[Bugfix] Disable TRTLLM attention when KV transfer is enabled (#33192)

Signed-off-by: Zhanqiu Hu <zh338@cornell.edu>

.buildkite/ci_config.yaml

@@ -1,7 +1,8 @@
 name: vllm_ci
 job_dirs:
-  - ".buildkite/test_areas"
   - ".buildkite/image_build"
+  - ".buildkite/test_areas"
+  - ".buildkite/hardware_tests"
 run_all_patterns:
   - "docker/Dockerfile"
   - "CMakeLists.txt"

.buildkite/hardware_tests/amd.yaml

@@ -0,0 +1,29 @@
+group: Hardware
+steps:
+  - label: "AMD: :docker: build image"
+    depends_on: []
+    device: amd_cpu
+    no_plugin: true
+    commands:
+    - >
+      docker build
+      --build-arg max_jobs=16
+      --build-arg REMOTE_VLLM=1
+      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942'
+      --build-arg VLLM_BRANCH=$BUILDKITE_COMMIT
+      --tag "rocm/vllm-ci:${BUILDKITE_COMMIT}"
+      -f docker/Dockerfile.rocm
+      --target test
+      --no-cache
+      --progress plain .
+    - docker push "rocm/vllm-ci:${BUILDKITE_COMMIT}"
+    env:
+      DOCKER_BUILDKIT: "1"
+    retry:
+      automatic:
+        - exit_status: -1  # Agent was lost
+          limit: 1
+        - exit_status: -10  # Agent was lost
+          limit: 1
+        - exit_status: 1  # Machine occasionally fail
+          limit: 1

.buildkite/hardware_tests/arm.yaml

@@ -0,0 +1,8 @@
+group: Hardware
+steps:
+  - label: "Arm CPU Test"
+    soft_fail: true
+    device: arm_cpu
+    no_plugin: true
+    commands: 
+    - bash .buildkite/scripts/hardware_ci/run-cpu-test-arm.sh

.buildkite/hardware_tests/ascend_npu.yaml

@@ -0,0 +1,10 @@
+group: Hardware
+depends_on: ~
+steps:
+  - label: "Ascend NPU Test"
+    soft_fail: true
+    timeout_in_minutes: 20
+    no_plugin: true
+    device: ascend_npu
+    commands: 
+    - bash .buildkite/scripts/hardware_ci/run-npu-test.sh

.buildkite/hardware_tests/gh200.yaml

@@ -0,0 +1,10 @@
+group: Hardware
+steps:
+  - label: "GH200 Test"
+    soft_fail: true
+    device: gh200
+    no_plugin: true
+    optional: true
+    commands: 
+    - nvidia-smi 
+    - bash .buildkite/scripts/hardware_ci/run-gh200-test.sh

.buildkite/hardware_tests/intel.yaml

@@ -0,0 +1,24 @@
+group: Hardware
+depends_on: ~
+steps:
+  - label: "Intel CPU Test"
+    soft_fail: true
+    device: intel_cpu
+    no_plugin: true
+    commands: 
+    - bash .buildkite/scripts/hardware_ci/run-cpu-test.sh
+  
+  - label: "Intel HPU Test"
+    soft_fail: true
+    device: intel_hpu
+    no_plugin: true
+    commands: 
+    - bash .buildkite/scripts/hardware_ci/run-hpu-test.sh
+
+  - label: "Intel GPU Test"
+    depends_on: []
+    soft_fail: true
+    device: intel_gpu
+    no_plugin: true
+    commands: 
+    - bash .buildkite/scripts/hardware_ci/run-xpu-test.sh

.buildkite/image_build/image_build.sh

@@ -1,56 +1,256 @@
 #!/bin/bash
-set -e
+set -euo pipefail
 
-if [[ $# -lt 8 ]]; then
-  echo "Usage: $0 <registry> <repo> <commit> <branch> <vllm_use_precompiled> <vllm_merge_base_commit> <cache_from> <cache_to>"
-  exit 1
+# replace invalid characters in Docker image tags and truncate to 128 chars
+clean_docker_tag() {
+    local input="$1"
+    echo "$input" | sed 's/[^a-zA-Z0-9._-]/_/g' | cut -c1-128
+}
+
+print_usage_and_exit() {
+    echo "Usage: $0 <registry> <repo> <commit> <branch> <vllm_use_precompiled> <vllm_merge_base_commit> <cache_from> <cache_to>"
+    exit 1
+}
+
+print_instance_info() {
+    echo ""
+    echo "=== Debug: Instance Information ==="
+    # Get IMDSv2 token
+    if TOKEN=$(curl -s -X PUT "http://169.254.169.254/latest/api/token" \
+            -H "X-aws-ec2-metadata-token-ttl-seconds: 21600" 2>/dev/null); then
+        AMI_ID=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
+            http://169.254.169.254/latest/meta-data/ami-id 2>/dev/null || echo "unknown")
+        INSTANCE_TYPE=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
+            http://169.254.169.254/latest/meta-data/instance-type 2>/dev/null || echo "unknown")
+        INSTANCE_ID=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
+            http://169.254.169.254/latest/meta-data/instance-id 2>/dev/null || echo "unknown")
+        AZ=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
+            http://169.254.169.254/latest/meta-data/placement/availability-zone 2>/dev/null || echo "unknown")
+        echo "AMI ID:        ${AMI_ID}"
+        echo "Instance Type: ${INSTANCE_TYPE}"
+        echo "Instance ID:   ${INSTANCE_ID}"
+        echo "AZ:            ${AZ}"
+    else
+        echo "Not running on EC2 or IMDS not available"
+    fi
+    # Check for warm cache AMI (marker file baked into custom AMI)
+    if [[ -f /etc/vllm-ami-info ]]; then
+        echo "Cache:         warm (custom vLLM AMI)"
+        cat /etc/vllm-ami-info
+    else
+        echo "Cache:         cold (standard AMI)"
+    fi
+    echo "==================================="
+    echo ""
+}
+
+setup_buildx_builder() {
+    echo "--- :buildkite: Setting up buildx builder"
+    if [[ -S "${BUILDKIT_SOCKET}" ]]; then
+        # Custom AMI with standalone buildkitd - use remote driver for warm cache
+        echo "✅ Found local buildkitd socket at ${BUILDKIT_SOCKET}"
+        echo "Using remote driver to connect to buildkitd (warm cache available)"
+        if docker buildx inspect baked-vllm-builder >/dev/null 2>&1; then
+            echo "Using existing baked-vllm-builder"
+            docker buildx use baked-vllm-builder
+        else
+            echo "Creating baked-vllm-builder with remote driver"
+            docker buildx create \
+                --name baked-vllm-builder \
+                --driver remote \
+                --use \
+                "unix://${BUILDKIT_SOCKET}"
+        fi
+        docker buildx inspect --bootstrap
+    elif docker buildx inspect "${BUILDER_NAME}" >/dev/null 2>&1; then
+        # Existing builder available
+        echo "Using existing builder: ${BUILDER_NAME}"
+        docker buildx use "${BUILDER_NAME}"
+        docker buildx inspect --bootstrap
+    else
+        # No local buildkitd, no existing builder - create new docker-container builder
+        echo "No local buildkitd found, using docker-container driver"
+        docker buildx create --name "${BUILDER_NAME}" --driver docker-container --use
+        docker buildx inspect --bootstrap
+    fi
+
+    # builder info
+    echo "Active builder:"
+    docker buildx ls | grep -E '^\*|^NAME' || docker buildx ls
+}
+
+check_and_skip_if_image_exists() {
+    if [[ -n "${IMAGE_TAG:-}" ]]; then
+        echo "--- :mag: Checking if image exists"
+        if docker manifest inspect "${IMAGE_TAG}" >/dev/null 2>&1; then
+            echo "Image already exists: ${IMAGE_TAG}"
+            echo "Skipping build"
+            exit 0
+        fi
+        echo "Image not found, proceeding with build"
+    fi
+}
+
+ecr_login() {
+    aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
+    aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin 936637512419.dkr.ecr.us-east-1.amazonaws.com
+}
+
+prepare_cache_tags() {
+    # resolve and set: CACHE_TO, CACHE_FROM, CACHE_FROM_BASE_BRANCH, CACHE_FROM_MAIN
+    TEST_CACHE_ECR="936637512419.dkr.ecr.us-east-1.amazonaws.com/vllm-ci-test-cache"
+    MAIN_CACHE_ECR="936637512419.dkr.ecr.us-east-1.amazonaws.com/vllm-ci-postmerge-cache"
+
+    if [[ "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
+        if [[ "$BUILDKITE_BRANCH" == "main" ]]; then
+            cache="${MAIN_CACHE_ECR}:latest"
+        else
+            clean_branch=$(clean_docker_tag "$BUILDKITE_BRANCH")
+            cache="${TEST_CACHE_ECR}:${clean_branch}"
+        fi
+        CACHE_TO="$cache"
+        CACHE_FROM="$cache"
+        CACHE_FROM_BASE_BRANCH="$cache"
+    else
+        CACHE_TO="${TEST_CACHE_ECR}:pr-${BUILDKITE_PULL_REQUEST}"
+        CACHE_FROM="${TEST_CACHE_ECR}:pr-${BUILDKITE_PULL_REQUEST}"
+        if [[ "$BUILDKITE_PULL_REQUEST_BASE_BRANCH" == "main" ]]; then
+            CACHE_FROM_BASE_BRANCH="${MAIN_CACHE_ECR}:latest"
+        else
+            clean_base=$(clean_docker_tag "$BUILDKITE_PULL_REQUEST_BASE_BRANCH")
+            CACHE_FROM_BASE_BRANCH="${TEST_CACHE_ECR}:${clean_base}"
+        fi
+    fi
+
+    CACHE_FROM_MAIN="${MAIN_CACHE_ECR}:latest"
+    export CACHE_TO CACHE_FROM CACHE_FROM_BASE_BRANCH CACHE_FROM_MAIN
+}
+
+resolve_parent_commit() {
+    if [[ -z "${PARENT_COMMIT:-}" ]]; then
+        PARENT_COMMIT=$(git rev-parse HEAD~1 2>/dev/null || echo "")
+        if [[ -n "${PARENT_COMMIT}" ]]; then
+            echo "Computed parent commit for cache fallback: ${PARENT_COMMIT}"
+            export PARENT_COMMIT
+        else
+            echo "Could not determine parent commit (may be first commit in repo)"
+        fi
+    else
+        echo "Using provided PARENT_COMMIT: ${PARENT_COMMIT}"
+    fi
+}
+
+print_bake_config() {
+    echo "--- :page_facing_up: Resolved bake configuration"
+    BAKE_CONFIG_FILE="bake-config-build-${BUILDKITE_BUILD_NUMBER:-local}.json"
+    docker buildx bake -f "${VLLM_BAKE_FILE_PATH}" -f "${CI_HCL_PATH}" --print "${TARGET}" | tee "${BAKE_CONFIG_FILE}" || true
+    echo "Saved bake config to ${BAKE_CONFIG_FILE}"
+    echo "--- :arrow_down: Uploading bake config to Buildkite"
+    buildkite-agent artifact upload "${BAKE_CONFIG_FILE}"
+}
+
+#################################
+#         Main Script           #
+#################################
+print_instance_info
+
+if [[ $# -lt 7 ]]; then
+    print_usage_and_exit
 fi
 
+# input args
 REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 BRANCH=$4
 VLLM_USE_PRECOMPILED=$5
 VLLM_MERGE_BASE_COMMIT=$6
-CACHE_FROM=$7
-CACHE_TO=$8
+IMAGE_TAG=$7
+IMAGE_TAG_LATEST=${8:-} # only used for main branch, optional
 
-# authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
-aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin 936637512419.dkr.ecr.us-east-1.amazonaws.com
+# build config
+TARGET="test-ci"
+VLLM_BAKE_FILE_PATH="${VLLM_BAKE_FILE_PATH:-docker/docker-bake.hcl}"
+BUILDER_NAME="${BUILDER_NAME:-vllm-builder}"
+CI_HCL_URL="${CI_HCL_URL:-https://raw.githubusercontent.com/vllm-project/ci-infra/main/docker/ci.hcl}"
+CI_HCL_PATH="/tmp/ci.hcl"
+BUILDKIT_SOCKET="/run/buildkit/buildkitd.sock"
 
-# docker buildx 
-docker buildx create --name vllm-builder --driver docker-container --use
-docker buildx inspect --bootstrap
-docker buildx ls
+prepare_cache_tags
+ecr_login
 
-# skip build if image already exists
-if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT) ]]; then
-  echo "Image not found, proceeding with build..."
-else
-  echo "Image found"
-  exit 0
+# Environment info (for docs and human readers)
+#   VLLM_CI_BRANCH      - ci-infra branch to use (default: main)
+#   VLLM_BAKE_FILE_PATH      - Path to vLLM's bake file (default: docker/docker-bake.hcl)
+#   BUILDER_NAME        - Name for buildx builder (default: vllm-builder)
+#
+# Build configuration (exported as environment variables for bake):
+export BUILDKITE_COMMIT
+export PARENT_COMMIT
+export IMAGE_TAG
+export IMAGE_TAG_LATEST
+export CACHE_FROM
+export CACHE_FROM_BASE_BRANCH
+export CACHE_FROM_MAIN
+export CACHE_TO
+export VLLM_USE_PRECOMPILED
+export VLLM_MERGE_BASE_COMMIT
+
+# print args
+echo "--- :mag: Arguments"
+echo "REGISTRY: ${REGISTRY}"
+echo "REPO: ${REPO}"
+echo "BUILDKITE_COMMIT: ${BUILDKITE_COMMIT}"
+echo "BRANCH: ${BRANCH}"
+echo "VLLM_USE_PRECOMPILED: ${VLLM_USE_PRECOMPILED}"
+echo "VLLM_MERGE_BASE_COMMIT: ${VLLM_MERGE_BASE_COMMIT}"
+echo "IMAGE_TAG: ${IMAGE_TAG}"
+echo "IMAGE_TAG_LATEST: ${IMAGE_TAG_LATEST}"
+
+# print build configuration
+echo "--- :mag: Build configuration"
+echo "TARGET: ${TARGET}"
+echo "vLLM bake file: ${VLLM_BAKE_FILE_PATH}"
+echo "BUILDER_NAME: ${BUILDER_NAME}"
+echo "CI_HCL_URL: ${CI_HCL_URL}"
+echo "BUILDKIT_SOCKET: ${BUILDKIT_SOCKET}"
+
+echo "--- :mag: Cache tags"
+echo "CACHE_TO: ${CACHE_TO}"
+echo "CACHE_FROM: ${CACHE_FROM}"
+echo "CACHE_FROM_BASE_BRANCH: ${CACHE_FROM_BASE_BRANCH}"
+echo "CACHE_FROM_MAIN: ${CACHE_FROM_MAIN}"
+
+check_and_skip_if_image_exists
+
+echo "--- :docker: Setting up Docker buildx bake"
+echo "Target: ${TARGET}"
+echo "vLLM bake file: ${VLLM_BAKE_FILE_PATH}"
+echo "CI HCL path: ${CI_HCL_PATH}"
+
+if [[ ! -f "${VLLM_BAKE_FILE_PATH}" ]]; then
+    echo "Error: vLLM bake file not found at ${VLLM_BAKE_FILE_PATH}"
+    echo "Make sure you're running from the vLLM repository root"
+    exit 1
 fi
 
-if [[ "${VLLM_USE_PRECOMPILED:-0}" == "1" ]]; then
-  merge_base_commit_build_args="--build-arg VLLM_MERGE_BASE_COMMIT=${VLLM_MERGE_BASE_COMMIT}"
-else
-  merge_base_commit_build_args=""
+echo "--- :arrow_down: Downloading ci.hcl"
+curl -sSfL -o "${CI_HCL_PATH}" "${CI_HCL_URL}"
+echo "Downloaded to ${CI_HCL_PATH}"
+
+if [[ ! -f "${CI_HCL_PATH}" ]]; then
+    echo "Error: ci.hcl not found at ${CI_HCL_PATH}"
+    exit 1
 fi
 
-# build
-docker buildx build --file docker/Dockerfile \
-  --build-arg max_jobs=16 \
-  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
-  --build-arg USE_SCCACHE=1 \
-  --build-arg TORCH_CUDA_ARCH_LIST="8.0 8.9 9.0 10.0" \
-  --build-arg FI_TORCH_CUDA_ARCH_LIST="8.0 8.9 9.0a 10.0a" \
-  --build-arg VLLM_USE_PRECOMPILED="${VLLM_USE_PRECOMPILED:-0}" \
-  ${merge_base_commit_build_args} \
-  --cache-from type=registry,ref=${CACHE_FROM},mode=max \
-  --cache-to type=registry,ref=${CACHE_TO},mode=max \
-  --tag ${REGISTRY}/${REPO}:${BUILDKITE_COMMIT} \
-  $( [[ "${BRANCH}" == "main" ]] && echo "--tag ${REGISTRY}/${REPO}:latest" ) \
-  --push \
-  --target test \
-  --progress plain .
+setup_buildx_builder
+
+resolve_parent_commit
+export PARENT_COMMIT
+
+print_bake_config
+
+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"

.buildkite/image_build/image_build.yaml

@@ -4,7 +4,8 @@ steps:
     key: image-build
     depends_on: []
     commands:
-    - .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $CACHE_FROM $CACHE_TO
+    - if [[ "$BUILDKITE_BRANCH" != "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG; fi
+    - if [[ "$BUILDKITE_BRANCH" == "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $VLLM_USE_PRECOMPILED $VLLM_MERGE_BASE_COMMIT $IMAGE_TAG $IMAGE_TAG_LATEST; fi
     retry:
       automatic:
         - exit_status: -1  # Agent was lost

.buildkite/lm-eval-harness/configs/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.yaml

@@ -0,0 +1,15 @@
+model_name: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.695
+  - name: "exact_match,flexible-extract"
+    value: 0.447
+limit: 1319
+num_fewshot: 5
+max_model_len: 262144
+enforce_eager: false
+apply_chat_template: true
+fewshot_as_multiturn: true
+trust_remote_code: true

.buildkite/lm-eval-harness/configs/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8.yaml

@@ -0,0 +1,19 @@
+model_name: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.7142
+  - name: "exact_match,flexible-extract"
+    value: 0.4579
+env_vars:
+  VLLM_USE_FLASHINFER_MOE_FP8: "1"
+  VLLM_FLASHINFER_MOE_BACKEND: "throughput"
+limit: 1319
+num_fewshot: 5
+max_model_len: 262144
+kv_cache_dtype: fp8
+enforce_eager: false
+apply_chat_template: true
+fewshot_as_multiturn: true
+trust_remote_code: true

.buildkite/lm-eval-harness/configs/models-large-hopper.txt

@@ -1 +1,2 @@
 Qwen3-235B-A22B-Instruct-2507-FP8.yaml
+NVIDIA-Nemotron-3-Nano-30B-A3B-FP8.yaml

.buildkite/lm-eval-harness/configs/models-large.txt

@@ -3,3 +3,4 @@ Meta-Llama-3-70B-Instruct.yaml
 Mixtral-8x7B-Instruct-v0.1.yaml
 Qwen2-57B-A14-Instruct.yaml
 DeepSeek-V2-Lite-Chat.yaml
+NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.yaml

.buildkite/lm-eval-harness/configs/models-small-rocm.txt

@@ -0,0 +1,5 @@
+Qwen2.5-1.5B-Instruct.yaml
+Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
+Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
+Qwen2.5-VL-3B-Instruct-FP8-dynamic.yaml
+Qwen1.5-MoE-W4A16-compressed-tensors.yaml

.buildkite/lm-eval-harness/run-lm-eval-chartqa-vllm-vlm-baseline.sh

@@ -2,7 +2,7 @@
 # We can use this script to compute baseline accuracy on chartqa for vllm.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install lm-eval==0.4.9
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``

.buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh

@@ -2,7 +2,7 @@
 # We can use this script to compute baseline accuracy on GSM for transformers.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d#egg=lm-eval[api]
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``

.buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh

@@ -3,7 +3,7 @@
 # We use this for fp8, which HF does not support.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d#egg=lm-eval[api]
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``

.buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh

@@ -3,7 +3,7 @@
 # We use this for fp8, which HF does not support.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d#egg=lm-eval[api]
+#   pip install "lm-eval[api]>=0.4.9.2"
 
 usage() {
     echo``

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -60,6 +60,7 @@ def launch_lm_eval(eval_config, tp_size):
         f"add_bos_token=true,"
         f"trust_remote_code={trust_remote_code},"
         f"max_model_len={max_model_len},"
+        "allow_deprecated_quantization=True,"
     )
 
     env_vars = eval_config.get("env_vars", None)

.buildkite/performance-benchmarks/README.md

@@ -7,7 +7,7 @@ vLLM also maintains a continuous performance benchmark under [perf.vllm.ai](http
 
 ## Performance benchmark quick overview
 
-**Benchmarking Coverage**: latency, throughput and fix-qps serving on B200, A100, H100, Intel® Xeon® Processors and Intel® Gaudi® 3 Accelerators with different models.
+**Benchmarking Coverage**: latency, throughput and fix-qps serving on B200, A100, H100, Intel® Xeon® Processors, Intel® Gaudi® 3 Accelerators and Arm® Neoverse™ with different models.
 
 **Benchmarking Duration**: about 1hr.
 
@@ -23,7 +23,7 @@ bash .buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh
 
 Runtime environment variables:
 
-- `ON_CPU`: set the value to '1' on Intel® Xeon® Processors. Default value is 0.
+- `ON_CPU`: set the value to '1' on Intel® Xeon® and Arm® Neoverse™ Processors. Default value is 0.
 - `SERVING_JSON`: JSON file to use for the serving tests. Default value is empty string (use default file).
 - `LATENCY_JSON`: JSON file to use for the latency tests. Default value is empty string (use default file).
 - `THROUGHPUT_JSON`: JSON file to use for the throughout tests. Default value is empty string (use default file).
@@ -34,8 +34,9 @@ Runtime environment variables:
 
 See [performance-benchmarks-descriptions.md](performance-benchmarks-descriptions.md) for detailed descriptions, and use `tests/latency-tests.json`, `tests/throughput-tests.json`, `tests/serving-tests.json` to configure the test cases.
 > NOTE: For Intel® Xeon® Processors, use `tests/latency-tests-cpu.json`, `tests/throughput-tests-cpu.json`, `tests/serving-tests-cpu.json` instead.
-For Intel® Gaudi® 3 Accelerators, use `tests/latency-tests-hpu.json`, `tests/throughput-tests-hpu.json`, `tests/serving-tests-hpu.json` instead.
->
+> For Intel® Gaudi® 3 Accelerators, use `tests/latency-tests-hpu.json`, `tests/throughput-tests-hpu.json`, `tests/serving-tests-hpu.json` instead.
+> For Arm® Neoverse™, use `tests/latency-tests-arm64-cpu.json`, `tests/throughput-tests-arm64-cpu.json`, `tests/serving-tests-arm64-cpu.json` instead.
+
 ### Latency test
 
 Here is an example of one test inside `latency-tests.json`:
@@ -175,19 +176,6 @@ If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.
 The raw benchmarking results (in the format of json files) are in the `Artifacts` tab of the benchmarking.
 
-The `compare-json-results.py` helps to compare benchmark results JSON files converted using `convert-results-json-to-markdown.py`.
-When run, benchmark script generates results under `benchmark/results` folder, along with the `benchmark_results.md` and `benchmark_results.json`.
-`compare-json-results.py` compares two `benchmark_results.json` files and provides performance ratio e.g. for Output Tput, Median TTFT and Median TPOT.  
-If only one benchmark_results.json is passed, `compare-json-results.py` compares different TP and PP configurations in the benchmark_results.json instead.
+#### Performance Results Comparison  
 
-Here is an example using the script to compare result_a and result_b with Model, Dataset name, input/output length, max concurrency and qps.
-`python3 compare-json-results.py -f results_a/benchmark_results.json -f results_b/benchmark_results.json`
-
-|   | Model | Dataset Name | Input Len | Output Len | # of max concurrency | qps  | results_a/benchmark_results.json | results_b/benchmark_results.json | perf_ratio        |
-|----|---------------------------------------|--------|-----|-----|------|-----|-----------|----------|----------|
-| 0  | meta-llama/Meta-Llama-3.1-8B-Instruct | random | 128 | 128 | 1000 | 1 | 142.633982                             | 156.526018                             | 1.097396 |
-| 1  | meta-llama/Meta-Llama-3.1-8B-Instruct | random | 128 | 128 | 1000 | inf| 241.620334                             | 294.018783                             | 1.216863 |
-
-A comparison diagram will be generated below the table.
-Here is an example to compare between 96c/results_gnr_96c_091_tp2pp3 and 128c/results_gnr_128c_091_tp2pp3
-<img width="1886" height="828" alt="image" src="https://github.com/user-attachments/assets/c02a43ef-25d0-4fd6-90e5-2169a28682dd" />
+Follow the instructions in [performance results comparison](https://docs.vllm.ai/en/latest/benchmarking/dashboard/#performance-results-comparison) to analyze performance results and the sizing guide.

.buildkite/performance-benchmarks/scripts/compare-json-results.py

@@ -1,8 +1,13 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from __future__ import annotations
+
 import argparse
+import html as _html
 import json
 import os
+from dataclasses import dataclass
 from importlib import util
 
 import pandas as pd
@@ -10,27 +15,49 @@ import pandas as pd
 pd.options.display.float_format = "{:.2f}".format
 plotly_found = util.find_spec("plotly.express") is not None
 
+DEFAULT_INFO_COLS = [
+    "Model",
+    "Dataset Name",
+    "Input Len",
+    "Output Len",
+    #    "TP Size",
+    #    "PP Size",
+    "# of max concurrency.",
+    "qps",
+]
 
+# Safety net: if any DataFrame leaks into to_html(), keep precision at 2.
+pd.set_option("display.precision", 2)
+pd.set_option("display.float_format", lambda x: f"{x:.2f}")
+
+
+# -----------------------------
+# Core data compare
+# -----------------------------
 def compare_data_columns(
-    files, name_column, data_column, info_cols, drop_column, debug=False
+    files: list[str],
+    name_column: str,
+    data_column: str,
+    info_cols: list[str],
+    drop_column: str,
+    debug: bool = False,
 ):
     """
     Align concatenation by keys derived from info_cols instead of row order.
     - Pick one canonical key list: subset of info_cols present in ALL files.
     - For each file: set index to those keys, aggregate duplicates
-    - (mean for metric, first for names).
+      (mean for metric, first for names).
     - Concat along axis=1 (indexes align), then reset_index so callers can
-    - group by columns.
+      group by columns.
     - If --debug, add a <file_label>_name column per file.
     """
     print("\ncompare_data_column:", data_column)
 
     frames = []
-    raw_data_cols = []
+    raw_data_cols: list[str] = []
     compare_frames = []
 
-    # 1) choose a canonical key list from info_cols that exists in ALL files
-    cols_per_file = []
+    cols_per_file: list[set] = []
     for f in files:
         try:
             df_tmp = pd.read_json(f, orient="records")
@@ -40,24 +67,20 @@ def compare_data_columns(
 
     key_cols = [c for c in info_cols if all(c in cset for cset in cols_per_file)]
     if not key_cols:
-        # soft fallback: use any info_cols present in the first file
         key_cols = [c for c in info_cols if c in list(cols_per_file[0])]
     if not key_cols:
         raise ValueError(
             "No common key columns found from info_cols across the input files."
         )
 
-    # 2) build a single "meta" block (keys as columns) once, aligned by the key index
     meta_added = False
 
     for file in files:
         df = pd.read_json(file, orient="records")
 
-        # Keep rows that actually have the compared metric (same as original behavior)
         if drop_column in df.columns:
             df = df.dropna(subset=[drop_column], ignore_index=True)
 
-        # Stabilize numeric key columns (harmless if missing)
         for c in (
             "Input Len",
             "Output Len",
@@ -69,32 +92,26 @@ def compare_data_columns(
             if c in df.columns:
                 df[c] = pd.to_numeric(df[c], errors="coerce")
 
-        # Ensure all key columns exist
         for c in key_cols:
             if c not in df.columns:
                 df[c] = pd.NA
 
-        # Set index = key_cols and aggregate duplicates → unique MultiIndex
         df_idx = df.set_index(key_cols, drop=False)
 
-        # meta (key columns), unique per key
         meta = df_idx[key_cols]
         if not meta.index.is_unique:
             meta = meta.groupby(level=key_cols, dropna=False).first()
 
-        # metric series for this file, aggregated to one row per key
         file_label = "/".join(file.split("/")[:-1]) or os.path.basename(file)
         s = df_idx[data_column]
         if not s.index.is_unique:
             s = s.groupby(level=key_cols, dropna=False).mean()
-        s.name = file_label  # column label like original
+        s.name = file_label
 
-        # add meta once (from first file) so keys are the leftmost columns
         if not meta_added:
             frames.append(meta)
             meta_added = True
 
-        # (NEW) debug: aligned test-name column per file
         if debug and name_column in df_idx.columns:
             name_s = df_idx[name_column]
             if not name_s.index.is_unique:
@@ -106,26 +123,19 @@ def compare_data_columns(
         raw_data_cols.append(file_label)
         compare_frames.append(s)
 
-        # Generalize ratio: for any file N>=2, add ratio (fileN / file1)
         if len(compare_frames) >= 2:
             base = compare_frames[0]
             current = compare_frames[-1]
             if "P99" in data_column or "Median" in data_column:
-                ratio = base / current  # for latency
+                ratio = base / current
             else:
                 ratio = current / base
-            ratio = ratio.mask(base == 0)  # avoid inf when baseline is 0
+            ratio = ratio.mask(base == 0)
             ratio.name = f"Ratio 1 vs {len(compare_frames)}"
             frames.append(ratio)
 
-    # 4) concat on columns with aligned MultiIndex;
-    # then reset_index to return keys as columns
-    concat_df = pd.concat(frames, axis=1)
-    concat_df = concat_df.reset_index(drop=True).reset_index()
-    if "index" in concat_df.columns:
-        concat_df = concat_df.drop(columns=["index"])
+    concat_df = pd.concat(frames, axis=1).reset_index(drop=True)
 
-    # Ensure key/info columns appear first (in your info_cols order)
     front = [c for c in info_cols if c in concat_df.columns]
     rest = [c for c in concat_df.columns if c not in front]
     concat_df = concat_df[front + rest]
@@ -134,20 +144,15 @@ def compare_data_columns(
     return concat_df, raw_data_cols
 
 
+# -----------------------------
+# Split helper
+# -----------------------------
 def split_json_by_tp_pp(
     input_file: str = "benchmark_results.json", output_root: str = "."
 ) -> list[str]:
-    """
-    Split a benchmark JSON into separate folders by (TP Size, PP Size).
-
-    Creates: <output_root>/tp{TP}_pp{PP}/benchmark_results.json
-    Returns: list of file paths written.
-    """
-    # Load JSON data into DataFrame
     with open(input_file, encoding="utf-8") as f:
         data = json.load(f)
 
-    # If the JSON is a dict with a list under common keys, use that list
     if isinstance(data, dict):
         for key in ("results", "serving_results", "benchmarks", "data"):
             if isinstance(data.get(key), list):
@@ -156,7 +161,6 @@ def split_json_by_tp_pp(
 
     df = pd.DataFrame(data)
 
-    # Keep only "serving" tests
     name_col = next(
         (c for c in ["Test name", "test_name", "Test Name"] if c in df.columns), None
     )
@@ -165,7 +169,6 @@ def split_json_by_tp_pp(
             df[name_col].astype(str).str.contains(r"serving", case=False, na=False)
         ].copy()
 
-    # Handle alias column names
     rename_map = {
         "tp_size": "TP Size",
         "tensor_parallel_size": "TP Size",
@@ -176,21 +179,14 @@ def split_json_by_tp_pp(
         columns={k: v for k, v in rename_map.items() if k in df.columns}, inplace=True
     )
 
-    # Ensure TP/PP columns exist (default to 1 if missing)
     if "TP Size" not in df.columns:
         df["TP Size"] = 1
     if "PP Size" not in df.columns:
         df["PP Size"] = 1
 
-    # make sure TP/PP are numeric ints with no NaN
-    df["TP Size"] = (
-        pd.to_numeric(df.get("TP Size", 1), errors="coerce").fillna(1).astype(int)
-    )
-    df["PP Size"] = (
-        pd.to_numeric(df.get("PP Size", 1), errors="coerce").fillna(1).astype(int)
-    )
+    df["TP Size"] = pd.to_numeric(df["TP Size"], errors="coerce").fillna(1).astype(int)
+    df["PP Size"] = pd.to_numeric(df["PP Size"], errors="coerce").fillna(1).astype(int)
 
-    # Split into separate folders
     saved_paths: list[str] = []
     for (tp, pp), group_df in df.groupby(["TP Size", "PP Size"], dropna=False):
         folder_name = os.path.join(output_root, f"tp{int(tp)}_pp{int(pp)}")
@@ -203,32 +199,9 @@ def split_json_by_tp_pp(
     return saved_paths
 
 
-def _add_limit_line(fig, y_value, label):
-    # Visible dashed line + annotation
-    fig.add_hline(
-        y=y_value,
-        line_dash="dash",
-        line_color="red" if "ttft" in label.lower() else "blue",
-        annotation_text=f"{label}: {y_value} ms",
-        annotation_position="top left",
-    )
-    # Optional: add a legend item (as a transparent helper trace)
-    if plot and plotly_found:
-        import plotly.graph_objects as go
-
-        fig.add_trace(
-            go.Scatter(
-                x=[None],
-                y=[None],
-                mode="lines",
-                line=dict(
-                    dash="dash", color="red" if "ttft" in label.lower() else "blue"
-                ),
-                name=f"{label}",
-            )
-        )
-
-
+# -----------------------------
+# Styling helpers
+# -----------------------------
 def _find_concurrency_col(df: pd.DataFrame) -> str:
     for c in [
         "# of max concurrency.",
@@ -239,7 +212,6 @@ def _find_concurrency_col(df: pd.DataFrame) -> str:
     ]:
         if c in df.columns:
             return c
-    # Fallback: guess an integer-like column (harmless if unused)
     for c in df.columns:
         if df[c].dtype.kind in "iu" and df[c].nunique() > 1 and df[c].min() >= 1:
             return c
@@ -248,8 +220,7 @@ def _find_concurrency_col(df: pd.DataFrame) -> str:
 
 def _highlight_threshold(
     df: pd.DataFrame, threshold: float
-) -> "pd.io.formats.style.Styler":
-    """Highlight numeric per-configuration columns with value <= threshold."""
+) -> pd.io.formats.style.Styler:
     conc_col = _find_concurrency_col(df)
     key_cols = [
         c
@@ -260,6 +231,7 @@ def _highlight_threshold(
         c for c in df.columns if c not in key_cols and not str(c).startswith("Ratio")
     ]
     conf_cols = [c for c in conf_cols if pd.api.types.is_numeric_dtype(df[c])]
+
     return df.style.map(
         lambda v: "background-color:#e6ffe6;font-weight:bold;"
         if pd.notna(v) and v <= threshold
@@ -268,7 +240,264 @@ def _highlight_threshold(
     )
 
 
-if __name__ == "__main__":
+def highlight_ratio_columns(styler: pd.io.formats.style.Styler):
+    ratio_cols = [c for c in styler.data.columns if "ratio" in str(c).lower()]
+    if not ratio_cols:
+        return styler
+
+    styler = styler.apply(
+        lambda _: ["background-color: #fff3b0"] * len(styler.data),
+        subset=ratio_cols,
+        axis=0,
+    )
+
+    styler = styler.set_table_styles(
+        [
+            {
+                "selector": f"th.col_heading.level0.col{i}",
+                "props": [("background-color", "#fff3b0")],
+            }
+            for i, col in enumerate(styler.data.columns)
+            if col in ratio_cols
+        ],
+        overwrite=False,
+    )
+    return styler
+
+
+def _apply_two_decimals(
+    styler: pd.io.formats.style.Styler,
+) -> pd.io.formats.style.Styler:
+    df = styler.data
+    num_cols = df.select_dtypes("number").columns
+    if len(num_cols) == 0:
+        return styler
+    return styler.format({c: "{:.2f}" for c in num_cols}, na_rep="")
+
+
+# -----------------------------
+# Valid max concurrency summary helpers
+# -----------------------------
+def _config_value_columns(df: pd.DataFrame, conc_col: str) -> list[str]:
+    key_cols = [
+        c
+        for c in ["Model", "Dataset Name", "Input Len", "Output Len"]
+        if c in df.columns
+    ]
+    exclude = set(key_cols + [conc_col, "qps", "QPS"])
+
+    cols: list[str] = []
+    for c in df.columns:
+        if c in exclude:
+            continue
+        lc = str(c).lower()
+        if lc.startswith("ratio"):
+            continue
+        if lc.endswith("_name") or lc == "test name" or lc == "test_name":
+            continue
+        if pd.api.types.is_numeric_dtype(df[c]):
+            cols.append(c)
+    return cols
+
+
+def _max_concurrency_ok(
+    df: pd.DataFrame, conc_col: str, cfg_col: str, threshold: float
+):
+    if df is None or conc_col not in df.columns or cfg_col not in df.columns:
+        return pd.NA
+
+    d = df[[conc_col, cfg_col]].copy()
+    d[conc_col] = pd.to_numeric(d[conc_col], errors="coerce")
+    d[cfg_col] = pd.to_numeric(d[cfg_col], errors="coerce")
+    d = d.dropna(subset=[conc_col, cfg_col])
+
+    if d.empty:
+        return pd.NA
+
+    ok = d[d[cfg_col] <= threshold]
+    if ok.empty:
+        return pd.NA
+
+    return ok[conc_col].max()
+
+
+def _value_at_concurrency(df: pd.DataFrame, conc_col: str, cfg_col: str, conc_value):
+    if (
+        df is None
+        or conc_col not in df.columns
+        or cfg_col not in df.columns
+        or pd.isna(conc_value)
+    ):
+        return pd.NA
+
+    d = df[[conc_col, cfg_col]].copy()
+    d[conc_col] = pd.to_numeric(d[conc_col], errors="coerce")
+    d[cfg_col] = pd.to_numeric(d[cfg_col], errors="coerce")
+
+    conc_value = pd.to_numeric(conc_value, errors="coerce")
+    if pd.isna(conc_value):
+        return pd.NA
+
+    hit = d[d[conc_col] == conc_value]
+    if hit.empty:
+        return pd.NA
+    return hit[cfg_col].iloc[0]
+
+
+def build_valid_max_concurrency_summary_html(
+    tput_group_df: pd.DataFrame | None,
+    ttft_group_df: pd.DataFrame | None,
+    tpot_group_df: pd.DataFrame | None,
+    conc_col: str,
+    args,
+) -> str:
+    if ttft_group_df is None and tpot_group_df is None:
+        return ""
+
+    ttft_cols = (
+        _config_value_columns(ttft_group_df, conc_col)
+        if ttft_group_df is not None
+        else []
+    )
+    tpot_cols = (
+        _config_value_columns(tpot_group_df, conc_col)
+        if tpot_group_df is not None
+        else []
+    )
+    tput_cols = (
+        _config_value_columns(tput_group_df, conc_col)
+        if tput_group_df is not None
+        else []
+    )
+
+    if ttft_group_df is not None and tpot_group_df is not None:
+        cfg_cols = [c for c in ttft_cols if c in tpot_cols]
+        if tput_group_df is not None:
+            cfg_cols = [c for c in cfg_cols if c in tput_cols] or cfg_cols
+    else:
+        cfg_cols = ttft_cols or tpot_cols
+
+    if not cfg_cols:
+        cfg_cols = sorted(set(ttft_cols) | set(tpot_cols) | set(tput_cols), key=str)
+
+    rows = []
+    for cfg in cfg_cols:
+        ttft_max = (
+            _max_concurrency_ok(ttft_group_df, conc_col, cfg, args.ttft_max_ms)
+            if ttft_group_df is not None
+            else pd.NA
+        )
+        tpot_max = (
+            _max_concurrency_ok(tpot_group_df, conc_col, cfg, args.tpot_max_ms)
+            if tpot_group_df is not None
+            else pd.NA
+        )
+        both = (
+            pd.NA
+            if (pd.isna(ttft_max) or pd.isna(tpot_max))
+            else min(ttft_max, tpot_max)
+        )
+
+        tput_at_both = (
+            _value_at_concurrency(tput_group_df, conc_col, cfg, both)
+            if tput_group_df is not None
+            else pd.NA
+        )
+        ttft_at_both = (
+            _value_at_concurrency(ttft_group_df, conc_col, cfg, both)
+            if ttft_group_df is not None
+            else pd.NA
+        )
+        tpot_at_both = (
+            _value_at_concurrency(tpot_group_df, conc_col, cfg, both)
+            if tpot_group_df is not None
+            else pd.NA
+        )
+
+        rows.append(
+            {
+                "Configuration": cfg,
+                f"Max {conc_col} (TTFT ≤ {args.ttft_max_ms:g} ms)": ttft_max,
+                f"Max {conc_col} (TPOT ≤ {args.tpot_max_ms:g} ms)": tpot_max,
+                f"Max {conc_col} (Both)": both,
+                "Output Tput @ Both (tok/s)": tput_at_both,
+                "TTFT @ Both (ms)": ttft_at_both,
+                "TPOT @ Both (ms)": tpot_at_both,
+            }
+        )
+
+    summary_df = pd.DataFrame(rows)
+
+    # --- Coerce numeric columns so Styler doesn't miss them due to object dtype ---
+    for c in summary_df.columns:
+        if c == "Configuration":
+            continue
+        summary_df[c] = pd.to_numeric(summary_df[c], errors="coerce")
+
+    both_col = f"Max {conc_col} (Both)"
+
+    # --- Strict 2-decimal formatting for ALL non-Configuration columns ---
+    formatters = {}
+    for c in summary_df.columns:
+        if c == "Configuration":
+            continue
+        # default argument binds per-column formatter correctly
+        formatters[c] = lambda v: "" if pd.isna(v) else f"{float(v):.2f}"
+
+    styler = summary_df.style.format(formatters)
+
+    def _green(v):
+        return "background-color:#e6ffe6;font-weight:bold;" if pd.notna(v) else ""
+
+    if both_col in summary_df.columns:
+        styler = styler.map(_green, subset=[both_col])
+
+    title = (
+        '<div style="font-size: 1.15em; font-weight: 700; margin: 12px 0 6px 0;">'
+        "Valid Max Concurrency Summary"
+        "</div>\n"
+    )
+    return title + styler.to_html(table_attributes='border="1" class="dataframe"')
+
+
+# -----------------------------
+# Plot helper
+# -----------------------------
+def _add_limit_line(fig, y_value: float, label: str):
+    fig.add_hline(
+        y=y_value,
+        line_dash="dash",
+        line_color="red" if "ttft" in label.lower() else "blue",
+        annotation_text=f"{label}: {y_value} ms",
+        annotation_position="top left",
+    )
+    if plotly_found:
+        import plotly.graph_objects as go
+
+        fig.add_trace(
+            go.Scatter(
+                x=[None],
+                y=[None],
+                mode="lines",
+                line=dict(
+                    dash="dash",
+                    color="red" if "ttft" in label.lower() else "blue",
+                ),
+                name=label,
+            )
+        )
+
+
+# -----------------------------
+# Refactored main + group-first report
+# -----------------------------
+@dataclass(frozen=True)
+class MetricPlan:
+    data_cols: list[str]
+    drop_column: str
+
+
+def build_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser()
     parser.add_argument(
         "-f", "--file", action="append", type=str, help="input file name"
@@ -308,149 +537,289 @@ if __name__ == "__main__":
         default=100.0,
         help="Reference limit for TPOT plots (ms)",
     )
+    return parser
 
-    args = parser.parse_args()
 
+def choose_metrics(latency: str) -> MetricPlan:
+    latency = (latency or "").lower()
     drop_column = "P99"
-    name_column = "Test name"
-    info_cols = [
-        "Model",
-        "Dataset Name",
-        "Input Len",
-        "Output Len",
-        "TP Size",
-        "PP Size",
-        "# of max concurrency.",
-        "qps",
-    ]
 
-    if "median" in args.latency:
-        data_cols_to_compare = ["Output Tput (tok/s)", "Median TTFT (ms)", "Median"]
-        html_msgs_for_data_cols = [
-            "Compare Output Tokens /n",
-            "Median TTFT /n",
-            "Median TPOT /n",
-        ]
-        drop_column = "P99"
-    elif "p99" in args.latency:
-        data_cols_to_compare = ["Output Tput (tok/s)", "P99 TTFT (ms)", "P99"]
-        html_msgs_for_data_cols = [
-            "Compare Output Tokens /n",
-            "P99 TTFT /n",
-            "P99 TPOT /n",
-        ]
+    if "median" in latency:
+        return MetricPlan(
+            data_cols=["Output Tput (tok/s)", "Median TTFT (ms)", "Median"],
+            drop_column=drop_column,
+        )
+
+    return MetricPlan(
+        data_cols=["Output Tput (tok/s)", "P99 TTFT (ms)", "P99"],
+        drop_column=drop_column,
+    )
+
+
+def prepare_input_files(args, info_cols: list[str]) -> tuple[list[str], list[str]]:
+    if not args.file:
+        raise ValueError("No input files provided. Use -f/--file.")
 
     if len(args.file) == 1:
         files = split_json_by_tp_pp(args.file[0], output_root="splits")
         info_cols = [c for c in info_cols if c not in ("TP Size", "PP Size")]
     else:
         files = args.file
+
+    return files, info_cols
+
+
+def get_y_axis_col(info_cols: list[str], xaxis: str) -> str:
+    y_axis_index = info_cols.index(xaxis) if xaxis in info_cols else 6
+    return info_cols[y_axis_index]
+
+
+def get_group_cols(output_df: pd.DataFrame, info_cols: list[str]) -> list[str]:
+    filtered_info_cols = info_cols[:4]
+    group_cols = [c for c in filtered_info_cols if c in output_df.columns]
+    if not group_cols:
+        raise ValueError(
+            f"No valid group-by columns. Expected subset: {filtered_info_cols}, "
+            f"but DataFrame has: {list(output_df.columns)}"
+        )
+    return group_cols
+
+
+def normalize_group_key(name):
+    return name if isinstance(name, tuple) else (name,)
+
+
+def group_filename(name, prefix: str = "perf_comparison_") -> str:
+    name_vals = normalize_group_key(name)
+    safe = ",".join(map(str, name_vals)).replace(",", "_").replace("/", "-")
+    return f"{prefix}{safe}.html"
+
+
+def build_group_suffix(group_cols: list[str], name) -> str:
+    name_vals = normalize_group_key(name)
+    return " , ".join(f"{col} : [ {val} ] " for col, val in zip(group_cols, name_vals))
+
+
+def render_metric_table_html(
+    display_group: pd.DataFrame,
+    metric_label: str,
+    group_suffix: str,
+    args,
+) -> str:
+    title = (
+        f'<div style="font-size: 1.25em; font-weight: 600; margin: 12px 0;">'
+        f"{_html.escape(metric_label)}"
+        f" — {_html.escape(group_suffix)}"
+        f"</div>\n"
+    )
+
+    metric_name = metric_label.lower()
+    if "ttft" in metric_name:
+        styler = _highlight_threshold(display_group, args.ttft_max_ms)
+    elif ("tpot" in metric_name) or ("median" in metric_name) or ("p99" in metric_name):
+        styler = _highlight_threshold(display_group, args.tpot_max_ms)
+    else:
+        styler = display_group.style
+
+    styler = _apply_two_decimals(styler)
+    styler = highlight_ratio_columns(styler)
+
+    return title + styler.to_html(table_attributes='border="1" class="dataframe"')
+
+
+def maybe_write_plot(
+    main_fh,
+    sub_fh,
+    group_df: pd.DataFrame,
+    raw_data_cols: list[str],
+    metric_label: str,
+    y_axis_col: str,
+    args,
+):
+    if not (args.plot and plotly_found):
+        return
+
+    import plotly.express as px
+
+    df = group_df[raw_data_cols].sort_values(by=y_axis_col)
+    df_melted = df.melt(
+        id_vars=y_axis_col,
+        var_name="Configuration",
+        value_name=metric_label,
+    )
+
+    fig = px.line(
+        df_melted,
+        x=y_axis_col,
+        y=metric_label,
+        color="Configuration",
+        title=f"{metric_label} vs {y_axis_col}",
+        markers=True,
+    )
+
+    # Ensure plot hover + y tick labels are also 2 decimals.
+    fig.update_traces(hovertemplate="%{y:.2f}<extra></extra>")
+    fig.update_yaxes(tickformat=".2f")
+
+    metric_name = metric_label.lower()
+    if "ttft" in metric_name:
+        _add_limit_line(fig, args.ttft_max_ms, "TTFT limit")
+    elif ("tpot" in metric_name) or ("median" in metric_name) or ("p99" in metric_name):
+        _add_limit_line(fig, args.tpot_max_ms, "TPOT limit")
+
+    html = fig.to_html(full_html=True, include_plotlyjs="cdn")
+    main_fh.write(html)
+    sub_fh.write(html)
+
+
+def build_group_keys(
+    df: pd.DataFrame, group_cols: list[str], sort_cols: list[str] | None = None
+):
+    if sort_cols:
+        df = df.sort_values(by=sort_cols)
+    gb = df.groupby(group_cols, dropna=False)
+    return [k for k, _ in gb]
+
+
+def write_report_group_first(
+    files: list[str], info_cols: list[str], plan: MetricPlan, args
+):
+    name_column = "Test name"
+    y_axis_col = get_y_axis_col(info_cols, args.xaxis)
+
     print("comparing : " + ", ".join(files))
-    debug = args.debug
-    plot = args.plot
-    # For Plot feature, assign y axis from one of info_cols
-    y_axis_index = info_cols.index(args.xaxis) if args.xaxis in info_cols else 6
-    with open("perf_comparison.html", "w") as text_file:
-        for i in range(len(data_cols_to_compare)):
-            output_df, raw_data_cols = compare_data_columns(
-                files,
-                name_column,
-                data_cols_to_compare[i],
-                info_cols,
-                drop_column,
-                debug=debug,
+
+    metric_cache: dict[str, tuple[pd.DataFrame, list[str]]] = {}
+    group_cols_canonical: list[str] | None = None
+
+    for metric_label in plan.data_cols:
+        output_df, raw_data_cols = compare_data_columns(
+            files,
+            name_column,
+            metric_label,
+            info_cols,
+            plan.drop_column,
+            debug=args.debug,
+        )
+
+        raw_data_cols = list(raw_data_cols)
+        raw_data_cols.insert(0, y_axis_col)
+
+        group_cols = get_group_cols(output_df, info_cols)
+        if group_cols_canonical is None:
+            group_cols_canonical = group_cols
+        else:
+            group_cols_canonical = [c for c in group_cols_canonical if c in group_cols]
+
+        metric_cache[metric_label] = (
+            output_df.sort_values(by=args.xaxis),
+            raw_data_cols,
+        )
+
+    if not group_cols_canonical:
+        raise ValueError("No canonical group columns found across metrics.")
+
+    first_metric = plan.data_cols[0]
+    first_df_sorted, _ = metric_cache[first_metric]
+    group_keys = build_group_keys(
+        first_df_sorted, group_cols_canonical, sort_cols=[args.xaxis]
+    )
+
+    metric_groupbys = {
+        metric_label: df.groupby(group_cols_canonical, dropna=False)
+        for metric_label, (df, _) in metric_cache.items()
+    }
+
+    with open("perf_comparison.html", "w", encoding="utf-8") as main_fh:
+        main_fh.write('<meta charset="utf-8">\n')
+        for gkey in group_keys:
+            gkey_tuple = normalize_group_key(gkey)
+            suffix = build_group_suffix(group_cols_canonical, gkey_tuple)
+            sub_path = group_filename(gkey_tuple)
+            group_header = (
+                '<div style="font-size: 1.4em; font-weight: 700; '
+                'margin: 18px 0 10px 0;">'
+                f"{_html.escape(suffix)}"
+                "</div>\n"
             )
 
