[Model] Add Inkling multi-depth MTP support [5/N]

Extend the merged Inkling MTP=1 implementation to multiple checkpoint depths with multi-step speculative decoding and KV-cache plumbing. Co-authored-by: Bugen Zhao <i@bugenzhao.com> Co-authored-by: Giancarlo Delfin <32987265+TheEpicDolphin@users.noreply.github.com> Co-authored-by: Isotr0py <Isotr0py@outlook.com> Co-authored-by: Isotr0py <mozf@inferact.ai> Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai> Co-authored-by: Roger Wang <hey@rogerw.io> Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: OpenAI Codex <codex@openai.com> Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
[Model] Add Inkling MTP=1 support [3/N] (#48869 )
2026-07-16 21:17:29 +00:00 · 2026-07-16 13:27:21 -07:00 · 2026-07-16 20:02:27 +00:00 · 2026-07-16 19:57:45 +00:00 · 2026-07-16 14:58:47 -04:00 · 2026-07-16 11:26:52 -07:00
1183 changed files with 66614 additions and 11192 deletions
@@ -8,6 +8,7 @@ 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"
@@ -1,5 +1,30 @@
 group: Hardware - AMD Build
+
+# ROCm image flow:
+# 1. Refresh the long-lived ROCm base image only when Dockerfile.rocm_base changes.
+# 2. Build ci_base from either the stable base or the freshly refreshed base.
+# 3. Build the per-commit ROCm CI image and smoke-test it before GPU jobs run.
 steps:
+  - label: "AMD: :docker: refresh ROCm base"
+    key: refresh-rocm-base-amd
+    depends_on: []
+    device: amd_cpu
+    no_plugin: true
+    commands:
+      - bash .buildkite/scripts/rocm/refresh-base-image.sh
+    env:
+      DOCKER_BUILDKIT: "1"
+      BUILDKIT_PROGRESS: "tty"
+      TERM: "xterm-256color"
+    retry:
+      automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
+        - exit_status: -1  # Agent was lost
+          limit: 1
+        - exit_status: -10  # Agent was lost
+          limit: 1
+
  # Ensure ci_base is up-to-date before building the test image.
  # Compares a content hash of ci_base-affecting files against the remote
  # image label. If hashes match the build is skipped (< 30 s); if they
@@ -7,19 +32,24 @@ steps:
  - label: "AMD: :docker: ensure ci_base"
    key: ensure-ci-base-amd
    soft_fail: false
-    depends_on: []
+    depends_on:
+      - refresh-rocm-base-amd
    device: amd_cpu
    no_plugin: true
    commands:
-      - bash .buildkite/scripts/ci-bake-rocm.sh ci-base-rocm-ci-with-deps
+      - bash .buildkite/scripts/rocm/build-ci-base.sh
    env:
      DOCKER_BUILDKIT: "1"
+      BUILDKIT_PROGRESS: "tty"
+      TERM: "xterm-256color"
      VLLM_BAKE_FILE: "docker/docker-bake-rocm.hcl"
      PYTORCH_ROCM_ARCH: "gfx90a;gfx942;gfx950"
      REMOTE_VLLM: "1"
      VLLM_BRANCH: "$BUILDKITE_COMMIT"
    retry:
      automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
        - exit_status: -1  # Agent was lost
          limit: 1
        - exit_status: -10  # Agent was lost
@@ -33,35 +63,12 @@ steps:
    device: amd_cpu
    no_plugin: true
    commands:
-      - |
-        if [[ "${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:${BUILDKITE_COMMIT}" -ec '
-          if [ ! -d /vllm-workspace ]; then echo Missing directory: /vllm-workspace >&2; exit 1; fi
-          if [ ! -d /vllm-workspace/tests ]; then echo Missing directory: /vllm-workspace/tests >&2; exit 1; fi
-          if [ ! -d /vllm-workspace/src/vllm ]; then echo Missing directory: /vllm-workspace/src/vllm >&2; exit 1; fi
-          if [ ! -x /vllm-workspace/src/vllm/vllm-rs ]; then echo Missing executable: /vllm-workspace/src/vllm/vllm-rs >&2; exit 1; fi
-          command -v python3
-          command -v uv
-          command -v pytest
-          if ! command -v amd-smi >/dev/null 2>&1 && ! command -v rocminfo >/dev/null 2>&1; then
-            echo No ROCm CLI found in image >&2
-            exit 1
-          fi
-          python3 - <<PY
-        import torch, vllm
-        print(torch.__version__)
-        print(vllm.__version__)
-        PY
-          echo AMD image smoke OK
-        '
+      - bash .buildkite/scripts/rocm/build-test-image.sh
+      - bash .buildkite/scripts/rocm/smoke-test-image.sh
    env:
      DOCKER_BUILDKIT: "1"
+      BUILDKIT_PROGRESS: "tty"
+      TERM: "xterm-256color"
      VLLM_BAKE_FILE: "docker/docker-bake-rocm.hcl"
      PYTORCH_ROCM_ARCH: "gfx90a;gfx942;gfx950"
      IMAGE_TAG: "rocm/vllm-ci:$BUILDKITE_COMMIT"
@@ -69,6 +76,8 @@ steps:
      VLLM_BRANCH: "$BUILDKITE_COMMIT"
    retry:
      automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
        - exit_status: -1  # Agent was lost
          limit: 1
        - exit_status: -10  # Agent was lost
@@ -141,9 +141,22 @@ steps:
  commands:
    - |
      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
-      pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
+      pytest -x -v -s tests/models/multimodal/generation --ignore=tests/models/multimodal/generation/test_pixtral.py --ignore=tests/models/multimodal/generation/test_qwen2_5_vl.py -m cpu_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB"
  parallelism: 4

+- label: CPU-Qwen2.5-VL Multimodal Tests
+  depends_on: []
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  # - vllm/
+  - vllm/model_executor/layers/rotary_embedding
+  - tests/models/multimodal/generation/
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 40m "
+      VLLM_CI_ENV=0 pytest -x -v -s tests/models/multimodal/generation/test_qwen2_5_vl.py"
+
 - label: "Arm CPU Test"
  depends_on: []
  soft_fail: false
@@ -18,7 +18,7 @@ steps:
  - label: "XPU example Test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 50
    optional: true
    device: intel_gpu
    agent_tags:
@@ -39,7 +39,7 @@ steps:
  - label: "XPU V1 test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 70
    optional: true
    device: intel_gpu
    agent_tags:
@@ -60,7 +60,7 @@ steps:
  - label: "XPU server test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 45
    optional: true
    device: intel_gpu
    agent_tags:
@@ -79,12 +79,18 @@ 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"
@@ -171,6 +177,18 @@ 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}"
@@ -254,3 +272,5 @@ 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
@@ -9,30 +9,31 @@ 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 [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64) ]]; then
-  echo "Image not found, proceeding with build..."
-else
+if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
  echo "Image found"
-  exit 0
+else
+  echo "Image not found, proceeding with build..."
+  # build for arm64 GPU targets: Grace/GH200 (sm_90),
+  # Blackwell/Thor (sm_100/sm_103/sm_110), and DGX Spark/GB10
+  # (sm_121, family-covered by 12.0 under CUDA 13)
+  docker build --file docker/Dockerfile \
+    --platform linux/arm64 \
+    --build-arg max_jobs=16 \
+    --build-arg nvcc_threads=4 \
+    --build-arg torch_cuda_arch_list="9.0 10.0 11.0 12.0" \
+    --build-arg USE_SCCACHE=1 \
+    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+    --tag "$IMAGE" \
+    --target test \
+    --progress plain .
+  # push
+  docker push "$IMAGE"
 fi

-# build for arm64 GPU targets: Grace/GH200 (sm_90) and DGX Spark/GB10
-# (sm_121, family-covered by 12.0 under CUDA 13)
-docker build --file docker/Dockerfile \
-  --platform linux/arm64 \
-  --build-arg max_jobs=16 \
-  --build-arg nvcc_threads=4 \
-  --build-arg torch_cuda_arch_list="9.0 12.0" \
-  --build-arg USE_SCCACHE=1 \
-  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64 \
-  --target test \
-  --progress plain .
-
-# push
-docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64
+.buildkite/scripts/annotate-image-build.sh "$IMAGE"
@@ -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 [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
-  echo "Image not found, proceeding with build..."
-else
+if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
  echo "Image found"
-  exit 0
+else
+  echo "Image not found, proceeding with build..."
+  # build
+  docker build --file docker/Dockerfile.cpu \
+    --build-arg max_jobs=16 \
+    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+    --build-arg VLLM_CPU_X86=true \
+    --tag "$IMAGE" \
+    --target vllm-test \
+    --progress plain .
+  # push
+  docker push "$IMAGE"
 fi

-# 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":"$BUILDKITE_COMMIT"-cpu
+.buildkite/scripts/annotate-image-build.sh "$IMAGE"
@@ -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 [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-arm64-cpu) ]]; then
-  echo "Image not found, proceeding with build..."
-else
+if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
  echo "Image found"
-  exit 0
+else
+  echo "Image not found, proceeding with build..."
+  # build
+  docker build --file docker/Dockerfile.cpu \
+    --build-arg max_jobs=16 \
+    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+    --tag "$IMAGE" \
+    --target vllm-test \
+    --progress plain .
+  # push
+  docker push "$IMAGE"
 fi

-# 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":"$BUILDKITE_COMMIT"-arm64-cpu
+.buildkite/scripts/annotate-image-build.sh "$IMAGE"
@@ -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 [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu) ]]; then
-  echo "Image not found, proceeding with build..."
-else
+if docker manifest inspect "$IMAGE" >/dev/null 2>&1; then
  echo "Image found"
-  exit 0
+else
+  echo "Image not found, proceeding with build..."
+  # build
+  docker build \
+    --file tests/pytorch_ci_hud_benchmark/Dockerfile.hpu \
+    --build-arg max_jobs=16 \
+    --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+    --tag "$IMAGE" \
+    --progress plain \
+    https://github.com/vllm-project/vllm-gaudi.git
+  # push
+  docker push "$IMAGE"
 fi

-# 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":"$BUILDKITE_COMMIT"-hpu
+.buildkite/scripts/annotate-image-build.sh "$IMAGE"
@@ -40,6 +40,7 @@ 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..."
@@ -66,3 +67,5 @@ 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 "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-xpu &> /dev/null; then
-  echo "Image not found, proceeding with build..."
-else
+if docker manifest inspect "$IMAGE" &> /dev/null; then
  echo "Image found"
-  exit 0
+else
+  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"
 fi

-# 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":"$BUILDKITE_COMMIT"-xpu
+.buildkite/scripts/annotate-image-build.sh "$IMAGE"
@@ -3,7 +3,7 @@ depends_on:
  - image-build-xpu
 steps:
 - label: XPU Sleep Mode
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
@@ -86,7 +86,7 @@ steps:
      pytest -v -s lora/test_punica_ops.py::test_add_lora_fused_moe_early_exit'

 - label: LoRA Punica FP8/XPU Ops
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
  device: intel_gpu
  agent_tags:
    label: production
@@ -3,7 +3,7 @@ depends_on:
  - image-build-xpu
 steps:
 - label: V1 Core + KV + Metrics
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: intel_gpu
  agent_tags:
    label: production
@@ -33,7 +33,7 @@ steps:
      pytest -v -s v1/executor'

 - label: V1 Sample + Logits
-  timeout_in_minutes: 30
+  timeout_in_minutes: 90
  device: intel_gpu
  agent_tags:
    label: production
@@ -72,9 +72,7 @@ steps:
      pytest -v -s v1/test_oracle.py &&
      pytest -v -s v1/test_request.py &&
      pytest -v -s v1/test_outputs.py &&
-      pytest -v -s v1/sample/test_topk_topp_sampler.py &&
-      pytest -v -s v1/sample/test_logprobs.py &&
-      pytest -v -s v1/sample/test_logprobs_e2e.py'
+      pytest -v -s v1/sample'

 - label: Basic Models Tests (Initialization)
  timeout_in_minutes: 60
@@ -154,7 +152,7 @@ steps:

 - label: Regression
  key: regression
-  timeout_in_minutes: 30
+  timeout_in_minutes: 50
  device: intel_gpu
  agent_tags:
    label: production
@@ -188,7 +186,7 @@ steps:

 - label: Metrics, Tracing (2 GPUs)
  key: metrics-tracing-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  num_devices: 2
  device: intel_gpu
  agent_tags:
@@ -224,7 +222,7 @@ steps:

 - label: Async Engine, Inputs, Utils, Worker
  key: async-engine-inputs-utils-worker
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
  device: intel_gpu
  agent_tags:
    label: production
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Model Tests (2 GPUs)
  key: distributed-model-tests-2-gpus
-  timeout_in_minutes: 50
+  timeout_in_minutes: 65
  device: intel_gpu
  agent_tags:
    label: production
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
  key: multi-modal-models-standard-1-qwen2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 70
  device: intel_gpu
  agent_tags:
    label: production
@@ -29,7 +29,7 @@ steps:

 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
  key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 45
+  timeout_in_minutes: 70
  device: intel_gpu
  agent_tags:
    label: production
@@ -52,7 +52,7 @@ steps:

 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
  key: multi-modal-models-standard-3-llava-qwen2-vl
-  timeout_in_minutes: 45
+  timeout_in_minutes: 65
  device: intel_gpu
  agent_tags:
    label: production
@@ -100,7 +100,7 @@ steps:

 - label: Multi-Modal Processor # 44min
  key: multi-modal-processor
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
  device: intel_gpu
  agent_tags:
    label: production
@@ -17,7 +17,7 @@ steps:
  - label: "XPU example Test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 50
    device: intel_gpu
    agent_tags:
      label: production
@@ -76,7 +76,7 @@ steps:
  - label: "XPU V1 test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 70
    device: intel_gpu
    agent_tags:
      label: production
@@ -99,12 +99,13 @@ 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/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]" &&
        pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py &&
        pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py'
  - label: "XPU server test"
    depends_on:
      - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 45
    device: intel_gpu
    agent_tags:
      label: production
@@ -1,6 +1,7 @@
 # For hf script, without -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh -m meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 -l 250 -t 8 -f 5
 model_name: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
+rocm_safetensors_load_strategy: lazy
 required_gpu_arch:
  - gfx942
  - gfx950
@@ -72,6 +72,11 @@ def launch_lm_eval(eval_config, tp_size):
    if moe_backend is not None:
        model_args += f"moe_backend={moe_backend},"

+    if current_platform.is_rocm():
+        rocm_load_strategy = eval_config.get("rocm_safetensors_load_strategy")
+        if rocm_load_strategy is not None:
+            model_args += f"safetensors_load_strategy={rocm_load_strategy},"
+
    env_vars = eval_config.get("env_vars", None)
    with scoped_env_vars(env_vars):
        results = lm_eval.simple_evaluate(
@@ -44,7 +44,7 @@ steps:
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)" release-wheels'
        env:
          DOCKER_BUILDKIT: "1"

@@ -58,7 +58,7 @@ steps:
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)" release-wheels'
        env:
          DOCKER_BUILDKIT: "1"

@@ -72,10 +72,53 @@ steps:
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)" release-wheels'
        env:
          DOCKER_BUILDKIT: "1"

+      - label: "Build wheel - macOS arm64 - CPU"
+        depends_on: ~
+        id: build-wheel-macos-arm64-cpu
+        agents:
+          queue: macmini
+        timeout_in_minutes: 90
+        commands:
+          - "bash .buildkite/scripts/build-macos-wheel.sh"
+          - "mkdir -p artifacts/transfer"
+          - "basename artifacts/dist/*.whl > artifacts/transfer/wheel.name"
+          - "shasum -a 256 artifacts/dist/*.whl > artifacts/transfer/wheel.sha256"
+          - "split -b 1m artifacts/dist/*.whl artifacts/transfer/wheel.part."
+        env:
+          BUILDKITE_ARTIFACT_UPLOAD_CONCURRENCY: "1"
+          BUILDKITE_NO_HTTP2: "true"
+        artifact_paths:
+          - "artifacts/transfer/*"
+
+      - label: "Upload wheel - macOS arm64 - CPU"
+        depends_on: build-wheel-macos-arm64-cpu
+        id: upload-wheel-macos-arm64-cpu
+        agents:
+          queue: small_cpu_queue_release
+        timeout_in_minutes: 15
+        commands:
+          - 'buildkite-agent artifact download "artifacts/transfer/*" .'
+          - "mkdir -p artifacts/dist artifacts/reassembled"
+          - 'wheel_name=$$(cat artifacts/transfer/wheel.name)'
+          - '[[ "$$wheel_name" =~ ^[A-Za-z0-9][A-Za-z0-9._+-]*\.whl$$ && "$$wheel_name" != *..* ]]'
+          - 'expected_sha=$$(awk ''{print $$1}'' artifacts/transfer/wheel.sha256)'
+          - '[[ "$$expected_sha" =~ ^[0-9a-f]{64}$$ ]]'
+          - 'cat artifacts/transfer/wheel.part.* > artifacts/reassembled/wheel'
+          - 'actual_sha=$$(shasum -a 256 artifacts/reassembled/wheel | awk ''{print $$1}'')'
+          - '[[ "$$actual_sha" == "$$expected_sha" ]]'
+          - 'mv artifacts/reassembled/wheel "artifacts/dist/$$wheel_name"'
+          - "aws sts get-caller-identity"
+          - "VLLM_WHEEL_PLATFORM=macos bash .buildkite/scripts/upload-nightly-wheels.sh"
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)"'
+        plugins:
+          - aws-assume-role-with-web-identity#v1.6.0:
+              role-arn: arn:aws:iam::936637512419:role/vllm-release-macos-wheel-uploader
+              region: us-west-2
+
      - label: "Build wheel - x86_64 - CUDA 12.9"
        depends_on: ~
        id: build-wheel-x86-cuda-12-9
@@ -86,7 +129,7 @@ steps:
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)" release-wheels'
        env:
          DOCKER_BUILDKIT: "1"

@@ -100,7 +143,7 @@ steps:
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)" release-wheels'
        env:
          DOCKER_BUILDKIT: "1"

@@ -114,7 +157,7 @@ steps:
          - "mkdir artifacts"
          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "s3://vllm-wheels/$$BUILDKITE_COMMIT/$(cd artifacts/dist && echo *.whl)" release-wheels'
        env:
          DOCKER_BUILDKIT: "1"

@@ -162,7 +205,7 @@ steps:
          # re-tag to default image tag and push, just in case arm64 build fails
          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)" release-images'

      - label: "Build release image - aarch64 - CUDA 13.0"
        depends_on: ~
@@ -187,7 +230,7 @@ steps:
              --progress plain \
              -f docker/Dockerfile .
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)" release-images'

      - label: "Build release image - x86_64 - CUDA 12.9"
        depends_on: ~
@@ -214,7 +257,7 @@ steps:
          # re-tag to default image tag and push, just in case arm64 build fails
          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129" release-images'

      - label: "Build release image - aarch64 - CUDA 12.9"
        depends_on: ~
@@ -238,7 +281,7 @@ steps:
              --progress plain \
              -f docker/Dockerfile .
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129" release-images'

      - label: "Build release image - x86_64 - CUDA 13.0 - Ubuntu 24.04"
        depends_on: ~
@@ -267,7 +310,7 @@ steps:
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"
          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404" release-images'

      - label: "Build release image - aarch64 - CUDA 13.0 - Ubuntu 24.04"
        depends_on: ~
@@ -294,7 +337,7 @@ steps:
              --progress plain \
              -f docker/Dockerfile .
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-ubuntu2404" release-images'

      - label: "Build release image - x86_64 - CUDA 12.9 - Ubuntu 24.04"
        depends_on: ~
@@ -322,7 +365,7 @@ steps:
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"
          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404" release-images'

      - label: "Build release image - aarch64 - CUDA 12.9 - Ubuntu 24.04"
        depends_on: ~
@@ -348,7 +391,26 @@ steps:
              --progress plain \
              -f docker/Dockerfile .
          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu129-ubuntu2404" release-images'
+
+      - 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" release-images'
+        env:
+          DOCKER_BUILDKIT: "1"

      - block: "Build release image for x86_64 CPU"
        key: block-cpu-release-image-build
@@ -366,7 +428,7 @@ steps:
          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_X86=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
          - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)" release-images'
        env:
          DOCKER_BUILDKIT: "1"

@@ -386,7 +448,7 @@ steps:
          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
          - "docker push public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:latest"
          - "docker push public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:$(buildkite-agent meta-data get release-version)"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "$$BUILDKITE_LABEL" "public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:$(buildkite-agent meta-data get release-version)" release-images'
        env:
          DOCKER_BUILDKIT: "1"

@@ -404,7 +466,7 @@ steps:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 --amend"
          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 13.0" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 13.0" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT" release-manifests'

      - label: "Create multi-arch manifest - CUDA 12.9"
        depends_on:
@@ -417,7 +479,7 @@ steps:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu129 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu129 --amend"
          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 12.9" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 12.9" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129" release-manifests'

      - label: "Create multi-arch manifest - CUDA 13.0 - Ubuntu 24.04"
        depends_on:
@@ -430,7 +492,7 @@ steps:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-ubuntu2404 --amend"
          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"
-          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 13.0 Ubuntu 24.04" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 13.0 Ubuntu 24.04" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-ubuntu2404" release-manifests'

      - label: "Create multi-arch manifest - CUDA 12.9 - Ubuntu 24.04"
        depends_on:
@@ -443,7 +505,17 @@ steps:
          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu129-ubuntu2404 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu129-ubuntu2404 --amend"
          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$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:$BUILDKITE_COMMIT-cu129-ubuntu2404"'
+          - 'bash .buildkite/scripts/annotate-build-artifact.sh "Manifest: CUDA 12.9 Ubuntu 24.04" "public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu129-ubuntu2404" release-manifests'
+
+      - 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" release-manifests'

      - label: "Publish nightly multi-arch image to DockerHub"
        depends_on:
@@ -819,6 +891,23 @@ 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
@@ -849,6 +938,7 @@ 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
@@ -3,7 +3,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 #
 # Append a build artifact line to the Buildkite annotation.
-# Usage: annotate-build-artifact.sh <label> <value>
+# Usage: annotate-build-artifact.sh <label> <value> <context>
 set -e
 echo "- **${1}**: \`${2}\`" | \
-  buildkite-agent annotate --append --style 'info' --context 'release-artifacts'
+  buildkite-agent annotate --append --style 'info' \
+    --context "${3:?context is required}"
@@ -0,0 +1,36 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Append the Docker image tag(s) an image-build step pushed to a Buildkite
+# annotation, so the built image tags show up on the build page instead of
+# being buried in the job logs.
+#
+# Usage: annotate-image-build.sh <image_tag> [<image_tag> ...]
+set -euo pipefail
+
+# buildkite-agent only exists on Buildkite agents; no-op elsewhere so the
+# image build scripts stay runnable locally.
+if ! command -v buildkite-agent >/dev/null 2>&1; then
+    echo "buildkite-agent not found; skipping image tag annotation"
+    exit 0
+fi
+
+label="${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' "$content" | \
+    buildkite-agent annotate --append --style 'info' --context 'docker-images'; then
+    echo "warning: failed to annotate build with image tags"
+fi
@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Build the macOS arm64 CPU wheel natively on a macOS agent (the `macmini`
+# queue) into artifacts/dist/ for upload-nightly-wheels.sh.
+
+set -euo pipefail
+
+# The Rust frontend build needs protoc.
+if ! command -v protoc >/dev/null 2>&1; then
+  brew install protobuf
+fi
+
+# upload-nightly-wheels.sh expects exactly one wheel.
+rm -rf artifacts/dist
+mkdir -p artifacts/dist
+
+export VLLM_TARGET_DEVICE=cpu
+export VLLM_REQUIRE_RUST_FRONTEND=1
+export MACOSX_DEPLOYMENT_TARGET=11.0
+# uv's CPython is universal2; force an arm64-only build and tag so the wheel
+# isn't mislabelled universal2 and installed on Intel Macs where import fails.
+export ARCHFLAGS="-arch arm64"
+export _PYTHON_HOST_PLATFORM="macosx-11.0-arm64"
+export CMAKE_BUILD_PARALLEL_LEVEL="${CMAKE_BUILD_PARALLEL_LEVEL:-4}"
+
+uv venv --python 3.12
+uv pip install -r requirements/build/cpu.txt --index-strategy unsafe-best-match
+uv build --wheel --no-build-isolation -o artifacts/dist
+
+ls -l artifacts/dist/*.whl
@@ -29,7 +29,11 @@ if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
        TORCH_INDEX_URL=""
    fi
 else
-    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu130"
+    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
 fi
 echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"

@@ -15,9 +15,9 @@ set -euo pipefail

 DEFAULT_REPO_SLUG="vllm-project/vllm"
 DEFAULT_CI_HCL_SOURCE="docker/ci-rocm.hcl"
-DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh"
+DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tools/install_protoc.sh rust-toolchain.toml tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh .buildkite/scripts/rocm/build-ci-base.sh"
 DEFAULT_CI_BASE_DOCKERFILE="docker/Dockerfile.rocm"
-DEFAULT_CI_BASE_DOCKERFILE_STAGES="base build_rixl build_rocshmem build_deepep mori_base ci_base"
+DEFAULT_CI_BASE_DOCKERFILE_STAGES="base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust-toolchain build_rixl build_rocshmem build_deepep mori_base ci_base"
 DEFAULT_CI_BASE_METADATA_VERSION="1"
 IMAGE_EXISTED_BEFORE_BUILD=0

@@ -285,7 +285,7 @@ get_content_arg_names() {
    fi | awk 'NF && !seen[$0]++'
 }

-compute_ci_base_content_hash() {
+compute_ci_base_content_hash_once() {
    local -a content_paths=()
    local -a content_args=()
    local dockerfile="${CI_BASE_DOCKERFILE:-}"
@@ -301,7 +301,8 @@ compute_ci_base_content_hash() {
        if [[ -n "${dockerfile}" ]]; then
            printf 'dockerfile:%s\n' "${dockerfile}"
            printf 'resolved-build-args:\n'
-            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}"
+            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}" \
+                || return 1
            if [[ -n "${stages}" ]]; then
                printf 'dockerfile-stages:%s\n' "${stages}"
                if [[ -f "${dockerfile}" ]]; then
@@ -314,6 +315,53 @@ compute_ci_base_content_hash() {
    } | sha256sum | cut -d' ' -f1
 }

+compute_ci_base_content_hash() {
+    local attempts="${CI_BASE_HASH_ATTEMPTS:-3}"
+    local delay_secs="${CI_BASE_HASH_RETRY_DELAY:-5}"
+    local attempt=0
+    local hash=""
+    local failed=0
+    local -a hashes=()
+
+    if [[ ! "${attempts}" =~ ^[1-9][0-9]*$ ]]; then
+        echo "Invalid CI_BASE_HASH_ATTEMPTS: ${attempts}" >&2
+        return 1
+    fi
+    if [[ ! "${delay_secs}" =~ ^[0-9]+$ ]]; then
+        echo "Invalid CI_BASE_HASH_RETRY_DELAY: ${delay_secs}" >&2
+        return 1
+    fi
+
+    for ((attempt = 1; attempt <= attempts; attempt++)); do
+        if ! hash=$(compute_ci_base_content_hash_once); then
+            echo "ci_base content hash calculation ${attempt}/${attempts} failed" >&2
+            failed=1
+        else
+            hashes+=("${hash}")
+            echo "ci_base content hash calculation ${attempt}/${attempts}: ${hash}" >&2
+        fi
+
+        if ((attempt < attempts)); then
+            sleep "${delay_secs}"
+        fi
+    done
+
+    if ((failed)) || ((${#hashes[@]} != attempts)); then
+        echo "Could not calculate a reliable ci_base content hash" >&2
+        return 1
+    fi
+
+    for hash in "${hashes[@]:1}"; do
+        if [[ "${hash}" != "${hashes[0]}" ]]; then
+            echo "ci_base content hash changed between calculations" >&2
+            printf '  observed: %s\n' "${hashes[@]}" >&2
+            return 1
+        fi
+    done
+
+    printf '%s\n' "${hashes[0]}"
+}
+
 extract_dockerfile_arg_default() {
    local dockerfile="$1"
    local arg_name="$2"
@@ -366,7 +414,11 @@ hash_dockerfile_arg_values() {
        printf 'arg:%s=%s\n' "${arg_name}" "${arg_value:-<empty>}"
        if [[ "${arg_name}" == "BASE_IMAGE" && -n "${arg_value}" ]]; then
            digest=$(resolve_image_digest "${arg_value}")
-            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest:-unknown}"
+            if [[ -z "${digest}" ]]; then
+                echo "Failed to resolve digest for BASE_IMAGE=${arg_value}" >&2
+                return 1
+            fi
+            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest}"
        fi
    done
 }
@@ -393,6 +445,16 @@ should_upload_wheel_artifacts() {
        || "${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"
@@ -733,6 +795,8 @@ configure_ci_base_image_refs() {

    if is_ci_base_target; then
        IMAGE_TAG="${primary_tag}"
+        CI_BASE_IMAGE="${primary_tag}"
+        export CI_BASE_IMAGE
        export IMAGE_TAG

        echo "ci_base primary image tag: ${CI_BASE_IMAGE_TAG}"
@@ -750,6 +814,10 @@ configure_ci_base_image_refs() {
            echo "ci_base stable alias will not be pushed for this build"
            echo "Set NIGHTLY=1 on ${CI_BASE_STABLE_BRANCH:-main} to refresh ${stable_tag}"
        fi
+        set_buildkite_metadata "rocm-ci-base-image" "${CI_BASE_IMAGE_TAG}"
+        set_buildkite_metadata "rocm-ci-base-image-content" "${content_tag}"
+        set_buildkite_metadata "rocm-ci-base-image-commit" "${CI_BASE_IMAGE_TAG_COMMIT_REF:-}"
+        set_buildkite_metadata "rocm-ci-base-image-stable" "${CI_BASE_IMAGE_TAG_STABLE:-}"
        return 0
    fi

