[Build] Update pre-commit pip-compile hook to use cu128 torch backend

Matches the test.txt lockfile change to CUDA 12.8.

Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>

.buildkite/image_build/image_build.sh

@@ -8,7 +8,7 @@ clean_docker_tag() {
 }
 
 print_usage_and_exit() {
-    echo "Usage: $0 <registry> <repo> <commit> <branch> <vllm_use_precompiled> <vllm_merge_base_commit> <cache_from> <cache_to>"
+    echo "Usage: $0 <registry> <repo> <commit> <branch> <image_tag> [<image_tag_latest>]"
     exit 1
 }
 
@@ -142,11 +142,16 @@ resolve_parent_commit() {
 
 print_bake_config() {
     echo "--- :page_facing_up: Resolved bake configuration"
-    BAKE_CONFIG_FILE="bake-config-build-${BUILDKITE_BUILD_NUMBER:-local}.json"
+    # Write to a temp directory to avoid polluting the repo root (which is the
+    # Docker build context). Files left in the repo root get COPY'd into the
+    # image and can cause duplicate artifact uploads from downstream steps.
+    local bake_tmp
+    bake_tmp="$(mktemp -d)"
+    BAKE_CONFIG_FILE="${bake_tmp}/bake-config-build-${BUILDKITE_BUILD_NUMBER:-local}.json"
     docker buildx bake -f "${VLLM_BAKE_FILE_PATH}" -f "${CI_HCL_PATH}" --print "${TARGET}" | tee "${BAKE_CONFIG_FILE}" || true
     echo "Saved bake config to ${BAKE_CONFIG_FILE}"
     echo "--- :arrow_down: Uploading bake config to Buildkite"
-    buildkite-agent artifact upload "${BAKE_CONFIG_FILE}"
+    (cd "$(dirname "${BAKE_CONFIG_FILE}")" && buildkite-agent artifact upload "$(basename "${BAKE_CONFIG_FILE}")")
 }
 
 #################################
@@ -154,7 +159,7 @@ print_bake_config() {
 #################################
 print_instance_info
 
-if [[ $# -lt 7 ]]; then
+if [[ $# -lt 5 ]]; then
     print_usage_and_exit
 fi
 
@@ -163,10 +168,8 @@ 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
+IMAGE_TAG=$5
+IMAGE_TAG_LATEST=${6:-} # only used for main branch, optional
 
 # build config
 TARGET="test-ci"
@@ -193,8 +196,6 @@ 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"
@@ -202,8 +203,6 @@ 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}"
 

.buildkite/image_build/image_build.yaml

@@ -5,8 +5,7 @@ steps:
     depends_on: []
     timeout_in_minutes: 600
     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 $IMAGE_TAG_LATEST; fi
+    - if [[ "$BUILDKITE_BRANCH" == "main" ]]; then .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $IMAGE_TAG $IMAGE_TAG_LATEST; else .buildkite/image_build/image_build.sh $REGISTRY $REPO $BUILDKITE_COMMIT $BRANCH $IMAGE_TAG; fi
     retry:
       automatic:
         - exit_status: -1  # Agent was lost

.buildkite/image_build/image_build_cpu.sh

@@ -11,10 +11,10 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu) ]]; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -24,13 +24,13 @@ fi
 # build
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
-  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
+  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
   --build-arg VLLM_CPU_AVX512BF16=true \
   --build-arg VLLM_CPU_AVX512VNNI=true \
   --build-arg VLLM_CPU_AMXBF16=true \
-  --tag $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu \
+  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
   --target vllm-test \
   --progress plain .
 
 # push
-docker push $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu
+docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu

.buildkite/image_build/image_build_cpu_arm64.sh

@@ -11,10 +11,10 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu) ]]; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -24,10 +24,10 @@ fi
 # build
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
-  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
-  --tag $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu \
+  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
   --target vllm-test \
   --progress plain .
 
 # push
-docker push $REGISTRY/$REPO:$BUILDKITE_COMMIT-cpu
+docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu

.buildkite/image_build/image_build_hpu.sh

@@ -11,10 +11,10 @@ REPO=$2
 BUILDKITE_COMMIT=$3
 
 # authenticate with AWS ECR
-aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin $REGISTRY
+aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$REGISTRY"
 
 # skip build if image already exists
-if [[ -z $(docker manifest inspect $REGISTRY/$REPO:$BUILDKITE_COMMIT-hpu) ]]; then
+if [[ -z $(docker manifest inspect "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu) ]]; then
   echo "Image not found, proceeding with build..."
 else
   echo "Image found"
@@ -25,10 +25,10 @@ fi
 docker build \
   --file tests/pytorch_ci_hud_benchmark/Dockerfile.hpu \
   --build-arg max_jobs=16 \
-  --build-arg buildkite_commit=$BUILDKITE_COMMIT \
-  --tag $REGISTRY/$REPO:$BUILDKITE_COMMIT-hpu \
+  --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
+  --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu \
   --progress plain \
   https://github.com/vllm-project/vllm-gaudi.git
 
 # push
-docker push $REGISTRY/$REPO:$BUILDKITE_COMMIT-hpu
+docker push "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-hpu

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

@@ -2,7 +2,7 @@
 # We can use this script to compute baseline accuracy on chartqa for vllm.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.9.2"
+#   pip install "lm-eval[api]>=0.4.11"
 
 usage() {
     echo``
@@ -41,4 +41,4 @@ lm_eval --model vllm-vlm \
   --tasks chartqa \
   --batch_size auto \
   --apply_chat_template \
-  --limit $LIMIT
+  --limit "$LIMIT"

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

@@ -2,7 +2,7 @@
 # We can use this script to compute baseline accuracy on GSM for transformers.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.9.2"
+#   pip install "lm-eval[api]>=0.4.11"
 
 usage() {
     echo``

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

@@ -3,7 +3,7 @@
 # We use this for fp8, which HF does not support.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.9.2"
+#   pip install "lm-eval[api]>=0.4.11"
 
 usage() {
     echo``

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

@@ -3,7 +3,7 @@
 # We use this for fp8, which HF does not support.
 #
 # Make sure you have lm-eval-harness installed:
-#   pip install "lm-eval[api]>=0.4.9.2"
+#   pip install "lm-eval[api]>=0.4.11"
 
 usage() {
     echo``
@@ -20,14 +20,11 @@ usage() {
     echo
 }
 
-while getopts "m:b:l:f:t:" OPT; do
+while getopts "m:l:f:t:" OPT; do
   case ${OPT} in
     m )
         MODEL="$OPTARG"
         ;;
-    b )
-        BATCH_SIZE="$OPTARG"
-        ;;
     l )
         LIMIT="$OPTARG"
         ;;

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

@@ -15,11 +15,11 @@ DTYPE_FILTER="${DTYPE_FILTER:-}"
 check_gpus() {
   if command -v nvidia-smi; then
     # check the number of GPUs and GPU type.
-    declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+    declare -g gpu_count=$(nvidia-smi --list-gpus | grep -c . || true)
   elif command -v amd-smi; then
-    declare -g gpu_count=$(amd-smi list | grep 'GPU' | wc -l)
+    declare -g gpu_count=$(amd-smi list | grep -c 'GPU' || true)
   elif command -v hl-smi; then
-    declare -g gpu_count=$(hl-smi --list | grep -i "Module ID" | wc -l)
+    declare -g gpu_count=$(hl-smi --list | grep -ci "Module ID" || true)
   fi
 
   if [[ $gpu_count -gt 0 ]]; then
@@ -47,7 +47,7 @@ check_cpus() {
   declare -g numa_count=$(lscpu | grep "NUMA node(s):" | awk '{print $3}')
   if [[ $numa_count -gt 0 ]]; then
     echo "NUMA found."
-    echo $numa_count
+    echo "$numa_count"
   else
     echo "Need at least 1 NUMA to run benchmarking."
     exit 1
@@ -434,7 +434,7 @@ run_serving_tests() {
 
       # iterate over different max_concurrency
       for max_concurrency in $max_concurrency_list; do
-        new_test_name=$test_name"_qps_"$qps"_concurrency_"$max_concurrency
+        new_test_name="${test_name}_qps_${qps}_concurrency_${max_concurrency}"
         echo " new test name $new_test_name"
         # pass the tensor parallel size, the compilation mode, and the optimization
         # level to the client so that they can be used on the benchmark dashboard
@@ -471,7 +471,7 @@ run_serving_tests() {
 
     # clean up
     if [[ "${DRY_RUN:-0}" != "1" ]]; then
-      kill -9 $server_pid
+      kill -9 "$server_pid"
       kill_gpu_processes
     fi
   done

.buildkite/release-pipeline.yaml

@@ -31,7 +31,7 @@ steps:
         commands:
           # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
           # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0' --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04  --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -70,7 +70,7 @@ steps:
         agents:
           queue: cpu_queue_postmerge
         commands:
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -123,7 +123,7 @@ steps:
           queue: cpu_queue_postmerge
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
           # re-tag to default image tag and push, just in case arm64 build fails
           - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-cu130"
@@ -137,7 +137,7 @@ steps:
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
           # compute capability 12.0 for RTX-50 series / RTX PRO 6000 Blackwell, 12.1 for DGX Spark
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=13.0.1 --build-arg BUILDER_CUDA_VERSION=13.0.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX 12.0 12.1' --build-arg INSTALL_KV_CONNECTORS=true --build-arg BUILD_BASE_IMAGE=nvidia/cuda:13.0.1-devel-ubuntu22.04 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130 --target vllm-openai --progress plain -f docker/Dockerfile ."
           - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)-cu130"
 
       - block: "Build release image for x86_64 CPU"

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

@@ -25,7 +25,7 @@ S3_REGION="${AWS_DEFAULT_REGION:-us-west-2}"
 S3_URL="http://${S3_BUCKET}.s3-website-${S3_REGION}.amazonaws.com"
 
 # Format ROCm version for path (e.g., "7.1" -> "rocm710")
-ROCM_VERSION_PATH="rocm$(echo ${ROCM_VERSION} | tr -d '.')"
+ROCM_VERSION_PATH="rocm$(echo "${ROCM_VERSION}" | tr -d '.')"
 ROCM_PATH="rocm/${BUILDKITE_COMMIT}/${ROCM_VERSION_PATH}"
 buildkite-agent annotate --style 'success' --context 'rocm-release-workflow' << EOF
 ## ROCm Wheel and Docker Image Releases

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

@@ -83,7 +83,7 @@ case "${1:-}" in
             exit 1
         fi
 
-        WHEEL_COUNT=$(ls artifacts/rocm-base-wheels/*.whl 2>/dev/null | wc -l)
+        WHEEL_COUNT=$(find artifacts/rocm-base-wheels -maxdepth 1 -name '*.whl' 2>/dev/null | wc -l)
         if [[ "$WHEEL_COUNT" -eq 0 ]]; then
             echo "ERROR: No wheels found in artifacts/rocm-base-wheels/" >&2
             exit 1
@@ -110,9 +110,9 @@ case "${1:-}" in
 
         echo ""
         echo "Downloaded wheels:"
-        ls -lh artifacts/rocm-base-wheels/
+        find artifacts/rocm-base-wheels -maxdepth 1 -name '*.whl' -exec ls -lh {} \;
 
-        WHEEL_COUNT=$(ls artifacts/rocm-base-wheels/*.whl 2>/dev/null | wc -l)
+        WHEEL_COUNT=$(find artifacts/rocm-base-wheels -maxdepth 1 -name '*.whl' 2>/dev/null | wc -l)
         echo ""
         echo "Total: $WHEEL_COUNT wheels"
         echo "========================================"

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

@@ -134,7 +134,7 @@ log_info "Fetching merged PRs from milestone '${MILESTONE}'..."
 
 # Store PR data in a temp file
 PR_DATA=$(mktemp)
-trap "rm -f $PR_DATA" EXIT
+trap 'rm -f "$PR_DATA"' EXIT
 
 if ! gh pr list --state merged --search "milestone:${MILESTONE}" \
     --limit 1000 \

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -27,7 +27,7 @@ function cpu_tests() {
   podman exec -it "$container_id" bash -c "
     export TORCH_COMPILE_DISABLE=1
     set -xve
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> $HOME/test_basic.log
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> "$HOME"/test_basic.log
 
   # Run basic model test
   podman exec -it "$container_id" bash -c "
@@ -43,7 +43,7 @@ function cpu_tests() {
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-google/gemma-1.1-2b-it]
     pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
     # TODO: Below test case tests/models/language/pooling/test_embedding.py::test_models[True-ssmits/Qwen2-7B-Instruct-embed-base] fails on ppc64le. Disabling it for time being.
-    # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> $HOME/test_rest.log
+    # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> "$HOME"/test_rest.log
 }
 
 # All of CPU tests are expected to be finished less than 40 mins.

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

@@ -16,5 +16,5 @@ echo "--- :docker: Building Docker image"
 docker build --progress plain --tag "$IMAGE_NAME" --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run --rm --cpuset-cpus=$CORE_RANGE --cpuset-mems=$NUMA_NODE -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN -e VLLM_CPU_KVCACHE_SPACE=16 -e VLLM_CPU_CI_ENV=1 -e VLLM_CPU_SIM_MULTI_NUMA=1 --shm-size=4g $IMAGE_NAME \
-        timeout $TIMEOUT_VAL bash -c "set -euox pipefail; echo \"--- Print packages\"; pip list; echo \"--- Running tests\"; ${TEST_COMMAND}"
+docker run --rm --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN -e VLLM_CPU_KVCACHE_SPACE=16 -e VLLM_CPU_CI_ENV=1 -e VLLM_CPU_SIM_MULTI_NUMA=1 --shm-size=4g "$IMAGE_NAME" \
+        timeout "$TIMEOUT_VAL" bash -c "set -euox pipefail; echo \"--- Print packages\"; pip list; echo \"--- Running tests\"; ${TEST_COMMAND}"

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

@@ -7,7 +7,7 @@ set -exuo pipefail
 # Try building the docker image
 image_name="hpu/upstream-vllm-ci:${BUILDKITE_COMMIT}"
 container_name="hpu-upstream-vllm-ci-${BUILDKITE_COMMIT}-container"
-cat <<EOF | docker build -t ${image_name} -f - .
+cat <<EOF | docker build -t "${image_name}" -f - .
 FROM gaudi-base-image:latest
 
 COPY ./ /workspace/vllm
@@ -39,12 +39,12 @@ EOF
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_containers() { docker rm -f ${container_name} || true; }
+remove_docker_containers() { docker rm -f "${container_name}" || true; }
 trap 'remove_docker_containers; exit $EXITCODE;' EXIT
 remove_docker_containers
 
 echo "Running HPU plugin v1 test"
-docker run --rm --runtime=habana --name=${container_name} --network=host \
+docker run --rm --runtime=habana --name="${container_name}" --network=host \
   -e HABANA_VISIBLE_DEVICES=all \
   -e VLLM_SKIP_WARMUP=true \
   -e PT_HPU_ENABLE_LAZY_COLLECTIVES=true \

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

@@ -41,6 +41,7 @@ get_config() {
         echo "Error: file '${TEST_RUN_CONFIG_FILE}' does not exist in the warehouse" >&2
         exit 1
     fi
+    # shellcheck source=/dev/null
     source "${TEST_RUN_CONFIG_FILE}"
     echo "Base docker image name that get from configuration: ${BASE_IMAGE_NAME}"
     return 0
@@ -48,9 +49,8 @@ get_config() {
 
 # get test running configuration.
 fetch_vllm_test_cfg
-get_config
 # Check if the function call was successful. If not, exit the script.
-if [ $? -ne 0 ]; then
+if ! get_config; then
   exit 1
 fi
 
@@ -62,14 +62,14 @@ agent_idx=$(echo "${BUILDKITE_AGENT_NAME}" | awk -F'-' '{print $(NF-1)}')
 echo "agent_idx: ${agent_idx}"
 builder_name="cachebuilder${agent_idx}"
 builder_cache_dir="/mnt/docker-cache${agent_idx}"
-mkdir -p ${builder_cache_dir}
+mkdir -p "${builder_cache_dir}"
 
