[CI][ROCm] Fix test_ocp_mx_wikitext_correctness reference value (#49690 )

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
[CI] Initialize DeepEP FP8 test weights (#49912 )
2026-07-27 21:47:07 +00:00 · 2026-07-28 05:29:08 +08:00 · 2026-07-27 16:01:34 -05:00 · 2026-07-27 15:58:26 -04:00 · 2026-07-27 15:56:52 -04:00 · 2026-07-27 19:48:35 +00:00
707 changed files with 38930 additions and 9304 deletions
@@ -0,0 +1,26 @@
+group: Benchmarks
+depends_on: 
+  - image-build-xpu
+steps:
+- label: Benchmarks CLI Test
+  key: benchmarks-cli-test
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/benchmarks/
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s benchmarks/'
@@ -2,6 +2,44 @@ group: Engine Intel
 depends_on:
  - image-build-xpu
 steps:
+- label: Engine
+  key: engine
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/compilation/
+  - vllm/config/
+  - vllm/engine/
+  - vllm/entrypoints/logger.py
+  - vllm/envs.py
+  - vllm/logger.py
+  - vllm/logging_utils/
+  - vllm/platforms/
+  - vllm/sequence.py
+  - vllm/triton_utils/
+  - vllm/utils/
+  - tests/engine
+  - tests/test_sequence
+  - tests/test_config
+  - tests/test_logger
+  - tests/test_vllm_port
+  - tests/test_jit_monitor.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s engine/test_arg_utils.py test_sequence.py test_logger.py test_vllm_port.py test_jit_monitor.py'
+
 - label: Engine (1 GPU)
  timeout_in_minutes: 30
  device: intel_gpu
@@ -23,3 +61,41 @@ steps:
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
      pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py'
+
+- label: V1 e2e (2 GPUs)
+  timeout_in_minutes: 30
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/compilation/
+    - vllm/config/
+    - vllm/distributed/
+    - vllm/engine/
+    - vllm/envs.py
+    - vllm/forward_context.py
+    - vllm/inputs/
+    - vllm/logger.py
+    - vllm/logging_utils/
+    - vllm/model_executor/
+    - vllm/multimodal/
+    - vllm/platforms/
+    - vllm/sampling_params.py
+    - vllm/transformers_utils/
+    - vllm/triton_utils/
+    - vllm/utils/
+    - vllm/v1/
+    - tests/v1/e2e/spec_decode
+  commands:
+  - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"'
@@ -125,13 +125,13 @@ steps:
      pytest -v -s v1/kv_offload &&
      pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'

- label: NixlConnector PD accuracy (2 GPUs)
+- label: NixlConnector PD accuracy (4 GPUs)
  timeout_in_minutes: 60
-  num_devices: 2
+  num_devices: 4
  device: intel_gpu
  agent_tags:
    label: production
-    gpu: 2+
+    gpu: 4+
    mem: 16+
  no_plugin: true
  working_dir: "."
@@ -148,7 +148,10 @@ steps:
    - >-
      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
      'cd tests &&
-      bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'
+      bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=1 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'

 - label: Regression
  key: regression
@@ -259,3 +262,25 @@ steps:
    pytest -v -s detokenizer &&
    pytest -v -s -m "not cpu_test" ./multimodal &&
    pytest -v -s utils_ --ignore=utils_/test_mem_utils.py'
+
+- label: Fusion Unit Tests
+  timeout_in_minutes: 30
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/compilation/
+    - tests/compile/passes/test_qk_norm_rope_fusion.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s compile/passes/test_qk_norm_rope_fusion.py'
@@ -0,0 +1,33 @@
+group: Model Executor Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Model Executor (Intel)
+  key: model-executor-intel
+  timeout_in_minutes: 45
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/engine/arg_utils.py
+  - vllm/config/model.py
+  - vllm/model_executor
+  - tests/model_executor
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'apt-get update && apt-get install -y curl libsodium23 &&
+      pip3 install tensorizer==2.10.1 &&
+      pip3 install runai-model-streamer[s3,gcs,azure]\>=0.15.7 &&
+      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      export PYTHONFAULTHANDLER=1 &&
+      cd tests &&
+      pytest -v -s model_executor -m "not slow_test" --ignore="model_executor/layers/test_rocm_unquantized_gemm.py" --deselect="tests/model_executor/model_loader/test_reload.py::test_kv_scale_reload"'
@@ -8,7 +8,7 @@ steps:
  agent_tags:
    label: production
    gpu: 2+
-    mem: 16+
+    mem: 24+
  no_plugin: true
  working_dir: "."
  env:
@@ -28,7 +28,9 @@ steps:
      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
      cd tests &&
      pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics" &&
+      pytest -v -s v1/e2e/general/test_context_length.py &&
      ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram" &&
+      pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0" &&
      pytest -v -s v1/e2e/general/test_min_tokens.py'

 - label: Model Runner V2 Examples (Intel)
@@ -60,3 +62,55 @@ steps:
      python3 basic/offline_inference/generate.py --model facebook/opt-125m &&
      python3 generate/multimodal/vision_language_offline.py --seed 0 &&
      python3 features/automatic_prefix_caching/prefix_caching_offline.py'
+
+- label: Model Runner V2 Distributed (2 GPUs)
+  timeout_in_minutes: 50
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/v1/worker/gpu/
+    - vllm/v1/worker/gpu_worker.py
+    - tests/basic_correctness/test_basic_correctness.py
+    - tests/v1/distributed/test_async_llm_dp.py
+    - tests/v1/distributed/test_eagle_dp.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd tests &&
+      TARGET_TEST_SUITE=L4 pytest -v -s basic_correctness/test_basic_correctness.py -m "distributed\(num_gpus=2\)" -k "not ray and not True"'
+
+- label: Model Runner V2 Spec Decode
+  timeout_in_minutes: 50
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/v1/worker/gpu/
+  - vllm/v1/worker/gpu_worker.py
+  - tests/v1/spec_decode/test_max_len.py
+  - tests/v1/spec_decode/test_rejection_sampler_utils.py
+  - tests/v1/e2e/spec_decode/test_spec_decode.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd tests &&
+      pytest -v -s v1/spec_decode/test_synthetic_rejection_sampler_utils.py'
@@ -0,0 +1,29 @@
+group: Samplers Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Samplers Test (FlashInfer)
+  key: samplers-test-flashinfer-intel
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - vllm/sampling_metadata.py
+  - tests/samplers
+  - tests/conftest.py
+  - vllm/entrypoints/generate/beam_search
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers'
@@ -7,6 +7,9 @@

 set -euo pipefail

+# The macmini queue uses persistent checkouts, so refresh tags for setuptools-scm.
+git fetch --tags --force origin
+
 # The Rust frontend build needs protoc.
 if ! command -v protoc >/dev/null 2>&1; then
  brew install protobuf
@@ -17,7 +17,7 @@ DEFAULT_REPO_SLUG="vllm-project/vllm"
 DEFAULT_CI_HCL_SOURCE="docker/ci-rocm.hcl"
 DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tools/install_protoc.sh rust-toolchain.toml tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh .buildkite/scripts/rocm/build-ci-base.sh"
 DEFAULT_CI_BASE_DOCKERFILE="docker/Dockerfile.rocm"
-DEFAULT_CI_BASE_DOCKERFILE_STAGES="base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust-toolchain build_rixl build_rocshmem build_deepep mori_base ci_base"
+DEFAULT_CI_BASE_DOCKERFILE_STAGES="base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust-toolchain build_nixl build_rocshmem build_deepep mori_base ci_base"
 DEFAULT_CI_BASE_METADATA_VERSION="1"
 IMAGE_EXISTED_BEFORE_BUILD=0

@@ -1159,8 +1159,8 @@ ci_base_metadata_pairs() {
    metadata_pair "vllm.rocm.nic_backend" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIC_BACKEND")"
    metadata_pair "vllm.rocm.ainic_version" "$(resolve_dockerfile_arg_value "${dockerfile}" "AINIC_VERSION")"
    metadata_pair "vllm.rocm.ubuntu_codename" "$(resolve_dockerfile_arg_value "${dockerfile}" "UBUNTU_CODENAME")"
-    metadata_pair "vllm.rocm.rixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_REPO")"
-    metadata_pair "vllm.rocm.rixl_commit" "${RIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_BRANCH")}"
+    metadata_pair "vllm.rocm.nixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_REPO")"
+    metadata_pair "vllm.rocm.nixl_commit" "${NIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_BRANCH")}"
    metadata_pair "vllm.rocm.ucx_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_REPO")"
    metadata_pair "vllm.rocm.ucx_commit" "${UCX_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_BRANCH")}"
    metadata_pair "vllm.rocm.rocshmem_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "ROCSHMEM_REPO")"
@@ -1169,7 +1169,7 @@ ci_base_metadata_pairs() {
    metadata_pair "vllm.rocm.deepep_commit" "${DEEPEP_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_BRANCH")}"
    metadata_pair "vllm.rocm.deepep_nic" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_NIC")"
    metadata_pair "vllm.rocm.deepep_rocm_arch" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_ROCM_ARCH")"
-    metadata_pair "vllm.rocm.rixl_cache_key" "${RIXL_CACHE_KEY:-}"
+    metadata_pair "vllm.rocm.nixl_cache_key" "${NIXL_CACHE_KEY:-}"
    metadata_pair "vllm.rocm.rocshmem_cache_key" "${ROCSHMEM_CACHE_KEY:-}"
    metadata_pair "vllm.rocm.deepep_cache_key" "${DEEPEP_CACHE_KEY:-}"

@@ -1686,7 +1686,7 @@ extract_dependency_pins() {
        return 0
    fi

-    for var in RIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
+    for var in NIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
        if [[ -n "${!var:-}" ]]; then
            echo "Using provided ${var}: ${!var}"
            continue
@@ -1706,30 +1706,30 @@ extract_dependency_pins() {
 compute_dependency_cache_keys() {
    local bake_dir=""
    local dockerfile_rocm=""
-    local rixl_branch=""
+    local nixl_branch=""
    local ucx_branch=""
    local rocshmem_branch=""
    local deepep_branch=""
-    local rixl_material=""
+    local nixl_material=""
    local rocshmem_material=""
    local deepep_material=""

    bake_dir=$(dirname "${VLLM_BAKE_FILE}")
    dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
-    rixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "RIXL_BRANCH")
+    nixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "NIXL_BRANCH")
    ucx_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "UCX_BRANCH")
    rocshmem_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "ROCSHMEM_BRANCH")
    deepep_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "DEEPEP_BRANCH")

-    if [[ -n "${rixl_branch}" && -n "${ucx_branch}" ]]; then
-        rixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_rixl")
-        RIXL_CACHE_KEY=$(
+    if [[ -n "${nixl_branch}" && -n "${ucx_branch}" ]]; then
+        nixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_nixl")
+        NIXL_CACHE_KEY=$(
            compose_dependency_cache_key \
-                "${rixl_branch}-ucx-${ucx_branch}" \
-                "${rixl_material}"
+                "${nixl_branch}-ucx-${ucx_branch}" \
+                "${nixl_material}"
        )
-        export RIXL_CACHE_KEY
-        echo "RIXL dependency cache key: ${RIXL_CACHE_KEY}"
+        export NIXL_CACHE_KEY
+        echo "NIXL dependency cache key: ${NIXL_CACHE_KEY}"
    fi

    if [[ -n "${rocshmem_branch}" ]]; then
@@ -1780,11 +1780,11 @@ dependency_cache_ref_for_target() {
    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"

    case "${target}" in
-        rixl-rocm-ci)
-            if [[ -n "${RIXL_CACHE_KEY:-}" ]]; then
-                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_CACHE_KEY}"
-            elif [[ -n "${RIXL_BRANCH:-}" ]]; then
-                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
+        nixl-rocm-ci)
+            if [[ -n "${NIXL_CACHE_KEY:-}" ]]; then
+                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_CACHE_KEY}"
+            elif [[ -n "${NIXL_BRANCH:-}" ]]; then
+                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
            fi
            ;;
        rocshmem-rocm-ci)
@@ -1815,7 +1815,7 @@ add_dependency_cache_target() {

 resolve_ci_base_dependency_targets() {
    local mode="${ROCM_DEP_CACHE_EXPORT_MODE:-missing}"
-    local rixl_ref=""
+    local nixl_ref=""
    local rocshmem_ref=""
    local deepep_ref=""

@@ -1824,7 +1824,7 @@ resolve_ci_base_dependency_targets() {
    case "${mode}" in
        always)
            echo "ROCM_DEP_CACHE_EXPORT_MODE=always; exporting all dependency caches serially"
-            for target in rixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
+            for target in nixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
                if [[ -n "$(dependency_cache_ref_for_target "${target}")" ]]; then
                    add_dependency_cache_target "${target}"
                fi
@@ -1844,13 +1844,13 @@ resolve_ci_base_dependency_targets() {
            ;;
    esac

-    if [[ "${mode}" != "always" && -n "${RIXL_CACHE_KEY:-}" ]]; then
-        rixl_ref=$(dependency_cache_ref_for_target "rixl-rocm-ci")
-        if dependency_cache_ref_exists "${rixl_ref}"; then
-            echo "RIXL dependency cache exists: ${rixl_ref}"
+    if [[ "${mode}" != "always" && -n "${NIXL_CACHE_KEY:-}" ]]; then
+        nixl_ref=$(dependency_cache_ref_for_target "nixl-rocm-ci")
+        if dependency_cache_ref_exists "${nixl_ref}"; then
+            echo "NIXL dependency cache exists: ${nixl_ref}"
        else
-            echo "RIXL dependency cache missing; will seed: ${rixl_ref}"
-            add_dependency_cache_target "rixl-rocm-ci"
+            echo "NIXL dependency cache missing; will seed: ${nixl_ref}"
+            add_dependency_cache_target "nixl-rocm-ci"
        fi
    fi

@@ -35,7 +35,7 @@ set -o pipefail
 : "${PY_COLORS:=1}"
 : "${ROCM_DOCKER_TTY:=1}"
 : "${PYTHONFAULTHANDLER:=1}"
-: "${PYTEST_TIMEOUT:=2100}"
+: "${PYTEST_TIMEOUT:=2400}"
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --color"* ]]; then
  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--color=yes"
 fi
@@ -45,9 +45,9 @@ fi
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations-min="* ]]; then
  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations-min=1.0"
 fi
-# Dump stacks after 15 minutes, then stop an individual test after 35 minutes.
+# Dump stacks after 25 minutes, then stop an individual test after 40 minutes.
 if [[ " ${PYTEST_ADDOPTS:-} " != *" faulthandler_timeout="* ]]; then
-  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }-o faulthandler_timeout=900"
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }-o faulthandler_timeout=1500"
 fi
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --timeout-method="* &&
  " ${PYTEST_ADDOPTS:-} " != *" --timeout-method "* ]]; then
@@ -387,6 +387,7 @@ initialize_native_environment() {
  local job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
  local job_id_suffix=""
  local native_root=""
+  local hf_fstype=""
  local hf_mount=""

  if [[ "$(id -u)" -ne 0 ]]; then
@@ -400,16 +401,19 @@ initialize_native_environment() {
  native_root="/tmp/vllm-native-${job_id}"
  TMPDIR="/tmp/vllm-${job_id_suffix}/tmp"
  VLLM_RPC_BASE_PATH="/tmp"
-  : "${TORCHINDUCTOR_CACHE_DIR:=${native_root}/cache/torchinductor}"
-  : "${TRITON_CACHE_DIR:=${native_root}/cache/triton}"
-  : "${VLLM_CACHE_ROOT:=${native_root}/cache/vllm}"
-  : "${XDG_CACHE_HOME:=${native_root}/cache/xdg}"
+  TORCHINDUCTOR_CACHE_DIR="${native_root}/cache/torchinductor"
+  TRITON_CACHE_DIR="${native_root}/cache/triton"
+  VLLM_CACHE_ROOT="${native_root}/cache/vllm"
+  XDG_CACHE_HOME="${native_root}/cache/xdg"
  : "${HF_HOME:=/home/buildkite-agent/huggingface}"
+  # datasets uses POSIX locks that are unsupported by the shared HF NFS cache.
+  # Keep processed datasets job-local while retaining the persistent Hub cache.
+  HF_DATASETS_CACHE="${native_root}/cache/huggingface/datasets"
  : "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
  : "${HF_HUB_ETAG_TIMEOUT:=60}"
  export TMPDIR VLLM_RPC_BASE_PATH
  export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
-  export HF_HOME HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
+  export HF_HOME HF_DATASETS_CACHE HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
  export PYTORCH_ROCM_ARCH=""

  mkdir -p "${TMPDIR}" \
@@ -417,7 +421,10 @@ initialize_native_environment() {
    "${TRITON_CACHE_DIR}" \
    "${VLLM_CACHE_ROOT}" \
    "${XDG_CACHE_HOME}" \
-    "${HF_HOME}" || return 1
+    "${HF_HOME}" \
+    "${HF_DATASETS_CACHE}" || return 1
+
+  echo "Native compile caches: VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT} TORCHINDUCTOR_CACHE_DIR=${TORCHINDUCTOR_CACHE_DIR}"

  if [[ "${VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE:-0}" == "1" ]]; then
    if ! command -v findmnt >/dev/null 2>&1; then
@@ -430,6 +437,18 @@ initialize_native_environment() {
      return 1
    fi
  fi
+
+  if command -v findmnt >/dev/null 2>&1; then
+    hf_fstype=$(findmnt -n -T "${HF_HOME}" -o FSTYPE 2>/dev/null || true)
+  fi
+  if [[ "${hf_fstype}" == nfs || "${hf_fstype}" == nfs4 ]]; then
+    # Keep hf-xet state local and avoid vectored writes on shared NFS.
+    export HF_XET_CACHE="${native_root}/cache/hf-xet"
+    export HF_XET_HIGH_PERFORMANCE=0
+    export HF_XET_RECONSTRUCTION_USE_VECTORED_WRITE=0
+    mkdir -p "${HF_XET_CACHE}" || return 1
+    echo "Configured hf-xet for shared ${hf_fstype} cache at ${HF_HOME}"
+  fi
 }

 run_native_preflight() {
@@ -1,10 +1,11 @@
 #!/bin/bash
 set -euox pipefail

-export VLLM_CPU_KVCACHE_SPACE=1 
+export VLLM_CPU_KVCACHE_SPACE=1
 export VLLM_CPU_CI_ENV=1
-# Reduce sub-processes for acceleration
-export TORCH_COMPILE_DISABLE=1 
+# Skip torch.compile via vLLM's --enforce-eager flag (passed below) instead of
+# TORCH_COMPILE_DISABLE=1, which torch 2.12 no longer treats as a silent no-op
+# when callers specify fullgraph=True.
 export VLLM_ENABLE_V1_MULTIPROCESSING=0

 SDE_ARCHIVE="sde-external-10.7.0-2026-02-18-lin.tar.xz"
@@ -49,15 +50,15 @@ wait_for_pid_and_check_log() {
 }

 # Test Sky Lake (AVX512F)
-./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_0.log 2>&1 &
+./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_0.log 2>&1 &
 PID_TEST_0=$!

 # Test Cascade Lake (AVX512F + VNNI)
-./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_1.log 2>&1 &
+./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_1.log 2>&1 &
 PID_TEST_1=$!

 # Test Cooper Lake (AVX512F + VNNI + BF16)
-./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_2.log 2>&1 &
+./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_2.log 2>&1 &
 PID_TEST_2=$!

 wait_for_pid_and_check_log $PID_TEST_0 test_0.log
@@ -40,7 +40,7 @@
 #####################################################################################################################################
 #                                                                                                                                   #
 # IMPORTANT:                                                                                                                        #
-#   * Currently AMD CI has MI250 agents, MI300 agents, MI325 agents, and MI355 agents. All upcoming feature improvements are        #
+#   * Currently AMD CI has MI250 agents, MI300 agents, and MI355 agents. All upcoming feature improvements are                      #
 #         tracked in: https://github.com/vllm-project/vllm/issues/34994                                                             #
 #                                                                                                                                   #
 #-----------------------------------------------------------------------------------------------------------------------------------#
@@ -81,10 +81,8 @@
 #                                                     the above test.) Also run if model initialization test file is modified.      #
 #   * [Language Models Tests (Extra Standard) %N]: Shard slow subset of standard language models tests. Only run when model         #
 #                                                  source is modified, or when specified test files are modified.                   #
-#   * [Language Models Tests (Hybrid) %N]: Install fast path packages for testing against transformers (mamba, conv1d) and to       #
-#                                          run plamo2 model in vLLM.                                                                #
-#   * [Language Models Test (Extended Generation)]: Install fast path packages for testing against transformers (mamba, conv1d)     #
-#                                                   and to run plamo2 model in vLLM.                                                #
+#   * [Language Models Tests (Hybrid) %N]: Install fast path packages for testing against transformers (mamba, conv1d).             #
+#   * [Language Models Test (Extended Generation)]: Install fast path packages for testing against transformers (mamba, conv1d).    #
 #   * [Multi-Modal Models (Standard) 1-4]:                                                                                          #
 #     - Do NOT remove `VLLM_WORKER_MULTIPROC_METHOD=spawn` setting as ROCm requires this for certain models to function.            #
 #   * [Transformers Nightly Models]: Whisper needs `VLLM_WORKER_MULTIPROC_METHOD=spawn` to avoid deadlock.                          #
@@ -171,20 +169,6 @@ steps:
  - pip install helion==1.1.0
  - pytest -v -s kernels/helion/

- label: Kernels Mamba Test # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
-  agent_pool: mi250_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - csrc/mamba/
-  - tests/kernels/mamba
-  - vllm/model_executor/layers/mamba/ops
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s kernels/mamba
-
 #------------------------------------------------------  mi250 · models / basic  -------------------------------------------------------#

 - label: Basic Models Test (Other CPU) # TBD
@@ -207,6 +191,7 @@ steps:
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
  agent_pool: mi250_1
+  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -240,6 +225,20 @@ steps:
  - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
  - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model

+- label: Multi-Modal Processor (CPU) %N # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
+  agent_pool: mi250_1
+  parallelism: 6
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  - tests/models/registry.py
+  commands:
+  - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+
 #------------------------------------------------------------  mi250 · v1  -------------------------------------------------------------#

 - label: Batch Invariance (H100-MI250) # TBD
@@ -351,21 +350,21 @@ steps:
  commands:
  - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"

- label: V1 attention (H100-MI250) # TBD
+- label: V1 others (CPU) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
  agent_pool: mi250_1
+  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
-  - vllm/config/attention.py
-  - vllm/model_executor/layers/attention
-  - vllm/v1/attention
-  - tests/v1/attention
-  - vllm/_aiter_ops.py
-  - vllm/envs.py
-  - vllm/platforms/rocm.py
+  - vllm/
+  - tests/v1
  commands:
-  - pytest -v -s v1/attention
+  - pytest -v -s -m 'cpu_test' v1/core
+  - pytest -v -s v1/structured_output
+  - pytest -v -s v1/test_serial_utils.py
+  - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
+  - pytest -v -s -m 'cpu_test' v1/metrics

 #-------------------------------------------------------------  mi250 · misc  ------------------------------------------------------------#

@@ -408,6 +407,19 @@ steps:
  - pytest -v -s transformers_utils
  - pytest -v -s config

+- label: Python-only Installation # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
+  agent_pool: mi250_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - tests/standalone_tests/python_only_compile.sh
+  - setup.py
+  - vllm/platforms/rocm.py
+  commands:
+  - bash standalone_tests/python_only_compile.sh
+
 #------------------------------------------------------------  mi250 · rust  -----------------------------------------------------------#

 - label: Rust Frontend Cargo Style + Clippy # TBD
@@ -445,6 +457,7 @@ steps:
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
  agent_pool: mi250_1
  no_gpu: true
+  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - .buildkite/scripts/docker-build-metadata-args.sh
@@ -507,7 +520,7 @@ steps:
  - tests/models/
  commands:
  - TARGET_TEST_SUITE=MI300 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m '(not slow_test)'
+  - HIP_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py -m '(not slow_test)'
  - pytest models/transformers/test_backend.py -v -s -m 'distributed(num_gpus=2)'
  - pytest models/language -v -s -m 'distributed(num_gpus=2)'
  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/generation/test_phi4siglip.py
@@ -657,6 +670,30 @@ steps:
  - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/passes/distributed/test_async_tp.py
  - pytest -v -s tests/compile/fusions_e2e/test_tp2_ar_rms.py::test_tp2_ar_rms_fusions

+- label: Distributed Compile + RPC Tests (2 GPUs) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_2
+  num_gpus: 2
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/compilation/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/worker/worker_base.py
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/compile/fullgraph/test_basic_correctness.py
+  - tests/compile/test_wrapper.py
+  - tests/entrypoints/llm/test_collective_rpc.py
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s ./compile/test_wrapper.py
+
 #-----------------------------------------------------------  mi300 · cuda  ------------------------------------------------------------#

 - label: Platform Tests # TBD
@@ -679,6 +716,7 @@ steps:
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
+  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -870,6 +908,71 @@ steps:
  commands:
  - torchrun --nproc-per-node=8 ../examples/features/torchrun/torchrun_dp_example_offline.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep

+- label: Distributed Torchrun + Shutdown Tests (2 GPUs) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_2
+  num_gpus: 2
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/worker/worker_base.py
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/distributed/
+  - tests/v1/shutdown
+  - tests/v1/worker/test_worker_memory_snapshot.py
+  - vllm/platforms/rocm.py
+  commands:
+  - 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'
+  - HIP_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
+  - pytest -v -s v1/worker/test_worker_memory_snapshot.py
+
+- label: Distributed Compile + Comm (4 GPUs) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_4
+  num_gpus: 4
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/distributed/
+  - tests/distributed/test_pynccl
+  - tests/distributed/test_events
+  - tests/compile/fullgraph/test_basic_correctness.py
+  - tests/distributed/test_symm_mem_allreduce.py
+  - tests/distributed/test_multiproc_executor.py
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s distributed/test_pynccl.py
+  - pytest -v -s distributed/test_events.py
+  - pytest -v -s distributed/test_symm_mem_allreduce.py
+  - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
+
+#----------------------------------------------------------  mi300 · engine  -----------------------------------------------------------#
+
+- label: Engine # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/engine
+  - tests/test_sequence
+  - tests/test_config
+  - tests/test_logger
+  - tests/test_vllm_port
+  commands:
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
+
 #--------------------------------------------------------  mi300 · entrypoints  --------------------------------------------------------#

 - label: Entrypoints Unit Tests # TBD
@@ -878,6 +981,7 @@ steps:
  dind: false
  agent_pool: mi300_1
  fast_check: true
+  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/entrypoints
@@ -982,6 +1086,7 @@ steps:
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
+  optional: true
  fast_check: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -1010,6 +1115,7 @@ steps:
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
+  optional: true
  fast_check: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -1024,6 +1130,7 @@ steps:
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
+  optional: true
  fast_check: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -1345,6 +1452,27 @@ steps:
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm.txt --tp-size=8

+- label: LM Eval Large Models (4xH100-4xMI300) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_4
+  num_gpus: 4
+  optional: true
+  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
+  source_file_dependencies:
+  - csrc/
+  - vllm/model_executor/layers/quantization
+  - vllm/model_executor/models/
+  - vllm/model_executor/model_loader/
+  - vllm/v1/attention/backends/
+  - vllm/v1/attention/selector.py
+  - vllm/_aiter_ops.py
+  - vllm/platforms/rocm.py
+  commands:
+  - export VLLM_USE_DEEP_GEMM=0
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm-fp8.txt --tp-size=4
+
 #---------------------------------------------------------  mi300 · examples  ----------------------------------------------------------#

 - label: Examples # TBD
@@ -1419,11 +1547,12 @@ steps:
  commands:
  - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT

- label: Kernels Core Operation Test # TBD
+- label: Kernels Core Operation Test %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
+  parallelism: 3
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - csrc/
@@ -1434,7 +1563,7 @@ steps:
  - vllm/_aiter_ops.py
  - vllm/platforms/rocm.py
  commands:
-  - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py kernels/test_top_k_per_row.py
+  - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py kernels/test_concat_mla_q.py kernels/test_top_k_per_row.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT

 - label: Kernels KDA Test # TBD
  timeout_in_minutes: 180
@@ -1452,6 +1581,21 @@ steps:
  commands:
  - pytest -v -s kernels/test_kda.py

+- label: Kernels Mamba Test # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - csrc/mamba/
+  - tests/kernels/mamba
+  - vllm/model_executor/layers/mamba/ops
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s kernels/mamba
+
 - label: Kernels MoE Test %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -1593,7 +1737,7 @@ steps:
  - set -x
  - export VLLM_USE_V2_MODEL_RUNNER=1
  - pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics"
-  - ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram"
+  - pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram"
  - pytest -v -s v1/e2e/general/test_context_length.py
  - pytest -v -s v1/e2e/general/test_min_tokens.py
  - pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0"
@@ -1797,6 +1941,37 @@ steps:
  - pip freeze | grep -E 'torch'
  - pytest -v -s models/language -m 'core_model and slow_test' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB

+- label: Language Models Test (Extended Generation) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+  - 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@v1.6.0'
+  - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
+
+- label: Language Models Tests (Hybrid) %N # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_1
+  parallelism: 2
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+  - 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@v1.6.0'
+  - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+
 #----------------------------------------------------  mi300 · models / multimodal  ----------------------------------------------------#

 - label: Multi-Modal Models (Extended Generation 1) # TBD
@@ -1899,20 +2074,32 @@ steps:
  commands:
  - pytest -v -s models/multimodal/processing/test_tensor_schema.py

- label: Multi-Modal Processor (CPU) %N # TBD
+- label: Multi-Modal Models (Extended Pooling) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal/pooling
+  commands:
+  - pytest -v -s models/multimodal/pooling -m 'not core_model'
+
+- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma" # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
-  parallelism: 4
  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
-  - tests/models/registry.py
  commands:
-  - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+  - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
+  - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model

 #-----------------------------------------------------  mi300 · models / quantized  -----------------------------------------------------#

@@ -1933,12 +2120,12 @@ steps:

 #--------------------------------------------------  mi300 · models / transformers  ---------------------------------------------------#

- label: Transformers Nightly Models (Shardable) %N # TBD
+- label: Transformers Nightly Models (Initialization) %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
-  parallelism: 4
+  parallelism: 6
  optional: true
  working_dir: "/vllm-workspace/"
  source_file_dependencies:
@@ -1954,6 +2141,27 @@ steps:
  commands:
  - pip install --upgrade git+https://github.com/huggingface/transformers
  - pytest -v -s tests/models/test_initialization.py --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+
+- label: Transformers Nightly Models (Processing) %N # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  dind: false
+  agent_pool: mi300_1
+  parallelism: 8
+  optional: true
+  working_dir: "/vllm-workspace/"
+  source_file_dependencies:
+  - vllm/model_executor/models/
+  - vllm/model_executor/model_loader/
+  - vllm/multimodal/
+  - vllm/model_executor/layers/
+  - vllm/v1/attention/backends/
+  - vllm/v1/attention/selector.py
+  - vllm/_aiter_ops.py
+  - vllm/platforms/rocm.py
+  - tests/models/
+  commands:
+  - pip install --upgrade git+https://github.com/huggingface/transformers
  - pytest -v -s tests/models/multimodal/processing/ --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB

 - label: Transformers Nightly Models (Single) # TBD
@@ -2472,11 +2680,12 @@ steps:
  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
  - pytest -v -s v1/kv_connector/extract_hidden_states_integration

- label: V1 attention (H100-MI300) # TBD
+- label: V1 attention (H100-MI300) %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
  dind: false
  agent_pool: mi300_1
+  parallelism: 2
  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -2488,7 +2697,7 @@ steps:
  - vllm/envs.py
  - vllm/platforms/rocm.py
  commands:
-  - pytest -v -s v1/attention
+  - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT

 - label: V1 Core + KV + Metrics # TBD
  timeout_in_minutes: 180
@@ -2517,23 +2726,6 @@ steps:
  # - export HSA_NO_SCRATCH_RECLAIM=1
  - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine

- label: V1 others (CPU) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  dind: false
-  agent_pool: mi300_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/v1
-  commands:
-  - pytest -v -s -m 'cpu_test' v1/core
-  - pytest -v -s v1/structured_output
-  - pytest -v -s v1/test_serial_utils.py
-  - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
-  - pytest -v -s -m 'cpu_test' v1/metrics
-
 - label: V1 Sample + Logits # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
@@ -2830,195 +3022,6 @@ steps:
  commands:
  - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt

-#########################################################################################################################################
-#                                                                                                                                       #
-#                                                         MI325 (gfx942) tests                                                          #
-#                                                                                                                                       #
-#########################################################################################################################################
-
-#----------------------------------------------------------  mi325 · compile  ----------------------------------------------------------#
-
- label: Distributed Compile + RPC Tests (2 GPUs) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_2
-  num_gpus: 2
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/compilation/
-  - vllm/distributed/
-  - vllm/engine/
-  - vllm/executor/
-  - vllm/worker/worker_base.py
-  - vllm/v1/engine/
-  - vllm/v1/worker/
-  - tests/compile/fullgraph/test_basic_correctness.py
-  - tests/compile/test_wrapper.py
-  - tests/entrypoints/llm/test_collective_rpc.py
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
-  - pytest -v -s ./compile/test_wrapper.py
-
-#--------------------------------------------------------  mi325 · distributed  --------------------------------------------------------#
-
- label: Distributed Torchrun + Shutdown Tests (2 GPUs) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_2
-  num_gpus: 2
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/distributed/
-  - vllm/engine/
-  - vllm/executor/
-  - vllm/worker/worker_base.py
-  - vllm/v1/engine/
-  - vllm/v1/worker/
-  - tests/distributed/
-  - tests/v1/shutdown
-  - tests/v1/worker/test_worker_memory_snapshot.py
-  - vllm/platforms/rocm.py
-  commands:
-  - 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'
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
-  - pytest -v -s v1/worker/test_worker_memory_snapshot.py
-
- label: Distributed Compile + Comm (4 GPUs) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_4
-  num_gpus: 4
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/distributed/
-  - tests/distributed/test_pynccl
-  - tests/distributed/test_events
-  - tests/compile/fullgraph/test_basic_correctness.py
-  - tests/distributed/test_symm_mem_allreduce.py
-  - tests/distributed/test_multiproc_executor.py
-  - vllm/platforms/rocm.py
-  commands:
-  - pytest -v -s compile/fullgraph/test_basic_correctness.py
-  - pytest -v -s distributed/test_pynccl.py
-  - pytest -v -s distributed/test_events.py
-  - pytest -v -s distributed/test_symm_mem_allreduce.py
-  - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
-
-#----------------------------------------------------------  mi325 · engine  -----------------------------------------------------------#
-
- label: Engine # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/engine
-  - tests/test_sequence
-  - tests/test_config
-  - tests/test_logger
-  - tests/test_vllm_port
-  commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
-
-#-----------------------------------------------------------  mi325 · evals  -----------------------------------------------------------#
-
- label: LM Eval Large Models (4xH100-4xMI325) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_4
-  num_gpus: 4
-  optional: true
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  - vllm/model_executor/models/
-  - vllm/model_executor/model_loader/
-  - vllm/v1/attention/backends/
-  - vllm/v1/attention/selector.py
-  - vllm/_aiter_ops.py
-  - vllm/platforms/rocm.py
-  commands:
-  - export VLLM_USE_DEEP_GEMM=0
-  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm-fp8.txt --tp-size=4
-
-#-----------------------------------------------------  mi325 · models / language  -----------------------------------------------------#
-
- label: Language Models Test (Extended Generation) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_1
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/models/language/generation
-  commands:
-  - 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@v1.6.0'
-  - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
-
- label: Language Models Tests (Hybrid) %N # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_1
-  parallelism: 2
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/models/language/generation
-  commands:
-  - 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@v1.6.0'
-  - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
-
-#----------------------------------------------------  mi325 · models / multimodal  ----------------------------------------------------#
-
- label: Multi-Modal Models (Extended Pooling) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/models/multimodal/pooling
-  commands:
-  - pytest -v -s models/multimodal/pooling -m 'not core_model'
-
- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma" # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/models/multimodal
-  commands:
-  - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
-  - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
-
-#-----------------------------------------------------------  mi325 · misc  ------------------------------------------------------------#
-
- label: Python-only Installation # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
-  agent_pool: mi325_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - tests/standalone_tests/python_only_compile.sh
-  - setup.py
-  - vllm/platforms/rocm.py
-  commands:
-  - bash standalone_tests/python_only_compile.sh
-
 #########################################################################################################################################
 #                                                                                                                                       #
 #                                                         MI355 (gfx950) tests                                                          #
@@ -3030,6 +3033,7 @@ steps:
 - label: Attention Benchmarks Smoke Test (B200-MI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  working_dir: "/vllm-workspace/"
@@ -3046,6 +3050,7 @@ steps:
 - label: Distributed Tests (2xH100-2xMI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  optional: true
@@ -3090,6 +3095,7 @@ steps:
 - label: Entrypoints Integration (API Server) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  fast_check: true
@@ -3107,6 +3113,7 @@ steps:
 - label: Entrypoints Integration (API Server OpenAI - Part 1) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  fast_check: true
  optional: true
@@ -3122,6 +3129,7 @@ steps:
 - label: Entrypoints Integration (API Server OpenAI - Part 2) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  fast_check: true
  optional: true
@@ -3138,6 +3146,7 @@ steps:
 - label: Entrypoints Integration (API Server Generate) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  fast_check: true
  optional: true
@@ -3158,6 +3167,7 @@ steps:
 - label: Entrypoints Integration (Speech to Text) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  fast_check: true
  working_dir: "/vllm-workspace/tests"
@@ -3171,6 +3181,7 @@ steps:
 - label: Entrypoints Integration (Multimodal)
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  fast_check: true
  working_dir: "/vllm-workspace/tests"
@@ -3184,6 +3195,7 @@ steps:
 - label: Entrypoints Integration (Pooling) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  fast_check: true
  working_dir: "/vllm-workspace/tests"
@@ -3199,6 +3211,7 @@ steps:
 - label: GPQA Eval (GPT-OSS) (2xB200-2xMI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  optional: true
@@ -3221,6 +3234,7 @@ steps:
 - label: LM Eval Qwen3-5 Models (B200-MI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  optional: true
@@ -3243,6 +3257,7 @@ steps:
 - label: LM Eval Small Models (2xB200-2xMI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  optional: true
@@ -3262,6 +3277,7 @@ steps:
 - label: Qwen3-30B-A3B-FP8-block Sync EPLB Accuracy (B200-MI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  working_dir: "/vllm-workspace"
@@ -3282,6 +3298,7 @@ steps:
 - label: LM Eval Large Models (4xH100-4xMI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_4
  num_gpus: 4
  optional: true
@@ -3304,6 +3321,7 @@ steps:
 - label: Examples # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  working_dir: "/vllm-workspace/examples"
  source_file_dependencies:
@@ -3339,6 +3357,7 @@ steps:
 - label: Kernels (B200-MI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  working_dir: "/vllm-workspace/"
  source_file_dependencies:
@@ -3364,6 +3383,7 @@ steps:
 - label: Kernels Attention Test %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  parallelism: 2
  working_dir: "/vllm-workspace/tests"
@@ -3381,6 +3401,7 @@ steps:
 - label: Kernels MoE Test %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  parallelism: 5
  working_dir: "/vllm-workspace/tests"
@@ -3401,6 +3422,7 @@ steps:
 - label: Kernels Quantization Test %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  parallelism: 2
  working_dir: "/vllm-workspace/tests"
@@ -3418,6 +3440,7 @@ steps:
 - label: Kernels FP8 MoE Test (2xH100-2xMI355) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  working_dir: "/vllm-workspace/tests"
@@ -3437,6 +3460,7 @@ steps:
 - label: Language Models Test (Extended Generation) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -3450,6 +3474,7 @@ steps:
 - label: Language Models Test (Extended Pooling) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3462,7 +3487,9 @@ steps:
 - label: Language Models Test (PPL) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
+  optional: true
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/model_executor/models/qwen3_5.py
@@ -3489,6 +3516,7 @@ steps:
 - label: Language Models Tests (Standard) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -3503,6 +3531,7 @@ steps:
 - label: Multi-Modal Models (Extended Generation 1) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3517,6 +3546,7 @@ steps:
 - label: Multi-Modal Models (Extended Generation 3) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3529,6 +3559,7 @@ steps:
 - label: Multi-Modal Models (Extended Pooling) # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3541,6 +3572,7 @@ steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2" # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3554,6 +3586,7 @@ steps:
 - label: "Multi-Modal Models (Standard) 4: other + whisper" # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3570,6 +3603,7 @@ steps:
 - label: Quantized Models Test # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -3586,6 +3620,7 @@ steps:
 - label: Quantization # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -3602,6 +3637,7 @@ steps:
 # - label: Quantized MoE Test (B200-MI355) # TBD
 #   timeout_in_minutes: 180
 #   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+#   dind: false
 #   agent_pool: mi355_1
 #   working_dir: "/vllm-workspace/"
 #   source_file_dependencies:
@@ -3627,10 +3663,12 @@ steps:

 #------------------------------------------------------------  mi355 · v1  -------------------------------------------------------------#

- label: V1 attention (B200-MI355) # TBD
+- label: V1 attention (B200-MI355) %N # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
+  parallelism: 2
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/config/attention.py
@@ -3641,11 +3679,12 @@ steps:
  - vllm/envs.py
  - vllm/platforms/rocm.py
  commands:
-  - pytest -v -s v1/attention
+  - pytest -v -s v1/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT

 - label: V1 Core + KV + Metrics # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3672,6 +3711,7 @@ steps:
 - label: V1 Sample + Logits # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -3692,6 +3732,7 @@ steps:
 - label: V1 Spec Decode # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
@@ -3705,6 +3746,7 @@ steps:
 - label: Weight Loading Multiple GPU # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  num_gpus: 2
  working_dir: "/vllm-workspace/tests"
@@ -3717,6 +3759,7 @@ steps:
 - label: Weight Loading Multiple GPU - Large Models # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_2
  working_dir: "/vllm-workspace/tests"
  num_gpus: 2
@@ -3732,6 +3775,7 @@ steps:
 - label: Regression # TBD
  timeout_in_minutes: 180
  mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
+  dind: false
  agent_pool: mi355_1
  optional: true
  working_dir: "/vllm-workspace/tests"
@@ -16,8 +16,9 @@ steps:
  parallelism: 2
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 95
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 125
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Correctness
  key: basic-correctness
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -18,7 +18,8 @@ steps:
  - pytest -v -s basic_correctness/test_cpu_offload.py
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Benchmarks CLI Test
  key: benchmarks-cli-test
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -15,6 +15,7 @@ steps:
    amd:
      dind: false
      device: mi300_1
+      timeout_in_minutes: 40
      depends_on:
      - image-build-amd

@@ -26,7 +26,10 @@ steps:
  - vllm/v1/cudagraph_dispatcher.py
  - vllm/config/compilation.py
  - vllm/compilation
+  - vllm/v1/worker/encoder_cudagraph.py
+  - vllm/v1/worker/encoder_cudagraph_defs.py
  commands:
    - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
    - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py
+    - pytest -v -s v1/cudagraph/test_encoder_cudagraph.py
@@ -17,7 +17,7 @@ steps:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 85
+      timeout_in_minutes: 60
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -68,7 +68,7 @@ steps:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 60
+      timeout_in_minutes: 40
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -94,7 +94,7 @@ steps:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 85
+      timeout_in_minutes: 60
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -120,7 +120,7 @@ steps:
    amd:
      dind: false
      device: mi300_4
-      timeout_in_minutes: 80
+      timeout_in_minutes: 55
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -177,7 +177,7 @@ steps:
    amd:
      dind: false
      device: mi300_2
-      timeout_in_minutes: 70
+      timeout_in_minutes: 45
      depends_on:
        - image-build-amd
      source_file_dependencies:
@@ -41,6 +41,7 @@ steps:
    amd:
      dind: false
      device: mi300_2
+      timeout_in_minutes: 45
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -28,8 +28,9 @@ steps:
  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 50
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 40
      depends_on:
      - image-build-amd

@@ -44,14 +45,14 @@ steps:
    - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 55
+      device: mi250_1
+      timeout_in_minutes: 45
      depends_on:
      - image-build-amd

 - label: e2e Scheduling (1 GPU)
  key: e2e-scheduling-1-gpu
-  timeout_in_minutes: 35
+  timeout_in_minutes: 53
  device: h200_18gb
  source_file_dependencies:
    - vllm/v1/
@@ -60,8 +61,8 @@ steps:
    - pytest -v -s v1/e2e/general/test_async_scheduling.py
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      device: mi250_1
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd

@@ -76,8 +77,8 @@ steps:
    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      device: mi250_1
+      timeout_in_minutes: 50
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -116,6 +117,7 @@ steps:
    amd:
      dind: false
      device: mi300_2
+      timeout_in_minutes: 30
      depends_on:
      - image-build-amd

@@ -30,16 +30,16 @@ steps:
  - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
  mirror:
    amd:
-      device: mi325_1
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd

 - label: Entrypoints Integration (API Server)
  key: entrypoints-integration-api-server
  device: h200_35gb
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -52,14 +52,16 @@ steps:
  - pytest -v -s entrypoints/scale_out
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
      depends_on:
      - image-build-amd

 - label: Entrypoints Integration (API Server OpenAI - Part 1)
  device: h200_35gb
  key: entrypoints-integration-api-server-openai-part-1
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -70,7 +72,8 @@ steps:
  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
@@ -78,7 +81,7 @@ steps:
 - label: Entrypoints Integration (API Server OpenAI - Part 2)
  device: h200_35gb
  key: entrypoints-integration-api-server-openai-part-2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 83
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -90,8 +93,9 @@ steps:
  - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 80
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 70
      depends_on:
      - image-build-amd

@@ -113,7 +117,8 @@ steps:
  - pytest -v -s entrypoints/anthropic
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
      timeout_in_minutes: 65
      depends_on:
      - image-build-amd
@@ -156,7 +161,7 @@ steps:
 - label: Entrypoints Integration (Pooling)
  device: h200_35gb
  key: entrypoints-integration-pooling
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
  working_dir: "/vllm-workspace/tests"
  source_file_dependencies:
  - vllm/
@@ -176,7 +181,9 @@ steps:
  - pytest -s entrypoints/openai/correctness/
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -18,6 +18,7 @@ steps:
    amd:
      dind: false
      device: mi300_1
+      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -0,0 +1,26 @@
+group: Fault Tolerance
+depends_on:
+  - image-build
+steps:
+- label: Fault Tolerance E2E (2xH100)
+  key: fault-tolerance-e2e-2xh100
+  timeout_in_minutes: 35
+  device: h100
+  num_devices: 2
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/v1/fault_tolerance/
+  - vllm/v1/worker/sentinel/
+  - vllm/entrypoints/serve/fault_tolerance/
+  - vllm/distributed/elastic_ep/
+  - vllm/distributed/device_communicators/
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/v1/fault_tolerance/
+  - tests/v1/distributed/test_external_lb_dp.py
+  commands:
+  # Base image has no nixl; install it or has_nixl_ep() skips the tests.
+  - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
+  - pytest -v -s v1/fault_tolerance/test_fault_tolerance_e2e.py
@@ -80,7 +80,8 @@ steps:
  parallelism: 2
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
      timeout_in_minutes: 90
      depends_on:
      - image-build-amd
@@ -118,7 +119,9 @@ steps:
  parallelism: 2
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 120
      source_file_dependencies:
      - csrc/quantization/
      - vllm/model_executor/layers/quantization
@@ -148,8 +151,9 @@ steps:
  parallelism: 5
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 65
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
      source_file_dependencies:
      - csrc/quantization/cutlass_w8a8/moe/
      - csrc/moe/
@@ -14,8 +14,9 @@ steps:
  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 55
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 45
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -141,7 +142,7 @@ steps:
    amd:
      dind: false
      device: mi300_8
-      timeout_in_minutes: 60
+      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
      commands:
@@ -336,7 +337,7 @@ steps:

 - label: LM Eval KV-Offload (2xH100)
  key: kv-offload-medium
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
  device: h100
  num_devices: 2
  source_file_dependencies:
@@ -346,7 +347,7 @@ steps:
  - vllm/v1/simple_kv_offload/
  - tests/evals/gsm8k/test_gsm8k_offloading.py
  commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b"
+  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b or deepseek-v2-lite"

 - label: LM Eval KV-Offload (4xH100)
  key: kv-offload-large
@@ -14,9 +14,10 @@ steps:
  parallelism: 4
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
      working_dir: "/vllm-workspace/tests"
-      timeout_in_minutes: 65
+      timeout_in_minutes: 85
      source_file_dependencies:
      - vllm/lora
      - tests/lora
@@ -46,4 +47,4 @@ steps:
    - pytest -v -s -x lora/test_qwen3_with_multi_loras.py
    - pytest -v -s -x lora/test_olmoe_tp.py
    - pytest -v -s -x lora/test_gptoss_tp.py
-    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
+    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
@@ -25,13 +25,13 @@ steps:
    amd:
      dind: false
      device: mi300_1
-      timeout_in_minutes: 75
+      timeout_in_minutes: 50
      depends_on:
      - image-build-amd

 - label: V1 Sample + Logits
  key: v1-sample-logits
-  timeout_in_minutes: 45
+  timeout_in_minutes: 83
  device: h200_18gb
  source_file_dependencies:
    - vllm/config/
@@ -59,7 +59,9 @@ steps:
    - pytest -v -s v1/test_outputs.py
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 70
      depends_on:
      - image-build-amd

@@ -90,6 +92,7 @@ steps:
    - tests/v1/kv_offload
    - tests/v1/simple_kv_offload
    - tests/v1/worker
+    - tests/v1/streaming_input
    - tests/v1/kv_connector/unit
    - tests/v1/ec_connector/unit
    - tests/v1/metrics
@@ -103,6 +106,7 @@ steps:
    - pytest -v -s v1/kv_offload
    - pytest -v -s v1/simple_kv_offload
    - pytest -v -s v1/worker
+    - pytest -v -s v1/streaming_input
    - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
    - pytest -v -s -m 'not cpu_test' v1/ec_connector/unit
    - pytest -v -s -m 'not cpu_test' v1/metrics
@@ -111,8 +115,9 @@ steps:
    - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 75
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
      depends_on:
      - image-build-amd

@@ -143,6 +148,7 @@ steps:
    - pytest -v -s -m 'cpu_test' v1/core
    - pytest -v -s v1/structured_output
    - pytest -v -s v1/test_serial_utils.py
+    - pytest -v -s v1/cudagraph/test_cudagraph_manager.py
    - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
    - pytest -v -s -m 'cpu_test' v1/metrics

@@ -206,7 +212,7 @@ steps:
  - vllm/multimodal
  - examples/
  commands:
-    - pip install tensorizer # for tensorizer test
+    - pip install --no-deps tensorizer # for tensorizer test
     # for basic
    - python3 basic/offline_inference/chat.py
    - python3 basic/offline_inference/generate.py --model facebook/opt-125m
@@ -230,7 +236,9 @@ steps:
    - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 75
      source_file_dependencies:
      - vllm/entrypoints
      - vllm/multimodal
@@ -266,10 +274,11 @@ steps:
  - pytest -v -s v1/tracing
  mirror:
    amd:
-      device: mi325_2
+      dind: false
+      device: mi300_2
+      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
-      optional: true

 - label: Python-only Installation
  key: python-only-installation
@@ -284,8 +293,8 @@ steps:
  - bash standalone_tests/python_only_compile.sh
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 45
+      device: mi250_1
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -5,7 +5,7 @@ steps:
 - label: Model Executor
  device: h200_35gb
  key: model-executor
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
  source_file_dependencies:
  - vllm/engine/arg_utils.py
  - vllm/config/model.py
@@ -30,6 +30,7 @@ steps:
    amd:
      dind: false
      device: mi300_1
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -41,7 +41,7 @@ steps:
  commands:
    - set -x
    - export VLLM_USE_V2_MODEL_RUNNER=1
-    - pip install tensorizer # for tensorizer test
+    - pip install --no-deps tensorizer # for tensorizer test
    - python3 basic/offline_inference/chat.py # for basic
    - python3 basic/offline_inference/generate.py --model facebook/opt-125m
    #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10  # TODO
@@ -42,10 +42,25 @@ steps:
    - pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 50
      depends_on:
      - image-build-amd

+- label: Inkling Unit Tests (B200)
+  key: inkling-unit-tests-b200
+  timeout_in_minutes: 40
+  device: b200-k8s
+  source_file_dependencies:
+  - vllm/models/inkling/
+  - vllm/cute_utils/
+  - cmake/external_projects/tml_fa4.cmake
+  - tests/models/inkling/
+  commands:
+    # FA4 kernel tests require SM100; the suite skips them elsewhere.
+    - pytest -v -s models/inkling
+
 - label: Basic Models Test (Other CPU) # 5min
  key: basic-models-test-other-cpu
  depends_on:
@@ -17,6 +17,7 @@ steps:
    amd:
      dind: false
      device: mi300_1
+      timeout_in_minutes: 45
      depends_on:
      - image-build-amd

@@ -39,6 +40,7 @@ steps:
    amd:
      dind: false
      device: mi300_1
+      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -61,7 +63,6 @@ steps:
  - tests/models/language/generation
  commands:
    # Install fast path packages for testing against transformers
-    # Note: also needed to run plamo2 model in vLLM
    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
    # Shard the hybrid language model tests that are numerically stable on Hopper.
@@ -69,8 +70,9 @@ steps:
  parallelism: 2
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      commands:
@@ -102,7 +104,6 @@ steps:
  - tests/models/language/generation
  commands:
    # Install fast path packages for testing against transformers
-    # Note: also needed to run plamo2 model in vLLM
    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
    - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
@@ -130,14 +131,15 @@ steps:
    - pytest -v -s models/language/pooling -m 'not core_model'
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 120
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 95
      depends_on:
      - image-build-amd

 - label: Language Models Test (MTEB)
  key: language-models-test-mteb
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
  device: h200_18gb
  optional: true
  source_file_dependencies:
@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
  key: multi-modal-models-standard-1-qwen2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -14,13 +14,15 @@ steps:
    - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
      depends_on:
      - image-build-amd

 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
  key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -31,7 +33,9 @@ steps:
    - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd

@@ -47,14 +51,15 @@ steps:
    - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
  mirror:
    amd:
-      device: mi325_1
+      device: mi250_1
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd

 - label: "Multi-Modal Models (Standard) 4: other + whisper"
  device: h200_35gb
  key: multi-modal-models-standard-4-other-whisper
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
  source_file_dependencies:
  - vllm/
  - tests/models/multimodal
@@ -65,7 +70,9 @@ 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
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 50
      depends_on:
      - image-build-amd

@@ -85,7 +92,7 @@ steps:

 - label: Multi-Modal Processor # 44min
  key: multi-modal-processor
-  timeout_in_minutes: 65
+  timeout_in_minutes: 98
  device: h200_18gb
  source_file_dependencies:
  - vllm/
@@ -109,6 +116,7 @@ steps:
    amd:
      dind: false
      device: mi300_1
+      timeout_in_minutes: 35
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -131,7 +139,9 @@ steps:
    - pytest -v -s models/multimodal/test_mapping.py
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 90
      depends_on:
      - image-build-amd

@@ -166,8 +176,9 @@ steps:
    - pytest -v -s models/multimodal/pooling -m 'not core_model'
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 75
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -5,7 +5,7 @@ steps:
 - label: PyTorch Compilation Unit Tests
  device: h200_35gb
  key: pytorch-compilation-unit-tests
-  timeout_in_minutes: 90
+  timeout_in_minutes: 150
  source_file_dependencies:
    - vllm/__init__.py
    - vllm/_aiter_ops.py
@@ -107,13 +107,6 @@ steps:
    - tests/compile/passes
  commands:
  - pytest -s -v compile/passes --ignore compile/passes/distributed
-  mirror:
-    amd:
-      dind: false
-      device: mi300_1
-      timeout_in_minutes: 65
-      depends_on:
-      - image-build-amd

 - label: PyTorch Fullgraph Smoke Test
  device: h200_35gb
@@ -236,6 +229,7 @@ steps:
    amd:
      dind: false
      device: mi300_1
+      timeout_in_minutes: 30
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -24,8 +24,7 @@ steps:
  - uv pip install --system conch-triton-kernels
  # The SM90-only checkpoint currently contains a removed weight_chan_scale
  # parameter. It was not exercised by the previous L4 job.
-  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py -k 'not test_compressed_tensors_w4a8_fp8' --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-  parallelism: 8
+  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py -k 'not test_compressed_tensors_w4a8_fp8'

 - label: Quantized Fusions
  device: h200_35gb
@@ -68,5 +67,4 @@ steps:
  - vllm/model_executor/layers/quantization
  - tests/models/quantization
  commands:
-    - pytest -v -s models/quantization --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-  parallelism: 3
+    - pytest -v -s models/quantization
@@ -19,8 +19,18 @@ steps:
    - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
  mirror:
    amd:
-      device: mi325_1
+      device: mi250_1
+      timeout_in_minutes: 40
      depends_on:
      - image-build-amd
+      source_file_dependencies:
+      - vllm/model_executor/layers
+      - vllm/sampling_metadata.py
+      - vllm/v1/sample/
+      - vllm/entrypoints/generate/beam_search/
+      - tests/samplers
+      - tests/conftest.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
      commands:
      - pytest -v -s samplers
@@ -14,8 +14,9 @@ steps:
    - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -53,8 +54,9 @@ steps:
    - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 65
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 75
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -92,10 +94,9 @@ steps:
    - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
  mirror:
    amd:
-      device: mi325_1
-      timeout_in_minutes: 55
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 35
      depends_on:
      - image-build-amd
      source_file_dependencies:
@@ -119,7 +120,8 @@ steps:
    - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
  mirror:
    amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
      timeout_in_minutes: 55
      depends_on:
      - image-build-amd
@@ -170,3 +172,19 @@ steps:
    - tests/v1/e2e/spec_decode/
  commands:
    - pytest -v -s v1/e2e/spec_decode -k "qwen3_5-hybrid"
+
+- label: Spec Decode DeepSeek MTP Parallel Load (B200)
+  key: spec-decode-deepseek-mtp-parallel-load-b200
+  timeout_in_minutes: 30
+  device: b200-k8s
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/v1/spec_decode/llm_base_proposer.py
+    - vllm/v1/spec_decode/eagle.py
+    - vllm/v1/worker/gpu/spec_decode/eagle/
+    - vllm/model_executor/models/deepseek_mtp.py
+    - vllm/model_executor/models/deepseek_v2.py
+    - tests/v1/e2e/spec_decode/test_mtp_parallel_load.py
+  commands:
+    - pytest -v -s v1/e2e/spec_decode/test_mtp_parallel_load.py
@@ -17,6 +17,7 @@ steps:
    amd:
      dind: false
      device: mi300_2
+      timeout_in_minutes: 35
      depends_on:
      - image-build-amd
      commands:
@@ -19,6 +19,7 @@ pull_request_rules:
  description: Comment on PR when pre-commit check fails
  conditions:
    - check-failure=pre-commit
+    - -check-cancelled=pre-commit
    - -closed
    - -draft
    - or:
@@ -181,6 +182,18 @@ pull_request_rules:
      add:
        - performance

+- name: label-quantization
+  description: Automatically apply quantization label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^vllm/model_executor/layers/quantization/
+      - title~=(?i)quant
+  actions:
+    label:
+      add:
+        - quantization
+
 - name: label-qwen
  description: Automatically apply qwen label
  conditions:
@@ -220,6 +233,31 @@ pull_request_rules:
      add:
        - gpt-oss

+- name: label-kimi
+  description: Automatically apply kimi label
+  conditions:
+    - label != stale
+    - or:
+      - files~=(?i)kimi
+      - files~=(?i)moonshot
+      - title~=(?i)(?:kimi|moonshot)
+  actions:
+    label:
+      add:
+        - kimi
+
+- name: label-k3
+  description: Automatically apply k3 label (launch triage; retire after ramp-down)
+  conditions:
+    - label != stale
+    - or:
+      - files~=(?i)kimi[-_]?k3
+      - title~=(?i)(?:kimi[-\s]?k3|\bk3\b)
+  actions:
+    label:
+      add:
+        - k3
+
 - name: label-nvidia
  description: Automatically apply nvidia label
  conditions:
@@ -130,6 +130,66 @@ jobs:
                  },
                ],
              },
+              kimi: {
+                keywords: [
+                  { term: "Kimi",     searchIn: "both" },
+                  { term: "Moonshot", searchIn: "both" },
+                ],
+                substrings: [
+                  { term: "moonshotai/", searchIn: "both" },
+                  { term: "kimi",        searchIn: "title" },
+                ],
+              },
+              k3: {
+                keywords: [
+                  { term: "Kimi K3", searchIn: "both" },
+                  { term: "K3",      searchIn: "title" },
+                ],
+                substrings: [
+                  { term: "moonshotai/kimi-k3", searchIn: "both" },
+                ],
+              },
+              quantization: {
+                keywords: [
+                  {
+                    term: "quantization",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "quantized",
+                    searchIn: "both"
+                  },
+                ],
+              },
+              "intel-gpu": {
+                // Keyword search - matches whole words only (with word boundaries)
+                keywords: [
+                  {
+                    term: "B50",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "B60",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "B70",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "intel gpu",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "Arc GPU",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "BMG",
+                    searchIn: "both"
+                  },
+                ],
+              },
              // Add more label configurations here as needed
              // example: {
              //   keywords: [...],
@@ -491,4 +551,4 @@ jobs:
              issue_number: context.issue.number,
              body: message,
            });
-            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);
+            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);
@@ -173,9 +173,6 @@ venv.bak/

 # mkdocs documentation
 /site
-docs/argparse
-docs/examples/*
-!docs/examples/README.md

 # mypy
 .mypy_cache/
@@ -3,6 +3,9 @@ MD007:
 MD013: false
 MD024:
  siblings_only: true
+MD025:
+  # Allow front matter title to be different from the first heading in the document.
+  front_matter_title: ""
 MD031:
  list_items: false
 MD033: false
@@ -260,10 +260,6 @@ repos:
    files: ^docker/(Dockerfile|versions\.json)$
    pass_filenames: false
    additional_dependencies: [dockerfile-parse]
-  - id: attention-backend-docs
-    name: Check attention backend documentation is up to date
-    entry: python tools/pre_commit/generate_attention_backend_docs.py --check
-    language: python
  - id: check-boolean-context-manager
    name: Check for boolean ops in with-statements
    entry: python tools/pre_commit/check_boolean_context_manager.py
@@ -68,8 +68,8 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.11.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.11.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.13.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.13.0")
 # TORCH_NIGHTLY=1 builds run against unpinned nightly wheels, so the supported-
 # version check would always warn. Only treat it as a nightly build when the
 # value is exactly "1" (the bootstrap exports TORCH_NIGHTLY=0 by default, which
@@ -114,6 +114,11 @@ find_package(Torch REQUIRED)
 # Supported NVIDIA architectures.
 # This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
 if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
+   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.4)
+  # Rubin (10.7) can run SM100 family code, but CUDA 13.4 also supports
+  # targeting it directly.
+  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.7;11.0;12.0")
+elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
   # starting from CUDA 12.9 and Blackwell (10.0), we use family-specific targets (10.0f, 12.0f, etc)
   # to support the whole generation without specifying all sub-architectures
@@ -420,7 +425,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")

    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
      cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
-        "9.0a;10.0f;10.1f;10.3f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
+        "9.0a;10.0f;10.1f;10.3f;10.7f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
    else()
      cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
        "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
@@ -695,7 +700,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")

  # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
  endif()
@@ -815,7 +820,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
  # require CUDA 12.8 or later
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
@@ -899,7 +904,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  endif()

  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
@@ -924,7 +929,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")

  # moe_data.cu is used by all CUTLASS MoE kernels.
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
  endif()
@@ -981,7 +986,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  # SM10x/11x FP4 kernels. MXFP4 experts quantization is currently compiled
  # only in this block; SM12x has separate NVFP4 matmul/MoE kernels above.
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
@@ -1047,7 +1052,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
  # Runtime dispatch is gated in
  # vllm/v1/attention/backends/mla/cutlass_mla.py.
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(MLA_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
  endif()
@@ -1358,6 +1358,10 @@ def main():
                            profile_memory=args.profile_memory,
                            warmup_ms=args.warmup_ms,
                            prefill_backend=pb,
+                            kv_lora_rank=args.kv_lora_rank,
+                            qk_nope_head_dim=args.qk_nope_head_dim,
+                            qk_rope_head_dim=args.qk_rope_head_dim,
+                            v_head_dim=args.v_head_dim,
                        )