@@ -1195,12 +1263,24 @@ uses_rocm_csrc_cache() {
    esac
 }

+uses_rocm_rust_cache() {
+    case "${TARGET}" in
+        rust-rocm-ci|test-rocm-ci|test-rocm-ci-with-wheel|test-rocm-ci-with-artifacts|export-wheel-rocm)
+            return 0
+            ;;
+        *)
+            return 1
+            ;;
+    esac
+}
+
 compute_rocm_csrc_content_hash() {
    local bake_dir=""
    local dockerfile_rocm=""
    local -a content_paths=(
        "requirements/common.txt"
        "requirements/rocm.txt"
+        "pyproject.toml"
        "setup.py"
        "CMakeLists.txt"
        "cmake"
@@ -1244,6 +1324,56 @@ compute_rocm_csrc_content_hash_if_needed() {
    echo "ROCm csrc content cache ref: ${ROCM_CSRC_CONTENT_CACHE_REF}"
 }

+compute_rocm_rust_content_hash() {
+    local bake_dir=""
+    local dockerfile_rocm=""
+    local -a content_paths=(
+        "requirements/build/rust.txt"
+        "rust/Cargo.lock"
+        "rust/Cargo.toml"
+        "rust/proto"
+        "rust/src"
+        "rust-toolchain.toml"
+        "tools/build_rust.py"
+        "tools/install_protoc.sh"
+        "build_rust.sh"
+    )
+    local -a content_args=()
+
+    bake_dir=$(dirname "${VLLM_BAKE_FILE}")
+    dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
+    mapfile -t content_args < <(
+        get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
+    )
+
+    {
+        printf 'rust-input-files-hash:%s\n' "$(compute_content_hash "${content_paths[@]}")"
+        printf 'dockerfile:%s\n' "${dockerfile_rocm}"
+        printf 'resolved-build-args:\n'
+        hash_dockerfile_arg_values "${dockerfile_rocm}" "${content_args[@]}"
+        printf 'dockerfile-stages:base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build\n'
+        if [[ -f "${dockerfile_rocm}" ]]; then
+            hash_dockerfile_stages "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build"
+        else
+            printf 'missing:%s\n' "${dockerfile_rocm}"
+        fi
+    } | sha256sum | cut -d' ' -f1
+}
+
+compute_rocm_rust_content_hash_if_needed() {
+    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
+
+    if [[ "${ROCM_RUST_CONTENT_CACHE:-1}" == "0" ]] || ! uses_rocm_rust_cache; then
+        return 0
+    fi
+
+    ROCM_RUST_CONTENT_HASH=$(compute_rocm_rust_content_hash)
+    ROCM_RUST_CONTENT_CACHE_REF="${cache_repo}:rust-rocm-input-${ROCM_RUST_CONTENT_HASH}"
+    export ROCM_RUST_CONTENT_HASH
+    export ROCM_RUST_CONTENT_CACHE_REF
+    echo "ROCm Rust content cache ref: ${ROCM_RUST_CONTENT_CACHE_REF}"
+}
+
 write_hcl_string_list_entries() {
    local indent="$1"
    local value=""
@@ -1301,6 +1431,7 @@ write_rocm_build_arg_override() {
                "${CI_BASE_DOCKERFILE_STAGES:-${DEFAULT_CI_BASE_DOCKERFILE_STAGES}}" \
                "${CI_BASE_CONTENT_ARGS:-}"
            get_content_arg_names "${dockerfile_rocm}" "base csrc-build" "${ROCM_CSRC_CONTENT_ARGS:-}"
+            get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
        } | awk 'NF && !seen[$0]++'
    )

@@ -1349,46 +1480,133 @@ validate_cache_export_mode() {
    esac
 }

+validate_content_cache_export_mode() {
+    local mode="$1"
+    local env_name="$2"
+
+    case "${mode}" in
+        missing|always|never)
+            ;;
+        *)
+            echo "Error: ${env_name} must be one of: missing, always, never"
+            exit 1
+            ;;
+    esac
+}
+
+should_export_content_cache_ref() {
+    local cache_ref="$1"
+    local cache_name="$2"
+    local mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
+
+    case "${mode}" in
+        always)
+            echo "${cache_name} content cache export mode is always; exporting ${cache_ref}"
+            return 0
+            ;;
+        never)
+            echo "${cache_name} content cache export mode is never; not exporting ${cache_ref}"
+            return 1
+            ;;
+        missing|"")
+            if docker buildx imagetools inspect "${cache_ref}" >/dev/null 2>&1; then
+                echo "${cache_name} content cache exists; not re-exporting ${cache_ref}"
+                return 1
+            fi
+            echo "${cache_name} content cache missing; will export ${cache_ref}"
+            return 0
+            ;;
+        *)
+            echo "Error: ROCM_CONTENT_CACHE_EXPORT_MODE must be one of: missing, always, never"
+            exit 1
+            ;;
+    esac
+}
+
 write_rocm_cache_override() {
    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
+    local content_cache_export_mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
    local csrc_cache_to_mode="${ROCM_CSRC_CACHE_TO_MODE:-max}"
+    local rust_cache_to_mode="${ROCM_RUST_CACHE_TO_MODE:-max}"
    local rocm_cache_to_mode="${ROCM_FINAL_CACHE_TO_MODE:-min}"
-    local -a content_cache_from=()
+    local -a csrc_content_cache_from=()
+    local -a rust_content_cache_from=()
+    local -a combined_content_cache_from=()
    local -a csrc_cache_to=()
+    local -a rust_cache_to=()
    local -a rocm_cache_to=()
    local -a export_wheel_cache_to=()
+    local export_csrc_cache=1
+    local export_rust_cache=1

-    if ! uses_rocm_csrc_cache; then
+    if ! uses_rocm_csrc_cache && ! uses_rocm_rust_cache; then
        return 0
    fi

+    validate_content_cache_export_mode \
+        "${content_cache_export_mode}" \
+        "ROCM_CONTENT_CACHE_EXPORT_MODE"
    validate_cache_export_mode "${csrc_cache_to_mode}" "ROCM_CSRC_CACHE_TO_MODE"
+    validate_cache_export_mode "${rust_cache_to_mode}" "ROCM_RUST_CACHE_TO_MODE"
    validate_cache_export_mode "${rocm_cache_to_mode}" "ROCM_FINAL_CACHE_TO_MODE"
+    echo "ROCm content cache export mode: ${content_cache_export_mode}"
    echo "ROCm csrc cache export mode: ${csrc_cache_to_mode}"
+    echo "ROCm Rust cache export mode: ${rust_cache_to_mode}"
    echo "ROCm final image cache export mode: ${rocm_cache_to_mode}"

    if [[ -n "${ROCM_CSRC_CONTENT_CACHE_REF:-}" ]]; then
-        content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
-        csrc_cache_to+=(
-            "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        csrc_content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
+        if should_export_content_cache_ref "${ROCM_CSRC_CONTENT_CACHE_REF}" "ROCm csrc"; then
+            csrc_cache_to+=(
+                "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        else
+            export_csrc_cache=0
+        fi
    fi

+    if [[ -n "${ROCM_RUST_CONTENT_CACHE_REF:-}" ]]; then
+        rust_content_cache_from+=("type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF}")
+        if should_export_content_cache_ref "${ROCM_RUST_CONTENT_CACHE_REF}" "ROCm Rust"; then
+            rust_cache_to+=(
+                "type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        else
+            export_rust_cache=0
+        fi
+    fi
+
+    combined_content_cache_from=("${csrc_content_cache_from[@]}" "${rust_content_cache_from[@]}")
+
    # Docker Hub cache exports are best-effort. A cache-only target failure can
    # otherwise cancel the sibling image target before its manifest is pushed.
    if [[ -n "${BUILDKITE_COMMIT:-}" ]]; then
-        csrc_cache_to+=(
-            "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        if [[ ${export_csrc_cache} -eq 1 ]]; then
+            csrc_cache_to+=(
+                "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        fi
+        if [[ ${export_rust_cache} -eq 1 ]]; then
+            rust_cache_to+=(
+                "type=registry,ref=${cache_repo}:rust-rocm-${BUILDKITE_COMMIT},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        fi
        rocm_cache_to+=(
            "type=registry,ref=${cache_repo}:rocm-${BUILDKITE_COMMIT},mode=${rocm_cache_to_mode},ignore-error=true"
        )
    fi

    if [[ -n "${ROCM_CACHE_BRANCH_TAG:-}" ]]; then
-        csrc_cache_to+=(
-            "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        if [[ ${export_csrc_cache} -eq 1 ]]; then
+            csrc_cache_to+=(
+                "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        fi
+        if [[ ${export_rust_cache} -eq 1 ]]; then
+            rust_cache_to+=(
+                "type=registry,ref=${cache_repo}:rust-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        fi
        rocm_cache_to+=(
            "type=registry,ref=${cache_repo}:rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rocm_cache_to_mode},ignore-error=true"
        )
@@ -1406,7 +1624,7 @@ target "csrc-rocm-ci" {
  cache-from = concat(
    get_cache_from_rocm_csrc(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${csrc_content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1414,11 +1632,23 @@ EOF
        cat <<EOF
 }

+target "rust-rocm-ci" {
+  cache-from = concat(
+    get_cache_from_rocm_rust(),
+EOF
+        write_hcl_string_list "    " "${rust_content_cache_from[@]}"
+        cat <<EOF
+  )
+EOF
+        write_hcl_string_list_attr "  " "cache-to" "${rust_cache_to[@]}"
+        cat <<EOF
+}
+
 target "test-rocm-ci" {
  cache-from = concat(
    get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1430,7 +1660,7 @@ target "export-wheel-rocm" {
  cache-from = concat(
    get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
        cat <<EOF
  )
 EOF
@@ -1720,8 +1950,8 @@ confirm_remote_image_push() {
        fi

        if [[ -z "${remote_revision}" \
-              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 \
-              && image_tag_is_commit_scoped ]]; then
+              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 ]] \
+            && image_tag_is_commit_scoped; then
            echo "Remote image exists under a commit-scoped tag; accepting push despite missing revision label."
            return 0
        fi
@@ -1779,7 +2009,10 @@ seed_dependency_caches_if_needed() {

        echo "--- :docker: Seeding ${target}"
        echo "Expected cache ref: ${cache_ref}"
-        docker buildx bake "${BAKE_FILES[@]}" --progress plain "${target}"
+        docker buildx bake \
+            "${BAKE_FILES[@]}" \
+            --progress "${BUILDKIT_PROGRESS:-plain}" \
+            "${target}"
        verify_dependency_cache_ref "${cache_ref}"
    done
 }
@@ -1807,7 +2040,10 @@ run_bake() {
    local build_rc=0

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

    if [[ ${build_rc} -eq 0 ]]; then
        echo "--- :white_check_mark: Build complete"
@@ -1845,36 +2081,57 @@ upload_wheel_artifacts_if_present() {
    local wheel_dir="./wheel-export"
    local artifact_dir="artifacts/vllm-rocm-install"
    local archive_name="vllm-rocm-install.tar.gz"
+    local metadata_dir="${wheel_dir}/.vllm-ci-artifact"
+    local native_base_image=""
    local whl=""
    local whl_name=""
+    local -a wheels=()

    if ! should_upload_wheel_artifacts; then
        return 0
    fi

-    if [[ ! -d "${wheel_dir}" ]] || ! ls "${wheel_dir}"/*.whl >/dev/null 2>&1; then
-        echo "No ROCm wheel artifacts found in ${wheel_dir}"
-        return 0
+    if [[ -d "${wheel_dir}" ]]; then
+        mapfile -t wheels < <(find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print)
+    fi
+    if [[ ${#wheels[@]} -ne 1 ]]; then
+        echo "Expected exactly one ROCm wheel in ${wheel_dir}; found ${#wheels[@]}" >&2
+        return 1
+    fi
+    whl="${wheels[0]}"
+    whl_name=$(basename "${whl}")
+    native_base_image="${CI_BASE_IMAGE_TAG_COMMIT_REF:-${CI_BASE_IMAGE:-}}"
+    if [[ -z "${native_base_image}" ]]; then
+        echo "Native ROCm artifact requires a ci_base image reference" >&2
+        return 1
    fi

    echo "--- :package: Uploading ROCm vLLM install artifact"
-    mkdir -p "${artifact_dir}"
+    rm -rf "${artifact_dir}" "${metadata_dir}"
+    mkdir -p "${artifact_dir}" "${metadata_dir}"
+
+    printf '%s\n' "${BUILDKITE_COMMIT:-local}" > "${metadata_dir}/commit.txt"
+    printf '%s\n' "${native_base_image}" > "${metadata_dir}/native-base-image.txt"
+    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${metadata_dir}/ci-base-image.txt"
+    printf '%s\n' "${IMAGE_TAG:-}" > "${metadata_dir}/fallback-image.txt"
+    printf '%s\n' "${whl_name}" > "${metadata_dir}/wheel-filename.txt"

    tar -C "${wheel_dir}" -czf "${artifact_dir}/${archive_name}" .
+    (
+        cd "${artifact_dir}"
+        sha256sum "${archive_name}" > "${archive_name}.sha256"
+    )
    echo "Created ${archive_name}: $(du -sh "${artifact_dir}/${archive_name}" | cut -f1)"
-    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${artifact_dir}/ci-base-image.txt"
-    printf '%s\n' "${IMAGE_TAG:-}" > "${artifact_dir}/fallback-image.txt"
-
-    for whl in "${wheel_dir}"/*.whl; do
-        [[ -f "${whl}" ]] || continue
-        whl_name=$(basename "${whl}")
-        cp "${whl}" "${artifact_dir}/${whl_name}"
-        echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"
-    done
+    cp "${metadata_dir}"/*.txt "${artifact_dir}/"
+    cp "${whl}" "${artifact_dir}/${whl_name}"
+    echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"

    if command -v buildkite-agent >/dev/null 2>&1; then
-        buildkite-agent artifact upload "${artifact_dir}/*"
+        buildkite-agent artifact upload "${artifact_dir}/*" || return 1
        echo "ROCm vLLM install artifacts uploaded to ${artifact_dir}/"
+    elif [[ "${BUILDKITE:-false}" == "true" ]]; then
+        echo "buildkite-agent not found; cannot upload required ROCm artifacts" >&2
+        return 1
    else
        echo "Not in Buildkite, skipping artifact upload"
    fi
@@ -1898,6 +2155,7 @@ main() {
    compute_dependency_cache_keys
    write_ci_base_label_override
    compute_rocm_csrc_content_hash_if_needed
+    compute_rocm_rust_content_hash_if_needed
    write_rocm_cache_override
    resolve_ci_base_dependency_targets
    print_bake_config
@@ -1905,6 +2163,11 @@ main() {
        echo "BAKE_PRINT_ONLY=1 set; skipping build"
        return 0
    fi
+    if should_upload_wheel_artifacts; then
+        # wheel-export is an output directory, not a BuildKit cache. Starting
+        # clean prevents a failed/retried export from packaging a stale wheel.
+        rm -rf ./wheel-export
+    fi
    seed_dependency_caches_if_needed
    run_bake
    upload_wheel_artifacts_if_present
@@ -1,7 +1,7 @@
 #!/bin/bash

-# This script runs tests inside the corresponding ROCm docker container.
-# It handles both single-node and multi-node test configurations.
+# This script runs ROCm tests either directly in a native CI pod or inside the
+# corresponding Docker container. Multi-node tests continue to use Docker.
 #
 # Multi-node detection: Instead of matching on fragile group names, we detect
 # multi-node jobs structurally by looking for the bracket command syntax
@@ -28,6 +28,34 @@
 ###############################################################################
 set -o pipefail

+: "${BUILDKIT_PROGRESS:=plain}"
+: "${TERM:=xterm-256color}"
+: "${FORCE_COLOR:=1}"
+: "${CLICOLOR_FORCE:=1}"
+: "${PY_COLORS:=1}"
+: "${ROCM_DOCKER_TTY:=1}"
+: "${PYTHONFAULTHANDLER:=1}"
+: "${PYTEST_TIMEOUT:=2100}"
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --color"* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--color=yes"
+fi
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations=25"
+fi
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations-min="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations-min=1.0"
+fi
+# Dump stacks after 15 minutes, then stop an individual test after 35 minutes.
+if [[ " ${PYTEST_ADDOPTS:-} " != *" faulthandler_timeout="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }-o faulthandler_timeout=900"
+fi
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --timeout-method="* &&
+  " ${PYTEST_ADDOPTS:-} " != *" --timeout-method "* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--timeout-method=thread"
+fi
+export BUILDKIT_PROGRESS TERM FORCE_COLOR CLICOLOR_FORCE PY_COLORS PYTEST_ADDOPTS PYTEST_TIMEOUT ROCM_DOCKER_TTY
+export PYTHONFAULTHANDLER
+
 # Export Python path for commands that run directly on the host. Containerized
 # tests set this to /vllm-workspace below so spawned Python processes do not
 # depend on their current working directory.
@@ -42,6 +70,28 @@ report_docker_usage() {
  docker system df || true
 }

+clear_ci_orchestration_env() {
+  unset -v \
+    VLLM_TEST_GROUP_NAME \
+    VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE \
+    VLLM_CI_ARTIFACT_STEP \
+    VLLM_TEST_CACHE \
+    VLLM_CI_EXECUTION_MODE \
+    VLLM_CI_WORKSPACE \
+    VLLM_CI_REQUIRE_WORKSPACE_MOUNT \
+    VLLM_TEST_COMMANDS \
+    VLLM_CI_BRANCH \
+    VLLM_CI_BASE_IMAGE \
+    VLLM_CI_FALLBACK_IMAGE \
+    VLLM_CI_DOCKER_DISABLED \
+    VLLM_CI_ARTIFACT_GLOB \
+    VLLM_CI_ARTIFACT_CHECKSUM_GLOB \
+    VLLM_CI_EXPECTED_GPU_COUNT \
+    VLLM_CI_USE_ARTIFACTS \
+    VLLM_CI_RESULTS_ROOT \
+    VLLM_ALLOW_DEPRECATED_BEAM_SEARCH
+}
+
 cleanup_network() {
  local max_nodes=${NUM_NODES:-2}
  for node in $(seq 0 $((max_nodes - 1))); do
@@ -134,7 +184,11 @@ prepare_artifact_image() {
  fi

  cp "${wheel_dir}"/*.whl "${context_dir}/wheels/" || return 1
-  tar -C "${wheel_dir}" --exclude='*.whl' -cf - . \
+  tar -C "${wheel_dir}" \
+    --exclude='*.whl' \
+    --exclude='.vllm-ci-artifact' \
+    --exclude='./.vllm-ci-artifact' \
+    -cf - . \
    | tar -C "${workspace_dir}" -xf - || return 1
  cat > "${context_dir}/Dockerfile" <<'EOF'
 ARG BASE_IMAGE
@@ -149,6 +203,7 @@ 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
@@ -156,6 +211,257 @@ EOF
  return 0
 }

+is_native_runtime() {
+  [[ "${AMD_CI_RUNTIME:-}" == "native" || "${NATIVE_CI:-}" == "true" ]]
+}
+
+validate_native_workspace() {
+  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+  local workspace_real=""
+  local checkout_real=""
+  local workspace_mount=""
+
+  mkdir -p "${workspace_dir}" || return 1
+  workspace_real=$(readlink -m "${workspace_dir}") || return 1
+  if [[ -n "${BUILDKITE_BUILD_CHECKOUT_PATH:-}" ]]; then
+    checkout_real=$(readlink -m "${BUILDKITE_BUILD_CHECKOUT_PATH}") || return 1
+    if [[ "${checkout_real}" == "${workspace_real}" \
+      || "${checkout_real}" == "${workspace_real}/"* \
+      || "${workspace_real}" == "${checkout_real}/"* ]]; then
+      echo "Refusing to replace ${workspace_real}; it overlaps the Buildkite checkout ${checkout_real}" >&2
+      return 1
+    fi
+  fi
+  if [[ "${VLLM_CI_REQUIRE_WORKSPACE_MOUNT:-1}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+      echo "findmnt is required to verify the native workspace mount" >&2
+      return 1
+    fi
+    workspace_mount=$(findmnt -n -T "${workspace_real}" -o TARGET 2>/dev/null || true)
+    if [[ "$(readlink -m "${workspace_mount:-/}")" != "${workspace_real}" ]]; then
+      echo "Native CI requires a dedicated volume mounted at ${workspace_real}" >&2
+      return 1
+    fi
+  fi
+}
+
+prepare_native_workspace() {
+  if [[ "${VLLM_CI_USE_ARTIFACTS:-0}" != "1" ]]; then
+    echo "Native CI requires VLLM_CI_USE_ARTIFACTS=1"
+    return 1
+  fi
+  if ! command -v buildkite-agent >/dev/null 2>&1; then
+    echo "buildkite-agent not found; cannot download ROCm wheel artifact"
+    return 1
+  fi
+  validate_native_workspace || return 1
+
+  local artifact_glob="${VLLM_CI_ARTIFACT_GLOB:-artifacts/vllm-rocm-install/vllm-rocm-install.tar.gz}"
+  local artifact_checksum_glob="${VLLM_CI_ARTIFACT_CHECKSUM_GLOB:-${artifact_glob}.sha256}"
+  local artifact_step="${VLLM_CI_ARTIFACT_STEP:-image-build-amd}"
+  local archive=""
+  local checksum=""
+  local download_dir=""
+  local metadata_dir=""
+  local recorded_base=""
+  local recorded_commit=""
+  local recorded_wheel=""
+  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+  local wheel_dir=""
+  local attempt=0
+  local attempt_dir=""
+  local -a archives=()
+  local -a checksums=()
+  local -a wheels=()
+
+  artifact_work_dir=$(mktemp -d -t vllm-rocm-artifact.XXXXXX) || return 1
+  wheel_dir="${artifact_work_dir}/wheels"
+  mkdir -p "${wheel_dir}" || return 1
+
+  echo "--- Downloading ROCm wheel artifact from ${artifact_step} (native in-pod)"
+  for attempt in 1 2 3; do
+    attempt_dir="${artifact_work_dir}/download-${attempt}"
+    rm -rf "${attempt_dir}" || return 1
+    mkdir -p "${attempt_dir}" || return 1
+    if buildkite-agent artifact download \
+      "${artifact_glob}" "${attempt_dir}" --step "${artifact_step}" \
+      && buildkite-agent artifact download \
+        "${artifact_checksum_glob}" "${attempt_dir}" --step "${artifact_step}"; then
+      download_dir="${attempt_dir}"
+      break
+    fi
+    echo "Artifact download attempt ${attempt}/3 failed"
+    if [[ "${attempt}" -lt 3 ]]; then
+      sleep $((attempt * 2))
+    fi
+  done
+  if [[ -z "${download_dir}" ]]; then
+    echo "Failed to download ${artifact_glob} and ${artifact_checksum_glob} from ${artifact_step}"
+    return 1
+  fi
+
+  mapfile -t archives < <(
+    find "${download_dir}" -name "vllm-rocm-install.tar.gz" -type f -print
+  )
+  mapfile -t checksums < <(
+    find "${download_dir}" -name "vllm-rocm-install.tar.gz.sha256" -type f -print
+  )
+  if [[ ${#archives[@]} -ne 1 || ${#checksums[@]} -ne 1 ]]; then
+    echo "Expected exactly one ROCm archive and checksum; found ${#archives[@]} archive(s) and ${#checksums[@]} checksum(s)" >&2
+    return 1
+  fi
+  archive="${archives[0]}"
+  checksum="${checksums[0]}"
+  if [[ "$(dirname "${archive}")" != "$(dirname "${checksum}")" ]]; then
+    echo "ROCm archive and checksum were downloaded to different directories" >&2
+    return 1
+  fi
+  (
+    cd "$(dirname "${archive}")"
+    sha256sum -c "$(basename "${checksum}")"
+  ) || return 1
+
+  tar --no-same-owner -xzf "${archive}" -C "${wheel_dir}" || return 1
+  mapfile -t wheels < <(
+    find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print
+  )
+  if [[ ${#wheels[@]} -ne 1 ]]; then
+    echo "ROCm artifact must contain exactly one top-level wheel; found ${#wheels[@]}" >&2
+    return 1
+  fi
+  metadata_dir="${wheel_dir}/.vllm-ci-artifact"
+  for metadata_file in commit.txt native-base-image.txt wheel-filename.txt; do
+    if [[ ! -s "${metadata_dir}/${metadata_file}" ]]; then
+      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
+      return 1
+    fi
+  done
+  for metadata_file in ci-base-image.txt fallback-image.txt; do
+    if [[ ! -f "${metadata_dir}/${metadata_file}" ]]; then
+      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
+      return 1
+    fi
+  done
+
+  recorded_commit=$(tr -d '\r\n' < "${metadata_dir}/commit.txt")
+  recorded_base=$(tr -d '\r\n' < "${metadata_dir}/native-base-image.txt")
+  recorded_wheel=$(tr -d '\r\n' < "${metadata_dir}/wheel-filename.txt")
+  if [[ -z "${BUILDKITE_COMMIT:-}" || "${recorded_commit}" != "${BUILDKITE_COMMIT}" ]]; then
+    echo "ROCm artifact commit ${recorded_commit} does not match ${BUILDKITE_COMMIT:-unset}" >&2
+    return 1
+  fi
+  if [[ -z "${VLLM_CI_BASE_IMAGE:-}" || "${recorded_base}" != "${VLLM_CI_BASE_IMAGE}" ]]; then
+    echo "ROCm artifact base ${recorded_base} does not match ${VLLM_CI_BASE_IMAGE:-unset}" >&2
+    return 1
+  fi
+  if [[ "${recorded_wheel}" != "$(basename "${wheels[0]}")" ]]; then
+    echo "ROCm artifact wheel manifest ${recorded_wheel} does not match $(basename "${wheels[0]}")" >&2
+    return 1
+  fi
+  for required_dir in tests .buildkite requirements; do
+    if [[ ! -d "${wheel_dir}/${required_dir}" ]]; then
+      echo "ROCm wheel artifact did not contain ${required_dir}/" >&2
+      return 1
+    fi
+  done
+
+  echo "--- Installing ROCm wheel into pod environment"
+  python3 -m pip install --no-deps --force-reinstall "${wheels[0]}" || return 1
+
+  echo "--- Preparing ${workspace_dir} from artifact"
+  find "${workspace_dir}" -mindepth 1 -maxdepth 1 -exec rm -rf -- {} + || return 1
+  tar -C "${wheel_dir}" \
+    --exclude='*.whl' \
+    --exclude='.vllm-ci-artifact' \
+    --exclude='./.vllm-ci-artifact' \
+    -cf - . | tar --no-same-owner -C "${workspace_dir}" -xf - || return 1
+  if [[ ! -d "${workspace_dir}/tests" ]]; then
+    echo "Failed to stage the native test workspace" >&2
+    return 1
+  fi
+
+  return 0
+}
+
+initialize_native_environment() {
+  local job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
+  local job_id_suffix=""
+  local native_root=""
+  local hf_mount=""
+
+  if [[ "$(id -u)" -ne 0 ]]; then
+    echo "Native ROCm CI currently requires the ci_base container to run as root" >&2
+    return 1
+  fi
+
+  job_id="${job_id//[^A-Za-z0-9_.-]/_}"
+  job_id_suffix="${job_id##*-}"
+  job_id_suffix="${job_id_suffix:0:12}"
+  native_root="/tmp/vllm-native-${job_id}"
+  TMPDIR="/tmp/vllm-${job_id_suffix}/tmp"
+  VLLM_RPC_BASE_PATH="/tmp"
+  : "${TORCHINDUCTOR_CACHE_DIR:=${native_root}/cache/torchinductor}"
+  : "${TRITON_CACHE_DIR:=${native_root}/cache/triton}"
+  : "${VLLM_CACHE_ROOT:=${native_root}/cache/vllm}"
+  : "${XDG_CACHE_HOME:=${native_root}/cache/xdg}"
+  : "${HF_HOME:=/home/buildkite-agent/huggingface}"
+  : "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
+  : "${HF_HUB_ETAG_TIMEOUT:=60}"
+  export TMPDIR VLLM_RPC_BASE_PATH
+  export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
+  export HF_HOME HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
+  export PYTORCH_ROCM_ARCH=""
+
+  mkdir -p "${TMPDIR}" \
+    "${TORCHINDUCTOR_CACHE_DIR}" \
+    "${TRITON_CACHE_DIR}" \
+    "${VLLM_CACHE_ROOT}" \
+    "${XDG_CACHE_HOME}" \
+    "${HF_HOME}" || return 1
+
+  if [[ "${VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE:-0}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+      echo "findmnt is required to verify the native Hugging Face cache mount" >&2
+      return 1
+    fi
+    hf_mount=$(findmnt -n -T "${HF_HOME}" -o TARGET 2>/dev/null || true)
+    if [[ -z "${hf_mount}" || "${hf_mount}" == "/" ]]; then
+      echo "Native CI requires a persistent volume mounted at or above ${HF_HOME}" >&2
+      return 1
+    fi
+  fi
+}
+
+run_native_preflight() {
+  local expected_gpus="${VLLM_CI_EXPECTED_GPU_COUNT:-1}"
+
+  if [[ ! "${expected_gpus}" =~ ^[0-9]+$ ]]; then
+    echo "Invalid VLLM_CI_EXPECTED_GPU_COUNT=${expected_gpus}" >&2
+    return 1
+  fi
+
+  python3 -c "import encodings, importlib.metadata as im, importlib.util as iu; [im.version(d) for d in ('transformers', 'torch', 'ray', 'sympy', 'markupsafe', 'vllm')]; missing=[m for m in ('torch.utils.model_zoo', 'transformers.models.nomic_bert', 'ray.dag', 'sympy.physics', 'markupsafe._speedups') if iu.find_spec(m) is None]; assert not missing, missing" || return 1
+
+  if [[ "${expected_gpus}" == "0" ]]; then
+    echo "Native CPU-only AMD job: skipping ROCm device validation"
+    return 0
+  fi
+
+  echo "--- ROCm info"
+  rocminfo || return 1
+  VLLM_CI_EXPECTED_GPU_COUNT="${expected_gpus}" python3 - <<'PY'
+import os
+
+import torch
+
+expected = int(os.environ["VLLM_CI_EXPECTED_GPU_COUNT"])
+assert torch.version.hip, "PyTorch is not a ROCm build"
+assert torch.cuda.is_available(), "ROCm GPU is not available to PyTorch"
+actual = torch.cuda.device_count()
+assert actual == expected, f"Expected {expected} ROCm GPU(s), found {actual}"
+PY
+}
+
 is_multi_node() {
  local cmds="$1"
  # Primary signal: NUM_NODES environment variable set by the pipeline
@@ -338,7 +644,58 @@ re_quote_pytest_markers() {
 # Main
 ###############################################################################