 # Try building the docker image
 cat <<EOF | DOCKER_BUILDKIT=1 docker build \
-    --add-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local:${PYPI_CACHE_HOST} \
-    --builder ${builder_name} --cache-from type=local,src=${builder_cache_dir} \
-                           --cache-to type=local,dest=${builder_cache_dir},mode=max \
-    --progress=plain --load -t ${image_name} -f - .
+    --add-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local:"${PYPI_CACHE_HOST}" \
+    --builder "${builder_name}" --cache-from type=local,src="${builder_cache_dir}" \
+                           --cache-to type=local,dest="${builder_cache_dir}",mode=max \
+    --progress=plain --load -t "${image_name}" -f - .
 FROM ${BASE_IMAGE_NAME}
 
 # Define environments
@@ -116,7 +116,7 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
     source /usr/local/Ascend/nnal/atb/set_env.sh && \
-    export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
+    export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/$(uname -i)-linux/devlib && \
     python3 -m pip install -v -e /workspace/vllm-ascend/ --extra-index https://download.pytorch.org/whl/cpu/
 
 ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
@@ -139,7 +139,7 @@ trap remove_docker_container EXIT
 # Generate corresponding --device args based on BUILDKITE_AGENT_NAME
 # Ascend NPU BUILDKITE_AGENT_NAME format is {hostname}-{agent_idx}-{npu_card_num}cards, and agent_idx starts from 1.
 #   e.g. atlas-a2-001-1-2cards means this is the 1-th agent on atlas-a2-001 host, and it has 2 NPU cards.
-#   returns --device /dev/davinci0 --device /dev/davinci1
+#   returns one argument per line: --device, /dev/davinciX, ...
 parse_and_gen_devices() {
     local input="$1"
     local index cards_num
@@ -151,29 +151,24 @@ parse_and_gen_devices() {
         return 1
     fi
 
-    local devices=""
     local i=0
     while (( i < cards_num )); do
         local dev_idx=$(((index - 1)*cards_num + i ))
-        devices="$devices --device /dev/davinci${dev_idx}"
+        printf '%s\n' "--device"
+        printf '%s\n' "/dev/davinci${dev_idx}"
         ((i++))
     done
-
-    # trim leading space
-    devices="${devices#"${devices%%[![:space:]]*}"}"
-    # Output devices: assigned to the caller variable
-    printf '%s' "$devices"
 }
 
-devices=$(parse_and_gen_devices "${BUILDKITE_AGENT_NAME}") || exit 1
+mapfile -t device_args < <(parse_and_gen_devices "${BUILDKITE_AGENT_NAME}") || exit 1
 
 # Run the image and execute the Out-Of-Tree (OOT) platform interface test case on Ascend NPU hardware.
 # This test checks whether the OOT platform interface is functioning properly in conjunction with
 # the hardware plugin vllm-ascend.
 model_cache_dir=/mnt/modelscope${agent_idx}
-mkdir -p ${model_cache_dir}
+mkdir -p "${model_cache_dir}"
 docker run \
-    ${devices} \
+    "${device_args[@]}" \
     --device /dev/davinci_manager \
     --device /dev/devmm_svm \
     --device /dev/hisi_hdc \
@@ -182,7 +177,7 @@ docker run \
     -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
     -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
     -v /etc/ascend_install.info:/etc/ascend_install.info \
-    -v ${model_cache_dir}:/root/.cache/modelscope \
+    -v "${model_cache_dir}":/root/.cache/modelscope \
     --entrypoint="" \
     --name "${container_name}" \
     "${image_name}" \

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

@@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.9.2" \
+    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.11" \
     && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 

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

@@ -61,7 +61,7 @@ echo "Results will be stored in: $RESULTS_DIR"
 echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
-    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.9.2" \
+    && python3 -m pip install --progress-bar off "lm-eval[api]>=0.4.11" \
     && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 

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

@@ -8,7 +8,7 @@ image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 
 # Try building the docker image
-docker build -t ${image_name} -f docker/Dockerfile.xpu .
+docker build -t "${image_name}" -f docker/Dockerfile.xpu .
 
 # Setup cleanup
 remove_docker_container() {

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

@@ -21,16 +21,16 @@ echo "Pushing original tag $ORIG_TAG_NAME$ORIG_TAG_SUFFIX to new nightly tag nam
 
 # pull original arch-dependent images from AWS ECR Public
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-x86_64$ORIG_TAG_SUFFIX
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-aarch64$ORIG_TAG_SUFFIX
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64"$ORIG_TAG_SUFFIX"
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-aarch64"$ORIG_TAG_SUFFIX"
 # tag arch-dependent images
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-x86_64$ORIG_TAG_SUFFIX vllm/vllm-openai:$TAG_NAME-x86_64
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$ORIG_TAG_NAME-aarch64$ORIG_TAG_SUFFIX vllm/vllm-openai:$TAG_NAME-aarch64
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-x86_64"$ORIG_TAG_SUFFIX" vllm/vllm-openai:"$TAG_NAME"-x86_64
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:"$ORIG_TAG_NAME"-aarch64"$ORIG_TAG_SUFFIX" vllm/vllm-openai:"$TAG_NAME"-aarch64
 # push arch-dependent images to DockerHub
-docker push vllm/vllm-openai:$TAG_NAME-x86_64
-docker push vllm/vllm-openai:$TAG_NAME-aarch64
+docker push vllm/vllm-openai:"$TAG_NAME"-x86_64
+docker push vllm/vllm-openai:"$TAG_NAME"-aarch64
 # push arch-independent manifest to DockerHub
-docker manifest create vllm/vllm-openai:$TAG_NAME vllm/vllm-openai:$TAG_NAME-x86_64 vllm/vllm-openai:$TAG_NAME-aarch64 --amend
-docker manifest create vllm/vllm-openai:$TAG_NAME-$BUILDKITE_COMMIT vllm/vllm-openai:$TAG_NAME-x86_64 vllm/vllm-openai:$TAG_NAME-aarch64 --amend
-docker manifest push vllm/vllm-openai:$TAG_NAME
-docker manifest push vllm/vllm-openai:$TAG_NAME-$BUILDKITE_COMMIT
+docker manifest create vllm/vllm-openai:"$TAG_NAME" vllm/vllm-openai:"$TAG_NAME"-x86_64 vllm/vllm-openai:"$TAG_NAME"-aarch64 --amend
+docker manifest create vllm/vllm-openai:"$TAG_NAME"-"$BUILDKITE_COMMIT" vllm/vllm-openai:"$TAG_NAME"-x86_64 vllm/vllm-openai:"$TAG_NAME"-aarch64 --amend
+docker manifest push vllm/vllm-openai:"$TAG_NAME"
+docker manifest push vllm/vllm-openai:"$TAG_NAME"-"$BUILDKITE_COMMIT"

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

@@ -67,7 +67,7 @@ start_nodes() {
         # 3. map the huggingface cache directory to the container
         # 3. assign ip addresses to the containers (head node: 192.168.10.10, worker nodes:
         #    starting from 192.168.10.11)
-        docker run -d $GPU_DEVICES --shm-size=10.24gb -e HF_TOKEN \
+        docker run -d "$GPU_DEVICES" --shm-size=10.24gb -e HF_TOKEN \
             -v ~/.cache/huggingface:/root/.cache/huggingface --name "node$node" \
             --network docker-net --ip 192.168.10.$((10 + $node)) --rm "$DOCKER_IMAGE" \
             /bin/bash -c "tail -f /dev/null"

.buildkite/scripts/run-prime-rl-test.sh

@@ -29,7 +29,7 @@ fi
 if ! command -v uv &> /dev/null; then
     echo "Installing UV package manager..."
     curl -LsSf https://astral.sh/uv/install.sh | sh
-    source $HOME/.local/bin/env
+    source "$HOME"/.local/bin/env
 fi
 
 # Clone Prime-RL repository at specific branch for reproducible tests

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh

@@ -51,14 +51,14 @@ for BACK in "${BACKENDS[@]}"; do
     --enable-eplb \
     --trust-remote-code \
     --max-model-len 2048 \
-    --all2all-backend $BACK \
-    --port $PORT &
+    --all2all-backend "$BACK" \
+    --port "$PORT" &
   SERVER_PID=$!
-  wait_for_server $PORT
+  wait_for_server "$PORT"
 
   TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
   OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
+  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
   python3 - <<PY
 import json; acc=json.load(open('${OUT}'))['accuracy']
 print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")

.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh

@@ -47,20 +47,20 @@ for BACK in "${BACKENDS[@]}"; do
   vllm serve "$MODEL" \
     --enforce-eager \
     --enable-eplb \
-    --all2all-backend $BACK \
+    --all2all-backend "$BACK" \
     --eplb-config '{"window_size":10, "step_interval":100, "num_redundant_experts":0, "log_balancedness":true}' \
-    --tensor-parallel-size ${TENSOR_PARALLEL_SIZE} \
-    --data-parallel-size ${DATA_PARALLEL_SIZE} \
+    --tensor-parallel-size "${TENSOR_PARALLEL_SIZE}" \
+    --data-parallel-size "${DATA_PARALLEL_SIZE}" \
     --enable-expert-parallel \
     --trust-remote-code \
     --max-model-len 2048 \
-    --port $PORT &
+    --port "$PORT" &
   SERVER_PID=$!
-  wait_for_server $PORT
+  wait_for_server "$PORT"
 
   TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
   OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
+  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
   python3 - <<PY
 import json; acc=json.load(open('${OUT}'))['accuracy']
 print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")

.buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh

@@ -51,20 +51,20 @@ for BACK in "${BACKENDS[@]}"; do
     --tensor-parallel-size 4 \
     --enable-expert-parallel \
     --enable-eplb \
-    --all2all-backend $BACK \
+    --all2all-backend "$BACK" \
     --eplb-config '{"window_size":200,"step_interval":600,"use_async":true}' \
     --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":1}' \
     --trust-remote-code \
     --max-model-len 2048 \
     --gpu-memory-utilization 0.9 \
     "${PLATFORM_ARGS[@]}" \
-    --port $PORT &
+    --port "$PORT" &
   SERVER_PID=$!
-  wait_for_server $PORT
+  wait_for_server "$PORT"
 
   TAG=$(echo "$MODEL" | tr '/: \\n' '_____')
   OUT="${OUT_DIR}/${TAG}_${BACK}.json"
-  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port $PORT --num-questions ${NUM_Q} --save-results ${OUT}
+  python3 tests/evals/gsm8k/gsm8k_eval.py --host http://127.0.0.1 --port "$PORT" --num-questions "${NUM_Q}" --save-results "${OUT}"
   python3 - <<PY
 import json; acc=json.load(open('${OUT}'))['accuracy']
 print(f"${MODEL} ${BACK}: accuracy {acc:.3f}")

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -9,10 +9,11 @@ ENV_FILE=$1
 
 # For testing on local vm, use `set -a` to export all variables
 source /etc/environment
-source $ENV_FILE
+# shellcheck source=/dev/null
+source "$ENV_FILE"
 
 remove_docker_container() { 
-    docker rm -f $CONTAINER_NAME || true;
+    docker rm -f "$CONTAINER_NAME" || true;
 }
 
 trap remove_docker_container EXIT
@@ -41,13 +42,13 @@ echo
 echo "starting docker...$CONTAINER_NAME"
 echo    
 docker run \
- -v $DOWNLOAD_DIR:$DOWNLOAD_DIR \
- --env-file $ENV_FILE \
+ -v "$DOWNLOAD_DIR":"$DOWNLOAD_DIR" \
+ --env-file "$ENV_FILE" \
  -e HF_TOKEN="$HF_TOKEN" \
- -e TARGET_COMMIT=$BUILDKITE_COMMIT \
- -e MODEL=$MODEL \
+ -e TARGET_COMMIT="$BUILDKITE_COMMIT" \
+ -e MODEL="$MODEL" \
  -e WORKSPACE=/workspace \
- --name $CONTAINER_NAME \
+ --name "$CONTAINER_NAME" \
  -d \
  --privileged \
  --network host \

.buildkite/scripts/tpu/run_bm.sh

@@ -42,21 +42,21 @@ echo "lanching vllm..."
 echo "logging to $VLLM_LOG"
 echo
 
-vllm serve $MODEL \
+vllm serve "$MODEL" \
  --seed 42 \
- --max-num-seqs $MAX_NUM_SEQS \
- --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
- --tensor-parallel-size $TENSOR_PARALLEL_SIZE \
+ --max-num-seqs "$MAX_NUM_SEQS" \
+ --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS" \
+ --tensor-parallel-size "$TENSOR_PARALLEL_SIZE" \
  --no-enable-prefix-caching \
- --download_dir $DOWNLOAD_DIR \
- --max-model-len $MAX_MODEL_LEN > "$VLLM_LOG" 2>&1 &
+ --download_dir "$DOWNLOAD_DIR" \
+ --max-model-len "$MAX_MODEL_LEN" > "$VLLM_LOG" 2>&1 &
 
 
 echo "wait for 20 minutes.."
 echo
 # sleep 1200
 # wait for 10 minutes...
-for i in {1..120}; do
+for _ in {1..120}; do
     # TODO: detect other type of errors.
     if grep -Fq "raise RuntimeError" "$VLLM_LOG"; then
         echo "Detected RuntimeError, exiting."
@@ -78,11 +78,11 @@ echo "logging to $BM_LOG"
 echo
 vllm bench serve \
     --backend vllm \
-    --model $MODEL  \
+    --model "$MODEL"  \
     --dataset-name sonnet \
     --dataset-path benchmarks/sonnet_4x.txt \
-    --sonnet-input-len $INPUT_LEN \
-    --sonnet-output-len $OUTPUT_LEN \
+    --sonnet-input-len "$INPUT_LEN" \
+    --sonnet-output-len "$OUTPUT_LEN" \
     --ignore-eos > "$BM_LOG"
 
 echo "completed..."

