracy
  key: qwen3-30b-a3b-fp8-dp4-async-eplb-accuracy
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 4
@@ -44,7 +44,7 @@ steps:
 - label: DeepSeek V2-Lite Prefetch Offload Accuracy (H100)
  key: deepseek-v2-lite-prefetch-offload-accuracy-h100
-  timeout_in_minutes: 20
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 1
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Engine
  key: engine
-  timeout_in_minutes: 30
+  timeout_in_minutes: 15
  device: h200_18gb
  source_file_dependencies:
  - vllm/compilation/
@@ -28,33 +28,30 @@ steps:
  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 60
      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
 - label: Engine (1 GPU)
  key: engine-1-gpu
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  source_file_dependencies:
    - vllm/v1/engine/
    - tests/v1/engine/
    - tests/v1/test_tensor_ipc_queue.py
  commands:
    - pytest -v -s v1/engine/test_preprocess_error_handling.py
    - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
    - pytest -v -s v1/test_tensor_ipc_queue.py
  mirror:
    amd:
-      device: mi250_1
+      device: mi325_1
-      timeout_in_minutes: 45
+      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
 - label: e2e Scheduling (1 GPU)
  key: e2e-scheduling-1-gpu
-  timeout_in_minutes: 53
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/
@@ -64,14 +61,14 @@ steps:
  mirror:
    amd:
      device: mi250_1
-      timeout_in_minutes: 55
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
 - label: e2e Core (1 GPU)
  device: h200_35gb
  key: e2e-core-1-gpu
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  source_file_dependencies:
    - vllm/v1/
    - tests/v1/e2e/general/
@@ -80,7 +77,7 @@ steps:
  mirror:
    amd:
      device: mi250_1
-      timeout_in_minutes: 50
+      timeout_in_minutes: 35
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -90,7 +87,7 @@ steps:
 - label: V1 e2e (2 GPUs)
  key: v1-e2e-2-gpus
-  timeout_in_minutes: 25 # TODO: Fix timeout after we have more confidence in the test stability
+  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
  optional: true
  num_devices: 2
  source_file_dependencies:
@@ -117,15 +114,13 @@ steps:
    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
  mirror:
    amd:
      dind: false
      device: mi300_2
      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
 - label: V1 e2e (4 GPUs)
  key: v1-e2e-4-gpus
-  timeout_in_minutes: 20 # TODO: Fix timeout after we have more confidence in the test stability
+  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
  optional: true
  num_devices: 4
  source_file_dependencies:
@@ -153,7 +148,7 @@ steps:
 - label: V1 e2e (4xH100)
  key: v1-e2e-4xh100
-  timeout_in_minutes: 35
+  timeout_in_minutes: 60
  device: h100
  num_devices: 4
  optional: true
@@ -3,9 +3,8 @@ depends_on:
  - image-build
 steps:
 - label: Entrypoints Unit Tests
  device: h200_35gb
  key: entrypoints-unit-tests
-  timeout_in_minutes: 25
+  timeout_in_minutes: 10
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/entrypoints
@@ -16,9 +15,8 @@ steps:
  - pytest -v -s entrypoints/weight_transfer
 - label: Entrypoints Integration (LLM)
  device: h200_35gb
  key: entrypoints-integration-llm
-  timeout_in_minutes: 60
+  timeout_in_minutes: 40
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -30,16 +28,16 @@ steps:
  - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      # TODO(akaratza): Test after Torch >= 2.12 bump
-      timeout_in_minutes: 55
+      soft_fail: true
      depends_on:
      - image-build-amd
 - label: Entrypoints Integration (API Server)
  key: entrypoints-integration-api-server
  device: h200_35gb
-  timeout_in_minutes: 75
+  timeout_in_minutes: 130
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -52,16 +50,13 @@ steps:
  - pytest -v -s entrypoints/scale_out
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
 - label: Entrypoints Integration (API Server OpenAI - Part 1)
  device: h200_35gb
  key: entrypoints-integration-api-server-openai-part-1
-  timeout_in_minutes: 68
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -72,16 +67,14 @@ steps:
  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 80
      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
 - label: Entrypoints Integration (API Server OpenAI - Part 2)
  device: h200_35gb
  key: entrypoints-integration-api-server-openai-part-2
-  timeout_in_minutes: 83
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -93,14 +86,12 @@ steps:
  - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 80
      timeout_in_minutes: 70
      depends_on:
      - image-build-amd
 - label: Entrypoints Integration (API Server Generate)
  device: h200_35gb
  key: entrypoints-integration-api-server-generate
  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
@@ -117,14 +108,12 @@ steps:
  - pytest -v -s entrypoints/anthropic
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 60
      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
 - label: Entrypoints Integration (Responses API)
  device: h200_35gb
  key: entrypoints-integration-responses-api
  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
@@ -137,7 +126,7 @@ steps:
 - label: Entrypoints Integration (Speech to Text)
  device: h200_35gb
  key: entrypoints-integration-speech_to_text
-  timeout_in_minutes: 45
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -149,7 +138,7 @@ steps:
 - label: Entrypoints Integration (Multimodal)
  device: h200_35gb
  key: entrypoints-integration-multimodal
-  timeout_in_minutes: 45
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -159,9 +148,8 @@ steps:
  - pytest -v -s entrypoints/multimodal
 - label: Entrypoints Integration (Pooling)
  device: h200_35gb
  key: entrypoints-integration-pooling
-  timeout_in_minutes: 75
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -172,7 +160,7 @@ steps:
 - label: OpenAI API Correctness
  key: openai-api-correctness
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - csrc/
@@ -181,9 +169,7 @@ steps:
  - pytest -s entrypoints/openai/correctness/
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: EPLB Algorithm
  key: eplb-algorithm
-  timeout_in_minutes: 20
+  timeout_in_minutes: 15
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -16,9 +16,7 @@ steps:
  - pytest -v -s distributed/test_eplb_utils.py
  mirror:
    amd:
      dind: false
      device: mi300_1
      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -29,7 +27,7 @@ steps:
 - label: EPLB Execution # 17min
  key: eplb-execution
-  timeout_in_minutes: 25
+  timeout_in_minutes: 27
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -41,7 +39,7 @@ steps:
 - label: Elastic EP Scaling Test
  key: elastic-ep-scaling-test
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h100
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
@@ -52,5 +50,4 @@ steps:
  - vllm/compilation/
  - tests/distributed/
  commands:
  - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
  - pytest -v -s distributed/test_elastic_ep.py
@@ -1,26 +0,0 @@
 group: Fault Tolerance
 depends_on:
  - image-build
 steps:
 - label: Fault Tolerance E2E (2xH100)
  key: fault-tolerance-e2e-2xh100
  timeout_in_minutes: 35
  device: h100
  num_devices: 2
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/v1/fault_tolerance/
  - vllm/v1/worker/sentinel/
  - vllm/entrypoints/serve/fault_tolerance/
  - vllm/distributed/elastic_ep/
  - vllm/distributed/device_communicators/
  - vllm/v1/engine/
  - vllm/v1/worker/
  - tests/v1/fault_tolerance/
  - tests/v1/distributed/test_external_lb_dp.py
  commands:
  # Base image has no nixl; install it or has_nixl_ep() skips the tests.
  - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
  # https://github.com/NVIDIA/nccl/issues/1838
  - export NCCL_CUMEM_HOST_ENABLE=0
  - pytest -v -s v1/fault_tolerance/test_fault_tolerance_e2e.py
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: vLLM IR Tests
  key: vllm-ir-tests
-  timeout_in_minutes: 35
+  timeout_in_minutes: 10
  device: h200_18gb
  working_dir: "/vllm-workspace/"
  source_file_dependencies:
@@ -15,21 +15,19 @@ steps:
    - pytest -v -s tests/kernels/ir
 - label: Kernels Core Operation Test
  device: h200_35gb
  key: kernels-core-operation-test
-  timeout_in_minutes: 120
+  timeout_in_minutes: 75
  source_file_dependencies:
  - csrc/
  - tests/kernels/core
  - tests/kernels/test_concat_mla_q.py
  - tests/kernels/test_fused_qk_norm_rope_gate.py
  commands:
-    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py
  parallelism: 3
 - label: Kernels MiniMax Reduce RMS Test (2 GPUs)
  key: kernels-minimax-reduce-rms-test-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 15
  num_devices: 2
  device: h100
  source_file_dependencies:
@@ -43,7 +41,7 @@ steps:
 - label: Deepseek V4 Kernel Test (H100)
  key: deepseek-v4-kernel-test-h100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 15
  device: h100
  source_file_dependencies:
  - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
@@ -56,54 +54,18 @@ steps:
 - label: Deepseek V4 Kernel Test (B200)
  key: deepseek-v4-kernel-test-b200
-  timeout_in_minutes: 20
+  timeout_in_minutes: 15
  device: b200-k8s
  source_file_dependencies:
  - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
  - vllm/models/deepseek_v4/common/ops/
  - vllm/models/deepseek_v4/nvidia/
  - tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
  - tests/models/test_deepseek_v4_mega_moe.py
  commands:
