Relax protobuf library version constraints (#33202)

Signed-off-by: Jeffrey Wang <jeffreywang@anyscale.com>
(cherry picked from commit a97b5e206d)

.buildkite/ci_config.yaml

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

.buildkite/hardware_tests/amd.yaml

@@ -1,28 +0,0 @@
-group: Hardware
-steps:
-  - label: "AMD: :docker: build image"
-    device: amd_cpu
-    no_plugin: true
-    commands:
-    - >
-      docker build
-      --build-arg max_jobs=16
-      --build-arg REMOTE_VLLM=1
-      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942'
-      --build-arg VLLM_BRANCH=$BUILDKITE_COMMIT
-      --tag "rocm/vllm-ci:${BUILDKITE_COMMIT}"
-      -f docker/Dockerfile.rocm
-      --target test
-      --no-cache
-      --progress plain .
-    - docker push "rocm/vllm-ci:${BUILDKITE_COMMIT}"
-    env:
-      DOCKER_BUILDKIT: "1"
-    retry:
-      automatic:
-        - exit_status: -1  # Agent was lost
-          limit: 1
-        - exit_status: -10  # Agent was lost
-          limit: 1
-        - exit_status: 1  # Machine occasionally fail
-          limit: 1

.buildkite/hardware_tests/arm.yaml

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

.buildkite/hardware_tests/ascend_npu.yaml

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

.buildkite/hardware_tests/gh200.yaml

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

.buildkite/hardware_tests/intel.yaml

@@ -1,23 +0,0 @@
-group: Hardware
-depends_on: ~
-steps:
-  - label: "Intel CPU Test"
-    soft_fail: true
-    device: intel_cpu
-    no_plugin: true
-    commands: 
-    - bash .buildkite/scripts/hardware_ci/run-cpu-test.sh
-  
-  - label: "Intel HPU Test"
-    soft_fail: true
-    device: intel_hpu
-    no_plugin: true
-    commands: 
-    - bash .buildkite/scripts/hardware_ci/run-hpu-test.sh
-
-  - label: "Intel GPU Test"
-    soft_fail: true
-    device: intel_gpu
-    no_plugin: true
-    commands: 
-    - bash .buildkite/scripts/hardware_ci/run-xpu-test.sh

.buildkite/image_build/image_build.sh

@@ -1,254 +1,56 @@
 #!/bin/bash
-set -euo pipefail
+set -e
 
-# replace invalid characters in Docker image tags and truncate to 128 chars
-clean_docker_tag() {
-    local input="$1"
-    echo "$input" | sed 's/[^a-zA-Z0-9._-]/_/g' | cut -c1-128
-}
-
-print_usage_and_exit() {
-    echo "Usage: $0 <registry> <repo> <commit> <branch> <vllm_use_precompiled> <vllm_merge_base_commit> <cache_from> <cache_to>"
-    exit 1
-}
-
-print_instance_info() {
-    echo ""
-    echo "=== Debug: Instance Information ==="
-    # Get IMDSv2 token
-    if TOKEN=$(curl -s -X PUT "http://169.254.169.254/latest/api/token" \
-            -H "X-aws-ec2-metadata-token-ttl-seconds: 21600" 2>/dev/null); then
-        AMI_ID=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
-            http://169.254.169.254/latest/meta-data/ami-id 2>/dev/null || echo "unknown")
-        INSTANCE_TYPE=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
-            http://169.254.169.254/latest/meta-data/instance-type 2>/dev/null || echo "unknown")
-        INSTANCE_ID=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
-            http://169.254.169.254/latest/meta-data/instance-id 2>/dev/null || echo "unknown")
-        AZ=$(curl -s -H "X-aws-ec2-metadata-token: $TOKEN" \
-            http://169.254.169.254/latest/meta-data/placement/availability-zone 2>/dev/null || echo "unknown")
-        echo "AMI ID:        ${AMI_ID}"
-        echo "Instance Type: ${INSTANCE_TYPE}"
-        echo "Instance ID:   ${INSTANCE_ID}"
-        echo "AZ:            ${AZ}"
-    else
-        echo "Not running on EC2 or IMDS not available"
-    fi
-    # Check for warm cache AMI (marker file baked into custom AMI)
-    if [[ -f /etc/vllm-ami-info ]]; then
-        echo "Cache:         warm (custom vLLM AMI)"
-        cat /etc/vllm-ami-info
-    else
-        echo "Cache:         cold (standard AMI)"
-    fi
-    echo "==================================="
-    echo ""
-}
-
-setup_buildx_builder() {
-    echo "--- :buildkite: Setting up buildx builder"
-    if [[ -S "${BUILDKIT_SOCKET}" ]]; then
-        # Custom AMI with standalone buildkitd - use remote driver for warm cache
-        echo "✅ Found local buildkitd socket at ${BUILDKIT_SOCKET}"
-        echo "Using remote driver to connect to buildkitd (warm cache available)"
-        if docker buildx inspect baked-vllm-builder >/dev/null 2>&1; then
-            echo "Using existing baked-vllm-builder"
-            docker buildx use baked-vllm-builder
-        else
-            echo "Creating baked-vllm-builder with remote driver"
-            docker buildx create \
-                --name baked-vllm-builder \
-                --driver remote \
-                --use \
-                "unix://${BUILDKIT_SOCKET}"
-        fi
-        docker buildx inspect --bootstrap
-    elif docker buildx inspect "${BUILDER_NAME}" >/dev/null 2>&1; then
-        # Existing builder available
-        echo "Using existing builder: ${BUILDER_NAME}"
-        docker buildx use "${BUILDER_NAME}"
-        docker buildx inspect --bootstrap
-    else
-        # No local buildkitd, no existing builder - create new docker-container builder
-        echo "No local buildkitd found, using docker-container driver"
-        docker buildx create --name "${BUILDER_NAME}" --driver docker-container --use
-        docker buildx inspect --bootstrap
-    fi
-
-    # builder info
-    echo "Active builder:"
-    docker buildx ls | grep -E '^\*|^NAME' || docker buildx ls
-}
-
-check_and_skip_if_image_exists() {
-    if [[ -n "${IMAGE_TAG:-}" ]]; then
-        echo "--- :mag: Checking if image exists"
-        if docker manifest inspect "${IMAGE_TAG}" >/dev/null 2>&1; then
-            echo "Image already exists: ${IMAGE_TAG}"
-            echo "Skipping build"
-            exit 0
-        fi
-        echo "Image not found, proceeding with build"
-    fi
-}
-
-ecr_login() {
-    aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
-    aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin 936637512419.dkr.ecr.us-east-1.amazonaws.com
-}
-
-prepare_cache_tags() {
-    # resolve and set: CACHE_TO, CACHE_FROM, CACHE_FROM_BASE_BRANCH, CACHE_FROM_MAIN
-    TEST_CACHE_ECR="936637512419.dkr.ecr.us-east-1.amazonaws.com/vllm-ci-test-cache"
-    MAIN_CACHE_ECR="936637512419.dkr.ecr.us-east-1.amazonaws.com/vllm-ci-postmerge-cache"
-
-    if [[ "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
-        if [[ "$BUILDKITE_BRANCH" == "main" ]]; then
-            cache="${MAIN_CACHE_ECR}:latest"
-        else
-            clean_branch=$(clean_docker_tag "$BUILDKITE_BRANCH")
-            cache="${TEST_CACHE_ECR}:${clean_branch}"
-        fi
-        CACHE_TO="$cache"
-        CACHE_FROM="$cache"
-        CACHE_FROM_BASE_BRANCH="$cache"
-    else
-        CACHE_TO="${TEST_CACHE_ECR}:pr-${BUILDKITE_PULL_REQUEST}"
-        CACHE_FROM="${TEST_CACHE_ECR}:pr-${BUILDKITE_PULL_REQUEST}"
-        if [[ "$BUILDKITE_PULL_REQUEST_BASE_BRANCH" == "main" ]]; then
-            CACHE_FROM_BASE_BRANCH="${MAIN_CACHE_ECR}:latest"
-        else
-            clean_base=$(clean_docker_tag "$BUILDKITE_PULL_REQUEST_BASE_BRANCH")
-            CACHE_FROM_BASE_BRANCH="${TEST_CACHE_ECR}:${clean_base}"
-        fi
-    fi
-
-    CACHE_FROM_MAIN="${MAIN_CACHE_ECR}:latest"
-    export CACHE_TO CACHE_FROM CACHE_FROM_BASE_BRANCH CACHE_FROM_MAIN
-}
-
-resolve_parent_commit() {
-    if [[ -z "${PARENT_COMMIT:-}" ]]; then
-        PARENT_COMMIT=$(git rev-parse HEAD~1 2>/dev/null || echo "")
-        if [[ -n "${PARENT_COMMIT}" ]]; then
-            echo "Computed parent commit for cache fallback: ${PARENT_COMMIT}"
-            export PARENT_COMMIT
-        else
-            echo "Could not determine parent commit (may be first commit in repo)"
-        fi
-    else
-        echo "Using provided PARENT_COMMIT: ${PARENT_COMMIT}"
-    fi
-}
-
-print_bake_config() {
-    echo "--- :page_facing_up: Resolved bake configuration"
-    BAKE_CONFIG_FILE="bake-config-build-${BUILDKITE_BUILD_NUMBER:-local}.json"
-    docker buildx bake -f "${VLLM_BAKE_FILE}" -f "${CI_HCL_PATH}" --print "${TARGET}" | tee "${BAKE_CONFIG_FILE}" || true
-    echo "Saved bake config to ${BAKE_CONFIG_FILE}"
-    echo "--- :arrow_down: Uploading bake config to Buildkite"
-    buildkite-agent artifact upload "${BAKE_CONFIG_FILE}"
-}
-
-#################################
-#         Main Script           #
-#################################
-print_instance_info
-
-if [[ $# -lt 7 ]]; then
-    print_usage_and_exit
+if [[ $# -lt 8 ]]; then
+  echo "Usage: $0 <registry> <repo> <commit> <branch> <vllm_use_precompiled> <vllm_merge_base_commit> <cache_from> <cache_to>"
+  exit 1
 fi
 
-# input args
 REGISTRY=$1
 REPO=$2
 BUILDKITE_COMMIT=$3
 BRANCH=$4
 VLLM_USE_PRECOMPILED=$5
 VLLM_MERGE_BASE_COMMIT=$6
-IMAGE_TAG=$7
-IMAGE_TAG_LATEST=${8:-} # only used for main branch, optional
+CACHE_FROM=$7
+CACHE_TO=$8
 
-# build config
-TARGET="test-ci"
-CI_HCL_URL="${CI_HCL_URL:-https://raw.githubusercontent.com/vllm-project/ci-infra/main/docker/ci.hcl}"
-VLLM_BAKE_FILE="${VLLM_BAKE_FILE:-docker/docker-bake.hcl}"
-BUILDER_NAME="${BUILDER_NAME:-vllm-builder}"
-CI_HCL_PATH="/tmp/ci.hcl"
-BUILDKIT_SOCKET="/run/buildkit/buildkitd.sock"
+# authenticate with AWS ECR
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
+aws ecr get-login-password --region us-east-1 | docker login --username AWS --password-stdin 936637512419.dkr.ecr.us-east-1.amazonaws.com
 
-prepare_cache_tags
-ecr_login
+# docker buildx 
+docker buildx create --name vllm-builder --driver docker-container --use
+docker buildx inspect --bootstrap
+docker buildx ls
 
-# Environment info (for docs and human readers)
-#   CI_HCL_URL          - URL to ci.hcl (default: from ci-infra main branch)
-#   VLLM_CI_BRANCH      - ci-infra branch to use (default: main)
-#   VLLM_BAKE_FILE      - Path to vLLM's bake file (default: docker/docker-bake.hcl)
-#   BUILDER_NAME        - Name for buildx builder (default: vllm-builder)
-#
-# Build configuration (exported as environment variables for bake):
-export BUILDKITE_COMMIT
-export PARENT_COMMIT
-export IMAGE_TAG
-export IMAGE_TAG_LATEST
-export CACHE_FROM
-export CACHE_FROM_BASE_BRANCH
-export CACHE_FROM_MAIN
-export CACHE_TO
-export VLLM_USE_PRECOMPILED
-export VLLM_MERGE_BASE_COMMIT
-
-# print args
-echo "--- :mag: Arguments"
-echo "REGISTRY: ${REGISTRY}"
-echo "REPO: ${REPO}"
-echo "BUILDKITE_COMMIT: ${BUILDKITE_COMMIT}"
-echo "BRANCH: ${BRANCH}"
-echo "VLLM_USE_PRECOMPILED: ${VLLM_USE_PRECOMPILED}"
-echo "VLLM_MERGE_BASE_COMMIT: ${VLLM_MERGE_BASE_COMMIT}"
-echo "IMAGE_TAG: ${IMAGE_TAG}"
-echo "IMAGE_TAG_LATEST: ${IMAGE_TAG_LATEST}"
-
-# print build configuration
-echo "--- :mag: Build configuration"
-echo "TARGET: ${TARGET}"
-echo "CI HCL URL: ${CI_HCL_URL}"
-echo "vLLM bake file: ${VLLM_BAKE_FILE}"
-echo "BUILDER_NAME: ${BUILDER_NAME}"
-echo "CI_HCL_PATH: ${CI_HCL_PATH}"
-echo "BUILDKIT_SOCKET: ${BUILDKIT_SOCKET}"
-
-echo "--- :mag: Cache tags"
-echo "CACHE_TO: ${CACHE_TO}"
-echo "CACHE_FROM: ${CACHE_FROM}"
-echo "CACHE_FROM_BASE_BRANCH: ${CACHE_FROM_BASE_BRANCH}"
-echo "CACHE_FROM_MAIN: ${CACHE_FROM_MAIN}"
-
-check_and_skip_if_image_exists
-
-echo "--- :docker: Setting up Docker buildx bake"
-echo "Target: ${TARGET}"
-echo "CI HCL URL: ${CI_HCL_URL}"
-echo "vLLM bake file: ${VLLM_BAKE_FILE}"
-
-if [[ ! -f "${VLLM_BAKE_FILE}" ]]; then
-    echo "Error: vLLM bake file not found at ${VLLM_BAKE_FILE}"
-    echo "Make sure you're running from the vLLM repository root"
-    exit 1
+# skip build if image already exists
+if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT) ]]; then
+  echo "Image not found, proceeding with build..."
+else
+  echo "Image found"
+  exit 0
 fi
 
-echo "--- :arrow_down: Downloading ci.hcl"
-curl -sSfL -o "${CI_HCL_PATH}" "${CI_HCL_URL}"
-echo "Downloaded to ${CI_HCL_PATH}"
+if [[ "${VLLM_USE_PRECOMPILED:-0}" == "1" ]]; then
+  merge_base_commit_build_args="--build-arg VLLM_MERGE_BASE_COMMIT=${VLLM_MERGE_BASE_COMMIT}"
+else
+  merge_base_commit_build_args=""
+fi
 
-setup_buildx_builder
-
-# Compute parent commit for cache fallback (if not already set)
-resolve_parent_commit
-export PARENT_COMMIT
-
-print_bake_config
-
-echo "--- :docker: Building ${TARGET}"
-docker --debug buildx bake -f "${VLLM_BAKE_FILE}" -f "${CI_HCL_PATH}" --progress plain "${TARGET}"
-
-echo "--- :white_check_mark: Build complete"
+# build
+docker buildx build --file docker/Dockerfile \
+  --build-arg max_jobs=16 \
+  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
+  --build-arg USE_SCCACHE=1 \
+  --build-arg TORCH_CUDA_ARCH_LIST="8.0 8.9 9.0 10.0" \
+  --build-arg FI_TORCH_CUDA_ARCH_LIST="8.0 8.9 9.0a 10.0a" \
+  --build-arg VLLM_USE_PRECOMPILED="${VLLM_USE_PRECOMPILED:-0}" \
+  ${merge_base_commit_build_args} \
+  --cache-from type=registry,ref=${CACHE_FROM},mode=max \
+  --cache-to type=registry,ref=${CACHE_TO},mode=max \
+  --tag ${REGISTRY}/${REPO}:${BUILDKITE_COMMIT} \
+  $( [[ "${BRANCH}" == "main" ]] && echo "--tag ${REGISTRY}/${REPO}:latest" ) \
+  --push \
+  --target test \
+  --progress plain .

.buildkite/image_build/image_build.yaml

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

.buildkite/test-amd.yaml

@@ -1131,7 +1131,7 @@ steps:
   - csrc/quantization/cutlass_w8a8/moe/
   - vllm/model_executor/layers/fused_moe/cutlass_moe.py
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_prepare_finalize.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/v1/attention/backends/mla/cutlass_mla.py

.buildkite/test-pipeline.yaml

@@ -1017,7 +1017,7 @@ steps:
   - csrc/quantization/cutlass_w8a8/moe/
   - vllm/model_executor/layers/fused_moe/cutlass_moe.py
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_prepare_finalize.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/v1/attention/backends/mla/cutlass_mla.py
@@ -1316,7 +1316,7 @@ steps:
   - pytest -v -s distributed/test_distributed_oot.py
   - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
-  - pytest -v -s plugins/lora_resolvers # unit tests for lora resolver plugins
+  - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 - label: Pipeline + Context Parallelism Test # 45min
   timeout_in_minutes: 60

.buildkite/test_areas/attention.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: V1 attention (H100)
   timeout_in_minutes: 30
-  device: h100
+  gpu: h100
   source_file_dependencies:
     - vllm/config/attention.py
     - vllm/model_executor/layers/attention
@@ -15,7 +15,7 @@ steps:
 
 - label: V1 attention (B200)
   timeout_in_minutes: 30
-  device: b200
+  gpu: b200
   source_file_dependencies:
     - vllm/config/attention.py
     - vllm/model_executor/layers/attention

.buildkite/test_areas/compile.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Fusion and Compile Tests (B200)
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
-  device: b200
+  gpu: b200
   source_file_dependencies:
   - csrc/quantization/fp4/
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
@@ -26,7 +26,7 @@ steps:
     - nvidia-smi
     - pytest -v -s tests/compile/test_fusion_attn.py
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
-    # this runner has 2 GPUs available even though num_devices=2 is not set
+    # this runner has 2 GPUs available even though num_gpus=2 is not set
     - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
     # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
     # Wrap with quotes to escape yaml
@@ -37,9 +37,9 @@ steps:
 - label: Fusion E2E (2 GPUs)(B200)
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
-  device: b200
+  gpu: b200
   optional: true
-  num_devices: 2
+  num_gpus: 2
   source_file_dependencies:
   - csrc/quantization/fp4/
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py

.buildkite/test_areas/distributed.yaml

@@ -5,7 +5,7 @@ steps:
 - label: Distributed Comm Ops
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
-  num_devices: 2
+  num_gpus: 2
   source_file_dependencies:
   - vllm/distributed
   - tests/distributed
@@ -18,7 +18,7 @@ steps:
 - label: Distributed (2 GPUs)
   timeout_in_minutes: 90
   working_dir: "/vllm-workspace/tests"
-  num_devices: 2
+  num_gpus: 2
   source_file_dependencies:
   - vllm/compilation/
   - vllm/distributed/
@@ -54,7 +54,7 @@ steps:
 - label: Distributed Tests (4 GPUs)
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
-  num_devices: 4
+  num_gpus: 4
   source_file_dependencies:
   - vllm/distributed/
   - tests/distributed/test_utils
@@ -103,8 +103,8 @@ steps:
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
-  device: h100
-  num_devices: 8
+  gpu: h100
+  num_gpus: 8
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - examples/offline_inference/torchrun_dp_example.py
@@ -120,9 +120,9 @@ steps:
   - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
 
 - label: Distributed Tests (4 GPUs)(A100)
-  device: a100
+  gpu: a100
   optional: true
-  num_devices: 4
+  num_gpus: 4
   source_file_dependencies:
   - vllm/
   commands:
@@ -133,34 +133,26 @@ steps:
   - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   - pytest -v -s -x lora/test_mixtral.py
 
-- label: Sequence Parallel Tests (H100)
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/"
-  device: h100
-  optional: true
-  num_devices: 2
-  commands:
-    - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-    # Run sequence parallel tests
-    - pytest -v -s tests/distributed/test_sequence_parallel.py
-    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
-
-- label: Distributed Tests (2 GPUs)(H100)
-  device: h100
+- label: Distributed Tests (2 GPUs)(H200)
+  gpu: h200
   optional: true
   working_dir: "/vllm-workspace/"
-  num_devices: 2
+  num_gpus: 2
   commands:
     - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_async_tp.py
+    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
+    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
+    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'
+    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/distributed/test_sequence_parallel.py
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
+    - CUDA_VISIBLE_DEVICES=1,2 VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 
 - label: Distributed Tests (2 GPUs)(B200)
-  device: b200
+  gpu: b200
   optional: true
   working_dir: "/vllm-workspace/"
-  num_devices: 2
+  num_gpus: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
     - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
@@ -169,9 +161,8 @@ steps:
 - label: 2 Node Test (4 GPUs)
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
-  num_devices: 2
+  num_gpus: 2
   num_nodes: 2
-  no_plugin: true
   source_file_dependencies:
   - vllm/distributed/
   - vllm/engine/
@@ -185,7 +176,7 @@ steps:
 - label: Distributed NixlConnector PD accuracy (4 GPUs)
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
-  num_devices: 4
+  num_gpus: 4
   source_file_dependencies:
     - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
     - tests/v1/kv_connector/nixl_integration/
@@ -193,21 +184,10 @@ steps:
     - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
-- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs)
-  timeout_in_minutes: 30
-  working_dir: "/vllm-workspace/tests"
-  num_devices: 4
-  source_file_dependencies:
-    - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
-    - tests/v1/kv_connector/nixl_integration/
-  commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
-    - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
-
 - label: Pipeline + Context Parallelism (4 GPUs))
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
-  num_devices: 4
+  num_gpus: 4
   source_file_dependencies:
   - vllm/distributed/
   - vllm/engine/
@@ -216,46 +196,4 @@ steps:
   - tests/distributed/
   commands:
   - pytest -v -s distributed/test_pp_cudagraph.py
-  - pytest -v -s distributed/test_pipeline_parallel.py
-
-- label: Hopper Fusion E2E Tests (H100)
-  timeout_in_minutes: 70
-  working_dir: "/vllm-workspace/"
-  device: h100
-  optional: true
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/test_fusion_attn.py
-  commands:
-    - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-    # skip Llama-4 since it does not fit on this device
-    - pytest -v -s tests/compile/test_fusion_attn.py -k 'not Llama-4'
-
-- label: Hopper Fusion Distributed E2E Tests (2xH100)
-  timeout_in_minutes: 70
-  working_dir: "/vllm-workspace/"
-  device: h100
-  optional: true
-  num_devices: 2
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/distributed/test_fusions_e2e.py
-  commands:
-    - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-    # Run all e2e fusion tests
-    - pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
+  - pytest -v -s distributed/test_pipeline_parallel.py

.buildkite/test_areas/e2e_integration.yaml

@@ -4,27 +4,27 @@ depends_on:
 steps:
 - label: DeepSeek V2-Lite Accuracy
   timeout_in_minutes: 60
-  device: h100
+  gpu: h100
   optional: true
-  num_devices: 4
+  num_gpus: 4
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
 
