Apply pre-commit formatting and use torch.accelerator API

- Apply ruff, clang-format formatting fixes - Replace torch.cuda.set_device/device_count/synchronize with torch.accelerator equivalents per project convention Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com> Signed-off-by: <>
Address review: add push_ar to benchmark and backend logging
2026-06-15 18:21:59 -04:00 · 2026-06-15 18:09:24 -04:00 · 2026-06-15 18:09:24 -04:00 · 2026-06-15 18:09:24 -04:00 · 2026-06-15 18:09:24 -04:00 · 2026-06-15 18:09:24 -04:00
2105 changed files with 62024 additions and 181440 deletions
@@ -2,17 +2,17 @@ name: vllm_intel_ci
 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"
  - "requirements/common.txt"
  - "requirements/xpu.txt"
+  - "requirements/build/cuda.txt"
+  - "requirements/test/cuda.txt"
  - "setup.py"
  - "csrc/"
  - "cmake/"
 run_all_exclude_patterns:
+  - "docker/Dockerfile."
  - "csrc/cpu/"
  - "csrc/rocm/"
  - "cmake/hipify.py"
@@ -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"
@@ -1,45 +1,18 @@
 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  # 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
  # differ ci_base is rebuilt and pushed automatically.
  - label: "AMD: :docker: ensure ci_base"
    key: ensure-ci-base-amd
-    soft_fail: false
-    depends_on:
-      - refresh-rocm-base-amd
+    depends_on: []
    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"
@@ -53,18 +26,40 @@ steps:

  - label: "AMD: :docker: build test image and artifacts"
    key: image-build-amd
-    soft_fail: false
    depends_on:
      - ensure-ci-base-amd
    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"
@@ -17,14 +17,12 @@ 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
  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
@@ -55,7 +53,7 @@ steps:
  - tests/models/language/pooling/
  commands:
    - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 50m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 40m "
      pytest -x -v -s tests/models/language/generation -m cpu_model
      pytest -x -v -s tests/models/language/pooling -m cpu_model"

@@ -70,15 +68,13 @@ steps:
  - vllm/v1/sample/ops/topk_topp_triton.py
  - vllm/v1/sample/ops/topk_topp_sampler.py
  - tests/v1/sample/test_topk_topp_sampler.py
-  - tests/v1/e2e/test_cpu_linear_attn_chunked_prefix.py
  commands:
    - |
      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
      uv pip install git+https://github.com/triton-lang/triton-cpu.git@270e696d
      VLLM_USE_V2_MODEL_RUNNER=1 pytest -x -v -s tests/models/language/generation/test_granite.py -m cpu_model
      # TODO: move to CPU-Kernel Tests once triton-cpu has a pre-built wheel
-      pytest -x -v -s tests/v1/sample/test_topk_topp_sampler.py::TestTritonTopkTopp
-      pytest -x -v -s tests/v1/e2e/test_cpu_linear_attn_chunked_prefix.py"
+      pytest -x -v -s tests/v1/sample/test_topk_topp_sampler.py::TestTritonTopkTopp"

 - label: CPU-Quantization Model Tests
  depends_on: []
@@ -93,13 +89,11 @@ steps:
  - vllm/model_executor/layers/fused_moe/experts/cpu_moe.py
  - tests/quantization/test_compressed_tensors.py
  - tests/quantization/test_cpu_wna16.py
-  - tests/quantization/test_cpu_w8a8.py
  commands:
    - |
      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
      pytest -x -v -s tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs
-      pytest -x -v -s tests/quantization/test_cpu_wna16.py
-      pytest -x -v -s tests/quantization/test_cpu_w8a8.py"
+      pytest -x -v -s tests/quantization/test_cpu_wna16.py"
      
 - label: CPU-Distributed Tests (PP+TP)
  depends_on: []
@@ -142,7 +136,7 @@ steps:
    - |
      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 -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
-  parallelism: 4
+  parallelism: 3

 - label: "Arm CPU Test"
  depends_on: []
@@ -1,80 +0,0 @@
-group: Intel
-steps:
-  - label: ":docker: Build XPU image"
-    soft_fail: true
-    optional: true
-    depends_on: []
-    key: image-build-xpu
-    commands:
-      - bash -lc '.buildkite/image_build/image_build_xpu.sh "public.ecr.aws/q9t5s3a7" "vllm-ci-test-repo" "$BUILDKITE_COMMIT"'
-    env:
-      DOCKER_BUILDKIT: "1"
-    retry:
-      automatic:
-        - exit_status: -1  # Agent was lost
-          limit: 2
-        - exit_status: -10  # Agent was lost
-          limit: 2
-  - label: "XPU example Test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    optional: true
-    device: intel_gpu
-    agent_tags:
-      label: production
-      gpu: 2+
-      mem: 24+
-    no_plugin: true
-    env:
-      REGISTRY: "public.ecr.aws/q9t5s3a7"
-      REPO: "vllm-ci-test-repo"
-    source_file_dependencies:
-      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
-      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh example'
-  - label: "XPU V1 test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    optional: true
-    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"
-    source_file_dependencies:
-      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
-      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh v1'
-  - label: "XPU server test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    optional: true
-    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"
-    source_file_dependencies:
-      - .buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml
-      - .buildkite/scripts/hardware_ci/run-intel-ci-test.sh
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'bash .buildkite/scripts/hardware_ci/run-intel-ci-test.sh server'
@@ -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 +9,29 @@ 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
-  echo "Image found"
-else
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64) ]]; then
  echo "Image not found, proceeding with build..."
-  # build for arm64 GPU targets: Grace/GH200 (sm_90),
-  # Blackwell/Thor (sm_100/sm_103/sm_110), 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 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"
+else
+  echo "Image found"
+  exit 0
 fi

-.buildkite/scripts/annotate-image-build.sh "$IMAGE"
+# build (Grace/GH200 is the arm64 GPU target; sm_90)
+docker build --file docker/Dockerfile \
+  --platform linux/arm64 \
+  --build-arg max_jobs=16 \
+  --build-arg nvcc_threads=4 \
+  --build-arg torch_cuda_arch_list="9.0" \
+  --build-arg USE_SCCACHE=1 \
+  --build-arg buildkite_commit="$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
-  echo "Image found"
-else
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
  echo "Image not found, proceeding with build..."
-  # build
-  docker build --file docker/Dockerfile.cpu \
-    --build-arg max_jobs=16 \
-    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-    --build-arg VLLM_CPU_X86=true \
-    --tag "$IMAGE" \
-    --target vllm-test \
-    --progress plain .
-  # push
-  docker push "$IMAGE"
+else
+  echo "Image found"
+  exit 0
 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
-  echo "Image found"
-else
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu) ]]; then
  echo "Image not found, proceeding with build..."
-  # build
-  docker build --file docker/Dockerfile.cpu \
-    --build-arg max_jobs=16 \
-    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-    --tag "$IMAGE" \
-    --target vllm-test \
-    --progress plain .
-  # push
-  docker push "$IMAGE"
+else
+  echo "Image found"
+  exit 0
 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
-  echo "Image found"
-else
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu) ]]; then
  echo "Image not found, proceeding with build..."
-  # build
-  docker build \
-    --file tests/pytorch_ci_hud_benchmark/Dockerfile.hpu \
-    --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"
+else
+  echo "Image found"
+  exit 0
 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 "$IMAGE" &> /dev/null; then
-  echo "Image found"
-else
+if ! docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-xpu &> /dev/null; then
  echo "Image not found, proceeding with build..."
-  # build
-  docker build \
-    --file docker/Dockerfile.xpu \
-    --build-arg max_jobs=16 \
-    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-    --tag "$IMAGE" \
-    --progress plain .
-  # push
-  docker push "$IMAGE"
+else
+  echo "Image found"
+  exit 0
 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
@@ -5,10 +5,6 @@ steps:
 - label: XPU Sleep Mode
  timeout_in_minutes: 30
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -23,5 +19,4 @@ steps:
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
-      pytest -v -s basic_correctness/test_cpu_offload.py &&
      pytest -v -s basic_correctness/test_mem.py::test_end_to_end'
@@ -5,10 +5,6 @@ steps:
 - label: Engine (1 GPU)
  timeout_in_minutes: 30
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -1,15 +1,11 @@
 group: Expert Parallelism
-depends_on:
+depends_on: 
  - image-build-xpu
 steps:
 - label: EPLB Algorithm
  key: eplb-algorithm
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -5,10 +5,6 @@ steps:
 - label: vLLM IR Tests
  timeout_in_minutes: 30
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -5,10 +5,6 @@ steps:
 - label: LoRA Runtime + Utils
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -38,10 +34,6 @@ steps:
 - label: LoRA Fused/MoE Kernels
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -62,10 +54,6 @@ steps:
 - label: LoRA Punica Kernels
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -81,17 +69,11 @@ 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::test_kernels_hidden_size &&
-      pytest -v -s lora/test_punica_ops.py::test_add_lora_fused_moe_early_exit'
+      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]"'

 - label: LoRA Punica FP8/XPU Ops
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -112,10 +94,6 @@ steps:
 - label: LoRA Models
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 2+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -130,19 +108,15 @@ steps:
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      (pytest -v -s lora/test_mixtral.py --deselect="tests/lora/test_mixtral.py::test_mixtral_lora[4]" || true) &&
      pytest -v -s lora/test_quant_model.py --deselect="tests/lora/test_quant_model.py::test_quant_model_lora[model0]" --deselect="tests/lora/test_quant_model.py::test_quant_model_lora[model1]" --deselect="tests/lora/test_quant_model.py::test_quant_model_tp_equality[model0]" &&
      pytest -v -s lora/test_transformers_model.py &&
      pytest -v -s lora/test_chatglm3_tp.py &&
-      pytest -v -s lora/test_llama_tp.py::test_llama_lora &&
      pytest -s -v lora/test_minicpmv_tp.py'

 - label: LoRA Multimodal
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -5,10 +5,6 @@ steps:
 - label: V1 Core + KV + Metrics
  timeout_in_minutes: 30
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -35,10 +31,6 @@ steps:
 - label: V1 Sample + Logits
  timeout_in_minutes: 30
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -65,46 +57,17 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'pip install lm_eval[api]>=0.4.12 &&
-      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      'export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
      cd tests &&
      pytest -v -s v1/logits_processors --ignore=v1/logits_processors/test_custom_online.py --ignore=v1/logits_processors/test_custom_offline.py &&
      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'
-
- label: Basic Models Tests (Initialization)
-  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]'
+      pytest -v -s v1/sample/test_topk_topp_sampler.py'

 - label: XPU CPU Offload
  timeout_in_minutes: 60
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -125,39 +88,10 @@ steps:
      pytest -v -s v1/kv_offload &&
      pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'

- label: NixlConnector PD accuracy (2 GPUs)
-  timeout_in_minutes: 60
-  num_devices: 2
-  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/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'
-
 - label: Regression
  key: regression
  timeout_in_minutes: 30
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -180,7 +114,7 @@ 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'

@@ -189,10 +123,6 @@ steps:
  timeout_in_minutes: 30
  num_devices: 2
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 2+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -224,10 +154,6 @@ steps:
  key: async-engine-inputs-utils-worker
  timeout_in_minutes: 30
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -1,62 +0,0 @@
-group: Model Runner V2 Intel
-depends_on:
-  - image-build-xpu
-steps:
- label: Model Runner V2 Core Tests (Intel)
-  timeout_in_minutes: 45
-  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
-    - vllm/v1/core/sched/
-    - vllm/v1/attention/
-    - tests/v1/engine/test_llm_engine.py
-    - tests/v1/e2e/
-  commands:
-    - >-
-      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
-      cd tests &&
-      pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics" &&
-      ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram" &&
-      pytest -v -s v1/e2e/general/test_min_tokens.py'
-
- label: Model Runner V2 Examples (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/v1/worker/gpu/
-    - vllm/v1/core/sched/
-    - vllm/v1/worker/gpu_worker.py
-    - examples/basic/offline_inference/
-    - examples/generate/multimodal/
-    - examples/features/
-  commands:
-    - >-
-      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
-      cd examples &&
-      python3 basic/offline_inference/chat.py &&
-      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'
@@ -1,27 +0,0 @@
-group: Models - Distributed
-depends_on: 
-  - image-build-xpu
-steps:
- label: Distributed Model Tests (2 GPUs)
-  key: distributed-model-tests-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/model_executor/model_loader/sharded_state_loader.py
-  - vllm/model_executor/models/
-  - tests/model_executor/model_loader/test_sharded_state_loader.py
-  commands:
-    - >-
-      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'cd tests &&
-      pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m "not slow_test"'
@@ -1,15 +1,11 @@
 group: Models - Multimodal
-depends_on:
+depends_on: 
  - image-build-xpu
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
  key: multi-modal-models-standard-1-qwen2
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -22,7 +18,7 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'pip install av &&
+      'pip install av git+https://github.com/TIGER-AI-Lab/Mantis.git &&
      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'
@@ -31,10 +27,6 @@ steps:
  key: multi-modal-models-standard-2-qwen3-gemma
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 16+
  no_plugin: true
  working_dir: "."
  env:
@@ -47,17 +39,14 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'cd tests &&
+      'pip install git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
      pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model'

 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
  key: multi-modal-models-standard-3-llava-qwen2-vl
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -70,7 +59,8 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'cd tests &&
+      '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'

@@ -78,10 +68,6 @@ steps:
  key: multi-modal-models-standard-4-other-whisper
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -94,7 +80,7 @@ steps:
  commands:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-      'pip install av &&
+      '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'

@@ -102,10 +88,6 @@ steps:
  key: multi-modal-processor
  timeout_in_minutes: 45
  device: intel_gpu
-  agent_tags:
-    label: production
-    gpu: 1+
-    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -119,9 +101,11 @@ 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
+      --deselect "tests/models/multimodal/processing/test_tensor_schema.py::test_model_tensor_schema[mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4]"
+      --deselect "tests/models/multimodal/processing/test_tensor_schema.py::test_model_tensor_schema[Qwen/Qwen2.5-Omni-7B-AWQ]"
      --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB'
  parallelism: 4
@@ -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]"'
-
@@ -19,10 +19,6 @@ steps:
      - image-build-xpu
    timeout_in_minutes: 30
    device: intel_gpu
-    agent_tags:
-      label: production
-      gpu: 2+
-      mem: 24+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -42,46 +38,17 @@ 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: 30
    device: intel_gpu
-    agent_tags:
-      label: production
-      gpu: 1+
-      mem: 24+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -99,17 +66,13 @@ steps:
        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/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/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 &&
        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: 30
    device: intel_gpu
-    agent_tags:
-      label: production
-      gpu: 1+
-      mem: 16+
    no_plugin: true
    env:
      REGISTRY: "public.ecr.aws/q9t5s3a7"
@@ -124,47 +87,3 @@ steps:
        cd tests &&
        pytest -v -s entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py &&
        pytest -v -s benchmarks/test_serve_cli.py'
-  - label: "XPU quantization test"
-    depends_on:
-      - image-build-xpu
-    timeout_in_minutes: 30
-    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"
-    source_file_dependencies:
-      - vllm/
-      - .buildkite/intel_jobs/test-intel.yaml 
-    commands:
-      - >-
-        bash .buildkite/scripts/hardware_ci/run-intel-test.sh
-        'cd tests &&
-        pytest -v -s quantization/test_auto_round.py'
-  - 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'
@@ -350,25 +350,6 @@ steps:
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"
          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"'

-      - label: ":docker: Build release image - x86_64 - XPU"
-        depends_on: ~
-        id: build-xpu-release-image
-        agents:
-          queue: cpu_queue_release
-        commands:
-          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - |
-            DOCKER_BUILDKIT=1 docker build \
-              $(bash .buildkite/scripts/docker-build-metadata-args.sh xpu) \
-              --build-arg GIT_REPO_CHECK=1 \
-              --target vllm-openai \
-              --progress plain \
-              -f docker/Dockerfile.xpu .
-          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-xpu"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-xpu"'
-        env:
-          DOCKER_BUILDKIT: "1"
-
      - block: "Build release image for x86_64 CPU"
        key: block-cpu-release-image-build
        depends_on: ~
@@ -464,16 +445,6 @@ steps:
          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"
          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 12.9 Ubuntu 24.04" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"'

-      - label: "Create manifest - XPU"
-        depends_on:
-          - build-xpu-release-image
-        id: create-manifest-xpu
-        agents:
-          queue: small_cpu_queue_release
-        commands:
-          - "bash .buildkite/scripts/xpu/create-xpu-ecr-manifest.sh"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: XPU" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-xpu"'
-
      - label: "Publish nightly multi-arch image to DockerHub"
        depends_on:
          - create-multi-arch-manifest
@@ -848,23 +819,6 @@ steps:
      DOCKER_BUILDKIT: "1"
      S3_BUCKET: "vllm-wheels"

-  - label: "Publish nightly XPU image to DockerHub"
-    depends_on:
-      - create-manifest-xpu
-    if: build.env("NIGHTLY") == "1"
-    agents:
-      queue: small_cpu_queue_release
-    commands:
-      - "bash .buildkite/scripts/xpu/push-nightly-builds-xpu.sh"
-      - "bash .buildkite/scripts/cleanup-nightly-builds.sh nightly- vllm/vllm-openai-xpu"
-    plugins:
-      - docker-login#v3.0.0:
-          username: vllmbot
-          password-env: DOCKERHUB_TOKEN
-    env:
-      DOCKER_BUILDKIT: "1"
-      DOCKERHUB_USERNAME: "vllmbot"
-
  - label: "Publish nightly ROCm image to DockerHub"
    depends_on:
      - build-rocm-release-image
@@ -895,7 +849,6 @@ steps:
      - create-multi-arch-manifest-cuda-12-9
      - create-multi-arch-manifest-ubuntu2404
      - create-multi-arch-manifest-cuda-12-9-ubuntu2404
-      - create-manifest-xpu
      - build-rocm-release-image
      - input-release-version
      # Wait for CPU builds if their block steps were unblocked, so publish
@@ -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
@@ -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/nightly/cu130"
-    else
-        TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
-    fi
+    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
 fi
 echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"

@@ -15,10 +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 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 tools/install_torchcodec_rocm.sh tests/vllm_test_utils"
 DEFAULT_CI_BASE_DOCKERFILE="docker/Dockerfile.rocm"
 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

 TARGET=""
@@ -393,16 +392,6 @@ 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"
@@ -536,22 +525,6 @@ get_remote_image_label_with_retry() {
    return 0
 }

-remote_ci_base_metadata_is_current() {
-    local image_ref="$1"
-    local metadata_version=""
-
-    metadata_version=$(get_remote_image_label "${image_ref}" "vllm.ci_base.metadata_version")
-    [[ "${metadata_version}" == "${CI_BASE_METADATA_VERSION:-${DEFAULT_CI_BASE_METADATA_VERSION}}" ]]
-}
-
-remote_ci_base_metadata_is_current_with_retry() {
-    local image_ref="$1"
-    local metadata_version=""
-
-    metadata_version=$(get_remote_image_label_with_retry "${image_ref}" "vllm.ci_base.metadata_version")
-    [[ "${metadata_version}" == "${CI_BASE_METADATA_VERSION:-${DEFAULT_CI_BASE_METADATA_VERSION}}" ]]
-}
-
 remote_image_exists() {
    local image_ref="$1"
    docker manifest inspect "${image_ref}" >/dev/null 2>&1
@@ -608,7 +581,6 @@ init_config() {
    CI_BASE_CONTENT_FILES="${CI_BASE_CONTENT_FILES:-${DEFAULT_CI_BASE_CONTENT_FILES}}"
    CI_BASE_DOCKERFILE="${CI_BASE_DOCKERFILE:-${DEFAULT_CI_BASE_DOCKERFILE}}"
    CI_BASE_DOCKERFILE_STAGES="${CI_BASE_DOCKERFILE_STAGES:-${DEFAULT_CI_BASE_DOCKERFILE_STAGES}}"
-    CI_BASE_METADATA_VERSION="${CI_BASE_METADATA_VERSION:-${DEFAULT_CI_BASE_METADATA_VERSION}}"
    CI_BASE_IMAGE_TAG="${CI_BASE_IMAGE_TAG:-rocm/vllm-dev:ci_base}"
    export PYTORCH_ROCM_ARCH