    - pytest -v -s kernels/test_fused_deepseek_v4_*.py
    - pytest -v -s models/test_deepseek_v4_mega_moe.py
 # Catch-all for test files at the tests/kernels root. This job collects
 # the whole root so new files are wired by default.
 # Files with dedicated jobs elsewhere in this file are excluded via --ignore
 # (test_kda, test_bf16x3_router_gemm_cutedsl and test_ll_bf16_gemm run in
 # their own jobs / Kernels (B200)).
 - label: Kernels Root Misc Test (B200)
  key: kernels-root-misc-test-b200
  timeout_in_minutes: 45
  device: b200-k8s
  source_file_dependencies:
  - csrc/
  - vllm/
  - tests/kernels/
  commands:
    - pytest -v -s kernels/
        --ignore=kernels/attention
        --ignore=kernels/core
        --ignore=kernels/helion
        --ignore=kernels/ir
        --ignore=kernels/mamba
        --ignore=kernels/moe
        --ignore=kernels/quantization
        --ignore=kernels/test_concat_mla_q.py
        --ignore=kernels/test_fused_qk_norm_rope_gate.py
        --ignore=kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
        --ignore=kernels/test_top_k_per_row.py
        --ignore=kernels/test_kda.py
        --ignore=kernels/test_bf16x3_router_gemm_cutedsl.py
        --ignore=kernels/test_ll_bf16_gemm.py
        --ignore=kernels/test_shuffle_rows.py
    # BROKEN on main, pending kernel fixes (B200):
    #   test_shuffle_rows.py (1: test_shuffle_rows_edge_cases)
 - label: Kernels Attention Test %N
  key: kernels-attention-test
-  timeout_in_minutes: 65
+  timeout_in_minutes: 35
  source_file_dependencies:
  - csrc/attention/
  - vllm/v1/attention
@@ -116,9 +78,8 @@ steps:
  parallelism: 2
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 55
      timeout_in_minutes: 90
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -145,7 +106,7 @@ steps:
 - label: Kernels Quantization Test %N
  key: kernels-quantization-test
-  timeout_in_minutes: 60
+  timeout_in_minutes: 90
  source_file_dependencies:
  - csrc/quantization/
  - vllm/model_executor/layers/quantization
@@ -155,9 +116,7 @@ steps:
  parallelism: 2
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 120
      source_file_dependencies:
      - csrc/quantization/
      - vllm/model_executor/layers/quantization
@@ -172,7 +131,7 @@ steps:
 - label: Kernels MoE Test %N
  key: kernels-moe-test
-  timeout_in_minutes: 50
+  timeout_in_minutes: 25
  source_file_dependencies:
  - csrc/quantization/cutlass_w8a8/moe/
  - csrc/moe/
@@ -187,9 +146,8 @@ steps:
  parallelism: 5
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 50
      timeout_in_minutes: 55
      source_file_dependencies:
      - csrc/quantization/cutlass_w8a8/moe/
      - csrc/moe/
@@ -204,9 +162,8 @@ steps:
      - image-build-amd
 - label: Kernels Mamba Test
  device: h200_35gb
  key: kernels-mamba-test
-  timeout_in_minutes: 40
+  timeout_in_minutes: 45
  source_file_dependencies:
  - csrc/mamba/
  - tests/kernels/mamba
@@ -214,9 +171,20 @@ steps:
  commands:
    - pytest -v -s kernels/mamba
 - label: Kernels KDA Test
  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - vllm/model_executor/layers/fla/ops/kda.py
  - vllm/model_executor/layers/fla/ops/chunk_delta_h.py
  - vllm/model_executor/layers/fla/ops/l2norm.py
  - tests/kernels/test_kda.py
  commands:
    - pytest -v -s kernels/test_kda.py
 - label: Kernels DeepGEMM Test (H100)
  key: kernels-deepgemm-test-h100
-  timeout_in_minutes: 35
+  timeout_in_minutes: 45
  device: h100
  num_devices: 1
  source_file_dependencies:
@@ -243,7 +211,7 @@ steps:
 - label: Kernels (B200)
  key: kernels-b200
-  timeout_in_minutes: 80
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  # optional: true
@@ -263,15 +231,6 @@ steps:
  - vllm/v1/attention/backends/mla/flashinfer_mla.py
  - vllm/v1/attention/selector.py
  - vllm/platforms/cuda.py
  - vllm/model_executor/kernels/linear/cute_dsl/ll_bf16.py
  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_dotprod.py
  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_splitk.py
  - vllm/cute_utils/
  - vllm/model_executor/layers/mamba/ops/gdn_chunk_cutedsl/
  - vllm/model_executor/layers/fused_moe/router/bf16x3_router_gemm_cutedsl.py
  - tests/kernels/mamba/test_gdn_prefill_cutedsl.py
  - tests/kernels/test_bf16x3_router_gemm_cutedsl.py
  - tests/kernels/test_ll_bf16_gemm.py
  - tests/kernels/test_top_k_per_row.py
  commands:
    - nvidia-smi
@@ -300,28 +259,24 @@ steps:
    - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py
    - pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py
    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
    - pytest -v -s tests/kernels/mamba/test_gdn_prefill_cutedsl.py
    - pytest -v -s tests/kernels/test_bf16x3_router_gemm_cutedsl.py
    - pytest -v -s tests/kernels/test_ll_bf16_gemm.py
    # e2e
    - pytest -v -s tests/models/quantization/test_nvfp4.py
 - label: Kernels Helion Test
  key: kernels-helion-test
-  timeout_in_minutes: 115
+  timeout_in_minutes: 30
  device: h100
  source_file_dependencies:
  - vllm/utils/import_utils.py
  - tests/kernels/helion/
  commands:
    - pip install helion==1.1.0
-    - pytest -v -s kernels/helion/ --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+    - pytest -v -s kernels/helion/
  parallelism: 2
 - label: Kernels FP8 MoE Test (1xH100)
  key: kernels-fp8-moe-test-1xh100
-  timeout_in_minutes: 40
+  timeout_in_minutes: 90
  device: h100
  num_devices: 1
  optional: true
@@ -338,7 +293,7 @@ steps:
 - label: Kernels FP8 MoE Test (2xH100)
  key: kernels-fp8-moe-test-2xh100
-  timeout_in_minutes: 45
+  timeout_in_minutes: 90
  device: h100
  num_devices: 2
  optional: true
@@ -348,7 +303,7 @@ steps:
 - label: Kernels Fp4 MoE Test (B200)
  key: kernels-fp4-moe-test-b200
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
  device: b200-k8s
  num_devices: 1
  optional: true
@@ -361,7 +316,7 @@ steps:
 - label: Kernels FusedMoE Layer Test (2 H100s)
  key: kernels-fusedmoe-layer-test-2-h100s
-  timeout_in_minutes: 30
+  timeout_in_minutes: 90
  device: h100
  num_devices: 2
  source_file_dependencies:
@@ -5,7 +5,7 @@ steps:
 - label: LM Eval Small Models
  device: h200_35gb
  key: lm-eval-small-models
-  timeout_in_minutes: 45
+  timeout_in_minutes: 75
  source_file_dependencies:
  - csrc/
  - vllm/model_executor/layers/quantization
@@ -14,9 +14,8 @@ steps:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 55
      timeout_in_minutes: 45
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -55,9 +54,9 @@ steps:
    - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
- label: LM Eval Small Models (1xB200)
+- label: LM Eval Small Models (2xB200)
-  key: lm-eval-small-models-1xb200
+  key: lm-eval-small-models-2xb200
-  timeout_in_minutes: 50
+  timeout_in_minutes: 120
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -66,10 +65,9 @@ steps:
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
- label: LM Eval Small Models Distributed (2xB200)
+- label: LM Eval Small Models (2xL4)
-  key: lm-eval-small-models-distributed-2xb200
+  key: lm-eval-small-models-tp
-  timeout_in_minutes: 120
+  timeout_in_minutes: 10
  device: b200-k8s
  num_devices: 2
  optional: true
  source_file_dependencies:
@@ -79,31 +77,9 @@ steps:
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small-tp.txt
 - label: LM Eval PCP (4xB200)
  key: lm-eval-pcp-4xb200
  timeout_in_minutes: 360
  device: b200-k8s
  num_devices: 4
  optional: true
  source_file_dependencies:
  - csrc/
  - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP2-PCP2-EP.yaml
  - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml
  - tests/evals/gsm8k/configs/models-pcp.txt
  - vllm/model_executor/layers/quantization
  - vllm/config/parallel.py
  - vllm/distributed/parallel_state.py
  - vllm/model_executor/layers/attention/mla_attention.py
  - vllm/model_executor/layers/attention/pcp.py
  - vllm/v1/worker/gpu/model_runner.py
  - vllm/v1/worker/gpu/pcp_manager.py
  autorun_on_main: true
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-pcp.txt
 - label: LM Eval Large Models EP (2xB200)
  key: lm-eval-large-models-ep-2xb200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 120
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -115,7 +91,7 @@ steps:
 - label: LM Eval Qwen3.5 Models (2xB200)
  key: lm-eval-qwen3-5-models-2xb200
-  timeout_in_minutes: 45
+  timeout_in_minutes: 120
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -126,13 +102,13 @@ steps:
  - vllm/transformers_utils/configs/qwen3_5_moe.py
  - vllm/model_executor/models/qwen3_next.py
  - vllm/model_executor/models/qwen3_next_mtp.py
-  - vllm/third_party/flash_linear_attention/ops/
+  - vllm/model_executor/layers/fla/ops/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt
 - label: LM Eval Large Models (8xH200)
  key: lm-eval-large-models-8xh200
-  timeout_in_minutes: 50
+  timeout_in_minutes: 60
  device: h200
  optional: true
  num_devices: 8
@@ -140,9 +116,8 @@ steps:
    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
  mirror:
    amd:
      dind: false
      device: mi300_8
-      timeout_in_minutes: 40
+      timeout_in_minutes: 180
      depends_on:
      - image-build-amd
      commands:
@@ -174,9 +149,9 @@ steps:
  commands:
    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor-dp-ep/config-b200.txt
- label: LM Eval Humming f16 (A100 - TEMPORARY)
+- label: LM Eval Humming (A100 - TEMPORARY)
-  key: lm-eval-humming-f16-a100
+  key: lm-eval-humming-a100
-  timeout_in_minutes: 75
+  timeout_in_minutes: 30
  device: a100
  optional: true
  num_devices: 1
@@ -184,29 +159,13 @@ steps:
  - vllm/model_executor/layers/quantization/humming.py
  - vllm/model_executor/layers/quantization/utils/humming_utils.py
  - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
-  - vllm/model_executor/layers/fused_moe/oracle/
+  - vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
  - vllm/model_executor/kernels/linear/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config.txt
- label: LM Eval Humming Act int8 (A100 - TEMPORARY)
+- label: LM Eval Humming (H100 - TEMPORARY)
-  key: lm-eval-humming-act-a100
+  key: lm-eval-humming-h100
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: a100
  optional: true
  num_devices: 1
  source_file_dependencies:
  - vllm/model_executor/layers/quantization/humming.py
  - vllm/model_executor/layers/quantization/utils/humming_utils.py
  - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
  - vllm/model_executor/layers/fused_moe/oracle/
  - vllm/model_executor/kernels/linear/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-int8.txt
 - label: LM Eval Humming f16 (H100 - TEMPORARY)
  key: lm-eval-humming-f16-h100
  timeout_in_minutes: 70
  device: h100
  optional: true
  num_devices: 1
@@ -214,30 +173,14 @@ steps:
  - vllm/model_executor/layers/quantization/humming.py
  - vllm/model_executor/layers/quantization/utils/humming_utils.py
  - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
-  - vllm/model_executor/layers/fused_moe/oracle/
+  - vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
  - vllm/model_executor/kernels/linear/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config.txt
 - label: LM Eval Humming Act fp8/int8 (H100 - TEMPORARY)
  key: lm-eval-humming-act-h100
  timeout_in_minutes: 70
  device: h100
  optional: true
  num_devices: 1
  source_file_dependencies:
  - vllm/model_executor/layers/quantization/humming.py
  - vllm/model_executor/layers/quantization/utils/humming_utils.py
  - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
  - vllm/model_executor/layers/fused_moe/oracle/
  - vllm/model_executor/kernels/linear/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-fp8.txt
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-int8.txt
- label: LM Eval Humming f16 (B200 - TEMPORARY)
+- label: LM Eval Humming (B200 - TEMPORARY)
-  key: lm-eval-humming-f16-b200
+  key: lm-eval-humming-b200
-  timeout_in_minutes: 50
+  timeout_in_minutes: 30
  device: b200-k8s
  optional: true
  num_devices: 1
@@ -245,30 +188,14 @@ steps:
  - vllm/model_executor/layers/quantization/humming.py
  - vllm/model_executor/layers/quantization/utils/humming_utils.py
  - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
-  - vllm/model_executor/layers/fused_moe/oracle/
+  - vllm/model_executor/layers/fused_moe/oracle/mxfp4.py
  - vllm/model_executor/kernels/linear/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config.txt
 - label: LM Eval Humming Act fp8/int8 (B200 - TEMPORARY)
  key: lm-eval-humming-act-b200
  timeout_in_minutes: 50
  device: b200-k8s
  optional: true
  num_devices: 1
  source_file_dependencies:
  - vllm/model_executor/layers/quantization/humming.py
  - vllm/model_executor/layers/quantization/utils/humming_utils.py
  - vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
  - vllm/model_executor/layers/fused_moe/oracle/
  - vllm/model_executor/kernels/linear/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-fp8.txt
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/humming/config-act-int8.txt
 - label: LM Eval TurboQuant KV Cache
  key: lm-eval-turboquant-kv-cache
-  timeout_in_minutes: 55
+  timeout_in_minutes: 75
  device: h200_18gb
  source_file_dependencies:
  - vllm/model_executor/layers/quantization/turboquant/
@@ -280,7 +207,7 @@ steps:
 - label: GPQA Eval (GPT-OSS) (2xH100)
  key: gpqa-eval-gpt-oss-2xh100
-  timeout_in_minutes: 35
+  timeout_in_minutes: 120
  device: h100
  optional: true
  num_devices: 2
@@ -294,7 +221,7 @@ steps:
 - label: GPQA Eval (GPT-OSS) (2xB200)
  key: gpqa-eval-gpt-oss-2xb200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 120
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -308,7 +235,7 @@ steps:
 - label: GPQA Eval (GPT-OSS) (DGX Spark)
  key: gpqa-eval-gpt-oss-spark
-  timeout_in_minutes: 35
+  timeout_in_minutes: 120
  device: dgx-spark
  optional: true
  num_devices: 1
@@ -322,50 +249,9 @@ steps:
    - uv pip install --system 'gpt-oss[eval]==0.0.5'
    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-spark.txt
 - label: LM Eval KV-Offload (1xH200)
  key: kv-offload-small
  timeout_in_minutes: 30
  device: h200_35gb
  source_file_dependencies:
  - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
  - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py
  - vllm/v1/kv_offload/
  - vllm/v1/simple_kv_offload/
  - tests/evals/gsm8k/test_gsm8k_offloading.py
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "nemotron-h-8b or gemma-4-e4b-it"
 - label: LM Eval KV-Offload (2xH100)
  key: kv-offload-medium
  timeout_in_minutes: 45
  device: h100
  num_devices: 2
  source_file_dependencies:
  - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
  - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py
  - vllm/v1/kv_offload/
  - vllm/v1/simple_kv_offload/
  - tests/evals/gsm8k/test_gsm8k_offloading.py
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b or deepseek-v2-lite"
 - label: LM Eval KV-Offload (4xH100)
  key: kv-offload-large
  timeout_in_minutes: 40
  device: h100
  num_devices: 4
  source_file_dependencies:
  - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
  - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py
  - vllm/v1/kv_offload/
  - vllm/v1/simple_kv_offload/
  - tests/evals/gsm8k/test_gsm8k_offloading.py
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "deepseek-v4-flash"
 - label: MRCR Eval Small Models
  device: h200_35gb
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
  source_file_dependencies:
  - tests/evals/mrcr/
  commands:
@@ -5,7 +5,7 @@ steps:
 - label: LoRA %N
  device: h200_35gb
  key: lora
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  source_file_dependencies:
  - vllm/lora
  - tests/lora
@@ -14,10 +14,9 @@ steps:
  parallelism: 4
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      working_dir: "/vllm-workspace/tests"
-      timeout_in_minutes: 85
+      timeout_in_minutes: 60
      source_file_dependencies:
      - vllm/lora
      - tests/lora
@@ -28,7 +27,7 @@ steps:
 - label: LoRA TP (Distributed)
  key: lora-tp-distributed
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
  num_devices: 4
  source_file_dependencies:
  - vllm/lora
@@ -47,4 +46,4 @@ steps:
    - pytest -v -s -x lora/test_qwen3_with_multi_loras.py
    - pytest -v -s -x lora/test_olmoe_tp.py
    - pytest -v -s -x lora/test_gptoss_tp.py
-    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
+    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
@@ -5,7 +5,7 @@ steps:
 - label: V1 Spec Decode
  device: h200_35gb
  key: v1-spec-decode
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  source_file_dependencies:
    - vllm/config/
    - vllm/distributed/
@@ -23,15 +23,14 @@ steps:
    - pytest -v -s -m 'not slow_test' v1/spec_decode
  mirror:
    amd:
      dind: false
      device: mi300_1
-      timeout_in_minutes: 50
+      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
 - label: V1 Sample + Logits
  key: v1-sample-logits
-  timeout_in_minutes: 83
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/config/
@@ -59,16 +58,13 @@ steps:
    - pytest -v -s v1/test_outputs.py
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 70
      depends_on:
      - image-build-amd
 - label: V1 Core + KV + Metrics
  device: h200_35gb
  key: v1-core-kv-metrics
-  timeout_in_minutes: 80
+  timeout_in_minutes: 30
  source_file_dependencies:
    - vllm/config/
    - vllm/distributed/
@@ -92,9 +88,7 @@ steps:
    - tests/v1/kv_offload
    - tests/v1/simple_kv_offload
    - tests/v1/worker
    - tests/v1/streaming_input
    - tests/v1/kv_connector/unit
    - tests/v1/ec_connector/unit
    - tests/v1/metrics
    - tests/entrypoints/openai/correctness/test_lmeval.py
  commands:
@@ -106,18 +100,15 @@ steps:
    - pytest -v -s v1/kv_offload
    - pytest -v -s v1/simple_kv_offload
    - pytest -v -s v1/worker
    - pytest -v -s v1/streaming_input
    - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
    - pytest -v -s -m 'not cpu_test' v1/ec_connector/unit
    - pytest -v -s -m 'not cpu_test' v1/metrics
    # Integration test for streaming correctness (requires special branch).
-    - pip install -U git+https://github.com/vllm-project/lm-evaluation-harness.git@streaming-api
+    - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
    - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 60
      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
@@ -148,8 +139,6 @@ steps:
    - pytest -v -s -m 'cpu_test' v1/core
    - pytest -v -s v1/structured_output
    - pytest -v -s v1/test_serial_utils.py
    - pytest -v -s v1/test_kv_cache_spec_registry.py
    - pytest -v -s v1/cudagraph/test_cudagraph_manager.py
    - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
    - pytest -v -s -m 'cpu_test' v1/metrics
@@ -183,7 +172,7 @@ steps:
 - label: Regression
  key: regression
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - vllm/config/
@@ -199,21 +188,21 @@ steps:
  - vllm/v1/
  - tests/test_regression
  commands:
-  - pip install 'modelscope<1.38'
+  - pip install modelscope
  - pytest -v -s test_regression.py
  working_dir: "/vllm-workspace/tests" # optional
 - label: Examples
  device: h200_35gb
  key: examples
-  timeout_in_minutes: 40
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/examples"
  source_file_dependencies:
  - vllm/entrypoints
  - vllm/multimodal
  - examples/
  commands:
-    - pip install --no-deps tensorizer # for tensorizer test
+    - pip install tensorizer # for tensorizer test
     # for basic
    - python3 basic/offline_inference/chat.py
    - python3 basic/offline_inference/generate.py --model facebook/opt-125m
@@ -237,9 +226,7 @@ steps:
    - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 75
      source_file_dependencies:
      - vllm/entrypoints
      - vllm/multimodal
@@ -250,7 +237,7 @@ steps:
 - label: Metrics, Tracing (2 GPUs)
  key: metrics-tracing-2-gpus
-  timeout_in_minutes: 25
+  timeout_in_minutes: 20
  num_devices: 2
  source_file_dependencies:
  - vllm/config/
@@ -266,7 +253,6 @@ steps:
  - vllm/utils/
  - vllm/v1/
  - tests/v1/tracing
  - tests/tracing/
  commands:
  - "pip install \
      'opentelemetry-sdk>=1.26.0' \
@@ -274,14 +260,12 @@ steps:
      'opentelemetry-exporter-otlp>=1.26.0' \
      'opentelemetry-semantic-conventions-ai>=0.4.1'"
  - pytest -v -s v1/tracing
  - pytest -v -s tracing
  mirror:
    amd:
-      dind: false
+      device: mi325_2
      device: mi300_2
      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
      optional: true
 - label: Python-only Installation
  key: python-only-installation
@@ -296,9 +280,8 @@ steps:
  - bash standalone_tests/python_only_compile.sh
  mirror:
    amd:
-      device: mi250_1
+      device: mi325_1
-      timeout_in_minutes: 55
+      timeout_in_minutes: 20
      soft_fail: true
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -309,7 +292,7 @@ steps:
 - label: Async Engine, Inputs, Utils, Worker
  device: h200_35gb
  key: async-engine-inputs-utils-worker
-  timeout_in_minutes: 25
+  timeout_in_minutes: 50
  source_file_dependencies:
  - vllm/assets/
  - vllm/config/
@@ -336,7 +319,7 @@ steps:
  key: async-engine-inputs-utils-worker-config-cpu
  depends_on:
  - image-build-cpu
-  timeout_in_minutes: 65
+  timeout_in_minutes: 30
  source_file_dependencies:
  - vllm/assets/
  - vllm/config/
@@ -368,7 +351,6 @@ steps:
  - tests/test_outputs.py
  - tests/test_pooling_params.py
  - tests/test_ray_env.py
  - tests/test_sampling_params.py
  - tests/multimodal
  - tests/renderers
  - tests/standalone_tests/lazy_imports.py
@@ -386,7 +368,6 @@ steps:
  - pytest -v -s test_outputs.py
  - pytest -v -s test_pooling_params.py
  - pytest -v -s test_ray_env.py
  - pytest -v -s test_sampling_params.py
  - pytest -v -s -m 'cpu_test' multimodal
  - pytest -v -s renderers
  - pytest -v -s reasoning
@@ -398,7 +379,7 @@ steps:
 - label: Batch Invariance (A100)
  key: batch-invariance-a100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
  device: a100
  source_file_dependencies:
    - vllm/v1/attention
@@ -408,11 +389,11 @@ steps:
    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
    - pip install pytest-timeout pytest-forked
    - pytest -v -s v1/determinism/test_batch_invariance.py
-    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k TRITON_MLA
+    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[TRITON_MLA]
 - label: Batch Invariance (H100)
  key: batch-invariance-h100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
  device: h100
  source_file_dependencies:
    - vllm/v1/attention
@@ -423,12 +404,12 @@ steps:
    - pip install pytest-timeout pytest-forked
    - pytest -v -s v1/determinism/test_batch_invariance.py
    - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
-    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k TRITON_MLA
+    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[TRITON_MLA]
-    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k FLASH_ATTN
+    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[FLASH_ATTN]
 - label: Batch Invariance (B200)
  key: batch-invariance-b200
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: b200-k8s
  source_file_dependencies:
    - vllm/v1/attention
@@ -439,18 +420,15 @@ steps:
    - pip install pytest-timeout pytest-forked
    - pytest -v -s v1/determinism/test_batch_invariance.py
    - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
-    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k TRITON_MLA
+    - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[TRITON_MLA]
-    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k FLASH_ATTN
+    - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle[FLASH_ATTN]
    - pytest -v -s v1/determinism/test_nvfp4_batch_invariant.py
    - pytest -v -s v1/determinism/test_nvfp4_batch_invariant_scaled_mm.py
-    - pytest -v -s v1/determinism/test_matmul_batch_invariant.py
+  
    - pytest -v -s v1/determinism/test_cutlass_batch_invariance.py
    - pytest -v -s v1/determinism/test_online_batch_invariance.py
 - label: Acceptance Length Test (Large Models) # optional
  device: h200_35gb
  key: acceptance-length-test-large-models
-  timeout_in_minutes: 20
+  timeout_in_minutes: 25
  gpu: h100
  optional: true
  num_gpus: 1
@@ -3,16 +3,13 @@ depends_on:
  - image-build
 steps:
 - label: Model Executor
  device: h200_35gb
  key: model-executor
-  timeout_in_minutes: 60
+  timeout_in_minutes: 35
  source_file_dependencies:
  - vllm/engine/arg_utils.py
  - vllm/config/model.py
  - vllm/model_executor
  - vllm/model_executor/warmup
  - tests/model_executor
  - tests/model_executor/test_jit_warmup.py
  - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
  commands:
    - apt-get update && apt-get install -y curl libsodium23
@@ -28,18 +25,14 @@ steps:
    - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py --timeout=900 --timeout-method=thread
  mirror:
    amd:
      dind: false
      device: mi300_1
      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      source_file_dependencies:
      - vllm/engine/arg_utils.py
      - vllm/config/model.py
      - vllm/model_executor
      - vllm/model_executor/warmup
      - tests/model_executor
      - tests/model_executor/test_jit_warmup.py
      - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
      - vllm/_aiter_ops.py
      - vllm/platforms/rocm.py
@@ -5,7 +5,7 @@ steps:
 - label: Model Runner V2 Core Tests
  device: h200_35gb
  key: model-runner-v2-core-tests
-  timeout_in_minutes: 35
+  timeout_in_minutes: 45
  source_file_dependencies:
  - vllm/v1/worker/gpu/
  - vllm/v1/worker/gpu_worker.py
@@ -18,7 +18,9 @@ steps:
  - set -x
  - export VLLM_USE_V2_MODEL_RUNNER=1
  - pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics"
-  - pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram"
+  # This requires eager until we sort out CG correctness issues.
  # TODO: remove ENFORCE_EAGER here after https://github.com/vllm-project/vllm/pull/32936 is merged.
  - ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram"
  - pytest -v -s v1/e2e/general/test_context_length.py
  - pytest -v -s v1/e2e/general/test_min_tokens.py
  # Temporary hack filter to exclude ngram spec decoding based tests.
@@ -27,7 +29,7 @@ steps:
 - label: Model Runner V2 Examples
  device: h200_35gb
  key: model-runner-v2-examples
-  timeout_in_minutes: 35
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/examples"
  source_file_dependencies:
    - vllm/v1/worker/gpu/
@@ -41,7 +43,7 @@ steps:
  commands:
    - set -x
    - export VLLM_USE_V2_MODEL_RUNNER=1
-    - pip install --no-deps tensorizer # for tensorizer test
+    - pip install tensorizer # for tensorizer test
    - python3 basic/offline_inference/chat.py # for basic
    - python3 basic/offline_inference/generate.py --model facebook/opt-125m
    #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10  # TODO
@@ -63,7 +65,7 @@ steps:
 - label: Model Runner V2 Distributed (2 GPUs)
  key: model-runner-v2-distributed-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -84,7 +86,7 @@ steps:
 - label: Model Runner V2 Pipeline Parallelism (4 GPUs)
  key: model-runner-v2-pipeline-parallelism-4-gpus
-  timeout_in_minutes: 50
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -4,8 +4,9 @@ depends_on:
 steps:
 - label: Basic Models Tests (Initialization)
  key: basic-models-tests-initialization
-  timeout_in_minutes: 25
+  timeout_in_minutes: 45
  device: h200_18gb
  torch_nightly: true
  source_file_dependencies:
  - vllm/
  - tests/models/test_initialization.py
@@ -13,11 +14,13 @@ steps:
  commands:
    # Run a subset of model initialization tests
    - pytest -v -s models/test_initialization.py::test_can_initialize_small_subset
  mirror:
    torch_nightly: {}
 - label: Basic Models Tests (Extra Initialization) %N
  device: h200_35gb
  key: basic-models-tests-extra-initialization
-  timeout_in_minutes: 100
+  timeout_in_minutes: 45
  source_file_dependencies:
  - vllm/model_executor/models/
  - tests/models/test_initialization.py
@@ -27,65 +30,36 @@ steps:
    # subset of supported models (the complement of the small subset in the above
    # test.) Also run if model initialization test file is modified
    - pytest -v -s models/test_initialization.py -k 'not test_can_initialize_small_subset' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
-  parallelism: 4
+  parallelism: 2
  mirror:
    torch_nightly: {}
 - label: Basic Models Tests (Other)
  device: h200_35gb
  key: basic-models-tests-other
-  timeout_in_minutes: 35
+  timeout_in_minutes: 45
  source_file_dependencies:
  - vllm/
  - tests/models/test_terratorch.py
-  - tests/models/transformers/test_backend.py
+  - tests/models/test_transformers.py
  - tests/models/test_registry.py
  commands:
-    - pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py
+    - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.py
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 50
      depends_on:
      - image-build-amd
 - label: Inkling Unit Tests (B200)
  key: inkling-unit-tests-b200
  timeout_in_minutes: 40
  device: b200-k8s
  source_file_dependencies:
  - vllm/models/inkling/
  - vllm/cute_utils/
  - cmake/external_projects/tml_fa4.cmake
  - tests/models/inkling/
  commands:
    # FA4 kernel tests require SM100; the suite skips them elsewhere.
    - pytest -v -s models/inkling
 - label: Kimi K3 Unit Tests (B200)
  key: kimi-k3-unit-tests-b200
  timeout_in_minutes: 40
  device: b200-k8s
  source_file_dependencies:
  - vllm/models/kimi_k3/
  - csrc/libtorch_stable/kimi_k3/
  - tests/models/kimi_k3/
  - tests/kernels/attention/test_kimi_k3_mla_fused_epilogue.py
  - tests/kernels/test_bf16_skinny_gemm.py
  commands:
    # The native NVIDIA Kimi K3 kernels require the SM100 family.
    - pytest -v -s models/kimi_k3 kernels/attention/test_kimi_k3_mla_fused_epilogue.py kernels/test_bf16_skinny_gemm.py
 - label: Basic Models Test (Other CPU) # 5min
  key: basic-models-test-other-cpu
  depends_on:
  - image-build-cpu
-  timeout_in_minutes: 20
+  timeout_in_minutes: 10
  source_file_dependencies:
  - vllm/
  - tests/models/test_utils.py
  - tests/models/test_vision.py
  - tests/models/test_adapters.py
  - tests/models/transformers/fusers/
  device: cpu-small
  commands:
-    - pytest -v -s models/test_utils.py models/test_vision.py models/test_adapters.py models/transformers/fusers/
+    - pytest -v -s models/test_utils.py models/test_vision.py
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Model Tests (2 GPUs)
  key: distributed-model-tests-2-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -17,7 +17,7 @@ steps:
  - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m '(not slow_test)'
  # Avoid importing model tests that cause CUDA reinitialization error
-  - pytest models/transformers/test_backend.py -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
  - pytest models/language -v -s -m 'distributed(num_gpus=2)'
  - pytest models/multimodal/generation/test_phi4siglip.py -v -s -m 'distributed(num_gpus=2)'
  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_phi4siglip.py
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Language Models Tests (Standard)
  key: language-models-tests-standard
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -14,17 +14,15 @@ steps:
    - pip freeze | grep -E 'torch'
    - pytest -v -s models/language -m 'core_model and (not slow_test)'
  mirror:
    torch_nightly: {}
    amd:
      dind: false
      device: mi300_1
      timeout_in_minutes: 45
      depends_on:
      - image-build-amd
 - label: Language Models Tests (Extra Standard) %N
  device: h200_35gb
  key: language-models-tests-extra-standard
-  timeout_in_minutes: 40
+  timeout_in_minutes: 45
  source_file_dependencies:
  - vllm/model_executor/models/
  - tests/models/language/pooling/test_embedding.py
@@ -37,10 +35,9 @@ steps:
    - pytest -v -s models/language -m 'core_model and slow_test' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
  parallelism: 2
  mirror:
    torch_nightly: {}
    amd:
      dind: false
      device: mi300_1
      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -54,25 +51,26 @@ steps:
      - tests/models/language/pooling/test_classification.py
      - vllm/_aiter_ops.py
      - vllm/platforms/rocm.py
 - label: Language Models Tests (Hybrid) %N
  device: h200_35gb
  key: language-models-tests-hybrid
-  timeout_in_minutes: 65
+  timeout_in_minutes: 75
  source_file_dependencies:
  - vllm/
  - tests/models/language/generation
  commands:
    # Install fast path packages for testing against transformers
    # Note: also needed to run plamo2 model in vLLM
    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
-    # Shard the hybrid language model tests that are numerically stable on Hopper.
+    # Shard hybrid language model tests
-    - pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+    - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
  parallelism: 2
  mirror:
    torch_nightly: {}
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 90
      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      commands:
@@ -80,37 +78,24 @@ steps:
      - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
      - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
 # Granite 4 hybrid generation is sensitive to hardware-specific Triton SSD
 # autotuning (https://github.com/vllm-project/vllm/issues/25194). Keep this one
 # correctness test on L4 until its H200 output matches the Transformers reference.
 - label: Language Models Tests (Granite L4 Compatibility)
  key: language-models-tests-granite-l4-compatibility
  timeout_in_minutes: 65
  source_file_dependencies:
  - vllm/
  - tests/models/language/generation
  commands:
    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
    - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
 - label: Language Models Test (Extended Generation) # 80min
  device: h200_35gb
  key: language-models-test-extended-generation
-  timeout_in_minutes: 65
+  timeout_in_minutes: 110
  optional: true
  source_file_dependencies:
  - vllm/
  - tests/models/language/generation
  commands:
    # Install fast path packages for testing against transformers
    # Note: also needed to run plamo2 model in vLLM
    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
    - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 - label: Language Models Test (PPL)
  key: language-models-test-ppl
-  timeout_in_minutes: 30
+  timeout_in_minutes: 110
  device: h200_18gb
  optional: true
  source_file_dependencies:
@@ -119,10 +104,10 @@ steps:
  commands:
    - pytest -v -s models/language/generation_ppl_test
- label: Language Models Test (Extended Pooling)
+- label: Language Models Test (Extended Pooling)  # 36min
  device: h200_35gb
  key: language-models-test-extended-pooling
-  timeout_in_minutes: 120
+  timeout_in_minutes: 50
  optional: true
  source_file_dependencies:
  - vllm/
@@ -131,15 +116,14 @@ steps:
    - pytest -v -s models/language/pooling -m 'not core_model'
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 100
      timeout_in_minutes: 95
      depends_on:
      - image-build-amd
 - label: Language Models Test (MTEB)
  key: language-models-test-mteb
-  timeout_in_minutes: 68
+  timeout_in_minutes: 110
  device: h200_18gb
  optional: true
  source_file_dependencies:
@@ -4,107 +4,104 @@ depends_on:
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
  key: multi-modal-models-standard-1-qwen2
-  timeout_in_minutes: 68
+  timeout_in_minutes: 45
  device: h200_18gb
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2"
    - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
  key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 75
+  timeout_in_minutes: 45
  device: h200_18gb
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
    - pytest -v -s models/multimodal/generation/test_mm_prefix_lm.py -m core_model
    - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
  device: h200_35gb
  key: multi-modal-models-standard-3-llava-qwen2-vl
-  timeout_in_minutes: 40
+  timeout_in_minutes: 45
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
    - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
  mirror:
    amd:
-      device: mi250_1
+      device: mi325_1
      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
  device: h200_35gb
  key: multi-modal-models-standard-4-other-whisper
-  timeout_in_minutes: 75
+  timeout_in_minutes: 45
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  commands:
-    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_mm_prefix_lm.py --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/generation/test_vit_cudagraph.py --ignore models/multimodal/processing
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --ignore models/multimodal/generation/test_whisper.py  --ignore models/multimodal/generation/test_memory_leak.py --ignore models/multimodal/generation/test_vit_cudagraph.py --ignore models/multimodal/processing
    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
    - pytest models/multimodal/generation/test_memory_leak.py -m core_model
    - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 50
      depends_on:
      - image-build-amd
- label: Multi-Modal Processor (CPU) %N
+- label: Multi-Modal Processor (CPU)
  key: multi-modal-processor-cpu
  depends_on:
  - image-build-cpu
-  timeout_in_minutes: 125
+  timeout_in_minutes: 60
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  - tests/models/registry.py
  device: cpu-medium
  commands:
-    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-  parallelism: 4
+    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
 - label: Multi-Modal Processor # 44min
  key: multi-modal-processor
-  timeout_in_minutes: 98
+  timeout_in_minutes: 60
  device: h200_18gb
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  - tests/models/registry.py
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/processing/test_tensor_schema.py
 - label: Multi-Modal Accuracy Eval (Small Models) # 50min
  device: h200_35gb
  key: multi-modal-accuracy-eval-small-models
-  timeout_in_minutes: 30
+  timeout_in_minutes: 70
  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
  source_file_dependencies:
  - vllm/multimodal/
@@ -114,9 +111,7 @@ steps:
  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
  mirror:
    amd:
      dind: false
      device: mi300_1
      timeout_in_minutes: 35
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -127,7 +122,6 @@ steps:
      - vllm/model_executor/model_loader/
 - label: Multi-Modal Models (Extended Generation 1)
  device: h200_35gb
  key: multi-modal-models-extended-generation-1
  optional: true
  source_file_dependencies:
@@ -135,13 +129,12 @@ steps:
  - tests/models/multimodal/generation
  - tests/models/multimodal/test_mapping.py
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
    - pytest -v -s models/multimodal/test_mapping.py
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 90
      depends_on:
      - image-build-amd
@@ -153,6 +146,7 @@ steps:
  - vllm/
  - tests/models/multimodal/generation
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 - label: Multi-Modal Models (Extended Generation 3)
@@ -163,6 +157,7 @@ steps:
  - vllm/
  - tests/models/multimodal/generation
  commands:
    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
 - label: Multi-Modal Models (Extended Pooling)
@@ -176,8 +171,7 @@ steps:
    - pytest -v -s models/multimodal/pooling -m 'not core_model'
  mirror:
    amd:
-      dind: false
+      device: mi325_1
      device: mi300_1
      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Plugin Tests (2 GPUs)
  key: plugin-tests-2-gpus
-  timeout_in_minutes: 35
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -37,11 +37,6 @@ steps:
  - pytest -v -s plugins_tests/test_stats_logger_plugins.py
  - pip uninstall dummy_stat_logger -y
  # end stat_logger plugins test
  # begin endpoint plugins test
  - pip install -e ./plugins/vllm_add_dummy_endpoint_plugin
  - pytest -v -s plugins_tests/test_endpoint_plugins.py
  - pip uninstall vllm_add_dummy_endpoint_plugin -y
  # end endpoint plugins test
  # other tests continue here:
  - pytest -v -s plugins_tests/test_scheduler_plugins.py
  - pip install -e ./plugins/vllm_add_dummy_model
@@ -5,7 +5,7 @@ steps:
 - label: PyTorch Compilation Unit Tests
  device: h200_35gb
  key: pytorch-compilation-unit-tests
-  timeout_in_minutes: 150
+  timeout_in_minutes: 10
  source_file_dependencies:
    - vllm/__init__.py
    - vllm/_aiter_ops.py
@@ -78,7 +78,7 @@ steps:
 - label: PyTorch Compilation Passes Unit Tests
  key: pytorch-compilation-passes-unit-tests
-  timeout_in_minutes: 45
+  timeout_in_minutes: 20
  source_file_dependencies:
    - vllm/__init__.py
    - vllm/_aiter_ops.py
@@ -107,11 +107,16 @@ steps:
    - tests/compile/passes
  commands:
  - pytest -s -v compile/passes --ignore compile/passes/distributed
  mirror:
    amd:
      device: mi300_1
      timeout_in_minutes: 180
      depends_on:
      - image-build-amd
 - label: PyTorch Fullgraph Smoke Test
  device: h200_35gb
  key: pytorch-fullgraph-smoke-test
-  timeout_in_minutes: 90
+  timeout_in_minutes: 35
  source_file_dependencies:
  - vllm/__init__.py
  - vllm/_aiter_ops.py
@@ -143,46 +148,11 @@ steps:
  # as it is a heavy test that is covered in other steps.
  # Use `find` to launch multiple instances of pytest so that
  # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_cudagraph.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