                        result = run_benchmark(config)
@@ -0,0 +1,176 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import statistics
+
+import torch
+from tabulate import tabulate
+
+from vllm.models.inkling.nvidia.ops import qkvr_prep
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+
+def make_inputs(tokens: int, tp_size: int, is_local: bool):
+    torch.manual_seed(0)
+    num_q_heads = 64 // tp_size
+    num_kv_heads = (16 if is_local else 8) // tp_size
+    head_dim = 128
+    d_rel = 16
+    rel_extent = 512 if is_local else 1024
+    page_size = 16
+    num_blocks = (tokens + page_size - 1) // page_size
+    q_width = num_q_heads * head_dim
+    kv_width = num_kv_heads * head_dim
+    r_width = num_q_heads * d_rel
+    device = "cuda"
+
+    qkvr = torch.randn(
+        tokens,
+        q_width + 2 * kv_width + r_width,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    k_weight = torch.randn(kv_width, 4, device=device, dtype=torch.bfloat16)
+    v_weight = torch.randn_like(k_weight)
+    q_norm_weight = torch.randn(head_dim, device=device, dtype=torch.bfloat16)
+    k_norm_weight = torch.randn_like(q_norm_weight)
+    rel_proj = torch.randn(d_rel, rel_extent, device=device, dtype=torch.bfloat16)
+    conv_cache = torch.zeros(
+        num_blocks,
+        num_kv_heads,
+        page_size,
+        2 * head_dim,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    key_cache = torch.empty(
+        num_blocks,
+        page_size,
+        num_kv_heads,
+        head_dim,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    value_cache = torch.empty_like(key_cache)
+    positions = torch.arange(tokens, device=device, dtype=torch.int64)
+    block_table = torch.arange(num_blocks, device=device, dtype=torch.int32)[None]
+    seq_idx = torch.zeros(tokens, device=device, dtype=torch.int32)
+    slots = torch.arange(tokens, device=device, dtype=torch.int64)
+    query_start = torch.zeros(tokens, device=device, dtype=torch.int32)
+    log_scaling = None
+    if not is_local:
+        effective_n = (positions + 1).to(torch.float32)
+        log_scaling = 1.0 + 0.1 * torch.log(torch.clamp(effective_n / 128000, min=1.0))
+    return (
+        qkvr,
+        k_weight,
+        v_weight,
+        q_norm_weight,
+        k_norm_weight,
+        rel_proj,
+        1e-6,
+        num_q_heads,
+        num_kv_heads,
+        head_dim,
+        d_rel,
+        conv_cache,
+        key_cache,
+        value_cache,
+        positions,
+        block_table,
+        seq_idx,
+        slots,
+        query_start,
+        slots,
+        0,
+        head_dim,
+        page_size,
+        log_scaling,
+    )
+
+
+def capture(implementation, inputs):
+    outputs = []
+
+    def run():
+        outputs[:] = implementation.fused_qkvr_prep(*inputs)
+
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        for _ in range(3):
+            run()
+    torch.cuda.current_stream().wait_stream(stream)
+    torch.accelerator.synchronize()
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        run()
+    torch.accelerator.synchronize()
+    return graph, outputs
+
+
+def time_graph(graph: torch.cuda.CUDAGraph, warmup: int, repeats: int) -> float:
+    for _ in range(warmup):
+        graph.replay()
+    torch.accelerator.synchronize()
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    start.record()
+    for _ in range(repeats):
+        graph.replay()
+    end.record()
+    end.synchronize()
+    return start.elapsed_time(end) * 1000 / repeats
+
+
+def benchmark(inputs, args) -> float:
+    graph, _ = capture(qkvr_prep, inputs)
+    return statistics.median(
+        time_graph(graph, args.warmup, args.repeats) for _ in range(args.trials)
+    )
+
+
+@torch.inference_mode()
+def main(args):
+    rows = []
+    for tp_size in args.tp_sizes:
+        for tokens in args.tokens:
+            for is_local in (True, False):
+                triton_us = benchmark(make_inputs(tokens, tp_size, is_local), args)
+                rows.append(
+                    [
+                        tp_size,
+                        tokens,
+                        "local" if is_local else "global",
+                        triton_us,
+                    ]
+                )
+
+    print("Inkling QKVR prep (CUDA graph, median latency)")
+    print(
+        tabulate(
+            rows,
+            headers=[
+                "TP",
+                "tokens",
+                "scope",
+                "Triton (us)",
+            ],
+            floatfmt=("d", "d", "", ".2f"),
+        )
+    )
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument(
+        "--tokens",
+        type=int,
+        nargs="+",
+        default=[1 << power for power in range(15)],
+    )
+    parser.add_argument("--tp-sizes", type=int, nargs="+", default=[4, 8])
+    parser.add_argument("--warmup", type=int, default=20)
+    parser.add_argument("--repeats", type=int, default=200)
+    parser.add_argument("--trials", type=int, default=5)
+    main(parser.parse_args())
@@ -154,7 +154,7 @@ def main(
                scale=scale,
                causal=True,
                alibi_slopes=None,
-                sliding_window=window_size,
+                sliding_window=window_size if sliding_window is not None else -1,
                block_table=block_tables,
                softcap=0,
                scheduler_metadata=metadata,
@@ -0,0 +1,267 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""End-to-end autoregressive decode benchmark: ReplaySSM vs the standard SSM kernel.
+
+Loads a hybrid Mamba2 model, replicates one prompt across the batch, and times a
+long greedy decode (CUDA graphs on) once with the standard kernel and once with
+ReplaySSM, then reports the per-step / throughput speedup. The two modes run in
+separate subprocesses so each gets a clean CUDA context.
+
+The FlashInfer FP4-MoE autotuner is disabled by default (it is unstable under
+CUDA-graph capture on the pre-release Blackwell FP4 path); pass
+--no-disable-flashinfer-autotune for non-FP4 models.
+
+Examples:
+    python e2e_decode_speedup.py --model-id nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
+    python e2e_decode_speedup.py --dtype auto --buffer-len 16 \
+        --model-id nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4   # B300 NVFP4
+"""
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+
+DEFAULT_PROMPT = "My cat wrote all this CUDA code for a new language model and"
+
+MODE_LABEL = {"standard": "standard", "replayssm": "ReplaySSM"}
+
+
+def parse_args():
+    p = argparse.ArgumentParser(
+        description="E2E decode speedup: ReplaySSM vs the standard SSM kernel."
+    )
+    p.add_argument("--model-id", default="nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16")
+    p.add_argument("--prompt", default=DEFAULT_PROMPT)
+    p.add_argument("--batch-size", type=int, default=256)
+    p.add_argument("--num-steps", type=int, default=1000)
+    p.add_argument("--warmup-steps", type=int, default=128)
+    p.add_argument("--repeats", type=int, default=1)
+    p.add_argument(
+        "--buffer-len", type=int, default=16, help="ReplaySSM input-buffer length."
+    )
+    p.add_argument(
+        "--dtype",
+        default="bfloat16",
+        choices=["bfloat16", "float16", "float32", "auto"],
+    )
+    p.add_argument("--gpu-memory-utilization", type=float, default=0.9)
+    p.add_argument("--max-model-len", type=int, default=None)
+    p.add_argument(
+        "--disable-flashinfer-autotune",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="Disable the FlashInfer FP4-MoE autotuner (default: on). "
+        "It is unstable under CUDA-graph capture on the "
+        "pre-release Blackwell FP4 path; pass "
+        "--no-disable-flashinfer-autotune for non-FP4 models.",
+    )
+    p.add_argument(
+        "--mamba-ssm-cache-dtype",
+        default="auto",
+        choices=["auto", "float32", "float16", "bfloat16"],
+        help="SSM state dtype (both modes). 'auto' = config-driven; "
+        "'float32' = fp32 state, 'bfloat16' = s16 state.",
+    )
+    p.add_argument(
+        "--baseline-ssm-config",
+        default="",
+        help="Pin the STANDARD baseline's SSM launch config as "
+        "'bsm,nw' via override_ssm_config (forces the in-process "
+        "engine so the override reaches the kernel). Empty = off.",
+    )
+    p.add_argument(
+        "--worker",
+        choices=["standard", "replayssm"],
+        default=None,
+        help=argparse.SUPPRESS,
+    )
+    return p.parse_args()
+
+
+def resolve_max_model_len(args) -> int:
+    if args.max_model_len is not None:
+        return args.max_model_len
+    return args.num_steps + 256
+
+
+def run_worker(args):
+    # override_ssm_config is a module global; it only reaches the model if the
+    # engine runs in-process (default V1 spawns a separate EngineCore). Force it.
+    if args.worker == "standard" and args.baseline_ssm_config:
+        os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
+
+    import torch
+
+    from vllm import LLM, SamplingParams
+
+    mode = args.worker
+    max_model_len = resolve_max_model_len(args)
+
+    llm_kwargs = dict(
+        model=args.model_id,
+        tensor_parallel_size=1,
+        dtype=args.dtype,
+        max_model_len=max_model_len,
+        trust_remote_code=True,
+        enable_prefix_caching=False,
+        enable_chunked_prefill=False,
+        max_num_seqs=args.batch_size,
+        max_num_batched_tokens=max(max_model_len, args.batch_size * 64),
+        enforce_eager=False,
+        disable_log_stats=True,
+        gpu_memory_utilization=args.gpu_memory_utilization,
+        # SSM state dtype (applies to both standard and ReplaySSM).
+        mamba_ssm_cache_dtype=args.mamba_ssm_cache_dtype,
+    )
+    if args.disable_flashinfer_autotune:
+        # FP4-MoE autotuner is unstable under CUDA-graph capture on Blackwell;
+        # re-enable (--no-disable-flashinfer-autotune) only for non-FP4 models.
+        llm_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    if mode == "replayssm":
+        llm_kwargs.update(use_replayssm=True, replayssm_buffer_len=args.buffer_len)
+
+    _ssm_cm = None
+    if mode == "standard" and args.baseline_ssm_config:
+        from vllm.model_executor.layers.mamba.ops.mamba_ssm import override_ssm_config
+
+        _bsm, _nw = (int(x) for x in args.baseline_ssm_config.split(","))
+        _ssm_cm = override_ssm_config((_bsm, _nw))
+        _ssm_cm.__enter__()  # active through LLM() graph capture + decode
+        print(
+            f"[{mode}] override_ssm_config -> (BLOCK_SIZE_M={_bsm}, num_warps={_nw})",
+            flush=True,
+        )
+
+    llm = LLM(**llm_kwargs)
+    prompts = [args.prompt] * args.batch_size
+
+    def timed_generate(n_tokens):
+        sp = SamplingParams(
+            n=1,
+            temperature=0.0,
+            ignore_eos=True,
+            min_tokens=n_tokens,
+            max_tokens=n_tokens,
+        )
+        if torch.accelerator.is_available():
+            torch.accelerator.synchronize()
+        t0 = time.perf_counter()
+        outs = llm.generate(prompts, sp, use_tqdm=False)
+        if torch.accelerator.is_available():
+            torch.accelerator.synchronize()
+        elapsed = time.perf_counter() - t0
+        produced = min(len(o.outputs[0].token_ids) for o in outs)
+        assert produced == n_tokens, f"expected {n_tokens} tokens, got {produced}"
+        return elapsed
+
+    timed_generate(args.warmup_steps)
+
+    best = None
+    for _ in range(args.repeats):
+        elapsed = timed_generate(args.num_steps)
+        tok_s = args.batch_size * args.num_steps / elapsed
+        per_step_ms = elapsed / args.num_steps * 1e3
+        print(
+            f"[{mode}] {elapsed:.3f}s  {tok_s:,.0f} tok/s  {per_step_ms:.3f} ms/step",
+            flush=True,
+        )
+        if best is None or elapsed < best["elapsed_s"]:
+            best = {
+                "mode": mode,
+                "elapsed_s": elapsed,
+                "tok_s": tok_s,
+                "per_step_ms": per_step_ms,
+            }
+
+    print("RESULT_JSON " + json.dumps(best), flush=True)
+    if _ssm_cm is not None:
+        _ssm_cm.__exit__(None, None, None)
+
+
+def run_one_mode(args, mode) -> dict:
+    cmd = [
+        sys.executable,
+        __file__,
+        "--worker",
+        mode,
+        "--model-id",
+        args.model_id,
+        "--prompt",
+        args.prompt,
+        "--batch-size",
+        str(args.batch_size),
+        "--num-steps",
+        str(args.num_steps),
+        "--warmup-steps",
+        str(args.warmup_steps),
+        "--repeats",
+        str(args.repeats),
+        "--buffer-len",
+        str(args.buffer_len),
+        "--dtype",
+        args.dtype,
+        "--gpu-memory-utilization",
+        str(args.gpu_memory_utilization),
+        "--mamba-ssm-cache-dtype",
+        args.mamba_ssm_cache_dtype,
+        "--baseline-ssm-config",
+        args.baseline_ssm_config,
+    ]
+    cmd.append(
+        "--disable-flashinfer-autotune"
+        if args.disable_flashinfer_autotune
+        else "--no-disable-flashinfer-autotune"
+    )
+    if args.max_model_len is not None:
+        cmd += ["--max-model-len", str(args.max_model_len)]
+
+    result = None
+    proc = subprocess.Popen(
+        cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1
+    )
+    for line in proc.stdout:
+        sys.stdout.write(line)
+        sys.stdout.flush()
+        if line.startswith("RESULT_JSON "):
+            result = json.loads(line[len("RESULT_JSON ") :])
+    proc.wait()
+    if proc.returncode != 0:
+        raise RuntimeError(f"mode '{mode}' worker exited with {proc.returncode}")
+    if result is None:
+        raise RuntimeError(f"mode '{mode}' produced no RESULT_JSON line")
+    return result
+
+
+def main():
+    args = parse_args()
+    if args.worker is not None:
+        run_worker(args)
+        return
+
+    print(
+        f"model={args.model_id}  batch_size={args.batch_size}  "
+        f"steps={args.num_steps}  buffer_len={args.buffer_len}  dtype={args.dtype}"
+    )
+
+    std = run_one_mode(args, "standard")
+    fla = run_one_mode(args, "replayssm")
+    speedup = std["per_step_ms"] / fla["per_step_ms"]
+
+    print()
+    header = f"{'mode':<10}{'ms/step':>12}{'tok/s':>16}{'wall (s)':>12}"
+    print(header)
+    print("-" * len(header))
+    for r in (std, fla):
+        print(
+            f"{MODE_LABEL[r['mode']]:<10}{r['per_step_ms']:>12.3f}"
+            f"{r['tok_s']:>16,.0f}{r['elapsed_s']:>12.3f}"
+        )
+    print("-" * len(header))
+    print(f"speedup (standard / ReplaySSM, per step): {speedup:.3f}x")
+
+
+if __name__ == "__main__":
+    main()
@@ -15,6 +15,7 @@ endif()
 #
 set(ENABLE_X86_ISA $ENV{VLLM_CPU_X86})
 set(ENABLE_ARM_BF16 $ENV{VLLM_CPU_ARM_BF16})
+set(ENABLE_ARM_I8MM $ENV{VLLM_CPU_ARM_I8MM})
 set(ENABLE_RVV_BF16 $ENV{VLLM_CPU_RVV_BF16})

 include_directories("${CMAKE_SOURCE_DIR}/csrc")
@@ -96,12 +97,14 @@ if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
    set(ENABLE_NUMA OFF)
    check_sysctl(hw.optional.neon ASIMD_FOUND)
    check_sysctl(hw.optional.arm.FEAT_BF16 ARM_BF16_FOUND)
+    check_sysctl(hw.optional.arm.FEAT_I8MM ARM_I8MM_FOUND)
 else()
    find_isa(${CPUINFO} "Power11" POWER11_FOUND)
    find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
    find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
    find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
    find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
+    find_isa(${CPUINFO} "i8mm" ARM_I8MM_FOUND) # Check for ARM I8MM support
    find_isa(${CPUINFO} "S390" S390_FOUND)
    find_isa(${CPUINFO} "zvfhmin" RVV_FP16_FOUND) # Check for RISC-V Vector FP16 support
    find_isa(${CPUINFO} "zvfbfmin" RVV_BF16_FOUND) # Check for RISC-V Vector BF16 support
@@ -111,6 +114,11 @@ else()
        set(ARM_BF16_FOUND ON)
        message(STATUS "ARM BF16 support enabled via VLLM_CPU_ARM_BF16 environment variable")
    endif()
+    if (ENABLE_ARM_I8MM)
+        set(ARM_I8MM_FOUND ON)
+        message(STATUS
+            "ARM I8MM support enabled via VLLM_CPU_ARM_I8MM environment variable")
+    endif()
    # Some kernels (e.g. Bianbu on Spacemit X100) do not report zvfbfmin
    # in /proc/cpuinfo despite hardware support. VLLM_CPU_RVV_BF16=1
    # overrides the detection result.
@@ -166,6 +174,11 @@ elseif (ASIMD_FOUND)
        message(WARNING "BF16 functionality is not available")
        set(MARCH_FLAGS "-march=armv8.2-a+dotprod+fp16")  
    endif()
+    if(ARM_I8MM_FOUND)
+        message(STATUS "I8MM extension detected")
+        string(APPEND MARCH_FLAGS "+i8mm")
+        add_compile_definitions(ARM_I8MM_SUPPORT)
+    endif()
    list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})     
 elseif (S390_FOUND)
    message(STATUS "S390 detected")
@@ -447,8 +460,13 @@ if (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
        "csrc/cpu/shm.cpp"
        "csrc/cpu/activation_lut_bf16.cpp"
        "csrc/cpu/cpu_tanhf_neon.hpp"
-        "csrc/cpu/cpu_fused_moe.cpp"
        ${VLLM_EXT_SRC})
+    if (ARM_BF16_FOUND)
+        set(VLLM_EXT_SRC "csrc/cpu/cpu_fused_moe.cpp" ${VLLM_EXT_SRC})
+        if (ARM_I8MM_FOUND)
+            set(VLLM_EXT_SRC "csrc/cpu/cpu_fused_moe_int8.cpp" ${VLLM_EXT_SRC})
+        endif()
+    endif()
 endif()

 if (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)	
@@ -68,6 +68,9 @@ endif()
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
    list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.0f")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+      list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.7f")
+    endif()
  else()
    list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.0a")
  endif()
@@ -60,6 +60,9 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
    # CUDA 12.9 has introduced "Family-Specific Architecture Features"
    # this supports all compute_10x family
    list(APPEND SUPPORT_ARCHS "10.0f")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+        list(APPEND SUPPORT_ARCHS "10.7f")
+    endif()
 elseif(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
    list(APPEND SUPPORT_ARCHS "10.0a")
 endif()
@@ -188,4 +191,3 @@ else()
    add_custom_target(_flashmla_C)
    add_custom_target(_flashmla_extension_C)
 endif()
-
@@ -17,7 +17,7 @@ else()
  FetchContent_Declare(
    fmha_sm100
    GIT_REPOSITORY https://github.com/vllm-project/MSA.git
-    GIT_TAG 2e63ec37a0fc29bc20f39cd1a52e0f5affc33a73
+    GIT_TAG 890aaa1a37a598ad17ccff0827fea21540d381fa
    GIT_PROGRESS TRUE
    CONFIGURE_COMMAND ""
    BUILD_COMMAND ""
@@ -55,7 +55,11 @@ message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")

 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0f" "${CUDA_ARCHS}")
-  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+    cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f;10.7f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
+  endif()
 else()
  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0a;12.1a" "${CUDA_ARCHS}")
  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0a;10.3a" "${CUDA_ARCHS}")
@@ -39,7 +39,7 @@ else()
  FetchContent_Declare(
          vllm-flash-attn
          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 168920233059c48de6199e2cda74003b2ce3d199
+          GIT_TAG ed4b7342bc8f0489dd9b649d5288867e35fc6a32
          GIT_PROGRESS TRUE
          # Don't share the vllm-flash-attn build between build types
          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
@@ -396,14 +396,24 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
  # match — e.g. SRC="12.0f" matches TGT="12.1a" since SM121 is in the SM12x
  # family. The output uses TGT's value to preserve the user's compilation flags.
  set(_CUDA_ARCHS)
+  # Resolve exact base matches before family fallbacks so a generic entry such
+  # as 10.0f cannot consume a 10.7 target that has a 10.7f source entry.
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "[af]$")
+      string(REGEX REPLACE "[af]$" "" _base "${_arch}")
+      if("${_base}" IN_LIST _TGT_CUDA_ARCHS)
+        list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
+        list(APPEND _CUDA_ARCHS "${_arch}")
+      endif()
+    endif()
+  endforeach()
+
  foreach(_arch ${_SRC_CUDA_ARCHS})
    if(_arch MATCHES "[af]$")
      list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
      string(REGEX REPLACE "[af]$" "" _base "${_arch}")
-      if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
-        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
-        list(APPEND _CUDA_ARCHS "${_arch}")
-      elseif("${_base}a" IN_LIST _TGT_CUDA_ARCHS)
+      if("${_base}a" IN_LIST _TGT_CUDA_ARCHS)
        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}a")
        list(APPEND _CUDA_ARCHS "${_base}a")
      elseif("${_base}f" IN_LIST _TGT_CUDA_ARCHS)
@@ -487,7 +497,7 @@ endfunction()

 function(cuda_archs_sm90plus OUT_CUDA_ARCHS TGT_CUDA_ARCHS)
  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0f;10.7f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
  else()
    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${TGT_CUDA_ARCHS}")
  endif()
@@ -172,4 +172,15 @@

 #endif  // __riscv_v

+// Power VSX
+#ifdef __powerpc__
+  // FP32Vec16::exp() in cpu_types_vsx.hpp delegates to FP32Vec8::exp(), which
+  // implements a vectorised 5-term minimax polynomial using VSX intrinsics.
+  #define DEFINE_FAST_EXP                                                     \
+    auto fast_exp = [&](const vec_op::FP32Vec16& vec)                         \
+                        __attribute__((always_inline)) { return vec.exp(); }; \
+    auto fast_exp_f16 = fast_exp;
+
+#endif  // __powerpc__
+
 #endif
@@ -1,5 +1,6 @@
 #include "cpu/cpu_types.hpp"
 #include "cpu/utils.hpp"
+#include "cpu/cpu_fused_moe_activations.hpp"
 #include "cpu/micro_gemm/cpu_micro_gemm_vec.hpp"
 #include "cpu/cpu_arch_macros.h"

@@ -43,193 +44,9 @@
  }()

 namespace {
-enum class FusedMOEAct {
-  SiluAndMul,
-  SwigluOAIAndMul,
-  GeluAndMul,
-  GeluTanhAndMul,
-};

-FusedMOEAct get_act_type(const std::string& act) {
-  if (act == "silu") {
-    return FusedMOEAct::SiluAndMul;
-  } else if (act == "swigluoai") {
-    return FusedMOEAct::SwigluOAIAndMul;
-  } else if (act == "gelu") {
-    return FusedMOEAct::GeluAndMul;
-  } else if (act == "gelu_tanh") {
-    return FusedMOEAct::GeluTanhAndMul;
-  } else {
-    TORCH_CHECK(false, "Invalid act type: " + act);
-  }
-}
-
-template <typename scalar_t>
-void swigluoai_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                       const int32_t m_size, const int32_t n_size,
-                       const int32_t input_stride,
-                       const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-#if !defined(__aarch64__)
-  // For GPT-OSS interleaved gate-up weights
-  alignas(64) static int32_t index[16] = {0,  2,  4,  6,  8,  10, 12, 14,
-                                          16, 18, 20, 22, 24, 26, 28, 30};
-  vec_op::INT32Vec16 index_vec(index);
-#endif
-  vec_op::FP32Vec16 gate_up_max_vec(7.0);
-  vec_op::FP32Vec16 up_min_vec(-7.0);
-  vec_op::FP32Vec16 alpha_vec(1.702);
-  vec_op::FP32Vec16 one_vec(1.0);
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < n_size; n += 32) {
-      // Note: AdvSIMD does not support gather loads
-#if defined(__aarch64__)
-      vec_op::FP32Vec16 gate_vec(vec_op::uninit);
-      vec_op::FP32Vec16 up_vec(vec_op::uninit);
-      vec_op::FP32Vec16::load_even_odd(input + n, gate_vec, up_vec);
-#else
-      vec_op::FP32Vec16 gate_vec(input + n, index_vec);
-      vec_op::FP32Vec16 up_vec(input + n + 1, index_vec);
-#endif
-      gate_vec = gate_vec.min(gate_up_max_vec);
-      up_vec = up_vec.clamp(up_min_vec, gate_up_max_vec);
-      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec * alpha_vec));
-      auto glu = gate_vec * sigmoid_vec;
-      auto gated_output_fp32 = (one_vec + up_vec) * glu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n / 2);
-    }
-    input += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void silu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                  const int32_t m_size, const int32_t n_size,
-                  const int32_t input_stride, const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec));
-      auto silu = gate_vec * sigmoid_vec;
-      auto gated_output_fp32 = up_vec * silu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void gelu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                  const int32_t m_size, const int32_t n_size,
-                  const int32_t input_stride, const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-  vec_op::FP32Vec16 w1_vec(M_SQRT1_2);
-  vec_op::FP32Vec16 w2_vec(0.5);
-  alignas(64) float temp[16];
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto er_input_vec = gate_vec * w1_vec;
-
-      er_input_vec.save(temp);
-      for (int32_t i = 0; i < 16; ++i) {
-        temp[i] = std::erf(temp[i]);
-      }
-      vec_op::FP32Vec16 er_vec(temp);
-      auto gelu = gate_vec * w2_vec * (one_vec + er_vec);
-      auto gated_output_fp32 = up_vec * gelu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                       const int32_t m_size, const int32_t n_size,
-                       const int32_t input_stride,
-                       const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-  vec_op::FP32Vec16 w1_vec(0.7978845608028654);
-  vec_op::FP32Vec16 w2_vec(0.5);
-  vec_op::FP32Vec16 w3_vec(0.044715);
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto gate_pow3_vec = gate_vec * gate_vec * gate_vec;
-      auto inner_vec = w1_vec * (gate_vec + w3_vec * gate_pow3_vec);
-      // Note: can't use fast_exp form because diffusiongemma will generate
-      // wrong results
-      auto tanh_vec = inner_vec.tanh();
-      auto gelu_tanh = gate_vec * w2_vec * (one_vec + tanh_vec);
-      auto gated_output_fp32 = up_vec * gelu_tanh;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-FORCE_INLINE void apply_gated_act(const FusedMOEAct act,
-                                  float* __restrict__ input,
-                                  scalar_t* __restrict__ output,
-                                  const int32_t m, const int32_t n,
-                                  const int32_t input_stride,
-                                  const int32_t output_stride) {
-  switch (act) {
-    case FusedMOEAct::SwigluOAIAndMul:
-      swigluoai_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::SiluAndMul:
-      silu_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::GeluAndMul:
-      gelu_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::GeluTanhAndMul:
-      gelu_tanh_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    default:
-      TORCH_CHECK(false, "Unsupported act type.");
-  }
-}
+using cpu_fused_moe_utils::apply_gated_act;
+using cpu_fused_moe_utils::FusedMOEAct;

 template <typename scalar_t, typename gemm_t>
 void prepack_moe_weight_impl(scalar_t* __restrict__ weight_ptr,
@@ -817,6 +634,7 @@ void fused_moe_impl(scalar_t* __restrict__ output, scalar_t* __restrict__ input,
    }
  }
 }
+
 }  // namespace