-            # For Plot feature, insert y axis from one of info_cols
-            raw_data_cols.insert(0, info_cols[y_axis_index])
+            main_fh.write(group_header)
+            with open(sub_path, "w", encoding="utf-8") as sub_fh:
+                sub_fh.write('<meta charset="utf-8">\n')
+                sub_fh.write(group_header)
+                tput_group_df = None
+                ttft_group_df = None
+                tpot_group_df = None
+                conc_col = args.xaxis
 
-            filtered_info_cols = info_cols[:-2]
-            existing_group_cols = [
-                c for c in filtered_info_cols if c in output_df.columns
-            ]
-            if not existing_group_cols:
-                raise ValueError(
-                    f"No valid group-by columns  "
-                    f"Expected subset: {filtered_info_cols}, "
-                    f"but DataFrame has: {list(output_df.columns)}"
+                for metric_label in plan.data_cols:
+                    gb = metric_groupbys[metric_label]
+                    df_sorted, raw_data_cols = metric_cache[metric_label]
+
+                    try:
+                        group_df = gb.get_group(gkey)
+                    except KeyError:
+                        missing = (
+                            '<div style="font-size: 1.1em; font-weight: 600; '
+                            'margin: 10px 0;">'
+                            f"{_html.escape(metric_label)} — missing for this group"
+                            "</div>\n"
+                        )
+
+                        main_fh.write(missing)
+                        sub_fh.write(missing)
+                        continue
+
+                    if conc_col not in group_df.columns:
+                        conc_col = _find_concurrency_col(group_df)
+
+                    mn = metric_label.lower().strip()
+                    if "tok/s" in mn:
+                        tput_group_df = group_df
+                    elif "ttft" in mn:
+                        ttft_group_df = group_df
+                    elif mn in ("p99", "median") or "tpot" in mn:
+                        tpot_group_df = group_df
+
+                    display_group = group_df.drop(
+                        columns=group_cols_canonical, errors="ignore"
+                    )
+
+                    html = render_metric_table_html(
+                        display_group, metric_label, suffix, args
+                    )
+                    main_fh.write(html)
+                    sub_fh.write(html)
+
+                    maybe_write_plot(
+                        main_fh,
+                        sub_fh,
+                        group_df=group_df,
+                        raw_data_cols=raw_data_cols,
+                        metric_label=metric_label,
+                        y_axis_col=y_axis_col,
+                        args=args,
+                    )
+
+                summary_html = build_valid_max_concurrency_summary_html(
+                    tput_group_df=tput_group_df,
+                    ttft_group_df=ttft_group_df,
+                    tpot_group_df=tpot_group_df,
+                    conc_col=conc_col,
+                    args=args,
                 )
-            # output_df_sorted = output_df.sort_values(by=existing_group_cols)
-            output_df_sorted = output_df.sort_values(by=args.xaxis)
-            output_groups = output_df_sorted.groupby(existing_group_cols, dropna=False)
-            for name, group in output_groups:
-                group_name = (
-                    ",".join(map(str, name)).replace(",", "_").replace("/", "-")
-                )
-                group_html_name = "perf_comparison_" + group_name + ".html"
+                if summary_html:
+                    main_fh.write(summary_html)
+                    sub_fh.write(summary_html)
 
-                metric_name = str(data_cols_to_compare[i]).lower()
-                if "tok/s" in metric_name:
-                    html = group.to_html()
-                elif "ttft" in metric_name:
-                    styler = _highlight_threshold(group, args.ttft_max_ms).format(
-                        {c: "{:.2f}" for c in group.select_dtypes("number").columns},
-                        na_rep="—",
-                    )
-                    html = styler.to_html(
-                        table_attributes='border="1" class="dataframe"'
-                    )
-                elif (
-                    "tpot" in metric_name
-                    or "median" in metric_name
-                    or "p99" in metric_name
-                ):
-                    styler = _highlight_threshold(group, args.tpot_max_ms).format(
-                        {c: "{:.2f}" for c in group.select_dtypes("number").columns},
-                        na_rep="—",
-                    )
-                    html = styler.to_html(
-                        table_attributes='border="1" class="dataframe"'
-                    )
 
-                text_file.write(html_msgs_for_data_cols[i])
-                text_file.write(html)
-                with open(group_html_name, "a+") as sub_text_file:
-                    sub_text_file.write(html_msgs_for_data_cols[i])
-                    sub_text_file.write(html)
+def main():
+    args = build_parser().parse_args()
+    info_cols = list(DEFAULT_INFO_COLS)
+    plan = choose_metrics(args.latency)
+    files, info_cols = prepare_input_files(args, info_cols)
+    write_report_group_first(files, info_cols, plan, args)
 
-                    if plot and plotly_found:
-                        import plotly.express as px
 
-                        df = group[raw_data_cols]
-                        df_sorted = df.sort_values(by=info_cols[y_axis_index])
-                        # Melt DataFrame for plotting
-                        df_melted = df_sorted.melt(
-                            id_vars=info_cols[y_axis_index],
-                            var_name="Configuration",
-                            value_name=data_cols_to_compare[i],
-                        )
-                        title = (
-                            data_cols_to_compare[i] + " vs " + info_cols[y_axis_index]
-                        )
-                        # Create Plotly line chart
-                        fig = px.line(
-                            df_melted,
-                            x=info_cols[y_axis_index],
-                            y=data_cols_to_compare[i],
-                            color="Configuration",
-                            title=title,
-                            markers=True,
-                        )
-
-                        # ---- Add threshold lines based on metric name ----
-                        if "ttft" in metric_name:
-                            _add_limit_line(fig, args.ttft_max_ms, "TTFT limit")
-                        elif (
-                            "tpot" in metric_name
-                            or "median" in metric_name
-                            or "p99" in metric_name
-                        ):
-                            _add_limit_line(fig, args.tpot_max_ms, "TPOT limit")
-
-                        # Export to HTML
-                        text_file.write(
-                            fig.to_html(full_html=True, include_plotlyjs="cdn")
-                        )
-                        sub_text_file.write(
-                            fig.to_html(full_html=True, include_plotlyjs="cdn")
-                        )
+if __name__ == "__main__":
+    main()

.buildkite/performance-benchmarks/scripts/convert-results-json-to-markdown.py

@@ -393,7 +393,7 @@ if __name__ == "__main__":
     with open(results_folder / md_file, "w") as f:
         results = read_markdown(
             "../.buildkite/performance-benchmarks/"
-            + "performance-benchmarks-descriptions.md"
+            "performance-benchmarks-descriptions.md"
         )
         results = results.format(
             latency_tests_markdown_table=latency_md_table,

.buildkite/performance-benchmarks/scripts/run-performance-benchmarks.sh

@@ -25,9 +25,9 @@ check_gpus() {
     echo "Need at least 1 GPU to run benchmarking."
     exit 1
   fi
-  
+
   declare -g arch_suffix=''
-  
+
   if command -v nvidia-smi; then
     declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
   elif command -v amd-smi; then
@@ -49,7 +49,11 @@ check_cpus() {
     echo "Need at least 1 NUMA to run benchmarking."
     exit 1
   fi
-  declare -g gpu_type="cpu"
+  if [[ "$(uname -m)" == "aarch64" ]] || [[ "$(uname -m)" == "arm64" ]]; then
+    declare -g gpu_type="arm64-cpu"
+  else
+    declare -g gpu_type="cpu"
+  fi
   echo "GPU type is $gpu_type"
 }
 
@@ -177,19 +181,20 @@ upload_to_buildkite() {
   $BUILDKITE_AGENT_COMMAND artifact upload "$RESULTS_FOLDER/*"
 }
 
-run_latency_tests() {
-  # run latency tests using `vllm bench latency` command
-  # $1: a json file specifying latency test cases
+run_benchmark_tests() {
+  # run benchmark tests using `vllm bench <test_type>` command
+  # $1: test type (latency or throughput)
+  # $2: a json file specifying test cases
 
-  local latency_test_file
-  latency_test_file=$1
+  local test_type=$1
+  local test_file=$2
 
-  # Iterate over latency tests
-  jq -c '.[]' "$latency_test_file" | while read -r params; do
+  # Iterate over tests
+  jq -c '.[]' "$test_file" | while read -r params; do
     # get the test name, and append the GPU type back to it.
     test_name=$(echo "$params" | jq -r '.test_name')
-    if [[ ! "$test_name" =~ ^latency_ ]]; then
-      echo "In latency-test.json, test_name must start with \"latency_\"."
+    if [[ ! "$test_name" =~ ^${test_type}_ ]]; then
+      echo "In ${test_type}-test.json, test_name must start with \"${test_type}_\"."
       exit 1
     fi
 
@@ -200,15 +205,15 @@ run_latency_tests() {
     fi
 
     # get arguments
-    latency_params=$(echo "$params" | jq -r '.parameters')
-    latency_args=$(json2args "$latency_params")
-    latency_environment_variables=$(echo "$params" | jq -r '.environment_variables')
-    latency_envs=$(json2envs "$latency_environment_variables")
+    bench_params=$(echo "$params" | jq -r '.parameters')
+    bench_args=$(json2args "$bench_params")
+    bench_environment_variables=$(echo "$params" | jq -r '.environment_variables')
+    bench_envs=$(json2envs "$bench_environment_variables")
 
     # check if there is enough GPU to run the test
-    tp=$(echo "$latency_params" | jq -r '.tensor_parallel_size')
-    if [ "$ON_CPU" == "1" ]; then
-      pp=$(echo "$latency_params" | jq -r '.pipeline_parallel_size')
+    tp=$(echo "$bench_params" | jq -r '.tensor_parallel_size')
+    if [[ "$ON_CPU" == "1" ]]; then
+      pp=$(echo "$bench_params" | jq -r '.pipeline_parallel_size // 1')
       world_size=$(($tp*$pp))
       if [[ $numa_count -lt $world_size  && -z "${REMOTE_HOST}" ]]; then
         echo "Required world-size $world_size but only $numa_count NUMA nodes found. Skip testcase $test_name."
@@ -221,97 +226,42 @@ run_latency_tests() {
       fi
     fi
 
-    latency_command=" $latency_envs vllm bench latency \
+    bench_command=" $bench_envs vllm bench $test_type \
       --output-json $RESULTS_FOLDER/${test_name}.json \
-      $latency_args"
+      $bench_args"
 
     echo "Running test case $test_name"
-    echo "Latency command: $latency_command"
+    echo "${test_type^} command: $bench_command"
 
-    # recoding benchmarking command ang GPU command
+    # recording benchmarking command and GPU command
     jq_output=$(jq -n \
-      --arg latency "$latency_command" \
+      --arg command "$bench_command" \
       --arg gpu "$gpu_type" \
+      --arg test_type "$test_type" \
       '{
-        latency_command: $latency,
+        ($test_type + "_command"): $command,
         gpu_type: $gpu
       }')
     echo "$jq_output" >"$RESULTS_FOLDER/$test_name.commands"
 
     # run the benchmark
-    eval "$latency_command"
+    eval "$bench_command"
 
     kill_gpu_processes
 
   done
 }
 
+run_latency_tests() {
+  run_benchmark_tests "latency" "$1"
+}
+
+run_startup_tests() {
+  run_benchmark_tests "startup" "$1"
+}
+
 run_throughput_tests() {
-  # run throughput tests using `vllm bench throughput`
-  # $1: a json file specifying throughput test cases
-
-  local throughput_test_file
-  throughput_test_file=$1
-
-  # Iterate over throughput tests
-  jq -c '.[]' "$throughput_test_file" | while read -r params; do
-    # get the test name, and append the GPU type back to it.
-    test_name=$(echo "$params" | jq -r '.test_name')
-    if [[ ! "$test_name" =~ ^throughput_ ]]; then
-      echo "In throughput-test.json, test_name must start with \"throughput_\"."
-      exit 1
-    fi
-
-    # if TEST_SELECTOR is set, only run the test cases that match the selector
-    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
-      echo "Skip test case $test_name."
-      continue
-    fi
-
-    # get arguments
-    throughput_params=$(echo "$params" | jq -r '.parameters')
-    throughput_args=$(json2args "$throughput_params")
-    throughput_environment_variables=$(echo "$params" | jq -r '.environment_variables')
-    throughput_envs=$(json2envs "$throughput_environment_variables")
-
-    # check if there is enough GPU to run the test
-    tp=$(echo "$throughput_params" | jq -r '.tensor_parallel_size')
-    if [ "$ON_CPU" == "1" ]; then
-      pp=$(echo "$throughput_params" | jq -r '.pipeline_parallel_size')
-      world_size=$(($tp*$pp))
-      if [[ $numa_count -lt $world_size  && -z "${REMOTE_HOST}" ]]; then
-        echo "Required world-size $world_size but only $numa_count NUMA nodes found. Skip testcase $test_name."
-        continue
-      fi
-    else
-      if [[ $gpu_count -lt $tp ]]; then
-        echo "Required tensor-parallel-size $tp but only $gpu_count GPU found. Skip testcase $test_name."
-        continue
-      fi
-    fi
-
-    throughput_command=" $throughput_envs vllm bench throughput \
-      --output-json $RESULTS_FOLDER/${test_name}.json \
-      $throughput_args"
-
-    echo "Running test case $test_name"
-    echo "Throughput command: $throughput_command"
-    # recoding benchmarking command ang GPU command
-    jq_output=$(jq -n \
-      --arg command "$throughput_command" \
-      --arg gpu "$gpu_type" \
-      '{
-        throughput_command: $command,
-        gpu_type: $gpu
-      }')
-    echo "$jq_output" >"$RESULTS_FOLDER/$test_name.commands"
-
-    # run the benchmark
-    eval "$throughput_command"
-
-    kill_gpu_processes
-
-  done
+  run_benchmark_tests "throughput" "$1"
 }
 
 run_serving_tests() {
@@ -393,8 +343,8 @@ run_serving_tests() {
 
     # check if there is enough resources to run the test
     tp=$(echo "$server_params" | jq -r '.tensor_parallel_size')
-    if [ "$ON_CPU" == "1" ]; then
-      pp=$(echo "$server_params" | jq -r '.pipeline_parallel_size')
+    if [[ "$ON_CPU" == "1" ]]; then
+      pp=$(echo "$server_params" | jq -r '.pipeline_parallel_size // 1')
       world_size=$(($tp*$pp))
       if [[ $numa_count -lt $world_size  && -z "${REMOTE_HOST}" ]]; then
         echo "Required world-size $world_size but only $numa_count NUMA nodes found. Skip testcase $test_name."
@@ -443,6 +393,11 @@ run_serving_tests() {
       fi
     fi
 
+    # save the compilation mode and optimization level on the serving results
+    # whenever they are set
+    compilation_config_mode=$(echo "$server_params" | jq -r '."compilation_config.mode" // empty')
+    optimization_level=$(echo "$server_params" | jq -r '.optimization_level // empty')
+
     # iterate over different QPS
     for qps in $qps_list; do
       # remove the surrounding single quote from qps
@@ -456,15 +411,15 @@ run_serving_tests() {
       for max_concurrency in $max_concurrency_list; do
         new_test_name=$test_name"_qps_"$qps"_concurrency_"$max_concurrency
         echo " new test name $new_test_name"
-        # pass the tensor parallel size to the client so that it can be displayed
-        # on the benchmark dashboard
+        # pass the tensor parallel size, the compilation mode, and the optimization
+        # level to the client so that they can be used on the benchmark dashboard
         client_command="vllm bench serve \
           --save-result \
           --result-dir $RESULTS_FOLDER \
           --result-filename ${new_test_name}.json \
           --request-rate $qps \
           --max-concurrency $max_concurrency \
-          --metadata "tensor_parallel_size=$tp" \
+          --metadata tensor_parallel_size=$tp compilation_config.mode=$compilation_config_mode optimization_level=$optimization_level \
           $client_args $client_remote_args "
 
         echo "Running test case $test_name with qps $qps"
@@ -496,9 +451,9 @@ run_serving_tests() {
 main() {
   local ARCH
   ARCH=''
-  if [ "$ON_CPU" == "1" ];then
-     check_cpus
-     ARCH='-cpu'
+  if [[ "$ON_CPU" == "1" ]]; then
+    check_cpus
+    ARCH="-$gpu_type"
   else
      check_gpus
      ARCH="$arch_suffix"
@@ -530,6 +485,7 @@ main() {
   # benchmarking
   run_serving_tests $QUICK_BENCHMARK_ROOT/tests/"${SERVING_JSON:-serving-tests$ARCH.json}"
   run_latency_tests $QUICK_BENCHMARK_ROOT/tests/"${LATENCY_JSON:-latency-tests$ARCH.json}"
+  run_startup_tests $QUICK_BENCHMARK_ROOT/tests/"${STARTUP_JSON:-startup-tests$ARCH.json}"
   run_throughput_tests $QUICK_BENCHMARK_ROOT/tests/"${THROUGHPUT_JSON:-throughput-tests$ARCH.json}"
 
   # postprocess benchmarking results

.buildkite/performance-benchmarks/tests/latency-tests-arm64-cpu.json

@@ -0,0 +1,26 @@
+[
+    {
+        "test_name": "latency_llama8B_tp1",
+        "environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+            "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+            "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+            "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "dtype": "bfloat16",
+            "distributed_executor_backend": "mp",
+            "block_size": 128,
+            "trust_remote_code": "",
+            "disable_log_stats": "",
+            "enforce_eager": "",
+            "max_num_batched_tokens": 2048,
+            "max_num_seqs": 256,
+            "num_iters_warmup": 5,
+            "num_iters": 15
+        }
+    }
+]

.buildkite/performance-benchmarks/tests/serving-tests-arm64-cpu.json

@@ -0,0 +1,130 @@
+{
+  "defaults": {
+    "qps_list": [
+      "inf"
+    ],
+    "max_concurrency_list": [
+      12,
+      16,
+      24,
+      32,
+      64,
+      128,
+      200
+    ],
+    "server_environment_variables": {
+      "VLLM_RPC_TIMEOUT": 100000,
+      "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+      "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+      "VLLM_CPU_SGL_KERNEL": 1,
+      "VLLM_CPU_KVCACHE_SPACE": 40
+    },
+    "server_parameters": {
+      "model": "meta-llama/Llama-3.1-8B-Instruct",
+      "tensor_parallel_size": 1,
+      "dtype": "bfloat16",
+      "distributed_executor_backend": "mp",
+      "block_size": 128,
+      "trust_remote_code": "",
+      "disable_log_stats": "",
+      "enforce_eager": "",
+      "max_num_batched_tokens": 2048,
+      "max_num_seqs": 256,
+      "load_format": "dummy"
+    },
+    "client_parameters": {
+      "model": "meta-llama/Llama-3.1-8B-Instruct",
+      "backend": "vllm",
+      "ignore-eos": "",
+      "num_prompts": 200
+    }
+  },
+  "tests": [
+    {
+      "test_name": "serving_llama8B_tp1_sharegpt",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "sharegpt",
+        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_sharegpt",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "sharegpt",
+        "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_128_128",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_random_128_128",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_128_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 2048
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_random_128_2048",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 2048
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp1_random_2048_128",
+      "server_parameters": {
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 2048,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_tp2_random_2048_128",
+      "server_parameters": {
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "dataset_name": "random",
+        "random-input-len": 2048,
+        "random-output-len": 128
+      }
+    }
+  ]
+}

.buildkite/performance-benchmarks/tests/serving-tests-cpu.json

@@ -19,10 +19,8 @@
       "block_size": 128,
       "trust_remote_code": "",
       "disable_log_stats": "",
-      "enforce_eager": "",
       "max_num_batched_tokens": 2048,
-      "max_num_seqs": 256,
-      "load_format": "dummy"
+      "max_num_seqs": 256
     },
     "client_parameters": {
       "model": "meta-llama/Llama-3.1-8B-Instruct",
@@ -151,6 +149,45 @@
         "random-output-len": 128
       }
     },
+    {
+      "test_name": "serving_llama8B_int4_tp1_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 1
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_int4_tp2_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 2
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
+    {
+      "test_name": "serving_llama8B_int4_tp4_random_128_128",
+      "server_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "tensor_parallel_size": 4
+      },
+      "client_parameters": {
+        "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "dataset_name": "random",
+        "random-input-len": 128,
+        "random-output-len": 128
+      }
+    },
     {
       "test_name": "serving_llama3B_tp1_random_128_128",
       "server_parameters": {

.buildkite/performance-benchmarks/tests/throughput-tests-arm64-cpu.json

@@ -0,0 +1,27 @@
+[
+    {
+        "test_name": "throughput_llama8B_tp1",
+        "environment_variables": {
+            "VLLM_RPC_TIMEOUT": 100000,
+            "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
+            "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
+            "VLLM_CPU_KVCACHE_SPACE": 40
+        },
+        "parameters": {
+            "model": "meta-llama/Llama-3.1-8B-Instruct",
+            "tensor_parallel_size": 1,
+            "load_format": "dummy",
+            "dtype": "bfloat16",
+            "distributed_executor_backend": "mp",
+            "block_size": 128,
+            "trust_remote_code": "",
+            "disable_log_stats": "",
+            "enforce_eager": "",
+            "max_num_batched_tokens": 2048,
+            "max_num_seqs": 256,
+            "dataset": "./ShareGPT_V3_unfiltered_cleaned_split.json",
+            "num_prompts": 200,
+            "backend": "vllm"
+        }
+    }
+]

.buildkite/release-pipeline.yaml

@@ -1,198 +1,713 @@
 steps:
-  # aarch64 + CUDA builds
-  - label: "Build arm64 wheel - CUDA 12.9"
-    depends_on: ~
-    id: build-wheel-arm64-cuda-12-9
-    agents:
-      queue: arm64_cpu_queue_postmerge
-    commands:
-      # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
-      # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
-      - "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-wheels.sh"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  - label: "Build arm64 wheel - CUDA 13.0"
-    depends_on: ~
-    id: build-wheel-arm64-cuda-13-0
-    agents:
-      queue: arm64_cpu_queue_postmerge
-    commands:
-      # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
-      # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
-      - "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-wheels.sh manylinux_2_35"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  # aarch64 build
-  - label: "Build arm64 CPU wheel"
-    depends_on: ~
-    id: build-wheel-arm64-cpu
-    agents:
-      queue: arm64_cpu_queue_postmerge
-    commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_BUILD_ACL=ON --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
-      - "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-wheels.sh manylinux_2_35"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  # x86 + CUDA builds
-  - label: "Build wheel - CUDA 12.9"
-    depends_on: ~
-    id: build-wheel-cuda-12-9
-    agents:
-      queue: cpu_queue_postmerge
-    commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
-      - "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-wheels.sh manylinux_2_31"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  - label: "Build wheel - CUDA 13.0"
-    depends_on: ~
-    id: build-wheel-cuda-13-0
-    agents:
-      queue: cpu_queue_postmerge
-    commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
-      - "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-wheels.sh manylinux_2_35"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  # x86 CPU wheel build
-  - label: "Build x86 CPU wheel"
-    depends_on: ~
-    id: build-wheel-x86-cpu
-    agents:
-      queue: cpu_queue_postmerge
-    commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
-      - "mkdir artifacts"
-      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/scripts/upload-wheels.sh manylinux_2_35"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  # Build release images (12.9)
-  - label: "Build release image (x86)"
-    depends_on: ~
-    id: build-release-image-x86
-    agents:
-      queue: cpu_queue_postmerge
-    commands:
-      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
-      # re-tag to default image tag and push, just in case arm64 build fails
-      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
-
-  - label: "Build release image (arm64)"
-    depends_on: ~
-    id: build-release-image-arm64
-    agents:
-      queue: arm64_cpu_queue_postmerge
-    commands:
-      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
-      - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
-
-  # Add job to create multi-arch manifest
-  - label: "Create multi-arch manifest"
-    depends_on:
-      - build-release-image-x86
-      - build-release-image-arm64
-    id: create-multi-arch-manifest
-    agents:
-      queue: cpu_queue_postmerge
-    commands:
-      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$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"
-
-  - label: "Annotate release workflow"
-    depends_on:
-      - create-multi-arch-manifest
-    id: annotate-release-workflow
-    agents:
-      queue: cpu_queue_postmerge
-    commands:
-      - "bash .buildkite/scripts/annotate-release.sh"
-
   - input: "Provide Release version here"
     id: input-release-version
     fields:
       - text: "What is the release version?"
         key: release-version
 
-  - block: "Build CPU release image"
-    key: block-cpu-release-image-build
+  - group: "Build Python wheels"
+    key: "build-wheels"
+    steps:
+      - label: "Build wheel - aarch64 - CUDA 12.9"
+        depends_on: ~
+        id: build-wheel-arm64-cuda-12-9
+        agents:
+          queue: arm64_cpu_queue_postmerge
+        commands:
+          # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
+          # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "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"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+      - label: "Build wheel - aarch64 - CUDA 13.0"
+        depends_on: ~
+        id: build-wheel-arm64-cuda-13-0
+        agents:
+          queue: arm64_cpu_queue_postmerge
+        commands:
+          # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
+          # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "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 manylinux_2_35"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+      - label: "Build wheel - aarch64 - CPU"
+        depends_on: ~
+        id: build-wheel-arm64-cpu
+        agents:
+          queue: arm64_cpu_queue_postmerge
+        commands:
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_BUILD_ACL=ON --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
+          - "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 manylinux_2_35"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+      - label: "Build wheel - x86_64 - CUDA 12.9"
+        depends_on: ~
+        id: build-wheel-x86-cuda-12-9
+        agents:
+          queue: cpu_queue_postmerge
+        commands:
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "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 manylinux_2_31"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+      - label: "Build wheel - x86_64 - CUDA 13.0"
+        depends_on: ~
+        id: build-wheel-x86-cuda-13-0
+        agents:
+          queue: cpu_queue_postmerge
+        commands:
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "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 manylinux_2_35"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+      - label: "Build wheel - x86_64 - CPU"
+        depends_on: ~
+        id: build-wheel-x86-cpu
+        agents:
+          queue: cpu_queue_postmerge
+        commands:
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
+          - "mkdir artifacts"
+          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+  - group: "Build release Docker images"
+    key: "build-release-images"
+    steps:
+      - label: "Build release image - x86_64 - CUDA 12.9"
+        depends_on: ~
+        id: build-release-image-x86
+        agents:
+          queue: cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
+          # re-tag to default image tag and push, just in case arm64 build fails
+          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
+
+      - label: "Build release image - aarch64 - CUDA 12.9"
+        depends_on: ~
+        id: build-release-image-arm64
+        agents:
+          queue: arm64_cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
+
+      - label: "Build release image - x86_64 - CUDA 13.0"
+        depends_on: ~
+        id: build-release-image-x86-cuda-13-0
+        agents:
+          queue: cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
+          # re-tag to default image tag and push, just in case arm64 build fails
+          - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
+
+      - label: "Build release image - aarch64 - CUDA 13.0"
+        depends_on: ~
+        id: build-release-image-arm64-cuda-13-0
+        agents:
+          queue: arm64_cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          # compute capability 12.0 for RTX-50 series / RTX PRO 6000 Blackwell, 12.1 for DGX Spark
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
+
+      - block: "Build release image for x86_64 CPU"
+        key: block-cpu-release-image-build
+        depends_on: ~
+
+      - label: "Build release image - x86_64 - CPU"
+        depends_on:
+          - block-cpu-release-image-build
+          - input-release-version
+        agents:
+          queue: cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(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:$(buildkite-agent meta-data get release-version)"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+      - block: "Build release image for arm64 CPU"
+        key: block-arm64-cpu-release-image-build
+        depends_on: ~
+
+      - label: "Build release image - arm64 - CPU"
+        depends_on: 
+          - block-arm64-cpu-release-image-build
+          - input-release-version
+        agents:
+          queue: arm64_cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:$(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:$(buildkite-agent meta-data get release-version)"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+  - group: "Publish release images"
+    key: "publish-release-images"
+    steps:
+      - label: "Create multi-arch manifest - CUDA 12.9"
+        depends_on:
+          - build-release-image-x86
+          - build-release-image-arm64
+        id: create-multi-arch-manifest
+        agents:
+          queue: small_cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$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:$BUILDKITE_COMMIT"
+
+      - label: "Annotate release workflow - CUDA 12.9"
+        depends_on:
+          - create-multi-arch-manifest
+        id: annotate-release-workflow
+        agents:
+          queue: small_cpu_queue_postmerge
+        commands:
+          - "bash .buildkite/scripts/annotate-release.sh"
+
+      - label: "Create multi-arch manifest - CUDA 13.0"
+        depends_on:
+          - build-release-image-x86-cuda-13-0
+          - build-release-image-arm64-cuda-13-0
+        id: create-multi-arch-manifest-cuda-13-0
+        agents:
+          queue: small_cpu_queue_postmerge
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - "docker manifest create public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64-cu130 --amend"
+          - "docker manifest push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
+
+      - label: "Publish nightly multi-arch image to DockerHub"
+        depends_on:
+          - create-multi-arch-manifest
+        if: build.env("NIGHTLY") == "1"
+        agents:
+          queue: small_cpu_queue_postmerge
+        commands:
+          - "bash .buildkite/scripts/push-nightly-builds.sh"
+          # Clean up old nightly builds (keep only last 14)
+          - "bash .buildkite/scripts/cleanup-nightly-builds.sh"
+        plugins:
+          - docker-login#v3.0.0:
+              username: vllmbot
+              password-env: DOCKERHUB_TOKEN
+        env:
+          DOCKER_BUILDKIT: "1"
+          DOCKERHUB_USERNAME: "vllmbot"
+
+      - label: "Publish nightly multi-arch image to DockerHub - CUDA 13.0"
+        depends_on:
+          - create-multi-arch-manifest-cuda-13-0
+        if: build.env("NIGHTLY") == "1"
+        agents:
+          queue: small_cpu_queue_postmerge
+        commands:
+          - "bash .buildkite/scripts/push-nightly-builds.sh cu130"
+          # Clean up old nightly builds (keep only last 14)
+          - "bash .buildkite/scripts/cleanup-nightly-builds.sh cu130-nightly-"
+        plugins:
+          - docker-login#v3.0.0:
+              username: vllmbot
+              password-env: DOCKERHUB_TOKEN
+        env:
+          DOCKER_BUILDKIT: "1"
+          DOCKERHUB_USERNAME: "vllmbot"
+
+  - group: "Publish wheels"
+    key: "publish-wheels"
+    steps:
+      - block: "Confirm update release wheels to PyPI (experimental, use with caution)?"
+        key: block-upload-release-wheels
+        depends_on:
+          - input-release-version
+          - build-wheels
+
+      - label: "Upload release wheels to PyPI"
+        depends_on:
+          - block-upload-release-wheels
+        id: upload-release-wheels
+        agents:
+          queue: small_cpu_queue_postmerge
+        commands:
+          - "bash .buildkite/scripts/upload-release-wheels-pypi.sh"
+
+  # =============================================================================
+  # ROCm Release Pipeline (x86_64 only)
+  # =============================================================================
+  #
+  # vLLM version is determined by the Buildkite checkout (like CUDA pipeline).
+  # To build a specific version, trigger the build from that branch/tag.
+  #
+  # Environment variables for ROCm builds (set via Buildkite UI or schedule):
+  #   ROCM_PYTHON_VERSION: Python version (default: 3.12)
+  #   PYTORCH_ROCM_ARCH: GPU architectures (default: gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151)
+  #   ROCM_UPLOAD_WHEELS: Upload to S3 (default: false for nightly, true for releases)
+  #   ROCM_FORCE_REBUILD: Force rebuild base wheels, ignore S3 cache (default: false)
+  #
+  # Note: ROCm version is determined by BASE_IMAGE in docker/Dockerfile.rocm_base
+  #       (currently rocm/dev-ubuntu-22.04:7.1-complete)
+  #
+  # =============================================================================
+
+  # ROCm Input Step - Collect build configuration (manual trigger only)
+  - input: "ROCm Wheel Release Build Configuration"
+    key: input-rocm-config
     depends_on: ~
+    if: build.source == "ui"
+    fields:
+      - text: "Python Version"
+        key: "rocm-python-version"
+        default: "3.12"
+        hint: "Python version (e.g., 3.12)"
+      - text: "GPU Architectures"
+        key: "rocm-pytorch-rocm-arch"
+        default: "gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151"
+        hint: "Semicolon-separated GPU architectures"
+      - select: "Upload Wheels to S3"
+        key: "rocm-upload-wheels"
+        default: "true"
+        options:
+          - label: "No - Build only (nightly/dev)"
+            value: "false"
+          - label: "Yes - Upload to S3 (release)"
+            value: "true"
+      - select: "Force Rebuild Base Wheels"
+        key: "rocm-force-rebuild"
+        default: "false"
+        hint: "Ignore S3 cache and rebuild base wheels from scratch"
+        options:
+          - label: "No - Use cached wheels if available"
+            value: "false"
+          - label: "Yes - Rebuild even if cache exists"
+            value: "true"
 
-  - label: "Build and publish CPU release image"
-    depends_on: block-cpu-release-image-build
-    agents:
-      queue: cpu_queue_postmerge
-    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 --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=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)"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  - block: "Build arm64 CPU release image"
-    key: block-arm64-cpu-release-image-build
-    depends_on: ~
-
-  - label: "Build and publish arm64 CPU release image"
-    depends_on: block-arm64-cpu-release-image-build
-    agents:
-      queue: arm64_cpu_queue_postmerge
-    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 --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)"
-    env:
-      DOCKER_BUILDKIT: "1"
-
-  - label: "Build and publish nightly multi-arch image to DockerHub"
+  # ROCm Job 1: Build ROCm Base Wheels (with S3 caching)
+  - label: ":rocm: Build ROCm Base Wheels"
+    id: build-rocm-base-wheels
     depends_on:
-      - create-multi-arch-manifest
-    if: build.env("NIGHTLY") == "1"
+      - step: input-rocm-config
+        allow_failure: true  # Allow failure so non-UI builds can proceed (input step is skipped)
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "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:$BUILDKITE_COMMIT-x86_64"
-      - "docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64"
-      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 vllm/vllm-openai:nightly-x86_64"
-      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 vllm/vllm-openai:nightly-aarch64"
-      - "docker push vllm/vllm-openai:nightly-x86_64"
-      - "docker push vllm/vllm-openai:nightly-aarch64"
-      - "docker manifest create vllm/vllm-openai:nightly vllm/vllm-openai:nightly-x86_64 vllm/vllm-openai:nightly-aarch64 --amend"
-      - "docker manifest create vllm/vllm-openai:nightly-$BUILDKITE_COMMIT vllm/vllm-openai:nightly-x86_64 vllm/vllm-openai:nightly-aarch64 --amend"
-      - "docker manifest push vllm/vllm-openai:nightly"
-      - "docker manifest push vllm/vllm-openai:nightly-$BUILDKITE_COMMIT"
-      # Clean up old nightly builds (keep only last 14)
-      - "bash .buildkite/scripts/cleanup-nightly-builds.sh"
-    plugins:
-      - docker-login#v3.0.0:
-          username: vllmbot
-          password-env: DOCKERHUB_TOKEN
+      # Set configuration and check cache
+      - |
+        set -euo pipefail
+
+        # Get values from meta-data (set by input step) or use defaults
+        PYTHON_VERSION="$$(buildkite-agent meta-data get rocm-python-version 2>/dev/null || echo '')"
+        export PYTHON_VERSION="$${PYTHON_VERSION:-3.12}"
+
+        PYTORCH_ROCM_ARCH="$$(buildkite-agent meta-data get rocm-pytorch-rocm-arch 2>/dev/null || echo '')"
+        export PYTORCH_ROCM_ARCH="$${PYTORCH_ROCM_ARCH:-gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151}"
+
+        # Check for force rebuild flag
+        ROCM_FORCE_REBUILD="$${ROCM_FORCE_REBUILD:-}"
+        if [ -z "$${ROCM_FORCE_REBUILD}" ]; then
+          ROCM_FORCE_REBUILD="$$(buildkite-agent meta-data get rocm-force-rebuild 2>/dev/null || echo '')"
+        fi
+
+        echo "========================================"
+        echo "ROCm Base Wheels Build Configuration"
+        echo "========================================"
+        echo "  PYTHON_VERSION: $${PYTHON_VERSION}"
+        echo "  PYTORCH_ROCM_ARCH: $${PYTORCH_ROCM_ARCH}"
+        echo "  ROCM_FORCE_REBUILD: $${ROCM_FORCE_REBUILD:-false}"
+        echo "========================================"
+
+        # Save resolved config for later jobs
+        buildkite-agent meta-data set "rocm-python-version" "$${PYTHON_VERSION}"
+        buildkite-agent meta-data set "rocm-pytorch-rocm-arch" "$${PYTORCH_ROCM_ARCH}"
+
+        # Check S3 cache for pre-built wheels
+        CACHE_KEY=$$(.buildkite/scripts/cache-rocm-base-wheels.sh key)
+        CACHE_PATH=$$(.buildkite/scripts/cache-rocm-base-wheels.sh path)
+        echo ""
+        echo "Cache key: $${CACHE_KEY}"
+        echo "Cache path: $${CACHE_PATH}"
+
+        # Save cache key for downstream jobs
+        buildkite-agent meta-data set "rocm-cache-key" "$${CACHE_KEY}"
+
+        CACHE_STATUS="miss"
+        if [ "$${ROCM_FORCE_REBUILD}" != "true" ]; then
+          CACHE_STATUS=$$(.buildkite/scripts/cache-rocm-base-wheels.sh check)
+        else
+          echo "Force rebuild requested, skipping cache check"
+        fi
+
+        if [ "$${CACHE_STATUS}" = "hit" ]; then
+          echo ""
+          echo "CACHE HIT! Downloading pre-built wheels..."
+          echo ""
+          .buildkite/scripts/cache-rocm-base-wheels.sh download
+
+          # Set the S3 path for the cached Docker image (for Job 2 to download)
+          S3_ARTIFACT_PATH="s3://$${S3_BUCKET}/rocm/cache/$${CACHE_KEY}"
+          buildkite-agent meta-data set "rocm-docker-image-s3-path" "$${S3_ARTIFACT_PATH}/rocm-base-image.tar.gz"
+
+          # Mark that we used cache (for Docker image handling)
+          buildkite-agent meta-data set "rocm-used-cache" "true"
+
+          echo ""
+          echo "Cache download complete. Skipping Docker build."
+          echo "Docker image will be downloaded from: $${S3_ARTIFACT_PATH}/rocm-base-image.tar.gz"
+        else
+          echo ""
+          echo "CACHE MISS. Building from scratch..."
+          echo ""
+
+          # Build full base image (for later vLLM build)
+          DOCKER_BUILDKIT=1 docker buildx build \
+            --file docker/Dockerfile.rocm_base \
+            --tag rocm/vllm-dev:base-$${BUILDKITE_BUILD_NUMBER} \
+            --build-arg PYTORCH_ROCM_ARCH="$${PYTORCH_ROCM_ARCH}" \
+            --build-arg PYTHON_VERSION="$${PYTHON_VERSION}" \
+            --build-arg USE_SCCACHE=1 \
+            --build-arg SCCACHE_BUCKET_NAME=vllm-build-sccache \
+            --build-arg SCCACHE_REGION_NAME=us-west-2 \
+            --build-arg SCCACHE_S3_NO_CREDENTIALS=0 \
+            --load \
+            .
+
+          # Build debs_wheel_release stage for wheel extraction
+          DOCKER_BUILDKIT=1 docker buildx build \
+            --file docker/Dockerfile.rocm_base \
+            --tag rocm-base-debs:$${BUILDKITE_BUILD_NUMBER} \
+            --target debs_wheel_release \
+            --build-arg PYTORCH_ROCM_ARCH="$${PYTORCH_ROCM_ARCH}" \
+            --build-arg PYTHON_VERSION="$${PYTHON_VERSION}" \
+            --build-arg USE_SCCACHE=1 \
+            --build-arg SCCACHE_BUCKET_NAME=vllm-build-sccache \
+            --build-arg SCCACHE_REGION_NAME=us-west-2 \
+            --build-arg SCCACHE_S3_NO_CREDENTIALS=0 \
+            --load \
+            .
+
+          # Extract wheels from Docker image
+          mkdir -p artifacts/rocm-base-wheels
+          container_id=$$(docker create rocm-base-debs:$${BUILDKITE_BUILD_NUMBER})
+          docker cp $${container_id}:/app/debs/. artifacts/rocm-base-wheels/
+          docker rm $${container_id}
+          echo "Extracted base wheels:"
+          ls -lh artifacts/rocm-base-wheels/
+
+          # Upload wheels to S3 cache for future builds
+          echo ""
+          echo "Uploading wheels to S3 cache..."
+          .buildkite/scripts/cache-rocm-base-wheels.sh upload
+
+          # Export base Docker image for reuse in vLLM build
+          mkdir -p artifacts/rocm-docker-image
+          docker save rocm/vllm-dev:base-$${BUILDKITE_BUILD_NUMBER} | gzip > artifacts/rocm-docker-image/rocm-base-image.tar.gz
+          echo "Docker image size:"
+          ls -lh artifacts/rocm-docker-image/
+
+          # Upload large Docker image to S3 (also cached by cache key)
+          S3_ARTIFACT_PATH="s3://$${S3_BUCKET}/rocm/cache/$${CACHE_KEY}"
+          echo "Uploading Docker image to $${S3_ARTIFACT_PATH}/"
+          aws s3 cp artifacts/rocm-docker-image/rocm-base-image.tar.gz "$${S3_ARTIFACT_PATH}/rocm-base-image.tar.gz"
+
+          # Save the S3 path for downstream jobs
+          buildkite-agent meta-data set "rocm-docker-image-s3-path" "$${S3_ARTIFACT_PATH}/rocm-base-image.tar.gz"
+
+          # Mark that we did NOT use cache
+          buildkite-agent meta-data set "rocm-used-cache" "false"
+
+          echo ""
+          echo "Build complete. Wheels cached for future builds."
+        fi
+    artifact_paths:
+      - "artifacts/rocm-base-wheels/*.whl"
     env:
       DOCKER_BUILDKIT: "1"
-      DOCKERHUB_USERNAME: "vllmbot"
+      S3_BUCKET: "vllm-wheels"
+
+  # ROCm Job 2: Build vLLM ROCm Wheel
+  - label: ":python: Build vLLM ROCm Wheel - x86_64"
+    id: build-rocm-vllm-wheel
+    depends_on:
+      - step: build-rocm-base-wheels
+        allow_failure: false
+    agents:
+      queue: cpu_queue_postmerge
+    timeout_in_minutes: 180
+    commands:
+      # Download artifacts and prepare Docker image
+      - |
+        set -euo pipefail
+
+        # Ensure git tags are up-to-date (Buildkite's default fetch doesn't update tags)
+        # This fixes version detection when tags are moved/force-pushed
+        echo "Fetching latest tags from origin..."
+        git fetch --tags --force origin
+        
+        # Log tag information for debugging version detection
+        echo "========================================"
+        echo "Git Tag Verification"
+        echo "========================================"
+        echo "Current HEAD: $(git rev-parse HEAD)"
+        echo "git describe --tags: $(git describe --tags 2>/dev/null || echo 'No tags found')"
+        echo ""
+        echo "Recent tags (pointing to commits near HEAD):"
+        git tag -l --sort=-creatordate | head -5
+        echo "setuptools_scm version detection:"
+        pip install -q setuptools_scm 2>/dev/null || true
+        python3 -c "import setuptools_scm; print('  Detected version:', setuptools_scm.get_version())" 2>/dev/null || echo "  (setuptools_scm not available in this environment)"
+        echo "========================================"
+
+        # Download wheel artifacts from current build
+        echo "Downloading wheel artifacts from current build"
+        buildkite-agent artifact download "artifacts/rocm-base-wheels/*.whl" .
+
+        # Download Docker image from S3 (too large for Buildkite artifacts)
+        DOCKER_IMAGE_S3_PATH="$$(buildkite-agent meta-data get rocm-docker-image-s3-path 2>/dev/null || echo '')"
+        if [ -z "$${DOCKER_IMAGE_S3_PATH}" ]; then
+          echo "ERROR: rocm-docker-image-s3-path metadata not found"
+          echo "This should have been set by the build-rocm-base-wheels job"
+          exit 1
+        fi
+        echo "Downloading Docker image from $${DOCKER_IMAGE_S3_PATH}"
+        mkdir -p artifacts/rocm-docker-image
+        aws s3 cp "$${DOCKER_IMAGE_S3_PATH}" artifacts/rocm-docker-image/rocm-base-image.tar.gz
+
+        # Load base Docker image and capture the tag
+        echo "Loading base Docker image..."
+        LOAD_OUTPUT=$$(gunzip -c artifacts/rocm-docker-image/rocm-base-image.tar.gz | docker load)
+        echo "$${LOAD_OUTPUT}"
+        # Extract the actual loaded image tag from "Loaded image: <tag>" output
+        # This avoids picking up stale images (like rocm/vllm-dev:nightly) already on the agent
+        BASE_IMAGE_TAG=$$(echo "$${LOAD_OUTPUT}" | grep "Loaded image:" | sed 's/Loaded image: //')
+        if [ -z "$${BASE_IMAGE_TAG}" ]; then
+          echo "ERROR: Failed to extract image tag from docker load output"
+          echo "Load output was: $${LOAD_OUTPUT}"
+          exit 1
+        fi
+        echo "Loaded base image: $${BASE_IMAGE_TAG}"
+
+        # Prepare base wheels for Docker build context
+        mkdir -p docker/context/base-wheels
+        touch docker/context/base-wheels/.keep
+        cp artifacts/rocm-base-wheels/*.whl docker/context/base-wheels/
+        echo "Base wheels for vLLM build:"
+        ls -lh docker/context/base-wheels/
+
+        # Get GPU architectures from meta-data
+        PYTORCH_ROCM_ARCH="$$(buildkite-agent meta-data get rocm-pytorch-rocm-arch 2>/dev/null || echo '')"
+        PYTORCH_ROCM_ARCH="$${PYTORCH_ROCM_ARCH:-gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151}"
+
+        echo "========================================"
+        echo "Building vLLM wheel with:"
+        echo "  BUILDKITE_COMMIT: $${BUILDKITE_COMMIT}"
+        echo "  BUILDKITE_BRANCH: $${BUILDKITE_BRANCH}"
+        echo "  PYTORCH_ROCM_ARCH: $${PYTORCH_ROCM_ARCH}"
+        echo "  BASE_IMAGE: $${BASE_IMAGE_TAG}"
+        echo "========================================"
+
+        # Build vLLM wheel using local checkout (REMOTE_VLLM=0)
+        DOCKER_BUILDKIT=1 docker build \
+          --file docker/Dockerfile.rocm \
+          --target export_vllm_wheel_release \
+          --output type=local,dest=rocm-dist \
+          --build-arg BASE_IMAGE="$${BASE_IMAGE_TAG}" \
+          --build-arg ARG_PYTORCH_ROCM_ARCH="$${PYTORCH_ROCM_ARCH}" \
+          --build-arg REMOTE_VLLM=0 \
+          --build-arg GIT_REPO_CHECK=1 \
+          --build-arg USE_SCCACHE=1 \
+          --build-arg SCCACHE_BUCKET_NAME=vllm-build-sccache \
+          --build-arg SCCACHE_REGION_NAME=us-west-2 \
+          --build-arg SCCACHE_S3_NO_CREDENTIALS=0 \
+          .
+
+        echo "Built vLLM wheel:"
+        ls -lh rocm-dist/*.whl
+
+        # Copy wheel to artifacts directory
+        mkdir -p artifacts/rocm-vllm-wheel
+        cp rocm-dist/*.whl artifacts/rocm-vllm-wheel/
+        echo "Final vLLM wheel:"
+        ls -lh artifacts/rocm-vllm-wheel/
+    artifact_paths:
+      - "artifacts/rocm-vllm-wheel/*.whl"
+    env:
+      DOCKER_BUILDKIT: "1"
+      S3_BUCKET: "vllm-wheels"
+
+  # ROCm Job 3: Upload Wheels to S3
+  - label: ":s3: Upload ROCm Wheels to S3"
+    id: upload-rocm-wheels
+    depends_on:
+      - step: build-rocm-vllm-wheel
+        allow_failure: false
+    agents:
+      queue: cpu_queue_postmerge
+    timeout_in_minutes: 60
+    commands:
+      # Download all wheel artifacts and run upload
+      - |
+        set -euo pipefail
+
+        # Check if upload is enabled (from env var, meta-data, or release branch)
+        ROCM_UPLOAD_WHEELS="$${ROCM_UPLOAD_WHEELS:-}"
+        if [ -z "$${ROCM_UPLOAD_WHEELS}" ]; then
+          # Try to get from meta-data (input form)
+          ROCM_UPLOAD_WHEELS="$$(buildkite-agent meta-data get rocm-upload-wheels 2>/dev/null || echo '')"
+        fi
+
+        echo "========================================"
+        echo "Upload check:"
+        echo "  ROCM_UPLOAD_WHEELS: $${ROCM_UPLOAD_WHEELS}"
+        echo "  BUILDKITE_BRANCH: $${BUILDKITE_BRANCH}"
+        echo "========================================"
+
+        # Skip upload if not enabled
+        if [ "$${ROCM_UPLOAD_WHEELS}" != "true" ]; then
+          echo "Skipping S3 upload (ROCM_UPLOAD_WHEELS != true, NIGHTLY != 1, not a release branch)"
+          echo "To enable upload, set 'Upload Wheels to S3' to 'Yes' in the build configuration"
+          exit 0
+        fi
+
+        echo "Upload enabled, proceeding..."
+
+        # Download artifacts from current build
+        echo "Downloading artifacts from current build"
+        buildkite-agent artifact download "artifacts/rocm-base-wheels/*.whl" .
+        buildkite-agent artifact download "artifacts/rocm-vllm-wheel/*.whl" .
+
+        # Run upload script
+        bash .buildkite/scripts/upload-rocm-wheels.sh
+    env:
+      DOCKER_BUILDKIT: "1"
+      S3_BUCKET: "vllm-wheels"
+
+  # ROCm Job 4: Annotate ROCm Wheel Release
+  - label: ":memo: Annotate ROCm wheel release"
+    id: annotate-rocm-release
+    depends_on:
+      - step: upload-rocm-wheels
+        allow_failure: true
+      - step: input-release-version
+        allow_failure: true
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "bash .buildkite/scripts/annotate-rocm-release.sh"
+    env:
+      S3_BUCKET: "vllm-wheels"
+
+  # ROCm Job 5: Generate Root Index for ROCm Wheels (for release only)
+  # This is the job to create https://wheels.vllm.ai/rocm/ index allowing
+  # users to install with `uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/`
+  - block: "Generate Root Index for ROCm Wheels for Release"
+    key: block-generate-root-index-rocm-wheels
+    depends_on: upload-rocm-wheels
+
+  - label: ":package: Generate Root Index for ROCm Wheels for Release"
+    depends_on: block-generate-root-index-rocm-wheels
+    id: generate-root-index-rocm-wheels
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "bash tools/vllm-rocm/generate-rocm-wheels-root-index.sh"
+    env:
+      S3_BUCKET: "vllm-wheels"
+      VARIANT: "rocm700"
+
+  # ROCm Job 5: Build ROCm Release Docker Image
+  - label: ":docker: Build release image - x86_64 - ROCm"
+    id: build-rocm-release-image
+    depends_on:
+      - step: build-rocm-base-wheels
+        allow_failure: false
+    agents:
+      queue: cpu_queue_postmerge
+    timeout_in_minutes: 60
+    commands:
+      - |
+        set -euo pipefail
+
+        # Login to ECR
+        aws ecr-public get-login-password --region us-east-1 | \
+          docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
+
+        # Download Docker image from S3 (set by build-rocm-base-wheels)
+        DOCKER_IMAGE_S3_PATH="$$(buildkite-agent meta-data get rocm-docker-image-s3-path 2>/dev/null || echo '')"
+        if [ -z "$${DOCKER_IMAGE_S3_PATH}" ]; then
+          echo "ERROR: rocm-docker-image-s3-path metadata not found"
+          exit 1
+        fi
+
+        echo "Downloading base image from $${DOCKER_IMAGE_S3_PATH}"
+        mkdir -p artifacts/rocm-docker-image
+        aws s3 cp "$${DOCKER_IMAGE_S3_PATH}" artifacts/rocm-docker-image/rocm-base-image.tar.gz
+
+        # Load base Docker image
+        echo "Loading base Docker image..."
+        LOAD_OUTPUT=$$(gunzip -c artifacts/rocm-docker-image/rocm-base-image.tar.gz | docker load)
+        BASE_IMAGE_TAG=$$(echo "$${LOAD_OUTPUT}" | grep "Loaded image:" | sed 's/Loaded image: //')
+        echo "Loaded base image: $${BASE_IMAGE_TAG}"
+
+        # Tag and push the base image to ECR
+        docker tag "$${BASE_IMAGE_TAG}" public.ecr.aws/q9t5s3a7/vllm-release-repo:$${BUILDKITE_COMMIT}-rocm-base
+        docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$${BUILDKITE_COMMIT}-rocm-base
+        echo "Pushed base image: public.ecr.aws/q9t5s3a7/vllm-release-repo:$${BUILDKITE_COMMIT}-rocm-base"
+
+        # Get GPU architectures from meta-data
+        PYTORCH_ROCM_ARCH="$$(buildkite-agent meta-data get rocm-pytorch-rocm-arch 2>/dev/null || echo '')"
+        PYTORCH_ROCM_ARCH="$${PYTORCH_ROCM_ARCH:-gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151}"
+
+        # Build vLLM ROCm release image using cached base
+        DOCKER_BUILDKIT=1 docker build \
+          --build-arg max_jobs=16 \
+          --build-arg BASE_IMAGE="$${BASE_IMAGE_TAG}" \
+          --build-arg ARG_PYTORCH_ROCM_ARCH="$${PYTORCH_ROCM_ARCH}" \
+          --build-arg USE_SCCACHE=1 \
+          --build-arg SCCACHE_BUCKET_NAME=vllm-build-sccache \
+          --build-arg SCCACHE_REGION_NAME=us-west-2 \
+          --build-arg SCCACHE_S3_NO_CREDENTIALS=0 \
+          --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$${BUILDKITE_COMMIT}-rocm \
+          --target vllm-openai \
+          --progress plain \
+          -f docker/Dockerfile.rocm .
+
+        # Push to ECR
+        docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$${BUILDKITE_COMMIT}-rocm
+        echo "Pushed: public.ecr.aws/q9t5s3a7/vllm-release-repo:$${BUILDKITE_COMMIT}-rocm"
+    env:
+      DOCKER_BUILDKIT: "1"
+      S3_BUCKET: "vllm-wheels"