 - label: Qwen3-30B-A3B-FP8-block Accuracy
   timeout_in_minutes: 60
-  device: h100
+  gpu: h100
   optional: true
-  num_devices: 4
+  num_gpus: 4
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
 
 - label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
   timeout_in_minutes: 60
-  device: b200
+  gpu: b200
   optional: true
-  num_devices: 2
+  num_gpus: 2
   working_dir: "/vllm-workspace"
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
@@ -33,11 +33,10 @@ steps:
   timeout_in_minutes: 30
   optional: true
   soft_fail: true
-  num_devices: 2
+  num_gpus: 2
   working_dir: "/vllm-workspace"
   source_file_dependencies:
   - vllm/
   - .buildkite/scripts/run-prime-rl-test.sh
   commands:
-    - nvidia-smi
     - bash .buildkite/scripts/run-prime-rl-test.sh

.buildkite/test_areas/engine.yaml

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

.buildkite/test_areas/expert_parallelism.yaml

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

.buildkite/test_areas/kernels.yaml

@@ -57,8 +57,8 @@ steps:
 
 - label: Kernels DeepGEMM Test (H100)
   timeout_in_minutes: 45
-  device: h100
-  num_devices: 1
+  gpu: h100
+  num_gpus: 1
   source_file_dependencies:
   - tools/install_deepgemm.sh
   - vllm/utils/deep_gemm.py
@@ -77,7 +77,7 @@ steps:
 - label: Kernels (B200)
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/"
-  device: b200
+  gpu: b200
   # optional: true
   source_file_dependencies:
   - csrc/quantization/fp4/
@@ -85,7 +85,7 @@ steps:
   - csrc/quantization/cutlass_w8a8/moe/
   - vllm/model_executor/layers/fused_moe/cutlass_moe.py
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py
-  - vllm/model_executor/layers/fused_moe/flashinfer_a2a_prepare_finalize.py
+  - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
   - vllm/v1/attention/backends/mla/cutlass_mla.py
@@ -114,55 +114,4 @@ steps:
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
-    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
-    # e2e
-    - pytest -v -s tests/models/quantization/test_nvfp4.py
-
-- label: Kernels Helion Test
-  timeout_in_minutes: 30
-  device: h100
-  source_file_dependencies:
-  - vllm/utils/import_utils.py
-  - tests/kernels/helion/
-  commands:
-    - pip install helion
-    - pytest -v -s kernels/helion/
-
- 
-- label: Kernels FP8 MoE Test (1 H100)
-  timeout_in_minutes: 90
-  device: h100
-  num_devices: 1
-  optional: true
-  commands:
-    - pytest -v -s kernels/moe/test_cutlass_moe.py
-    - pytest -v -s kernels/moe/test_flashinfer.py
-    - pytest -v -s kernels/moe/test_gpt_oss_triton_kernels.py
-    - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py
-    - pytest -v -s kernels/moe/test_moe.py
-    # - pytest -v -s kernels/moe/test_block_fp8.py - failing on main
-    - pytest -v -s kernels/moe/test_block_int8.py
-    - pytest -v -s kernels/moe/test_triton_moe_no_act_mul.py
-    - pytest -v -s kernels/moe/test_triton_moe_ptpc_fp8.py
-
-- label: Kernels FP8 MoE Test (2 H100s)
-  timeout_in_minutes: 90
-  device: h100
-  num_devices: 2
-  optional: true
-  commands:
-    - pytest -v -s kernels/moe/test_deepep_deepgemm_moe.py
-    - pytest -v -s kernels/moe/test_deepep_moe.py
-    - pytest -v -s kernels/moe/test_pplx_cutlass_moe.py
-    # - pytest -v -s kernels/moe/test_pplx_moe.py - failing on main
-  
-- label: Kernels Fp4 MoE Test (B200)
-  timeout_in_minutes: 60
-  device: b200
-  num_devices: 1
-  optional: true
-  commands:
-    - pytest -v -s kernels/moe/test_cutedsl_moe.py
-    - pytest -v -s kernels/moe/test_flashinfer_moe.py
-    - pytest -v -s kernels/moe/test_nvfp4_moe.py
-    - pytest -v -s kernels/moe/test_ocp_mx_moe.py
+    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py

.buildkite/test_areas/lm_eval.yaml

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

.buildkite/test_areas/lora.yaml

@@ -14,7 +14,7 @@ steps:
 
 - label: LoRA TP (Distributed)
   timeout_in_minutes: 30
-  num_devices: 4
+  num_gpus: 4
   source_file_dependencies:
   - vllm/lora
   - tests/lora

.buildkite/test_areas/misc.yaml

@@ -31,7 +31,7 @@ steps:
   source_file_dependencies:
     - vllm/
     - tests/v1
-  device: cpu
+  no_gpu: true
   commands:
     # split the test to avoid interference
     - pytest -v -s -m 'cpu_test' v1/core
@@ -82,7 +82,7 @@ steps:
 
 - label: Metrics, Tracing (2 GPUs)
   timeout_in_minutes: 20
-  num_devices: 2
+  num_gpus: 2
   source_file_dependencies:
   - vllm/
   - tests/v1/tracing
@@ -127,7 +127,7 @@ steps:
   - tests/tool_parsers
   - tests/transformers_utils
   - tests/config
-  device: cpu
+  no_gpu: true
   commands:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
@@ -142,7 +142,7 @@ steps:
 - label: GPT-OSS Eval (B200)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
-  device: b200
+  gpu: b200
   optional: true
   source_file_dependencies:
   - tests/evals/gpt_oss
@@ -155,7 +155,7 @@ steps:
 
 - label: Batch Invariance (H100)
   timeout_in_minutes: 25
-  device: h100
+  gpu: h100
   source_file_dependencies:
     - vllm/v1/attention
     - vllm/model_executor/layers

.buildkite/test_areas/models_basic.yaml

@@ -44,7 +44,7 @@ steps:
   - vllm/
   - tests/models/test_utils.py
   - tests/models/test_vision.py
-  device: cpu
+  no_gpu: true
   commands:
     - pytest -v -s models/test_utils.py models/test_vision.py
 

.buildkite/test_areas/models_distributed.yaml

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

.buildkite/test_areas/models_multimodal.yaml

@@ -18,7 +18,7 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
-  device: cpu
+  no_gpu: true
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py

.buildkite/test_areas/plugins.yaml

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

.buildkite/test_areas/quantization.yaml

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

.buildkite/test_areas/weight_loading.yaml

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

benchmarks/kernels/benchmark_cutlass_moe_nvfp4.py

@@ -197,7 +197,7 @@ def bench_run(
         )
 
         kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            MoEPrepareAndFinalizeNoEP(defer_input_quant=True),
             CutlassExpertsFp4(
                 make_dummy_moe_config(),
                 quant_config=quant_config,
@@ -242,7 +242,7 @@ def bench_run(
         )
 
         kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            MoEPrepareAndFinalizeNoEP(defer_input_quant=True),
             CutlassExpertsFp4(
                 make_dummy_moe_config(),
                 quant_config=quant_config,

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -10,6 +10,8 @@ from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe import fused_topk
 from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import (
+    _moe_permute,
+    _moe_unpermute_and_reduce,
     moe_permute,
     moe_unpermute,
 )
@@ -39,6 +41,7 @@ def benchmark_permute(
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
     num_iters: int = 100,
+    use_customized_permute: bool = False,
 ) -> float:
     # init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
@@ -61,14 +64,29 @@ def benchmark_permute(
         input_gating.copy_(gating_output[i])
 
     def run():
-        moe_permute(
-            qhidden_states,
-            a1q_scale=None,
-            topk_ids=topk_ids,
-            n_expert=num_experts,
-            expert_map=None,
-            align_block_size=align_block_size,
-        )
+        if use_customized_permute:
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            ) = moe_permute(
+                qhidden_states,
+                a1q_scale=None,
+                topk_ids=topk_ids,
+                n_expert=num_experts,
+                expert_map=None,
+                align_block_size=align_block_size,
+            )
+        else:
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            ) = _moe_permute(qhidden_states, None, topk_ids, num_experts, None, 16)
 
     # JIT compilation & warmup
     run()
@@ -113,9 +131,11 @@ def benchmark_unpermute(
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
     num_iters: int = 100,
+    use_customized_permute: bool = False,
 ) -> float:
     # init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    output_hidden_states = torch.empty_like(hidden_states)
     if use_fp8_w8a8:
         align_block_size = 128  # deepgemm needs 128 m aligned block
         qhidden_states, scale = _fp8_quantize(hidden_states, None, None)
@@ -130,37 +150,78 @@ def benchmark_unpermute(
     )
 
     def prepare():
-        (
-            permuted_hidden_states,
-            _,
-            first_token_off,
-            inv_perm_idx,
-            _,
-        ) = moe_permute(
-            qhidden_states,
-            a1q_scale=None,
-            topk_ids=topk_ids,
-            n_expert=num_experts,
-            expert_map=None,
-            align_block_size=align_block_size,
-        )
-        # convert to fp16/bf16 as gemm output
-        return (
-            permuted_hidden_states.to(dtype),
-            first_token_off,
-            inv_perm_idx,
-        )
+        if use_customized_permute:
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            ) = moe_permute(
+                qhidden_states,
+                a1q_scale=None,
+                topk_ids=topk_ids,
+                n_expert=num_experts,
+                expert_map=None,
+                align_block_size=align_block_size,
+            )
+            # convert to fp16/bf16 as gemm output
+            return (
+                permuted_hidden_states.to(dtype),
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            )
+        else:
+            (
+                permuted_qhidden_states,
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            ) = _moe_permute(
+                qhidden_states, None, topk_ids, num_experts, None, block_m=16
+            )
+            # convert to fp16/bf16 as gemm output
+            return (
+                permuted_qhidden_states.to(dtype),
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            )
 
     def run(input: tuple):
-        (permuted_hidden_states, first_token_off, inv_perm_idx) = input
-        output = torch.empty_like(hidden_states)
-        moe_unpermute(
-            output,
-            permuted_hidden_states,
-            topk_weights,
-            inv_perm_idx,
-            first_token_off,
-        )
+        if use_customized_permute:
+            (
+                permuted_hidden_states,
+                first_token_off,
+                inv_perm_idx,
+                m_indices,
+            ) = input
+            output = torch.empty_like(hidden_states)
+            moe_unpermute(
+                output,
+                permuted_hidden_states,
+                topk_weights,
+                inv_perm_idx,
+                first_token_off,
+            )
+        else:
+            (
+                permuted_hidden_states,
+                a1q_scale,
+                sorted_token_ids,
+                expert_ids,
+                inv_perm,
+            ) = input
+            _moe_unpermute_and_reduce(
+                output_hidden_states,
+                permuted_hidden_states,
+                inv_perm,
+                topk_weights,
+                True,
+            )
 
     # JIT compilation & warmup
     input = prepare()
@@ -215,7 +276,8 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
-    ) -> tuple[float, float]:
+        use_customized_permute: bool = False,
+    ) -> tuple[dict[str, int], float]:
         set_random_seed(self.seed)
 
         permute_time = benchmark_permute(
@@ -227,6 +289,7 @@ class BenchmarkWorker:
             use_fp8_w8a8,
             use_int8_w8a16,
             num_iters=100,
+            use_customized_permute=use_customized_permute,
         )
         unpermute_time = benchmark_unpermute(
             num_tokens,
@@ -237,6 +300,7 @@ class BenchmarkWorker:
             use_fp8_w8a8,
             use_int8_w8a16,
             num_iters=100,
+            use_customized_permute=use_customized_permute,
         )
         return permute_time, unpermute_time
 
@@ -283,6 +347,7 @@ def main(args: argparse.Namespace):
     dtype = torch.float16 if current_platform.is_rocm() else config.dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
+    use_customized_permute = args.use_customized_permute
 
     if args.batch_size is None:
         batch_sizes = [
@@ -334,6 +399,7 @@ def main(args: argparse.Namespace):
                 dtype,
                 use_fp8_w8a8,
                 use_int8_w8a16,
+                use_customized_permute,
             )
             for batch_size in batch_sizes
         ],
@@ -353,6 +419,7 @@ if __name__ == "__main__":
     parser.add_argument(
         "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
     )
+    parser.add_argument("--use-customized-permute", action="store_true")
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, required=False)
     parser.add_argument("--trust-remote-code", action="store_true")

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -22,8 +22,8 @@ from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 mp.set_start_method("spawn", force=True)
 
-assert current_platform.is_cuda() or current_platform.is_rocm(), (
-    "Only support tune w8a8 block fp8 kernel on CUDA/ROCm device."
+assert current_platform.is_cuda(), (
+    "Only support tune w8a8 block fp8 kernel on CUDA device."
 )
 
 DTYPE_MAP = {

csrc/cache_kernels.cu

@@ -24,12 +24,6 @@
 typedef __hip_bfloat16 __nv_bfloat16;
 #endif
 
-#if defined(__gfx942__)
-constexpr float kFp8ScaleDivisor = 224.f;
-#else
-constexpr float kFp8ScaleDivisor = 448.f;
-#endif
-
 void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
                  int64_t block_size_in_bytes,
                  const torch::Tensor& block_mapping) {
@@ -407,7 +401,8 @@ __global__ void concat_and_cache_ds_mla_kernel(
   }
 
   // Compute the scale for the tile
-  float tile_scale = fmaxf(max_abs / kFp8ScaleDivisor, FLT_MIN);
+  float tile_scale = max_abs / 448.f;
+  tile_scale = fmaxf(tile_scale, FLT_MIN);
 
   // The first lane of each half-warp writes the scale to kv_cache
   if ((lane_idx == 0) || (lane_idx == 16)) {
@@ -476,8 +471,11 @@ __global__ void indexer_k_quant_and_cache_kernel(
 #endif
   }
 
-  float scale = fmaxf(amax, 1e-4) / kFp8ScaleDivisor;
-
+#if defined(__gfx942__)
+  float scale = fmaxf(amax, 1e-4) / 224.0f;
+#else
+  float scale = fmaxf(amax, 1e-4) / 448.0f;
+#endif
   if (use_ue8m0) {
     scale = exp2f(ceilf(log2f(scale)));
   }

csrc/cpu/dnnl_kernels.cpp

@@ -360,14 +360,13 @@ void onednn_scaled_mm(
     const std::optional<torch::Tensor>& azp,      // [M] or [1]
     const std::optional<torch::Tensor>& azp_adj,  // [M] or [1]
     const std::optional<torch::Tensor>& bias,     // [N]
-    const torch::Tensor& handler_tensor) {
+    int64_t handler) {
   CPU_KERNEL_GUARD_IN(onednn_scaled_mm)
   TORCH_CHECK(a.dim() == 2);
   TORCH_CHECK(a.is_contiguous());
   TORCH_CHECK(c.is_contiguous());
   W8A8MatMulPrimitiveHandler* ptr =
-      reinterpret_cast<W8A8MatMulPrimitiveHandler*>(
-          handler_tensor.item<int64_t>());
+      reinterpret_cast<W8A8MatMulPrimitiveHandler*>(handler);
   const int32_t* azp_ptr = nullptr;
   if (azp.has_value()) {
     azp_ptr = azp->data_ptr<int32_t>();
@@ -520,14 +519,13 @@ int64_t create_onednn_mm_handler(const torch::Tensor& b,
 
 void onednn_mm(torch::Tensor& c,        // [M, OC], row-major
                const torch::Tensor& a,  // [M, IC], row-major
-               const std::optional<torch::Tensor>& bias,
-               const torch::Tensor& handler_tensor) {
+               const std::optional<torch::Tensor>& bias, int64_t handler) {
   CPU_KERNEL_GUARD_IN(onednn_mm)
   TORCH_CHECK(a.dim() == 2);
   TORCH_CHECK(a.stride(-1) == 1);
   TORCH_CHECK(c.stride(-1) == 1);
   MatMulPrimitiveHandler* ptr =
-      reinterpret_cast<MatMulPrimitiveHandler*>(handler_tensor.item<int64_t>());
+      reinterpret_cast<MatMulPrimitiveHandler*>(handler);
 
 // ACL matmuls expect contiguous source tensors
 #ifdef VLLM_USE_ACL

csrc/cpu/torch_bindings.cpp

@@ -19,14 +19,13 @@ void onednn_scaled_mm(torch::Tensor& c, const torch::Tensor& a,
                       const std::optional<torch::Tensor>& azp,
                       const std::optional<torch::Tensor>& azp_adj,
                       const std::optional<torch::Tensor>& bias,
-                      const torch::Tensor& handler_tensor);
+                      int64_t handler);
 
 int64_t create_onednn_mm_handler(const torch::Tensor& b,
                                  int64_t primitive_cache_size);
 
 void onednn_mm(torch::Tensor& c, const torch::Tensor& a,
-               const std::optional<torch::Tensor>& bias,
-               const torch::Tensor& handler_tensor);
+               const std::optional<torch::Tensor>& bias, int64_t handler);
 
 bool is_onednn_acl_supported();
 
@@ -197,7 +196,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // oneDNN GEMM
   ops.def(
       "onednn_mm(Tensor! c, Tensor a, Tensor? bias, "
-      "Tensor handler_tensor) -> ()");
+      "int handler) -> ()");
   ops.impl("onednn_mm", torch::kCPU, &onednn_mm);
 
   // Check if oneDNN was built with ACL backend
@@ -213,7 +212,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // oneDNN scaled_mm for W8A8 with static per-tensor activation quantization
   ops.def(
       "onednn_scaled_mm(Tensor! c, Tensor a, Tensor a_scales, Tensor? azp, "
-      "Tensor? azp_adj, Tensor? bias, Tensor handler_tensor) -> ()");
+      "Tensor? azp_adj, Tensor? bias, int handler) -> ()");
   ops.impl("onednn_scaled_mm", torch::kCPU, &onednn_scaled_mm);
 
   // Compute int8 quantized tensor for given scaling factor.

docs/configuration/optimization.md

@@ -47,10 +47,6 @@ You can tune the performance by adjusting `max_num_batched_tokens`:
 - For optimal throughput, we recommend setting `max_num_batched_tokens > 8192` especially for smaller models on large GPUs.
 - If `max_num_batched_tokens` is the same as `max_model_len`, that's almost the equivalent to the V0 default scheduling policy (except that it still prioritizes decodes).
 
-!!! warning
-    When chunked prefill is disabled, `max_num_batched_tokens` must be greater than `max_model_len`.  
-    In that case, if `max_num_batched_tokens < max_model_len`, vLLM may crash at server start‑up.
-
 ```python
 from vllm import LLM
 

docs/contributing/model/basic.md

@@ -71,7 +71,7 @@ class MyModel(nn.Module):
 ```python
 def forward(
     self,
-    input_ids: torch.Tensor | None,
+    input_ids: torch.Tensor,
     positions: torch.Tensor,
     intermediate_tensors: IntermediateTensors | None = None,
     inputs_embeds: torch.Tensor | None = None,

docs/contributing/model/multimodal.md

@@ -43,73 +43,28 @@ Further update the model as follows:
                 )
     ```
 
-- Remove the embedding part from the [forward][torch.nn.Module.forward] method:
-    - Move the multi-modal embedding to [embed_multimodal][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_multimodal].
-    - The text embedding and embedding merge are handled automatically by a default implementation of [embed_input_ids][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_input_ids]. It does not need to be overridden in most cases.
+- Implement [embed_multimodal][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_multimodal] that returns the embeddings from running the multimodal inputs through the multimodal tokenizer of the model. Below we provide a boilerplate of a typical implementation pattern, but feel free to adjust it to your own needs.
 
-    ```diff
-      def forward(
-          self,
-          input_ids: torch.Tensor | None,
-    -     pixel_values: torch.Tensor,
-          positions: torch.Tensor,
-          intermediate_tensors: IntermediateTensors | None = None,
-          inputs_embeds: torch.Tensor | None = None,
-      ) -> torch.Tensor:
-    -     if inputs_embeds is None:
-    -         inputs_embeds = self.get_input_embeddings()(input_ids)
-    -
-    -     if pixel_values is not None:
-    -         image_features = self.get_image_features(
-    -             pixel_values=pixel_values,
-    -         )
-    -         special_image_mask = self.get_placeholder_mask(
-    -             input_ids,
-    -             inputs_embeds=inputs_embeds,
-    -             image_features=image_features,
-    -         )
-    -         inputs_embeds = inputs_embeds.masked_scatter(
-    -             special_image_mask,
-    -             image_features,
-    -         )
+    ??? code
 
-           hidden_states = self.language_model(
-               input_ids,
-               positions,
-               intermediate_tensors,
-               inputs_embeds=inputs_embeds,
-           )
-         ...
-  
-    +  def embed_multimodal(
-    +      self,
-    +      pixel_values: torch.Tensor,
-    +  ) -> MultiModalEmbeddings | None:
-    +      return self.get_image_features(
-    +          pixel_values=pixel_values,
-    +      )
-    ```
+        ```python
+            def _process_image_input(self, image_input: YourModelImageInputs) -> torch.Tensor:
+                image_features = self.vision_encoder(image_input)
+                return self.multi_modal_projector(image_features)
 
-    Below we provide a boilerplate of a typical implementation pattern of [embed_multimodal][vllm.model_executor.models.interfaces.SupportsMultiModal.embed_multimodal], but feel free to adjust it to your own needs.
+            def embed_multimodal(
+                self,
+                **kwargs: object,
+            ) -> MultiModalEmbeddings | None:
+                # Validate the multimodal input keyword arguments
+                image_input = self._parse_and_validate_image_input(**kwargs)
+                if image_input is None:
+                    return None
 
-    ```python
-    def _process_image_input(self, image_input: YourModelImageInputs) -> torch.Tensor:
-        image_features = self.vision_encoder(image_input)
-        return self.multi_modal_projector(image_features)
-
-    def embed_multimodal(
-        self,
-        **kwargs: object,
-    ) -> MultiModalEmbeddings | None:
-        # Validate the multimodal input keyword arguments
-        image_input = self._parse_and_validate_image_input(**kwargs)
-        if image_input is None:
-            return None
-
-        # Run multimodal inputs through encoder and projector
-        vision_embeddings = self._process_image_input(image_input)
-        return vision_embeddings
-    ```
+                # Run multimodal inputs through encoder and projector
+                vision_embeddings = self._process_image_input(image_input)
+                return vision_embeddings
+        ```
 
 !!! important
     The returned `multimodal_embeddings` must be either a **3D [torch.Tensor][]** of shape `(num_items, feature_size, hidden_size)`, or a **list / tuple of 2D [torch.Tensor][]'s** of shape `(feature_size, hidden_size)`, so that `multimodal_embeddings[i]` retrieves the embeddings generated from the `i`-th multimodal data item (e.g, image) of the request.

docs/design/lora_resolver_plugins.md

@@ -10,7 +10,7 @@ receives a request for a LoRA adapter that hasn't been loaded yet, the resolver
 to locate and load the adapter from their configured storage locations. This enables:
 
 - **Dynamic LoRA Loading**: Load adapters on-demand without server restarts
-- **Multiple Storage Backends**: Support for filesystem, S3, and custom backends. The built-in `lora_filesystem_resolver` requires a local storage path, while the built-in `hf_hub_resolver` will pull LoRA adapters from Huggingface Hub and proceed in an identical manner. In general, custom resolvers can be implemented to fetch from any source.
+- **Multiple Storage Backends**: Support for filesystem, S3, and custom backends. The built-in `lora_filesystem_resolver` requires a local storage path, but custom resolvers can be implemented to fetch from any source.
 - **Automatic Discovery**: Seamless integration with existing LoRA workflows
 - **Scalable Deployment**: Centralized adapter management across multiple vLLM instances
 

docs/design/moe_kernel_features.md

@@ -36,7 +36,8 @@ th {
 | pplx | batched | fp8,int8 | G,A,T | Y | Y | [`PplxPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.pplx_prepare_finalize.PplxPrepareAndFinalize] |
 | deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
 | deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
-| flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferA2APrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize.FlashInferA2APrepareAndFinalize] |
+| flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferAllToAllMoEPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize.FlashInferAllToAllMoEPrepareAndFinalize] |
+| flashinfer<sup>4</sup> | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferCutlassMoEPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize.FlashInferCutlassMoEPrepareAndFinalize] |
 | MoEPrepareAndFinalizeNoEP<sup>5</sup> | standard | fp8,int8 | G,A,T | N | Y | [`MoEPrepareAndFinalizeNoEP`][vllm.model_executor.layers.fused_moe.prepare_finalize.MoEPrepareAndFinalizeNoEP] |
 | BatchedPrepareAndFinalize<sup>5</sup> | batched | fp8,int8 | G,A,T | N | Y | [`BatchedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.fused_batched_moe.BatchedPrepareAndFinalize] |
 

docs/features/lora.md

@@ -159,12 +159,10 @@ Alternatively, you can use the LoRAResolver plugin to dynamically load LoRA adap
 
 You can set up multiple LoRAResolver plugins if you want to load LoRA adapters from different sources. For example, you might have one resolver for local files and another for S3 storage. vLLM will load the first LoRA adapter that it finds.
 