-# --- GPU initialization ---
+if is_native_runtime; then
+  echo "--- Native in-pod ROCm CI (AMD_CI_RUNTIME=${AMD_CI_RUNTIME:-unset}, NATIVE_CI=${NATIVE_CI:-unset})"
+  artifact_work_dir=""
+
+  cleanup_native_workspace() {
+    if [[ -n "${artifact_work_dir}" ]]; then
+      rm -rf "${artifact_work_dir}"
+    fi
+  }
+  trap cleanup_native_workspace EXIT
+
+  if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
+    commands="${VLLM_TEST_COMMANDS}"
+    commands_source="env"
+  else
+    commands="$*"
+    commands_source="argv"
+    if [[ -z "$commands" ]]; then
+      echo "Error: No test commands provided for native CI." >&2
+      exit 1
+    fi
+  fi
+
+  if [[ "$commands_source" == "argv" ]]; then
+    commands=$(re_quote_pytest_markers "$commands")
+  fi
+
+  if is_multi_node "$commands"; then
+    echo "Native CI does not support multi-node jobs yet."
+    exit 1
+  fi
+
+  if ! initialize_native_environment; then
+    echo "Failed to initialize the native test environment"
+    exit 1
+  fi
+  if ! prepare_native_workspace; then
+    echo "Failed to prepare native test workspace"
+    exit 1
+  fi
+
+  export PYTHONPATH="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+
+  echo "Native test commands: $commands"
+  run_native_preflight || exit 1
+  # Keep AMD CI orchestration variables out of vLLM's runtime environment.
+  clear_ci_orchestration_env
+  /bin/bash -o pipefail -c "${commands}"
+  handle_pytest_exit "$?"
+fi
+
+# --- GPU initialization for legacy Docker execution ---
 echo "--- ROCm info"
 rocminfo

@@ -440,25 +797,27 @@ 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 "${checkout}" ]]; then
-  checkout="."
+standalone_merge_base_env=()
+if [[ "$commands" == *python_only_compile.sh* ]]; then
+  # The ROCm test image often ships /vllm-workspace without .git. Resolve the
+  # wheels.vllm.ai commit from the agent checkout for this test only.
+  vllm_standalone_merge_base=""
+  checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
+  if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
+    checkout="."
+  fi
+  # Pass safe.directory per-command because Buildkite uses mixed user IDs.
+  if git -c "safe.directory=${checkout}" -C "${checkout}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+    vllm_standalone_merge_base="$(
+      git -c "safe.directory=${checkout}" -C "${checkout}" merge-base HEAD origin/main 2>/dev/null || true
+    )"
+  fi
+  if [[ -z "${vllm_standalone_merge_base}" ]]; then
+    vllm_standalone_merge_base="${BUILDKITE_COMMIT:-}"
+  fi
+  echo "INFO: passing CI_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
+  standalone_merge_base_env=(-e "CI_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}")
 fi
-# 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"

@@ -489,6 +848,7 @@ else
 fi

 # --- Route: multi-node vs single-node ---
+clear_ci_orchestration_env
 if is_multi_node "$commands"; then
  echo "--- Multi-node job detected"
  export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
@@ -535,6 +895,13 @@ 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
@@ -551,7 +918,7 @@ else
  fi

  docker run \
-    -t -i \
+    "${docker_run_terminal_args[@]}" \
    --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
    $RDMA_FLAGS \
    --network=host \
@@ -566,6 +933,13 @@ else
    -e AWS_SECRET_ACCESS_KEY \
    -e BUILDKITE_PARALLEL_JOB \
    -e BUILDKITE_PARALLEL_JOB_COUNT \
+    -e TERM \
+    -e FORCE_COLOR \
+    -e CLICOLOR_FORCE \
+    -e PY_COLORS \
+    -e PYTHONFAULTHANDLER \
+    -e PYTEST_ADDOPTS \
+    -e PYTEST_TIMEOUT \
    -v "${HF_CACHE}:${HF_MOUNT}" \
    -e "HF_HOME=${HF_MOUNT}" \
    -e "PYTHONPATH=${MYPYTHONPATH}" \
@@ -575,7 +949,7 @@ else
    -e "VLLM_CACHE_ROOT=${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}" \
+    "${standalone_merge_base_env[@]}" \
    --name "${container_name}" \
    "${image_name}" \
    /bin/bash -c "${CONTAINER_PREFLIGHT} && ${commands}"
@@ -35,6 +35,7 @@ case "${test_suite}" in
    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
    pytest -v -s v1/structured_output
    pytest -v -s v1/test_serial_utils.py
+    pytest -v -s v1/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]"
    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py
    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py
    ;;
@@ -130,6 +130,22 @@ 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
@@ -0,0 +1,32 @@
+#!/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: ${CI_BASE_PUSH_STABLE_TAG}"
+    fi
+
+    bash .buildkite/scripts/ci-bake-rocm.sh ci-base-rocm-ci-with-deps
+}
+
+main "$@"
@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+# Build the ROCm CI test image or wheel artifact.
+#
+# When Dockerfile.rocm_base changes, always build the full image so downstream
+# ROCm tests can validate the freshly rebuilt base -> ci_base -> ci image chain.
+
+set -euo pipefail
+
+metadata_get() {
+    local key="$1"
+    if command -v buildkite-agent >/dev/null 2>&1; then
+        buildkite-agent meta-data get "${key}" 2>/dev/null || true
+    fi
+}
+
+use_ci_base_if_present() {
+    local ci_base_image=""
+
+    ci_base_image="$(metadata_get rocm-ci-base-image)"
+    if [[ -z "${ci_base_image}" ]]; then
+        return 1
+    fi
+
+    export CI_BASE_IMAGE="${ci_base_image}"
+    echo "Using ROCm ci_base image selected by the preceding build step: ${CI_BASE_IMAGE}"
+}
+
+use_refreshed_base_if_present() {
+    local base_refreshed=""
+
+    base_refreshed="$(metadata_get rocm-base-refresh)"
+    if [[ "${base_refreshed}" != "1" ]]; then
+        return 1
+    fi
+
+    export BASE_IMAGE
+    export IMAGE_TAG_LATEST
+
+    BASE_IMAGE="$(metadata_get rocm-base-image)"
+    IMAGE_TAG_LATEST="$(metadata_get rocm-ci-image-descriptive)"
+
+    echo "Using refreshed ROCm base image for test image: ${BASE_IMAGE}"
+    if [[ -n "${IMAGE_TAG_LATEST}" ]]; then
+        echo "Also tagging full ROCm CI image as: ${IMAGE_TAG_LATEST}"
+    fi
+
+    return 0
+}
+
+main() {
+    local base_refreshed=0
+
+    use_ci_base_if_present || true
+
+    if use_refreshed_base_if_present; then
+        base_refreshed=1
+    fi
+
+    if [[ "${ROCM_CI_ARTIFACT_ONLY:-0}" == "1" && "${base_refreshed}" != "1" ]]; then
+        echo "ROCM_CI_ARTIFACT_ONLY=1; building ROCm wheel artifact only"
+        IMAGE_TAG="" bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-artifacts
+        return
+    fi
+
+    bash .buildkite/scripts/ci-bake-rocm.sh test-rocm-ci-with-wheel
+}
+
+main "$@"
@@ -0,0 +1,513 @@
+#!/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 "$@"
@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# Fast structural smoke test for the full ROCm CI image.
+
+set -euo pipefail
+
+image_ref="${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
+'
@@ -21,16 +21,20 @@ export CARGO_HOME="${CARGO_HOME:-$HOME/.cargo}"
 export RUSTUP_HOME="${RUSTUP_HOME:-$HOME/.rustup}"
 export PATH="$CARGO_HOME/bin:$PATH"

+PROTOC_VERSION="${PROTOC_VERSION:-31.1}"
+CARGO_BINSTALL_VERSION="${CARGO_BINSTALL_VERSION:-1.20.1}"
+UV_VERSION="${UV_VERSION:-0.11.28}"
+PYO3_PYTHON_VERSION="${PYO3_PYTHON_VERSION:-3.12}"
+
+CARGO_SORT_VERSION_REQ="${CARGO_SORT_VERSION_REQ:-2}"
+CARGO_DENY_VERSION_REQ="${CARGO_DENY_VERSION_REQ:-0.20}"
+CARGO_NEXTEST_VERSION_REQ="${CARGO_NEXTEST_VERSION_REQ:-0.9}"
+
 log_section() {
  echo "--- $*"
 }

 install_protoc() {
-  if command -v protoc >/dev/null 2>&1; then
-    return
-  fi
-
-  local version="${PROTOC_VERSION:-31.1}"
  local arch
  case "$(uname -m)" in
    x86_64)
@@ -45,16 +49,17 @@ install_protoc() {
      ;;
  esac

-  local url="https://github.com/protocolbuffers/protobuf/releases/download/v${version}/protoc-${version}-linux-${arch}.zip"
+  local url="https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-${arch}.zip"
  local tmp_dir
  tmp_dir="$(mktemp -d)"

-  log_section "Installing protoc ${version}"
+  log_section "Installing protoc ${PROTOC_VERSION}"
  curl -L --proto '=https' --tlsv1.2 -sSf "$url" -o "$tmp_dir/protoc.zip"
  mkdir -p "$CARGO_HOME/bin"
  unzip -q "$tmp_dir/protoc.zip" bin/protoc 'include/*' -d "$CARGO_HOME"
  chmod +x "$CARGO_HOME/bin/protoc"
  rm -rf "$tmp_dir"
+  protoc --version
 }

 rust_toolchain() {
@@ -75,66 +80,48 @@ install_rust_toolchain() {
 }

 install_cargo_binstall() {
-  if command -v cargo-binstall >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-binstall"
+  log_section "Installing cargo-binstall ${CARGO_BINSTALL_VERSION}"
  curl -L --proto '=https' --tlsv1.2 -sSf \
-    https://raw.githubusercontent.com/cargo-bins/cargo-binstall/main/install-from-binstall-release.sh \
-    | bash
+    "https://raw.githubusercontent.com/cargo-bins/cargo-binstall/v${CARGO_BINSTALL_VERSION}/install-from-binstall-release.sh" \
+    | env BINSTALL_VERSION="$CARGO_BINSTALL_VERSION" bash
+  cargo-binstall -V
 }

 install_cargo_sort() {
-  if command -v cargo-sort >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-sort"
-  install_cargo_binstall
-  cargo binstall --no-confirm cargo-sort
+  log_section "Installing cargo-sort ${CARGO_SORT_VERSION_REQ}"
+  cargo binstall --no-confirm --force "cargo-sort@${CARGO_SORT_VERSION_REQ}"
 }

 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
+  log_section "Installing cargo-deny ${CARGO_DENY_VERSION_REQ}"
+  cargo binstall --no-confirm --force "cargo-deny@${CARGO_DENY_VERSION_REQ}"
 }

 install_cargo_nextest() {
-  if command -v cargo-nextest >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing cargo-nextest"
-  install_cargo_binstall
-  cargo binstall --no-confirm --secure cargo-nextest
+  log_section "Installing cargo-nextest ${CARGO_NEXTEST_VERSION_REQ}"
+  cargo binstall \
+    --no-confirm \
+    --force \
+    --secure \
+    "cargo-nextest@${CARGO_NEXTEST_VERSION_REQ}"
 }

 install_uv() {
-  if command -v uv >/dev/null 2>&1; then
-    return
-  fi
-
-  log_section "Installing uv"
-  curl -LsSf --proto '=https' --tlsv1.2 https://astral.sh/uv/install.sh \
+  log_section "Installing uv ${UV_VERSION}"
+  curl -L --proto '=https' --tlsv1.2 -sSf \
+    "https://github.com/astral-sh/uv/releases/download/${UV_VERSION}/uv-installer.sh" \
    | env UV_INSTALL_DIR="$CARGO_HOME/bin" sh
+  uv --version
 }

 setup_pyo3_python() {
-  local python_version="${PYO3_PYTHON_VERSION:-3.12}"
-
-  log_section "Installing Python ${python_version} for PyO3 tests"
-  uv python install "$python_version"
+  log_section "Installing Python ${PYO3_PYTHON_VERSION} for PyO3 tests"
+  uv python install "$PYO3_PYTHON_VERSION"
  PYO3_PYTHON="$(uv python find \
    --managed-python \
    --no-project \
    --resolve-links \
-    "$python_version")"
+    "$PYO3_PYTHON_VERSION")"
  export PYO3_PYTHON

  local python_libdir
@@ -156,6 +143,7 @@ PY
 }

 run_style_clippy() {
+  install_cargo_binstall
  install_cargo_sort
  install_cargo_deny

@@ -168,8 +156,8 @@ run_style_clippy() {
  log_section "Checking Rust dependency bans"
  cargo deny \
    --manifest-path rust/Cargo.toml \
-    check \
    --config rust/deny.toml \
+    check \
    bans

  log_section "Running clippy"
@@ -186,6 +174,7 @@ run_style_clippy() {
 run_tests() {
  install_uv
  setup_pyo3_python
+  install_cargo_binstall
  install_cargo_nextest

  log_section "Running cargo nextest"
@@ -23,6 +23,7 @@ NC='\033[0m' # No Color
 # Default configuration
 PIPELINE="ci"
 DRY_RUN=true
+TORCH_NIGHTLY=false

 usage() {
    cat <<EOF
@@ -34,12 +35,14 @@ 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)
-    --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)
+    --torch-nightly  Also build and run the full suite against torch nightly
+                     (sets TORCH_NIGHTLY=1)
+    --help           Show this help message

 Prerequisites:
    - bk CLI installed: brew tap buildkite/buildkite && brew install buildkite/buildkite/bk
@@ -49,6 +52,7 @@ 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
 }
@@ -96,6 +100,10 @@ while [[ $# -gt 0 ]]; do
            MESSAGE="$2"
            shift 2
            ;;
+        --torch-nightly)
+            TORCH_NIGHTLY=true
+            shift
+            ;;
        --help|-h)
            usage
            ;;
@@ -171,11 +179,17 @@ 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: RUN_ALL=1, NIGHTLY=1"
+log_info "Environment: ${BUILD_ENV[*]}"
 echo ""

 # Build the command
@@ -187,9 +201,10 @@ 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 [[ "$DRY_RUN" == true ]]; then
    echo "=========================================="
@@ -210,8 +225,14 @@ if [[ "$DRY_RUN" == true ]]; then
    echo "    --commit '$(escape_for_shell "${COMMIT}")' \\"
    echo "    --branch '$(escape_for_shell "${BRANCH}")' \\"
    echo "    --message '$(escape_for_shell "${MESSAGE}")' \\"
-    echo "    --env 'RUN_ALL=1' \\"
-    echo "    --env 'NIGHTLY=1'"
+    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[$i]}")' \\"
+        fi
+    done
    echo ""
    echo "=========================================="
    echo -e "${YELLOW}To actually trigger this build, run:${NC}"
@@ -6,8 +6,14 @@ set -ex
 # manylinux platform tag with auditwheel.
 # Index generation is handled separately by generate-and-upload-nightly-index.sh.

-# shellcheck source=lib/manylinux.sh
-source .buildkite/scripts/lib/manylinux.sh
+# auditwheel is Linux-only; macOS wheels already carry a valid tag, so skip the
+# manylinux retag for them.
+WHEEL_PLATFORM="${VLLM_WHEEL_PLATFORM:-linux}"
+
+if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
+  # shellcheck source=lib/manylinux.sh
+  source .buildkite/scripts/lib/manylinux.sh
+fi

 BUCKET="vllm-wheels"
 SUBPATH=$BUILDKITE_COMMIT
@@ -27,8 +33,10 @@ wheel="${wheel_files[0]}"

 # ========= detect manylinux tag and rename ==========

-wheel="$(apply_manylinux_tag "$wheel")"
-echo "Renamed wheel to: $wheel"
+if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
+  wheel="$(apply_manylinux_tag "$wheel")"
+  echo "Renamed wheel to: $wheel"
+fi

 # Extract the version from the wheel
 version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
@@ -0,0 +1,13 @@
+#!/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}:${PLATFORM_TAG}
@@ -0,0 +1,21 @@
+#!/bin/bash
+
+set -ex
+
+ORIG_TAG_NAME="$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: 30
+  timeout_in_minutes: 85
  device: h100
  source_file_dependencies:
    - vllm/config/attention.py
@@ -12,11 +12,12 @@ steps:
    - vllm/v1/attention
    - tests/v1/attention
  commands:
-    - pytest -v -s v1/attention
+    - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 70
+      timeout_in_minutes: 95
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -30,7 +31,7 @@ steps:

 - label: V1 attention (B200)
  key: v1-attention-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 80
  device: b200-k8s
  source_file_dependencies:
    - vllm/config/attention.py
@@ -38,4 +39,5 @@ steps:
    - vllm/v1/attention
    - tests/v1/attention
  commands:
-    - pytest -v -s v1/attention
+    - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Correctness
  key: basic-correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -19,6 +19,6 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 50
+      timeout_in_minutes: 70
      depends_on:
      - image-build-amd
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Benchmarks CLI Test
  key: benchmarks-cli-test
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -13,6 +13,7 @@ steps:
  - pytest -v -s benchmarks/
  mirror:
    amd:
+      dind: false
      device: mi300_1
      depends_on:
      - image-build-amd
@@ -23,7 +24,7 @@ steps:
  num_gpus: 2
  optional: true
  working_dir: "/vllm-workspace/"
-  timeout_in_minutes: 10
+  timeout_in_minutes: 20
  source_file_dependencies:
  - benchmarks/attention_benchmarks/
  - vllm/v1/attention/
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Sequence Parallel Correctness Tests (2 GPUs)
  key: sequence-parallel-correctness-tests-2-gpus
-  timeout_in_minutes: 50
+  timeout_in_minutes: 80
  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: 50
+  timeout_in_minutes: 75
  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: 50
+  timeout_in_minutes: 30
  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: 50
+  timeout_in_minutes: 30
  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: 20
+  timeout_in_minutes: 45
  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: 20
+  timeout_in_minutes: 30
  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: 15
+  timeout_in_minutes: 25
  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: 30
+  timeout_in_minutes: 25
  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: 20
+  timeout_in_minutes: 35
  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: 40
+  timeout_in_minutes: 30
  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: 20
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  num_devices: 2
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Platform Tests
  key: platform-tests
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - vllm/envs.py
@@ -19,7 +19,7 @@ steps:

 - label: Cudagraph
  key: cudagraph
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  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: 30
+  timeout_in_minutes: 55
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -15,8 +15,9 @@ steps:
    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
+      dind: false
      device: mi300_4
-      timeout_in_minutes: 110
+      timeout_in_minutes: 85
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -29,7 +30,7 @@ steps:

 - label: Distributed FlashInfer NixlConnector PD accuracy (4 GPUs)
  key: distributed-flashinfer-nixlconnector-pd-accuracy-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -39,6 +40,19 @@ 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
@@ -52,8 +66,9 @@ steps:
    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
+      dind: false
      device: mi300_4
-      timeout_in_minutes: 50
+      timeout_in_minutes: 60
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -66,7 +81,7 @@ steps:

 - label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
  key: crosslayer-kv-layout-distributed-nixlconnector-pd-accuracy-tests-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -77,8 +92,9 @@ steps:
    - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
+      dind: false
      device: mi300_4
-      timeout_in_minutes: 110
+      timeout_in_minutes: 85
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -91,7 +107,7 @@ steps:

 - label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs)
  key: hybrid-ssm-nixlconnector-pd-accuracy-tests-4-gpus
-  timeout_in_minutes: 25
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -102,8 +118,9 @@ steps:
    - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
  mirror:
    amd:
+      dind: false
      device: mi300_4
-      timeout_in_minutes: 60
+      timeout_in_minutes: 80
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -130,7 +147,7 @@ steps:

 - label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
  key: multiconnector-nixl-offloading-pd-accuracy-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -145,7 +162,7 @@ steps:

 - label: NixlConnector PD + Spec Decode acceptance (2 GPUs)
  key: nixlconnector-pd-spec-decode-acceptance-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: a100
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
@@ -158,8 +175,9 @@ steps:
    - bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
  mirror:
    amd:
+      dind: false
      device: mi300_2
-      timeout_in_minutes: 60
+      timeout_in_minutes: 70
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -169,11 +187,11 @@ steps:
        - 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
+        - KV_CACHE_MEMORY_BYTES=8G 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: 30
+  timeout_in_minutes: 25
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -185,3 +203,25 @@ steps:
  commands:
    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
    - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
+
+# P TP 4 - D DPEP 4 test case for DSv4-Flash
+- label: DSv4-Flash Disaggregated DP EP
+  key: dsv4-flash-disaggregated
+  timeout_in_minutes: 60
+  device: h200
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 8
+  env:
+    ENABLE_HMA_FLAG: "1"
+    DP_EP: "1"
+    GPU_MEMORY_UTILIZATION: "0.85"
+    PREFILLER_TP_SIZE: "4"
+    DECODER_TP_SIZE: "4"
+    PREFILL_BLOCK_SIZE: "256"
+    DECODE_BLOCK_SIZE: "256"
+    MODEL_NAMES: "deepseek-ai/DeepSeek-V4-Flash"
+    VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--kv-cache-dtype,fp8"
+  commands:
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Comm Ops
  key: distributed-comm-ops
-  timeout_in_minutes: 20
+  timeout_in_minutes: 25
  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: 20
+  timeout_in_minutes: 35
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -39,6 +39,7 @@ steps:
  - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
  mirror:
    amd:
+      dind: false
      device: mi300_2
      depends_on:
      - image-build-amd
@@ -55,7 +56,7 @@ steps:

 - label: Distributed Compile + RPC Tests (2 GPUs)
  key: distributed-compile-rpc-tests-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 65
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -78,7 +79,7 @@ steps:

 - label: Distributed Torchrun + Shutdown Tests (2 GPUs)
  key: distributed-torchrun-shutdown-tests-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -133,7 +134,7 @@ steps:

 - label: Distributed DP Tests (4 GPUs)
  key: distributed-dp-tests-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -154,7 +155,7 @@ steps:

 - label: Distributed Compile + Comm (4 GPUs)
  key: distributed-compile-comm-4-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 70
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -176,7 +177,7 @@ steps:

 - label: Distributed Tests (8xH100)
  key: distributed-tests-8xh100
-  timeout_in_minutes: 10
+  timeout_in_minutes: 20
  device: h100
  num_devices: 8
  working_dir: "/vllm-workspace/tests"
@@ -212,7 +213,7 @@ steps:

 - label: Distributed Tests (2xH100-2xMI300)
  key: distributed-tests-2xh100-2xmi300
-  timeout_in_minutes: 15
+  timeout_in_minutes: 30
  device: h100
  optional: true
  working_dir: "/vllm-workspace/"
@@ -259,7 +260,7 @@ steps:

 - label: Pipeline + Context Parallelism (4 GPUs)
  key: pipeline-context-parallelism-4-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 55
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -274,7 +275,7 @@ steps:

 - label: RayExecutorV2 (4 GPUs)
  key: rayexecutorv2-4-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 45
  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: 10
+  timeout_in_minutes: 20
  device: cpu-small
  source_file_dependencies:
    - .buildkite/release-pipeline.yaml
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: DeepSeek V2-Lite Sync EPLB Accuracy (4xH100)
  key: deepseek-v2-lite-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
  device: h100
  optional: true
  num_devices: 4
@@ -14,7 +14,7 @@ steps:

 - label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (4xH100)
  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-4xh100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
  device: h100
  optional: true
  num_devices: 4
@@ -24,7 +24,7 @@ steps:

 - label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (2xB200)
  key: qwen3-30b-a3b-fp8-block-sync-eplb-accuracy-2xb200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 20
  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: 60
+  timeout_in_minutes: 25
  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: 60
+  timeout_in_minutes: 20
  device: h100
  optional: true
  num_devices: 1
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Engine
  key: engine
-  timeout_in_minutes: 15
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/compilation/
@@ -29,13 +29,13 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 60
+      timeout_in_minutes: 50
      depends_on:
      - image-build-amd

 - label: Engine (1 GPU)
  key: engine-1-gpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  source_file_dependencies:
    - vllm/v1/engine/
    - tests/v1/engine/
@@ -45,13 +45,13 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 40
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd

 - label: e2e Scheduling (1 GPU)
  key: e2e-scheduling-1-gpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 35
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/
@@ -60,15 +60,15 @@ steps:
    - pytest -v -s v1/e2e/general/test_async_scheduling.py
  mirror:
    amd:
-      device: mi250_1
-      timeout_in_minutes: 60
+      device: mi325_1
+      timeout_in_minutes: 70
      depends_on:
      - image-build-amd

 - label: e2e Core (1 GPU)
  device: h200_35gb
  key: e2e-core-1-gpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  source_file_dependencies:
    - vllm/v1/
    - tests/v1/e2e/general/
@@ -76,8 +76,8 @@ steps:
    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
  mirror:
    amd:
-      device: mi250_1
-      timeout_in_minutes: 35
+      device: mi325_1
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -87,7 +87,7 @@ steps:

 - label: V1 e2e (2 GPUs)
  key: v1-e2e-2-gpus
-  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
+  timeout_in_minutes: 25 # TODO: Fix timeout after we have more confidence in the test stability
  optional: true
  num_devices: 2
  source_file_dependencies:
@@ -114,13 +114,14 @@ steps:
    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
  mirror:
    amd:
+      dind: false
      device: mi300_2
      depends_on:
      - image-build-amd

 - label: V1 e2e (4 GPUs)
  key: v1-e2e-4-gpus
-  timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
+  timeout_in_minutes: 20 # TODO: Fix timeout after we have more confidence in the test stability
  optional: true
  num_devices: 4
  source_file_dependencies:
@@ -148,7 +149,7 @@ steps:

 - label: V1 e2e (4xH100)
  key: v1-e2e-4xh100
-  timeout_in_minutes: 60
+  timeout_in_minutes: 35
  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: 10
+  timeout_in_minutes: 25
  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: 40
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -37,7 +37,7 @@ steps:
 - label: Entrypoints Integration (API Server)
  key: entrypoints-integration-api-server
  device: h200_35gb
-  timeout_in_minutes: 130
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -56,7 +56,7 @@ steps:

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

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

@@ -126,7 +126,7 @@ steps:
 - label: Entrypoints Integration (Speech to Text)
  device: h200_35gb
  key: entrypoints-integration-speech_to_text
-  timeout_in_minutes: 50
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -138,7 +138,7 @@ steps:
 - label: Entrypoints Integration (Multimodal)
  device: h200_35gb
  key: entrypoints-integration-multimodal
-  timeout_in_minutes: 50
+  timeout_in_minutes: 45
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -160,7 +160,7 @@ steps:

 - label: OpenAI API Correctness
  key: openai-api-correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
  - csrc/
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: EPLB Algorithm
  key: eplb-algorithm
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -16,6 +16,7 @@ steps:
  - pytest -v -s distributed/test_eplb_utils.py
  mirror:
    amd:
+      dind: false
      device: mi300_1
      depends_on:
      - image-build-amd
@@ -27,7 +28,7 @@ steps:

 - label: EPLB Execution # 17min
  key: eplb-execution
-  timeout_in_minutes: 27
+  timeout_in_minutes: 25
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -39,7 +40,7 @@ steps:

 - label: Elastic EP Scaling Test
  key: elastic-ep-scaling-test
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  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: 10
+  timeout_in_minutes: 35
  device: h200_18gb
  working_dir: "/vllm-workspace/"
  source_file_dependencies:
@@ -16,18 +16,19 @@ steps:

 - label: Kernels Core Operation Test
  key: kernels-core-operation-test
-  timeout_in_minutes: 75
+  timeout_in_minutes: 120
  source_file_dependencies:
  - csrc/
  - tests/kernels/core
  - tests/kernels/test_concat_mla_q.py
  - tests/kernels/test_fused_qk_norm_rope_gate.py
  commands:
-    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py
+    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 3

 - label: Kernels MiniMax Reduce RMS Test (2 GPUs)
  key: kernels-minimax-reduce-rms-test-2-gpus
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
  num_devices: 2
  device: h100
  source_file_dependencies:
@@ -41,7 +42,7 @@ steps:

 - label: Deepseek V4 Kernel Test (H100)
  key: deepseek-v4-kernel-test-h100
-  timeout_in_minutes: 15
+  timeout_in_minutes: 30
  device: h100
  source_file_dependencies:
  - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
@@ -54,7 +55,7 @@ steps:

 - label: Deepseek V4 Kernel Test (B200)
  key: deepseek-v4-kernel-test-b200
-  timeout_in_minutes: 15
+  timeout_in_minutes: 20
  device: b200-k8s
  source_file_dependencies:
  - csrc/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
@@ -65,7 +66,7 @@ steps:

 - label: Kernels Attention Test %N
  key: kernels-attention-test
-  timeout_in_minutes: 35
+  timeout_in_minutes: 65
  source_file_dependencies:
  - csrc/attention/
  - vllm/v1/attention
@@ -79,7 +80,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 55
+      timeout_in_minutes: 90
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -106,7 +107,7 @@ steps:

 - label: Kernels Quantization Test %N
  key: kernels-quantization-test
-  timeout_in_minutes: 90
+  timeout_in_minutes: 60
  source_file_dependencies:
  - csrc/quantization/
  - vllm/model_executor/layers/quantization
@@ -131,7 +132,7 @@ steps:

 - label: Kernels MoE Test %N
  key: kernels-moe-test
-  timeout_in_minutes: 25
+  timeout_in_minutes: 50
  source_file_dependencies:
  - csrc/quantization/cutlass_w8a8/moe/
  - csrc/moe/
@@ -147,7 +148,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 50
+      timeout_in_minutes: 65
      source_file_dependencies:
      - csrc/quantization/cutlass_w8a8/moe/
      - csrc/moe/
@@ -163,7 +164,7 @@ steps:

 - label: Kernels Mamba Test
  key: kernels-mamba-test
-  timeout_in_minutes: 45
+  timeout_in_minutes: 40
  source_file_dependencies:
  - csrc/mamba/
  - tests/kernels/mamba
@@ -172,19 +173,19 @@ steps:
    - pytest -v -s kernels/mamba

 - label: Kernels KDA Test
-  timeout_in_minutes: 20
+  timeout_in_minutes: 25
  device: h200_18gb
  source_file_dependencies:
-  - vllm/model_executor/layers/fla/ops/kda.py
-  - vllm/model_executor/layers/fla/ops/chunk_delta_h.py
-  - vllm/model_executor/layers/fla/ops/l2norm.py
+  - vllm/third_party/flash_linear_attention/ops/kda.py
+  - vllm/third_party/flash_linear_attention/ops/chunk_delta_h.py
+  - vllm/third_party/flash_linear_attention/ops/l2norm.py
  - tests/kernels/test_kda.py
  commands:
    - pytest -v -s kernels/test_kda.py

 - label: Kernels DeepGEMM Test (H100)
  key: kernels-deepgemm-test-h100
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
  device: h100
  num_devices: 1
  source_file_dependencies:
@@ -211,7 +212,7 @@ steps:

 - label: Kernels (B200)
  key: kernels-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 80
  working_dir: "/vllm-workspace/"
  device: b200-k8s
  # optional: true
@@ -231,6 +232,10 @@ steps:
  - vllm/v1/attention/backends/mla/flashinfer_mla.py
  - vllm/v1/attention/selector.py
  - vllm/platforms/cuda.py
+  - vllm/model_executor/kernels/linear/cute_dsl/ll_bf16.py
+  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_dotprod.py
+  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_splitk.py
+  - tests/kernels/test_ll_bf16_gemm.py
  - tests/kernels/test_top_k_per_row.py
  commands:
    - nvidia-smi
@@ -259,24 +264,26 @@ steps:
    - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py
    - pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py
    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
+    - pytest -v -s tests/kernels/test_ll_bf16_gemm.py
    # e2e
    - pytest -v -s tests/models/quantization/test_nvfp4.py

 - label: Kernels Helion Test
  key: kernels-helion-test
-  timeout_in_minutes: 30
+  timeout_in_minutes: 115
  device: h100
  source_file_dependencies:
  - vllm/utils/import_utils.py
  - tests/kernels/helion/
  commands:
    - pip install helion==1.1.0
-    - pytest -v -s kernels/helion/
+    - pytest -v -s kernels/helion/ --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2

 
 - label: Kernels FP8 MoE Test (1xH100)
  key: kernels-fp8-moe-test-1xh100
-  timeout_in_minutes: 90
+  timeout_in_minutes: 40
  device: h100
  num_devices: 1
  optional: true
@@ -293,7 +300,7 @@ steps:

 - label: Kernels FP8 MoE Test (2xH100)
  key: kernels-fp8-moe-test-2xh100
-  timeout_in_minutes: 90
+  timeout_in_minutes: 45
  device: h100
  num_devices: 2
  optional: true
@@ -303,7 +310,7 @@ steps:

 - label: Kernels Fp4 MoE Test (B200)
  key: kernels-fp4-moe-test-b200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
  device: b200-k8s
  num_devices: 1
  optional: true
@@ -316,7 +323,7 @@ steps:

 - label: Kernels FusedMoE Layer Test (2 H100s)
  key: kernels-fusedmoe-layer-test-2-h100s
-  timeout_in_minutes: 90
+  timeout_in_minutes: 30
  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: 75
+  timeout_in_minutes: 45
  source_file_dependencies:
  - csrc/
  - vllm/model_executor/layers/quantization
@@ -56,7 +56,7 @@ steps:

 - label: LM Eval Small Models (1xB200)
  key: lm-eval-small-models-1xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 50
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -80,7 +80,7 @@ steps:

 - label: LM Eval Large Models EP (2xB200)
  key: lm-eval-large-models-ep-2xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 60
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -92,7 +92,7 @@ steps:

 - label: LM Eval Qwen3.5 Models (2xB200)
  key: lm-eval-qwen3-5-models-2xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 45
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -103,13 +103,13 @@ steps:
  - vllm/transformers_utils/configs/qwen3_5_moe.py
  - vllm/model_executor/models/qwen3_next.py
  - vllm/model_executor/models/qwen3_next_mtp.py
-  - vllm/model_executor/layers/fla/ops/
+  - vllm/third_party/flash_linear_attention/ops/
  commands:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt

 - label: LM Eval Large Models (8xH200)
  key: lm-eval-large-models-8xh200
-  timeout_in_minutes: 60
+  timeout_in_minutes: 50
  device: h200
  optional: true
  num_devices: 8
@@ -117,8 +117,9 @@ steps:
    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
  mirror:
    amd:
+      dind: false
      device: mi300_8
-      timeout_in_minutes: 180
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      commands:
@@ -152,7 +153,7 @@ steps:

 - label: LM Eval Humming f16 (A100 - TEMPORARY)
  key: lm-eval-humming-f16-a100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 75
  device: a100
  optional: true
  num_devices: 1
@@ -167,7 +168,7 @@ steps:

 - label: LM Eval Humming Act int8 (A100 - TEMPORARY)
  key: lm-eval-humming-act-a100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 45
  device: a100
  optional: true
  num_devices: 1
@@ -182,7 +183,7 @@ steps:

 - label: LM Eval Humming f16 (H100 - TEMPORARY)
  key: lm-eval-humming-f16-h100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 70
  device: h100
  optional: true
  num_devices: 1
@@ -197,7 +198,7 @@ steps:

 - label: LM Eval Humming Act fp8/int8 (H100 - TEMPORARY)
  key: lm-eval-humming-act-h100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 70
  device: h100
  optional: true
  num_devices: 1
@@ -213,7 +214,7 @@ steps:

 - label: LM Eval Humming f16 (B200 - TEMPORARY)
  key: lm-eval-humming-f16-b200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 50
  device: b200-k8s
  optional: true
  num_devices: 1
@@ -228,7 +229,7 @@ steps:

 - label: LM Eval Humming Act fp8/int8 (B200 - TEMPORARY)
  key: lm-eval-humming-act-b200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 50
  device: b200-k8s
  optional: true
  num_devices: 1
@@ -244,7 +245,7 @@ steps:

 - label: LM Eval TurboQuant KV Cache
  key: lm-eval-turboquant-kv-cache
-  timeout_in_minutes: 75
+  timeout_in_minutes: 55
  device: h200_18gb
  source_file_dependencies:
  - vllm/model_executor/layers/quantization/turboquant/
@@ -256,7 +257,7 @@ steps:

 - label: GPQA Eval (GPT-OSS) (2xH100)
  key: gpqa-eval-gpt-oss-2xh100
-  timeout_in_minutes: 120
+  timeout_in_minutes: 35
  device: h100
  optional: true
  num_devices: 2
@@ -270,7 +271,7 @@ steps:

 - label: GPQA Eval (GPT-OSS) (2xB200)
  key: gpqa-eval-gpt-oss-2xb200
-  timeout_in_minutes: 120
+  timeout_in_minutes: 30
  device: b200-k8s
  optional: true
  num_devices: 2
@@ -284,7 +285,7 @@ steps:

 - label: GPQA Eval (GPT-OSS) (DGX Spark)
  key: gpqa-eval-gpt-oss-spark
-  timeout_in_minutes: 120
+  timeout_in_minutes: 35
  device: dgx-spark
  optional: true
  num_devices: 1
@@ -298,9 +299,50 @@ steps:
    - uv pip install --system 'gpt-oss[eval]==0.0.5'
    - pytest -s -v evals/gpt_oss/test_gpqa_correctness.py --config-list-file=configs/models-spark.txt

+- label: LM Eval KV-Offload (1xH200)
+  key: kv-offload-small
+  timeout_in_minutes: 30
+  device: h200_35gb
+  source_file_dependencies:
+  - vllm/distributed/kv_transfer/kv_connector/v1/offloading/
+  - vllm/distributed/kv_transfer/kv_connector/v1/simple_cpu_offload_connector.py
+  - vllm/v1/kv_offload/
+  - vllm/v1/simple_kv_offload/
+  - tests/evals/gsm8k/test_gsm8k_offloading.py
+  commands:
+  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "nemotron-h-8b or gemma-4-e4b-it"
+
+- label: LM Eval KV-Offload (2xH100)
+  key: kv-offload-medium
+  timeout_in_minutes: 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: 30
+  timeout_in_minutes: 25
  source_file_dependencies:
  - tests/evals/mrcr/
  commands:
@@ -5,7 +5,7 @@ steps:
 - label: LoRA %N
  device: h200_35gb
  key: lora
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  source_file_dependencies:
  - vllm/lora
  - tests/lora
@@ -16,7 +16,7 @@ steps:
    amd:
      device: mi325_1
      working_dir: "/vllm-workspace/tests"
-      timeout_in_minutes: 60
+      timeout_in_minutes: 65
      source_file_dependencies:
      - vllm/lora
      - tests/lora
@@ -27,7 +27,7 @@ steps:

 - label: LoRA TP (Distributed)
  key: lora-tp-distributed
-  timeout_in_minutes: 30
+  timeout_in_minutes: 60
  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: 30
+  timeout_in_minutes: 40
  source_file_dependencies:
    - vllm/config/
    - vllm/distributed/
@@ -23,14 +23,15 @@ steps:
    - pytest -v -s -m 'not slow_test' v1/spec_decode
  mirror:
    amd:
+      dind: false
      device: mi300_1
-      timeout_in_minutes: 65
+      timeout_in_minutes: 75
      depends_on:
      - image-build-amd

 - label: V1 Sample + Logits
  key: v1-sample-logits
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: h200_18gb
  source_file_dependencies:
    - vllm/config/
@@ -64,7 +65,7 @@ steps:

 - label: V1 Core + KV + Metrics
  key: v1-core-kv-metrics
-  timeout_in_minutes: 30
+  timeout_in_minutes: 60
  source_file_dependencies:
    - vllm/config/
    - vllm/distributed/
@@ -89,6 +90,7 @@ steps:
    - tests/v1/simple_kv_offload
    - tests/v1/worker
    - tests/v1/kv_connector/unit
+    - tests/v1/ec_connector/unit
    - tests/v1/metrics
    - tests/entrypoints/openai/correctness/test_lmeval.py
  commands:
@@ -101,6 +103,7 @@ steps:
    - pytest -v -s v1/simple_kv_offload
    - pytest -v -s v1/worker
    - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
+    - pytest -v -s -m 'not cpu_test' v1/ec_connector/unit
    - pytest -v -s -m 'not cpu_test' v1/metrics
    # Integration test for streaming correctness (requires special branch).
    - pip install -U git+https://github.com/vllm-project/lm-evaluation-harness.git@streaming-api
@@ -108,7 +111,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 60
+      timeout_in_minutes: 75
      depends_on:
      - image-build-amd

@@ -172,7 +175,7 @@ steps:

 - label: Regression
  key: regression
-  timeout_in_minutes: 20
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/config/
@@ -195,7 +198,7 @@ steps:
 - label: Examples
  device: h200_35gb
  key: examples
-  timeout_in_minutes: 45
+  timeout_in_minutes: 40
  working_dir: "/vllm-workspace/examples"
  source_file_dependencies:
  - vllm/entrypoints
@@ -237,7 +240,7 @@ steps:

 - label: Metrics, Tracing (2 GPUs)
  key: metrics-tracing-2-gpus
-  timeout_in_minutes: 20
+  timeout_in_minutes: 25
  num_devices: 2
  source_file_dependencies:
  - vllm/config/
@@ -281,7 +284,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 20
+      timeout_in_minutes: 45
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -292,7 +295,7 @@ steps:
 - label: Async Engine, Inputs, Utils, Worker
  device: h200_35gb
  key: async-engine-inputs-utils-worker
-  timeout_in_minutes: 50
+  timeout_in_minutes: 25
  source_file_dependencies:
  - vllm/assets/
  - vllm/config/
@@ -319,7 +322,7 @@ steps:
  key: async-engine-inputs-utils-worker-config-cpu
  depends_on:
  - image-build-cpu
-  timeout_in_minutes: 30
+  timeout_in_minutes: 65
  source_file_dependencies:
  - vllm/assets/
  - vllm/config/
@@ -351,6 +354,7 @@ 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
@@ -368,6 +372,7 @@ 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
@@ -379,7 +384,7 @@ steps:

 - label: Batch Invariance (A100)
  key: batch-invariance-a100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  device: a100
  source_file_dependencies:
    - vllm/v1/attention
@@ -393,7 +398,7 @@ steps:

 - label: Batch Invariance (H100)
  key: batch-invariance-h100
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  device: h100
  source_file_dependencies:
    - vllm/v1/attention
@@ -409,7 +414,7 @@ steps:

 - label: Batch Invariance (B200)
  key: batch-invariance-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 35
  device: b200-k8s
  source_file_dependencies:
    - vllm/v1/attention
@@ -428,7 +433,7 @@ steps:
 - label: Acceptance Length Test (Large Models) # optional
  device: h200_35gb
  key: acceptance-length-test-large-models
-  timeout_in_minutes: 25
+  timeout_in_minutes: 20
  gpu: h100
  optional: true
  num_gpus: 1
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Model Executor
  key: model-executor
-  timeout_in_minutes: 35
+  timeout_in_minutes: 45
  source_file_dependencies:
  - vllm/engine/arg_utils.py
  - vllm/config/model.py
@@ -25,6 +25,7 @@ steps:
    - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py --timeout=900 --timeout-method=thread
  mirror:
    amd:
+      dind: false
      device: mi300_1
      depends_on:
      - image-build-amd
@@ -5,7 +5,7 @@ steps:
 - label: Model Runner V2 Core Tests
  device: h200_35gb
  key: model-runner-v2-core-tests
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
  source_file_dependencies:
  - vllm/v1/worker/gpu/
  - vllm/v1/worker/gpu_worker.py
@@ -27,7 +27,7 @@ steps:
 - label: Model Runner V2 Examples
  device: h200_35gb
  key: model-runner-v2-examples
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
  working_dir: "/vllm-workspace/examples"
  source_file_dependencies:
    - vllm/v1/worker/gpu/
@@ -63,7 +63,7 @@ steps:

 - label: Model Runner V2 Distributed (2 GPUs)
  key: model-runner-v2-distributed-2-gpus
-  timeout_in_minutes: 45
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -84,7 +84,7 @@ steps:

 - label: Model Runner V2 Pipeline Parallelism (4 GPUs)
  key: model-runner-v2-pipeline-parallelism-4-gpus
-  timeout_in_minutes: 60
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  num_devices: 4
  source_file_dependencies:
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Models Tests (Initialization)
  key: basic-models-tests-initialization
-  timeout_in_minutes: 45
+  timeout_in_minutes: 25
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -17,7 +17,7 @@ steps:
 - label: Basic Models Tests (Extra Initialization) %N
  device: h200_35gb
  key: basic-models-tests-extra-initialization
-  timeout_in_minutes: 45
+  timeout_in_minutes: 100
  source_file_dependencies:
  - vllm/model_executor/models/
  - tests/models/test_initialization.py
@@ -27,12 +27,12 @@ steps:
    # subset of supported models (the complement of the small subset in the above
    # test.) Also run if model initialization test file is modified
    - pytest -v -s models/test_initialization.py -k 'not test_can_initialize_small_subset' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
-  parallelism: 2
+  parallelism: 4

 - label: Basic Models Tests (Other)
  device: h200_35gb
  key: basic-models-tests-other
-  timeout_in_minutes: 45
+  timeout_in_minutes: 35
  source_file_dependencies:
  - vllm/
  - tests/models/test_terratorch.py
@@ -50,7 +50,7 @@ steps:
  key: basic-models-test-other-cpu
  depends_on:
  - image-build-cpu
-  timeout_in_minutes: 10
+  timeout_in_minutes: 20
  source_file_dependencies:
  - vllm/
  - tests/models/test_utils.py
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Model Tests (2 GPUs)
  key: distributed-model-tests-2-gpus
-  timeout_in_minutes: 50
+  timeout_in_minutes: 60
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Language Models Tests (Standard)
  key: language-models-tests-standard
-  timeout_in_minutes: 25
+  timeout_in_minutes: 30
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -15,13 +15,14 @@ steps:
    - pytest -v -s models/language -m 'core_model and (not slow_test)'
  mirror:
    amd:
+      dind: false
      device: mi300_1
      depends_on:
      - image-build-amd

 - label: Language Models Tests (Extra Standard) %N
  key: language-models-tests-extra-standard
-  timeout_in_minutes: 45
+  timeout_in_minutes: 40
  source_file_dependencies:
  - vllm/model_executor/models/
  - tests/models/language/pooling/test_embedding.py
@@ -35,6 +36,7 @@ steps:
  parallelism: 2
  mirror:
    amd:
+      dind: false
      device: mi300_1
      depends_on:
      - image-build-amd
@@ -52,7 +54,7 @@ steps:

 - label: Language Models Tests (Hybrid) %N
  key: language-models-tests-hybrid
-  timeout_in_minutes: 75
+  timeout_in_minutes: 65
  source_file_dependencies:
  - vllm/
  - tests/models/language/generation
@@ -67,7 +69,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 90
+      timeout_in_minutes: 70
      depends_on:
      - image-build-amd
      commands:
@@ -78,7 +80,7 @@ steps:
 - label: Language Models Test (Extended Generation) # 80min
  device: h200_35gb
  key: language-models-test-extended-generation
-  timeout_in_minutes: 110
+  timeout_in_minutes: 65
  optional: true
  source_file_dependencies:
  - vllm/
@@ -92,7 +94,7 @@ steps:

 - label: Language Models Test (PPL)
  key: language-models-test-ppl
-  timeout_in_minutes: 110
+  timeout_in_minutes: 30
  device: h200_18gb
  optional: true
  source_file_dependencies:
@@ -101,10 +103,10 @@ steps:
  commands:
    - pytest -v -s models/language/generation_ppl_test

- label: Language Models Test (Extended Pooling)  # 36min
+- label: Language Models Test (Extended Pooling)
  device: h200_35gb
  key: language-models-test-extended-pooling
-  timeout_in_minutes: 50
+  timeout_in_minutes: 120
  optional: true
  source_file_dependencies:
  - vllm/
@@ -114,13 +116,13 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 100
+      timeout_in_minutes: 120
      depends_on:
      - image-build-amd

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

 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
  key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 45
+  timeout_in_minutes: 50
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -38,7 +38,7 @@ 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: 45
+  timeout_in_minutes: 40
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
@@ -54,7 +54,7 @@ steps:
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
  device: h200_35gb
  key: multi-modal-models-standard-4-other-whisper
-  timeout_in_minutes: 45
+  timeout_in_minutes: 50
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
@@ -69,22 +69,23 @@ steps:
      depends_on:
      - image-build-amd

- label: Multi-Modal Processor (CPU)
+- label: Multi-Modal Processor (CPU) %N
  key: multi-modal-processor-cpu
  depends_on:
  - image-build-cpu
-  timeout_in_minutes: 60
+  timeout_in_minutes: 125
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
  - tests/models/registry.py
  device: cpu-medium
  commands:
-    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
+    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+  parallelism: 4

 - label: Multi-Modal Processor # 44min
  key: multi-modal-processor
-  timeout_in_minutes: 60
+  timeout_in_minutes: 65
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -96,7 +97,7 @@ steps:
 - label: Multi-Modal Accuracy Eval (Small Models) # 50min
  device: h200_35gb
  key: multi-modal-accuracy-eval-small-models
-  timeout_in_minutes: 70
+  timeout_in_minutes: 30
  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
  source_file_dependencies:
  - vllm/multimodal/
@@ -106,6 +107,7 @@ steps:
  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
  mirror:
    amd:
+      dind: false
      device: mi300_1
      depends_on:
      - image-build-amd
@@ -164,7 +166,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 60
+      timeout_in_minutes: 75
      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: 60
+  timeout_in_minutes: 35
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  source_file_dependencies:
@@ -37,6 +37,11 @@ steps:
  - pytest -v -s plugins_tests/test_stats_logger_plugins.py
  - pip uninstall dummy_stat_logger -y
  # end stat_logger plugins test
+  # begin endpoint plugins test
+  - pip install -e ./plugins/vllm_add_dummy_endpoint_plugin
+  - pytest -v -s plugins_tests/test_endpoint_plugins.py
+  - pip uninstall vllm_add_dummy_endpoint_plugin -y
+  # end endpoint plugins test
  # other tests continue here:
  - pytest -v -s plugins_tests/test_scheduler_plugins.py
  - pip install -e ./plugins/vllm_add_dummy_model
@@ -5,7 +5,7 @@ steps:
 - label: PyTorch Compilation Unit Tests
  device: h200_35gb
  key: pytorch-compilation-unit-tests
-  timeout_in_minutes: 10
+  timeout_in_minutes: 90
  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: 20
+  timeout_in_minutes: 45
  source_file_dependencies:
    - vllm/__init__.py
    - vllm/_aiter_ops.py
@@ -109,14 +109,15 @@ steps:
  - pytest -s -v compile/passes --ignore compile/passes/distributed
  mirror:
    amd:
+      dind: false
      device: mi300_1
-      timeout_in_minutes: 180
+      timeout_in_minutes: 65
      depends_on:
      - image-build-amd

 - label: PyTorch Fullgraph Smoke Test
  key: pytorch-fullgraph-smoke-test
-  timeout_in_minutes: 35
+  timeout_in_minutes: 60
  source_file_dependencies:
  - vllm/__init__.py
  - vllm/_aiter_ops.py
@@ -152,7 +153,7 @@ steps:

 - label: PyTorch Fullgraph
  key: pytorch-fullgraph
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  device: h200_18gb
  source_file_dependencies:
  - vllm/__init__.py
@@ -197,6 +198,7 @@ steps:
  - bash standalone_tests/pytorch_nightly_dependency.sh
  mirror:
    amd:
+      dind: false
      device: mi300_1
      depends_on:
      - image-build-amd
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Quantization
  key: quantization
-  timeout_in_minutes: 90
+  timeout_in_minutes: 60
  source_file_dependencies:
  - csrc/
  - vllm/model_executor/layers/quantization
@@ -23,7 +23,7 @@ steps:

 - label: Quantized Fusions
  key: quantized-fusions
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  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: 60
+  timeout_in_minutes: 120
  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: 60
+  timeout_in_minutes: 50
  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: 90
+  timeout_in_minutes: 30
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -15,28 +15,30 @@ 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 invalid"
+  - pytest -v -s entrypoints/openai/chat_completion/test_chat_logit_bias_validation.py -k "not multiple"

  # - 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"
-  # - pytest -v -s entrypoints/openai/test_return_token_ids.py
-  # - pytest -v -s entrypoints/openai/test_uds.py
+  # 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 v1/sample/test_logprobs_e2e.py -k "test_prompt_logprobs_e2e_server"

 - label: Rust Frontend Serve/Admin Coverage
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -45,22 +47,27 @@ steps:
  - vllm/entrypoints/serve/
  - vllm/v1/engine/
  - tests/utils.py
-  # - tests/entrypoints/serve/dev/rpc/test_collective_rpc.py
+  - tests/entrypoints/serve/dev/rpc/test_collective_rpc.py
  - tests/entrypoints/scale_out/token_in_token_out/test_serving_tokens.py
  - tests/entrypoints/serve/instrumentator/test_basic.py
  - tests/entrypoints/serve/instrumentator/test_metrics.py
  # - tests/entrypoints/serve/dev/test_sleep.py
+  - tests/entrypoints/serve/tokenize/test_tokenization.py
  commands:
  - export VLLM_USE_RUST_FRONTEND=1
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  # - pytest -v -s entrypoints/serve/dev/rpc/test_collective_rpc.py
+  - 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/instrumentator/test_basic.py -k "not show_version and not server_load"
+  # test_generate_logprobs expects Python-style top_logprobs truncation (dedup sampled + cap at max(k, 1)).
  - pytest -v -s entrypoints/scale_out/token_in_token_out/test_serving_tokens.py -k "not stream and not lora and not test_generate_logprobs and not stop_string_workflow"
  - pytest -v -s entrypoints/serve/instrumentator/test_metrics.py -k "text and not show and not run_batch and not test_metrics_counts and not test_metrics_exist"
  # - pytest -v -s entrypoints/serve/dev/test_sleep.py
+  # /tokenizer_info is not implemented in the Rust frontend (the CLI flag is accepted as a no-op).
+  - pytest -v -s entrypoints/serve/tokenize/test_tokenization.py -k "not tokenizer_info"

 - label: Rust Frontend Core Correctness
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h200_18gb
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -74,7 +81,7 @@ steps:
  - pytest -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine

 - label: Rust Frontend Tool Use
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - rust/
@@ -88,7 +95,7 @@ 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: 30
+  timeout_in_minutes: 25
  num_devices: 4
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -4,7 +4,7 @@ steps:
 - label: Rust Frontend Cargo Style + Clippy
  key: rust-frontend-cargo-style-clippy
  depends_on: []
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  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: 30
+  timeout_in_minutes: 20
  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: 75
+  timeout_in_minutes: 40
  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: mi250_1
+      device: mi325_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: 30
+  timeout_in_minutes: 25
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
@@ -15,7 +15,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 45
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -29,7 +29,7 @@ steps:

 - label: Spec Decode Eagle Nightly B200
  key: spec-decode-eagle-nightly-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 25
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -41,7 +41,7 @@ steps:

 - label: Spec Decode Speculators + MTP
  key: spec-decode-speculators-mtp
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
@@ -82,7 +82,7 @@ steps:
  
 - label: Spec Decode Ngram + Suffix
  key: spec-decode-ngram-suffix
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
@@ -93,7 +93,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 65
+      timeout_in_minutes: 55
      # TODO(akaratza): Test after Torch >= 2.12 bump
      soft_fail: true
      depends_on:
@@ -109,7 +109,7 @@ steps:

 - label: Spec Decode Draft Model
  key: spec-decode-draft-model
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/spec_decode/
@@ -120,7 +120,7 @@ steps:
  mirror:
    amd:
      device: mi325_1
-      timeout_in_minutes: 50
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -134,7 +134,7 @@ steps:

 - label: Spec Decode Draft Model Nightly B200
  key: spec-decode-draft-model-nightly-b200
-  timeout_in_minutes: 30
+  timeout_in_minutes: 40
  device: b200-k8s
  optional: true
  source_file_dependencies:
@@ -146,7 +146,7 @@ steps:

 - label: Speculators Correctness
  key: speculators-correctness
-  timeout_in_minutes: 60
+  timeout_in_minutes: 30
  device: h100
  optional: true
  num_devices: 1
@@ -159,7 +159,7 @@ steps:
    - pytest -v -s v1/spec_decode/test_speculators_correctness.py -m slow_test

 - label: Spec Decode MTP hybrid (B200)
-  timeout_in_minutes: 30
+  timeout_in_minutes: 20
  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: 45
+  timeout_in_minutes: 50
  working_dir: "/vllm-workspace/tests"
  num_devices: 2
  optional: true
@@ -15,6 +15,7 @@ steps:
    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
  mirror:
    amd:
+      dind: false
      device: mi300_2
      depends_on:
      - image-build-amd
@@ -172,7 +172,7 @@ mkdocs.yaml @hmellor
 # Kernels
 /vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
 /vllm/v1/attention/ops/triton_unified_attention.py @tdoublep
-/vllm/model_executor/layers/fla @ZJY0516 @vadiklyutiy
+/vllm/third_party/flash_linear_attention @ZJY0516 @vadiklyutiy

 # ROCm related: specify owner with write access to notify AMD folks for careful code review
 /vllm/**/*rocm* @tjtanaa @dllehr-amd
@@ -323,7 +323,7 @@ jobs:
            // {users} will be replaced with @mentions
            const ccConfig = {
              rocm: {
-                users: ['hongxiayang', 'tjtanaa', 'vllmellm'],
+                users: ['hongxiayang', 'tjtanaa', 'vllmellm', 'giuseppegrossi'],
                message: 'CC {users} for ROCm-related issue',
              },
              mistral: {
@@ -28,7 +28,8 @@ jobs:
            pull_number: context.payload.pull_request.number,
          });

-          const hasReadyLabel = pr.labels.some(l => l.name === 'ready');
+          const readyLabels = ['ready', 'ready-run-all-tests'];
+          const hasReadyLabel = pr.labels.some(l => readyLabels.includes(l.name));
          const hasVerifiedLabel = pr.labels.some(l => l.name === 'verified');

          const { data: mergedPRs } = await github.rest.search.issuesAndPullRequests({
@@ -40,7 +41,7 @@ jobs:
          if (hasReadyLabel || hasVerifiedLabel || mergedCount >= 4) {
            core.info(`Check passed: verified label=${hasVerifiedLabel}, ready label=${hasReadyLabel}, 4+ merged PRs=${mergedCount >= 4}`);
          } else {
-            core.setFailed(`PR must have the 'verified' or 'ready' (which also triggers tests) label or the author must have at least 4 merged PRs (found ${mergedCount}).`);
+            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}).`);
          }

  pre-commit:
@@ -18,6 +18,9 @@ vllm/third_party/deep_gemm/
 # fmha_sm100 vendored package built from source
 vllm/third_party/fmha_sm100/

+# tml-fa4 vendored package built from source
+vllm/third_party/tml_fa4/
+
 # triton jit
 .triton

@@ -210,7 +210,7 @@ repos:
    name: Check SPDX headers
    entry: python tools/pre_commit/check_spdx_header.py
    language: python
-    types: [python]
+    types_or: [python, rust, proto]
  - id: check-root-lazy-imports
    name: Check root lazy imports
    entry: python tools/pre_commit/check_init_lazy_imports.py
@@ -1,2 +0,0 @@
-collect_env.py
-vllm/model_executor/layers/fla/ops/*.py
@@ -29,6 +29,7 @@ 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
@@ -66,23 +67,38 @@ VLLM_USE_PRECOMPILED=1 uv pip install -e . --torch-backend=auto
 uv pip install -e . --torch-backend=auto
 ```

-### Running tests
+### Tests

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

 ```bash
-# 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
+# Install test dependencies (use cuda.in on non-x86_64):
+uv pip install -r requirements/test/cuda.in

-# Run a specific test file (use .venv/bin/python directly;
-# `source activate` does not persist in non-interactive shells):
+# Run a specific test file:
 .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).
@@ -107,23 +123,18 @@ Use [Google-style docstrings](https://google.github.io/styleguide/pyguide.html#3

 ### Coding style guidelines

-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.
+- 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.
 - Assume the reader is familiar with vLLM.