 void prepack_moe_weight(
@@ -864,7 +682,7 @@ void cpu_fused_moe(
  const int32_t input_size_2 = w2.size(2);
  const int32_t output_size_2 = w2.size(1);
  const int32_t topk_num = topk_id.size(1);
-  const FusedMOEAct act_type = get_act_type(act);
+  const FusedMOEAct act_type = cpu_fused_moe_utils::get_act_type(act);
  cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
  TORCH_CHECK(!skip_weighted || topk_num == 1,
              "skip_weighted is only supported for topk=1 on CPU");
@@ -0,0 +1,204 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#ifndef CPU_FUSED_MOE_ACTIVATIONS_HPP
+#define CPU_FUSED_MOE_ACTIVATIONS_HPP
+
+#include <cmath>
+#include <cstdint>
+#include <string>
+
+#include "cpu/cpu_arch_macros.h"
+#include "cpu/utils.hpp"
+
+namespace cpu_fused_moe_utils {
+enum class FusedMOEAct {
+  SiluAndMul,
+  SwigluOAIAndMul,
+  GeluAndMul,
+  GeluTanhAndMul,
+};
+
+inline FusedMOEAct get_act_type(const std::string& act) {
+  if (act == "silu") {
+    return FusedMOEAct::SiluAndMul;
+  } else if (act == "swigluoai") {
+    return FusedMOEAct::SwigluOAIAndMul;
+  } else if (act == "gelu") {
+    return FusedMOEAct::GeluAndMul;
+  } else if (act == "gelu_tanh") {
+    return FusedMOEAct::GeluTanhAndMul;
+  } else {
+    TORCH_CHECK(false, "Invalid act type: " + act);
+  }
+}
+
+template <typename scalar_t>
+void swigluoai_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                       const int32_t m_size, const int32_t n_size,
+                       const int32_t input_stride,
+                       const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+#if !defined(__aarch64__)
+  // For GPT-OSS interleaved gate-up weights
+  alignas(64) static int32_t index[16] = {0,  2,  4,  6,  8,  10, 12, 14,
+                                          16, 18, 20, 22, 24, 26, 28, 30};
+  vec_op::INT32Vec16 index_vec(index);
+#endif
+  vec_op::FP32Vec16 gate_up_max_vec(7.0);
+  vec_op::FP32Vec16 up_min_vec(-7.0);
+  vec_op::FP32Vec16 alpha_vec(1.702);
+  vec_op::FP32Vec16 one_vec(1.0);
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < n_size; n += 32) {
+      // Note: AdvSIMD does not support gather loads
+#if defined(__aarch64__)
+      vec_op::FP32Vec16 gate_vec(vec_op::uninit);
+      vec_op::FP32Vec16 up_vec(vec_op::uninit);
+      vec_op::FP32Vec16::load_even_odd(input + n, gate_vec, up_vec);
+#else
+      vec_op::FP32Vec16 gate_vec(input + n, index_vec);
+      vec_op::FP32Vec16 up_vec(input + n + 1, index_vec);
+#endif
+      gate_vec = gate_vec.min(gate_up_max_vec);
+      up_vec = up_vec.clamp(up_min_vec, gate_up_max_vec);
+      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec * alpha_vec));
+      auto glu = gate_vec * sigmoid_vec;
+      auto gated_output_fp32 = (one_vec + up_vec) * glu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n / 2);
+    }
+    input += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void silu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                  const int32_t m_size, const int32_t n_size,
+                  const int32_t input_stride, const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec));
+      auto silu = gate_vec * sigmoid_vec;
+      auto gated_output_fp32 = up_vec * silu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void gelu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                  const int32_t m_size, const int32_t n_size,
+                  const int32_t input_stride, const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+  vec_op::FP32Vec16 w1_vec(M_SQRT1_2);
+  vec_op::FP32Vec16 w2_vec(0.5);
+  alignas(64) float temp[16];
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto er_input_vec = gate_vec * w1_vec;
+
+      er_input_vec.save(temp);
+      for (int32_t i = 0; i < 16; ++i) {
+        temp[i] = std::erf(temp[i]);
+      }
+      vec_op::FP32Vec16 er_vec(temp);
+      auto gelu = gate_vec * w2_vec * (one_vec + er_vec);
+      auto gated_output_fp32 = up_vec * gelu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                       const int32_t m_size, const int32_t n_size,
+                       const int32_t input_stride,
+                       const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+  vec_op::FP32Vec16 w1_vec(0.7978845608028654);
+  vec_op::FP32Vec16 w2_vec(0.5);
+  vec_op::FP32Vec16 w3_vec(0.044715);
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto gate_pow3_vec = gate_vec * gate_vec * gate_vec;
+      auto inner_vec = w1_vec * (gate_vec + w3_vec * gate_pow3_vec);
+      // Note: can't use fast_exp form because diffusiongemma will generate
+      // wrong results
+      auto tanh_vec = inner_vec.tanh();
+      auto gelu_tanh = gate_vec * w2_vec * (one_vec + tanh_vec);
+      auto gated_output_fp32 = up_vec * gelu_tanh;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+FORCE_INLINE void apply_gated_act(const FusedMOEAct act,
+                                  float* __restrict__ input,
+                                  scalar_t* __restrict__ output,
+                                  const int32_t m, const int32_t n,
+                                  const int32_t input_stride,
+                                  const int32_t output_stride) {
+  switch (act) {
+    case FusedMOEAct::SwigluOAIAndMul:
+      swigluoai_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::SiluAndMul:
+      silu_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::GeluAndMul:
+      gelu_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::GeluTanhAndMul:
+      gelu_tanh_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    default:
+      TORCH_CHECK(false, "Unsupported act type.");
+  }
+}
+}  // namespace cpu_fused_moe_utils
+
+#endif
@@ -0,0 +1,647 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#include "cpu/cpu_arch_macros.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <cstring>
+#include <optional>
+#include <string>
+
+#include "cpu/cpu_fused_moe_activations.hpp"
+#include "cpu/cpu_types.hpp"
+#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
+#include "cpu/utils.hpp"
+
+#if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT)
+  #include "cpu/micro_gemm/cpu_micro_gemm_int8_neon.hpp"
+  #define NEON_DISPATCH(SCALAR_TYPE, ...)                                   \
+    case cpu_utils::ISA::NEON: {                                            \
+      using gemm_t =                                                        \
+          cpu_micro_gemm::MicroGemmINT8<cpu_utils::ISA::NEON, SCALAR_TYPE>; \
+      return __VA_ARGS__();                                                 \
+    }
+#else
+  #define NEON_DISPATCH(SCALAR_TYPE, ...) case cpu_utils::ISA::NEON:
+#endif
+
+#define CPU_INT8_ISA_DISPATCH_IMPL(ISA_TYPE, SCALAR_TYPE, ...) \
+  [&] {                                                        \
+    switch (ISA_TYPE) {                                        \
+      NEON_DISPATCH(SCALAR_TYPE, __VA_ARGS__)                  \
+      default: {                                               \
+        TORCH_CHECK(false, "Invalid CPU ISA type.");           \
+      }                                                        \
+    }                                                          \
+  }()
+
+namespace {
+using cpu_fused_moe_utils::apply_gated_act;
+using cpu_fused_moe_utils::FusedMOEAct;
+
+template <typename gemm_t>
+void prepack_moe_weight_int8_impl(const int8_t* __restrict__ weight_ptr,
+                                  int8_t* __restrict__ packed_weight_ptr,
+                                  const int32_t expert_num,
+                                  const int32_t output_size,
+                                  const int32_t input_size,
+                                  const int64_t expert_stride) {
+#pragma omp parallel for
+  for (int32_t e_idx = 0; e_idx < expert_num; ++e_idx) {
+    gemm_t::pack_weight(weight_ptr + expert_stride * e_idx,
+                        packed_weight_ptr + expert_stride * e_idx, output_size,
+                        input_size);
+  }
+}
+
+// INT8 MoE kernel, based on the original BF16 kernel in cpu_fused_moe.cpp
+template <typename scalar_t, typename gemm_t>
+void fused_moe_int8_impl(
+    scalar_t* __restrict__ output, const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ w13, const int8_t* __restrict__ w2,
+    const float* __restrict__ w13_scales, const float* __restrict__ w2_scales,
+    scalar_t* __restrict__ w13_bias, scalar_t* __restrict__ w2_bias,
+    const float* __restrict__ topk_weights, const int32_t* __restrict__ topk_id,
+    const FusedMOEAct act_type, const int32_t token_num,
+    const int32_t expert_num, const int32_t topk_num,
+    const int32_t input_size_13, const int32_t output_size_13,
+    const int32_t input_size_2, const int32_t output_size_2,
+    const bool skip_weighted) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  constexpr int32_t gemm_n_tile_size = gemm_t::NSize;
+  constexpr int32_t gemm_m_tile_size = gemm_t::MaxMSize;
+  constexpr int32_t min_w13_n_tile_size = 2 * gemm_n_tile_size;
+
+  TORCH_CHECK_EQ(input_size_13 % gemm_t::K, 0);
+  TORCH_CHECK_EQ(input_size_2 % gemm_t::K, 0);
+  TORCH_CHECK_EQ(output_size_13 % min_w13_n_tile_size, 0);
+  TORCH_CHECK_EQ(output_size_2 % gemm_n_tile_size, 0);
+  TORCH_CHECK_EQ(output_size_13 / 2, input_size_2);
+
+  const int32_t thread_num = cpu_utils::get_max_threads();
+  const int32_t w13_input_buffer_size = cpu_utils::round_up<64>(
+      gemm_m_tile_size * input_size_13 * sizeof(int8_t));
+  const int32_t w2_input_buffer_size =
+      cpu_utils::round_up<64>(gemm_m_tile_size * input_size_2 * sizeof(int8_t));
+
+  const int32_t w13_n_tile_size = [&]() {
+    const int64_t cache_size = cpu_utils::get_available_l2_size();
+    const int32_t n_size_cache_limit =
+        (cache_size - w13_input_buffer_size) /
+        (gemm_m_tile_size * sizeof(float) + input_size_13 * sizeof(int8_t));
+    const int32_t n_size_thread_limit =
+        output_size_13 / std::max(1, thread_num / topk_num);
+    const int32_t n_size = cpu_utils::round_down<min_w13_n_tile_size>(
+        std::min(n_size_cache_limit, n_size_thread_limit));
+    return std::max(n_size, min_w13_n_tile_size);
+  }();
+
+  const int32_t w2_n_tile_size = [&]() {
+    const int64_t cache_size = cpu_utils::get_available_l2_size();
+    const int32_t n_size_cache_limit =
+        (cache_size - w2_input_buffer_size) / (input_size_2 * sizeof(int8_t));
+    const int32_t n_size_thread_limit =
+        output_size_2 / std::max(1, thread_num / topk_num);
+    const int32_t n_size = cpu_utils::round_down<gemm_n_tile_size>(
+        std::min(n_size_cache_limit, n_size_thread_limit));
+    return std::max(n_size, gemm_n_tile_size);
+  }();
+
+  int32_t common_buffer_offset = 0;
+  const int32_t token_num_per_group_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(expert_num * sizeof(int32_t));
+  const int32_t cu_token_num_per_group_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>((expert_num + 1) * sizeof(int32_t));
+  const int32_t expanded_token_num = token_num * topk_num;
+  const int32_t expand_token_id_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(int32_t));
+  const int32_t expand_token_id_index_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(int32_t));
+  const int32_t input_quant_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(token_num * input_size_13 * sizeof(int8_t));
+  const int32_t input_scale_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(token_num * sizeof(float));
+  const int32_t w13_gemm_output_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(
+      expanded_token_num * input_size_2 * sizeof(scalar_t));
+  const int32_t w13_output_scale_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(float));
+  const int32_t w2_gemm_output_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(
+      expanded_token_num * output_size_2 * sizeof(float));
+
+  int32_t gemm_thread_buffer_offset = 0;
+  const int32_t gemm_input_buffer_offset = gemm_thread_buffer_offset;
+  gemm_thread_buffer_offset +=
+      std::max(w13_input_buffer_size, w2_input_buffer_size);
+  const int32_t gemm_output_buffer_offset = gemm_thread_buffer_offset;
+  gemm_thread_buffer_offset += cpu_utils::round_up<64>(
+      gemm_m_tile_size * std::max(w13_n_tile_size, w2_n_tile_size) *
+      sizeof(int32_t));
+
+  const int32_t ws_output_buffer_offset = 0;
+  const int32_t ws_thread_buffer_size =
+      cpu_utils::round_up<64>(output_size_2 * sizeof(float));
+  const int32_t thread_buffer_size =
+      std::max(gemm_thread_buffer_offset, ws_thread_buffer_size);
+  const int32_t buffer_size =
+      common_buffer_offset + thread_buffer_size * thread_num;
+  cpu_utils::ScratchPadManager::get_scratchpad_manager()->realloc(buffer_size);
+  uint8_t* common_buffer_start =
+      cpu_utils::ScratchPadManager::get_scratchpad_manager()
+          ->get_data<uint8_t>();
+  uint8_t* thread_buffer_start = common_buffer_start + common_buffer_offset;
+
+  int32_t* __restrict__ token_num_per_group_buffer = reinterpret_cast<int32_t*>(
+      common_buffer_start + token_num_per_group_buffer_offset);
+  int32_t* __restrict__ cu_token_num_per_group_buffer =
+      reinterpret_cast<int32_t*>(common_buffer_start +
+                                 cu_token_num_per_group_buffer_offset);
+  int32_t* __restrict__ expand_token_id_buffer = reinterpret_cast<int32_t*>(
+      common_buffer_start + expand_token_id_buffer_offset);
+  int32_t* __restrict__ expand_token_id_index_buffer =
+      reinterpret_cast<int32_t*>(common_buffer_start +
+                                 expand_token_id_index_buffer_offset);
+  int8_t* __restrict__ input_quant_buffer = reinterpret_cast<int8_t*>(
+      common_buffer_start + input_quant_buffer_offset);
+  float* __restrict__ input_scale_buffer =
+      reinterpret_cast<float*>(common_buffer_start + input_scale_buffer_offset);
+
+  std::memset(token_num_per_group_buffer, 0, expert_num * sizeof(int32_t));
+  for (int32_t i = 0; i < expanded_token_num; ++i) {
+    ++token_num_per_group_buffer[topk_id[i]];
+  }
+
+  int32_t token_num_sum = 0;
+  cu_token_num_per_group_buffer[0] = 0;
+  int32_t* token_index_buffer = cu_token_num_per_group_buffer + 1;
+  for (int32_t i = 0; i < expert_num; ++i) {
+    token_index_buffer[i] = token_num_sum;
+    token_num_sum += token_num_per_group_buffer[i];
+  }
+
+  for (int32_t i = 0; i < token_num; ++i) {
+    const int32_t* curr_topk_id = topk_id + i * topk_num;
+    int32_t* curr_index_buffer = expand_token_id_index_buffer + i * topk_num;
+    for (int32_t j = 0; j < topk_num; ++j) {
+      const int32_t curr_expert_id = curr_topk_id[j];
+      const int32_t curr_index = token_index_buffer[curr_expert_id]++;
+      expand_token_id_buffer[curr_index] = i;
+      curr_index_buffer[j] = curr_index;
+    }
+  }
+
+// quantize inputs
+#pragma omp parallel for
+  for (int32_t token_idx = 0; token_idx < token_num; ++token_idx) {
+    gemm_t::quantize_row(input + token_idx * input_size_13,
+                         input_quant_buffer + token_idx * input_size_13,
+                         input_scale_buffer[token_idx], input_size_13);
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+// w13 GEMM + act
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      const int32_t task_num_per_expert =
+          (output_size_13 + w13_n_tile_size - 1) / w13_n_tile_size;
+      const int32_t task_num = task_num_per_expert * expert_num;
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      int8_t* __restrict__ gemm_input_buffer =
+          reinterpret_cast<int8_t*>(thread_buffer + gemm_input_buffer_offset);
+      float* __restrict__ gemm_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + gemm_output_buffer_offset);
+      auto* __restrict__ w13_gemm_output_buffer = reinterpret_cast<scalar_t*>(
+          common_buffer_start + w13_gemm_output_buffer_offset);
+      gemm_t gemm;
+
+      const int32_t w13_n_group_stride =
+          gemm_t::WeightOCGroupSize * input_size_13;
+      const int32_t w13_n_tile_stride = gemm_n_tile_size * input_size_13;
+
+      for (;;) {
+        const int32_t task_id = counter_ptr->acquire_counter();
+        if (task_id >= task_num) {
+          break;
+        }
+        const int32_t curr_expert_id = task_id / task_num_per_expert;
+        const int32_t curr_output_group_id = task_id % task_num_per_expert;
+        const int32_t curr_token_num =
+            token_num_per_group_buffer[curr_expert_id];
+        if (curr_token_num == 0) {
+          continue;
+        }
+
+        const int32_t actual_n_tile_size =
+            std::min(w13_n_tile_size,
+                     output_size_13 - curr_output_group_id * w13_n_tile_size);
+        const int32_t* __restrict__ curr_expand_token_id_buffer =
+            expand_token_id_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id];
+        scalar_t* __restrict__ curr_w13_gemm_output_buffer =
+            w13_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * input_size_2 +
+            curr_output_group_id * w13_n_tile_size / 2;
+
+        const int8_t* w13_weight_ptr_0 = nullptr;
+        const int8_t* w13_weight_ptr_1 = nullptr;
+        const float* w13_scale_ptr_0 = nullptr;
+        const float* w13_scale_ptr_1 = nullptr;
+        scalar_t* w13_bias_ptr_0 = nullptr;
+        scalar_t* w13_bias_ptr_1 = nullptr;
+        if (act_type == FusedMOEAct::SwigluOAIAndMul) {
+          const int32_t output_offset = curr_output_group_id * w13_n_tile_size;
+          w13_weight_ptr_0 = w13 +
+                             curr_expert_id * input_size_13 * output_size_13 +
+                             output_offset * input_size_13;
+          w13_weight_ptr_1 =
+              w13_weight_ptr_0 + actual_n_tile_size / 2 * input_size_13;
+          w13_scale_ptr_0 =
+              w13_scales + curr_expert_id * output_size_13 + output_offset;
+          w13_scale_ptr_1 = w13_scale_ptr_0 + actual_n_tile_size / 2;
+          if (w13_bias != nullptr) {
+            w13_bias_ptr_0 =
+                w13_bias + curr_expert_id * output_size_13 + output_offset;
+            w13_bias_ptr_1 = w13_bias_ptr_0 + actual_n_tile_size / 2;
+          }
+        } else {
+          const int32_t output_offset =
+              curr_output_group_id * (w13_n_tile_size / 2);
+          w13_weight_ptr_0 = w13 +
+                             curr_expert_id * input_size_13 * output_size_13 +
+                             output_offset * input_size_13;
+          w13_weight_ptr_1 =
+              w13_weight_ptr_0 + output_size_13 / 2 * input_size_13;
+          w13_scale_ptr_0 =
+              w13_scales + curr_expert_id * output_size_13 + output_offset;
+          w13_scale_ptr_1 = w13_scale_ptr_0 + output_size_13 / 2;
+          if (w13_bias != nullptr) {
+            w13_bias_ptr_0 =
+                w13_bias + curr_expert_id * output_size_13 + output_offset;
+            w13_bias_ptr_1 = w13_bias_ptr_0 + output_size_13 / 2;
+          }
+        }
+
+        for (int32_t token_idx = 0; token_idx < curr_token_num;
+             token_idx += gemm_m_tile_size) {
+          const int32_t actual_token_num =
+              std::min(gemm_m_tile_size, curr_token_num - token_idx);
+          const int8_t* input_rows[gemm_m_tile_size];
+          alignas(64) float input_scales[gemm_m_tile_size];
+          // gather and pack
+          for (int32_t i = 0; i < actual_token_num; ++i) {
+            const int32_t curr_token_id = curr_expand_token_id_buffer[i];
+            input_rows[i] = input_quant_buffer + curr_token_id * input_size_13;
+            input_scales[i] = input_scale_buffer[curr_token_id];
+          }
+          gemm_t::pack_input_from_rows(input_rows, gemm_input_buffer,
+                                       actual_token_num, input_size_13);
+          curr_expand_token_id_buffer += actual_token_num;
+
+          const int8_t* w13_weight_ptr_0_iter = w13_weight_ptr_0;
+          const int8_t* w13_weight_ptr_1_iter = w13_weight_ptr_1;
+          const float* w13_scale_ptr_0_iter = w13_scale_ptr_0;
+          const float* w13_scale_ptr_1_iter = w13_scale_ptr_1;
+          scalar_t* w13_bias_ptr_0_iter = w13_bias_ptr_0;
+          scalar_t* w13_bias_ptr_1_iter = w13_bias_ptr_1;
+          float* w13_output_buffer_0_iter = gemm_output_buffer;
+          float* w13_output_buffer_1_iter =
+              gemm_output_buffer + actual_n_tile_size / 2;
+
+          for (int32_t i = 0; i < actual_n_tile_size;
+               i += min_w13_n_tile_size) {
+            auto* output_0_int32 =
+                reinterpret_cast<int32_t*>(w13_output_buffer_0_iter);
+            gemm.gemm(gemm_input_buffer, w13_weight_ptr_0_iter, output_0_int32,
+                      actual_token_num, input_size_13, w13_n_group_stride,
+                      actual_n_tile_size);
+            gemm_t::dequantize_tile(output_0_int32, w13_output_buffer_0_iter,
+                                    input_scales, w13_scale_ptr_0_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    actual_n_tile_size);
+            if (w13_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  w13_output_buffer_0_iter, w13_output_buffer_0_iter,
+                  w13_bias_ptr_0_iter, actual_token_num, actual_n_tile_size,
+                  actual_n_tile_size);
+              w13_bias_ptr_0_iter += gemm_n_tile_size;
+            }
+
+            auto* output_1_int32 =
+                reinterpret_cast<int32_t*>(w13_output_buffer_1_iter);
+            gemm.gemm(gemm_input_buffer, w13_weight_ptr_1_iter, output_1_int32,
+                      actual_token_num, input_size_13, w13_n_group_stride,
+                      actual_n_tile_size);
+            gemm_t::dequantize_tile(output_1_int32, w13_output_buffer_1_iter,
+                                    input_scales, w13_scale_ptr_1_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    actual_n_tile_size);
+            if (w13_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  w13_output_buffer_1_iter, w13_output_buffer_1_iter,
+                  w13_bias_ptr_1_iter, actual_token_num, actual_n_tile_size,
+                  actual_n_tile_size);
+              w13_bias_ptr_1_iter += gemm_n_tile_size;
+            }
+
+            w13_weight_ptr_0_iter += w13_n_tile_stride;
+            w13_weight_ptr_1_iter += w13_n_tile_stride;
+            w13_scale_ptr_0_iter += gemm_n_tile_size;
+            w13_scale_ptr_1_iter += gemm_n_tile_size;
+            w13_output_buffer_0_iter += gemm_n_tile_size;
+            w13_output_buffer_1_iter += gemm_n_tile_size;
+          }
+
+          apply_gated_act(act_type, gemm_output_buffer,
+                          curr_w13_gemm_output_buffer, actual_token_num,
+                          actual_n_tile_size, actual_n_tile_size, input_size_2);
+          curr_w13_gemm_output_buffer += gemm_m_tile_size * input_size_2;
+        }
+      }
+    }
+  }
+
+  auto* __restrict__ w13_gemm_output_buffer = reinterpret_cast<scalar_t*>(
+      common_buffer_start + w13_gemm_output_buffer_offset);
+  float* __restrict__ w13_output_scale_buffer = reinterpret_cast<float*>(
+      common_buffer_start + w13_output_scale_buffer_offset);
+
+// quantize w2 inputs - in place
+#pragma omp parallel for
+  for (int32_t token_idx = 0; token_idx < expanded_token_num; ++token_idx) {
+    scalar_t* input_row = w13_gemm_output_buffer + token_idx * input_size_2;
+    int8_t* output_row = reinterpret_cast<int8_t*>(input_row);
+    gemm_t::quantize_row(input_row, output_row,
+                         w13_output_scale_buffer[token_idx], input_size_2);
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+// w2 gemm
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      const int32_t task_num_per_expert =
+          (output_size_2 + w2_n_tile_size - 1) / w2_n_tile_size;
+      const int32_t task_num = task_num_per_expert * expert_num;
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      int8_t* __restrict__ gemm_input_buffer =
+          reinterpret_cast<int8_t*>(thread_buffer + gemm_input_buffer_offset);
+      float* __restrict__ gemm_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + gemm_output_buffer_offset);
+      float* __restrict__ w2_gemm_output_buffer = reinterpret_cast<float*>(
+          common_buffer_start + w2_gemm_output_buffer_offset);
+      gemm_t gemm;
+
+      const int32_t w2_n_group_stride =
+          gemm_t::WeightOCGroupSize * input_size_2;
+      const int32_t w2_n_tile_stride = gemm_n_tile_size * input_size_2;
+
+      for (;;) {
+        const int32_t task_id = counter_ptr->acquire_counter();
+        if (task_id >= task_num) {
+          break;
+        }
+        const int32_t curr_expert_id = task_id / task_num_per_expert;
+        const int32_t curr_output_group_id = task_id % task_num_per_expert;
+        const int32_t curr_token_num =
+            token_num_per_group_buffer[curr_expert_id];
+        if (curr_token_num == 0) {
+          continue;
+        }
+
+        const int32_t actual_n_tile_size =
+            std::min(w2_n_tile_size,
+                     output_size_2 - curr_output_group_id * w2_n_tile_size);
+        scalar_t* __restrict__ curr_w13_gemm_output_buffer =
+            w13_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * input_size_2;
+        float* __restrict__ curr_w13_output_scale_buffer =
+            w13_output_scale_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id];
+        float* __restrict__ curr_w2_gemm_output_buffer =
+            w2_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * output_size_2 +
+            curr_output_group_id * w2_n_tile_size;
+        const int8_t* __restrict__ w2_weight_ptr =
+            w2 + curr_expert_id * output_size_2 * input_size_2 +
+            curr_output_group_id * w2_n_tile_size * input_size_2;
+        const float* __restrict__ w2_scale_ptr =
+            w2_scales + curr_expert_id * output_size_2 +
+            curr_output_group_id * w2_n_tile_size;
+        scalar_t* w2_bias_ptr = nullptr;
+        if (w2_bias != nullptr) {
+          w2_bias_ptr = w2_bias + curr_expert_id * output_size_2 +
+                        curr_output_group_id * w2_n_tile_size;
+        }
+
+        for (int32_t token_idx = 0; token_idx < curr_token_num;
+             token_idx += gemm_m_tile_size) {
+          const int32_t actual_token_num =
+              std::min(gemm_m_tile_size, curr_token_num - token_idx);
+          const int8_t* input_rows[gemm_m_tile_size];
+          alignas(64) float input_scales[gemm_m_tile_size];
+          for (int32_t i = 0; i < actual_token_num; ++i) {
+            input_rows[i] = reinterpret_cast<const int8_t*>(
+                curr_w13_gemm_output_buffer + i * input_size_2);
+            input_scales[i] = curr_w13_output_scale_buffer[i];
+          }
+          gemm_t::pack_input_from_rows(input_rows, gemm_input_buffer,
+                                       actual_token_num, input_size_2);
+
+          const int8_t* w2_weight_ptr_iter = w2_weight_ptr;
+          const float* w2_scale_ptr_iter = w2_scale_ptr;
+          scalar_t* w2_bias_ptr_iter = w2_bias_ptr;
+          float* curr_w2_gemm_output_buffer_iter = curr_w2_gemm_output_buffer;
+          for (int32_t i = 0; i < actual_n_tile_size; i += gemm_n_tile_size) {
+            auto* output_int32 = reinterpret_cast<int32_t*>(gemm_output_buffer);
+            gemm.gemm(gemm_input_buffer, w2_weight_ptr_iter, output_int32,
+                      actual_token_num, input_size_2, w2_n_group_stride,
+                      gemm_n_tile_size);
+            gemm_t::dequantize_tile(output_int32, gemm_output_buffer,
+                                    input_scales, w2_scale_ptr_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    gemm_n_tile_size);
+            if (w2_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  gemm_output_buffer, gemm_output_buffer, w2_bias_ptr_iter,
+                  actual_token_num, gemm_n_tile_size, gemm_n_tile_size);
+              w2_bias_ptr_iter += gemm_n_tile_size;
+            }
+            for (int32_t m_idx = 0; m_idx < actual_token_num; ++m_idx) {
+              std::memcpy(
+                  curr_w2_gemm_output_buffer_iter + m_idx * output_size_2,
+                  gemm_output_buffer + m_idx * gemm_n_tile_size,
+                  gemm_n_tile_size * sizeof(float));
+            }
+
+            w2_weight_ptr_iter += w2_n_tile_stride;
+            w2_scale_ptr_iter += gemm_n_tile_size;
+            curr_w2_gemm_output_buffer_iter += gemm_n_tile_size;
+          }
+
+          curr_w13_gemm_output_buffer += gemm_m_tile_size * input_size_2;
+          curr_w13_output_scale_buffer += gemm_m_tile_size;
+          curr_w2_gemm_output_buffer += gemm_m_tile_size * output_size_2;
+        }
+      }
+    }
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      float* __restrict__ ws_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + ws_output_buffer_offset);
+      float* __restrict__ w2_gemm_output_buffer = reinterpret_cast<float*>(
+          common_buffer_start + w2_gemm_output_buffer_offset);
+
+      for (;;) {
+        const int32_t token_id = counter_ptr->acquire_counter();
+        if (token_id >= token_num) {
+          break;
+        }
+        int32_t* __restrict__ curr_expand_token_id_index_buffer =
+            expand_token_id_index_buffer + token_id * topk_num;
+        const float* __restrict__ curr_weight =
+            topk_weights + token_id * topk_num;
+        const float first_weight = skip_weighted ? 1.0f : curr_weight[0];
+        scalar_t* __restrict__ curr_output_buffer =
+            output + token_id * output_size_2;
+
+        if (topk_num > 1) {
+          int32_t w2_output_idx = curr_expand_token_id_index_buffer[0];
+          float* w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          float* ws_output_buffer_iter = ws_output_buffer;
+          vec_op::FP32Vec16 weight_vec(first_weight);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            (vec * weight_vec).save(ws_output_buffer_iter);
+            w2_output_iter += 16;
+            ws_output_buffer_iter += 16;
+          }
+
+          for (int32_t idx = 1; idx < topk_num - 1; ++idx) {
+            w2_output_idx = curr_expand_token_id_index_buffer[idx];
+            w2_output_iter =
+                w2_gemm_output_buffer + w2_output_idx * output_size_2;
+            ws_output_buffer_iter = ws_output_buffer;
+            weight_vec = vec_op::FP32Vec16(curr_weight[idx]);
+            for (int32_t i = 0; i < output_size_2; i += 16) {
+              vec_op::FP32Vec16 vec(w2_output_iter);
+              vec_op::FP32Vec16 sum(ws_output_buffer_iter);
+              (sum + vec * weight_vec).save(ws_output_buffer_iter);
+              w2_output_iter += 16;
+              ws_output_buffer_iter += 16;
+            }
+          }
+
+          const int32_t last_idx = topk_num - 1;
+          w2_output_idx = curr_expand_token_id_index_buffer[last_idx];
+          w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          ws_output_buffer_iter = ws_output_buffer;
+          scalar_t* curr_output_buffer_iter = curr_output_buffer;
+          weight_vec = vec_op::FP32Vec16(curr_weight[last_idx]);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            vec_op::FP32Vec16 sum(ws_output_buffer_iter);
+            scalar_vec_t(sum + vec * weight_vec).save(curr_output_buffer_iter);
+            w2_output_iter += 16;
+            ws_output_buffer_iter += 16;
+            curr_output_buffer_iter += 16;
+          }
+        } else {
+          const int32_t w2_output_idx = curr_expand_token_id_index_buffer[0];
+          float* w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          scalar_t* curr_output_buffer_iter = curr_output_buffer;
+          vec_op::FP32Vec16 weight_vec(first_weight);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            scalar_vec_t(vec * weight_vec).save(curr_output_buffer_iter);
+            w2_output_iter += 16;
+            curr_output_buffer_iter += 16;
+          }
+        }
+      }
+    }
+  }
+}
+}  // namespace
+
+void prepack_moe_weight_int8(
+    const torch::Tensor& weight,  // [expert_num, output_size, input_size]
+    torch::Tensor& packed_weight, const std::string& isa) {
+  TORCH_CHECK(weight.is_contiguous());
+  const int32_t expert_num = weight.size(0);
+  const int32_t output_size = weight.size(1);
+  const int32_t input_size = weight.size(2);
+  const int64_t expert_stride = weight.stride(0);
+  const cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
+  TORCH_CHECK_EQ(output_size % 32, 0);
+
+  CPU_INT8_ISA_DISPATCH_IMPL(isa_type, c10::BFloat16, [&]() {
+    TORCH_CHECK_EQ(input_size % gemm_t::K, 0);
+    prepack_moe_weight_int8_impl<gemm_t>(
+        weight.data_ptr<int8_t>(), packed_weight.data_ptr<int8_t>(), expert_num,
+        output_size, input_size, expert_stride);
+  });
+}
+
+void cpu_fused_moe_int8(torch::Tensor& output, const torch::Tensor& input,
+                        const torch::Tensor& w13, const torch::Tensor& w2,
+                        const torch::Tensor& w13_scale,
+                        const torch::Tensor& w2_scale,
+                        const std::optional<torch::Tensor>& w13_bias,
+                        const std::optional<torch::Tensor>& w2_bias,
+                        const torch::Tensor& topk_weights,
+                        const torch::Tensor& topk_id, const bool skip_weighted,
+                        const std::string& act, const std::string& isa) {
+  const int32_t token_num = input.size(0);
+  const int32_t input_size_13 = input.size(1);
+  const int64_t input_stride = input.stride(0);
+  TORCH_CHECK_EQ(input_stride, input_size_13);
+  const int32_t expert_num = w13.size(0);
+  const int32_t output_size_13 = w13.size(1);
+  const int32_t input_size_2 = w2.size(2);
+  const int32_t output_size_2 = w2.size(1);
+  const int32_t topk_num = topk_id.size(1);
+  const FusedMOEAct act_type = cpu_fused_moe_utils::get_act_type(act);
+  const cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
+  TORCH_CHECK(!skip_weighted || topk_num == 1,
+              "skip_weighted is only supported for topk=1 on CPU");
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "cpu_fused_moe_int8", [&]() {
+        CPU_INT8_ISA_DISPATCH_IMPL(isa_type, scalar_t, [&]() {
+          fused_moe_int8_impl<scalar_t, gemm_t>(
+              output.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
+              w13.data_ptr<int8_t>(), w2.data_ptr<int8_t>(),
+              w13_scale.data_ptr<float>(), w2_scale.data_ptr<float>(),
+              w13_bias.has_value() ? w13_bias->data_ptr<scalar_t>() : nullptr,
+              w2_bias.has_value() ? w2_bias->data_ptr<scalar_t>() : nullptr,
+              topk_weights.data_ptr<float>(), topk_id.data_ptr<int32_t>(),
+              act_type, token_num, expert_num, topk_num, input_size_13,
+              output_size_13, input_size_2, output_size_2, skip_weighted);
+        });
+      });
+}
@@ -287,7 +287,7 @@ struct FP32Vec4 : public Vec<FP32Vec4> {

  explicit FP32Vec4(__vector float data) : reg(data) {}

-  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
+  FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
 };

 struct FP32Vec8 : public Vec<FP32Vec8> {
@@ -316,7 +316,7 @@ struct FP32Vec8 : public Vec<FP32Vec8> {

  explicit FP32Vec8(f32x4x2_t data) : reg(data) {}

-  explicit FP32Vec8(const FP32Vec8& data) {
+  FP32Vec8(const FP32Vec8& data) {
    reg.val[0] = data.reg.val[0];
    reg.val[1] = data.reg.val[1];
  }
@@ -593,7 +593,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
  explicit FP32Vec16(bool, const float* ptr) : FP32Vec16(ptr) {}
  explicit FP32Vec16(f32x4x4_t data) : reg(data) {}

-  explicit FP32Vec16(const FP32Vec16& data) {
+  FP32Vec16(const FP32Vec16& data) {
    reg.val[0] = data.reg.val[0];
    reg.val[1] = data.reg.val[1];
    reg.val[2] = data.reg.val[2];
@@ -747,6 +747,15 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                                vec_abs(reg.val[2]), vec_abs(reg.val[3])}));
  }

+  FP32Vec16 exp() const {
+    FP32Vec8 lo(f32x4x2_t{reg.val[0], reg.val[1]});
+    FP32Vec8 hi(f32x4x2_t{reg.val[2], reg.val[3]});
+    auto lo_e = lo.exp();
+    auto hi_e = hi.exp();
+    return FP32Vec16(f32x4x4_t{lo_e.reg.val[0], lo_e.reg.val[1],
+                               hi_e.reg.val[0], hi_e.reg.val[1]});
+  }
+
  float reduce_max() {
    __vector float max01 = vec_max(reg.val[0], reg.val[1]);
    __vector float max23 = vec_max(reg.val[2], reg.val[3]);
@@ -31,6 +31,9 @@ class MicroGemm {
  }
 };