-You can either install existing plugins or implement your own. By default, vLLM comes with a [resolver plugin to load LoRA adapters from a local directory, as well as a resolver plugin to load LoRA adapters from repositories on Hugging Face Hub](https://github.com/vllm-project/vllm/tree/main/vllm/plugins/lora_resolvers)
-To enable either of these resolvers, you must `set VLLM_ALLOW_RUNTIME_LORA_UPDATING` to True.
-
-- To leverage a local directory, set `VLLM_PLUGINS` to include `lora_filesystem_resolver` and set `VLLM_LORA_RESOLVER_CACHE_DIR` to a local directory. When vLLM receives a request using a LoRA adapter `foobar`,
-it will first look in the local directory for a directory `foobar`, and attempt to load the contents of that directory as a LoRA adapter. If successful, the request will complete as normal and that adapter will then be available for normal use on the server.
-- To leverage repositories on Hugging Face Hub, set `VLLM_PLUGINS` to include `lora_hf_hub_resolver` and set `VLLM_LORA_RESOLVER_HF_REPO_LIST` to a comma separated list of repository IDs on Hugging Face Hub. When vLLM receives a request for the LoRA adapter `my/repo/subpath`, it will download the adapter at the `subpath` of `my/repo` if it exists and contains an `adapter_config.json`, then build a request to the cached dir for the adapter, similar to the `lora_filesystem_resolver`. Please note that enabling remote downloads is insecure and not intended for use in production environments.
+You can either install existing plugins or implement your own. By default, vLLM comes with a [resolver plugin to load LoRA adapters from a local directory.](https://github.com/vllm-project/vllm/tree/main/vllm/plugins/lora_resolvers)
+To enable this resolver, set `VLLM_ALLOW_RUNTIME_LORA_UPDATING` to True, set `VLLM_PLUGINS` to include `lora_filesystem_resolver`, and then set `VLLM_LORA_RESOLVER_CACHE_DIR` to a local directory. When vLLM receives a request using a LoRA adapter `foobar`,
+it will first look in the local directory for a directory `foobar`, and attempt to load the contents of that directory as a LoRA adapter. If successful, the request will complete as normal and
+that adapter will then be available for normal use on the server.
 
 Alternatively, follow these example steps to implement your own plugin:
 

docs/models/supported_models.md

@@ -674,7 +674,6 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `GLM4VForCausalLM`<sup>^</sup> | GLM-4V | T + I | `zai-org/glm-4v-9b`, `zai-org/cogagent-9b-20241220`, etc. | ✅︎ | ✅︎ |
 | `Glm4vForConditionalGeneration` | GLM-4.1V-Thinking | T + I<sup>E+</sup> + V<sup>E+</sup> | `zai-org/GLM-4.1V-9B-Thinking`, etc. | ✅︎ | ✅︎ |
 | `Glm4vMoeForConditionalGeneration` | GLM-4.5V | T + I<sup>E+</sup> + V<sup>E+</sup> | `zai-org/GLM-4.5V`, etc. | ✅︎ | ✅︎ |
-| `GlmOcrForConditionalGeneration` | GLM-OCR | T + I<sup>E+</sup>  | `zai-org/GLM-OCR`, etc. | ✅︎ | ✅︎ |
 | `GraniteSpeechForConditionalGeneration` | Granite Speech | T + A | `ibm-granite/granite-speech-3.3-8b` | ✅︎ | ✅︎ |
 | `H2OVLChatModel` | H2OVL | T + I<sup>E+</sup> | `h2oai/h2ovl-mississippi-800m`, `h2oai/h2ovl-mississippi-2b`, etc. | | ✅︎ |
 | `HunYuanVLForConditionalGeneration` | HunyuanOCR | T + I<sup>E+</sup> | `tencent/HunyuanOCR`, etc. | ✅︎ | ✅︎ |

examples/offline_inference/qwen3_omni/only_thinker.py

@@ -2,7 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 This example shows how to use vLLM for running offline inference
-with the correct prompt format on Qwen3-Omni (thinker only).
+with the correct prompt format on Qwen2.5-Omni (thinker only).
 """
 
 from typing import NamedTuple
@@ -112,51 +112,23 @@ def get_multi_audios_query() -> QueryResult:
     )
 
 
-def get_multi_images_query() -> QueryResult:
-    question = "What are the differences between these two images?"
-    prompt = (
-        f"<|im_start|>system\n{default_system}<|im_end|>\n"
-        "<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>"
-        "<|vision_start|><|image_pad|><|vision_end|>"
-        f"{question}<|im_end|>\n"
-        f"<|im_start|>assistant\n"
-    )
-    return QueryResult(
-        inputs={
-            "prompt": prompt,
-            "multi_modal_data": {
-                "image": [
-                    convert_image_mode(ImageAsset("cherry_blossom").pil_image, "RGB"),
-                    convert_image_mode(ImageAsset("stop_sign").pil_image, "RGB"),
-                ],
-            },
-        },
-        limit_mm_per_prompt={
-            "image": 2,
-        },
-    )
-
-
 query_map = {
     "mixed_modalities": get_mixed_modalities_query,
     "use_audio_in_video": get_use_audio_in_video_query,
     "multi_audios": get_multi_audios_query,
-    "multi_images": get_multi_images_query,
 }
 
 
 def main(args):
-    model_name = args.model
+    model_name = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
     query_result = query_map[args.query_type]()
 
     llm = LLM(
         model=model_name,
-        max_model_len=args.max_model_len,
+        max_model_len=12800,
         max_num_seqs=5,
         limit_mm_per_prompt=query_result.limit_mm_per_prompt,
         seed=args.seed,
-        tensor_parallel_size=args.tensor_parallel_size,
-        gpu_memory_utilization=args.gpu_memory_utilization,
     )
 
     # We set temperature to 0.2 so that outputs can be different
@@ -189,31 +161,6 @@ def parse_args():
         default=0,
         help="Set the seed when initializing `vllm.LLM`.",
     )
-    parser.add_argument(
-        "--model",
-        type=str,
-        default="Qwen/Qwen3-Omni-30B-A3B-Instruct",
-        help="Model name or path.",
-    )
-    parser.add_argument(
-        "--tensor-parallel-size",
-        "-tp",
-        type=int,
-        default=1,
-        help="Tensor parallel size for distributed inference.",
-    )
-    parser.add_argument(
-        "--gpu-memory-utilization",
-        type=float,
-        default=0.9,
-        help="GPU memory utilization (0.0 to 1.0).",
-    )
-    parser.add_argument(
-        "--max-model-len",
-        type=int,
-        default=12800,
-        help="Maximum model context length.",
-    )
 
     return parser.parse_args()
 

examples/offline_inference/vision_language.py

@@ -566,42 +566,6 @@ def run_glm4_5v_fp8(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
-# GLM-OCR
-def run_glm_ocr(questions: list[str], modality: str) -> ModelRequestData:
-    model_name = "zai-org/GLM-OCR"
-
-    engine_args = EngineArgs(
-        model=model_name,
-        max_model_len=4096,
-        max_num_seqs=2,
-        mm_processor_kwargs={
-            "size": {"shortest_edge": 12544, "longest_edge": 47040000},
-            "fps": 1,
-        },
-        limit_mm_per_prompt={modality: 1},
-        enforce_eager=True,
-    )
-
-    if modality == "image":
-        placeholder = "<|begin_of_image|><|image|><|end_of_image|>"
-    elif modality == "video":
-        placeholder = "<|begin_of_video|><|video|><|end_of_video|>"
-
-    prompts = [
-        (
-            "[gMASK]<sop><|system|>\nYou are a helpful assistant.<|user|>\n"
-            f"{placeholder}"
-            f"{question}<|assistant|>assistant\n"
-        )
-        for question in questions
-    ]
-
-    return ModelRequestData(
-        engine_args=engine_args,
-        prompts=prompts,
-    )
-
-
 # H2OVL-Mississippi
 def run_h2ovl(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -1925,32 +1889,6 @@ def run_step3(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
-# StepVL10B
-def run_step_vl(questions: list[str], modality: str) -> ModelRequestData:
-    assert modality == "image"
-
-    model_name = "stepfun-ai/Step3-VL-10B"
-    engine_args = EngineArgs(
-        model=model_name,
-        max_num_batched_tokens=4096,
-        tensor_parallel_size=1,
-        trust_remote_code=True,
-        limit_mm_per_prompt={modality: 1},
-        reasoning_parser="deepseek_r1",
-    )
-
-    prompts = [
-        "<｜begin▁of▁sentence｜> You are a helpful assistant.<|BOT|>user\n "
-        f"<im_patch>{question} <|EOT|><|BOT|>assistant\n<think>\n"
-        for question in questions
-    ]
-
-    return ModelRequestData(
-        engine_args=engine_args,
-        prompts=prompts,
-    )
-
-
 # omni-research/Tarsier-7b
 def run_tarsier(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -2024,7 +1962,6 @@ model_example_map = {
     "glm4_1v": run_glm4_1v,
     "glm4_5v": run_glm4_5v,
     "glm4_5v_fp8": run_glm4_5v_fp8,
-    "glm_ocr": run_glm_ocr,
     "h2ovl_chat": run_h2ovl,
     "hunyuan_vl": run_hunyuan_vl,
     "hyperclovax_seed_vision": run_hyperclovax_seed_vision,
@@ -2069,7 +2006,6 @@ model_example_map = {
     "skywork_chat": run_skyworkr1v,
     "smolvlm": run_smolvlm,
     "step3": run_step3,
-    "stepvl": run_step_vl,
     "tarsier": run_tarsier,
     "tarsier2": run_tarsier2,
 }
@@ -2077,7 +2013,6 @@ model_example_map = {
 
 MODELS_NEED_VIDEO_METADATA = [
     "glm4_1v",
-    "glm_ocr",
     "glm4_5v",
     "glm4_5v_fp8",
     "molmo2",

examples/offline_inference/vision_language_multi_image.py

@@ -1182,32 +1182,6 @@ def load_step3(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
-def load_step_vl(question: str, image_urls: list[str]) -> ModelRequestData:
-    model_name = "stepfun-ai/Step3-VL-10B"
-
-    engine_args = EngineArgs(
-        model=model_name,
-        max_num_batched_tokens=4096,
-        limit_mm_per_prompt={"image": len(image_urls)},
-        hf_overrides={"vision_config": {"enable_patch": False}},
-        trust_remote_code=True,
-        reasoning_parser="deepseek_r1",
-    )
-
-    prompt = (
-        "<｜begin▁of▁sentence｜> You are a helpful assistant.<|BOT|>user\n "
-        f"{'<im_patch>' * len(image_urls)}{question}<|EOT|><|BOT|>"
-        "assistant\n<think>\n"
-    )
-    image_data = [fetch_image(url) for url in image_urls]
-
-    return ModelRequestData(
-        engine_args=engine_args,
-        prompt=prompt,
-        image_data=image_data,
-    )
-
-
 def load_tarsier(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "omni-research/Tarsier-7b"
 
@@ -1400,7 +1374,6 @@ model_example_map = {
     "rvl": load_r_vl,
     "smolvlm": load_smolvlm,
     "step3": load_step3,
-    "stepvl": load_step_vl,
     "tarsier": load_tarsier,
     "tarsier2": load_tarsier2,
     "glm4_5v": load_glm4_5v,