 # Hopper-only DeepSeek-V2-Lite cases in this file require two 29.3-GiB model
 # instances and cannot fit a 35GB MIG slice. L4 retains the original coverage:
 # those SM90 cases skip while the architecture-compatible cases still run.
 - label: PyTorch Fullgraph CUDAGraph (L4 Compatibility)
  key: pytorch-fullgraph-cudagraph-l4-compatibility
  timeout_in_minutes: 60
  source_file_dependencies:
  - vllm/__init__.py
  - vllm/_aiter_ops.py
  - vllm/_custom_ops.py
  - vllm/compilation/
  - vllm/config/
  - vllm/distributed/
  - vllm/engine/
  - vllm/env_override.py
  - vllm/envs.py
  - vllm/forward_context.py
  - vllm/inputs/
  - vllm/ir/
  - vllm/kernels/
  - vllm/logger.py
  - vllm/model_executor/
  - vllm/multimodal/
  - vllm/platforms/
  - vllm/plugins/
  - vllm/sampling_params.py
  - vllm/sequence.py
  - vllm/transformers_utils/
  - vllm/triton_utils/
  - vllm/utils/
  - vllm/v1/
  - tests/compile
  commands:
  - pytest -s -v compile/fullgraph/test_full_cudagraph.py
 - label: PyTorch Fullgraph
  key: pytorch-fullgraph
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/__init__.py
@@ -227,9 +197,7 @@ steps:
  - bash standalone_tests/pytorch_nightly_dependency.sh
  mirror:
    amd:
      dind: false
      device: mi300_1
      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -3,11 +3,8 @@ depends_on:
  - image-build
 steps:
 - label: Quantization
  device: h200_35gb
  key: quantization
-  timeout_in_minutes: 75
+  timeout_in_minutes: 90
  env:
    VLLM_USE_V2_MODEL_RUNNER: "0"
  source_file_dependencies:
  - csrc/
  - vllm/model_executor/layers/quantization
@@ -22,14 +19,11 @@ steps:
  # TODO(jerryzh168): resolve the above comment
  - uv pip install --system torchao==0.17.0 --index-url https://download.pytorch.org/whl/cu130
  - uv pip install --system conch-triton-kernels
-  # The SM90-only checkpoint currently contains a removed weight_chan_scale
+  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
  # parameter. It was not exercised by the previous L4 job.
  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py -k 'not test_compressed_tensors_w4a8_fp8'
 - label: Quantized Fusions
  device: h200_35gb
  key: quantized-fusions
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  source_file_dependencies:
  - tests/fusion
  - vllm/model_executor/layers/fusion
@@ -41,7 +35,7 @@ steps:
 - label: Quantized MoE Test (B200)
  key: quantized-moe-test-b200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  source_file_dependencies:
@@ -58,11 +52,8 @@ steps:
    - pytest -s -v tests/quantization/test_blackwell_moe.py
 - label: Quantized Models Test
  device: h200_35gb
  key: quantized-models-test
-  timeout_in_minutes: 65
+  timeout_in_minutes: 60
  env:
    VLLM_USE_V2_MODEL_RUNNER: "0"
  source_file_dependencies:
  - vllm/model_executor/layers/quantization
  - tests/models/quantization
@@ -3,7 +3,7 @@ depends_on:
  - image-build
 steps:
 - label: Rust Frontend OpenAI Coverage
-  timeout_in_minutes: 30
+  timeout_in_minutes: 90
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -15,30 +15,28 @@ steps:
  - tests/utils.py
  - tests/benchmarks/test_serve_cli.py
  - tests/entrypoints/openai/chat_completion/test_chat_completion.py
-  - tests/entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py
+  # - tests/entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py
  # - tests/entrypoints/openai/completion/test_prompt_validation.py
  - tests/entrypoints/openai/completion/test_shutdown.py
-  - tests/entrypoints/openai/test_return_token_ids.py
+  # - tests/entrypoints/openai/test_return_token_ids.py
-  - tests/entrypoints/openai/test_uds.py
+  # - tests/entrypoints/openai/test_uds.py
  - tests/v1/sample/test_logprobs_e2e.py
  commands:
  - export VLLM_USE_RUST_FRONTEND=1
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
  - pytest -v -s benchmarks/test_serve_cli.py -k "not insecure and not (test_bench_serve and not test_bench_serve_chat)"
-  - pytest -v -s entrypoints/openai/chat_completion/test_chat_completion.py -k "not test_invalid_json_schema and not test_invalid_regex and not test_kv_transfer_prompt_token_ids_round_trip and not test_kv_transfer_prompt_token_ids_streaming"
+  - pytest -v -s entrypoints/openai/chat_completion/test_chat_completion.py -k "not test_invalid_json_schema and not test_invalid_regex"
-  - pytest -v -s entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py -k "not multiple"
+  # - pytest -v -s entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py -k "not invalid"
  # - pytest -v -s entrypoints/openai/completion/test_prompt_validation.py -k "not prompt_embeds"
  - pytest -v -s entrypoints/openai/completion/test_shutdown.py -k "not engine_failure and not test_abort_timeout_exits_quickly"
-  # test_comparison streams differently: Rust emits a separate first (prompt_token_ids) chunk and
+  # - pytest -v -s entrypoints/openai/test_return_token_ids.py
-  # finish chunk without logprobs, while the test reads `logprobs.tokens` on every chunk.
+  # - pytest -v -s entrypoints/openai/test_uds.py
  - pytest -v -s entrypoints/openai/test_return_token_ids.py -k "not test_comparison"
  - pytest -v -s entrypoints/openai/test_uds.py
  - pytest -v -s v1/sample/test_logprobs_e2e.py -k "test_prompt_logprobs_e2e_server"
 - label: Rust Frontend Serve/Admin Coverage
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -47,27 +45,22 @@ steps:
  - vllm/entrypoints/serve/
  - vllm/v1/engine/
  - tests/utils.py
-  - tests/entrypoints/serve/dev/rpc/test_collective_rpc.py
+  # - tests/entrypoints/serve/dev/rpc/test_collective_rpc.py
  - tests/entrypoints/scale_out/token_in_token_out/test_serving_tokens.py
  - tests/entrypoints/serve/instrumentator/test_basic.py
  - tests/entrypoints/serve/instrumentator/test_metrics.py
  # - tests/entrypoints/serve/dev/test_sleep.py
  - tests/entrypoints/serve/tokenize/test_tokenization.py
  commands:
  - export VLLM_USE_RUST_FRONTEND=1
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py
+  # - pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py
  # server_load can be flaky under the Rust frontend; keep it excluded for now.
  - pytest -v -s entrypoints/serve/instrumentator/test_basic.py -k "not show_version and not server_load"
  # test_generate_logprobs expects Python-style top_logprobs truncation (dedup sampled + cap at max(k, 1)).
  - pytest -v -s entrypoints/scale_out/token_in_token_out/test_serving_tokens.py -k "not stream and not lora and not test_generate_logprobs and not stop_string_workflow"
  - pytest -v -s entrypoints/serve/instrumentator/test_metrics.py -k "text and not show and not run_batch and not test_metrics_counts and not test_metrics_exist"
  # - pytest -v -s entrypoints/serve/dev/test_sleep.py
  # /tokenizer_info is not implemented in the Rust frontend (the CLI flag is accepted as a no-op).
  - pytest -v -s entrypoints/serve/tokenize/test_tokenization.py -k "not tokenizer_info"
 - label: Rust Frontend Core Correctness
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -81,8 +74,7 @@ steps:
  - pytest -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 - label: Rust Frontend Tool Use
-  device: h200_35gb
+  timeout_in_minutes: 60
  timeout_in_minutes: 25
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - rust/
@@ -96,7 +88,7 @@ steps:
  - pytest -v -s tool_use --ignore=tool_use/mistral --models llama3.2 -k "not test_response_format_with_tool_choice_required and not test_parallel_tool_calls_false and not test_tool_call_and_choice"
 - label: Rust Frontend Distributed
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
  num_devices: 4
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -4,7 +4,7 @@ steps:
 - label: Rust Frontend Cargo Style + Clippy
  key: rust-frontend-cargo-style-clippy
  depends_on: []
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: cpu-medium
  no_plugin: true
  source_file_dependencies:
@@ -18,7 +18,7 @@ steps:
 - label: Rust Frontend Cargo Tests
  key: rust-frontend-cargo-tests
  depends_on: []
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: cpu-medium
  no_plugin: true
  source_file_dependencies:
@@ -5,7 +5,7 @@ steps:
 - label: Samplers Test
  device: h200_35gb
  key: samplers-test
-  timeout_in_minutes: 40
+  timeout_in_minutes: 75
  source_file_dependencies:
  - vllm/model_executor/layers
  - vllm/sampling_metadata.py
@@ -20,17 +20,7 @@ steps:
  mirror:
    amd:
      device: mi250_1
      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
      source_file_dependencies:
      - vllm/model_executor/layers
      - vllm/sampling_metadata.py
      - vllm/v1/sample/
      - vllm/entrypoints/generate/beam_search/
      - tests/samplers
      - tests/conftest.py
      - vllm/_aiter_ops.py
      - vllm/platforms/rocm.py
      commands:
      - pytest -v -s samplers
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Spec Decode Eagle
  key: spec-decode-eagle
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
@@ -14,9 +14,8 @@ steps:
    - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 45
      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -30,7 +29,7 @@ steps:
 - label: Spec Decode Eagle Nightly B200
  key: spec-decode-eagle-nightly-b200
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -42,7 +41,7 @@ steps:
 - label: Spec Decode Speculators + MTP
  key: spec-decode-speculators-mtp
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
@@ -54,9 +53,8 @@ steps:
    - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 65
      timeout_in_minutes: 75
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -84,21 +82,20 @@ steps:
 - label: Spec Decode Ngram + Suffix
  key: spec-decode-ngram-suffix
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
    - vllm/v1/worker/gpu/spec_decode/
    - tests/v1/e2e/spec_decode/
    - tests/spec_decode/
  commands:
    - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
    - python3 spec_decode/test_custom_proposer.py
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 65
-      timeout_in_minutes: 35
+      # TODO(akaratza): Test after Torch >= 2.12 bump
      soft_fail: true
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -112,7 +109,7 @@ steps:
 - label: Spec Decode Draft Model
  key: spec-decode-draft-model
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
@@ -122,9 +119,8 @@ steps:
    - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
  mirror:
    amd:
-      dind: false
+      device: mi325_1
-      device: mi300_1
+      timeout_in_minutes: 50
      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -138,7 +134,7 @@ steps:
 - label: Spec Decode Draft Model Nightly B200
  key: spec-decode-draft-model-nightly-b200
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -150,7 +146,7 @@ steps:
 - label: Speculators Correctness
  key: speculators-correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 1
@@ -163,7 +159,7 @@ steps:
    - pytest -v -s v1/spec_decode/test_speculators_correctness.py -m slow_test
 - label: Spec Decode MTP hybrid (B200)
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -174,31 +170,3 @@ steps:
    - tests/v1/e2e/spec_decode/
  commands:
    - pytest -v -s v1/e2e/spec_decode -k "qwen3_5-hybrid"
 - label: Spec Decode DeepSeek MTP Parallel Load (B200)
  key: spec-decode-deepseek-mtp-parallel-load-b200
  timeout_in_minutes: 30
  device: b200-k8s
  optional: true
  num_devices: 2
  source_file_dependencies:
    - vllm/v1/spec_decode/llm_base_proposer.py
    - vllm/v1/spec_decode/eagle.py
    - vllm/v1/worker/gpu/spec_decode/eagle/
    - vllm/model_executor/models/deepseek_mtp.py
    - vllm/model_executor/models/deepseek_v2.py
    - tests/v1/e2e/spec_decode/test_mtp_parallel_load.py
  commands:
    - pytest -v -s v1/e2e/spec_decode/test_mtp_parallel_load.py
 - label: Spec Decode Acceptance Rates Nightly
  key: spec-decode-acceptance-rates-nightly
  timeout_in_minutes: 60
  device: h200_35gb
  optional: true
  source_file_dependencies:
    - vllm/v1/spec_decode/
    - vllm/v1/worker/gpu/spec_decode/
    - tests/v1/e2e/spec_decode/
  commands:
    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "acceptance_rates"
@@ -1,14 +0,0 @@
 group: Torch ABI
 depends_on:
  - image-build
 steps:
 - label: Torch Stable ABI Audit
  key: torch-stable-abi-audit
  timeout_in_minutes: 5
  source_file_dependencies:
  - .buildkite/check-torch-abi.py
  - csrc/
  - cmake/
  - setup.py
  commands:
  - python3 /vllm-workspace/.buildkite/check-torch-abi.py
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Weight Loading Multiple GPU  # 33min
  key: weight-loading-multiple-gpu
-  timeout_in_minutes: 50
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  optional: true
@@ -15,9 +15,7 @@ steps:
    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
  mirror:
    amd:
      dind: false
      device: mi300_2
      timeout_in_minutes: 35
      depends_on:
      - image-build-amd
      commands:
@@ -47,7 +47,6 @@
 # Rust Frontend
 /rust/ @BugenZhao @njhill
 /rust/src/bench @esmeetu
 /build_rust.sh @BugenZhao @njhill
 /rust-toolchain.toml @BugenZhao @njhill
 /.buildkite/test_areas/rust* @BugenZhao @njhill
@@ -120,7 +119,7 @@
 # Transformers modeling backend
 /vllm/model_executor/models/transformers @hmellor
-/tests/models/transformers @hmellor
+/tests/models/test_transformers.py @hmellor
 # Docs
 /docs/mkdocs @hmellor
@@ -173,7 +172,7 @@ mkdocs.yaml @hmellor
 # Kernels
 /vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
 /vllm/v1/attention/ops/triton_unified_attention.py @tdoublep
-/vllm/third_party/flash_linear_attention @ZJY0516 @vadiklyutiy
+/vllm/model_executor/layers/fla @ZJY0516 @vadiklyutiy
 # ROCm related: specify owner with write access to notify AMD folks for careful code review
 /vllm/**/*rocm* @tjtanaa @dllehr-amd
@@ -19,7 +19,6 @@ pull_request_rules:
  description: Comment on PR when pre-commit check fails
  conditions:
    - check-failure=pre-commit
    - -check-cancelled=pre-commit
    - -closed
    - -draft
    - or:
@@ -182,18 +181,6 @@ pull_request_rules:
      add:
        - performance
 - name: label-quantization
  description: Automatically apply quantization label
  conditions:
    - label != stale
    - or:
      - files~=^vllm/model_executor/layers/quantization/