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

@@ -76,16 +76,15 @@ mkdir -p "$INDICES_OUTPUT_DIR"
 # this indices have relative paths that could work as long as it is next to the wheel directory in s3
 # i.e., the wheels are always in s3://vllm-wheels/<commit>/
 # and indices can be placed in /<commit>/, or /nightly/, or /<version>/
-if [[ ! -z "$DEFAULT_VARIANT_ALIAS" ]]; then
-    alias_arg="--alias-to-default $DEFAULT_VARIANT_ALIAS"
-else
-    alias_arg=""
+alias_args=()
+if [[ -n "$DEFAULT_VARIANT_ALIAS" ]]; then
+    alias_args=(--alias-to-default "$DEFAULT_VARIANT_ALIAS")
 fi
 
 # HACK: we do not need regex module here, but it is required by pre-commit hook
 # To avoid any external dependency, we simply replace it back to the stdlib re module
 sed -i 's/import regex as re/import re/g' .buildkite/scripts/generate-nightly-index.py
-$PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "commit $BUILDKITE_COMMIT" $alias_arg
+$PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "commit $BUILDKITE_COMMIT" "${alias_args[@]}"
 
 # copy indices to /<commit>/ unconditionally
 echo "Uploading indices to $S3_COMMIT_PREFIX"
@@ -100,9 +99,9 @@ fi
 # re-generate and copy to /<pure_version>/ only if it does not have "dev" in the version
 if [[ "$version" != *"dev"* ]]; then
     echo "Re-generating indices for /$pure_version/"
-    rm -rf "$INDICES_OUTPUT_DIR/*"
+    rm -rf "${INDICES_OUTPUT_DIR:?}/*"
     mkdir -p "$INDICES_OUTPUT_DIR"
     # wheel-dir is overridden to be the commit directory, so that the indices point to the correct wheel path
-    $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$pure_version" --wheel-dir "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "version $pure_version" $alias_arg
+    $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$pure_version" --wheel-dir "$SUBPATH" --current-objects "$obj_json" --output-dir "$INDICES_OUTPUT_DIR" --comment "version $pure_version" "${alias_args[@]}"
     aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/$pure_version/"
 fi

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

@@ -7,7 +7,7 @@ SUBPATH=$BUILDKITE_COMMIT
 S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
 
 RELEASE_VERSION=$(buildkite-agent meta-data get release-version)
-GIT_VERSION=$(git describe --exact-match --tags $BUILDKITE_COMMIT 2>/dev/null)
+GIT_VERSION=$(git describe --exact-match --tags "$BUILDKITE_COMMIT" 2>/dev/null)
 
 echo "Release version from Buildkite: $RELEASE_VERSION"
 
@@ -55,7 +55,7 @@ mkdir -p $DIST_DIR
 aws s3 cp --recursive --exclude "*" --include "vllm-${PURE_VERSION}*.whl" --exclude "*dev*" --exclude "*rc[0-9]*" "$S3_COMMIT_PREFIX" $DIST_DIR
 echo "Wheels copied to local directory"
 # generate source tarball
-git archive --format=tar.gz --output="$DIST_DIR/vllm-${PURE_VERSION}.tar.gz" $BUILDKITE_COMMIT
+git archive --format=tar.gz --output="$DIST_DIR/vllm-${PURE_VERSION}.tar.gz" "$BUILDKITE_COMMIT"
 ls -la $DIST_DIR
 
 # upload wheels to PyPI (only default variant, i.e. files without '+' in the name)
@@ -65,6 +65,6 @@ if [[ -z "$PYPI_WHEEL_FILES" ]]; then
   exit 1
 fi
 
-python3 -m twine check $PYPI_WHEEL_FILES
-python3 -m twine upload --non-interactive --verbose $PYPI_WHEEL_FILES
+python3 -m twine check "$PYPI_WHEEL_FILES"
+python3 -m twine upload --non-interactive --verbose "$PYPI_WHEEL_FILES"
 echo "Wheels uploaded to PyPI"

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

@@ -55,7 +55,7 @@ mkdir -p all-rocm-wheels
 cp artifacts/rocm-base-wheels/*.whl all-rocm-wheels/ 2>/dev/null || true
 cp artifacts/rocm-vllm-wheel/*.whl all-rocm-wheels/ 2>/dev/null || true
 
-WHEEL_COUNT=$(ls all-rocm-wheels/*.whl 2>/dev/null | wc -l)
+WHEEL_COUNT=$(find all-rocm-wheels -maxdepth 1 -name '*.whl' 2>/dev/null | wc -l)
 echo "Total wheels to upload: $WHEEL_COUNT"
 
 if [ "$WHEEL_COUNT" -eq 0 ]; then
@@ -115,7 +115,7 @@ if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]] |
 fi
 
 # Extract version from vLLM wheel and update version-specific index
-VLLM_WHEEL=$(ls all-rocm-wheels/vllm*.whl 2>/dev/null | head -1)
+VLLM_WHEEL=$(find all-rocm-wheels -maxdepth 1 -name 'vllm*.whl' 2>/dev/null | head -1)
 if [ -n "$VLLM_WHEEL" ]; then
     VERSION=$(unzip -p "$VLLM_WHEEL" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
     echo "Version in wheel: $VERSION"

.buildkite/test_areas/distributed.yaml

@@ -197,6 +197,17 @@ steps:
     - 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: CrossLayer KV layout 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
+    - CROSS_LAYERS_BLOCKS=True 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"

.buildkite/test_areas/misc.yaml

@@ -123,6 +123,7 @@ steps:
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/test_pooling_params.py
+  - tests/test_ray_env.py
   - tests/multimodal
   - tests/renderers
   - tests/standalone_tests/lazy_imports.py
@@ -136,6 +137,7 @@ steps:
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s test_pooling_params.py
+  - pytest -v -s test_ray_env.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s renderers
   - pytest -v -s tokenizers_

.dockerignore

@@ -2,7 +2,6 @@
 /build
 dist
 vllm/*.so
-.deps/
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

.github/CODEOWNERS

@@ -2,7 +2,9 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @njhill @22quinn
+/vllm/compilation @zou3519 @youkaichao @ProExpertProg
+/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
+/vllm/lora @jeejeelee
 /vllm/model_executor/layers/attention @LucasWilkinson
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
@@ -11,18 +13,34 @@
 /vllm/model_executor/layers/batch_invariant.py @yewentao256 
 /vllm/multimodal @DarkLight1337 @ywang96 @NickLucche @tjtanaa
 /vllm/vllm_flash_attn @LucasWilkinson
-/vllm/lora @jeejeelee
-/vllm/reasoning @aarnphm @chaunceyjiang
-/vllm/entrypoints @aarnphm @chaunceyjiang
-/vllm/tool_parsers @aarnphm @chaunceyjiang
-/vllm/compilation @zou3519 @youkaichao @ProExpertProg
-/vllm/distributed/kv_transfer @NickLucche @ApostaC @orozery
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
 /vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
-/vllm/config/cache.py @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
+/vllm/config/cache.py @heheda12345
+
+# Entrypoints
+/vllm/entrypoints/anthropic @mgoin @DarkLight1337
+/vllm/entrypoints/cli @hmellor @mgoin @DarkLight1337 @russellb
+/vllm/entrypoints/mcp @heheda12345
+/vllm/entrypoints/openai @aarnphm @chaunceyjiang @DarkLight1337 @russellb
+/vllm/entrypoints/openai/realtime @njhill
+/vllm/entrypoints/openai/speech_to_text @NickLucche
+/vllm/entrypoints/pooling @noooop
+/vllm/entrypoints/sagemaker @DarkLight1337
+/vllm/entrypoints/serve @njhill
+/vllm/entrypoints/*.py @njhill
+/vllm/entrypoints/chat_utils.py @DarkLight1337
+/vllm/entrypoints/llm.py @DarkLight1337
+
+# Input/Output Processing
+/vllm/sampling_params.py @njhill @NickLucche
+/vllm/pooling_params.py @noooop @DarkLight1337
+/vllm/tokenizers @DarkLight1337 @njhill
+/vllm/renderers @DarkLight1337 @njhill
+/vllm/reasoning @aarnphm @chaunceyjiang
+/vllm/tool_parsers @aarnphm @chaunceyjiang
 
 # vLLM V1
 /vllm/v1/attention @LucasWilkinson
@@ -115,8 +133,8 @@ mkdocs.yaml @hmellor
 /vllm/model_executor/models/mixtral*.py @patrickvonplaten
 /vllm/model_executor/models/voxtral*.py @patrickvonplaten
 /vllm/model_executor/models/pixtral*.py @patrickvonplaten
+/vllm/tokenizers/mistral.py @patrickvonplaten
 /vllm/transformers_utils/configs/mistral.py @patrickvonplaten
-/vllm/transformers_utils/tokenizers/mistral.py @patrickvonplaten
 
 # Kernels
 /vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
@@ -152,9 +170,7 @@ mkdocs.yaml @hmellor
 /examples/pooling @noooop
 /tests/models/*/pooling* @noooop
 /tests/entrypoints/pooling @noooop
-/vllm/entrypoints/pooling @noooop
 /vllm/config/pooler.py @noooop
-/vllm/pooling_params.py @noooop
 /vllm/model_executor/layers/pooler @noooop
 
 # Security guide and policies

.pre-commit-config.yaml

@@ -38,7 +38,7 @@ repos:
   rev: 0.9.1
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:

benchmarks/auto_tune/auto_tune.sh

@@ -46,10 +46,10 @@ echo "VLLM_LOGGING_LEVEL=$VLLM_LOGGING_LEVEL"
 echo "RESULT_FILE=$RESULT"
 echo "====================== AUTO TUNEPARAMETERS ===================="
 
-rm -rf $LOG_FOLDER
-rm -rf $PROFILE_PATH
-mkdir -p $LOG_FOLDER
-mkdir -p $PROFILE_PATH
+rm -rf "$LOG_FOLDER"
+rm -rf "$PROFILE_PATH"
+mkdir -p "$LOG_FOLDER"
+mkdir -p "$PROFILE_PATH"
 
 cd "$BASE/vllm"
 
@@ -114,7 +114,7 @@ start_server() {
 
     # wait for 10 minutes...
     server_started=0
-    for i in {1..60}; do
+    for _ in {1..60}; do
         # This line checks whether the server is still alive or not,
         # since that we should always have permission to send signal to the server process.
         kill -0 $server_pid 2> /dev/null || break
@@ -145,12 +145,12 @@ run_benchmark() {
     local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
     echo "vllm_log: $vllm_log"
     echo
-    rm -f $vllm_log
+    rm -f "$vllm_log"
     pkill -if "vllm serve" || true
 
     echo "starting server..."
     # Call start_server without a profile_dir to avoid profiling overhead
-    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log ""
+    start_server "$gpu_memory_utilization" "$max_num_seqs" "$max_num_batched_tokens" "$vllm_log" ""
     result=$?
     if [[ "$result" -eq 1 ]]; then
         echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
@@ -168,15 +168,15 @@ run_benchmark() {
     # --profile flag is removed from this call
     vllm bench serve \
         --backend vllm \
-        --model $MODEL  \
+        --model "$MODEL"  \
         --dataset-name random \
         --random-input-len $adjusted_input_len \
-        --random-output-len $OUTPUT_LEN \
+        --random-output-len "$OUTPUT_LEN" \
         --ignore-eos \
         --disable-tqdm \
         --request-rate inf \
         --percentile-metrics ttft,tpot,itl,e2el \
-        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --goodput e2el:"$MAX_LATENCY_ALLOWED_MS" \
         --num-prompts 1000 \
         --random-prefix-len $prefix_len \
         --host "$HOSTNAME" \
@@ -195,20 +195,20 @@ run_benchmark() {
         request_rate=$((${throughput%.*} + 1))
         while ((request_rate > 0)); do
             # clear prefix cache
-            curl -X POST http://${HOSTNAME}:8004/reset_prefix_cache
+            curl -X POST http://"${HOSTNAME}":8004/reset_prefix_cache
             sleep 5
             bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
             vllm bench serve \
                 --backend vllm \
-                --model $MODEL  \
+                --model "$MODEL"  \
                 --dataset-name random \
                 --random-input-len $adjusted_input_len \
-                --random-output-len $OUTPUT_LEN \
+                --random-output-len "$OUTPUT_LEN" \
                 --ignore-eos \
                 --disable-tqdm \
                 --request-rate $request_rate \
                 --percentile-metrics ttft,tpot,itl,e2el \
-                --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+                --goodput e2el:"$MAX_LATENCY_ALLOWED_MS" \
                 --num-prompts 100 \
                 --random-prefix-len $prefix_len \
                 --host "$HOSTNAME" \
@@ -255,7 +255,7 @@ gpu_memory_utilization=0.98
 find_gpu_memory_utilization=0
 while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
     # Pass empty string for profile_dir argument
-    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log" ""
+    start_server "$gpu_memory_utilization" "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log" ""
     result=$?
     if [[ "$result" -eq 0 ]]; then
         find_gpu_memory_utilization=1
@@ -274,7 +274,7 @@ fi
 
 for num_seqs in "${num_seqs_list[@]}"; do
     for num_batched_tokens in "${num_batched_tokens_list[@]}"; do
-        run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization
+        run_benchmark "$num_seqs" "$num_batched_tokens" "$gpu_memory_utilization"
     done
 done
 echo "finish permutations"
@@ -285,7 +285,7 @@ echo "finish permutations"
 if (( $(echo "$best_throughput > 0" | bc -l) )); then
     echo
     echo "Benchmark tuning finished. Now running profiling on the best configuration found..."
-    echo "Best config: max_num_seqs: $best_max_num_seqs, max_num_batched_tokens: $best_num_batched_tokens, throughput: $best_throughput"
+    echo "Best config: max_num_seqs: $best_max_num_seqs, max_num_batched_tokens: $best_num_batched_tokens, throughput: $best_throughput, goodput: $best_goodput"
     echo
 
     vllm_log="$LOG_FOLDER/vllm_log_BEST_PROFILE.txt"
@@ -293,7 +293,7 @@ if (( $(echo "$best_throughput > 0" | bc -l) )); then
 
     # Start server with the best params and profiling ENABLED
     echo "Starting server for profiling..."
-    start_server $gpu_memory_utilization $best_max_num_seqs $best_num_batched_tokens "$vllm_log" "$PROFILE_PATH"
+    start_server "$gpu_memory_utilization" "$best_max_num_seqs" "$best_num_batched_tokens" "$vllm_log" "$PROFILE_PATH"
 
     # Run benchmark with the best params and the --profile flag
     echo "Running benchmark with profiling..."
@@ -301,15 +301,15 @@ if (( $(echo "$best_throughput > 0" | bc -l) )); then
     adjusted_input_len=$(( INPUT_LEN - prefix_len ))
     vllm bench serve \
         --backend vllm \
-        --model $MODEL \
+        --model "$MODEL" \
         --dataset-name random \
         --random-input-len $adjusted_input_len \
-        --random-output-len $OUTPUT_LEN \
+        --random-output-len "$OUTPUT_LEN" \
         --ignore-eos \
         --disable-tqdm \
-        --request-rate $best_request_rate \
+        --request-rate "$best_request_rate" \
         --percentile-metrics ttft,tpot,itl,e2el \
-        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --goodput e2el:"$MAX_LATENCY_ALLOWED_MS" \
         --num-prompts 100 \
         --random-prefix-len $prefix_len \
         --host "$HOSTNAME" \

benchmarks/auto_tune/batch_auto_tune.sh

@@ -64,7 +64,7 @@ for i in $(seq 0 $(($num_runs - 1))); do
   else
     STATUS="FAILURE"
     ((FAILURE_COUNT++))
-    FAILED_RUNS+=("Run #$((i+1)): $(echo $run_object | jq -c .)")
+    FAILED_RUNS+=("Run #$((i+1)): $(echo "$run_object" | jq -c .)")
   fi
 