examples/others/logging_configuration.md

@@ -157,37 +157,6 @@ VLLM_CONFIGURE_LOGGING=0 \
     vllm serve mistralai/Mistral-7B-v0.1 --max-model-len 2048
 ```
 
-### Example 4: Disable access logs for health check endpoints
-
-In production environments, health check endpoints like `/health`, `/metrics`,
-and `/ping` are frequently called by load balancers and monitoring systems,
-generating a large volume of repetitive access logs. To reduce log noise while
-keeping logs for other endpoints, use the `--disable-access-log-for-endpoints`
-option.
-
-**Disable access logs for health and metrics endpoints:**
-
-```bash
-vllm serve mistralai/Mistral-7B-v0.1 --max-model-len 2048 \
-    --disable-access-log-for-endpoints /health,/metrics,/ping
-```
-
-**Common endpoints to consider filtering:**
-
-| Endpoint   | Description            | Typical Caller                                       |
-| ---------- | ---------------------- | ---------------------------------------------------- |
-| `/health`  | Health check           | Kubernetes liveness/readiness probes, load balancers |
-| `/metrics` | Prometheus metrics     | Prometheus scraper (every 15-60s)                    |
-| `/ping`    | SageMaker health check | SageMaker infrastructure                             |
-| `/load`    | Server load metrics    | Custom monitoring                                    |
-
-**Notes:**
-
-- This option only affects uvicorn access logs, not vLLM application logs
-- Specify multiple endpoints by separating them with commas (no spaces)
-- The filter uses exact path matching, query parameters are ignored (e.g., `/health?verbose=true` matches `/health`)
-- If you need to completely disable all access logs, use `--disable-uvicorn-access-log` instead
-
 ## Additional resources
 
 - [`logging.config` Dictionary Schema Details](https://docs.python.org/3/library/logging.config.html#dictionary-schema-details)

pyproject.toml

@@ -9,7 +9,7 @@ requires = [
     "torch == 2.9.1",
     "wheel",
     "jinja2",
-    "grpcio-tools>=1.76.0",
+    "grpcio-tools",
 ]
 build-backend = "setuptools.build_meta"
 
@@ -44,7 +44,6 @@ vllm = "vllm.entrypoints.cli.main:main"
 
 [project.entry-points."vllm.general_plugins"]
 lora_filesystem_resolver = "vllm.plugins.lora_resolvers.filesystem_resolver:register_filesystem_resolver"
-lora_hf_hub_resolver = "vllm.plugins.lora_resolvers.hf_hub_resolver:register_hf_hub_resolver"
 
 [tool.setuptools_scm]
 # no extra settings needed, presence enables setuptools-scm

requirements/build.txt

@@ -9,5 +9,5 @@ wheel
 jinja2>=3.1.6
 regex
 build
-protobuf>=6.33.2
-grpcio-tools>=1.76.0
+protobuf
+grpcio-tools

requirements/common.txt

@@ -9,7 +9,7 @@ blake3
 py-cpuinfo
 transformers >= 4.56.0, < 5
 tokenizers >= 0.21.1  # Required for fast incremental detokenization.
-protobuf >= 6.30.0 # Required by LlamaTokenizer, gRPC.
+protobuf # Required by LlamaTokenizer, gRPC.
 fastapi[standard] >= 0.115.0 # Required by FastAPI's form models in the OpenAI API server's audio transcriptions endpoint.
 aiohttp
 openai >= 1.99.1  # For Responses API with reasoning content
@@ -51,5 +51,5 @@ openai-harmony >= 0.0.3  # Required for gpt-oss
 anthropic >= 0.71.0
 model-hosting-container-standards >= 0.1.13, < 1.0.0
 mcp
-grpcio>=1.76.0
-grpcio-reflection>=1.76.0
+grpcio
+grpcio-reflection

requirements/cpu.txt

@@ -3,7 +3,7 @@
 
 setuptools==77.0.3 # this version can reuse CMake build dir
 
-numba == 0.63.1; platform_machine != "s390x" # Required for N-gram speculative decoding
+numba == 0.61.2; platform_machine != "s390x" # Required for N-gram speculative decoding
 
 # Dependencies for CPUs
 torch==2.10.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"

requirements/cuda.txt

@@ -1,7 +1,7 @@
 # Common dependencies
 -r common.txt
 
-numba == 0.63.1 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 
 # Dependencies for NVIDIA GPUs
 ray[cgraph]>=2.48.0 # Ray Compiled Graph, required for pipeline parallelism in V1.

requirements/nightly_torch_test.txt

@@ -40,7 +40,7 @@ buildkite-test-collector==0.1.9
 genai_perf>=0.0.8
 tritonclient>=2.51.0
 
-numba == 0.63.1 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
 runai-model-streamer[s3,gcs]==0.15.3
 fastsafetensors>=0.1.10

requirements/rocm.txt

@@ -1,7 +1,7 @@
 # Common dependencies
 -r common.txt
 
-numba == 0.63.1 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 
 # Dependencies for AMD GPUs
 datasets
@@ -15,4 +15,4 @@ setuptools-scm>=8
 runai-model-streamer[s3,gcs]==0.15.3
 conch-triton-kernels==1.2.1
 timm>=1.0.17
-grpcio-tools>=1.76.0
+grpcio-tools>=1.76.0

requirements/test.in

@@ -49,7 +49,7 @@ genai_perf>=0.0.8
 tritonclient>=2.51.0
 
 arctic-inference == 0.1.1 # Required for suffix decoding test
-numba == 0.63.1 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
 runai-model-streamer[s3,gcs]==0.15.3
 fastsafetensors>=0.1.10

requirements/test.txt

@@ -456,7 +456,7 @@ lightning-utilities==0.14.3
     #   lightning
     #   pytorch-lightning
     #   torchmetrics
-llvmlite==0.46.0
+llvmlite==0.44.0
     # via numba
 lm-eval==0.4.9.2
     # via -r requirements/test.in
@@ -527,7 +527,7 @@ nltk==3.9.1
     # via rouge-score
 num2words==0.5.14
     # via -r requirements/test.in
-numba==0.63.1
+numba==0.61.2
     # via
     #   -r requirements/test.in
     #   librosa

requirements/xpu.txt

@@ -9,7 +9,7 @@ setuptools>=77.0.3,<81.0.0
 wheel
 jinja2>=3.1.6
 datasets # for benchmark scripts
-numba == 0.63.1 # Required for N-gram speculative decoding
+numba == 0.61.2 # Required for N-gram speculative decoding
 --extra-index-url=https://download.pytorch.org/whl/xpu
 torch==2.9.0+xpu
 torchaudio

tests/entrypoints/openai/test_chat_error.py

@@ -42,7 +42,6 @@ class MockModelConfig:
     tokenizer_revision = None
     multimodal_config = MultiModalConfig()
     hf_config = MockHFConfig()
-    hf_text_config = MockHFConfig()
     logits_processor_pattern = None
     logits_processors: list[str] | None = None
     diff_sampling_param: dict | None = None

tests/entrypoints/openai/test_serving_chat.py

@@ -518,7 +518,6 @@ class MockModelConfig:
     tokenizer_revision = None
     multimodal_config = MultiModalConfig()
     hf_config = MockHFConfig()
-    hf_text_config = MockHFConfig()
     logits_processors: list[str] | None = None
     logits_processor_pattern = None
     diff_sampling_param: dict | None = None

tests/entrypoints/openai/test_transcription_validation_whisper.py

@@ -112,14 +112,18 @@ async def test_long_audio_request(mary_had_lamb, whisper_client):
 @pytest.mark.asyncio
 async def test_completion_endpoints(whisper_client):
     # text to text model
-    with pytest.raises(openai.NotFoundError):
-        await whisper_client.chat.completions.create(
-            model=MODEL_NAME,
-            messages=[{"role": "system", "content": "You are a helpful assistant."}],
-        )
+    res = await whisper_client.chat.completions.create(
+        model=MODEL_NAME,
+        messages=[{"role": "system", "content": "You are a helpful assistant."}],
+    )
+    err = res.error
+    assert err["code"] == 400
+    assert err["message"] == "The model does not support Chat Completions API"
 
-    with pytest.raises(openai.NotFoundError):
-        await whisper_client.completions.create(model=MODEL_NAME, prompt="Hello")
+    res = await whisper_client.completions.create(model=MODEL_NAME, prompt="Hello")
+    err = res.error
+    assert err["code"] == 400
+    assert err["message"] == "The model does not support Completions API"
 
 
 @pytest.mark.asyncio

tests/entrypoints/openai/test_translation_validation.py

@@ -9,7 +9,6 @@ import json
 import httpx
 import librosa
 import numpy as np
-import openai
 import pytest
 import pytest_asyncio
 import soundfile as sf
@@ -53,11 +52,12 @@ async def test_non_asr_model(foscolo, rocm_aiter_fa_attention):
         model_name, _get_server_args(rocm_aiter_fa_attention)
     ) as remote_server:
         client = remote_server.get_async_client()
-
-        with pytest.raises(openai.NotFoundError):
-            await client.audio.translations.create(
-                model=model_name, file=foscolo, temperature=0.0
-            )
+        res = await client.audio.translations.create(
+            model=model_name, file=foscolo, temperature=0.0
+        )
+        err = res.error
+        assert err["code"] == 400 and not res.text
+        assert err["message"] == "The model does not support Translations API"
 
 
 @pytest.mark.asyncio

tests/entrypoints/pooling/classify/test_online.py

@@ -401,7 +401,7 @@ async def test_score(server: RemoteOpenAIServer, model_name: str):
             "documents": "pong",
         },
     )
-    assert response.json()["detail"] == "Not Found"
+    assert response.json()["error"]["type"] == "BadRequestError"
 
 
 @pytest.mark.asyncio
@@ -416,7 +416,7 @@ async def test_rerank(server: RemoteOpenAIServer, model_name: str):
             "documents": ["pong"],
         },
     )
-    assert response.json()["detail"] == "Not Found"
+    assert response.json()["error"]["type"] == "BadRequestError"
 
 
 @pytest.mark.asyncio

tests/kernels/moe/modular_kernel_tools/common.py

@@ -22,9 +22,6 @@ from vllm.distributed import (
 )
 from vllm.forward_context import set_forward_context
 from vllm.model_executor.layers.fused_moe import fused_topk
-from vllm.model_executor.layers.fused_moe.all2all_utils import (
-    maybe_make_prepare_finalize,
-)
 from vllm.model_executor.layers.fused_moe.config import (
     FusedMoEConfig,
     FusedMoEParallelConfig,
@@ -43,6 +40,7 @@ from .mk_objects import (
     TestMoEQuantConfig,
     expert_info,
     make_fused_experts,
+    make_prepare_finalize,
     prepare_finalize_info,
 )
 from .parallel_utils import ProcessGroupInfo
@@ -605,12 +603,10 @@ def make_modular_kernel(
         routing_method=RoutingMethodType.DeepSeekV3,
     )
 
-    prepare_finalize = maybe_make_prepare_finalize(
-        moe=moe,
-        quant_config=quant_config,
-        allow_new_interface=True,
+    # make modular kernel
+    prepare_finalize = make_prepare_finalize(
+        config.prepare_finalize_type, config.all2all_backend(), moe, quant_config
     )
-    assert prepare_finalize is not None
 
     fused_experts = make_fused_experts(
         config.fused_experts_type,

tests/kernels/moe/modular_kernel_tools/mk_objects.py

@@ -7,6 +7,9 @@ import torch
 # Fused experts and PrepareFinalize imports
 import vllm.model_executor.layers.fused_moe.modular_kernel as mk
 from vllm.model_executor.layers.fused_moe import TritonExperts
+from vllm.model_executor.layers.fused_moe.all2all_utils import (
+    maybe_make_prepare_finalize,
+)
 from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
     BatchedDeepGemmExperts,
 )
@@ -252,12 +255,13 @@ if has_pplx():
     )
 
 if has_flashinfer_cutlass_fused_moe() and current_platform.has_device_capability(100):
-    from vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize import (  # noqa: E501
-        FlashInferCutlassMoEPrepareAndFinalize,
-    )
     from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (
         FlashInferExperts,
     )
+    from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize import (  # noqa: E501
+        FlashInferCutlassMoEPrepareAndFinalize,
+        create_flashinfer_prepare_finalize,
+    )
 
     register_prepare_and_finalize(
         FlashInferCutlassMoEPrepareAndFinalize,
@@ -425,6 +429,24 @@ if cutlass_fp4_supported() or has_flashinfer_cutlass_fused_moe():
     ]
 
 
+def make_prepare_finalize(
+    prepare_finalize_type: mk.FusedMoEPrepareAndFinalize,
+    backend: str | None,
+    moe: FusedMoEConfig,
+    quant_config: FusedMoEQuantConfig,
+) -> mk.FusedMoEPrepareAndFinalize:
+    if backend != "naive" and backend is not None:
+        prepare_finalize = maybe_make_prepare_finalize(moe, quant_config)
+        assert prepare_finalize is not None
+        return prepare_finalize
+    elif prepare_finalize_type == FlashInferCutlassMoEPrepareAndFinalize:
+        return create_flashinfer_prepare_finalize(
+            use_dp=moe.moe_parallel_config.dp_size > 1
+        )
+    else:
+        return MoEPrepareAndFinalizeNoEP()
+
+
 def _slice(rank: int, num_local_experts: int, t: torch.Tensor) -> torch.Tensor:
     s = rank * num_local_experts
     e = s + num_local_experts

tests/kernels/moe/test_flashinfer.py

@@ -294,7 +294,12 @@ def test_flashinfer_cutlass_moe_fp8_no_graph(
         )
 
         kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            MoEPrepareAndFinalizeNoEP(
+                defer_input_quant=FlashInferExperts.expects_unquantized_inputs(
+                    moe_config=moe_config,
+                    quant_config=quant_config,
+                )
+            ),
             FlashInferExperts(
                 moe_config=moe_config,
                 quant_config=quant_config,

tests/kernels/moe/test_flashinfer_moe.py

@@ -106,7 +106,12 @@ def test_flashinfer_fp4_moe_no_graph(
         )
 
         flashinfer_experts = FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            MoEPrepareAndFinalizeNoEP(
+                defer_input_quant=FlashInferExperts.expects_unquantized_inputs(
+                    moe_config=moe_config,
+                    quant_config=quant_config,
+                )
+            ),
             FlashInferExperts(moe_config=moe_config, quant_config=quant_config),
         )
 

tests/kernels/moe/test_nvfp4_moe.py

@@ -90,7 +90,7 @@ def test_cutlass_fp4_moe_no_graph(
         )
 
         kernel = mk.FusedMoEModularKernel(
-            MoEPrepareAndFinalizeNoEP(),
+            MoEPrepareAndFinalizeNoEP(defer_input_quant=True),
             CutlassExpertsFp4(
                 moe_config=make_dummy_moe_config(),
                 quant_config=quant_config,

tests/models/language/pooling/test_token_classification.py

@@ -1,8 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import random
-
-import numpy as np
 import pytest
 import torch
 from transformers import AutoModelForTokenClassification
@@ -11,20 +8,6 @@ from tests.models.utils import softmax
 from vllm.platforms import current_platform
 
 
-@pytest.fixture(autouse=True)
-def seed_everything():
-    """Seed all random number generators for reproducibility."""
-    seed = 0
-    random.seed(seed)
-    np.random.seed(seed)
-    torch.manual_seed(seed)
-    if torch.cuda.is_available():
-        torch.cuda.manual_seed_all(seed)
-    torch.backends.cudnn.deterministic = True
-    torch.backends.cudnn.benchmark = False
-    yield
-
-
 @pytest.mark.parametrize("model", ["boltuix/NeuroBERT-NER"])
 # The float32 is required for this tiny model to pass the test.
 @pytest.mark.parametrize("dtype", ["float"])
@@ -68,7 +51,6 @@ def test_bert_models(
 
 @pytest.mark.parametrize("model", ["disham993/electrical-ner-ModernBERT-base"])
 @pytest.mark.parametrize("dtype", ["float"])
-@pytest.mark.flaky(reruns=3)
 @torch.inference_mode
 def test_modernbert_models(
     hf_runner,
@@ -77,15 +59,6 @@ def test_modernbert_models(
     model: str,
     dtype: str,
 ) -> None:
-    # NOTE: https://github.com/vllm-project/vllm/pull/32403
-    # `disham993/electrical-ner-ModernBERT-base` is a randomly initialized
-    # model, which can cause numerical precision variance and edge cases.
-    # We use @flaky(reruns=3) to mitigate intermittent failures.
-    print(
-        f"\n[NOTE] Testing {model} (randomly initialized weights) - "
-        "flaky tolerance enabled due to numerical precision variance."
-    )
-
     with vllm_runner(model, max_model_len=None, dtype=dtype) as vllm_model:
         vllm_outputs = vllm_model.token_classify(example_prompts)
 

tests/models/multimodal/generation/test_common.py

@@ -458,20 +458,6 @@ VLM_TEST_SETTINGS = {
         ],
         marks=[large_gpu_mark(min_gb=32)],
     ),
-    "glm_ocr": VLMTestInfo(
-        models=["zai-org/GLM-OCR"],
-        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
-        prompt_formatter=lambda img_prompt: f"[gMASK]<|user|>\n{img_prompt}<|assistant|>\n",  # noqa: E501
-        img_idx_to_prompt=lambda idx: "<|begin_of_image|><|image|><|end_of_image|>",
-        video_idx_to_prompt=lambda idx: "<|begin_of_video|><|video|><|end_of_video|>",
-        max_model_len=2048,
-        max_num_seqs=2,
-        get_stop_token_ids=lambda tok: [151329, 151336, 151338],
-        num_logprobs=10,
-        image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
-        auto_cls=AutoModelForImageTextToText,
-        marks=[large_gpu_mark(min_gb=32)],
-    ),
     "h2ovl": VLMTestInfo(
         models=[
             "h2oai/h2ovl-mississippi-800m",

tests/models/multimodal/generation/test_vit_backend_functionality.py

@@ -91,19 +91,6 @@ MODEL_CONFIGS: dict[str, dict[str, Any]] = {
         "use_processor": True,
         "question": "What is the content of each image?",
     },
-    "glm_ocr": {
-        "model_name": "zai-org/GLM-OCR",
-        "interface": "llm_generate",
-        "max_model_len": 131072,
-        "max_num_seqs": 2,
-        "sampling_params": {
-            "temperature": 0.0,
-            "max_tokens": 256,
-            "stop_token_ids": None,
-        },
-        "use_processor": True,
-        "question": "Text Recognition:",
-    },
     "keye_vl": {
         "model_name": "Kwai-Keye/Keye-VL-8B-Preview",
         "interface": "llm_generate",

tests/models/multimodal/processing/test_common.py

@@ -122,7 +122,6 @@ MM_DATA_PATCHES = {
     "ernie4_5_moe_vl": qwen3_vl_patch_mm_data,
     "glm4v": glm4_1v_patch_mm_data,
     "glm4v_moe": glm4_1v_patch_mm_data,
-    "glm_ocr": glm4_1v_patch_mm_data,
     "glmasr": glmasr_patch_mm_data,
     "molmo2": qwen3_vl_patch_mm_data,
     "qwen3_vl": qwen3_vl_patch_mm_data,

tests/models/multimodal/test_mapping.py

@@ -30,12 +30,7 @@ def create_dummy_model(repo: str, model_arch: str) -> PreTrainedModel:
     model_cls: PreTrainedModel = getattr(transformers, model_arch)
     config = AutoConfig.from_pretrained(repo)
     with torch.device("meta"):
-        model = model_cls._from_config(config)
-    # TODO(hmellor): Remove this once Transformers has fixed tied weights on meta device
-    # https://github.com/huggingface/transformers/issues/43522
-    if getattr(config.get_text_config(), "tie_word_embeddings", False):
-        model.tie_weights()
-    return model
+        return model_cls._from_config(config)
 
 
 def model_architectures_for_test() -> list[str]:

tests/models/registry.py

@@ -256,7 +256,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     ),
     "Exaone4ForCausalLM": _HfExamplesInfo("LGAI-EXAONE/EXAONE-4.0-32B"),
     "ExaoneMoEForCausalLM": _HfExamplesInfo(
-        "LGAI-EXAONE/K-EXAONE-236B-A23B", min_transformers_version="5.1.0"
+        "LGAI-EXAONE/K-EXAONE-236B-A23B", min_transformers_version="5.0.0"
     ),
     "Fairseq2LlamaForCausalLM": _HfExamplesInfo("mgleize/fairseq2-dummy-Llama-3.2-1B"),
     "FalconForCausalLM": _HfExamplesInfo("tiiuae/falcon-7b"),
@@ -273,7 +273,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "Glm4MoeForCausalLM": _HfExamplesInfo("zai-org/GLM-4.5"),
     "Glm4MoeLiteForCausalLM": _HfExamplesInfo(
         "zai-org/GLM-4.7-Flash",
-        min_transformers_version="5.0.0",
+        min_transformers_version="5.0.0.dev",
+        is_available_online=False,
     ),
     "GPT2LMHeadModel": _HfExamplesInfo("openai-community/gpt2", {"alias": "gpt2"}),
     "GPTBigCodeForCausalLM": _HfExamplesInfo(
@@ -650,7 +651,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     # [Decoder-only]
     "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
     "AudioFlamingo3ForConditionalGeneration": _HfExamplesInfo(
-        "nvidia/audio-flamingo-3-hf", min_transformers_version="5.0.0"
+        "nvidia/audio-flamingo-3-hf", min_transformers_version="5.0.0.dev"
     ),
     "AyaVisionForConditionalGeneration": _HfExamplesInfo("CohereLabs/aya-vision-8b"),
     "BagelForConditionalGeneration": _HfExamplesInfo("ByteDance-Seed/BAGEL-7B-MoT"),
@@ -693,7 +694,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "GlmAsrForConditionalGeneration": _HfExamplesInfo(
         "zai-org/GLM-ASR-Nano-2512",
         trust_remote_code=True,
-        min_transformers_version="5.0.0",
+        min_transformers_version="5.0",
     ),
     "GraniteVision": _HfExamplesInfo("ibm-granite/granite-vision-3.3-2b"),
     "GraniteSpeechForConditionalGeneration": _HfExamplesInfo(
@@ -706,11 +707,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     ),
     "Glm4vForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.1V-9B-Thinking"),
     "Glm4vMoeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V"),
-    "GlmOcrForConditionalGeneration": _HfExamplesInfo(
-        "zai-org/GLM-OCR",
-        is_available_online=False,
-        min_transformers_version="5.1.0",
-    ),
     "H2OVLChatModel": _HfExamplesInfo(
         "h2oai/h2ovl-mississippi-800m",
         trust_remote_code=True,
@@ -1053,7 +1049,7 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
     "ExaoneMoeMTP": _HfExamplesInfo(
         "LGAI-EXAONE/K-EXAONE-236B-A23B",
         speculative_model="LGAI-EXAONE/K-EXAONE-236B-A23B",
-        min_transformers_version="5.1.0",
+        min_transformers_version="5.0.0",
     ),
     "Glm4MoeMTPModel": _HfExamplesInfo(
         "zai-org/GLM-4.5",
@@ -1062,13 +1058,7 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
     "Glm4MoeLiteMTPModel": _HfExamplesInfo(
         "zai-org/GLM-4.7-Flash",
         speculative_model="zai-org/GLM-4.7-Flash",
-        min_transformers_version="5.0.0",
-    ),
-    "GlmOcrMTPModel": _HfExamplesInfo(
-        "zai-org/GLM-OCR",
-        speculative_model="zai-org/GLM-OCR",
         is_available_online=False,
-        min_transformers_version="5.1.0",
     ),
     "LongCatFlashMTPModel": _HfExamplesInfo(
         "meituan-longcat/LongCat-Flash-Chat",
@@ -1095,27 +1085,27 @@ _SPECULATIVE_DECODING_EXAMPLE_MODELS = {
 
 _TRANSFORMERS_BACKEND_MODELS = {
     "TransformersEmbeddingModel": _HfExamplesInfo(
-        "BAAI/bge-base-en-v1.5", min_transformers_version="5.0.0"
+        "BAAI/bge-base-en-v1.5", min_transformers_version="5.0.0.dev"
     ),
     "TransformersForSequenceClassification": _HfExamplesInfo(
         "papluca/xlm-roberta-base-language-detection",
-        min_transformers_version="5.0.0",
+        min_transformers_version="5.0.0.dev",
     ),
     "TransformersForCausalLM": _HfExamplesInfo(
         "hmellor/Ilama-3.2-1B", trust_remote_code=True
     ),
     "TransformersMultiModalForCausalLM": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
     "TransformersMoEForCausalLM": _HfExamplesInfo(
-        "allenai/OLMoE-1B-7B-0924", min_transformers_version="5.0.0"
+        "allenai/OLMoE-1B-7B-0924", min_transformers_version="5.0.0.dev"
     ),
     "TransformersMultiModalMoEForCausalLM": _HfExamplesInfo(
-        "Qwen/Qwen3-VL-30B-A3B-Instruct", min_transformers_version="5.0.0"
+        "Qwen/Qwen3-VL-30B-A3B-Instruct", min_transformers_version="5.0.0.dev"
     ),
     "TransformersMoEEmbeddingModel": _HfExamplesInfo(
-        "Qwen/Qwen3-30B-A3B", min_transformers_version="5.0.0"
+        "Qwen/Qwen3-30B-A3B", min_transformers_version="5.0.0.dev"
     ),
     "TransformersMoEForSequenceClassification": _HfExamplesInfo(
-        "Qwen/Qwen3-30B-A3B", min_transformers_version="5.0.0"
+        "Qwen/Qwen3-30B-A3B", min_transformers_version="5.0.0.dev"
     ),
     "TransformersMultiModalEmbeddingModel": _HfExamplesInfo("google/gemma-3-4b-it"),
     "TransformersMultiModalForSequenceClassification": _HfExamplesInfo(

tests/models/test_transformers.py

@@ -78,7 +78,7 @@ def test_models(
     from packaging.version import Version
 
     installed = Version(transformers.__version__)
-    required = Version("5.0.0")
+    required = Version("5.0.0.dev")
     if model == "allenai/OLMoE-1B-7B-0924" and installed < required:
         pytest.skip(
             "MoE models with the Transformers modeling backend require "

tests/plugins/lora_resolvers/test_hf_hub_resolver.py

@@ -1,107 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import os
-
-import pytest
-from huggingface_hub.constants import HF_HUB_CACHE
-
-from vllm.plugins.lora_resolvers.hf_hub_resolver import HfHubResolver
-
-LORA_LIB_MODEL_NAME = "ibm-granite/granite-3.3-8b-instruct"
-# Repo with multiple LoRAs contained in it
-LORA_LIB = "ibm-granite/granite-3.3-8b-rag-agent-lib"
-LORA_NAME = "ibm-granite/granite-3.3-8b-rag-agent-lib/answerability_prediction_lora"  # noqa: E501
-NON_LORA_SUBPATH = "ibm-granite/granite-3.3-8b-rag-agent-lib/README.md"
-LIB_DOWNLOAD_DIR = os.path.join(
-    HF_HUB_CACHE, "models--ibm-granite--granite-3.3-8b-rag-agent-lib"
-)
-INVALID_REPO_NAME = "thisrepodoesnotexist"
-
-# Repo with only one LoRA in the root dir
-LORA_REPO_MODEL_NAME = "meta-llama/Llama-2-7b-hf"
-LORA_REPO = "yard1/llama-2-7b-sql-lora-test"
-REPO_DOWNLOAD_DIR = os.path.join(
-    HF_HUB_CACHE, "models--yard1--llama-2-7b-sql-lora-test"
-)
-
-
-@pytest.mark.asyncio
-async def test_hf_resolver_with_direct_path():
-    hf_resolver = HfHubResolver([LORA_REPO])
-    assert hf_resolver is not None
-
-    lora_request = await hf_resolver.resolve_lora(LORA_REPO_MODEL_NAME, LORA_REPO)
-    assert lora_request.lora_name == LORA_REPO
-    assert REPO_DOWNLOAD_DIR in lora_request.lora_path
-    assert "adapter_config.json" in os.listdir(lora_request.lora_path)
-
-
-@pytest.mark.asyncio
-async def test_hf_resolver_with_nested_paths():
-    hf_resolver = HfHubResolver([LORA_LIB])
-    assert hf_resolver is not None
-
-    lora_request = await hf_resolver.resolve_lora(LORA_LIB_MODEL_NAME, LORA_NAME)
-    assert lora_request is not None
-    assert lora_request.lora_name == LORA_NAME
-    assert LIB_DOWNLOAD_DIR in lora_request.lora_path
-    assert "adapter_config.json" in os.listdir(lora_request.lora_path)
-
-
-@pytest.mark.asyncio
-async def test_hf_resolver_with_multiple_repos():
-    hf_resolver = HfHubResolver([LORA_LIB, LORA_REPO])
-    assert hf_resolver is not None
-
-    lora_request = await hf_resolver.resolve_lora(LORA_LIB_MODEL_NAME, LORA_NAME)
-    assert lora_request is not None
-    assert lora_request.lora_name == LORA_NAME
-    assert LIB_DOWNLOAD_DIR in lora_request.lora_path
-    assert "adapter_config.json" in os.listdir(lora_request.lora_path)
-
-
-@pytest.mark.asyncio
-async def test_missing_adapter():
-    hf_resolver = HfHubResolver([LORA_LIB])
-    assert hf_resolver is not None
-
-    missing_lora_request = await hf_resolver.resolve_lora(LORA_LIB_MODEL_NAME, "foobar")
-    assert missing_lora_request is None
-
-
-@pytest.mark.asyncio
-async def test_nonlora_adapter():
-    hf_resolver = HfHubResolver([LORA_LIB])