.buildkite/scripts/annotate-release.sh

@@ -11,27 +11,32 @@ fi
 buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
 To download the wheel (by commit):
 \`\`\`
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_31_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux_2_31_aarch64.whl .
 
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
+(Optional) For CUDA 13.0:
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux_2_35_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux_2_35_aarch64.whl .
+
+(Optional) For CPU:
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cpu-cp38-abi3-manylinux_2_35_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cpu-cp38-abi3-manylinux_2_35_aarch64.whl .
 \`\`\`
 
-To download the wheel (by version):
-\`\`\`
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.whl .
-
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu129/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu130/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux1_x86_64.whl .
-\`\`\`
 
 To download and upload the image:
 
 \`\`\`
+Download images:
+
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64
 docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm
+
+Tag and push images:
 
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64 vllm/vllm-openai:x86_64
 docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:latest-x86_64
@@ -39,16 +44,69 @@ docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
 docker push vllm/vllm-openai:latest-x86_64
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
 
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64-cu130 vllm/vllm-openai:x86_64-cu130
+docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:latest-x86_64-cu130
+docker tag vllm/vllm-openai:x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130
+docker push vllm/vllm-openai:latest-x86_64-cu130
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130
+
 docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64 vllm/vllm-openai:aarch64
 docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:latest-aarch64
 docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 docker push vllm/vllm-openai:latest-aarch64
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64-cu130 vllm/vllm-openai:aarch64-cu130
+docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:latest-aarch64-cu130
+docker tag vllm/vllm-openai:aarch64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker push vllm/vllm-openai:latest-aarch64-cu130
+docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-rocm
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:latest
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:v${RELEASE_VERSION}-rocm
+docker push vllm/vllm-openai-rocm:latest
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-rocm
+
+Create multi-arch manifest:
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:latest-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+docker push vllm/vllm-openai-rocm:latest-base
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+
 docker manifest rm vllm/vllm-openai:latest
 docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64
 docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 docker manifest push vllm/vllm-openai:latest
 docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
+
+docker manifest rm vllm/vllm-openai:latest-cu130
+docker manifest create vllm/vllm-openai:latest-cu130 vllm/vllm-openai:latest-x86_64-cu130 vllm/vllm-openai:latest-aarch64-cu130
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION}-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64-cu130 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64-cu130
+docker manifest push vllm/vllm-openai:latest-cu130
+docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}-cu130
+
+# CPU images (vllm/vllm-openai-cpu)
+docker pull public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION}
+docker pull public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION}
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v${RELEASE_VERSION} vllm/vllm-openai-cpu:x86_64
+docker tag vllm/vllm-openai-cpu:x86_64 vllm/vllm-openai-cpu:latest-x86_64
+docker tag vllm/vllm-openai-cpu:x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
+docker push vllm/vllm-openai-cpu:latest-x86_64
+docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-arm64-cpu-release-repo:v${RELEASE_VERSION} vllm/vllm-openai-cpu:arm64
+docker tag vllm/vllm-openai-cpu:arm64 vllm/vllm-openai-cpu:latest-arm64
+docker tag vllm/vllm-openai-cpu:arm64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+docker push vllm/vllm-openai-cpu:latest-arm64
+docker push vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+
+docker manifest rm vllm/vllm-openai-cpu:latest || true
+docker manifest create vllm/vllm-openai-cpu:latest vllm/vllm-openai-cpu:latest-x86_64 vllm/vllm-openai-cpu:latest-arm64
+docker manifest create vllm/vllm-openai-cpu:v${RELEASE_VERSION} vllm/vllm-openai-cpu:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai-cpu:v${RELEASE_VERSION}-arm64
+docker manifest push vllm/vllm-openai-cpu:latest
+docker manifest push vllm/vllm-openai-cpu:v${RELEASE_VERSION}
 \`\`\`
-EOF 
+EOF

.buildkite/scripts/annotate-rocm-release.sh

@@ -0,0 +1,112 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Generate Buildkite annotation for ROCm wheel release
+set -ex
+
+# Get build configuration from meta-data
+# Extract ROCm version dynamically from Dockerfile.rocm_base
+# BASE_IMAGE format: rocm/dev-ubuntu-22.04:7.0-complete -> extracts "7.0"
+ROCM_VERSION=$(grep -E '^ARG BASE_IMAGE=' docker/Dockerfile.rocm_base | sed -E 's/.*:([0-9]+\.[0-9]+).*/\1/' || echo "unknown")
+PYTHON_VERSION=$(buildkite-agent meta-data get rocm-python-version 2>/dev/null || echo "3.12")
+PYTORCH_ROCM_ARCH=$(buildkite-agent meta-data get rocm-pytorch-rocm-arch 2>/dev/null || echo "gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
+
+# TODO: Enable the nightly build for ROCm
+# Get release version, default to 1.0.0.dev for nightly/per-commit builds
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version 2>/dev/null || echo "")
+if [ -z "${RELEASE_VERSION}" ]; then
+  RELEASE_VERSION="1.0.0.dev"
+fi
+
+# S3 URLs
+S3_BUCKET="${S3_BUCKET:-vllm-wheels}"
+S3_REGION="${AWS_DEFAULT_REGION:-us-west-2}"
+S3_URL="http://${S3_BUCKET}.s3-website-${S3_REGION}.amazonaws.com"
+
+# Format ROCm version for path (e.g., "7.1" -> "rocm710")
+ROCM_VERSION_PATH="rocm$(echo ${ROCM_VERSION} | tr -d '.')"
+ROCM_PATH="rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}"
+buildkite-agent annotate --style 'success' --context 'rocm-release-workflow' << EOF
+## ROCm Wheel and Docker Image Releases
+### Build Configuration
+| Setting | Value |
+|---------|-------|
+| **ROCm Version** | ${ROCM_VERSION} |
+| **Python Version** | ${PYTHON_VERSION} |
+| **GPU Architectures** | ${PYTORCH_ROCM_ARCH} |
+| **Branch** | \`${BUILDKITE_BRANCH}\` |
+| **Commit** | \`${BUILDKITE_COMMIT}\` |
+
+### :package: Installation
+
+**Install from this build (by commit):**
+
+\`\`\`bash
+pip install vllm --extra-index-url ${S3_URL}/${ROCM_PATH}/ --trusted-host ${S3_BUCKET}.s3-website-${S3_REGION}.amazonaws.com
+
+# Example for ROCm ${ROCM_VERSION}:
+pip install vllm --extra-index-url ${S3_URL}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/ --trusted-host ${S3_BUCKET}.s3-website-${S3_REGION}.amazonaws.com
+\`\`\`
+
+**Install from nightly (if published):**
+
+\`\`\`bash
+pip install vllm --extra-index-url ${S3_URL}/rocm/nightly/ --trusted-host ${S3_BUCKET}.s3-website-${S3_REGION}.amazonaws.com
+\`\`\`
+
+### :floppy_disk: Download Wheels Directly
+
+\`\`\`bash
+# List all ROCm wheels
+aws s3 ls s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/
+# Download specific wheels
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/vllm-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/torch-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/triton-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/triton-kernels-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/torchvision-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/torchaudio-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/amdsmi-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/aiter-*.whl .
+aws s3 cp s3://${S3_BUCKET}/rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}/flash-attn-*.whl .
+\`\`\`
+
+### :gear: Included Packages
+- **vllm**: vLLM with ROCm support
+- **torch**: PyTorch built for ROCm ${ROCM_VERSION}
+- **triton**: Triton
+- **triton-kernels**: Triton kernels
+- **torchvision**: TorchVision for ROCm PyTorch
+- **torchaudio**: Torchaudio for ROCm PyTorch
+- **amdsmi**: AMD SMI Python bindings
+- **aiter**: Aiter for ROCm
+- **flash-attn**: Flash Attention for ROCm
+
+### :warning: Notes
+- These wheels are built for **ROCm ${ROCM_VERSION}** and will NOT work with CUDA GPUs
+- Supported GPU architectures: ${PYTORCH_ROCM_ARCH}
+- Platform: Linux x86_64 only
+
+### :package: Docker Image Release
+
+To download and upload the image:
+
+\`\`\`
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm-base vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:latest-base
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}-base vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+docker push vllm/vllm-openai-rocm:latest-base
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}-base
+
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-rocm vllm/vllm-openai-rocm:${BUILDKITE_COMMIT}
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:latest
+docker tag vllm/vllm-openai-rocm:${BUILDKITE_COMMIT} vllm/vllm-openai-rocm:v${RELEASE_VERSION}
+docker push vllm/vllm-openai-rocm:latest
+docker push vllm/vllm-openai-rocm:v${RELEASE_VERSION}
+\`\`\`
+
+EOF

.buildkite/scripts/cache-rocm-base-wheels.sh

@@ -0,0 +1,140 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Cache helper for ROCm base wheels
+#
+# This script manages caching of pre-built ROCm base wheels (torch, triton, etc.)
+# to avoid rebuilding them when Dockerfile.rocm_base hasn't changed.
+#
+# Usage:
+#   cache-rocm-base-wheels.sh check    - Check if cache exists, outputs "hit" or "miss"
+#   cache-rocm-base-wheels.sh upload   - Upload wheels to cache
+#   cache-rocm-base-wheels.sh download - Download wheels from cache
+#   cache-rocm-base-wheels.sh key      - Output the cache key
+#
+# Environment variables:
+#   S3_BUCKET          - S3 bucket name (default: vllm-wheels)
+#   PYTHON_VERSION     - Python version (affects cache key)
+#   PYTORCH_ROCM_ARCH  - GPU architectures (affects cache key)
+#
+# Note: ROCm version is determined by BASE_IMAGE in Dockerfile.rocm_base,
+#       so changes to ROCm version are captured by the Dockerfile hash.
+
+set -euo pipefail
+
+BUCKET="${S3_BUCKET:-vllm-wheels}"
+DOCKERFILE="docker/Dockerfile.rocm_base"
+CACHE_PREFIX="rocm/cache"
+
+# Generate hash from Dockerfile content + build args
+generate_cache_key() {
+    # Include Dockerfile content
+    if [[ ! -f "$DOCKERFILE" ]]; then
+        echo "ERROR: Dockerfile not found: $DOCKERFILE" >&2
+        exit 1
+    fi
+    local dockerfile_hash=$(sha256sum "$DOCKERFILE" | cut -c1-16)
+
+    # Include key build args that affect the output
+    # These should match the ARGs in Dockerfile.rocm_base that change the build output
+    # Note: ROCm version is determined by BASE_IMAGE in the Dockerfile, so it's captured by dockerfile_hash
+    local args_string="${PYTHON_VERSION:-}|${PYTORCH_ROCM_ARCH:-}"
+    local args_hash=$(echo "$args_string" | sha256sum | cut -c1-8)
+
+    echo "${dockerfile_hash}-${args_hash}"
+}
+
+CACHE_KEY=$(generate_cache_key)
+CACHE_PATH="s3://${BUCKET}/${CACHE_PREFIX}/${CACHE_KEY}/"
+
+case "${1:-}" in
+    check)
+        echo "Checking cache for key: ${CACHE_KEY}" >&2
+        echo "Cache path: ${CACHE_PATH}" >&2
+        echo "Variables used in cache key:" >&2
+        echo "  PYTHON_VERSION: ${PYTHON_VERSION:-<not set>}" >&2
+        echo "  PYTORCH_ROCM_ARCH: ${PYTORCH_ROCM_ARCH:-<not set>}" >&2
+
+        # Check if cache exists by listing objects
+        # We look for at least one .whl file
+        echo "Running: aws s3 ls ${CACHE_PATH}" >&2
+        S3_OUTPUT=$(aws s3 ls "${CACHE_PATH}" 2>&1) || true
+        echo "S3 ls output:" >&2
+        echo "$S3_OUTPUT" | head -5 >&2
+
+        if echo "$S3_OUTPUT" | grep -q "\.whl"; then
+            echo "hit"
+        else
+            echo "miss"
+        fi
+        ;;
+
+    upload)
+        echo "========================================"
+        echo "Uploading wheels to cache"
+        echo "========================================"
+        echo "Cache key: ${CACHE_KEY}"
+        echo "Cache path: ${CACHE_PATH}"
+        echo ""
+
+        if [[ ! -d "artifacts/rocm-base-wheels" ]]; then
+            echo "ERROR: artifacts/rocm-base-wheels directory not found" >&2
+            exit 1
+        fi
+
+        WHEEL_COUNT=$(ls artifacts/rocm-base-wheels/*.whl 2>/dev/null | wc -l)
+        if [[ "$WHEEL_COUNT" -eq 0 ]]; then
+            echo "ERROR: No wheels found in artifacts/rocm-base-wheels/" >&2
+            exit 1
+        fi
+
+        echo "Uploading $WHEEL_COUNT wheels..."
+        aws s3 cp --recursive artifacts/rocm-base-wheels/ "${CACHE_PATH}"
+
+        echo ""
+        echo "Cache upload complete!"
+        echo "========================================"
+        ;;
+
+    download)
+        echo "========================================"
+        echo "Downloading wheels from cache"
+        echo "========================================"
+        echo "Cache key: ${CACHE_KEY}"
+        echo "Cache path: ${CACHE_PATH}"
+        echo ""
+
+        mkdir -p artifacts/rocm-base-wheels
+        aws s3 cp --recursive "${CACHE_PATH}" artifacts/rocm-base-wheels/
+
+        echo ""
+        echo "Downloaded wheels:"
+        ls -lh artifacts/rocm-base-wheels/
+
+        WHEEL_COUNT=$(ls artifacts/rocm-base-wheels/*.whl 2>/dev/null | wc -l)
+        echo ""
+        echo "Total: $WHEEL_COUNT wheels"
+        echo "========================================"
+        ;;
+
+    key)
+        echo "${CACHE_KEY}"
+        ;;
+
+    path)
+        echo "${CACHE_PATH}"
+        ;;
+
+    *)
+        echo "Usage: $0 {check|upload|download|key|path}" >&2
+        echo "" >&2
+        echo "Commands:" >&2
+        echo "  check    - Check if cache exists, outputs 'hit' or 'miss'" >&2
+        echo "  upload   - Upload wheels from artifacts/rocm-base-wheels/ to cache" >&2
+        echo "  download - Download wheels from cache to artifacts/rocm-base-wheels/" >&2
+        echo "  key      - Output the cache key" >&2
+        echo "  path     - Output the full S3 cache path" >&2
+        exit 1
+        ;;
+esac

.buildkite/scripts/cherry-pick-from-milestone.sh

@@ -0,0 +1,242 @@
+#!/bin/bash
+#
+# cherry-pick-from-milestone.sh
+# Find commits from a GitHub milestone that are missing from the current branch
+# and output them in chronological order for cherry-picking.
+#
+# Usage: ./cherry-pick-from-milestone.sh <milestone> [--dry-run] [--execute]
+#
+
+set -euo pipefail
+
+# Colors for output
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+NC='\033[0m' # No Color
+
+usage() {
+    cat <<EOF
+Usage: $(basename "$0") <milestone> [options]
+
+Find commits from a GitHub milestone that need to be cherry-picked into the current branch.
+
+Arguments:
+    milestone       The GitHub milestone name (e.g., v0.14.0)
+
+Options:
+    --dry-run       Show the cherry-pick commands without executing (default)
+    --execute       Actually execute the cherry-picks
+    --main-branch   Specify the main branch name (default: main)
+    --help          Show this help message
+
+Examples:
+    $(basename "$0") v0.14.0
+    $(basename "$0") v0.14.0 --dry-run
+    $(basename "$0") v0.14.0 --execute
+    $(basename "$0") v0.14.0 --main-branch master
+EOF
+    exit 1
+}
+
+log_info() {
+    echo -e "${BLUE}[INFO]${NC} $1"
+}
+
+log_success() {
+    echo -e "${GREEN}[OK]${NC} $1"
+}
+
+log_warn() {
+    echo -e "${YELLOW}[WARN]${NC} $1"
+}
+
+log_error() {
+    echo -e "${RED}[ERROR]${NC} $1" >&2
+}
+
+# Default values
+MILESTONE=""
+DRY_RUN=true
+MAIN_BRANCH="main"
+
+# Parse arguments
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --dry-run)
+            DRY_RUN=true
+            shift
+            ;;
+        --execute)
+            DRY_RUN=false
+            shift
+            ;;
+        --main-branch)
+            MAIN_BRANCH="$2"
+            shift 2
+            ;;
+        --help|-h)
+            usage
+            ;;
+        -*)
+            log_error "Unknown option: $1"
+            usage
+            ;;
+        *)
+            if [[ -z "$MILESTONE" ]]; then
+                MILESTONE="$1"
+            else
+                log_error "Unexpected argument: $1"
+                usage
+            fi
+            shift
+            ;;
+    esac
+done
+
+# Validate milestone argument
+if [[ -z "$MILESTONE" ]]; then
+    log_error "Milestone is required"
+    usage
+fi
+
+# Check if we're in a git repository
+if ! git rev-parse --is-inside-work-tree &>/dev/null; then
+    log_error "Not in a git repository"
+    exit 1
+fi
+
+# Check if gh CLI is available
+if ! command -v gh &>/dev/null; then
+    log_error "GitHub CLI (gh) is not installed"
+    exit 1
+fi
+
+# Check if authenticated with gh
+if ! gh auth status &>/dev/null; then
+    log_error "Not authenticated with GitHub CLI. Run 'gh auth login' first."
+    exit 1
+fi
+
+CURRENT_BRANCH=$(git branch --show-current)
+log_info "Current branch: ${CURRENT_BRANCH}"
+log_info "Main branch: ${MAIN_BRANCH}"
+log_info "Milestone: ${MILESTONE}"
+echo ""
+
+# Fetch latest from remote
+log_info "Fetching latest from remote..."
+git fetch origin "$MAIN_BRANCH" --quiet
+
+# Get merged PRs from the milestone, sorted by merge date
+log_info "Fetching merged PRs from milestone '${MILESTONE}'..."
+
+# Store PR data in a temp file
+PR_DATA=$(mktemp)
+trap "rm -f $PR_DATA" EXIT
+
+if ! gh pr list --state merged --search "milestone:${MILESTONE}" \
+    --limit 1000 \
+    --json number,title,mergeCommit,mergedAt \
+    --jq 'sort_by(.mergedAt) | .[] | "\(.mergeCommit.oid)\t\(.number)\t\(.title)"' > "$PR_DATA" 2>/dev/null; then
+    log_error "Failed to fetch PRs from milestone '${MILESTONE}'"
+    log_error "This could be due to:"
+    log_error "  - Milestone does not exist"
+    log_error "  - Network/authentication issues"
+    log_error "  - Invalid milestone name format"
+    exit 1
+fi
+
+if [[ ! -s "$PR_DATA" ]]; then
+    log_warn "No merged PRs found for milestone '${MILESTONE}'"
+    exit 0
+fi
+
+TOTAL_PRS=$(wc -l < "$PR_DATA")
+log_info "Found ${TOTAL_PRS} merged PR(s) in milestone"
+echo ""
+
+# Find commits that are missing from current branch
+MISSING_COMMITS=()
+MISSING_INFO=()
+
+while IFS=$'\t' read -r sha pr_number title; do
+    # Skip if SHA is empty or null
+    if [[ -z "$sha" || "$sha" == "null" ]]; then
+        log_warn "PR #${pr_number} has no merge commit SHA, skipping"
+        continue
+    fi
+    
+    # Check if this commit is already in the current branch
+    if git merge-base --is-ancestor "$sha" HEAD 2>/dev/null; then
+        log_success "PR #${pr_number} already in branch: ${title:0:60}"
+    else
+        log_warn "PR #${pr_number} MISSING: ${title:0:60}"
+        MISSING_COMMITS+=("$sha")
+        MISSING_INFO+=("$sha PR #${pr_number}: ${title}")
+    fi
+done < "$PR_DATA"
+
+echo ""
+
+if [[ ${#MISSING_COMMITS[@]} -eq 0 ]]; then
+    log_success "All PRs from milestone '${MILESTONE}' are already in the current branch!"
+    exit 0
+fi
+
+log_info "Found ${#MISSING_COMMITS[@]} missing commit(s) to cherry-pick"
+echo ""
+
+# Output the cherry-pick commands
+echo "=========================================="
+echo "Cherry-pick commands (in chronological order):"
+echo "=========================================="
+echo ""
+
+for info in "${MISSING_INFO[@]}"; do
+    echo "# $info"
+done
+echo ""
+
+echo "# Run these commands to cherry-pick all missing commits:"
+echo "git cherry-pick ${MISSING_COMMITS[*]}"
+echo ""
+
+# Or one by one
+echo "# Or cherry-pick one at a time:"
+for sha in "${MISSING_COMMITS[@]}"; do
+    echo "git cherry-pick $sha"
+done
+echo ""
+
+# Execute if requested
+if [[ "$DRY_RUN" == false ]]; then
+    echo "=========================================="
+    log_info "Executing cherry-picks..."
+    echo "=========================================="
+    
+    for i in "${!MISSING_COMMITS[@]}"; do
+        sha="${MISSING_COMMITS[$i]}"
+        info="${MISSING_INFO[$i]}"
+        
+        echo ""
+        log_info "Cherry-picking: $info"
+        
+        if git cherry-pick "$sha"; then
+            log_success "Successfully cherry-picked $sha"
+        else
+            log_error "Failed to cherry-pick $sha"
+            log_error "Resolve conflicts and run 'git cherry-pick --continue', or 'git cherry-pick --abort' to cancel"
+            exit 1
+        fi
+    done
+    
+    echo ""
+    log_success "All cherry-picks completed successfully!"
+else
+    echo "=========================================="
+    echo -e "${YELLOW}Dry run mode - no changes made${NC}"
+    echo "Run with --execute to perform the cherry-picks"
+    echo "=========================================="
+fi

.buildkite/scripts/cleanup-nightly-builds.sh

@@ -3,7 +3,14 @@
 set -ex
 
 # Clean up old nightly builds from DockerHub, keeping only the last 14 builds
-# This script uses DockerHub API to list and delete old tags with "nightly-" prefix
+# This script uses DockerHub API to list and delete old tags with specified prefix
+# Usage: cleanup-nightly-builds.sh [TAG_PREFIX]
+# Example: cleanup-nightly-builds.sh "nightly-" or cleanup-nightly-builds.sh "cu130-nightly-"
+
+# Get tag prefix from argument, default to "nightly-" if not provided
+TAG_PREFIX="${1:-nightly-}"
+
+echo "Cleaning up tags with prefix: $TAG_PREFIX"
 
 # DockerHub API endpoint for vllm/vllm-openai repository
 REPO_API_URL="https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags"
@@ -45,7 +52,7 @@ get_all_tags() {
         set -x
         
         # Get both last_updated timestamp and tag name, separated by |
-        local tags=$(echo "$response" | jq -r '.results[] | select(.name | startswith("nightly-")) | "\(.last_updated)|\(.name)"')
+        local tags=$(echo "$response" | jq -r --arg prefix "$TAG_PREFIX" '.results[] | select(.name | startswith($prefix)) | "\(.last_updated)|\(.name)"')
         
         if [ -z "$tags" ]; then
             break

.buildkite/scripts/generate-nightly-index.py

@@ -16,6 +16,18 @@ from urllib.parse import quote
 
 import regex as re
 
+
+def normalize_package_name(name: str) -> str:
+    """
+    Normalize package name according to PEP 503.
+    https://peps.python.org/pep-0503/#normalized-names
+
+    Replace runs of underscores, hyphens, and periods with a single hyphen,
+    and lowercase the result.
+    """
+    return re.sub(r"[-_.]+", "-", name).lower()
+
+
 if not sys.version_info >= (3, 12):
     raise RuntimeError("This script requires Python 3.12 or higher.")
 
@@ -78,7 +90,13 @@ def parse_from_filename(file: str) -> WheelFileInfo:
             version = version.removesuffix("." + variant)
     else:
         if "+" in version:
-            version, variant = version.split("+")
+            version_part, suffix = version.split("+", 1)
+            # Only treat known patterns as variants (rocmXXX, cuXXX, cpu)
+            # Git hashes and other suffixes are NOT variants
+            if suffix.startswith(("rocm", "cu", "cpu")):
+                variant = suffix
+                version = version_part
+            # Otherwise keep the full version string (variant stays None)
 
     return WheelFileInfo(
         package_name=package_name,
@@ -94,7 +112,7 @@ def parse_from_filename(file: str) -> WheelFileInfo:
 
 def generate_project_list(subdir_names: list[str], comment: str = "") -> str:
     """
-    Generate project list HTML content linking to each project & variant sub-directory.
+    Generate project list HTML content linking to each project & variant subdirectory.
     """
     href_tags = []
     for name in sorted(subdir_names):
@@ -150,23 +168,23 @@ def generate_index_and_metadata(
         comment (str | None): Optional comment to include in the generated HTML files.
 
     First, parse all wheel files to extract metadata.
-    We need to collect all wheel files for each variant, and generate an index for it (in a sub-directory).
+    We need to collect all wheel files for each variant, and generate an index for it (in a subdirectory).
     The index for the default variant (if any) is generated in the root index directory.
 
     If `default_variant` is provided, all wheels must have variant suffixes, and the default variant index
     is purely a copy of the corresponding variant index, with only the links adjusted.
     Otherwise, all wheels without variant suffixes are treated as the default variant.
 
-    If `alias_to_default` is provided, an additional alias sub-directory is created, it has the same content
+    If `alias_to_default` is provided, an additional alias subdirectory is created, it has the same content
     as the default variant index, but the links are adjusted accordingly.
 
     Index directory structure:
         index_base_dir/ (hosted at wheels.vllm.ai/{nightly,$commit,$version}/)
-            index.html  # project list, linking to "vllm/" and other packages, and all variant sub-directories
+            index.html  # project list, linking to "vllm/" and other packages, and all variant subdirectories
             vllm/
                 index.html # package index, pointing to actual files in wheel_base_dir (relative path)
                 metadata.json # machine-readable metadata for all wheels in this package
-            cpu/ # cpu variant sub-directory
+            cpu/ # cpu variant subdirectory
                 index.html
                 vllm/
                     index.html
@@ -176,7 +194,7 @@ def generate_index_and_metadata(
                 vllm/
                     index.html
                     metadata.json
-            cu130/ # cu130 variant sub-directory
+            cu130/ # cu130 variant subdirectory
                 index.html
                 vllm/
                     index.html
@@ -206,6 +224,26 @@ def generate_index_and_metadata(
         print("No wheel files found, skipping index generation.")
         return
 
+    # For ROCm builds: inherit variant from vllm wheel
+    # All ROCm wheels should share the same variant as vllm
+    rocm_variant = None
+    for file in parsed_files:
+        if (
+            file.package_name == "vllm"
+            and file.variant
+            and file.variant.startswith("rocm")
+        ):
+            rocm_variant = file.variant
+            print(f"Detected ROCm variant from vllm: {rocm_variant}")
+            break
+
+    # Apply ROCm variant to all wheels without a variant
+    if rocm_variant:
+        for file in parsed_files:
+            if file.variant is None:
+                file.variant = rocm_variant
+                print(f"Inherited variant '{rocm_variant}' for {file.filename}")
+
     # Group by variant
     variant_to_files: dict[str, list[WheelFileInfo]] = {}
     for file in parsed_files:
@@ -256,8 +294,8 @@ def generate_index_and_metadata(
 
         variant_dir.mkdir(parents=True, exist_ok=True)
 
-        # gather all package names in this variant
-        packages = set(f.package_name for f in files)
+        # gather all package names in this variant (normalized per PEP 503)
+        packages = set(normalize_package_name(f.package_name) for f in files)
         if variant == "default":
             # these packages should also appear in the "project list"
             # generate after all variants are processed
@@ -269,8 +307,10 @@ def generate_index_and_metadata(
                 f.write(project_list_str)
 
         for package in packages:
-            # filter files belonging to this package only
-            package_files = [f for f in files if f.package_name == package]
+            # filter files belonging to this package only (compare normalized names)
+            package_files = [
+                f for f in files if normalize_package_name(f.package_name) == package
+            ]
             package_dir = variant_dir / package
             package_dir.mkdir(parents=True, exist_ok=True)
             index_str, metadata_str = generate_package_index_and_metadata(
@@ -291,6 +331,7 @@ if __name__ == "__main__":
     """
     Arguments:
         --version <version> : version string for the current build (e.g., commit hash)
+        --wheel-dir <wheel_directory> : directory containing wheel files (default to be same as `version`)
         --current-objects <path_to_json> : path to JSON file containing current S3 objects listing in this version directory
         --output-dir <output_directory> : directory to store generated index files
         --alias-to-default <alias_variant_name> : (optional) alias variant name for the default variant
@@ -318,6 +359,12 @@ if __name__ == "__main__":
         required=True,
         help="Directory to store generated index files",
     )
+    parser.add_argument(
+        "--wheel-dir",
+        type=str,
+        default=None,
+        help="Directory containing wheel files (default to be same as `version`)",
+    )
     parser.add_argument(
         "--alias-to-default",
         type=str,
@@ -334,8 +381,13 @@ if __name__ == "__main__":
     args = parser.parse_args()
 
     version = args.version
-    if "/" in version or "\\" in version:
-        raise ValueError("Version string must not contain slashes.")
+    # Allow rocm/ prefix, reject other slashes and all backslashes
+    if "\\" in version:
+        raise ValueError("Version string must not contain backslashes.")
+    if "/" in version and not version.startswith("rocm/"):
+        raise ValueError(
+            "Version string must not contain slashes (except for 'rocm/' prefix)."
+        )
     current_objects_path = Path(args.current_objects)
     output_dir = Path(args.output_dir)
     if not output_dir.exists():
@@ -372,7 +424,7 @@ if __name__ == "__main__":
 
     print(f"Found {len(wheel_files)} wheel files for version {version}: {wheel_files}")
 
-    # keep only "official" files for a non-nightly version (specifed by cli args)
+    # keep only "official" files for a non-nightly version (specified by cli args)
     PY_VERSION_RE = re.compile(r"^\d+\.\d+\.\d+([a-zA-Z0-9.+-]*)?$")
     if PY_VERSION_RE.match(version):
         # upload-wheels.sh ensures no "dev" is in args.version
@@ -384,9 +436,25 @@ if __name__ == "__main__":
         print("Nightly version detected, keeping all wheel files.")
 
     # Generate index and metadata, assuming wheels and indices are stored as:
-    # s3://vllm-wheels/{version}/<wheel files>
+    # s3://vllm-wheels/{wheel_dir}/<wheel files>
     # s3://vllm-wheels/<anything>/<index files>
-    wheel_base_dir = Path(output_dir).parent / version
+    #
+    # For ROCm builds, version is "rocm/{commit}" and indices are uploaded to:
+    #   - rocm/{commit}/  (same as wheels)
+    #   - rocm/nightly/
+    #   - rocm/{version}/
+    # All these are under the "rocm/" prefix, so relative paths should be
+    # relative to "rocm/", not the bucket root.
+    if args.wheel_dir:
+        # Explicit wheel-dir provided (e.g., for version-specific indices pointing to commit dir)
+        wheel_dir = args.wheel_dir.strip().rstrip("/")
+    elif version.startswith("rocm/"):
+        # For rocm/commit, wheel_base_dir should be just the commit part
+        # so relative path from rocm/0.12.0/rocm710/vllm/ -> ../../../{commit}/
+        wheel_dir = version.split("/", 1)[1]
+    else:
+        wheel_dir = version
+    wheel_base_dir = Path(output_dir).parent / wheel_dir
     index_base_dir = Path(output_dir)
 
     generate_index_and_metadata(

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -44,6 +44,17 @@ cleanup_docker() {
   fi
 }
 
+cleanup_network() {
+  for node in $(seq 0 $((NUM_NODES-1))); do
+    if docker pr -a -q -f name="node${node}" | grep -q .; then
+      docker stop "node${node}"
+    fi
+  done
+  if docker network ls | grep docker-net; then
+    docker network rm docker-net
+  fi
+}
+
 # Call the cleanup docker function
 cleanup_docker
 
@@ -76,7 +87,7 @@ mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
 
 commands=$@
-echo "Commands:$commands"
+echo "Raw commands: $commands"
 
 commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"pytest -v -s basic_correctness/test_basic_correctness.py"}
 
@@ -141,7 +152,6 @@ if [[ $commands == *" entrypoints/openai "* ]]; then
   --ignore=entrypoints/openai/test_audio.py \
   --ignore=entrypoints/openai/test_shutdown.py \
   --ignore=entrypoints/openai/test_completion.py \
-  --ignore=entrypoints/openai/test_sleep.py \
   --ignore=entrypoints/openai/test_models.py \
   --ignore=entrypoints/openai/test_lora_adapters.py \
   --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
@@ -159,6 +169,9 @@ if [[ $commands == *" entrypoints/llm "* ]]; then
   --ignore=entrypoints/llm/test_prompt_validation.py "}
 fi
 
+commands=$(echo "$commands" | sed 's/ \\ / /g')
+echo "Final commands: $commands"
+
 # --ignore=entrypoints/openai/test_encoder_decoder.py \
 # --ignore=entrypoints/openai/test_embedding.py \
 # --ignore=entrypoints/openai/test_oot_registration.py
@@ -166,7 +179,6 @@ fi
 # --ignore=entrypoints/openai/test_models.py <= Fails on MI250 but passes on MI300 as of 2025-03-13
 
 
-PARALLEL_JOB_COUNT=8
 MYPYTHONPATH=".."
 
 # Test that we're launching on the machine that has
@@ -177,45 +189,34 @@ if [[ -z "$render_gid" ]]; then
   exit 1
 fi
 
-# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
-if [[ $commands == *"--shard-id="* ]]; then
-  # assign job count as the number of shards used
-  commands=$(echo "$commands" | sed -E "s/--num-shards[[:blank:]]*=[[:blank:]]*[0-9]*/--num-shards=${PARALLEL_JOB_COUNT} /g" | sed 's/ \\ / /g')
-  for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
-    # assign shard-id for each shard
-    commands_gpu=$(echo "$commands" | sed -E "s/--shard-id[[:blank:]]*=[[:blank:]]*[0-9]*/--shard-id=${GPU} /g" | sed 's/ \\ / /g')
-    echo "Shard ${GPU} commands:$commands_gpu"
-    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
-    docker run \
-        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
-        --network=host \
-        --shm-size=16gb \
-        --group-add "$render_gid" \
-        --rm \
-        -e HIP_VISIBLE_DEVICES="${GPU}" \
-        -e HF_TOKEN \
-        -e AWS_ACCESS_KEY_ID \
-        -e AWS_SECRET_ACCESS_KEY \
-        -v "${HF_CACHE}:${HF_MOUNT}" \
-        -e "HF_HOME=${HF_MOUNT}" \
-        -e "PYTHONPATH=${MYPYTHONPATH}" \
-        --name "${container_name}_${GPU}" \
-        "${image_name}" \
-        /bin/bash -c "${commands_gpu}" \
-        |& while read -r line; do echo ">>Shard $GPU: $line"; done &
-    PIDS+=($!)
-  done
-  #wait for all processes to finish and collect exit codes
-  for pid in "${PIDS[@]}"; do
-    wait "${pid}"
-    STATUS+=($?)
-  done
-  for st in "${STATUS[@]}"; do
-    if [[ ${st} -ne 0 ]]; then
-      echo "One of the processes failed with $st"
-      exit "${st}"
-    fi
-  done
+if [[ $commands == *"VLLM_TEST_GROUP_NAME=mi325_4-2-node-tests-4-gpus-in-total"* ]]; then
+
+  export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
+
+  if [[ "$commands" =~ ^(.*)"["(.*)"] && ["(.*)"]"$ ]]; then
+      prefix=$( echo "${BASH_REMATCH[1]}" | sed 's/;//g')
+      echo "PREFIX: ${prefix}"
+      export composite_command="(command rocm-smi || true)"
+      myIFS=$IFS
+      IFS=','
+      read -ra node0 <<< ${BASH_REMATCH[2]}
+      read -ra node1 <<< ${BASH_REMATCH[3]}
+      IFS=$myIFS
+      for i in "${!node0[@]}";do 
+        command_node_0=$(echo ${node0[i]} | sed 's/\"//g')
+        command_node_1=$(echo ${node1[i]} | sed 's/\"//g')
+        
+        export commands="./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 ${image_name} '${command_node_0}' '${command_node_1}'"
+        echo "COMMANDS: ${commands}"
+        composite_command=$(echo "${composite_command} && ${commands}")
+      done
+      /bin/bash -c "${composite_command}"
+      cleanup_network
+  else
+      echo "Failed to parse node commands! Exiting."
+      cleanup_network
+      exit 111
+  fi
 else
   echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \

.buildkite/scripts/hardware_ci/run-cpu-test.sh

@@ -50,6 +50,7 @@ function cpu_tests() {
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
+    pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
     pytest -x -v -s tests/kernels/test_onednn.py"
 
   # Run basic model test
@@ -83,7 +84,7 @@ function cpu_tests() {
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -x -s -v \
-    tests/lora/test_qwen2vl.py"
+    tests/lora/test_qwenvl.py"
 
   # online serving: tp+pp
   docker exec cpu-test-"$NUMA_NODE" bash -c '

.buildkite/scripts/hardware_ci/run-hpu-test.sh

@@ -5,7 +5,9 @@
 set -exuo pipefail
 
 # Try building the docker image
-cat <<EOF | docker build -t hpu-plugin-v1-test-env -f - .
+image_name="hpu/upstream-vllm-ci:${BUILDKITE_COMMIT}"
+container_name="hpu-upstream-vllm-ci-${BUILDKITE_COMMIT}-container"
+cat <<EOF | docker build -t ${image_name} -f - .
 FROM gaudi-base-image:latest
 
 COPY ./ /workspace/vllm
@@ -15,7 +17,8 @@ WORKDIR /workspace/vllm
 ENV no_proxy=localhost,127.0.0.1
 ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
 
-RUN VLLM_TARGET_DEVICE=empty pip install .
+RUN bash -c 'pip install -r <(sed "/^torch/d" requirements/build.txt)'
+RUN VLLM_TARGET_DEVICE=empty pip install --no-build-isolation -e .
 RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
 
 # install development dependencies (for testing)
@@ -36,15 +39,20 @@ EOF
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_containers() { docker rm -f hpu-plugin-v1-test || true; }
+remove_docker_containers() { docker rm -f ${container_name} || true; }
 trap 'remove_docker_containers; exit $EXITCODE;' EXIT
 remove_docker_containers
 
 echo "Running HPU plugin v1 test"
-docker run --rm --runtime=habana --name=hpu-plugin-v1-test --network=host \
+docker run --rm --runtime=habana --name=${container_name} --network=host \
   -e HABANA_VISIBLE_DEVICES=all \
-  hpu-plugin-v1-test-env \
-  /bin/bash "/workspace/vllm-gaudi/tests/upstream_tests/ci_tests.sh"
+  -e VLLM_SKIP_WARMUP=true \
+  -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \
+  -e PT_HPU_LAZY_MODE=1 \
+  "${image_name}" \
+  /bin/bash -c '
+  cd vllm; timeout 120s python -u examples/offline_inference/basic/generate.py --model facebook/opt-125m
+'
 
 EXITCODE=$?
 if [ $EXITCODE -eq 0 ]; then

.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh

@@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
+    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.9.2" \
     && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
+    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.9.2" \
     && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -38,15 +38,16 @@ docker run \
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    python3 examples/offline_inference/basic/generate.py --model Intel/Qwen2.5-0.5B-W4A16-G128-AutoRound-LLMC-TEST-ONLY --enforce-eager
-    VLLM_ATTENTION_BACKEND=TRITON_ATTN python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
+    python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
+    python3 examples/offline_inference/basic/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
     cd tests
-    pytest -v -s v1/core
+    pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py
     pytest -v -s v1/engine
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
     pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py --ignore=v1/spec_decode/test_speculators_eagle3.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py
     pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py
     pytest -v -s v1/test_serial_utils.py
 '

.buildkite/scripts/push-nightly-builds.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+
+set -ex
+
+# Get tag variant from argument, default to empty if not provided, should be something like "cu130".
+# Due to limits in cleanup script, we must move variants to use separate tags like "cu130-nightly",
+# otherwise they will be cleaned up together with the main "nightly" tags.
+
+TAG_VARIANT="$1"
+if [ -n "$TAG_VARIANT" ]; then
+    ORIG_TAG_SUFFIX="-$TAG_VARIANT"
+    TAG_NAME="$TAG_VARIANT-nightly"
+else
+    ORIG_TAG_SUFFIX=""
+    TAG_NAME="nightly"
+fi
+
+ORIG_TAG_NAME="$BUILDKITE_COMMIT"
+
+echo "Pushing original tag $ORIG_TAG_NAME$ORIG_TAG_SUFFIX to new nightly tag name: $TAG_NAME"
+
+# pull original arch-dependent images from AWS ECR Public
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-x86_64$ORIG_TAG_SUFFIX
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-aarch64$ORIG_TAG_SUFFIX
+# tag arch-dependent images
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-x86_64$ORIG_TAG_SUFFIX vllm/vllm-openai:$TAG_NAME-x86_64
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-aarch64$ORIG_TAG_SUFFIX vllm/vllm-openai:$TAG_NAME-aarch64
+# push arch-dependent images to DockerHub
+docker push vllm/vllm-openai:$TAG_NAME-x86_64
+docker push vllm/vllm-openai:$TAG_NAME-aarch64
+# push arch-independent manifest to DockerHub
+docker manifest create vllm/vllm-openai:$TAG_NAME vllm/vllm-openai:$TAG_NAME-x86_64 vllm/vllm-openai:$TAG_NAME-aarch64 --amend
+docker manifest create vllm/vllm-openai:$TAG_NAME-$BUILDKITE_COMMIT vllm/vllm-openai:$TAG_NAME-x86_64 vllm/vllm-openai:$TAG_NAME-aarch64 --amend
+docker manifest push vllm/vllm-openai:$TAG_NAME
+docker manifest push vllm/vllm-openai:$TAG_NAME-$BUILDKITE_COMMIT