      - title~=(?i)quant
  actions:
    label:
      add:
        - quantization
 - name: label-qwen
  description: Automatically apply qwen label
  conditions:
@@ -233,31 +220,6 @@ pull_request_rules:
      add:
        - gpt-oss
 - name: label-kimi
  description: Automatically apply kimi label
  conditions:
    - label != stale
    - or:
      - files~=(?i)kimi
      - files~=(?i)moonshot
      - title~=(?i)(?:kimi|moonshot)
  actions:
    label:
      add:
        - kimi
 - name: label-k3
  description: Automatically apply k3 label (launch triage; retire after ramp-down)
  conditions:
    - label != stale
    - or:
      - files~=(?i)kimi[-_]?k3
      - title~=(?i)(?:kimi[-\s]?k3|\bk3\b)
  actions:
    label:
      add:
        - k3
 - name: label-nvidia
  description: Automatically apply nvidia label
  conditions:
@@ -130,66 +130,6 @@ jobs:
                  },
                ],
              },
              kimi: {
                keywords: [
                  { term: "Kimi",     searchIn: "both" },
                  { term: "Moonshot", searchIn: "both" },
                ],
                substrings: [
                  { term: "moonshotai/", searchIn: "both" },
                  { term: "kimi",        searchIn: "title" },
                ],
              },
              k3: {
                keywords: [
                  { term: "Kimi K3", searchIn: "both" },
                  { term: "K3",      searchIn: "title" },
                ],
                substrings: [
                  { term: "moonshotai/kimi-k3", searchIn: "both" },
                ],
              },
              quantization: {
                keywords: [
                  {
                    term: "quantization",
                    searchIn: "both"
                  },
                  {
                    term: "quantized",
                    searchIn: "both"
                  },
                ],
              },
              "intel-gpu": {
                // Keyword search - matches whole words only (with word boundaries)
                keywords: [
                  {
                    term: "B50",
                    searchIn: "both"
                  },
                  {
                    term: "B60",
                    searchIn: "both"
                  },
                  {
                    term: "B70",
                    searchIn: "both"
                  },
                  {
                    term: "intel gpu",
                    searchIn: "both"
                  },
                  {
                    term: "Arc GPU",
                    searchIn: "both"
                  },
                  {
                    term: "BMG",
                    searchIn: "both"
                  },
                ],
              },
              // Add more label configurations here as needed
              // example: {
              //   keywords: [...],
@@ -383,7 +323,7 @@ jobs:
            // {users} will be replaced with @mentions
            const ccConfig = {
              rocm: {
-                users: ['hongxiayang', 'tjtanaa', 'vllmellm', 'giuseppegrossi'],
+                users: ['hongxiayang', 'tjtanaa', 'vllmellm'],
                message: 'CC {users} for ROCm-related issue',
              },
              mistral: {
@@ -551,4 +491,4 @@ jobs:
              issue_number: context.issue.number,
              body: message,
            });
-            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);
+            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);
@@ -80,9 +80,9 @@ jobs:
                  '',
                  '\u{1f4ac} Join our developer Slack at https://slack.vllm.ai to discuss your PR in `#pr-reviews`, coordinate on features in `#feat-` channels, or join special interest groups in `#sig-` channels.',
                  '',
-                  'PRs do not trigger a full CI run by default. Reviewers with write access and configured trusted contributors can comment `/ci run` whenever CI signals are needed.',
+                  'PRs do not trigger a full CI run by default. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.',
                  '',
-                  'Once the PR is approved or has the `ready` label, the PR author can also use `/ci run` or `/ci retry`. New commits do not start CI automatically.',
+                  'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.',
                  '',
                  'If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.',
                  '',
@@ -28,8 +28,7 @@ jobs:
            pull_number: context.payload.pull_request.number,
          });
-          const readyLabels = ['ready', 'ready-run-all-tests'];
+          const hasReadyLabel = pr.labels.some(l => l.name === 'ready');
          const hasReadyLabel = pr.labels.some(l => readyLabels.includes(l.name));
          const hasVerifiedLabel = pr.labels.some(l => l.name === 'verified');
          const { data: mergedPRs } = await github.rest.search.issuesAndPullRequests({
@@ -41,7 +40,7 @@ jobs:
          if (hasReadyLabel || hasVerifiedLabel || mergedCount >= 4) {
            core.info(`Check passed: verified label=${hasVerifiedLabel}, ready label=${hasReadyLabel}, 4+ merged PRs=${mergedCount >= 4}`);
          } else {
-            core.setFailed(`PR must have the 'verified', 'ready', or 'ready-run-all-tests' label to run pre-commit, or the author must have at least 4 merged PRs (found ${mergedCount}).`);
+            core.setFailed(`PR must have the 'verified' or 'ready' (which also triggers tests) label or the author must have at least 4 merged PRs (found ${mergedCount}).`);
          }
  pre-commit:
@@ -50,7 +49,7 @@ jobs:
    runs-on: [self-hosted, linux, x64, vllm-runners]
    steps:
    - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
-    - uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
+    - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
      with:
        python-version: "3.12"
    # Provide shellcheck on PATH so tools/pre_commit/shellcheck.sh skips its
@@ -1,40 +0,0 @@
 name: Run CI from PR comment
 on:
  issue_comment:
    types: [created]
 concurrency:
  group: run-ci-comment-${{ github.event.issue.number }}
  cancel-in-progress: false
 permissions:
  contents: read
  issues: write
  pull-requests: write
 jobs:
  run-ci-command:
    if: >-
      github.event.issue.pull_request &&
      (github.event.comment.body == '/ci run' ||
      github.event.comment.body == '/ci retry')
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
        with:
          ref: ${{ github.event.repository.default_branch }}
          persist-credentials: false
      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
        with:
          python-version: "3.12"
      - name: Authorize and run CI command
        run: >-
          uv run --no-project --python 3.12
          .github/workflows/scripts/run_ci_command.py
        env:
          BUILDKITE_API_TOKEN: ${{ secrets.BUILDKITE_API_TOKEN }}
          BUILDKITE_ORGANIZATION: vllm
          BUILDKITE_PIPELINE: ci
          CI_TRUSTED_USERS: ${{ vars.CI_TRUSTED_USERS }}
          GH_TOKEN: ${{ github.token }}
@@ -1,581 +0,0 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import json
 import os
 import sys
 import urllib.error
 import urllib.parse
 import urllib.request
 from collections.abc import Mapping, Sequence
 from typing import Any
 COMMAND_RUN_CI = "/ci run"
 COMMAND_RETRY_FAILED = "/ci retry"
 READY_LABELS = {"ready", "ready-run-all-tests"}
 TRUSTED_PERMISSIONS = {"admin", "maintain", "write"}
 ACTIVE_BUILD_STATES = {
    "blocked",
    "creating",
    "scheduled",
    "running",
    "failing",
    "canceling",
    "waiting",
    "waiting_failed",
 }
 RETRY_STATES = "failed,timed_out,expired"
 class ApiError(RuntimeError):
    def __init__(self, status: int | None, message: str) -> None:
        super().__init__(message)
        self.status = status
 class HttpTransport:
    def request(
        self,
        url: str,
        *,
        body: Mapping[str, Any] | None = None,
        headers: Mapping[str, str] | None = None,
        method: str = "GET",
    ) -> Any:
        data = None if body is None else json.dumps(body).encode()
        request = urllib.request.Request(
            url,
            data=data,
            headers=dict(headers or {}),
            method=method,
        )
        try:
            with urllib.request.urlopen(request, timeout=30) as response:
                response_body = response.read().decode()
        except urllib.error.HTTPError as error:
            response_body = error.read().decode()
            message = self._error_message(response_body, error.reason)
            raise ApiError(
                error.code,
                f"API returned {error.code}: {message}",
            ) from error
        except urllib.error.URLError as error:
            raise ApiError(None, f"API request failed: {error.reason}") from error
        if not response_body:
            return None
        try:
            return json.loads(response_body)
        except json.JSONDecodeError as error:
            raise ApiError(None, "API returned a non-JSON response.") from error
    @staticmethod
    def _error_message(response_body: str, fallback: str) -> str:
        try:
            parsed = json.loads(response_body)
        except json.JSONDecodeError:
            return fallback
        return str(parsed.get("message", fallback))
 class GitHubClient:
    def __init__(
        self,
        token: str,
        repository: str,
        transport: HttpTransport | None = None,
    ) -> None:
        if not token:
            raise RuntimeError("GH_TOKEN is not set.")
        self.owner, self.repo = repository.split("/", maxsplit=1)
        self.transport = transport or HttpTransport()
        self.headers = {
            "Accept": "application/vnd.github+json",
            "Authorization": f"Bearer {token}",
            "Content-Type": "application/json",
            "User-Agent": "vllm-ci-command",
            "X-GitHub-Api-Version": "2022-11-28",
        }
    def _request(
        self,
        path: str,
        *,
        body: Mapping[str, Any] | None = None,
        method: str = "GET",
    ) -> Any:
        return self.transport.request(
            f"https://api.github.com{path}",
            body=body,
            headers=self.headers,
            method=method,
        )
    def _repo_path(self, suffix: str) -> str:
        owner = urllib.parse.quote(self.owner, safe="")
        repo = urllib.parse.quote(self.repo, safe="")
        return f"/repos/{owner}/{repo}{suffix}"
    def _paginate(self, path: str) -> list[dict[str, Any]]:
        results: list[dict[str, Any]] = []
        separator = "&" if "?" in path else "?"
        for page in range(1, 101):
            response = self._request(f"{path}{separator}per_page=100&page={page}")
            if not isinstance(response, list):
                raise ApiError(None, "GitHub API returned an invalid list response.")
            results.extend(response)
            if len(response) < 100:
                return results
        raise ApiError(None, "GitHub API pagination exceeded 10,000 results.")
    def get_pr(self, number: int) -> dict[str, Any]:
        return self._request(self._repo_path(f"/pulls/{number}"))
    def get_permission(self, actor: str) -> str:
        username = urllib.parse.quote(actor, safe="")
        try:
            response = self._request(
                self._repo_path(f"/collaborators/{username}/permission")
            )
        except ApiError as error:
            if error.status == 404:
                return "none"
            raise
        return str(response["permission"])
    def get_review_decision(self, number: int) -> str | None:
        query = """
          query($owner: String!, $repo: String!, $number: Int!) {
            repository(owner: $owner, name: $repo) {
              pullRequest(number: $number) {
                reviewDecision
              }
            }
          }
        """
        response = self._request(
            "/graphql",
            body={
                "query": query,
                "variables": {
                    "number": number,
                    "owner": self.owner,
                    "repo": self.repo,
                },
            },
            method="POST",
        )
        return response["data"]["repository"]["pullRequest"]["reviewDecision"]
    def list_reviews(self, number: int) -> list[dict[str, Any]]:
        return self._paginate(self._repo_path(f"/pulls/{number}/reviews"))
    def list_reactions(self, comment_id: int) -> list[dict[str, Any]]:
        return self._paginate(
            self._repo_path(f"/issues/comments/{comment_id}/reactions")
        )
    def add_reaction(self, comment_id: int, content: str) -> None:
        self._request(
            self._repo_path(f"/issues/comments/{comment_id}/reactions"),
            body={"content": content},
            method="POST",
        )
    def add_comment(self, issue_number: int, body: str) -> None:
        self._request(
            self._repo_path(f"/issues/{issue_number}/comments"),
            body={"body": body},
            method="POST",
        )
 class BuildkiteClient:
    def __init__(
        self,
        token: str,
        organization: str,
        pipeline: str,
        transport: HttpTransport | None = None,
    ) -> None:
        self.token = token
        self.transport = transport or HttpTransport()
        organization = urllib.parse.quote(organization, safe="")
        pipeline = urllib.parse.quote(pipeline, safe="")
        self.base_url = (
            "https://api.buildkite.com/v2/organizations/"
            f"{organization}/pipelines/{pipeline}/builds"
        )
    def _request(
        self,
        *,
        body: Mapping[str, Any] | None = None,
        method: str = "GET",
        path: str = "",
        query: Sequence[tuple[str, str]] = (),
    ) -> Any:
        if not self.token:
            raise RuntimeError("The BUILDKITE_API_TOKEN repository secret is not set.")