   RUN_OUTPUT=$(<"$RUN_OUTPUT_FILE")

benchmarks/benchmark_topk_topp.py

@@ -0,0 +1,471 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark comparing Triton vs PyTorch sort-based top-k/top-p implementations.
+
+Compares:
+- apply_top_k_top_p_triton (Triton binary search)
+- apply_top_k_top_p (PyTorch sort-based)
+
+Scenarios:
+- top_k only (whole batch, partial batch)
+- top_p only (whole batch, partial batch)
+- mix of top_k and top_p
+"""
+
+import argparse
+import gc
+from dataclasses import dataclass
+
+import torch
+
+from vllm.v1.sample.ops.topk_topp_sampler import apply_top_k_top_p_pytorch
+from vllm.v1.sample.ops.topk_topp_triton import (
+    apply_top_k_top_p_triton,
+    reset_buffer_cache,
+)
+
+
+@dataclass
+class BenchmarkConfig:
+    """Configuration for a benchmark run."""
+
+    name: str
+    batch_size: int
+    vocab_size: int
+    # k and p can be tensors or None
+    k_values: torch.Tensor | None  # [batch_size] or None
+    p_values: torch.Tensor | None  # [batch_size] or None
+    description: str
+    ops_pct: float = 0.0  # Percentage of ops relative to batch size
+
+
+def calculate_ops_pct(
+    k_values: torch.Tensor | None,
+    p_values: torch.Tensor | None,
+    vocab_size: int,
+    batch_size: int,
+) -> float:
+    """
+    Calculate the percentage of active top-k and top-p operations.
+
+    Returns percentage where 100% = batch_size ops.
+    E.g., if all rows have both top-k and top-p active, returns 200%.
+    """
+    active_ops = 0
+
+    if k_values is not None:
+        # Count rows where k < vocab_size (active top-k filtering)
+        active_ops += (k_values < vocab_size).sum().item()
+
+    if p_values is not None:
+        # Count rows where p < 1.0 (active top-p filtering)
+        active_ops += (p_values < 1.0).sum().item()
+
+    return (active_ops / batch_size) * 100 if batch_size > 0 else 0.0
+
+
+def create_logits(
+    batch_size: int, vocab_size: int, device: str = "cuda"
+) -> torch.Tensor:
+    """Create random logits mimicking a realistic LLM distribution.
+
+    Uses a Zipf-like probability distribution (rank^-1.1) converted to logits
+    via log, then randomly permuted per row. This produces a peaked distribution
+    where a small number of tokens capture most probability mass, similar to
+    real model outputs.
+    """
+    # Create Zipf-like probabilities: p(rank) ~ rank^(-alpha)
+    ranks = torch.arange(1, vocab_size + 1, dtype=torch.float32, device=device)
+    probs = ranks.pow(-1.1)
+    probs = probs / probs.sum()
+
+    # Convert to logits (log-probabilities, unnormalized is fine)
+    base_logits = probs.log()
+
+    # Broadcast to batch and randomly permute each row
+    logits = base_logits.unsqueeze(0).expand(batch_size, -1).clone()
+    for i in range(batch_size):
+        logits[i] = logits[i, torch.randperm(vocab_size, device=device)]
+
+    return logits
+
+
+def measure_memory() -> tuple[int, int]:
+    """Return (allocated, reserved) memory in bytes."""
+    torch.cuda.synchronize()
+    return torch.cuda.memory_allocated(), torch.cuda.max_memory_allocated()
+
+
+def reset_memory_stats():
+    """Reset peak memory statistics."""
+    reset_buffer_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.empty_cache()
+    gc.collect()
+
+
+def benchmark_function(
+    func,
+    logits: torch.Tensor,
+    k: torch.Tensor | None,
+    p: torch.Tensor | None,
+    warmup_iters: int = 5,
+    benchmark_iters: int = 20,
+) -> tuple[float, int]:
+    """
+    Benchmark a function and return (avg_time_ms, peak_memory_bytes).
+
+    Returns average time in milliseconds and peak memory usage.
+    """
+    # Warmup
+    for _ in range(warmup_iters):
+        logits_copy = logits.clone()
+        func(logits_copy, k, p)
+    torch.cuda.synchronize()
+
+    # Reset memory stats before benchmark
+    reset_memory_stats()
+
+    # Benchmark
+    start_events = [
+        torch.cuda.Event(enable_timing=True) for _ in range(benchmark_iters)
+    ]
+    end_events = [torch.cuda.Event(enable_timing=True) for _ in range(benchmark_iters)]
+
+    for i in range(benchmark_iters):
+        logits_copy = logits.clone()
+        start_events[i].record()
+        func(logits_copy, k, p)
+        end_events[i].record()
+
+    torch.cuda.synchronize()
+
+    # Calculate timing
+    times = [
+        start_events[i].elapsed_time(end_events[i]) for i in range(benchmark_iters)
+    ]
+    avg_time = sum(times) / len(times)
+
+    # Get peak memory
+    _, peak_memory = measure_memory()
+
+    return avg_time, peak_memory
+
+
+def create_benchmark_configs(
+    batch_sizes: list[int],
+    vocab_sizes: list[int],
+    device: str = "cuda",
+) -> list[BenchmarkConfig]:
+    """Create all benchmark configurations."""
+    configs = []
+
+    for vocab_size in vocab_sizes:
+        for batch_size in batch_sizes:
+            # 1. Top-k only - whole batch (all rows have k < vocab_size)
+            k_all = torch.full((batch_size,), 50, dtype=torch.int32, device=device)
+            configs.append(
+                BenchmarkConfig(
+                    name=f"topk_whole_b{batch_size}_v{vocab_size // 1000}k",
+                    batch_size=batch_size,
+                    vocab_size=vocab_size,
+                    k_values=k_all,
+                    p_values=None,
+                    description=f"Top-k only (whole batch, k=50), "
+                    f"batch={batch_size}, vocab={vocab_size}",
+                    ops_pct=calculate_ops_pct(k_all, None, vocab_size, batch_size),
+                )
+            )
+
+            # 2. Top-k only - partial batch (half have k=50, half have k=vocab_size)
+            k_partial = torch.full((batch_size,), 50, dtype=torch.int32, device=device)
+            k_partial[batch_size // 2 :] = vocab_size  # No filtering for second half
+            configs.append(
+                BenchmarkConfig(
+                    name=f"topk_partial_b{batch_size}_v{vocab_size // 1000}k",
+                    batch_size=batch_size,
+                    vocab_size=vocab_size,
+                    k_values=k_partial,
+                    p_values=None,
+                    description=f"Top-k only (partial batch, 50% k=50, 50% k=vocab), "
+                    f"batch={batch_size}, vocab={vocab_size}",
+                    ops_pct=calculate_ops_pct(k_partial, None, vocab_size, batch_size),
+                )
+            )
+
+            # 3. Top-p only - whole batch (all rows have p < 1.0)
+            p_all = torch.full((batch_size,), 0.9, dtype=torch.float32, device=device)
+            configs.append(
+                BenchmarkConfig(
+                    name=f"topp_whole_b{batch_size}_v{vocab_size // 1000}k",
+                    batch_size=batch_size,
+                    vocab_size=vocab_size,
+                    k_values=None,
+                    p_values=p_all,
+                    description=f"Top-p only (whole batch, p=0.9), "
+                    f"batch={batch_size}, vocab={vocab_size}",
+                    ops_pct=calculate_ops_pct(None, p_all, vocab_size, batch_size),
+                )
+            )
+
+            # 4. Top-p only - partial batch (half have p=0.9, half have p=1.0)
+            p_partial = torch.full(
+                (batch_size,), 0.9, dtype=torch.float32, device=device
+            )
+            p_partial[batch_size // 2 :] = 1.0  # No filtering for second half
+            configs.append(
+                BenchmarkConfig(
+                    name=f"topp_partial_b{batch_size}_v{vocab_size // 1000}k",
+                    batch_size=batch_size,
+                    vocab_size=vocab_size,
+                    k_values=None,
+                    p_values=p_partial,
+                    description=f"Top-p only (partial batch, 50% p=0.9, 50% p=1.0), "
+                    f"batch={batch_size}, vocab={vocab_size}",
+                    ops_pct=calculate_ops_pct(None, p_partial, vocab_size, batch_size),
+                )
+            )
+
+            # 5. Mix of top-k and top-p (both applied to whole batch)
+            k_mix = torch.full((batch_size,), 100, dtype=torch.int32, device=device)
+            p_mix = torch.full((batch_size,), 0.9, dtype=torch.float32, device=device)
+            configs.append(
+                BenchmarkConfig(
+                    name=f"topk_topp_whole_b{batch_size}_v{vocab_size // 1000}k",
+                    batch_size=batch_size,
+                    vocab_size=vocab_size,
+                    k_values=k_mix,
+                    p_values=p_mix,
+                    description=f"Top-k + Top-p (whole batch, k=100, p=0.9), "
+                    f"batch={batch_size}, vocab={vocab_size}",
+                    ops_pct=calculate_ops_pct(k_mix, p_mix, vocab_size, batch_size),
+                )
+            )
+
+            # 6. Mix with partial application (some rows k only, some p only, some both)
+            k_mixed = torch.full(
+                (batch_size,), vocab_size, dtype=torch.int32, device=device
+            )
+            p_mixed = torch.full((batch_size,), 1.0, dtype=torch.float32, device=device)
+            # First third: k only
+            third = batch_size // 3
+            k_mixed[:third] = 50
+            # Second third: p only
+            p_mixed[third : 2 * third] = 0.5
+            # Last third: both k and p
+            k_mixed[2 * third :] = 100
+            p_mixed[2 * third :] = 0.9
+            configs.append(
+                BenchmarkConfig(
+                    name=f"mixed_partial_b{batch_size}_v{vocab_size // 1000}k",
+                    batch_size=batch_size,
+                    vocab_size=vocab_size,
+                    k_values=k_mixed,
+                    p_values=p_mixed,
+                    description=f"Mixed partial (1/3 k=50, 1/3 p=0.9, 1/3 both), "
+                    f"batch={batch_size}, vocab={vocab_size}",
+                    ops_pct=calculate_ops_pct(k_mixed, p_mixed, vocab_size, batch_size),
+                )
+            )
+
+    return configs
+
+
+def format_memory(bytes_val: int) -> str:
+    """Format memory in human-readable form."""
+    if bytes_val >= 1024**3:
+        return f"{bytes_val / (1024**3):.2f} GB"
+    elif bytes_val >= 1024**2:
+        return f"{bytes_val / (1024**2):.2f} MB"
+    elif bytes_val >= 1024:
+        return f"{bytes_val / 1024:.2f} KB"
+    return f"{bytes_val} B"
+
+
+def run_benchmark(
+    configs: list[BenchmarkConfig],
+    warmup_iters: int = 5,
+    benchmark_iters: int = 20,
+    verbose: bool = True,
+):
+    """Run all benchmarks and print results."""
+    results = []
+
+    print("=" * 100)
+    print("Top-k/Top-p Benchmark: Triton vs PyTorch Sort-based")
+    print("=" * 100)
+    print()
+
+    for config in configs:
+        if verbose:
+            print(f"Running: {config.description}")
+
+        # Create fresh logits for this config
+        logits = create_logits(config.batch_size, config.vocab_size)
+
+        # Benchmark Triton
+        reset_memory_stats()
+        triton_time, triton_mem = benchmark_function(
+            apply_top_k_top_p_triton,
+            logits,
+            config.k_values,
+            config.p_values,
+            warmup_iters,
+            benchmark_iters,
+        )
+
+        # Benchmark PyTorch
+        reset_memory_stats()
+        pytorch_time, pytorch_mem = benchmark_function(
+            apply_top_k_top_p_pytorch,
+            logits,
+            config.k_values,
+            config.p_values,
+            warmup_iters,
+            benchmark_iters,
+        )
+
+        speedup = pytorch_time / triton_time if triton_time > 0 else float("inf")
+        mem_ratio = pytorch_mem / triton_mem if triton_mem > 0 else float("inf")
+
+        result = {
+            "config": config,
+            "triton_time_ms": triton_time,
+            "pytorch_time_ms": pytorch_time,
+            "triton_mem": triton_mem,
+            "pytorch_mem": pytorch_mem,
+            "speedup": speedup,
+            "mem_ratio": mem_ratio,
+        }
+        results.append(result)
+
+        if verbose:
+            print(f"  Triton:  {triton_time:.3f} ms, {format_memory(triton_mem)}")
+            print(f"  PyTorch: {pytorch_time:.3f} ms, {format_memory(pytorch_mem)}")
+            print(f"  Speedup: {speedup:.2f}x, Memory ratio: {mem_ratio:.2f}x")
+            print()
+
+        # Clean up
+        del logits
+        reset_memory_stats()
+
+    return results
+
+
+def print_summary_table(results: list[dict]):
+    """Print a summary table of results."""
+    print()
+    print("=" * 130)
+    print("SUMMARY TABLE")
+    print("=" * 130)
+    print()
+
+    # Header
+    header = (
+        f"{'Scenario':<40} {'Batch':>6} {'Vocab':>7} {'Ops%':>6} "
+        f"{'Triton (ms)':>12} {'PyTorch (ms)':>13} {'Speedup':>8} "
+        f"{'Tri Mem':>10} {'Pyt Mem':>10}"
+    )
+    print(header)
+    print("-" * 130)
+
+    # Group by scenario type
+    current_vocab = None
+    for result in results:
+        config = result["config"]
+
+        # Add separator between vocab sizes
+        if current_vocab != config.vocab_size:
+            if current_vocab is not None:
+                print("-" * 130)
+            current_vocab = config.vocab_size
+
+        scenario = config.name.split("_b")[0]  # Extract scenario name
+        print(
+            f"{scenario:<40} {config.batch_size:>6} {config.vocab_size:>7} "
+            f"{config.ops_pct:>5.0f}% "
+            f"{result['triton_time_ms']:>12.3f} {result['pytorch_time_ms']:>13.3f} "
+            f"{result['speedup']:>7.2f}x "
+            f"{format_memory(result['triton_mem']):>10} "
+            f"{format_memory(result['pytorch_mem']):>10}"
+        )
+
+    print("=" * 130)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Benchmark Triton vs PyTorch sort-based top-k/top-p implementations"
+    )
+    parser.add_argument(
+        "--batch-sizes",
+        type=int,
+        nargs="+",
+        default=[1, 4, 16, 64, 128, 512, 1024, 2048],
+        help="Batch sizes to test (default: 1 4 16 64)",
+    )
+    parser.add_argument(
+        "--vocab-sizes",
+        type=int,
+        nargs="+",
+        default=[32768, 131072],  # 32k, 128k
+        help="Vocabulary sizes to test (default: 32768 131072)",
+    )
+    parser.add_argument(
+        "--warmup-iters",
+        type=int,
+        default=5,
+        help="Number of warmup iterations (default: 5)",
+    )
+    parser.add_argument(
+        "--benchmark-iters",
+        type=int,
+        default=20,
+        help="Number of benchmark iterations (default: 20)",
+    )
+    parser.add_argument(
+        "--quiet",
+        action="store_true",
+        help="Only print summary table",
+    )
+
+    args = parser.parse_args()
+
+    # Print configuration
+    print(f"Batch sizes: {args.batch_sizes}")
+    print(f"Vocab sizes: {args.vocab_sizes}")
+    print(f"Warmup iterations: {args.warmup_iters}")
+    print(f"Benchmark iterations: {args.benchmark_iters}")
+    print()
+
+    # Check CUDA
+    if not torch.cuda.is_available():
+        print("ERROR: CUDA is not available. This benchmark requires a GPU.")
+        return
+
+    device_name = torch.cuda.get_device_name(0)
+    print(f"GPU: {device_name}")
+    print()
+
+    # Create configs
+    configs = create_benchmark_configs(
+        args.batch_sizes,
+        args.vocab_sizes,
+    )
+
+    # Run benchmarks
+    results = run_benchmark(
+        configs,
+        warmup_iters=args.warmup_iters,
+        benchmark_iters=args.benchmark_iters,
+        verbose=not args.quiet,
+    )
+
+    # Print summary
+    print_summary_table(results)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/kernels/benchmark_moe.py