.buildkite/scripts/run-multi-node-test.sh

@@ -2,6 +2,17 @@
 
 set -euox pipefail
 
+# To detect ROCm
+# Check multiple indicators:
+if [ -e /dev/kfd ] || \
+    [ -d /opt/rocm ] || \
+    command -v rocm-smi &> /dev/null || \
+    [ -n "${ROCM_HOME:-}" ]; then
+    IS_ROCM=1
+else
+    IS_ROCM=0
+fi
+
 if [[ $# -lt 4 ]]; then
     echo "Usage: .buildkite/scripts/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
     exit 1
@@ -26,13 +37,18 @@ for command in "${COMMANDS[@]}"; do
     echo "$command"
 done
 
+
 start_network() {
     docker network create --subnet=192.168.10.0/24 docker-net
 }
 
 start_nodes() {
     for node in $(seq 0 $(($NUM_NODES-1))); do
-        GPU_DEVICES='"device='
+        if [ "$IS_ROCM" -eq 1 ]; then
+            GPU_DEVICES='--device /dev/kfd --device /dev/dri -e HIP_VISIBLE_DEVICES='
+        else
+            GPU_DEVICES='--gpus "device='
+        fi
         for node_gpu in $(seq 0 $(($NUM_GPUS - 1))); do
             DEVICE_NUM=$(($node * $NUM_GPUS + $node_gpu))
             GPU_DEVICES+=$(($DEVICE_NUM))
@@ -40,7 +56,9 @@ start_nodes() {
                 GPU_DEVICES+=','
             fi
         done
-        GPU_DEVICES+='"'
+        if [ "$IS_ROCM" -eq 0 ]; then
+            GPU_DEVICES+='"'
+        fi
 
         # start the container in detached mode
         # things to note:
@@ -49,7 +67,7 @@ start_nodes() {
         # 3. map the huggingface cache directory to the container
         # 3. assign ip addresses to the containers (head node: 192.168.10.10, worker nodes:
         #    starting from 192.168.10.11)
-        docker run -d --gpus "$GPU_DEVICES" --shm-size=10.24gb -e HF_TOKEN \
+        docker run -d $GPU_DEVICES --shm-size=10.24gb -e HF_TOKEN \
             -v ~/.cache/huggingface:/root/.cache/huggingface --name "node$node" \
             --network docker-net --ip 192.168.10.$((10 + $node)) --rm "$DOCKER_IMAGE" \
             /bin/bash -c "tail -f /dev/null"

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh

@@ -43,7 +43,6 @@ trap cleanup EXIT
 
 for BACK in "${BACKENDS[@]}"; do
   VLLM_DEEP_GEMM_WARMUP=skip \
-  VLLM_ALL2ALL_BACKEND=$BACK \
   vllm serve "$MODEL" \
     --enforce-eager \
     --tensor-parallel-size 2 \
@@ -52,6 +51,7 @@ for BACK in "${BACKENDS[@]}"; do
     --enable-eplb \
     --trust-remote-code \
     --max-model-len 2048 \
+    --all2all-backend $BACK \
     --port $PORT &
   SERVER_PID=$!
   wait_for_server $PORT

.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh

@@ -44,10 +44,10 @@ trap cleanup EXIT
 
 for BACK in "${BACKENDS[@]}"; do
   VLLM_DEEP_GEMM_WARMUP=skip \
-  VLLM_ALL2ALL_BACKEND=$BACK \
   vllm serve "$MODEL" \
     --enforce-eager \
     --enable-eplb \
+    --all2all-backend $BACK \
     --eplb-config '{"window_size":10, "step_interval":100, "num_redundant_experts":0, "log_balancedness":true}' \
     --tensor-parallel-size ${TENSOR_PARALLEL_SIZE} \
     --data-parallel-size ${DATA_PARALLEL_SIZE} \

.buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh

@@ -18,15 +18,18 @@ wait_for_server() {
 
 MODEL="Qwen/Qwen3-Next-80B-A3B-Instruct"
 
-# Set BACKENDS based on platform
+# Set BACKENDS and platform-specific args based on platform
 if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
   # ROCm platform
   BACKENDS=("allgather_reducescatter")
   # Disable MOE padding for ROCm since it is causing eplb to fail
   export VLLM_ROCM_MOE_PADDING=0
+  PLATFORM_ARGS=("--no-async-scheduling")
+  echo "Disabled async scheduling for ROCm platform due to issues with spec decode."
 else
   # Non-ROCm platform (CUDA/other)
   BACKENDS=("deepep_high_throughput" "deepep_low_latency")
+  PLATFORM_ARGS=()
 fi
 
 cleanup() {
@@ -43,17 +46,18 @@ trap cleanup EXIT
 
 for BACK in "${BACKENDS[@]}"; do
   VLLM_DEEP_GEMM_WARMUP=skip \
-  VLLM_ALL2ALL_BACKEND=$BACK \
   vllm serve "$MODEL" \
     --enforce-eager \
     --tensor-parallel-size 4 \
     --enable-expert-parallel \
     --enable-eplb \
+    --all2all-backend $BACK \
     --eplb-config '{"window_size":200,"step_interval":600,"use_async":true}' \
     --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}' \
     --trust-remote-code \
     --max-model-len 2048 \
     --gpu-memory-utilization 0.9 \
+    "${PLATFORM_ARGS[@]}" \
     --port $PORT &
   SERVER_PID=$!
   wait_for_server $PORT

.buildkite/scripts/trigger-ci-build.sh

@@ -0,0 +1,227 @@
+#!/bin/bash
+#
+# trigger-ci-build.sh
+# Trigger a Buildkite CI build using the bk CLI for the current commit and branch
+# with RUN_ALL=1 and NIGHTLY=1 environment variables.
+#
+# Usage: ./trigger-ci-build.sh [options]
+#
+# Requires: bk CLI (https://buildkite.com/docs/platform/cli)
+#
+# SAFETY: Dry-run by default. Use --execute to actually trigger a build.
+#
+
+set -euo pipefail
+
+# Colors for output
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+NC='\033[0m' # No Color
+
+# Default configuration
+PIPELINE="ci"
+DRY_RUN=true
+
+usage() {
+    cat <<EOF
+Usage: $(basename "$0") [options]
+
+Trigger a Buildkite CI build using the bk CLI for the current commit and branch.
+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
+
+Prerequisites:
+    - bk CLI installed: brew tap buildkite/buildkite && brew install buildkite/buildkite/bk
+    - bk configured: bk configure
+
+Examples:
+    $(basename "$0")                        # Dry-run, show what would happen
+    $(basename "$0") --execute              # Actually trigger the build
+    $(basename "$0") --pipeline ci-shadow   # Dry-run with different pipeline
+EOF
+    exit 1
+}
+
+log_info() {
+    echo -e "${BLUE}[INFO]${NC} $1"
+}
+
+log_success() {
+    echo -e "${GREEN}[OK]${NC} $1"
+}
+
+log_warn() {
+    echo -e "${YELLOW}[WARN]${NC} $1"
+}
+
+log_error() {
+    echo -e "${RED}[ERROR]${NC} $1" >&2
+}
+
+# Parse arguments
+COMMIT=""
+BRANCH=""
+MESSAGE=""
+
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --execute)
+            DRY_RUN=false
+            shift
+            ;;
+        --pipeline)
+            PIPELINE="$2"
+            shift 2
+            ;;
+        --commit)
+            COMMIT="$2"
+            shift 2
+            ;;
+        --branch)
+            BRANCH="$2"
+            shift 2
+            ;;
+        --message)
+            MESSAGE="$2"
+            shift 2
+            ;;
+        --help|-h)
+            usage
+            ;;
+        -*)
+            log_error "Unknown option: $1"
+            usage
+            ;;
+        *)
+            log_error "Unexpected argument: $1"
+            usage
+            ;;
+    esac
+done
+
+# Check if bk CLI is installed
+if ! command -v bk &>/dev/null; then
+    log_error "Buildkite CLI (bk) is not installed"
+    echo ""
+    echo "Install with:"
+    echo "  brew tap buildkite/buildkite && brew install buildkite/buildkite/bk"
+    echo ""
+    echo "Then configure:"
+    echo "  bk configure"
+    exit 1
+fi
+
+# Check if we're in a git repository
+if ! git rev-parse --is-inside-work-tree &>/dev/null; then
+    log_error "Not in a git repository"
+    exit 1
+fi
+
+# Get current commit and branch if not overridden
+if [[ -z "$COMMIT" ]]; then
+    COMMIT=$(git rev-parse HEAD)
+fi
+
+if [[ -z "$BRANCH" ]]; then
+    BRANCH=$(git branch --show-current)
+    if [[ -z "$BRANCH" ]]; then
+        # Detached HEAD state - try to get branch from ref
+        BRANCH=$(git rev-parse --abbrev-ref HEAD)
+    fi
+fi
+
+# Generate default message if not provided
+if [[ -z "$MESSAGE" ]]; then
+    COMMIT_MSG=$(git log -1 --pretty=format:"%s" "$COMMIT" 2>/dev/null || echo "Manual build")
+    MESSAGE="[Manual] ${COMMIT_MSG}"
+fi
+
+# Safety check: Verify the commit exists on the remote
+log_info "Verifying commit exists on remote..."
+git fetch origin --quiet 2>/dev/null || true
+
+# Check if commit is reachable from any remote branch
+REMOTE_BRANCHES=$(git branch -r --contains "$COMMIT" 2>/dev/null || true)
+if [[ -z "$REMOTE_BRANCHES" ]]; then
+    log_error "Commit ${COMMIT} does not exist on any remote branch!"
+    echo ""
+    echo "The CI system will fail to checkout this commit."
+    echo "Please push your changes first:"
+    echo ""
+    echo "  git push origin ${BRANCH}"
+    echo ""
+    exit 1
+fi
+
+log_success "Commit found on remote branches:"
+echo "$REMOTE_BRANCHES" | head -5 | sed 's/^/  /'
+if [[ $(echo "$REMOTE_BRANCHES" | wc -l) -gt 5 ]]; then
+    echo "  ... and more"
+fi
+echo ""
+
+log_info "Pipeline: ${PIPELINE}"
+log_info "Branch: ${BRANCH}"
+log_info "Commit: ${COMMIT}"
+log_info "Message: ${MESSAGE}"
+log_info "Environment: RUN_ALL=1, NIGHTLY=1"
+echo ""
+
+# Build the command
+CMD=(bk build create
+    -y
+    -w
+    -i
+    --pipeline "${PIPELINE}"
+    --commit "${COMMIT}"
+    --branch "${BRANCH}"
+    --message "${MESSAGE}"
+    --env "RUN_ALL=1"
+    --env "NIGHTLY=1"
+)
+
+if [[ "$DRY_RUN" == true ]]; then
+    echo "=========================================="
+    log_warn "DRY-RUN MODE - No build will be triggered"
+    echo "=========================================="
+    echo ""
+    echo "Command that would be executed:"
+    echo ""
+    # Escape single quotes in values for safe shell display
+    escape_for_shell() {
+        printf '%s' "$1" | sed "s/'/'\\\\''/g"
+    }
+    echo "  bk build create \\"
+    echo "    -y \\"
+    echo "    -w \\"
+    echo "    -i \\"
+    echo "    --pipeline '$(escape_for_shell "${PIPELINE}")' \\"
+    echo "    --commit '$(escape_for_shell "${COMMIT}")' \\"
+    echo "    --branch '$(escape_for_shell "${BRANCH}")' \\"
+    echo "    --message '$(escape_for_shell "${MESSAGE}")' \\"
+    echo "    --env 'RUN_ALL=1' \\"
+    echo "    --env 'NIGHTLY=1'"
+    echo ""
+    echo "=========================================="
+    echo -e "${YELLOW}To actually trigger this build, run:${NC}"
+    echo ""
+    echo "  $0 --execute"
+    echo "=========================================="
+    exit 0
+fi
+
+log_info "Triggering build..."
+
+# Execute the command - bk will print the URL and open browser
+"${CMD[@]}"

.buildkite/scripts/upload-nightly-wheels.sh

@@ -102,6 +102,7 @@ if [[ "$version" != *"dev"* ]]; then
     echo "Re-generating indices for /$pure_version/"
     rm -rf "$INDICES_OUTPUT_DIR/*"
     mkdir -p "$INDICES_OUTPUT_DIR"
-    $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$pure_version" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "version $pure_version" $alias_arg
+    # wheel-dir is overridden to be the commit directory, so that the indices point to the correct wheel path
+    $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$pure_version" --wheel-dir "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "version $pure_version" $alias_arg
     aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/$pure_version/"
 fi

.buildkite/scripts/upload-release-wheels-pypi.sh

@@ -0,0 +1,70 @@
+#!/usr/bin/env bash
+
+set -e
+
+BUCKET="vllm-wheels"
+SUBPATH=$BUILDKITE_COMMIT
+S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
+
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version)
+GIT_VERSION=$(git describe --exact-match --tags $BUILDKITE_COMMIT 2>/dev/null)
+
+echo "Release version from Buildkite: $RELEASE_VERSION"
+
+if [[ -z "$GIT_VERSION" ]]; then
+    echo "[FATAL] Not on a git tag, cannot create release."
+    exit 1
+else
+    echo "Git version for commit $BUILDKITE_COMMIT: $GIT_VERSION"
+fi
+# sanity check for version mismatch
+if [[ "$RELEASE_VERSION" != "$GIT_VERSION" ]]; then
+  if [[ "$FORCE_RELEASE_IGNORE_VERSION_MISMATCH" == "true" ]]; then
+    echo "[WARNING] Force release and ignore version mismatch"
+  else
+    echo "[FATAL] Release version from Buildkite does not match Git version."
+    exit 1
+  fi
+fi
+PURE_VERSION=${RELEASE_VERSION#v} # remove leading 'v'
+
+# check pypi token
+if [[ -z "$PYPI_TOKEN" ]]; then
+  echo "[FATAL] PYPI_TOKEN is not set."
+  exit 1
+else
+  export TWINE_USERNAME="__token__"
+  export TWINE_PASSWORD="$PYPI_TOKEN"
+fi
+
+set -x # avoid printing secrets above
+
+# install twine from pypi
+python3 -m venv /tmp/vllm-release-env
+source /tmp/vllm-release-env/bin/activate
+pip install twine
+python3 -m twine --version
+
+# copy release wheels to local directory
+DIST_DIR=/tmp/vllm-release-dist
+echo "Existing wheels on S3:"
+aws s3 ls "$S3_COMMIT_PREFIX"
+echo "Copying wheels to local directory"
+mkdir -p $DIST_DIR
+# include only wheels for the release version, ignore all files with "dev" or "rc" in the name (without excluding 'aarch64')
+aws s3 cp --recursive --exclude "*" --include "vllm-${PURE_VERSION}*.whl" --exclude "*dev*" --exclude "*rc[0-9]*" "$S3_COMMIT_PREFIX" $DIST_DIR
+echo "Wheels copied to local directory"
+# generate source tarball
+git archive --format=tar.gz --output="$DIST_DIR/vllm-${PURE_VERSION}.tar.gz" $BUILDKITE_COMMIT
+ls -la $DIST_DIR
+
+# upload wheels to PyPI (only default variant, i.e. files without '+' in the name)
+PYPI_WHEEL_FILES=$(find $DIST_DIR -name "vllm-${PURE_VERSION}*.whl" -not -name "*+*")
+if [[ -z "$PYPI_WHEEL_FILES" ]]; then
+  echo "No default variant wheels found, quitting..."
+  exit 1
+fi
+
+python3 -m twine check $PYPI_WHEEL_FILES
+python3 -m twine upload --non-interactive --verbose $PYPI_WHEEL_FILES
+echo "Wheels uploaded to PyPI"

.buildkite/scripts/upload-rocm-wheels.sh

@@ -0,0 +1,151 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Upload ROCm wheels to S3 with proper index generation
+#
+# Required environment variables:
+#   AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY (or IAM role)
+#   S3_BUCKET (default: vllm-wheels)
+#
+# S3 path structure:
+#   s3://vllm-wheels/rocm/{commit}/     - All wheels for this commit
+#   s3://vllm-wheels/rocm/nightly/      - Index pointing to latest nightly
+#   s3://vllm-wheels/rocm/{version}/    - Index for release versions
+
+set -ex
+
+# ======== Configuration ========
+BUCKET="${S3_BUCKET:-vllm-wheels}"
+ROCM_SUBPATH="rocm/${BUILDKITE_COMMIT}"
+S3_COMMIT_PREFIX="s3://$BUCKET/$ROCM_SUBPATH/"
+INDICES_OUTPUT_DIR="rocm-indices"
+PYTHON="${PYTHON_PROG:-python3}"
+
+# ROCm uses manylinux_2_35 (Ubuntu 22.04 based)
+MANYLINUX_VERSION="manylinux_2_35"
+
+echo "========================================"
+echo "ROCm Wheel Upload Configuration"
+echo "========================================"
+echo "S3 Bucket: $BUCKET"
+echo "S3 Path: $ROCM_SUBPATH"
+echo "Commit: $BUILDKITE_COMMIT"
+echo "Branch: $BUILDKITE_BRANCH"
+echo "========================================"
+
+# ======== Part 0: Setup Python ========
+
+# Detect if python3.12+ is available
+has_new_python=$($PYTHON -c "print(1 if __import__('sys').version_info >= (3,12) else 0)" 2>/dev/null || echo 0)
+if [[ "$has_new_python" -eq 0 ]]; then
+    # Use new python from docker
+    # Use --user to ensure files are created with correct ownership (not root)
+    docker pull python:3-slim
+    PYTHON="docker run --rm --user $(id -u):$(id -g) -v $(pwd):/app -w /app python:3-slim python3"
+fi
+
+echo "Using python interpreter: $PYTHON"
+echo "Python version: $($PYTHON --version)"
+
+# ======== Part 1: Collect and prepare wheels ========
+
+# Collect all wheels
+mkdir -p all-rocm-wheels
+cp artifacts/rocm-base-wheels/*.whl all-rocm-wheels/ 2>/dev/null || true
+cp artifacts/rocm-vllm-wheel/*.whl all-rocm-wheels/ 2>/dev/null || true
+
+WHEEL_COUNT=$(ls all-rocm-wheels/*.whl 2>/dev/null | wc -l)
+echo "Total wheels to upload: $WHEEL_COUNT"
+
+if [ "$WHEEL_COUNT" -eq 0 ]; then
+    echo "ERROR: No wheels found to upload!"
+    exit 1
+fi
+
+# Rename linux to manylinux in wheel filenames
+for wheel in all-rocm-wheels/*.whl; do
+    if [[ "$wheel" == *"linux"* ]] && [[ "$wheel" != *"manylinux"* ]]; then
+        new_wheel="${wheel/linux/$MANYLINUX_VERSION}"
+        mv -- "$wheel" "$new_wheel"
+        echo "Renamed: $(basename "$wheel") -> $(basename "$new_wheel")"
+    fi
+done
+
+echo ""
+echo "Wheels to upload:"
+ls -lh all-rocm-wheels/
+
+# ======== Part 2: Upload wheels to S3 ========
+
+echo ""
+echo "Uploading wheels to $S3_COMMIT_PREFIX"
+for wheel in all-rocm-wheels/*.whl; do
+    aws s3 cp "$wheel" "$S3_COMMIT_PREFIX"
+done
+
+# ======== Part 3: Generate and upload indices ========
+
+# List existing wheels in commit directory
+echo ""
+echo "Generating indices..."
+obj_json="rocm-objects.json"
+aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$ROCM_SUBPATH/" --delimiter / --output json > "$obj_json"
+
+mkdir -p "$INDICES_OUTPUT_DIR"
+
+# Use the existing generate-nightly-index.py
+# HACK: Replace regex module with stdlib re (same as CUDA script)
+sed -i 's/import regex as re/import re/g' .buildkite/scripts/generate-nightly-index.py
+
+$PYTHON .buildkite/scripts/generate-nightly-index.py \
+    --version "$ROCM_SUBPATH" \
+    --current-objects "$obj_json" \
+    --output-dir "$INDICES_OUTPUT_DIR" \
+    --comment "ROCm commit $BUILDKITE_COMMIT"
+
+# Upload indices to commit directory
+echo "Uploading indices to $S3_COMMIT_PREFIX"
+aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
+
+# Update rocm/nightly/ if on main branch and not a PR
+if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]] || [[ "$NIGHTLY" == "1" ]]; then
+    echo "Updating rocm/nightly/ index..."
+    aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/rocm/nightly/"
+fi
+
+# Extract version from vLLM wheel and update version-specific index
+VLLM_WHEEL=$(ls all-rocm-wheels/vllm*.whl 2>/dev/null | head -1)
+if [ -n "$VLLM_WHEEL" ]; then
+    VERSION=$(unzip -p "$VLLM_WHEEL" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
+    echo "Version in wheel: $VERSION"
+    PURE_VERSION="${VERSION%%+*}"
+    PURE_VERSION="${PURE_VERSION%%.rocm}"
+    echo "Pure version: $PURE_VERSION"
+
+    if [[ "$VERSION" != *"dev"* ]]; then
+        echo "Updating rocm/$PURE_VERSION/ index..."
+        aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/rocm/$PURE_VERSION/"
+    fi
+fi
+
+# ======== Part 4: Summary ========
+
+echo ""
+echo "========================================"
+echo "ROCm Wheel Upload Complete!"
+echo "========================================"
+echo ""
+echo "Wheels available at:"
+echo "  s3://$BUCKET/$ROCM_SUBPATH/"
+echo ""
+echo "Install command (by commit):"
+echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/$ROCM_SUBPATH/"
+echo ""
+if [[ "$BUILDKITE_BRANCH" == "main" ]] || [[ "$NIGHTLY" == "1" ]]; then
+    echo "Install command (nightly):"
+    echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/rocm/nightly/"
+fi
+echo ""
+echo "Wheel count: $WHEEL_COUNT"
+echo "========================================"

.buildkite/test-amd.yaml

@@ -71,6 +71,7 @@ steps:
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/multimodal
+  - tests/renderers
   - tests/standalone_tests/lazy_imports.py
   - tests/tokenizers_
   - tests/tool_parsers
@@ -82,6 +83,7 @@ steps:
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s -m 'cpu_test' multimodal
+  - pytest -v -s renderers
   - pytest -v -s tokenizers_
   - pytest -v -s tool_parsers
   - pytest -v -s transformers_utils
@@ -128,7 +130,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/instrumentator --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
 
 - label: Entrypoints Integration Test (LLM) # 30min
   timeout_in_minutes: 40
@@ -148,7 +150,7 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
-- label: Entrypoints Integration Test (API Server) # 100min
+- label: Entrypoints Integration Test (API Server 1) # 100min
   timeout_in_minutes: 130
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
@@ -162,10 +164,28 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/openai/test_collective_rpc.py # PYTHONPATH is needed to import custom Worker extension
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_collective_rpc.py --ignore=entrypoints/openai/tool_parsers/
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
   - pytest -v -s entrypoints/test_chat_utils.py
 
+- label: Entrypoints Integration Test (API Server 2)
+  timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
+  agent_pool: mi325_1
+  # grade: Blocking
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/sleep
+  - tests/entrypoints/rpc
+  - tests/tool_use
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/sleep
+  - pytest -v -s tool_use
+  - PYTHONPATH=/vllm-workspace  pytest -v -s entrypoints/rpc
+
 - label: Entrypoints Integration Test (Pooling)
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
@@ -181,6 +201,21 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
 
+- label: Entrypoints Integration Test (Responses API)
+  timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
+  agent_pool: mi325_1
+  # grade: Blocking
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai/responses
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai/responses
+
 - label: Distributed Tests (4 GPUs) # 35min
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
@@ -201,6 +236,9 @@ steps:
   - tests/v1/engine/test_engine_core_client.py
   - tests/distributed/test_symm_mem_allreduce.py
   commands:
+  # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
+  # TODO: Remove when the bug is fixed in a future ROCm release
+  - export TORCH_NCCL_BLOCKING_WAIT=1
   # test with torchrun tp=2 and external_dp=2
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   # test with torchrun tp=2 and pp=2
@@ -249,9 +287,10 @@ steps:
   - vllm/v1/executor/uniproc_executor.py
   - vllm/v1/worker/gpu_worker.py
   commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  #- export NCCL_CUMEM_HOST_ENABLE=0
   # test with torchrun tp=2 and dp=4 with ep
+  # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
+  # TODO: Remove when the bug is fixed in a future ROCm release
+  - export TORCH_NCCL_BLOCKING_WAIT=1
   - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
 
 - label: EPLB Algorithm Test # 5min
@@ -331,7 +370,9 @@ steps:
 - label: V1 Test e2e + engine # 65min
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
+  # The test uses 4 GPUs, but we schedule it on 8-GPU machines for stability.
+  # See discussion here: https://github.com/vllm-project/vllm/pull/31040
+  agent_pool: mi325_8
   # grade: Blocking
   source_file_dependencies:
     - vllm/
@@ -389,6 +430,8 @@ steps:
   timeout_in_minutes: 30
   gpu: h100
   source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
     - vllm/v1/attention
     - tests/v1/attention
   commands:
@@ -413,10 +456,12 @@ steps:
   timeout_in_minutes: 30
   gpu: b200
   source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
     - vllm/v1/attention
     - tests/v1/attention
   commands:
-    - VLLM_DISABLE_FLASHINFER_PREFILL=1 pytest -v -s v1/attention # TODO: FI prefill is bugged and causes incorrectness, fix this
+    - pytest -v -s v1/attention
 
 - label: V1 Test others (CPU) # 5 mins
   mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
@@ -492,13 +537,12 @@ steps:
   - tests/samplers
   - tests/conftest.py
   commands:
-    - pytest -v -s samplers
-    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
+    - pytest -v -s -m 'not skip_v1' samplers
 
 - label: LoRA Test %N # 20min each
   timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
   - vllm/lora
@@ -560,9 +604,11 @@ steps:
   - tests/compile
   commands:
   - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
-    # Limit to no custom ops to reduce running time
-    # Wrap with quotes to escape yaml and avoid starting -k string with a -
-  - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
+    # # Limit to no custom ops to reduce running time
+    # # Wrap with quotes to escape yaml and avoid starting -k string with a -
+    # - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
+    # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
+    # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
 
 - label: Cudagraph test
   timeout_in_minutes: 20
@@ -592,12 +638,13 @@ steps:
 - label: Kernels Attention Test %N # 23min
   timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
   - csrc/attention/
-  - vllm/attention
   - vllm/v1/attention
+    # TODO: remove this dependency (https://github.com/vllm-project/vllm/issues/32267)
+  - vllm/model_executor/layers/attention
   - tests/kernels/attention
   commands:
     - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
@@ -606,7 +653,7 @@ steps:
 - label: Kernels Quantization Test %N # 64min
   timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
   - csrc/quantization/
@@ -619,7 +666,7 @@ steps:
 - label: Kernels MoE Test %N # 40min
   timeout_in_minutes: 60
   mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
   - csrc/quantization/cutlass_w8a8/moe/
@@ -665,6 +712,17 @@ steps:
     - pytest -v -s kernels/moe/test_batched_deepgemm.py
     - pytest -v -s kernels/attention/test_deepgemm_attention.py
 
+- label: Kernels Helion Test
+  timeout_in_minutes: 30
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_1
+  source_file_dependencies:
+  - vllm/utils/import_utils.py
+  - tests/kernels/helion/
+  commands:
+    - pip install helion
+    - pytest -v -s kernels/helion/
+
 - label: Model Executor Test # 23min
   timeout_in_minutes: 35
   torch_nightly: true
@@ -686,7 +744,7 @@ steps:
 - label: Benchmarks # 11min
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   working_dir: "/vllm-workspace/.buildkite"
   source_file_dependencies:
@@ -697,7 +755,7 @@ steps:
 - label: Benchmarks CLI Test # 7min
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
   - vllm/
@@ -707,7 +765,7 @@ steps:
 
 - label: Quantization Test # 70min
   timeout_in_minutes: 90
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
@@ -722,7 +780,7 @@ steps:
   # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
   # we can only upgrade after this is resolved
   # TODO(jerryzh168): resolve the above comment
-  - uv pip install --system torchao==0.13.0
+  - uv pip install --system torchao==0.14.1
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
@@ -736,7 +794,7 @@ steps:
   - vllm/model_executor/layers/quantization
   autorun_on_main: true
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
 
 - label: OpenAI API correctness # 10min
   timeout_in_minutes: 15
@@ -747,21 +805,11 @@ steps:
   - csrc/
   - vllm/entrypoints/openai/
   - vllm/model_executor/models/whisper.py
+  - tools/
   commands: # LMEval+Transcription WER check
-  # Transcription WER check is skipped because encoder-decoder models are not supported on ROCm, see https://github.com/vllm-project/vllm/issues/27442
+  - bash ../tools/install_torchcodec_rocm.sh || exit 1
   - pytest -s entrypoints/openai/correctness/
 
-- label: OpenAI-Compatible Tool Use # 23 min
-  timeout_in_minutes: 35
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_1
-  # grade: Blocking
-  fast_check: false
-  source_file_dependencies:
-    - vllm/
-    - tests/tool_use
-  commands:
-    - pytest -v -s tool_use
 
 #####  models test  #####
 
@@ -781,7 +829,7 @@ steps:
 - label: Basic Models Tests (Extra Initialization) %N
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -827,7 +875,7 @@ steps:
 
 - label: Language Models Tests (Standard)
   timeout_in_minutes: 25
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   torch_nightly: true
@@ -842,7 +890,7 @@ steps:
 - label: Language Models Tests (Extra Standard) %N
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -854,6 +902,7 @@ steps:
     # Shard slow subset of standard language models tests. Only run when model
     # source is modified, or when specified test files are modified
     - pip freeze | grep -E 'torch'
+    - export TORCH_NCCL_BLOCKING_WAIT=1
     - pytest -v -s models/language -m 'core_model and slow_test' \
              --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT \
              --shard-id=$$BUILDKITE_PARALLEL_JOB
@@ -862,7 +911,7 @@ steps:
 - label: Language Models Tests (Hybrid) %N
   timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_8
+  agent_pool: mi325_1
   # grade: Blocking
   torch_nightly: true
   source_file_dependencies:
@@ -871,7 +920,7 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     # Shard hybrid language model tests
     - pytest -v -s models/language/generation \
@@ -892,7 +941,7 @@ steps:
   commands:
     # Install fast path packages for testing against transformers
     # Note: also needed to run plamo2 model in vLLM
-    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/AndreasKaratzas/mamba@fix-rocm-7.0-warp-size-constexpr'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
@@ -957,7 +1006,7 @@ steps:
     - pytest -v -s models/multimodal/processing
 
 - label: Multi-Modal Models Test (Standard) # 60min
-  timeout_in_minutes: 80
+  timeout_in_minutes: 100
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
@@ -966,13 +1015,16 @@ steps:
   - vllm/
   - tests/models/multimodal
   commands:
+    - export MIOPEN_DEBUG_CONV_DIRECT=0
+    - export MIOPEN_DEBUG_CONV_GEMM=0
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pip freeze | grep -E 'torch'
-    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing --ignore models/multimodal/pooling/test_prithvi_mae.py
+    - pytest -v -s models/multimodal/pooling/test_prithvi_mae.py -m core_model
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
-- label: Multi-Modal Accuracy Eval (Small Models) # 150min - 180min
-  timeout_in_minutes: 180
+- label: Multi-Modal Accuracy Eval (Small Models) # 5min
+  timeout_in_minutes: 10
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
@@ -982,7 +1034,9 @@ steps:
   - vllm/inputs/
   - vllm/v1/core/
   commands:
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
+  - export MIOPEN_DEBUG_CONV_DIRECT=0
+  - export MIOPEN_DEBUG_CONV_GEMM=0
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt
 
 - label: Multi-Modal Models Test (Extended) 1 # 60min
   timeout_in_minutes: 120
@@ -994,10 +1048,13 @@ steps:
   - vllm/
   - tests/models/multimodal
   commands:
+    - export MIOPEN_DEBUG_CONV_DIRECT=0
+    - export MIOPEN_DEBUG_CONV_GEMM=0
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal -m 'not core_model' --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing
 
-- label: Multi-Modal Models Test (Extended) 2
+- label: Multi-Modal Models Test (Extended) 2 #60min
+  timeout_in_minutes: 120
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
@@ -1006,6 +1063,8 @@ steps:
   - vllm/
   - tests/models/multimodal
   commands:
+    - export MIOPEN_DEBUG_CONV_DIRECT=0
+    - export MIOPEN_DEBUG_CONV_GEMM=0
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 
@@ -1019,6 +1078,8 @@ steps:
   - vllm/
   - tests/models/multimodal
   commands:
+    - export MIOPEN_DEBUG_CONV_DIRECT=0
+    - export MIOPEN_DEBUG_CONV_GEMM=0
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
 
@@ -1073,13 +1134,13 @@ steps:
   - csrc/quantization/cutlass_w8a8/moe/
   - vllm/model_executor/layers/fused_moe/cutlass_moe.py
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/v1/attention/backends/mla/cutlass_mla.py
   - vllm/v1/attention/backends/mla/flashinfer_mla.py
+  - vllm/v1/attention/selector.py
   - vllm/platforms/cuda.py
-  - vllm/attention/selector.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
@@ -1122,7 +1183,6 @@ steps:
   - tests/compile/test_fusion_attn.py
   - tests/compile/test_silu_mul_quant_fusion.py
   - tests/compile/distributed/test_fusion_all_reduce.py
-  - tests/compile/distributed/test_fusions_e2e.py
   - tests/compile/fullgraph/test_full_graph.py
   commands:
     - nvidia-smi
@@ -1130,33 +1190,16 @@ steps:
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
     - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # Wrap with quotes to escape yaml
-    - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
+
+    # # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
+    # # Wrap with quotes to escape yaml
+    # - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
+    # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
+    # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
+
     # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
     - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
 
-- label: Blackwell Fusion E2E Tests # 30 min
-  timeout_in_minutes: 40
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/distributed/test_fusions_e2e.py
-  commands:
-    - nvidia-smi
-    # Run all e2e fusion tests
-    - pytest -v -s tests/compile/distributed/test_fusions_e2e.py
-
 - label: Blackwell GPT-OSS Eval
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
@@ -1196,7 +1239,7 @@ steps:
   - csrc/
   - vllm/model_executor/layers/quantization
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
 
 #####  1 GPU test  #####
 #####  multi gpus test  #####
@@ -1219,7 +1262,7 @@ steps:
 
 - label: 2 Node Tests (4 GPUs in total) # 16min
   timeout_in_minutes: 30
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdmultinode]
   agent_pool: mi325_4
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
@@ -1233,16 +1276,16 @@ steps:
   - tests/distributed/
   - tests/examples/offline_inference/data_parallel.py
   commands:
-  - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
-    - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
-    - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=0 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
+  - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)  | grep 'Same node test passed'   | grep 'Node count test passed'
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py 
+    - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py 
+    - python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=0 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
   - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
-    - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
-    - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py 
+    - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py 
+    - python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=1 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code
 
 - label: Distributed Tests (2 GPUs) # 68min
   timeout_in_minutes: 90
@@ -1268,6 +1311,9 @@ steps:
   - tests/v1/shutdown
   - tests/v1/worker/test_worker_memory_snapshot.py
   commands:
+  # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
+  # TODO: Remove when the bug is fixed in a future ROCm release
+  - export TORCH_NCCL_BLOCKING_WAIT=1
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
@@ -1407,7 +1453,7 @@ steps:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt
 
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
   # grade: Blocking
   timeout_in_minutes: 30
@@ -1417,8 +1463,22 @@ steps:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
     - tests/v1/kv_connector/nixl_integration/
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/tp_config_sweep_accuracy_test.sh
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
+    - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: DP EP NixlConnector PD accuracy tests (Distributed) # 15min
+  mirror_hardwares: [amdexperimental, amdproduction]
+  agent_pool: mi325_4
+  # grade: Blocking
+  timeout_in_minutes: 15
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
+    - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 ##### multi gpus test #####
 ##### A100 test #####
@@ -1433,6 +1493,9 @@ steps:
   source_file_dependencies:
   - vllm/
   commands:
+  # Work around HIP bug tracked here: https://github.com/ROCm/hip/issues/3876
+  # TODO: Remove when the bug is fixed in a future ROCm release
+  - export TORCH_NCCL_BLOCKING_WAIT=1
   # NOTE: don't test llama model here, it seems hf implementation is buggy
   # see https://github.com/vllm-project/vllm/pull/5689 for details
   - pytest -v -s distributed/test_custom_all_reduce.py
@@ -1487,10 +1550,13 @@ steps:
     - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
     - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
     #- pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
-    - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
+    # - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
+    # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
+    # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
+
     - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/distributed/test_sequence_parallel.py
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - HIP_VISIBLE_DEVICES=0,1 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
+    - HIP_VISIBLE_DEVICES=0,1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=allgather_reducescatter --disable-nccl-for-dp-synchronization
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### B200 test #####
@@ -1514,7 +1580,7 @@ steps:
   - csrc/
   - vllm/model_executor/layers/quantization
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
 
 - label: LM Eval Large Models (4 Card)
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -1569,6 +1635,8 @@ steps:
   - .buildkite/scripts/run-prime-rl-test.sh
   commands:
     - bash .buildkite/scripts/run-prime-rl-test.sh
+
+##### EPLB Accuracy Tests #####
 - label: DeepSeek V2-Lite Accuracy
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
@@ -1602,17 +1670,6 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
 
-- label: DeepSeek V2-Lite Async EPLB Accuracy
-  timeout_in_minutes: 60
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
-  # grade: Blocking
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_async_eplb.sh 0.25 1319 8030
 
 - label: Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy
   timeout_in_minutes: 60

.buildkite/test-pipeline.yaml

@@ -64,6 +64,7 @@ steps:
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/multimodal
+  - tests/renderers
   - tests/standalone_tests/lazy_imports.py
   - tests/tokenizers_
   - tests/tool_parsers
@@ -75,6 +76,7 @@ steps:
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s -m 'cpu_test' multimodal
+  - pytest -v -s renderers
   - pytest -v -s tokenizers_
   - pytest -v -s tool_parsers
   - pytest -v -s transformers_utils
@@ -114,7 +116,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
 
 - label: Entrypoints Integration Test (LLM) # 30min
   timeout_in_minutes: 40
@@ -132,7 +134,7 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
-- label: Entrypoints Integration Test (API Server) # 100min
+- label: Entrypoints Integration Test (API Server 1) # 100min
   timeout_in_minutes: 130
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
@@ -144,10 +146,26 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/openai/test_collective_rpc.py # PYTHONPATH is needed to import custom Worker extension
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_collective_rpc.py --ignore=entrypoints/openai/tool_parsers/
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
   - pytest -v -s entrypoints/test_chat_utils.py
 
+- label: Entrypoints Integration Test (API Server 2)
+  timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/sleep
+  - tests/entrypoints/rpc
+  - tests/tool_use
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/sleep
+  - PYTHONPATH=/vllm-workspace  pytest -v -s entrypoints/rpc
+  - pytest -v -s tool_use
+
 - label: Entrypoints Integration Test (Pooling)
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
@@ -161,6 +179,18 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
 
+- label: Entrypoints Integration Test (Responses API)
+  timeout_in_minutes: 50
+  mirror_hardwares: [amdexperimental]
+  working_dir: "/vllm-workspace/tests"
+  fast_check: true
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai/responses
+  commands:
+  - pytest -v -s entrypoints/openai/responses
+
 - label: Distributed Tests (4 GPUs) # 35min
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
@@ -303,7 +333,10 @@ steps:
     # TODO: accuracy does not match, whether setting
     # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
     - pytest -v -s v1/e2e
-    - pytest -v -s v1/engine
+    # Run this test standalone for now;
+    # need to untangle use (implicit) use of spawn/fork across the tests.
+    - pytest -v -s v1/engine/test_preprocess_error_handling.py
+    - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
 
 - label: V1 Test entrypoints # 35min
   timeout_in_minutes: 50
@@ -329,7 +362,7 @@ steps:
     - pytest -v -s v1/sample
     - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
-    - pytest -v -s v1/spec_decode
+    - pytest -v -s -m 'not slow_test' v1/spec_decode
     - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/metrics
     - pytest -v -s v1/test_oracle.py
@@ -343,6 +376,8 @@ steps:
   timeout_in_minutes: 30
   gpu: h100
   source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
     - vllm/v1/attention
     - tests/v1/attention
   commands:
@@ -365,10 +400,12 @@ steps:
   timeout_in_minutes: 30
   gpu: b200
   source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
     - vllm/v1/attention
     - tests/v1/attention
   commands:
-    - VLLM_DISABLE_FLASHINFER_PREFILL=1 pytest -v -s v1/attention # TODO: FI prefill is bugged and causes incorrectness, fix this
+    - pytest -v -s v1/attention
 
 - label: V1 Test others (CPU) # 5 mins
   source_file_dependencies:
@@ -500,9 +537,11 @@ steps:
   commands:
     # fp8 kv scales not supported on sm89, tested on Blackwell instead
   - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
-    # Limit to no custom ops to reduce running time
-    # Wrap with quotes to escape yaml and avoid starting -k string with a -
-  - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
+    # # Limit to no custom ops to reduce running time
+    # # Wrap with quotes to escape yaml and avoid starting -k string with a -
+    # - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
+    # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
+    # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
 
 - label: Cudagraph test
   timeout_in_minutes: 20
@@ -531,8 +570,9 @@ steps:
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - csrc/attention/
-  - vllm/attention
   - vllm/v1/attention
+    # TODO: remove this dependency (https://github.com/vllm-project/vllm/issues/32267)
+  - vllm/model_executor/layers/attention
   - tests/kernels/attention
   commands:
     - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
@@ -593,6 +633,56 @@ steps:
     - pytest -v -s kernels/moe/test_batched_deepgemm.py
     - pytest -v -s kernels/attention/test_deepgemm_attention.py
 
+- label: Kernels Helion Test
+  timeout_in_minutes: 30
+  gpu: h100
+  source_file_dependencies:
+  - vllm/utils/import_utils.py
+  - tests/kernels/helion/
+  commands:
+    - pip install helion
+    - pytest -v -s kernels/helion/
+
+  
+- label: Kernels FP8 MoE Test (1 H100)
+  timeout_in_minutes: 90
+  gpu: h100
+  num_gpus: 1
+  optional: true
+  commands:
+    - pytest -v -s kernels/moe/test_cutlass_moe.py
+    - pytest -v -s kernels/moe/test_flashinfer.py
+    - pytest -v -s kernels/moe/test_gpt_oss_triton_kernels.py
+    - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py
+    - pytest -v -s kernels/moe/test_moe.py
+    # - pytest -v -s kernels/moe/test_block_fp8.py - failing on main
+    - pytest -v -s kernels/moe/test_block_int8.py
+    - pytest -v -s kernels/moe/test_triton_moe_no_act_mul.py
+    - pytest -v -s kernels/moe/test_triton_moe_ptpc_fp8.py
+
+- label: Kernels FP8 MoE Test (2 H100s)
+  timeout_in_minutes: 90
+  gpu: h100
+  num_gpus: 2
+  optional: true
+  commands:
+    - pytest -v -s kernels/moe/test_deepep_deepgemm_moe.py
+    - pytest -v -s kernels/moe/test_deepep_moe.py
+    - pytest -v -s kernels/moe/test_pplx_cutlass_moe.py
+    # - pytest -v -s kernels/moe/test_pplx_moe.py - failing on main
+  
+- label: Kernels Fp4 MoE Test (B200)
+  timeout_in_minutes: 60
+  gpu: b200
+  num_gpus: 1
+  optional: true
+  commands:
+    - pytest -v -s kernels/moe/test_cutedsl_moe.py
+    - pytest -v -s kernels/moe/test_flashinfer_moe.py
+    - pytest -v -s kernels/moe/test_nvfp4_moe.py
+    - pytest -v -s kernels/moe/test_ocp_mx_moe.py
+
+
 - label: Model Executor Test # 23min
   timeout_in_minutes: 35
   torch_nightly: true
@@ -642,7 +732,7 @@ steps:
   # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
   # we can only upgrade after this is resolved
   # TODO(jerryzh168): resolve the above comment
-  - uv pip install --system torchao==0.13.0 --index-url https://download.pytorch.org/whl/cu129
+  - uv pip install --system torchao==0.14.1 --index-url https://download.pytorch.org/whl/cu129
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
@@ -654,7 +744,7 @@ steps:
   - vllm/model_executor/layers/quantization
   autorun_on_main: true
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
 
 - label: OpenAI API correctness # 22min
   timeout_in_minutes: 30
@@ -666,16 +756,6 @@ steps:
   commands: # LMEval+Transcription WER check
   - pytest -s entrypoints/openai/correctness/
 
-- label: OpenAI-Compatible Tool Use # 23 min
-  timeout_in_minutes: 35
-  mirror_hardwares: [amdexperimental]
-  fast_check: false
-  source_file_dependencies:
-    - vllm/
-    - tests/tool_use
-  commands:
-    - pytest -v -s tool_use
-
 #####  models test  #####
 
 - label: Basic Models Tests (Initialization)
@@ -930,24 +1010,23 @@ steps:
     # Whisper needs spawn method to avoid deadlock
     - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
-- label: Blackwell Test # 21 min
+- label: Blackwell Test # 23 min
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
   gpu: b200
-  # optional: true
   source_file_dependencies:
   - csrc/quantization/fp4/
   - csrc/attention/mla/
   - csrc/quantization/cutlass_w8a8/moe/
   - vllm/model_executor/layers/fused_moe/cutlass_moe.py
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/v1/attention/backends/mla/cutlass_mla.py
   - vllm/v1/attention/backends/mla/flashinfer_mla.py
+  - vllm/v1/attention/selector.py
   - vllm/platforms/cuda.py
-  - vllm/attention/selector.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
@@ -971,6 +1050,8 @@ steps:
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
+    # e2e
+    - pytest -v -s tests/models/quantization/test_nvfp4.py
 
 - label: Blackwell Fusion and Compile Tests # 30 min
   timeout_in_minutes: 40
@@ -990,7 +1071,6 @@ steps:
   - tests/compile/test_fusion_attn.py
   - tests/compile/test_silu_mul_quant_fusion.py
   - tests/compile/distributed/test_fusion_all_reduce.py
-  - tests/compile/distributed/test_fusions_e2e.py
   - tests/compile/fullgraph/test_full_graph.py
   commands:
     - nvidia-smi
@@ -998,33 +1078,15 @@ steps:
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
     - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # Wrap with quotes to escape yaml
-    - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
+    #  # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
+    #  # Wrap with quotes to escape yaml
+    #  - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
+    # Old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293
+    # in favor of new tests in fusions_e2e. We avoid replicating the new jobs in this file as it's deprecated.
+
     # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
     - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
 
-- label: Blackwell Fusion E2E Tests # 30 min
-  timeout_in_minutes: 40
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/distributed/test_fusions_e2e.py
-  commands:
-    - nvidia-smi
-    # Run all e2e fusion tests
-    - pytest -v -s tests/compile/distributed/test_fusions_e2e.py
-
 - label: Blackwell GPT-OSS Eval
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
@@ -1064,7 +1126,7 @@ steps:
   - csrc/
   - vllm/model_executor/layers/quantization
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
 
 #####  1 GPU test  #####
 #####  multi gpus test  #####
@@ -1096,17 +1158,18 @@ steps:
   - vllm/model_executor/models/
   - tests/distributed/
   - tests/examples/offline_inference/data_parallel.py
+  - .buildkite/scripts/run-multi-node-test.sh
   commands:
   - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
     - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
     - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
-    - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=0 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
+    - python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=0 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
   - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
     - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
     - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
-    - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
+    - python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=1 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code
 
 - label: Distributed Tests (2 GPUs) # 68min
   timeout_in_minutes: 90
@@ -1195,7 +1258,7 @@ steps:
   - pytest -v -s distributed/test_distributed_oot.py
   - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
-  - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
+  - pytest -v -s plugins/lora_resolvers # unit tests for lora resolver plugins
 
 - label: Pipeline + Context Parallelism Test # 45min
   timeout_in_minutes: 60
@@ -1223,6 +1286,8 @@ steps:
     # FIXIT: find out which code initialize cuda before running the test
     # before the fix, we need to use spawn to test it
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    # Alot of these tests are on the edge of OOMing
+    - export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
     # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
@@ -1256,8 +1321,8 @@ steps:
   commands:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt
 
-- label: NixlConnector PD accuracy tests (Distributed) # 30min
-  timeout_in_minutes: 30
+- label: NixlConnector PD accuracy tests (Distributed) # 40min
+  timeout_in_minutes: 40
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -1265,7 +1330,18 @@ steps:
     - tests/v1/kv_connector/nixl_integration/
   commands:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/tp_config_sweep_accuracy_test.sh
+    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: DP EP NixlConnector PD accuracy tests (Distributed) # 15min
+  timeout_in_minutes: 15
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 
 ##### multi gpus test #####
@@ -1285,6 +1361,20 @@ steps:
   - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   - pytest -v -s -x lora/test_mixtral.py
 
+- label: Acceptance Length Test (Large Models) # optional
+  timeout_in_minutes: 120
+  gpu: h100
+  optional: true
+  num_gpus: 1
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/v1/spec_decode/
+  - vllm/model_executor/models/mlp_speculator.py
+  - tests/v1/spec_decode/test_acceptance_length.py
+  commands:
+    - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
+    - pytest -v -s v1/spec_decode/test_acceptance_length.py -m slow_test
+
 - label: LM Eval Large Models # optional
   gpu: a100
   optional: true
@@ -1310,22 +1400,39 @@ steps:
     - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
     - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
 