+template <cpu_utils::ISA isa, typename scalar_t>
+class MicroGemmINT8;
+
 template <int32_t n_size, typename scalar_t>
 FORCE_INLINE void default_epilogue(float* __restrict__ c_ptr,
                                   scalar_t* __restrict__ d_ptr,
@@ -0,0 +1,424 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#ifndef CPU_MICRO_GEMM_INT8_NEON_HPP
+#define CPU_MICRO_GEMM_INT8_NEON_HPP
+
+#include <algorithm>
+#include <cstdint>
+
+#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
+
+#include <arm_bf16.h>
+#include <arm_neon.h>
+#include <c10/util/BFloat16.h>
+#include <c10/util/Exception.h>
+#include <c10/util/Half.h>
+
+namespace cpu_micro_gemm {
+
+namespace neon_smmla {
+
+constexpr int32_t K = 8;
+constexpr int32_t Cols = 2;
+constexpr int32_t TileSize = K * Cols;
+
+FORCE_INLINE float32x4x2_t load_as_f32(const float* input) {
+  float32x4x2_t result;
+  result.val[0] = vld1q_f32(input);
+  result.val[1] = vld1q_f32(input + 4);
+  return result;
+}
+
+FORCE_INLINE float32x4x2_t load_as_f32(const c10::Half* input) {
+  const auto input_vec = vld1q_f16(reinterpret_cast<const float16_t*>(input));
+  float32x4x2_t result;
+  result.val[0] = vcvt_f32_f16(vget_low_f16(input_vec));
+  result.val[1] = vcvt_f32_f16(vget_high_f16(input_vec));
+  return result;
+}
+
+FORCE_INLINE float32x4x2_t load_as_f32(const c10::BFloat16* input) {
+  const auto input_vec = vld1q_bf16(reinterpret_cast<const bfloat16_t*>(input));
+  float32x4x2_t result;
+  result.val[0] = vcvt_f32_bf16(vget_low_bf16(input_vec));
+  result.val[1] = vcvt_f32_bf16(vget_high_bf16(input_vec));
+  return result;
+}
+
+FORCE_INLINE void store_acc_rowpair(const int32x4_t acc01,
+                                    const int32x4_t acc23,
+                                    const int32x4_t acc45,
+                                    const int32x4_t acc67,
+                                    int32_t* __restrict__ c_ptr,
+                                    const int64_t ldc, const int32_t m_rows) {
+  if (m_rows == 0) {
+    return;
+  }
+
+  vst1q_s32(c_ptr, vcombine_s32(vget_low_s32(acc01), vget_low_s32(acc23)));
+  vst1q_s32(c_ptr + 4, vcombine_s32(vget_low_s32(acc45), vget_low_s32(acc67)));
+
+  if (m_rows == 2) {
+    vst1q_s32(c_ptr + ldc,
+              vcombine_s32(vget_high_s32(acc01), vget_high_s32(acc23)));
+    vst1q_s32(c_ptr + ldc + 4,
+              vcombine_s32(vget_high_s32(acc45), vget_high_s32(acc67)));
+  }
+}
+
+FORCE_INLINE void gemm_micro_smmla_8x8_packed_a(
+    const int8_t* __restrict__ a_packed, const int8_t* __restrict__ b_packed,
+    int32_t* __restrict__ c_ptr, const int32_t m, const int32_t k_size,
+    const int64_t ldc) {
+  const int32x4_t zero = vdupq_n_s32(0);
+  int32x4_t acc0101 = zero, acc0123 = zero, acc0145 = zero, acc0167 = zero;
+  int32x4_t acc2301 = zero, acc2323 = zero, acc2345 = zero, acc2367 = zero;
+  int32x4_t acc4501 = zero, acc4523 = zero, acc4545 = zero, acc4567 = zero;
+  int32x4_t acc6701 = zero, acc6723 = zero, acc6745 = zero, acc6767 = zero;
+
+  const int8_t* __restrict__ a_tile = a_packed;
+  const int8_t* __restrict__ b_tile = b_packed;
+
+#pragma GCC unroll 8
+  for (int32_t k_idx = 0; k_idx < k_size; k_idx += K) {
+    const int8x16_t a_tile01 = vld1q_s8(a_tile);
+    const int8x16_t a_tile23 = vld1q_s8(a_tile + TileSize);
+    const int8x16_t a_tile45 = vld1q_s8(a_tile + 2 * TileSize);
+    const int8x16_t a_tile67 = vld1q_s8(a_tile + 3 * TileSize);
+
+    const int8x16_t b_tile01 = vld1q_s8(b_tile);
+    const int8x16_t b_tile23 = vld1q_s8(b_tile + TileSize);
+    const int8x16_t b_tile45 = vld1q_s8(b_tile + 2 * TileSize);
+    const int8x16_t b_tile67 = vld1q_s8(b_tile + 3 * TileSize);
+
+    acc0101 = vmmlaq_s32(acc0101, a_tile01, b_tile01);
+    acc2301 = vmmlaq_s32(acc2301, a_tile23, b_tile01);
+    acc4501 = vmmlaq_s32(acc4501, a_tile45, b_tile01);
+    acc6701 = vmmlaq_s32(acc6701, a_tile67, b_tile01);
+
+    acc0123 = vmmlaq_s32(acc0123, a_tile01, b_tile23);
+    acc2323 = vmmlaq_s32(acc2323, a_tile23, b_tile23);
+    acc4523 = vmmlaq_s32(acc4523, a_tile45, b_tile23);
+    acc6723 = vmmlaq_s32(acc6723, a_tile67, b_tile23);
+
+    acc0145 = vmmlaq_s32(acc0145, a_tile01, b_tile45);
+    acc2345 = vmmlaq_s32(acc2345, a_tile23, b_tile45);
+    acc4545 = vmmlaq_s32(acc4545, a_tile45, b_tile45);
+    acc6745 = vmmlaq_s32(acc6745, a_tile67, b_tile45);
+
+    acc0167 = vmmlaq_s32(acc0167, a_tile01, b_tile67);
+    acc2367 = vmmlaq_s32(acc2367, a_tile23, b_tile67);
+    acc4567 = vmmlaq_s32(acc4567, a_tile45, b_tile67);
+    acc6767 = vmmlaq_s32(acc6767, a_tile67, b_tile67);
+
+    a_tile += 4 * TileSize;
+    b_tile += 4 * TileSize;
+  }
+
+  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc,
+                    std::min(2, m));
+  store_acc_rowpair(acc2301, acc2323, acc2345, acc2367, c_ptr + 2 * ldc, ldc,
+                    std::min(2, std::max(0, m - 2)));
+  store_acc_rowpair(acc4501, acc4523, acc4545, acc4567, c_ptr + 4 * ldc, ldc,
+                    std::min(2, std::max(0, m - 4)));
+  store_acc_rowpair(acc6701, acc6723, acc6745, acc6767, c_ptr + 6 * ldc, ldc,
+                    std::min(2, std::max(0, m - 6)));
+}
+
+FORCE_INLINE void gemm_micro_smmla_4x16_packed_a(
+    const int8_t* __restrict__ a_packed, const int8_t* __restrict__ b_packed,
+    int32_t* __restrict__ c_ptr, const int32_t m, const int32_t k_size,
+    const int64_t b_n_group_stride, const int64_t ldc) {
+  const int32_t m_rows_01 = std::min(2, m);
+  const int32_t m_rows_23 = std::min(2, std::max(0, m - 2));
+  const int32x4_t zero = vdupq_n_s32(0);
+
+  int32x4_t acc0101 = zero, acc0123 = zero, acc0145 = zero, acc0167 = zero;
+  int32x4_t acc2301 = zero, acc2323 = zero, acc2345 = zero, acc2367 = zero;
+  int32x4_t acc0189 = zero, acc011011 = zero, acc011213 = zero,
+            acc011415 = zero;
+  int32x4_t acc2389 = zero, acc231011 = zero, acc231213 = zero,
+            acc231415 = zero;
+
+  const int8_t* __restrict__ a_tile = a_packed;
+  // note: b packs 8 panels contiguously, so we need 2 b_tile ptrs
+  // for the 4x16 microkernel
+  const int8_t* __restrict__ b_tile0 = b_packed;
+  const int8_t* __restrict__ b_tile1 = b_packed + b_n_group_stride;
+
+#pragma GCC unroll 8
+  for (int32_t k_idx = 0; k_idx < k_size; k_idx += K) {
+    const int8x16_t a_tile01 = vld1q_s8(a_tile);
+    const int8x16_t a_tile23 = vld1q_s8(a_tile + TileSize);
+    const int8x16_t b_tile01 = vld1q_s8(b_tile0);
+    const int8x16_t b_tile23 = vld1q_s8(b_tile0 + TileSize);
+    const int8x16_t b_tile45 = vld1q_s8(b_tile0 + 2 * TileSize);
+    const int8x16_t b_tile67 = vld1q_s8(b_tile0 + 3 * TileSize);
+    const int8x16_t b_tile89 = vld1q_s8(b_tile1);
+    const int8x16_t b_tile1011 = vld1q_s8(b_tile1 + TileSize);
+    const int8x16_t b_tile1213 = vld1q_s8(b_tile1 + 2 * TileSize);
+    const int8x16_t b_tile1415 = vld1q_s8(b_tile1 + 3 * TileSize);
+
+    acc0101 = vmmlaq_s32(acc0101, a_tile01, b_tile01);
+    acc2301 = vmmlaq_s32(acc2301, a_tile23, b_tile01);
+    acc0123 = vmmlaq_s32(acc0123, a_tile01, b_tile23);
+    acc2323 = vmmlaq_s32(acc2323, a_tile23, b_tile23);
+
+    acc0145 = vmmlaq_s32(acc0145, a_tile01, b_tile45);
+    acc2345 = vmmlaq_s32(acc2345, a_tile23, b_tile45);
+    acc0167 = vmmlaq_s32(acc0167, a_tile01, b_tile67);
+    acc2367 = vmmlaq_s32(acc2367, a_tile23, b_tile67);
+
+    acc0189 = vmmlaq_s32(acc0189, a_tile01, b_tile89);
+    acc2389 = vmmlaq_s32(acc2389, a_tile23, b_tile89);
+    acc011011 = vmmlaq_s32(acc011011, a_tile01, b_tile1011);
+    acc231011 = vmmlaq_s32(acc231011, a_tile23, b_tile1011);
+
+    acc011213 = vmmlaq_s32(acc011213, a_tile01, b_tile1213);
+    acc231213 = vmmlaq_s32(acc231213, a_tile23, b_tile1213);
+    acc011415 = vmmlaq_s32(acc011415, a_tile01, b_tile1415);
+    acc231415 = vmmlaq_s32(acc231415, a_tile23, b_tile1415);
+
+    a_tile += 2 * TileSize;
+    b_tile0 += 4 * TileSize;
+    b_tile1 += 4 * TileSize;
+  }
+
+  // rows 0-1, columns 0-7
+  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc, m_rows_01);
+  // rows 0-1, columns 8-15
+  store_acc_rowpair(acc0189, acc011011, acc011213, acc011415, c_ptr + 8, ldc,
+                    m_rows_01);
+  // rows 2-3, columns 0-7
+  store_acc_rowpair(acc2301, acc2323, acc2345, acc2367, c_ptr + 2 * ldc, ldc,
+                    m_rows_23);
+  // rows 2-3, columns 8-15
+  store_acc_rowpair(acc2389, acc231011, acc231213, acc231415,
+                    c_ptr + 2 * ldc + 8, ldc, m_rows_23);
+}
+
+}  // namespace neon_smmla
+
+template <typename scalar_t>
+class MicroGemmINT8<cpu_utils::ISA::NEON, scalar_t> {
+ public:
+  static constexpr int32_t K = neon_smmla::K;
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
+  static constexpr int32_t MaxMSize = 8;
+  static constexpr int32_t NSize = 32;
+  static constexpr int32_t WeightOCGroupSize = Nr;
+  static_assert(MaxMSize % Mr == 0);
+
+  static FORCE_INLINE void quantize_row(const scalar_t* input, int8_t* output,
+                                        float& scale, const int32_t size) {
+    TORCH_CHECK_EQ(size % K, 0);
+    float32x4_t max_vec = vdupq_n_f32(0.0f);
+
+    for (int32_t i = 0; i < size; i += K) {
+      const float32x4x2_t input_vec = neon_smmla::load_as_f32(input + i);
+      max_vec = vmaxq_f32(max_vec, vabsq_f32(input_vec.val[0]));
+      max_vec = vmaxq_f32(max_vec, vabsq_f32(input_vec.val[1]));
+    }
+
+    const float abs_max = std::max(vmaxvq_f32(max_vec), 1.0e-7f);
+    scale = abs_max / 127.0f;
+    const float32x4_t inv_scale_vec = vdupq_n_f32(127.0f / abs_max);
+
+    for (int32_t i = 0; i < size; i += K) {
+      const float32x4x2_t input_vec = neon_smmla::load_as_f32(input + i);
+      const int32x4_t output_low =
+          vcvtnq_s32_f32(vmulq_f32(input_vec.val[0], inv_scale_vec));
+      const int32x4_t output_high =
+          vcvtnq_s32_f32(vmulq_f32(input_vec.val[1], inv_scale_vec));
+      const int16x8_t output_s16 =
+          vcombine_s16(vqmovn_s32(output_low), vqmovn_s32(output_high));
+      vst1_s8(output + i, vqmovn_s16(output_s16));
+    }
+  }
+
+  // with current code, fusing this into the gemm micro kernel didn't move the
+  // needle
+  static FORCE_INLINE void dequantize_tile(
+      int32_t* input, float* output, const float* __restrict__ input_scales,
+      const float* __restrict__ weight_scales, const int32_t m, const int32_t n,
+      const int32_t stride) {
+    TORCH_CHECK_EQ(n % 4, 0);
+    for (int32_t m_idx = 0; m_idx < m; ++m_idx) {
+      const float32x4_t input_scale_vec = vdupq_n_f32(input_scales[m_idx]);
+      for (int32_t n_idx = 0; n_idx < n; n_idx += 4) {
+        const int32x4_t input_vec = vld1q_s32(input + m_idx * stride + n_idx);
+        const float32x4_t weight_scale_vec = vld1q_f32(weight_scales + n_idx);
+        const float32x4_t output_vec =
+            vmulq_f32(vcvtq_f32_s32(input_vec),
+                      vmulq_f32(input_scale_vec, weight_scale_vec));
+        vst1q_f32(output + m_idx * stride + n_idx, output_vec);
+      }
+    }
+  }
+
+  // physical layout [
+  //  M / (8 or 4); Mr is 8 or 4
+  //  K / 8; K for smmla is 8
+  //  4,   ; 4 row-pairs for each 8 rows
+  //  2,   ; row-pair is 2 rows
+  //  4    ; 4 elements per row
+  // ]
+  static void pack_input_from_rows(const int8_t* const* __restrict__ rows,
+                                   int8_t* __restrict__ a_packed,
+                                   const int32_t m, const int32_t k) {
+    TORCH_CHECK(m > 0 && m <= MaxMSize);
+    TORCH_CHECK(k % K == 0);
+    const int8x8_t zero = vdup_n_s8(0);
+
+    for (int32_t row_base = 0; row_base < m; row_base += Mr) {
+      const int32_t panel_m = std::min(Mr, m - row_base);
+      const int8_t* const* panel_rows = rows + row_base;
+      int8_t* __restrict__ out = a_packed + row_base * k;
+
+      // fast path for full 8-row panels (fast path for 4-row panels didn't move
+      // the needle)
+      if (panel_m == Mr) {
+        const int8_t* __restrict__ row0 = panel_rows[0];
+        const int8_t* __restrict__ row1 = panel_rows[1];
+        const int8_t* __restrict__ row2 = panel_rows[2];
+        const int8_t* __restrict__ row3 = panel_rows[3];
+        const int8_t* __restrict__ row4 = panel_rows[4];
+        const int8_t* __restrict__ row5 = panel_rows[5];
+        const int8_t* __restrict__ row6 = panel_rows[6];
+        const int8_t* __restrict__ row7 = panel_rows[7];
+        int32_t k_idx = 0;
+        for (; k_idx + 2 * K <= k; k_idx += 2 * K) {
+          int8_t* __restrict__ block0 = out;
+          int8_t* __restrict__ block1 = out + 4 * neon_smmla::TileSize;
+
+          int8x16_t a0 = vld1q_s8(row0 + k_idx);
+          int8x16_t a1 = vld1q_s8(row1 + k_idx);
+          vst1q_s8(block0, vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1, vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row2 + k_idx);
+          a1 = vld1q_s8(row3 + k_idx);
+          vst1q_s8(block0 + neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row4 + k_idx);
+          a1 = vld1q_s8(row5 + k_idx);
+          vst1q_s8(block0 + 2 * neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + 2 * neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row6 + k_idx);
+          a1 = vld1q_s8(row7 + k_idx);
+          vst1q_s8(block0 + 3 * neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + 3 * neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          out += 8 * neon_smmla::TileSize;
+        }
+
+        for (; k_idx < k; k_idx += K) {
+          int8x8_t a0 = vld1_s8(row0 + k_idx);
+          int8x8_t a1 = vld1_s8(row1 + k_idx);
+          vst1q_s8(out, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row2 + k_idx);
+          a1 = vld1_s8(row3 + k_idx);
+          vst1q_s8(out + neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row4 + k_idx);
+          a1 = vld1_s8(row5 + k_idx);
+          vst1q_s8(out + 2 * neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row6 + k_idx);
+          a1 = vld1_s8(row7 + k_idx);
+          vst1q_s8(out + 3 * neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          out += 4 * neon_smmla::TileSize;
+        }
+        continue;
+      }
+
+      const int32_t row_pairs = (panel_m <= 4) ? 2 : Mr / 2;
+      for (int32_t k_idx = 0; k_idx < k; k_idx += K) {
+        for (int32_t pair_idx = 0; pair_idx < row_pairs; ++pair_idx) {
+          const int32_t row_idx = pair_idx * 2;
+          const int8x8_t row0 =
+              (row_idx < panel_m) ? vld1_s8(panel_rows[row_idx] + k_idx) : zero;
+          const int8x8_t row1 = (row_idx + 1 < panel_m)
+                                    ? vld1_s8(panel_rows[row_idx + 1] + k_idx)
+                                    : zero;
+          vst1q_s8(out, vcombine_s8(row0, row1));
+          out += neon_smmla::TileSize;
+        }
+      }
+    }
+  }
+
+  // physical layout [
+  //  N / 8; Nr is 8
+  //  K / 8; K for smmla is 8
+  //  4,   ; 4 col-pairs for each 8 cols
+  //  2,   ; col-pair is 2 cols
+  //  4    ; 4 elements per col
+  // ]
+  static void pack_weight(const int8_t* __restrict__ weight,
+                          int8_t* __restrict__ packed_weight,
+                          const int32_t output_size, const int32_t input_size) {
+    TORCH_CHECK(output_size % NSize == 0);
+    TORCH_CHECK(input_size % K == 0);
+
+    for (int32_t o_idx = 0; o_idx < output_size; o_idx += Nr) {
+      int8_t* __restrict__ dst = packed_weight + o_idx * input_size;
+      for (int32_t k_idx = 0; k_idx < input_size; k_idx += K) {
+        for (int32_t pair_idx = 0; pair_idx < Nr;
+             pair_idx += neon_smmla::Cols) {
+          const int8_t* __restrict__ row0 =
+              weight + (o_idx + pair_idx) * input_size + k_idx;
+          const int8_t* __restrict__ row1 = row0 + input_size;
+          vst1q_s8(dst, vcombine_s8(vld1_s8(row0), vld1_s8(row1)));
+          dst += neon_smmla::TileSize;
+        }
+      }
+    }
+  }
+
+  void gemm(const int8_t* __restrict__ a_packed,
+            const int8_t* __restrict__ b_packed, int32_t* __restrict__ c,
+            const int32_t m, const int32_t k, const int64_t b_n_group_stride,
+            const int64_t ldc) const {
+    TORCH_CHECK(m > 0 && m <= MaxMSize);
+    TORCH_CHECK(k % K == 0);
+
+    for (int32_t n_idx = 0; n_idx < NSize; n_idx += NrGemv) {
+      const int8_t* __restrict__ b_panel = b_packed + n_idx * k;
+
+      for (int32_t row_base = 0; row_base < m; row_base += Mr) {
+        const int32_t panel_m = std::min(Mr, m - row_base);
+        const int8_t* __restrict__ a_panel = a_packed + row_base * k;
+        int32_t* __restrict__ c_panel = c + row_base * ldc + n_idx;
+
+        if (panel_m <= 4) {
+          neon_smmla::gemm_micro_smmla_4x16_packed_a(
+              a_panel, b_panel, c_panel, panel_m, k, b_n_group_stride, ldc);
+        } else {
+          neon_smmla::gemm_micro_smmla_8x8_packed_a(a_panel, b_panel, c_panel,
+                                                    panel_m, k, ldc);
+          neon_smmla::gemm_micro_smmla_8x8_packed_a(
+              a_panel, b_panel + b_n_group_stride, c_panel + Nr, panel_m, k,
+              ldc);
+        }
+      }
+    }
+  }
+};
+
+}  // namespace cpu_micro_gemm
+
+#endif
@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 #ifndef CPU_MICRO_GEMM_NEON_HPP
 #define CPU_MICRO_GEMM_NEON_HPP

@@ -16,9 +19,6 @@ namespace {
 constexpr int32_t K = 4;
 constexpr int32_t Cols = 2;
 constexpr int32_t TileSize = K * Cols;
-constexpr int32_t Mr = 8;
-constexpr int32_t Nr = 8;
-constexpr int32_t Nr_gemv = 16;

 // a = [a0, a1, a2, a3], b = [b0, b1, b2, b3] -> [a0, a1, b0, b1]
 FORCE_INLINE float32x4_t zip1_f32x4(const float32x4_t a, const float32x4_t b) {
@@ -132,7 +132,7 @@ FORCE_INLINE void gemm_micro_bfmmla_8x8_packed_a(
    acc6767 = vbfmmlaq_f32(acc6767, a_tile67, b_tile67);

    a_tile += 4 * TileSize;
-    b_tile += Nr * K;
+    b_tile += 4 * TileSize;
  }

  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc,
@@ -205,8 +205,8 @@ FORCE_INLINE void gemm_micro_bfmmla_4x16_packed_a(
    acc231415 = vbfmmlaq_f32(acc231415, a_tile23, b_tile1415);

    a_tile += 2 * TileSize;
-    b_tile0 += Nr * K;
-    b_tile1 += Nr * K;
+    b_tile0 += 4 * TileSize;
+    b_tile1 += 4 * TileSize;
  }

  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc, m_rows_01);
@@ -223,6 +223,9 @@ FORCE_INLINE void gemm_micro_bfmmla_4x16_packed_a(
 template <typename scalar_t>
 class MicroGemm<cpu_utils::ISA::NEON, scalar_t> {
 public:
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
  static constexpr int32_t MaxMSize = 8;
  static constexpr int32_t NSize = 32;
  static constexpr int32_t WeightOCGroupSize = Nr;
@@ -246,6 +249,9 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
 public:
  using scalar_t = c10::BFloat16;

+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
  static constexpr int32_t MaxMSize = 8;
  static constexpr int32_t NSize = 32;
  static constexpr int32_t WeightOCGroupSize = Nr;
@@ -253,7 +259,7 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {

 public:
  // physical layout [
-  //  M / 8; Mr is 8
+  //  M / (8 or 4); Mr is 8 or 4
  //  K / 4; K for bfmmla is 4
  //  4,   ; 4 row-pairs for each 8 rows
  //  2,   ; row-pair is 2 rows
@@ -439,7 +445,7 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
    (void)lda;  // A is packed, so lda is not needed
    TORCH_CHECK_EQ(k % K, 0);

-    for (int32_t n_idx = 0; n_idx < NSize; n_idx += Nr_gemv) {
+    for (int32_t n_idx = 0; n_idx < NSize; n_idx += NrGemv) {
      const bfloat16_t* __restrict__ b_panel =
          reinterpret_cast<const bfloat16_t*>(b_ptr) + n_idx * k;

@@ -451,6 +451,90 @@ void causal_conv1d_update_kernel_impl(
  });
 }

+template <typename scalar_t>
+void causal_conv1d_update_multi_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    scalar_t* __restrict__ conv_states,
+    const scalar_t* __restrict__ weight,
+    const scalar_t* __restrict__ bias,
+    const int32_t* __restrict__ num_accepted_tokens,
+    const int32_t* __restrict__ conv_indices,
+    bool silu_activation,
+    int64_t batch,
+    int64_t dim,
+    int64_t seqlen,
+    int64_t width,
+    int64_t state_len,
+    int64_t conv_state_slot_stride) {
+  constexpr int64_t BLOCK_N = block_size_n() * 2;
+  const int64_t NB = div_up(dim, BLOCK_N);
+
+  AT_DISPATCH_BOOL2(bias != nullptr, has_bias, silu_activation, has_silu, [&] {
+    at::parallel_for(0, batch * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t bs{0}, nb{0};
+      data_index_init(begin, bs, batch, nb, NB);
+
+      for (int64_t i = begin; i < end; ++i) {
+        const int64_t nb_start = nb * BLOCK_N;
+        const int64_t nb_size = std::min(dim - nb_start, BLOCK_N);
+        const int32_t conv_state_index = conv_indices[bs];
+        const int32_t history_offset = num_accepted_tokens[bs] - 1;
+
+        switch (width << 4 | nb_size >> 4) {
+          case 0x42:
+            tinygemm_kernel<scalar_t, 4, 32, has_bias, has_silu>::apply(
+                input + bs * seqlen * dim + nb_start,
+                weight + nb_start * width,
+                out + bs * seqlen * dim + nb_start,
+                has_bias ? bias + nb_start : nullptr,
+                conv_states + conv_state_index * conv_state_slot_stride +
+                    history_offset * dim + nb_start,
+                true,
+                seqlen,
+                dim,
+                true);
+            break;
+          case 0x44:
+            tinygemm_kernel<scalar_t, 4, 64, has_bias, has_silu>::apply(
+                input + bs * seqlen * dim + nb_start,
+                weight + nb_start * width,
+                out + bs * seqlen * dim + nb_start,
+                has_bias ? bias + nb_start : nullptr,
+                conv_states + conv_state_index * conv_state_slot_stride +
+                    history_offset * dim + nb_start,
+                true,
+                seqlen,
+                dim,
+                true);
+            break;
+          default:
+            TORCH_CHECK(false, "Unexpected block size, ", width, " x ", nb_size);
+        }
+
+        data_index_step(bs, batch, nb, NB);
+      }
+    });
+  });
+
+  at::parallel_for(0, batch, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t bs = begin; bs < end; ++bs) {
+      const int32_t conv_state_index = conv_indices[bs];
+      const int32_t num_accepted = num_accepted_tokens[bs];
+      scalar_t* state = conv_states + conv_state_index * conv_state_slot_stride;
+
+      std::memmove(
+          state,
+          state + num_accepted * dim,
+          (state_len - seqlen) * dim * sizeof(scalar_t));
+      std::memcpy(
+          state + (state_len - seqlen) * dim,
+          input + bs * seqlen * dim,
+          seqlen * dim * sizeof(scalar_t));
+    }
+  });
+}
+
 }  // anonymous namespace

 // from [dim, width] or [N, K]
@@ -545,7 +629,7 @@ at::Tensor get_block_indices(const std::optional<at::Tensor>& offsets, int64_t n
 //   query_start_loc: (batch + 1) int32
 //   cache_indices: (batch)  int32
 //   has_initial_state: (batch) bool
-//   conv_states: (..., dim, width - 1) itype
+//   conv_states: (..., dim, state_len) itype, where state_len >= width - 1
 //   activation: either None or "silu" or "swish"
 //   pad_slot_id: int
 //
@@ -586,11 +670,14 @@ at::Tensor causal_conv1d_fwd_cpu(
    CHECK_EQ(conv_states_val.scalar_type(), scalar_type);
    CHECK_GE(padded_batch, batch);
    CHECK_EQ(conv_states_val.size(1), dim);
-    CHECK_EQ(conv_states_val.size(2), width - 1);
+    const int64_t state_len = conv_states_val.size(2);
+    CHECK_GE(state_len, width - 1);

    // adjust `conv_states` to be contiguous on `dim`
    // should happen only once
    if (conv_states_val.stride(-2) != 1) {
+      TORCH_CHECK(state_len == width - 1,
+          "causal_conv1d_fwd_cpu: wide conv_states must be contiguous on dim.");
      auto conv_states_copy = conv_states_val.clone();
      conv_states_val.as_strided_({padded_batch, dim, width - 1}, {(width - 1) * dim, 1, dim});
      conv_states_val.copy_(conv_states_copy);
@@ -651,14 +738,14 @@ at::Tensor causal_conv1d_fwd_cpu(

 // API aligned with GPUs
 //
-//   x: (batch, dim) or (batch, dim, seqlen)
+//   x: (batch, dim) or (batch, seqlen, dim)
 //   conv_state: (..., dim, state_len), where state_len >= width - 1
 //   weight: (dim, width)
 //   bias: (dim,)
-//   cache_seqlens: (batch,), dtype int32.
+//   num_accepted_tokens: (batch,), dtype int32.
 //   conv_state_indices: (batch,), dtype int32
 //   pad_slot_id: int
-//   out: (batch, dim) or (batch, dim, seqlen)
+//   out: (batch, dim) or (batch, seqlen, dim)
 //
 at::Tensor causal_conv1d_update_cpu(
    const at::Tensor& x,
@@ -666,7 +753,7 @@ at::Tensor causal_conv1d_update_cpu(
    const at::Tensor& weight,
    const std::optional<at::Tensor>& bias,
    bool silu_activation,
-    const std::optional<at::Tensor>& cache_seqlens,
+    const std::optional<at::Tensor>& num_accepted_tokens,
    const std::optional<at::Tensor>& conv_state_indices,
    int64_t pad_slot_id,
    bool is_vnni) {
@@ -674,13 +761,13 @@ at::Tensor causal_conv1d_update_cpu(
  CHECK_CONTIGUOUS(weight);
  auto packed_w = is_vnni ? weight : causal_conv1d_weight_pack(weight);

-  // TODO: add multi-token prediction support
-  TORCH_CHECK(x.dim() == 2, "causal_conv1d_update_cpu: expect x to be 2D tensor.");
-  TORCH_CHECK(!cache_seqlens.has_value(), "causal_conv1d_update_cpu: don't support cache_seqlens.");
+  TORCH_CHECK(
+      x.dim() == 2 || x.dim() == 3,
+      "causal_conv1d_update_cpu: expect x to be 2D or 3D tensor.");

  int64_t batch = x.size(0);
-  int64_t dim = x.size(1);
-  int64_t seqlen = 1;
+  int64_t dim = x.dim() == 2 ? x.size(1) : x.size(2);
+  int64_t seqlen = x.dim() == 2 ? 1 : x.size(1);
  int64_t width = weight.size(-1);

  const auto scalar_type = x.scalar_type();
@@ -690,10 +777,84 @@ at::Tensor causal_conv1d_update_cpu(

  CHECK_EQ(conv_states.scalar_type(), scalar_type);
  CHECK_EQ(conv_states.size(1), dim);
-  CHECK_EQ(conv_states.size(2), width - 1);
+  const int64_t state_len = conv_states.size(2);
+  CHECK_GE(state_len, width - 1);
+
+  if (x.dim() == 3) {
+    TORCH_CHECK(
+        num_accepted_tokens.has_value(),
+        "causal_conv1d_update_cpu: num_accepted_tokens is required for 3D x.");
+    TORCH_CHECK(
+        conv_state_indices.has_value(),
+        "causal_conv1d_update_cpu: conv_state_indices is required for 3D x.");
+    CHECK_OPTIONAL_SHAPE_DTYPE(num_accepted_tokens, batch, at::kInt);
+    TORCH_CHECK(
+        width == 4,
+        "causal_conv1d_update_cpu: support only width of 4 for 3D x.");
+    TORCH_CHECK(
+        seqlen > 0,
+        "causal_conv1d_update_cpu: expect non-empty sequence for 3D x.");
+    TORCH_CHECK(
+        state_len >= seqlen,
+        "causal_conv1d_update_cpu: state_len must be >= seqlen for 3D x.");
+    TORCH_CHECK(
+        conv_states.stride(-2) == 1 && conv_states.stride(-1) == dim,
+        "causal_conv1d_update_cpu: 3D x requires SD conv_states layout.");
+
+    const int32_t* accepted_counts =
+        num_accepted_tokens.value().data_ptr<int32_t>();
+    const int32_t* indices = conv_state_indices.value().data_ptr<int32_t>();
+    const int64_t num_slots = conv_states.size(0);
+    for (int64_t bs = 0; bs < batch; ++bs) {
+      const int32_t num_accepted = accepted_counts[bs];
+      const int32_t conv_state_index = indices[bs];
+      TORCH_CHECK(
+          conv_state_index != pad_slot_id,
+          "causal_conv1d_update_cpu: 3D x does not support pad slots.");
+      TORCH_CHECK(
+          conv_state_index >= 0 && conv_state_index < num_slots,
+          "causal_conv1d_update_cpu: conv_state_indices out of range.");
+      TORCH_CHECK(
+          num_accepted >= 1 && num_accepted <= seqlen,
+          "causal_conv1d_update_cpu: num_accepted_tokens must be in [1, "
+          "seqlen].");
+      TORCH_CHECK(
+          num_accepted - 1 + width - 1 <= state_len,
+          "causal_conv1d_update_cpu: history window exceeds conv_states.");
+    }
+
+    int64_t conv_state_slot_stride = conv_states.stride(0);
+    at::Tensor out = at::empty_like(x);
+    AT_DISPATCH_REDUCED_FLOATING_TYPES(
+        scalar_type, "causal_conv1d_update_multi_kernel_impl", [&] {
+          causal_conv1d_update_multi_kernel_impl<scalar_t>(
+              out.data_ptr<scalar_t>(),
+              x.data_ptr<scalar_t>(),
+              conv_states.data_ptr<scalar_t>(),
+              packed_w.data_ptr<scalar_t>(),
+              conditional_data_ptr<scalar_t>(bias),
+              accepted_counts,
+              indices,
+              silu_activation,
+              batch,
+              dim,
+              seqlen,
+              width,
+              state_len,
+              conv_state_slot_stride);
+        });
+    return out;
+  }
+
+  TORCH_CHECK(
+      !num_accepted_tokens.has_value(),
+      "causal_conv1d_update_cpu: num_accepted_tokens is only supported for 3D "
+      "x.");