-    assert hf_resolver is not None
-
-    readme_request = await hf_resolver.resolve_lora(
-        LORA_LIB_MODEL_NAME, NON_LORA_SUBPATH
-    )
-    assert readme_request is None
-
-
-@pytest.mark.asyncio
-async def test_invalid_repo():
-    hf_resolver = HfHubResolver([LORA_LIB])
-    assert hf_resolver is not None
-
-    invalid_repo_req = await hf_resolver.resolve_lora(
-        INVALID_REPO_NAME,
-        f"{INVALID_REPO_NAME}/foo",
-    )
-    assert invalid_repo_req is None
-
-
-@pytest.mark.asyncio
-async def test_trailing_slash():
-    hf_resolver = HfHubResolver([LORA_LIB])
-    assert hf_resolver is not None
-
-    lora_request = await hf_resolver.resolve_lora(
-        LORA_LIB_MODEL_NAME,
-        f"{LORA_NAME}/",
-    )
-    assert lora_request is not None
-    assert lora_request.lora_name == f"{LORA_NAME}/"
-    assert LIB_DOWNLOAD_DIR in lora_request.lora_path
-    assert "adapter_config.json" in os.listdir(lora_request.lora_path)

tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_gemma_embedding.py

@@ -36,7 +36,7 @@ class MyGemma2Embedding(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

tests/test_access_log_filter.py

@@ -1,371 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-Tests for the UvicornAccessLogFilter class.
-"""
-
-import logging
-
-from vllm.logging_utils.access_log_filter import (
-    UvicornAccessLogFilter,
-    create_uvicorn_log_config,
-)
-
-
-class TestUvicornAccessLogFilter:
-    """Test cases for UvicornAccessLogFilter."""
-
-    def test_filter_allows_all_when_no_excluded_paths(self):
-        """Filter should allow all logs when no paths are excluded."""
-        filter = UvicornAccessLogFilter(excluded_paths=[])
-
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/v1/completions", "1.1", 200),
-            exc_info=None,
-        )
-
-        assert filter.filter(record) is True
-
-    def test_filter_allows_all_when_excluded_paths_is_none(self):
-        """Filter should allow all logs when excluded_paths is None."""
-        filter = UvicornAccessLogFilter(excluded_paths=None)
-
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/health", "1.1", 200),
-            exc_info=None,
-        )
-
-        assert filter.filter(record) is True
-
-    def test_filter_excludes_health_endpoint(self):
-        """Filter should exclude /health endpoint when configured."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health"])
-
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/health", "1.1", 200),
-            exc_info=None,
-        )
-
-        assert filter.filter(record) is False
-
-    def test_filter_excludes_metrics_endpoint(self):
-        """Filter should exclude /metrics endpoint when configured."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/metrics"])
-
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/metrics", "1.1", 200),
-            exc_info=None,
-        )
-
-        assert filter.filter(record) is False
-
-    def test_filter_allows_non_excluded_endpoints(self):
-        """Filter should allow endpoints not in the excluded list."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health", "/metrics"])
-
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "POST", "/v1/completions", "1.1", 200),
-            exc_info=None,
-        )
-
-        assert filter.filter(record) is True
-
-    def test_filter_excludes_multiple_endpoints(self):
-        """Filter should exclude multiple configured endpoints."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health", "/metrics", "/ping"])
-
-        # Test /health
-        record_health = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/health", "1.1", 200),
-            exc_info=None,
-        )
-        assert filter.filter(record_health) is False
-
-        # Test /metrics
-        record_metrics = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/metrics", "1.1", 200),
-            exc_info=None,
-        )
-        assert filter.filter(record_metrics) is False
-
-        # Test /ping
-        record_ping = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/ping", "1.1", 200),
-            exc_info=None,
-        )
-        assert filter.filter(record_ping) is False
-
-    def test_filter_with_query_parameters(self):
-        """Filter should exclude endpoints even with query parameters."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health"])
-
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/health?verbose=true", "1.1", 200),
-            exc_info=None,
-        )
-
-        assert filter.filter(record) is False
-
-    def test_filter_different_http_methods(self):
-        """Filter should exclude endpoints regardless of HTTP method."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/ping"])
-
-        # Test GET
-        record_get = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "GET", "/ping", "1.1", 200),
-            exc_info=None,
-        )
-        assert filter.filter(record_get) is False
-
-        # Test POST
-        record_post = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg='%s - "%s %s HTTP/%s" %d',
-            args=("127.0.0.1:12345", "POST", "/ping", "1.1", 200),
-            exc_info=None,
-        )
-        assert filter.filter(record_post) is False
-
-    def test_filter_with_different_status_codes(self):
-        """Filter should exclude endpoints regardless of status code."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health"])
-
-        for status_code in [200, 500, 503]:
-            record = logging.LogRecord(
-                name="uvicorn.access",
-                level=logging.INFO,
-                pathname="",
-                lineno=0,
-                msg='%s - "%s %s HTTP/%s" %d',
-                args=("127.0.0.1:12345", "GET", "/health", "1.1", status_code),
-                exc_info=None,
-            )
-            assert filter.filter(record) is False
-
-
-class TestCreateUvicornLogConfig:
-    """Test cases for create_uvicorn_log_config function."""
-
-    def test_creates_valid_config_structure(self):
-        """Config should have required logging configuration keys."""
-        config = create_uvicorn_log_config(excluded_paths=["/health"])
-
-        assert "version" in config
-        assert config["version"] == 1
-        assert "disable_existing_loggers" in config
-        assert "formatters" in config
-        assert "handlers" in config
-        assert "loggers" in config
-        assert "filters" in config
-
-    def test_config_includes_access_log_filter(self):
-        """Config should include the access log filter."""
-        config = create_uvicorn_log_config(excluded_paths=["/health", "/metrics"])
-
-        assert "access_log_filter" in config["filters"]
-        filter_config = config["filters"]["access_log_filter"]
-        assert filter_config["()"] == UvicornAccessLogFilter
-        assert filter_config["excluded_paths"] == ["/health", "/metrics"]
-
-    def test_config_applies_filter_to_access_handler(self):
-        """Config should apply the filter to the access handler."""
-        config = create_uvicorn_log_config(excluded_paths=["/health"])
-
-        assert "access" in config["handlers"]
-        assert "filters" in config["handlers"]["access"]
-        assert "access_log_filter" in config["handlers"]["access"]["filters"]
-
-    def test_config_with_custom_log_level(self):
-        """Config should respect custom log level."""
-        config = create_uvicorn_log_config(
-            excluded_paths=["/health"], log_level="debug"
-        )
-
-        assert config["loggers"]["uvicorn"]["level"] == "DEBUG"
-        assert config["loggers"]["uvicorn.access"]["level"] == "DEBUG"
-        assert config["loggers"]["uvicorn.error"]["level"] == "DEBUG"
-
-    def test_config_with_empty_excluded_paths(self):
-        """Config should work with empty excluded paths."""
-        config = create_uvicorn_log_config(excluded_paths=[])
-
-        assert config["filters"]["access_log_filter"]["excluded_paths"] == []
-
-    def test_config_with_none_excluded_paths(self):
-        """Config should work with None excluded paths."""
-        config = create_uvicorn_log_config(excluded_paths=None)
-
-        assert config["filters"]["access_log_filter"]["excluded_paths"] == []
-
-
-class TestIntegration:
-    """Integration tests for the access log filter."""
-
-    def test_filter_with_real_logger(self):
-        """Test filter works with a real Python logger simulating uvicorn."""
-        # Create a logger with our filter (simulating uvicorn.access)
-        logger = logging.getLogger("uvicorn.access")
-        logger.setLevel(logging.INFO)
-
-        # Clear any existing handlers
-        logger.handlers = []
-
-        # Create a custom handler that tracks messages
-        logged_messages: list[str] = []
-
-        class TrackingHandler(logging.Handler):
-            def emit(self, record):
-                logged_messages.append(record.getMessage())
-
-        handler = TrackingHandler()
-        handler.setLevel(logging.INFO)
-        filter = UvicornAccessLogFilter(excluded_paths=["/health", "/metrics"])
-        handler.addFilter(filter)
-        logger.addHandler(handler)
-
-        # Log using uvicorn's format with args tuple
-        # Format: '%s - "%s %s HTTP/%s" %d'
-        logger.info(
-            '%s - "%s %s HTTP/%s" %d',
-            "127.0.0.1:12345",
-            "GET",
-            "/health",
-            "1.1",
-            200,
-        )
-        logger.info(
-            '%s - "%s %s HTTP/%s" %d',
-            "127.0.0.1:12345",
-            "GET",
-            "/v1/completions",
-            "1.1",
-            200,
-        )
-        logger.info(
-            '%s - "%s %s HTTP/%s" %d',
-            "127.0.0.1:12345",
-            "GET",
-            "/metrics",
-            "1.1",
-            200,
-        )
-        logger.info(
-            '%s - "%s %s HTTP/%s" %d',
-            "127.0.0.1:12345",
-            "POST",
-            "/v1/chat/completions",
-            "1.1",
-            200,
-        )
-
-        # Verify only non-excluded endpoints were logged
-        assert len(logged_messages) == 2
-        assert "/v1/completions" in logged_messages[0]
-        assert "/v1/chat/completions" in logged_messages[1]
-
-    def test_filter_allows_non_uvicorn_access_logs(self):
-        """Test filter allows logs from non-uvicorn.access loggers."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health"])
-
-        # Log record from a different logger name
-        record = logging.LogRecord(
-            name="uvicorn.error",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg="Some error message about /health",
-            args=(),
-            exc_info=None,
-        )
-
-        # Should allow because it's not from uvicorn.access
-        assert filter.filter(record) is True
-
-    def test_filter_handles_malformed_args(self):
-        """Test filter handles log records with unexpected args format."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health"])
-
-        # Log record with insufficient args
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg="Some message",
-            args=("only", "two"),
-            exc_info=None,
-        )
-
-        # Should allow because args doesn't have expected format
-        assert filter.filter(record) is True
-
-    def test_filter_handles_non_tuple_args(self):
-        """Test filter handles log records with non-tuple args."""
-        filter = UvicornAccessLogFilter(excluded_paths=["/health"])
-
-        # Log record with None args
-        record = logging.LogRecord(
-            name="uvicorn.access",
-            level=logging.INFO,
-            pathname="",
-            lineno=0,
-            msg="Some message without args",
-            args=None,
-            exc_info=None,
-        )
-
-        # Should allow because args is None
-        assert filter.filter(record) is True

tests/v1/e2e/test_spec_decode.py

@@ -455,7 +455,7 @@ def test_eagle_correctness(
         from packaging.version import Version
 
         installed = Version(transformers.__version__)
-        required = Version("5.0.0")
+        required = Version("5.0.0.dev")
         if installed < required:
             pytest.skip(
                 "Eagle3 with the Transformers modeling backend requires "

tests/v1/kv_connector/unit/test_offloading_connector.py

@@ -16,7 +16,6 @@ from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorRole
 from vllm.distributed.kv_transfer.kv_connector.v1.offloading_connector import (
     OffloadingConnector,
     OffloadingConnectorMetadata,
-    OffloadingConnectorStats,
 )
 from vllm.forward_context import ForwardContext
 from vllm.utils.hashing import sha256
@@ -87,14 +86,7 @@ class MockOffloadingHandler(OffloadingHandler):
             if job_id in self.waiting_jobs:
                 self.waiting_jobs.remove(job_id)
                 self.completed_jobs.append(job_id)
-                result = TransferResult(
-                    job_id=job_id,
-                    success=True,
-                    transfer_size=None,
-                    transfer_time=None,
-                    transfer_type=None,
-                )
-                self.completed_transfers.append(result)
+                self.completed_transfers.append((job_id, True))
 
     def wait(self, job_ids: set[int]) -> None:
         self.flushed_jobs |= job_ids
@@ -728,144 +720,3 @@ def test_concurrent_lookups_of_the_same_prefix(request_runner):
 
     # second request will use the GPU prefix cache
     assert transfer_jobs == list(runner.offloading_spec.handler.transfer_specs)
-
-
-class TestOffloadingConnectorStats:
-    """Tests for OffloadingConnector stats reconstruction and operations."""
-
-    def test_build_kv_connector_stats_with_none(self):
-        """Test that build_kv_connector_stats returns empty stats when given None."""
-        stats = OffloadingConnector.build_kv_connector_stats(data=None)
-
-        assert stats is not None
-        assert isinstance(stats, OffloadingConnectorStats)
-        assert len(stats.data) == 0
-        assert stats.is_empty()
-
-    def test_build_kv_connector_stats_with_empty_dict(self):
-        """Test that build_kv_connector_stats returns empty stats with empty dict."""
-        stats = OffloadingConnector.build_kv_connector_stats(data={})
-
-        assert stats is not None
-        assert isinstance(stats, OffloadingConnectorStats)
-        assert len(stats.data) == 0
-        assert stats.is_empty()
-
-    def test_build_kv_connector_stats_reconstructs_offload_stats(self):
-        """Test that OffloadingConnector stats are properly reconstructed with
-        correct data."""
-        serialized_data = {
-            "CPU_to_GPU": [
-                {"op_size": 16, "op_time": 1.0},
-                {"op_size": 8, "op_time": 0.5},
-            ],
-            "GPU_to_CPU": [
-                {"op_size": 1, "op_time": 0.1},
-                {"op_size": 2, "op_time": 0.2},
-            ],
-        }
-
-        stats = OffloadingConnector.build_kv_connector_stats(data=serialized_data)
-
-        offload_connector_stats = stats
-        assert isinstance(offload_connector_stats, OffloadingConnectorStats)
-        assert offload_connector_stats.data["CPU_to_GPU"] == [
-            {"op_size": 16, "op_time": 1.0},
-            {"op_size": 8, "op_time": 0.5},
-        ]
-        assert offload_connector_stats.data["GPU_to_CPU"] == [
-            {"op_size": 1, "op_time": 0.1},
-            {"op_size": 2, "op_time": 0.2},
-        ]
-
-    def test_aggregate_same_connector(self):
-        """Test aggregating stats from the same connector type."""
-        stats1 = OffloadingConnectorStats(
-            data={
-                "CPU_to_GPU": [
-                    {"op_size": 16, "op_time": 1.0},
-                    {"op_size": 8, "op_time": 0.5},
-                ],
-                "GPU_to_CPU": [
-                    {"op_size": 1, "op_time": 0.1},
-                    {"op_size": 2, "op_time": 0.2},
-                ],
-            }
-        )
-
-        stats2 = OffloadingConnectorStats(
-            data={
-                "CPU_to_GPU": [
-                    {"op_size": 3, "op_time": 0.2},
-                    {"op_size": 7, "op_time": 0.9},
-                ],
-                "GPU_to_CPU": [{"op_size": 16, "op_time": 2}],
-            }
-        )
-
-        result = stats1.aggregate(stats2)
-
-        assert result is stats1  # Should return self
-        offload_connector_stats = result
-        assert offload_connector_stats.data["CPU_to_GPU"] == [
-            {"op_size": 16, "op_time": 1.0},
-            {"op_size": 8, "op_time": 0.5},
-            {"op_size": 3, "op_time": 0.2},
-            {"op_size": 7, "op_time": 0.9},
-        ]
-        assert offload_connector_stats.data["GPU_to_CPU"] == [
-            {"op_size": 1, "op_time": 0.1},
-            {"op_size": 2, "op_time": 0.2},
-            {"op_size": 16, "op_time": 2},
-        ]
-
-    def test_reduce(self):
-        """Test that reduce() correctly reduces all nested connector stats."""
-        stats = OffloadingConnectorStats(
-            data={
-                "CPU_to_GPU": [
-                    {"op_size": 16, "op_time": 1.0},
-                    {"op_size": 8, "op_time": 0.5},
-                    {"op_size": 3, "op_time": 0.2},
-                    {"op_size": 7, "op_time": 0.9},
-                ],
-                "GPU_to_CPU": [
-                    {"op_size": 1, "op_time": 0.1},
-                    {"op_size": 2, "op_time": 0.2},
-                    {"op_size": 16, "op_time": 2},
-                ],
-            }
-        )
-
-        reduced = stats.reduce()
-
-        assert isinstance(reduced, dict)
-        # Check that the stats were reduced (should have aggregated values)
-        assert "CPU_to_GPU_total_bytes" in reduced
-        assert "CPU_to_GPU_total_time" in reduced
-        assert "GPU_to_CPU_total_bytes" in reduced
-        assert "GPU_to_CPU_total_time" in reduced
-        assert reduced["CPU_to_GPU_total_bytes"] == 34
-        assert reduced["CPU_to_GPU_total_time"] == 2.6
-        assert reduced["GPU_to_CPU_total_time"] == 2.3
-        assert reduced["GPU_to_CPU_total_bytes"] == 19
-
-    def test_reset(self):
-        """Test that reset() resets all nested connector stats."""
-        offload_connector_stats = OffloadingConnectorStats(
-            data={
-                "CPU_to_GPU": [
-                    {"op_size": 3, "op_time": 0.2},
-                    {"op_size": 7, "op_time": 0.9},
-                ],
-                "GPU_to_CPU": [{"op_size": 16, "op_time": 2}],
-            }
-        )
-
-        assert not offload_connector_stats.is_empty()
-
-        offload_connector_stats.reset()
-
-        # After reset, stats should be empty
-        assert offload_connector_stats.is_empty()
-        assert len(offload_connector_stats.data) == 0

tests/v1/kv_offload/test_cpu_gpu.py

@@ -168,30 +168,15 @@ def test_transfer(
     orig_dst_caches = [x.clone() for x in handler.dst_tensors]
 
     # call transfer function
-    start_time = time.time()
     assert handler.transfer_async(1, (src_spec, dst_spec))
-    assert set({x.job_id for x in handler._transfers}) == {1}
+    assert set({x[0] for x in handler._transfers}) == {1}
 
     # wait for transfer to complete
     end_time = time.time() + 10
     while time.time() < end_time:
         finished = handler.get_finished()
         if finished:
-            assert finished[0].job_id == 1
-            assert finished[0].success
-            assert (
-                finished[0].transfer_type == ("GPU", "CPU")
-                if gpu_to_cpu
-                else ("CPU", "GPU")
-            )
-            assert (
-                finished[0].transfer_size
-                == handler.total_block_size_in_bytes
-                * handler.dst_block_size_factor
-                * len(dst_blocks)
-            )
-            assert finished[0].transfer_time > 0
-            assert finished[0].transfer_time < (time.time() - start_time)
+            assert finished == [(1, True)]
             break
         time.sleep(0.1)
 

vllm/_custom_ops.py

@@ -2845,13 +2845,13 @@ if hasattr(torch.ops._C, "int8_scaled_mm_with_quant"):
 
 class CPUDNNLGEMMHandler:
     def __init__(self) -> None:
-        self.handler_tensor: torch.Tensor | None = None
+        self.handler: int | None = None
         self.n = -1
         self.k = -1
 
     def __del__(self):
-        if self.handler_tensor is not None:
-            torch.ops._C.release_dnnl_matmul_handler(self.handler_tensor.item())
+        if self.handler is not None:
+            torch.ops._C.release_dnnl_matmul_handler(self.handler)
 
 
 _supports_onednn = bool(hasattr(torch.ops._C, "create_onednn_mm_handler"))
@@ -2867,10 +2867,8 @@ def create_onednn_mm(
 ) -> CPUDNNLGEMMHandler:
     handler = CPUDNNLGEMMHandler()
     handler.k, handler.n = weight.size()
-    # store the handler pointer in a tensor it doesn't get inlined
-    handler.handler_tensor = torch.tensor(
-        torch.ops._C.create_onednn_mm_handler(weight, primitive_cache_size),
-        dtype=torch.int64,
+    handler.handler = torch.ops._C.create_onednn_mm_handler(
+        weight, primitive_cache_size
     )
     return handler
 
@@ -2882,7 +2880,7 @@ def onednn_mm(
 ) -> torch.Tensor:
     output = torch.empty((*x.shape[0:-1], dnnl_handler.n), dtype=x.dtype)
     torch.ops._C.onednn_mm(
-        output, x.reshape(-1, dnnl_handler.k), bias, dnnl_handler.handler_tensor
+        output, x.reshape(-1, dnnl_handler.k), bias, dnnl_handler.handler
     )
 
     return output
@@ -2898,17 +2896,8 @@ def create_onednn_scaled_mm(
 ) -> CPUDNNLGEMMHandler:
     handler = CPUDNNLGEMMHandler()
     handler.k, handler.n = weight.size()
-    # store the handler pointer in a tensor so it doesn't get inlined
-    handler.handler_tensor = torch.tensor(
-        torch.ops._C.create_onednn_scaled_mm_handler(
-            weight,
-            weight_scales,
-            output_type,
-            dynamic_quant,
-            use_azp,
-            primitive_cache_size,
-        ),
-        dtype=torch.int64,
+    handler.handler = torch.ops._C.create_onednn_scaled_mm_handler(
+        weight, weight_scales, output_type, dynamic_quant, use_azp, primitive_cache_size
     )
     return handler
 
@@ -2961,13 +2950,7 @@ def onednn_scaled_mm(
     bias: torch.Tensor | None,
 ) -> torch.Tensor:
     torch.ops._C.onednn_scaled_mm(
-        output,
-        x,
-        input_scale,
-        input_zp,
-        input_zp_adj,
-        bias,
-        dnnl_handler.handler_tensor,
+        output, x, input_scale, input_zp, input_zp_adj, bias, dnnl_handler.handler
     )
 
     return output

vllm/config/compilation.py

@@ -280,10 +280,9 @@ class DynamicShapesConfig:
     until this change picked up https://github.com/pytorch/pytorch/pull/169239.
     """
 
-    assume_32_bit_indexing: bool = False
+    assume_32_bit_indexing: bool = True
     """
     whether all tensor sizes can use 32 bit indexing.
-    `True` requires PyTorch 2.10+
     """
 
     def compute_hash(self) -> str:

vllm/config/speculative.py

@@ -34,7 +34,6 @@ MTPModelTypes = Literal[
     "mimo_mtp",
     "glm4_moe_mtp",
     "glm4_moe_lite_mtp",
-    "glm_ocr_mtp",
     "ernie_mtp",
     "exaone_moe_mtp",
     "qwen3_next_mtp",
@@ -222,17 +221,6 @@ class SpeculativeConfig:
                 }
             )
 
-        if hf_config.architectures[0] == "GlmOcrForConditionalGeneration":
-            hf_config.model_type = "glm_ocr_mtp"
-            n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
-            hf_config.update(
-                {
-                    "num_hidden_layers": 0,
-                    "n_predict": n_predict,
-                    "architectures": ["GlmOcrMTPModel"],
-                }
-            )
-
         if hf_config.model_type == "ernie4_5_moe":
             hf_config.model_type = "ernie_mtp"
         if hf_config.model_type == "ernie_mtp":

vllm/distributed/device_communicators/all2all.py

@@ -59,7 +59,7 @@ class NaiveAll2AllManager(All2AllManagerBase):
 
         return buffer
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
@@ -84,34 +84,6 @@ class NaiveAll2AllManager(All2AllManagerBase):
 
         return hidden_states, router_logits
 
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        if extra_tensors is not None:
-            raise NotImplementedError(
-                "extra_tensors is not supported for NaiveAll2AllManager"
-            )
-        sp_size = self.tp_group.world_size if is_sequence_parallel else 1
-        dp_metadata = get_forward_context().dp_metadata
-        assert dp_metadata is not None
-        cu_tokens_across_sp_cpu = dp_metadata.cu_tokens_across_sp(sp_size)
-
-        hidden_states = self.naive_multicast(
-            hidden_states, cu_tokens_across_sp_cpu, is_sequence_parallel
-        )
-        topk_weights = self.naive_multicast(
-            topk_weights, cu_tokens_across_sp_cpu, is_sequence_parallel
-        )
-        topk_ids = self.naive_multicast(
-            topk_ids, cu_tokens_across_sp_cpu, is_sequence_parallel
-        )
-        return hidden_states, topk_weights, topk_ids
-
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:
@@ -142,7 +114,7 @@ class AgRsAll2AllManager(All2AllManagerBase):
     def __init__(self, cpu_group):
         super().__init__(cpu_group)
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
@@ -176,46 +148,6 @@ class AgRsAll2AllManager(All2AllManagerBase):
             return (gathered_tensors[0], gathered_tensors[1], gathered_tensors[2:])
         return gathered_tensors[0], gathered_tensors[1]
 
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        """
-        Gather hidden_states and router_logits from all dp ranks.
-        """
-        dp_metadata = get_forward_context().dp_metadata
-        assert dp_metadata is not None
-        sizes = dp_metadata.get_chunk_sizes_across_dp_rank()
-        assert sizes is not None
-        dist_group = get_ep_group() if is_sequence_parallel else get_dp_group()
-        assert sizes[dist_group.rank_in_group] == hidden_states.shape[0]
-
-        tensors_to_gather = [hidden_states, topk_weights, topk_ids]
-        if extra_tensors is not None:
-            tensors_to_gather.extend(extra_tensors)
-
-        gathered_tensors = dist_group.all_gatherv(
-            tensors_to_gather,
-            dim=0,
-            sizes=sizes,
-        )
-
-        hidden_states = gathered_tensors[0]
-        topk_weights = gathered_tensors[1]