@@ -100,13 +100,38 @@ def benchmark_config(
     dtype: torch.dtype,
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
+    use_int4_w4a16: bool = False,
     num_iters: int = 100,
     block_quant_shape: list[int] = None,
     use_deep_gemm: bool = False,
 ) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
-    if use_int8_w8a16:
+    if use_int4_w4a16:
+        # Int4 packed weights: 2 int4 values per uint8 byte
+        # K dimension is packed (halved)
+        intermediate_size = shard_intermediate_size // 2  # after silu_and_mul
+        w1 = torch.randint(
+            0,
+            255,
+            (
+                num_experts,
+                shard_intermediate_size,
+                hidden_size // 2,  # int4 packing
+            ),
+            dtype=torch.uint8,
+        )
+        w2 = torch.randint(
+            0,
+            255,
+            (
+                num_experts,
+                hidden_size,
+                intermediate_size // 2,  # int4 packing
+            ),
+            dtype=torch.uint8,
+        )
+    elif use_int8_w8a16:
         w1 = torch.randint(
             -127,
             127,
@@ -140,7 +165,20 @@ def benchmark_config(
     w2_scale = None
     a1_scale = None
     a2_scale = None
-    if use_int8_w8a16:
+    if use_int4_w4a16:
+        if block_quant_shape is None:
+            raise ValueError("block_quant_shape is required for int4_w4a16")
+        group_size = block_quant_shape[1]
+        # Scales shape: (E, N, K // group_size) in fp16
+        w1_scale = torch.rand(
+            (num_experts, shard_intermediate_size, hidden_size // group_size),
+            dtype=dtype,
+        )
+        w2_scale = torch.rand(
+            (num_experts, hidden_size, intermediate_size // group_size),
+            dtype=dtype,
+        )
+    elif use_int8_w8a16:
         w1_scale = torch.randn(
             (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
         )
@@ -199,6 +237,7 @@ def benchmark_config(
             a1_scale=a1_scale,
             a2_scale=a2_scale,
             block_shape=block_quant_shape,
+            weight_dtype="int4" if use_int4_w4a16 else None,
         )
 
         deep_gemm_experts = None
@@ -481,6 +520,7 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
+        use_int4_w4a16: bool = False,
         block_quant_shape: list[int] = None,
         use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
@@ -488,7 +528,10 @@ class BenchmarkWorker:
 
         set_random_seed(self.seed)
         dtype_str = _get_config_dtype_str(
-            dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
+            dtype,
+            use_int8_w8a16=use_int8_w8a16,
+            use_fp8_w8a8=use_fp8_w8a8,
+            use_int4_w4a16=use_int4_w4a16,
         )
         # NOTE(woosuk): The current naming convention uses w2.shape[2], which
         # is the intermediate size after silu_and_mul.
@@ -519,6 +562,7 @@ class BenchmarkWorker:
             dtype,
             use_fp8_w8a8,
             use_int8_w8a16,
+            use_int4_w4a16=use_int4_w4a16,
             num_iters=100,
             block_quant_shape=block_quant_shape,
             use_deep_gemm=use_deep_gemm,
@@ -535,6 +579,7 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
+        use_int4_w4a16: bool,
         search_space: list[dict[str, int]],
         block_quant_shape: list[int],
         use_deep_gemm: bool,
@@ -545,7 +590,7 @@ class BenchmarkWorker:
         best_config = None
         best_time = float("inf")
         if current_platform.is_rocm():
-            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16 or use_int4_w4a16)
             search_space = prune_rocm_search_space(
                 num_tokens,
                 shard_intermediate_size,
@@ -574,6 +619,7 @@ class BenchmarkWorker:
                         dtype,
                         use_fp8_w8a8,
                         use_int8_w8a16,
+                        use_int4_w4a16,
                         num_iters=20,
                         block_quant_shape=block_quant_shape,
                         use_deep_gemm=use_deep_gemm,
@@ -621,6 +667,7 @@ def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
             else {}
         ),
         **({"kpack": config["kpack"]} if "kpack" in config else {}),
+        **({"SPLIT_K": config["SPLIT_K"]} if "SPLIT_K" in config else {}),
     }
 
 
@@ -633,11 +680,15 @@ def save_configs(
     dtype: torch.dtype,
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
+    use_int4_w4a16: bool,
     block_quant_shape: list[int],
     save_dir: str,
 ) -> None:
     dtype_str = _get_config_dtype_str(
-        dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
+        dtype,
+        use_int8_w8a16=use_int8_w8a16,
+        use_fp8_w8a8=use_fp8_w8a8,
+        use_int4_w4a16=use_int4_w4a16,
     )
 
     # NOTE(woosuk): The current naming convention uses w2.shape[2], which
@@ -739,6 +790,38 @@ def get_model_params(config):
     return E, topk, intermediate_size, hidden_size
 
 
+def get_quantization_group_size(config) -> int | None:
+    """Extract the quantization group size from the HF model config.
+
+    This reads directly from the HuggingFace config object (as returned by
+    ``get_config()``), not from vLLM's quantization config classes.
+
+    Supports AWQ/GPTQ-style configs (direct 'group_size' key) and
+    compressed-tensors configs (nested inside 'config_groups').
+    """
+    quantization_config = getattr(config, "quantization_config", {})
+    if not isinstance(quantization_config, dict):
+        return None
+    # AWQ / GPTQ style: group_size is a top-level key
+    gs = quantization_config.get("group_size")
+    if gs is not None:
+        return gs
+    # compressed-tensors style: group_size is nested in config_groups
+    config_groups = quantization_config.get("config_groups", {})
+    if not isinstance(config_groups, dict):
+        return None
+    for group_cfg in config_groups.values():
+        if not isinstance(group_cfg, dict):
+            continue
+        weights = group_cfg.get("weights", {})
+        if not isinstance(weights, dict):
+            continue
+        gs = weights.get("group_size")
+        if gs is not None:
+            return gs
+    return None
+
+
 def main(args: argparse.Namespace):
     print(args)
 
@@ -757,7 +840,20 @@ 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_int4_w4a16 = args.dtype == "int4_w4a16"
     block_quant_shape = get_weight_block_size_safety(config)
+    if use_int4_w4a16:
+        group_size = get_quantization_group_size(config)
+        if group_size is None:
+            raise ValueError(
+                "Could not determine group_size from model config. "
+                "The model's quantization_config must contain a 'group_size' "
+                "field (AWQ/GPTQ) or 'config_groups.*.weights.group_size' "
+                "(compressed-tensors)."
+            )
+        # For int4_w4a16, block_shape = [0, group_size]
+        # block_shape[0]=0 means no block quantization on N dimension
+        block_quant_shape = [0, group_size]
 
     if args.batch_size is None:
         batch_sizes = [
@@ -811,8 +907,20 @@ def main(args: argparse.Namespace):
         return ray.get(outputs)
 
     if args.tune:
-        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
-        search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
+        # int4_w4a16 weights are uint8-packed, not fp16; treat like fp8 for
+        # search space generation (no matrix_instr_nonkdim/kpack exploration).
+        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16 or use_int4_w4a16)
+        # For int4_w4a16, the group_size constraint on BLOCK_SIZE_K does not
+        # apply: the gptq_awq kernel handles arbitrary BLOCK_SIZE_K regardless
+        # of group_size. Skip block_quant_shape filtering to keep the full
+        # search space (e.g. BLOCK_SIZE_K=64 with group_size=128).
+        tune_block_quant_shape = None if use_int4_w4a16 else block_quant_shape
+        search_space = get_configs_compute_bound(is_fp16, tune_block_quant_shape)
+        if use_int4_w4a16:
+            # SPLIT_K is a required kernel constexpr for gptq_awq kernel;
+            # only SPLIT_K=1 is used at runtime, so fix it during tuning.
+            for cfg in search_space:
+                cfg["SPLIT_K"] = 1
         print(f"Start tuning over {len(search_space)} configurations...")
         if use_deep_gemm:
             raise ValueError(
@@ -832,6 +940,7 @@ def main(args: argparse.Namespace):
                     dtype,
                     use_fp8_w8a8,
                     use_int8_w8a16,
+                    use_int4_w4a16,
                     search_space,
                     block_quant_shape,
                     use_deep_gemm,
@@ -851,6 +960,7 @@ def main(args: argparse.Namespace):
             dtype,
             use_fp8_w8a8,
             use_int8_w8a16,
+            use_int4_w4a16,
             block_quant_shape,
             args.save_dir,
         )
@@ -869,6 +979,7 @@ def main(args: argparse.Namespace):
                     dtype,
                     use_fp8_w8a8,
                     use_int8_w8a16,
+                    use_int4_w4a16,
                     block_quant_shape,
                     use_deep_gemm,
                 )
@@ -891,7 +1002,10 @@ if __name__ == "__main__":
     )
     parser.add_argument("--enable-expert-parallel", "-enable-ep", action="store_true")
     parser.add_argument(
-        "--dtype", type=str, choices=["auto", "fp8_w8a8", "int8_w8a16"], default="auto"
+        "--dtype",
+        type=str,
+        choices=["auto", "fp8_w8a8", "int8_w8a16", "int4_w4a16"],
+        default="auto",
     )
     parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument(

benchmarks/run_structured_output_benchmark.sh

@@ -71,7 +71,7 @@ while [[ $# -gt 0 ]]; do
       usage
       ;;
     *)
-      echo "Unknown argument: $1\n"
+      printf "Unknown argument: %s\n" "$1"
       usage
       ;;
   esac
@@ -84,15 +84,17 @@ mkdir -p "$OUTPUT_DIR"
 QPS_VALUES=(25 20 15 10 5 1)
 
 # Common parameters
-COMMON_PARAMS="--backend $BACKEND \
-               --model $MODEL \
-               --dataset $DATASET \
-               --structured-output-ratio $STRUCTURED_OUTPUT_RATIO \
-               --save-results \
-               --result-dir $OUTPUT_DIR \
-               --output-len $MAX_NEW_TOKENS \
-               --port $PORT \
-               --tokenizer-mode $TOKENIZER_MODE"
+COMMON_PARAMS=(
+  --backend "$BACKEND"
+  --model "$MODEL"
+  --dataset "$DATASET"
+  --structured-output-ratio "$STRUCTURED_OUTPUT_RATIO"
+  --save-results
+  --result-dir "$OUTPUT_DIR"
+  --output-len "$MAX_NEW_TOKENS"
+  --port "$PORT"
+  --tokenizer-mode "$TOKENIZER_MODE"
+)
 
 echo "Starting structured output benchmark with model: $MODEL"
 echo "Backend: $BACKEND"
@@ -109,17 +111,17 @@ for qps in "${QPS_VALUES[@]}"; do
   GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown")
 
   # Construct filename for this run
-  FILENAME="${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json"
+  FILENAME="${BACKEND}_${qps}qps_$(basename "$MODEL")_${DATASET}_${GIT_HASH}_${GIT_BRANCH}.json"
 
   NUM_PROMPTS=$(echo "$TOTAL_SECONDS * $qps" | bc)
   NUM_PROMPTS=${NUM_PROMPTS%.*}  # Remove fractional part
   echo "Running benchmark with $NUM_PROMPTS prompts"
 
   # Run the benchmark
-  python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \
-    --request-rate $qps \
+  python "$SCRIPT_DIR/benchmark_serving_structured_output.py" "${COMMON_PARAMS[@]}" \
+    --request-rate "$qps" \
     --result-filename "$FILENAME" \
-    --num-prompts $NUM_PROMPTS
+    --num-prompts "$NUM_PROMPTS"
 
   echo "Completed benchmark with QPS: $qps"
   echo "----------------------------------------"

cmake/cpu_extension.cmake

@@ -18,6 +18,7 @@ set(ENABLE_AVX512 $ENV{VLLM_CPU_AVX512})
 set(ENABLE_AVX512BF16 $ENV{VLLM_CPU_AVX512BF16})
 set(ENABLE_AVX512VNNI $ENV{VLLM_CPU_AVX512VNNI})
 set(ENABLE_AMXBF16 $ENV{VLLM_CPU_AMXBF16})
+set(ENABLE_ARM_BF16 $ENV{VLLM_CPU_ARM_BF16})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
 
@@ -115,6 +116,10 @@ else()
         set(AVX512_FOUND ON)
         message(STATUS "AVX512 support enabled via VLLM_CPU_AVX512 environment variable")
     endif()
+    if (ENABLE_ARM_BF16)
+        set(ARM_BF16_FOUND ON)
+        message(STATUS "ARM BF16 support enabled via VLLM_CPU_ARM_BF16 environment variable")
+    endif()
 endif()
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)

cmake/external_projects/flashmla.cmake

@@ -19,7 +19,7 @@ else()
   FetchContent_Declare(
         flashmla
         GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG c2afa9cb93e674d5a9120a170a6da57b89267208
+        GIT_TAG 692917b1cda61b93ac9ee2d846ec54e75afe87b1
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""

csrc/cuda_view.cu

@@ -2,33 +2,58 @@
 #include <torch/cuda.h>
 #include <cuda_runtime.h>
 
-// This function assumes that `cpu_tensor` is a CPU tensor allocated with pinned
-// memory, and that UVA (Unified Virtual Addressing) is enabled.
+// This function assumes that `cpu_tensor` is a CPU tensor,
+// and that UVA (Unified Virtual Addressing) is enabled.
 torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor) {
   TORCH_CHECK(cpu_tensor.device().is_cpu(), "Input tensor must be on CPU");
 
-  // Get raw host pointer from CPU tensor
-  void* host_ptr = cpu_tensor.data_ptr();
+  // handle empty tensor
+  if (cpu_tensor.numel() == 0) {
+    return torch::empty(cpu_tensor.sizes(),
+                        cpu_tensor.options().device(torch::kCUDA));
+  }
+
+  if (cpu_tensor.is_pinned()) {
+    // If CPU tensor is pinned, directly get the device pointer.
+    void* host_ptr = const_cast<void*>(cpu_tensor.data_ptr());
+    void* device_ptr = nullptr;
+    cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+    TORCH_CHECK(err == cudaSuccess,
+                "cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+
+    return torch::from_blob(
+        device_ptr, cpu_tensor.sizes(), cpu_tensor.strides(),
+        [base = cpu_tensor](void*) {},  // keep cpu tensor alive
+        cpu_tensor.options().device(torch::kCUDA));
+  }
+
+  // If CPU tensor is not pinned, allocate a new pinned memory buffer.
+  torch::Tensor contiguous_cpu = cpu_tensor.contiguous();
+  size_t nbytes = contiguous_cpu.nbytes();
+
+  void* host_ptr = nullptr;
+  cudaError_t err = cudaHostAlloc(&host_ptr, nbytes, cudaHostAllocMapped);
+  if (err != cudaSuccess) {
+    AT_ERROR("cudaHostAlloc failed: ", cudaGetErrorString(err));
+  }
+
+  err = cudaMemcpy(host_ptr, contiguous_cpu.data_ptr(), nbytes,
+                   cudaMemcpyDefault);
+  if (err != cudaSuccess) {
+    cudaFreeHost(host_ptr);
+    AT_ERROR("cudaMemcpy failed: ", cudaGetErrorString(err));
+  }
 
-  // Get a device pointer corresponding to the pinned host memory
   void* device_ptr = nullptr;
-  cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
-  TORCH_CHECK(err == cudaSuccess,
-              "cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+  err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+  if (err != cudaSuccess) {
+    cudaFreeHost(host_ptr);
+    AT_ERROR("cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+  }
 
-  // We'll use the same sizes, strides, and dtype as the CPU tensor.
-  // TODO: check if layout is respected.
-  auto sizes = cpu_tensor.sizes();
-  auto strides = cpu_tensor.strides();
-  auto options = cpu_tensor.options().device(torch::kCUDA);
+  auto deleter = [host_ptr](void*) { cudaFreeHost(host_ptr); };
 
-  // use default no-op deleter, since the memory is owned by the original CPU
-  // tensor
-  torch::Tensor cuda_tensor =
-      torch::from_blob(device_ptr, sizes, strides, options);
-
-  TORCH_CHECK(cuda_tensor.device().is_cuda(),
-              "Resulting tensor is not on CUDA device");
-
-  return cuda_tensor;
-}
+  return torch::from_blob(device_ptr, contiguous_cpu.sizes(),
+                          contiguous_cpu.strides(), deleter,
+                          contiguous_cpu.options().device(torch::kCUDA));
+}

csrc/rocm/skinny_gemms.cu

@@ -1568,8 +1568,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         {
   #endif
           unsigned int kOff = k + (thrd * A_CHUNK);
-          unsigned int kOffcp =
-              k_str + kOff;  // min__(K - A_CHUNK, k_str + kOff);
+          unsigned int kOffcp = min__(K - A_CHUNK, k_str + kOff);
           for (unsigned int n = 0; n < N; n += CHUNKK * sprdN) {
             __builtin_amdgcn_global_load_lds(
                 (int*)(&A[min__(

csrc/topk.cu

@@ -349,7 +349,7 @@ void setup_kernel_smem_once() {
 void large_context_topk(
     const torch::Tensor& logits, torch::Tensor& indices,
     const torch::Tensor& seq_lens,
-    c10::optional<torch::Tensor> row_starts = c10::nullopt) {
+    std::optional<torch::Tensor> row_starts = std::nullopt) {
   TORCH_CHECK(logits.is_cuda(), "logits must be a CUDA tensor");
   TORCH_CHECK(indices.is_cuda(), "indices must be a CUDA tensor");
   TORCH_CHECK(seq_lens.is_cuda(), "seq_lens must be a CUDA tensor");

docker/Dockerfile

@@ -22,7 +22,12 @@
 #   docker buildx bake -f docker/docker-bake.hcl -f docker/versions.json
 # =============================================================================
 