        url = f"{self.base_url}{path}"
        if query:
            url = f"{url}?{urllib.parse.urlencode(query)}"
        return self.transport.request(
            url,
            body=body,
            headers={
                "Authorization": f"Bearer {self.token}",
                "Content-Type": "application/json",
                "User-Agent": "vllm-ci-command",
            },
            method=method,
        )
    def list_builds(
        self,
        commit: str,
        *,
        metadata: tuple[str, str] | None = None,
    ) -> list[dict[str, Any]]:
        query = [
            ("commit", commit),
            ("exclude_jobs", "true"),
            ("exclude_pipeline", "true"),
            ("per_page", "100"),
        ]
        if metadata:
            key, value = metadata
            query.append((f"meta_data[{key}]", value))
        response = self._request(query=query)
        if not isinstance(response, list):
            raise ApiError(None, "Buildkite API returned an invalid build list.")
        return response
    def create_build(self, body: Mapping[str, Any]) -> dict[str, Any]:
        return self._request(body=body, method="POST")
    def retry_failed_jobs(
        self,
        build_number: int,
        states: str,
    ) -> dict[str, Any]:
        number = urllib.parse.quote(str(build_number), safe="")
        return self._request(
            body={"states": states},
            method="PUT",
            path=f"/{number}/retry_failed_jobs",
        )
 def parse_command(body: str) -> str | None:
    if body in {COMMAND_RUN_CI, COMMAND_RETRY_FAILED}:
        return body
    return None
 def parse_trusted_users(value: str = "") -> set[str]:
    return {
        user.casefold() for item in value.split(",") for user in item.split() if user
    }
 def has_ready_label(pr: Mapping[str, Any]) -> bool:
    return any(label["name"] in READY_LABELS for label in pr["labels"])
 def is_trusted_permission(permission: str) -> bool:
    return permission in TRUSTED_PERMISSIONS
 def authorize(
    *,
    actor: str,
    permission: str,
    pr: Mapping[str, Any],
    trusted_approval: bool = False,
    trusted_users: set[str] | None = None,
 ) -> tuple[bool, str]:
    trusted_users = trusted_users or set()
    if is_trusted_permission(permission):
        return True, f"repository {permission} permission"
    if actor.casefold() in trusted_users:
        return True, "configured trusted contributor"
    if actor.casefold() != pr["user"]["login"].casefold():
        return (
            False,
            "Only reviewers with write access can run CI before it is "
            "delegated to the PR author.",
        )
    if pr["draft"]:
        return False, "PR authors cannot run CI while the PR is a draft."
    if has_ready_label(pr):
        return True, "ready label"
    if trusted_approval:
        return True, "approval from a trusted reviewer"
    return (
        False,
        "A reviewer with write access must run `/ci run`, approve the PR, "
        "or add the `ready` label first.",
    )
 def has_trusted_approval(
    github: GitHubClient,
    number: int,
    trusted_users: set[str],
 ) -> bool:
    if github.get_review_decision(number) != "APPROVED":
        return False
    latest_review_states: dict[str, tuple[str, str]] = {}
    for review in github.list_reviews(number):
        user = review.get("user") or {}
        login = user.get("login")
        state = review.get("state")
        if login and state in {"APPROVED", "CHANGES_REQUESTED", "DISMISSED"}:
            latest_review_states[login.casefold()] = (login, state)
    for login, state in latest_review_states.values():
        if state != "APPROVED":
            continue
        if login.casefold() in trusted_users:
            return True
        if is_trusted_permission(github.get_permission(login)):
            return True
    return False
 def is_build_for_pr(build: Mapping[str, Any], pr_number: int) -> bool:
    pull_request = build.get("pull_request")
    if isinstance(pull_request, Mapping):
        build_pr_number = pull_request.get("id", pull_request.get("number"))
        if build_pr_number is not None:
            return str(build_pr_number) == str(pr_number)
    metadata = build.get("meta_data") or {}
    return str(metadata.get("github-pr-number")) == str(pr_number)
 def is_active_build(build: Mapping[str, Any]) -> bool:
    return bool(build.get("blocked")) or build.get("state") in ACTIVE_BUILD_STATES
 def select_latest_build(
    builds: Sequence[dict[str, Any]],
    pr_number: int,
 ) -> dict[str, Any] | None:
    matching = [build for build in builds if is_build_for_pr(build, pr_number)]
    return max(matching, key=lambda build: build.get("created_at", ""), default=None)
 def create_build_payload(
    *,
    actor: str,
    comment_id: int,
    pr: Mapping[str, Any],
 ) -> dict[str, Any]:
    return {
        "commit": pr["head"]["sha"],
        "branch": pr["head"]["ref"],
        "message": f"PR #{pr['number']} {COMMAND_RUN_CI} by @{actor}",
        "pull_request_id": pr["number"],
        "pull_request_base_branch": pr["base"]["ref"],
        "pull_request_repository": pr["head"]["repo"]["clone_url"],
        "pull_request_labels": [label["name"] for label in pr["labels"]],
        "ignore_pipeline_branch_filters": True,
        "env": {
            "VLLM_CI_GITHUB_COMMENT_ID": str(comment_id),
            "VLLM_CI_TRIGGERED_BY": actor,
        },
        "meta_data": {
            "github-comment-id": str(comment_id),
            "github-pr-number": str(pr["number"]),
            "github-triggered-by": actor,
        },
    }
 def add_reaction_safely(
    github: GitHubClient,
    comment_id: int,
    content: str,
 ) -> None:
    try:
        github.add_reaction(comment_id, content)
    except Exception as error:
        print(f"Could not add {content} reaction: {error}", file=sys.stderr)
 def is_already_handled(github: GitHubClient, comment_id: int) -> bool:
    return any(
        reaction.get("content") in {"rocket", "-1"}
        and (reaction.get("user") or {}).get("login") == "github-actions[bot]"
        for reaction in github.list_reactions(comment_id)
    )
 def handle_run_ci(
    *,
    actor: str,
    buildkite: BuildkiteClient,
    comment_id: int,
    github: GitHubClient,
    pr: Mapping[str, Any],
 ) -> str:
    duplicate_builds = buildkite.list_builds(
        pr["head"]["sha"],
        metadata=("github-comment-id", str(comment_id)),
    )
    duplicate = select_latest_build(duplicate_builds, pr["number"])
    if duplicate:
        return f"CI was already requested by this comment: {duplicate['web_url']}"
    current_builds = buildkite.list_builds(pr["head"]["sha"])
    active_build = next(
        (
            build
            for build in current_builds
            if is_build_for_pr(build, pr["number"]) and is_active_build(build)
        ),
        None,
    )
    if active_build:
        return f"CI is already running for this commit: {active_build['web_url']}"
    current_pr = github.get_pr(pr["number"])
    if current_pr["state"] != "open" or current_pr["head"]["sha"] != pr["head"]["sha"]:
        return (
            "The PR head changed while processing the command. Comment `/ci run` again."
        )
    build = buildkite.create_build(
        create_build_payload(
            actor=actor,
            comment_id=comment_id,
            pr=current_pr,
        )
    )
    return (
        f"Triggered [Buildkite CI #{build['number']}]({build['web_url']}) "
        f"for commit `{current_pr['head']['sha'][:12]}`."
    )
 def handle_retry_failed(
    *,
    buildkite: BuildkiteClient,
    pr: Mapping[str, Any],
 ) -> str:
    builds = buildkite.list_builds(pr["head"]["sha"])
    build = select_latest_build(builds, pr["number"])
    if not build:
        return "No CI build exists for the current PR commit. Use `/ci run` first."
    if not build.get("finished_at") or is_active_build(build):
        return f"CI is still running for this commit: {build['web_url']}"
    retried = buildkite.retry_failed_jobs(build["number"], RETRY_STATES)
    if retried["retried_jobs_count"] == 0:
        return (
            f"No failed, timed-out, or expired jobs need retrying: {build['web_url']}"
        )
    return (
        f"Queued {retried['retried_jobs_count']} failed job(s) for retry in "
        f"[Buildkite CI #{build['number']}]({build['web_url']})."
    )
 def run(
    event: Mapping[str, Any],
    github: GitHubClient,
    buildkite: BuildkiteClient,
    trusted_users_value: str = "",
 ) -> None:
    command = parse_command(event["comment"]["body"])
    if not command or "pull_request" not in event["issue"]:
        return
    issue_number = event["issue"]["number"]
    comment_id = event["comment"]["id"]
    actor = event["comment"]["user"]["login"]
    if is_already_handled(github, comment_id):
        print(f"Comment {comment_id} was already handled.")
        return
    add_reaction_safely(github, comment_id, "eyes")
    try:
        pr = github.get_pr(issue_number)
        permission = github.get_permission(actor)
        if pr["state"] != "open":
            github.add_comment(issue_number, "CI commands require an open PR.")