  // adjust `conv_states` to be contiguous on `dim`
  if (conv_states.stride(-2) != 1) {
+    TORCH_CHECK(state_len == width - 1,
+        "causal_conv1d_update_cpu: wide conv_states must be contiguous on dim.");
    int64_t num_cache_lines = conv_states.size(0);
    auto conv_states_copy = conv_states.clone();
    conv_states.as_strided_({num_cache_lines, dim, width - 1}, {(width - 1) * dim, 1, dim});
@@ -147,7 +147,7 @@ at::Tensor causal_conv1d_fwd_cpu(
 at::Tensor causal_conv1d_update_cpu(
    const at::Tensor& x, const at::Tensor& conv_states,
    const at::Tensor& weight, const std::optional<at::Tensor>& bias,
-    bool silu_activation, const std::optional<at::Tensor>& cache_seqlens,
+    bool silu_activation, const std::optional<at::Tensor>& num_accepted_tokens,
    const std::optional<at::Tensor>& conv_state_indices, int64_t pad_slot_id,
    bool is_vnni);

@@ -207,6 +207,20 @@ void cpu_fused_moe(torch::Tensor& output, const torch::Tensor& input,
                   const torch::Tensor& topk_id, const bool skip_weighted,
                   const std::string& act, const std::string& isa);

+void prepack_moe_weight_int8(const torch::Tensor& weight,
+                             torch::Tensor& packed_weight,
+                             const std::string& isa);
+
+void cpu_fused_moe_int8(torch::Tensor& output, const torch::Tensor& input,
+                        const torch::Tensor& w13, const torch::Tensor& w2,
+                        const torch::Tensor& w13_scale,
+                        const torch::Tensor& w2_scale,
+                        const std::optional<torch::Tensor>& w13_bias,
+                        const std::optional<torch::Tensor>& w2_bias,
+                        const torch::Tensor& topk_weights,
+                        const torch::Tensor& topk_id, const bool skip_weighted,
+                        const std::string& act, const std::string& isa);
+
 void compute_slot_mapping_kernel_impl(const torch::Tensor query_start_loc,
                                      const torch::Tensor positions,
                                      const torch::Tensor block_table,
@@ -502,7 +516,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def(
      "causal_conv1d_update_cpu(Tensor x, Tensor(a!) conv_states, Tensor "
      "weight, Tensor? bias, bool silu_activation,"
-      "Tensor? cache_seqlens, Tensor? conv_state_indices, int pad_slot_id, "
+      "Tensor? num_accepted_tokens, Tensor? conv_state_indices, int "
+      "pad_slot_id, "
      "bool is_vnni) -> Tensor");
  ops.impl("causal_conv1d_update_cpu", torch::kCPU, &causal_conv1d_update_cpu);
 #endif
@@ -596,8 +611,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif

  // fused moe
-#if defined(__AVX512F__) || \
-    (defined(__aarch64__) && !defined(__APPLE__) && defined(ARM_BF16_SUPPORT))
+#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) && !defined(__APPLE__))
  ops.def(
      "prepack_moe_weight(Tensor weight, Tensor(a1!) packed_weight, str isa) "
      "-> ()");
@@ -608,7 +622,22 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
      "bool skip_weighted, "
      "str act, str isa) -> ()");
  ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
-#endif
+#endif  // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) &&
+        // !defined(__APPLE__))
+#if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT) && \
+    !defined(__APPLE__)
+  ops.def(
+      "prepack_moe_weight_int8(Tensor weight, Tensor(a1!) packed_weight, "
+      "str isa) -> ()");
+  ops.impl("prepack_moe_weight_int8", torch::kCPU, &prepack_moe_weight_int8);
+  ops.def(
+      "cpu_fused_moe_int8(Tensor(a0!) output, Tensor input, Tensor w13, "
+      "Tensor w2, Tensor w13_scale, Tensor w2_scale, Tensor? w13_bias, "
+      "Tensor? w2_bias, Tensor topk_weights, Tensor topk_id, bool "
+      "skip_weighted, str act, str isa) -> ()");
+  ops.impl("cpu_fused_moe_int8", torch::kCPU, &cpu_fused_moe_int8);
+#endif  // #if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT) &&
+        // !defined(__APPLE__)
  ops.def(
      "mla_decode_kvcache("
      "   Tensor! out, Tensor query, Tensor kv_cache,"
@@ -1025,6 +1025,9 @@ __global__ void gather_and_maybe_dequant_cache(
    batch_offset += offset;
    int32_t block_table_id = batch_offset / block_size;
    int32_t slot_id = batch_offset % block_size;
+    // seq_starts may push the block index past the end of the batch's block
+    // table row.
+    if (block_table_id >= block_table_stride) continue;
    int32_t block_table_offset = batch_id * block_table_stride + block_table_id;
    int32_t block_id = block_table[block_table_offset];
    int64_t cache_offset =
@@ -9,14 +9,16 @@ void topk_softmax(torch::stable::Tensor& topk_weights,
                  torch::stable::Tensor& topk_indices,
                  torch::stable::Tensor& token_expert_indices,
                  torch::stable::Tensor& gating_output, bool renormalize,
-                  std::optional<torch::stable::Tensor> bias);
+                  std::optional<torch::stable::Tensor> bias,
+                  std::optional<torch::stable::Tensor> is_padding);

 void topk_sigmoid(torch::stable::Tensor& topk_weights,
                  torch::stable::Tensor& topk_indices,
                  torch::stable::Tensor& token_expert_indices,
                  torch::stable::Tensor& gating_output, bool renormalize,
                  std::optional<torch::stable::Tensor> bias,
-                  double routed_scaling_factor);
+                  double routed_scaling_factor,
+                  std::optional<torch::stable::Tensor> is_padding);

 void topk_softplus_sqrt(
    torch::stable::Tensor& topk_weights, torch::stable::Tensor& topk_indices,
@@ -25,7 +27,8 @@ void topk_softplus_sqrt(
    double routed_scaling_factor,
    const std::optional<torch::stable::Tensor>& correction_bias,
    const std::optional<torch::stable::Tensor>& input_ids,
-    const std::optional<torch::stable::Tensor>& tid2eid);
+    const std::optional<torch::stable::Tensor>& tid2eid,
+    const std::optional<torch::stable::Tensor>& is_padding);

 void moe_sum(torch::stable::Tensor& input, torch::stable::Tensor& output,
             std::optional<torch::stable::Tensor> topk_ids,
@@ -174,7 +174,8 @@ __launch_bounds__(TPB) __global__ void moeTopK(
    const int end_expert,
    const bool renormalize,
    const float* bias,
-    const double routed_scaling_factor)
+    const double routed_scaling_factor,
+    const bool* is_padding)
 {

    using cub_kvp = cub::KeyValuePair<int, float>;
@@ -228,12 +229,14 @@ __launch_bounds__(TPB) __global__ void moeTopK(
            const int expert = result_kvp.key;
            const bool node_uses_expert = expert >= start_expert && expert < end_expert;
            const bool should_process_row = row_is_active && node_uses_expert;
+            const bool is_pad_row = is_padding != nullptr && is_padding[block_row];

            const int idx = k * block_row + k_idx;
            // Return the unbiased scores for output weights
            output[idx] = inputs_after_softmax[thread_read_offset + expert];
-            indices[idx] = should_process_row ? (expert - start_expert) : num_experts;
-            assert(indices[idx] >= 0);
+            indices[idx] = is_pad_row ? static_cast<IndType>(-1)
+                                       : (should_process_row ? (expert - start_expert) : num_experts);
+            assert(is_pad_row || indices[idx] >= 0);
            source_rows[idx] = k_idx * num_rows + block_row;
            if (renormalize) {
                selected_sum += inputs_after_softmax[thread_read_offset + expert];
@@ -277,7 +280,7 @@ template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG, int WA
 __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
    void topkGating(const InputType* input, const bool* finished, float* output, const int num_rows, IndType* indices,
        int* source_rows, const int k, const int start_expert, const int end_expert, const bool renormalize,
-        const float* bias, const double routed_scaling_factor)
+        const float* bias, const double routed_scaling_factor, const bool* is_padding)
 {
    static_assert(std::is_same_v<InputType, float> || std::is_same_v<InputType, __nv_bfloat16> ||
                      std::is_same_v<InputType, __half>,
@@ -545,12 +548,14 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
            // Add a guard to ignore experts not included by this node
            const bool node_uses_expert = expert >= start_expert && expert < end_expert;
            const bool should_process_row = row_is_active && node_uses_expert;
+            const bool is_pad_row = is_padding != nullptr && is_padding[thread_row];

            // The lead thread from each sub-group will write out the final results to global memory. (This will be a
            // single) thread per row of the input/output matrices.
            const int idx = k * thread_row + k_idx;
            output[idx] = max_val;
-            indices[idx] = should_process_row ? (expert - start_expert) : NUM_EXPERTS;
+            indices[idx] = is_pad_row ? static_cast<IndType>(-1)
+                                       : (should_process_row ? (expert - start_expert) : NUM_EXPERTS);
            source_rows[idx] = k_idx * num_rows + thread_row;
            if (renormalize) {
                selected_sum += max_val;
@@ -605,7 +610,7 @@ struct TopkConstants
 template <int EXPERTS, int WARPS_PER_TB, int WARP_SIZE_PARAM, int MAX_BYTES_PER_LDG, typename IndType, typename InputType, ScoringFunc SF>
 void topkGatingLauncherHelper(const InputType* input, const bool* finished, float* output, IndType* indices,
    int* source_row, const int num_rows, const int k, const int start_expert, const int end_expert, const bool renormalize,
-    const float* bias, const double routed_scaling_factor, cudaStream_t stream)
+    const float* bias, const double routed_scaling_factor, cudaStream_t stream, const bool* is_padding)
 {
    static constexpr int BYTES_PER_LDG = MIN(MAX_BYTES_PER_LDG, sizeof(InputType) * EXPERTS);
    using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG, WARP_SIZE_PARAM, InputType>;
@@ -616,7 +621,7 @@ void topkGatingLauncherHelper(const InputType* input, const bool* finished, floa

    dim3 block_dim(WARP_SIZE_PARAM, WARPS_PER_TB);
    topkGating<VPT, EXPERTS, WARPS_PER_TB, BYTES_PER_LDG, WARP_SIZE_PARAM, IndType, InputType, SF><<<num_blocks, block_dim, 0, stream>>>(
-        input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert, renormalize, bias, routed_scaling_factor);
+        input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert, renormalize, bias, routed_scaling_factor, is_padding);
 }

 #ifndef USE_ROCM
@@ -627,7 +632,7 @@ void topkGatingLauncherHelper(const InputType* input, const bool* finished, floa
                             IndType, InputType, SF>(                         \
        gating_output, nullptr, topk_weights, topk_indices,                   \
        token_expert_indices, num_tokens, topk, 0, num_experts, renormalize,  \
-        bias, routed_scaling_factor, stream);
+        bias, routed_scaling_factor, stream, is_padding);
 #else
  #define LAUNCH_TOPK(NUM_EXPERTS, WARPS_PER_TB, MAX_BYTES)                    \
    if (WARP_SIZE == 64) {                                                     \
@@ -635,13 +640,13 @@ void topkGatingLauncherHelper(const InputType* input, const bool* finished, floa
                               IndType, InputType, SF>(                        \
          gating_output, nullptr, topk_weights, topk_indices,                  \
          token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
-          bias, routed_scaling_factor, stream);                                                       \
+          bias, routed_scaling_factor, stream, is_padding);                                            \
    } else if (WARP_SIZE == 32) {                                              \
      topkGatingLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 32, MAX_BYTES,       \
                               IndType, InputType, SF>(                        \
          gating_output, nullptr, topk_weights, topk_indices,                  \
          token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
-          bias, routed_scaling_factor, stream);                                                       \
+          bias, routed_scaling_factor, stream, is_padding);                                            \
    } else {                                                                   \
      assert(false &&                                                          \
             "Unsupported warp size. Only 32 and 64 are supported for ROCm");  \
@@ -661,7 +666,8 @@ void topkGatingKernelLauncher(
    const bool renormalize,
    const float* bias,
    const double routed_scaling_factor,
-    cudaStream_t stream) {
+    cudaStream_t stream,
+    const bool* is_padding) {
    static constexpr int WARPS_PER_TB = 4;
    static constexpr int BYTES_PER_LDG_POWER_OF_2 = 16;
 #ifndef USE_ROCM
@@ -736,7 +742,7 @@ void topkGatingKernelLauncher(
            }
            moeTopK<TPB><<<num_tokens, TPB, 0, stream>>>(
                workspace, nullptr, topk_weights, topk_indices, token_expert_indices,
-                num_experts, topk, 0, num_experts, renormalize, bias, routed_scaling_factor);
+                num_experts, topk, 0, num_experts, renormalize, bias, routed_scaling_factor, is_padding);
        }
    }
 }
@@ -755,7 +761,8 @@ void dispatch_topk_launch(
    int num_tokens, int num_experts, int topk, bool renormalize,
    std::optional<torch::stable::Tensor> bias,
    double routed_scaling_factor,
-    cudaStream_t stream)
+    cudaStream_t stream,
+    std::optional<torch::stable::Tensor> is_padding)
 {
    const float* bias_ptr = nullptr;
    if (bias.has_value()) {
@@ -769,6 +776,18 @@ void dispatch_topk_launch(
      bias_ptr = bias_tensor.const_data_ptr<float>();
    }

+    const bool* is_padding_ptr = nullptr;
+    if (is_padding.has_value()) {
+      const torch::stable::Tensor& is_padding_tensor = is_padding.value();
+      STD_TORCH_CHECK(is_padding_tensor.scalar_type() == torch::headeronly::ScalarType::Bool,
+                      "is_padding tensor must be bool");
+      STD_TORCH_CHECK(is_padding_tensor.dim() == 1, "is_padding tensor must be 1D");
+      STD_TORCH_CHECK(is_padding_tensor.size(0) == num_tokens,
+                      "is_padding size mismatch, expected: ", num_tokens);
+      STD_TORCH_CHECK(is_padding_tensor.is_contiguous(), "is_padding tensor must be contiguous");
+      is_padding_ptr = is_padding_tensor.const_data_ptr<bool>();
+    }
+
    if (topk_indices.scalar_type() == torch::headeronly::ScalarType::Int) {
        vllm::moe::topkGatingKernelLauncher<int, ComputeType, SF>(
            reinterpret_cast<const ComputeType*>(gating_output.const_data_ptr()),
@@ -777,7 +796,7 @@ void dispatch_topk_launch(
            token_expert_indices.mutable_data_ptr<int>(),
            softmax_workspace.mutable_data_ptr<float>(),
            num_tokens, num_experts, topk, renormalize,
-            bias_ptr, routed_scaling_factor, stream);
+            bias_ptr, routed_scaling_factor, stream, is_padding_ptr);
    } else if (topk_indices.scalar_type() == torch::headeronly::ScalarType::UInt32) {
        vllm::moe::topkGatingKernelLauncher<uint32_t, ComputeType, SF>(
            reinterpret_cast<const ComputeType*>(gating_output.const_data_ptr()),
@@ -786,7 +805,7 @@ void dispatch_topk_launch(
            token_expert_indices.mutable_data_ptr<int>(),
            softmax_workspace.mutable_data_ptr<float>(),
            num_tokens, num_experts, topk, renormalize,
-            bias_ptr, routed_scaling_factor, stream);
+            bias_ptr, routed_scaling_factor, stream, is_padding_ptr);
    } else {
        STD_TORCH_CHECK(topk_indices.scalar_type() == torch::headeronly::ScalarType::Long);
        vllm::moe::topkGatingKernelLauncher<int64_t, ComputeType, SF>(
@@ -796,7 +815,7 @@ void dispatch_topk_launch(
            token_expert_indices.mutable_data_ptr<int>(),
            softmax_workspace.mutable_data_ptr<float>(),
            num_tokens, num_experts, topk, renormalize,
-            bias_ptr, routed_scaling_factor, stream);
+            bias_ptr, routed_scaling_factor, stream, is_padding_ptr);
    }
 }

@@ -806,7 +825,8 @@ void topk_softmax(
    torch::stable::Tensor& token_expert_indices,        // [num_tokens, topk]
    torch::stable::Tensor& gating_output,               // [num_tokens, num_experts]
    bool renormalize,
-    std::optional<torch::stable::Tensor> bias)
+    std::optional<torch::stable::Tensor> bias,
+    std::optional<torch::stable::Tensor> is_padding)
 {
    const int num_experts = gating_output.size(-1);
    const auto num_tokens = gating_output.numel() / num_experts;
@@ -825,15 +845,15 @@ void topk_softmax(
    if (gating_output.scalar_type() == torch::headeronly::ScalarType::Float) {
        dispatch_topk_launch<float, vllm::moe::SCORING_SOFTMAX>(gating_output, topk_weights, topk_indices,
            token_expert_indices, softmax_workspace, num_tokens, num_experts, topk, renormalize,
-            bias, 1.0, stream);
+            bias, 1.0, stream, is_padding);
    } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::Half) {
        dispatch_topk_launch<__half, vllm::moe::SCORING_SOFTMAX>(gating_output, topk_weights, topk_indices,
            token_expert_indices, softmax_workspace, num_tokens, num_experts, topk, renormalize,
-            bias, 1.0, stream);
+            bias, 1.0, stream, is_padding);
    } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
        dispatch_topk_launch<__nv_bfloat16, vllm::moe::SCORING_SOFTMAX>(gating_output, topk_weights, topk_indices,
            token_expert_indices, softmax_workspace, num_tokens, num_experts, topk, renormalize,
-            bias, 1.0, stream);
+            bias, 1.0, stream, is_padding);
    } else {
        STD_TORCH_CHECK(false, "Unsupported gating_output data type: ", gating_output.scalar_type());
    }
@@ -846,7 +866,8 @@ void topk_sigmoid(
    torch::stable::Tensor& gating_output,               // [num_tokens, num_experts]
    bool renormalize,
    std::optional<torch::stable::Tensor> bias,
-    double routed_scaling_factor)
+    double routed_scaling_factor,
+    std::optional<torch::stable::Tensor> is_padding)
 {
    const int num_experts = gating_output.size(-1);
    const auto num_tokens = gating_output.numel() / num_experts;
@@ -865,15 +886,15 @@ void topk_sigmoid(
    if (gating_output.scalar_type() == torch::headeronly::ScalarType::Float) {
        dispatch_topk_launch<float, vllm::moe::SCORING_SIGMOID>(gating_output, topk_weights, topk_indices,
            token_expert_indices, workspace, num_tokens, num_experts, topk, renormalize,
-            bias, routed_scaling_factor, stream);
+            bias, routed_scaling_factor, stream, is_padding);
    } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::Half) {
        dispatch_topk_launch<__half, vllm::moe::SCORING_SIGMOID>(gating_output, topk_weights, topk_indices,
            token_expert_indices, workspace, num_tokens, num_experts, topk, renormalize,
-            bias, routed_scaling_factor, stream);
+            bias, routed_scaling_factor, stream, is_padding);
    } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
        dispatch_topk_launch<__nv_bfloat16, vllm::moe::SCORING_SIGMOID>(gating_output, topk_weights, topk_indices,
            token_expert_indices, workspace, num_tokens, num_experts, topk, renormalize,
-            bias, routed_scaling_factor, stream);
+            bias, routed_scaling_factor, stream, is_padding);
    } else {
        STD_TORCH_CHECK(false, "Unsupported gating_output data type: ", gating_output.scalar_type());
    }
@@ -80,22 +80,27 @@ __launch_bounds__(128) __global__
                                  OutIndType* indices, int num_rows,
                                  int num_experts, float routed_scaling_factor,
                                  const HashIndType* input_ids,
-                                  const HashIndType* tid2eid) {
+                                  const HashIndType* tid2eid,
+                                  const bool* is_padding) {
  const int warp = (blockIdx.x * blockDim.x + threadIdx.x) / 32;
  const int lane = threadIdx.x % 32;
  if (warp >= num_rows) return;
  const int64_t token_id = load_index_as_int64(input_ids, warp);
+  const bool is_pad_row = is_padding != nullptr && is_padding[warp];

  #if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
  cudaGridDependencySynchronize();
  #endif
  int expert = 0;
  float weight = 0.f;
-  if (lane < 6) {
+  if (lane < 6 && !is_pad_row) {
    // only load and calculate for 6 experts
    expert = static_cast<int>(tid2eid[token_id * 6 + lane]);
    const float x = input[warp * num_experts + expert];
    weight = sqrtf(fmaxf(x, 0.f) + __logf(1.f + __expf(-fabsf(x))));
+    if (isnan(weight)) {
+      weight = 0.f;
+    }
  }
  float weight_sum = weight;
  #pragma unroll
@@ -111,7 +116,8 @@ __launch_bounds__(128) __global__
    const int offset = warp * 6 + lane;
    output[offset] =
        weight * routed_scaling_factor / (weight_sum > 0.f ? weight_sum : 1.f);
-    indices[offset] = static_cast<OutIndType>(expert);
+    indices[offset] = !is_pad_row ? static_cast<OutIndType>(expert)
+                                  : static_cast<OutIndType>(-1);
  }
 }

@@ -120,7 +126,8 @@ void launchDsv4HashTopk(const float* input, float* output, OutIndType* indices,
                        int num_rows, int num_experts,
                        double routed_scaling_factor,
                        const HashIndType* input_ids,
-                        const HashIndType* tid2eid, cudaStream_t stream) {
+                        const HashIndType* tid2eid, cudaStream_t stream,
+                        const bool* is_padding) {
  if (num_rows == 0) return;
  auto* kernel = &dsv4HashTopkSoftplusSqrt<OutIndType, HashIndType>;
  cudaLaunchConfig_t config = {};
@@ -134,7 +141,7 @@ void launchDsv4HashTopk(const float* input, float* output, OutIndType* indices,
  config.numAttrs = 1;
  const float scale = static_cast<float>(routed_scaling_factor);
  cudaLaunchKernelEx(&config, kernel, input, output, indices, num_rows,
-                     num_experts, scale, input_ids, tid2eid);
+                     num_experts, scale, input_ids, tid2eid, is_padding);
 }
 #endif

@@ -166,7 +173,8 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
        const int num_rows, IndType* indices, int* source_rows, const int k,
        const int start_expert, const int end_expert, const bool renormalize,
        double routed_scaling_factor, const float* correction_bias,
-        const HashIndType* input_ids, const HashIndType* tid2eid) {
+        const HashIndType* input_ids, const HashIndType* tid2eid,
+        const bool* is_padding) {
  static_assert(std::is_same_v<InputType, float> ||
                    std::is_same_v<InputType, __nv_bfloat16> ||
                    std::is_same_v<InputType, __half>,
@@ -231,6 +239,7 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
    return;
  }
  const bool row_is_active = finished ? !finished[thread_row] : true;
+  const bool is_pad_row = is_padding != nullptr && is_padding[thread_row];

  // We finally start setting up the read pointers for each thread. First, each
  // thread jumps to the start of the row it will read.
@@ -249,9 +258,12 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
  cudaGridDependencySynchronize();
 #endif

-  // NOTE(zhuhaoran): dispatch different input types loading, BF16/FP16 convert
-  // to float
-  if constexpr (std::is_same_v<InputType, float>) {
+  if (is_pad_row) {
+#pragma unroll
+    for (int ii = 0; ii < VPT; ++ii) {
+      row_chunk[ii] = 0.f;
+    }
+  } else if constexpr (std::is_same_v<InputType, float>) {
    using VecType = AlignedArray<float, ELTS_PER_LDG>;
    VecType* row_chunk_vec_ptr = reinterpret_cast<VecType*>(&row_chunk);
    const VecType* vec_thread_read_ptr =
@@ -315,12 +327,22 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
  if constexpr (USE_HASH) {
    const int64_t token_id = load_index_as_int64(input_ids, thread_row);
    const int64_t token_expert_offset = token_id * static_cast<int64_t>(k);
+    if (!is_pad_row) {
 #pragma unroll
-    for (int ii = 0; ii < VPT; ++ii) {
-      float val = row_chunk[ii];
-      float val_b = val * beta;
-      val = (val_b > threshold) ? val : (__logf(1.0f + __expf(val_b))) / beta;
-      row_chunk[ii] = sqrtf(val);
+      for (int ii = 0; ii < VPT; ++ii) {
+        float val = row_chunk[ii];
+        float val_b = val * beta;
+        val = (val_b > threshold) ? val : (__logf(1.0f + __expf(val_b))) / beta;
+        val = sqrtf(val);
+
+        // Dummy/padding tokens can result in NaN values, so
+        // clamp them to 0.0. Note: this clamp could likely be removed if
+        // 'is_padding' is made mandatory
+        if (isnan(val)) {
+          val = 0.f;
+        }
+        row_chunk[ii] = val;
+      }
    }
    float selected_sum = 0.f;
 #pragma unroll
@@ -335,7 +357,8 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
                               group_id * THREADS_PER_ROW * ELTS_PER_LDG +
                               local_id;
        if (expert == expert_idx) {
-          indices[idx] = static_cast<IndType>(expert);
+          indices[idx] = !is_pad_row ? static_cast<IndType>(expert)
+                                     : static_cast<IndType>(-1);
          selected_sum += row_chunk[ii];
          break;
        }
@@ -379,23 +402,31 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
 #endif
    return;
  } else {
+    if (!is_pad_row) {
 #pragma unroll
-    for (int ii = 0; ii < VPT; ++ii) {
-      float val = row_chunk[ii];
-      float val_b = val * beta;
-      // Compute softplus: log(1 + exp(val)) with numerical stability
-      // When val > threshold, softplus(x) ≈ x to avoid exp overflow
-      val = (val_b > threshold) ? val : (__logf(1.0f + __expf(val_b))) / beta;
-      val = sqrtf(val);
-      if (correction_bias) {
-        const int group_id = ii / ELTS_PER_LDG;
-        const int local_id = ii % ELTS_PER_LDG;
-        const int expert_idx = first_elt_read_by_thread +
-                               group_id * THREADS_PER_ROW * ELTS_PER_LDG +
-                               local_id;
-        val = val + correction_bias[expert_idx];
+      for (int ii = 0; ii < VPT; ++ii) {
+        float val = row_chunk[ii];
+        float val_b = val * beta;
+        // Compute softplus: log(1 + exp(val)) with numerical stability
+        // When val > threshold, softplus(x) ≈ x to avoid exp overflow
+        val = (val_b > threshold) ? val : (__logf(1.0f + __expf(val_b))) / beta;
+        val = sqrtf(val);
+        // Dummy/padding tokens can result in NaN values, so
+        // clamp them to 0.0. Note: this clamp could likely be removed if
+        // 'is_padding' is made mandatory
+        if (isnan(val)) {
+          val = 0.f;
+        }
+        if (correction_bias) {
+          const int group_id = ii / ELTS_PER_LDG;
+          const int local_id = ii % ELTS_PER_LDG;
+          const int expert_idx = first_elt_read_by_thread +
+                                 group_id * THREADS_PER_ROW * ELTS_PER_LDG +
+                                 local_id;
+          val = val + correction_bias[expert_idx];
+        }
+        row_chunk[ii] = val;
      }
-      row_chunk[ii] = val;
    }

    // Original TopK path: find top-k experts by score
@@ -450,18 +481,19 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
        // Add a guard to ignore experts not included by this node
        const bool node_uses_expert =
            expert >= start_expert && expert < end_expert;
-        const bool should_process_row = row_is_active && node_uses_expert;
+        const bool should_process_row =
+            row_is_active && node_uses_expert && !is_pad_row;

        // The lead thread from each sub-group will write out the final results
        // to global memory. (This will be a single) thread per row of the
        // input/output matrices.
        const int idx = k * thread_row + k_idx;
-        if (correction_bias != nullptr) {
+        if (correction_bias != nullptr && should_process_row) {
          max_val -= correction_bias[expert];
        }
        output[idx] = max_val;
        indices[idx] =
-            should_process_row ? (expert - start_expert) : NUM_EXPERTS;
+            !is_pad_row ? expert - start_expert : static_cast<IndType>(-1);
        source_rows[idx] = k_idx * num_rows + thread_row;
        if (renormalize) {
          selected_sum += max_val;
@@ -544,7 +576,7 @@ void topkGatingSoftplusSqrtLauncherHelper(
    const int start_expert, const int end_expert, const bool renormalize,
    double routed_scaling_factor, const float* correction_bias,
    const bool use_hash, const HashIndType* input_ids,
-    const HashIndType* tid2eid, cudaStream_t stream) {
+    const HashIndType* tid2eid, cudaStream_t stream, const bool* is_padding) {
  static constexpr int BYTES_PER_LDG =
      MIN(MAX_BYTES_PER_LDG, sizeof(InputType) * EXPERTS);
  using Constants =
@@ -573,12 +605,12 @@ void topkGatingSoftplusSqrtLauncherHelper(
    cudaLaunchKernelEx(&config, kernel, input, finished, output, num_rows,
                       indices, source_row, k, start_expert, end_expert,
                       renormalize, routed_scaling_factor, correction_bias,
-                       input_ids, tid2eid);
+                       input_ids, tid2eid, is_padding);
 #else
    kernel<<<num_blocks, block_dim, 0, stream>>>(
        input, finished, output, num_rows, indices, source_row, k, start_expert,
        end_expert, renormalize, routed_scaling_factor, correction_bias,
-        input_ids, tid2eid);
+        input_ids, tid2eid, is_padding);
 #endif
  })
 }
@@ -592,7 +624,7 @@ void topkGatingSoftplusSqrtLauncherHelper(
        gating_output, nullptr, topk_weights, topk_indices,                    \
        token_expert_indices, num_tokens, topk, 0, num_experts, renormalize,   \
        routed_scaling_factor, correction_bias, use_hash, input_ids, tid2eid,  \
-        stream);
+        stream, is_padding);
 #else
  #define LAUNCH_SOFTPLUS_SQRT(NUM_EXPERTS, WARPS_PER_TB, MAX_BYTES)           \
    if (WARP_SIZE == 64) {                                                     \
@@ -601,14 +633,14 @@ void topkGatingSoftplusSqrtLauncherHelper(
          gating_output, nullptr, topk_weights, topk_indices,                  \
          token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
          routed_scaling_factor, correction_bias, use_hash, input_ids,         \
-          tid2eid, stream);                                                    \
+          tid2eid, stream, is_padding);                                        \
    } else if (WARP_SIZE == 32) {                                              \
      topkGatingSoftplusSqrtLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 32,      \
                                           MAX_BYTES>(                         \
          gating_output, nullptr, topk_weights, topk_indices,                  \
          token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
          routed_scaling_factor, correction_bias, use_hash, input_ids,         \
-          tid2eid, stream);                                                    \
+          tid2eid, stream, is_padding);                                        \
    } else {                                                                   \
      assert(false &&                                                          \
             "Unsupported warp size. Only 32 and 64 are supported for ROCm");  \
@@ -622,14 +654,14 @@ void topkGatingSoftplusSqrtKernelLauncher(
    const int topk, const bool renormalize, double routed_scaling_factor,
    const float* correction_bias, const bool use_hash,
    const HashIndType* input_ids, const HashIndType* tid2eid,
-    cudaStream_t stream) {
+    cudaStream_t stream, const bool* is_padding) {
 #ifndef USE_ROCM
  if constexpr (std::is_same_v<InputType, float>) {
    if (use_hash && topk == 6 && renormalize &&
        (num_experts == 256 || num_experts == 384)) {
      launchDsv4HashTopk<IndType, HashIndType>(
          gating_output, topk_weights, topk_indices, num_tokens, num_experts,
-          routed_scaling_factor, input_ids, tid2eid, stream);
+          routed_scaling_factor, input_ids, tid2eid, stream, is_padding);
      return;
    }
  }
@@ -728,7 +760,8 @@ void dispatch_topk_softplus_sqrt_launch(
    int num_experts, int topk, bool renormalize, double routed_scaling_factor,
    const std::optional<torch::stable::Tensor>& correction_bias,
    const std::optional<torch::stable::Tensor>& input_ids,
-    const std::optional<torch::stable::Tensor>& tid2eid, cudaStream_t stream) {
+    const std::optional<torch::stable::Tensor>& tid2eid, cudaStream_t stream,
+    const std::optional<torch::stable::Tensor>& is_padding) {
  const float* bias_ptr = nullptr;
  if (correction_bias.has_value()) {
    bias_ptr = correction_bias.value().const_data_ptr<float>();
@@ -737,6 +770,22 @@ void dispatch_topk_softplus_sqrt_launch(
  auto launch = [&](auto* topk_indices_ptr) {
    using OutIndType =
        typename std::remove_pointer<decltype(topk_indices_ptr)>::type;
+
+    const bool* is_padding_ptr = nullptr;
+    if (is_padding.has_value()) {
+      const torch::stable::Tensor& is_padding_tensor = is_padding.value();
+      STD_TORCH_CHECK(is_padding_tensor.scalar_type() ==
+                          torch::headeronly::ScalarType::Bool,
+                      "is_padding tensor must be bool");
+      STD_TORCH_CHECK(is_padding_tensor.dim() == 1,
+                      "is_padding tensor must be 1D");
+      STD_TORCH_CHECK(is_padding_tensor.size(0) == num_tokens,
+                      "is_padding size mismatch, expected: ", num_tokens);
+      STD_TORCH_CHECK(is_padding_tensor.is_contiguous(),
+                      "is_padding tensor must be contiguous");
+      is_padding_ptr = is_padding_tensor.const_data_ptr<bool>();
+    }
+
    if (tid2eid.has_value()) {
      STD_TORCH_CHECK(input_ids.has_value(),
                      "input_ids is required for hash MoE");
@@ -751,7 +800,7 @@ void dispatch_topk_softplus_sqrt_launch(
            topk_indices_ptr, token_expert_indices.mutable_data_ptr<int>(),
            num_tokens, num_experts, topk, renormalize, routed_scaling_factor,
            bias_ptr, true, input_ids.value().const_data_ptr<int64_t>(),
-            tid2eid.value().const_data_ptr<int64_t>(), stream);
+            tid2eid.value().const_data_ptr<int64_t>(), stream, is_padding_ptr);
      } else {
        STD_TORCH_CHECK(tid2eid.value().scalar_type() ==
                        torch::headeronly::ScalarType::Int);
@@ -761,7 +810,7 @@ void dispatch_topk_softplus_sqrt_launch(
            topk_indices_ptr, token_expert_indices.mutable_data_ptr<int>(),
            num_tokens, num_experts, topk, renormalize, routed_scaling_factor,
            bias_ptr, true, input_ids.value().const_data_ptr<int>(),
-            tid2eid.value().const_data_ptr<int>(), stream);
+            tid2eid.value().const_data_ptr<int>(), stream, is_padding_ptr);
      }
    } else {
      vllm::moe::topkGatingSoftplusSqrtKernelLauncher<OutIndType, ComputeType>(
@@ -769,7 +818,7 @@ void dispatch_topk_softplus_sqrt_launch(
          topk_indices_ptr, token_expert_indices.mutable_data_ptr<int>(),
          num_tokens, num_experts, topk, renormalize, routed_scaling_factor,
          bias_ptr, false, static_cast<const OutIndType*>(nullptr),
-          static_cast<const OutIndType*>(nullptr), stream);
+          static_cast<const OutIndType*>(nullptr), stream, is_padding_ptr);
    }
  };