@@ -663,10 +635,6 @@ load_ci_hcl() {
    echo "Copied ${CI_HCL_SOURCE} to ${CI_HCL_PATH}"
 }

-init_bake_files() {
-    BAKE_FILES=(-f "${VLLM_BAKE_FILE}" -f "${CI_HCL_PATH}")
-}
-
 compute_ci_base_hash_if_needed() {
    if [[ -z "${CI_BASE_CONTENT_FILES:-}" ]]; then
        return 0
@@ -708,14 +676,12 @@ configure_ci_base_image_refs() {
    fi

    content_tag=$(ci_base_tag_with_suffix "${stable_tag}" "${CI_BASE_CONTENT_HASH}")
-    CI_BASE_IMAGE_TAG_CONTENT_REF="${content_tag}"
    if [[ -n "${BUILDKITE_COMMIT:-}" ]]; then
        commit_tag=$(ci_base_tag_with_suffix "${stable_tag}" "${BUILDKITE_COMMIT}")
+        CI_BASE_IMAGE_TAG_COMMIT="${commit_tag}"
+        export CI_BASE_IMAGE_TAG_COMMIT
    fi
-    CI_BASE_IMAGE_TAG_COMMIT_REF="${commit_tag}"

-    # *_REF is the logical tag recorded in metadata. *_EXTRA is only passed to
-    # bake when that tag is not already the primary tag, avoiding duplicates.
    if should_push_stable_ci_base_tag; then
        primary_tag="${content_tag}"
        CI_BASE_IMAGE_TAG_STABLE="${stable_tag}"
@@ -725,35 +691,19 @@ configure_ci_base_image_refs() {
    fi
    CI_BASE_IMAGE_TAG="${primary_tag}"
    if [[ "${primary_tag}" == "${content_tag}" ]]; then
-        CI_BASE_IMAGE_TAG_CONTENT_EXTRA=""
+        CI_BASE_IMAGE_TAG_CONTENT=""
    else
-        CI_BASE_IMAGE_TAG_CONTENT_EXTRA="${content_tag}"
+        CI_BASE_IMAGE_TAG_CONTENT="${content_tag}"
    fi
-    if [[ -n "${commit_tag}" && "${commit_tag}" != "${primary_tag}" ]]; then
-        CI_BASE_IMAGE_TAG_COMMIT_EXTRA="${commit_tag}"
-    else
-        CI_BASE_IMAGE_TAG_COMMIT_EXTRA=""
-    fi
-    export CI_BASE_IMAGE_TAG
-    export CI_BASE_IMAGE_TAG_COMMIT_EXTRA
-    export CI_BASE_IMAGE_TAG_CONTENT_EXTRA
-    export CI_BASE_IMAGE_TAG_CONTENT_REF
-    export CI_BASE_IMAGE_TAG_COMMIT_REF
-    export CI_BASE_IMAGE_TAG_STABLE
+    export CI_BASE_IMAGE_TAG CI_BASE_IMAGE_TAG_CONTENT CI_BASE_IMAGE_TAG_STABLE

    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: ${CI_BASE_IMAGE_TAG}"
-        if [[ -n "${commit_tag}" ]]; then
-            if [[ "${commit_tag}" == "${primary_tag}" ]]; then
-                echo "ci_base commit image tag: ${commit_tag} (primary)"
-            else
-                echo "ci_base commit image tag: ${commit_tag}"
-            fi
+        if [[ -n "${CI_BASE_IMAGE_TAG_COMMIT:-}" ]]; then
+            echo "ci_base commit image tag: ${CI_BASE_IMAGE_TAG_COMMIT}"
        fi
        echo "ci_base content image tag: ${content_tag}"
        if [[ -n "${CI_BASE_IMAGE_TAG_STABLE}" ]]; then
@@ -762,10 +712,6 @@ 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:-}"
-        set_buildkite_metadata "rocm-ci-base-image-stable" "${CI_BASE_IMAGE_TAG_STABLE:-}"
        return 0
    fi

@@ -782,8 +728,8 @@ ci_base_candidate_refs() {
    printf '%s\n' \
        "${IMAGE_TAG:-}" \
        "${CI_BASE_IMAGE_TAG:-}" \
-        "${CI_BASE_IMAGE_TAG_COMMIT_EXTRA:-}" \
-        "${CI_BASE_IMAGE_TAG_CONTENT_EXTRA:-}" \
+        "${CI_BASE_IMAGE_TAG_COMMIT:-}" \
+        "${CI_BASE_IMAGE_TAG_CONTENT:-}" \
        "${CI_BASE_IMAGE_TAG_STABLE:-}" \
        | awk 'NF && !seen[$0]++'
 }
@@ -797,10 +743,6 @@ find_matching_ci_base_ref() {
        remote_image_exists "${candidate}" || continue
        candidate_hash=$(get_remote_image_label "${candidate}" "vllm.ci_base.content_hash")
        if [[ "${candidate_hash}" == "${CI_BASE_CONTENT_HASH}" ]]; then
-            if ! remote_ci_base_metadata_is_current "${candidate}"; then
-                echo "Found matching ci_base content hash but stale metadata: ${candidate}" >&2
-                continue
-            fi
            printf '%s\n' "${candidate}"
            return 0
        fi
@@ -875,10 +817,6 @@ maybe_skip_existing_image() {
        if [[ -n "${remote_hash}" ]]; then
            echo "Remote ci_base content hash: ${remote_hash:0:16}..."
            if [[ "${remote_hash}" == "${CI_BASE_CONTENT_HASH}" ]]; then
-                if ! remote_ci_base_metadata_is_current "${IMAGE_TAG}"; then
-                    echo "Content hashes match but ci_base metadata is stale; rebuilding to refresh metadata"
-                    return 0
-                fi
                if ! refresh_ci_base_tags_from_ref "${IMAGE_TAG}"; then
                    echo "ci_base tag refresh failed; rebuilding to push expected tags"
                    return 0
@@ -1060,104 +998,12 @@ prepare_git_cache_metadata() {
    fi
 }

-ci_base_metadata_pairs() {
-    local dockerfile="${CI_BASE_DOCKERFILE:-${DEFAULT_CI_BASE_DOCKERFILE}}"
-    local stages="${CI_BASE_DOCKERFILE_STAGES:-${DEFAULT_CI_BASE_DOCKERFILE_STAGES}}"
-    local content_files="${CI_BASE_CONTENT_FILES:-${DEFAULT_CI_BASE_CONTENT_FILES}}"
-    local content_files_hash=""
-    local base_image=""
-    local base_image_digest=""
-    local git_branch=""
-    local -a content_paths=()
-    local -a content_args=()
-
-    read -r -a content_paths <<< "${content_files}"
-    if [[ ${#content_paths[@]} -gt 0 ]]; then
-        content_files_hash=$(compute_content_hash "${content_paths[@]}")
-    fi
-    mapfile -t content_args < <(
-        get_content_arg_names "${dockerfile}" "${stages}" "${CI_BASE_CONTENT_ARGS:-}"
-    )
-
-    base_image=$(resolve_dockerfile_arg_value "${dockerfile}" "BASE_IMAGE")
-    if [[ -n "${base_image}" ]]; then
-        base_image_digest=$(resolve_image_digest "${base_image}")
-    fi
-    git_branch="${BUILDKITE_BRANCH:-${VLLM_BRANCH:-}}"
-
-    metadata_pair "vllm.ci_base.metadata_version" "${CI_BASE_METADATA_VERSION:-${DEFAULT_CI_BASE_METADATA_VERSION}}"
-    metadata_pair "vllm.ci_base.content_hash" "${CI_BASE_CONTENT_HASH:-}"
-    metadata_pair "vllm.ci_base.content_files_hash" "${content_files_hash}"
-    metadata_pair "vllm.ci_base.content_files" "${content_files}"
-    metadata_pair "vllm.ci_base.content_args" "$(join_words "${content_args[@]}")"
-    metadata_pair "vllm.ci_base.dockerfile" "${dockerfile}"
-    metadata_pair "vllm.ci_base.dockerfile_stages" "${stages}"
-    metadata_pair "vllm.ci_base.image.primary" "${CI_BASE_IMAGE_TAG:-}"
-    metadata_pair "vllm.ci_base.image.content" "${CI_BASE_IMAGE_TAG_CONTENT_REF:-${CI_BASE_IMAGE_TAG_CONTENT_EXTRA:-}}"
-    metadata_pair "vllm.ci_base.image.commit" "${CI_BASE_IMAGE_TAG_COMMIT_REF:-${CI_BASE_IMAGE_TAG_COMMIT_EXTRA:-}}"
-    metadata_pair "vllm.ci_base.image.stable" "${CI_BASE_IMAGE_TAG_STABLE:-}"
-    metadata_pair "vllm.ci_base.git_commit" "${BUILDKITE_COMMIT:-}"
-    metadata_pair "vllm.ci_base.git_branch" "${git_branch}"
-    metadata_pair "vllm.ci_base.vllm_branch" "${VLLM_BRANCH:-}"
-    metadata_pair "vllm.ci_base.stable_branch" "${CI_BASE_STABLE_BRANCH:-main}"
-
-    metadata_pair "vllm.rocm.base_image" "${base_image}"
-    metadata_pair "vllm.rocm.base_image_digest" "${base_image_digest}"
-    metadata_pair "vllm.rocm.pytorch_rocm_arch" "${PYTORCH_ROCM_ARCH:-}"
-    metadata_pair "vllm.rocm.nic_backend" "$(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 "${dockerfile}" "UBUNTU_CODENAME")"
-    metadata_pair "vllm.rocm.rixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_REPO")"
-    metadata_pair "vllm.rocm.rixl_commit" "${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 "${dockerfile}" "ROCSHMEM_REPO")"
-    metadata_pair "vllm.rocm.rocshmem_commit" "${ROCSHMEM_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "ROCSHMEM_BRANCH")}"
-    metadata_pair "vllm.rocm.deepep_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_REPO")"
-    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 "${dockerfile}" "DEEPEP_ROCM_ARCH")"
-    metadata_pair "vllm.rocm.rixl_cache_key" "${RIXL_CACHE_KEY:-}"
-    metadata_pair "vllm.rocm.rocshmem_cache_key" "${ROCSHMEM_CACHE_KEY:-}"
-    metadata_pair "vllm.rocm.deepep_cache_key" "${DEEPEP_CACHE_KEY:-}"
-
-    metadata_pair "vllm.buildkite.build_number" "${BUILDKITE_BUILD_NUMBER:-}"
-    metadata_pair "vllm.buildkite.build_id" "${BUILDKITE_BUILD_ID:-}"
-}
-
-write_ci_base_metadata_annotations() {
-    local metadata="$1"
-    local key=""
-    local value=""
-    local annotation=""
-
-    [[ -n "${metadata}" ]] || return 0
-    while IFS=$'\t' read -r key value; do
-        [[ -n "${key}" && -n "${value}" ]] || continue
-        annotation="manifest:${key}=${value}"
-        printf '    "%s",\n' "$(hcl_escape_string "${annotation}")"
-    done <<< "${metadata}"
-}
-
-write_ci_base_metadata_labels() {
-    local metadata="$1"
-    local key=""
-    local value=""
-
-    [[ -n "${metadata}" ]] || return 0
-    while IFS=$'\t' read -r key value; do
-        [[ -n "${key}" && -n "${value}" ]] || continue
-        printf '    "%s" = "%s"\n' \
-            "$(hcl_escape_string "${key}")" \
-            "$(hcl_escape_string "${value}")"
-    done <<< "${metadata}"
-}
-
 write_ci_base_label_override() {
    local target_name=""
-    local metadata=""
    local -a ci_base_targets=()

+    BAKE_FILES=(-f "${VLLM_BAKE_FILE}" -f "${CI_HCL_PATH}")
+
    if [[ -z "${CI_BASE_CONTENT_HASH:-}" ]]; then
        return 0
    fi
@@ -1173,23 +1019,16 @@ write_ci_base_label_override() {
        return 0
    fi

-    metadata=$(ci_base_metadata_pairs)
-
    : > "${CI_BASE_LABEL_OVERRIDE_PATH}"
    for target_name in "${ci_base_targets[@]}"; do
        cat >> "${CI_BASE_LABEL_OVERRIDE_PATH}" <<EOF
 target "${target_name}" {
  annotations = [
    "manifest:org.opencontainers.image.revision=",
-EOF
-        write_ci_base_metadata_annotations "${metadata}" >> "${CI_BASE_LABEL_OVERRIDE_PATH}"
-        cat >> "${CI_BASE_LABEL_OVERRIDE_PATH}" <<EOF
  ]
  labels = {
    "org.opencontainers.image.revision" = ""
-EOF
-        write_ci_base_metadata_labels "${metadata}" >> "${CI_BASE_LABEL_OVERRIDE_PATH}"
-        cat >> "${CI_BASE_LABEL_OVERRIDE_PATH}" <<EOF
+    "vllm.ci_base.content_hash" = "${CI_BASE_CONTENT_HASH}"
  }
 }

@@ -1197,7 +1036,7 @@ EOF
    done

    BAKE_FILES+=(-f "${CI_BASE_LABEL_OVERRIDE_PATH}")
-    echo "Appended ci_base metadata label override for targets: ${ci_base_targets[*]}"
+    echo "Appended ci_base content-hash label override for targets: ${ci_base_targets[*]}"
 }

 uses_rocm_csrc_cache() {
@@ -1280,18 +1119,6 @@ hcl_escape_string() {
    printf '%s' "${value}"
 }

-join_words() {
-    local IFS=" "
-    printf '%s' "$*"
-}
-
-metadata_pair() {
-    local key="$1"
-    local value="${2:-}"
-
-    printf '%s\t%s\n' "${key}" "${value}"
-}
-
 write_hcl_string_list() {
    local indent="$1"
    shift
@@ -1714,13 +1541,7 @@ confirm_remote_image_push() {

        remote_hash=$(get_remote_image_label_with_retry "${image_ref}" "vllm.ci_base.content_hash")
        if [[ -n "${remote_hash}" && "${remote_hash}" == "${CI_BASE_CONTENT_HASH}" ]]; then
-            if remote_ci_base_metadata_is_current_with_retry "${image_ref}"; then
-                return 0
-            fi
-
-            echo "Remote image exists with the expected ci_base content hash but stale metadata."
-            echo "  expected metadata version: ${CI_BASE_METADATA_VERSION:-${DEFAULT_CI_BASE_METADATA_VERSION}}"
-            return 1
+            return 0
        fi

        echo "Remote image exists but does not have the expected ci_base content hash."
@@ -1795,10 +1616,7 @@ seed_dependency_caches_if_needed() {

        echo "--- :docker: Seeding ${target}"
        echo "Expected cache ref: ${cache_ref}"
-        docker buildx bake \
-            "${BAKE_FILES[@]}" \
-            --progress "${BUILDKIT_PROGRESS:-plain}" \
-            "${target}"
+        docker buildx bake "${BAKE_FILES[@]}" --progress plain "${target}"
        verify_dependency_cache_ref "${cache_ref}"
    done
 }
@@ -1826,10 +1644,7 @@ run_bake() {
    local build_rc=0

    echo "--- :docker: Building ${TARGET}"
-    docker buildx bake \
-        "${BAKE_FILES[@]}" \
-        --progress "${BUILDKIT_PROGRESS:-plain}" \
-        "${BAKE_TARGETS[@]}" || build_rc=$?
+    docker buildx bake "${BAKE_FILES[@]}" --progress plain "${BAKE_TARGETS[@]}" || build_rc=$?

    if [[ ${build_rc} -eq 0 ]]; then
        echo "--- :white_check_mark: Build complete"
@@ -1909,16 +1724,15 @@ main() {
    print_header
    validate_inputs
    load_ci_hcl
-    init_bake_files
    compute_ci_base_hash_if_needed
    configure_ci_base_image_refs
    maybe_skip_existing_image
    setup_builder
    prepare_git_cache_metadata
+    write_ci_base_label_override
    extract_dependency_pins
    write_rocm_build_arg_override
    compute_dependency_cache_keys
-    write_ci_base_label_override
    compute_rocm_csrc_content_hash_if_needed
    write_rocm_cache_override
    resolve_ci_base_dependency_targets
@@ -28,17 +28,6 @@
 ###############################################################################
 set -o pipefail

-: "${BUILDKIT_PROGRESS:=plain}"
-: "${TERM:=xterm-256color}"
-: "${FORCE_COLOR:=1}"
-: "${CLICOLOR_FORCE:=1}"
-: "${PY_COLORS:=1}"
-: "${ROCM_DOCKER_TTY:=1}"
-if [[ " ${PYTEST_ADDOPTS:-} " != *" --color"* ]]; then
-  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--color=yes"
-fi
-export BUILDKIT_PROGRESS TERM FORCE_COLOR CLICOLOR_FORCE PY_COLORS PYTEST_ADDOPTS ROCM_DOCKER_TTY
-
 # Export Python path for commands that run directly on the host. Containerized
 # tests set this to /vllm-workspace below so spawned Python processes do not
 # depend on their current working directory.
@@ -160,7 +149,6 @@ EOF
  echo "--- Building local ROCm test image"
  docker build \
    --pull=false \
-    --progress "${BUILDKIT_PROGRESS}" \
    --build-arg "BASE_IMAGE=${base_image}" \
    -t "${artifact_image}" \
    "${context_dir}" || return 1
@@ -379,20 +367,6 @@ remove_docker_container() {
 }
 trap remove_docker_container EXIT

-# python_only_compile.sh runs `python setup.py develop` and needs the full repo tree
-# under /vllm-workspace (Dockerfile.rocm test stage: mkdir src && mv vllm).
-# The ROCm wheel artifact tarball only ships a thin tree (tests, etc.), so
-# artifact images cannot satisfy that test — use the full rocm/vllm-ci image.
-_cmd_probe="${VLLM_TEST_COMMANDS:-}"
-if [[ -z "${_cmd_probe}" ]]; then
-  _cmd_probe="$*"
-fi
-if [[ "${VLLM_CI_USE_ARTIFACTS:-0}" == "1" && "${_cmd_probe}" == *python_only_compile.sh* ]]; then
-  echo "INFO: disabling VLLM_CI_USE_ARTIFACTS for python_only_compile (requires full /vllm-workspace tree)"
-  export VLLM_CI_USE_ARTIFACTS=0
-fi
-unset -v _cmd_probe
-
 if ! prepare_artifact_image; then
  echo "Using full ROCm CI image: ${image_name}"
  docker pull "${image_name}" || exit 1
@@ -452,26 +426,6 @@ fi

 echo "Final commands: $commands"

-# The ROCm test image often ships /vllm-workspace without .git (artifact tarball unpack).
-# tests/standalone_tests/python_only_compile.sh uses merge-base(HEAD, origin/main) for
-# wheels.vllm.ai; compute on the agent (full git checkout) and pass into the container.
-vllm_standalone_merge_base=""
-checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
-if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
-  checkout="."
-fi
-# Pass safe.directory per-command (-c) because buildkite runs will always fail
-# the next check on git 2.35.2+ due to mixed uses of root and buildkite-agent/uids.
-if git -c "safe.directory=${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 VLLM_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
-
 MYPYTHONPATH="/vllm-workspace"

 container_job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-0}}"
@@ -547,37 +501,14 @@ 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=${HF_HUB_ETAG_TIMEOUT}" \
@@ -585,11 +516,6 @@ else
    -e AWS_SECRET_ACCESS_KEY \
    -e BUILDKITE_PARALLEL_JOB \
    -e BUILDKITE_PARALLEL_JOB_COUNT \
-    -e TERM \
-    -e FORCE_COLOR \
-    -e CLICOLOR_FORCE \
-    -e PY_COLORS \
-    -e PYTEST_ADDOPTS \
    -v "${HF_CACHE}:${HF_MOUNT}" \
    -e "HF_HOME=${HF_MOUNT}" \
    -e "PYTHONPATH=${MYPYTHONPATH}" \
@@ -599,7 +525,6 @@ else
    -e "VLLM_CACHE_ROOT=${CONTAINER_CACHE_ROOT}/vllm" \
    -e "XDG_CACHE_HOME=${CONTAINER_CACHE_ROOT}/xdg" \
    -e "PYTORCH_ROCM_ARCH=" \
-    -e "VLLM_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}" \
    --name "${container_name}" \
    "${image_name}" \
    /bin/bash -c "${CONTAINER_PREFLIGHT} && ${commands}"
@@ -8,7 +8,7 @@ set -ex
 CORE_RANGE=${CORE_RANGE:-0-31}
 OMP_CORE_RANGE=${OMP_CORE_RANGE:-0-31}