-##### H200 test #####
-- label: Distributed Tests (H200) # optional
-  gpu: h200
+- label: Sequence Parallel Tests (H100) # 60 min
+  timeout_in_minutes: 60
+  working_dir: "/vllm-workspace/"
+  gpu: h100
+  optional: true
+  num_gpus: 2
+  commands:
+    - export VLLM_TEST_CLEAN_GPU_MEMORY=1
+    # Run sequence parallel tests
+    - pytest -v -s tests/distributed/test_sequence_parallel.py
+    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
+
+- label: Distributed Tests (H100) # optional
+  gpu: h100
   optional: true
   working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
     - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_async_tp.py
-    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/distributed/test_sequence_parallel.py
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
+    - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
+##### H200 test #####
+
+- label: LM Eval Large Models (H200) # optional
+  timeout_in_minutes: 60
+  gpu: h200
+  optional: true
+  num_gpus: 8
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
+
 ##### B200 test #####
 - label: Distributed Tests (B200) # optional
   gpu: b200
@@ -1348,6 +1455,7 @@ steps:
   - vllm/
   - .buildkite/scripts/run-prime-rl-test.sh
   commands:
+    - nvidia-smi
     - bash .buildkite/scripts/run-prime-rl-test.sh
 
 - label: DeepSeek V2-Lite Accuracy
@@ -1376,3 +1484,26 @@ steps:
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
+
+##### MoE Refactor (Temporary) Tests #####
+
+- label: MoE Refactor Integration Test (H100 - TEMPORARY) # optional
+  gpu: h100
+  optional: true
+  num_gpus: 2
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-h100.txt
+  
+- label: MoE Refactor Integration Test (B200 - TEMPORARY) # optional
+  gpu: b200
+  optional: true
+  num_gpus: 2
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-b200.txt
+
+- label: MoE Refactor Integration Test (B200 DP - TEMPORARY) # optional
+  gpu: b200
+  optional: true
+  num_gpus: 2
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor-dp-ep/config-b200.txt

.buildkite/test_areas/attention.yaml

@@ -4,8 +4,10 @@ depends_on:
 steps:
 - label: V1 attention (H100)
   timeout_in_minutes: 30
-  gpu: h100
+  device: h100
   source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
     - vllm/v1/attention
     - tests/v1/attention
   commands:
@@ -13,9 +15,11 @@ steps:
 
 - label: V1 attention (B200)
   timeout_in_minutes: 30
-  gpu: b200
+  device: b200
   source_file_dependencies:
+    - vllm/config/attention.py
+    - vllm/model_executor/layers/attention
     - vllm/v1/attention
     - tests/v1/attention
   commands:
-    - VLLM_DISABLE_FLASHINFER_PREFILL=1 pytest -v -s v1/attention # TODO: FI prefill is bugged and causes incorrectness, fix this
+    - pytest -v -s v1/attention

.buildkite/test_areas/compile.yaml

@@ -2,56 +2,196 @@ group: Compile
 depends_on: 
   - image-build
 steps:
-- label: Fusion and Compile Tests (B200)
-  timeout_in_minutes: 40
+- label: Sequence Parallel Tests (2 GPUs)
+  timeout_in_minutes: 50
   working_dir: "/vllm-workspace/"
-  gpu: b200
+  num_devices: 2
   source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
+  - vllm/model_executor/layers/
+  - vllm/compilation/
   - vllm/v1/worker/
   - vllm/v1/cudagraph_dispatcher.py
+  - tests/distributed/test_sequence_parallel.py
+  commands:
+  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
+  - pytest -v -s tests/distributed/test_sequence_parallel.py
+
+- label: Sequence Parallel Tests (2xH100)
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/"
+  device: h100
+  optional: true
+  num_devices: 2
+  commands:
+  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
+  - pytest -v -s tests/distributed/test_sequence_parallel.py
+
+- label: Distributed Compile Unit Tests (2xH100)
+  timeout_in_minutes: 40
+  working_dir: "/vllm-workspace/"
+  device: h100
+  num_devices: 2
+  source_file_dependencies:
   - vllm/compilation/
-  # can affect pattern matching
+  - vllm/model_executor/layers
+  - tests/compile/distributed/test_fusion_all_reduce.py
+  - tests/compile/distributed/test_sequence_parallelism.py
+  - tests/compile/distributed/test_async_tp.py
+  commands:
+  - export VLLM_TEST_CLEAN_GPU_MEMORY=1
+  - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
+  - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
+  - pytest -v -s tests/compile/distributed/test_async_tp.py
+
+- label: Fusion and Compile Unit Tests (B200)
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/"
+  device: b200
+  source_file_dependencies:
+  - csrc/quantization/fp4/
+  - vllm/model_executor/layers/quantization/
   - vllm/model_executor/layers/layernorm.py
   - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
+  - vllm/model_executor/layers/attention/attention.py
+  - vllm/v1/attention/backends/flashinfer.py
+  - vllm/compilation/ # TODO(luka) limit to vllm/compilation/passes
   - tests/compile/test_fusion_attn.py
   - tests/compile/test_silu_mul_quant_fusion.py
   - tests/compile/distributed/test_fusion_all_reduce.py
-  - tests/compile/distributed/test_fusions_e2e.py
   - tests/compile/fullgraph/test_full_graph.py
   commands:
+    # b200 runners are limited, so we limit the tests to the minimum set only supported on Blackwell
     - nvidia-smi
-    - pytest -v -s tests/compile/test_fusion_attn.py
+    - pytest -v -s tests/compile/test_fusion_attn.py -k FLASHINFER
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
-    # this runner has 2 GPUs available even though num_gpus=2 is not set
+    # this runner has 2 GPUs available even though num_devices=2 is not set
     - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # Wrap with quotes to escape yaml
-    - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
     # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
+    # TODO(luka) move to H100 once pass tests run on H100
     - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
 
-- label: Fusion E2E (2 GPUs)(B200)
-  timeout_in_minutes: 40
+- label: Fusion E2E Quick (H100)
+  timeout_in_minutes: 15
   working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true
-  num_gpus: 2
+  device: h100
+  num_devices: 1
   source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/distributed/test_fusions_e2e.py
+    - csrc/quantization/
+    - vllm/model_executor/
+    - vllm/v1/attention/
+    - vllm/compilation/
+    - tests/compile/fusions_e2e/
   commands:
     - nvidia-smi
-    # Run all e2e fusion tests
-    - pytest -v -s tests/compile/distributed/test_fusions_e2e.py
+    # Run all models and attn backends but only Inductor partition and native custom ops
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
+    # Qwen requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3"
 
+- label: Fusion E2E Config Sweep (H100)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/"
+  device: h100
+  num_devices: 1
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/compilation/
+    # can affect pattern matching
+    - vllm/model_executor/layers/layernorm.py
+    - vllm/model_executor/layers/activation.py
+    - vllm/model_executor/layers/attention/attention.py
+    - vllm/model_executor/layers/quantization/input_quant_fp8.py
+    - tests/compile/fusions_e2e/
+  commands:
+    - nvidia-smi
+    # Run just llama3 (fp8) for all config combinations
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "llama-3"
+
+- label: Fusion E2E Config Sweep (B200)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/"
+  device: b200
+  num_devices: 1
+  optional: true
+  commands:
+    - nvidia-smi
+    # Run all models and attn backends but only Inductor partition and native custom ops
+    # -k "inductor_partition and not +rms_norm and not +quant_fp8"
+    # Qwen requires +quant_fp8 as -quant_fp8 rms+quant fusion is not supported
+    # -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3"
+    # Run just llama3 (fp8 & fp4) for all config combinations
+    # -k "llama-3"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp1_quant.py -k "inductor_partition and not +rms_norm and not +quant_fp8" -k "inductor_partition and not +rms_norm and +quant_fp8 and qwen3" -k "llama-3"
+
+- label: Fusion E2E TP2 Quick (H100)
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/"
+  device: h100
+  num_devices: 2
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/model_executor/
+    - vllm/v1/attention/
+    - vllm/compilation/
+    - tests/compile/fusions_e2e/
+  commands:
+    - nvidia-smi
+    # Run all models and attn backends but only Inductor partition and native custom ops
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and not +quant_fp8"
+
+- label: Fusion E2E TP2 AR-RMS Config Sweep (H100)
+  timeout_in_minutes: 40
+  working_dir: "/vllm-workspace/"
+  device: h100
+  num_devices: 2
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/compilation/
+    # can affect pattern matching
+    - vllm/model_executor/layers/layernorm.py
+    - vllm/model_executor/layers/activation.py
+    - vllm/model_executor/layers/attention/attention.py
+    - vllm/model_executor/layers/quantization/input_quant_fp8.py
+    - tests/compile/fusions_e2e/
+  commands:
+    - nvidia-smi
+    # Run just llama3 (fp4 & fp8 & bf16) for all config combinations
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "llama-3"
+
+- label: Fusion E2E TP2 AsyncTP Config Sweep (H100)
+  timeout_in_minutes: 40
+  working_dir: "/vllm-workspace/"
+  device: h100
+  num_devices: 2
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/compilation/
+    # can affect pattern matching
+    - vllm/model_executor/layers/layernorm.py
+    - vllm/model_executor/layers/activation.py
+    - vllm/model_executor/layers/attention/attention.py
+    - vllm/model_executor/layers/quantization/input_quant_fp8.py
+    - tests/compile/fusions_e2e/
+  commands:
+    - nvidia-smi
+    # Run just llama3 (fp8 & bf16) for all config combinations
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "llama-3"
+
+- label: Fusion E2E TP2 (B200)
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/"
+  device: b200
+  num_devices: 2
+  source_file_dependencies:
+    - csrc/quantization/
+    - vllm/model_executor/
+    - vllm/v1/attention/
+    - vllm/compilation/
+    - tests/compile/fusions_e2e/
+  commands:
+    - nvidia-smi
+    # Run all models and attn backends but only Inductor partition and native custom ops
+    # for ar-rms-quant-fp4, also sweep llama3
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py -k "inductor_partition and not +rms_norm and not +quant_fp8" -k "Llama-3.1-8B-Instruct-FP4"
+    - pytest -v -s tests/compile/fusions_e2e/test_tp2_async_tp.py -k "inductor_partition and not +rms_norm and not +quant_fp8"

.buildkite/test_areas/distributed.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Distributed Comm Ops
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
+  num_devices: 2
   source_file_dependencies:
   - vllm/distributed
   - tests/distributed
@@ -16,9 +16,9 @@ steps:
   - pytest -v -s distributed/test_shm_storage.py
 
 - label: Distributed (2 GPUs)
-  timeout_in_minutes: 90
+  timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
+  num_devices: 2
   source_file_dependencies:
   - vllm/compilation/
   - vllm/distributed/
@@ -47,14 +47,13 @@ steps:
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
-  - pytest -v -s distributed/test_sequence_parallel.py
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s v1/worker/test_worker_memory_snapshot.py
 
 - label: Distributed Tests (4 GPUs)
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 4
+  num_devices: 4
   source_file_dependencies:
   - vllm/distributed/
   - tests/distributed/test_utils
@@ -103,8 +102,8 @@ steps:
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
-  gpu: h100
-  num_gpus: 8
+  device: h100
+  num_devices: 8
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - examples/offline_inference/torchrun_dp_example.py
@@ -120,9 +119,9 @@ steps:
   - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
 
 - label: Distributed Tests (4 GPUs)(A100)
-  gpu: a100
+  device: a100
   optional: true
-  num_gpus: 4
+  num_devices: 4
   source_file_dependencies:
   - vllm/
   commands:
@@ -133,26 +132,22 @@ steps:
   - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   - pytest -v -s -x lora/test_mixtral.py
 
-- label: Distributed Tests (2 GPUs)(H200)
-  gpu: h200
+- label: Distributed Tests (2 GPUs)(H100)
+  timeout_in_minutes: 15
+  device: h100
   optional: true
   working_dir: "/vllm-workspace/"
-  num_gpus: 2
+  num_devices: 2
   commands:
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_async_tp.py
-    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/distributed/test_sequence_parallel.py
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
+    - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
 - label: Distributed Tests (2 GPUs)(B200)
-  gpu: b200
+  device: b200
   optional: true
   working_dir: "/vllm-workspace/"
-  num_gpus: 2
+  num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
     - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
@@ -161,8 +156,9 @@ steps:
 - label: 2 Node Test (4 GPUs)
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
+  num_devices: 2
   num_nodes: 2
+  no_plugin: true
   source_file_dependencies:
   - vllm/distributed/
   - vllm/engine/
@@ -171,23 +167,34 @@ steps:
   - tests/distributed/
   - tests/examples/offline_inference/data_parallel.py
   commands:
-    - ./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:0bec63fa317e1fbd62e19b0fc31c43c81bf89077 "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=0 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py" "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code"
+    - ./.buildkite/scripts/run-multi-node-test.sh /vllm-workspace/tests 2 2 $IMAGE_TAG "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=0 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py && VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py" "VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed' && NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed' && python3 ../examples/offline_inference/data_parallel.py -dp=2 -tp=1 --dp-num-nodes=2 --dp-node-rank=1 --dp-master-addr=192.168.10.10 --dp-master-port=12345 --enforce-eager --trust-remote-code"
 
 - label: Distributed NixlConnector PD accuracy (4 GPUs)
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 4
+  num_devices: 4
   source_file_dependencies:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
     - tests/v1/kv_connector/nixl_integration/
   commands:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - bash v1/kv_connector/nixl_integration/tp_config_sweep_accuracy_test.sh
+    - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+
+- label: DP EP Distributed NixlConnector PD accuracy tests (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_connector.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
+    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: Pipeline + Context Parallelism (4 GPUs))
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 4
+  num_devices: 4
   source_file_dependencies:
   - vllm/distributed/
   - vllm/engine/
@@ -196,4 +203,4 @@ steps:
   - tests/distributed/
   commands:
   - pytest -v -s distributed/test_pp_cudagraph.py
-  - pytest -v -s distributed/test_pipeline_parallel.py
+  - pytest -v -s distributed/test_pipeline_parallel.py

.buildkite/test_areas/e2e_integration.yaml

@@ -4,27 +4,27 @@ depends_on:
 steps:
 - label: DeepSeek V2-Lite Accuracy
   timeout_in_minutes: 60
-  gpu: h100
+  device: h100
   optional: true
-  num_gpus: 4
+  num_devices: 4
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
 
 - label: Qwen3-30B-A3B-FP8-block Accuracy
   timeout_in_minutes: 60
-  gpu: h100
+  device: h100
   optional: true
-  num_gpus: 4
+  num_devices: 4
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
 
 - label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
   timeout_in_minutes: 60
-  gpu: b200
+  device: b200
   optional: true
-  num_gpus: 2
+  num_devices: 2
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
@@ -32,28 +32,12 @@ steps:
 - label: Prime-RL Integration (2 GPUs)
   timeout_in_minutes: 30
   optional: true
-  num_gpus: 2
+  soft_fail: true
+  num_devices: 2
   working_dir: "/vllm-workspace"
   source_file_dependencies:
   - vllm/
   - .buildkite/scripts/run-prime-rl-test.sh
   commands:
+    - nvidia-smi
     - bash .buildkite/scripts/run-prime-rl-test.sh
-
-- label: DeepSeek V2-Lite Async EPLB Accuracy
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_async_eplb.sh 0.25 1319 8030
-
-- label: Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040

.buildkite/test_areas/engine.yaml

@@ -23,4 +23,8 @@ steps:
     # TODO: accuracy does not match, whether setting
     # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
     - pytest -v -s v1/e2e
-    - pytest -v -s v1/engine
+    # Run this test standalone for now;
+    # need to untangle use (implicit) use of spawn/fork across the tests.
+    - pytest -v -s v1/engine/test_preprocess_error_handling.py
+    # Run the rest of v1/engine tests
+    - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py

.buildkite/test_areas/entrypoints.yaml

@@ -10,7 +10,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
 
 - label: Entrypoints Integration (LLM)
   timeout_in_minutes: 40
@@ -25,7 +25,7 @@ steps:
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
-- label: Entrypoints Integration (API Server)
+- label: Entrypoints Integration (API Server 1)
   timeout_in_minutes: 130
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -34,10 +34,24 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/openai/test_collective_rpc.py # PYTHONPATH is needed to import custom Worker extension
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_collective_rpc.py --ignore=entrypoints/openai/tool_parsers/
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
   - pytest -v -s entrypoints/test_chat_utils.py
 
+- label: Entrypoints Integration (API Server 2)
+  timeout_in_minutes: 130
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/tool_use
+  - tests/entrypoints/sleep
+  - tests/entrypoints/instrumentator
+  - tests/entrypoints/rpc
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
+  - pytest -v -s entrypoints/instrumentator
+  - pytest -v -s entrypoints/sleep
+  - pytest -v -s tool_use
 
 - label: Entrypoints Integration (Pooling)
   timeout_in_minutes: 50
@@ -49,6 +63,14 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
 
+- label: Entrypoints Integration (Responses API)
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai/responses
+  commands:
+  - pytest -v -s entrypoints/openai/responses
 
 - label: Entrypoints V1
   timeout_in_minutes: 50

.buildkite/test_areas/expert_parallelism.yaml

@@ -14,7 +14,7 @@ steps:
 - label: EPLB Execution
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 4
+  num_devices: 4
   source_file_dependencies:
   - vllm/distributed/eplb
   - tests/distributed/test_eplb_execute.py

.buildkite/test_areas/kernels.yaml

@@ -15,8 +15,9 @@ steps:
   timeout_in_minutes: 35
   source_file_dependencies:
   - csrc/attention/
-  - vllm/attention
   - vllm/v1/attention
+    # TODO: remove this dependency (https://github.com/vllm-project/vllm/issues/32267)
+  - vllm/model_executor/layers/attention
   - tests/kernels/attention
   commands:
     - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
@@ -57,8 +58,8 @@ steps:
 
 - label: Kernels DeepGEMM Test (H100)
   timeout_in_minutes: 45
-  gpu: h100
-  num_gpus: 1
+  device: h100
+  num_devices: 1
   source_file_dependencies:
   - tools/install_deepgemm.sh
   - vllm/utils/deep_gemm.py
@@ -77,7 +78,7 @@ steps:
 - label: Kernels (B200)
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
-  gpu: b200
+  device: b200
   # optional: true
   source_file_dependencies:
   - csrc/quantization/fp4/
@@ -85,13 +86,13 @@ steps:
   - csrc/quantization/cutlass_w8a8/moe/
   - vllm/model_executor/layers/fused_moe/cutlass_moe.py
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/v1/attention/backends/mla/cutlass_mla.py
   - vllm/v1/attention/backends/mla/flashinfer_mla.py
+  - vllm/v1/attention/selector.py
   - vllm/platforms/cuda.py
-  - vllm/attention/selector.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
@@ -114,4 +115,55 @@ steps:
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
-    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
+    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
+    # e2e
+    - pytest -v -s tests/models/quantization/test_nvfp4.py
+
+- label: Kernels Helion Test
+  timeout_in_minutes: 30
+  device: h100
+  source_file_dependencies:
+  - vllm/utils/import_utils.py
+  - tests/kernels/helion/
+  commands:
+    - pip install helion
+    - pytest -v -s kernels/helion/
+
+ 
+- label: Kernels FP8 MoE Test (1 H100)
+  timeout_in_minutes: 90
+  device: h100
+  num_devices: 1
+  optional: true
+  commands:
+    - pytest -v -s kernels/moe/test_cutlass_moe.py
+    - pytest -v -s kernels/moe/test_flashinfer.py
+    - pytest -v -s kernels/moe/test_gpt_oss_triton_kernels.py
+    - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py
+    - pytest -v -s kernels/moe/test_moe.py
+    # - pytest -v -s kernels/moe/test_block_fp8.py - failing on main
+    - pytest -v -s kernels/moe/test_block_int8.py
+    - pytest -v -s kernels/moe/test_triton_moe_no_act_mul.py
+    - pytest -v -s kernels/moe/test_triton_moe_ptpc_fp8.py
+
+- label: Kernels FP8 MoE Test (2 H100s)
+  timeout_in_minutes: 90
+  device: h100
+  num_devices: 2
+  optional: true
+  commands:
+    - pytest -v -s kernels/moe/test_deepep_deepgemm_moe.py
+    - pytest -v -s kernels/moe/test_deepep_moe.py
+    - pytest -v -s kernels/moe/test_pplx_cutlass_moe.py
+    # - pytest -v -s kernels/moe/test_pplx_moe.py - failing on main
+  
+- label: Kernels Fp4 MoE Test (B200)
+  timeout_in_minutes: 60
+  device: b200
+  num_devices: 1
+  optional: true
+  commands:
+    - pytest -v -s kernels/moe/test_cutedsl_moe.py
+    - pytest -v -s kernels/moe/test_flashinfer_moe.py
+    - pytest -v -s kernels/moe/test_nvfp4_moe.py
+    - pytest -v -s kernels/moe/test_ocp_mx_moe.py

.buildkite/test_areas/lm_eval.yaml

@@ -9,12 +9,12 @@ steps:
   - vllm/model_executor/layers/quantization
   autorun_on_main: true
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
 
 - label: LM Eval Large Models (4 GPUs)(A100)
-  gpu: a100
+  device: a100
   optional: true
-  num_gpus: 4
+  num_devices: 4
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - csrc/
@@ -24,9 +24,9 @@ steps:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
 - label: LM Eval Large Models (4 GPUs)(H100)
-  gpu: h100
+  device: h100
   optional: true
-  num_gpus: 4
+  num_devices: 4
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - csrc/
@@ -37,10 +37,39 @@ steps:
 
 - label: LM Eval Small Models (B200)
   timeout_in_minutes: 120
-  gpu: b200
+  device: b200
   optional: true
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt --tp-size=1
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
+
+- label: LM Eval Large Models (H200)
+  timeout_in_minutes: 60
+  device: h200
+  optional: true
+  num_devices: 8
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
+
+- label: MoE Refactor Integration Test (H100 - TEMPORARY)
+  device: h100
+  optional: true
+  num_devices: 2
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-h100.txt
+  
+- label: MoE Refactor Integration Test (B200 - TEMPORARY)
+  device: b200
+  optional: true
+  num_devices: 2
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor/config-b200.txt
+
+- label: MoE Refactor Integration Test (B200 DP - TEMPORARY)
+  device: b200
+  optional: true
+  num_devices: 2
+  commands:
+    - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=evals/gsm8k/configs/moe-refactor-dp-ep/config-b200.txt

.buildkite/test_areas/lora.yaml

@@ -14,7 +14,7 @@ steps:
 
 - label: LoRA TP (Distributed)
   timeout_in_minutes: 30
-  num_gpus: 4
+  num_devices: 4
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -22,6 +22,8 @@ steps:
     # FIXIT: find out which code initialize cuda before running the test
     # before the fix, we need to use spawn to test it
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    # Alot of these tests are on the edge of OOMing
+    - export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
     # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py

.buildkite/test_areas/misc.yaml

@@ -16,7 +16,7 @@ steps:
     - pytest -v -s v1/sample
     - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
-    - pytest -v -s v1/spec_decode
+    - pytest -v -s -m 'not slow_test' v1/spec_decode
     - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/metrics
     - pytest -v -s v1/test_oracle.py
@@ -27,11 +27,12 @@ steps:
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
 - label: V1 Others (CPU)
-  depends_on: ~
+  depends_on:
+    - image-build-cpu
   source_file_dependencies:
     - vllm/
     - tests/v1
-  no_gpu: true
+  device: cpu
   commands:
     # split the test to avoid interference
     - pytest -v -s -m 'cpu_test' v1/core
@@ -82,7 +83,7 @@ steps:
 
 - label: Metrics, Tracing (2 GPUs)
   timeout_in_minutes: 20
-  num_gpus: 2
+  num_devices: 2
   source_file_dependencies:
   - vllm/
   - tests/v1/tracing
@@ -114,24 +115,27 @@ steps:
   - pytest -v -s utils_
 
 - label: Async Engine, Inputs, Utils, Worker, Config (CPU)
-  depends_on: ~
+  depends_on: 
+  - image-build-cpu
   timeout_in_minutes: 30
   source_file_dependencies:
   - vllm/
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/multimodal
+  - tests/renderers
   - tests/standalone_tests/lazy_imports.py
   - tests/tokenizers_
   - tests/tool_parsers
   - tests/transformers_utils
   - tests/config
-  no_gpu: true
+  device: cpu
   commands:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s -m 'cpu_test' multimodal
+  - pytest -v -s renderers
   - pytest -v -s tokenizers_
   - pytest -v -s tool_parsers
   - pytest -v -s transformers_utils
@@ -140,7 +144,7 @@ steps:
 - label: GPT-OSS Eval (B200)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
-  gpu: b200
+  device: b200
   optional: true
   source_file_dependencies:
   - tests/evals/gpt_oss
@@ -153,7 +157,7 @@ steps:
 
 - label: Batch Invariance (H100)
   timeout_in_minutes: 25
-  gpu: h100
+  device: h100
   source_file_dependencies:
     - vllm/v1/attention
     - vllm/model_executor/layers
@@ -162,4 +166,18 @@ steps:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pip install pytest-timeout pytest-forked
     - pytest -v -s v1/determinism/test_batch_invariance.py
-    - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
+    - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
+  
+- label: Acceptance Length Test (Large Models) # optional
+  timeout_in_minutes: 25
+  gpu: h100
+  optional: true
+  num_gpus: 1
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/v1/spec_decode/
+  - vllm/model_executor/models/mlp_speculator.py
+  - tests/v1/spec_decode/test_acceptance_length.py
+  commands:
+    - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
+    - pytest -v -s v1/spec_decode/test_acceptance_length.py -m slow_test

.buildkite/test_areas/models_basic.yaml

@@ -9,6 +9,7 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/test_initialization.py
+  - tests/models/registry.py
   commands:
     # Run a subset of model initialization tests
     - pytest -v -s models/test_initialization.py::test_can_initialize_small_subset
@@ -20,6 +21,7 @@ steps:
   source_file_dependencies:
   - vllm/model_executor/models/
   - tests/models/test_initialization.py
+  - tests/models/registry.py
   commands:
     # Only when vLLM model source is modified - test initialization of a large
     # subset of supported models (the complement of the small subset in the above
@@ -37,12 +39,14 @@ steps:
     - pytest -v -s models/test_transformers.py models/test_registry.py
 
 - label: Basic Models Test (Other CPU) # 5min
+  depends_on: 
+  - image-build-cpu
   timeout_in_minutes: 10
   source_file_dependencies:
   - vllm/
   - tests/models/test_utils.py
   - tests/models/test_vision.py
-  no_gpu: true
+  device: cpu
   commands:
     - pytest -v -s models/test_utils.py models/test_vision.py
 

.buildkite/test_areas/models_distributed.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Distributed Model Tests (2 GPUs)
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
+  num_devices: 2
   source_file_dependencies:
   - vllm/model_executor/model_loader/sharded_state_loader.py
   - vllm/model_executor/models/

.buildkite/test_areas/models_multimodal.yaml

@@ -14,11 +14,13 @@ steps:
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
 - label: Multi-Modal Processor Test (CPU)
+  depends_on: 
+  - image-build-cpu
   timeout_in_minutes: 60
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
-  no_gpu: true
+  device: cpu
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py

.buildkite/test_areas/plugins.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Plugin Tests (2 GPUs)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
+  num_devices: 2
   source_file_dependencies:
   - vllm/plugins/
   - tests/plugins/

.buildkite/test_areas/pytorch.yaml

@@ -13,10 +13,12 @@ steps:
   # tests covered elsewhere.
   # Use `find` to launch multiple instances of pytest so that
   # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/ -maxdepth 1 -name 'test_*.py' -exec pytest -s -v {} \\;"
+  # However, find does not normally propagate error codes, so we combine it with xargs
+  # (using -0 for proper path handling)
+  - "find compile/ -maxdepth 1 -name 'test_*.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
 
 - label: PyTorch Fullgraph Smoke Test
-  timeout_in_minutes: 30
+  timeout_in_minutes: 35
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -28,16 +30,13 @@ steps:
   - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -exec pytest -s -v {} \\;"
 
 - label: PyTorch Fullgraph
-  timeout_in_minutes: 40
+  timeout_in_minutes: 30
   source_file_dependencies:
   - vllm/
   - tests/compile
   commands:
     # fp8 kv scales not supported on sm89, tested on Blackwell instead
   - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
-    # Limit to no custom ops to reduce running time
-    # Wrap with quotes to escape yaml and avoid starting -k string with a -
-  - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
 
 - label: Pytorch Nightly Dependency Override Check # 2min
   # if this test fails, it means the nightly torch version is not compatible with some

.buildkite/test_areas/quantization.yaml

@@ -16,14 +16,14 @@ steps:
   # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
   # we can only upgrade after this is resolved
   # TODO(jerryzh168): resolve the above comment
-  - uv pip install --system torchao==0.13.0 --index-url https://download.pytorch.org/whl/cu129
+  - uv pip install --system torchao==0.14.1 --index-url https://download.pytorch.org/whl/cu129
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
 - label: Quantized MoE Test (B200)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
-  gpu: b200
+  device: b200
   source_file_dependencies:
   - tests/quantization/test_blackwell_moe.py
   - vllm/model_executor/models/deepseek_v2.py

.buildkite/test_areas/tool_use.yaml

@@ -1,13 +0,0 @@
-group: Tool use
-depends_on: 
-  - image-build
-steps:
-- label: OpenAI-Compatible Tool Use
-  timeout_in_minutes: 35
-  mirror_hardwares: [amdexperimental]
-  fast_check: false
-  source_file_dependencies:
-    - vllm/
-    - tests/tool_use
-  commands:
-    - pytest -v -s tool_use

.buildkite/test_areas/weight_loading.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Weight Loading Multiple GPU  # 33min
   timeout_in_minutes: 45
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
+  num_devices: 2
   optional: true
   source_file_dependencies:
   - vllm/
@@ -15,8 +15,8 @@ steps:
 
 - label: Weight Loading Multiple GPU - Large Models # optional
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
-  gpu: a100
+  num_devices: 2
+  device: a100
   optional: true
   source_file_dependencies:
   - vllm/

.github/CODEOWNERS

@@ -2,9 +2,8 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/attention @LucasWilkinson
-/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @njhill
 /vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @njhill @22quinn
+/vllm/model_executor/layers/attention @LucasWilkinson
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
 /vllm/model_executor/layers/mamba @tdoublep
@@ -15,8 +14,9 @@
 /vllm/lora @jeejeelee
 /vllm/reasoning @aarnphm @chaunceyjiang
 /vllm/entrypoints @aarnphm @chaunceyjiang
+/vllm/tool_parsers @aarnphm @chaunceyjiang
 /vllm/compilation @zou3519 @youkaichao @ProExpertProg
-/vllm/distributed/kv_transfer @NickLucche @ApostaC
+/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
@@ -26,15 +26,18 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # vLLM V1
 /vllm/v1/attention @LucasWilkinson
+/vllm/v1/attention/backend.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @njhill
 /vllm/v1/attention/backends/mla @pavanimajety
 /vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
-/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC
+/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /vllm/v1/sample @22quinn @houseroad @njhill
 /vllm/v1/spec_decode @benchislett @luccafong
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
 /vllm/v1/kv_cache_interface.py @heheda12345
-/vllm/v1/offloading @ApostaC
+/vllm/v1/kv_offload @ApostaC @orozery
+/vllm/v1/worker/gpu/kv_connector.py @orozery
+/vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery
 
 # Model runner V2
 /vllm/v1/worker/gpu @WoosukKwon
@@ -53,13 +56,13 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
-/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC
+/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 /tests/models/language/generation/test_hybrid.py @tdoublep
 /tests/v1/kv_connector/nixl_integration @NickLucche
-/tests/v1/kv_connector @ApostaC
-/tests/v1/offloading @ApostaC
+/tests/v1/kv_connector @ApostaC @orozery
+/tests/v1/kv_offload @ApostaC @orozery
 /tests/v1/determinism @yewentao256 
 
 # Transformers modeling backend
@@ -116,15 +119,15 @@ mkdocs.yaml @hmellor
 /vllm/transformers_utils/tokenizers/mistral.py @patrickvonplaten
 
 # Kernels
-/vllm/attention/ops/chunked_prefill_paged_decode.py @tdoublep
-/vllm/attention/ops/triton_unified_attention.py @tdoublep
+/vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
+/vllm/v1/attention/ops/triton_unified_attention.py @tdoublep
 
 # ROCm related: specify owner with write access to notify AMD folks for careful code review
 /vllm/**/*rocm* @tjtanaa
 /docker/Dockerfile.rocm* @gshtras @tjtanaa
 /vllm/v1/attention/backends/rocm*.py @gshtras @tjtanaa
 /vllm/v1/attention/backends/mla/rocm*.py @gshtras @tjtanaa
-/vllm/attention/ops/rocm*.py @gshtras @tjtanaa
+/vllm/v1/attention/ops/rocm*.py @gshtras @tjtanaa
 /vllm/model_executor/layers/fused_moe/rocm*.py @gshtras @tjtanaa
 /csrc/rocm @gshtras @tjtanaa
 /requirements/*rocm* @tjtanaa
@@ -152,7 +155,7 @@ mkdocs.yaml @hmellor
 /vllm/entrypoints/pooling @noooop
 /vllm/config/pooler.py @noooop
 /vllm/pooling_params.py @noooop
-/vllm/model_executor/layers/pooler.py @noooop
+/vllm/model_executor/layers/pooler @noooop
 
 # Security guide and policies
 /docs/usage/security.md @russellb

.github/mergify.yml

@@ -222,10 +222,10 @@ pull_request_rules:
       - files~=^csrc/rocm/
       - files~=^docker/Dockerfile.rocm
       - files~=^requirements/rocm.*\.txt
-      - files~=^vllm/attention/backends/rocm.*\.py
-      - files~=^vllm/attention/ops/rocm.*\.py
       - files~=^vllm/model_executor/layers/fused_moe/rocm.*\.py
+      - files~=^vllm/v1/attention/backends/rocm.*\.py
       - files~=^vllm/v1/attention/backends/mla/rocm.*\.py
+      - files~=^vllm/v1/attention/ops/rocm.*\.py
       - files~=^tests/kernels/.*_rocm.*\.py
       - files=vllm/platforms/rocm.py
       - title~=(?i)AMD
@@ -235,6 +235,20 @@ pull_request_rules:
       add:
         - rocm
 
+- name: label-cpu
+  description: Automatically apply cpu label
+  conditions:
+    - label != stale
+    - files~=^(?!.*kv_offload)(?!.*cpu_offload).*\bcpu.*
+  actions:
+    label:
+      add:
+        - cpu
+    assign:
+      users:
+        - "fadara01"
+        - "aditew01"
+
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:
@@ -335,6 +349,18 @@ pull_request_rules:
       add:
         - tool-calling
 
+- name: auto-rebase if approved, ready, and 40 commits behind main
+  conditions:
+    - base = main
+    - label=ready
+    - "#approved-reviews-by >= 1"
+    - "#commits-behind >= 40"
+    - -closed
+    - -draft
+    - -conflict
+  actions:
+    rebase: {}
+
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
     - label != stale
@@ -388,6 +414,18 @@ pull_request_rules:
       remove:
         - needs-rebase
 
+- name: label-bug
+  description: Automatically apply bug label
+  conditions:
+    - label != stale
+    - or:
+      - title~=(?i)\bbug\b
+      - title~=(?i)\bbugfix\b
+  actions:
+    label:
+      add:
+        - bug
+
 - name: label-kv-connector
   description: Automatically apply kv-connector label
   conditions:

.github/workflows/macos-smoke-test.yml

@@ -29,8 +29,9 @@ jobs:
 
       - name: Install dependencies and build vLLM
         run: |
+          uv pip install -r requirements/cpu-build.txt --index-strategy unsafe-best-match
           uv pip install -r requirements/cpu.txt --index-strategy unsafe-best-match
-          uv pip install -e .
+          uv pip install -e . --no-build-isolation
         env:
           CMAKE_BUILD_PARALLEL_LEVEL: 4
 

.gitignore

@@ -7,6 +7,9 @@ vllm/vllm_flash_attn/*
 # OpenAI triton kernels copied from source
 vllm/third_party/triton_kernels/*
 
+# FlashMLA interface copied from source
+vllm/third_party/flashmla/flash_mla_interface.py
+
 # triton jit
 .triton
 
@@ -191,6 +194,9 @@ CLAUDE.md
 AGENTS.md
 .codex/
 
+# Cursor
+.cursor/
+
 # DS Store
 .DS_Store
 
@@ -227,3 +233,8 @@ ep_kernels_workspace/
 
 # Allow tracked library source folders under submodules (e.g., benchmarks/lib)
 !vllm/benchmarks/lib/
+
+# Generated gRPC protobuf files (compiled at build time from vllm_engine.proto)
+vllm/grpc/vllm_engine_pb2.py
+vllm/grpc/vllm_engine_pb2_grpc.py
+vllm/grpc/vllm_engine_pb2.pyi

.pre-commit-config.yaml

@@ -147,6 +147,17 @@ repos:
     entry: python tools/pre_commit/validate_config.py
     language: python
     additional_dependencies: [regex]
+  - id: validate-docker-versions
+    name: Validate docker/versions.json matches Dockerfile
+    entry: python tools/generate_versions_json.py --check
+    language: python
+    files: ^docker/(Dockerfile|versions\.json)$
+    pass_filenames: false
+    additional_dependencies: [dockerfile-parse]
+  - id: attention-backend-docs
+    name: Check attention backend documentation is up to date
+    entry: python tools/pre_commit/generate_attention_backend_docs.py --check
+    language: python
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -56,8 +56,8 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.9.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.9.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.9.1")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.9.1")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -282,6 +282,7 @@ endif()
 set(VLLM_EXT_SRC
   "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
   "csrc/cache_kernels.cu"
+  "csrc/cache_kernels_fused.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
   "csrc/attention/merge_attn_states.cu"
@@ -357,6 +358,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # marlin arches for fp16 output
   cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
+  # marlin has limited support for turing
+  cuda_archs_loose_intersection(MARLIN_SM75_ARCHS "7.5" "${CUDA_ARCHS}")
   # marlin arches for bf16 output (we need 9.0 for bf16 atomicAdd PTX)
   cuda_archs_loose_intersection(MARLIN_BF16_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
   # marlin arches for fp8 input
@@ -364,15 +367,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # - sm90 and sm100 don't support QMMA.16832.F32.E4M3.E4M3 SAAS instruction
   # so we only enable fp8 computation for SM89 (e.g. RTX 40x0)  and 12.0 (e.g. RTX 50x0)
   cuda_archs_loose_intersection(MARLIN_FP8_ARCHS "8.9;12.0" "${CUDA_ARCHS}")
+  # marlin arches for other files
+  cuda_archs_loose_intersection(MARLIN_OTHER_ARCHS "7.5;8.0+PTX" "${CUDA_ARCHS}")
 
-  if (MARLIN_ARCHS)
+  if (MARLIN_OTHER_ARCHS)
 
     #
     # For the Marlin kernels we automatically generate sources for various
     # preselected input type pairs and schedules.
     # Generate sources:
     set(MARLIN_GEN_SCRIPT
-      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/marlin/generate_kernels.py)
     file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
     list(JOIN CUDA_ARCHS "," CUDA_ARCHS_STR)
     set(MARLIN_GEN_SCRIPT_HASH_AND_ARCH "${MARLIN_GEN_SCRIPT_HASH}(ARCH:${CUDA_ARCHS_STR})")
@@ -406,28 +411,42 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       message(STATUS "Marlin generation script has not changed, skipping generation.")
     endif()
 
-    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/sm80_kernel_*_float16.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
-      CUDA_ARCHS "${MARLIN_ARCHS}")
-    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-      set_source_files_properties(${MARLIN_TEMPLATE_KERNEL_SRC}
-        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
-    endif()
-    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
+    if (MARLIN_ARCHS)
+      file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/marlin/sm80_kernel_*_float16.cu")
+      set_gencode_flags_for_srcs(
+        SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
+        CUDA_ARCHS "${MARLIN_ARCHS}")
+      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+        set_source_files_properties(${MARLIN_TEMPLATE_KERNEL_SRC}
+          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+      endif()
+      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
 
-    file(GLOB MARLIN_TEMPLATE_BF16_KERNEL_SRC "csrc/quantization/gptq_marlin/sm80_kernel_*_bfloat16.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_TEMPLATE_BF16_KERNEL_SRC}"
-      CUDA_ARCHS "${MARLIN_BF16_ARCHS}")
-    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-      set_source_files_properties(${MARLIN_TEMPLATE_BF16_KERNEL_SRC}
-        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+      file(GLOB MARLIN_TEMPLATE_BF16_KERNEL_SRC "csrc/quantization/marlin/sm80_kernel_*_bfloat16.cu")
+      set_gencode_flags_for_srcs(
+        SRCS "${MARLIN_TEMPLATE_BF16_KERNEL_SRC}"
+        CUDA_ARCHS "${MARLIN_BF16_ARCHS}")
+      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+        set_source_files_properties(${MARLIN_TEMPLATE_BF16_KERNEL_SRC}
+          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+      endif()
+      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_BF16_KERNEL_SRC})
+    endif()
+
+    if (MARLIN_SM75_ARCHS) 
+      file(GLOB MARLIN_TEMPLATE_SM75_KERNEL_SRC "csrc/quantization/marlin/sm75_kernel_*.cu")
+      set_gencode_flags_for_srcs(
+        SRCS "${MARLIN_TEMPLATE_SM75_KERNEL_SRC}"
+        CUDA_ARCHS "${MARLIN_SM75_ARCHS}")
+      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+        set_source_files_properties(${MARLIN_TEMPLATE_SM75_KERNEL_SRC}
+          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+      endif()
+      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_SM75_KERNEL_SRC})
     endif()
-    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_BF16_KERNEL_SRC})
 
     if (MARLIN_FP8_ARCHS) 
-      file(GLOB MARLIN_TEMPLATE_FP8_KERNEL_SRC "csrc/quantization/gptq_marlin/sm89_kernel_*.cu")
+      file(GLOB MARLIN_TEMPLATE_FP8_KERNEL_SRC "csrc/quantization/marlin/sm89_kernel_*.cu")
       set_gencode_flags_for_srcs(
         SRCS "${MARLIN_TEMPLATE_FP8_KERNEL_SRC}"
         CUDA_ARCHS "${MARLIN_FP8_ARCHS}")
@@ -439,21 +458,20 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
 
     set(MARLIN_SRCS
-       "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
-       "csrc/quantization/gptq_marlin/gptq_marlin.cu"
-       "csrc/quantization/gptq_marlin/marlin_int4_fp8_preprocess.cu"
-       "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
-       "csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
+       "csrc/quantization/marlin/marlin.cu"
+       "csrc/quantization/marlin/marlin_int4_fp8_preprocess.cu"
+       "csrc/quantization/marlin/gptq_marlin_repack.cu"
+       "csrc/quantization/marlin/awq_marlin_repack.cu")
     set_gencode_flags_for_srcs(
       SRCS "${MARLIN_SRCS}"
-      CUDA_ARCHS "${MARLIN_ARCHS}")
+      CUDA_ARCHS "${MARLIN_OTHER_ARCHS}")
     if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-      set_source_files_properties("csrc/quantization/gptq_marlin/gptq_marlin.cu"
+      set_source_files_properties(${MARLIN_SRCS}
         PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
     endif()
     list(APPEND VLLM_EXT_SRC "${MARLIN_SRCS}")
 
-    message(STATUS "Building Marlin kernels for archs: ${MARLIN_ARCHS}")
+    message(STATUS "Building Marlin kernels for archs: ${MARLIN_OTHER_ARCHS}")
   else()
     message(STATUS "Not building Marlin kernels as no compatible archs found"
                    " in CUDA target architectures")
@@ -781,24 +799,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   else()
     cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
-    message(STATUS "Building blockwise_scaled_group_mm_sm100 for archs: ${SCALED_MM_ARCHS}")
-  else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-      message(STATUS "Not building blockwise_scaled_group_mm_sm100 kernels as CUDA Compiler version is "
-                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
-                     "if you intend on running FP8 quantized MoE models on Blackwell.")
-    else()
-      message(STATUS "Not building blockwise_scaled_group_mm_sm100 as no compatible archs found "
-                     "in CUDA target architectures")
-    endif()
-  endif()
 
   #
   # Machete kernels
@@ -980,12 +980,16 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # note that we always set `use_atomic_add=False` for moe marlin now,
   # so we don't need 9.0 for bf16 atomicAdd PTX
   cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
+  # moe marlin has limited support for turing
+  cuda_archs_loose_intersection(MARLIN_MOE_SM75_ARCHS "7.5" "${CUDA_ARCHS}")
   # moe marlin arches for fp8 input
   # - sm80 doesn't support fp8 computation
   # - sm90 and sm100 don't support QMMA.16832.F32.E4M3.E4M3 SAAS instruction
   # so we only enable fp8 computation for SM89 (e.g. RTX 40x0)  and 12.0 (e.g. RTX 50x0)
   cuda_archs_loose_intersection(MARLIN_MOE_FP8_ARCHS "8.9;12.0" "${CUDA_ARCHS}")
-  if (MARLIN_MOE_ARCHS)
+  # moe marlin arches for other files
+  cuda_archs_loose_intersection(MARLIN_MOE_OTHER_ARCHS "7.5;8.0+PTX" "${CUDA_ARCHS}")
+  if (MARLIN_MOE_OTHER_ARCHS)
 
     #
     # For the Marlin MOE kernels we automatically generate sources for various
@@ -1026,16 +1030,29 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
     endif()
 
-    file(GLOB MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/sm80_kernel_*.cu")
-    list(APPEND MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/ops.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_MOE_SRC}"
-      CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
-    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-      set_source_files_properties(${MARLIN_MOE_SRC}
-        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+    if (MARLIN_MOE_ARCHS)
+      file(GLOB MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/sm80_kernel_*.cu")
+      set_gencode_flags_for_srcs(
+        SRCS "${MARLIN_MOE_SRC}"
+        CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
+      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+        set_source_files_properties(${MARLIN_MOE_SRC}
+          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+      endif()
+      list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SRC})
+    endif()
+
+    if (MARLIN_MOE_SM75_ARCHS) 
+      file(GLOB MARLIN_MOE_SM75_SRC "csrc/moe/marlin_moe_wna16/sm75_kernel_*.cu")
+      set_gencode_flags_for_srcs(
+        SRCS "${MARLIN_MOE_SM75_SRC}"
+        CUDA_ARCHS "${MARLIN_MOE_SM75_ARCHS}")
+      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+        set_source_files_properties(${MARLIN_MOE_SM75_SRC}
+          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+      endif()
+      list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SM75_SRC})
     endif()
-    list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SRC})
 
     if (MARLIN_MOE_FP8_ARCHS)
       file(GLOB MARLIN_MOE_FP8_SRC "csrc/moe/marlin_moe_wna16/sm89_kernel_*.cu")
@@ -1049,7 +1066,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_FP8_SRC})
     endif()
 
-    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
+    set(MARLIN_MOE_OTHER_SRC "csrc/moe/marlin_moe_wna16/ops.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${MARLIN_MOE_OTHER_SRC}"
+      CUDA_ARCHS "${MARLIN_MOE_OTHER_ARCHS}")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
+      set_source_files_properties(${MARLIN_MOE_OTHER_SRC}
+        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
+    endif()
+    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_OTHER_SRC}")
+
+    message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_OTHER_ARCHS}")
   else()
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
                    " in CUDA target architectures")

README.md

@@ -14,51 +14,8 @@ Easy, fast, and cheap LLM serving for everyone
 | <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
----
-Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundation.org/pytorch-conference/) and [Ray Summit, November 3-5](https://www.anyscale.com/ray-summit/2025) in San Francisco for our latest updates on vLLM and to meet the vLLM team! Register now for the largest vLLM community events of the year!
-
----
-
-*Latest News* 🔥
-
-- [2025/11] We hosted [vLLM Bangkok Meetup](https://luma.com/v0f647nv). We explored vLLM and LMCache inference and low-resource language adaptation with speakers from Embedded LLM, AMD, and Red Hat. Please find the meetup slides [here](https://drive.google.com/drive/folders/1H0DS57F8HQ5q3kSOSoRmucPJWL3E0A_X?usp=sharing).
-- [2025/11] We hosted [the first vLLM Europe Meetup in Zurich](https://luma.com/0gls27kb) focused on quantization, distributed inference, and reinforcement learning at scale with speakers from Mistral, IBM, and Red Hat. Please find the meetup slides [here](https://docs.google.com/presentation/d/1UC9PTLCHYXQpOmJDSFg6Sljra3iVXzc09DeEI7dnxMc/edit?usp=sharing) and recording [here](https://www.youtube.com/watch?v=6m6ZE6yVEDI)
-- [2025/11] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/xSrYXjNgr1HbCP4ExYNG1w) focusing on distributed inference and diverse accelerator support with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1nQJ8ZkLSjKxvu36sSHaceVXtttbLvvu-?usp=drive_link).
-- [2025/10] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/__xb4OyOsImz-9eAVrdlcg) focused on hands-on vLLM inference optimization! Please find the meetup slides [here](https://drive.google.com/drive/folders/1KqwjsFJLfEsC8wlDugnrR61zsWHt94Q6).
-- [2025/09] We hosted [vLLM Toronto Meetup](https://luma.com/e80e0ymm) focused on tackling inference at scale and speculative decoding with speakers from NVIDIA and Red Hat! Please find the meetup slides [here](https://docs.google.com/presentation/d/1IYJYmJcu9fLpID5N5RbW_vO0XLo0CGOR14IXOjB61V8/edit?usp=sharing).
-- [2025/08] We hosted [vLLM Shenzhen Meetup](https://mp.weixin.qq.com/s/k8ZBO1u2_2odgiKWH_GVTQ) focusing on the ecosystem around vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Ua2SVKVSu-wp5vou_6ElraDt2bnKhiEA).
-- [2025/08] We hosted [vLLM Singapore Meetup](https://www.sginnovate.com/event/vllm-sg-meet). We shared V1 updates, disaggregated serving and MLLM speedups with speakers from Embedded LLM, AMD, WekaIO, and A*STAR. Please find the meetup slides [here](https://drive.google.com/drive/folders/1ncf3GyqLdqFaB6IeB834E5TZJPLAOiXZ?usp=sharing).
-- [2025/08] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/pDmAXHcN7Iqc8sUKgJgGtg) focusing on building, developing, and integrating with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1OvLx39wnCGy_WKq8SiVKf7YcxxYI3WCH).
-- [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
-- [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
-
-<details>
-<summary>Previous News</summary>
-
-- [2025/08] We hosted [vLLM Korea Meetup](https://luma.com/cgcgprmh) with Red Hat and Rebellions! We shared the latest advancements in vLLM along with project spotlights from the vLLM Korea community. Please find the meetup slides [here](https://drive.google.com/file/d/1bcrrAE1rxUgx0mjIeOWT6hNe2RefC5Hm/view).
-- [2025/08] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/dgkWg1WFpWGO2jCdTqQHxA) focusing on large-scale LLM deployment! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Pid6NSFLU43DZRi0EaTcPgXsAzDvbBqF) and the recording [here](https://www.chaspark.com/#/live/1166916873711665152).
-- [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
-- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
-- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
-- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
-- [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
-- [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
-- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
-- [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
-- [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
-- [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
-- [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users!
-- [2024/09] We hosted [the sixth vLLM meetup](https://lu.ma/87q3nvnh) with NVIDIA! Please find the meetup slides [here](https://docs.google.com/presentation/d/1wrLGwytQfaOTd5wCGSPNhoaW3nq0E-9wqyP7ny93xRs/edit?usp=sharing).
-- [2024/07] We hosted [the fifth vLLM meetup](https://lu.ma/lp0gyjqr) with AWS! Please find the meetup slides [here](https://docs.google.com/presentation/d/1RgUD8aCfcHocghoP3zmXzck9vX3RCI9yfUAB2Bbcl4Y/edit?usp=sharing).
-- [2024/07] In partnership with Meta, vLLM officially supports Llama 3.1 with FP8 quantization and pipeline parallelism! Please check out our blog post [here](https://blog.vllm.ai/2024/07/23/llama31.html).
-- [2024/06] We hosted [the fourth vLLM meetup](https://lu.ma/agivllm) with Cloudflare and BentoML! Please find the meetup slides [here](https://docs.google.com/presentation/d/1iJ8o7V2bQEi0BFEljLTwc5G1S10_Rhv3beed5oB0NJ4/edit?usp=sharing).
-- [2024/04] We hosted [the third vLLM meetup](https://robloxandvllmmeetup2024.splashthat.com/) with Roblox! Please find the meetup slides [here](https://docs.google.com/presentation/d/1A--47JAK4BJ39t954HyTkvtfwn0fkqtsL8NGFuslReM/edit?usp=sharing).
-- [2024/01] We hosted [the second vLLM meetup](https://lu.ma/ygxbpzhl) with IBM! Please find the meetup slides [here](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing).
-- [2023/10] We hosted [the first vLLM meetup](https://lu.ma/first-vllm-meetup) with a16z! Please find the meetup slides [here](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing).
-- [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
-- [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
-
-</details>
+🔥 We have built a vllm website to help you get started with vllm. Please visit [vllm.ai](https://vllm.ai) to learn more.
+For events, please visit [vllm.ai/events](https://vllm.ai/events) to join us.
 