            return
        trusted_users = parse_trusted_users(trusted_users_value)
        should_check_approval = (
            not is_trusted_permission(permission)
            and actor.casefold() not in trusted_users
            and actor.casefold() == pr["user"]["login"].casefold()
            and not pr["draft"]
            and not has_ready_label(pr)
        )
        trusted_approval = should_check_approval and has_trusted_approval(
            github,
            issue_number,
            trusted_users,
        )
        allowed, reason = authorize(
            actor=actor,
            permission=permission,
            pr=pr,
            trusted_approval=trusted_approval,
            trusted_users=trusted_users,
        )
        if not allowed:
            add_reaction_safely(github, comment_id, "-1")
            github.add_comment(issue_number, f"@{actor}, {reason}")
            return
        print(f"Authorized @{actor}: {reason}")
        if command == COMMAND_RUN_CI:
            message = handle_run_ci(
                actor=actor,
                buildkite=buildkite,
                comment_id=comment_id,
                github=github,
                pr=pr,
            )
        else:
            message = handle_retry_failed(buildkite=buildkite, pr=pr)
        add_reaction_safely(github, comment_id, "rocket")
        github.add_comment(issue_number, message)
    except Exception:
        add_reaction_safely(github, comment_id, "confused")
        raise
 def main() -> None:
    event_path = os.environ["GITHUB_EVENT_PATH"]
    with open(event_path, encoding="utf-8") as event_file:
        event = json.load(event_file)
    if not parse_command(event["comment"]["body"]):
        return
    github = GitHubClient(
        os.environ.get("GH_TOKEN", ""),
        os.environ["GITHUB_REPOSITORY"],
    )
    buildkite = BuildkiteClient(
        os.environ.get("BUILDKITE_API_TOKEN", ""),
        os.environ.get("BUILDKITE_ORGANIZATION", "vllm"),
        os.environ.get("BUILDKITE_PIPELINE", "ci"),
    )
    run(
        event,
        github,
        buildkite,
        os.environ.get("CI_TRUSTED_USERS", ""),
    )
 if __name__ == "__main__":
    main()
@@ -1,363 +0,0 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import unittest
 from typing import Any
 from run_ci_command import (
    COMMAND_RETRY_FAILED,
    COMMAND_RUN_CI,
    RETRY_STATES,
    BuildkiteClient,
    authorize,
    create_build_payload,
    has_trusted_approval,
    is_active_build,
    is_build_for_pr,
    parse_command,
    parse_trusted_users,
    run,
    select_latest_build,
 )
 def make_pr(**overrides: Any) -> dict[str, Any]:
    pr = {
        "base": {"ref": "main"},
        "draft": False,
        "head": {
            "ref": "feature",
            "repo": {"clone_url": "https://github.com/contributor/vllm.git"},
            "sha": "0123456789abcdef",
        },
        "labels": [],
        "number": 42,
        "state": "open",
        "user": {"login": "author"},
    }
    pr.update(overrides)
    return pr
 def make_event(command: str, actor: str = "reviewer") -> dict[str, Any]:
    return {
        "comment": {
            "body": command,
            "id": 99,
            "user": {"login": actor},
        },
        "issue": {
            "number": 42,
            "pull_request": {},
        },
    }
 class FakeGitHub:
    def __init__(
        self,
        *,
        permission: str = "write",
        permissions: dict[str, str] | None = None,
        pr: dict[str, Any] | None = None,
        review_decision: str = "REVIEW_REQUIRED",
        reviews: list[dict[str, Any]] | None = None,
    ) -> None:
        self.comments: list[str] = []
        self.permission = permission
        self.permissions = permissions or {}
        self.pr = pr or make_pr()
        self.reactions: list[str] = []
        self.review_decision = review_decision
        self.reviews = reviews or []
    def get_pr(self, number: int) -> dict[str, Any]:
        return self.pr
    def get_permission(self, actor: str) -> str:
        return self.permissions.get(actor, self.permission)
    def get_review_decision(self, number: int) -> str:
        return self.review_decision
    def list_reviews(self, number: int) -> list[dict[str, Any]]:
        return self.reviews
    def list_reactions(self, comment_id: int) -> list[dict[str, Any]]:
        return []
    def add_reaction(self, comment_id: int, content: str) -> None:
        self.reactions.append(content)
    def add_comment(self, issue_number: int, body: str) -> None:
        self.comments.append(body)
 class FakeBuildkite:
    def __init__(
        self,
        build_lists: list[list[dict[str, Any]]] | None = None,
    ) -> None:
        self.build_lists = build_lists or []
        self.created_builds: list[dict[str, Any]] = []
        self.list_calls: list[tuple[str, tuple[str, str] | None]] = []
        self.retry_calls: list[tuple[int, str]] = []
    def list_builds(
        self,
        commit: str,
        *,
        metadata: tuple[str, str] | None = None,
    ) -> list[dict[str, Any]]:
        self.list_calls.append((commit, metadata))
        return self.build_lists.pop(0)
    def create_build(self, body: dict[str, Any]) -> dict[str, Any]:
        self.created_builds.append(body)
        return {
            "number": 123,
            "web_url": "https://buildkite.example/builds/123",
        }
    def retry_failed_jobs(
        self,
        build_number: int,
        states: str,
    ) -> dict[str, Any]:
        self.retry_calls.append((build_number, states))
        return {"retried_jobs_count": 3}
 class FakeTransport:
    def __init__(self, response: Any) -> None:
        self.calls: list[dict[str, Any]] = []
        self.response = response
    def request(self, url: str, **kwargs: Any) -> Any:
        self.calls.append({"url": url, **kwargs})
        return self.response
 class RunCiCommandTest(unittest.TestCase):
    def test_only_exact_ci_commands_are_accepted(self) -> None:
        self.assertEqual(parse_command(COMMAND_RUN_CI), COMMAND_RUN_CI)
        self.assertEqual(
            parse_command(COMMAND_RETRY_FAILED),
            COMMAND_RETRY_FAILED,
        )
        self.assertIsNone(parse_command("/ci run please"))
        self.assertIsNone(parse_command(" /ci run"))
    def test_write_access_authorizes_reviewers_and_authors(self) -> None:
        allowed, _ = authorize(
            actor="reviewer",
            permission="write",
            pr=make_pr(),
        )
        self.assertTrue(allowed)
    def test_configured_trusted_contributors_can_run_ci(self) -> None:
        trusted_users = parse_trusted_users("trusted-one, TRUSTED-TWO")
        allowed, _ = authorize(
            actor="trusted-two",
            permission="read",
            pr=make_pr(),
            trusted_users=trusted_users,
        )
        self.assertTrue(allowed)
    def test_authors_need_an_approval_or_ready_label(self) -> None:
        pending, _ = authorize(
            actor="author",
            permission="read",
            pr=make_pr(),
        )
        approved, _ = authorize(
            actor="author",
            permission="read",
            pr=make_pr(),
            trusted_approval=True,
        )
        ready, _ = authorize(
            actor="author",
            permission="read",
            pr=make_pr(labels=[{"name": "ready"}]),
        )
        self.assertFalse(pending)
        self.assertTrue(approved)
        self.assertTrue(ready)
    def test_non_author_contributors_without_write_are_denied(self) -> None:
        allowed, _ = authorize(
            actor="contributor",
            permission="read",
            pr=make_pr(),
            trusted_approval=True,
        )
        self.assertFalse(allowed)
    def test_authors_cannot_use_ready_state_on_draft_prs(self) -> None:
        allowed, _ = authorize(
            actor="author",
            permission="read",
            pr=make_pr(draft=True, labels=[{"name": "ready"}]),
            trusted_approval=True,
        )
        self.assertFalse(allowed)
    def test_only_trusted_reviewers_can_delegate_through_approval(self) -> None:
        approved_review = {
            "state": "APPROVED",
            "user": {"login": "reviewer"},
        }
        trusted = FakeGitHub(
            permission="read",
            permissions={"reviewer": "write"},
            review_decision="APPROVED",
            reviews=[approved_review],
        )
        untrusted = FakeGitHub(
            permission="read",
            review_decision="APPROVED",
            reviews=[approved_review],
        )
        self.assertTrue(has_trusted_approval(trusted, 42, set()))
        self.assertFalse(has_trusted_approval(untrusted, 42, set()))
    def test_build_matching_is_scoped_to_the_pr(self) -> None:
        self.assertTrue(is_build_for_pr({"pull_request": {"id": 42}}, 42))
        self.assertFalse(is_build_for_pr({"pull_request": {"id": 43}}, 42))
        self.assertTrue(
            is_build_for_pr(
                {"meta_data": {"github-pr-number": "42"}},
                42,
            )
        )
    def test_latest_build_selection_ignores_other_prs(self) -> None:
        latest = select_latest_build(
            [
                {
                    "created_at": "2026-07-28T02:00:00Z",
                    "number": 3,
                    "pull_request": {"id": 43},
                },
                {
                    "created_at": "2026-07-28T01:00:00Z",
                    "number": 2,
                    "pull_request": {"id": 42},
                },
                {
                    "created_at": "2026-07-28T00:00:00Z",
                    "number": 1,
                    "pull_request": {"id": 42},
                },
            ],
            42,
        )
        self.assertEqual(latest["number"], 2)
    def test_active_build_states_prevent_duplicate_runs(self) -> None:
        self.assertTrue(is_active_build({"state": "scheduled"}))
        self.assertTrue(is_active_build({"state": "running"}))
        self.assertTrue(is_active_build({"state": "waiting"}))
        self.assertTrue(is_active_build({"blocked": True, "state": "passed"}))
        self.assertFalse(is_active_build({"state": "failed"}))
    def test_build_payload_preserves_pr_context(self) -> None:
        payload = create_build_payload(
            actor="reviewer",
            comment_id=99,
            pr=make_pr(labels=[{"name": "ready"}, {"name": "v1"}]),
        )
        self.assertEqual(
            payload,
            {
                "commit": "0123456789abcdef",
                "branch": "feature",
                "message": "PR #42 /ci run by @reviewer",
                "pull_request_id": 42,
                "pull_request_base_branch": "main",
                "pull_request_repository": ("https://github.com/contributor/vllm.git"),
                "pull_request_labels": ["ready", "v1"],
                "ignore_pipeline_branch_filters": True,
                "env": {
                    "VLLM_CI_GITHUB_COMMENT_ID": "99",
                    "VLLM_CI_TRIGGERED_BY": "reviewer",
                },
                "meta_data": {
                    "github-comment-id": "99",
                    "github-pr-number": "42",
                    "github-triggered-by": "reviewer",
                },
            },
        )
    def test_ci_run_dispatches_build_with_current_pr_metadata(self) -> None:
        github = FakeGitHub()
        buildkite = FakeBuildkite([[], []])
        run(make_event(COMMAND_RUN_CI), github, buildkite)
        self.assertEqual(len(buildkite.created_builds), 1)
        self.assertEqual(
            buildkite.created_builds[0]["message"],
            "PR #42 /ci run by @reviewer",
        )
        self.assertEqual(github.reactions, ["eyes", "rocket"])
        self.assertIn("Buildkite CI #123", github.comments[0])
    def test_unapproved_authors_are_denied_without_buildkite(self) -> None:
        github = FakeGitHub(
            permission="read",
            pr=make_pr(),
            review_decision="REVIEW_REQUIRED",
        )
        buildkite = FakeBuildkite()
        run(make_event(COMMAND_RUN_CI, "author"), github, buildkite)
        self.assertEqual(buildkite.list_calls, [])
        self.assertEqual(github.reactions, ["eyes", "-1"])
        self.assertIn("approve the PR", github.comments[0])
    def test_ci_retry_uses_latest_current_sha_build(self) -> None:
        github = FakeGitHub(
            permission="read",
            pr=make_pr(labels=[{"name": "ready"}]),
        )
        buildkite = FakeBuildkite(
            [
                [
                    {
                        "created_at": "2026-07-28T01:00:00Z",
                        "finished_at": "2026-07-28T02:00:00Z",
                        "number": 123,
                        "pull_request": {"id": 42},
                        "state": "failed",
                        "web_url": "https://buildkite.example/builds/123",
                    }
                ]
            ]
        )
        run(make_event(COMMAND_RETRY_FAILED, "author"), github, buildkite)
        self.assertEqual(buildkite.retry_calls, [(123, RETRY_STATES)])
        self.assertIn("Queued 3 failed job", github.comments[0])
    def test_buildkite_retry_uses_retry_failed_jobs_endpoint(self) -> None:
        transport = FakeTransport({"retried_jobs_count": 2})
        client = BuildkiteClient(
            "secret",
            "vllm",
            "ci",
            transport=transport,
        )
        client.retry_failed_jobs(123, RETRY_STATES)
        call = transport.calls[0]
        self.assertEqual(call["method"], "PUT")
        self.assertTrue(call["url"].endswith("/123/retry_failed_jobs"))
        self.assertEqual(call["body"], {"states": RETRY_STATES})
 if __name__ == "__main__":
    unittest.main()
@@ -18,9 +18,6 @@ vllm/third_party/deep_gemm/
 # fmha_sm100 vendored package built from source
 vllm/third_party/fmha_sm100/
 # tml-fa4 vendored package built from source
 vllm/third_party/tml_fa4/
 # triton jit
 .triton
@@ -173,6 +170,9 @@ venv.bak/
 # mkdocs documentation
 /site
 docs/argparse
 docs/examples/*
 !docs/examples/README.md
 # mypy
 .mypy_cache/
@@ -3,9 +3,6 @@ MD007:
 MD013: false
 MD024:
  siblings_only: true
 MD025:
  # Allow front matter title to be different from the first heading in the document.
  front_matter_title: ""
 MD031:
  list_items: false
 MD033: false
@@ -4,7 +4,7 @@ default_install_hook_types:
 default_stages:
  - pre-commit # Run locally
  - manual # Run in CI
-exclude: 'vllm/third_party/.*|vllm/models/kimi_k3/nvidia/ops/third_party/.*|vllm/models/kimi_k3/amd/ops/third_party/.*'
+exclude: 'vllm/third_party/.*'
 repos:
 - repo: https://github.com/astral-sh/ruff-pre-commit
  rev: v0.14.0
@@ -30,7 +30,7 @@ repos:
  - id: markdownlint-cli2
    language_version: lts
    args: [--fix]
-    exclude: (^|/)CLAUDE\.md$
+    exclude: ^CLAUDE\.md$
 - repo: https://github.com/rhysd/actionlint
  rev: v1.7.7
  hooks:
@@ -210,7 +210,7 @@ repos:
    name: Check SPDX headers
    entry: python tools/pre_commit/check_spdx_header.py
    language: python
-    types_or: [python, rust, proto]
+    types: [python]
  - id: check-root-lazy-imports
    name: Check root lazy imports
    entry: python tools/pre_commit/check_init_lazy_imports.py
@@ -260,6 +260,10 @@ repos:
    files: ^docker/(Dockerfile|versions\.json)$
    pass_filenames: false
    additional_dependencies: [dockerfile-parse]
  - id: attention-backend-docs
    name: Check attention backend documentation is up to date
    entry: python tools/pre_commit/generate_attention_backend_docs.py --check
    language: python
  - id: check-boolean-context-manager
    name: Check for boolean ops in with-statements
    entry: python tools/pre_commit/check_boolean_context_manager.py
@@ -0,0 +1,2 @@
 collect_env.py
 vllm/model_executor/layers/fla/ops/*.py
--- a/Show More
+++ b/Show More
		`@@ -0,0 +1,2 @@`
							`collect_env.py`
							`vllm/model_executor/layers/fla/ops/*.py`