-export CMAKE_BUILD_PARALLEL_LEVEL=32
+export CMAKE_BUILD_PARALLEL_LEVEL=16

 # Setup cleanup
 remove_docker_container() {
@@ -37,10 +37,8 @@ function cpu_tests() {
    pytest -x -v -s tests/kernels/test_onednn.py
    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/moe/test_cpu_int4_moe.py
-    pytest -x -v -s tests/kernels/mamba/test_cpu_short_conv.py"
+    pytest -x -v -s tests/kernels/moe/test_moe.py -k test_cpu_fused_moe_basic
+    pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py"

  # skip tests requiring model downloads if HF_TOKEN is not set
  # due to rate-limits
@@ -64,6 +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
@@ -7,49 +7,10 @@ set -euox pipefail
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
-AGENT_SLOT=${AGENT_SLOT:-}
-IMAGE_NAME="cpu-test-${NUMA_NODE}${AGENT_SLOT:+-${AGENT_SLOT}}"
+IMAGE_NAME="cpu-test-$NUMA_NODE"
 TIMEOUT_VAL=$1
 TEST_COMMAND=$2

-# Disk hygiene knobs. Reclaim space only once the Docker root filesystem crosses
-# DISK_USAGE_THRESHOLD percent, and cap the shared BuildKit cache at
-# BUILDKIT_CACHE_MAX so subsequent builds keep reusing the hottest layers.
-DISK_USAGE_THRESHOLD=${DISK_USAGE_THRESHOLD:-70}
-BUILDKIT_CACHE_MAX=${BUILDKIT_CACHE_MAX:-80GB}
-
-# Reclaim disk only when the host is under pressure. We trim (not purge) the
-# shared BuildKit cache so cross-job/cross-agent reuse stays intact, and only
-# touch dangling images; other agents' uniquely tagged images are left alone.
-prune_if_disk_pressure() {
-    local docker_root disk_usage
-    docker_root=$(docker info -f '{{.DockerRootDir}}' 2>/dev/null || true)
-    if [ -z "$docker_root" ]; then
-        return 0
-    fi
-    disk_usage=$(df "$docker_root" 2>/dev/null | tail -1 | awk '{print $5}' | tr -d '%')
-    if [ "${disk_usage:-0}" -gt "$DISK_USAGE_THRESHOLD" ]; then
-        echo "--- :broom: Disk usage ${disk_usage}% exceeds ${DISK_USAGE_THRESHOLD}%, reclaiming space"
-        docker image prune -f || true
-        docker builder prune -f --keep-storage="$BUILDKIT_CACHE_MAX" || true
-    else
-        echo "Disk usage ${disk_usage:-unknown}% within ${DISK_USAGE_THRESHOLD}% threshold; skipping prune"
-    fi
-}
-
-# Always drop this agent's image once the job ends (the default builder never
-# uses it as a cache source, so removing it costs no rebuild speed), then
-# reclaim space if needed. Guard every docker call with `|| true` so the trap
-# never overrides the test's exit code.
-cleanup() {
-    docker image rm -f "$IMAGE_NAME" || true
-    prune_if_disk_pressure
-}
-trap cleanup EXIT
-
-# Free space up front so a nearly-full host doesn't fail the build.
-prune_if_disk_pressure
-
 # building the docker image
 echo "--- :docker: Building Docker image"
 docker build --progress plain --tag "$IMAGE_NAME" --target vllm-test -f docker/Dockerfile.cpu .
@@ -1,52 +0,0 @@
-#!/bin/bash
-
-set -euo pipefail
-
-test_suite="${1:-}"
-
-if [[ -z "${test_suite}" ]]; then
-  echo "Usage: $0 <example|v1|server>" >&2
-  exit 1
-fi
-
-case "${test_suite}" in
-  example)
-    pip install tblib==3.1.0
-
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
-    python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
-    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
-    ;;
-  v1)
-    cd tests
-
-    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/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
-    ;;
-  server)
-    pip install av
-    cd tests
-
-    pytest -v -s entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py
-    pytest -v -s benchmarks/test_serve_cli.py
-    ;;
-  *)
-    echo "Unknown Intel test suite: ${test_suite}" >&2
-    exit 1
-    ;;
-esac
@@ -243,10 +243,8 @@ container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head

 # ---- Command source selection ----
 commands=""
-commands_source=""
 if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
  commands="${VLLM_TEST_COMMANDS}"
-  commands_source="env"
  echo "Commands sourced from VLLM_TEST_COMMANDS (quoting preserved)"
 elif [[ $# -gt 0 ]]; then
  all_yaml=true
@@ -305,12 +303,8 @@ if [[ -z "$commands" ]]; then
 fi

 echo "Raw commands: $commands"
-if [[ "$commands_source" != "env" ]]; then
-  commands=$(re_quote_pytest_markers "$commands")
-  echo "After re-quoting: $commands"
-else
-  echo "Skipping re-quoting for VLLM_TEST_COMMANDS input"
-fi
+commands=$(re_quote_pytest_markers "$commands")
+echo "After re-quoting: $commands"
 commands=$(apply_intel_test_overrides "$commands")
 echo "Final commands: $commands"

@@ -360,7 +354,7 @@ 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 "${HOME}/.cache/huggingface:/root/.cache/huggingface" \
    --entrypoint='' \
    -e HF_TOKEN \
    -e ZE_AFFINITY_MASK \
@@ -369,7 +363,7 @@ export HF_TOKEN ZE_AFFINITY_MASK
    -e CMDS \
    --name "${container_name}" \
    "${IMAGE}" \
-    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"' \
+    bash -c 'set -e; 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

@@ -4,11 +4,6 @@

 set -euo pipefail

-if python3 -c "import torch; raise SystemExit(0 if torch.version.hip is not None else 1)"; then
-    uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-    exit 0
-fi
-
 REQUIREMENTS_FILE="${KV_CONNECTORS_REQUIREMENTS:-/vllm-workspace/requirements/kv_connectors.txt}"

 uv pip install --system -r "${REQUIREMENTS_FILE}"
@@ -130,22 +130,6 @@ docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${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: ${CI_BASE_PUSH_STABLE_TAG}"
-    fi
-
-    bash .buildkite/scripts/ci-bake-rocm.sh ci-base-rocm-ci-with-deps
-}
-
-main "$@"
@@ -1,57 +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_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 CI_BASE_IMAGE
-    export IMAGE_TAG_LATEST
-
-    BASE_IMAGE="$(metadata_get rocm-base-image)"
-    CI_BASE_IMAGE="$(metadata_get rocm-ci-base-image)"
-    IMAGE_TAG_LATEST="$(metadata_get rocm-ci-image-descriptive)"
-
-    echo "Using refreshed ROCm base image for test image: ${BASE_IMAGE}"
-    echo "Using refreshed ROCm ci_base image for test image: ${CI_BASE_IMAGE}"
-    if [[ -n "${IMAGE_TAG_LATEST}" ]]; then
-        echo "Also tagging full ROCm CI image as: ${IMAGE_TAG_LATEST}"
-    fi
-
-    return 0
-}
-
-main() {
-    local base_refreshed=0
-
-    if use_refreshed_base_if_present; then
-        base_refreshed=1
-    fi
-
-    if [[ "${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
-            # 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]}"
+            docker exec "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="$(git rev-parse --show-toplevel 2>/dev/null)"; then
-  :
-else
-  SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd -P)"
-  ROOT_DIR="$(cd -- "${SCRIPT_DIR}/../.." && pwd -P)"
-fi
+ROOT_DIR="$(git rev-parse --show-toplevel)"
 cd "$ROOT_DIR"

 export CARGO_TERM_COLOR="${CARGO_TERM_COLOR:-always}"
@@ -95,16 +90,6 @@ install_cargo_sort() {
  cargo binstall --no-confirm cargo-sort
 }

-install_cargo_deny() {
-  if command -v cargo-deny >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-deny"
-  install_cargo_binstall
-  cargo binstall --no-confirm cargo-deny
-}
-
 install_cargo_nextest() {
  if command -v cargo-nextest >/dev/null 2>&1; then
    return
@@ -157,7 +142,6 @@ PY

 run_style_clippy() {
  install_cargo_sort
-  install_cargo_deny

  log_section "Checking Rust formatting"
  cargo fmt --manifest-path rust/Cargo.toml --all -- --check
@@ -165,13 +149,6 @@ run_style_clippy() {
  log_section "Checking Cargo.toml ordering"
  cargo sort --workspace --check rust

-  log_section "Checking Rust dependency bans"
-  cargo deny \
-    --manifest-path rust/Cargo.toml \
-    --config rust/deny.toml \
-    check \
-    bans
-
  log_section "Running clippy"
  cargo clippy \
    --manifest-path rust/Cargo.toml \
@@ -33,14 +33,6 @@ if [[ -n "${ATTENTION_BACKEND:-}" ]]; then
  EXTRA_ARGS+=(--attention-backend "${ATTENTION_BACKEND}")
 fi

-# ROCm: run eager to avoid intermittent HIP-graph decode corruption.
-# See https://github.com/ROCm/clr/issues/279
-# TODO(aarushjain29): Revert after TheRock 7.14
-if command -v rocm-smi &> /dev/null || command -v amd-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
-  echo "ROCm platform detected: adding --enforce-eager to avoid HIP-graph decode corruption"
-  EXTRA_ARGS+=(--enforce-eager)
-fi
-
 cleanup() {
  if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
    kill "${SERVER_PID}" 2>/dev/null || true
@@ -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)
-    --pipeline       Buildkite pipeline slug (default: ${PIPELINE})
-    --commit         Override commit SHA (default: current HEAD)
-    --branch         Override branch name (default: current branch)
-    --message        Custom build message (default: auto-generated)
-    --torch-nightly  Also build and run the full suite against torch nightly
-                     (sets TORCH_NIGHTLY=1)
-    --help           Show this help message
+    --execute       Actually trigger the build (default: dry-run)
+    --pipeline      Buildkite pipeline slug (default: ${PIPELINE})
+    --commit        Override commit SHA (default: current HEAD)
+    --branch        Override branch name (default: current branch)
+    --message       Custom build message (default: auto-generated)
+    --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=$(( ${#BUILD_ENV[@]} - 1 ))
-    for i in "${!BUILD_ENV[@]}"; do
-        if [[ $i -eq $last_idx ]]; then
-            echo "    --env '$(escape_for_shell "${BUILD_ENV[$i]}")'"
-        else
-            echo "    --env '$(escape_for_shell "${BUILD_ENV[$i]}")' \\"
-        fi
-    done
+    echo "    --env 'RUN_ALL=1' \\"
+    echo "    --env 'NIGHTLY=1'"
    echo ""
    echo "=========================================="
    echo -e "${YELLOW}To actually trigger this build, run:${NC}"
@@ -1,13 +0,0 @@
-#!/bin/bash
-
-set -euo pipefail
-
-REGISTRY="public.ecr.aws/q9t5s3a7"
-REPO="vllm-release-repo"
-ARCH_TAG="${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
@@ -16,7 +16,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 95
+      timeout_in_minutes: 70
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -30,7 +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
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Correctness
  key: basic-correctness
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -16,9 +16,3 @@ steps:
  - pytest -v -s basic_correctness/test_mem.py
  - pytest -v -s basic_correctness/test_basic_correctness.py
  - pytest -v -s basic_correctness/test_cpu_offload.py
-  mirror:
-    amd:
-      device: mi325_1
-      timeout_in_minutes: 70
-      depends_on:
-      - image-build-amd
@@ -4,18 +4,13 @@ depends_on:
 steps:
 - label: Benchmarks CLI Test
  key: benchmarks-cli-test
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - vllm/
  - tests/benchmarks/
  commands:
  - pytest -v -s benchmarks/
-  mirror:
-    amd:
-      device: mi300_1
-      depends_on:
-      - image-build-amd

 - label: Attention Benchmarks Smoke Test (B200)
  key: attention-benchmarks-smoke-test-b200
@@ -23,9 +18,9 @@ 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/
  commands:
-  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k"
+  - python3 benchmarks/attention_benchmarks/benchmark.py --backends flash flashinfer --batch-specs "8q1s1k" --repeats 1 --warmup-iters 1
@@ -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
@@ -2,9 +2,9 @@ group: CUDA
 depends_on: 
  - image-build
 steps:
- label: Platform Tests
-  key: platform-tests
-  timeout_in_minutes: 20
+- label: Platform Tests (CUDA)
+  key: platform-tests-cuda
+  timeout_in_minutes: 15
  device: h200_18gb
  source_file_dependencies:
  - vllm/envs.py
@@ -19,7 +19,7 @@ steps:

 - label: Cudagraph
  key: cudagraph
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  source_file_dependencies:
  - tests/v1/cudagraph
  - vllm/v1/cudagraph_dispatcher.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:
@@ -13,23 +13,9 @@ steps:
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-  mirror:
-    amd:
-      device: mi300_4
-      timeout_in_minutes: 85
-      depends_on:
-        - image-build-amd
-      source_file_dependencies:
-        - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
-        - tests/v1/kv_connector/nixl_integration/
-        - vllm/platforms/rocm.py
-      commands:
-        - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-        - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
 - 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:
@@ -39,19 +25,6 @@ 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
@@ -63,23 +36,10 @@ steps:
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-  mirror:
-    amd:
-      device: mi300_4
-      timeout_in_minutes: 60
-      depends_on:
-        - image-build-amd
-      source_file_dependencies:
-        - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
-        - tests/v1/kv_connector/nixl_integration/
-        - vllm/platforms/rocm.py
-      commands:
-        - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-        - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh

 - 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:
@@ -88,23 +48,10 @@ steps:
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-  mirror:
-    amd:
-      device: mi300_4
-      timeout_in_minutes: 85
-      depends_on:
-        - image-build-amd
-      source_file_dependencies:
-        - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
-        - tests/v1/kv_connector/nixl_integration/
-        - vllm/platforms/rocm.py
-      commands:
-        - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-        - CROSS_LAYERS_BLOCKS=True ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh

 - 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:
@@ -113,19 +60,6 @@ steps:
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-  mirror:
-    amd:
-      device: mi300_4
-      timeout_in_minutes: 80
-      depends_on:
-        - image-build-amd
-      source_file_dependencies:
-        - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
-        - tests/v1/kv_connector/nixl_integration/
-        - vllm/platforms/rocm.py
-      commands:
-        - 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: Hybrid SSM NixlConnector PD prefix cache test (2 GPUs)
  key: hybrid-ssm-nixlconnector-pd-prefix-cache-2-gpus
@@ -143,7 +77,7 @@ 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:
@@ -158,7 +92,7 @@ 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
@@ -169,24 +103,10 @@ steps:
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
-  mirror:
-    amd:
-      device: mi300_2
-      timeout_in_minutes: 70
-      depends_on:
-        - image-build-amd
-      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/rocm.py
-      commands:
-        - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-        - 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:
@@ -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:
@@ -37,25 +37,10 @@ steps:
  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
  - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
-  mirror:
-    amd:
-      device: mi300_2
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - vllm/distributed/
-      - vllm/engine/
-      - vllm/executor/
-      - vllm/worker/worker_base.py
-      - vllm/v1/engine/
-      - vllm/v1/worker/
-      - tests/v1/distributed
-      - tests/entrypoints/openai/test_multi_api_servers.py
-      - vllm/platforms/rocm.py

 - 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:
@@ -78,7 +63,7 @@ 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:
@@ -133,7 +118,7 @@ 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:
@@ -154,7 +139,7 @@ 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:
@@ -174,9 +159,9 @@ steps:
  # test multi-node TP with multiproc executor (simulated on single node)
  - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node

- label: Distributed Tests (8xH100)
-  key: distributed-tests-8xh100
-  timeout_in_minutes: 20
+- label: Distributed Tests (8 GPUs)(H100)
+  key: distributed-tests-8-gpus-h100
+  timeout_in_minutes: 10
  device: h100
  num_devices: 8
  working_dir: "/vllm-workspace/tests"
@@ -195,8 +180,8 @@ steps:
  # test with torchrun tp=2 and dp=4 with ep
  - torchrun --nproc-per-node=8 ../examples/features/torchrun/torchrun_dp_example_offline.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep

- label: Distributed Tests (4xA100)
-  key: distributed-tests-4xa100
+- label: Distributed Tests (4 GPUs)(A100)
+  key: distributed-tests-4-gpus-a100
  device: a100
  optional: true
  num_devices: 4
@@ -210,9 +195,9 @@ steps:
  - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
  - pytest -v -s -x lora/test_mixtral.py

- label: Distributed Tests (2xH100-2xMI300)
-  key: distributed-tests-2xh100-2xmi300
-  timeout_in_minutes: 30
+- label: Distributed Tests (2 GPUs)(H100)
+  key: distributed-tests-2-gpus-h100
+  timeout_in_minutes: 15
  device: h100
  optional: true
  working_dir: "/vllm-workspace/"
@@ -225,15 +210,15 @@ steps:
    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 pytest -v -s tests/distributed/test_weight_transfer.py
    - pytest -v -s tests/distributed/test_packed_tensor.py

- label: Distributed Tests (2xB200)
-  key: distributed-tests-2xb200
+- label: Distributed Tests (2 GPUs)(B200)
+  key: distributed-tests-2-gpus-b200
  device: b200-k8s
  optional: true
  working_dir: "/vllm-workspace/"
  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

@@ -259,7 +244,7 @@ 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:
@@ -274,7 +259,7 @@ 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
@@ -2,9 +2,9 @@ group: E2E Integration
 depends_on: 
  - image-build
 steps:
- label: DeepSeek V2-Lite Sync EPLB Accuracy (4xH100)
-  key: deepseek-v2-lite-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 25
+- label: DeepSeek V2-Lite Sync EPLB Accuracy
+  key: deepseek-v2-lite-sync-eplb-accuracy
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 4
@@ -12,9 +12,9 @@ steps:
  commands:
  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010

- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (4xH100)
-  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 25
+- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy
+  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 4
@@ -22,9 +22,9 @@ steps:
  commands:
  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020

- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (2xB200)
-  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-2xb200
-  timeout_in_minutes: 20
+- label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (B200)
+  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-b200
+  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/
@@ -29,13 +29,13 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 50
+      timeout_in_minutes: 60
      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/
@@ -45,13 +45,13 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 55
+      timeout_in_minutes: 40
      depends_on:
      - image-build-amd

 - label: e2e Scheduling (1 GPU)
  key: e2e-scheduling-1-gpu
-  timeout_in_minutes: 35
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/
@@ -60,34 +60,24 @@ steps:
    - pytest -v -s v1/e2e/general/test_async_scheduling.py
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      device: mi250_1
+      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/
  commands:
    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
-  mirror:
-    amd:
-      device: mi325_1
-      timeout_in_minutes: 60
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - vllm/v1/
-      - tests/v1/e2e/general/
-      - vllm/platforms/rocm.py

 - 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:
@@ -112,15 +102,10 @@ steps:
  commands:
    # Only run tests that need exactly 2 GPUs
    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
-  mirror:
-    amd:
-      device: mi300_2
-      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:
@@ -148,7 +133,7 @@ steps:

 - label: V1 e2e (4xH100)
  key: v1-e2e-4xh100
-  timeout_in_minutes: 35
+  timeout_in_minutes: 60
  device: h100
  num_devices: 4
  optional: true
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Entrypoints Unit Tests
  key: entrypoints-unit-tests
-  timeout_in_minutes: 25
+  timeout_in_minutes: 10
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/entrypoints
@@ -16,7 +16,7 @@ steps:

 - label: Entrypoints Integration (LLM)
  key: entrypoints-integration-llm
-  timeout_in_minutes: 60
+  timeout_in_minutes: 40
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -29,25 +29,21 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
      depends_on:
      - image-build-amd

 - label: Entrypoints Integration (API Server)
  key: entrypoints-integration-api-server
  device: h200_35gb
-  timeout_in_minutes: 50
+  timeout_in_minutes: 130
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
  - tests/entrypoints/serve
-  - tests/entrypoints/scale_out
  commands:
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
  - pytest -v -s entrypoints/serve --ignore=entrypoints/serve/dev/rpc
  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc
-  - pytest -v -s entrypoints/scale_out
  mirror:
    amd:
      device: mi325_1
@@ -56,7 +52,7 @@ steps:

 - label: Entrypoints Integration (API Server OpenAI - Part 1)
  key: entrypoints-integration-api-server-openai-part-1
-  timeout_in_minutes: 45
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -68,13 +64,13 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 65
+      timeout_in_minutes: 80
      depends_on:
      - image-build-amd

 - label: Entrypoints Integration (API Server OpenAI - Part 2)
  key: entrypoints-integration-api-server-openai-part-2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -109,7 +105,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 65
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd

@@ -126,7 +122,7 @@ 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/
@@ -138,7 +134,7 @@ 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/
@@ -160,7 +156,7 @@ steps:

 - label: OpenAI API Correctness
  key: openai-api-correctness
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - csrc/
@@ -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:
@@ -14,20 +14,10 @@ steps:
  commands:
  - pytest -v -s distributed/test_eplb_algo.py
  - pytest -v -s distributed/test_eplb_utils.py
-  mirror:
-    amd:
-      device: mi300_1
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - vllm/distributed/eplb
-      - tests/distributed/test_eplb_algo.py
-      - tests/distributed/test_eplb_utils.py
-      - vllm/platforms/rocm.py

 - 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:
@@ -39,7 +29,7 @@ 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
@@ -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:
@@ -16,7 +16,7 @@ steps:

 - label: Kernels Core Operation Test
  key: kernels-core-operation-test
-  timeout_in_minutes: 120
+  timeout_in_minutes: 75
  source_file_dependencies:
  - csrc/
  - tests/kernels/core
@@ -27,7 +27,7 @@ steps:

 - 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:
@@ -41,20 +41,18 @@ 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
  - vllm/models/deepseek_v4/common/ops/
  - tests/kernels/test_fused_deepseek_v4_qnorm_rope_kv_insert.py
-  - tests/kernels/test_top_k_per_row.py # it runs on Blackwell too - some kernels have arch-specific optimizations
  commands:
    - pytest -v -s kernels/test_fused_deepseek_v4_*.py
-    - pytest -v -s kernels/test_top_k_per_row.py

 - 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
@@ -65,7 +63,7 @@ steps:

 - 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
@@ -76,20 +74,6 @@ steps:
  commands:
    - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
  parallelism: 2
-  mirror:
-    amd:
-      device: mi325_1
-      timeout_in_minutes: 90
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - csrc/attention/
-      - vllm/v1/attention
-      - vllm/model_executor/layers/attention
-      - tests/kernels/attention
-      - vllm/_aiter_ops.py
-      - vllm/envs.py
-      - vllm/platforms/rocm.py

 - label: Kernels Attention DiffKV Test (H100)
  key: kernels-attention-diffkv-test-h100
@@ -106,7 +90,7 @@ 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
@@ -120,7 +104,6 @@ steps:
      source_file_dependencies:
      - csrc/quantization/
      - vllm/model_executor/layers/quantization
-      - vllm/config/
      - tests/kernels/quantization
      - tests/kernels/quantization/test_rocm_skinny_gemms.py
      - vllm/_aiter_ops.py
@@ -131,7 +114,7 @@ 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/
@@ -144,26 +127,10 @@ steps:
    - pytest -v -s kernels/moe --ignore=kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
    - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
  parallelism: 5
-  mirror:
-    amd:
-      device: mi325_1
-      timeout_in_minutes: 65
-      source_file_dependencies:
-      - csrc/quantization/cutlass_w8a8/moe/
-      - csrc/moe/
-      - tests/kernels/moe
-      - vllm/model_executor/layers/fused_moe/
-      - vllm/distributed/device_communicators/
-      - vllm/envs.py
-      - vllm/config
-      - vllm/_aiter_ops.py
-      - vllm/platforms/rocm.py
-      depends_on:
-      - image-build-amd

 - label: Kernels Mamba Test
  key: kernels-mamba-test
-  timeout_in_minutes: 40
+  timeout_in_minutes: 45
  source_file_dependencies:
  - csrc/mamba/
  - tests/kernels/mamba
@@ -172,7 +139,7 @@ steps:
    - pytest -v -s kernels/mamba

 - label: Kernels KDA Test
-  timeout_in_minutes: 25
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - vllm/model_executor/layers/fla/ops/kda.py
@@ -184,7 +151,7 @@ steps:

 - 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:
@@ -211,7 +178,7 @@ steps:

 - label: Kernels (B200)
  key: kernels-b200
-  timeout_in_minutes: 80
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  # optional: true
@@ -264,7 +231,7 @@ steps:

 - 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
@@ -274,9 +241,9 @@ steps:
    - pytest -v -s kernels/helion/

 
- label: Kernels FP8 MoE Test (1xH100)
-  key: kernels-fp8-moe-test-1xh100
-  timeout_in_minutes: 40
+- label: Kernels FP8 MoE Test (1 H100)
+  key: kernels-fp8-moe-test-1-h100
+  timeout_in_minutes: 90
  device: h100
  num_devices: 1
  optional: true
@@ -291,9 +258,9 @@ steps:
    - pytest -v -s kernels/moe/test_triton_moe_no_act_mul.py
    - pytest -v -s kernels/moe/test_triton_moe_ptpc_fp8.py

- label: Kernels FP8 MoE Test (2xH100)
-  key: kernels-fp8-moe-test-2xh100
-  timeout_in_minutes: 45
+- label: Kernels FP8 MoE Test (2 H100s)
+  key: kernels-fp8-moe-test-2-h100s
+  timeout_in_minutes: 90
  device: h100
  num_devices: 2
  optional: true
@@ -303,7 +270,7 @@ 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
@@ -316,7 +283,7 @@ 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
@@ -28,8 +28,7 @@ steps:
      - vllm/_aiter_ops.py
      - vllm/platforms/rocm.py

-# - label: LM Eval Large Models (4xA100)
-#   key: lm-eval-large-models-4xa100
+# - label: LM Eval Large Models (4 GPUs)(A100)
 #   device: a100
 #   optional: true
 #   num_devices: 4
@@ -41,8 +40,8 @@ steps:
 #   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4

- label: LM Eval Large Models (4xH100)
-  key: lm-eval-large-models-4xh100
+- label: LM Eval Large Models (4 GPUs)(H100)
+  key: lm-eval-large-models-4-gpus-h100
  device: h100
  optional: true
  num_devices: 4
@@ -54,9 +53,9 @@ 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)
-  key: lm-eval-small-models-1xb200
-  timeout_in_minutes: 50
+- label: LM Eval Small Models (B200)
+  key: lm-eval-small-models-b200
+  timeout_in_minutes: 120
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -65,23 +64,10 @@ 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)
-  key: lm-eval-small-models-distributed-2xb200
+- label: LM Eval Large Models (B200, EP)
+  key: lm-eval-large-models-b200-ep
  timeout_in_minutes: 120
  device: b200-k8s
-  num_devices: 2
-  optional: true
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  autorun_on_main: true
-  commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small-tp.txt
-
- label: LM Eval Large Models EP (2xB200)
-  key: lm-eval-large-models-ep-2xb200
-  timeout_in_minutes: 60
-  device: b200-k8s
  optional: true
  num_devices: 2
  source_file_dependencies:
@@ -90,9 +76,9 @@ steps:
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell-ep.txt

- label: LM Eval Qwen3.5 Models (2xB200)
-  key: lm-eval-qwen3-5-models-2xb200
-  timeout_in_minutes: 45
+- label: LM Eval Qwen3.5 Models (B200)
+  key: lm-eval-qwen3-5-models-b200
+  timeout_in_minutes: 120
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -107,24 +93,14 @@ steps:
  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
+- label: LM Eval Large Models (H200)
+  key: lm-eval-large-models-h200
+  timeout_in_minutes: 60
  device: h200
  optional: true
  num_devices: 8
  commands:
    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
-  mirror:
-    amd:
-      device: mi300_8
-      timeout_in_minutes: 60
-      depends_on:
-      - image-build-amd
-      commands:
-      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-      - export PYTORCH_ROCM_ARCH=gfx942 # Limit Quark compilation to save time
-      - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-mi3xx.txt

 - label: MoE Refactor Integration Test (H100 - TEMPORARY)
  key: moe-refactor-integration-test-h100-temporary
@@ -150,101 +126,10 @@ 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)
-  key: lm-eval-humming-f16-a100
-  timeout_in_minutes: 75
-  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.txt
-
- label: LM Eval Humming Act int8 (A100 - TEMPORARY)
-  key: lm-eval-humming-act-a100
-  timeout_in_minutes: 45
-  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
-  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.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)
-  key: lm-eval-humming-f16-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.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/
@@ -254,9 +139,9 @@ steps:
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/models-turboquant.txt

- label: GPQA Eval (GPT-OSS) (2xH100)
-  key: gpqa-eval-gpt-oss-2xh100
-  timeout_in_minutes: 35
+- label: GPQA Eval (GPT-OSS) (H100)
+  key: gpqa-eval-gpt-oss-h100
+  timeout_in_minutes: 120
  device: h100
  optional: true
  num_devices: 2
@@ -268,9 +153,9 @@ 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-h100.txt

- label: GPQA Eval (GPT-OSS) (2xB200)
-  key: gpqa-eval-gpt-oss-2xb200
-  timeout_in_minutes: 30
+- label: GPQA Eval (GPT-OSS) (B200)
+  key: gpqa-eval-gpt-oss-b200
+  timeout_in_minutes: 120
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -282,66 +167,9 @@ 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-b200.txt

- label: GPQA Eval (GPT-OSS) (DGX Spark)
-  key: gpqa-eval-gpt-oss-spark
-  timeout_in_minutes: 35
-  device: dgx-spark
-  optional: true
-  num_devices: 1
-  depends_on:
-  - arm64-image-build
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  - tests/evals/gpt_oss/
-  commands:
-    - 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: 30
-  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"
-
- 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:
@@ -5,29 +5,18 @@ steps:
 - label: LoRA %N
  device: h200_35gb
  key: lora
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  source_file_dependencies:
  - vllm/lora
  - tests/lora
  commands:
    - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_qwen3_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemodel_lora.py 
  parallelism: 4
-  mirror:
-    amd:
-      device: mi325_1
-      working_dir: "/vllm-workspace/tests"
-      timeout_in_minutes: 65
-      source_file_dependencies:
-      - vllm/lora
-      - tests/lora
-      - vllm/platforms/rocm.py
-      depends_on:
-      - image-build-amd


 - label: LoRA TP (Distributed)
  key: lora-tp-distributed
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
  num_devices: 4
  source_file_dependencies:
  - vllm/lora
@@ -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/
@@ -21,16 +21,10 @@ steps:
    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
    # TODO: create another `optional` test group for slow tests
    - pytest -v -s -m 'not slow_test' v1/spec_decode
-  mirror:
-    amd:
-      device: mi300_1
-      timeout_in_minutes: 75
-      depends_on:
-      - image-build-amd

 - label: V1 Sample + Logits
  key: v1-sample-logits
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
    - vllm/config/
@@ -64,7 +58,7 @@ steps:

 - label: V1 Core + KV + Metrics
  key: v1-core-kv-metrics
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
  source_file_dependencies:
    - vllm/config/
    - vllm/distributed/
@@ -103,14 +97,8 @@ steps:
    - pytest -v -s -m 'not cpu_test' v1/kv_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:
-      device: mi325_1
-      timeout_in_minutes: 75
-      depends_on:
-      - image-build-amd

 - label: V1 Others (CPU)
  key: v1-others-cpu
@@ -172,7 +160,7 @@ steps:

 - label: Regression
  key: regression
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - vllm/config/
@@ -188,14 +176,14 @@ 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
@@ -224,20 +212,10 @@ steps:
    - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle --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 2048
    # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
    - 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:
-      device: mi325_1
-      source_file_dependencies:
-      - vllm/entrypoints
-      - vllm/multimodal
-      - examples/
-      - vllm/platforms/rocm.py
-      depends_on:
-      - image-build-amd

 - 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/
@@ -260,12 +238,6 @@ steps:
      'opentelemetry-exporter-otlp>=1.26.0' \
      'opentelemetry-semantic-conventions-ai>=0.4.1'"
  - pytest -v -s v1/tracing
-  mirror:
-    amd:
-      device: mi325_2
-      depends_on:
-      - image-build-amd
-      optional: true

 - label: Python-only Installation
  key: python-only-installation
@@ -278,21 +250,11 @@ steps:
  - setup.py
  commands:
  - bash standalone_tests/python_only_compile.sh
-  mirror:
-    amd:
-      device: mi325_1
-      timeout_in_minutes: 45
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - tests/standalone_tests/python_only_compile.sh
-      - setup.py
-      - vllm/platforms/rocm.py

 - 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/
@@ -319,7 +281,7 @@ 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/
@@ -351,7 +313,6 @@ 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
@@ -369,10 +330,9 @@ 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
+  - pytest -v -s reasoning --ignore=reasoning/test_seedoss_reasoning_parser.py --ignore=reasoning/test_glm4_moe_reasoning_parser.py
  - pytest -v -s tool_parsers
  - pytest -v -s tokenizers_
  - pytest -v -s parser
@@ -381,7 +341,7 @@ steps:

 - label: Batch Invariance (A100)
  key: batch-invariance-a100
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  device: a100
  source_file_dependencies:
    - vllm/v1/attention
@@ -395,7 +355,7 @@ steps:

 - label: Batch Invariance (H100)
  key: batch-invariance-h100
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  device: h100
  source_file_dependencies:
    - vllm/v1/attention
@@ -411,7 +371,7 @@ steps:

 - label: Batch Invariance (B200)
  key: batch-invariance-b200
-  timeout_in_minutes: 35
+  timeout_in_minutes: 30
  device: b200-k8s
  source_file_dependencies:
    - vllm/v1/attention
@@ -430,7 +390,7 @@ steps:
 - 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
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Model Executor
  key: model-executor
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
  source_file_dependencies:
  - vllm/engine/arg_utils.py
  - vllm/config/model.py
@@ -23,16 +23,3 @@ steps:
    # calls that the signal method cannot interrupt.
    - pytest -v -s model_executor -m '(not slow_test)' --timeout=900 --timeout-method=thread
    - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py --timeout=900 --timeout-method=thread
-  mirror:
-    amd:
-      device: mi300_1
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - vllm/engine/arg_utils.py
-      - vllm/config/model.py
-      - vllm/model_executor
-      - tests/model_executor
-      - 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/
@@ -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
@@ -28,34 +31,30 @@ steps:
    # 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: 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
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
+    - pytest -v -s models/test_terratorch.py models/test_transformers.py models/test_registry.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/transformers/fusers/
  device: cpu-small
  commands:
-    - pytest -v -s models/test_utils.py models/test_vision.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/
@@ -14,14 +14,11 @@ steps:
    - pip freeze | grep -E 'torch'
    - pytest -v -s models/language -m 'core_model and (not slow_test)'
  mirror:
-    amd:
-      device: mi300_1
-      depends_on:
-      - image-build-amd
+    torch_nightly: {}

 - label: Language Models Tests (Extra Standard) %N
  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
@@ -34,25 +31,11 @@ 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:
-    amd:
-      device: mi300_1
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - vllm/model_executor/models/
-      - vllm/model_executor/model_loader/
-      - vllm/model_executor/layers/
-      - vllm/v1/attention/backends/
-      - vllm/v1/attention/selector.py
-      - tests/models/language/pooling/test_embedding.py
-      - tests/models/language/generation/test_common.py
-      - tests/models/language/pooling/test_classification.py
-      - vllm/_aiter_ops.py
-      - vllm/platforms/rocm.py
+    torch_nightly: {}

 - label: Language Models Tests (Hybrid) %N
  key: language-models-tests-hybrid
-  timeout_in_minutes: 65
+  timeout_in_minutes: 75
  source_file_dependencies:
  - vllm/
  - tests/models/language/generation
@@ -65,9 +48,10 @@ steps:
    - 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:
      device: mi325_1
-      timeout_in_minutes: 70
+      timeout_in_minutes: 90
      depends_on:
      - image-build-amd
      commands:
@@ -78,7 +62,7 @@ steps:
 - 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/
@@ -92,7 +76,7 @@ steps:

 - 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:
@@ -104,7 +88,7 @@ steps:
 - label: Language Models Test (Extended Pooling)  # 36min
  device: h200_35gb
  key: language-models-test-extended-pooling
-  timeout_in_minutes: 70
+  timeout_in_minutes: 50
  optional: true
  source_file_dependencies:
  - vllm/
@@ -120,7 +104,7 @@ steps:

 - label: Language Models Test (MTEB)
  key: language-models-test-mteb
-  timeout_in_minutes: 45
+  timeout_in_minutes: 110
  device: h200_18gb
  optional: true
  source_file_dependencies:
@@ -10,6 +10,7 @@ steps:
  - 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:
@@ -20,15 +21,16 @@ steps:

 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
  key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 50
+  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
+    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
  mirror:
    amd:
      device: mi325_1
@@ -38,11 +40,12 @@ steps:
 - 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:
@@ -54,49 +57,46 @@ steps:
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
  device: h200_35gb
  key: multi-modal-models-standard-4-other-whisper
-  timeout_in_minutes: 50
+  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
-    - pytest -v -s models/multimodal/generation/test_vit_cudagraph.py -m core_model
+    - 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/processing
    - 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:
-      device: mi325_1
-      depends_on:
-      - image-build-amd

 - 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:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
    - 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: 65
+  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/
@@ -104,17 +104,6 @@ steps:
  - vllm/v1/core/
  commands:
  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
-  mirror:
-    amd:
-      device: mi300_1
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - vllm/multimodal/
-      - vllm/inputs/
-      - vllm/v1/core/
-      - vllm/platforms/rocm.py
-      - vllm/model_executor/model_loader/

 - label: Multi-Modal Models (Extended Generation 1)
  key: multi-modal-models-extended-generation-1
@@ -124,6 +113,7 @@ 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:
@@ -140,6 +130,7 @@ 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)
@@ -150,6 +141,7 @@ 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)
@@ -164,7 +156,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 75
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -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:
@@ -27,21 +27,12 @@ steps:
  - pip install -e ./plugins/bge_m3_sparse_plugin
  - pytest -v -s plugins_tests/test_bge_m3_sparse_io_processor_plugins.py
  - pip uninstall bge_m3_sparse_plugin -y
-  # test colbert_query io_processor plugin
-  - pip install -e ./plugins/colbert_query_plugin
-  - pytest -v -s plugins_tests/test_colbert_query_io_processor_plugins.py
-  - pip uninstall colbert_query_plugin -y
  # end io_processor plugins test
  # begin stat_logger plugins test
  - pip install -e ./plugins/vllm_add_dummy_stat_logger
  - 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: 90
+  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
@@ -107,16 +107,10 @@ steps:
    - tests/compile/passes
  commands:
  - pytest -s -v compile/passes --ignore compile/passes/distributed
-  mirror:
-    amd:
-      device: mi300_1
-      timeout_in_minutes: 65
-      depends_on:
-      - image-build-amd

 - label: PyTorch Fullgraph Smoke Test
  key: pytorch-fullgraph-smoke-test
-  timeout_in_minutes: 60
+  timeout_in_minutes: 35
  source_file_dependencies:
  - vllm/__init__.py
  - vllm/_aiter_ops.py
@@ -152,7 +146,7 @@ steps:

 - label: PyTorch Fullgraph
  key: pytorch-fullgraph
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/__init__.py
@@ -195,11 +189,3 @@ steps:
  - requirements/test/nightly-torch.txt
  commands:
  - bash standalone_tests/pytorch_nightly_dependency.sh
-  mirror:
-    amd:
-      device: mi300_1
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - requirements/test/nightly-torch.txt
-      - vllm/platforms/rocm.py
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Quantization
  key: quantization
-  timeout_in_minutes: 60
+  timeout_in_minutes: 90
  source_file_dependencies:
  - csrc/
  - vllm/model_executor/layers/quantization
@@ -23,7 +23,7 @@ steps:

 - label: Quantized Fusions
  key: quantized-fusions
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  source_file_dependencies:
  - tests/fusion
  - vllm/model_executor/layers/fusion
@@ -35,7 +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:
@@ -53,7 +53,7 @@ steps:

 - label: Quantized Models Test
  key: quantized-models-test
-  timeout_in_minutes: 50
+  timeout_in_minutes: 60
  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_uds.py
+  # - tests/entrypoints/openai/test_return_token_ids.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"
-  - 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_completion.py
+  # - 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
-  # finish chunk without logprobs, while the test reads `logprobs.tokens` on every chunk.
-  - 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 entrypoints/openai/test_return_token_ids.py
+  # - 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/scale_out/token_in_token_out/test_serving_tokens.py
+  # - tests/entrypoints/serve/dev/rpc/test_collective_rpc.py
+  - tests/entrypoints/serve/disagg/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
-  # server_load can be flaky under the Rust frontend; keep it excluded for now.
+  # - pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py
  - 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/disagg/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:
@@ -81,7 +74,7 @@ steps:
  - pytest -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine

 - label: Rust Frontend Tool Use
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - rust/
@@ -95,7 +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:
@@ -106,13 +99,9 @@ steps:
  - vllm/v1/engine/
  - vllm/v1/worker/
  - tests/utils.py
-  - tests/v1/distributed/test_external_lb_dp.py
-  - tests/v1/distributed/test_hybrid_lb_dp.py
  - tests/v1/distributed/test_internal_lb_dp.py
  commands:
  - export VLLM_USE_RUST_FRONTEND=1
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
  - export NCCL_CUMEM_HOST_ENABLE=0
  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py -k "not 4 and not server_info"
-  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py -k "not 4 and not server_info"
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py -k "not 4 and not server_info"
@@ -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
@@ -19,7 +19,7 @@ steps:
    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
  mirror:
    amd:
-      device: mi325_1
+      device: mi250_1
      depends_on:
      - image-build-amd
      commands:
@@ -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/
@@ -12,24 +12,10 @@ steps:
    - tests/v1/e2e/spec_decode/
  commands:
    - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
-  mirror:
-    amd:
-      device: mi325_1
-      timeout_in_minutes: 60
-      depends_on:
-      - image-build-amd
-      source_file_dependencies:
-      - vllm/v1/spec_decode/
-      - vllm/v1/worker/gpu/spec_decode/
-      - vllm/model_executor/model_loader/
-      - vllm/v1/sample/
-      - vllm/model_executor/layers/
-      - tests/v1/e2e/spec_decode/
-      - vllm/platforms/rocm.py

 - 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:
@@ -41,7 +27,7 @@ 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/
@@ -82,7 +68,7 @@ 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/
@@ -93,9 +79,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 55
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
+      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -109,7 +93,7 @@ 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/
@@ -120,7 +104,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 55
+      timeout_in_minutes: 50
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -134,7 +118,7 @@ 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:
@@ -146,7 +130,7 @@ steps:

 - label: Speculators Correctness
  key: speculators-correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 60
  device: h100
  optional: true
  num_devices: 1
@@ -159,7 +143,7 @@ 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:
@@ -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
@@ -13,13 +13,6 @@ steps:
  - tests/weight_loading
  commands:
    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
-  mirror:
-    amd:
-      device: mi300_2
-      depends_on:
-      - image-build-amd
-      commands:
-      - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt

 # - label: Weight Loading Multiple GPU - Large Models # optional
 #   working_dir: "/vllm-workspace/tests"
@@ -1,35 +0,0 @@
---
-name: ci-fails-buildkite
-description: Fetch and diagnose vLLM Buildkite CI failure logs. Use when investigating failing CI jobs on a PR or build, when the user pastes a buildkite.com URL, or asks to fetch/diagnose CI logs.
---
-
-# Diagnosing vLLM Buildkite CI Failures
-
-Buildkite logs are public; no login needed.
-
-`.buildkite/scripts/ci-fetch-log.sh` saves each log as `ci-<build>-<job-name>.log`, stripped of timestamps and ANSI codes. Existing files are kept; set `CI_FETCH_LOG_FORCE=1` to refetch.
-
-## Fetching logs
-
-```bash
-# All failed jobs in a PR's latest build (current branch's PR if omitted):
-.buildkite/scripts/ci-fetch-log.sh --pr <PR>
-
-# All failed jobs in a build (--soft also includes soft-failed jobs;
-# --all fetches every finished job):
-.buildkite/scripts/ci-fetch-log.sh "https://buildkite.com/vllm/ci/builds/<N>"
-
-# One job — `gh pr checks` URLs (#<job_uuid>) and web UI URLs (?sid=) both
-# work; pass "-" as a second argument to stream to stdout:
-.buildkite/scripts/ci-fetch-log.sh "https://buildkite.com/vllm/ci/builds/<N>#<job_uuid>"
-```
-
-To clean an already-downloaded log with `.buildkite/scripts/ci-clean-log.sh`:
-
-```bash
-./ci-clean-log.sh ci.log
-```
-
-## Reference
-
-See [docs/contributing/ci/failures.md](../../../docs/contributing/ci/failures.md) for the full guide: filing CI failure issues, investigating/bisecting, reproducing flaky tests, and daily triage.
@@ -2,16 +2,17 @@
 # for more info about CODEOWNERS file

 # This lists cover the "core" components of vLLM that require careful review
-/vllm/compilation @zou3519 @youkaichao @ProExpertProg @BoyuanFeng
-/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery @xuechendi @ivanium
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg @BoyuanFeng @vadiklyutiy
+/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery @xuechendi
 /vllm/lora @jeejeelee
 /vllm/model_executor/layers/attention @LucasWilkinson @MatthewBonanni
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety @zyongye
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety @zyongye
 /vllm/model_executor/layers/mamba @tdoublep @tomeras91
-/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py @tdoublep @ZJY0516 @vadiklyutiy
+/vllm/model_executor/layers/mamba/gdn_linear_attn.py @tdoublep @ZJY0516 @vadiklyutiy
+/vllm/model_executor/layers/rotary_embedding.py @vadiklyutiy
 /vllm/model_executor/model_loader @22quinn
-/vllm/model_executor/layers/batch_invariant.py @yewentao256
+/vllm/model_executor/layers/batch_invariant.py @yewentao256 
 /vllm/ir @ProExpertProg
 /vllm/kernels/ @ProExpertProg @tjtanaa
 /vllm/kernels/helion @ProExpertProg @zou3519
@@ -23,7 +24,7 @@
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
 /vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @yewentao256 @ProExpertProg
-/vllm/config/cache.py @heheda12345 @ivanium
+/vllm/config/cache.py @heheda12345

 # Config utils
 /vllm/config/utils.py @hmellor
@@ -67,17 +68,16 @@
 /vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety @vadiklyutiy
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
 /vllm/v1/attention/backends/gdn_attn.py @ZJY0516 @vadiklyutiy
-/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery @ivanium
+/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /vllm/v1/sample @22quinn @houseroad @njhill
 /vllm/v1/spec_decode @benchislett @luccafong @MatthewBonanni
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
-/vllm/v1/kv_cache_interface.py @heheda12345 @ivanium
+/vllm/v1/kv_cache_interface.py @heheda12345
 /vllm/v1/kv_offload @ApostaC @orozery
-/vllm/v1/simple_kv_offload @ivanium
 /vllm/v1/engine @njhill
 /vllm/v1/executor @njhill
 /vllm/v1/worker @njhill
-/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche @ivanium
+/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche

 # Model runner V2
 /vllm/v1/worker/gpu @WoosukKwon @njhill @yewentao256
@@ -104,14 +104,13 @@
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
-/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery @ivanium
+/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 /tests/models/language/generation/test_hybrid.py @tdoublep @tomeras91
 /tests/v1/kv_connector/nixl_integration @NickLucche
-/tests/v1/kv_connector @ApostaC @orozery @ivanium
+/tests/v1/kv_connector @ApostaC @orozery
 /tests/v1/kv_offload @ApostaC @orozery
-/tests/v1/simple_kv_offload @ivanium
 /tests/v1/determinism @yewentao256
 /tests/reasoning @aarnphm @chaunceyjiang @sfeng33 @bbrowning
 /tests/tool_parsers @aarnphm @chaunceyjiang @sfeng33 @bbrowning
@@ -119,7 +118,17 @@

 # Transformers modeling backend
 /vllm/model_executor/models/transformers @hmellor
-/tests/models/transformers @hmellor
+/tests/models/test_transformers.py @hmellor
+
+# Observability
+/vllm/config/observability.py @markmc
+/vllm/v1/metrics @markmc
+/tests/v1/metrics @markmc
+/vllm/tracing.py @markmc
+/tests/v1/tracing/test_tracing.py @markmc
+/vllm/config/kv_events.py @markmc
+/vllm/distributed/kv_events.py @markmc
+/tests/distributed/test_events.py @markmc

 # Docs
 /docs/mkdocs @hmellor
@@ -1,7 +0,0 @@
-# Custom self-hosted runner labels (e.g. the autoscaling vllm-runners pool) so
-# actionlint doesn't flag them as unknown in `runs-on`.
-self-hosted-runner:
-  labels:
-    - vllm-runners
-    # Not yet in actionlint's known-label set.
-    - macos-26
@@ -388,13 +388,9 @@ pull_request_rules:
    - or:
      - files~=^tests/tool_use/
      - files~=^tests/tool_parsers/
-      - files~=^tests/parser/
-      - files~=^tests/reasoning/
      - files~=^tests/entrypoints/openai/.*tool.*
      - files~=^tests/entrypoints/anthropic/.*tool.*
      - files~=^vllm/tool_parsers/
-      - files~=^vllm/parser/
-      - files~=^vllm/reasoning/
      - files=docs/features/tool_calling.md
      - files~=^examples/tool_calling/
  actions:
@@ -327,7 +327,7 @@ jobs:
                message: 'CC {users} for ROCm-related issue',
              },
              mistral: {
-                users: ['patrickvonplaten', 'juliendenize', 'andylolu2', 'NickLucche'],
+                users: ['patrickvonplaten', 'juliendenize', 'andylolu2'],
                message: 'CC {users} for Mistral-related issue',
              },
              // Add more label -> user mappings here
@@ -11,25 +11,13 @@ permissions:

 jobs:
  macos-m1-smoke-test:
-    # macos-26 (the supported target) is still a preview runner, so gate on GA
-    # macos-15 and keep macos-26 non-blocking.
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - os: macos-15
-            required: true
-          - os: macos-26
-            required: false
-    name: macos-m1-smoke-test (${{ matrix.os }})
-    runs-on: ${{ matrix.os }}
-    continue-on-error: ${{ !matrix.required }}
+    runs-on: macos-latest
    timeout-minutes: 30

    steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@v6.0.1

-      - uses: astral-sh/setup-uv@37802adc94f370d6bfd71619e3f0bf239e1f3b78 # v7.6.0
+      - uses: astral-sh/setup-uv@v7
        with:
          enable-cache: true
          cache-dependency-glob: |
@@ -84,11 +72,14 @@ jobs:
          # Test health endpoint
          curl -f http://localhost:8000/health

-          # Long prompt: hits the split-KV path that short prompts skip (#46769).
-          PAYLOAD=$(python -c "import json; print(json.dumps({'model': 'Qwen/Qwen3-0.6B', 'prompt': 'The quick brown fox jumps over the lazy dog. ' * 24, 'max_tokens': 16}))")
-          curl -f --max-time 120 http://localhost:8000/v1/completions \
+          # Test completion
+          curl -f http://localhost:8000/v1/completions \
            -H "Content-Type: application/json" \
-            -d "$PAYLOAD"
+            -d '{
+              "model": "Qwen/Qwen3-0.6B",
+              "prompt": "Hello",
+              "max_tokens": 5
+            }'

          # Cleanup
          kill "$SERVER_PID"
@@ -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 => readyLabels.includes(l.name));
+          const hasReadyLabel = pr.labels.some(l => l.name === 'ready');
          const hasVerifiedLabel = pr.labels.some(l => l.name === 'verified');

          const { data: mergedPRs } = await github.rest.search.issuesAndPullRequests({
@@ -41,22 +40,18 @@ 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 (the ready labels also trigger tests) 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:
    needs: pre-run-check
    if: always() && (needs.pre-run-check.result == 'success' || needs.pre-run-check.result == 'skipped')
-    runs-on: [self-hosted, linux, x64, vllm-runners]
+    runs-on: ubuntu-latest
    steps:
-    - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+    - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
    - uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
      with:
        python-version: "3.12"
-    # Provide shellcheck on PATH so tools/pre_commit/shellcheck.sh skips its
-    # wget + tar -xJ self-download, which the self-hosted runner image lacks
-    # (no wget/xz). Pinned to shellcheck 0.10.0 to match the script's "stable".
-    - run: python -m pip install shellcheck-py==0.10.0.1
    - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
    - run: echo "::add-matcher::.github/workflows/matchers/markdownlint.json"
    - run: echo "::add-matcher::.github/workflows/matchers/mypy.json"
@@ -199,9 +199,7 @@ cython_debug/
 .vscode/

 # Claude
-.claude/*
-!.claude/skills/
-!.claude/skills/**
+.claude/

 # Codex
 .codex/
@@ -131,19 +131,6 @@ repos:
        --python-version, "3.12",
      ]
      files: ^requirements/(common|xpu|test/xpu)\.(in|txt)$
-    - id: pip-compile
-      alias: pip-compile-cpu
-      name: pip-compile-cpu
-      args: [
-        requirements/test/cuda.in,
-        -o, requirements/test/cpu.txt,
-        --index-strategy, unsafe-best-match,
-        --torch-backend, cpu,
-        --python-platform, x86_64-manylinux_2_28,
-        --python-version, "3.12",
-      ]
-      files: ^requirements/(common|cpu|test/(cuda|cpu))\.(in|txt)$
-      exclude: ^requirements/test/cuda\.txt$
    - id: pip-compile
      alias: pip-compile-docs
      name: pip-compile-docs
@@ -29,7 +29,6 @@ Do not open one-off PRs for tiny edits (single typo, isolated style change, one
 - PR descriptions for AI-assisted work **must** include:
    - Why this is not duplicating an existing PR.
    - Test commands run and results.
-    - Model evaluation results when the change affects output, accuracy, or serving.
    - Clear statement that AI assistance was used.

 ### Fail-closed behavior
@@ -67,38 +66,23 @@ VLLM_USE_PRECOMPILED=1 uv pip install -e . --torch-backend=auto
 uv pip install -e . --torch-backend=auto
 ```

-### Tests
+### Running tests

 > Requires [Environment setup](#environment-setup) and [Installing dependencies](#installing-dependencies).

 ```bash
-# Install test dependencies (use cuda.in on non-x86_64):
-uv pip install -r requirements/test/cuda.in
+# Install test dependencies.
+# requirements/test/cuda.txt is pinned to x86_64; on other platforms, use the
+# unpinned source file instead:
+uv pip install -r requirements/test/cuda.in    # resolves for current platform
+# Or on x86_64:
+uv pip install -r requirements/test/cuda.txt

-# Run a specific test file:
+# Run a specific test file (use .venv/bin/python directly;
+# `source activate` does not persist in non-interactive shells):
 .venv/bin/python -m pytest tests/path/to/test_file.py -v
 ```

-When adding tests:
-
- **Design before you write.** Answer four questions first: what is the module
-  for, what is its I/O contract, what failure am I guarding against, and what is
-  the cheapest level that catches it (unit over integration over e2e)?
- **Reuse before create.** Extend existing test files, `conftest.py` fixtures, and
-  helpers; add a new file only when no nearby suite fits.
- **Test behavior with intent.** Assert observable outcomes through public APIs;
-  state why in the name or docstring. Skip trivial wiring; flaky tests are worse
-  than no tests.
- **Keep it minimal.** One behavior per test and the smallest setup that
-  triggers it; if the test diff dwarfs the code change, cut scope.
- **No one-off kernel benchmarks in `tests/`.** Put kernel perf work in
-  `benchmarks/kernels/`; prove correctness in existing pytest suites.
- **Run model evals for model-affecting changes.** Search `tests/evals/` or use
-  `vllm bench` and include results in the PR — do not wait for reviewers to ask.
-
-For model-specific requirements, see
-[`docs/contributing/model/tests.md`](docs/contributing/model/tests.md).
-
 ### Running linters

 > Requires [Environment setup](#environment-setup).
@@ -123,18 +107,34 @@ Use [Google-style docstrings](https://google.github.io/styleguide/pyguide.html#3

 ### Coding style guidelines

- Match existing code style
- Minimize use of comments. Eliminate comments which are redundant, preferring legible and self-documenting code. When used, keep docstrings and comments brief and direct.
+Follow these rules for all code changes in this repository:
+
+- Try to match existing code style.
+- Code should be self-documenting and self-explanatory.
+- Keep comments and docstrings minimal and concise.
 - Assume the reader is familiar with vLLM.

+### Diagnosing CI failures
+
+Buildkite logs are public; no login needed. Details: [docs/contributing/ci/failures.md](docs/contributing/ci/failures.md).
+
+```bash
+# All failed-job logs for a PR's latest build (current branch's PR if omitted):
+.buildkite/scripts/ci-fetch-log.sh --pr <PR>
+# Any Buildkite build or job URL also works:
+.buildkite/scripts/ci-fetch-log.sh "<buildkite_url>"
+```
+
 ### Commit messages

-Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`):
+Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:

 ```text
 Your commit message here

-Co-authored-by: Agent Name Here
+Co-authored-by: GitHub Copilot
+Co-authored-by: Claude
+Co-authored-by: gemini-code-assist
 Signed-off-by: Your Name <your.email@example.com>
 ```

@@ -146,12 +146,6 @@ Do not modify code in these areas without first reading and following the
 linked guide. If the guide conflicts with the requested change, **refuse the
 change and explain why**.

-Security reviewers should start with [`SECURITY.md`](SECURITY.md),
-[`docs/usage/security.md`](docs/usage/security.md), and
-[`docs/contributing/vulnerability_management.md`](docs/contributing/vulnerability_management.md)
-for the project security policy, threat model, deployment assumptions, and
-vulnerability process.
-
 - **Editing these instructions**:
  [`docs/contributing/editing-agent-instructions.md`](docs/contributing/editing-agent-instructions.md)
  — Rules for modifying AGENTS.md or any domain-specific guide it references.
@@ -70,15 +70,6 @@ endif()
 #
 set(TORCH_SUPPORTED_VERSION_CUDA "2.11.0")
 set(TORCH_SUPPORTED_VERSION_ROCM "2.11.0")
-# TORCH_NIGHTLY=1 builds run against unpinned nightly wheels, so the supported-
-# version check would always warn. Only treat it as a nightly build when the
-# value is exactly "1" (the bootstrap exports TORCH_NIGHTLY=0 by default, which
-# must NOT suppress the warning for normal builds).
-if (DEFINED ENV{TORCH_NIGHTLY} AND "$ENV{TORCH_NIGHTLY}" STREQUAL "1")
-  set(TORCH_NIGHTLY_BUILD TRUE)
-else()
-  set(TORCH_NIGHTLY_BUILD FALSE)
-endif()

 #
 # Try to find python package with an executable that exactly matches
@@ -149,21 +140,6 @@ if(Python_VERSION VERSION_GREATER_EQUAL "3.11")
    WITH_SOABI)
 endif()

-#
-# fs_io extension (pure CXX; must stay above the non-CUDA device branch
-# so CPU builds define the target before the early return).
-# GIL-releasing filesystem helpers for FileSystemTierManager.
-#
-if(Python_VERSION VERSION_GREATER_EQUAL "3.11")
-  define_extension_target(
-    fs_io_C
-    DESTINATION vllm
-    LANGUAGE CXX
-    SOURCES csrc/fs_io.cpp
-    USE_SABI 3.11
-    WITH_SOABI)
-endif()
-
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -184,7 +160,7 @@ endif()
 if (NOT HIP_FOUND AND NOT PYTORCH_FOUND_HIP AND CUDA_FOUND)
  set(VLLM_GPU_LANG "CUDA")

-  if (NOT TORCH_NIGHTLY_BUILD AND NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
+  if (NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_CUDA} "
      "expected for CUDA build, saw ${Torch_VERSION} instead.")
  endif()
@@ -197,7 +173,7 @@ elseif(HIP_FOUND OR PYTORCH_FOUND_HIP)
  enable_language(HIP)

  # ROCm 5.X and 6.X
-  if (NOT TORCH_NIGHTLY_BUILD AND ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
+  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
      Torch_VERSION VERSION_LESS ${TORCH_SUPPORTED_VERSION_ROCM})
    message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
      "expected for ROCm build, saw ${Torch_VERSION} instead.")
@@ -294,16 +270,6 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
  #
  set(CMAKE_${VLLM_GPU_LANG}_FLAGS "${CMAKE_${VLLM_GPU_LANG}_FLAGS} -Wno-unused-result -Wno-unused-value")
  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-result -Wno-unused-value")
-
-  # When using LTO then *.cpp files must be compiled with same compiler as used linker
-  # So if HIP uses clang linker we also must use it
-  # Otherwise symbols will be missing from .so
-  if (CMAKE_CXX_FLAGS MATCHES "\-flto")
-    if(NOT CMAKE_CXX_COMPILER_ID STREQUAL CMAKE_HIP_COMPILER_ID)
-      message(FATAL_ERROR "LTO is enabled for ROCm build, but the C++ compiler (${CMAKE_CXX_COMPILER_ID}) and HIP compiler (${CMAKE_HIP_COMPILER_ID}) are different which is not supported. "
-                          "Please ensure they are same by setting CXX=${CMAKE_HIP_COMPILER} environment variable. Or alternatively disable LTO.")
-    endif()
-   endif()
 endif()

 #
@@ -353,33 +319,112 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 endif()

 #
-# Legacy _C extension (ROCm only — CUDA ops migrated to _C_stable_libtorch)
+# _C extension
 #

-if(VLLM_GPU_LANG STREQUAL "HIP")
-  set(VLLM_EXT_SRC
-    "csrc/torch_bindings.cpp"
-    "csrc/custom_quickreduce.cu")
+set(VLLM_EXT_SRC
+  "csrc/quantization/activation_kernels.cu"
+  "csrc/push_all_reduce.cu"
+  "csrc/torch_bindings.cpp")

-  message(STATUS "Enabling C extension.")
-  define_extension_target(
-    _C
-    DESTINATION vllm
-    LANGUAGE ${VLLM_GPU_LANG}
-    SOURCES ${VLLM_EXT_SRC}
-    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
-    ARCHITECTURES ${VLLM_GPU_ARCHES}
-    INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
-    INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
-    USE_SABI 3
-    WITH_SOABI)
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")

-  # If CUTLASS is compiled on NVCC >= 12.5, it by default uses
-  # cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
-  # driver API. This causes problems when linking with earlier versions of CUDA.
-  # Setting this variable sidesteps the issue by calling the driver directly.
-  target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
-endif() # _C HIP endif
+  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
+  set(CUTLASS_REVISION "v4.4.2")
+
+  # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
+  if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
+    set(VLLM_CUTLASS_SRC_DIR $ENV{VLLM_CUTLASS_SRC_DIR})
+  endif()
+
+  if(VLLM_CUTLASS_SRC_DIR)
+    if(NOT IS_ABSOLUTE VLLM_CUTLASS_SRC_DIR)
+      get_filename_component(VLLM_CUTLASS_SRC_DIR "${VLLM_CUTLASS_SRC_DIR}" ABSOLUTE)
+    endif()
+    message(STATUS "The VLLM_CUTLASS_SRC_DIR is set, using ${VLLM_CUTLASS_SRC_DIR} for compilation")
+    FetchContent_Declare(cutlass SOURCE_DIR ${VLLM_CUTLASS_SRC_DIR})
+  else()
+    FetchContent_Declare(
+        cutlass
+        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
+        # Please keep this in sync with CUTLASS_REVISION line above.
+        GIT_TAG ${CUTLASS_REVISION}
+        GIT_PROGRESS TRUE
+
+        # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
+        # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
+        # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
+        GIT_SHALLOW TRUE
+    )
+  endif()
+  FetchContent_MakeAvailable(cutlass)
+
+  set_gencode_flags_for_srcs(
+    SRCS "${VLLM_EXT_SRC}"
+    CUDA_ARCHS "${CUDA_ARCHS}")
+
+  # Expert-specialization MXFP8 blockscaled grouped kernels (SM100+).
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
+    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(ES_MXFP8_GROUPED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+  endif()
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND ES_MXFP8_GROUPED_MM_ARCHS)
+    set(ES_MXFP8_GROUPED_MM_SRCS
+      "csrc/libtorch_stable/moe/mxfp8_moe/cutlass_mxfp8_grouped_mm.cu"
+      "csrc/libtorch_stable/moe/mxfp8_moe/mxfp8_experts_quant.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${ES_MXFP8_GROUPED_MM_SRCS}"
+      CUDA_ARCHS "${ES_MXFP8_GROUPED_MM_ARCHS}")
+    list(APPEND VLLM_STABLE_EXT_SRC "${ES_MXFP8_GROUPED_MM_SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_ES_MXFP8_GROUPED_MM_SM100=1")
+    message(STATUS "Building ES MXFP8 grouped kernels for archs: ${ES_MXFP8_GROUPED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8
+        AND ES_MXFP8_GROUPED_MM_ARCHS)
+      message(STATUS "Not building ES MXFP8 grouped kernels as CUDA Compiler version is "
+                     "not >= 12.8.")
+    else()
+      message(STATUS "Not building ES MXFP8 grouped kernels as no compatible archs found "
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
+
+
+# if CUDA endif
+endif()
+
+if (VLLM_GPU_LANG STREQUAL "HIP")
+  # Add QuickReduce kernels (ROCm-only; not part of stable ABI migration).
+  # TODO: Remove the cuda_view when ROCm upgrade to torch 2.11.
+  list(APPEND VLLM_EXT_SRC
+    "csrc/custom_quickreduce.cu"
+    "csrc/cuda_view.cu"
+    "csrc/libtorch_stable/cuda_utils_kernels.cu"
+  )
+# if ROCM endif
+endif()
+
+message(STATUS "Enabling C extension.")
+define_extension_target(
+  _C
+  DESTINATION vllm
+  LANGUAGE ${VLLM_GPU_LANG}
+  SOURCES ${VLLM_EXT_SRC}
+  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+  ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  USE_SABI 3
+  WITH_SOABI)
+
+# If CUTLASS is compiled on NVCC >= 12.5, it by default uses
+# cudaGetDriverEntryPointByVersion as a wrapper to avoid directly calling the
+# driver API. This causes problems when linking with earlier versions of CUDA.
+# Setting this variable sidesteps the issue by calling the driver directly.
+target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)

 if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  #
@@ -387,11 +432,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  #
  set(VLLM_STABLE_EXT_SRC
    "csrc/libtorch_stable/torch_bindings.cpp"
-    "csrc/libtorch_stable/cuda_view.cu"
-    "csrc/libtorch_stable/cuda_utils_kernels.cu"
    "csrc/libtorch_stable/activation_kernels.cu"
-    "csrc/libtorch_stable/ngram_embedding_kernels.cu"
-    "csrc/libtorch_stable/quantization/activation_kernels.cu"
    "csrc/libtorch_stable/quantization/w8a8/int8/scaled_quant.cu"
    "csrc/libtorch_stable/quantization/w8a8/fp8/common.cu"
    "csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu"
@@ -409,63 +450,18 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
    "csrc/libtorch_stable/sampler.cu"
    "csrc/libtorch_stable/topk.cu"
    "csrc/libtorch_stable/mamba/selective_scan_fwd.cu"
+    "csrc/libtorch_stable/attention/paged_attention_v1.cu"
+    "csrc/libtorch_stable/attention/paged_attention_v2.cu"
+    "csrc/libtorch_stable/cache_kernels.cu"
    "csrc/libtorch_stable/cache_kernels.cu"
    "csrc/libtorch_stable/cache_kernels_fused.cu"
    "csrc/libtorch_stable/custom_all_reduce.cu"
    "csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu")

-  if(VLLM_GPU_LANG STREQUAL "CUDA" AND
-     DEFINED CMAKE_CUDA_COMPILER_VERSION AND
-     CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.0)
-
-    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-      cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
-        "9.0a;10.0f;10.1f;10.3f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
-    else()
-      cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
-        "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-    endif()
-
-    if(COOPERATIVE_TOPK_ARCHS)
-      list(APPEND VLLM_GPU_FLAGS "-DVLLM_ENABLE_COOPERATIVE_TOPK=1")
-
-    endif()
-  endif()
-
  if(VLLM_GPU_LANG STREQUAL "CUDA")
-    SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
-
-    # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-    set(CUTLASS_REVISION "v4.4.2")
-
-    # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
-    if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
-      set(VLLM_CUTLASS_SRC_DIR $ENV{VLLM_CUTLASS_SRC_DIR})
-    endif()
-
-    if(VLLM_CUTLASS_SRC_DIR)
-      if(NOT IS_ABSOLUTE VLLM_CUTLASS_SRC_DIR)
-        get_filename_component(VLLM_CUTLASS_SRC_DIR "${VLLM_CUTLASS_SRC_DIR}" ABSOLUTE)
-      endif()
-      message(STATUS "The VLLM_CUTLASS_SRC_DIR is set, using ${VLLM_CUTLASS_SRC_DIR} for compilation")
-      FetchContent_Declare(cutlass SOURCE_DIR ${VLLM_CUTLASS_SRC_DIR})
-    else()
-      FetchContent_Declare(
-          cutlass
-          GIT_REPOSITORY https://github.com/nvidia/cutlass.git
-          # Please keep this in sync with CUTLASS_REVISION line above.
-          GIT_TAG ${CUTLASS_REVISION}
-          GIT_PROGRESS TRUE
-
-          # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
-          # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
-          # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
-          GIT_SHALLOW TRUE
-      )
-    endif()
-    FetchContent_MakeAvailable(cutlass)
-
    list(APPEND VLLM_STABLE_EXT_SRC
+      "csrc/libtorch_stable/cuda_view.cu"
+      "csrc/libtorch_stable/cuda_utils_kernels.cu"
      "csrc/libtorch_stable/cutlass_extensions/common.cpp"
      "csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_entry.cu"
      "csrc/libtorch_stable/quantization/fp4/nvfp4_quant_entry.cu"
@@ -546,14 +542,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
      SRCS "${VLLM_STABLE_EXT_SRC}"
      CUDA_ARCHS "${CUDA_ARCHS}")

-    if(COOPERATIVE_TOPK_ARCHS)
-      list(APPEND VLLM_STABLE_EXT_SRC
-        "csrc/libtorch_stable/cooperative_topk.cu")
-      set_gencode_flags_for_srcs(
-        SRCS "csrc/libtorch_stable/cooperative_topk.cu"
-        CUDA_ARCHS "${COOPERATIVE_TOPK_ARCHS}")
-    endif()
-
    # Only build Marlin kernels if we are building for at least some compatible archs.
    # Keep building Marlin for 9.0 as there are some group sizes and shapes that
    # are not supported by Machete yet.
@@ -899,9 +887,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  endif()

  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f" "${CUDA_ARCHS}")
  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.3a" "${CUDA_ARCHS}")
  endif()
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
    set(CUTLASS_MOE_SM100_SRCS "csrc/libtorch_stable/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
@@ -971,6 +959,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
      SRCS "${FP4_SM120_SRCS}"
      CUDA_ARCHS "${FP4_SM120_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${FP4_SM120_SRCS}")
+    target_compile_definitions(_C PRIVATE ENABLE_NVFP4_SM120=1)
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM120=1")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM120=1")
    message(STATUS "Building SM12x NVFP4 for archs: ${FP4_SM120_ARCHS}")
@@ -1003,6 +992,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
      SRCS "${FP4_SM100_SRCS}"
      CUDA_ARCHS "${FP4_SM100_ARCHS}")
    list(APPEND VLLM_STABLE_EXT_SRC "${FP4_SM100_SRCS}")
+    target_compile_definitions(_C PRIVATE ENABLE_NVFP4_SM100=1)
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM100=1")
    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
    message(STATUS "Building SM10x/11x NVFP4/MXFP4 for archs: ${FP4_SM100_ARCHS}")
@@ -1096,24 +1086,25 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
    USE_SABI 3
    WITH_SOABI)

-  # Set TORCH_TARGET_VERSION for stable ABI compatibility.
-  # This ensures we only use C-shim APIs available in PyTorch 2.11.
-  # _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
-  # which is currently set to 2.11.
-  target_compile_definitions(_C_stable_libtorch PRIVATE
-    TORCH_TARGET_VERSION=0x020B000000000000ULL)
-
  # Needed to use cuda/hip APIs from C-shim
  if(VLLM_GPU_LANG STREQUAL "CUDA")
+    # Set TORCH_TARGET_VERSION for stable ABI compatibility.
+    # This ensures we only use C-shim APIs available in PyTorch 2.11.
+    # _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
+    # which is currently set to 2.11.
+    target_compile_definitions(_C_stable_libtorch PRIVATE
+      TORCH_TARGET_VERSION=0x020B000000000000ULL)
    target_compile_definitions(_C_stable_libtorch PRIVATE USE_CUDA)
-    if(COOPERATIVE_TOPK_ARCHS)
-      target_compile_definitions(_C_stable_libtorch PRIVATE
-        VLLM_ENABLE_COOPERATIVE_TOPK=1)
-    endif()
    # Needed by CUTLASS kernels
    target_compile_definitions(_C_stable_libtorch PRIVATE
      CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
  elseif(VLLM_GPU_LANG STREQUAL "HIP")
+    # Set TORCH_TARGET_VERSION for stable ABI compatibility.
+    # This ensures we only use C-shim APIs available in PyTorch 2.10.
+    # _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
+    # which is currently set to 2.10.
+    target_compile_definitions(_C_stable_libtorch PRIVATE
+      TORCH_TARGET_VERSION=0x020A000000000000ULL)
    target_compile_definitions(_C_stable_libtorch PRIVATE USE_ROCM)
  endif()

@@ -1321,20 +1312,25 @@ define_extension_target(
  USE_SABI 3
  WITH_SOABI)

-# Set TORCH_TARGET_VERSION for stable ABI compatibility.
-# This ensures we only use C-shim APIs available in PyTorch 2.11.
-# _moe_C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
-# which is currently set to 2.11.
-target_compile_definitions(_moe_C_stable_libtorch PRIVATE
-  TORCH_TARGET_VERSION=0x020B000000000000ULL)
-
 # Needed to use cuda/hip APIs from C-shim
 if(VLLM_GPU_LANG STREQUAL "CUDA")
+  # Set TORCH_TARGET_VERSION for stable ABI compatibility.
+  # This ensures we only use C-shim APIs available in PyTorch 2.11.
+  # _moe_C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
+  # which is currently set to 2.11.
+  target_compile_definitions(_moe_C_stable_libtorch PRIVATE
+    TORCH_TARGET_VERSION=0x020B000000000000ULL)
  target_compile_definitions(_moe_C_stable_libtorch PRIVATE USE_CUDA)
  # Needed by CUTLASS kernels
  target_compile_definitions(_moe_C_stable_libtorch PRIVATE
    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
 elseif(VLLM_GPU_LANG STREQUAL "HIP")
+  # Set TORCH_TARGET_VERSION for stable ABI compatibility.
+  # This ensures we only use C-shim APIs available in PyTorch 2.10.
+  # _moe_C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
+  # which is currently set to 2.10.
+  target_compile_definitions(_moe_C_stable_libtorch PRIVATE
+    TORCH_TARGET_VERSION=0x020A000000000000ULL)
  target_compile_definitions(_moe_C_stable_libtorch PRIVATE USE_ROCM)
 endif()

@@ -53,16 +53,6 @@ from common import (
 from vllm.v1.worker.workspace import init_workspace_manager


-def _str2bool(v) -> bool:
-    if isinstance(v, bool):
-        return v
-    if v.lower() in ("true", "1", "yes", "t"):
-        return True
-    if v.lower() in ("false", "0", "no", "f"):
-        return False
-    raise argparse.ArgumentTypeError(f"expected a boolean, got {v!r}")
-
-
 def run_standard_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
    """Run standard attention benchmark (Flash/Triton/FlashInfer)."""
    from runner import run_attention_benchmark
@@ -495,20 +485,6 @@ def main():
        help="Prefill backends to compare (fa2, fa3, fa4). "
        "Uses the first decode backend for impl construction.",
    )
-    parser.add_argument(
-        "--fp8-output-scale",
-        type=float,
-        help="Static per-tensor scale enabling the MLA prefill FP8-output "
-        "comparison on FA4 (fused write vs standalone post-quant).",
-    )
-    parser.add_argument(
-        "--fuse-quant-op",
-        nargs="+",
-        type=_str2bool,
-        help="FP8-output write path(s) to run: false = bf16 attention + "
-        "standalone static-FP8 quant, true = FA4 writes FP8 directly. "
-        "Default: both.",
-    )

    # Batch specifications
    parser.add_argument(
@@ -642,12 +618,6 @@ def main():
        # Prefill backends (e.g., ["fa3", "fa4"])
        args.prefill_backends = yaml_config.get("prefill_backends", None)

-        # FP8 output benchmark knobs; CLI wins.
-        if args.fp8_output_scale is None:
-            args.fp8_output_scale = yaml_config.get("fp8_output_scale", None)
-        if args.fuse_quant_op is None:
-            args.fuse_quant_op = yaml_config.get("fuse_quant_op", None)
-
        # Check for special modes
        args.mode = yaml_config.get("mode", None)

@@ -817,59 +787,8 @@ def main():
            "skipped (timings are placeholder zeros).[/]"
        )

-    # FA4 fused FP8 output vs standalone post-quant, on the same fa4 kernel:
-    # the delta is the post-quant kernel the fused path removes.
-    fp8_output_scale = getattr(args, "fp8_output_scale", None)
-    if fp8_output_scale is not None:
-        decode_backend = backends[0]
-        fuse_variants = args.fuse_quant_op or [False, True]
-        label_of = {False: "post_quant", True: "fused"}
-        console.print(
-            f"[yellow]FP8 output comparison @ scale={fp8_output_scale} "
-            f"(prefill=fa4, decode impl={decode_backend})[/]"
-        )
-        fp8_results = []
-        total = len(fuse_variants) * len(args.batch_specs)
-        with tqdm(total=total, desc="FP8 output benchmarking") as pbar:
-            for spec in args.batch_specs:
-                for fuse in fuse_variants:
-                    config = BenchmarkConfig(
-                        backend=decode_backend,
-                        batch_spec=spec,
-                        num_layers=args.num_layers,
-                        head_dim=args.head_dim,
-                        num_q_heads=args.num_q_heads,
-                        num_kv_heads=args.num_kv_heads,
-                        block_size=args.block_size,
-                        device=args.device,
-                        repeats=args.repeats,