 ### Commit messages

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

 ```text
 Your commit message here

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

@@ -70,6 +70,15 @@ 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
@@ -175,7 +184,7 @@ endif()
 if (NOT HIP_FOUND AND NOT PYTORCH_FOUND_HIP AND CUDA_FOUND)
  set(VLLM_GPU_LANG "CUDA")

-  if (NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
+  if (NOT TORCH_NIGHTLY_BUILD AND 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()
@@ -188,7 +197,7 @@ elseif(HIP_FOUND OR PYTORCH_FOUND_HIP)
  enable_language(HIP)

  # ROCm 5.X and 6.X
-  if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
+  if (NOT TORCH_NIGHTLY_BUILD AND 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.")
@@ -381,6 +390,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
    "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"
@@ -1354,6 +1364,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
  set(VLLM_ROCM_EXT_SRC
    "csrc/rocm/torch_bindings.cpp"
    "csrc/rocm/skinny_gemms.cu"
+    "csrc/rocm/skinny_gemms_int4.cu"
    "csrc/rocm/attention.cu")

  set(VLLM_ROCM_HAS_GFX1100 OFF)
@@ -1397,6 +1408,7 @@ if (VLLM_GPU_LANG STREQUAL "CUDA")
    include(cmake/external_projects/fmha_sm100.cmake)
    include(cmake/external_projects/flashmla.cmake)
    include(cmake/external_projects/qutlass.cmake)
+    include(cmake/external_projects/tml_fa4.cmake)

    # vllm-flash-attn should be last as it overwrites some CMake functions
    include(cmake/external_projects/vllm_flash_attn.cmake)
@@ -75,7 +75,11 @@ def run_mla_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
    from mla_runner import run_mla_benchmark as run_mla

    return run_mla(
-        config.backend, config, prefill_backend=config.prefill_backend, **kwargs
+        config.backend,
+        config,
+        prefill_backend=config.prefill_backend,
+        sparse_mla_force_mqa=config.sparse_mla_force_mqa,
+        **kwargs,
    )


@@ -592,6 +596,30 @@ def main():
        default="profile",
        help="Output file name for ncu profile (default: 'profile').",
    )
+    parser.add_argument(
+        "--torch-profile",
+        action="store_true",
+        default=False,
+        help="Collect a PyTorch profiler Chrome trace for each benchmark run.",
+    )
+    parser.add_argument(
+        "--torch-profile-dir",
+        default=None,
+        help="Directory for PyTorch profiler traces.",
+    )
+    parser.add_argument(
+        "--torch-profile-iters",
+        type=int,
+        default=3,
+        help="Number of forward passes to record per PyTorch profiler trace.",
+    )
+    parser.add_argument(
+        "--sparse-mla-mha-variants",
+        nargs="+",
+        default=None,
+        choices=["dense_mha", "mqa"],
+        help="Sparse MLA variants to run in mha_vs_mqa mode. Defaults to both.",
+    )

    # Parameter sweep (use YAML config for advanced sweeps)
    parser.add_argument(
@@ -641,6 +669,7 @@ def main():

        # Prefill backends (e.g., ["fa3", "fa4"])
        args.prefill_backends = yaml_config.get("prefill_backends", None)
+        args.prefill_backend = yaml_config.get("prefill_backend", None)

        # FP8 output benchmark knobs; CLI wins.
        if args.fp8_output_scale is None:
@@ -683,6 +712,9 @@ def main():
            args.num_q_heads = model.get("num_q_heads", args.num_q_heads)
            args.num_kv_heads = model.get("num_kv_heads", args.num_kv_heads)
            args.block_size = model.get("block_size", args.block_size)
+            args.max_model_len = model.get(
+                "max_model_len", getattr(args, "max_model_len", None)
+            )
            # MLA-specific dimensions
            args.kv_lora_rank = model.get("kv_lora_rank", args.kv_lora_rank)
            args.qk_nope_head_dim = model.get("qk_nope_head_dim", args.qk_nope_head_dim)
@@ -701,6 +733,21 @@ def main():
            args.cuda_graphs = yaml_config["cuda_graphs"]
        if "ncu_profile" in yaml_config:
            args.ncu_profile = yaml_config["ncu_profile"]
+        if "torch_profile" in yaml_config:
+            args.torch_profile = yaml_config["torch_profile"]
+        if "torch_profile_dir" in yaml_config:
+            args.torch_profile_dir = yaml_config["torch_profile_dir"]
+        if "torch_profile_iters" in yaml_config:
+            args.torch_profile_iters = yaml_config["torch_profile_iters"]
+        args.sparse_mla_topk_pattern = yaml_config.get(
+            "sparse_mla_topk_pattern", "random"
+        )
+        args.sparse_mla_dense_mha_max_seq_len = yaml_config.get(
+            "sparse_mla_dense_mha_max_seq_len", None
+        )
+        args.sparse_mla_mha_variants = yaml_config.get(
+            "sparse_mla_mha_variants", args.sparse_mla_mha_variants
+        )

        # Parameter sweep configuration
        if "parameter_sweep" in yaml_config:
@@ -842,8 +889,6 @@ def main():
                        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,
@@ -1063,6 +1108,133 @@ def main():
                    f"\n  [yellow]Prefill always faster for batch_size={bs}[/]"
                )

+    # Handle MHA vs MQA comparison mode for sparse MLA
+    elif hasattr(args, "mode") and args.mode == "mha_vs_mqa":
+        console.print("[yellow]Mode: MHA vs MQA comparison for sparse MLA[/]")
+
+        sparse_mla_topk_pattern = getattr(args, "sparse_mla_topk_pattern", "random")
+        dense_mha_max_seq_len = getattr(args, "sparse_mla_dense_mha_max_seq_len", None)
+        prefill_backend = getattr(args, "prefill_backend", None)
+        if prefill_backend:
+            console.print(f"Prefill backend: {prefill_backend}")
+        available_variants = [
+            ("dense_mha", False, "dense"),
+            ("mqa", True, "auto"),
+        ]
+        requested_variants = getattr(args, "sparse_mla_mha_variants", None)
+        if requested_variants is not None:
+            valid_variants = {label for label, _, _ in available_variants}
+            invalid_variants = sorted(set(requested_variants) - valid_variants)
+            if invalid_variants:
+                raise ValueError(
+                    "Invalid sparse_mla_mha_variants entries: "
+                    f"{invalid_variants}. Valid variants are: "
+                    f"{sorted(valid_variants)}"
+                )
+            requested_variant_set = set(requested_variants)
+            variants = [
+                variant
+                for variant in available_variants
+                if variant[0] in requested_variant_set
+            ]
+        else:
+            variants = available_variants
+        formatter = ResultsFormatter(console)
+        total = 0
+        for spec in args.batch_specs:
+            q_len = max(request.q_len for request in parse_batch_spec(spec))
+            for variant_label, _, _ in variants:
+                if (
+                    variant_label == "dense_mha"
+                    and dense_mha_max_seq_len is not None
+                    and q_len > dense_mha_max_seq_len
+                ):
+                    continue
+                total += len(backends)
+
+        with tqdm(total=total, desc="Benchmarking") as pbar:
+            for spec in args.batch_specs:
+                q_len = max(request.q_len for request in parse_batch_spec(spec))
+                for backend in backends:
+                    for variant_label, force_mqa, mha_mode in variants:
+                        if (
+                            variant_label == "dense_mha"
+                            and dense_mha_max_seq_len is not None
+                            and q_len > dense_mha_max_seq_len
+                        ):
+                            continue
+                        config = BenchmarkConfig(
+                            backend=f"{backend}_{variant_label}",
+                            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,
+                            max_model_len=getattr(args, "max_model_len", None),
+                            kv_cache_dtype=args.kv_cache_dtype,
+                            profile_memory=args.profile_memory,
+                            use_cuda_graphs=args.cuda_graphs,
+                            ncu_profile=args.ncu_profile,
+                            torch_profile=args.torch_profile,
+                            torch_profile_dir=args.torch_profile_dir,
+                            torch_profile_iters=args.torch_profile_iters,
+                            warmup_ms=args.warmup_ms,
+                            kv_lora_rank=getattr(args, "kv_lora_rank", None),
+                            qk_nope_head_dim=getattr(args, "qk_nope_head_dim", None),
+                            qk_rope_head_dim=getattr(args, "qk_rope_head_dim", None),
+                            v_head_dim=getattr(args, "v_head_dim", None),
+                            sparse_mla_force_mqa=force_mqa,
+                            sparse_mla_mha_mode=mha_mode,
+                            sparse_mla_dense_mha_max_seq_len=dense_mha_max_seq_len,
+                            sparse_mla_topk_pattern=sparse_mla_topk_pattern,
+                            prefill_backend=prefill_backend,
+                        )
+
+                        # run_mla_benchmark needs the real backend name
+                        from mla_runner import run_mla_benchmark as run_mla
+
+                        run_label = f"{backend}_{variant_label} {spec}"
+                        pbar.set_postfix_str(run_label)
+
+                        try:
+                            result = run_mla(
+                                backend,
+                                config,
+                                prefill_backend=prefill_backend,
+                                sparse_mla_force_mqa=force_mqa,
+                            )
+                        except Exception as e:
+                            result = BenchmarkResult(
+                                config=config,
+                                mean_time=float("inf"),
+                                median_time=float("inf"),
+                                std_time=0,
+                                min_time=float("inf"),
+                                max_time=float("inf"),
+                                error=str(e),
+                            )
+
+                        all_results.append(result)
+                        if args.output_csv:
+                            formatter.save_csv(all_results, args.output_csv)
+                        if args.output_json:
+                            formatter.save_json(all_results, args.output_json)
+
+                        if not result.success:
+                            console.print(
+                                f"[red]Error {backend}_{variant_label} "
+                                f"{spec}: {result.error}[/]"
+                            )
+
+                        pbar.update(1)
+
+        # Display results with variant labels as separate "backends"
+        console.print("\n[bold green]MHA vs MQA Results:[/]")
+        variant_backends = [f"{b}_{v}" for b in backends for v, _, _ in variants]
+        formatter.print_table(all_results, variant_backends)
+
    # Handle model parameter sweep mode
    elif hasattr(args, "model_parameter_sweep") and args.model_parameter_sweep:
        # Model parameter sweep
@@ -4,8 +4,10 @@
 """Common utilities for attention benchmarking."""

 import csv
+import gc
 import json
 import math
+from collections.abc import Sequence
 from dataclasses import asdict, dataclass
 from pathlib import Path
 from typing import Any
@@ -44,10 +46,13 @@ def run_do_bench(
    kwargs: dict[str, Any] = {"return_mode": "all"}
    if use_cuda_graphs:
        result = triton.testing.do_bench_cudagraph(benchmark_fn, **kwargs)
+        gc.collect()
+        torch.accelerator.empty_cache()
    else:
        if warmup_ms is not None:
            kwargs["warmup"] = warmup_ms
        result = triton.testing.do_bench(benchmark_fn, **kwargs)
+    torch.accelerator.synchronize()
    return result


@@ -91,42 +96,6 @@ except ImportError:
    AttentionLayerBase = object  # Fallback


-class MockKVBProj:
-    """Mock KV projection layer for MLA prefill mode.
-
-    Mimics ColumnParallelLinear behavior for kv_b_proj in MLA backends.
-    Projects kv_c_normed to [qk_nope_head_dim + v_head_dim] per head.
-    """
-
-    def __init__(self, num_heads: int, qk_nope_head_dim: int, v_head_dim: int):
-        self.num_heads = num_heads
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.v_head_dim = v_head_dim
-        self.out_dim = qk_nope_head_dim + v_head_dim
-        self.weight = torch.empty(0, dtype=torch.bfloat16)
-
-    def __call__(self, x: torch.Tensor) -> tuple[torch.Tensor]:
-        """
-        Project kv_c_normed to output space.
-
-        Args:
-            x: Input tensor [num_tokens, kv_lora_rank]
-
-        Returns:
-            Tuple containing output tensor
-                [num_tokens, num_heads, qk_nope_head_dim + v_head_dim]
-        """
-        num_tokens = x.shape[0]
-        result = torch.randn(
-            num_tokens,
-            self.num_heads,
-            self.out_dim,
-            device=x.device,
-            dtype=x.dtype,
-        )
-        return (result,)  # Return as tuple to match ColumnParallelLinear API
-
-
 class MockIndexer:
    """Mock Indexer for sparse MLA backends.

@@ -158,6 +127,60 @@ class MockIndexer:
        )
        self.topk_indices_buffer[:num_tokens] = indices

+    def fill_indices(
+        self,
+        num_tokens: int,
+        max_kv_len: int,
+        pattern: str = "random",
+        requests: Sequence[Any] | None = None,
+    ):
+        if pattern == "random":
+            self.fill_random_indices(num_tokens, max_kv_len)
+            return
+        if pattern == "prefix":
+            indices = torch.arange(
+                self.topk_tokens,
+                dtype=torch.int32,
+                device=self.topk_indices_buffer.device,
+            )
+            indices = (indices % max_kv_len).expand(num_tokens, -1)
+            self.topk_indices_buffer[:num_tokens] = indices
+            return
+        if pattern == "sliding_window":
+            if requests is None:
+                start = max(max_kv_len - self.topk_tokens, 0)
+                indices = torch.arange(
+                    start,
+                    start + self.topk_tokens,
+                    dtype=torch.int32,
+                    device=self.topk_indices_buffer.device,
+                )
+                indices = indices.clamp(max=max_kv_len - 1).expand(num_tokens, -1)
+                self.topk_indices_buffer[:num_tokens] = indices
+                return
+
+            rows = []
+            offsets = torch.arange(
+                self.topk_tokens,
+                dtype=torch.int32,
+                device=self.topk_indices_buffer.device,
+            ) - (self.topk_tokens - 1)
+            for request in requests:
+                q_len = request.q_len
+                kv_len = request.kv_len
+                context_len = kv_len - q_len
+                positions = torch.arange(
+                    context_len,
+                    kv_len,
+                    dtype=torch.int32,
+                    device=self.topk_indices_buffer.device,
+                )
+                row_indices = positions[:, None] + offsets[None, :]
+                rows.append(row_indices.clamp(min=0, max=kv_len - 1))
+            self.topk_indices_buffer[:num_tokens] = torch.cat(rows, dim=0)
+            return
+        raise ValueError(f"Unknown sparse MLA topk pattern: {pattern}")
+

 class MockLayer(AttentionLayerBase):
    """Mock attention layer with scale parameters and impl.
@@ -252,10 +275,14 @@ class BenchmarkConfig:
    num_kv_heads: int
    block_size: int
    device: str
+    max_model_len: int | None = None
    dtype: torch.dtype = torch.float16
    profile_memory: bool = False
-    use_cuda_graphs: bool = False
+    use_cuda_graphs: bool = True
    ncu_profile: bool = False
+    torch_profile: bool = False
+    torch_profile_dir: str | None = None
+    torch_profile_iters: int = 3
    warmup_ms: int | None = None

    # "auto" or "fp8"
@@ -271,6 +298,10 @@ class BenchmarkConfig:
    # Backend-specific tuning
    num_kv_splits: int | None = None  # CUTLASS MLA
    reorder_batch_threshold: int | None = None  # FlashAttn MLA, FlashMLA
+    sparse_mla_force_mqa: bool = False  # Force MQA path for sparse MLA
+    sparse_mla_mha_mode: str = "auto"  # "auto" or "dense"
+    sparse_mla_dense_mha_max_seq_len: int | None = None
+    sparse_mla_topk_pattern: str = "random"  # "random", "prefix", "sliding_window"
    num_splits: int | None = None  # FlashAttention split-K (0=auto, 1=disabled)


@@ -0,0 +1,474 @@
+# Sparse MLA benchmark: forward_mha vs forward_mqa
+#
+# Usage:
+#   python benchmark.py --config configs/mla_sparse_mha_vs_mqa.yaml
+#
+# Heatmap grid:
+#   - batch_size: 1, 2, 4, 8, 16, 32
+#   - seq_len: 32, 64, 128, 256, 512, 1024, 2048
+#   - q_len: powers of two through seq_len
+#
+# Specs with q_len < seq_len include context; the q_len == seq_len diagonal
+# covers pure prefill.
+# The model shape below is the DP case. For the TP8 run, manually change
+# model.num_q_heads from 128 to 16 before rerunning this benchmark.
+
+mode: mha_vs_mqa
+
+model:
+  name: "deepseek-v3"
+  num_layers: 60
+  num_q_heads: 128
+  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
+  max_model_len: 2048
+
+batch_specs:
+  # Batch size 1
+  # seq_len = 32
+  - "1q1s32"
+  - "1q2s32"
+  - "1q4s32"
+  - "1q8s32"
+  - "1q16s32"
+  - "1q32"
+  # seq_len = 64
+  - "1q1s64"
+  - "1q2s64"
+  - "1q4s64"
+  - "1q8s64"
+  - "1q16s64"
+  - "1q32s64"
+  - "1q64"
+  # seq_len = 128
+  - "1q1s128"
+  - "1q2s128"
+  - "1q4s128"
+  - "1q8s128"
+  - "1q16s128"
+  - "1q32s128"
+  - "1q64s128"
+  - "1q128"
+  # seq_len = 256
+  - "1q1s256"
+  - "1q2s256"
+  - "1q4s256"
+  - "1q8s256"
+  - "1q16s256"
+  - "1q32s256"
+  - "1q64s256"
+  - "1q128s256"
+  - "1q256"
+  # seq_len = 512
+  - "1q1s512"
+  - "1q2s512"
+  - "1q4s512"
+  - "1q8s512"
+  - "1q16s512"
+  - "1q32s512"
+  - "1q64s512"
+  - "1q128s512"
+  - "1q256s512"
+  - "1q512"
+  # seq_len = 1024
+  - "1q1s1024"
+  - "1q2s1024"
+  - "1q4s1024"
+  - "1q8s1024"
+  - "1q16s1024"
+  - "1q32s1024"
+  - "1q64s1024"
+  - "1q128s1024"
+  - "1q256s1024"
+  - "1q512s1024"
+  - "1q1024"
+  # seq_len = 2048
+  - "1q1s2048"
+  - "1q2s2048"
+  - "1q4s2048"
+  - "1q8s2048"
+  - "1q16s2048"
+  - "1q32s2048"
+  - "1q64s2048"
+  - "1q128s2048"
+  - "1q256s2048"
+  - "1q512s2048"
+  - "1q1024s2048"
+  - "1q2048"
+
+  # Batch size 2
+  # seq_len = 32
+  - "2q1s32"
+  - "2q2s32"
+  - "2q4s32"
+  - "2q8s32"
+  - "2q16s32"
+  - "2q32"
+  # seq_len = 64
+  - "2q1s64"
+  - "2q2s64"
+  - "2q4s64"
+  - "2q8s64"
+  - "2q16s64"
+  - "2q32s64"
+  - "2q64"
+  # seq_len = 128
+  - "2q1s128"
+  - "2q2s128"
+  - "2q4s128"
+  - "2q8s128"
+  - "2q16s128"
+  - "2q32s128"
+  - "2q64s128"
+  - "2q128"
+  # seq_len = 256
+  - "2q1s256"
+  - "2q2s256"
+  - "2q4s256"
+  - "2q8s256"
+  - "2q16s256"
+  - "2q32s256"
+  - "2q64s256"
+  - "2q128s256"
+  - "2q256"
+  # seq_len = 512
+  - "2q1s512"
+  - "2q2s512"
+  - "2q4s512"
+  - "2q8s512"
+  - "2q16s512"
+  - "2q32s512"
+  - "2q64s512"
+  - "2q128s512"
+  - "2q256s512"
+  - "2q512"
+  # seq_len = 1024
+  - "2q1s1024"
+  - "2q2s1024"
+  - "2q4s1024"
+  - "2q8s1024"
+  - "2q16s1024"
+  - "2q32s1024"
+  - "2q64s1024"
+  - "2q128s1024"
+  - "2q256s1024"
+  - "2q512s1024"
+  - "2q1024"
+  # seq_len = 2048
+  - "2q1s2048"
+  - "2q2s2048"
+  - "2q4s2048"
+  - "2q8s2048"
+  - "2q16s2048"
+  - "2q32s2048"
+  - "2q64s2048"
+  - "2q128s2048"
+  - "2q256s2048"
+  - "2q512s2048"
+  - "2q1024s2048"
+  - "2q2048"
+
+  # Batch size 4
+  # seq_len = 32
+  - "4q1s32"
+  - "4q2s32"
+  - "4q4s32"
+  - "4q8s32"
+  - "4q16s32"
+  - "4q32"
+  # seq_len = 64
+  - "4q1s64"
+  - "4q2s64"
+  - "4q4s64"
+  - "4q8s64"
+  - "4q16s64"
+  - "4q32s64"
+  - "4q64"
+  # seq_len = 128
+  - "4q1s128"
+  - "4q2s128"
+  - "4q4s128"
+  - "4q8s128"
+  - "4q16s128"
+  - "4q32s128"
+  - "4q64s128"
+  - "4q128"
+  # seq_len = 256
+  - "4q1s256"
+  - "4q2s256"
+  - "4q4s256"
+  - "4q8s256"
+  - "4q16s256"
+  - "4q32s256"
+  - "4q64s256"
+  - "4q128s256"
+  - "4q256"
+  # seq_len = 512
+  - "4q1s512"
+  - "4q2s512"
+  - "4q4s512"
+  - "4q8s512"
+  - "4q16s512"
+  - "4q32s512"
+  - "4q64s512"
+  - "4q128s512"
+  - "4q256s512"
+  - "4q512"
+  # seq_len = 1024
+  - "4q1s1024"
+  - "4q2s1024"
+  - "4q4s1024"
+  - "4q8s1024"
+  - "4q16s1024"
+  - "4q32s1024"
+  - "4q64s1024"
+  - "4q128s1024"
+  - "4q256s1024"
+  - "4q512s1024"
+  - "4q1024"
+  # seq_len = 2048
+  - "4q1s2048"
+  - "4q2s2048"
+  - "4q4s2048"
+  - "4q8s2048"
+  - "4q16s2048"
+  - "4q32s2048"
+  - "4q64s2048"
+  - "4q128s2048"
+  - "4q256s2048"
+  - "4q512s2048"
+  - "4q1024s2048"
+  - "4q2048"
+
+  # Batch size 8
+  # seq_len = 32
+  - "8q1s32"
+  - "8q2s32"
+  - "8q4s32"
+  - "8q8s32"
+  - "8q16s32"
+  - "8q32"
+  # seq_len = 64
+  - "8q1s64"
+  - "8q2s64"
+  - "8q4s64"
+  - "8q8s64"
+  - "8q16s64"
+  - "8q32s64"
+  - "8q64"
+  # seq_len = 128
+  - "8q1s128"
+  - "8q2s128"
+  - "8q4s128"
+  - "8q8s128"
+  - "8q16s128"
+  - "8q32s128"
+  - "8q64s128"
+  - "8q128"
+  # seq_len = 256
+  - "8q1s256"
+  - "8q2s256"
+  - "8q4s256"
+  - "8q8s256"
+  - "8q16s256"
+  - "8q32s256"
+  - "8q64s256"
+  - "8q128s256"
+  - "8q256"
+  # seq_len = 512
+  - "8q1s512"
+  - "8q2s512"
+  - "8q4s512"
+  - "8q8s512"
+  - "8q16s512"
+  - "8q32s512"
+  - "8q64s512"
+  - "8q128s512"
+  - "8q256s512"
+  - "8q512"
+  # seq_len = 1024
+  - "8q1s1024"
+  - "8q2s1024"
+  - "8q4s1024"
+  - "8q8s1024"
+  - "8q16s1024"
+  - "8q32s1024"
+  - "8q64s1024"
+  - "8q128s1024"
+  - "8q256s1024"
+  - "8q512s1024"
+  - "8q1024"
+  # seq_len = 2048
+  - "8q1s2048"
+  - "8q2s2048"
+  - "8q4s2048"
+  - "8q8s2048"
+  - "8q16s2048"
+  - "8q32s2048"
+  - "8q64s2048"
+  - "8q128s2048"
+  - "8q256s2048"
+  - "8q512s2048"
+  - "8q1024s2048"
+  - "8q2048"
+
+  # Batch size 16
+  # seq_len = 32
+  - "16q1s32"
+  - "16q2s32"
+  - "16q4s32"
+  - "16q8s32"
+  - "16q16s32"
+  - "16q32"
+  # seq_len = 64
+  - "16q1s64"
+  - "16q2s64"
+  - "16q4s64"
+  - "16q8s64"
+  - "16q16s64"
+  - "16q32s64"
+  - "16q64"
+  # seq_len = 128
+  - "16q1s128"
+  - "16q2s128"
+  - "16q4s128"
+  - "16q8s128"
+  - "16q16s128"
+  - "16q32s128"
+  - "16q64s128"
+  - "16q128"
+  # seq_len = 256
+  - "16q1s256"
+  - "16q2s256"
+  - "16q4s256"
+  - "16q8s256"
+  - "16q16s256"
+  - "16q32s256"
+  - "16q64s256"
+  - "16q128s256"
+  - "16q256"
+  # seq_len = 512
+  - "16q1s512"
+  - "16q2s512"
+  - "16q4s512"
+  - "16q8s512"
+  - "16q16s512"
+  - "16q32s512"
+  - "16q64s512"
+  - "16q128s512"
+  - "16q256s512"
+  - "16q512"
+  # seq_len = 1024
+  - "16q1s1024"
+  - "16q2s1024"
+  - "16q4s1024"
+  - "16q8s1024"
+  - "16q16s1024"
+  - "16q32s1024"
+  - "16q64s1024"
+  - "16q128s1024"
+  - "16q256s1024"
+  - "16q512s1024"
+  - "16q1024"
+  # seq_len = 2048
+  - "16q1s2048"
+  - "16q2s2048"
+  - "16q4s2048"
+  - "16q8s2048"
+  - "16q16s2048"
+  - "16q32s2048"
+  - "16q64s2048"
+  - "16q128s2048"
+  - "16q256s2048"
+  - "16q512s2048"
+  - "16q1024s2048"
+  - "16q2048"
+
+  # Batch size 32
+  # seq_len = 32
+  - "32q1s32"
+  - "32q2s32"
+  - "32q4s32"
+  - "32q8s32"
+  - "32q16s32"
+  - "32q32"
+  # seq_len = 64
+  - "32q1s64"
+  - "32q2s64"
+  - "32q4s64"
+  - "32q8s64"
+  - "32q16s64"
+  - "32q32s64"
+  - "32q64"
+  # seq_len = 128
+  - "32q1s128"
+  - "32q2s128"
+  - "32q4s128"
+  - "32q8s128"
+  - "32q16s128"
+  - "32q32s128"
+  - "32q64s128"
+  - "32q128"
+  # seq_len = 256
+  - "32q1s256"
+  - "32q2s256"
+  - "32q4s256"
+  - "32q8s256"
+  - "32q16s256"
+  - "32q32s256"
+  - "32q64s256"
+  - "32q128s256"
+  - "32q256"
+  # seq_len = 512
+  - "32q1s512"
+  - "32q2s512"
+  - "32q4s512"
+  - "32q8s512"
+  - "32q16s512"
+  - "32q32s512"
+  - "32q64s512"
+  - "32q128s512"
+  - "32q256s512"
+  - "32q512"
+  # seq_len = 1024
+  - "32q1s1024"
+  - "32q2s1024"
+  - "32q4s1024"
+  - "32q8s1024"
+  - "32q16s1024"
+  - "32q32s1024"
+  - "32q64s1024"
+  - "32q128s1024"
+  - "32q256s1024"
+  - "32q512s1024"
+  - "32q1024"
+  # seq_len = 2048
+  - "32q1s2048"
+  - "32q2s2048"
+  - "32q4s2048"
+  - "32q8s2048"
+  - "32q16s2048"
+  - "32q32s2048"
+  - "32q64s2048"
+  - "32q128s2048"
+  - "32q256s2048"
+  - "32q512s2048"
+  - "32q1024s2048"
+  - "32q2048"
+
+backends:
+  - FLASHMLA_SPARSE
+
+device: "cuda:0"
+profile_memory: false
+sparse_mla_dense_mha_max_seq_len: 2048
+sparse_mla_topk_pattern: "random"
+
+output:
+  csv: "benchmark_output/mla_sparse_mha_vs_mqa.csv"
+  json: "benchmark_output/mla_sparse_mha_vs_mqa.json"
@@ -9,6 +9,8 @@ needing full VllmConfig integration.
 """

 import statistics
+import tempfile
+from pathlib import Path

 import numpy as np
 import torch
@@ -17,7 +19,6 @@ from common import (
    BenchmarkResult,
    MockHfConfig,
    MockIndexer,
-    MockKVBProj,
    MockLayer,
    run_do_bench,
    run_ncu_profile,
@@ -33,8 +34,59 @@ from vllm.config import (
    VllmConfig,
    set_current_vllm_config,
 )