 ---
 
@@ -118,50 +75,6 @@ Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
 We welcome and value any contributions and collaborations.
 Please check out [Contributing to vLLM](https://docs.vllm.ai/en/latest/contributing/index.html) for how to get involved.
 
-## Sponsors
-
-vLLM is a community project. Our compute resources for development and testing are supported by the following organizations. Thank you for your support!
-
-<!-- Note: Please sort them in alphabetical order. -->
-<!-- Note: Please keep these consistent with docs/community/sponsors.md -->
-Cash Donations:
-
-- a16z
-- Dropbox
-- Sequoia Capital
-- Skywork AI
-- ZhenFund
-
-Compute Resources:
-
-- Alibaba Cloud
-- AMD
-- Anyscale
-- Arm
-- AWS
-- Crusoe Cloud
-- Databricks
-- DeepInfra
-- Google Cloud
-- IBM
-- Intel
-- Lambda Lab
-- Nebius
-- Novita AI
-- NVIDIA
-- Red Hat
-- Replicate
-- Roblox
-- RunPod
-- Trainy
-- UC Berkeley
-- UC San Diego
-- Volcengine
-
-Slack Sponsor: Anyscale
-
-We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). We plan to use the fund to support the development, maintenance, and adoption of vLLM.
-
 ## Citation
 
 If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs/2309.06180):
@@ -182,7 +95,7 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 - For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
 - For coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
 - For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
-- For collaborations and partnerships, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@lists.berkeley.edu)
+- For collaborations and partnerships, please contact us at [collaboration@vllm.ai](mailto:collaboration@vllm.ai)
 <!-- --8<-- [end:contact-us] -->
 
 ## Media Kit

RELEASE.md

@@ -1,47 +1,30 @@
 # Releasing vLLM
 
-vLLM releases offer a reliable version of the code base, packaged into a binary format that can be conveniently accessed via PyPI. These releases also serve as key milestones for the development team to communicate with the community about newly available features, improvements, and upcoming changes that could affect users, including potential breaking changes.
+vLLM releases offer a reliable version of the code base, packaged into a binary format that can be conveniently accessed via [PyPI](https://pypi.org/project/vllm). These releases also serve as key milestones for the development team to communicate with the community about newly available features, improvements, and upcoming changes that could affect users, including potential breaking changes.
 
-## Release Versioning
+## Release Cadence and Versioning
 
-vLLM uses a “right-shifted” versioning scheme where a new patch release is out every 2 weeks. And patch releases contain features and bug fixes (as opposed to semver where patch release contains only backwards-compatible bug fixes). When critical fixes need to be made, special release post1 is released.
+We aim to have a regular release every 2 weeks. Since v0.12.0, regular releases increment the minor version rather than patch version. The list of past releases can be found [here](https://vllm.ai/releases).
 
-* _major_ major architectural milestone and when incompatible API changes are made, similar to PyTorch 2.0.
-* _minor_ major features
-* _patch_ features and backwards-compatible bug fixes
-* _post1_ or _patch-1_ backwards-compatible bug fixes, either explicit or implicit post release
+Our version numbers are expressed in the form `vX.Y.Z`, where `X` is the major version, `Y` is the minor version, and `Z` is the patch version. They are incremented according to the following rules:
 
-## Release Cadence
+* _Major_ releases are reserved for architectural milestones involving sweeping API changes, similar to PyTorch 2.0.
+* _Minor_ releases correspond to regular releases, which include new features, bug fixes and other backwards-compatible changes.
+* _Patch_ releases correspond to special releases for new models, as well as emergency patches for critical performance, functionality and security issues.
 
-Patch release is released on bi-weekly basis. Post release 1-3 days after patch release and uses same branch as patch release.
-Following is the release cadence for year 2025. All future release dates below are tentative. Please note: Post releases are optional.
+This versioning scheme is similar to [SemVer](https://semver.org/) for compatibility purposes, except that backwards compatibility is only guaranteed for a limited number of minor releases (see our [deprecation policy](https://docs.vllm.ai/en/latest/contributing/deprecation_policy) for details).
 
-| Release Date | Patch release versions | Post Release versions |
-| --- | --- | --- |
-| Jan 2025 | 0.7.0 | --- |
-| Feb 2025 | 0.7.1, 0.7.2, 0.7.3  | --- |
-| Mar 2025 | 0.7.4, 0.7.5 | --- |
-| Apr 2025 | 0.7.6, 0.7.7 | --- |
-| May 2025 | 0.7.8, 0.7.9 | --- |
-| Jun 2025 | 0.7.10, 0.7.11 | --- |
-| Jul 2025 | 0.7.12, 0.7.13 | --- |
-| Aug 2025 | 0.7.14, 0.7.15 | --- |
-| Sep 2025 | 0.7.16, 0.7.17 | --- |
-| Oct 2025 | 0.7.18, 0.7.19 | --- |
-| Nov 2025 | 0.7.20, 0.7.21 | --- |
-| Dec 2025 | 0.7.22, 0.7.23 | --- |
-
-## Release branch
+## Release Branch
 
 Each release is built from a dedicated release branch.
 
-* For _major_, _minor_, _patch_ releases, the release branch cut is performed 1-2 days before release is live.
-* For post releases, previously cut release branch is reused
-* Release builds are triggered via push to RC tag like vX.Y.Z-rc1 . This enables us to build and test multiple RCs for each release.
-* Final tag : vX.Y.Z does not trigger the build but used for Release notes and assets.
-* After branch cut is created we monitor the main branch for any reverts and apply these reverts to a release branch.
+* For _major_ and _minor_ releases, the release branch cut is performed 1-2 days before release is live.
+* For _patch_ releases, previously cut release branch is reused.
+* Release builds are triggered via push to RC tag like `vX.Y.Z-rc1`. This enables us to build and test multiple RCs for each release.
+* Final tag: `vX.Y.Z` does not trigger the build but used for Release notes and assets.
+* After branch cut is created, we monitor the main branch for any reverts and apply these reverts to a release branch.
 
-## Release Cherry-Pick Criteria
+### Cherry-Pick Criteria
 
 After branch cut, we approach finalizing the release branch with clear criteria on what cherry picks are allowed in. Note: a cherry pick is a process to land a PR in the release branch after branch cut. These are typically limited to ensure that the team has sufficient time to complete a thorough round of testing on a stable code base.
 

benchmarks/attention_benchmarks/README.md

@@ -0,0 +1,266 @@
+# vLLM Attention Benchmarking Suite
+
+Fast, flexible benchmarking for vLLM attention and MLA backends with an extended batch specification grammar.
+
+## Quick Start
+
+```bash
+cd benchmarks/attention_benchmarks
+
+# Run a pre-configured benchmark
+python benchmark.py --config configs/mla_decode.yaml
+python benchmark.py --config configs/mla_mixed_batch.yaml
+python benchmark.py --config configs/speculative_decode.yaml
+python benchmark.py --config configs/standard_attention.yaml
+python benchmark.py --config configs/reorder_threshold.yaml
+
+# Or run custom benchmarks
+python benchmark.py \
+    --backends flash flashinfer \
+    --batch-specs "q2k" "8q1s1k" "2q2k_32q1s1k" \
+    --output-csv results.csv
+```
+
+## Simplified Batch Specification Grammar
+
+Express workloads concisely using query length and sequence length:
+
+```python
+"q2k"              # 2048-token prefill (q_len=2048, seq_len=2048)
+"q1s1k"            # Decode: 1 token with 1K sequence
+"8q1s1k"           # 8 decode requests
+"q4s1k"            # 4-token extend (e.g., spec decode)
+"2q2k_32q1s1k"     # Mixed: 2 prefills + 32 decodes
+"16q4s1k"          # 16 spec decode (4 tokens each)
+```
+
+### Grammar Rule
+
+```text
+Format: (<count>?) q<q_len>(k?) (s<seq_len>(k?))?
+
+- count:   Number of identical requests (optional, default=1)
+- q_len:   Query length (number of new tokens)
+- seq_len: Total sequence length (optional, defaults to q_len for prefill)
+- 'k':     Multiplies value by 1024
+
+Mixed batches: Use _ to combine (e.g., "2q2k_32q1s1k")
+```
+
+**Note**: Decode, prefill, and spec decode are just different query lengths - no special syntax needed!
+
+## Pre-configured Benchmarks
+
+The suite includes several pre-configured YAML benchmark configurations:
+
+### MLA Decode Benchmark
+
+Tests pure decode performance across MLA backends with varying batch sizes and sequence lengths.
+
+```bash
+python benchmark.py --config configs/mla_decode.yaml
+```
+
+### MLA Mixed Batch Benchmark
+
+Tests chunked prefill performance with mixed prefill + decode batches.
+
+```bash
+python benchmark.py --config configs/mla_mixed_batch.yaml
+```
+
+### Speculative Decoding Benchmark
+
+Tests speculative decode scenarios (K-token verification) and reorder_batch_threshold optimization.
+
+```bash
+python benchmark.py --config configs/speculative_decode.yaml
+```
+
+### Standard Attention Benchmark
+
+Tests standard attention backends (Flash/Triton/FlashInfer) with pure prefill, decode, and mixed batches.
+
+```bash
+python benchmark.py --config configs/standard_attention.yaml
+```
+
+### Reorder Threshold Study
+
+**Question:** At what query length does the prefill pipeline become faster than the decode pipeline?
+
+Tests query lengths from 1-1024 across 9 batch sizes to find the crossover point. Uses `decode_vs_prefill` mode to compare both pipelines for each query length.
+
+```bash
+python benchmark.py --config configs/reorder_threshold.yaml
+```
+
+---
+
+## Universal Benchmark
+
+The `benchmark.py` script handles **all** backends - both standard attention and MLA.
+
+### Standard Attention (Flash/Triton/FlashInfer)
+
+```bash
+python benchmark.py \
+    --backends flash triton flashinfer \
+    --batch-specs "q2k" "8q1s1k" "2q2k_32q1s1k" \
+    --num-layers 10 \
+    --repeats 5 \
+    --output-csv results.csv
+```
+
+### MLA Backends
+
+```bash
+# Compare all MLA backends
+python benchmark.py \
+    --backends cutlass_mla flashinfer_mla flashattn_mla flashmla \
+    --batch-specs "64q1s1k" "64q1s4k" \
+    --output-csv mla_results.csv
+```
+
+### Parameter Sweeps
+
+Use `--sweep-param` and `--sweep-values` to run parameter sweeps from the CLI:
+
+#### CUTLASS MLA num-splits Optimization
+
+**Question:** What is the optimal `num_kv_splits` for CUTLASS MLA?
+
+```bash
+python benchmark.py \
+    --backend cutlass_mla \
+    --batch-specs "64q1s1k" "64q1s4k" "64q1s16k" \
+    --sweep-param num_kv_splits \
+    --sweep-values 1 2 4 8 16 \
+    --output-json optimal_splits.json
+```
+
+#### Reorder Batch Threshold Optimization
+
+**Question:** What's the optimal `reorder_batch_threshold` for speculative decoding?
+
+```bash
+python benchmark.py \
+    --backend flashmla \
+    --batch-specs "q4s1k" "q8s2k" \
+    --sweep-param reorder_batch_threshold \
+    --sweep-values 1 4 16 64 256 512 \
+    --output-csv threshold_sweep.csv
+```
+
+### All Command-Line Options
+
+```text
+--config CONFIG                     # Path to YAML config file (overrides other args)
+--backends BACKEND [BACKEND ...]    # flash, triton, flashinfer, cutlass_mla,
+                                    # flashinfer_mla, flashattn_mla, flashmla
+--backend BACKEND                   # Single backend (alternative to --backends)
+--batch-specs SPEC [SPEC ...]       # Batch specifications using extended grammar
+
+# Model configuration
+--num-layers N                      # Number of layers
+--head-dim N                        # Head dimension
+--num-q-heads N                     # Query heads
+--num-kv-heads N                    # KV heads
+--block-size N                      # Block size
+
+# Benchmark settings
+--device DEVICE                     # Device (default: cuda:0)
+--repeats N                         # Repetitions
+--warmup-iters N                    # Warmup iterations
+--profile-memory                    # Profile memory usage
+
+# Parameter sweeps
+--sweep-param PARAM                 # Parameter name to sweep (e.g., num_kv_splits,
+                                    # reorder_batch_threshold)
+--sweep-values N [N ...]            # Values to sweep for the parameter
+
+# Output
+--output-csv FILE                   # Save to CSV
+--output-json FILE                  # Save to JSON
+```
+
+## Hardware Requirements
+
+| Backend | Hardware |
+|---------|----------|
+| Flash/Triton/FlashInfer | Any CUDA GPU |
+| CUTLASS MLA | Blackwell (SM100+) |
+| FlashAttn MLA | Hopper (SM90+) |
+| FlashMLA | Hopper (SM90+) |
+| FlashInfer-MLA | Any CUDA GPU |
+
+## Using MLA Runner Directly
+
+All MLA backends are available through `mla_runner.run_mla_benchmark()`:
+
+```python
+from mla_runner import run_mla_benchmark
+from common import BenchmarkConfig
+
+config = BenchmarkConfig(
+    backend="cutlass_mla",
+    batch_spec="64q1s4k",
+    num_layers=10,
+    head_dim=576,
+    num_q_heads=128,
+    num_kv_heads=1,
+    block_size=128,
+    device="cuda:0",
+    repeats=5,
+    warmup_iters=3,
+)
+
+# CUTLASS MLA with specific num_kv_splits
+result = run_mla_benchmark("cutlass_mla", config, num_kv_splits=4)
+print(f"Time: {result.mean_time:.6f}s")
+
+# FlashInfer-MLA
+result = run_mla_benchmark("flashinfer_mla", config)
+
+# FlashAttn MLA (Hopper SM90+)
+result = run_mla_benchmark("flashattn_mla", config, reorder_batch_threshold=64)
+
+# FlashMLA (Hopper SM90+)
+result = run_mla_benchmark("flashmla", config, reorder_batch_threshold=64)
+```
+
+## Python API
+
+```python
+from batch_spec import parse_batch_spec, format_batch_spec, get_batch_stats
+from common import BenchmarkConfig, BenchmarkResult, ResultsFormatter
+
+# Parse batch specs
+requests = parse_batch_spec("2q2k_q4s1k_32q1s1k")
+print(format_batch_spec(requests))
+# "2 prefill (2x2k), 1 extend (1xq4kv1k), 32 decode (32x1k)"
+
+# Get batch statistics
+stats = get_batch_stats(requests)
+print(f"Total tokens: {stats['total_tokens']}")
+print(f"Num decode: {stats['num_decode']}, Num prefill: {stats['num_prefill']}")
+
+# Format results
+formatter = ResultsFormatter()
+formatter.save_csv(results, "output.csv")
+formatter.save_json(results, "output.json")
+```
+
+## Tips
+
+**1. Warmup matters** - Use `--warmup-iters 10` for stable results
+
+**2. Multiple repeats** - Use `--repeats 20` for low variance
+
+**3. Save results** - Always use `--output-csv` or `--output-json`
+
+**4. Test incrementally** - Start with `--num-layers 1 --repeats 1`
+
+**5. Extended grammar** - Leverage spec decode, chunked prefill patterns
+
+**6. Parameter sweeps** - Use `--sweep-param` and `--sweep-values` to find optimal values

benchmarks/attention_benchmarks/__init__.py

@@ -0,0 +1,44 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""vLLM Attention Benchmarking Suite."""
+
+from .batch_spec import (
+    BatchRequest,
+    format_batch_spec,
+    get_batch_stats,
+    parse_batch_spec,
+    reorder_for_flashinfer,
+    split_by_type,
+)
+from .common import (
+    BenchmarkConfig,
+    BenchmarkResult,
+    MockLayer,
+    MockModelConfig,
+    ResultsFormatter,
+    get_attention_scale,
+    is_mla_backend,
+    setup_mla_dims,
+)
+
+__all__ = [
+    # Batch specification
+    "BatchRequest",
+    "parse_batch_spec",
+    "format_batch_spec",
+    "reorder_for_flashinfer",
+    "split_by_type",
+    "get_batch_stats",
+    # Benchmarking infrastructure
+    "BenchmarkConfig",
+    "BenchmarkResult",
+    "ResultsFormatter",
+    # Mock objects
+    "MockLayer",
+    "MockModelConfig",
+    # Utilities
+    "setup_mla_dims",
+    "get_attention_scale",
+    "is_mla_backend",
+]

benchmarks/attention_benchmarks/batch_spec.py

@@ -0,0 +1,231 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Simplified batch specification grammar for attention benchmarks.
+
+Grammar (underscore-separated segments):
+  Format: (<count>?) q<q_len>(k?) (s<seq_len>(k?))?
+
+  - count: Number of identical requests (optional, default=1)
+  - q_len: Query length (number of new tokens)
+  - seq_len: Total sequence length (optional, defaults to q_len for prefill)
+  - 'k' suffix: Multiplies value by 1024
+
+Common patterns:
+  - Prefill:  q_len == seq_len  (e.g., "q2k" → 2048 new tokens, 2048 seq)
+  - Decode:   q_len == 1        (e.g., "q1s1k" → 1 token, 1024 seq length)
+  - Extend:   q_len < seq_len   (e.g., "q4s1k" → 4 tokens, 1024 seq length)
+
+Examples:
+  q2k              -> [(2048, 2048)]           # Prefill: 2048 tokens
+  q1s1k            -> [(1, 1024)]              # Decode: 1 token, 1K sequence
+  8q1s1k           -> [(1, 1024)] * 8          # 8 decode requests
+  q4s1k            -> [(4, 1024)]              # 4-token extend (spec decode)
+  2q1k_32q1s1k     -> [(1024, 1024)] * 2 + [(1, 1024)] * 32  # Mixed batch
+  16q4s1k          -> [(4, 1024)] * 16         # 16 spec decode requests
+"""
+
+from collections import Counter
+from dataclasses import dataclass
+
+import regex as re
+
+
+@dataclass
+class BatchRequest:
+    """Represents a single request in a batch."""
+
+    q_len: int  # Query length (number of new tokens)
+    kv_len: int  # Total KV cache length
+
+    @property
+    def is_decode(self) -> bool:
+        """True if this is a decode request (q_len == 1)."""
+        return self.q_len == 1
+
+    @property
+    def is_prefill(self) -> bool:
+        """True if this is a pure prefill (q_len == kv_len)."""
+        return self.q_len == self.kv_len
+
+    @property
+    def is_extend(self) -> bool:
+        """True if this is context extension (q_len > 1, kv_len > q_len)."""
+        return self.q_len > 1 and self.kv_len > self.q_len
+
+    @property
+    def context_len(self) -> int:
+        """Context length (KV cache - query)."""
+        return self.kv_len - self.q_len
+
+    def as_tuple(self) -> tuple[int, int]:
+        """Return as (q_len, kv_len) tuple for compatibility."""
+        return (self.q_len, self.kv_len)
+
+
+def _parse_size(size_str: str, k_suffix: str) -> int:
+    """Parse size string with optional 'k' suffix."""
+    size = int(size_str)
+    return size * 1024 if k_suffix == "k" else size
+
+
+def parse_batch_spec(spec: str) -> list[BatchRequest]:
+    """
+    Parse batch specification string into list of BatchRequest objects.
+
+    Grammar: (<count>?) q<q_len>(k?) (s<seq_len>(k?))?
+
+    Args:
+        spec: Batch specification string (see module docstring for grammar)
+
+    Returns:
+        List of BatchRequest objects
+
+    Raises:
+        ValueError: If spec format is invalid
+    """
+    requests = []
+
+    for seg in spec.split("_"):
+        # Unified pattern: (<count>?) q<q_len>(k?) (s<seq_len>(k?))?
+        m = re.match(r"^(?:(\d+))?q(\d+)(k?)(?:s(\d+)(k?))?$", seg)
+        if m:
+            cnt = int(m.group(1)) if m.group(1) else 1
+            q_len = _parse_size(m.group(2), m.group(3))
+            kv_len = _parse_size(m.group(4), m.group(5)) if m.group(4) else q_len
+            requests.extend([BatchRequest(q_len=q_len, kv_len=kv_len)] * cnt)
+            continue
+
+        raise ValueError(f"Invalid batch spec segment: '{seg}'")
+
+    return requests
+
+
+def format_batch_spec(requests: list[BatchRequest]) -> str:
+    """
+    Format list of BatchRequest into human-readable string.
+
+    Groups requests by type and provides counts and sizes.
+
+    Args:
+        requests: List of BatchRequest objects
+
+    Returns:
+        Formatted string describing the batch
+    """
+    kinds = {
+        "prefill": [],
+        "extend": [],
+        "decode": [],
+    }
+
+    for req in requests:
+        tup = (req.q_len, req.kv_len)
+        if req.is_prefill:
+            kinds["prefill"].append(tup)
+        elif req.is_extend:
+            kinds["extend"].append(tup)
+        elif req.is_decode:
+            kinds["decode"].append(tup)
+
+    parts = []
+    for kind in ["prefill", "extend", "decode"]:
+        lst = kinds[kind]
+        if not lst:
+            continue
+
+        cnt_total = len(lst)
+        ctr = Counter(lst)
+        inner = []
+
+        for (q, kv), cnt in ctr.items():
+            if kind == "prefill":
+                size = f"{q // 1024}k" if q % 1024 == 0 else str(q)
+                inner.append(f"{cnt}x{size}")
+            elif kind == "decode":
+                size = f"{kv // 1024}k" if kv % 1024 == 0 else str(kv)
+                inner.append(f"{cnt}x{size}")
+            else:  # extend
+                qstr = f"{q // 1024}k" if q % 1024 == 0 else str(q)
+                kstr = f"{kv // 1024}k" if kv % 1024 == 0 else str(kv)
+                inner.append(f"{cnt}xq{qstr}kv{kstr}")
+
+        parts.append(f"{cnt_total} {kind} ({', '.join(inner)})")
+
+    return ", ".join(parts)
+
+
+def reorder_for_flashinfer(requests: list[BatchRequest]) -> list[BatchRequest]:
+    """
+    Reorder requests for FlashInfer: decode first, then prefill.
+
+    FlashInfer expects decode requests before prefill requests for
+    optimal performance.
+
+    Args:
+        requests: Original list of BatchRequest
+
+    Returns:
+        Reordered list with decode requests first
+    """
+    decodes = [r for r in requests if r.is_decode]
+    non_decodes = [r for r in requests if not r.is_decode]
+    return decodes + non_decodes
+
+
+def split_by_type(
+    requests: list[BatchRequest],
+) -> dict[str, list[BatchRequest]]:
+    """
+    Split requests by type for analysis.
+
+    Args:
+        requests: List of BatchRequest
+
+    Returns:
+        Dict with keys: 'decode', 'prefill', 'extend'
+    """
+    result = {
+        "decode": [],
+        "prefill": [],
+        "extend": [],
+    }
+
+    for req in requests:
+        if req.is_decode:
+            result["decode"].append(req)
+        elif req.is_prefill:
+            result["prefill"].append(req)
+        elif req.is_extend:
+            result["extend"].append(req)
+
+    return result
+
+
+def get_batch_stats(requests: list[BatchRequest]) -> dict:
+    """
+    Compute statistics about a batch.
+
+    Args:
+        requests: List of BatchRequest
+
+    Returns:
+        Dict with batch statistics
+    """
+    by_type = split_by_type(requests)
+
+    return {
+        "total_requests": len(requests),
+        "num_decode": len(by_type["decode"]),
+        "num_prefill": len(by_type["prefill"]),
+        "num_extend": len(by_type["extend"]),
+        "total_tokens": sum(r.q_len for r in requests),
+        "total_kv_cache": sum(r.kv_len for r in requests),
+        "max_q_len": max((r.q_len for r in requests), default=0),
+        "max_kv_len": max((r.kv_len for r in requests), default=0),
+        "avg_q_len": sum(r.q_len for r in requests) / len(requests) if requests else 0,
+        "avg_kv_len": (
+            sum(r.kv_len for r in requests) / len(requests) if requests else 0
+        ),
+    }

benchmarks/attention_benchmarks/benchmark.py

@@ -0,0 +1,886 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Universal vLLM Attention Benchmark
+
+Benchmark any attention backend with the extended grammar.
+Supports standard attention (Flash/Triton/FlashInfer) and MLA backends.
+
+Examples:
+    # Standard attention
+    python benchmark.py --backends flash flashinfer --batch-specs "q2k" "8q1s1k"
+
+    # MLA backends
+    python benchmark.py --backends cutlass_mla flashinfer_mla --batch-specs "64q1s1k"
+
+    # Parameter sweep (CLI)
+    python benchmark.py --backend cutlass_mla \
+                        --batch-specs "64q1s1k" \
+                        --sweep-param num_kv_splits \
+                        --sweep-values 1 4 8 16
+
+    # Parameter sweep (YAML config - recommended)
+    python benchmark.py --config configs/cutlass_numsplits.yaml
+"""
+
+import argparse
+import sys
+from dataclasses import replace
+from pathlib import Path
+
+import yaml
+from rich.console import Console
+from tqdm import tqdm
+
+sys.path.insert(0, str(Path(__file__).parent.parent.parent))
+
+from batch_spec import parse_batch_spec
+from common import (
+    BenchmarkConfig,
+    BenchmarkResult,
+    ModelParameterSweep,
+    ParameterSweep,
+    ResultsFormatter,
+    is_mla_backend,
+)
+
+
+def run_standard_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
+    """Run standard attention benchmark (Flash/Triton/FlashInfer)."""
+    from runner import run_attention_benchmark
+
+    return run_attention_benchmark(config)
+
+
+def run_mla_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
+    """Run MLA benchmark with appropriate backend."""
+    from mla_runner import run_mla_benchmark as run_mla
+
+    return run_mla(config.backend, config, **kwargs)
+
+
+def run_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
+    """
+    Run a single benchmark with proper backend selection.
+
+    Args:
+        config: BenchmarkConfig with backend, batch_spec, and model params
+        **kwargs: Additional arguments passed to MLA benchmarks
+
+    Returns:
+        BenchmarkResult (may have error field set on failure)
+    """
+    try:
+        if is_mla_backend(config.backend):
+            return run_mla_benchmark(config, **kwargs)
+        else:
+            return run_standard_attention_benchmark(config)
+    except Exception as e:
+        return BenchmarkResult(
+            config=config,
+            mean_time=float("inf"),
+            std_time=0,
+            min_time=float("inf"),
+            max_time=float("inf"),
+            error=str(e),
+        )
+
+
+def run_model_parameter_sweep(
+    backends: list[str],
+    batch_specs: list[str],
+    base_config_args: dict,
+    sweep: ModelParameterSweep,
+    console: Console,
+) -> list[BenchmarkResult]:
+    """
+    Run model parameter sweep for given backends and batch specs.
+
+    Args:
+        backends: List of backend names
+        batch_specs: List of batch specifications
+        base_config_args: Base configuration arguments (num_layers, head_dim, etc.)
+        sweep: ModelParameterSweep configuration
+        console: Rich console for output
+
+    Returns:
+        List of BenchmarkResult objects
+    """
+    all_results = []
+
+    console.print(
+        f"[yellow]Model sweep mode: testing {sweep.param_name} = {sweep.values}[/]"
+    )
+
+    total = len(backends) * len(batch_specs) * len(sweep.values)
+
+    with tqdm(total=total, desc="Benchmarking") as pbar:
+        for backend in backends:
+            for spec in batch_specs:
+                for value in sweep.values:
+                    # Create config with modified model parameter
+                    config_args = base_config_args.copy()
+                    config_args[sweep.param_name] = value
+
+                    # Create config with original backend for running
+                    clean_config = BenchmarkConfig(
+                        backend=backend, batch_spec=spec, **config_args
+                    )
+
+                    # Run benchmark
+                    result = run_benchmark(clean_config)
+
+                    # Replace backend with labeled version for display
+                    backend_label = sweep.get_label(backend, value)
+                    labeled_config = replace(result.config, backend=backend_label)
+                    result = replace(result, config=labeled_config)
+                    all_results.append(result)
+
+                    if not result.success:
+                        console.print(
+                            f"[red]Error {backend} {spec} {sweep.param_name}="
+                            f"{value}: {result.error}[/]"
+                        )
+
+                    pbar.update(1)
+
+    # Display sweep results - create separate table for each parameter value
+    console.print("\n[bold green]Model Parameter Sweep Results:[/]")
+    formatter = ResultsFormatter(console)
+
+    # Group results by parameter value and extract backend mapping
+    by_param_value = {}
+    backend_mapping = {}  # Maps labeled backend -> original backend
+
+    for r in all_results:
+        # Extract original backend and param value from labeled backend
+        # The label format is: {backend}_{param_name}_{value}
+        # We need to reverse engineer this
+        labeled_backend = r.config.backend
+
+        # Try each backend to find which one this result belongs to
+        for backend in backends:
+            for value in sweep.values:
+                expected_label = sweep.get_label(backend, value)
+                if labeled_backend == expected_label:
+                    backend_mapping[labeled_backend] = backend
+                    param_value = str(value)
+
+                    if param_value not in by_param_value:
+                        by_param_value[param_value] = []
+                    by_param_value[param_value].append(r)
+                    break
+
+    # Create a table for each parameter value
+    sorted_param_values = sorted(
+        by_param_value.keys(), key=lambda x: int(x) if x.isdigit() else x
+    )
+
+    for param_value in sorted_param_values:
+        console.print(f"\n[bold cyan]{sweep.param_name} = {param_value}[/]")
+        param_results = by_param_value[param_value]
+
+        # Create modified results with original backend names
+        modified_results = []
+        for r in param_results:
+            # Get the original backend name from our mapping
+            original_backend = backend_mapping[r.config.backend]
+            modified_config = replace(r.config, backend=original_backend)
+            modified_result = replace(r, config=modified_config)
+            modified_results.append(modified_result)
+
+        # Print table with original backend names
+        formatter.print_table(modified_results, backends, compare_to_fastest=True)
+
+    # Show optimal backend for each (param_value, batch_spec) combination
+    console.print(
+        f"\n[bold cyan]Optimal backend for each ({sweep.param_name}, batch_spec):[/]"
+    )
+
+    # Group by (param_value, batch_spec)
+    by_param_and_spec = {}
+    for r in all_results:
+        if r.success:
+            # Find which (backend, value) this result corresponds to
+            labeled_backend = r.config.backend
+            for backend in backends:
+                for value in sweep.values:
+                    expected_label = sweep.get_label(backend, value)
+                    if labeled_backend == expected_label:
+                        param_value = str(value)
+                        spec = r.config.batch_spec
+                        key = (param_value, spec)
+
+                        if key not in by_param_and_spec:
+                            by_param_and_spec[key] = []
+                        by_param_and_spec[key].append(r)
+                        break
+
+    # Sort by param value then spec
+    sorted_keys = sorted(
+        by_param_and_spec.keys(),
+        key=lambda x: (int(x[0]) if x[0].isdigit() else x[0], x[1]),
+    )
+
+    current_param_value = None
+    for param_value, spec in sorted_keys:
+        # Print header when param value changes
+        if param_value != current_param_value:
+            console.print(f"\n  [bold]{sweep.param_name}={param_value}:[/]")
+            current_param_value = param_value
+
+        results = by_param_and_spec[(param_value, spec)]
+        best = min(results, key=lambda r: r.mean_time)
+
+        # Extract original backend name using the mapping
+        backend_name = backend_mapping[best.config.backend]
+
+        # Show all backends' times for comparison
+        times_str = " | ".join(
+            [
+                f"{backend_mapping[r.config.backend]}: {r.mean_time:.6f}s"
+                for r in sorted(results, key=lambda r: r.mean_time)
+            ]
+        )
+
+        console.print(
+            f"    {spec:12s} -> [bold green]{backend_name:15s}[/] ({times_str})"
+        )
+
+    return all_results
+
+
+def run_parameter_sweep(
+    backends: list[str],
+    batch_specs: list[str],
+    base_config_args: dict,
+    sweep: ParameterSweep,
+    console: Console,
+) -> list[BenchmarkResult]:
+    """
+    Run parameter sweep for given backends and batch specs.
+
+    Args:
+        backends: List of backend names
+        batch_specs: List of batch specifications
+        base_config_args: Base configuration arguments (num_layers, head_dim, etc.)
+        sweep: ParameterSweep configuration
+        console: Rich console for output
+
+    Returns:
+        List of BenchmarkResult objects
+    """
+    all_results = []
+
+    # Build list of values to sweep (including auto if requested)
+    sweep_values = list(sweep.values)
+    if sweep.include_auto:
+        sweep_values.append("auto")
+
+    console.print(f"[yellow]Sweep mode: testing {sweep.param_name} = {sweep_values}[/]")
+
+    total = len(backends) * len(batch_specs) * len(sweep_values)
+
+    with tqdm(total=total, desc="Benchmarking") as pbar:
+        for backend in backends:
+            for spec in batch_specs:
+                for value in sweep_values:
+                    # Create config with original backend for running
+                    config = BenchmarkConfig(
+                        backend=backend, batch_spec=spec, **base_config_args
+                    )
+
+                    # Prepare kwargs for benchmark runner
+                    kwargs = {}
+                    if value != "auto":
+                        kwargs[sweep.param_name] = value
+
+                    # Run benchmark
+                    result = run_benchmark(config, **kwargs)
+
+                    # Replace backend with labeled version for display
+                    backend_label = sweep.get_label(backend, value)
+                    labeled_config = replace(result.config, backend=backend_label)
+                    result = replace(result, config=labeled_config)
+                    all_results.append(result)
+
+                    if not result.success:
+                        console.print(
+                            f"[red]Error {backend} {spec} {sweep.param_name}="
+                            f"{value}: {result.error}[/]"
+                        )
+
+                    pbar.update(1)
+
+    # Display sweep results
+    console.print("\n[bold green]Sweep Results:[/]")
+    backend_labels = [sweep.get_label(b, v) for b in backends for v in sweep_values]
+    formatter = ResultsFormatter(console)
+    formatter.print_table(all_results, backend_labels)
+
+    # Show optimal values
+    console.print(f"\n[bold cyan]Optimal {sweep.param_name} per batch spec:[/]")
+    by_spec = {}
+    for r in all_results:
+        if r.success:
+            spec = r.config.batch_spec
+            if spec not in by_spec:
+                by_spec[spec] = []
+            by_spec[spec].append(r)
+
+    for spec in sorted(by_spec.keys()):
+        results = by_spec[spec]
+        best = min(results, key=lambda r: r.mean_time)
+        console.print(
+            f"  {spec}: [bold green]{best.config.backend}[/] ({best.mean_time:.6f}s)"
+        )
+
+    return all_results
+
+
+def load_config_from_yaml(config_path: str) -> dict:
+    """Load configuration from YAML file."""
+    with open(config_path) as f:
+        return yaml.safe_load(f)
+
+
+def generate_batch_specs_from_ranges(ranges: list[dict]) -> list[str]:
+    """
+    Generate batch specs from range specifications.
+
+    Args:
+        ranges: List of range specifications, each containing:
+            - template: Batch spec template (e.g., "q{q_len}kv1k")
+            - q_len: Dict with start, stop, step, end_inclusive (optional)
+            - Other parameters can also be ranges
+
+    Returns:
+        List of generated batch spec strings
+
+    Example:
+        ranges = [
+            {
+                "template": "q{q_len}kv1k",
+                "q_len": {
+                    "start": 1,
+                    "stop": 16,
+                    "step": 1,
+                    "end_inclusive": true  # Optional, defaults to true
+                }
+            }
+        ]
+        Returns: ["q1kv1k", "q2kv1k", ..., "q16kv1k"]
+    """
+    all_specs = []
+
+    for range_spec in ranges:
+        template = range_spec.get("template")
+        if not template:
+            raise ValueError("Range specification must include 'template'")
+
+        # Extract all range parameters from the spec
+        range_params = {}
+        for key, value in range_spec.items():
+            if key == "template":
+                continue
+            if isinstance(value, dict) and "start" in value:
+                # This is a range specification
+                start = value["start"]
+                stop = value["stop"]
+                step = value.get("step", 1)
+                # Check if end should be inclusive (default: True)
+                end_inclusive = value.get("end_inclusive", True)
+
+                # Adjust stop based on end_inclusive
+                if end_inclusive:
+                    range_params[key] = list(range(start, stop + 1, step))
+                else:
+                    range_params[key] = list(range(start, stop, step))
+            else:
+                # This is a fixed value
+                range_params[key] = [value]
+
+        # Generate all combinations (Cartesian product)
+        if range_params:
+            import itertools
+
+            param_names = list(range_params.keys())
+            param_values = [range_params[name] for name in param_names]
+
+            for values in itertools.product(*param_values):
+                params = dict(zip(param_names, values))
+                spec = template.format(**params)
+                all_specs.append(spec)
+        else:
+            # No parameters, just use template as-is
+            all_specs.append(template)
+
+    return all_specs
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Universal vLLM attention benchmark",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=__doc__,
+    )
+
+    # Config file
+    parser.add_argument(
+        "--config",
+        help="Path to YAML config file (overrides other args)",
+    )
+
+    # Backend selection
+    parser.add_argument(
+        "--backends",
+        nargs="+",
+        help="Backends to benchmark (flash, triton, flashinfer, cutlass_mla, "
+        "flashinfer_mla, flashattn_mla, flashmla)",
+    )
+    parser.add_argument(
+        "--backend",
+        help="Single backend (alternative to --backends)",
+    )
+
+    # Batch specifications
+    parser.add_argument(
+        "--batch-specs",
+        nargs="+",
+        default=["q2k", "8q1s1k"],
+        help="Batch specifications using extended grammar",
+    )
+
+    # Model config
+    parser.add_argument("--num-layers", type=int, default=10, help="Number of layers")
+    parser.add_argument("--head-dim", type=int, default=128, help="Head dimension")
+    parser.add_argument("--num-q-heads", type=int, default=32, help="Query heads")
+    parser.add_argument("--num-kv-heads", type=int, default=8, help="KV heads")
+    parser.add_argument("--block-size", type=int, default=16, help="Block size")
+
+    # Benchmark settings
+    parser.add_argument("--device", default="cuda:0", help="Device")
+    parser.add_argument("--repeats", type=int, default=1, help="Repetitions")
+    parser.add_argument("--warmup-iters", type=int, default=3, help="Warmup iterations")
+    parser.add_argument("--profile-memory", action="store_true", help="Profile memory")
+
+    # Parameter sweep (use YAML config for advanced sweeps)
+    parser.add_argument(
+        "--sweep-param",
+        help="Parameter name to sweep (e.g., num_kv_splits, reorder_batch_threshold)",
+    )
+    parser.add_argument(
+        "--sweep-values",
+        type=int,
+        nargs="+",
+        help="Values to sweep for the parameter",
+    )
+
+    # Output
+    parser.add_argument("--output-csv", help="Save to CSV")
+    parser.add_argument("--output-json", help="Save to JSON")
+
+    args = parser.parse_args()
+
+    console = Console()
+    console.print("[bold cyan]vLLM Attention Benchmark[/]")
+
+    # Load config from YAML if provided
+    if args.config:
+        console.print(f"[yellow]Loading config from: {args.config}[/]")
+        yaml_config = load_config_from_yaml(args.config)
+
+        # Show description if available
+        if "description" in yaml_config:
+            console.print(f"[dim]{yaml_config['description']}[/]")
+
+        # Override args with YAML values
+        # (YAML takes precedence unless CLI arg was explicitly set)
+        # Backend(s)
+        if "backend" in yaml_config:
+            args.backend = yaml_config["backend"]
+            args.backends = None
+        elif "backends" in yaml_config:
+            args.backends = yaml_config["backends"]
+            args.backend = None
+
+        # Check for special modes
+        if "mode" in yaml_config:
+            args.mode = yaml_config["mode"]
+        else:
+            args.mode = None
+
+        # Batch specs and sizes
+        # Support both explicit batch_specs and generated batch_spec_ranges
+        if "batch_spec_ranges" in yaml_config:
+            # Generate batch specs from ranges
+            generated_specs = generate_batch_specs_from_ranges(
+                yaml_config["batch_spec_ranges"]
+            )
+            # Combine with any explicit batch_specs
+            if "batch_specs" in yaml_config:
+                args.batch_specs = yaml_config["batch_specs"] + generated_specs
+            else:
+                args.batch_specs = generated_specs
+            console.print(
+                f"[dim]Generated {len(generated_specs)} batch specs from ranges[/]"
+            )
+        elif "batch_specs" in yaml_config:
+            args.batch_specs = yaml_config["batch_specs"]
+
+        if "batch_sizes" in yaml_config:
+            args.batch_sizes = yaml_config["batch_sizes"]
+        else:
+            args.batch_sizes = None
+
+        # Model config
+        if "model" in yaml_config:
+            model = yaml_config["model"]
+            args.num_layers = model.get("num_layers", args.num_layers)
+            args.head_dim = model.get("head_dim", args.head_dim)
+            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)
+
+        # Benchmark settings
+        if "benchmark" in yaml_config:
+            bench = yaml_config["benchmark"]
+            args.device = bench.get("device", args.device)
+            args.repeats = bench.get("repeats", args.repeats)
+            args.warmup_iters = bench.get("warmup_iters", args.warmup_iters)
+            args.profile_memory = bench.get("profile_memory", args.profile_memory)
+
+        # Parameter sweep configuration
+        if "parameter_sweep" in yaml_config:
+            sweep_config = yaml_config["parameter_sweep"]
+            args.parameter_sweep = ParameterSweep(
+                param_name=sweep_config["param_name"],
+                values=sweep_config["values"],
+                include_auto=sweep_config.get("include_auto", False),
+                label_format=sweep_config.get(
+                    "label_format", "{backend}_{param_name}_{value}"
+                ),
+            )
+        else:
+            args.parameter_sweep = None
+
+        # Model parameter sweep configuration
+        if "model_parameter_sweep" in yaml_config:
+            sweep_config = yaml_config["model_parameter_sweep"]
+            args.model_parameter_sweep = ModelParameterSweep(
+                param_name=sweep_config["param_name"],
+                values=sweep_config["values"],
+                label_format=sweep_config.get(
+                    "label_format", "{backend}_{param_name}_{value}"
+                ),
+            )
+        else:
+            args.model_parameter_sweep = None
+
+        # Output
+        if "output" in yaml_config:
+            output = yaml_config["output"]
+            if "csv" in output and not args.output_csv:
+                args.output_csv = output["csv"]
+            if "json" in output and not args.output_json:
+                args.output_json = output["json"]
+
+        console.print()
+
+    # Handle CLI-based parameter sweep (if not from YAML)
+    if (
+        (not hasattr(args, "parameter_sweep") or args.parameter_sweep is None)
+        and args.sweep_param
+        and args.sweep_values
+    ):
+        args.parameter_sweep = ParameterSweep(
+            param_name=args.sweep_param,
+            values=args.sweep_values,
+            include_auto=False,
+            label_format="{backend}_{param_name}_{value}",
+        )
+
+    # Determine backends
+    backends = args.backends or ([args.backend] if args.backend else ["flash"])
+    console.print(f"Backends: {', '.join(backends)}")
+    console.print(f"Batch specs: {', '.join(args.batch_specs)}")
+    console.print()
+
+    # Run benchmarks
+    all_results = []
+
+    # Handle special mode: decode_vs_prefill comparison
+    if hasattr(args, "mode") and args.mode == "decode_vs_prefill":
+        console.print("[yellow]Mode: Decode vs Prefill pipeline comparison[/]")
+        console.print(
+            "[dim]For each query length, testing both decode and prefill pipelines[/]"
+        )
+        console.print("[dim]Using batched execution for optimal performance[/]")
+
+        # Extract batch sizes from config
+        batch_sizes = getattr(args, "batch_sizes", [1])
+        backend = backends[0]  # Use first backend (should only be one)
+
+        # Calculate total benchmarks
+        total = len(batch_sizes)
+
+        with tqdm(total=total, desc="Benchmarking") as pbar:
+            for batch_size in batch_sizes:
+                # Prepare all configs for this batch size
+                configs_with_thresholds = []
+
+                for spec in args.batch_specs:
+                    # Parse the batch spec to get query length
+                    requests = parse_batch_spec(spec)
+                    if not requests:
+                        console.print(
+                            f"[red]Error: Could not parse batch spec '{spec}'[/]"
+                        )
+                        continue
+
+                    # Get query length from first request
+                    query_length = requests[0].q_len
+
+                    # Create batch spec for this batch size
+                    # For batch_size > 1, we need to prepend the count
+                    batch_spec = f"{batch_size}{spec}" if batch_size > 1 else spec
+
+                    # Create base config (without backend name)
+                    base_config = BenchmarkConfig(
+                        backend=backend,  # Will be overridden later
+                        batch_spec=batch_spec,
+                        num_layers=args.num_layers,
+                        head_dim=args.head_dim,
+                        num_q_heads=args.num_q_heads,
+                        num_kv_heads=args.num_kv_heads,
+                        block_size=args.block_size,
+                        device=args.device,
+                        repeats=args.repeats,
+                        warmup_iters=args.warmup_iters,
+                        profile_memory=args.profile_memory,
+                    )
+
+                    # Add decode pipeline config
+                    decode_threshold = query_length
+                    config_decode = replace(
+                        base_config,
+                        backend=f"{backend}_decode_qlen{query_length}_bs{batch_size}",
+                    )
+                    configs_with_thresholds.append((config_decode, decode_threshold))
+
+                    # Add prefill pipeline config if query_length > 1
+                    if query_length > 1:
+                        prefill_threshold = query_length - 1
+                        config_prefill = replace(
+                            base_config,
+                            backend=f"{backend}_prefill_qlen{query_length}"
+                            f"_bs{batch_size}",
+                        )
+                        configs_with_thresholds.append(
+                            (config_prefill, prefill_threshold)
+                        )
+
+                # Run all benchmarks for this batch size in one go (batched mode)
+                try:
+                    from mla_runner import run_mla_benchmark as run_mla
+
+                    # Use batched API: pass list of (config, threshold) tuples
+                    timing_results = run_mla(backend, configs_with_thresholds)
+
+                    # Create BenchmarkResult objects from timing results
+                    for (config, _), timing in zip(
+                        configs_with_thresholds, timing_results
+                    ):
+                        result = BenchmarkResult(
+                            config=config,
+                            mean_time=timing["mean"],
+                            std_time=timing["std"],
+                            min_time=timing["min"],
+                            max_time=timing["max"],
+                            throughput_tokens_per_sec=timing.get("throughput", None),
+                        )
+                        all_results.append(result)
+
+                except Exception as e:
+                    import traceback
+
+                    console.print(
+                        f"[red]Error running batched benchmarks for "
+                        f"batch_size={batch_size}: {e}[/]"
+                    )
+                    console.print("[red]Traceback:[/]")
+                    traceback.print_exc()
+                    # Add error results for all configs
+                    for config, _ in configs_with_thresholds:
+                        result = BenchmarkResult(
+                            config=config,
+                            mean_time=float("inf"),
+                            std_time=0,
+                            min_time=float("inf"),
+                            max_time=float("inf"),
+                            error=str(e),
+                        )
+                        all_results.append(result)
+
+                pbar.update(1)
+
+        # Display decode vs prefill results
+        console.print("\n[bold green]Decode vs Prefill Results:[/]")
+
+        # Group by batch size
+        by_batch_size = {}
+        for r in all_results:
+            if r.success:
+                # Extract batch size from backend name
+                parts = r.config.backend.split("_")
+                bs_part = [p for p in parts if p.startswith("bs")]
+                if bs_part:
+                    bs = int(bs_part[0][2:])
+                    if bs not in by_batch_size:
+                        by_batch_size[bs] = []
+                    by_batch_size[bs].append(r)
+
+        # For each batch size, analyze crossover point
+        for bs in sorted(by_batch_size.keys()):
+            console.print(f"\n[bold cyan]Batch size: {bs}[/]")
+            results = by_batch_size[bs]
+
+            # Group by query length
+            by_qlen = {}
+            for r in results:
+                parts = r.config.backend.split("_")
+                qlen_part = [p for p in parts if p.startswith("qlen")]
+                if qlen_part:
+                    qlen = int(qlen_part[0][4:])
+                    if qlen not in by_qlen:
+                        by_qlen[qlen] = {}
+
+                    pipeline = "decode" if "decode" in r.config.backend else "prefill"
+                    by_qlen[qlen][pipeline] = r
+
+            # Find crossover point
+            last_decode_faster = None
+            for qlen in sorted(by_qlen.keys()):
+                pipelines = by_qlen[qlen]
+                if "decode" in pipelines and "prefill" in pipelines:
+                    decode_time = pipelines["decode"].mean_time
+                    prefill_time = pipelines["prefill"].mean_time
+                    faster = "decode" if decode_time < prefill_time else "prefill"
+
+                    speedup = (
+                        prefill_time / decode_time
+                        if decode_time < prefill_time
+                        else decode_time / prefill_time
+                    )
+
+                    console.print(
+                        f"  qlen={qlen:3d}: decode={decode_time:.6f}s, "
+                        f"prefill={prefill_time:.6f}s -> "
+                        f"[bold]{faster}[/] ({speedup:.2f}x)"
+                    )
+
+                    if faster == "decode":
+                        last_decode_faster = qlen
+
+            if last_decode_faster is not None:
+                optimal_threshold = last_decode_faster
+                console.print(
+                    f"\n  [bold green]Optimal threshold for batch_size={bs}: "
+                    f"{optimal_threshold}[/]"
+                )
+                console.print(
+                    f"  [dim](Use decode pipeline for query_length <= "
+                    f"{optimal_threshold})[/]"
+                )
+            else:
+                console.print(
+                    f"\n  [yellow]Prefill always faster for batch_size={bs}[/]"
+                )
+
+    # Handle model parameter sweep mode
+    elif hasattr(args, "model_parameter_sweep") and args.model_parameter_sweep:
+        # Model parameter sweep
+        base_config_args = {
+            "num_layers": args.num_layers,
+            "head_dim": args.head_dim,
+            "num_q_heads": args.num_q_heads,
+            "num_kv_heads": args.num_kv_heads,
+            "block_size": args.block_size,
+            "device": args.device,
+            "repeats": args.repeats,
+            "warmup_iters": args.warmup_iters,
+            "profile_memory": args.profile_memory,
+        }
+        all_results = run_model_parameter_sweep(
+            backends,
+            args.batch_specs,
+            base_config_args,
+            args.model_parameter_sweep,
+            console,
+        )
+
+    # Handle parameter sweep mode (unified)
+    elif hasattr(args, "parameter_sweep") and args.parameter_sweep:
+        # Unified parameter sweep
+        base_config_args = {
+            "num_layers": args.num_layers,
+            "head_dim": args.head_dim,
+            "num_q_heads": args.num_q_heads,
+            "num_kv_heads": args.num_kv_heads,
+            "block_size": args.block_size,
+            "device": args.device,
+            "repeats": args.repeats,
+            "warmup_iters": args.warmup_iters,
+            "profile_memory": args.profile_memory,
+        }
+        all_results = run_parameter_sweep(
+            backends, args.batch_specs, base_config_args, args.parameter_sweep, console
+        )
+
+    else:
+        # Normal mode: compare backends
+        total = len(backends) * len(args.batch_specs)
+
+        with tqdm(total=total, desc="Benchmarking") as pbar:
+            for spec in args.batch_specs:
+                for backend in backends:
+                    config = BenchmarkConfig(
+                        backend=backend,
+                        batch_spec=spec,
+                        num_layers=args.num_layers,
+                        head_dim=args.head_dim,
+                        num_q_heads=args.num_q_heads,
+                        num_kv_heads=args.num_kv_heads,
+                        block_size=args.block_size,
+                        device=args.device,
+                        repeats=args.repeats,
+                        warmup_iters=args.warmup_iters,
+                        profile_memory=args.profile_memory,
+                    )
+
+                    result = run_benchmark(config)
+                    all_results.append(result)
+
+                    if not result.success:
+                        console.print(f"[red]Error {backend} {spec}: {result.error}[/]")
+
+                    pbar.update(1)
+
+        # Display results
+        console.print("\n[bold green]Results:[/]")
+        formatter = ResultsFormatter(console)
+        formatter.print_table(all_results, backends)
+
+    # Save results
+    if all_results:
+        formatter = ResultsFormatter(console)
+        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 __name__ == "__main__":
+    main()