-ARG CUDA_VERSION=12.9.1
+ARG CUDA_VERSION=12.8.1
+# BUILDER_CUDA_VERSION controls the CUDA toolkit used to compile csrc/ and
+# extensions (DeepGEMM, EP kernels).  It can differ from CUDA_VERSION, which
+# is the CUDA version shipped in the final runtime image and used to select
+# the matching PyTorch wheel.
+ARG BUILDER_CUDA_VERSION=12.9.1
 ARG PYTHON_VERSION=3.12
 
 # By parameterizing the base images, we allow third-party to use their own
@@ -36,7 +41,7 @@ ARG PYTHON_VERSION=3.12
 # compatibility with other Linux OSes. The main reason for this is that the
 # glibc version is baked into the distro, and binaries built with one glibc
 # version are not backwards compatible with OSes that use an earlier version.
-ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
+ARG BUILD_BASE_IMAGE=nvidia/cuda:${BUILDER_CUDA_VERSION}-devel-ubuntu20.04
 # Using cuda base image with minimal dependencies necessary for JIT compilation (FlashInfer, DeepGEMM, EP kernels)
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04
 
@@ -92,8 +97,13 @@ ARG INSTALL_KV_CONNECTORS=false
 FROM ${BUILD_BASE_IMAGE} AS base
 
 ARG CUDA_VERSION
+ARG BUILDER_CUDA_VERSION
 ARG PYTHON_VERSION
 
+# Override the CUDA_VERSION env var inherited from the nvidia base image
+# (which equals BUILDER_CUDA_VERSION) so that $CUDA_VERSION in RUN commands
+# resolves to the runtime CUDA version used for PyTorch wheel selection.
+ENV CUDA_VERSION=${CUDA_VERSION}
 ENV DEBIAN_FRONTEND=noninteractive
 
 # Install system dependencies including build tools
@@ -133,7 +143,7 @@ ENV UV_LINK_MODE=copy
 RUN gcc --version
 
 # Ensure CUDA compatibility library is loaded
-RUN echo "/usr/local/cuda-$(echo "$CUDA_VERSION" | cut -d. -f1,2)/compat/" > /etc/ld.so.conf.d/cuda-compat.conf && ldconfig
+RUN echo "/usr/local/cuda-$(echo "$BUILDER_CUDA_VERSION" | cut -d. -f1,2)/compat/" > /etc/ld.so.conf.d/cuda-compat.conf && ldconfig
 
 # ============================================================
 # SLOW-CHANGING DEPENDENCIES BELOW
@@ -309,7 +319,7 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
 # Build DeepGEMM, pplx-kernels, DeepEP - runs in PARALLEL with csrc-build
 # This stage is independent and doesn't affect csrc cache
 FROM base AS extensions-build
-ARG CUDA_VERSION
+ARG BUILDER_CUDA_VERSION
 
 # This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
 ENV UV_HTTP_TIMEOUT=500
@@ -325,7 +335,7 @@ COPY tools/install_deepgemm.sh /tmp/install_deepgemm.sh
 RUN --mount=type=cache,target=/root/.cache/uv \
     mkdir -p /tmp/deepgemm/dist && \
     VLLM_DOCKER_BUILD_CONTEXT=1 TORCH_CUDA_ARCH_LIST="9.0a 10.0a" /tmp/install_deepgemm.sh \
-        --cuda-version "${CUDA_VERSION}" \
+        --cuda-version "${BUILDER_CUDA_VERSION}" \
         ${DEEPGEMM_GIT_REF:+--ref "$DEEPGEMM_GIT_REF"} \
         --wheel-dir /tmp/deepgemm/dist || \
     echo "DeepGEMM build skipped (CUDA version requirement not met)"

docker/Dockerfile.cpu

@@ -20,6 +20,7 @@
 #   VLLM_CPU_AVX512BF16=false (default)|true (for cross-compilation)
 #   VLLM_CPU_AVX512VNNI=false (default)|true (for cross-compilation)
 #   VLLM_CPU_AMXBF16=false (default)|true (for cross-compilation)
+#   VLLM_CPU_ARM_BF16=false (default)|true (for cross-compilation)
 #
 
 ######################### COMMON BASE IMAGE #########################
@@ -108,9 +109,22 @@ ENV VLLM_CPU_AVX512VNNI=${VLLM_CPU_AVX512VNNI}
 # Support for building with AMXBF16 ISA: docker build --build-arg VLLM_CPU_AMXBF16="true" ...
 ARG VLLM_CPU_AMXBF16=1
 ENV VLLM_CPU_AMXBF16=${VLLM_CPU_AMXBF16}
+# Support for cross-compilation with ARM BF16 ISA: docker build --build-arg VLLM_CPU_ARM_BF16="true" ...
+ARG VLLM_CPU_ARM_BF16=0
+ENV VLLM_CPU_ARM_BF16=${VLLM_CPU_ARM_BF16}
 
 WORKDIR /vllm-workspace
 
+# Validate build arguments - prevent mixing incompatible ISA flags
+RUN if [ "$TARGETARCH" = "arm64" ] && { [ "$VLLM_CPU_AVX2" != "0" ] || [ "$VLLM_CPU_AVX512" != "0" ] || [ "$VLLM_CPU_AVX512BF16" != "0" ] || [ "$VLLM_CPU_AVX512VNNI" != "0" ]; }; then \
+        echo "ERROR: Cannot use x86-specific ISA flags (AVX2, AVX512, etc.) when building for ARM64 (--platform=linux/arm64)"; \
+        exit 1; \
+    fi && \
+    if [ "$TARGETARCH" = "amd64" ] && [ "$VLLM_CPU_ARM_BF16" != "0" ]; then \
+        echo "ERROR: Cannot use ARM-specific ISA flags (ARM_BF16) when building for x86_64 (--platform=linux/amd64)"; \
+        exit 1; \
+    fi
+
 # Copy build requirements
 COPY requirements/cpu-build.txt requirements/build.txt
 
@@ -224,6 +238,7 @@ ARG VLLM_CPU_AVX512
 ARG VLLM_CPU_AVX512BF16
 ARG VLLM_CPU_AVX512VNNI
 ARG VLLM_CPU_AMXBF16
+ARG VLLM_CPU_ARM_BF16
 ARG PYTHON_VERSION
 
 LABEL ai.vllm.build.target-arch="${TARGETARCH}"
@@ -233,6 +248,7 @@ LABEL ai.vllm.build.cpu-avx512="${VLLM_CPU_AVX512:-false}"
 LABEL ai.vllm.build.cpu-avx512bf16="${VLLM_CPU_AVX512BF16:-false}"
 LABEL ai.vllm.build.cpu-avx512vnni="${VLLM_CPU_AVX512VNNI:-false}"
 LABEL ai.vllm.build.cpu-amxbf16="${VLLM_CPU_AMXBF16:-false}"
+LABEL ai.vllm.build.cpu-arm-bf16="${VLLM_CPU_ARM_BF16:-false}"
 LABEL ai.vllm.build.python-version="${PYTHON_VERSION:-3.12}"
 
 ENTRYPOINT ["vllm", "serve"]

docker/versions.json

@@ -2,6 +2,9 @@
   "_comment": "Auto-generated from Dockerfile ARGs. Do not edit manually. Run: python tools/generate_versions_json.py",
   "variable": {
     "CUDA_VERSION": {
+      "default": "12.8.1"
+    },
+    "BUILDER_CUDA_VERSION": {
       "default": "12.9.1"
     },
     "PYTHON_VERSION": {
@@ -11,7 +14,7 @@
       "default": "nvidia/cuda:12.9.1-devel-ubuntu20.04"
     },
     "FINAL_BASE_IMAGE": {
-      "default": "nvidia/cuda:12.9.1-base-ubuntu22.04"
+      "default": "nvidia/cuda:12.8.1-base-ubuntu22.04"
     },
     "GET_PIP_URL": {
       "default": "https://bootstrap.pypa.io/get-pip.py"

docs/design/cuda_graphs.md

@@ -182,6 +182,7 @@ The following table lists backends that support full CUDA Graphs at the time of
 | FlashInfer | `UNIFORM_SINGLE_TOKEN_DECODE` | Will be set to `UNIFORM_BATCH` when using TRTLLM attention on Blackwell |
 | FlashMLA | `UNIFORM_BATCH` | |
 | FlashInferMLA | `UNIFORM_BATCH` | |
+| FlashInferMLASparse | `UNIFORM_BATCH` | |
 | AITER MLA | `UNIFORM_SINGLE_TOKEN_DECODE` | |
 | CUTLASS MLA | `UNIFORM_SINGLE_TOKEN_DECODE` | |
 | Mamba attention| `UNIFORM_SINGLE_TOKEN_DECODE` | |

docs/features/batch_invariance.md

@@ -109,6 +109,7 @@ Batch invariance has been tested and verified on the following models:
 - **Qwen2.5**: `Qwen/Qwen2.5-0.5B-Instruct`, `Qwen/Qwen2.5-1.5B-Instruct`, `Qwen/Qwen2.5-3B-Instruct`, `Qwen/Qwen2.5-7B-Instruct`, `Qwen/Qwen2.5-14B-Instruct`, `Qwen/Qwen2.5-32B-Instruct`
 - **Llama 3**: `meta-llama/Llama-3.1-8B-Instruct`, `meta-llama/Llama-3.2-1B-Instruct`
 - **GPT-OSS**: `openai/gpt-oss-20b`, `openai/gpt-oss-120b`
+- **Mistral**: `mistralai/Mistral-7B-v0.3`
 
 Other models may also work, but these have been explicitly validated. If you encounter issues with a specific model, please report them on the [GitHub issue tracker](https://github.com/vllm-project/vllm/issues/new/choose).
 

docs/features/quantization/fp8.md

@@ -84,7 +84,7 @@ Since simple RTN does not require data for weight quantization and the activatio
 Install `vllm` and `lm-evaluation-harness` for evaluation:
 
 ```bash
-pip install vllm "lm-eval[api]>=0.4.9.2"
+pip install vllm "lm-eval[api]>=0.4.11"
 ```
 
 Load and run the model in `vllm`:

docs/features/quantization/int4.md

@@ -18,7 +18,7 @@ pip install llmcompressor
 Additionally, install `vllm` and `lm-evaluation-harness` for evaluation:
 
 ```bash
-pip install vllm "lm-eval[api]>=0.4.9.2"
+pip install vllm "lm-eval[api]>=0.4.11"
 ```
 
 ## Quantization Process

docs/features/quantization/int8.md

@@ -23,7 +23,7 @@ pip install llmcompressor
 Additionally, install `vllm` and `lm-evaluation-harness` for evaluation:
 
 ```bash
-pip install vllm "lm-eval[api]>=0.4.9.2"
+pip install vllm "lm-eval[api]>=0.4.11"
 ```
 
 ## Quantization Process

docs/features/quantization/quark.md

@@ -20,7 +20,7 @@ for more installation details.
 Additionally, install `vllm` and `lm-evaluation-harness` for evaluation:
 
 ```bash
-pip install vllm "lm-eval[api]>=0.4.9.2"
+pip install vllm "lm-eval[api]>=0.4.11"
 ```
 
 ## Quantization Process

docs/getting_started/installation/cpu.arm.inc.md

@@ -172,25 +172,78 @@ docker pull public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:${VLLM_COMMIT}-arm64-
 
 # --8<-- [end:pre-built-images]
 # --8<-- [start:build-image-from-source]
+
+## Building for your target ARM CPU
+
 ```bash
 docker build -f docker/Dockerfile.cpu \
-        --tag vllm-cpu-env .
+        --platform=linux/arm64 \
+        --build-arg VLLM_CPU_ARM_BF16=<false (default)|true> \
+        --tag vllm-cpu-env \
+        --target vllm-openai .
+```
 
-# Launching OpenAI server
+!!! note "Auto-detection by default"
+    By default, ARM CPU instruction sets (BF16, NEON, etc.) are automatically detected from the build system's CPU flags. The `VLLM_CPU_ARM_BF16` build argument is used for cross-compilation:
+
+    - `VLLM_CPU_ARM_BF16=true` - Force-enable ARM BF16 support (build with BF16 regardless of build system capabilities)
+    - `VLLM_CPU_ARM_BF16=false` - Rely on auto-detection (default)
+
+### Examples
+
+**Auto-detection build (native ARM)**
+
+```bash
+# Building on ARM64 system - platform auto-detected
+docker build -f docker/Dockerfile.cpu \
+        --tag vllm-cpu-arm64 \
+        --target vllm-openai .
+```
+
+**Cross-compile for ARM with BF16 support**
+
+```bash
+# Building on ARM64 for newer ARM CPUs with BF16
+docker build -f docker/Dockerfile.cpu \
+        --build-arg VLLM_CPU_ARM_BF16=true \
+        --tag vllm-cpu-arm64-bf16 \
+        --target vllm-openai .
+```
+
+**Cross-compile from x86_64 to ARM64 with BF16**
+
+```bash
+# Requires Docker buildx with ARM emulation (QEMU)
+docker buildx build -f docker/Dockerfile.cpu \
+        --platform=linux/arm64 \
+        --build-arg VLLM_CPU_ARM_BF16=true \
+        --build-arg max_jobs=4 \
+        --tag vllm-cpu-arm64-bf16 \
+        --target vllm-openai \
+        --load .
+```
+
+!!! note "ARM BF16 requirements"
+    ARM BF16 support requires ARMv8.6-A or later (FEAT_BF16). Supported on AWS Graviton3/4, AmpereOne, and other recent ARM processors.
+
+## Launching the OpenAI server
+
+```bash
 docker run --rm \
-            --privileged=true \
+            --security-opt seccomp=unconfined \
+            --cap-add SYS_NICE \
             --shm-size=4g \
             -p 8000:8000 \
             -e VLLM_CPU_KVCACHE_SPACE=<KV cache space> \
             -e VLLM_CPU_OMP_THREADS_BIND=<CPU cores for inference> \
-            vllm-cpu-env \
-            --model=meta-llama/Llama-3.2-1B-Instruct \
+            vllm-cpu-arm64 \
+            meta-llama/Llama-3.2-1B-Instruct \
             --dtype=bfloat16 \
             other vLLM OpenAI server arguments
 ```
 
-!!! tip
-    An alternative of `--privileged=true` is `--cap-add SYS_NICE --security-opt seccomp=unconfined`.
+!!! tip "Alternative to --privileged"
+    Instead of `--privileged=true`, use `--cap-add SYS_NICE --security-opt seccomp=unconfined` for better security.
 
 # --8<-- [end:build-image-from-source]
 # --8<-- [start:extra-information]

docs/models/pooling_models.md

@@ -311,20 +311,31 @@ An OpenAI client example can be found here: [examples/pooling/embed/openai_embed
 
 [ColBERT](https://arxiv.org/abs/2004.12832) (Contextualized Late Interaction over BERT) is a retrieval model that uses per-token embeddings and MaxSim scoring for document ranking. Unlike single-vector embedding models, ColBERT retains token-level representations and computes relevance scores through late interaction, providing better accuracy while being more efficient than cross-encoders.
 