-        topk_ids = gathered_tensors[2]
-
-        if extra_tensors is None:
-            return hidden_states, topk_weights, topk_ids
-
-        return hidden_states, topk_weights, topk_ids, gathered_tensors[3:]
-
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:
@@ -284,7 +216,7 @@ class PPLXAll2AllManager(All2AllManagerBase):
             pplx.AllToAll.internode if self.internode else pplx.AllToAll.intranode,
         )
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
@@ -293,19 +225,6 @@ class PPLXAll2AllManager(All2AllManagerBase):
     ) -> tuple[torch.Tensor, torch.Tensor]:
         raise NotImplementedError
 
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        raise NotImplementedError
-
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:
@@ -345,7 +264,7 @@ class DeepEPAll2AllManagerBase(All2AllManagerBase):
     def get_handle(self, kwargs):
         raise NotImplementedError
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
@@ -354,19 +273,6 @@ class DeepEPAll2AllManagerBase(All2AllManagerBase):
     ) -> tuple[torch.Tensor, torch.Tensor]:
         raise NotImplementedError
 
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        raise NotImplementedError
-
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:

vllm/distributed/device_communicators/base_device_communicator.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import threading
+from typing import Any
 from weakref import WeakValueDictionary
 
 import torch
@@ -63,32 +64,13 @@ class All2AllManagerBase:
         # and reuse it for the same config.
         raise NotImplementedError
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
         is_sequence_parallel: bool = False,
         extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        # Subclasses should either:
-        # - implement handling for extra_tensors, or
-        # - raise a clear error if extra_tensors is not supported.
-        raise NotImplementedError
-
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
+    ) -> Any:
         # Subclasses should either:
         # - implement handling for extra_tensors, or
         # - raise a clear error if extra_tensors is not supported.
@@ -298,7 +280,7 @@ class DeviceCommunicatorBase:
         for module in moe_modules:
             module.maybe_init_modular_kernel()
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
@@ -312,29 +294,8 @@ class DeviceCommunicatorBase:
         Dispatch the hidden states and router logits to the appropriate device.
         This is a no-op in the base class.
         """
-        if extra_tensors is not None:
-            return hidden_states, router_logits, extra_tensors
         return hidden_states, router_logits
 
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        """
-        Dispatch the hidden states and topk weights/ids to the appropriate device.
-        This is a no-op in the base class.
-        """
-        if extra_tensors is not None:
-            return hidden_states, topk_weights, topk_ids, extra_tensors
-        return hidden_states, topk_weights, topk_ids
-
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:

vllm/distributed/device_communicators/cpu_communicator.py

@@ -130,65 +130,29 @@ class CpuCommunicator(DeviceCommunicatorBase):
     ) -> dict[str, torch.Tensor | Any]:
         return self.dist_module.recv_tensor_dict(src)
 
-    def dispatch_router_logits(
+    def dispatch(  # type: ignore[override]
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
         is_sequence_parallel: bool = False,
         extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        """
-        Dispatch the hidden states and router logits to the appropriate device.
-        This is a no-op in the base class.
-        """
-
-        assert self.all2all_manager is not None
-        return self.all2all_manager.dispatch_router_logits(
-            hidden_states,
-            router_logits,
-            is_sequence_parallel,
-            extra_tensors,
-        )
-
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        """
-        Dispatch the hidden states and topk weights/ids to the appropriate device.
-        This is a no-op in the base class.
-        """
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         assert self.all2all_manager is not None
         return self.all2all_manager.dispatch(
             hidden_states,
-            topk_weights,
-            topk_ids,
+            router_logits,
             is_sequence_parallel,
-            extra_tensors=extra_tensors,
+            extra_tensors,  # type: ignore[call-arg]
         )
 
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:
-        """
-        Combine the hidden states and router logits from the appropriate device.
-        This is a no-op in the base class.
-        """
         assert self.all2all_manager is not None
-        return self.all2all_manager.combine(
-            hidden_states,
-            is_sequence_parallel,
+        hidden_states = self.all2all_manager.combine(
+            hidden_states, is_sequence_parallel
         )
+        return hidden_states
 
 
 class _CPUSHMDistributed:

vllm/distributed/device_communicators/cuda_communicator.py

@@ -322,7 +322,7 @@ class CudaCommunicator(DeviceCommunicatorBase):
 
         return output_list
 
-    def dispatch_router_logits(
+    def dispatch(  # type: ignore[override]
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
@@ -332,52 +332,19 @@ class CudaCommunicator(DeviceCommunicatorBase):
         tuple[torch.Tensor, torch.Tensor]
         | tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]
     ):
-        """
-        Dispatch the hidden states and router logits to the appropriate device.
-        This is a no-op in the base class.
-        """
-
-        assert self.all2all_manager is not None
-        return self.all2all_manager.dispatch_router_logits(
-            hidden_states,
-            router_logits,
-            is_sequence_parallel,
-            extra_tensors,
-        )
-
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        """
-        Dispatch the hidden states and topk weights/ids to the appropriate device.
-        This is a no-op in the base class.
-        """
         assert self.all2all_manager is not None
         return self.all2all_manager.dispatch(
             hidden_states,
-            topk_weights,
-            topk_ids,
+            router_logits,
             is_sequence_parallel,
-            extra_tensors=extra_tensors,
+            extra_tensors,  # type: ignore[call-arg]
         )
 
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:
-        """
-        Combine the hidden states and router logits from the appropriate device.
-        This is a no-op in the base class.
-        """
         assert self.all2all_manager is not None
-        return self.all2all_manager.combine(
-            hidden_states,
-            is_sequence_parallel,
+        hidden_states = self.all2all_manager.combine(
+            hidden_states, is_sequence_parallel
         )
+        return hidden_states

vllm/distributed/device_communicators/mnnvl_compat.py

@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from typing import Any
-
 import torch.distributed as dist
 from flashinfer.comm.mnnvl import CommBackend as CommBackend
 
@@ -25,14 +23,5 @@ class CustomCommunicator(CommBackend):
         dist.all_gather_object(gathered, data, group=self._group)
         return gathered
 
-    # NOTE(rob): CommBackend is an abstract class, and bcast/barrier
-    # are unimplemented on vLLM side. If we need to utilize these
-    # methods in the future, can create a concrete implementation.
-    def bcast(self, data: Any, root: int) -> Any:
-        raise NotImplementedError
-
-    def barrier(self) -> None:
-        raise NotImplementedError
-
     def Split(self, color: int, key: int) -> "CustomCommunicator":
         return self

vllm/distributed/device_communicators/xpu_communicator.py

@@ -196,62 +196,26 @@ class XpuCommunicator(DeviceCommunicatorBase):
     def broadcast(self, input_: torch.Tensor, src: int = 0) -> None:
         dist.broadcast(input_, src=src, group=self.device_group)
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
         is_sequence_parallel: bool = False,
         extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        """
-        Dispatch the hidden states and router logits to the appropriate device.
-        This is a no-op in the base class.
-        """
-
-        assert self.all2all_manager is not None
-        return self.all2all_manager.dispatch_router_logits(
-            hidden_states,
-            router_logits,
-            is_sequence_parallel,
-            extra_tensors,
-        )
-
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-    ):
-        """
-        Dispatch the hidden states and topk weights/ids to the appropriate device.
-        This is a no-op in the base class.
-        """
+    ) -> tuple[torch.Tensor, torch.Tensor]:
         assert self.all2all_manager is not None
         return self.all2all_manager.dispatch(
             hidden_states,
-            topk_weights,
-            topk_ids,
+            router_logits,
             is_sequence_parallel,
-            extra_tensors=extra_tensors,
+            extra_tensors,  # type: ignore[call-arg]
         )
 
     def combine(
         self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
     ) -> torch.Tensor:
-        """
-        Combine the hidden states and router logits from the appropriate device.
-        This is a no-op in the base class.
-        """
         assert self.all2all_manager is not None
-        return self.all2all_manager.combine(
-            hidden_states,
-            is_sequence_parallel,
+        hidden_states = self.all2all_manager.combine(
+            hidden_states, is_sequence_parallel
         )
+        return hidden_states

vllm/distributed/kv_transfer/kv_connector/utils.py

@@ -322,7 +322,7 @@ class TpKVTopology:
         # Figure out whether the first dimension of the cache is K/V
         # or num_blocks. This is used to register the memory regions correctly.
         kv_cache_shape = self.attn_backend.get_kv_cache_shape(
-            num_blocks=1, block_size=16, num_kv_heads=4, head_size=1
+            num_blocks=1, block_size=16, num_kv_heads=1, head_size=1
         )
         # Non-MLA backends caches have 5 dims [2, num_blocks, H,N,D],
         # we just mock num_blocks to 1 for the dimension check below.

vllm/distributed/kv_transfer/kv_connector/v1/metrics.py

@@ -124,7 +124,7 @@ class KVConnectorPromMetrics:
         self._counter_cls = metric_types[Counter]
         self._histogram_cls = metric_types[Histogram]
         self._labelnames = labelnames
-        self.per_engine_labelvalues = per_engine_labelvalues
+        self._per_engine_labelvalues = per_engine_labelvalues
 
     def make_per_engine(self, metric: PromMetric) -> dict[int, PromMetric]:
         """
@@ -134,7 +134,7 @@ class KVConnectorPromMetrics:
         """
         return {
             idx: metric.labels(*labelvalues)
-            for idx, labelvalues in self.per_engine_labelvalues.items()
+            for idx, labelvalues in self._per_engine_labelvalues.items()
         }
 
     def observe(self, transfer_stats_data: dict[str, Any], engine_idx: int = 0):

vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

@@ -302,7 +302,7 @@ class NixlConnector(KVConnectorBase_V1):
     @property
     def prefer_cross_layer_blocks(self) -> bool:
         backend = get_current_attn_backend(self._vllm_config)
-        if backend.get_name() not in (
+        if backend().get_name() not in (
             "FLASH_ATTN",
             "FLASHINFER",
         ):

vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py

@@ -17,12 +17,6 @@ from vllm.distributed.kv_transfer.kv_connector.v1 import (
     KVConnectorRole,
 )
 from vllm.distributed.kv_transfer.kv_connector.v1.base import KVConnectorMetadata
-from vllm.distributed.kv_transfer.kv_connector.v1.metrics import (
-    KVConnectorPromMetrics,
-    KVConnectorStats,
-    PromMetric,
-    PromMetricT,
-)
 from vllm.forward_context import ForwardContext
 from vllm.logger import init_logger
 from vllm.v1.attention.backend import AttentionBackend, AttentionMetadata
@@ -34,11 +28,7 @@ from vllm.v1.kv_offload.abstract import OffloadingManager
 from vllm.v1.kv_offload.factory import OffloadingSpecFactory
 from vllm.v1.kv_offload.mediums import GPULoadStoreSpec
 from vllm.v1.kv_offload.spec import OffloadingSpec
-from vllm.v1.kv_offload.worker.worker import (
-    OffloadingWorker,
-    TransferSpec,
-    TransferType,
-)
+from vllm.v1.kv_offload.worker.worker import OffloadingWorker, TransferSpec
 from vllm.v1.outputs import KVConnectorOutput
 from vllm.v1.request import Request
 
@@ -47,66 +37,6 @@ ReqId = str
 logger = init_logger(__name__)
 
 
-@dataclass
-class OffloadingOperationMetrics:
-    op_size: int
-    op_time: float
-
-
-@dataclass
-class OffloadingConnectorStats(KVConnectorStats):
-    def __post_init__(self):
-        if not self.data:
-            # Empty container init, no data is passed in.
-            self.reset()
-
-    def reset(self):
-        self.data: dict[str, list[OffloadingOperationMetrics]] = {}
-
-    def aggregate(self, other: KVConnectorStats) -> KVConnectorStats:
-        if not other.is_empty():
-            for k, v in other.data.items():
-                if k not in self.data:
-                    self.data[k] = v
-                else:
-                    accumulator = self.data[k]
-                    assert isinstance(accumulator, list)
-                    accumulator.extend(v)
-        return self
-
-    def reduce(self) -> dict[str, int | float]:
-        """
-        Reduce the observations collected during a time interval to one or
-        more representative values (eg avg/median/sum of the series).
-        This is meant to be called by the logger to produce a summary of the
-        stats for the last time interval.
-        """
-        return_dict: dict[str, int | float] = {}
-        for transfer_type, ops_list in self.data.items():
-            assert isinstance(ops_list, list)
-            total_bytes = 0
-            total_time = 0
-            for op in ops_list:
-                assert isinstance(op, dict)
-                total_bytes += op["op_size"]
-                total_time += op["op_time"]
-            return_dict[f"{transfer_type}_total_bytes"] = total_bytes
-            return_dict[f"{transfer_type}_total_time"] = total_time
-        return return_dict
-
-    def is_empty(self) -> bool:
-        return not self.data
-
-    def record_transfer(self, num_bytes: int, time: float, transfer_type: TransferType):
-        src, dst = transfer_type
-        transfer_type_key = src + "_to_" + dst
-        op = OffloadingOperationMetrics(num_bytes, time)
-        if transfer_type_key in self.data:
-            self.data[transfer_type_key].append(op)
-        else:
-            self.data[transfer_type_key] = [op]
-
-
 @dataclass
 class OffloadingConnectorMetadata(KVConnectorMetadata):
     reqs_to_load: dict[ReqId, TransferSpec]
@@ -213,33 +143,6 @@ class OffloadingConnector(KVConnectorBase_V1):
         assert self.connector_scheduler is not None
         return self.connector_scheduler.take_events()
 
-    def get_kv_connector_stats(self) -> KVConnectorStats | None:
-        if self.connector_worker is None:
-            return None  # We only emit stats from the worker-side
-        return self.connector_worker.get_kv_connector_stats()
-
-    @classmethod
-    def build_kv_connector_stats(
-        cls, data: dict[str, Any] | None = None
-    ) -> KVConnectorStats | None:
-        return (
-            OffloadingConnectorStats(data=data)
-            if data is not None
-            else OffloadingConnectorStats()
-        )
-
-    @classmethod
-    def build_prom_metrics(
-        cls,
-        vllm_config: VllmConfig,
-        metric_types: dict[type[PromMetric], type[PromMetricT]],
-        labelnames: list[str],
-        per_engine_labelvalues: dict[int, list[object]],
-    ) -> KVConnectorPromMetrics:
-        return OffloadPromMetrics(
-            vllm_config, metric_types, labelnames, per_engine_labelvalues
-        )
-
 
 class OffloadingConnectorScheduler:
     """Implementation of Scheduler side methods"""
@@ -564,8 +467,7 @@ class OffloadingConnectorWorker:
 
         self._job_counter = 0
 
-        self.kv_connector_stats = OffloadingConnectorStats()
-        # req_id -> (job_id, store)
+        # job_id -> (req_id, store)
         self._jobs: dict[int, tuple[ReqId, bool]] = {}
         # req_id -> active job IDs
         self._load_job: dict[ReqId, int] = {}
@@ -657,21 +559,10 @@ class OffloadingConnectorWorker:
         """
         finished_sending = set()
         finished_recving = set()
-        for transfer_result in self.worker.get_finished():
+        for job_id, success in self.worker.get_finished():
             # we currently do not support job failures
-            job_id = transfer_result.job_id
-            assert transfer_result.success
+            assert success
             req_id, store = self._jobs.pop(job_id)
-            if (
-                transfer_result.transfer_time
-                and transfer_result.transfer_size is not None
-                and transfer_result.transfer_type is not None
-            ):
-                self.kv_connector_stats.record_transfer(
-                    num_bytes=transfer_result.transfer_size,
-                    time=transfer_result.transfer_time,
-                    transfer_type=transfer_result.transfer_type,
-                )
             if store:
                 req_jobs = self._store_jobs[req_id]
                 req_jobs.remove(job_id)
@@ -697,104 +588,3 @@ class OffloadingConnectorWorker:
                 del self._store_jobs[req_id]
 
         return finished_sending, finished_recving
-
-    def get_kv_connector_stats(self) -> KVConnectorStats | None:
-        """
-        Get the KV transfer stats for the connector.
-        """
-
-        if self.kv_connector_stats.is_empty():
-            return None
-        # Clear stats for next iteration
-        kv_connector_stats = self.kv_connector_stats
-        self.kv_connector_stats = OffloadingConnectorStats()
-        return kv_connector_stats
-
-
-class OffloadPromMetrics(KVConnectorPromMetrics):
-    def __init__(
-        self,
-        vllm_config: VllmConfig,
-        metric_types: dict[type[PromMetric], type[PromMetricT]],
-        labelnames: list[str],
-        per_engine_labelvalues: dict[int, list[object]],
-    ):
-        super().__init__(vllm_config, metric_types, labelnames, per_engine_labelvalues)
-        # (engine_idx, transfer_tupe) -> (metric with bounded labels)
-        self.histogram_transfer_size: dict[tuple[int, str], PromMetricT] = {}
-        self.counter_kv_bytes: dict[tuple[int, str], PromMetricT] = {}
-        self.counter_kv_transfer_time: dict[tuple[int, str], PromMetricT] = {}
-        buckets = [  # In bytes
-            1e6,
-            5e6,
-            10e6,
-            20e6,
-            40e6,
-            60e6,
-            80e6,
-            100e6,
-            150e6,
-            200e6,
-        ]
-
-        self._counter_kv_bytes = self._counter_cls(
-            name="vllm:kv_offload_total_bytes",
-            documentation="Number of bytes offloaded by KV connector",
-            labelnames=labelnames + ["transfer_type"],
-        )
-
-        self._counter_kv_transfer_time = self._counter_cls(
-            name="vllm:kv_offload_total_time",
-            documentation="Total time measured by all KV offloading operations",
-            labelnames=labelnames + ["transfer_type"],
-        )
-
-        self._histogram_transfer_size = self._histogram_cls(
-            name="vllm:kv_offload_size",
-            documentation="Histogram of KV offload transfer size, in bytes.",
-            buckets=buckets[:],
-            labelnames=labelnames + ["transfer_type"],
-        )
-
-    def observe(self, transfer_stats_data: dict[str, Any], engine_idx: int = 0):
-        """
-        Observe transfer statistics from the new data structure.
-        transfer_stats_data is expected to be a dict where:
-        - keys are transfer type strings (e.g., "cpu_to_gpu", "gpu_to_cpu")
-        - values are lists of OffloadingOperationMetrics objects
-        """
-
-        for transfer_type, ops in transfer_stats_data.items():
-            # Cache:
-            if (engine_idx, transfer_type) not in self.histogram_transfer_size:
-                self.histogram_transfer_size[(engine_idx, transfer_type)] = (
-                    self._histogram_transfer_size.labels(
-                        *(self.per_engine_labelvalues[engine_idx] + [transfer_type])
-                    )
-                )
-                self.counter_kv_bytes[(engine_idx, transfer_type)] = (
-                    self._counter_kv_bytes.labels(
-                        *(self.per_engine_labelvalues[engine_idx] + [transfer_type])
-                    )
-                )
-                self.counter_kv_transfer_time[(engine_idx, transfer_type)] = (
-                    self._counter_kv_transfer_time.labels(
-                        *(self.per_engine_labelvalues[engine_idx] + [transfer_type])
-                    )
-                )
-
-            # Process ops:
-            assert isinstance(ops, list)
-            for op in ops:  # ops is a list of serialized OffloadingOperationMetrics
-                assert isinstance(op, dict)
-                # Observe size histogram
-                self.histogram_transfer_size[(engine_idx, transfer_type)].observe(
-                    op["op_size"]
-                )
-
-                # Increment byte and time counters
-                self.counter_kv_bytes[(engine_idx, transfer_type)].inc(op["op_size"])
-
-                self.counter_kv_transfer_time[(engine_idx, transfer_type)].inc(
-                    op["op_time"]
-                )

vllm/distributed/parallel_state.py

@@ -1000,7 +1000,7 @@ class GroupCoordinator:
         if self.device_communicator is not None:
             self.device_communicator.prepare_communication_buffer_for_model(model)
 
-    def dispatch_router_logits(
+    def dispatch(
         self,
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
@@ -1011,7 +1011,7 @@ class GroupCoordinator:
         | tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]
     ):
         if self.device_communicator is not None:
-            return self.device_communicator.dispatch_router_logits(
+            return self.device_communicator.dispatch(  # type: ignore[call-arg]
                 hidden_states,
                 router_logits,
                 is_sequence_parallel,
@@ -1020,28 +1020,6 @@ class GroupCoordinator:
         else:
             return hidden_states, router_logits
 
-    def dispatch(
-        self,
-        hidden_states: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        is_sequence_parallel: bool = False,
-        extra_tensors: list[torch.Tensor] | None = None,
-    ) -> (
-        tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
-        | tuple[torch.Tensor, torch.Tensor, torch.Tensor]
-    ):
-        if self.device_communicator is not None:
-            return self.device_communicator.dispatch(
-                hidden_states,
-                topk_weights,
-                topk_ids,
-                is_sequence_parallel,
-                extra_tensors,
-            )
-        else:
-            return hidden_states, topk_weights, topk_ids
-
     def combine(
         self, hidden_states, is_sequence_parallel: bool = False
     ) -> torch.Tensor:

vllm/entrypoints/openai/api_server.py

@@ -33,10 +33,14 @@ from starlette.types import ASGIApp, Message, Receive, Scope, Send
 import vllm.envs as envs
 from vllm.engine.arg_utils import AsyncEngineArgs
 from vllm.engine.protocol import EngineClient
+from vllm.entrypoints.anthropic.serving import AnthropicServingMessages
 from vllm.entrypoints.chat_utils import load_chat_template
 from vllm.entrypoints.launcher import serve_http
 from vllm.entrypoints.logger import RequestLogger
+from vllm.entrypoints.mcp.tool_server import DemoToolServer, MCPToolServer, ToolServer
+from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat
 from vllm.entrypoints.openai.cli_args import make_arg_parser, validate_parsed_serve_args
+from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion
 from vllm.entrypoints.openai.engine.protocol import (
     ErrorInfo,
     ErrorResponse,
@@ -46,6 +50,12 @@ from vllm.entrypoints.openai.models.protocol import BaseModelPath
 from vllm.entrypoints.openai.models.serving import (
     OpenAIServingModels,
 )
+from vllm.entrypoints.openai.responses.serving import OpenAIServingResponses
+from vllm.entrypoints.openai.translations.serving import (
+    OpenAIServingTranscription,
+    OpenAIServingTranslation,
+)
+from vllm.entrypoints.serve.disagg.serving import ServingTokens
 from vllm.entrypoints.serve.elastic_ep.middleware import (
     ScalingMiddleware,
 )
@@ -60,7 +70,6 @@ from vllm.entrypoints.utils import (
 from vllm.exceptions import VLLMValidationError
 from vllm.logger import init_logger
 from vllm.reasoning import ReasoningParserManager
-from vllm.tasks import POOLING_TASKS, SupportedTask
 from vllm.tool_parsers import ToolParserManager
 from vllm.usage.usage_lib import UsageContext
 from vllm.utils.argparse_utils import FlexibleArgumentParser
@@ -255,39 +264,6 @@ def load_log_config(log_config_file: str | None) -> dict | None:
         return None
 
 
-def get_uvicorn_log_config(args: Namespace) -> dict | None:
-    """
-    Get the uvicorn log config based on the provided arguments.
-
-    Priority:
-    1. If log_config_file is specified, use it
-    2. If disable_access_log_for_endpoints is specified, create a config with
-       the access log filter
-    3. Otherwise, return None (use uvicorn defaults)
-    """
-    # First, try to load from file if specified
-    log_config = load_log_config(args.log_config_file)
-    if log_config is not None:
-        return log_config
-
-    # If endpoints to filter are specified, create a config with the filter
-    if args.disable_access_log_for_endpoints:
-        from vllm.logging_utils import create_uvicorn_log_config
-
-        # Parse comma-separated string into list
-        excluded_paths = [
-            p.strip()
-            for p in args.disable_access_log_for_endpoints.split(",")