@@ -793,7 +842,8 @@ void topk_softplus_sqrt(
    bool renormalize, double routed_scaling_factor,
    const std::optional<torch::stable::Tensor>& correction_bias,
    const std::optional<torch::stable::Tensor>& input_ids,
-    const std::optional<torch::stable::Tensor>& tid2eid) {
+    const std::optional<torch::stable::Tensor>& tid2eid,
+    const std::optional<torch::stable::Tensor>& is_padding) {
  const int num_experts = gating_output.size(-1);
  const auto num_tokens = gating_output.numel() / num_experts;
  const int topk = topk_weights.size(-1);
@@ -806,21 +856,22 @@ void topk_softplus_sqrt(
    dispatch_topk_softplus_sqrt_launch<float>(
        gating_output.const_data_ptr<float>(), topk_weights, topk_indices,
        token_expert_indices, num_tokens, num_experts, topk, renormalize,
-        routed_scaling_factor, correction_bias, input_ids, tid2eid, stream);
+        routed_scaling_factor, correction_bias, input_ids, tid2eid, stream,
+        is_padding);
  } else if (gating_output.scalar_type() ==
             torch::headeronly::ScalarType::Half) {
    dispatch_topk_softplus_sqrt_launch<__half>(
        reinterpret_cast<const __half*>(gating_output.const_data_ptr()),
        topk_weights, topk_indices, token_expert_indices, num_tokens,
        num_experts, topk, renormalize, routed_scaling_factor, correction_bias,
-        input_ids, tid2eid, stream);
+        input_ids, tid2eid, stream, is_padding);
  } else if (gating_output.scalar_type() ==
             torch::headeronly::ScalarType::BFloat16) {
    dispatch_topk_softplus_sqrt_launch<__nv_bfloat16>(
        reinterpret_cast<const __nv_bfloat16*>(gating_output.const_data_ptr()),
        topk_weights, topk_indices, token_expert_indices, num_tokens,
        num_experts, topk, renormalize, routed_scaling_factor, correction_bias,
-        input_ids, tid2eid, stream);
+        input_ids, tid2eid, stream, is_padding);
  } else {
    STD_TORCH_CHECK(false, "Unsupported gating_output data type: ",
                    gating_output.scalar_type());
@@ -8,19 +8,19 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_moe_C, m) {
  m.def(
      "topk_softmax(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
      "token_expert_indices, Tensor gating_output, bool renormalize, Tensor? "
-      "bias) -> ()");
+      "bias, Tensor? is_padding) -> ()");

  // Apply topk sigmoid to the gating outputs.
  m.def(
      "topk_sigmoid(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
      "token_expert_indices, Tensor gating_output, bool renormalize, "
-      "Tensor? bias, float routed_scaling_factor) -> ()");
+      "Tensor? bias, float routed_scaling_factor, Tensor? is_padding) -> ()");

  m.def(
      "topk_softplus_sqrt(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
      "token_expert_indices, Tensor gating_output, bool renormalize, float "
      "routed_scaling_factor, Tensor? "
-      "bias, Tensor? input_ids, Tensor? tid2eid) -> ()");
+      "bias, Tensor? input_ids, Tensor? tid2eid, Tensor? is_padding) -> ()");

  // Calculate the result of moe by summing up the partial results
  // from all selected experts. topk_ids/expert_map are optional and, when
@@ -22,17 +22,25 @@ template <typename AllReduceKernel, typename T>
 __global__ __quickreduce_launch_bounds_two_shot__ static void
 allreduce_prototype_twoshot(T const* A, T* B, uint32_t N, uint32_t num_blocks,
                            int rank, uint8_t** dbuffer_list,
-                            uint32_t data_offset, uint32_t flag_color,
+                            uint32_t data_offset, uint32_t* d_flag_counters,
                            int64_t data_size_per_phase) {
  int block = blockIdx.x;
  int grid = gridDim.x;

+  // Load this block's counter from device memory and advance it on-device,
+  // so the color keeps changing across graph replays instead of being frozen.
+  uint32_t flag_color = d_flag_counters[blockIdx.x];
+
  while (block < num_blocks) {
    AllReduceKernel::run(A, B, N, block, rank, dbuffer_list, data_offset,
                         flag_color, data_size_per_phase);
    block += grid;
    flag_color++;
  }
+  // All threads compute the same final value; one writer per block is enough.
+  if (threadIdx.x == 0 && threadIdx.y == 0) {
+    d_flag_counters[blockIdx.x] = flag_color;
+  }
 }

 #define TWOSHOT_DISPATCH(__codec)                                           \
@@ -42,21 +50,21 @@ allreduce_prototype_twoshot(T const* A, T* B, uint32_t N, uint32_t num_blocks,
    hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                       dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                       num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
  } else if (world_size == 4) {                                             \
    using LineCodec = __codec<T, 4>;                                        \
    using AllReduceKernel = AllReduceTwoshot<T, LineCodec, cast_bf2half>;   \
    hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                       dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                       num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
  } else if (world_size == 8) {                                             \
    using LineCodec = __codec<T, 8>;                                        \
    using AllReduceKernel = AllReduceTwoshot<T, LineCodec, cast_bf2half>;   \
    hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                       dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                       num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
  }

 // INT3 only retains good performance on TP2 (world_size == 2). On TP4/TP8
@@ -69,7 +77,7 @@ allreduce_prototype_twoshot(T const* A, T* B, uint32_t N, uint32_t num_blocks,
    hipLaunchKernelGGL((allreduce_prototype_twoshot<AllReduceKernel, T>),   \
                       dim3(grid), dim3(kBlockTwoShot), 0, stream, A, B, N, \
                       num_blocks, rank, dbuffer_list, data_offset,         \
-                       flag_color, this->kMaxProblemSize);                  \
+                       d_flag_counters, this->kMaxProblemSize);             \
  } else {                                                                  \
    throw std::runtime_error(                                               \
        "INT3 quick all-reduce is only supported for world_size == 2 "      \
@@ -94,7 +102,7 @@ struct DeviceComms {
  static int constexpr kMaxWorldSize = 8;

  bool initialized = false;
-  uint32_t flag_color = 1;
+  uint32_t* d_flag_counters = nullptr;
  int world_size;
  int rank;

@@ -128,6 +136,16 @@ struct DeviceComms {
    // Clear the flags buffer.
    HIP_CHECK(hipMemset(dbuffer, 0, flags_buffer_size));

+    // One flag-color counter per block, advanced by the kernel. Start at 1
+    // to stay clear of the flags buffer we just zeroed.
+    HIP_CHECK(hipMalloc(&d_flag_counters, kMaxNumBlocks * sizeof(uint32_t)));
+    {
+      std::vector<uint32_t> init_color(kMaxNumBlocks, 1u);
+      HIP_CHECK(hipMemcpy(d_flag_counters, init_color.data(),
+                          kMaxNumBlocks * sizeof(uint32_t),
+                          hipMemcpyHostToDevice));
+    }
+
    // Device-side list of IPC buffers.
    buffer_list.resize(world_size);
    HIP_CHECK(hipMalloc(&dbuffer_list, world_size * sizeof(uint8_t*)));
@@ -144,6 +162,12 @@ struct DeviceComms {
  hipIpcMemHandle_t const get_handle() { return buffer_ipc_handle; }

  void destroy() {
+    // Allocated before `initialized` flips true, so free it on its own guard
+    // to avoid a leak if init fails partway through.
+    if (d_flag_counters) {
+      HIP_CHECK(hipFree(d_flag_counters));
+      d_flag_counters = nullptr;
+    }
    if (initialized) {
      for (int i = 0; i < world_size; i++) {
        if (i != rank) {
@@ -211,8 +235,6 @@ struct DeviceComms {
        break;
    }
    HIP_CHECK(cudaGetLastError());
-    // Rotate the flag color.
-    flag_color += divceil(N, grid);
  }
 };

@@ -22,9 +22,13 @@
 #   docker buildx bake -f docker/docker-bake.hcl -f docker/versions.json
 # =============================================================================

-ARG CUDA_VERSION=13.0.2
+ARG CUDA_VERSION=13.0.3
 ARG PYTHON_VERSION=3.12
 ARG UBUNTU_VERSION=22.04
+# DeepEPv2 requires NCCL >= 2.30.4 (GIN backend).
+# This version is only used for CUDA 13+ builds; CUDA 12 falls back to
+# the default NCCL version shipped with the base image.
+ARG NCCL_VERSION=2.30.7

 # By parameterizing the base images, we allow third-party to use their own
 # base images. One use case is hermetic builds with base images stored in
@@ -477,10 +481,17 @@ WORKDIR /workspace
 # Build DeepEP wheels
 COPY tools/ep_kernels/install_python_libraries.sh /tmp/install_python_libraries.sh
 # Defaults moved here from tools/ep_kernels/install_python_libraries.sh for centralized version management
-ARG DEEPEP_COMMIT_HASH=73b6ea4
+ARG DEEPEP_COMMIT_HASH=d4f41e4e93
 ARG NVSHMEM_VER
+ARG NCCL_VERSION
 RUN --mount=type=cache,target=/opt/uv/cache \
    mkdir -p /tmp/ep_kernels_workspace/dist && \
+    CUDA_MAJOR=$(echo $CUDA_VERSION | cut -d. -f1) && \
+    if [ "$CUDA_MAJOR" -ge 13 ] && [ -n "$NCCL_VERSION" ]; then \
+        echo "nvidia-nccl-cu${CUDA_MAJOR}==${NCCL_VERSION}" \
+            > /tmp/nccl-override.txt && \
+        export UV_OVERRIDE=/tmp/nccl-override.txt; \
+    fi && \
    export TORCH_CUDA_ARCH_LIST='9.0a 10.0a' && \
    /tmp/install_python_libraries.sh \
        --workspace /tmp/ep_kernels_workspace \
@@ -644,6 +655,7 @@ FROM ${FINAL_BASE_IMAGE} AS vllm-base

 ARG CUDA_VERSION
 ARG PYTHON_VERSION
+ARG NCCL_VERSION
 ARG DEADSNAKES_MIRROR_URL
 ARG DEADSNAKES_GPGKEY_URL
 ARG GET_PIP_URL
@@ -696,7 +708,6 @@ RUN apt-get update -y \
 # Install CUDA development tools for runtime JIT compilation
 # (FlashInfer, DeepGEMM, EP kernels all require compilation at runtime)
 RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
-    CUDA_VERSION_SHORT=$(echo $CUDA_VERSION | cut -d. -f1,2) && \
    apt-get update -y && \
    apt-get install -y --no-install-recommends --allow-change-held-packages \
        cuda-nvcc-${CUDA_VERSION_DASH} \
@@ -709,12 +720,6 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
        libnuma-dev \
        # numactl CLI for NUMA binding at runtime
        numactl && \
-    # Fixes nccl_allocator requiring nccl.h at runtime
-    # https://github.com/vllm-project/vllm/blob/1336a1ea244fa8bfd7e72751cabbdb5b68a0c11a/vllm/distributed/device_communicators/pynccl_allocator.py#L22
-    # NCCL packages don't use the cuda-MAJOR-MINOR naming convention,
-    # so we pin the version to match our CUDA version
-    NCCL_VER=$(apt-cache madison libnccl-dev | grep "+cuda${CUDA_VERSION_SHORT}" | head -1 | awk -F'|' '{gsub(/^ +| +$/, "", $2); print $2}') && \
-    apt-get install -y --no-install-recommends --allow-change-held-packages libnccl-dev=${NCCL_VER} libnccl2=${NCCL_VER} && \
    rm -rf /var/lib/apt/lists/*

 # Install uv for faster pip installs
@@ -734,6 +739,18 @@ RUN mkdir -p "${UV_PYTHON_INSTALL_DIR}" "${UV_CACHE_DIR}" \
    && chgrp -R 0 /opt/uv \
    && chmod -R g+rwX,a+rX /opt/uv

+# DeepEPv2 GIN requires NCCL >= 2.30.4 at both compile and runtime. torch pins
+# an older version as a transitive dep; this override forces uv to use our
+# pinned version whenever nvidia-nccl-cu* is resolved. Empty on CUDA 12 (no-op).
+RUN CUDA_MAJOR=$(echo $CUDA_VERSION | cut -d. -f1) && \
+    if [ "$CUDA_MAJOR" -ge 13 ]; then \
+        echo "nvidia-nccl-cu${CUDA_MAJOR}==${NCCL_VERSION}" \
+            > /etc/uv-overrides.txt; \
+    else \
+        touch /etc/uv-overrides.txt; \
+    fi
+ENV UV_OVERRIDE=/etc/uv-overrides.txt
+
 # ----------------------------------------------------------------------
 # Non-root support (opt-in)
 # ----------------------------------------------------------------------
@@ -793,7 +810,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \
 # Install FlashInfer JIT cache (requires CUDA-version-specific index URL)
 # https://docs.flashinfer.ai/installation.html
 # From versions.json: .flashinfer.version
-ARG FLASHINFER_VERSION=0.6.14
+ARG FLASHINFER_VERSION=0.6.15.post1
 RUN --mount=type=cache,target=/opt/uv/cache \
    uv pip install --system flashinfer-jit-cache==${FLASHINFER_VERSION} \
        --index-url https://flashinfer.ai/whl/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
@@ -339,18 +339,17 @@ COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/rust /rust
 COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/rust-toolchain.toml /rust-toolchain.toml
 COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1

-# RIXL/UCX build stages
-FROM base AS build_rixl
-ARG RIXL_BRANCH="39be1de8"
-ARG RIXL_REPO="https://github.com/ROCm/RIXL.git"
-ARG UCX_BRANCH="bfb51733"
+# NIXL/UCX build stages
+FROM base AS build_nixl
+ARG NIXL_BRANCH="231d56753047c989062a5cb2ac703a1ad761c7d2"
+ARG NIXL_REPO="https://github.com/ai-dynamo/nixl.git"
+ARG UCX_BRANCH="96e58a16039f6d7d213bc967b8069238742c5194"
 ARG UCX_REPO="https://github.com/openucx/ucx.git"
 ENV ROCM_PATH=/opt/rocm
 ENV UCX_HOME=/usr/local/ucx
-ENV RIXL_HOME=/usr/local/rixl
-ENV RIXL_BENCH_HOME=/usr/local/rixl_bench
+ENV NIXL_HOME=/usr/local/nixl

-# RIXL build system dependences and RDMA support
+# NIXL build system dependencies and RDMA support
 RUN apt-get -y update && apt-get -y install autoconf libtool pkg-config \
    libgrpc-dev \
    libgrpc++-dev \
@@ -368,7 +367,8 @@ RUN apt-get -y update && apt-get -y install autoconf libtool pkg-config \
    && rm -rf /var/lib/apt/lists/*

 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system meson auditwheel patchelf tomlkit
+    uv pip install --system meson meson-python pybind11 pyyaml types-PyYAML \
+        auditwheel build patchelf pytest tomlkit "setuptools>=80.9.0"

 RUN --mount=type=cache,target=/root/.cache/ccache \
    cd /usr/local/src && \
@@ -396,30 +396,50 @@ ENV PATH=/usr/local/ucx/bin:$PATH
 ENV LD_LIBRARY_PATH=${UCX_HOME}/lib:${LD_LIBRARY_PATH}

 RUN --mount=type=cache,target=/root/.cache/ccache \
-    git clone ${RIXL_REPO} /opt/rixl && \
-    cd /opt/rixl && \
-    git checkout ${RIXL_BRANCH} && \
+    git clone ${NIXL_REPO} /opt/nixl && \
+    cd /opt/nixl && \
+    git checkout ${NIXL_BRANCH} && \
    CC="ccache gcc" CXX="ccache g++" \
-    meson setup build --prefix=${RIXL_HOME} \
+    meson setup build --prefix=${NIXL_HOME} \
                     -Ducx_path=${UCX_HOME} \
-                     -Drocm_path=${ROCM_PATH} && \
+                     -Dwheel_variant=rocm \
+                     -Dbuild_tests=false \
+                     -Dbuild_examples=false && \
    cd build && \
    ninja -j$(nproc) && \
-    ninja install
+    ninja install && \
+    echo "${NIXL_HOME}/lib/$(uname -m)-linux-gnu" \
+        > /etc/ld.so.conf.d/nixl.conf && \
+    echo "${NIXL_HOME}/lib/$(uname -m)-linux-gnu/plugins" \
+        >> /etc/ld.so.conf.d/nixl.conf && \
+    ldconfig

-# Generate RIXL wheel
+# Generate the ROCm NIXL wheel. Upstream's generic wheel helper detects CUDA,
+# so configure the ROCm wheel variant directly through Meson.
 # Exclude libcore and libpull from auditwheel: transitive dependencies
 # that are not shipped in the wheel and vary across base images.
-RUN cd /opt/rixl && \
-    sed -i "s/--exclude 'libamdhip64\*'/--exclude 'libamdhip64*' --exclude 'libcore*' --exclude 'libpull*'/" \
-        contrib/build-wheel.sh && \
-    mkdir -p /app/install && \
-    _ucx_install_dir=${UCX_HOME} \
-    ./contrib/build-wheel.sh \
-        --output-dir /app/install \
-        --rocm-dir ${ROCM_PATH} \
+RUN cd /opt/nixl && \
+    ./contrib/tomlutil.py --wheel-name nixl-rocm pyproject.toml && \
+    CC="ccache gcc" CXX="ccache g++" \
+    uv build --wheel --no-build-isolation --out-dir /tmp/nixl_wheels \
+        --python ${PYTHON_VERSION} \
+        -Csetup-args=-Ducx_path=${UCX_HOME} \
+        -Csetup-args=-Dwheel_variant=rocm \
+        -Csetup-args=-Dbuild_tests=false \
+        -Csetup-args=-Dbuild_examples=false && \
+    mkdir -p /tmp/nixl_wheels/repaired /app/install && \
+    auditwheel repair \
+        --exclude 'libamdhip64*' \
+        --exclude 'libcore*' \
+        --exclude 'libpull*' \
+        /tmp/nixl_wheels/nixl_rocm*.whl \
+        --plat manylinux_2_34_$(uname -m) \
+        --wheel-dir /tmp/nixl_wheels/repaired && \
+    ./contrib/wheel_add_ucx_plugins.py \
        --ucx-plugins-dir ${UCX_HOME}/lib/ucx \
-        --nixl-plugins-dir ${RIXL_HOME}/lib/x86_64-linux-gnu/plugins
+        --nixl-plugins-dir ${NIXL_HOME}/lib/$(uname -m)-linux-gnu/plugins \
+        /tmp/nixl_wheels/repaired/*.whl && \
+    cp /tmp/nixl_wheels/repaired/*.whl /app/install

 # ROCShmem build stage - split from DeepEP so changing DEEPEP_BRANCH does not
 # invalidate the slow ROCShmem build.
@@ -660,10 +680,10 @@ RUN if [ "${DEEPEP_NIC}" = "cx7" ] || [ "${DEEPEP_NIC}" = "io" ]; then \
    ninja && ninja install && ldconfig && rm -rf /tmp/rdma-core; \
 fi

-# Install RIXL + DeepEP wheels.
-RUN --mount=type=bind,from=build_rixl,src=/app/install,target=/rixl_install \
+# Install NIXL + DeepEP wheels.
+RUN --mount=type=bind,from=build_nixl,src=/app/install,target=/nixl_install \
    --mount=type=bind,from=build_deepep,src=/app/deep_install,target=/deep_install \
-    uv pip install --system /rixl_install/*.whl /deep_install/*.whl
+    uv pip install --system /nixl_install/*.whl /deep_install/*.whl

 # Copy ROCShmem runtime libraries.
 COPY --from=build_rocshmem /opt/rocshmem /opt/rocshmem
@@ -724,6 +744,8 @@ ENV MIOPEN_DEBUG_CONV_GEMM=0
 # Use legacy IPC mode for HSA to avoid GPU memory pinning issues with UCX rocm_ipc.
 # See: https://github.com/ROCm/rocm-libraries/issues/6266
 ENV HSA_ENABLE_IPC_MODE_LEGACY=1
+ENV UCX_RMA_PPLN_ENABLE=y
+ENV UCX_ROCM_COPY_SIGPOOL_MAX_ELEMS=inf

 # ROCm profiler limits workaround.
 RUN echo "ROCTRACER_MAX_EVENTS=10000000" > ${COMMON_WORKDIR}/libkineto.conf
@@ -796,9 +818,9 @@ RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
    && pip uninstall -y vllm \
    && uv pip install --system *.whl

-# Install RIXL wheel
-RUN --mount=type=bind,from=build_rixl,src=/app/install,target=/rixl_install \
-    uv pip install --system /rixl_install/*.whl
+# Install NIXL ROCm wheel
+RUN --mount=type=bind,from=build_nixl,src=/app/install,target=/nixl_install \
+    uv pip install --system /nixl_install/*.whl

 ARG COMMON_WORKDIR
 ARG BASE_IMAGE
@@ -813,6 +835,8 @@ COPY --from=export_vllm /docker ${COMMON_WORKDIR}/vllm/docker
 # Use legacy IPC mode for HSA to avoid GPU memory pinning issues with UCX rocm_ipc
 # See: https://github.com/ROCm/rocm-libraries/issues/6266
 ENV HSA_ENABLE_IPC_MODE_LEGACY=1
+ENV UCX_RMA_PPLN_ENABLE=y
+ENV UCX_ROCM_COPY_SIGPOOL_MAX_ELEMS=inf

 ENV TOKENIZERS_PARALLELISM=false

@@ -9,7 +9,7 @@ ARG PYTORCH_AUDIO_BRANCH="v2.9.0"
 ARG PYTORCH_AUDIO_REPO="https://github.com/pytorch/audio.git"
 ARG FA_BRANCH="0e60e394"
 ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
-ARG AITER_BRANCH="v0.1.16.post3"
+ARG AITER_BRANCH="v0.1.16.post5"
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 ARG MORI_BRANCH="v1.1.0"
 ARG MORI_REPO="https://github.com/ROCm/mori.git"
@@ -30,7 +30,7 @@ ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
 ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151
 ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
 ENV AITER_ROCM_ARCH=gfx942;gfx950
-ENV MORI_GPU_ARCHS=gfx942;gfx950
+# Note: Do not set MORI_GPU_ARCHS here, it is automatically inferred at runtime

 # Required for RCCL in ROCm7.1
 ENV HSA_NO_SCRATCH_RECLAIM=1
@@ -86,6 +86,29 @@ RUN --mount=type=cache,target=/root/.cache/uv \
    mkdir -p /tmp/hf-xet/dist && \
    cp dist/*.whl /tmp/hf-xet/dist/

+# Build LLVM 20 from source for llvmlite (system repos ship LLVM 21 which
+# llvmlite v0.47 does not support; only SystemZ target is needed).
+FROM base AS llvm20-build
+ARG LLVM_VERSION=20.1.8
+WORKDIR /tmp
+RUN microdnf install -y ninja-build gcc gcc-c++ python3 xz && \
+    curl -LO https://github.com/llvm/llvm-project/releases/download/llvmorg-${LLVM_VERSION}/llvm-project-${LLVM_VERSION}.src.tar.xz && \
+    tar -xf llvm-project-${LLVM_VERSION}.src.tar.xz && \
+    cmake -G Ninja -S llvm-project-${LLVM_VERSION}.src/llvm -B build \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_INSTALL_PREFIX=/opt/llvm20 \
+        -DLLVM_TARGETS_TO_BUILD="SystemZ" \
+        -DLLVM_ENABLE_RTTI=ON \
+        -DLLVM_BUILD_TOOLS=OFF \
+        -DLLVM_BUILD_UTILS=ON \
+        -DLLVM_BUILD_EXAMPLES=OFF \
+        -DLLVM_BUILD_TESTS=OFF \
+        -DLLVM_INCLUDE_TESTS=OFF \
+        -DLLVM_INCLUDE_EXAMPLES=OFF \
+        -DLLVM_INCLUDE_BENCHMARKS=OFF && \
+    ninja -C build install && \
+    rm -rf build llvm-project-${LLVM_VERSION}.src*
+
 # Build numba
 FROM python-install AS numba-builder

@@ -96,11 +119,13 @@ WORKDIR /tmp

 # Clone all required dependencies
 RUN --mount=type=cache,target=/root/.cache/uv \
+    --mount=type=bind,from=llvm20-build,source=/opt/llvm20,target=/opt/llvm20 \
    microdnf install ninja-build gcc gcc-c++ -y && \
    git clone --recursive https://github.com/numba/llvmlite.git -b v0.47.0 && \
    git clone --recursive https://github.com/numba/numba.git -b ${NUMBA_VERSION} && \
    cd llvmlite && \
    uv pip install 'cmake<4' 'setuptools<70' numpy && \
+    CMAKE_PREFIX_PATH=/opt/llvm20 LLVM_CONFIG=/opt/llvm20/bin/llvm-config \
    python setup.py bdist_wheel && \
    cd ../numba && \
    if ! grep '#include "dynamic_annotations.h"' numba/_dispatcher.cpp; then \
@@ -158,7 +183,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     NUMBA_WHL_FILE=$(ls /tmp/numba-wheels/*.whl) && \
     OPENCV_WHL_FILE=$(ls /tmp/opencv-wheels/*.whl) && \
     uv pip install -v \
-        $ARROW_WHL_FILE \
        $VISION_WHL_FILE \
        $HF_XET_WHL_FILE \
        $LLVM_WHL_FILE \
@@ -59,7 +59,7 @@ variable "PYTORCH_ROCM_ARCH" {
 }

 # Pre-built CI base image (Tier 1). Per-PR builds pull this instead of
-# rebuilding RIXL/DeepEP/torchcodec from scratch. The ci_base stage in
+# rebuilding NIXL/DeepEP/torchcodec from scratch. The ci_base stage in
 # Dockerfile.rocm inherits from base, so CI_BASE_IMAGE only affects the test
 # stage and is irrelevant when building --target ci_base itself.
 variable "CI_BASE_IMAGE" {
@@ -75,7 +75,7 @@ variable "CI_MAX_JOBS" {
 # Upstream dependency commit pins -- extracted from Dockerfile.rocm by
 # ci-bake-rocm.sh at build time. Empty defaults are safe: the cache
 # functions produce no entries when the variable is empty.
-variable "RIXL_BRANCH" {
+variable "NIXL_BRANCH" {
  default = ""
 }

@@ -91,7 +91,7 @@ variable "DEEPEP_BRANCH" {
  default = ""
 }

-variable "RIXL_CACHE_KEY" {
+variable "NIXL_CACHE_KEY" {
  default = ""
 }

@@ -236,7 +236,7 @@ function "get_cache_to_rocm_rust" {
  ])
 }

-# Cache functions for upstream dependency stages (RIXL/UCX, ROCShmem, DeepEP).
+# Cache functions for upstream dependency stages (NIXL/UCX, ROCShmem, DeepEP).
 # These stages are pinned to specific upstream commit hashes, so cache keys use
 # those hashes rather than the Buildkite commit. This means the cache persists
 # across all vLLM commits as long as the upstream dependency pins don't change.
@@ -244,16 +244,16 @@ function "get_cache_to_rocm_rust" {
 function "get_cache_from_rocm_deps" {
  params = []
  result = compact([
-    RIXL_CACHE_KEY != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rixl-rocm-${RIXL_CACHE_KEY}" : (RIXL_BRANCH != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rixl-rocm-${RIXL_BRANCH}-ucx-${UCX_BRANCH}" : ""),
+    NIXL_CACHE_KEY != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:nixl-rocm-${NIXL_CACHE_KEY}" : (NIXL_BRANCH != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:nixl-rocm-${NIXL_BRANCH}-ucx-${UCX_BRANCH}" : ""),
    ROCSHMEM_CACHE_KEY != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rocshmem-rocm-${ROCSHMEM_CACHE_KEY}" : (ROCSHMEM_BRANCH != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rocshmem-rocm-${ROCSHMEM_BRANCH}" : ""),
    DEEPEP_CACHE_KEY != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:deepep-rocm-${DEEPEP_CACHE_KEY}" : (DEEPEP_BRANCH != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:deepep-rocm-${DEEPEP_BRANCH}-rocshmem-${ROCSHMEM_BRANCH}" : ""),
  ])
 }

-function "get_cache_to_rocm_rixl" {
+function "get_cache_to_rocm_nixl" {
  params = []
  result = compact([
-    RIXL_CACHE_KEY != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rixl-rocm-${RIXL_CACHE_KEY},mode=min" : (RIXL_BRANCH != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:rixl-rocm-${RIXL_BRANCH}-ucx-${UCX_BRANCH},mode=min" : ""),
+    NIXL_CACHE_KEY != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:nixl-rocm-${NIXL_CACHE_KEY},mode=min" : (NIXL_BRANCH != "" ? "type=registry,ref=${DOCKERHUB_CACHE_REPO}:nixl-rocm-${NIXL_BRANCH}-ucx-${UCX_BRANCH},mode=min" : ""),
  ])
 }

@@ -372,11 +372,11 @@ variable "CI_BASE_IMAGE_TAG_STABLE" {
 # in the registry cache keyed by its upstream commit hash. When ci_base rebuilds
 # (e.g., requirements change), these stages are cache hits if their upstream
 # pins haven't changed -- saving ~35min of compilation.
-target "rixl-rocm-ci" {
+target "nixl-rocm-ci" {
  inherits   = ["_common-rocm", "_ci-rocm"]
-  target     = "build_rixl"
+  target     = "build_nixl"
  cache-from = get_cache_from_rocm_deps()
-  cache-to   = get_cache_to_rocm_rixl()
+  cache-to   = get_cache_to_rocm_nixl()
  output     = ["type=cacheonly"]
 }

@@ -396,7 +396,7 @@ target "deepep-rocm-ci" {
  output     = ["type=cacheonly"]
 }