-vLLM supports ColBERT models for reranking tasks, automatically applying MaxSim scoring for query-document relevance:
+vLLM supports ColBERT models with multiple encoder backbones:
+
+| Architecture | Backbone | Example HF Models |
+|---|---|---|
+| `HF_ColBERT` | BERT | `answerdotai/answerai-colbert-small-v1`, `colbert-ir/colbertv2.0` |
+| `ColBERTModernBertModel` | ModernBERT | `lightonai/GTE-ModernColBERT-v1` |
+| `ColBERTJinaRobertaModel` | Jina XLM-RoBERTa | `jinaai/jina-colbert-v2` |
+
+**BERT-based ColBERT** models work out of the box:
 
 ```shell
 vllm serve answerdotai/answerai-colbert-small-v1
 ```
 
-Currently supports ColBERT models with standard BERT encoders (e.g., `answerdotai/answerai-colbert-small-v1`, `colbert-ir/colbertv2.0`).
-
-ColBERT models with modified encoder architectures are not yet supported, including BERT variants with rotary embeddings (e.g., `jinaai/jina-colbert-v2`) or other custom encoders (e.g., `LiquidAI/LFM2-ColBERT-350M`).
-
-If your standard BERT ColBERT model's config doesn't specify the architecture as `HF_ColBERT`, override it with:
+For **non-BERT backbones**, use `--hf-overrides` to set the correct architecture:
 
 ```shell
-vllm serve your-colbert-model --hf-overrides '{"architectures": ["HF_ColBERT"]}'
+# ModernBERT backbone
+vllm serve lightonai/GTE-ModernColBERT-v1 \
+    --hf-overrides '{"architectures": ["ColBERTModernBertModel"]}'
+
+# Jina XLM-RoBERTa backbone
+vllm serve jinaai/jina-colbert-v2 \
+    --hf-overrides '{"architectures": ["ColBERTJinaRobertaModel"]}' \
+    --trust-remote-code
 ```
 
 Then you can use the rerank endpoint:
@@ -363,6 +374,77 @@ curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{
 
 An example can be found here: [examples/pooling/score/colbert_rerank_online.py](../../examples/pooling/score/colbert_rerank_online.py)
 
+### ColQwen3 Multi-Modal Late Interaction Models
+
+ColQwen3 is based on [ColPali](https://arxiv.org/abs/2407.01449), which extends ColBERT's late interaction approach to **multi-modal** inputs. While ColBERT operates on text-only token embeddings, ColPali/ColQwen3 can embed both **text and images** (e.g. PDF pages, screenshots, diagrams) into per-token L2-normalized vectors and compute relevance via MaxSim scoring. ColQwen3 specifically uses Qwen3-VL as its vision-language backbone.
+
+| Architecture | Backbone | Example HF Models |
+|---|---|---|
+| `ColQwen3` | Qwen3-VL | `TomoroAI/tomoro-colqwen3-embed-4b`, `TomoroAI/tomoro-colqwen3-embed-8b` |
+| `OpsColQwen3Model` | Qwen3-VL | `OpenSearch-AI/Ops-Colqwen3-4B`, `OpenSearch-AI/Ops-Colqwen3-8B` |
+
+Start the server:
+
+```shell
+vllm serve TomoroAI/tomoro-colqwen3-embed-4b --max-model-len 4096
+```
+
+Then you can use the rerank endpoint:
+
+```shell
+curl -s http://localhost:8000/rerank -H "Content-Type: application/json" -d '{
+    "model": "TomoroAI/tomoro-colqwen3-embed-4b",
+    "query": "What is machine learning?",
+    "documents": [
+        "Machine learning is a subset of artificial intelligence.",
+        "Python is a programming language.",
+        "Deep learning uses neural networks."
+    ]
+}'
+```
+
+Or the score endpoint:
+
+```shell
+curl -s http://localhost:8000/score -H "Content-Type: application/json" -d '{
+    "model": "TomoroAI/tomoro-colqwen3-embed-4b",
+    "text_1": "What is the capital of France?",
+    "text_2": ["The capital of France is Paris.", "Python is a programming language."]
+}'
+```
+
+You can also get the raw token embeddings using the pooling endpoint with `token_embed` task:
+
+```shell
+curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{
+    "model": "TomoroAI/tomoro-colqwen3-embed-4b",
+    "input": "What is machine learning?",
+    "task": "token_embed"
+}'
+```
+
+For **image inputs**, use the chat-style `messages` field so that the vLLM multimodal processor handles them correctly:
+
+```shell
+curl -s http://localhost:8000/pooling -H "Content-Type: application/json" -d '{
+    "model": "TomoroAI/tomoro-colqwen3-embed-4b",
+    "messages": [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64>"}},
+                {"type": "text", "text": "Describe the image."}
+            ]
+        }
+    ]
+}'
+```
+
+Examples can be found here:
+
+- Multi-vector retrieval: [examples/pooling/token_embed/colqwen3_token_embed_online.py](../../examples/pooling/token_embed/colqwen3_token_embed_online.py)
+- Reranking: [examples/pooling/score/colqwen3_rerank_online.py](../../examples/pooling/score/colqwen3_rerank_online.py)
+
 ### BAAI/bge-m3
 
 The `BAAI/bge-m3` model comes with extra weights for sparse and colbert embeddings but unfortunately in its `config.json`

docs/models/supported_models.md

@@ -181,6 +181,7 @@ Some model architectures are supported via vLLM plugins. These plugins extend vL
 | Architecture | Models | Plugin Repository |
 |--------------|--------|-------------------|
 | `BartForConditionalGeneration` | BART | [bart-plugin](https://github.com/vllm-project/bart-plugin) |
+| `Florence2ForConditionalGeneration` | Florence-2 | [bart-plugin](https://github.com/vllm-project/bart-plugin) |
 
 For other model architectures not natively supported, in particular for Encoder-Decoder models, we recommend following a similar pattern by implementing support through the plugin system.
 
@@ -728,6 +729,8 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `OpenPanguVLForConditionalGeneration` | openpangu-VL | T + I<sup>E+</sup> + V<sup>E+</sup> |`FreedomIntelligence/openPangu-VL-7B` | ✅︎ | ✅︎ |
 | `Ovis` | Ovis2, Ovis1.6 | T + I<sup>+</sup> | `AIDC-AI/Ovis2-1B`, `AIDC-AI/Ovis1.6-Llama3.2-3B`, etc. | | ✅︎ |
 | `Ovis2_5` | Ovis2.5 | T + I<sup>+</sup> + V | `AIDC-AI/Ovis2.5-9B`, etc. | | |
+| `Ovis2_6ForCausalLM` | Ovis2.6 | T + I<sup>+</sup> + V | `AIDC-AI/Ovis2.6-2B`, etc. | | |
+| `Ovis2_6_MoeForCausalLM` | Ovis2.6 | T + I<sup>+</sup> + V | `AIDC-AI/Ovis2.6-30B-A3B`, etc. | | |
 | `PaddleOCRVLForConditionalGeneration` | Paddle-OCR | T + I<sup>+</sup> | `PaddlePaddle/PaddleOCR-VL`, etc. | | |
 | `PaliGemmaForConditionalGeneration` | PaliGemma, PaliGemma 2 | T + I<sup>E</sup> | `google/paligemma-3b-pt-224`, `google/paligemma-3b-mix-224`, `google/paligemma2-3b-ft-docci-448`, etc. | ✅︎ | ✅︎ |
 | `Phi3VForCausalLM` | Phi-3-Vision, Phi-3.5-Vision | T + I<sup>E+</sup> | `microsoft/Phi-3-vision-128k-instruct`, `microsoft/Phi-3.5-vision-instruct`, etc. | | ✅︎ |

docs/usage/v1_guide.md

@@ -137,6 +137,7 @@ Please note that prefix caching is not yet supported for any of the above models
 Whisper is supported natively. Other encoder-decoder models are supported via the plugin system:
 
 - **BART**: `BartForConditionalGeneration` is supported via the official [bart-plugin](https://github.com/vllm-project/bart-plugin).
+- **Florence-2**: `Florence2ForConditionalGeneration` is supported via the official [bart-plugin](https://github.com/vllm-project/bart-plugin).
 
 For other encoder-decoder models (e.g., `MllamaForConditionalGeneration`), we recommend
 following a similar pattern by implementing support through the [plugin system](../design/plugin_system.md).

examples/online_serving/data_parallel_pause_resume.py

@@ -0,0 +1,135 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Test pause/resume with Data Parallel (DP) via HTTP API.
+
+This example demonstrates coordinated pause/resume across multiple DP ranks.
+The pause synchronizes across all DP engines via all-reduce.
+
+Prerequisites:
+    Start a vLLM server with data parallelism:
+
+    $ VLLM_SERVER_DEV_MODE=1 vllm serve facebook/opt-125m \
+        --enforce-eager \
+        --data-parallel-size 4 \
+        --tensor-parallel-size 1
+
+    Then run this script:
+
+    $ python data_parallel_pause_resume.py
+
+The test verifies pause works by:
+1. Starting a streaming generation request
+2. Pausing the server mid-generation
+3. Sleeping for PAUSE_DURATION seconds
+4. Resuming the server
+5. Verifying there was a gap in token generation matching the pause duration
+"""
+
+import argparse
+import threading
+import time
+
+import requests
+from openai import OpenAI
+
+BASE_URL = "http://localhost:8000"
+MODEL_NAME = "facebook/opt-125m"
+PAUSE_DURATION = 3.0
+
+
+def pause_generation(base_url: str, mode: str = "keep") -> None:
+    """Pause generation via HTTP endpoint."""
+    url = f"{base_url}/pause"
+    response = requests.post(url, params={"mode": mode}, timeout=60)
+    response.raise_for_status()
+    print("Server paused")
+
+
+def resume_generation(base_url: str) -> None:
+    """Resume generation via HTTP endpoint."""
+    url = f"{base_url}/resume"
+    response = requests.post(url, timeout=60)
+    response.raise_for_status()
+    print("Server resumed")
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-url", default=BASE_URL)
+    parser.add_argument("--model", default=MODEL_NAME)
+    args = parser.parse_args()
+
+    client = OpenAI(
+        base_url=f"{args.base_url}/v1",
+        api_key="EMPTY",
+    )
+
+    prompt = "Write a long story about a dragon. Once upon a time"
+    token_times: list[float] = []
+    pause_token_idx = 0
+    pause_triggered = threading.Event()
+
+    def generator_thread():
+        """Stream tokens and record timestamps."""
+        stream = client.completions.create(
+            model=args.model,
+            prompt=prompt,
+            max_tokens=50,
+            stream=True,
+        )
+        for chunk in stream:
+            if chunk.choices[0].text:
+                token_times.append(time.monotonic())
+                token_count = len(token_times)
+                print(f"Token {token_count}: {chunk.choices[0].text!r}")
+
+                # Signal controller after some tokens
+                if token_count >= 5 and not pause_triggered.is_set():
+                    pause_triggered.set()
+
+    def controller_thread():
+        """Pause and resume the server."""
+        nonlocal pause_token_idx
+
+        # Wait for some tokens
+        pause_triggered.wait()
+
+        print(f"\nPausing server (keep mode) at token {len(token_times)}...")
+        pause_generation(args.base_url, mode="keep")
+        pause_token_idx = len(token_times)
+        print(f"Sleeping for {PAUSE_DURATION}s...")
+
+        time.sleep(PAUSE_DURATION)
+
+        print("Resuming server...")
+        resume_generation(args.base_url)
+        print("Resumed!\n")
+
+    # Run both threads
+    gen_thread = threading.Thread(target=generator_thread)
+    ctrl_thread = threading.Thread(target=controller_thread)
+
+    gen_thread.start()
+    ctrl_thread.start()
+
+    gen_thread.join()
+    ctrl_thread.join()
+
+    # Check gap at the pause point
+    if pause_token_idx < len(token_times):
+        pause_gap = token_times[pause_token_idx] - token_times[pause_token_idx - 1]
+        print(
+            f"\nGap after pause (token {pause_token_idx} -> "
+            f"{pause_token_idx + 1}): {pause_gap:.3f}s"
+        )
+        if pause_gap >= PAUSE_DURATION * 0.9:
+            print("Test passed! Pause synchronized across DP ranks.")
+        else:
+            print(f"Test failed! Expected ~{PAUSE_DURATION}s gap, got {pause_gap:.3f}s")
+    else:
+        print("Test failed! No tokens were generated after resuming.")
+
+
+if __name__ == "__main__":
+    main()

examples/online_serving/disaggregated_encoder/disagg_1e1p1d_example.sh

@@ -8,7 +8,7 @@ declare -a PIDS=()
 ###############################################################################
 MODEL="${MODEL:-Qwen/Qwen2.5-VL-3B-Instruct}"
 LOG_PATH="${LOG_PATH:-./logs}"
-mkdir -p $LOG_PATH
+mkdir -p "$LOG_PATH"
 
 ENCODE_PORT="${ENCODE_PORT:-19534}"
 PREFILL_PORT="${PREFILL_PORT:-19535}"
@@ -84,10 +84,10 @@ trap cleanup TERM
 
 # clear previous cache
 echo "remove previous ec cache folder"
-rm -rf $EC_SHARED_STORAGE_PATH
+rm -rf "$EC_SHARED_STORAGE_PATH"
 
 echo "make ec cache folder"
-mkdir -p $EC_SHARED_STORAGE_PATH
+mkdir -p "$EC_SHARED_STORAGE_PATH"
 
 ###############################################################################
 # Encoder worker
@@ -100,7 +100,7 @@ CUDA_VISIBLE_DEVICES="$GPU_E" vllm serve "$MODEL" \
     --no-enable-prefix-caching \
     --max-num-batched-tokens 114688 \
     --max-num-seqs 128 \
-    --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+    --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
     --ec-transfer-config '{
         "ec_connector": "ECExampleConnector",
         "ec_role": "ec_producer",
@@ -124,7 +124,7 @@ vllm serve "$MODEL" \
     --enforce-eager \
     --enable-request-id-headers \
     --max-num-seqs 128 \
-    --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+    --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
     --ec-transfer-config '{
         "ec_connector": "ECExampleConnector",
         "ec_role": "ec_consumer",
@@ -152,7 +152,7 @@ vllm serve "$MODEL" \
     --enforce-eager \
     --enable-request-id-headers \
     --max-num-seqs 128 \
-    --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+    --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
     --kv-transfer-config '{
         "kv_connector": "NixlConnector",
         "kv_role": "kv_consumer"
@@ -162,9 +162,9 @@ vllm serve "$MODEL" \
 PIDS+=($!)
 
 # Wait for workers
-wait_for_server $ENCODE_PORT
-wait_for_server $PREFILL_PORT
-wait_for_server $DECODE_PORT
+wait_for_server "$ENCODE_PORT"
+wait_for_server "$PREFILL_PORT"
+wait_for_server "$DECODE_PORT"
 
 ###############################################################################
 # Proxy
@@ -179,7 +179,7 @@ python disagg_epd_proxy.py \
 
 PIDS+=($!)
 
-wait_for_server $PROXY_PORT
+wait_for_server "$PROXY_PORT"
 echo "All services are up!"
 
 ###############################################################################
@@ -187,14 +187,14 @@ echo "All services are up!"
 ###############################################################################
 echo "Running benchmark (stream)..."
 vllm bench serve \
-  --model               $MODEL \
+  --model               "$MODEL" \
   --backend             openai-chat \
   --endpoint            /v1/chat/completions \
   --dataset-name        hf \
   --dataset-path        lmarena-ai/VisionArena-Chat \
   --seed                0 \
-  --num-prompts         $NUM_PROMPTS \
-  --port                $PROXY_PORT
+  --num-prompts         "$NUM_PROMPTS" \
+  --port                "$PROXY_PORT"
 
 PIDS+=($!)
 