benchmarks/attention_benchmarks/common.py

@@ -0,0 +1,503 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""Common utilities for attention benchmarking."""
+
+import csv
+import json
+import math
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from rich.console import Console
+from rich.table import Table
+
+# Mock classes for vLLM attention infrastructure
+
+
+class MockHfConfig:
+    """Mock HuggingFace config that satisfies vLLM's requirements."""
+
+    def __init__(self, mla_dims: dict):
+        self.num_attention_heads = mla_dims["num_q_heads"]
+        self.num_key_value_heads = mla_dims["num_kv_heads"]
+        self.hidden_size = mla_dims["head_dim"] * mla_dims["num_q_heads"]
+        self.model_type = "deepseek_v2"
+        self.is_encoder_decoder = False
+        self.kv_lora_rank = mla_dims["kv_lora_rank"]
+        self.qk_nope_head_dim = mla_dims["qk_nope_head_dim"]
+        self.qk_rope_head_dim = mla_dims["qk_rope_head_dim"]
+        self.v_head_dim = mla_dims["v_head_dim"]
+        self.qk_head_dim = mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"]
+
+    def get_text_config(self):
+        return self
+
+
+# Import AttentionLayerBase at module level to avoid circular dependencies
+try:
+    from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
+
+    _HAS_ATTENTION_LAYER_BASE = True
+except ImportError:
+    _HAS_ATTENTION_LAYER_BASE = False
+    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
+
+    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 MockLayer(AttentionLayerBase):
+    """Mock attention layer with scale parameters and impl.
+
+    Inherits from AttentionLayerBase so it passes isinstance checks
+    in get_layers_from_vllm_config when FlashInfer prefill is enabled.
+    """
+
+    def __init__(self, device: torch.device, impl=None, kv_cache_spec=None):
+        # Don't call super().__init__() as AttentionLayerBase doesn't have __init__
+        self._k_scale = torch.tensor(1.0, device=device)
+        self._v_scale = torch.tensor(1.0, device=device)
+        self._q_scale = torch.tensor(1.0, device=device)
+        # Scalar floats for kernels that need them
+        self._k_scale_float = float(self._k_scale.item())
+        self._v_scale_float = float(self._v_scale.item())
+        self._q_scale_float = float(self._q_scale.item())
+        # AttentionImpl for metadata builders to query
+        self.impl = impl
+        # KV cache spec for get_kv_cache_spec
+        self._kv_cache_spec = kv_cache_spec
+
+    def get_attn_backend(self):
+        """Get the attention backend class (required by AttentionLayerBase)."""
+        # Return None as this is just a mock layer for benchmarking
+        return None
+
+    def get_kv_cache_spec(self):
+        """Get the KV cache spec (required by AttentionLayerBase)."""
+        return self._kv_cache_spec
+
+
+class MockModelConfig:
+    """Mock model configuration."""
+
+    def __init__(
+        self,
+        num_q_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        dtype: torch.dtype = torch.float16,
+        max_model_len: int = 32768,
+    ):
+        self._n_q = num_q_heads
+        self._n_kv = num_kv_heads
+        self._d = head_dim
+        self.dtype = dtype
+        self.max_model_len = max_model_len
+
+    def get_num_attention_heads(self, _=None) -> int:
+        return self._n_q
+
+    def get_num_kv_heads(self, _=None) -> int:
+        return self._n_kv
+
+    def get_head_size(self) -> int:
+        return self._d
+
+    def get_num_layers(self) -> int:
+        """Mock method for layer count queries."""
+        return 1
+
+    def get_sliding_window_for_layer(self, _layer_idx: int):
+        """Mock method for sliding window queries."""
+        return None
+
+    def get_logits_soft_cap_for_layer(self, _layer_idx: int):
+        """Mock method for logits soft cap queries."""
+        return None
+
+    def get_sm_scale_for_layer(self, _layer_idx: int) -> float:
+        """Mock method for SM scale queries."""
+        return 1.0 / (self.get_head_size() ** 0.5)
+
+
+class MockParallelConfig:
+    """Mock parallel configuration."""
+
+    pass
+
+
+class MockCompilationConfig:
+    """Mock compilation configuration."""
+
+    def __init__(self):
+        self.full_cuda_graph = False
+        self.static_forward_context = {}
+
+
+class MockVLLMConfig:
+    """Mock VLLM configuration."""
+
+    def __init__(self):
+        self.compilation_config = MockCompilationConfig()
+
+
+class MockRunner:
+    """Mock GPU runner for metadata builders."""
+
+    def __init__(
+        self,
+        seq_lens: np.ndarray,
+        query_start_locs: np.ndarray,
+        device: torch.device,
+        num_q_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        dtype: torch.dtype,
+    ):
+        self.model_config = MockModelConfig(num_q_heads, num_kv_heads, head_dim, dtype)
+        self.parallel_config = MockParallelConfig()
+        self.vllm_config = MockVLLMConfig()
+        self.seq_lens_np = seq_lens
+        self.query_start_loc_np = query_start_locs
+        self.device = device
+        self.attention_chunk_size = None
+        self.num_query_heads = num_q_heads
+        self.num_kv_heads = num_kv_heads
+        self.dtype = dtype
+
+
+@dataclass
+class ParameterSweep:
+    """Configuration for sweeping a backend parameter."""
+
+    param_name: str  # Name of the backend parameter to sweep
+    values: list[Any]  # List of values to test
+    include_auto: bool = False  # Also test with param unset (auto mode)
+    label_format: str = "{backend}_{param_name}_{value}"  # Result label template
+
+    def get_label(self, backend: str, value: Any) -> str:
+        """Generate a label for a specific parameter value."""
+        return self.label_format.format(
+            backend=backend, param_name=self.param_name, value=value
+        )
+
+
+@dataclass
+class ModelParameterSweep:
+    """Configuration for sweeping a model configuration parameter."""
+
+    param_name: str  # Name of the model config parameter to sweep (e.g., "num_q_heads")
+    values: list[Any]  # List of values to test
+    label_format: str = "{backend}_{param_name}_{value}"  # Result label template
+
+    def get_label(self, backend: str, value: Any) -> str:
+        """Generate a label for a specific parameter value."""
+        return self.label_format.format(
+            backend=backend, param_name=self.param_name, value=value
+        )
+
+
+@dataclass
+class BenchmarkConfig:
+    """Configuration for a single benchmark run."""
+
+    backend: str
+    batch_spec: str
+    num_layers: int
+    head_dim: int
+    num_q_heads: int
+    num_kv_heads: int
+    block_size: int
+    device: str
+    dtype: torch.dtype = torch.float16
+    repeats: int = 1
+    warmup_iters: int = 3
+    profile_memory: bool = False
+    use_cuda_graphs: bool = False
+
+    # MLA-specific
+    kv_lora_rank: int | None = None
+    qk_nope_head_dim: int | None = None
+    qk_rope_head_dim: int | None = None
+    v_head_dim: int | None = None
+
+    # Backend-specific tuning
+    num_kv_splits: int | None = None  # CUTLASS MLA
+    reorder_batch_threshold: int | None = None  # FlashAttn MLA, FlashMLA
+
+
+@dataclass
+class BenchmarkResult:
+    """Results from a single benchmark run."""
+
+    config: BenchmarkConfig
+    mean_time: float  # seconds
+    std_time: float  # seconds
+    min_time: float  # seconds
+    max_time: float  # seconds
+    throughput_tokens_per_sec: float | None = None
+    memory_allocated_mb: float | None = None
+    memory_reserved_mb: float | None = None
+    error: str | None = None
+
+    @property
+    def success(self) -> bool:
+        """Whether benchmark completed successfully."""
+        return self.error is None
+
+    def to_dict(self) -> dict[str, Any]:
+        """Convert to dictionary for serialization."""
+        return {
+            "config": asdict(self.config),
+            "mean_time": self.mean_time,
+            "std_time": self.std_time,
+            "min_time": self.min_time,
+            "max_time": self.max_time,
+            "throughput_tokens_per_sec": self.throughput_tokens_per_sec,
+            "memory_allocated_mb": self.memory_allocated_mb,
+            "memory_reserved_mb": self.memory_reserved_mb,
+            "error": self.error,
+        }
+
+
+class ResultsFormatter:
+    """Format and display benchmark results."""
+
+    def __init__(self, console: Console | None = None):
+        self.console = console or Console()
+
+    def print_table(
+        self,
+        results: list[BenchmarkResult],
+        backends: list[str],
+        compare_to_fastest: bool = True,
+    ):
+        """
+        Print results as a rich table.
+
+        Args:
+            results: List of BenchmarkResult
+            backends: List of backend names being compared
+            compare_to_fastest: Show percentage comparison to fastest
+        """
+        # Group by batch spec
+        by_spec = {}
+        for r in results:
+            spec = r.config.batch_spec
+            if spec not in by_spec:
+                by_spec[spec] = {}
+            by_spec[spec][r.config.backend] = r
+
+        # Create shortened backend names for display
+        def shorten_backend_name(name: str) -> str:
+            """Shorten long backend names for table display."""
+            # Remove common prefixes
+            name = name.replace("flashattn_mla", "famla")
+            name = name.replace("flashinfer_mla", "fimla")
+            name = name.replace("flashmla", "fmla")
+            name = name.replace("cutlass_mla", "cmla")
+            name = name.replace("numsplits", "ns")
+            return name
+
+        table = Table(title="Attention Benchmark Results")
+        table.add_column("Batch\nSpec", no_wrap=True)
+
+        multi = len(backends) > 1
+        for backend in backends:
+            short_name = shorten_backend_name(backend)
+            # Time column
+            col_time = f"{short_name}\nTime (s)"
+            table.add_column(col_time, justify="right", no_wrap=False)
+            if multi and compare_to_fastest:
+                # Relative performance column
+                col_rel = f"{short_name}\nvs Best"
+                table.add_column(col_rel, justify="right", no_wrap=False)
+
+        # Add rows
+        for spec in sorted(by_spec.keys()):
+            spec_results = by_spec[spec]
+            times = {b: r.mean_time for b, r in spec_results.items() if r.success}
+            best_time = min(times.values()) if times else 0.0
+
+            row = [spec]
+            for backend in backends:
+                if backend in spec_results:
+                    r = spec_results[backend]
+                    if r.success:
+                        row.append(f"{r.mean_time:.6f}")
+                        if multi and compare_to_fastest:
+                            pct = (
+                                (r.mean_time / best_time * 100) if best_time > 0 else 0
+                            )
+                            pct_str = f"{pct:.1f}%"
+                            if r.mean_time == best_time:
+                                pct_str = f"[bold green]{pct_str}[/]"
+                            row.append(pct_str)
+                    else:
+                        row.append("[red]ERROR[/]")
+                        if multi and compare_to_fastest:
+                            row.append("-")
+                else:
+                    row.append("-")
+                    if multi and compare_to_fastest:
+                        row.append("-")
+
+            table.add_row(*row)
+
+        self.console.print(table)
+
+    def save_csv(self, results: list[BenchmarkResult], path: str):
+        """Save results to CSV file."""
+        if not results:
+            return
+
+        path_obj = Path(path)
+        path_obj.parent.mkdir(parents=True, exist_ok=True)
+
+        with open(path, "w", newline="") as f:
+            writer = csv.DictWriter(
+                f,
+                fieldnames=[
+                    "backend",
+                    "batch_spec",
+                    "num_layers",
+                    "mean_time",
+                    "std_time",
+                    "throughput",
+                    "memory_mb",
+                ],
+            )
+            writer.writeheader()
+            for r in results:
+                writer.writerow(
+                    {
+                        "backend": r.config.backend,
+                        "batch_spec": r.config.batch_spec,
+                        "num_layers": r.config.num_layers,
+                        "mean_time": r.mean_time,
+                        "std_time": r.std_time,
+                        "throughput": r.throughput_tokens_per_sec or 0,
+                        "memory_mb": r.memory_allocated_mb or 0,
+                    }
+                )
+
+        self.console.print(f"[green]Saved CSV results to {path}[/]")
+
+    def save_json(self, results: list[BenchmarkResult], path: str):
+        """Save results to JSON file."""
+        path_obj = Path(path)
+        path_obj.parent.mkdir(parents=True, exist_ok=True)
+
+        data = [r.to_dict() for r in results]
+        with open(path, "w") as f:
+            json.dump(data, f, indent=2, default=str)
+
+        self.console.print(f"[green]Saved JSON results to {path}[/]")
+
+
+def setup_mla_dims(model_name: str = "deepseek-v3") -> dict:
+    """
+    Get MLA dimensions for known models.
+
+    Args:
+        model_name: Model identifier
+
+    Returns:
+        Dict with MLA dimension configuration
+    """
+    configs = {
+        "deepseek-v2": {
+            "kv_lora_rank": 512,
+            "qk_nope_head_dim": 128,
+            "qk_rope_head_dim": 64,
+            "v_head_dim": 128,
+            "num_q_heads": 128,
+            "num_kv_heads": 1,
+            "head_dim": 576,
+        },
+        "deepseek-v3": {
+            "kv_lora_rank": 512,
+            "qk_nope_head_dim": 128,
+            "qk_rope_head_dim": 64,
+            "v_head_dim": 128,
+            "num_q_heads": 128,
+            "num_kv_heads": 1,
+            "head_dim": 576,
+        },
+        "deepseek-v2-lite": {
+            "kv_lora_rank": 512,
+            "qk_nope_head_dim": 128,
+            "qk_rope_head_dim": 64,
+            "v_head_dim": 128,
+            "num_q_heads": 16,
+            "num_kv_heads": 1,
+            "head_dim": 576,
+        },
+    }
+
+    if model_name not in configs:
+        raise ValueError(
+            f"Unknown model '{model_name}'. Known models: {list(configs.keys())}"
+        )
+
+    return configs[model_name]
+
+
+def get_attention_scale(head_dim: int) -> float:
+    """Compute attention scale factor (1/sqrt(d))."""
+    return 1.0 / math.sqrt(head_dim)
+
+
+def is_mla_backend(backend: str) -> bool:
+    """
+    Check if backend is an MLA backend using the backend's is_mla() property.
+
+    Args:
+        backend: Backend name (e.g., "CUTLASS_MLA", "FLASHINFER_MLA")
+
+    Returns:
+        True if the backend is an MLA backend, False otherwise
+    """
+    from vllm.v1.attention.backends.registry import AttentionBackendEnum
+
+    try:
+        backend_class = AttentionBackendEnum[backend.upper()].get_class()
+        return backend_class.is_mla()
+    except (KeyError, ValueError, ImportError):
+        return False

benchmarks/attention_benchmarks/configs/mla_decode.yaml

@@ -0,0 +1,61 @@
+# MLA decode-only benchmark configuration
+
+model:
+  name: "deepseek-v3"
+  num_layers: 60
+  num_q_heads: 128
+  num_kv_heads: 1  # MLA uses single latent KV
+  head_dim: 576
+  kv_lora_rank: 512
+  qk_nope_head_dim: 128
+  qk_rope_head_dim: 64
+  v_head_dim: 128
+  block_size: 128  # CUTLASS MLA and FlashAttn MLA use 128
+
+batch_specs:
+  # Small batches, varying sequence lengths
+  - "16q1s512"     # 16 requests, 512 KV cache
+  - "16q1s1k"      # 16 requests, 1k KV cache
+  - "16q1s2k"      # 16 requests, 2k KV cache
+  - "16q1s4k"      # 16 requests, 4k KV cache
+
+  # Medium batches
+  - "32q1s1k"      # 32 requests, 1k KV cache
+  - "32q1s2k"      # 32 requests, 2k KV cache
+  - "32q1s4k"      # 32 requests, 4k KV cache
+  - "32q1s8k"      # 32 requests, 8k KV cache
+
+  # Large batches
+  - "64q1s1k"      # 64 requests, 1k KV cache
+  - "64q1s2k"      # 64 requests, 2k KV cache
+  - "64q1s4k"      # 64 requests, 4k KV cache
+  - "64q1s8k"      # 64 requests, 8k KV cache
+
+  # Very large batches
+  - "128q1s1k"     # 128 requests, 1k KV cache
+  - "128q1s2k"     # 128 requests, 2k KV cache
+
+  # Long context
+  - "32q1s16k"     # 32 requests, 16k KV cache
+  - "32q1s32k"     # 32 requests, 32k KV cache
+
+backends:
+  - cutlass_mla
+  - flashinfer_mla
+  - flashattn_mla  # Hopper only
+  - flashmla        # Hopper only
+
+device: "cuda:0"
+repeats: 5
+warmup_iters: 3
+profile_memory: true
+
+# Backend-specific tuning
+cutlass_mla:
+  num_kv_splits: auto  # or specific value like 4, 8, 16
+
+flashattn_mla:
+  reorder_batch_threshold: 512
+
+flashmla:
+  reorder_batch_threshold: 1

benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

@@ -0,0 +1,60 @@
+# MLA mixed batch benchmark (prefill + decode)
+# Tests chunked prefill performance
+
+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
+
+batch_specs:
+  # Small prefill + decode
+  - "1q1k_8q1s1k"           # 1 prefill + 8 decode
+  - "2q2k_16q1s1k"          # 2 prefill + 16 decode
+  - "4q1k_32q1s2k"          # 4 prefill + 32 decode
+
+  # Medium prefill + decode
+  - "2q4k_32q1s2k"          # 2 medium prefill + 32 decode
+  - "4q4k_64q1s2k"          # 4 medium prefill + 64 decode
+  - "8q2k_64q1s4k"          # 8 prefill + 64 decode
+
+  # Large prefill + decode (chunked prefill stress test)
+  - "2q8k_32q1s1k"          # 2 large prefill + 32 decode
+  - "1q16k_16q1s2k"         # 1 very large prefill + 16 decode
+  - "2q16k_32q1s4k"         # 2 very large prefill + 32 decode
+
+  # Context extension + decode
+  - "2q1kkv2k_16q1s1k"       # 2 extend + 16 decode
+  - "4q2kkv4k_32q1s2k"       # 4 extend + 32 decode
+  - "2q1kkv8k_32q1s2k"       # 2 large extend + 32 decode
+
+  # Explicitly chunked prefill
+  - "q8k"           # 8k prefill with chunking hint
+  - "q16k"          # 16k prefill with chunking hint
+  - "2q8k_32q1s2k"    # 2 chunked prefill + 32 decode
+
+  # High decode ratio (realistic serving)
+  - "1q2k_63q1s1k"          # 1 prefill + 63 decode
+  - "2q2k_62q1s2k"          # 2 prefill + 62 decode
+  - "4q4k_60q1s4k"          # 4 prefill + 60 decode
+
+backends:
+  - cutlass_mla
+  - flashinfer_mla
+  - flashattn_mla   # Hopper only
+  - flashmla        # Hopper only
+
+device: "cuda:0"
+repeats: 5
+warmup_iters: 3
+profile_memory: true
+
+# Analyze chunked prefill workspace size impact
+chunked_prefill:
+  test_workspace_sizes: [4096, 8192, 16384, 32768, 65536]

benchmarks/attention_benchmarks/configs/reorder_threshold.yaml

@@ -0,0 +1,88 @@
+# Study 4: What is optimal reorder_batch_threshold for MLA backends supporting query length > 1?
+# Question: At what query length does prefill pipeline become faster than decode pipeline?
+# Methodology: For each query length, compare decode vs prefill performance to find crossover point
+# Applies to: FlashAttn MLA, FlashMLA
+
+description: "Decode vs Prefill pipeline crossover analysis"
+
+# Test FlashAttn MLA
+backend: flashattn_mla
+
+# Mode: decode_vs_prefill comparison (special sweep mode)
+# For each batch spec, we'll test both decode and prefill pipelines
+mode: "decode_vs_prefill"
+
+# Query lengths to test (from old benchmark_mla_threshold.py methodology)
+# Each query length will be tested with BOTH decode and prefill pipelines:
+#   - decode: threshold >= query_length (forces decode pipeline)
+#   - prefill: threshold < query_length (forces prefill pipeline)
+#
+# We use q<N>s1k format which creates q_len=N, seq_len=1024 requests
+# This tests different query lengths with fixed sequence length context
+#
+# Using batch_spec_ranges for automatic generation:
+batch_spec_ranges:
+  - template: "q{q_len}s1k"
+    q_len:
+      start: 1
+      stop: 16
+      step: 1
+      end_inclusive: false
+  - template: "q{q_len}s1k"
+    q_len:
+      start: 16
+      stop: 64
+      step: 2
+      end_inclusive: false
+  - template: "q{q_len}s1k"
+    q_len:
+      start: 64
+      stop: 1024
+      step: 4
+      end_inclusive: true
+
+# Batch sizes to test (from old script)
+batch_sizes:
+  - 1
+  - 2
+  - 4
+  - 8
+  - 16
+  - 32
+  - 64
+  - 128
+  - 256
+
+# Model configuration (DeepSeek V2/V3 defaults)
+model:
+  num_layers: 10
+  head_dim: 576
+  num_q_heads: 128
+  num_kv_heads: 1
+  block_size: 128
+
+# Benchmark settings
+benchmark:
+  device: "cuda:0"
+  repeats: 15          # More repeats for spec decode variance
+  warmup_iters: 5
+  profile_memory: false
+
+# Output
+output:
+  csv: "reorder_threshold_results.csv"
+  json: "reorder_threshold_results.json"
+
+# Expected outcome (reproduces old benchmark_mla_threshold.py study):
+# - For each batch size, find the crossover point where prefill becomes faster than decode
+# - Show decode vs prefill performance across all query lengths
+# - Determine optimal reorder_batch_threshold based on last query length where decode is faster
+# - Understand how crossover point varies with batch size
+# - Provide data-driven guidance for default threshold value
+#
+# Methodology (from old script):
+# - Each query length tested with BOTH pipelines:
+#     * decode: threshold >= query_length (forces decode pipeline)
+#     * prefill: threshold < query_length (forces prefill pipeline)
+# - Compare which is faster to find crossover point
+#

benchmarks/attention_benchmarks/configs/speculative_decode.yaml

@@ -0,0 +1,62 @@
+# Speculative decoding benchmark configuration
+# Tests reorder_batch_threshold optimization
+
+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
+
+batch_specs:
+  # Pure speculative decode (K-token verification)
+  - "q2s1k"      # 2-token spec, 1k KV
+  - "q4s1k"      # 4-token spec, 1k KV
+  - "q8s1k"      # 8-token spec, 1k KV
+  - "q16s1k"     # 16-token spec, 1k KV
+
+  # Speculative with different context lengths
+  - "q4s2k"      # 4-token spec, 2k KV
+  - "q4s4k"      # 4-token spec, 4k KV
+  - "q8s2k"      # 8-token spec, 2k KV
+  - "q8s4k"      # 8-token spec, 4k KV
+
+  # Mixed: speculative + regular decode
+  - "32q4s1k"                    # 32 spec requests
+  - "16q4s1k_16q1s1k"              # 16 spec + 16 regular
+  - "8q8s2k_24q1s2k"               # 8 spec (8-tok) + 24 regular
+
+  # Mixed: speculative + prefill + decode
+  - "2q1k_16q4s1k_16q1s1k"         # 2 prefill + 16 spec + 16 decode
+  - "4q2k_32q4s2k_32q1s2k"         # 4 prefill + 32 spec + 32 decode
+
+  # Large batches with speculation
+  - "64q4s1k"                    # 64 spec requests
+  - "32q8s2k"                    # 32 spec (8-token)
+  - "16q16s4k"                   # 16 spec (16-token)
+
+# Backends that support query length > 1
+backends:
+  - flashattn_mla    # reorder_batch_threshold = 512
+  - flashmla          # reorder_batch_threshold = 1 (tunable)
+
+# FlashInfer-MLA also supports uniform spec-as-decode but with different mechanism
+# - flashinfer_mla
+
+# Benchmark settings
+benchmark:
+  device: "cuda:0"
+  repeats: 10  # More repeats for statistical significance
+  warmup_iters: 5
+  profile_memory: false
+
+# Test these threshold values for optimization
+parameter_sweep:
+  param_name: "reorder_batch_threshold"
+  values: [1, 2, 4, 8, 16, 32, 64, 128, 256, 512]
+  include_auto: false
+  label_format: "{backend}_threshold_{value}"

benchmarks/attention_benchmarks/configs/standard_attention.yaml

@@ -0,0 +1,40 @@
+# Standard attention backend benchmark configuration
+
+model:
+  num_layers: 32
+  num_q_heads: 32
+  num_kv_heads: 8  # GQA with 4:1 ratio
+  head_dim: 128
+  block_size: 16
+
+batch_specs:
+  # Pure prefill
+  - "q512"      # Small prefill (512 tokens)
+  - "q2k"       # Medium prefill (2048 tokens)
+  - "q4k"       # Large prefill (4096 tokens)
+  - "q8k"       # Very large prefill (8192 tokens)
+
+  # Pure decode
+  - "8q1s1k"      # 8 requests, 1k KV cache each
+  - "16q1s2k"     # 16 requests, 2k KV cache each
+  - "32q1s1k"     # 32 requests, 1k KV cache each
+  - "64q1s4k"     # 64 requests, 4k KV cache each
+
+  # Mixed prefill/decode
+  - "2q2k_8q1s1k"      # 2 prefill + 8 decode
+  - "4q1k_16q1s2k"     # 4 prefill + 16 decode
+  - "2q4k_32q1s1k"     # 2 large prefill + 32 decode
+
+  # Context extension
+  - "q1ks2k"          # 1k query, 2k sequence (chunked prefill)
+  - "2q1ks4k"         # 2 requests: 1k query, 4k sequence
+
+backends:
+  - flash
+  - triton
+  - flashinfer
+
+device: "cuda:0"
+repeats: 5
+warmup_iters: 3
+profile_memory: false

benchmarks/attention_benchmarks/mla_runner.py

@@ -0,0 +1,836 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+MLA benchmark runner - shared utilities for MLA benchmarks.
+
+This module provides helpers for running MLA backends without
+needing full VllmConfig integration.
+"""
+
+import importlib
+
+import numpy as np
+import torch
+from batch_spec import parse_batch_spec
+from common import (
+    BenchmarkResult,
+    MockHfConfig,
+    MockKVBProj,
+    MockLayer,
+    setup_mla_dims,
+)
+
+from vllm.config import (
+    CacheConfig,
+    CompilationConfig,
+    ModelConfig,
+    ParallelConfig,
+    SchedulerConfig,
+    VllmConfig,
+    set_current_vllm_config,
+)
+
+# ============================================================================
+# VllmConfig Creation
+# ============================================================================
+
+
+def _add_mock_methods_to_model_config(model_config: ModelConfig) -> None:
+    """
+    Add mock methods for layer-specific queries to ModelConfig.
+
+    These methods are needed by metadata builders but aren't normally
+    present on ModelConfig when used in benchmark contexts.
+    """
+    import types
+
+    model_config.get_num_layers = types.MethodType(lambda self: 1, model_config)
+    model_config.get_sliding_window_for_layer = types.MethodType(
+        lambda self, _i: None, model_config
+    )
+    model_config.get_logits_soft_cap_for_layer = types.MethodType(
+        lambda self, _i: None, model_config
+    )
+    model_config.get_sm_scale_for_layer = types.MethodType(
+        lambda self, _i: 1.0 / model_config.get_head_size() ** 0.5, model_config
+    )
+
+
+def create_minimal_vllm_config(
+    model_name: str = "deepseek-v3",
+    block_size: int = 128,
+    max_num_seqs: int = 256,
+    mla_dims: dict | None = None,
+) -> VllmConfig:
+    """
+    Create minimal VllmConfig for MLA benchmarks.
+
+    Args:
+        model_name: Model name (deepseek-v2, deepseek-v3, etc.) - used if mla_dims not
+                    provided
+        block_size: KV cache block size
+        max_num_seqs: Maximum number of sequences
+        mla_dims: Optional custom MLA dimensions dict. If not provided, uses
+                  setup_mla_dims(model_name)
+
+    Returns:
+        VllmConfig for benchmarking
+    """
+    # Get MLA dimensions - use provided or load from model name
+    if mla_dims is None:
+        mla_dims = setup_mla_dims(model_name)
+
+    # Create mock HF config first (avoids downloading from HuggingFace)
+    mock_hf_config = MockHfConfig(mla_dims)
+
+    # Create a temporary minimal config.json to avoid HF downloads
+    # This ensures consistent ModelConfig construction without network access
+    import json
+    import os
+    import shutil
+    import tempfile
+
+    minimal_config = {
+        "architectures": ["DeepseekV2ForCausalLM"],
+        "model_type": "deepseek_v2",
+        "num_attention_heads": mla_dims["num_q_heads"],
+        "num_key_value_heads": mla_dims["num_kv_heads"],
+        "hidden_size": mla_dims["head_dim"] * mla_dims["num_q_heads"],
+        "torch_dtype": "bfloat16",
+        "max_position_embeddings": 163840,  # DeepSeek V3 default
+        "rope_theta": 10000.0,
+        "vocab_size": 128256,
+    }
+
+    # Create temporary directory with config.json
+    temp_dir = tempfile.mkdtemp(prefix="vllm_bench_")
+    config_path = os.path.join(temp_dir, "config.json")
+    with open(config_path, "w") as f:
+        json.dump(minimal_config, f)
+
+    try:
+        # Create model config using local path - no HF downloads
+        model_config = ModelConfig(
+            model=temp_dir,  # Use local temp directory
+            tokenizer=None,
+            tokenizer_mode="auto",
+            trust_remote_code=True,
+            dtype="bfloat16",
+            seed=0,
+            max_model_len=32768,
+            quantization=None,
+            quantization_param_path=None,
+            enforce_eager=False,
+            max_context_len_to_capture=None,
+            max_seq_len_to_capture=8192,
+            max_logprobs=20,
+            disable_sliding_window=False,
+            skip_tokenizer_init=True,
+            served_model_name=None,
+            limit_mm_per_prompt=None,
+            use_async_output_proc=True,
+            config_format="auto",
+        )
+    finally:
+        # Clean up temporary directory
+        shutil.rmtree(temp_dir, ignore_errors=True)
+
+    # Override with our mock config
+    model_config.hf_config = mock_hf_config
+    model_config.hf_text_config = mock_hf_config
+
+    # Add mock methods for layer-specific queries
+    _add_mock_methods_to_model_config(model_config)
+
+    # Create sub-configs
+    cache_config = CacheConfig(
+        block_size=block_size,
+        gpu_memory_utilization=0.9,
+        swap_space=0,
+        cache_dtype="auto",
+        enable_prefix_caching=False,
+    )
+
+    scheduler_config = SchedulerConfig(
+        max_num_seqs=max_num_seqs,
+        max_num_batched_tokens=8192,
+        max_model_len=32768,
+        is_encoder_decoder=False,
+        enable_chunked_prefill=True,
+    )
+
+    parallel_config = ParallelConfig(
+        tensor_parallel_size=1,
+    )
+
+    compilation_config = CompilationConfig()
+
+    return VllmConfig(
+        model_config=model_config,
+        cache_config=cache_config,
+        parallel_config=parallel_config,
+        scheduler_config=scheduler_config,
+        compilation_config=compilation_config,
+    )
+
+
+# ============================================================================
+# Backend Configuration
+# ============================================================================
+
+
+# Backend name to class name prefix mapping
+_BACKEND_NAME_MAP = {
+    "flashattn_mla": "FlashAttnMLA",
+    "flashmla": "FlashMLA",
+    "flashinfer_mla": "FlashInferMLA",
+    "cutlass_mla": "CutlassMLA",
+}
+
+# Special properties that differ from defaults
+_BACKEND_PROPERTIES = {
+    "flashmla": {
+        "query_format": "concat",  # Single concatenated tensor (vs tuple)
+        "block_size": 64,  # FlashMLA uses fixed block size
+    },
+    "flashinfer_mla": {
+        "block_size": 64,  # FlashInfer MLA only supports 32 or 64
+    },
+}
+
+
+def _get_backend_config(backend: str) -> dict:
+    """
+    Get backend configuration using naming conventions.
+
+    All MLA backends follow the pattern:
+    - Module: vllm.v1.attention.backends.mla.{backend}
+    - Impl: {Name}Impl
+    - Metadata: {Name}Metadata (or MLACommonMetadata)
+    - DecodeMetadata: {Name}DecodeMetadata (or MLACommonDecodeMetadata)
+    - MetadataBuilder: {Name}MetadataBuilder
+    """
+    if backend not in _BACKEND_NAME_MAP:
+        raise ValueError(f"Unknown backend: {backend}")
+
+    name = _BACKEND_NAME_MAP[backend]
+    props = _BACKEND_PROPERTIES.get(backend, {})
+
+    # Check if backend uses common metadata (FlashInfer, CUTLASS)
+    uses_common = backend in ("flashinfer_mla", "cutlass_mla")
+
+    return {
+        "module": f"vllm.v1.attention.backends.mla.{backend}",
+        "impl_class": f"{name}Impl",
+        "metadata_class": "MLACommonMetadata" if uses_common else f"{name}Metadata",
+        "decode_metadata_class": "MLACommonDecodeMetadata"
+        if uses_common
+        else f"{name}DecodeMetadata",
+        "builder_class": f"{name}MetadataBuilder",
+        "query_format": props.get("query_format", "tuple"),
+        "block_size": props.get("block_size", None),
+    }
+
+
+# ============================================================================
+# Metadata Building Helpers
+# ============================================================================
+
+
+def _build_attention_metadata(
+    requests: list,
+    block_size: int,
+    device: torch.device,
+    builder_instance,
+) -> tuple:
+    """
+    Build attention metadata from batch requests.
+
+    Args:
+        requests: List of BatchRequest objects
+        block_size: KV cache block size
+        device: Target device
+        builder_instance: Metadata builder instance
+
+    Returns:
+        Tuple of (metadata, kv_cache_num_blocks)
+    """
+    q_lens = [r.q_len for r in requests]
+    kv_lens = [r.kv_len for r in requests]
+    total_q = sum(q_lens)
+    max_kv = max(kv_lens)
+
+    # Build query start locations
+    q_start_cpu = torch.tensor(
+        [0] + [sum(q_lens[: i + 1]) for i in range(len(q_lens))],
+        dtype=torch.int32,
+    )
+    q_start_gpu = q_start_cpu.to(device)
+
+    # Build sequence lengths
+    seq_lens_cpu = torch.tensor(kv_lens, dtype=torch.int32)
+    seq_lens_gpu = seq_lens_cpu.to(device)
+
+    # Build num_computed_tokens (context length for each request)
+    context_lens = [kv_len - q_len for q_len, kv_len in zip(q_lens, kv_lens)]
+    num_computed_tokens_cpu = torch.tensor(context_lens, dtype=torch.int32)
+
+    # Build block table
+    num_blocks_per_req = [(kv + block_size - 1) // block_size for kv in kv_lens]
+    max_num_blocks = max(num_blocks_per_req)
+
+    block_table_cpu = np.zeros((len(requests), max_num_blocks), dtype=np.int32)
+    current_block = 0
+    for i, num_blocks in enumerate(num_blocks_per_req):
+        for j in range(num_blocks):
+            block_table_cpu[i, j] = current_block
+            current_block += 1
+
+    block_table_gpu = torch.from_numpy(block_table_cpu).to(device)
+
+    # Build slot mapping
+    slot_mapping_list = []
+    for i, (q_len, kv_len, num_blocks) in enumerate(
+        zip(q_lens, kv_lens, num_blocks_per_req)
+    ):
+        context_len = kv_len - q_len
+        for j in range(q_len):
+            token_kv_idx = context_len + j
+            block_idx = token_kv_idx // block_size
+            offset_in_block = token_kv_idx % block_size
+            global_block_id = block_table_cpu[i, block_idx]
+            slot_id = global_block_id * block_size + offset_in_block
+            slot_mapping_list.append(slot_id)
+
+    slot_mapping = torch.tensor(slot_mapping_list, dtype=torch.int64, device=device)
+
+    # Create CommonAttentionMetadata
+    from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+
+    common_attn_metadata = CommonAttentionMetadata(
+        num_reqs=len(requests),
+        max_query_len=max(q_lens),
+        max_seq_len=max_kv,
+        num_actual_tokens=total_q,
+        query_start_loc=q_start_gpu,
+        query_start_loc_cpu=q_start_cpu,
+        seq_lens=seq_lens_gpu,
+        _seq_lens_cpu=seq_lens_cpu,
+        _num_computed_tokens_cpu=num_computed_tokens_cpu,
+        slot_mapping=slot_mapping,
+        block_table_tensor=block_table_gpu,
+        dcp_local_seq_lens=None,
+    )
+
+    # Use the production build() method
+    metadata = builder_instance.build(
+        common_prefix_len=0,
+        common_attn_metadata=common_attn_metadata,
+        fast_build=False,
+    )
+
+    return metadata, current_block
+
+
+def _create_input_tensors(
+    total_q: int,
+    mla_dims: dict,
+    query_format: str,
+    device: torch.device,
+    dtype: torch.dtype,
+):
+    """
+    Create input tensors for both decode and prefill modes.
+
+    MLA requires different tensor formats for decode vs prefill:
+    - Decode: Uses kv_lora_rank (512) dimension
+    - Prefill: Uses qk_nope_head_dim (128) to stay under FlashAttention's 256 limit
+
+    Args:
+        total_q: Total number of query tokens
+        mla_dims: MLA dimension configuration
+        query_format: Either "tuple" or "concat"
+        device: Target device
+        dtype: Tensor dtype
+
+    Returns:
+        Tuple of (decode_inputs, prefill_inputs)
+        - decode_inputs: Query tensor(s) for decode mode
+        - prefill_inputs: Dict with 'q', 'k_c_normed', 'k_pe', 'k_scale' for prefill
+    """
+    if query_format == "tuple":
+        # Decode mode format: (q_nope, q_pe) where q_nope has kv_lora_rank dim
+        q_nope_decode = torch.randn(
+            total_q,
+            mla_dims["num_q_heads"],
+            mla_dims["kv_lora_rank"],
+            device=device,
+            dtype=dtype,
+        )
+        q_pe = torch.randn(
+            total_q,
+            mla_dims["num_q_heads"],
+            mla_dims["qk_rope_head_dim"],
+            device=device,
+            dtype=dtype,
+        )
+        decode_inputs = (q_nope_decode, q_pe)
+
+        # For prefill, we need q with qk_nope_head_dim instead of kv_lora_rank
+        q_nope_prefill = torch.randn(
+            total_q,
+            mla_dims["num_q_heads"],
+            mla_dims["qk_nope_head_dim"],
+            device=device,
+            dtype=dtype,
+        )
+        prefill_q = torch.cat([q_nope_prefill, q_pe], dim=-1)
+    else:  # concat
+        decode_inputs = torch.randn(
+            total_q,
+            mla_dims["num_q_heads"],
+            mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"],
+            device=device,
+            dtype=dtype,
+        )
+        # For prefill with concat format
+        prefill_q = torch.randn(
+            total_q,
+            mla_dims["num_q_heads"],
+            mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
+            device=device,
+            dtype=dtype,
+        )
+
+    # Create additional inputs needed for prefill forward
+    k_c_normed = torch.randn(
+        total_q,
+        mla_dims["kv_lora_rank"],
+        device=device,
+        dtype=dtype,
+    )
+    k_pe = torch.randn(
+        total_q,
+        1,  # Single head for MLA
+        mla_dims["qk_rope_head_dim"],
+        device=device,
+        dtype=dtype,
+    )
+    k_scale = torch.ones(1, device=device, dtype=torch.float32)
+
+    output = torch.zeros(
+        total_q,
+        mla_dims["num_q_heads"] * mla_dims["v_head_dim"],
+        device=device,
+        dtype=dtype,
+    )
+
+    prefill_inputs = {
+        "q": prefill_q,
+        "k_c_normed": k_c_normed,
+        "k_pe": k_pe,
+        "k_scale": k_scale,
+        "output": output,
+    }
+
+    return decode_inputs, prefill_inputs
+
+
+# ============================================================================
+# Backend Initialization
+# ============================================================================
+
+
+def _create_backend_impl(
+    backend_cfg: dict,
+    mla_dims: dict,
+    vllm_config: VllmConfig,
+    device: torch.device,
+):
+    """
+    Create backend implementation instance.
+
+    Args:
+        backend_cfg: Backend configuration dict
+        mla_dims: MLA dimension configuration
+        vllm_config: VllmConfig instance
+        device: Target device
+
+    Returns:
+        Tuple of (impl, layer, builder_instance)
+    """
+    # Import backend classes
+    backend_module = importlib.import_module(backend_cfg["module"])
+    impl_class = getattr(backend_module, backend_cfg["impl_class"])
+
+    # 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"],
+    )
+
+    # Create impl
+    impl = impl_class(
+        num_heads=mla_dims["num_q_heads"],
+        head_size=mla_dims["head_dim"],
+        scale=scale,
+        num_kv_heads=mla_dims["num_kv_heads"],
+        alibi_slopes=None,
+        sliding_window=None,
+        kv_cache_dtype="auto",
+        logits_soft_cap=None,
+        attn_type="decoder",
+        kv_sharing_target_layer_name=None,
+        q_lora_rank=None,
+        kv_lora_rank=mla_dims["kv_lora_rank"],
+        qk_nope_head_dim=mla_dims["qk_nope_head_dim"],
+        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,
+    )
+
+    # Initialize DCP attributes
+    if not hasattr(impl, "dcp_world_size") or impl.dcp_world_size in (None, -1):
+        impl.dcp_world_size = 1
+        impl.dcp_rank = 0
+
+    # Create KV cache spec for MockLayer
+    from vllm.v1.kv_cache_interface import FullAttentionSpec
+
+    kv_cache_spec = FullAttentionSpec(
+        block_size=backend_cfg["block_size"] or vllm_config.cache_config.block_size,
+        num_kv_heads=1,  # MLA uses 1 KV head
+        head_size=576,  # MLA head dim
+        dtype=torch.bfloat16,
+    )
+
+    # Create mock layer
+    layer = MockLayer(device, impl=impl, kv_cache_spec=kv_cache_spec)
+
+    # Create builder instance if needed
+    builder_instance = None
+    if backend_cfg["builder_class"]:
+        builder_class = getattr(backend_module, backend_cfg["builder_class"])
+
+        # Populate static_forward_context so builder can find the layer
+        # MockLayer inherits from AttentionLayerBase, so isinstance checks pass
+        vllm_config.compilation_config.static_forward_context = {"placeholder": layer}
+
+        builder_instance = builder_class(
+            kv_cache_spec=kv_cache_spec,
+            layer_names=["placeholder"],
+            vllm_config=vllm_config,
+            device=device,
+        )
+
+    return impl, layer, builder_instance
+
+
+# ============================================================================
+# Config Helpers
+# ============================================================================
+
+
+def _extract_mla_dims_from_config(config) -> dict | None:
+    """
+    Extract MLA dimensions from BenchmarkConfig if all required fields are present.
+
+    Args:
+        config: BenchmarkConfig instance
+
+    Returns:
+        Dict with MLA dimensions if all fields are provided, None otherwise
+    """
+    # Check if all MLA-specific fields are provided
+    if all(
+        [
+            config.kv_lora_rank is not None,
+            config.qk_nope_head_dim is not None,
+            config.qk_rope_head_dim is not None,
+            config.v_head_dim is not None,
+        ]
+    ):
+        return {
+            "kv_lora_rank": config.kv_lora_rank,
+            "qk_nope_head_dim": config.qk_nope_head_dim,
+            "qk_rope_head_dim": config.qk_rope_head_dim,
+            "v_head_dim": config.v_head_dim,
+            "num_q_heads": config.num_q_heads,
+            "num_kv_heads": config.num_kv_heads,
+            "head_dim": config.head_dim,
+        }
+    # Fallback: if MLA fields not fully specified, try to construct from basic fields
+    elif config.head_dim == 576:
+        # This looks like a DeepSeek MLA config, use standard dimensions with custom
+        # head count
+        return {
+            "kv_lora_rank": 512,
+            "qk_nope_head_dim": 128,
+            "qk_rope_head_dim": 64,
+            "v_head_dim": 128,
+            "num_q_heads": config.num_q_heads,
+            "num_kv_heads": config.num_kv_heads,
+            "head_dim": config.head_dim,
+        }
+    return None
+
+
+# ============================================================================
+# Benchmark Execution
+# ============================================================================
+
+
+def _run_single_benchmark(
+    config,
+    impl,
+    layer,
+    builder_instance,
+    backend_cfg: dict,
+    mla_dims: dict,
+    device: torch.device,
+) -> BenchmarkResult:
+    """
+    Run a single benchmark iteration.
+
+    Args:
+        config: BenchmarkConfig instance
+        impl: Backend implementation instance
+        layer: MockLayer instance
+        builder_instance: Metadata builder instance
+        backend_cfg: Backend configuration dict
+        mla_dims: MLA dimension configuration
+        device: Target device
+
+    Returns:
+        BenchmarkResult with timing statistics
+    """
+    # Parse batch spec
+    requests = parse_batch_spec(config.batch_spec)
+    q_lens = [r.q_len for r in requests]
+    total_q = sum(q_lens)
+
+    # Determine block size
+    block_size = backend_cfg["block_size"] or config.block_size
+
+    # Build metadata
+    metadata, num_blocks = _build_attention_metadata(
+        requests, block_size, device, builder_instance
+    )
+
+    # Create KV cache
+    kv_cache = torch.zeros(
+        num_blocks,
+        block_size,
+        mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"],
+        device=device,
+        dtype=torch.bfloat16,
+    )
+
+    # Create input tensors for both decode and prefill modes
+    decode_inputs, prefill_inputs = _create_input_tensors(
+        total_q,
+        mla_dims,
+        backend_cfg["query_format"],
+        device,
+        torch.bfloat16,
+    )
+
+    # Determine which forward method to use based on metadata
+    if metadata.decode is not None:
+        forward_fn = lambda: impl._forward_decode(
+            decode_inputs, kv_cache, metadata, layer
+        )
+    elif metadata.prefill is not None:
+        forward_fn = lambda: impl._forward_prefill(
+            prefill_inputs["q"],
+            prefill_inputs["k_c_normed"],
+            prefill_inputs["k_pe"],
+            kv_cache,
+            metadata,
+            prefill_inputs["k_scale"],
+            prefill_inputs["output"],
+        )
+    else:
+        raise RuntimeError("Metadata has neither decode nor prefill metadata")
+
+    # Warmup
+    for _ in range(config.warmup_iters):
+        forward_fn()
+    torch.cuda.synchronize()
+
+    # Benchmark
+    times = []
+    for _ in range(config.repeats):
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+
+        start.record()
+        for _ in range(config.num_layers):
+            forward_fn()
+        end.record()
+
+        torch.cuda.synchronize()
+        elapsed_ms = start.elapsed_time(end)
+        times.append(elapsed_ms / 1000.0 / config.num_layers)
+
+    mean_time = float(np.mean(times))
+    return BenchmarkResult(
+        config=config,
+        mean_time=mean_time,
+        std_time=float(np.std(times)),
+        min_time=float(np.min(times)),
+        max_time=float(np.max(times)),
+        throughput_tokens_per_sec=total_q / mean_time if mean_time > 0 else 0,
+    )
+
+
+def _run_mla_benchmark_batched(
+    backend: str,
+    configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
+) -> list[BenchmarkResult]:
+    """
+    Unified batched MLA benchmark runner for all backends.
+
+    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla
+
+    This function reuses backend initialization across multiple benchmarks
+    to avoid setup/teardown overhead.
+
+    Args:
+        backend: Backend name
+        configs_with_params: List of (config, threshold, num_splits) tuples
+            - threshold: reorder_batch_threshold (FlashAttn/FlashMLA only)
+            - num_splits: num_kv_splits (CUTLASS only)
+
+    Returns:
+        List of BenchmarkResult objects
+    """
+    if not configs_with_params:
+        return []
+
+    backend_cfg = _get_backend_config(backend)
+    device = torch.device(configs_with_params[0][0].device)
+    torch.cuda.set_device(device)
+
+    # Determine block size
+    config_block_size = configs_with_params[0][0].block_size
+    block_size = backend_cfg["block_size"] or config_block_size
+
+    # Extract MLA dimensions from the first config
+    first_config = configs_with_params[0][0]
+    mla_dims = _extract_mla_dims_from_config(first_config)
+
+    # If config didn't provide MLA dims, fall back to default model
+    if mla_dims is None:
+        mla_dims = setup_mla_dims("deepseek-v3")
+
+    # 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,
+        mla_dims=mla_dims,  # Use custom dims from config or default
+    )
+
+    results = []
+
+    with set_current_vllm_config(vllm_config):
+        # Create backend impl, layer, and builder (reused across benchmarks)
+        impl, layer, builder_instance = _create_backend_impl(
+            backend_cfg, mla_dims, vllm_config, device
+        )
+
+        # Run each benchmark with the shared impl
+        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:
+                original_threshold = builder_instance.reorder_batch_threshold
+                builder_instance.reorder_batch_threshold = threshold
+
+            # Set num_splits for CUTLASS
+            original_num_splits = None
+            if num_splits is not None and hasattr(impl, "_num_kv_splits"):
+                original_num_splits = impl._num_kv_splits
+                impl._num_kv_splits = num_splits
+
+            try:
+                result = _run_single_benchmark(
+                    config,
+                    impl,
+                    layer,
+                    builder_instance,
+                    backend_cfg,
+                    mla_dims,
+                    device,
+                )
+                results.append(result)
+
+            finally:
+                # Restore original threshold
+                if original_threshold is not None:
+                    builder_instance.reorder_batch_threshold = original_threshold
+
+                # Restore original num_splits
+                if original_num_splits is not None:
+                    impl._num_kv_splits = original_num_splits
+
+    return results
+
+
+# ============================================================================
+# Public API
+# ============================================================================
+
+
+def run_mla_benchmark(
+    backend: str,
+    config,
+    reorder_batch_threshold: int | None = None,
+    num_kv_splits: int | None = None,
+) -> BenchmarkResult | list[BenchmarkResult]:
+    """
+    Unified MLA benchmark runner for all backends.
+
+    Works for: flashattn_mla, flashmla, flashinfer_mla, cutlass_mla
+
+    Always uses batched execution internally for optimal performance.
+
+    Args:
+        backend: Backend name (flashattn_mla, flashmla, flashinfer_mla, cutlass_mla)
+        config: BenchmarkConfig or list of (BenchmarkConfig, param) tuples
+        reorder_batch_threshold: Threshold override for FlashAttn/FlashMLA
+                                 (single config mode only)
+        num_kv_splits: Number of KV splits for CUTLASS (single config mode only)
+
+    Returns:
+        BenchmarkResult (single mode) or list of BenchmarkResult (batched mode)
+    """
+    # Normalize to batched mode: (config, threshold, num_splits)
+    if isinstance(config, list):
+        # Already in batched format
+        if len(config) > 0 and isinstance(config[0], tuple):
+            # Format: [(cfg, param), ...] where param is threshold or num_splits
+            if backend in ("flashattn_mla", "flashmla"):
+                configs_with_params = [(cfg, param, None) for cfg, param in config]
+            else:  # cutlass_mla or flashinfer_mla
+                configs_with_params = [(cfg, None, param) for cfg, param in config]
+        else:
+            # Format: [cfg, ...] - just configs
+            configs_with_params = [(cfg, None, None) for cfg in config]
+        return_single = False
+    else:
+        # Single config: convert to batched format
+        configs_with_params = [(config, reorder_batch_threshold, num_kv_splits)]
+        return_single = True
+
+    # Use unified batched execution
+    results = _run_mla_benchmark_batched(backend, configs_with_params)
+
+    # Return single result or list based on input
+    return results[0] if return_single else results