-# Builds only the ci_base stage (RIXL, DeepEP, torchcodec, etc.)
+# Builds only the ci_base stage (NIXL, DeepEP, torchcodec, etc.)
 # Invoked by the ensure-ci-base step when the content hash of ci_base-affecting
 # files drifts from the remote image label. Per-PR builds then pull the result
 # as CI_BASE_IMAGE instead of rebuilding those slow layers on every commit.
@@ -412,7 +412,7 @@ target "ci-base-rocm-ci" {
      CI_BASE_IMAGE_TAG_CONTENT_EXTRA != "" ? "type=registry,ref=${CI_BASE_IMAGE_TAG_CONTENT_EXTRA}" : "",
      CI_BASE_IMAGE_TAG_STABLE != "" ? "type=registry,ref=${CI_BASE_IMAGE_TAG_STABLE}" : "",
    ]),
-    # Import upstream dependency caches so RIXL/ROCShmem/DeepEP stages
+    # Import upstream dependency caches so NIXL/ROCShmem/DeepEP stages
    # are cache hits even when ci_base itself needs rebuilding.
    get_cache_from_rocm_deps(),
  )
@@ -424,5 +424,5 @@ target "ci-base-rocm-ci" {
 # Group for ci_base builds -- exports dependency stage caches alongside the
 # ci_base image so future rebuilds can reuse them independently.
 group "ci-base-rocm-ci-with-deps" {
-  targets = ["rixl-rocm-ci", "rocshmem-rocm-ci", "deepep-rocm-ci", "ci-base-rocm-ci"]
+  targets = ["nixl-rocm-ci", "rocshmem-rocm-ci", "deepep-rocm-ci", "ci-base-rocm-ci"]
 }
@@ -53,7 +53,7 @@ variable "CI_BASE_IMAGE" {
 # Upstream dependency commit pins. Plain local bake builds use the Dockerfile
 # ARG defaults. ci-bake-rocm.sh resolves those defaults (plus any env
 # overrides) and writes a small HCL override before invoking CI targets.
-variable "RIXL_BRANCH" {
+variable "NIXL_BRANCH" {
  default = ""
 }

@@ -106,7 +106,7 @@ target "test-rocm" {
  output   = ["type=docker"]
 }

-# CI base image target - builds only the ci_base stage (RIXL, DeepEP,
+# CI base image target - builds only the ci_base stage (NIXL, DeepEP,
 # torchcodec, requirements, etc.). Used by the weekly scheduled build and
 # the auto-rebuild trigger when requirements change in a PR.
 target "ci-base-rocm" {
@@ -2,7 +2,7 @@
  "_comment": "Auto-generated from Dockerfile ARGs. Do not edit manually. Run: python tools/generate_versions_json.py",
  "variable": {
    "CUDA_VERSION": {
-      "default": "13.0.2"
+      "default": "13.0.3"
    },
    "PYTHON_VERSION": {
      "default": "3.12"
@@ -10,11 +10,14 @@
    "UBUNTU_VERSION": {
      "default": "22.04"
    },
+    "NCCL_VERSION": {
+      "default": "2.30.7"
+    },
    "BUILD_BASE_IMAGE": {
-      "default": "nvidia/cuda:13.0.2-devel-ubuntu22.04"
+      "default": "nvidia/cuda:13.0.3-devel-ubuntu22.04"
    },
    "FINAL_BASE_IMAGE": {
-      "default": "nvidia/cuda:13.0.2-base-ubuntu22.04"
+      "default": "nvidia/cuda:13.0.3-base-ubuntu22.04"
    },
    "BUILD_OS": {
      "default": "ubuntu"
@@ -56,7 +59,7 @@
      "default": "cuda"
    },
    "DEEPEP_COMMIT_HASH": {
-      "default": "73b6ea4"
+      "default": "d4f41e4e93"
    },
    "GIT_REPO_CHECK": {
      "default": "0"
@@ -68,7 +71,7 @@
      "default": "true"
    },
    "FLASHINFER_VERSION": {
-      "default": "0.6.14"
+      "default": "0.6.15.post1"
    },
    "GDRCOPY_CUDA_VERSION": {
      "default": "12.8"
@@ -56,7 +56,9 @@ nav:
  - API Reference:
    - api/README.md
    - api/vllm
-  - CLI Reference: cli
+  - CLI Reference:
+    - cli/README.md
+    - vllm: cli
  - Community:
    - community/*
    - Governance: governance
@@ -1,10 +1,8 @@
 nav:
-  - README.md
-  - serve.md
-  - chat.md
-  - complete.md
-  - run-batch.md
-  - vllm bench:
-    - bench/**/*.md
-  - vllm launch:
-    - launch/**/*.md
+  - "*.md"
+  - bench:
+    - bench/*.md
+    - sweep:
+      - bench/sweep/*.md
+  - launch:
+    - launch/*.md
@@ -1,9 +0,0 @@
-# vllm bench latency
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_latency.inc.md"
@@ -1,55 +0,0 @@
-# vllm bench mm-processor
-
-## Overview
-
-`vllm bench mm-processor` profiles the multimodal input processor pipeline of
-vision-language models. It measures per-stage latency from the HuggingFace
-processor through to the encoder forward pass, helping you identify
-preprocessing bottlenecks and understand how different image resolutions or
-item counts affect end-to-end request time.
-
-The benchmark supports two data sources: synthetic random multimodal inputs
-(`random-mm`) and HuggingFace datasets (`hf`). Warmup requests are run before
-measurement to ensure stable results.
-
-## Quick Start
-
-```bash
-vllm bench mm-processor \
-  --model Qwen/Qwen2-VL-7B-Instruct \
-  --dataset-name random-mm \
-  --num-prompts 50 \
-  --random-input-len 300 \
-  --random-output-len 40 \
-  --random-mm-base-items-per-request 2 \
-  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
-  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}'
-```
-
-## Measured Stages
-
-| Stage | Description |
-| ----- | ----------- |
-| `get_mm_hashes_secs` | Time spent hashing multimodal inputs |
-| `get_cache_missing_items_secs` | Time spent looking up the processor cache |
-| `apply_hf_processor_secs` | Time spent in the HuggingFace processor |
-| `merge_mm_kwargs_secs` | Time spent merging multimodal kwargs |
-| `apply_prompt_updates_secs` | Time spent updating prompt tokens |
-| `preprocessor_total_secs` | Total preprocessing time |
-| `encoder_forward_secs` | Time spent in the encoder model forward pass |
-| `num_encoder_calls` | Number of encoder invocations per request |
-
-The benchmark also reports end-to-end latency (TTFT + decode time) per
-request. Use `--metric-percentiles` to select which percentiles to report
-(default: p99) and `--output-json` to save results.
-
-For more examples (HF datasets, warmup, JSON output), see
-[Benchmarking CLI — Multimodal Processor Benchmark](../../benchmarking/cli.md#multimodal-processor-benchmark).
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_mm_processor.inc.md"
@@ -1,9 +0,0 @@
-# vllm bench serve
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_serve.inc.md"
@@ -1,9 +0,0 @@
-# vllm bench sweep plot
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_sweep_plot.inc.md"
@@ -1,9 +0,0 @@
-# vllm bench sweep plot_pareto
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_sweep_plot_pareto.inc.md"
@@ -1,9 +0,0 @@
-# vllm bench sweep serve
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_sweep_serve.inc.md"
@@ -1,9 +0,0 @@
-# vllm bench sweep serve_workload
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_sweep_serve_workload.inc.md"
@@ -1,9 +0,0 @@
-# vllm bench throughput
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/bench_throughput.inc.md"
@@ -1,5 +0,0 @@
-# vllm chat
-
-## Arguments
-
--8<-- "docs/generated/argparse/chat.inc.md"
@@ -1,5 +0,0 @@
-# vllm complete
-
-## Arguments
-
--8<-- "docs/generated/argparse/complete.inc.md"
@@ -1,10 +0,0 @@
-<!-- markdownlint-disable MD041 -->
-When passing JSON CLI arguments, the following sets of arguments are equivalent:
-
- `--json-arg '{"key1": "value1", "key2": {"key3": "value2"}}'`
- `--json-arg.key1 value1 --json-arg.key2.key3 value2`
-
-Additionally, list elements can be passed individually using `+`:
-
- `--json-arg '{"key4": ["value3", "value4", "value5"]}'`
- `--json-arg.key4+ value3 --json-arg.key4+='value4,value5'`
@@ -1,22 +0,0 @@
-# vllm launch render
-
-## Overview
-
-`vllm launch render` starts a GPU-less rendering server for preprocessing and
-postprocessing only.
-
-```bash
-vllm launch render meta-llama/Llama-3.2-1B-Instruct --port 8100
-```
-
-This command reuses the standard serving parser, so model, frontend,
-networking, and related CLI options follow the same conventions as
-[`vllm serve`](../serve.md).
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/launch_render.inc.md"
@@ -1,9 +0,0 @@
-# vllm run-batch
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/run-batch.inc.md"
@@ -1,9 +0,0 @@
-# vllm serve
-
-## JSON CLI Arguments
-
--8<-- "docs/cli/json_tip.inc.md"
-
-## Arguments
-
--8<-- "docs/generated/argparse/serve.inc.md"
@@ -11,12 +11,4 @@ Engine arguments control the behavior of the vLLM engine.

 The engine argument classes, [EngineArgs][vllm.engine.arg_utils.EngineArgs] and [AsyncEngineArgs][vllm.engine.arg_utils.AsyncEngineArgs], are a combination of the configuration classes defined in [vllm.config][]. Therefore, if you are interested in developer documentation, we recommend looking at these configuration classes as they are the source of truth for types, defaults and docstrings.

--8<-- "docs/cli/json_tip.inc.md"
-
-## `EngineArgs`
-
--8<-- "docs/generated/argparse/engine_args.inc.md"
-
-## `AsyncEngineArgs`
-
--8<-- "docs/generated/argparse/async_engine_args.inc.md"
+--8<-- "gen:engine-args"
@@ -195,7 +195,7 @@ Provide a fast duration→token estimate to improve streaming usage statistics:
 The API server takes care of basic audio I/O and optional chunking before building prompts:

 - Resampling: Input audio is resampled to `SpeechToTextConfig.sample_rate` using `AudioResampler`.
- Chunking: If `SpeechToTextConfig.allow_audio_chunking` is True and the duration exceeds `max_audio_clip_s`, the server splits the audio into overlapping chunks and generates a prompt per chunk. Overlap is controlled by `overlap_chunk_second`.
+- Chunking: If `SpeechToTextConfig.allow_audio_chunking` is True and the duration exceeds `max_audio_clip_s`, the server splits the audio into chunks and generates a prompt per chunk. There is no overlap between chunks, overlap_chunk_second controls the size of the search window used to find the split point.
 - Energy-aware splitting: When `min_energy_split_window_size` is set, the server finds low-energy regions to minimize cutting within words.

 Relevant server logic:
@@ -8,6 +8,26 @@ toc_depth: 2

 --8<-- "docs/getting_started/installation/gpu.md:pre-built-images"

+## Persist the compile cache across containers
+
+Mounting the Hugging Face cache keeps model weights across containers, but each
+new container still starts with an empty `VLLM_CACHE_ROOT` (default
+`~/.cache/vllm`) and recompiles the model's `torch.compile` artifacts. Mount a
+named volume at that path to reuse the inductor, Triton, and AOT artifacts from
+the second container onward:
+
+```bash
+docker run --rm --gpus all \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    -v vllm-cache:/root/.cache/vllm \
+    -p 8000:8000 \
+    vllm/vllm-openai:latest \
+    meta-llama/Llama-3.1-8B-Instruct
+```
+
+See [Faster Startup](../configuration/optimization.md#faster-startup) for the
+mechanism and for what invalidates the cache.
+
 ## Run as a non-root user

 The CUDA `vllm/vllm-openai` image runs as root by default for backward
@@ -1,15 +1,9 @@
 # Attention Backend Feature Support

-This document is auto-generated by `tools/pre_commit/generate_attention_backend_docs.py`.
-It shows the feature support for each registered attention backend
-based on the checks in `AttentionBackend.validate_configuration()`.
-
-**Do not edit this file manually.** Run the following command to
-regenerate it:
-
-```bash
-python tools/pre_commit/generate_attention_backend_docs.py
-```
+The priority and feature tables on this page are auto-generated from the
+attention backend registry by
+`docs/mkdocs/gen_files/generate_attention_backends.py`, based on the checks in
+`AttentionBackend.validate_configuration()`.

 ## Setting the Attention Backend

@@ -98,40 +92,11 @@ Priority is **1 = highest** (tried first).

 ### Standard Attention (MHA, MQA, GQA)

-**Blackwell (SM 10.x):**
-
-| Priority | Backend |
-| -------- | ------- |
-| 1 | `FLASHINFER` |
-| 2 | `FLASH_ATTN` |
-| 3 | `TRITON_ATTN` |
-| 4 | `FLEX_ATTENTION` |
-| 5 | `TURBOQUANT` |
-
-**Ampere/Hopper (SM 8.x-9.x):**
-
-| Priority | Backend |
-| -------- | ------- |
-| 1 | `FLASH_ATTN` |
-| 2 | `FLASHINFER` |
-| 3 | `TRITON_ATTN` |
-| 4 | `FLEX_ATTENTION` |
-| 5 | `TURBOQUANT` |
+--8<-- "gen:priority-standard"

 ### MLA Attention (DeepSeek-style)

-**Blackwell (SM 10.x):**
-
-| Priority | Backend |
-| -------- | ------- |
-| 1 | `FLASHINFER_MLA` |
-| 2 | `TOKENSPEED_MLA` |
-| 3 | `CUTLASS_MLA` |
-| 4 | `FLASH_ATTN_MLA` |
-| 5 | `FLASHMLA` |
-| 6 | `TRITON_MLA` |
-| 7 | `FLASHINFER_MLA_SPARSE`**\*** |
-| 8 | `FLASHMLA_SPARSE` |
+--8<-- "gen:priority-mla"

 > **\*** For sparse MLA, FP8 KV cache always prefers `FLASHINFER_MLA_SPARSE`. With BF16 KV cache, `FLASHINFER_MLA_SPARSE` is preferred for low query-head counts (<= 16), while `FLASHMLA_SPARSE` is preferred otherwise.
 >
@@ -157,24 +122,7 @@ Priority is **1 = highest** (tried first).

 ## Standard Attention (MHA, MQA, GQA) Backends

-| Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | --------- | --- | --------------- | ------------ |
-| `CPU_ATTN` | | fp16, bf16, fp32 | `auto`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256, 512 | ❌ | ✅ | ❌ | ❌ | All | N/A |
-| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ❌ | ✅ | ❌ | ✅ | Decoder | 8.x-9.x |
-| `FLASHINFER` | XQA† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ❌ | ❌ | ❌ | ✅ | Decoder | 9.0 |
-| `FLASHINFER` | trtllm-gen† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `nvfp4` | 16, 32, 64, 128, 256, 512, 1024 | 64, 128, 256, 512 | ✅ | ✅ | ❌ | ✅ | Decoder | 10.x |
-| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ❌ | ✅ | All | ≥8.0 |
-| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | 9.x |
-| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ✅ | ✅ | ❌ | ✅ | All | ≥10.0 |
-| `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
-| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder Only | Any |
-| `HPC_ATTN` | | fp16, bf16 | `auto`, `bfloat16`, `fp8_e4m3` | 64 | 128 | ❌ | ❌ | ❌ | ❌ | Decoder | ≥9.0 |
-| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ✅ | ✅ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | ❌ | All | N/A |
-| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ❌ | ✅ | ✅ | ❌ | Decoder, Encoder, Encoder Only | N/A |
-| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2`, `int4_per_token_head`, `int8_per_token_head`, `fp8_per_token_head` | %16 | Any | ✅ | ✅ | ✅ | ❌ | All | Any |
-| `TRITON_ATTN_DIFFKV` | | fp16, bf16 | `auto`, `bfloat16` | Any | Any | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
-| `TURBOQUANT` | | fp16, bf16 | `turboquant_k8v4`, `turboquant_4bit_nc`, `turboquant_k3v4_nc`, `turboquant_3bit_nc` | 16, 32, 64, 128 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
+--8<-- "gen:table-standard"

 > **†** FlashInfer Native is the regular FlashInfer path. XQA is the SM90 decode path exposed through FlashInfer's TRTLLM decode API. trtllm-gen is used on SM100 and supports sinks. Disable XQA/trtllm-gen via `--attention-config.use_trtllm_attention=0`.
 >
@@ -188,9 +136,7 @@ automatic priority lists above. A lightning indexer scores KV blocks, the
 top-k blocks (plus fixed init/local blocks) are selected, and attention
 attends only to those blocks; index keys live in a separate side cache.

-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | --------- | --- | --------------- | ------------ |
-| `MINIMAX_M3_SPARSE` | bf16, fp16 | `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 128 | 128 | ❌ | ❌ | ❌ | ❌ | Decoder | Any |
+--8<-- "gen:table-minimax"

 ## MLA (Multi-head Latent Attention) Backends

@@ -203,38 +149,20 @@ To explicitly select a prefill backend, use
 Otherwise, the prefill backend is selected automatically at runtime based on
 hardware and configuration.

-| Backend | Description | Dtypes | Compute Cap. | Notes |
-| ------- | ----------- | ------ | ------------ | ----- |
-| `FLASH_ATTN`‡ | FlashAttention varlen (FA2/FA3/FA4) | fp16, bf16 | Any | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) (FA2/FA3/FA4) or (qk_nope_head_dim=64, qk_rope_head_dim=64, v_head_dim=128) (FA2/FA3/FA4) or (qk_nope_head_dim=192, qk_rope_head_dim=64, v_head_dim=256) (FA2/FA3 only) |
-| `TRTLLM_RAGGED` | TensorRT-LLM ragged attention | fp16, bf16 | 10.x | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) or (qk_nope_head_dim=192, qk_rope_head_dim=64, v_head_dim=256) only |
-| `FLASHINFER` | FlashInfer CUTLASS backend | fp16, bf16 | 10.x | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) only |
-| `TOKENSPEED_MLA` | | fp16, bf16 | 10.x | (qk_nope_head_dim=128, qk_rope_head_dim=64, v_head_dim=128) only |
+--8<-- "gen:table-mla-prefill"

 > **‡** Automatic selection tries FlashAttention first. On Blackwell
 > (SM100), the fallback order is TRT-LLM Ragged, FlashInfer, then
-> TokenSpeed MLA. On other GPUs, only FlashAttention is considered.
+> TokenSpeed MLA; for (qk_nope_head_dim=192, qk_rope_head_dim=64,
+> v_head_dim=256) TRT-LLM Ragged is tried before FlashAttention. On other
+> GPUs, only FlashAttention is considered.

 ### Decode Backends

 MLA decode backends are selected using the standard
 `-ac.backend=<BACKEND>` argument (e.g., `FLASHMLA`, `TRITON_MLA`).

-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | ------ | --------- | --- | --------------- | ------------ |
-| `CUTLASS_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA_SPARSE_SM120` | bf16 | `auto`, `fp8`, `fp8_e4m3`, `fp8_ds_mla` | 64, 256 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | 12.x |
-| `FLASHMLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
-| `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
-| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
-| `FLASH_ATTN_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x |
-| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %1 | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 1, 64 | Any | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `TOKENSPEED_MLA` | fp16, bf16 | `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `TRITON_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ❌ | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
-| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
+--8<-- "gen:table-mla-decode"

 ### DeepSeek V4 Decode Backends

@@ -245,8 +173,4 @@ pipeline (compressor + SWA + indexer, 256-token blocks, head 512);
 default on NVIDIA is `FLASHINFER_MLA_SPARSE_DSV4` on SM12x and
 `FLASHMLA_SPARSE_DSV4` on other supported CUDA architectures.

-| Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Non-Causal | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
-| ------- | ------ | --------- | ----------- | ---------- | ---- | ---------- | ------ | --------- | --- | --------------- | ------------ |
-| `FLASHINFER_MLA_SPARSE_DSV4` | bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_ds_mla` | 256 | 512 | ✅ | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x, 12.x |
-| `FLASHMLA_SPARSE_DSV4` | bf16 | `auto`, `fp8_ds_mla`, `fp8` | 256 | 512 | ✅ | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
-| `ROCM_FLASHMLA_SPARSE_DSV4` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+--8<-- "gen:table-mla-v4-decode"
@@ -129,6 +129,7 @@ Models opt-in to encoder CUDA Graphs by implementing the [SupportsEncoderCudaGra
 | `DeepseekOCRForCausalLM` | `DeepSeek-OCR` | ✅︎ | ❌︎ | ✅︎ |
 | `Gemma3ForConditionalGeneration` | `Gemma3` | ✅︎ | ❌︎ | ❌︎ |
 | `Glm4vForConditionalGeneration` | `GLM-4.1V, GLM-4.6V-Flash` | ✅︎ | ✅︎ | ❌︎ |
+| `Gemma4ForConditionalGeneration` | `Gemma-4` | ✅︎ | ✅︎ | ❌︎ |
 | `InternVLChatModel` | `InternVL3.5`, `InternVL3`, `InternVL2.5`, `InternVL2` | ✅︎ | ✅︎ | ❌︎ |
 | `KimiVLForConditionalGeneration` | `Kimi-VL` | ✅︎ | ❌︎ | ❌︎ |
 | `Llama4ForConditionalGeneration` | `Llama 4` | ✅︎ | ❌︎ | ❌︎ |
@@ -122,8 +122,6 @@ For example:

 --8<-- "vllm/model_executor/layers/mamba/mamba_mixer2.py:mixer2_gated_rms_norm"

--8<-- "vllm/model_executor/models/plamo2.py:plamo2_mamba_mixer"
-
 --8<-- "vllm/model_executor/layers/mamba/short_conv.py:short_conv"
 ```

@@ -242,4 +242,4 @@ See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-modular-all2a

 ## FusedMoEExpertsModular

-See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-experts-kernels) for a list of all the available modular experts.
+See [Fused MoE Kernel features](./moe_kernel_features.md#fused-experts-kernels) for a list of all the available modular experts.
@@ -47,7 +47,8 @@ sequenceDiagram
    else only one side present
        PWriter->>PWriter: stash and wait, self-poll only when blocks unmatched
    end
-    PWriter->>PWriter: ensure D handshake (one-time)
+    PWriter->>PWriter: _ensure_handshake to D (async; defer WRITE)
+    PWriter->>PWriter: handshake callback re-queues on _deferred_push_inbox, wake
    PWriter->>DWriter: NIXL WRITE direct to D GPU + completion notif

    note over DWorker,DWriter: D side - completion accounting
@@ -100,15 +101,20 @@ event:
   D, completion notifs after a WRITE, late-arriving ``PUSH_REG``)
   even when there is no new metadata to act on.
 3. **Handshake-completion callback** (background handshake executor
-   thread) — when a deferred D→P handshake finishes successfully, the
-   future's done-callback re-enqueues the registration onto
-   ``_reg_send_inbox`` and sets the wake so the corresponding
-   ``send_notif`` runs on the writer (we never call ``send_notif`` from
-   the executor thread). On this second pass ``_ensure_handshake``
-   returns ``None`` (the agent is now connected), so the writer sends
-   the ``PUSH_REG`` directly. If the handshake *failed*, the callback
-   fails the request instead of re-enqueuing, so there is no retry
-   loop.
+   thread) — both handshakes run on the executor and never block the
+   writer; their done-callbacks re-enqueue the deferred op and set the
+   wake, since neither ``send_notif`` nor the NIXL WRITE may run off the
+   writer thread:
+    * the **D→P** handshake (before sending ``PUSH_REG``) re-enqueues the
+      registration onto ``_reg_send_inbox``;
+    * the **P→D** handshake (before a WRITE) re-enqueues the matched
+      ``(req_id, blocks, reg_data)`` onto ``_deferred_push_inbox``.
+
+   On this second pass ``_ensure_handshake`` returns ``None`` (the agent
+   is now connected), so the writer sends the ``PUSH_REG`` / issues the
+   WRITE directly. If a handshake *failed*, the callback fails or drops
+   the request instead of re-enqueuing, so there is no retry loop (see
+   Failure handling).

 In addition to event-driven wakes, the writer self-polls at
 ``_PUSH_WRITER_POLL_INTERVAL_MS = 1.0`` ms while there are P-side
@@ -227,6 +233,13 @@ Two per-request timers are armed on the scheduler:
 * **D-side ``send_notif`` failure when shipping the PUSH_REG to P** —
  identical handling: ``_handle_failed_transfer`` marks the recv as
  failed.
+* **P-side handshake failure (P→D handshake before a WRITE)** — the
+  future's done-callback logs ``push_handshake_failed`` and drops the
+  request without re-queuing. It deliberately does *not* call
+  ``_handle_failed_transfer`` (there is no ``_recving_metadata`` entry to
+  invalidate on the producer side, same reasoning as the WRITE-submission
+  failure below). P's blocks are reclaimed by the ``_kv_lease_duration``
+  lease and D's stale registration by its watchdog.
 * **P-side WRITE submission failure** — the WRITE handle (if any) is
  released and ``xfer_stats.record_failed_transfer()`` bumps the
  failure counter. We deliberately do not call
@@ -246,10 +259,11 @@ existing NIXL connector:
  class — all subclasses of the existing base classes;
 * one dedicated background thread per worker;
 * a few cross-thread queues, each with a single consumer (the writer);
-  most have one producer, except ``_reg_send_inbox``, which is fed both
-  by the engine main thread (new registrations) and by the
-  handshake-completion callback (registrations replayed after their
-  D→P handshake finishes);
+  most have one producer, except the two replay queues fed by both the
+  engine main thread and a handshake-completion callback:
+  ``_reg_send_inbox`` (registrations replayed after their D→P handshake)
+  and ``_deferred_push_inbox`` (matched pushes replayed after their P→D
+  handshake);
 * one new notification type (`PUSH_REG:<msgpack>`).

 Behavior on the engine main thread is otherwise unchanged. The writer
@@ -157,7 +157,7 @@ Object keys follow the same run-configuration digest scheme as the filesystem ti

 The P2P tier (`type: "p2p"`) shares completed KV blocks between vLLM instances over RDMA via NIXL. Each instance binds a control socket on `host:port` and exchanges blocks directly with peers — no shared filesystem required.

-PYTHONHASHSEED environment variable must be set to the same fixed value on all nodes.
+The `PYTHONHASHSEED` environment variable must be set to the same fixed value (e.g. `"0"`) on all nodes so that block content hashes match across instances (see [Cross-Process Sharing](#cross-process-sharing)). This is enforced: a P2P instance started without `PYTHONHASHSEED` set fails at startup, and each peer's value is verified during the connect handshake — a peer advertising a different `PYTHONHASHSEED` is rejected.

 | Key | Required | Default | Notes |
 | --- | --- | --- | --- |
@@ -176,6 +176,71 @@ Rather than embedding `host`/`port` in each `secondary_tiers` entry, set them on
 - `VLLM_P2P_SIDE_CHANNEL_HOST` (default `localhost`): address the P2P control socket binds to. It is used **verbatim** as both the bind address and the identity peers dial back — there is no auto-detection (this mirrors `VLLM_NIXL_SIDE_CHANNEL_HOST`). The default binds the loopback interface only, so peers on another host cannot reach it. **For any cross-host P2P deployment you must set this explicitly to the node's routable IP** (e.g. the pod IP) before launching `vllm serve` — otherwise remote peers will fail to connect. The NIXL agent name is a separate per-process identifier, so peers sharing a `host:port` never collide.
 - `VLLM_P2P_SIDE_CHANNEL_PORT` (default `5710`): base port for the P2P control socket. The port actually bound is `VLLM_P2P_SIDE_CHANNEL_PORT + data_parallel_index` — one socket per DP replica, matching NIXL (for DP=1 the offset is 0). The peer's port is passed as `remote_port` in `kv_transfer_params`; the router/EPP that selects the DP rank (e.g. via the `X-data-parallel-rank` header) computes `remote_port = base + rank`. The DP-index offset separates replicas *within* one deployment; two co-located *deployments* (a prefiller and a decoder on the same host) still need distinct base ports (e.g. decoder base `5711`) to avoid a bind collision.

+#### Orchestration-Layer Protocol
+
+The P2P tier does not decide *which* peer to pull from — that is the orchestration layer's job (the router/EPP and its scheduler). The orchestrator drives every transfer through a request's `kv_transfer_params` dict: it picks the request's role, allocates a unique transaction ID, and supplies the remote peer's address. All block lookup, hash matching, and NIXL transfer happen at the tier level below; the orchestrator only sets the correct role keys and enforces the allowed combinations.
+
+Every vLLM instance is a symmetric **peer**. Per request it acts as a **consumer** (pulls KV blocks from a remote peer's CPU cache instead of computing locally) or a **producer** (serves blocks from its own CPU cache to remote consumers) — or both, on the same session, for different requests. Roles are chosen per request by the keys below; there are no fixed prefiller/decoder processes.
+
+Three role keys are defined, each mapping to a sub-dict. All are optional; a request with none of them uses the tier only as a local CPU cache.
+
+Each key names the **remote counterpart** this peer transfers with (not this
+peer's own role), so the name reads as "the remote ___ I transfer with".
+
+| Key | Set on | Value fields | Meaning |
+| --- | --- | --- | --- |
+| `remote_decoder` | prefill producer request | `kv_request_id` | Peer computes KV and keeps it available in CPU cache for the remote decoder to pull. |
+| `remote_prefiller` | decode consumer request | `kv_request_id`, `remote_host`, `remote_port` | Peer pulls KV from the remote prefiller at the given address (classic P/D disaggregation). |
+| `remote_kv_source` | P2P consumer request | `kv_request_id`, `remote_host`, `remote_port` | Peer looks up and pulls whatever blocks the remote source currently holds in CPU cache. |
+
+Field semantics:
+
+- `kv_request_id` (str): unique transaction ID allocated by the orchestrator and pushed to every peer involved in the transfer; used to correlate the lookup, fetch, and transfer-done messages. The producer is implicit — it serves whatever block hashes it currently holds in its CPU cache for that ID.
+- `remote_host` (str): IP/hostname of the remote peer's control socket to query. Must be the peer's routable node IP (see [Environment Variables](#environment-variables)).
+- `remote_port` (int): the peer's bound control-socket port, i.e. `base + data_parallel_index` for the selected DP rank.
+
+Allowed and forbidden combinations:
+
+- **`remote_decoder` + `remote_kv_source`** is the only legal multi-key combination: a prefill producer may *also* act as a P2P consumer for the same request — skipping prefix prefill by pulling cached blocks from a source while still keeping its own computed blocks available for a downstream decoder.
+- Forbidden: `remote_prefiller` + `remote_decoder` (contradictory roles), `remote_prefiller` + `remote_kv_source` (two competing fetch sources), and all three together.
+
+Minimal examples (values that would appear in the request's `kv_transfer_params`):
+
+```python
+# Prefill producer — compute and keep KV for a remote decoder to pull
+kv_transfer_params = {"remote_decoder": {"kv_request_id": "<unique-transfer-id>"}}
+
+# Decode consumer — pull KV from a specific prefiller (classic P/D)
+kv_transfer_params = {
+    "remote_prefiller": {
+        "kv_request_id": "<unique-transfer-id>",
+        "remote_host": "<prefiller-node-ip>",
+        "remote_port": 5710,
+    }
+}
+
+# P2P consumer — pull whatever the source already has cached
+kv_transfer_params = {
+    "remote_kv_source": {
+        "kv_request_id": "<unique-transfer-id>",
+        "remote_host": "<source-node-ip>",
+        "remote_port": 5710,
+    }
+}
+```
+
+Runtime handshake for a P2P (or P/D) pull, once the orchestrator has set the keys above:
+
+1. Both peers already have listener threads on their control sockets (see [Environment Variables](#environment-variables)).
+2. **Lookup.** The consumer's tiering manager does per-block lookups; in P2P mode the tier returns `None` and registers the key. At `on_schedule_end` the consumer sends one **`LookupMsg`** (`kv_request_id` + block hashes) to the peer, per request, per step.
+3. The producer matches those hashes against its local CPU cache and replies with a **`LookupRespMsg`** carrying the hit block hashes.
+4. **Resolve.** Retried lookups now return hit / miss / in-flight. The consumer calls `submit_load` for hits only, allocating CPU slots only for hits.
+5. The consumer sends a **`FetchMsg`** (`kv_request_id`, block hashes, destination block indexes).
+6. The producer performs the **NIXL WRITE** transfer and sends **`TransferDone`** with a success status.
+7. On `get_finished`, hits are loaded into GPU as ordinary cache hits; misses are recomputed by the engine.
+
+In classic **P/D mode** (`remote_prefiller` set, no `remote_kv_source`), the lookup phase (steps 2–4) is skipped: the decode consumer assumes the prefiller holds all of the request's blocks, so every block `lookup()` returns an immediate hit and the consumer jumps straight to the **`FetchMsg`** in step 5. The `LookupMsg`/`LookupRespMsg` round-trip only happens in P2P mode, where the consumer does not know in advance which blocks the peer has cached.
+
 ## Tuning Tips

 - `cpu_bytes_to_use`: a bigger CPU tier means fewer trips to slower secondary tiers and a higher hit rate. The value is total across all workers, not per-worker. Leave headroom for the rest of the host workload.
--- a/Show More
+++ b/Show More