-                        warmup_iters=args.warmup_iters,
-                        profile_memory=args.profile_memory,
-                        kv_cache_dtype=args.kv_cache_dtype,
-                        use_cuda_graphs=args.cuda_graphs,
-                        prefill_backend="fa4",
-                    )
-                    result = run_benchmark(
-                        config, output_scale=fp8_output_scale, fuse_quant_op=fuse
-                    )
-                    label = label_of[fuse]
-                    labeled_config = replace(result.config, backend=label)
-                    result = replace(result, config=labeled_config)
-                    fp8_results.append(result)
-
-                    if not result.success:
-                        console.print(f"[red]Error {label} {spec}: {result.error}[/]")
-
-                    pbar.update(1)
-
-        console.print("\n[bold green]FP8 Output Results:[/]")
-        formatter = ResultsFormatter(console)
-        labels = [label_of[f] for f in fuse_variants]
-        formatter.print_table(fp8_results, labels, compare_to_fastest=True)
-        all_results = fp8_results
-
    # Handle special mode: decode_vs_prefill comparison
-    elif hasattr(args, "mode") and args.mode == "decode_vs_prefill":
+    if hasattr(args, "mode") and args.mode == "decode_vs_prefill":
        console.print("[yellow]Mode: Decode vs Prefill pipeline comparison[/]")
        console.print(
            "[dim]For each query length, testing both decode and prefill pipelines[/]"
@@ -1,44 +0,0 @@
-# MLA prefill FP8-output microbenchmark (FA4).
-# Compares the fused FP8 write against bf16 attention + a standalone static-FP8
-# quant; the delta is the post-quant kernel the fused path removes.
-# DeepSeek-Coder-V2-Lite dims; FA4 needs SM100/110.
-#
-# Usage:
-#   python benchmark.py --config configs/mla_fa4_fp8_output.yaml
-
-description: "MLA prefill FA4 fused-FP8 output vs post-quant"
-
-model:
-  name: "deepseek-v2-lite"
-  num_layers: 27
-  num_q_heads: 16
-  num_kv_heads: 1
-  head_dim: 576
-  kv_lora_rank: 512
-  qk_nope_head_dim: 128
-  qk_rope_head_dim: 64
-  v_head_dim: 128
-  block_size: 128
-
-# Pure prefill (q_len == kv_len) so every token goes through forward_mha.
-batch_specs:
-  - "q512"
-  - "q1k"
-  - "q2k"
-  - "q4k"
-  - "q8k"
-  - "2q4k"
-  - "4q4k"
-  - "8q4k"
-
-# Only used to construct the MLA impl; the pure-prefill specs skip decode.
-decode_backends:
-  - CUTLASS_MLA
-
-# Sweep the two FP8 write paths (prefill backend is fixed to fa4).
-fp8_output_scale: 0.1
-fuse_quant_op: [false, true]
-
-device: "cuda:0"
-repeats: 50
-warmup_iters: 10
@@ -708,8 +708,6 @@ def _run_single_benchmark(
    device: torch.device,
    indexer=None,
    kv_cache_dtype: str | None = None,
-    output_scale: float | None = None,
-    fuse_quant_op: bool = False,
 ) -> BenchmarkResult:
    """
    Run a single benchmark iteration.
@@ -723,11 +721,6 @@ def _run_single_benchmark(
        mla_dims: MLA dimension configuration
        device: Target device
        indexer: Optional MockIndexer for sparse backends
-        output_scale: Static per-tensor FP8 scale for prefill output. None
-            keeps the plain bf16 output (no quantization).
-        fuse_quant_op: With output_scale set, True lets the prefill kernel write
-            FP8 directly; False runs bf16 attention then a standalone static-FP8
-            quant. The delta isolates the saved post-quant kernel.

    Returns:
        BenchmarkResult with timing statistics
@@ -831,55 +824,23 @@ def _run_single_benchmark(
            num_prefill, mla_dims, query_fmt, device, torch.bfloat16
        )

-    # Prefill FP8 output: fused (kernel writes e4m3) vs separate post-quant.
-    prefill_fp8_output = None
-    prefill_output_scale = None
-    prefill_quant_op = None
-    if has_prefill and output_scale is not None:
-        from vllm.platforms import current_platform
-
-        prefill_output_scale = torch.tensor(
-            [output_scale], device=device, dtype=torch.float32
-        )
-        if fuse_quant_op:
-            prefill_fp8_output = torch.empty_like(
-                prefill_inputs["output"], dtype=current_platform.fp8_dtype()
-            )
-        else:
-            from vllm.model_executor.layers.quantization.input_quant_fp8 import (
-                QuantFP8,
-            )
-            from vllm.model_executor.layers.quantization.utils.quant_utils import (
-                GroupShape,
-            )
-
-            prefill_quant_op = QuantFP8(static=True, group_shape=GroupShape.PER_TENSOR)
-
-    fused_output = output_scale is not None and fuse_quant_op
-
    # Build forward function (runs a single decode/prefill pass)
    def forward_fn():
        results = []
        if has_decode:
            results.append(impl.forward_mqa(decode_inputs, kv_cache, metadata, layer))
        if has_prefill:
-            out = impl.forward_mha(
-                prefill_inputs["q"],
-                prefill_inputs["k_c_normed"],
-                prefill_inputs["k_pe"],
-                kv_cache,
-                metadata,
-                prefill_inputs["k_scale"],
-                prefill_fp8_output if fused_output else prefill_inputs["output"],
-                prefill_output_scale if fused_output else None,
-            )
-            if fused_output:
-                out = prefill_fp8_output
-            elif prefill_quant_op is not None:
-                out, _ = prefill_quant_op(
-                    prefill_inputs["output"], prefill_output_scale
+            results.append(
+                impl.forward_mha(
+                    prefill_inputs["q"],
+                    prefill_inputs["k_c_normed"],
+                    prefill_inputs["k_pe"],
+                    kv_cache,
+                    metadata,
+                    prefill_inputs["k_scale"],
+                    prefill_inputs["output"],
                )
-            results.append(out)
+            )
        return results[0] if len(results) == 1 else tuple(results)

    def benchmark_fn():
@@ -920,8 +881,6 @@ def _run_mla_benchmark_batched(
    configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
    index_topk: int = 2048,
    prefill_backend: str | None = None,
-    output_scale: float | None = None,
-    fuse_quant_op: bool = False,
 ) -> list[BenchmarkResult]:
    """
    Unified batched MLA benchmark runner for all backends.
@@ -1061,8 +1020,6 @@ def _run_mla_benchmark_batched(
                    device,
                    indexer=indexer,
                    kv_cache_dtype=kv_cache_dtype,
-                    output_scale=output_scale,
-                    fuse_quant_op=fuse_quant_op,
                )
                results.append(result)

@@ -1090,8 +1047,6 @@ def run_mla_benchmark(
    num_kv_splits: int | None = None,
    index_topk: int = 2048,
    prefill_backend: str | None = None,
-    output_scale: float | None = None,
-    fuse_quant_op: bool = False,
 ) -> BenchmarkResult | list[BenchmarkResult]:
    """
    Unified MLA benchmark runner for all backends.
@@ -1111,9 +1066,6 @@ def run_mla_benchmark(
        index_topk: Topk value for sparse MLA backends (default 2048)
        prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
            When set, forces the specified FlashAttention version for prefill.
-        output_scale: Static per-tensor FP8 scale for prefill output (None = bf16).
-        fuse_quant_op: With output_scale set, fuse the FP8 write into the prefill
-            kernel vs a standalone post-quant kernel. See _run_single_benchmark.

    Returns:
        BenchmarkResult (single mode) or list of BenchmarkResult (batched mode)
@@ -1138,12 +1090,7 @@ def run_mla_benchmark(

    # Use unified batched execution
    results = _run_mla_benchmark_batched(
-        backend,
-        configs_with_params,
-        index_topk,
-        prefill_backend=prefill_backend,
-        output_scale=output_scale,
-        fuse_quant_op=fuse_quant_op,
+        backend, configs_with_params, index_topk, prefill_backend=prefill_backend
    )

    # Return single result or list based on input
@@ -12,7 +12,7 @@ from dataclasses import dataclass, field
 import aiohttp
 import huggingface_hub.constants
 from tqdm.asyncio import tqdm
-from transformers import AutoTokenizer, PythonBackend, TokenizersBackend
+from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast

 # NOTE(simon): do not import vLLM here so the benchmark script
 # can run without vLLM installed.
@@ -609,7 +609,7 @@ def get_tokenizer(
    tokenizer_mode: str = "auto",
    trust_remote_code: bool = False,
    **kwargs,
-) -> PythonBackend | TokenizersBackend:
+) -> PreTrainedTokenizer | PreTrainedTokenizerFast:
    if pretrained_model_name_or_path is not None and not os.path.exists(
        pretrained_model_name_or_path
    ):
@@ -1,358 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Benchmark and regression-test pinned (page-locked) CPU memory for vLLM.
-
-Verifies that enabling pinned memory does not regress throughput or latency
-compared to unpinned memory.  Each condition runs in an isolated ``spawn``
-subprocess so both start from a cold CUDA context, giving an unbiased
-comparison.
-
-Usage
-----
-Run all tests with the default model::
-
-    python benchmarks/benchmark_pin_memory.py -v
-
-Override the model and optional max-model-len::
-
-    python benchmarks/benchmark_pin_memory.py --model unsloth/Qwen3-1.7B -v
-    python benchmarks/benchmark_pin_memory.py --model unsloth/Qwen3-1.7B \
-        --max-model-len 8192 -v
-
-Run only throughput or latency tests::
-
-    python benchmarks/benchmark_pin_memory.py -v -k test_throughput
-    python benchmarks/benchmark_pin_memory.py -v -k test_latency
-
-Run only the v1 or v2 runner variant::
-
-    python benchmarks/benchmark_pin_memory.py -v -k v1
-    python benchmarks/benchmark_pin_memory.py -v -k v2
-
-Note: on WSL2, v1 runner tests are skipped because pin memory is not available
-for the v1 runner without cpu_offload_gb.  Run on other platforms to exercise v1.
-"""
-
-import argparse
-import json
-import multiprocessing
-import sys
-import tempfile
-
-import pytest
-
-# Allow up to 2% degradation.  Both benchmark runs start from an identical
-# cold CUDA context (separate spawn subprocesses), so the measured difference
-# reflects the genuine pin_memory overhead rather than cold/warm ordering bias.
-_THROUGHPUT_TOLERANCE = 0.98
-_THROUGHPUT_NUM_REQUESTS = 200
-_THROUGHPUT_INPUT_LEN = 128
-_THROUGHPUT_OUTPUT_LEN = 512
-_THROUGHPUT_MAX_NUM_SEQS = 128
-
-# Latency benchmark constants — match latency.py defaults.
-_LATENCY_TOLERANCE = 1.02  # Allow up to 2% latency regression.
-_LATENCY_BATCH_SIZE = 64
-_LATENCY_INPUT_LEN = 32
-_LATENCY_OUTPUT_LEN = 128
-_LATENCY_WARMUP_ITERS = 5
-_LATENCY_BENCH_ITERS = 15
-
-_DEFAULT_MODEL = "unsloth/Qwen3-1.7B"
-_DEFAULT_MAX_MODEL_LEN = 16384
-
-
-def _benchmark_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(add_help=False)
-    parser.add_argument("--model", default=_DEFAULT_MODEL)
-    parser.add_argument("--max-model-len", type=int, default=_DEFAULT_MAX_MODEL_LEN)
-    args, _ = parser.parse_known_args()
-    return args
-
-
-@pytest.fixture
-def model() -> str:
-    return _benchmark_args().model
-
-
-@pytest.fixture
-def max_model_len() -> int:
-    return _benchmark_args().max_model_len
-
-
-def _skip_if_pin_memory_not_available(engine_args_kwargs: dict) -> None:
-    """Skip the current pytest test if pin_memory is unavailable for this config."""
-    import vllm.utils.platform_utils as pu
-    from vllm.config import set_current_vllm_config
-    from vllm.engine.arg_utils import EngineArgs
-
-    vllm_config = EngineArgs(**engine_args_kwargs).create_engine_config()
-    with set_current_vllm_config(vllm_config):
-        pu.is_pin_memory_available.cache_clear()
-        if not pu.is_pin_memory_available():
-            import os
-
-            runner = "v2" if os.environ.get("VLLM_USE_V2_MODEL_RUNNER") == "1" else "v1"
-            model = engine_args_kwargs.get("model", "unknown")
-            print(
-                f"\033[33mSKIP: pin_memory not available for "
-                f"{runner} runner, model={model}\033[0m"
-            )
-            pytest.skip("pin_memory not available for this configuration")
-
-
-def _throughput_worker(
-    pin: bool,
-    engine_args_kwargs: dict,
-    q: "multiprocessing.Queue[float]",
-    v2_mode: bool = False,
-) -> None:
-    """Run throughput benchmark in a fresh spawn subprocess.
-
-    Delegates to vllm/benchmarks/throughput.py main() using the random dataset,
-    so the methodology matches the official benchmark.  Results are written to a
-    temp JSON file and forwarded through the queue as tokens/s.
-
-    v2_mode: when True, monkeypatches is_uva_available() to always return True
-    so the v2 model runner's UVA buffers remain functional even when pin=False.
-    This isolates the non-UVA pin_memory paths in v2.
-    """
-    import vllm.utils.platform_utils as pu
-    from vllm.platforms import current_platform
-
-    pu.is_pin_memory_available.cache_clear()
-    pu.is_uva_available.cache_clear()
-    type(current_platform).is_pin_memory_available = classmethod(lambda cls: pin)
-    if v2_mode:
-        pu.is_uva_available = lambda: True
-
-    from vllm.benchmarks.throughput import add_cli_args
-    from vllm.benchmarks.throughput import main as throughput_main
-
-    parser = argparse.ArgumentParser()
-    add_cli_args(parser)
-    args = parser.parse_args([])
-
-    for key, val in engine_args_kwargs.items():
-        setattr(args, key, val)
-    args.max_num_seqs = _THROUGHPUT_MAX_NUM_SEQS
-    args.dataset_name = "random"
-    args.input_len = _THROUGHPUT_INPUT_LEN
-    args.output_len = _THROUGHPUT_OUTPUT_LEN
-    # Nullify defaults that conflict with explicit input/output_len.
-    args.random_input_len = None
-    args.random_output_len = None
-    args.random_prefix_len = None
-    args.num_prompts = _THROUGHPUT_NUM_REQUESTS
-    args.seed = 0
-    args.disable_detokenize = True
-
-    with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
-        tmp_path = f.name
-    args.output_json = tmp_path
-
-    throughput_main(args)
-
-    with open(tmp_path) as f:
-        results = json.load(f)
-    q.put(results["tokens_per_second"])
-
-
-def _run_throughput_benchmark(
-    pin: bool,
-    engine_args_kwargs: dict,
-    v2_mode: bool = False,
-) -> float:
-    ctx = multiprocessing.get_context("spawn")
-    q = ctx.Queue()
-    p = ctx.Process(
-        target=_throughput_worker,
-        args=(pin, engine_args_kwargs, q, v2_mode),
-    )
-    p.start()
-    p.join()
-    if p.exitcode != 0:
-        raise RuntimeError(
-            f"Throughput benchmark subprocess (pin={pin}) exited with code {p.exitcode}"
-        )
-    return q.get()
-
-
-def _latency_worker(
-    pin: bool,
-    engine_args_kwargs: dict,
-    q: "multiprocessing.Queue[dict]",
-    v2_mode: bool = False,
-) -> None:
-    """Run latency benchmark in a fresh spawn subprocess.
-
-    Follows latency.py methodology: fixed batch of dummy token IDs, warmup
-    iterations to reach steady state, then timed iterations reduced to avg
-    and percentiles.  Results are written to a temp JSON file by latency_main
-    and forwarded through the queue.
-    """
-    import vllm.utils.platform_utils as pu
-    from vllm.platforms import current_platform
-
-    pu.is_pin_memory_available.cache_clear()
-    pu.is_uva_available.cache_clear()
-    type(current_platform).is_pin_memory_available = classmethod(lambda cls: pin)
-    if v2_mode:
-        pu.is_uva_available = lambda: True
-
-    from vllm.benchmarks.latency import add_cli_args
-    from vllm.benchmarks.latency import main as latency_main
-
-    parser = argparse.ArgumentParser()
-    add_cli_args(parser)
-    args = parser.parse_args([])
-
-    for key, val in engine_args_kwargs.items():
-        setattr(args, key, val)
-    args.input_len = _LATENCY_INPUT_LEN
-    args.output_len = _LATENCY_OUTPUT_LEN
-    args.batch_size = _LATENCY_BATCH_SIZE
-    args.num_iters_warmup = _LATENCY_WARMUP_ITERS
-    args.num_iters = _LATENCY_BENCH_ITERS
-    args.profile = False
-    args.disable_detokenize = True
-
-    with tempfile.NamedTemporaryFile(mode="w", suffix=".json", delete=False) as f:
-        tmp_path = f.name
-    args.output_json = tmp_path
-
-    latency_main(args)
-
-    with open(tmp_path) as f:
-        results = json.load(f)
-    q.put(results)
-
-
-def _run_latency_benchmark(
-    pin: bool,
-    engine_args_kwargs: dict,
-    v2_mode: bool = False,
-) -> dict:
-    ctx = multiprocessing.get_context("spawn")
-    q = ctx.Queue()
-    p = ctx.Process(
-        target=_latency_worker,
-        args=(pin, engine_args_kwargs, q, v2_mode),
-    )
-    p.start()
-    p.join()
-    if p.exitcode != 0:
-        raise RuntimeError(
-            f"Latency benchmark subprocess (pin={pin}) exited with code {p.exitcode}"
-        )
-    return q.get()
-
-
-@pytest.mark.parametrize(
-    "test_v2_runner",
-    [
-        pytest.param(False, id="v1"),
-        pytest.param(True, id="v2"),
-    ],
-)
-class TestPinnedMemory:
-    """Verify pinned memory yields >= throughput vs unpinned via real vLLM inference."""
-
-    def test_throughput(self, monkeypatch, test_v2_runner, model, max_model_len):
-        """Benchmark throughput with pin_memory forced on then off.
-
-        Delegates to vllm/benchmarks/throughput.py main() with the random
-        dataset.  Each condition runs in an isolated spawn subprocess so both
-        start from a cold CUDA context, giving an unbiased comparison.
-        """
-        monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
-        monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if test_v2_runner else "0")
-
-        engine_args_kwargs = dict(
-            model=model,
-            gpu_memory_utilization=0.88,
-            max_model_len=max_model_len,
-            enable_prefix_caching=False,
-        )
-
-        _skip_if_pin_memory_not_available(engine_args_kwargs)
-
-        unpinned_tps = _run_throughput_benchmark(
-            False, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-        pinned_tps = _run_throughput_benchmark(
-            True, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-
-        pct_diff = (pinned_tps - unpinned_tps) / unpinned_tps * 100
-        runner = "v2" if test_v2_runner else "v1"
-        print(
-            f"\n=== Throughput results ({runner} runner, {model}) ==="
-            f"\npin_memory=True:  {pinned_tps:.1f} tok/s"
-            f"\npin_memory=False: {unpinned_tps:.1f} tok/s"
-            f"\nDifference: {pct_diff:+.1f}% (pinned vs unpinned)"
-        )
-
-        assert pinned_tps >= unpinned_tps * _THROUGHPUT_TOLERANCE, (
-            f"Pinned throughput ({pinned_tps:.1f} tok/s) fell more than "
-            f"{(1.0 - _THROUGHPUT_TOLERANCE) * 100:.1f}% below "
-            f"unpinned ({unpinned_tps:.1f} tok/s)."
-        )
-
-    def test_latency(self, monkeypatch, test_v2_runner, model, max_model_len):
-        """Benchmark per-batch latency with pin_memory forced on then off.
-
-        Follows vllm/benchmarks/latency.py: fixed dummy-token batch, warmup
-        iterations to reach steady state, then timed iterations reduced to avg
-        and percentiles.  Subprocesses run serially so each gets a cold CUDA
-        context without GPU memory pressure from the other run.
-        """
-        monkeypatch.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
-        monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1" if test_v2_runner else "0")
-
-        engine_args_kwargs = dict(
-            model=model,
-            gpu_memory_utilization=0.88,
-            max_model_len=max_model_len,
-            enable_prefix_caching=False,
-        )
-
-        _skip_if_pin_memory_not_available(engine_args_kwargs)
-
-        unpinned = _run_latency_benchmark(
-            False, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-        pinned = _run_latency_benchmark(
-            True, engine_args_kwargs, v2_mode=test_v2_runner
-        )
-
-        pct_diff = (
-            (pinned["avg_latency"] - unpinned["avg_latency"])
-            / unpinned["avg_latency"]
-            * 100
-        )
-        runner = "v2" if test_v2_runner else "v1"
-        print(
-            f"\n=== Latency results ({runner} runner, {model}) ==="
-            f"\npin_memory=True:  avg={pinned['avg_latency']:.3f}s"
-            f"  p50={pinned['percentiles']['50']:.3f}s"