+from vllm.model_executor.layers.linear import ColumnParallelLinear
 from vllm.v1.attention.backends.mla.prefill.registry import MLAPrefillBackendEnum

+
+def _safe_profile_name(value: str) -> str:
+    return "".join(c if c.isalnum() or c in "._-" else "_" for c in value)
+
+
+def _create_kv_b_proj(
+    mla_dims: dict,
+    device: torch.device,
+):
+    kv_b_proj = ColumnParallelLinear(
+        mla_dims["kv_lora_rank"],
+        mla_dims["num_q_heads"]
+        * (mla_dims["qk_nope_head_dim"] + mla_dims["v_head_dim"]),
+        bias=False,
+        params_dtype=torch.bfloat16,
+        quant_config=None,
+        prefix="benchmark.kv_b_proj",
+    ).to(device)
+    with torch.no_grad():
+        kv_b_proj.weight.copy_(torch.randn_like(kv_b_proj.weight))
+    return kv_b_proj
+
+
+def _ensure_single_rank_model_parallel() -> None:
+    import torch.distributed as dist
+
+    from vllm.distributed import (
+        ensure_model_parallel_initialized,
+        init_distributed_environment,
+        model_parallel_is_initialized,
+    )
+
+    if not dist.is_available():
+        return
+    if not dist.is_initialized():
+        with tempfile.NamedTemporaryFile(
+            prefix="vllm_bench_dist_", delete=False
+        ) as init_file:
+            distributed_init_method = f"file://{init_file.name}"
+        init_distributed_environment(
+            world_size=1,
+            rank=0,
+            distributed_init_method=distributed_init_method,
+            local_rank=0,
+            backend="nccl",
+        )
+    if not model_parallel_is_initialized():
+        ensure_model_parallel_initialized(1, 1)
+
+
 # ============================================================================
 # VllmConfig Creation
 # ============================================================================
@@ -66,10 +118,12 @@ def create_minimal_vllm_config(
    block_size: int = 128,
    max_num_seqs: int = 256,
    max_num_batched_tokens: int = 8192,
+    max_model_len: int = 32768,
    mla_dims: dict | None = None,
    index_topk: int | None = None,
    prefill_backend: str | None = None,
    kv_cache_dtype: str = "auto",
+    sparse_mla_force_mqa: bool = False,
 ) -> VllmConfig:
    """
    Create minimal VllmConfig for MLA benchmarks.
@@ -86,6 +140,8 @@ def create_minimal_vllm_config(
        prefill_backend: Prefill backend name (e.g., "fa3", "fa4", "flashinfer",
                        "trtllm"). Configures the attention config to force
                        the specified prefill backend.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+                    forward_mqa (even prefill tokens).

    Returns:
        VllmConfig for benchmarking
@@ -131,7 +187,7 @@ def create_minimal_vllm_config(
            trust_remote_code=True,
            dtype="bfloat16",
            seed=0,
-            max_model_len=32768,
+            max_model_len=max_model_len,
            quantization=None,
            enforce_eager=False,
            max_logprobs=20,
@@ -163,7 +219,7 @@ def create_minimal_vllm_config(
    scheduler_config = SchedulerConfig(
        max_num_seqs=max_num_seqs,
        max_num_batched_tokens=max(max_num_batched_tokens, max_num_seqs),
-        max_model_len=32768,
+        max_model_len=max_model_len,
        is_encoder_decoder=False,
        enable_chunked_prefill=True,
    )
@@ -192,6 +248,9 @@ def create_minimal_vllm_config(
                "flash_attn_version"
            ]

+    if sparse_mla_force_mqa:
+        vllm_config.attention_config.sparse_mla_force_mqa = True
+
    return vllm_config


@@ -548,12 +607,7 @@ def _create_backend_impl(
    # Calculate scale
    scale = 1.0 / np.sqrt(mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"])

-    # Create mock kv_b_proj layer for prefill mode
-    mock_kv_b_proj = MockKVBProj(
-        num_heads=mla_dims["num_q_heads"],
-        qk_nope_head_dim=mla_dims["qk_nope_head_dim"],
-        v_head_dim=mla_dims["v_head_dim"],
-    )
+    kv_b_proj = _create_kv_b_proj(mla_dims, device)

    # Create indexer for sparse backends
    indexer = None
@@ -584,7 +638,7 @@ def _create_backend_impl(
        "qk_rope_head_dim": mla_dims["qk_rope_head_dim"],
        "qk_head_dim": mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
        "v_head_dim": mla_dims["v_head_dim"],
-        "kv_b_proj": mock_kv_b_proj,
+        "kv_b_proj": kv_b_proj,
    }

    # Add indexer for sparse backends
@@ -785,14 +839,35 @@ def _run_single_benchmark(
    # Fill indexer with random indices for sparse backends
    is_sparse = backend_cfg.get("is_sparse", False)
    if is_sparse and indexer is not None:
-        indexer.fill_random_indices(total_q, max_kv_len)
+        indexer.fill_indices(
+            total_q,
+            max_kv_len,
+            getattr(config, "sparse_mla_topk_pattern", "random"),
+        )

    # Determine which forward methods to use based on metadata.
-    # Sparse MLA backends always use forward_mqa
-    has_decode = is_sparse or getattr(metadata, "decode", None) is not None
-    has_prefill = not is_sparse and getattr(metadata, "prefill", None) is not None
+    # Non-sparse backends use .decode/.prefill sub-objects.
+    # Sparse backends use num_decode_tokens/num_prefills directly.
+    #
+    # sparse_mla_force_mqa overrides: even for prefill metadata, use MQA.
+    force_mqa = getattr(config, "sparse_mla_force_mqa", False)
+    force_dense_mha = getattr(config, "sparse_mla_mha_mode", "auto") == "dense"
+    if force_mqa:
+        has_decode = True
+        has_prefill = False
+    elif is_sparse:
+        has_decode = metadata.num_decode_tokens > 0
+        has_prefill = metadata.num_prefills > 0
+    else:
+        has_decode = metadata.decode is not None
+        has_prefill = metadata.prefill is not None
    if not has_decode and not has_prefill:
        raise RuntimeError("Metadata has neither decode nor prefill metadata")
+    if is_sparse and force_dense_mha and not has_prefill:
+        raise RuntimeError(
+            "Sparse MLA dense_mha benchmark did not produce prefill metadata. "
+            "Check reorder_batch_threshold/path forcing."
+        )

    num_decode = (
        metadata.num_decode_tokens
@@ -871,7 +946,6 @@ def _run_single_benchmark(
                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
@@ -898,6 +972,48 @@ def _run_single_benchmark(
            throughput_tokens_per_sec=0.0,
        )

+    if config.torch_profile:
+        profile_dir = Path(
+            config.torch_profile_dir or "benchmark_outputs/torch_profiles"
+        )
+        profile_dir.mkdir(parents=True, exist_ok=True)
+        trace_name = _safe_profile_name(f"{config.backend}_{config.batch_spec}")
+        trace_path = profile_dir / f"{trace_name}.json"
+        iters = max(config.torch_profile_iters, 1)
+
+        forward_fn()
+        torch.accelerator.synchronize()
+        with torch.profiler.profile(
+            activities=[
+                torch.profiler.ProfilerActivity.CPU,
+                torch.profiler.ProfilerActivity.CUDA,
+            ],
+            record_shapes=True,
+            profile_memory=True,
+            with_stack=False,
+        ) as prof:
+            for _ in range(iters):
+                forward_fn()
+                torch.accelerator.synchronize()
+                prof.step()
+        prof.export_chrome_trace(str(trace_path))
+        print(f"Saved PyTorch profiler trace to {trace_path}")
+        print(
+            prof.key_averages().table(
+                sort_by="cuda_time_total",
+                row_limit=25,
+            )
+        )
+        return BenchmarkResult(
+            config=config,
+            mean_time=0.0,
+            median_time=0.0,
+            std_time=0.0,
+            min_time=0.0,
+            max_time=0.0,
+            throughput_tokens_per_sec=0.0,
+        )
+
    all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)

    # Convert ms to seconds per layer
@@ -920,6 +1036,7 @@ def _run_mla_benchmark_batched(
    configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
    index_topk: int = 2048,
    prefill_backend: str | None = None,
+    sparse_mla_force_mqa: bool = False,
    output_scale: float | None = None,
    fuse_quant_op: bool = False,
 ) -> list[BenchmarkResult]:
@@ -940,6 +1057,8 @@ def _run_mla_benchmark_batched(
        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.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+            forward_mqa (even prefill tokens).

    Returns:
        List of BenchmarkResult objects
@@ -980,21 +1099,41 @@ def _run_mla_benchmark_batched(
        sum(r.q_len for r in parse_batch_spec(cfg.batch_spec))
        for cfg, *_ in configs_with_params
    )
+    max_model_len = max(
+        max_total_q,
+        max(
+            getattr(cfg, "max_model_len", None) or 32768
+            for cfg, *_ in configs_with_params
+        ),
+    )

    # Create and set vLLM config for MLA (reused across all benchmarks)
    vllm_config = create_minimal_vllm_config(
        model_name="deepseek-v3",  # Used only for model path
        block_size=block_size,
        max_num_batched_tokens=max_total_q,
+        max_model_len=max_model_len,
        mla_dims=mla_dims,  # Use custom dims from config or default
        index_topk=index_topk if is_sparse else None,
        prefill_backend=prefill_backend,
        kv_cache_dtype=kv_cache_dtype,
+        sparse_mla_force_mqa=sparse_mla_force_mqa,
    )

    results = []

+    # Initialize workspace manager (needed by metadata builders)
+    from vllm.v1.worker.workspace import (
+        init_workspace_manager,
+        is_workspace_manager_initialized,
+    )
+
+    if not is_workspace_manager_initialized():
+        init_workspace_manager(device)
+
    with set_current_vllm_config(vllm_config):
+        _ensure_single_rank_model_parallel()
+
        # Create backend impl, layer, builder, and indexer (reused across benchmarks)
        impl, layer, builder_instance, indexer = _create_backend_impl(
            backend_cfg,
@@ -1040,9 +1179,20 @@ def _run_mla_benchmark_batched(
        for config, threshold, num_splits in configs_with_params:
            # Set threshold for this benchmark (FlashAttn/FlashMLA only)
            original_threshold = None
-            if threshold is not None and builder_instance:
+            effective_threshold = threshold
+            force_dense_mha = (
+                is_sparse
+                and getattr(config, "sparse_mla_mha_mode", "auto") == "dense"
+                and not getattr(config, "sparse_mla_force_mqa", False)
+            )
+            if force_dense_mha:
+                # Sparse MLA normally treats q_len <= 1 as decode. Use an
+                # impossible threshold so dense_mha benchmarks actually run
+                # the prefill/MHA path, including q_len=1 short extends.
+                effective_threshold = -1
+            if effective_threshold is not None and builder_instance:
                original_threshold = builder_instance.reorder_batch_threshold
-                builder_instance.reorder_batch_threshold = threshold
+                builder_instance.reorder_batch_threshold = effective_threshold

            # Set num_splits for CUTLASS
            original_num_splits = None
@@ -1090,6 +1240,7 @@ def run_mla_benchmark(
    num_kv_splits: int | None = None,
    index_topk: int = 2048,
    prefill_backend: str | None = None,
+    sparse_mla_force_mqa: bool = False,
    output_scale: float | None = None,
    fuse_quant_op: bool = False,
 ) -> BenchmarkResult | list[BenchmarkResult]:
@@ -1111,6 +1262,8 @@ 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.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+            forward_mqa (even prefill tokens).
        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.
@@ -1142,6 +1295,7 @@ def run_mla_benchmark(
        configs_with_params,
        index_topk,
        prefill_backend=prefill_backend,
+        sparse_mla_force_mqa=sparse_mla_force_mqa,
        output_scale=output_scale,
        fuse_quant_op=fuse_quant_op,
    )
@@ -12,7 +12,7 @@ from dataclasses import dataclass, field
 import aiohttp
 import huggingface_hub.constants
 from tqdm.asyncio import tqdm
-from transformers import AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast
+from transformers import AutoTokenizer, PythonBackend, TokenizersBackend

 # 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,
-) -> PreTrainedTokenizer | PreTrainedTokenizerFast:
+) -> PythonBackend | TokenizersBackend:
    if pretrained_model_name_or_path is not None and not os.path.exists(
        pretrained_model_name_or_path
    ):
@@ -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.platforms import current_platform
+from vllm.utils.platform_utils import get_device_name_as_file_name

 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 = current_platform.get_device_name().replace(" ", "_")
+            device_name = get_device_name_as_file_name()
            tuned_config_file_name = (
                f"E={num_experts},N={inter_dim},device_name={device_name},"
                f"dtype=int4_w4a16,backend=flydsl.json"
@@ -0,0 +1,201 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark the RDNAHybridW4A16LinearKernel across decode and prefill shapes.
+
+Usage:
+    python benchmark_int4_gemm.py
+    python benchmark_int4_gemm.py --models Qwen/Qwen3-4B
+    python benchmark_int4_gemm.py --group-size 128
+"""
+
+import argparse
+import copy
+import itertools
+import os
+
+import torch
+
+from vllm.triton_utils import triton
+
+# ---------------------------------------------------------------------------
+# Weight shapes: [K, N], TP_SPLIT_DIM
+# ---------------------------------------------------------------------------
+WEIGHT_SHAPES = {
+    "Qwen/Qwen3-4B": [
+        ([2560, 3840], 1),  # qkv_proj
+        ([2560, 2560], 0),  # o_proj
+        ([2560, 19456], 1),  # gate_up_proj
+        ([9728, 2560], 0),  # down_proj
+    ],
+    "Qwen/Qwen2.5-7B-Instruct": [
+        ([3584, 4608], 1),
+        ([3584, 3584], 0),
+        ([3584, 37888], 1),
+        ([18944, 3584], 0),
+    ],
+    "trymirai/SmolLM2-1.7B-Instruct-AWQ": [
+        ([2048, 6144], 1),  # qkv_proj
+        ([2048, 2048], 0),  # o_proj
+        ([2048, 16384], 1),  # gate_up_proj
+        ([8192, 2048], 0),  # down_proj
+    ],
+    "RedHatAI/Qwen3-8B-quantized.w4a16": [
+        ([4096, 6144], 1),  # qkv_proj
+        ([4096, 4096], 0),  # o_proj
+        ([4096, 24576], 1),  # gate_up_proj
+        ([12288, 4096], 0),  # down_proj
+    ],
+}
+
+
+# ---------------------------------------------------------------------------
+# Weight packing
+# ---------------------------------------------------------------------------
+def prepare_hybrid_weights(K, N, group_size, device="cuda"):
+    """Create random weights for benchmarking.
+
+    Returns (w_q_skinny, w_s_skinny, w_fp16, w_zp). The triton path derives
+    its int32 view from w_q_skinny, so no separate int32 buffer is returned.
+    """
+    num_groups = K // group_size
+
+    # Random packed weights — actual values don't matter for throughput
+    w_q_skinny_i32 = torch.randint(
+        0, 2**31, (N, K // 8), dtype=torch.int32, device=device
+    )
+    w_q_skinny = w_q_skinny_i32.view(torch.int8).contiguous()
+    w_s_skinny = torch.randn(N, num_groups, dtype=torch.float16, device=device) * 0.01
+
+    # Raw per-group zero-points for asymmetric benchmarks
+    w_zp = torch.randint(0, 16, (N, num_groups), dtype=torch.int32, device=device).to(
+        torch.float16
+    )
+
+    # FP16 baseline for F.linear
+    w_fp16 = torch.randn(N, K, dtype=torch.float16, device=device) * 0.01
+
+    return w_q_skinny, w_s_skinny, w_fp16, w_zp
+
+
+# ---------------------------------------------------------------------------
+# Benchmark
+# ---------------------------------------------------------------------------
+PROVIDERS = ["torch-fp16", "hybrid-w4a16", "hybrid-w4a16-zp"]
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096],
+        x_log=False,
+        line_arg="provider",
+        line_vals=PROVIDERS,
+        line_names=PROVIDERS,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="FP16 vs Hybrid W4A16",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K, group_size, weights):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.float16
+    a = torch.randn((M, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-fp16":
+        w_fp16 = weights["w_fp16"]
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, w_fp16),
+            quantiles=quantiles,
+        )
+    elif provider in ("hybrid-w4a16", "hybrid-w4a16-zp"):
+        from vllm.model_executor.kernels.linear.mixed_precision import (
+            rdna_hybrid_w4a16 as _k,
+        )
+
+        _rdna_hybrid_w4a16_apply_impl = _k._rdna_hybrid_w4a16_apply_impl
+        from vllm.utils.platform_utils import num_compute_units
+
+        w = weights
+        cu_count = num_compute_units()
+        use_zp = provider == "hybrid-w4a16-zp"
+
+        def run():
+            return _rdna_hybrid_w4a16_apply_impl(
+                a,
+                w["w_q_skinny"],
+                w["w_s_skinny"],
+                w["w_zp"] if use_zp else None,
+                None,  # bias
+                cu_count,
+                group_size,
+            )
+
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            run,
+            quantiles=quantiles,
+        )
+    else:
+        return 0.0, 0.0, 0.0
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    KN_model_names = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            KN_model_names.append(KN)
+    return KN_model_names
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Benchmark RDNAHybridW4A16LinearKernel"
+    )
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["Qwen/Qwen3-4B"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    parser.add_argument("--group-size", type=int, default=128)
+    parser.add_argument("--save-path", type=str, default=None)
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        group_size = args.group_size
+        print(f"\n{'=' * 70}")
+        print(f"{model}, N={N} K={K}, group_size={group_size}")
+        print(f"{'=' * 70}")
+
+        w_q_skinny, w_s_skinny, w_fp16, w_zp = prepare_hybrid_weights(K, N, group_size)
+
+        weights = {
+            "w_q_skinny": w_q_skinny,
+            "w_s_skinny": w_s_skinny,
+            "w_fp16": w_fp16,
+            "w_zp": w_zp,
+        }
+
+        save_path = args.save_path or f"bench_int4_res_n{N}_k{K}"
+        os.makedirs(save_path, exist_ok=True)
+        benchmark.run(
+            print_data=True,
+            show_plots=False,
+            save_path=save_path,
+            N=N,
+            K=K,
+            group_size=group_size,
+            weights=weights,
+        )
+
+    print("\nBenchmark finished!")
@@ -0,0 +1,108 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Benchmark ReLUSquaredActivation: custom CUDA kernel vs forward_native, both
+# eager and under torch.compile (Inductor fuses relu+square into one kernel).
+
+import itertools
+
+import torch
+import torch.nn.functional as F
+
+import vllm.model_executor.layers.activation  # noqa: F401
+from vllm.benchmarks.lib.utils import default_vllm_config
+from vllm.triton_utils import triton
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
+
+# Capped so the largest tensor stays under 2**31 elements: the shared activation
+# kernel computes the per-token pointer offset (blockIdx.x * d) in 32-bit, which
+# overflows for tensors with >2**32 elements. Realistic token counts are well
+# below this; the kernel-vs-native gap is already clear at these sizes.
+batch_size_range = [1, 16, 128]
+seq_len_range = [1, 16, 64, 1024]
+intermediate_size = [3072, 9728, 12288]
+configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
+
+
+@default_vllm_config()
+def benchmark_relu_squared(
+    batch_size: int,
+    seq_len: int,
+    intermediate_size: int,
+    provider: str,
+    dtype: torch.dtype,
+):
+    device = "cuda"
+    num_tokens = batch_size * seq_len
+    set_random_seed(42)
+    torch.set_default_device(device)
+
+    x = torch.randn(num_tokens, intermediate_size, dtype=dtype, device=device)
+    out = torch.empty_like(x)
+
+    def native(x: torch.Tensor) -> torch.Tensor:
+        return torch.square(F.relu(x))
+
+    # Verify the custom kernel matches the native implementation before timing.
+    ref = native(x)
+    torch.ops._C.relu_squared(out, x)
+    torch.testing.assert_close(out, ref)
+
+    if provider == "custom":
+        # Custom CUDA kernel — single fused kernel.
+        fn = lambda: torch.ops._C.relu_squared(out, x)
+    elif provider == "native":
+        # forward_native, eager — relu and square as separate ops.
+        fn = lambda: native(x)
+    elif provider == "native_compiled":
+        # forward_native under torch.compile — Inductor fuses relu+square.
+        # This is the real production baseline (custom ops are off when
+        # Inductor is enabled), so it is the comparison reviewers care about.
+        compiled = torch.compile(native)
+        compiled(x)  # warm up / trigger compilation before timing
+        fn = lambda: compiled(x)
+
+    ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+        fn, quantiles=[0.5, 0.2, 0.8]
+    )
+    return ms, max_ms, min_ms
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark ReLUSquaredActivation: custom kernel vs native."
+    )
+    parser.add_argument(
+        "--dtype",
+        type=str,
+        choices=["half", "bfloat16", "float"],
+        default="bfloat16",
+    )
+    args = parser.parse_args()
+
+    dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
+
+    perf_report = triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["batch_size", "seq_len", "intermediate_size"],
+            x_vals=configs,
+            line_arg="provider",
+            line_vals=["custom", "native_compiled", "native"],
+            line_names=[
+                "Custom Kernel",
+                "Native (torch.compile)",
+                "Native (eager)",
+            ],
+            styles=[("blue", "-"), ("green", "-"), ("red", "-")],
+            ylabel="ms",
+            plot_name="relu_squared-eager-performance",
+            args={},
+        )
+    )
+
+    perf_report(
+        lambda batch_size, seq_len, intermediate_size, provider: benchmark_relu_squared(
+            batch_size, seq_len, intermediate_size, provider, dtype
+        )
+    ).run(print_data=True)
@@ -19,6 +19,7 @@ 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)

@@ -264,7 +265,7 @@ def save_configs(
    input_type="fp8",
 ) -> None:
    os.makedirs(save_path, exist_ok=True)
-    device_name = current_platform.get_device_name().replace(" ", "_")
+    device_name = get_device_name_as_file_name()
    json_file_name = (
        f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
        f"block_shape=[{block_n},{block_k}].json"
@@ -0,0 +1,241 @@
+#!/bin/bash
+set -eoux pipefail
+
+########################################
+# Resolve repo root (IMPORTANT)
+########################################
+REPO_ROOT="$(pwd)"
+
+cd "$REPO_ROOT"
+
+########################################
+# DevPI configuration
+########################################
+
+IBM_DEVPI_URL=${IBM_DEVPI_URL:-"https://wheels.developerfirst.ibm.com/ppc64le/linux/+simple/"}
+RHOAI_INDEX_URL=${RHOAI_INDEX_URL:-"https://console.redhat.com/api/pypi/public-rhai/rhoai/3.4/cpu-ubi9/simple/"}
+
+########################################
+# wheel dir
+########################################
+
+WHEEL_DIR=${WHEEL_DIR:-"/tmp/wheels"}
+mkdir -p "$WHEEL_DIR"
+
+########################################
+# Helpers
+########################################
+try_install_from_devpi() {
+    local pkg=$1
+    uv pip install \
+        --extra-index-url "${IBM_DEVPI_URL}" \
+        --index-strategy unsafe-best-match \
+        --no-build-isolation \
+        "${pkg}"
+}
+
+########################################
+# Package Versions
+########################################
+cd "$REPO_ROOT"
+TORCH_VERSION=${TORCH_VERSION:-$(grep -E '^torch==.+==\s*"ppc64le"' requirements/cpu.txt | grep -Eo '\b[0-9\.]+\b' || true)}
+TORCH_VERSION=${TORCH_VERSION:-2.11.0}
+
+TORCHVISION_VERSION=${TORCHVISION_VERSION:-0.26.0}
+TORCHAUDIO_VERSION=${TORCHAUDIO_VERSION:-${TORCH_VERSION}}
+
+export TORCH_VERSION
+export TORCHVISION_VERSION
+export TORCHAUDIO_VERSION
+export OPENCV_VERSION=${OPENCV_VERSION:-4.13.0.92}
+export XGRAMMAR_VERSION=${XGRAMMAR_VERSION:-0.2.1}
+
+########################################
+# install system dependencies
+########################################
+
+rpm -ivh https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm || true
+
+microdnf install -y \
+    python3.12 python3.12-devel python3.12-pip gcc \
+    git jq gcc-toolset-14 gcc-toolset-14-libatomic-devel \
+    automake libtool clang-devel openssl-devel \
+    harfbuzz-devel kmod lcms2-devel libimagequant-devel libjpeg-turbo-devel \
+    llvm15-devel libraqm-devel libtiff-devel libwebp-devel libxcb-devel \
+    ninja-build openjpeg2-devel pkgconfig \