-            if p.strip()
-        ]
-        return create_uvicorn_log_config(
-            excluded_paths=excluded_paths,
-            log_level=args.uvicorn_log_level,
-        )
-
-    return None
-
-
 class AuthenticationMiddleware:
     """
     Pure ASGI middleware that authenticates each request by checking
@@ -504,7 +480,7 @@ def _log_non_streaming_response(response_body: list) -> None:
         logger.info("response_body={<binary_data>}")
 
 
-def build_app(args: Namespace, supported_tasks: tuple["SupportedTask", ...]) -> FastAPI:
+def build_app(args: Namespace) -> FastAPI:
     if args.disable_fastapi_docs:
         app = FastAPI(
             openapi_url=None, docs_url=None, redoc_url=None, lifespan=lifespan
@@ -514,44 +490,52 @@ def build_app(args: Namespace, supported_tasks: tuple["SupportedTask", ...]) ->
     else:
         app = FastAPI(lifespan=lifespan)
     app.state.args = args
-    app.include_router(router)
-
     from vllm.entrypoints.serve import register_vllm_serve_api_routers
 
     register_vllm_serve_api_routers(app)
+    from vllm.entrypoints.openai.chat_completion.api_router import (
+        attach_router as register_chat_api_router,
+    )
 
+    register_chat_api_router(app)
+
+    from vllm.entrypoints.openai.responses.api_router import (
+        attach_router as register_responses_api_router,
+    )
+
+    register_responses_api_router(app)
+    from vllm.entrypoints.openai.translations.api_router import (
+        attach_router as register_translations_api_router,
+    )
+
+    register_translations_api_router(app)
+
+    from vllm.entrypoints.openai.completion.api_router import (
+        attach_router as register_completion_api_router,
+    )
+
+    register_completion_api_router(app)
+    from vllm.entrypoints.anthropic.api_router import (
+        attach_router as register_anthropic_api_router,
+    )
+
+    register_anthropic_api_router(app)
     from vllm.entrypoints.openai.models.api_router import (
         attach_router as register_models_api_router,
     )
 
     register_models_api_router(app)
+    from vllm.entrypoints.sagemaker.routes import register_sagemaker_routes
 
-    from vllm.entrypoints.sagemaker.api_router import (
-        attach_router as register_sagemaker_api_router,
-    )
-
-    register_sagemaker_api_router(app, supported_tasks)
-
-    if "generate" in supported_tasks:
-        from vllm.entrypoints.openai.generate.api_router import (
-            register_generate_api_routers,
-        )
-
-        register_generate_api_routers(app)
-
-    if "transcription" in supported_tasks:
-        from vllm.entrypoints.openai.translations.api_router import (
-            attach_router as register_translations_api_router,
-        )
-
-        register_translations_api_router(app)
-
-    if any(task in POOLING_TASKS for task in supported_tasks):
-        from vllm.entrypoints.pooling import register_pooling_api_routers
-
-        register_pooling_api_routers(app, supported_tasks)
+    register_sagemaker_routes(router)
+    app.include_router(router)
 
     app.root_path = args.root_path
+
+    from vllm.entrypoints.pooling import register_pooling_api_routers
+
+    register_pooling_api_routers(app)
+
     app.add_middleware(
         CORSMiddleware,
         allow_origins=args.allowed_origins,
@@ -656,7 +640,6 @@ async def init_app_state(
     engine_client: EngineClient,
     state: State,
     args: Namespace,
-    supported_tasks: tuple["SupportedTask", ...],
 ) -> None:
     vllm_config = engine_client.vllm_config
 
@@ -678,9 +661,28 @@ async def init_app_state(
     state.log_stats = not args.disable_log_stats
     state.vllm_config = vllm_config
     state.args = args
+    supported_tasks = await engine_client.get_supported_tasks()
+    logger.info("Supported tasks: %s", supported_tasks)
+
     resolved_chat_template = load_chat_template(args.chat_template)
 
+    if args.tool_server == "demo":
+        tool_server: ToolServer | None = DemoToolServer()
+        assert isinstance(tool_server, DemoToolServer)
+        await tool_server.init_and_validate()
+    elif args.tool_server:
+        tool_server = MCPToolServer()
+        await tool_server.add_tool_server(args.tool_server)
+    else:
+        tool_server = None
+
     # Merge default_mm_loras into the static lora_modules
+    default_mm_loras = (
+        vllm_config.lora_config.default_mm_loras
+        if vllm_config.lora_config is not None
+        else {}
+    )
+
     default_mm_loras = (
         vllm_config.lora_config.default_mm_loras
         if vllm_config.lora_config is not None
@@ -694,6 +696,66 @@ async def init_app_state(
         lora_modules=lora_modules,
     )
     await state.openai_serving_models.init_static_loras()
+    state.openai_serving_responses = (
+        OpenAIServingResponses(
+            engine_client,
+            state.openai_serving_models,
+            request_logger=request_logger,
+            chat_template=resolved_chat_template,
+            chat_template_content_format=args.chat_template_content_format,
+            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
+            enable_auto_tools=args.enable_auto_tool_choice,
+            tool_parser=args.tool_call_parser,
+            tool_server=tool_server,
+            reasoning_parser=args.structured_outputs_config.reasoning_parser,
+            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
+            enable_force_include_usage=args.enable_force_include_usage,
+            enable_log_outputs=args.enable_log_outputs,
+            log_error_stack=args.log_error_stack,
+        )
+        if "generate" in supported_tasks
+        else None
+    )
+    state.openai_serving_chat = (
+        OpenAIServingChat(
+            engine_client,
+            state.openai_serving_models,
+            args.response_role,
+            request_logger=request_logger,
+            chat_template=resolved_chat_template,
+            chat_template_content_format=args.chat_template_content_format,
+            default_chat_template_kwargs=args.default_chat_template_kwargs,
+            trust_request_chat_template=args.trust_request_chat_template,
+            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
+            enable_auto_tools=args.enable_auto_tool_choice,
+            exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
+            tool_parser=args.tool_call_parser,
+            reasoning_parser=args.structured_outputs_config.reasoning_parser,
+            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
+            enable_force_include_usage=args.enable_force_include_usage,
+            enable_log_outputs=args.enable_log_outputs,
+            enable_log_deltas=args.enable_log_deltas,
+            log_error_stack=args.log_error_stack,
+        )
+        if "generate" in supported_tasks
+        else None
+    )
+    # Warm up chat template processing to avoid first-request latency
+    if state.openai_serving_chat is not None:
+        await state.openai_serving_chat.warmup()
+    state.openai_serving_completion = (
+        OpenAIServingCompletion(
+            engine_client,
+            state.openai_serving_models,
+            request_logger=request_logger,
+            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
+            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
+            enable_force_include_usage=args.enable_force_include_usage,
+            log_error_stack=args.log_error_stack,
+        )
+        if "generate" in supported_tasks
+        else None
+    )
     state.openai_serving_tokenization = OpenAIServingTokenization(
         engine_client,
         state.openai_serving_models,
@@ -703,27 +765,64 @@ async def init_app_state(
         trust_request_chat_template=args.trust_request_chat_template,
         log_error_stack=args.log_error_stack,
     )
-
-    if "generate" in supported_tasks:
-        from vllm.entrypoints.openai.generate.api_router import init_generate_state
-
-        await init_generate_state(
-            engine_client, state, args, request_logger, supported_tasks
+    state.openai_serving_transcription = (
+        OpenAIServingTranscription(
+            engine_client,
+            state.openai_serving_models,
+            request_logger=request_logger,
+            log_error_stack=args.log_error_stack,
+            enable_force_include_usage=args.enable_force_include_usage,
         )
-
-    if "transcription" in supported_tasks:
-        from vllm.entrypoints.openai.translations.api_router import (
-            init_transcription_state,
+        if "transcription" in supported_tasks
+        else None
+    )
+    state.openai_serving_translation = (
+        OpenAIServingTranslation(
+            engine_client,
+            state.openai_serving_models,
+            request_logger=request_logger,
+            log_error_stack=args.log_error_stack,
+            enable_force_include_usage=args.enable_force_include_usage,
         )
-
-        init_transcription_state(
-            engine_client, state, args, request_logger, supported_tasks
+        if "transcription" in supported_tasks
+        else None
+    )
+    state.anthropic_serving_messages = (
+        AnthropicServingMessages(
+            engine_client,
+            state.openai_serving_models,
+            args.response_role,
+            request_logger=request_logger,
+            chat_template=resolved_chat_template,
+            chat_template_content_format=args.chat_template_content_format,
+            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
+            enable_auto_tools=args.enable_auto_tool_choice,
+            tool_parser=args.tool_call_parser,
+            reasoning_parser=args.structured_outputs_config.reasoning_parser,
+            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
+            enable_force_include_usage=args.enable_force_include_usage,
         )
+        if "generate" in supported_tasks
+        else None
+    )
+    state.serving_tokens = (
+        ServingTokens(
+            engine_client,
+            state.openai_serving_models,
+            request_logger=request_logger,
+            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
+            log_error_stack=args.log_error_stack,
+            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
+            enable_log_outputs=args.enable_log_outputs,
+            force_no_detokenize=args.tokens_only,
+        )
+        if "generate" in supported_tasks
+        else None
+    )
 
-    if any(task in POOLING_TASKS for task in supported_tasks):
-        from vllm.entrypoints.pooling import init_pooling_state
+    from vllm.entrypoints.pooling import init_pooling_state
 
-        init_pooling_state(engine_client, state, args, request_logger, supported_tasks)
+    await init_pooling_state(engine_client, state, args)
 
     state.enable_server_load_tracking = args.enable_server_load_tracking
     state.server_load_metrics = 0
@@ -831,8 +930,8 @@ async def run_server_worker(
     if args.reasoning_parser_plugin and len(args.reasoning_parser_plugin) > 3:
         ReasoningParserManager.import_reasoning_parser(args.reasoning_parser_plugin)
 
-    # Get uvicorn log config (from file or with endpoint filter)
-    log_config = get_uvicorn_log_config(args)
+    # Load logging config for uvicorn if specified
+    log_config = load_log_config(args.log_config_file)
     if log_config is not None:
         uvicorn_kwargs["log_config"] = log_config
 
@@ -840,11 +939,9 @@ async def run_server_worker(
         args,
         client_config=client_config,
     ) as engine_client:
-        supported_tasks = await engine_client.get_supported_tasks()
-        logger.info("Supported tasks: %s", supported_tasks)
+        app = build_app(args)
 
-        app = build_app(args, supported_tasks)
-        await init_app_state(engine_client, app.state, args, supported_tasks)
+        await init_app_state(engine_client, app.state, args)
 
         logger.info(
             "Starting vLLM API server %d on %s",

vllm/entrypoints/openai/chat_completion/serving.py

@@ -44,7 +44,6 @@ from vllm.entrypoints.openai.engine.protocol import (
     DeltaMessage,
     DeltaToolCall,
     ErrorResponse,
-    FunctionCall,
     PromptTokenUsageInfo,
     RequestResponseMetadata,
     ToolCall,
@@ -68,7 +67,6 @@ from vllm.entrypoints.openai.parser.harmony_utils import (
 from vllm.entrypoints.openai.utils import maybe_filter_parallel_tool_calls
 from vllm.entrypoints.utils import get_max_tokens, should_include_usage
 from vllm.inputs.data import TokensPrompt
-from vllm.inputs.parse import get_prompt_components
 from vllm.logger import init_logger
 from vllm.logprobs import Logprob
 from vllm.outputs import CompletionOutput, RequestOutput
@@ -145,6 +143,11 @@ class OpenAIServingChat(OpenAIServing):
         self.enable_prompt_tokens_details = enable_prompt_tokens_details
         self.enable_force_include_usage = enable_force_include_usage
         self.default_sampling_params = self.model_config.get_diff_sampling_param()
+        if self.model_config.hf_config.model_type == "kimi_k2":
+            self.tool_call_id_type = "kimi_k2"
+        else:
+            self.tool_call_id_type = "random"
+
         self.use_harmony = self.model_config.hf_config.model_type == "gpt_oss"
         if self.use_harmony:
             if "stop_token_ids" not in self.default_sampling_params:
@@ -153,16 +156,6 @@ class OpenAIServingChat(OpenAIServing):
                 get_stop_tokens_for_assistant_actions()
             )
 
-        # Handle tool call ID type for Kimi K2 (supporting test mocking via overrides)
-        hf_overrides = getattr(self.model_config, "hf_overrides", None)
-        if self.model_config.hf_text_config.model_type == "kimi_k2" or (
-            isinstance(hf_overrides, dict)
-            and hf_overrides.get("model_type") == "kimi_k2"
-        ):
-            self.tool_call_id_type = "kimi_k2"
-        else:
-            self.tool_call_id_type = "random"
-
         # NOTE(woosuk): While OpenAI's chat completion API supports browsing
         # for some models, currently vLLM doesn't support it. Please use the
         # Responses API instead.
@@ -254,8 +247,8 @@ class OpenAIServingChat(OpenAIServing):
                 # because of issues with pydantic we need to potentially
                 # re-serialize the tool_calls field of the request
                 # for more info: see comment in `maybe_serialize_tool_calls`
-                maybe_serialize_tool_calls(request)  # type: ignore[arg-type]
-                truncate_tool_call_ids(request)  # type: ignore[arg-type]
+                maybe_serialize_tool_calls(request)
+                truncate_tool_call_ids(request)
                 validate_request_params(request)
 
             # Check if tool parsing is unavailable (common condition)
@@ -375,18 +368,20 @@ class OpenAIServingChat(OpenAIServing):
         generators: list[AsyncGenerator[RequestOutput, None]] = []
         try:
             for i, engine_prompt in enumerate(engine_prompts):
-                prompt_text, _, _ = get_prompt_components(engine_prompt)
-
+                prompt_text, _, _ = self._get_prompt_components(engine_prompt)
                 # If we are creating sub requests for multiple prompts, ensure that they
                 # have unique request ids.
                 sub_request_id = (
                     request_id if len(engine_prompts) == 1 else f"{request_id}_{i}"
                 )
 
+                if self.default_sampling_params is None:
+                    self.default_sampling_params = {}
+
                 max_tokens = get_max_tokens(
                     max_model_len=self.max_model_len,
                     request=request,
-                    prompt=engine_prompt,
+                    input_length=len(engine_prompt["prompt_token_ids"]),
                     default_sampling_params=self.default_sampling_params,
                 )
 
@@ -459,7 +454,6 @@ class OpenAIServingChat(OpenAIServing):
 
         # Streaming response
         tokenizer = self.renderer.tokenizer
-        assert tokenizer is not None
 
         if request.stream:
             return self.chat_completion_stream_generator(
@@ -638,11 +632,9 @@ class OpenAIServingChat(OpenAIServing):
         request_id: str,
         model_name: str,
         conversation: list[ConversationMessage],
-        tokenizer: TokenizerLike,
+        tokenizer: TokenizerLike | None,
         request_metadata: RequestResponseMetadata,
     ) -> AsyncGenerator[str, None]:
-        from vllm.tokenizers.mistral import MistralTokenizer
-
         created_time = int(time.time())
         chunk_object_type: Final = "chat.completion.chunk"
         first_iteration = True
@@ -706,7 +698,7 @@ class OpenAIServingChat(OpenAIServing):
                 )
                 reasoning_parser = self.reasoning_parser(
                     tokenizer,
-                    chat_template_kwargs=chat_template_kwargs or {},  # type: ignore[call-arg]
+                    chat_template_kwargs=chat_template_kwargs,  # type: ignore[call-arg]
                 )
         except RuntimeError as e:
             logger.exception("Error in reasoning parser creation.")
@@ -963,17 +955,8 @@ class OpenAIServingChat(OpenAIServing):
                                     index=i,
                                 )
                             else:
-                                # Generate ID based on tokenizer type
-                                if isinstance(tokenizer, MistralTokenizer):
-                                    tool_call_id = MistralToolCall.generate_random_id()
-                                else:
-                                    tool_call_id = make_tool_call_id(
-                                        id_type=self.tool_call_id_type,
-                                        func_name=tool_choice_function_name,
-                                        idx=history_tool_call_cnt,
-                                    )
                                 delta_tool_call = DeltaToolCall(
-                                    id=tool_call_id,
+                                    id=make_tool_call_id(),
                                     type="function",
                                     function=DeltaFunctionCall(
                                         name=tool_choice_function_name,
@@ -1404,11 +1387,9 @@ class OpenAIServingChat(OpenAIServing):
         request_id: str,
         model_name: str,
         conversation: list[ConversationMessage],
-        tokenizer: TokenizerLike,
+        tokenizer: TokenizerLike | None,
         request_metadata: RequestResponseMetadata,
     ) -> ErrorResponse | ChatCompletionResponse:
-        from vllm.tokenizers.mistral import MistralTokenizer
-
         created_time = int(time.time())
         final_res: RequestOutput | None = None
 
@@ -1543,85 +1524,39 @@ class OpenAIServingChat(OpenAIServing):
             tool_call_class = (
                 MistralToolCall if isinstance(tokenizer, MistralTokenizer) else ToolCall
             )
-            if self.use_harmony:
-                # Harmony models already have parsed content and tool_calls
-                # through parse_chat_output. Respect its output directly.
-                message = ChatMessage(
-                    role=role,
-                    reasoning=reasoning,
-                    content=content,
-                    tool_calls=tool_calls if tool_calls else [],
-                )
-
-            elif (not self.enable_auto_tools or not self.tool_parser) and (
+            if (not self.enable_auto_tools or not self.tool_parser) and (
                 not isinstance(request.tool_choice, ChatCompletionNamedToolChoiceParam)
                 and request.tool_choice != "required"
             ):
                 message = ChatMessage(role=role, reasoning=reasoning, content=content)
 
+            # if the request uses tools and specified a tool choice
             elif (
                 request.tool_choice
                 and type(request.tool_choice) is ChatCompletionNamedToolChoiceParam
             ):
                 assert tool_calls is not None and len(tool_calls) > 0
-                tool_call_class_items = []
-                for idx, tc in enumerate(tool_calls):
-                    # Use native ID if available (e.g., Kimi K2),
-                    # otherwise generate ID with correct id_type
-                    if tc.id:
-                        tool_call_class_items.append(
-                            tool_call_class(id=tc.id, function=tc)
-                        )
-                    else:
-                        # Generate ID using the correct format (kimi_k2 or random),
-                        # but leave it to the class if it's Mistral to preserve
-                        # 9-char IDs
-                        if isinstance(tokenizer, MistralTokenizer):
-                            tool_call_class_items.append(tool_call_class(function=tc))
-                        else:
-                            generated_id = make_tool_call_id(
-                                id_type=self.tool_call_id_type,
-                                func_name=tc.name,
-                                idx=history_tool_call_cnt + idx,
-                            )
-                            tool_call_class_items.append(
-                                tool_call_class(id=generated_id, function=tc)
-                            )
-                    history_tool_call_cnt += 1
                 message = ChatMessage(
                     role=role,
                     reasoning=reasoning,
                     content="",
-                    tool_calls=tool_call_class_items,
+                    tool_calls=[tool_call_class(function=tc) for tc in tool_calls],
                 )
 
             elif request.tool_choice and request.tool_choice == "required":
                 tool_call_class_items = []
                 assert tool_calls is not None and len(tool_calls) > 0
-                for idx, tool_call in enumerate(tool_calls):
-                    # Use native ID if available,
-                    # otherwise generate ID with correct id_type
-                    if tool_call.id:
-                        tool_call_class_items.append(
-                            tool_call_class(id=tool_call.id, function=tool_call)
-                        )
-                    else:
-                        # Generate ID using the correct format (kimi_k2 or random),
-                        # but leave it to the class if it's Mistral to preserve
-                        # 9-char IDs
-                        if isinstance(tokenizer, MistralTokenizer):
-                            tool_call_class_items.append(
-                                tool_call_class(function=tool_call)
-                            )
-                        else:
-                            generated_id = make_tool_call_id(
+                for tool_call in tool_calls:
+                    tool_call_class_items.append(
+                        tool_call_class(
+                            id=make_tool_call_id(
                                 id_type=self.tool_call_id_type,
                                 func_name=tool_call.name,
-                                idx=history_tool_call_cnt + idx,
-                            )
-                            tool_call_class_items.append(
-                                tool_call_class(id=generated_id, function=tool_call)
-                            )
+                                idx=history_tool_call_cnt,
+                            ),
+                            function=tool_call,
+                        )
+                    )
                     history_tool_call_cnt += 1
                 message = ChatMessage(
                     role=role,
@@ -1647,35 +1582,17 @@ class OpenAIServingChat(OpenAIServing):
                 # call. The same is not true for named function calls
                 auto_tools_called = tool_calls is not None and len(tool_calls) > 0
                 if tool_calls:
-                    tool_call_items = []
-                    for idx, tc in enumerate(tool_calls):
-                        # Use native ID if available (e.g., Kimi K2),
-                        # otherwise generate ID with correct id_type
-                        if tc.id:
-                            tool_call_items.append(
-                                tool_call_class(id=tc.id, function=tc)
-                            )
-                        else:
-                            # Generate ID using the correct format (kimi_k2 or random),
-                            # but leave it to the class if it's Mistral to preserve
-                            # 9-char IDs
-                            if isinstance(tokenizer, MistralTokenizer):
-                                tool_call_items.append(tool_call_class(function=tc))
-                            else:
-                                generated_id = make_tool_call_id(
-                                    id_type=self.tool_call_id_type,
-                                    func_name=tc.name,
-                                    idx=history_tool_call_cnt + idx,
-                                )
-                                tool_call_items.append(
-                                    tool_call_class(id=generated_id, function=tc)
-                                )
-                        history_tool_call_cnt += 1
                     message = ChatMessage(
                         role=role,
                         reasoning=reasoning,
                         content=content,
-                        tool_calls=tool_call_items,
+                        tool_calls=[
+                            ToolCall(
+                                function=tc,
+                                type="function",
+                            )
+                            for tc in tool_calls
+                        ],
                     )
 
                 else:
@@ -1784,11 +1701,13 @@ class OpenAIServingChat(OpenAIServing):
                 elif choice.message.tool_calls:
                     # For tool calls, log the function name and arguments
                     tool_call_descriptions = []
-                    for tc in choice.message.tool_calls:  # type: ignore
-                        function_call: FunctionCall = tc.function  # type: ignore
-                        tool_call_descriptions.append(
-                            f"{function_call.name}({function_call.arguments})"
-                        )