-            f"  p99={pinned['percentiles']['99']:.3f}s"
-            f"\npin_memory=False: avg={unpinned['avg_latency']:.3f}s"
-            f"  p50={unpinned['percentiles']['50']:.3f}s"
-            f"  p99={unpinned['percentiles']['99']:.3f}s"
-            f"\nDifference: {pct_diff:+.1f}% (pinned vs unpinned)"
-        )
-
-        assert pinned["avg_latency"] <= unpinned["avg_latency"] * _LATENCY_TOLERANCE, (
-            f"Pinned avg latency ({pinned['avg_latency']:.3f}s) exceeded "
-            f"unpinned ({unpinned['avg_latency']:.3f}s) by more than "
-            f"{(_LATENCY_TOLERANCE - 1.0) * 100:.1f}%."
-        )
-
-
-if __name__ == "__main__":
-    _parser = argparse.ArgumentParser(add_help=False)
-    _parser.add_argument("--model", default=_DEFAULT_MODEL)
-    _parser.add_argument("--max-model-len", type=int, default=_DEFAULT_MAX_MODEL_LEN)
-    _, _remaining = _parser.parse_known_args()
-    sys.exit(pytest.main([__file__] + _remaining))
@@ -33,6 +33,7 @@ from vllm.distributed.device_communicators.custom_all_reduce import CustomAllred
 from vllm.distributed.device_communicators.flashinfer_all_reduce import (
    FlashInferAllReduce,
 )
+from vllm.distributed.device_communicators.push_all_reduce import PushAllReduce
 from vllm.distributed.device_communicators.pynccl import (
    PyNcclCommunicator,
    register_nccl_symmetric_ops,
@@ -80,6 +81,7 @@ class CommunicatorBenchmark:

        # Initialize communicators
        self.custom_allreduce = None
+        self.push_ar_comm = None
        self.pynccl_comm = None
        self.symm_mem_comm = None
        self.symm_mem_comm_multimem = None
@@ -106,6 +108,23 @@ class CommunicatorBenchmark:
            )
            self.custom_allreduce = None

+        try:
+            self.push_ar_comm = PushAllReduce(
+                group=self.cpu_group,
+                device=self.device,
+                max_size=self.max_size_override,
+            )
+            if not self.push_ar_comm.disabled:
+                logger.info("Rank %s: PushAllReduce initialized", self.rank)
+            else:
+                logger.info("Rank %s: PushAllReduce disabled", self.rank)
+                self.push_ar_comm = None
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize PushAllReduce: %s", self.rank, e
+            )
+            self.push_ar_comm = None
+
        try:
            self.pynccl_comm = PyNcclCommunicator(
                group=self.cpu_group, device=self.device
@@ -216,6 +235,19 @@ class CommunicatorBenchmark:
                )
            )

+        if self.push_ar_comm is not None:
+            comm = self.push_ar_comm
+            communicators.append(
+                (
+                    "push_ar",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use(t),
+                    comm.capture(),
+                    {},
+                    None,
+                )
+            )
+
        if self.pynccl_comm is not None:
            comm = self.pynccl_comm
            communicators.append(
@@ -17,7 +17,7 @@ from vllm.model_executor.layers.fused_moe.fused_flydsl_moe import fused_flydsl_m
 from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import (  # noqa: E501
    compressed_tensors_moe_w4a16_flydsl,
 )
-from vllm.utils.platform_utils import get_device_name_as_file_name
+from vllm.platforms import current_platform

 RoutingBuffers = tuple[
    torch.Tensor,  # sorted_token_ids
@@ -259,7 +259,7 @@ def tune_flydsl_moe_w4a16(
                            )
                            us_best = us
                            tuned_config[str(num_tokens)] = tile_config
-            device_name = get_device_name_as_file_name()
+            device_name = current_platform.get_device_name().replace(" ", "_")
            tuned_config_file_name = (
                f"E={num_experts},N={inter_dim},device_name={device_name},"
                f"dtype=int4_w4a16,backend=flydsl.json"
@@ -80,17 +80,13 @@ _FI_MAX_SIZES = {
    2: 64 * MiB,  # 64MB
    4: 64 * MiB,  # 64MB
    8: 64 * MiB,  # 64MB
-    16: 64 * MiB,  # 64MB (multi-node)
 }

 # Global workspace tensors for FlashInfer (keyed by backend name)
 _FI_WORKSPACES: dict = {}

-# Backends to benchmark. trtllm is single-node only and can hang cross-node, so
-# multi-node sweeps can restrict to mnnvl via FI_BACKENDS=mnnvl.
-FLASHINFER_BACKENDS = [
-    b for b in os.environ.get("FI_BACKENDS", "trtllm,mnnvl").split(",") if b
-]
+# Backends to benchmark
+FLASHINFER_BACKENDS = ["trtllm", "mnnvl"]


 def setup_flashinfer_workspace(
@@ -999,10 +995,7 @@ def main():
    rank = int(os.environ["RANK"])
    world_size = int(os.environ["WORLD_SIZE"])

-    # Use LOCAL_RANK for the device so multi-node runs (global rank >= GPUs per
-    # node) map to a valid local GPU; falls back to global rank single-node.
-    local_rank = int(os.environ.get("LOCAL_RANK", rank))
-    device = torch.device(f"cuda:{local_rank}")
+    device = torch.device(f"cuda:{rank}")
    torch.accelerator.set_device_index(device)
    torch.set_default_device(device)

@@ -391,19 +391,16 @@ def get_configs_compute_bound(use_fp16, block_quant_shape) -> list[dict[str, int
        config = dict(zip(keys, config_values))
        configs.append(config)

-    # Drop configs incompatible with fp8 block quantization. A tile must align
-    # to the quant-block scale grid, i.e. tile and block must divide one
-    # another. The kernel indexes scales per element (offs_bn // group_n,
-    # k_start // group_k), so a tile narrower than the block (e.g. N=64 with
-    # block_n=128) is valid -- and often faster at small batch. An exact
-    # multiple was required before, which dropped those smaller tiles entirely.
+    # Remove configs that are not compatible with fp8 block quantization
+    # BLOCK_SIZE_K must be a multiple of block_k
+    # BLOCK_SIZE_N must be a multiple of block_n
    if block_quant_shape is not None and not use_fp16:
        block_n, block_k = block_quant_shape[0], block_quant_shape[1]
        for config in configs[:]:
-            bn, bk = config["BLOCK_SIZE_N"], config["BLOCK_SIZE_K"]
-            n_aligned = bn % block_n == 0 or block_n % bn == 0
-            k_aligned = bk % block_k == 0 or block_k % bk == 0
-            if not (n_aligned and k_aligned):
+            if (
+                config["BLOCK_SIZE_K"] % block_k != 0
+                or config["BLOCK_SIZE_N"] % block_n != 0
+            ):
                configs.remove(config)
    return configs

@@ -19,11 +19,13 @@ from vllm.utils.torch_utils import (
 logger = init_logger(__name__)

 NUM_BLOCKS = 128 * 1024
+PARTITION_SIZE = 512
 PARTITION_SIZE_ROCM = 256


@torch.inference_mode()
 def main(
+    version: str,
    num_seqs: int,
    seq_len: int,
    num_query_heads: int,
@@ -80,20 +82,27 @@ def main(

    # Prepare for the paged attention kernel.
    output = torch.empty_like(query)
-    num_partitions = (max_seq_len + PARTITION_SIZE_ROCM - 1) // PARTITION_SIZE_ROCM
-    tmp_output = torch.empty(
-        size=(num_seqs, num_query_heads, num_partitions, head_size),
-        dtype=output.dtype,
-        device=output.device,
-    )
-    exp_sums = torch.empty(
-        size=(num_seqs, num_query_heads, num_partitions),
-        dtype=torch.float32,
-        device=output.device,
-    )
-    max_logits = torch.empty_like(exp_sums)
+    if version == "v2":
+        if current_platform.is_rocm():
+            global PARTITION_SIZE
+            if not args.custom_paged_attn and not current_platform.is_navi():
+                PARTITION_SIZE = 1024
+            else:
+                PARTITION_SIZE = PARTITION_SIZE_ROCM
+        num_partitions = (max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE
+        tmp_output = torch.empty(
+            size=(num_seqs, num_query_heads, num_partitions, head_size),
+            dtype=output.dtype,
+            device=output.device,
+        )
+        exp_sums = torch.empty(
+            size=(num_seqs, num_query_heads, num_partitions),
+            dtype=torch.float32,
+            device=output.device,
+        )
+        max_logits = torch.empty_like(exp_sums)

-    def run_benchmark(num_iters: int, profile: bool = False) -> float:
+    def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
        torch.accelerator.synchronize()
        if profile:
            torch.cuda.cudart().cudaProfilerStart()
@@ -103,26 +112,67 @@ def main(
        k_scale = v_scale = torch.tensor(1.0, dtype=torch.float32, device=device)

        for _ in range(num_iters):
-            ops.paged_attention_rocm(
-                output,
-                exp_sums,
-                max_logits,
-                tmp_output,
-                query,
-                key_cache,
-                value_cache,
-                num_kv_heads,
-                scale,
-                block_tables,
-                seq_lens,
-                None,
-                block_size,
-                max_seq_len,
-                alibi_slopes,
-                kv_cache_dtype,
-                k_scale,
-                v_scale,
-            )
+            if version == "v1":
+                ops.paged_attention_v1(
+                    output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    num_kv_heads,
+                    scale,
+                    block_tables,
+                    seq_lens,
+                    block_size,
+                    max_seq_len,
+                    alibi_slopes,
+                    kv_cache_dtype,
+                    k_scale,
+                    v_scale,
+                )
+            elif version == "v2":
+                if not args.custom_paged_attn:
+                    ops.paged_attention_v2(
+                        output,
+                        exp_sums,
+                        max_logits,
+                        tmp_output,
+                        query,
+                        key_cache,
+                        value_cache,
+                        num_kv_heads,
+                        scale,
+                        block_tables,
+                        seq_lens,
+                        block_size,
+                        max_seq_len,
+                        alibi_slopes,
+                        kv_cache_dtype,
+                        k_scale,
+                        v_scale,
+                    )
+                else:
+                    ops.paged_attention_rocm(
+                        output,
+                        exp_sums,
+                        max_logits,
+                        tmp_output,
+                        query,
+                        key_cache,
+                        value_cache,
+                        num_kv_heads,
+                        scale,
+                        block_tables,
+                        seq_lens,
+                        None,
+                        block_size,
+                        max_seq_len,
+                        alibi_slopes,
+                        kv_cache_dtype,
+                        k_scale,
+                        v_scale,
+                    )
+            else:
+                raise ValueError(f"Invalid version: {version}")
        torch.accelerator.synchronize()

        end_time = time.perf_counter()
@@ -132,6 +182,7 @@ def main(

    # Warmup.
    print("Warming up...")
+    run_benchmark = run_cuda_benchmark
    run_benchmark(num_iters=3, profile=False)

    # Benchmark.
@@ -144,13 +195,12 @@ def main(

 if __name__ == "__main__":
    logger.warning(
-        "This script benchmarks the ROCm paged attention kernel. "
+        "This script benchmarks the paged attention kernel. "
        "By default this is no longer used in vLLM inference."
    )
-    if not current_platform.is_rocm():
-        raise RuntimeError("This benchmark requires the ROCm platform.")

    parser = FlexibleArgumentParser(description="Benchmark the paged attention kernel.")
+    parser.add_argument("--version", type=str, choices=["v1", "v2"], default="v2")
    parser.add_argument("--batch-size", type=int, default=8)
    parser.add_argument("--seq-len", type=int, default=4096)
    parser.add_argument("--num-query-heads", type=int, default=64)
@@ -158,7 +208,7 @@ if __name__ == "__main__":
    parser.add_argument(
        "--head-size",
        type=int,
-        choices=[64, 128],
+        choices=[64, 80, 96, 112, 120, 128, 192, 256],
        default=128,
    )
    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
@@ -174,7 +224,11 @@ if __name__ == "__main__":
        choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"],
        default="auto",
        help="Data type for kv cache storage. If 'auto', will use model "
-        "data type. ROCm (AMD GPU) supports fp8 (=fp8_e4m3)",
+        "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
+        "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)",
+    )
+    parser.add_argument(
+        "--custom-paged-attn", action="store_true", help="Use custom paged attention"
    )
    args = parser.parse_args()
    print(args)
@@ -182,6 +236,7 @@ if __name__ == "__main__":
    if args.num_query_heads % args.num_kv_heads != 0:
        raise ValueError("num_query_heads must be divisible by num_kv_heads")
    main(
+        version=args.version,
        num_seqs=args.batch_size,
        seq_len=args.seq_len,
        num_query_heads=args.num_query_heads,
@@ -19,7 +19,6 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton
 from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.platform_utils import get_device_name_as_file_name

 mp.set_start_method("spawn", force=True)

@@ -265,7 +264,7 @@ def save_configs(
    input_type="fp8",
 ) -> None:
    os.makedirs(save_path, exist_ok=True)
-    device_name = get_device_name_as_file_name()
+    device_name = current_platform.get_device_name().replace(" ", "_")
    json_file_name = (
        f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
        f"block_shape=[{block_n},{block_k}].json"
@@ -7,7 +7,6 @@ import time
 import numpy as np
 import torch

-from vllm.platforms import CpuArchEnum, current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm.utils.torch_utils import set_random_seed

@@ -15,15 +14,17 @@ from vllm.utils.torch_utils import set_random_seed
 try:
    from vllm._custom_ops import cpu_fused_moe, cpu_prepack_moe_weight
 except (ImportError, AttributeError) as e:
+    print("ERROR: CPU fused MoE operations are not available on this platform.")
+    print("This benchmark requires x86 CPU with proper vLLM CPU extensions compiled.")
+    print(
+        "The cpu_fused_moe kernel is typically available on Linux x86_64 "
+        "with AVX2/AVX512."
+    )
    print(f"Import error: {e}")
    sys.exit(1)

 # ISA selection following test_cpu_fused_moe.py pattern
-ISA_CHOICES = ["vec"]
-if torch.cpu._is_amx_tile_supported():
-    ISA_CHOICES.append("amx")
-if current_platform.get_cpu_architecture() == CpuArchEnum.ARM:
-    ISA_CHOICES.append("neon")
+ISA_CHOICES = ["amx", "vec"] if torch.cpu._is_amx_tile_supported() else ["vec"]


@torch.inference_mode()
@@ -144,7 +145,7 @@ if __name__ == "__main__":
        "--isa",
        type=str,
        choices=ISA_CHOICES,
-        default="vec",
+        default=ISA_CHOICES[0],
        help=f"ISA to use (available: {ISA_CHOICES})",
    )
    parser.add_argument("--seed", type=int, default=0)
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Alexander MatveevandClaude Opus 4.6	fa6b6a83ec	Apply pre-commit formatting and use torch.accelerator API - Apply ruff, clang-format formatting fixes - Replace torch.cuda.set_device/device_count/synchronize with torch.accelerator equivalents per project convention Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com> Signed-off-by: <>	2026-06-15 18:21:59 -04:00
Alexander MatveevandClaude Opus 4.6	fd44100bb0	Address review: add push_ar to benchmark and backend logging - Add PushAllReduce to benchmark_device_communicators.py for comparing against other allreduce implementations - Add PUSH_AR to _log_all_reduce_backend_selection in cuda_communicator.py for visibility in dispatch logging Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00
Alexander MatveevandClaude Opus 4.6	bc50e5fc2e	Address review: use registered envs.VLLM_DISABLE_PUSH_ALLREDUCE Replace direct os.environ.get(_DISABLE_ENV_VAR) == "1" check with envs.VLLM_DISABLE_PUSH_ALLREDUCE to use the centrally registered env var from envs.py, which provides validation and caching. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00
Alexander MatveevandClaude Opus 4.6	625bea3939	Address review: register VLLM_DISABLE_PUSH_ALLREDUCE in envs.py Register the push allreduce feature toggle env var in the central envs.py registry so it is validated on startup and follows the standard vllm env var pattern. Default is False (push allreduce enabled); set to 1 to disable. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00
Alexander MatveevandClaude Opus 4.6	17488a743f	Address review: add architecture-specific threshold selection The push threshold map was labeled as sm100-specific but applied unconditionally to all architectures. Now: - PUSH_THRESHOLD_SM100 is only used on Blackwell (compute capability 10.x) - PUSH_THRESHOLD_DEFAULT provides conservative 512 KB thresholds for architectures without tuned values - _THRESHOLD_BY_ARCH maps GPU major compute capability to threshold tables - A log message is emitted when falling back to conservative defaults Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00
Alexander MatveevandClaude Opus 4.6	54b43935a8	Address review: add type annotation and cleanup for push_ar_comm - Add PushAllReduce \| None type annotation on push_ar_comm to be consistent with other communicator fields (ca_comm, qr_comm, etc.) - Add push_ar_comm.close() + None assignment in destroy() method to match the cleanup pattern for other communicators - Add lazy import of PushAllReduce alongside other communicator imports Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00
Alexander MatveevandClaude Opus 4.6	e0601e1b94	Address review: bind test sockets to localhost instead of all interfaces Fix CodeQL security warning by binding test helper sockets to "localhost" instead of "" (all interfaces). These sockets are only used for finding a free port for torch distributed init in tests. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00
Alexander MatveevandClaude Opus 4.6	c55216ebbf	Address review: add CUDA error checking and runtime buffer overflow guard - Add PUSH_AR_CUDACHECK macro wrapping all CUDA API calls (cudaGetDevice, cudaDeviceGetAttribute, cudaMalloc, cudaMemset, cudaIpcGetMemHandle, cudaIpcOpenMemHandle) to match the CUDACHECK pattern in custom_all_reduce.cuh - Replace assert(input_bytes <= push_buffer_bytes_) with a runtime std::runtime_error check that is not compiled out under -DNDEBUG - Add #include <stdexcept> and #include <string> for the runtime check Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00
Alexander MatveevandAlexander Matveev	5fd9bd4f04	perf: add push-based allreduce for small tensor reductions Port SGLang's push-based 2-buffer allreduce protocol into vLLM as a new communicator backend for small-message reductions. The push protocol eliminates the two explicit cross-GPU NVLink barrier round-trips used by the existing barrier-based CustomAllreduce, replacing them with a sentinel-based data arrival detection mechanism and double-buffered epoch alternation. Key advantages over the barrier-based approach: - Zero barriers: data arrival IS the synchronization (positive-zero sentinel) - Single NVLink round-trip instead of two barrier exchanges + remote reads - All SMs active (SM_count CTAs vs 2 CTAs) for higher NVLink bandwidth - No cudaMemcpy to IPC staging buffer in eager mode - PDL (griddepcontrol) support for kernel overlap on sm_90+ The new PushAllReduce is inserted in the CudaCommunicator dispatch chain above the existing CustomAllreduce for messages below a size threshold (~720 KB at TP=8). Larger messages continue to use the barrier-based path. The existing CustomAllreduce code is not modified. Measured results on DeepSeek-V4-Pro (61 layers, TP=8, 8x NVIDIA B200, BS=1, decode with ISL=4, OSL=33024): - Throughput: +2.14% (84.06 vs 82.30 tokens/s) - TPOT: -2.09% (11.90 vs 12.15 ms/token) Correctness verified via lm_eval gsm8k 5-shot with no regression (exact_match delta within statistical noise). The feature can be disabled at runtime via VLLM_DISABLE_PUSH_ALLREDUCE=1 to fall back to the barrier-based path. Signed-off-by: Alexander Matveev <amatveev@redhat.com> Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>	2026-06-15 18:09:24 -04:00