benchmarks/attention_benchmarks/runner.py

@@ -0,0 +1,481 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+Standard attention benchmark runner - shared utilities for non-MLA benchmarks.
+
+This module provides helpers for running standard attention backends
+(FlashAttention, Triton, FlashInfer) with real vLLM integration.
+"""
+
+import types
+
+import numpy as np
+import torch
+from batch_spec import parse_batch_spec, reorder_for_flashinfer
+from common import BenchmarkConfig, BenchmarkResult, MockLayer, get_attention_scale
+
+from vllm.config import (
+    CacheConfig,
+    CompilationConfig,
+    DeviceConfig,
+    LoadConfig,
+    ModelConfig,
+    ParallelConfig,
+    SchedulerConfig,
+    VllmConfig,
+)
+from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.kv_cache_interface import FullAttentionSpec
+
+# ============================================================================
+# Backend Configuration
+# ============================================================================
+
+
+_BACKEND_CONFIG = {
+    "flash": {
+        "module": "vllm.v1.attention.backends.flash_attn",
+        "backend_class": "FlashAttentionBackend",
+        "dtype": torch.float16,
+        "cache_layout": "standard",
+        # ^ [2, num_blocks, block_size, num_kv_heads, head_dim]
+    },
+    "triton": {
+        "module": "vllm.v1.attention.backends.triton_attn",
+        "backend_class": "TritonAttentionBackend",
+        "dtype": torch.float32,
+        "cache_layout": "standard",
+    },
+    "flashinfer": {
+        "module": "vllm.v1.attention.backends.flashinfer",
+        "backend_class": "FlashInferBackend",
+        "dtype": torch.float16,
+        "cache_layout": "flashinfer",
+        # ^ [num_blocks, 2, block_size, num_kv_heads, head_dim]
+    },
+}
+
+
+def _get_backend_config(backend: str) -> dict:
+    if backend not in _BACKEND_CONFIG:
+        raise ValueError(
+            f"Unknown backend: {backend}. "
+            f"Available: {', '.join(_BACKEND_CONFIG.keys())}"
+        )
+    return _BACKEND_CONFIG[backend]
+
+
+# ============================================================================
+# Metadata Building Helpers
+# ============================================================================
+
+
+def _build_common_attn_metadata(
+    q_lens: list[int],
+    kv_lens: list[int],
+    block_size: int,
+    device: torch.device,
+) -> CommonAttentionMetadata:
+    """Build CommonAttentionMetadata from query/kv lengths."""
+    batch_size = len(q_lens)
+    total_tokens = sum(q_lens)
+
+    query_start_loc = torch.zeros(batch_size + 1, dtype=torch.int32, device=device)
+    query_start_loc[1:] = torch.tensor(q_lens, dtype=torch.int32, device=device).cumsum(
+        0
+    )
+    query_start_loc_cpu = query_start_loc.cpu()
+
+    seq_lens = torch.tensor(kv_lens, dtype=torch.int32, device=device)
+    seq_lens_cpu = seq_lens.cpu()
+    max_seq_len = int(seq_lens_cpu.max())
+
+    context_lens = [kv - q for kv, q in zip(kv_lens, q_lens)]
+    num_computed_tokens_cpu = torch.tensor(context_lens, dtype=torch.int32)
+
+    max_blocks = (max(kv_lens) + block_size - 1) // block_size
+    num_blocks = batch_size * max_blocks
+    block_table_tensor = torch.arange(
+        num_blocks, dtype=torch.int32, device=device
+    ).view(batch_size, max_blocks)
+    slot_mapping = torch.arange(total_tokens, dtype=torch.int64, device=device)
+
+    max_query_len = max(q_lens)
+
+    return CommonAttentionMetadata(
+        query_start_loc=query_start_loc,
+        query_start_loc_cpu=query_start_loc_cpu,
+        seq_lens=seq_lens,
+        seq_lens_cpu=seq_lens_cpu,
+        num_computed_tokens_cpu=num_computed_tokens_cpu,
+        num_reqs=batch_size,
+        num_actual_tokens=total_tokens,
+        max_query_len=max_query_len,
+        max_seq_len=max_seq_len,
+        block_table_tensor=block_table_tensor,
+        slot_mapping=slot_mapping,
+        causal=True,
+    )
+
+
+def _create_vllm_config(
+    config: BenchmarkConfig,
+    dtype: torch.dtype,
+    max_num_blocks: int,
+) -> VllmConfig:
+    """Create a VllmConfig for benchmarking with mock model methods."""
+    model_config = ModelConfig(
+        model="meta-llama/Meta-Llama-3-8B",
+        tokenizer="meta-llama/Meta-Llama-3-8B",
+        trust_remote_code=False,
+        dtype=dtype,
+        seed=0,
+        max_model_len=1024,
+    )
+
+    cache_config = CacheConfig(
+        block_size=config.block_size,
+        cache_dtype="auto",
+        swap_space=0,
+    )
+    cache_config.num_gpu_blocks = max_num_blocks
+    cache_config.num_cpu_blocks = 0
+
+    parallel_config = ParallelConfig(tensor_parallel_size=1)
+    scheduler_config = SchedulerConfig(
+        max_num_seqs=256,
+        max_num_batched_tokens=8192,
+        max_model_len=8192,
+        is_encoder_decoder=False,
+        enable_chunked_prefill=True,
+    )
+    device_config = DeviceConfig()
+    load_config = LoadConfig()
+    compilation_config = CompilationConfig()
+
+    # Add mock methods for benchmark config values
+    model_config.get_num_layers = types.MethodType(
+        lambda self: config.num_layers, model_config
+    )
+    model_config.get_sliding_window_for_layer = types.MethodType(
+        lambda self, i: None, model_config
+    )
+    model_config.get_logits_soft_cap_for_layer = types.MethodType(
+        lambda self, i: 0.0, model_config
+    )
+    model_config.get_sm_scale_for_layer = types.MethodType(
+        lambda self, i: 1.0 / config.head_dim**0.5, model_config
+    )
+    model_config.get_num_attention_heads = types.MethodType(
+        lambda self, parallel_config=None: config.num_q_heads, model_config
+    )
+    model_config.get_num_kv_heads = types.MethodType(
+        lambda self, parallel_config=None: config.num_kv_heads, model_config
+    )
+    model_config.get_head_size = types.MethodType(
+        lambda self: config.head_dim, model_config
+    )
+    model_config.get_sliding_window = types.MethodType(lambda self: None, model_config)
+
+    return VllmConfig(
+        model_config=model_config,
+        cache_config=cache_config,
+        parallel_config=parallel_config,
+        scheduler_config=scheduler_config,
+        device_config=device_config,
+        load_config=load_config,
+        compilation_config=compilation_config,
+    )
+
+
+# ============================================================================
+# Backend Initialization
+# ============================================================================
+
+
+def _create_backend_impl(
+    backend_cfg: dict,
+    config: BenchmarkConfig,
+    device: torch.device,
+):
+    """Create backend implementation instance."""
+    import importlib
+
+    backend_module = importlib.import_module(backend_cfg["module"])
+    backend_class = getattr(backend_module, backend_cfg["backend_class"])
+
+    scale = get_attention_scale(config.head_dim)
+    dtype = backend_cfg["dtype"]
+
+    impl = backend_class.get_impl_cls()(
+        num_heads=config.num_q_heads,
+        head_size=config.head_dim,
+        scale=scale,
+        num_kv_heads=config.num_kv_heads,
+        alibi_slopes=None,
+        sliding_window=None,
+        kv_cache_dtype="auto",
+    )
+
+    kv_cache_spec = FullAttentionSpec(
+        block_size=config.block_size,
+        num_kv_heads=config.num_kv_heads,
+        head_size=config.head_dim,
+        dtype=dtype,
+    )
+
+    layer = MockLayer(device, kv_cache_spec=kv_cache_spec)
+
+    return backend_class, impl, layer, dtype
+
+
+def _create_metadata_builder(
+    backend_class,
+    kv_cache_spec: FullAttentionSpec,
+    vllm_config: VllmConfig,
+    device: torch.device,
+):
+    """Create metadata builder instance."""
+    return backend_class.get_builder_cls()(
+        kv_cache_spec=kv_cache_spec,
+        layer_names=["layer_0"],
+        vllm_config=vllm_config,
+        device=device,
+    )
+
+
+# ============================================================================
+# Tensor Creation Helpers
+# ============================================================================
+
+
+def _create_input_tensors(
+    config: BenchmarkConfig,
+    total_q: int,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> tuple:
+    """Create Q, K, V input tensors for all layers."""
+    q_list = [
+        torch.randn(
+            total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
+        )
+        for _ in range(config.num_layers)
+    ]
+    k_list = [
+        torch.randn(
+            total_q, config.num_kv_heads, config.head_dim, device=device, dtype=dtype
+        )
+        for _ in range(config.num_layers)
+    ]
+    v_list = [
+        torch.randn(
+            total_q, config.num_kv_heads, config.head_dim, device=device, dtype=dtype
+        )
+        for _ in range(config.num_layers)
+    ]
+    return q_list, k_list, v_list
+
+
+def _create_kv_cache(
+    config: BenchmarkConfig,
+    max_num_blocks: int,
+    cache_layout: str,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> list:
+    """Create KV cache tensors for all layers."""
+    if cache_layout == "flashinfer":
+        # FlashInfer layout: [num_blocks, 2, block_size, num_kv_heads, head_dim]
+        cache_list = [
+            torch.zeros(
+                max_num_blocks,
+                2,
+                config.block_size,
+                config.num_kv_heads,
+                config.head_dim,
+                device=device,
+                dtype=dtype,
+            )
+            for _ in range(config.num_layers)
+        ]
+    else:
+        # Standard layout: [2, num_blocks, block_size, num_kv_heads, head_dim]
+        cache_list = [
+            torch.zeros(
+                2,
+                max_num_blocks,
+                config.block_size,
+                config.num_kv_heads,
+                config.head_dim,
+                device=device,
+                dtype=dtype,
+            )
+            for _ in range(config.num_layers)
+        ]
+    return cache_list
+
+
+# ============================================================================
+# Benchmark Execution
+# ============================================================================
+
+
+def _run_single_benchmark(
+    config: BenchmarkConfig,
+    impl,
+    layer,
+    q_list: list,
+    k_list: list,
+    v_list: list,
+    cache_list: list,
+    attn_metadata,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> tuple:
+    """Run single benchmark iteration with warmup and timing loop."""
+    total_q = q_list[0].shape[0]
+    out = torch.empty(
+        total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
+    )
+
+    # Warmup
+    for _ in range(config.warmup_iters):
+        for i in range(config.num_layers):
+            impl.forward(
+                layer,
+                q_list[i],
+                k_list[i],
+                v_list[i],
+                cache_list[i],
+                attn_metadata,
+                output=out,
+            )
+    torch.cuda.synchronize()
+
+    # Benchmark
+    times = []
+    for _ in range(config.repeats):
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
+
+        start.record()
+        for i in range(config.num_layers):
+            impl.forward(
+                layer,
+                q_list[i],
+                k_list[i],
+                v_list[i],
+                cache_list[i],
+                attn_metadata,
+                output=out,
+            )
+        end.record()
+
+        torch.cuda.synchronize()
+        elapsed_ms = start.elapsed_time(end)
+        times.append(elapsed_ms / 1000.0 / config.num_layers)  # seconds per layer
+
+    mem_stats = {}
+    if config.profile_memory:
+        mem_stats = {
+            "allocated_mb": torch.cuda.memory_allocated(device) / 1024**2,
+            "reserved_mb": torch.cuda.memory_reserved(device) / 1024**2,
+        }
+
+    return times, mem_stats
+
+
+# ============================================================================
+# Public API
+# ============================================================================
+
+
+def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
+    """
+    Run standard attention benchmark with real kernels.
+
+    Supports: flash, triton, flashinfer
+
+    Args:
+        config: Benchmark configuration
+
+    Returns:
+        BenchmarkResult with timing and memory statistics
+    """
+    device = torch.device(config.device)
+    torch.cuda.set_device(device)
+
+    backend_cfg = _get_backend_config(config.backend)
+
+    requests = parse_batch_spec(config.batch_spec)
+
+    if config.backend == "flashinfer":
+        requests = reorder_for_flashinfer(requests)
+
+    q_lens = [r.q_len for r in requests]
+    kv_lens = [r.kv_len for r in requests]
+    total_q = sum(q_lens)
+    max_kv = max(kv_lens)
+
+    max_num_blocks = (max_kv + config.block_size - 1) // config.block_size
+
+    backend_class, impl, layer, dtype = _create_backend_impl(
+        backend_cfg, config, device
+    )
+
+    common_metadata = _build_common_attn_metadata(
+        q_lens, kv_lens, config.block_size, device
+    )
+
+    kv_cache_spec = FullAttentionSpec(
+        block_size=config.block_size,
+        num_kv_heads=config.num_kv_heads,
+        head_size=config.head_dim,
+        dtype=dtype,
+    )
+
+    vllm_config = _create_vllm_config(config, dtype, max_num_blocks)
+
+    builder = _create_metadata_builder(
+        backend_class, kv_cache_spec, vllm_config, device
+    )
+
+    attn_metadata = builder.build(
+        common_prefix_len=0,
+        common_attn_metadata=common_metadata,
+    )
+
+    q_list, k_list, v_list = _create_input_tensors(config, total_q, device, dtype)
+
+    cache_list = _create_kv_cache(
+        config, max_num_blocks, backend_cfg["cache_layout"], device, dtype
+    )
+
+    times, mem_stats = _run_single_benchmark(
+        config,
+        impl,
+        layer,
+        q_list,
+        k_list,
+        v_list,
+        cache_list,
+        attn_metadata,
+        device,
+        dtype,
+    )
+
+    mean_time = np.mean(times)
+    throughput = total_q / mean_time if mean_time > 0 else 0
+
+    return BenchmarkResult(
+        config=config,
+        mean_time=mean_time,
+        std_time=np.std(times),
+        min_time=np.min(times),
+        max_time=np.max(times),
+        throughput_tokens_per_sec=throughput,
+        memory_allocated_mb=mem_stats.get("allocated_mb"),
+        memory_reserved_mb=mem_stats.get("reserved_mb"),
+    )

benchmarks/benchmark_batch_invariance.py

@@ -104,7 +104,6 @@ def run_benchmark_with_batch_invariant(
     random.seed(seed)
 
     # Set environment variables
-    os.environ["VLLM_ATTENTION_BACKEND"] = backend
     if batch_invariant:
         os.environ["VLLM_BATCH_INVARIANT"] = "1"
     else:
@@ -140,6 +139,7 @@ def run_benchmark_with_batch_invariant(
             max_model_len=max_model_len,
             dtype="bfloat16",
             tensor_parallel_size=tp_size,
+            attention_config={"backend": backend},
             enable_prefix_caching=False,
         )
         init_time = time.perf_counter() - start_init

benchmarks/benchmark_ngram_proposer.py

@@ -135,7 +135,6 @@ def benchmark_batched_propose(args):
             block_sizes=[16],
         )
         dummy_input_batch._req_ids = list(str(id) for id in range(args.num_req))
-        dummy_input_batch.spec_decode_unsupported_reqs = ()
         dummy_input_batch.num_tokens_no_spec = [args.num_token] * args.num_req
         dummy_input_batch.token_ids_cpu = np.random.randint(
             0, 20, (args.num_req, args.num_token)
@@ -151,10 +150,8 @@ def benchmark_batched_propose(args):
             start = time.time()
             runner.drafter.propose(
                 sampled_token_ids,
-                dummy_input_batch.req_ids,
                 dummy_input_batch.num_tokens_no_spec,
                 dummy_input_batch.token_ids_cpu,
-                dummy_input_batch.spec_decode_unsupported_reqs,
             )
             end = time.time()
             print(f"Iteration time (s): {end - start}")

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -343,7 +343,9 @@ def bench(
         return bench_int8(dtype, m, k, n, label, sub_label)
     if dtype == torch.float8_e4m3fn:
         return bench_fp8(dtype, m, k, n, label, sub_label)
-    raise ValueError("unsupported type")
+    raise ValueError(
+        f"Unsupported dtype {dtype}: should be one of torch.int8, torch.float8_e4m3fn."
+    )
 
 
 # runner

benchmarks/kernels/bench_nvfp4_quant.py

@@ -0,0 +1,210 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+from vllm.scalar_type import scalar_types
+from vllm.triton_utils import triton
+from vllm.utils.flashinfer import flashinfer_fp4_quantize
+
+if not current_platform.has_device_capability(100):
+    raise RuntimeError("NVFP4 requires compute capability of 10.0 (Blackwell)")
+
+FLOAT4_E2M1_MAX = scalar_types.float4_e2m1f.max()
+FLOAT8_E4M3_MAX = torch.finfo(torch.float8_e4m3fn).max
+
+PROVIDER_CFGS = {
+    "vllm": dict(backend="vllm", is_sf_swizzled_layout=False, enabled=True),
+    "vllm-swizzle": dict(backend="vllm", is_sf_swizzled_layout=True, enabled=True),
+    "flashinfer": dict(backend="flashinfer", is_sf_swizzled_layout=False, enabled=True),
+    "flashinfer-swizzle": dict(
+        backend="flashinfer", is_sf_swizzled_layout=True, enabled=True
+    ),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def compute_global_scale(tensor: torch.Tensor) -> torch.Tensor:
+    """Compute global scale for FP4 quantization."""
+    amax = torch.abs(tensor).max().to(torch.float32)
+    return FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / amax
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192],
+        x_log=False,
+        line_arg="provider",
+        line_vals=_enabled,
+        line_names=_enabled,
+        ylabel="us (lower is better)",
+        plot_name="NVFP4 Input Quantization Latency (us)",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    # Create input tensor
+    a = torch.randn((M, K), device=device, dtype=dtype)
+
+    # Compute global scale for activation
+    a_global_scale = compute_global_scale(a)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    cfg = PROVIDER_CFGS[provider]
+
+    if cfg["backend"] == "vllm":
+        # vLLM's FP4 quantization
+        if cfg["is_sf_swizzled_layout"]:
+            ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+                lambda: ops.scaled_fp4_quant(
+                    a, a_global_scale, is_sf_swizzled_layout=True
+                ),
+                quantiles=quantiles,
+            )
+        else:
+            ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+                lambda: ops.scaled_fp4_quant(
+                    a, a_global_scale, is_sf_swizzled_layout=False
+                ),
+                quantiles=quantiles,
+            )
+    elif cfg["backend"] == "flashinfer":
+        # FlashInfer's FP4 quantization
+        if cfg["is_sf_swizzled_layout"]:
+            ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+                lambda: flashinfer_fp4_quantize(
+                    a, a_global_scale, is_sf_swizzled_layout=True
+                ),
+                quantiles=quantiles,
+            )
+        else:
+            ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+                lambda: flashinfer_fp4_quantize(
+                    a, a_global_scale, is_sf_swizzled_layout=False
+                ),
+                quantiles=quantiles,
+            )
+
+    # Convert ms to us for better readability at small batch sizes
+    to_us = lambda t_ms: t_ms * 1000
+    return to_us(ms), to_us(max_ms), to_us(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    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)
+            out.append(KN)
+    return out
+
+
+def _test_accuracy_once(
+    M: int, K: int, dtype: torch.dtype, device: str, is_sf_swizzled_layout: bool
+):
+    """Test accuracy between vLLM and FlashInfer FP4 quantization."""
+    # Create input tensor
+    a = torch.randn((M, K), device=device, dtype=dtype)
+
+    # Compute global scale
+    a_global_scale = compute_global_scale(a)
+
+    # vLLM quantization
+    vllm_fp4, vllm_scale = ops.scaled_fp4_quant(
+        a, a_global_scale, is_sf_swizzled_layout=is_sf_swizzled_layout
+    )
+
+    # FlashInfer quantization (with swizzled layout to match vLLM's output)
+    flashinfer_fp4, flashinfer_scale = flashinfer_fp4_quantize(
+        a, a_global_scale, is_sf_swizzled_layout=is_sf_swizzled_layout
+    )
+    flashinfer_scale = flashinfer_scale.view(torch.float8_e4m3fn)
+
+    # Compare outputs
+    torch.testing.assert_close(
+        vllm_fp4,
+        flashinfer_fp4,
+    )
+    # Compare scales
+    torch.testing.assert_close(
+        vllm_scale,
+        flashinfer_scale,
+    )
+    print(
+        f"M={M}, K={K}, dtype={dtype}, is_sf_swizzled_layout={is_sf_swizzled_layout}: PASSED"  # noqa: E501
+    )
+
+
+def test_accuracy():
+    """Run accuracy tests across various shapes."""
+    print("\n" + "=" * 60)
+    print("Running accuracy tests: vLLM vs FlashInfer")
+    print("=" * 60)
+
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    # Test various batch sizes and hidden dimensions
+    Ms = [1, 1024]
+    Ks = [4096]
+
+    for is_sf_swizzled_layout in [True, False]:
+        for M in Ms:
+            for K in Ks:
+                _test_accuracy_once(M, K, dtype, device, is_sf_swizzled_layout)
+
+    print("\nAll accuracy tests passed!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Benchmark NVFP4 quantization: vLLM vs FlashInfer"
+    )
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.3-70B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    parser.add_argument(
+        "--save-path",
+        type=str,
+        default=None,
+        help="Path to save benchmark results",
+    )
+    parser.add_argument(
+        "--accuracy",
+        action="store_true",
+        help="Run accuracy tests",
+    )
+    args = parser.parse_args()
+
+    if args.accuracy:
+        test_accuracy()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"\n{model}, N={N} K={K}")
+        benchmark.run(
+            print_data=True,
+            save_path=args.save_path,
+            N=N,
+            K=K,
+        )
+
+    print("\nBenchmark finished!")

benchmarks/kernels/benchmark_activation.py

@@ -7,14 +7,13 @@ import itertools
 import torch
 
 import vllm.model_executor.layers.activation  # noqa F401
-from vllm.model_executor.custom_op import CustomOp
-from vllm.platforms import current_platform
+from vllm.model_executor.custom_op import op_registry
 from vllm.triton_utils import triton
 from vllm.utils.argparse_utils import FlexibleArgumentParser
-from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
+from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE, set_random_seed
 
-batch_size_range = [1, 16, 32, 64, 128]
-seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
+batch_size_range = [1, 16, 128]
+seq_len_range = [1, 16, 64, 1024, 4096]
 intermediate_size = [3072, 9728, 12288]
 configs = list(itertools.product(batch_size_range, seq_len_range, intermediate_size))
 
@@ -30,18 +29,18 @@ def benchmark_activation(
     device = "cuda"
     num_tokens = batch_size * seq_len
     dim = intermediate_size
-    current_platform.seed_everything(42)
+    set_random_seed(42)
     torch.set_default_device(device)
 
     if func_name == "gelu_and_mul":
-        layer = CustomOp.op_registry[func_name](approximate="none")
+        layer = op_registry[func_name](approximate="none")
     elif func_name == "gelu_and_mul_tanh":
-        layer = CustomOp.op_registry["gelu_and_mul"](approximate="tanh")
+        layer = op_registry["gelu_and_mul"](approximate="tanh")
     elif func_name == "fatrelu_and_mul":
         threshold = 0.5
-        layer = CustomOp.op_registry[func_name](threshold)
+        layer = op_registry[func_name](threshold)
     else:
-        layer = CustomOp.op_registry[func_name]()
+        layer = op_registry[func_name]()
 
     x = torch.randn(num_tokens, dim, dtype=dtype, device=device)
     compiled_layer = torch.compile(layer.forward_native)

benchmarks/kernels/benchmark_bitblas.py

@@ -1,244 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-# Copyright (c) Microsoft Corporation.
-# Licensed under the MIT License.
-
-from packaging import version
-
-from vllm.model_executor.layers.quantization.utils.bitblas_utils import (
-    MINIMUM_BITBLAS_VERSION,
-)
-
-try:
-    import bitblas
-
-    if version.parse(bitblas.__version__) < version.parse(MINIMUM_BITBLAS_VERSION):
-        raise ImportError(
-            "bitblas version is wrong. Please "
-            f"install bitblas>={MINIMUM_BITBLAS_VERSION}"
-        )
-except ImportError as e:
-    bitblas_import_exception = e
-    raise ValueError(
-        "Trying to use the bitblas backend, but could not import"
-        f"with the following error: {bitblas_import_exception}. "
-        "Please install bitblas through the following command: "
-        f"`pip install bitblas>={MINIMUM_BITBLAS_VERSION}`"
-    ) from bitblas_import_exception
-
-from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target
-
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-
-parser = FlexibleArgumentParser(
-    description="Benchmark BitBLAS int4 on a specific target."
-)
-
-# Add arguments to the parser
-parser.add_argument(
-    "--target",
-    type=str,
-    default=auto_detect_nvidia_target(),
-    help="Specify the target device for benchmarking.",
-)
-parser.add_argument(
-    "--group_size", type=int, default=None, help="Group size for grouped quantization."
-)
-parser.add_argument(
-    "--A_dtype",
-    type=str,
-    default="float16",
-    choices=["float16", "float32", "float64", "int32", "int8"],
-    help="Data type of activation A.",
-)
-parser.add_argument(
-    "--W_dtype",
-    type=str,
-    default="int4",
-    choices=[
-        "float16",
-        "float32",
-        "float64",
-        "int32",
-        "int8",
-        "int4",
-        "int2",
-        "int1",
-        "nf4",
-        "fp4_e2m1",
-    ],
-    help="Data type of weight W.",
-)
-parser.add_argument(
-    "--accum_dtype",
-    type=str,
-    default="float16",
-    choices=["float16", "int32"],
-    help="Data type for accumulation.",
-)
-parser.add_argument(
-    "--out_dtype",
-    type=str,
-    default="float16",
-    choices=["float16", "float32", "int32", "int8"],
-    help="Data type for output.",
-)
-parser.add_argument(
-    "--layout",
-    type=str,
-    default="nt",
-    choices=["nt", "nn"],
-    help="Matrix layout, 'nt' for non-transpose A and transpose W.",
-)
-parser.add_argument(
-    "--with_bias", action="store_true", help="Include bias in the benchmark."
-)
-parser.add_argument(
-    "--with_scaling",
-    action="store_true",
-    help="Include scaling factor in the quantization.",
-)
-parser.add_argument(
-    "--with_zeros", action="store_true", help="Include zeros in the quantization."
-)
-parser.add_argument(
-    "--zeros_mode",
-    type=str,
-    default=None,
-    choices=["original", "rescale", "quantized"],
-    help="Specify the mode for calculating zeros.",
-)
-
-# Parse the arguments
-args = parser.parse_args()
-
-# Assign arguments to variables
-target = args.target
-A_dtype = args.A_dtype
-W_dtype = args.W_dtype
-accum_dtype = args.accum_dtype
-out_dtype = args.out_dtype
-layout = args.layout
-with_bias = args.with_bias
-group_size = args.group_size
-with_scaling = args.with_scaling
-with_zeros = args.with_zeros
-zeros_mode = args.zeros_mode
-
-# Define a list of shared arguments that repeat in every config
-shared_args = [
-    A_dtype,
-    W_dtype,
-    out_dtype,
-    accum_dtype,
-    layout,
-    with_bias,
-    group_size,
-    with_scaling,
-    with_zeros,
-    zeros_mode,
-]
-
-# Define just the (M, K, N) shapes in a more compact list
-shapes = [
-    # square test
-    (1, 16384, 16384),
-    # BLOOM-176B
-    (1, 43008, 14336),
-    (1, 14336, 14336),
-    (1, 57344, 14336),
-    (1, 14336, 57344),
-    # OPT-65B
-    (1, 9216, 9216),
-    (1, 36864, 9216),
-    (1, 9216, 36864),
-    (1, 22016, 8192),
-    # LLAMA-70B/65B
-    (1, 8192, 22016),
-    (1, 8192, 8192),
-    (1, 28672, 8192),
-    (1, 8192, 28672),
-    # square test
-    (16384, 16384, 16384),
-    # BLOOM-176B
-    (8192, 43008, 14336),
-    (8192, 14336, 14336),
-    (8192, 57344, 14336),
-    (8192, 14336, 57344),
-    # OPT-65B
-    (8192, 9216, 9216),
-    (8192, 36864, 9216),
-    (8192, 9216, 36864),
-    (8192, 22016, 8192),
-    # LLAMA-70B/65B
-    (8192, 8192, 22016),
-    (8192, 8192, 8192),
-    (8192, 28672, 8192),
-    (8192, 8192, 28672),
-]
-
-# Build test shapes with all the shared arguments
-test_shapes = [(MatmulConfig, Matmul, (*shape, *shared_args)) for shape in shapes]
-
-benchmark_sets = []
-benchmark_sets.extend(test_shapes)
-
-benchmark_results = {}
-for config_class, operator, input_args in benchmark_sets:
-    config = config_class(*input_args)
-    matmul = operator(config, target=target, enable_tuning=True)
-    kernel_latency = matmul.profile_latency()
-
-    print("Time cost is: {:.3f} ms".format(kernel_latency))
-
-    profile_config = {
-        f"{operator.__name__}-{'-'.join([str(i) for i in input_args])}": {
-            "BitBLAS_top20_latency": kernel_latency,
-        }
-    }
-
-    benchmark_results.update(profile_config)
-
-# Define headers for the table
-headers = [
-    "PrimFunc",
-    "Input Arguments",
-    "BitBLAS Top20 Latency",
-]
-
-# Calculate column widths for pretty printing
-col_widths = [0, 0, 0]
-for config_key, values in benchmark_results.items():
-    args_split = config_key.split("-")
-    func_name = args_split[0]
-    input_args_str = "-".join(args_split[1:])
-    col_widths[0] = max(col_widths[0], len(func_name) + 2, len(headers[0]) + 2)
-    col_widths[1] = max(col_widths[1], len(input_args_str) + 2, len(headers[1]) + 2)
-    col_widths[2] = max(
-        col_widths[2],
-        len(f"{values['BitBLAS_top20_latency']:.3f} ms") + 2,
-        len(headers[2]) + 2,
-    )
-    # break only if you want to measure widths from a single example;
-    # otherwise, let it loop over all items.
-
-# Print header
-for i, header in enumerate(headers):
-    headers[i] = header.ljust(col_widths[i])
-print("".join(headers))
-print("-" * sum(col_widths))
-
-# Print rows
-for config_key, values in benchmark_results.items():
-    args_split = config_key.split("-")
-    func_name = args_split[0]
-    input_args_str = "-".join(args_split[1:])
-    row = [
-        func_name,
-        input_args_str,
-        f"{values['BitBLAS_top20_latency']:.3f} ms",
-    ]
-    row_str = "".join(
-        [str(cell).ljust(col_widths[idx]) for idx, cell in enumerate(row)]
-    )
-    print(row_str)

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -6,15 +6,20 @@ kernel. Both kernels take in fp8 quantized weights and 16-bit activations,
 but use different quantization strategies and backends.
 """
 
-import nvtx
 import torch
 
+import vllm.model_executor.layers.fused_moe.modular_kernel as mk
+from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
+from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
+from vllm.model_executor.layers.fused_moe.prepare_finalize import (
+    MoEPrepareAndFinalizeNoEP,
+)
 from vllm.platforms import current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.v1.worker.workspace import init_workspace_manager
 
 # Weight shapes for different models: [num_experts, topk, hidden_size,
 # intermediate_size]
@@ -58,6 +63,7 @@ def bench_run(
     per_out_ch: bool,
     mkn: tuple[int, int, int],
 ):
+    init_workspace_manager(torch.cuda.current_device())
     (m, k, n) = mkn
 
     dtype = torch.half
@@ -120,85 +126,6 @@ def bench_run(
     # Force per-tensor quantization for all cases
     per_act_token = False
 
-    # Create stride tensors for CUTLASS
-    ab_strides1 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-    ab_strides2 = torch.full((num_experts,), n, dtype=torch.int64, device=device)
-    c_strides1 = torch.full((num_experts,), 2 * n, dtype=torch.int64, device=device)
-    c_strides2 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-
-    def run_triton_moe(
-        a: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        w1_scale: torch.Tensor,
-        w2_scale: torch.Tensor,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            fused_experts(
-                a,
-                w1,
-                w2,
-                topk_weights,
-                topk_ids,
-                quant_config=quant_config,
-            )
-
-    def run_cutlass_moe_fp8(
-        a: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides1: torch.Tensor,
-        c_strides2: torch.Tensor,
-        w1_scale: torch.Tensor,
-        w2_scale: torch.Tensor,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            with nvtx.annotate("cutlass_moe_fp8", color="blue"):
-                cutlass_moe_fp8(
-                    a=a,
-                    w1_q=w1,
-                    w2_q=w2,
-                    topk_weights=topk_weights,
-                    topk_ids=topk_ids,
-                    ab_strides1=ab_strides1,
-                    ab_strides2=ab_strides2,
-                    c_strides1=c_strides1,
-                    c_strides2=c_strides2,
-                    quant_config=quant_config,
-                    activation="silu",
-                    global_num_experts=num_experts,
-                )
-
     # Pre-create quantization config to avoid creating it inside CUDA graph
     quant_config = fp8_w8a8_moe_quant_config(
         w1_scale=w1_scale,
@@ -209,23 +136,31 @@ def bench_run(
         per_out_ch_quant=per_out_ch,
     )
 
+    fn = mk.FusedMoEModularKernel(
+        MoEPrepareAndFinalizeNoEP(),
+        CutlassExpertsFp8(
+            moe_config=make_dummy_moe_config(
+                num_experts=num_experts,
+                hidden_dim=k,
+                intermediate_size_per_partition=n,
+                in_dtype=a.dtype,
+            ),
+            quant_config=quant_config,
+        ),
+    )
+
     # Create CUDA graphs for CUTLASS (match benchmark_moe.py pattern exactly)
     cutlass_stream = torch.cuda.Stream()
     cutlass_graph = torch.cuda.CUDAGraph()
     with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
         # Capture 10 invocations like benchmark_moe.py
         for _ in range(10):
-            cutlass_moe_fp8(
-                a=a,
-                w1_q=w1_fp8q_cutlass,
-                w2_q=w2_fp8q_cutlass,
-                topk_weights=topk_weights,
-                topk_ids=topk_ids,
-                ab_strides1=ab_strides1,
-                ab_strides2=ab_strides2,
-                c_strides1=c_strides1,
-                c_strides2=c_strides2,
-                quant_config=quant_config,
+            fn(
+                a,
+                w1_fp8q_cutlass,
+                w2_fp8q_cutlass,
+                topk_weights,
+                topk_ids,
                 activation="silu",
                 global_num_experts=num_experts,
             )
@@ -297,6 +232,10 @@ def bench_run(
 
 
 def main(args):
+    # Initialize workspace manager (required for CUTLASS MoE kernels)
+    device = torch.device("cuda:0")
+    init_workspace_manager(device)
+
     print("Benchmarking models:")
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")

benchmarks/kernels/benchmark_cutlass_moe_nvfp4.py

@@ -11,16 +11,24 @@ import nvtx
 import torch
 import torch.utils.benchmark as benchmark
 
+import vllm.model_executor.layers.fused_moe.modular_kernel as mk
+from tests.kernels.moe.utils import make_dummy_moe_config
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
 from vllm.model_executor.layers.fused_moe.config import (
     fp8_w8a8_moe_quant_config,
     nvfp4_moe_quant_config,
 )
-from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4
+from vllm.model_executor.layers.fused_moe.cutlass_moe import (
+    CutlassExpertsFp4,
+)
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
+from vllm.model_executor.layers.fused_moe.prepare_finalize import (
+    MoEPrepareAndFinalizeNoEP,
+)
 from vllm.scalar_type import scalar_types
 from vllm.utils.argparse_utils import FlexibleArgumentParser
+from vllm.v1.worker.workspace import init_workspace_manager
 
 WEIGHT_SHAPES_MOE = {
     "nvidia/DeepSeek-R1-FP4": [
@@ -187,19 +195,23 @@ def bench_run(
             g1_alphas=w1_gs,
             g2_alphas=w2_gs,
         )
+
+        kernel = mk.FusedMoEModularKernel(
+            MoEPrepareAndFinalizeNoEP(),
+            CutlassExpertsFp4(
+                make_dummy_moe_config(),
+                quant_config=quant_config,
+            ),
+        )
+
         for _ in range(num_repeats):
             with nvtx.annotate("cutlass_moe_fp4", color="green"):
-                cutlass_moe_fp4(
-                    a=a,
-                    w1_fp4=w1_fp4,
-                    w2_fp4=w2_fp4,
+                kernel(
+                    hidden_states=a,
+                    w1=w1_fp4,
+                    w2=w2_fp4,
                     topk_weights=topk_weights,
                     topk_ids=topk_ids,
-                    m=m,
-                    n=n,
-                    k=k,
-                    e=num_experts,
-                    quant_config=quant_config,
                 )
 
     def run_cutlass_from_graph(
@@ -229,20 +241,23 @@ def bench_run(
             g2_alphas=w2_gs,
         )
 
+        kernel = mk.FusedMoEModularKernel(
+            MoEPrepareAndFinalizeNoEP(),
+            CutlassExpertsFp4(
+                make_dummy_moe_config(),
+                quant_config=quant_config,
+            ),
+        )
+
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
-            return cutlass_moe_fp4(
-                a=a,
-                w1_fp4=w1_fp4,
-                w2_fp4=w2_fp4,
+            return kernel(
+                hidden_states=a,
+                w1=w1_fp4,
+                w2=w2_fp4,
                 topk_weights=topk_weights,
                 topk_ids=topk_ids,
-                m=m,
-                n=n,
-                k=k,
-                e=num_experts,
-                quant_config=quant_config,
             )
 
     def run_triton_from_graph(
@@ -441,6 +456,10 @@ def bench_run(
 
 
 def main(args):
+    # Initialize workspace manager (required for CUTLASS MoE kernels)
+    device = torch.device("cuda:0")
+    init_workspace_manager(device)
+
     print("Benchmarking models:")
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")

benchmarks/kernels/benchmark_device_communicators.py

@@ -293,7 +293,7 @@ class CommunicatorBenchmark:
                     graph = torch.cuda.CUDAGraph()
                     graph_pool = torch.cuda.graph_pool_handle()
                     set_graph_pool_id(graph_pool)
-                    with torch.cuda.graph(graph, pool=graph_pool):
+                    with torch.cuda.graph(graph, pool=graph_pool, stream=stream):
                         for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
                             allreduce_fn(graph_input)
 

benchmarks/kernels/benchmark_fused_topk.py

@@ -0,0 +1,99 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import itertools
+
+import torch
+
+from vllm.model_executor.layers.fused_moe.router.fused_topk_router import fused_topk
+from vllm.triton_utils import triton
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+num_tokens_range = [2**i for i in range(0, 8, 2)]
+num_experts_range = [16, 32, 64, 128, 256, 512]
+topk_range = [3, 4]
+configs = list(itertools.product(num_tokens_range, num_experts_range, topk_range))
+
+
+def torch_topk(
+    gating_output: torch.Tensor,
+    topk: int,
+    renormalize: bool,
+    scoring_func: str = "softmax",
+):
+    if scoring_func == "softmax":
+        scores = torch.softmax(gating_output.float(), dim=-1)
+    else:
+        scores = torch.sigmoid(gating_output.float())
+    topk_weights, topk_ids = torch.topk(scores, k=topk, dim=-1)
+
+    if renormalize:
+        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+
+    return topk_weights, topk_ids
+
+
+def get_benchmark(scoring_func):
+    @triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["num_tokens", "num_experts", "topk"],
+            x_vals=[list(_) for _ in configs],
+            line_arg="provider",
+            line_vals=["torch", "vllm"],
+            line_names=["Torch", "vLLM"],
+            styles=[("blue", "-"), ("red", "-")],
+            ylabel="us",
+            plot_name=f"fused-topk-perf-{scoring_func}",
+            args={},
+        )
+    )
+    def benchmark(num_tokens, num_experts, topk, provider):
+        dtype = torch.bfloat16
+        hidden_size = 1024
+        renormalize = True
+        hidden_states = torch.randn(
+            (num_tokens, hidden_size), dtype=dtype, device="cuda"
+        )
+        gating_output = torch.randn(
+            (num_tokens, num_experts), dtype=dtype, device="cuda"
+        )
+
+        quantiles = [0.5, 0.2, 0.8]
+
+        if provider == "torch":
+            ms, min_ms, max_ms = triton.testing.do_bench(
+                lambda: torch_topk(
+                    gating_output=gating_output,
+                    topk=topk,
+                    renormalize=renormalize,
+                    scoring_func=scoring_func,
+                ),
+                quantiles=quantiles,
+            )
+        else:
+            ms, min_ms, max_ms = triton.testing.do_bench(
+                lambda: fused_topk(
+                    hidden_states=hidden_states,
+                    gating_output=gating_output,
+                    topk=topk,
+                    renormalize=renormalize,
+                    scoring_func=scoring_func,
+                ),
+                quantiles=quantiles,
+            )
+
+        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+    return benchmark
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(description="Benchmark the MoE topk kernel.")
+    parser.add_argument("--scoring-func", type=str, default="softmax")
+    parser.add_argument("--save-path", type=str, default="./configs/fused_topk/")
+    args = parser.parse_args()
+
+    # Get the benchmark function
+    benchmark = get_benchmark(args.scoring_func)
+    # Run performance benchmark
+    benchmark.run(print_data=True, save_path=args.save_path)