+                    for tc in choice.message.tool_calls:
+                        if hasattr(tc.function, "name") and hasattr(
+                            tc.function, "arguments"
+                        ):
+                            tool_call_descriptions.append(
+                                f"{tc.function.name}({tc.function.arguments})"
+                            )
                     tool_calls_str = ", ".join(tool_call_descriptions)
                     output_text = f"[tool_calls: {tool_calls_str}]"
 
@@ -1976,7 +1895,7 @@ class OpenAIServingChat(OpenAIServing):
         # because of issues with pydantic we need to potentially
         # re-serialize the tool_calls field of the request
         # for more info: see comment in `maybe_serialize_tool_calls`
-        maybe_serialize_tool_calls(request)  # type: ignore[arg-type]
+        maybe_serialize_tool_calls(request)
 
         # Add system message.
         # NOTE: In Chat Completion API, browsing is enabled by default
@@ -1994,7 +1913,7 @@ class OpenAIServingChat(OpenAIServing):
         # Add developer message.
         if request.tools:
             dev_msg = get_developer_message(
-                tools=request.tools if should_include_tools else None  # type: ignore[arg-type]
+                tools=request.tools if should_include_tools else None
             )
             messages.append(dev_msg)
 

vllm/entrypoints/openai/cli_args.py

@@ -85,12 +85,6 @@ class FrontendArgs:
     """Log level for uvicorn."""
     disable_uvicorn_access_log: bool = False
     """Disable uvicorn access log."""
-    disable_access_log_for_endpoints: str | None = None
-    """Comma-separated list of endpoint paths to exclude from uvicorn access
-    logs. This is useful to reduce log noise from high-frequency endpoints
-    like health checks. Example: "/health,/metrics,/ping".
-    When set, access logs for requests to these paths will be suppressed
-    while keeping logs for other endpoints."""
     allow_credentials: bool = False
     """Allow credentials."""
     allowed_origins: list[str] = field(default_factory=lambda: ["*"])
@@ -250,11 +244,6 @@ class FrontendArgs:
             del frontend_kwargs["middleware"]["nargs"]
         frontend_kwargs["middleware"]["default"] = []
 
-        # Special case: disable_access_log_for_endpoints is a single
-        # comma-separated string, not a list
-        if "nargs" in frontend_kwargs["disable_access_log_for_endpoints"]:
-            del frontend_kwargs["disable_access_log_for_endpoints"]["nargs"]
-
         # Special case: Tool call parser shows built-in options.
         valid_tool_parsers = list(ToolParserManager.list_registered())
         parsers_str = ",".join(valid_tool_parsers)

vllm/entrypoints/openai/completion/serving.py

@@ -36,7 +36,6 @@ from vllm.entrypoints.renderer import RenderConfig
 from vllm.entrypoints.utils import get_max_tokens, should_include_usage
 from vllm.exceptions import VLLMValidationError
 from vllm.inputs.data import EmbedsPrompt, TokensPrompt, is_embeds_prompt
-from vllm.inputs.parse import get_prompt_components
 from vllm.logger import init_logger
 from vllm.logprobs import Logprob
 from vllm.outputs import RequestOutput
@@ -163,12 +162,25 @@ class OpenAIServingCompletion(OpenAIServing):
         generators: list[AsyncGenerator[RequestOutput, None]] = []
         try:
             for i, engine_prompt in enumerate(engine_prompts):
-                prompt_text, _, _ = get_prompt_components(engine_prompt)
+                prompt_text, prompt_token_ids, prompt_embeds = (
+                    self._get_prompt_components(engine_prompt)
+                )
+
+                input_length = None
+                if prompt_token_ids is not None:
+                    input_length = len(prompt_token_ids)
+                elif prompt_embeds is not None:
+                    input_length = len(prompt_embeds)
+                else:
+                    raise NotImplementedError
+
+                if self.default_sampling_params is None:
+                    self.default_sampling_params = {}
 
                 max_tokens = get_max_tokens(
                     max_model_len=self.max_model_len,
                     request=request,
-                    prompt=engine_prompt,
+                    input_length=input_length,
                     default_sampling_params=self.default_sampling_params,
                 )
 

vllm/entrypoints/openai/engine/protocol.py

@@ -218,10 +218,6 @@ def get_logits_processors(
 
 
 class FunctionCall(OpenAIBaseModel):
-    # Internal field to preserve native tool call ID from tool parser.
-    # Excluded from serialization to maintain OpenAI API compatibility
-    # (function object should only contain 'name' and 'arguments').
-    id: str | None = Field(default=None, exclude=True)
     name: str
     arguments: str
 

vllm/entrypoints/openai/engine/serving.py

@@ -64,12 +64,13 @@ from vllm.entrypoints.openai.translations.protocol import (
 from vllm.entrypoints.pooling.classify.protocol import (
     ClassificationChatRequest,
     ClassificationCompletionRequest,
+    ClassificationRequest,
     ClassificationResponse,
 )
 from vllm.entrypoints.pooling.embed.protocol import (
-    EmbeddingBytesResponse,
     EmbeddingChatRequest,
     EmbeddingCompletionRequest,
+    EmbeddingRequest,
     EmbeddingResponse,
 )
 from vllm.entrypoints.pooling.pooling.protocol import (
@@ -94,14 +95,11 @@ from vllm.entrypoints.serve.tokenize.protocol import (
     TokenizeCompletionRequest,
     TokenizeResponse,
 )
-from vllm.entrypoints.utils import (
-    _validate_truncation_size,
-    get_max_tokens,
-    sanitize_message,
-)
+from vllm.entrypoints.utils import _validate_truncation_size, sanitize_message
 from vllm.exceptions import VLLMValidationError
 from vllm.inputs.data import PromptType, TokensPrompt
 from vllm.inputs.parse import (
+    PromptComponents,
     get_prompt_components,
     is_explicit_encoder_decoder_prompt,
 )
@@ -172,7 +170,6 @@ AnyResponse: TypeAlias = (
     CompletionResponse
     | ChatCompletionResponse
     | EmbeddingResponse
-    | EmbeddingBytesResponse
     | TranscriptionResponse
     | TokenizeResponse
     | PoolingResponse
@@ -186,21 +183,51 @@ RequestT = TypeVar("RequestT", bound=AnyRequest)
 
 
 @dataclass(kw_only=True)
-class ServeContext(Generic[RequestT]):
+class RequestProcessingMixin:
+    """
+    Mixin for request processing,
+    handling prompt preparation and engine input.
+    """
+
+    engine_prompts: list[TokensPrompt] | None = field(default_factory=list)
+
+
+@dataclass(kw_only=True)
+class ResponseGenerationMixin:
+    """
+    Mixin for response generation,
+    managing result generators and final batch results.
+    """
+
+    result_generator: (
+        AsyncGenerator[tuple[int, RequestOutput | PoolingRequestOutput], None] | None
+    ) = None
+    final_res_batch: list[RequestOutput | PoolingRequestOutput] = field(
+        default_factory=list
+    )
+
+    model_config = ConfigDict(arbitrary_types_allowed=True)
+
+
+@dataclass(kw_only=True)
+class ServeContext(RequestProcessingMixin, ResponseGenerationMixin, Generic[RequestT]):
     request: RequestT
     raw_request: Request | None = None
     model_name: str
     request_id: str
     created_time: int = field(default_factory=lambda: int(time.time()))
     lora_request: LoRARequest | None = None
-    engine_prompts: list[TokensPrompt] | None = None
 
-    result_generator: AsyncGenerator[tuple[int, PoolingRequestOutput], None] | None = (
-        None
-    )
-    final_res_batch: list[PoolingRequestOutput] = field(default_factory=list)
 
-    model_config = ConfigDict(arbitrary_types_allowed=True)
+@dataclass(kw_only=True)
+class ClassificationServeContext(ServeContext[ClassificationRequest]):
+    pass
+
+
+@dataclass(kw_only=True)
+class EmbeddingServeContext(ServeContext[EmbeddingRequest]):
+    chat_template: str | None = None
+    chat_template_content_format: ChatTemplateContentFormatOption
 
 
 class OpenAIServing:
@@ -578,7 +605,10 @@ class OpenAIServing:
         self,
         ctx: ServeContext,
     ) -> AnyResponse | ErrorResponse:
-        async for response in self._pipeline(ctx):
+        generation: AsyncGenerator[AnyResponse | ErrorResponse, None]
+        generation = self._pipeline(ctx)
+
+        async for response in generation:
             return response
 
         return self.create_error_response("No response yielded from pipeline")
@@ -637,7 +667,9 @@ class OpenAIServing:
         ctx: ServeContext,
     ) -> ErrorResponse | None:
         """Schedule the request and get the result generator."""
-        generators: list[AsyncGenerator[PoolingRequestOutput, None]] = []
+        generators: list[
+            AsyncGenerator[RequestOutput | PoolingRequestOutput, None]
+        ] = []
 
         try:
             trace_headers = (
@@ -691,7 +723,7 @@ class OpenAIServing:
                 return self.create_error_response("Engine prompts not available")
 
             num_prompts = len(ctx.engine_prompts)
-            final_res_batch: list[PoolingRequestOutput | None]
+            final_res_batch: list[RequestOutput | PoolingRequestOutput | None]
             final_res_batch = [None] * num_prompts
 
             if ctx.result_generator is None:
@@ -979,7 +1011,7 @@ class OpenAIServing:
 
     def _validate_input(
         self,
-        request: object,
+        request: AnyRequest,
         input_ids: list[int],
         input_text: str,
     ) -> TokensPrompt:
@@ -1290,7 +1322,7 @@ class OpenAIServing:
         priority: int = 0,
         **kwargs,
     ):
-        prompt_text, _, _ = get_prompt_components(engine_prompt)
+        prompt_text, _, _ = self._get_prompt_components(engine_prompt)
 
         orig_priority = priority
         sub_request = 0
@@ -1341,12 +1373,10 @@ class OpenAIServing:
             # yield context
 
             # Create inputs for the next turn.
-            # Render the next prompt token ids and update sampling_params.
+            # Render the next prompt token ids.
             if isinstance(context, (HarmonyContext, StreamingHarmonyContext)):
-                token_ids = context.render_for_completion()
-                engine_prompt = TokensPrompt(prompt_token_ids=token_ids)
-
-                sampling_params.max_tokens = self.max_model_len - len(token_ids)
+                prompt_token_ids = context.render_for_completion()
+                engine_prompt = TokensPrompt(prompt_token_ids=prompt_token_ids)
             elif isinstance(context, ParsableContext):
                 engine_prompts = await self._render_next_turn(
                     context.request,
@@ -1358,19 +1388,19 @@ class OpenAIServing:
                     context.chat_template_content_format,
                 )
                 engine_prompt = engine_prompts[0]
-                prompt_text, _, _ = get_prompt_components(engine_prompt)
-
-                sampling_params.max_tokens = get_max_tokens(
-                    self.max_model_len,
-                    context.request,
-                    engine_prompt,
-                    self.default_sampling_params,  # type: ignore
-                )
+                prompt_text, _, _ = self._get_prompt_components(engine_prompt)
 
+            # Update the sampling params.
+            sampling_params.max_tokens = self.max_model_len - len(
+                engine_prompt["prompt_token_ids"]
+            )
             # OPTIMIZATION
             priority = orig_priority - 1
             sub_request += 1
 
+    def _get_prompt_components(self, prompt: PromptType) -> PromptComponents:
+        return get_prompt_components(prompt)
+
     def _log_inputs(
         self,
         request_id: str,
@@ -1381,7 +1411,7 @@ class OpenAIServing:
         if self.request_logger is None:
             return
 
-        prompt, prompt_token_ids, prompt_embeds = get_prompt_components(inputs)
+        prompt, prompt_token_ids, prompt_embeds = self._get_prompt_components(inputs)
 
         self.request_logger.log_inputs(
             request_id,
@@ -1495,7 +1525,6 @@ class OpenAIServing:
                 # extract_tool_calls() returns a list of tool calls.
                 function_calls.extend(
                     FunctionCall(
-                        id=tool_call.id,
                         name=tool_call.function.name,
                         arguments=tool_call.function.arguments,
                     )

vllm/entrypoints/openai/generate/api_router.py

@@ -1,166 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from typing import TYPE_CHECKING
-
-from fastapi import FastAPI
-
-if TYPE_CHECKING:
-    from argparse import Namespace
-
-    from starlette.datastructures import State
-
-    from vllm.engine.protocol import EngineClient
-    from vllm.entrypoints.logger import RequestLogger
-    from vllm.tasks import SupportedTask
-else:
-    RequestLogger = object
-
-
-def register_generate_api_routers(app: FastAPI):
-    from vllm.entrypoints.openai.chat_completion.api_router import (
-        attach_router as register_chat_api_router,
-    )
-
-    register_chat_api_router(app)
-
-    from vllm.entrypoints.openai.responses.api_router import (
-        attach_router as register_responses_api_router,
-    )
-
-    register_responses_api_router(app)
-
-    from vllm.entrypoints.openai.completion.api_router import (
-        attach_router as register_completion_api_router,
-    )
-
-    register_completion_api_router(app)
-
-    from vllm.entrypoints.anthropic.api_router import (
-        attach_router as register_anthropic_api_router,
-    )
-
-    register_anthropic_api_router(app)
-
-
-async def init_generate_state(
-    engine_client: "EngineClient",
-    state: "State",
-    args: "Namespace",
-    request_logger: RequestLogger | None,
-    supported_tasks: tuple["SupportedTask", ...],
-):
-    from vllm.entrypoints.anthropic.serving import AnthropicServingMessages
-    from vllm.entrypoints.chat_utils import load_chat_template
-    from vllm.entrypoints.mcp.tool_server import (
-        DemoToolServer,
-        MCPToolServer,
-        ToolServer,
-    )
-    from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat
-    from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion
-    from vllm.entrypoints.openai.responses.serving import OpenAIServingResponses
-    from vllm.entrypoints.serve.disagg.serving import ServingTokens
-
-    if args.tool_server == "demo":
-        tool_server: ToolServer | None = DemoToolServer()
-        assert isinstance(tool_server, DemoToolServer)
-        await tool_server.init_and_validate()
-    elif args.tool_server:
-        tool_server = MCPToolServer()
-        await tool_server.add_tool_server(args.tool_server)
-    else:
-        tool_server = None
-    resolved_chat_template = load_chat_template(args.chat_template)
-
-    state.openai_serving_responses = (
-        OpenAIServingResponses(
-            engine_client,
-            state.openai_serving_models,
-            request_logger=request_logger,
-            chat_template=resolved_chat_template,
-            chat_template_content_format=args.chat_template_content_format,
-            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
-            enable_auto_tools=args.enable_auto_tool_choice,
-            tool_parser=args.tool_call_parser,
-            tool_server=tool_server,
-            reasoning_parser=args.structured_outputs_config.reasoning_parser,
-            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
-            enable_force_include_usage=args.enable_force_include_usage,
-            enable_log_outputs=args.enable_log_outputs,
-            log_error_stack=args.log_error_stack,
-        )
-        if "generate" in supported_tasks
-        else None
-    )
-    state.openai_serving_chat = (
-        OpenAIServingChat(
-            engine_client,
-            state.openai_serving_models,
-            args.response_role,
-            request_logger=request_logger,
-            chat_template=resolved_chat_template,
-            chat_template_content_format=args.chat_template_content_format,
-            default_chat_template_kwargs=args.default_chat_template_kwargs,
-            trust_request_chat_template=args.trust_request_chat_template,
-            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
-            enable_auto_tools=args.enable_auto_tool_choice,
-            exclude_tools_when_tool_choice_none=args.exclude_tools_when_tool_choice_none,
-            tool_parser=args.tool_call_parser,
-            reasoning_parser=args.structured_outputs_config.reasoning_parser,
-            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
-            enable_force_include_usage=args.enable_force_include_usage,
-            enable_log_outputs=args.enable_log_outputs,
-            enable_log_deltas=args.enable_log_deltas,
-            log_error_stack=args.log_error_stack,
-        )
-        if "generate" in supported_tasks
-        else None
-    )
-    # Warm up chat template processing to avoid first-request latency
-    if state.openai_serving_chat is not None:
-        await state.openai_serving_chat.warmup()
-    state.openai_serving_completion = (
-        OpenAIServingCompletion(
-            engine_client,
-            state.openai_serving_models,
-            request_logger=request_logger,
-            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
-            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
-            enable_force_include_usage=args.enable_force_include_usage,
-            log_error_stack=args.log_error_stack,
-        )
-        if "generate" in supported_tasks
-        else None
-    )
-    state.anthropic_serving_messages = (
-        AnthropicServingMessages(
-            engine_client,
-            state.openai_serving_models,
-            args.response_role,
-            request_logger=request_logger,
-            chat_template=resolved_chat_template,
-            chat_template_content_format=args.chat_template_content_format,
-            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
-            enable_auto_tools=args.enable_auto_tool_choice,
-            tool_parser=args.tool_call_parser,
-            reasoning_parser=args.structured_outputs_config.reasoning_parser,
-            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
-            enable_force_include_usage=args.enable_force_include_usage,
-        )
-        if "generate" in supported_tasks
-        else None
-    )
-    state.serving_tokens = (
-        ServingTokens(
-            engine_client,
-            state.openai_serving_models,
-            request_logger=request_logger,
-            return_tokens_as_token_ids=args.return_tokens_as_token_ids,
-            log_error_stack=args.log_error_stack,
-            enable_prompt_tokens_details=args.enable_prompt_tokens_details,
-            enable_log_outputs=args.enable_log_outputs,
-            force_no_detokenize=args.tokens_only,
-        )
-        if "generate" in supported_tasks
-        else None
-    )

vllm/entrypoints/openai/responses/serving.py

@@ -63,7 +63,6 @@ from vllm.engine.protocol import EngineClient
 from vllm.entrypoints.chat_utils import (
     ChatCompletionMessageParam,
     ChatTemplateContentFormatOption,
-    make_tool_call_id,
 )
 from vllm.entrypoints.logger import RequestLogger
 from vllm.entrypoints.mcp.tool_server import ToolServer
@@ -116,7 +115,6 @@ from vllm.entrypoints.openai.responses.utils import (
     extract_tool_types,
     should_continue_final_message,
 )
-from vllm.entrypoints.utils import get_max_tokens
 from vllm.exceptions import VLLMValidationError
 from vllm.inputs.data import TokensPrompt
 from vllm.logger import init_logger
@@ -252,17 +250,6 @@ class OpenAIServingResponses(OpenAIServing):
             self.default_sampling_params["stop_token_ids"].extend(
                 get_stop_tokens_for_assistant_actions()
             )
-
-        # Handle tool call ID type for Kimi K2 (supporting test mocking via overrides)
-        hf_overrides = getattr(self.model_config, "hf_overrides", None)
-        if self.model_config.hf_text_config.model_type == "kimi_k2" or (
-            isinstance(hf_overrides, dict)
-            and hf_overrides.get("model_type") == "kimi_k2"
-        ):
-            self.tool_call_id_type = "kimi_k2"
-        else:
-            self.tool_call_id_type = "random"
-
         self.enable_auto_tools = enable_auto_tools
         # set up tool use
         self.tool_parser = self._get_tool_parser(
@@ -436,11 +423,8 @@ class OpenAIServingResponses(OpenAIServing):
                 if maybe_error is not None:
                     return maybe_error
 
-                default_max_tokens = get_max_tokens(
-                    self.max_model_len,
-                    request,
-                    engine_prompt,
-                    self.default_sampling_params,
+                default_max_tokens = self.max_model_len - len(
+                    engine_prompt["prompt_token_ids"]
                 )
 
                 sampling_params = request.to_sampling_params(
@@ -970,28 +954,25 @@ class OpenAIServingResponses(OpenAIServing):
             enable_auto_tools=self.enable_auto_tools,
             tool_parser_cls=self.tool_parser,
         )
-
-        if content or (self.use_harmony and tool_calls):
-            res_text_part = None
-            if content:
-                res_text_part = ResponseOutputText(
-                    text=content,
-                    annotations=[],  # TODO
-                    type="output_text",
-                    logprobs=(
-                        self._create_response_logprobs(
-                            token_ids=final_output.token_ids,
-                            logprobs=final_output.logprobs,
-                            tokenizer=tokenizer,
-                            top_logprobs=request.top_logprobs,
-                        )
-                        if request.is_include_output_logprobs()
-                        else None
-                    ),
-                )
+        if content:
+            output_text = ResponseOutputText(
+                text=content,
+                annotations=[],  # TODO
+                type="output_text",
+                logprobs=(
+                    self._create_response_logprobs(
+                        token_ids=final_output.token_ids,
+                        logprobs=final_output.logprobs,
+                        tokenizer=tokenizer,
+                        top_logprobs=request.top_logprobs,
+                    )
+                    if request.is_include_output_logprobs()
+                    else None
+                ),
+            )
             message_item = ResponseOutputMessage(
                 id=f"msg_{random_uuid()}",
-                content=[res_text_part] if res_text_part else [],
+                content=[output_text],
                 role="assistant",
                 status="completed",
                 type="message",
@@ -1003,28 +984,17 @@ class OpenAIServingResponses(OpenAIServing):
         if message_item:
             outputs.append(message_item)
         if tool_calls:
-            # We use a simple counter for history_tool_call_count because
-            # we don't track the history of tool calls in the Responses API yet.
-            # This means that the tool call index will start from 0 for each
-            # request.
-            tool_call_items = []
-            for history_tool_call_cnt, tool_call in enumerate(tool_calls):
-                tool_call_items.append(
-                    ResponseFunctionToolCall(
-                        id=f"fc_{random_uuid()}",
-                        call_id=tool_call.id
-                        if tool_call.id
-                        else make_tool_call_id(
-                            id_type=self.tool_call_id_type,
-                            func_name=tool_call.name,
-                            idx=history_tool_call_cnt,
-                        ),
-                        type="function_call",
-                        status="completed",
-                        name=tool_call.name,
-                        arguments=tool_call.arguments,
-                    )
+            tool_call_items = [
+                ResponseFunctionToolCall(
+                    id=f"fc_{random_uuid()}",
+                    call_id=f"call_{random_uuid()}",
+                    type="function_call",
+                    status="completed",
+                    name=tool_call.name,
+                    arguments=tool_call.arguments,
                 )
+                for tool_call in tool_calls
+            ]
             outputs.extend(tool_call_items)
         return outputs
 

vllm/entrypoints/openai/translations/api_router.py

@@ -3,7 +3,7 @@
 
 
 from http import HTTPStatus
-from typing import TYPE_CHECKING, Annotated
+from typing import Annotated
 
 from fastapi import APIRouter, FastAPI, Form, Request
 from fastapi.responses import JSONResponse, StreamingResponse
@@ -25,17 +25,6 @@ from vllm.entrypoints.utils import (
 )
 from vllm.logger import init_logger
 
-if TYPE_CHECKING:
-    from argparse import Namespace
-
-    from starlette.datastructures import State
-
-    from vllm.engine.protocol import EngineClient
-    from vllm.entrypoints.logger import RequestLogger
-    from vllm.tasks import SupportedTask
-else:
-    RequestLogger = object
-
 logger = init_logger(__name__)
 
 router = APIRouter()
@@ -126,34 +115,3 @@ async def create_translations(
 
 def attach_router(app: FastAPI):
     app.include_router(router)
-
-
-def init_transcription_state(
-    engine_client: "EngineClient",
-    state: "State",
-    args: "Namespace",
-    request_logger: RequestLogger | None,
-    supported_tasks: tuple["SupportedTask", ...],
-):
-    state.openai_serving_transcription = (
-        OpenAIServingTranscription(
-            engine_client,
-            state.openai_serving_models,
-            request_logger=request_logger,
-            log_error_stack=args.log_error_stack,
-            enable_force_include_usage=args.enable_force_include_usage,
-        )
-        if "transcription" in supported_tasks
-        else None
-    )
-    state.openai_serving_translation = (
-        OpenAIServingTranslation(
-            engine_client,
-            state.openai_serving_models,
-            request_logger=request_logger,
-            log_error_stack=args.log_error_stack,
-            enable_force_include_usage=args.enable_force_include_usage,
-        )
-        if "transcription" in supported_tasks
-        else None
-    )