+    tcl-devel tk-devel xsimd-devel zeromq-devel zlib-devel patchelf file openblas openblas-devel protobuf numactl numactl-devel openmpi openmpi-devel
+    
+rpm -ivh --nodeps \
+    https://mirror.stream.centos.org/9-stream/CRB/ppc64le/os/Packages/protobuf-lite-devel-3.14.0-17.el9.ppc64le.rpm
+
+rpm -ivh --nodeps \
+    https://mirror.stream.centos.org/9-stream/CRB/ppc64le/os/Packages/protobuf-devel-3.14.0-17.el9.ppc64le.rpm
+
+rpm -ivh --nodeps \
+    https://mirror.stream.centos.org/9-stream/CRB/ppc64le/os/Packages/protobuf-compiler-3.14.0-17.el9.ppc64le.rpm
+
+########################################
+# Python 3.12 virtual environment
+########################################
+
+python3.12 -m venv /opt/vllm
+source /opt/vllm/bin/activate
+
+export PATH=/opt/vllm/bin:$PATH
+
+python --version
+
+########################################
+# install build tools (stable uv)
+########################################
+
+pip install -U pip setuptools-rust
+pip install uv
+pip install "setuptools<70" build wheel cmake auditwheel
+uv pip install "setuptools<70" cython meson-python pybind11 "sympy>=1.13.3" --no-build-isolation
+
+########################################
+# Rust
+########################################
+
+curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
+source /root/.cargo/env
+
+########################################
+# Compiler env
+########################################
+
+source /opt/rh/gcc-toolset-14/enable
+
+export PATH=/usr/lib64/llvm15/bin:$PATH
+export LLVM_CONFIG=/usr/lib64/llvm15/bin/llvm-config
+export CMAKE_ARGS="-DPython3_EXECUTABLE=python"
+
+export MAX_JOBS=${MAX_JOBS:-$(nproc)}
+export GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=1
+
+########################################
+# Install Packages From Devpi
+########################################
+uv pip install numpy==2.3.5 pillow==12.2.0 --extra-index-url "$IBM_DEVPI_URL"
+try_install_from_devpi "opencv-python-headless==${OPENCV_VERSION}"
+try_install_from_devpi "torch==${TORCH_VERSION}"
+try_install_from_devpi "torchvision==${TORCHVISION_VERSION}"
+
+########################################
+# torch audio
+########################################
+
+TEMP_BUILD_DIR=$(mktemp -d)
+cd "${TEMP_BUILD_DIR}"
+export BUILD_SOX=1 BUILD_KALDI=1 BUILD_RNNT=1 USE_FFMPEG=0 USE_ROCM=0 USE_CUDA=0
+export TORCHAUDIO_TEST_ALLOW_SKIP_IF_NO_FFMPEG=1
+git clone --recursive https://github.com/pytorch/audio.git -b v${TORCHAUDIO_VERSION}
+cd audio
+#patching 
+sed -i '
+s|_CSRC_DIR / "_torchaudio.cpp"|str(_CSRC_DIR / "_torchaudio.cpp")|;
+s|_CSRC_DIR / "utils.cpp"|str(_CSRC_DIR / "utils.cpp")|;
+s|sources=\[_CSRC_DIR / s for s in sources\]|sources=[str(_CSRC_DIR / s) for s in sources]|;
+' tools/setup_helpers/extension.py
+MAX_JOBS=${MAX_JOBS:-$(nproc)} \
+BUILD_VERSION=${TORCHAUDIO_VERSION} \
+uv build --wheel --out-dir "${WHEEL_DIR}" --no-build-isolation
+uv pip install "${WHEEL_DIR}"/torchaudio*.whl
+cd "${REPO_ROOT}"
+rm -rf "${TEMP_BUILD_DIR}"
+
+########################################
+# Xgrammar
+########################################
+uv pip install \
+   "scikit-build-core==0.11.6" \
+   "pyproject-metadata<0.8" \
+    pathspec \
+    packaging \
+    distro \
+   "setuptools<70" \
+    setuptools_scm \
+    cmake \
+    ninja \
+    pybind11 \
+    nanobind
+uv pip install apache-tvm-ffi==0.1.12 \
+  --no-build-isolation \
+  --no-cache
+
+TEMP_BUILD_DIR=$(mktemp -d)
+
+pushd "${TEMP_BUILD_DIR}"
+
+export CFLAGS="-fno-lto -mcpu=power9"
+export CXXFLAGS="-fno-lto -mcpu=power9"
+export LDFLAGS="-fno-lto"
+export PATH=/opt/vllm/bin:$PATH
+
+export Python_EXECUTABLE=/opt/vllm/bin/python3
+export Python3_EXECUTABLE=/opt/vllm/bin/python3
+export PYTHON_EXECUTABLE=/opt/vllm/bin/python3
+
+export Python_ROOT_DIR=/opt/vllm
+export Python3_ROOT_DIR=/opt/vllm
+
+git clone \
+    --recursive \
+    https://github.com/mlc-ai/xgrammar \
+    -b "v${XGRAMMAR_VERSION}"
+
+cd xgrammar
+
+cp cmake/config.cmake .
+export PYTHONPATH=/opt/vllm/lib64/python3.12/site-packages:/opt/vllm/lib/python3.12/site-packages:${PYTHONPATH:-}
+
+uv build \
+    --wheel \
+    --out-dir "${WHEEL_DIR}" \
+    --no-build-isolation
+
+uv pip install "${WHEEL_DIR}"/xgrammar*.whl -v
+
+popd
+
+rm -rf "${TEMP_BUILD_DIR}"
+cd "${REPO_ROOT}"
+
+########################################
+# RHOAI Binary Downloads
+########################################
+pip download \
+    --index-url "${RHOAI_INDEX_URL}" \
+    --only-binary=:all: \
+    --no-deps \
+    llvmlite==0.47.0 \
+    -d "${WHEEL_DIR}"
+
+pip download \
+    --index-url "${RHOAI_INDEX_URL}" \
+    --only-binary=:all: \
+    --no-deps \
+    Numba==0.65.0 \
+    -d "${WHEEL_DIR}"
+
+########################################
+# install built wheels
+########################################
+uv pip install setuptools_scm maturin setuptools-rust ninja scikit-build-core pybind11 nanobind \
+    --no-build-isolation
+uv pip install "${WHEEL_DIR}"/*.whl
+
+########################################
+# install remaining deps
+########################################
+
+sed -i.bak -e 's/.*torch.*//g' pyproject.toml requirements/*.txt
+
+uv pip install "setuptools>=78.1.1" --no-build-isolation
+
+export PKG_CONFIG_PATH=/usr/local/lib/pkgconfig:/usr/local/lib64/pkgconfig:/usr/lib64/pkgconfig
+
+uv pip install -r requirements/common.txt \
+               -r requirements/cpu.txt \
+               -r requirements/build/cpu.txt --index-strategy unsafe-best-match
@@ -19,7 +19,7 @@ else()
  FetchContent_Declare(
        flashmla
        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG a6ec2ba7bd0a7dff98b3f4d3e6b52b159c48d78b
+        GIT_TAG a8f794d1251cbfd88a5011445dd5582289c727e4
        GIT_PROGRESS TRUE
        CONFIGURE_COMMAND ""
        BUILD_COMMAND ""
@@ -35,7 +35,7 @@ set(FLASHMLA_VENDOR_DIR "${CMAKE_SOURCE_DIR}/vllm/third_party/flashmla")
 file(MAKE_DIRECTORY "${FLASHMLA_VENDOR_DIR}")
 file(READ "${flashmla_SOURCE_DIR}/flash_mla/flash_mla_interface.py"
     FLASHMLA_INTERFACE_CONTENT)
-string(REPLACE "import flash_mla.cuda as flash_mla_cuda"
+string(REPLACE "flash_mla_cuda = torch.ops._flashmla_C"
               "import vllm._flashmla_C\nflash_mla_cuda = torch.ops._flashmla_C"
               FLASHMLA_INTERFACE_CONTENT
               "${FLASHMLA_INTERFACE_CONTENT}")
@@ -72,7 +72,7 @@ if(FLASH_MLA_ARCHS)
    list(APPEND VLLM_FLASHMLA_GPU_FLAGS "--expt-relaxed-constexpr" "--expt-extended-lambda" "--use_fast_math")

    set(FlashMLA_SOURCES
-        ${flashmla_SOURCE_DIR}/csrc/torch_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/api/api.cpp

        # Misc kernels for decoding
        ${flashmla_SOURCE_DIR}/csrc/smxx/decode/get_decoding_sched_meta/get_decoding_sched_meta.cu
@@ -128,6 +128,7 @@ if(FLASH_MLA_ARCHS)

    set(FlashMLA_Extension_INCLUDES
        ${flashmla_SOURCE_DIR}/csrc
+        ${flashmla_SOURCE_DIR}/csrc/kerutils/include
        ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
        ${flashmla_SOURCE_DIR}/csrc/cutlass/include
        ${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
@@ -152,15 +153,18 @@ if(FLASH_MLA_ARCHS)
        USE_SABI 3
        WITH_SOABI)

-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    # Also enable C++20 for the FlashMLA sources (required for std::span, requires, etc.)
+    # Enable C++20 for the FlashMLA sources (required for std::span, requires, etc.)
    target_compile_options(_flashmla_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>
        $<$<COMPILE_LANGUAGE:CXX>:-std=c++20>
        $<$<COMPILE_LANGUAGE:CUDA>:-std=c++20>)

+    # _flashmla_C is now ABI-stable torch 2.11+
+    target_compile_definitions(_flashmla_C PRIVATE
+        TORCH_TARGET_VERSION=0x020B000000000000ULL)
+    if(VLLM_GPU_LANG STREQUAL "CUDA")
+        target_compile_definitions(_flashmla_C PRIVATE USE_CUDA)
+    endif()
+
    define_extension_target(
        _flashmla_extension_C
        DESTINATION vllm
@@ -172,11 +176,12 @@ if(FLASH_MLA_ARCHS)
        USE_SABI 3
        WITH_SOABI)

-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    target_compile_options(_flashmla_extension_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
+    # _flashmla_extension_C is now ABI-stable w/ torch 2.11+
+    target_compile_definitions(_flashmla_extension_C PRIVATE
+        TORCH_TARGET_VERSION=0x020B000000000000ULL)
+    if(VLLM_GPU_LANG STREQUAL "CUDA")
+        target_compile_definitions(_flashmla_extension_C PRIVATE USE_CUDA)
+    endif()
 else()
    message(STATUS "FlashMLA will not compile: unsupported CUDA architecture ${CUDA_ARCHS}")
    # Create empty targets for setup.py on unsupported systems
@@ -17,7 +17,7 @@ else()
  FetchContent_Declare(
    fmha_sm100
    GIT_REPOSITORY https://github.com/vllm-project/MSA.git
-    GIT_TAG fee783153f3efe57e3e933c5cb7e267a7cebcfb5
+    GIT_TAG 2e63ec37a0fc29bc20f39cd1a52e0f5affc33a73
    GIT_PROGRESS TRUE
    CONFIGURE_COMMAND ""
    BUILD_COMMAND ""
@@ -0,0 +1,50 @@
+include(FetchContent)
+
+if(DEFINED ENV{TML_FA4_SRC_DIR})
+  set(TML_FA4_SRC_DIR $ENV{TML_FA4_SRC_DIR})
+endif()
+
+if(TML_FA4_SRC_DIR)
+  FetchContent_Declare(
+    tml_fa4
+    SOURCE_DIR ${TML_FA4_SRC_DIR}
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND "")
+else()
+  FetchContent_Declare(
+    tml_fa4
+    GIT_REPOSITORY https://github.com/vllm-project/tml-fa4.git
+    GIT_TAG 13374f0c855acc1add1bf30444bd67aebbc24a8e
+    GIT_PROGRESS TRUE
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND "")
+endif()
+
+FetchContent_GetProperties(tml_fa4)
+if(NOT tml_fa4_POPULATED)
+  FetchContent_Populate(tml_fa4)
+endif()
+message(STATUS "tml-fa4 is available at ${tml_fa4_SOURCE_DIR}")
+
+add_custom_target(tml_fa4)
+
+# Install into a private namespace so this implementation cannot shadow the
+# flash_attn package used by vLLM's standard attention backends.
+install(CODE "
+  file(GLOB_RECURSE TML_FA4_PY_FILES
+    \"${tml_fa4_SOURCE_DIR}/flash_attn/cute/*.py\")
+  foreach(SRC_FILE \${TML_FA4_PY_FILES})
+    file(RELATIVE_PATH REL_PATH
+      \"${tml_fa4_SOURCE_DIR}/flash_attn/cute\" \${SRC_FILE})
+    set(DST_FILE
+      \"\${CMAKE_INSTALL_PREFIX}/vllm/third_party/tml_fa4/\${REL_PATH}\")
+    get_filename_component(DST_DIR \${DST_FILE} DIRECTORY)
+    file(MAKE_DIRECTORY \${DST_DIR})
+    file(READ \${SRC_FILE} FILE_CONTENTS)
+    string(REPLACE
+      \"flash_attn.cute\"
+      \"vllm.third_party.tml_fa4\"
+      FILE_CONTENTS \"\${FILE_CONTENTS}\")
+    file(WRITE \${DST_FILE} \"\${FILE_CONTENTS}\")
+  endforeach()
+" COMPONENT tml_fa4)
@@ -39,7 +39,7 @@ else()
  FetchContent_Declare(
          vllm-flash-attn
          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 2c839c33742309ec41e620bf837495ec9926c56e
+          GIT_TAG caaa4eb59845388a20b1f435ecaafb4bd9517ad8
          GIT_PROGRESS TRUE
          # Don't share the vllm-flash-attn build between build types
          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
@@ -323,8 +323,6 @@ class AttentionImpl<ISA::VSX, scalar_t, head_dim> {
      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
      const int64_t block_size, const int64_t block_size_stride,
      const float k_inv = 0.0f, const float v_inv = 0.0f) {
-    // k_inv and v_inv are unused on VSX: FP8 KV cache is not supported on
-    // PowerPC. The parameters are present to match the common interface.
 #pragma omp parallel for collapse(2)
    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
@@ -4,7 +4,7 @@
 #if defined(__x86_64__)
  // x86 implementation
  #include "cpu_types_x86.hpp"
-#elif defined(__POWER9_VECTOR__)
+#elif defined(__powerpc__)
  // ppc implementation
  #include "cpu_types_vsx.hpp"
 #elif defined(__s390x__)
@@ -41,4 +41,4 @@ inline int get_max_threads() {
 }
 }  // namespace cpu_utils

-#endif
+#endif
@@ -344,53 +344,133 @@ struct FP32Vec8 : public Vec<FP32Vec8> {

    return result;
  }
-
  FP32Vec8 exp() const {
-    // TODO: Vectorize this
-    AliasReg ar;
-    ar.reg = reg;
-    f32x4x4_t ret;
-    ret.val[0][0] = std::exp(ar.values[0]);
-    ret.val[0][1] = std::exp(ar.values[1]);
-    ret.val[0][2] = std::exp(ar.values[2]);
-    ret.val[0][3] = std::exp(ar.values[3]);
-    ret.val[1][0] = std::exp(ar.values[4]);
-    ret.val[1][1] = std::exp(ar.values[5]);
-    ret.val[1][2] = std::exp(ar.values[6]);
-    ret.val[1][3] = std::exp(ar.values[7]);
-    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+    f32x4x2_t out;
+    const __vector float log2e = vec_splats(1.44269504088896341f);
+    const __vector float one = vec_splats(1.0f);
+    const __vector float min_x = vec_splats(-87.3f);
+    const __vector float max_x = vec_splats(88.7f);
+
+    // 5th-degree minimax polynomial for 2^r (r in [0,1))
+    const __vector float c1 = vec_splats(0.6931471805599453f);
+    const __vector float c2 = vec_splats(0.240226506959101f);
+    const __vector float c3 = vec_splats(0.05550410866482158f);
+    const __vector float c4 = vec_splats(0.009618129107628477f);
+    const __vector float c5 = vec_splats(0.0013333558146428443f);
+
+    for (int i = 0; i < 2; i++) {
+      __vector float x = reg.val[i];
+      x = vec_max(x, min_x);
+      x = vec_min(x, max_x);
+
+      __vector float y = vec_mul(x, log2e);
+
+      __vector float kf = vec_floor(y);
+      __vector float r = vec_sub(y, kf);
+
+      // Convert float to signed integer. Use vec_cts for PowerPC AltiVec
+      // compatibility.
+      __vector signed int k = vec_cts(kf, 0);
+      const __vector signed int min_k = vec_splats((signed int)-126);
+      const __vector signed int max_k = vec_splats((signed int)127);
+      k = vec_min(vec_max(k, min_k), max_k);
+
+      // Build 2^k from exponent bits
+      __vector signed int exp_int = vec_add(k, vec_splats((signed int)127));
+      __vector unsigned int bits = (__vector unsigned int)exp_int;
+      bits = vec_sl(bits, vec_splats((unsigned int)23));
+      __vector float pow2k = (__vector float)bits;
+
+      // Improved minimax polynomial
+      __vector float poly = vec_madd(c5, r, c4);
+      poly = vec_madd(poly, r, c3);
+      poly = vec_madd(poly, r, c2);
+      poly = vec_madd(poly, r, c1);
+      poly = vec_madd(poly, r, one);
+
+      out.val[i] = vec_mul(pow2k, poly);
+    }
+    return FP32Vec8(out);
  }

  FP32Vec8 tanh() const {
-    // TODO: Vectorize this
-    AliasReg ar;
-    ar.reg = reg;
-    f32x4x4_t ret;
-    ret.val[0][0] = std::tanh(ar.values[0]);
-    ret.val[0][1] = std::tanh(ar.values[1]);
-    ret.val[0][2] = std::tanh(ar.values[2]);
-    ret.val[0][3] = std::tanh(ar.values[3]);
-    ret.val[1][0] = std::tanh(ar.values[4]);
-    ret.val[1][1] = std::tanh(ar.values[5]);
-    ret.val[1][2] = std::tanh(ar.values[6]);
-    ret.val[1][3] = std::tanh(ar.values[7]);
-    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+    const __vector float one = vec_splats(1.0f);
+    const __vector float two = vec_splats(2.0f);
+    const __vector float zero = vec_splats(0.0f);
+    const __vector float sat = vec_splats(9.0f);
+
+    f32x4x2_t out;
+
+    for (int i = 0; i < 2; i++) {
+      __vector float x = reg.val[i];
+      __vector float ax = vec_abs(x);
+
+      __vector bool int mask = vec_cmpge(x, zero);
+      __vector float sign = vec_sel(vec_splats(-1.0f), one, mask);
+
+      __vector bool int saturated = vec_cmpge(ax, sat);
+
+      __vector float two_x = vec_mul(x, two);
+      f32x4x2_t tmp;
+      tmp.val[0] = two_x;
+      tmp.val[1] = two_x;
+      FP32Vec8 temp_vec(tmp);
+      vector float e = temp_vec.exp().reg.val[0];
+
+      vector float num = vec_sub(e, one);
+      vector float den = vec_add(e, one);
+      vector float t = vec_div(num, den);
+
+      out.val[i] = vec_sel(t, sign, saturated);
+    }
+    return FP32Vec8(out);
  }

  FP32Vec8 er() const {
-    // TODO: Vectorize this
-    AliasReg ar;
-    ar.reg = reg;
-    f32x4x4_t ret;
-    ret.val[0][0] = std::erf(ar.values[0]);
-    ret.val[0][1] = std::erf(ar.values[1]);
-    ret.val[0][2] = std::erf(ar.values[2]);
-    ret.val[0][3] = std::erf(ar.values[3]);
-    ret.val[1][0] = std::erf(ar.values[4]);
-    ret.val[1][1] = std::erf(ar.values[5]);
-    ret.val[1][2] = std::erf(ar.values[6]);
-    ret.val[1][3] = std::erf(ar.values[7]);
-    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+    const vector float a1 = vec_splats(0.254829592f);
+    const vector float a2 = vec_splats(-0.284496736f);
+    const vector float a3 = vec_splats(1.421413741f);
+    const vector float a4 = vec_splats(-1.453152027f);
+    const vector float a5 = vec_splats(1.061405429f);
+    const vector float p = vec_splats(0.3275911f);
+    const vector float one = vec_splats(1.0f);
+    const vector float zero = vec_splats(0.0f);
+    const vector float sat = vec_splats(6.0f);
+
+    f32x4x2_t ret;
+
+    for (int i = 0; i < 2; i++) {
+      vector float x = reg.val[i];
+      vector float ax = vec_abs(x);
+
+      vector bool int mask = vec_cmpge(x, zero);
+      vector float sign = vec_sel(vec_splats(-1.0f), one, mask);
+
+      vector bool int saturated = vec_cmpge(ax, sat);
+
+      vector float t = vec_div(one, vec_madd(p, ax, one));
+
+      vector float poly = a5;
+      poly = vec_madd(poly, t, a4);
+      poly = vec_madd(poly, t, a3);
+      poly = vec_madd(poly, t, a2);
+      poly = vec_madd(poly, t, a1);
+      poly = vec_mul(poly, t);
+
+      vector float x_squared = vec_mul(x, x);
+      vector float neg_x_squared = vec_mul(vec_splats(-1.0f), x_squared);
+      f32x4x2_t tmp;
+      tmp.val[0] = neg_x_squared;
+      tmp.val[1] = neg_x_squared;
+      FP32Vec8 exp_input(tmp);
+      vector float exp_term = exp_input.exp().reg.val[0];
+
+      vector float y = vec_nmsub(poly, exp_term, one);
+      vector float erf_val = vec_mul(sign, y);
+
+      ret.val[i] = vec_sel(erf_val, sign, saturated);
+    }
+    return FP32Vec8(ret);
  }

  FP32Vec8 operator*(const FP32Vec8& b) const {
@@ -1116,6 +1116,164 @@ void fused_sigmoid_gating_delta_rule_update_kernel_impl(
  });
 }

+// Speculative-decode variant: processes a varlen batch where each sequence has
+// ``q_len`` draft tokens, runs the recurrence sequentially over those tokens
+// (inside the kernel, so one dispatch handles the whole draft block), reads the
+// initial state from cache slot ``num_accepted-1`` and stores the state *after*
+// token ``t`` into cache slot ``t`` (multi-slot rollback, matching the GPU
+// kernel). Parallelized over (sequence, v_head); the per-sequence token loop is
+// sequential as required by the recurrence.
+template <typename scalar_t, typename param_t>
+void fused_sigmoid_gating_delta_rule_update_spec_kernel_impl(
+    const scalar_t* __restrict__ q_ptr,  // [T, HK, EK]
+    const scalar_t* __restrict__ k_ptr,  // [T, HK, EK]
+    const scalar_t* __restrict__ v_ptr,  // [T, HV, EV]
+    const param_t* __restrict__ A_log_ptr,
+    const scalar_t* __restrict__ a_ptr,  // [T, HV]
+    const scalar_t* __restrict__ dt_bias_ptr,
+    const scalar_t* __restrict__ b_ptr,  // [T, HV]
+    const int32_t* __restrict__ spec_indices_ptr,  // [N, S]
+    const int32_t* __restrict__ num_accepted_ptr,  // [N]
+    const int32_t* __restrict__ cu_seqlens_ptr,     // [N + 1]
+    float* __restrict__ state_ptr,
+    scalar_t* __restrict__ o_ptr,  // [T, HV, EV]
+    float* __restrict__ qk_scale_buf,  // [2, T, HK]
+    int64_t total_tokens,
+    int64_t batch_size,
+    int64_t spec_stride,
+    int64_t num_heads,
+    int64_t head_dim,
+    int64_t v_num_heads,
+    int64_t v_head_dim,
+    int64_t q_strideT,
+    int64_t q_strideH,
+    int64_t k_strideT,
+    int64_t k_strideH,
+    int64_t v_strideT,
+    int64_t v_strideH,
+    int64_t state_slot_stride,
+    bool use_qk_l2norm_in_kernel,
+    double softplus_threshold) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int64_t VecSize = bVec::size();
+  constexpr int64_t fVecSize = fVec::size();
+  int64_t group_size = v_num_heads / num_heads;
+  double scale = 1 / std::sqrt((double)head_dim);
+  fVec scale_vec = fVec((float)scale);
+
+  if (use_qk_l2norm_in_kernel) {
+    float eps = 1e-5f;
+    at::parallel_for(0, total_tokens * num_heads, 0, [&](int64_t begin, int64_t end) {
+      for (int64_t i = begin; i < end; ++i) {
+        int64_t ti = i / num_heads;
+        int64_t ni = i % num_heads;
+        const scalar_t* qp = q_ptr + ti * q_strideT + ni * q_strideH;
+        const scalar_t* kp = k_ptr + ti * k_strideT + ni * k_strideH;
+        float sq = 0.f, sk = 0.f;
+        for (int64_t d = 0; d < head_dim; ++d) {
+          float qv = (float)qp[d];
+          sq += qv * qv;
+          float kv = (float)kp[d];
+          sk += kv * kv;
+        }
+        qk_scale_buf[ti * num_heads + ni] = 1.f / std::sqrt(sq + eps);
+        qk_scale_buf[total_tokens * num_heads + ti * num_heads + ni] = 1.f / std::sqrt(sk + eps);
+      }
+    });
+  }
+
+  at::parallel_for(0, batch_size * v_num_heads, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t idx = begin; idx < end; ++idx) {
+      int64_t bi = idx / v_num_heads;
+      int64_t ni = idx % v_num_heads;
+      int64_t kh = ni / group_size;
+      int64_t q_start = cu_seqlens_ptr[bi];
+      int64_t q_len = cu_seqlens_ptr[bi + 1] - q_start;
+      if (q_len <= 0) {
+        continue;
+      }
+      int64_t acc = (int64_t)num_accepted_ptr[bi];
+      // Clamp acc-1 to >=0: when num_accepted is 0 the unclamped index reads
+      // out of bounds and yields an arbitrary prev_slot used to index the SSM
+      // state. Mirrors the GPU guard tl.maximum(num_accepted - 1, 0).
+      int64_t prev_slot =
+          (int64_t)spec_indices_ptr[bi * spec_stride + (acc > 0 ? acc - 1 : 0)];
+      for (int64_t t = 0; t < q_len; ++t) {
+        int64_t cur_slot = (int64_t)spec_indices_ptr[bi * spec_stride + t];
+        int64_t token = q_start + t;
+        const float* src = state_ptr + prev_slot * state_slot_stride + ni * head_dim * v_head_dim;
+        float* dst = state_ptr + cur_slot * state_slot_stride + ni * head_dim * v_head_dim;
+        float g_val = -std::exp((float)A_log_ptr[ni]) *
+            softplus((float)a_ptr[token * v_num_heads + ni] + (float)dt_bias_ptr[ni], softplus_threshold);
+        float g_val_exp = std::exp(g_val);
+        fVec g_val_exp_vec = fVec(g_val_exp);
+        float beta_val = 1.f / (1.f + std::exp(-(float)b_ptr[token * v_num_heads + ni]));
+        fVec beta_vec = fVec(beta_val);
+        int64_t q_offset = token * q_strideT + kh * q_strideH;
+        int64_t k_offset = token * k_strideT + kh * k_strideH;
+        float q_scale = use_qk_l2norm_in_kernel ? qk_scale_buf[token * num_heads + kh] : 1.f;
+        float k_scale =
+            use_qk_l2norm_in_kernel ? qk_scale_buf[total_tokens * num_heads + token * num_heads + kh] : 1.f;
+        int64_t v_offset = token * v_strideT + ni * v_strideH;
+        int64_t o_offset = (token * v_num_heads + ni) * v_head_dim;
+        int64_t dvi = 0;
+        for (; dvi <= v_head_dim - VecSize; dvi += VecSize) {
+          fVec kv_mem_vec0 = fVec(0.f);
+          fVec kv_mem_vec1 = fVec(0.f);
+          for (int di = 0; di < head_dim; ++di) {
+            fVec k_val_vec = fVec((float)k_ptr[k_offset + di] * k_scale);
+            fVec sv0 = fVec::loadu(src + di * v_head_dim + dvi);
+            fVec sv1 = fVec::loadu(src + di * v_head_dim + dvi + fVecSize);
+            kv_mem_vec0 = kv_mem_vec0 + sv0 * g_val_exp_vec * k_val_vec;
+            kv_mem_vec1 = kv_mem_vec1 + sv1 * g_val_exp_vec * k_val_vec;
+          }
+          bVec v_bvec = bVec::loadu(v_ptr + v_offset + dvi);
+          fVec v_vec0, v_vec1;
+          std::tie(v_vec0, v_vec1) = at::vec::convert_to_float(v_bvec);
+          fVec dt_vec0 = (v_vec0 - kv_mem_vec0) * beta_vec;
+          fVec dt_vec1 = (v_vec1 - kv_mem_vec1) * beta_vec;
+          fVec o_vec0 = fVec(0.f);
+          fVec o_vec1 = fVec(0.f);
+          for (int di = 0; di < head_dim; ++di) {
+            fVec q_vec = fVec((float)q_ptr[q_offset + di] * q_scale);
+            fVec k_vec = fVec((float)k_ptr[k_offset + di] * k_scale);
+            fVec sv0 = fVec::loadu(src + di * v_head_dim + dvi);
+            fVec sv1 = fVec::loadu(src + di * v_head_dim + dvi + fVecSize);
+            sv0 = sv0 * g_val_exp_vec + k_vec * dt_vec0;
+            sv1 = sv1 * g_val_exp_vec + k_vec * dt_vec1;
+            o_vec0 = o_vec0 + sv0 * q_vec * scale_vec;
+            o_vec1 = o_vec1 + sv1 * q_vec * scale_vec;
+            sv0.store(dst + di * v_head_dim + dvi);
+            sv1.store(dst + di * v_head_dim + dvi + fVecSize);
+          }
+          bVec o_vec = at::vec::convert_from_float<scalar_t>(o_vec0, o_vec1);
+          o_vec.store(o_ptr + o_offset + dvi);
+        }
+        for (; dvi < v_head_dim; ++dvi) {
+          float kv_mem_val = 0.f;
+          for (int di = 0; di < head_dim; ++di) {
+            float k_val = (float)k_ptr[k_offset + di] * k_scale;
+            kv_mem_val += src[di * v_head_dim + dvi] * g_val_exp * k_val;
+          }
+          float v_val = (float)v_ptr[v_offset + dvi];
+          float dt_val = (v_val - kv_mem_val) * beta_val;
+          float o_val = 0.f;
+          for (int di = 0; di < head_dim; ++di) {
+            float q_val = (float)q_ptr[q_offset + di] * q_scale;
+            float k_val = (float)k_ptr[k_offset + di] * k_scale;
+            float ns = src[di * v_head_dim + dvi] * g_val_exp + k_val * dt_val;
+            dst[di * v_head_dim + dvi] = ns;
+            o_val += ns * q_val * scale;
+          }
+          o_ptr[o_offset + dvi] = (scalar_t)o_val;
+        }
+        prev_slot = cur_slot;
+      }
+    }
+  });
+}
+
 template <typename scalar_t>
 void fused_gdn_gating_kernel_impl(
    float* __restrict__ A_log,
@@ -1500,6 +1658,103 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
  return core_attn_out;
 }

+// Speculative-decode update (multi-token, multi-slot rollback).
+// q: [T, HK, EK]  k: [T, HK, EK]  v: [T, HV, EV]
+// a: [T, HV]  b: [T, HV]
+// initial_state_source: [N_slots, HV, EK, EV] FP32 (updated in place)
+// spec_state_indices: [batch, S] INT32 (S = num_spec + 1)
+// num_accepted_tokens: [batch] INT32
+// cu_seqlens: [batch + 1] INT32
+// Returns output: [T, HV, EV]
+at::Tensor fused_sigmoid_gating_delta_rule_update_spec_cpu(
+    const at::Tensor& A_log,
+    const at::Tensor& dt_bias,
+    const at::Tensor& q,
+    const at::Tensor& k,
+    const at::Tensor& v,
+    const at::Tensor& a,
+    const at::Tensor& b,
+    at::Tensor& initial_state_source,
+    const at::Tensor& spec_state_indices,
+    const at::Tensor& num_accepted_tokens,
+    const at::Tensor& cu_seqlens,
+    bool use_qk_l2norm_in_kernel,
+    double softplus_beta = 1.0,
+    double softplus_threshold = 20.0) {
+  CHECK_DIM(3, q);
+  CHECK_DIM(3, v);
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(q);
+  int64_t total_tokens = q.size(0);
+  int64_t num_heads = q.size(1);
+  int64_t head_dim = q.size(2);
+  int64_t v_num_heads = v.size(1);
+  int64_t v_head_dim = v.size(2);
+  int64_t batch_size = cu_seqlens.size(0) - 1;
+  int64_t spec_stride = spec_state_indices.stride(0);
+  CHECK_INPUT_SHAPE_DTYPE<true>(k, {total_tokens, num_heads, head_dim}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(v, {total_tokens, v_num_heads, v_head_dim}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(a, {total_tokens, v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(b, {total_tokens, v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(dt_bias, {v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(num_accepted_tokens, {batch_size}, at::kInt);
+  CHECK_INPUT_SHAPE_DTYPE<true>(cu_seqlens, {batch_size + 1}, at::kInt);
+  CHECK_EQ(v_num_heads % num_heads, 0);
+  TORCH_CHECK(A_log.sizes() == at::IntArrayRef({v_num_heads}));
+  CHECK_INPUT_SHAPE_DTYPE<true>(
+      initial_state_source,
+      {initial_state_source.size(0), v_num_heads, head_dim, v_head_dim},
+      at::kFloat);
+  TORCH_CHECK(initial_state_source.size(0) >= batch_size,
+      "initial_state_source capacity too small: size(0)=",
+      initial_state_source.size(0), ", batch_size=", batch_size);
+
+  int64_t q_strideT = q.stride(0);
+  int64_t q_strideH = q.stride(1);
+  int64_t k_strideT = k.stride(0);
+  int64_t k_strideH = k.stride(1);
+  int64_t v_strideT = v.stride(0);
+  int64_t v_strideH = v.stride(1);
+  int64_t state_slot_stride = initial_state_source.stride(0);
+
+  at::Tensor o = at::empty({total_tokens, v_num_heads, v_head_dim}, q.options());
+  at::Tensor qk_scale_buf = at::empty({2, total_tokens, num_heads}, at::kFloat);
+
+  CPU_DISPATCH_REDUCED_FLOATING_TYPES_EXT(
+      q.scalar_type(), A_log.scalar_type(), "fused_sigmoid_gating_delta_rule_update_spec_kernel_impl", [&] {
+        fused_sigmoid_gating_delta_rule_update_spec_kernel_impl<scalar_t, param_t>(
+            q.data_ptr<scalar_t>(),
+            k.data_ptr<scalar_t>(),
+            v.data_ptr<scalar_t>(),
+            A_log.data_ptr<param_t>(),
+            a.data_ptr<scalar_t>(),
+            dt_bias.data_ptr<scalar_t>(),
+            b.data_ptr<scalar_t>(),
+            spec_state_indices.data_ptr<int32_t>(),
+            num_accepted_tokens.data_ptr<int32_t>(),
+            cu_seqlens.data_ptr<int32_t>(),
+            initial_state_source.data_ptr<float>(),
+            o.data_ptr<scalar_t>(),
+            qk_scale_buf.data_ptr<float>(),
+            total_tokens,
+            batch_size,
+            spec_stride,
+            num_heads,
+            head_dim,
+            v_num_heads,
+            v_head_dim,
+            q_strideT,
+            q_strideH,
+            k_strideT,
+            k_strideH,
+            v_strideT,
+            v_strideH,
+            state_slot_stride,
+            use_qk_l2norm_in_kernel,
+            softplus_threshold);
+      });
+  return o;
+}
+
 // A_log: [num_v_heads]
 // a: [batch, num_v_heads]
 // b: [batch, num_v_heads]
@@ -120,6 +120,14 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
    bool use_qk_l2norm_in_kernel, double softplus_beta = 1.0,
    double softplus_threshold = 20.0);

+at::Tensor fused_sigmoid_gating_delta_rule_update_spec_cpu(
+    const at::Tensor& A_log, const at::Tensor& dt_bias, const at::Tensor& q,
+    const at::Tensor& k, const at::Tensor& v, const at::Tensor& a,
+    const at::Tensor& b, at::Tensor& initial_state_source,
+    const at::Tensor& spec_state_indices, const at::Tensor& num_accepted_tokens,
+    const at::Tensor& cu_seqlens, bool use_qk_l2norm_in_kernel,
+    double softplus_beta = 1.0, double softplus_threshold = 20.0);
+
 std::tuple<at::Tensor, at::Tensor> fused_gdn_gating_cpu(
    const at::Tensor& A_log, const at::Tensor& a, const at::Tensor& b,
    const at::Tensor& dt_bias);
@@ -508,6 +516,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      "softplus_threshold=20.0) -> Tensor");
  ops.impl("fused_sigmoid_gating_delta_rule_update_cpu", torch::kCPU,
           &fused_sigmoid_gating_delta_rule_update_cpu);
+  ops.def(
+      "fused_sigmoid_gating_delta_rule_update_spec_cpu(Tensor A_log, Tensor "
+      "dt_bias, Tensor q, Tensor k, Tensor v, Tensor a, Tensor b, "
+      "Tensor(a!) initial_state_source, Tensor spec_state_indices, "
+      "Tensor num_accepted_tokens, Tensor cu_seqlens, bool "
+      "use_qk_l2norm_in_kernel, float softplus_beta=1.0, float "
+      "softplus_threshold=20.0) -> Tensor");
+  ops.impl("fused_sigmoid_gating_delta_rule_update_spec_cpu", torch::kCPU,
+           &fused_sigmoid_gating_delta_rule_update_spec_cpu);
  ops.def(
      "fused_gdn_gating_cpu(Tensor A_log, Tensor a, Tensor b, Tensor dt_bias) "
      "-> (Tensor, Tensor)");
@@ -553,7 +570,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif

  // fused moe
-#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
+#if defined(__AVX512F__) || \
+    (defined(__aarch64__) && !defined(__APPLE__) && defined(ARM_BF16_SUPPORT))
  ops.def(
      "prepack_moe_weight(Tensor weight, Tensor(a1!) packed_weight, str isa) "
      "-> ()");
@@ -564,7 +582,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      "bool skip_weighted, "
      "str act, str isa) -> ()");
  ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
-#endif  // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
+#endif
  ops.def(
      "mla_decode_kvcache("
      "   Tensor! out, Tensor query, Tensor kv_cache,"
@@ -76,14 +76,14 @@ inline int64_t get_available_l2_size() {
    if (l2_cache_size == 0) {
      l2_cache_size = 256 * 1024;
    }
-    return static_cast<int64_t>(l2_cache_size) >> 1;  // use 50% of L2 cache
+    return static_cast<int64_t>(l2_cache_size) >> 1;
  }();
  return size;
 #else
  static int64_t size = []() {
    auto caps = at::cpu::get_cpu_capabilities();
    const uint32_t l2_cache_size = caps.at("l2_cache_size").toInt();
-    return l2_cache_size >> 1;  // use 50% of L2 cache
+    return l2_cache_size >> 1;
  }();
  return size;
 #endif
@@ -669,6 +669,14 @@ __device__ __forceinline__ T gelu_quick_kernel(const T& x) {
  return (T)(((float)x) / (1.0f + expf(-1.702f * (float)x)));
 }

+template <typename T>
+__device__ __forceinline__ T relu_squared_kernel(const T& x) {
+  // relu(x)^2 — introduced in https://arxiv.org/abs/2109.08668v2
+  const float f = (float)x;
+  const float val = f > 0.0f ? f : 0.0f;
+  return (T)(val * val);
+}
+
 }  // namespace vllm

 void gelu_new(torch::stable::Tensor& out,    // [..., d]
@@ -688,3 +696,9 @@ void gelu_quick(torch::stable::Tensor& out,    // [..., d]
 {
  LAUNCH_ACTIVATION_KERNEL(vllm::gelu_quick_kernel);
 }
+
+void relu_squared(torch::stable::Tensor& out,    // [..., d]
+                  torch::stable::Tensor& input)  // [..., d]
+{
+  LAUNCH_ACTIVATION_KERNEL(vllm::relu_squared_kernel);
+}
@@ -21,6 +21,21 @@
 #define VLLM_STABLE_DISPATCH_FP8_CASE(enum_type, ...) \
  THO_PRIVATE_CASE_TYPE_USING_HINT(enum_type, fp8_t, __VA_ARGS__)

+// Same idea, for dispatching on an int32/int64 index tensor (e.g. topk_ids)
+// nested inside a value-type dispatch. Named 'idx_t' instead of 'scalar_t'.
+#define VLLM_STABLE_DISPATCH_IDX_CASE(enum_type, ...) \
+  THO_PRIVATE_CASE_TYPE_USING_HINT(enum_type, idx_t, __VA_ARGS__)
+
+#define VLLM_STABLE_DISPATCH_CASE_IDX_TYPES(...)                     \
+  VLLM_STABLE_DISPATCH_IDX_CASE(torch::headeronly::ScalarType::Int,  \
+                                __VA_ARGS__)                         \
+  VLLM_STABLE_DISPATCH_IDX_CASE(torch::headeronly::ScalarType::Long, \
+                                __VA_ARGS__)
+
+#define VLLM_STABLE_DISPATCH_IDX_TYPES(TYPE, NAME, ...) \
+  THO_DISPATCH_SWITCH(TYPE, NAME,                       \
+                      VLLM_STABLE_DISPATCH_CASE_IDX_TYPES(__VA_ARGS__))
+
 #define VLLM_STABLE_DISPATCH_CASE_FLOATING_TYPES(...)                  \
  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Float, __VA_ARGS__) \
  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Half, __VA_ARGS__)  \
@@ -1,13 +1,18 @@
 // SPDX-License-Identifier: Apache-2.0
 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 //
-// Router GEMM: activation(T) x weight(fp32) -> fp32, H=3072, E=256, M<=32.
+// Router GEMM: activation(T) x weight(fp32) -> fp32, M<=32, for the
+// supported (E, H) pairs listed at the bottom of this file.
 // Supports bf16 or fp32 activation; weight is always fp32.
 // Adapted from dsv3_router_gemm_float_out.cu.
+// (E=256, H=6144) bf16 uses a B300-tuned wide-block geometry; see
+// invokeFp32RouterGemm.

 #include <cuda_bf16.h>
 #include <cuda_runtime.h>

+#include <type_traits>
+
 // ---------------------------------------------------------------------------
 // Load helpers
 // ---------------------------------------------------------------------------
@@ -73,94 +78,113 @@ __device__ __forceinline__ void load_activation<__nv_bfloat16, 8>(
 // InputT : type of activation (float or __nv_bfloat16)
 // Weight is always fp32; output is always fp32.
 // VPT = 16 / sizeof(InputT):  4 for fp32, 8 for bf16
-template <typename InputT, int kBlockSize, int kNumTokens, int kNumExperts,
-          int kHiddenDim>
-__global__ __launch_bounds__(128, 1) void fp32_router_gemm_kernel(
-    float* out, InputT const* mat_a, float const* mat_b) {
+// Each block computes kEPB expert columns; wider blocks / kEPB > 1 are
+// selected per (shape, M) in invokeFp32RouterGemm (B300-tuned, see below).
+// kTGroups > 1 splits the tokens across groups of kBlockSize threads within
+// the block: all groups scan the same weight K-slices (group 0 misses to
+// DRAM, later groups hit L1) so weight traffic stays 1x, while per-thread
+// accumulator registers drop by kTGroups (at M=16 the 32 fp32 accumulators
+// push the kernel to 128 regs/thread and 1 block/SM).
+template <typename InputT, int kBlockSize, int kNumTokens, int kEPB,
+          int kNumExperts, int kHiddenDim, int kTGroups = 1>
+__global__ __launch_bounds__(
+    kBlockSize* kTGroups, 1) void fp32_router_gemm_kernel(float* out,
+                                                          InputT const* mat_a,
+                                                          float const* mat_b) {
  constexpr int VPT = 16 / sizeof(InputT);
  constexpr int k_elems_per_k_iteration = VPT * kBlockSize;
  constexpr int k_iterations = kHiddenDim / k_elems_per_k_iteration;
+  static_assert(kHiddenDim % k_elems_per_k_iteration == 0);
+  static_assert(kNumTokens % kTGroups == 0);
  constexpr int kWarpSize = 32;
-  constexpr int kNumWarps = kBlockSize / kWarpSize;
+  constexpr int kNumWarps = kBlockSize / kWarpSize;  // per token group
+  constexpr int kMG = kNumTokens / kTGroups;         // tokens per group

-  int const n_idx = blockIdx.x;
-  int const tid = threadIdx.x;
+  int const e_base = blockIdx.x * kEPB;
+  int const tid = threadIdx.x % kBlockSize;
+  int const m0 = (threadIdx.x / kBlockSize) * kMG;
  int const warpId = tid / kWarpSize;
  int const laneId = tid % kWarpSize;

-  float acc[kNumTokens] = {};
-  __shared__ float sm_reduction[kNumTokens][kNumWarps];
-
-  float const* b_col = mat_b + n_idx * kHiddenDim;
-
-  int k_bases[k_iterations];
-#pragma unroll
-  for (int ki = 0; ki < k_iterations; ki++) {
-    k_bases[ki] = ki * k_elems_per_k_iteration + tid * VPT;
-  }
+  float acc[kMG][kEPB] = {};
+  __shared__ float sm_reduction[kNumTokens][kEPB][kNumWarps];

 #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
  cudaGridDependencySynchronize();
+  // Fire the PDL trigger right after our own wait instead of at kernel end:
+  // a gridsync-ing consumer is unaffected (its wait always targets full grid
+  // completion), while a consumer that reads none of our outputs (e.g. the
+  // NVFP4 activation quant, which reads the same hidden_states) can launch
+  // now and fully overlap this kernel's body.
+  cudaTriggerProgrammaticLaunchCompletion();
 #endif

+#pragma unroll
  for (int ki = 0; ki < k_iterations; ki++) {
-    int const k_base = k_bases[ki];
+    int const k_base = ki * k_elems_per_k_iteration + tid * VPT;

-    float b_float[VPT];
-    load_weight<VPT>(b_col + k_base, b_float);
+    float b_float[kEPB][VPT];
+#pragma unroll
+    for (int e = 0; e < kEPB; e++) {
+      load_weight<VPT>(mat_b + (e_base + e) * kHiddenDim + k_base, b_float[e]);
+    }

 #pragma unroll
-    for (int m_idx = 0; m_idx < kNumTokens; m_idx++) {
+    for (int m_idx = 0; m_idx < kMG; m_idx++) {
      float a_float[VPT];
-      load_activation<InputT, VPT>(mat_a + m_idx * kHiddenDim + k_base,
-                                   a_float);
+      load_activation<InputT, VPT>(
+          mat_a + (size_t)(m0 + m_idx) * kHiddenDim + k_base, a_float);
 #pragma unroll
-      for (int k = 0; k < VPT; k++) {
-        acc[m_idx] += a_float[k] * b_float[k];
+      for (int e = 0; e < kEPB; e++) {
+#pragma unroll
+        for (int k = 0; k < VPT; k++) {
+          acc[m_idx][e] += a_float[k] * b_float[e][k];
+        }
      }
    }
  }

  // Warp-level butterfly reduction
 #pragma unroll
-  for (int m = 0; m < kNumTokens; m++) {
-    float sum = acc[m];
-    sum += __shfl_xor_sync(0xffffffff, sum, 16);
-    sum += __shfl_xor_sync(0xffffffff, sum, 8);
-    sum += __shfl_xor_sync(0xffffffff, sum, 4);
-    sum += __shfl_xor_sync(0xffffffff, sum, 2);
-    sum += __shfl_xor_sync(0xffffffff, sum, 1);
-    if (laneId == 0) sm_reduction[m][warpId] = sum;
+  for (int m = 0; m < kMG; m++) {
+#pragma unroll
+    for (int e = 0; e < kEPB; e++) {
+      float sum = acc[m][e];
+      sum += __shfl_xor_sync(0xffffffff, sum, 16);
+      sum += __shfl_xor_sync(0xffffffff, sum, 8);
+      sum += __shfl_xor_sync(0xffffffff, sum, 4);
+      sum += __shfl_xor_sync(0xffffffff, sum, 2);
+      sum += __shfl_xor_sync(0xffffffff, sum, 1);
+      if (laneId == 0) sm_reduction[m0 + m][e][warpId] = sum;
+    }
  }

  __syncthreads();

-  if (tid == 0) {
+  // Parallel finalize: one thread per (m, e) output.
+  for (int idx = threadIdx.x; idx < kNumTokens * kEPB;
+       idx += kBlockSize * kTGroups) {
+    int const m = idx / kEPB;
+    int const e = idx % kEPB;
+    float final_sum = 0.0f;
 #pragma unroll
-    for (int m = 0; m < kNumTokens; m++) {
-      float final_sum = 0.0f;
-#pragma unroll
-      for (int w = 0; w < kNumWarps; w++) final_sum += sm_reduction[m][w];
-      out[m * kNumExperts + n_idx] = final_sum;
-    }
+    for (int w = 0; w < kNumWarps; w++) final_sum += sm_reduction[m][e][w];
+    out[m * kNumExperts + e_base + e] = final_sum;
  }
-
-#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
-  cudaTriggerProgrammaticLaunchCompletion();
-#endif
 }

 // ---------------------------------------------------------------------------
 // Launcher
 // ---------------------------------------------------------------------------

-template <typename InputT, int kNumTokens, int kNumExperts, int kHiddenDim>
-void invokeFp32RouterGemm(float* output, InputT const* mat_a,
-                          float const* mat_b, cudaStream_t stream) {
-  constexpr int kBlockSize = 128;
+template <typename InputT, int kBlockSize, int kEPB, int kNumTokens,
+          int kNumExperts, int kHiddenDim, int kTGroups = 1>
+static void launchFp32RouterGemm(float* output, InputT const* mat_a,
+                                 float const* mat_b, cudaStream_t stream) {
+  static_assert(kNumExperts % kEPB == 0);
  cudaLaunchConfig_t config;
-  config.gridDim = kNumExperts;
-  config.blockDim = kBlockSize;
+  config.gridDim = kNumExperts / kEPB;
+  config.blockDim = kBlockSize * kTGroups;
  config.dynamicSmemBytes = 0;
  config.stream = stream;
  cudaLaunchAttribute attrs[1];
@@ -168,15 +192,112 @@ void invokeFp32RouterGemm(float* output, InputT const* mat_a,
  attrs[0].val.programmaticStreamSerializationAllowed = 1;
  config.numAttrs = 1;
  config.attrs = attrs;
-  cudaLaunchKernelEx(&config,
-                     fp32_router_gemm_kernel<InputT, kBlockSize, kNumTokens,
-                                             kNumExperts, kHiddenDim>,
-                     output, mat_a, mat_b);
+  cudaLaunchKernelEx(
+      &config,
+      fp32_router_gemm_kernel<InputT, kBlockSize, kNumTokens, kEPB, kNumExperts,
+                              kHiddenDim, kTGroups>,
+      output, mat_a, mat_b);
+}
+
+static bool isBlackwellFamily() {
+  static int sm = []() {
+    int dev = 0, major = 0, minor = 0;
+    cudaGetDevice(&dev);
+    cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, dev);
+    cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, dev);
+    return major * 10 + minor;
+  }();
+  return sm >= 100;
+}
+
+template <typename InputT, int kNumTokens, int kNumExperts, int kHiddenDim>
+void invokeFp32RouterGemm(float* output, InputT const* mat_a,
+                          float const* mat_b, cudaStream_t stream) {
+  // Geometry tuned on B300 per supported shape, bf16 activation, under a
+  // production-fidelity harness (CUDA-graph replay, per-layer cold weights).
+  // GLM-5.2 (E=256, H=6144):
+  //   M <= 4        : BS=768, EPB=1 (2.7us vs cast+cuBLAS 8.1us at M=1)
+  //   M in [5, 15]
+  //   or odd        : BS=384, EPB=2 (crossover vs BS=768 measured in (4, 8))
+  //   M >= 16, even : BS=192, EPB=2, 2 token groups (M=16 4.79us vs 5.04,
+  //                   M=24 5.71 vs 6.38, M=32 6.79 vs 7.72; M=12 loses at
+  //                   0.97x, so the boundary is 16).
+  // Only enabled on the Blackwell family where it was validated; Hopper and
+  // other shapes / fp32 activation keep the legacy geometry.
+  if constexpr (std::is_same_v<InputT, __nv_bfloat16> && kNumExperts == 256 &&
+                kHiddenDim == 6144) {
+    if (!isBlackwellFamily()) {
+      launchFp32RouterGemm<InputT, 128, 1, kNumTokens, kNumExperts, kHiddenDim>(
+          output, mat_a, mat_b, stream);
+      return;
+    }
+    if constexpr (kNumTokens <= 4) {
+      launchFp32RouterGemm<InputT, 768, 1, kNumTokens, kNumExperts, kHiddenDim>(
+          output, mat_a, mat_b, stream);
+    } else if constexpr (kNumTokens >= 16 && kNumTokens % 2 == 0) {
+      launchFp32RouterGemm<InputT, 192, 2, kNumTokens, kNumExperts, kHiddenDim,
+                           2>(output, mat_a, mat_b, stream);
+    } else {
+      launchFp32RouterGemm<InputT, 384, 2, kNumTokens, kNumExperts, kHiddenDim>(
+          output, mat_a, mat_b, stream);
+    }
+  } else if constexpr (std::is_same_v<InputT, __nv_bfloat16> &&
+                       kNumExperts == 128 && kHiddenDim == 6144) {
+    // MiniMax-M3. Legacy 128/1 only fills 128 blocks and pays the same
+    // accumulator register cliffs; B300 sweep:
+    //   even M in [6, 10] : BS=384, EPB=1, 2 token groups (1.26-1.43x)
+    //   even M >= 12      : BS=192, EPB=1, 2 token groups (1.59-1.66x at
+    //                       M >= 18; re-measured on B300+B200: 192 also wins
+    //                       M=12/14 by 5-11%% on both, ties 384 at 16)
+    //   M <= 5 / odd      : BS=384, EPB=1 (1.03-1.19x)
+    if (!isBlackwellFamily()) {
+      launchFp32RouterGemm<InputT, 128, 1, kNumTokens, kNumExperts, kHiddenDim>(
+          output, mat_a, mat_b, stream);
+      return;
+    }
+    if constexpr (kNumTokens >= 12 && kNumTokens % 2 == 0) {
+      launchFp32RouterGemm<InputT, 192, 1, kNumTokens, kNumExperts, kHiddenDim,
+                           2>(output, mat_a, mat_b, stream);
+    } else if constexpr (kNumTokens >= 6 && kNumTokens % 2 == 0) {
+      launchFp32RouterGemm<InputT, 384, 1, kNumTokens, kNumExperts, kHiddenDim,
+                           2>(output, mat_a, mat_b, stream);
+    } else {
+      launchFp32RouterGemm<InputT, 384, 1, kNumTokens, kNumExperts, kHiddenDim>(
+          output, mat_a, mat_b, stream);
+    }
+  } else if constexpr (std::is_same_v<InputT, __nv_bfloat16> &&
+                       kNumExperts == 256 && kHiddenDim == 3072) {
+    // MiniMax-M2/M2.5. The 3.1MB weight is latency-floor bound at small M
+    // (legacy already optimal); token groups win only at even M >= 8
+    // (1.05-1.17x). EPB crossover measured between 12 and 16.
+    if (!isBlackwellFamily()) {
+      launchFp32RouterGemm<InputT, 128, 1, kNumTokens, kNumExperts, kHiddenDim>(
+          output, mat_a, mat_b, stream);
+      return;
+    }
+    if constexpr (kNumTokens >= 14 && kNumTokens % 2 == 0) {
+      // M=14 originally measured 0.91x and stayed on legacy; two fresh
+      // sweeps (B300 dev1 + B200) both put 192/2/tg2 ahead by 3.5-4%%.
+      launchFp32RouterGemm<InputT, 192, 2, kNumTokens, kNumExperts, kHiddenDim,
+                           2>(output, mat_a, mat_b, stream);
+    } else if constexpr (kNumTokens >= 8 && kNumTokens <= 12 &&
+                         kNumTokens % 2 == 0) {
+      launchFp32RouterGemm<InputT, 192, 1, kNumTokens, kNumExperts, kHiddenDim,
+                           2>(output, mat_a, mat_b, stream);
+    } else {
+      launchFp32RouterGemm<InputT, 128, 1, kNumTokens, kNumExperts, kHiddenDim>(
+          output, mat_a, mat_b, stream);
+    }
+  } else {
+    launchFp32RouterGemm<InputT, 128, 1, kNumTokens, kNumExperts, kHiddenDim>(
+        output, mat_a, mat_b, stream);
+  }
 }

 // ---------------------------------------------------------------------------
 // Explicit instantiations: M=1..32, for both input types, for the supported
-// (E, H) pairs:  (256, 3072) [MiniMax-M2/M2.5]  and  (128, 6144) [MiniMax-M3].
+// (E, H) pairs:  (256, 3072) [MiniMax-M2/M2.5],  (128, 6144) [MiniMax-M3]
+// and  (256, 6144) [GLM-5.2].
 // ---------------------------------------------------------------------------

 #define INSTANTIATE(T, M, E, H)                                    \
@@ -221,6 +342,8 @@ INSTANTIATE_ALL(float, 256, 3072)
 INSTANTIATE_ALL(__nv_bfloat16, 256, 3072)
 INSTANTIATE_ALL(float, 128, 6144)
 INSTANTIATE_ALL(__nv_bfloat16, 128, 6144)
+INSTANTIATE_ALL(float, 256, 6144)
+INSTANTIATE_ALL(__nv_bfloat16, 256, 6144)

 #undef INSTANTIATE_ALL
 #undef INSTANTIATE
@@ -25,10 +25,12 @@ inline int getSMVersion() {
 static constexpr int FP32_MAX_TOKENS = 32;

 // Supported (hidden_dim, num_experts) pairs (must match the instantiations in
-// fp32_router_gemm.cu): (3072, 256) for MiniMax-M2/M2.5, (6144, 128) for M3.
+// fp32_router_gemm.cu): (3072, 256) for MiniMax-M2/M2.5, (6144, 128) for M3,
+// (6144, 256) for GLM-5.2.
 static inline bool fp32_router_gemm_supported(int hidden_dim, int num_experts) {
  return (hidden_dim == 3072 && num_experts == 256) ||
-         (hidden_dim == 6144 && num_experts == 128);
+         (hidden_dim == 6144 && num_experts == 128) ||
+         (hidden_dim == 6144 && num_experts == 256);
 }

 // Forward declarations — 4 template params must match fp32_router_gemm.cu
@@ -77,6 +79,9 @@ void dispatchFp32RouterGemm(int num_experts, int hidden_dim, int num_tokens,
  } else if (num_experts == 128 && hidden_dim == 6144) {
    Fp32LoopUnroller<InputT, 128, 6144, 1, FP32_MAX_TOKENS>::unroll(
        num_tokens, output, mat_a, mat_b, stream);
+  } else if (num_experts == 256 && hidden_dim == 6144) {
+    Fp32LoopUnroller<InputT, 256, 6144, 1, FP32_MAX_TOKENS>::unroll(
+        num_tokens, output, mat_a, mat_b, stream);
  } else {
    throw std::invalid_argument(
        "fp32_router_gemm: unsupported (hidden_dim, num_experts) pair");
@@ -111,7 +116,7 @@ void fp32_router_gemm(
  STD_TORCH_CHECK(
      fp32_router_gemm_supported(hidden_dim, num_experts),
      "fp32_router_gemm: supported (hidden_dim, num_experts) pairs are "
-      "(3072, 256) and (6144, 128)");
+      "(3072, 256), (6144, 128) and (6144, 256)");
  STD_TORCH_CHECK(num_tokens <= FP32_MAX_TOKENS,
                  "fp32_router_gemm: num_tokens must be in [0, 32]");
  STD_TORCH_CHECK(
--- a/Show More
+++ b/Show More