@@ -202,10 +202,10 @@ PIDS+=($!)
 # Single request with local image
 ###############################################################################
 echo "Running single request with local image (non-stream)..."
-curl http://127.0.0.1:${PROXY_PORT}/v1/chat/completions \
+curl http://127.0.0.1:"${PROXY_PORT}"/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
-    "model": "'${MODEL}'",
+    "model": "'"${MODEL}"'",
     "messages": [
     {"role": "system", "content": "You are a helpful assistant."},
     {"role": "user", "content": [

examples/online_serving/disaggregated_encoder/disagg_1e1pd_example.sh

@@ -8,7 +8,7 @@ declare -a PIDS=()
 ###############################################################################
 MODEL="${MODEL:-Qwen/Qwen2.5-VL-3B-Instruct}"
 LOG_PATH="${LOG_PATH:-./logs}"
-mkdir -p $LOG_PATH
+mkdir -p "$LOG_PATH"
 
 ENCODE_PORT="${ENCODE_PORT:-19534}"
 PREFILL_DECODE_PORT="${PREFILL_DECODE_PORT:-19535}"
@@ -78,10 +78,10 @@ trap cleanup TERM
 
 # clear previous cache
 echo "remove previous ec cache folder"
-rm -rf $EC_SHARED_STORAGE_PATH
+rm -rf "$EC_SHARED_STORAGE_PATH"
 
 echo "make ec cache folder"
-mkdir -p $EC_SHARED_STORAGE_PATH
+mkdir -p "$EC_SHARED_STORAGE_PATH"
 
 ###############################################################################
 # Encoder worker
@@ -94,7 +94,7 @@ CUDA_VISIBLE_DEVICES="$GPU_E" vllm serve "$MODEL" \
     --no-enable-prefix-caching \
     --max-num-batched-tokens 114688 \
     --max-num-seqs 128 \
-    --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+    --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
     --ec-transfer-config '{
         "ec_connector": "ECExampleConnector",
         "ec_role": "ec_producer",
@@ -115,7 +115,7 @@ CUDA_VISIBLE_DEVICES="$GPU_PD" vllm serve "$MODEL" \
     --enforce-eager \
     --enable-request-id-headers \
     --max-num-seqs 128 \
-    --allowed-local-media-path ${GIT_ROOT}/tests/v1/ec_connector/integration \
+    --allowed-local-media-path "${GIT_ROOT}"/tests/v1/ec_connector/integration \
     --ec-transfer-config '{
         "ec_connector": "ECExampleConnector",
         "ec_role": "ec_consumer",
@@ -128,8 +128,8 @@ CUDA_VISIBLE_DEVICES="$GPU_PD" vllm serve "$MODEL" \
 PIDS+=($!)
 
 # Wait for workers
-wait_for_server $ENCODE_PORT
-wait_for_server $PREFILL_DECODE_PORT
+wait_for_server "$ENCODE_PORT"
+wait_for_server "$PREFILL_DECODE_PORT"
 
 ###############################################################################
 # Proxy
@@ -144,7 +144,7 @@ python disagg_epd_proxy.py \
 
 PIDS+=($!)
 
-wait_for_server $PROXY_PORT
+wait_for_server "$PROXY_PORT"
 echo "All services are up!"
 
 ###############################################################################
@@ -152,14 +152,14 @@ echo "All services are up!"
 ###############################################################################
 echo "Running benchmark (stream)..."
 vllm bench serve \
-  --model               $MODEL \
+  --model               "$MODEL" \
   --backend             openai-chat \
   --endpoint            /v1/chat/completions \
   --dataset-name        hf \
   --dataset-path        lmarena-ai/VisionArena-Chat \
   --seed                0 \
-  --num-prompts         $NUM_PROMPTS \
-  --port                $PROXY_PORT
+  --num-prompts         "$NUM_PROMPTS" \
+  --port                "$PROXY_PORT"
 
 PIDS+=($!)
 
@@ -167,10 +167,10 @@ PIDS+=($!)
 # Single request with local image
 ###############################################################################
 echo "Running single request with local image (non-stream)..."
-curl http://127.0.0.1:${PROXY_PORT}/v1/chat/completions \
+curl http://127.0.0.1:"${PROXY_PORT}"/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
-    "model": "'${MODEL}'",
+    "model": "'"${MODEL}"'",
     "messages": [
     {"role": "system", "content": "You are a helpful assistant."},
     {"role": "user", "content": [

examples/online_serving/disaggregated_prefill.sh

@@ -54,7 +54,7 @@ wait_for_server() {
 # You can also adjust --kv-ip and --kv-port for distributed inference.
 
 # prefilling instance, which is the KV producer
-CUDA_VISIBLE_DEVICES=0 vllm serve $MODEL_NAME \
+CUDA_VISIBLE_DEVICES=0 vllm serve "$MODEL_NAME" \
     --host 0.0.0.0 \
     --port 8100 \
     --max-model-len 100 \
@@ -64,7 +64,7 @@ CUDA_VISIBLE_DEVICES=0 vllm serve $MODEL_NAME \
     '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":"1e9","kv_port":"14579","kv_connector_extra_config":{"proxy_ip":"'"$VLLM_HOST_IP"'","proxy_port":"30001","http_ip":"'"$VLLM_HOST_IP"'","http_port":"8100","send_type":"PUT_ASYNC"}}' &
 
 # decoding instance, which is the KV consumer  
-CUDA_VISIBLE_DEVICES=1 vllm serve $MODEL_NAME \
+CUDA_VISIBLE_DEVICES=1 vllm serve "$MODEL_NAME" \
     --host 0.0.0.0 \
     --port 8200 \
     --max-model-len 100 \

examples/online_serving/disaggregated_serving/disagg_proxy_demo.py

@@ -328,9 +328,9 @@ class Proxy:
         if instance_type == "decode" and instance in self.decode_instances:
             self.decode_instances.remove(instance)
             self.decode_cycler = itertools.cycle(self.decode_instances)
-        if instance_type == "prefill" and instance in self.decode_instances:
+        if instance_type == "prefill" and instance in self.prefill_instances:
             self.prefill_instances.remove(instance)
-            self.prefill_cycler = itertools.cycle(self.decode_instances)
+            self.prefill_cycler = itertools.cycle(self.prefill_instances)
 
 
 class RoundRobinSchedulingPolicy(SchedulingPolicy):

examples/online_serving/disaggregated_serving/kv_events.sh

@@ -34,7 +34,7 @@ wait_for_server() {
     done" && return 0 || return 1
 }
 
-vllm serve $MODEL_NAME \
+vllm serve "$MODEL_NAME" \
     --port 8100 \
     --max-model-len 100 \
     --enforce-eager \

examples/online_serving/disaggregated_serving/mooncake_connector/run_mooncake_connector.sh

@@ -143,7 +143,7 @@ main() {
     IFS=',' read -ra BOOTSTRAP_PORT_ARRAY <<< "$BOOTSTRAP_PORTS"
     IFS=',' read -ra DECODE_PORT_ARRAY <<< "$DECODE_PORTS"
 
-    proxy_param=""
+    proxy_args=()
 
     # =============================================================================
     # Launch Prefill Servers (X Producers)
@@ -156,12 +156,12 @@ main() {
         local bootstrap_port=${BOOTSTRAP_PORT_ARRAY[$i]}
 
         echo "  Prefill server $((i+1)): GPU $gpu_id, Port $port, Bootstrap Port $bootstrap_port"
-        VLLM_MOONCAKE_BOOTSTRAP_PORT=$bootstrap_port CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
-        --port $port \
+        VLLM_MOONCAKE_BOOTSTRAP_PORT=$bootstrap_port CUDA_VISIBLE_DEVICES=$gpu_id vllm serve "$MODEL" \
+        --port "$port" \
         --kv-transfer-config \
         "{\"kv_connector\":\"MooncakeConnector\",\"kv_role\":\"kv_producer\"}" > prefill$((i+1)).log 2>&1 &
         PIDS+=($!)
-        proxy_param="${proxy_param} --prefill http://0.0.0.0:${port} $bootstrap_port"
+        proxy_args+=(--prefill "http://0.0.0.0:${port}" "$bootstrap_port")
     done
 
     # =============================================================================
@@ -174,12 +174,12 @@ main() {
         local port=${DECODE_PORT_ARRAY[$i]}
 
         echo "  Decode server $((i+1)): GPU $gpu_id, Port $port"
-        CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
-        --port $port \
+        CUDA_VISIBLE_DEVICES=$gpu_id vllm serve "$MODEL" \
+        --port "$port" \
         --kv-transfer-config \
         "{\"kv_connector\":\"MooncakeConnector\",\"kv_role\":\"kv_consumer\"}" > decode$((i+1)).log 2>&1 &
         PIDS+=($!)
-        proxy_param="${proxy_param} --decode http://0.0.0.0:${port}"
+        proxy_args+=(--decode "http://0.0.0.0:${port}")
     done
 
     # =============================================================================
@@ -187,7 +187,7 @@ main() {
     # =============================================================================
     echo ""
     echo "Starting proxy server on port $PROXY_PORT..."
-    python3 mooncake_connector_proxy.py $proxy_param --port $PROXY_PORT > proxy.log 2>&1 &
+    python3 mooncake_connector_proxy.py "${proxy_args[@]}" --port "$PROXY_PORT" > proxy.log 2>&1 &
     PIDS+=($!)
 
     # =============================================================================
@@ -196,9 +196,10 @@ main() {
     echo ""
     echo "Waiting for all servers to start..."
     for port in "${PREFILL_PORT_ARRAY[@]}" "${DECODE_PORT_ARRAY[@]}"; do
-        if ! wait_for_server $port; then
+        if ! wait_for_server "$port"; then
             echo "Failed to start server on port $port"
             cleanup
+            # shellcheck disable=SC2317
             exit 1
         fi
     done
@@ -209,8 +210,8 @@ main() {
     # =============================================================================
     # Run Benchmark
     # =============================================================================
-    vllm bench serve --port $PROXY_PORT --seed $(date +%s) \
-        --backend vllm --model $MODEL \
+    vllm bench serve --port "$PROXY_PORT" --seed "$(date +%s)" \
+        --backend vllm --model "$MODEL" \
         --dataset-name random --random-input-len 7500 --random-output-len 200 \
         --num-prompts 200 --burstiness 100 --request-rate 2 | tee benchmark.log
 

examples/online_serving/disaggregated_serving_p2p_nccl_xpyd/disagg_example_p2p_nccl_xpyd.sh

@@ -166,10 +166,10 @@ main() {
         local kv_port=$((21001 + i))
 
         echo "  Prefill server $((i+1)): GPU $gpu_id, Port $port, KV Port $kv_port"
-        CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
+        CUDA_VISIBLE_DEVICES=$gpu_id vllm serve "$MODEL" \
         --enforce-eager \
         --host 0.0.0.0 \
-        --port $port \
+        --port "$port" \
         --tensor-parallel-size 1 \
         --seed 1024 \
         --dtype float16 \
@@ -194,10 +194,10 @@ main() {
         local kv_port=$((22001 + i))
 
         echo "  Decode server $((i+1)): GPU $gpu_id, Port $port, KV Port $kv_port"
-        CUDA_VISIBLE_DEVICES=$gpu_id vllm serve $MODEL \
+        CUDA_VISIBLE_DEVICES=$gpu_id vllm serve "$MODEL" \
         --enforce-eager \
         --host 0.0.0.0 \
-        --port $port \
+        --port "$port" \
         --tensor-parallel-size 1 \
         --seed 1024 \
         --dtype float16 \
@@ -217,9 +217,10 @@ main() {
     echo ""
     echo "Waiting for all servers to start..."
     for port in "${PREFILL_PORT_ARRAY[@]}" "${DECODE_PORT_ARRAY[@]}"; do
-        if ! wait_for_server $port; then
+        if ! wait_for_server "$port"; then
             echo "Failed to start server on port $port"
             cleanup
+            # shellcheck disable=SC2317
             exit 1
         fi
     done
@@ -231,8 +232,8 @@ main() {
     # Run Benchmark
     # =============================================================================
     cd ../../../benchmarks/
-    vllm bench serve --port 10001 --seed $(date +%s) \
-        --model $MODEL \
+    vllm bench serve --port 10001 --seed "$(date +%s)" \
+        --model "$MODEL" \
         --dataset-name random --random-input-len 7500 --random-output-len 200 \
         --num-prompts 200 --burstiness 100 --request-rate 2 | tee benchmark.log
 

examples/online_serving/elastic_ep/bench.sh

@@ -50,8 +50,8 @@ while [[ $# -gt 0 ]]; do
 done
 
 vllm bench serve \
-    --model $MODEL_NAME \
-    --host $HOST \
-    --port $PORT \
-    --num-prompts $NUM_PROMPTS \
-    --request-rate $REQUEST_RATE
+    --model "$MODEL_NAME" \
+    --host "$HOST" \
+    --port "$PORT" \
+    --num-prompts "$NUM_PROMPTS" \
+    --request-rate "$REQUEST_RATE"

examples/online_serving/elastic_ep/serve_deepseek_v2.sh

@@ -57,15 +57,15 @@ echo "Starting vLLM server for $MODEL_NAME with data parallel size: $DATA_PARALL
 export RAY_DEDUP_LOGS=0
 export VLLM_USE_DEEP_GEMM=1
 
-vllm serve $MODEL_NAME \
-    --data-parallel-size $DATA_PARALLEL_SIZE \
-    --data-parallel-size-local $DATA_PARALLEL_SIZE \
+vllm serve "$MODEL_NAME" \
+    --data-parallel-size "$DATA_PARALLEL_SIZE" \
+    --data-parallel-size-local "$DATA_PARALLEL_SIZE" \
     --data-parallel-backend ray \
     --enforce-eager \
     --enable-expert-parallel \
     --enable-eplb \
     --all2all-backend pplx \
-    --num-redundant-experts $REDUNDANT_EXPERTS \
+    --num-redundant-experts "$REDUNDANT_EXPERTS" \
     --trust-remote-code \
-    --host $HOST \
-    --port $PORT
+    --host "$HOST" \
+    --port "$PORT"

examples/online_serving/multi-node-serving.sh

@@ -57,8 +57,7 @@ case "$subcommand" in
 
     # Retry until the worker node connects to the head node or the timeout expires.
     for (( i=0; i < $ray_init_timeout; i+=5 )); do
-      ray start --address=$ray_address:$ray_port --block "${start_params[@]}"
-      if [ $? -eq 0 ]; then
+      if ray start --address="$ray_address":"$ray_port" --block "${start_params[@]}"; then
         echo "Worker: Ray runtime started with head address $ray_address:$ray_port"
         exit 0
       fi
@@ -95,12 +94,12 @@ case "$subcommand" in
     fi
 
     # Start the Ray head node.
-    ray start --head --port=$ray_port "${start_params[@]}"
+    ray start --head --port="$ray_port" "${start_params[@]}"
 
     # Poll Ray until every worker node is active.
     for (( i=0; i < $ray_init_timeout; i+=5 )); do
-        active_nodes=`python3 -c 'import ray; ray.init(); print(sum(node["Alive"] for node in ray.nodes()))'`
-        if [ $active_nodes -eq $ray_cluster_size ]; then
+        active_nodes=$(python3 -c 'import ray; ray.init(); print(sum(node["Alive"] for node in ray.nodes()))')
+        if [ "$active_nodes" -eq "$ray_cluster_size" ]; then
           echo "All ray workers are active and the ray cluster is initialized successfully."
           exit 0
         fi

examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_example_nixl.sh

@@ -22,11 +22,10 @@ check_hf_token() {
 
 check_num_gpus() {
     # can you check if the number of GPUs are >=2 via nvidia-smi/rocm-smi?
-    which rocm-smi > /dev/null 2>&1
-    if [ $? -ne 0 ]; then
+    if ! which rocm-smi > /dev/null 2>&1; then
 	num_gpus=$(nvidia-smi --query-gpu=name --format=csv,noheader | wc -l)
     else
-	num_gpus=$(rocm-smi --showid | grep Instinct | wc -l)
+	num_gpus=$(rocm-smi --showid | grep -c Instinct)
     fi
 
     if [ "$num_gpus" -lt 2 ]; then
@@ -39,8 +38,7 @@ check_num_gpus() {
 
 ensure_python_library_installed() {
     echo "Checking if $1 is installed..."
-    python3 -c "import $1" > /dev/null 2>&1
-    if [ $? -ne 0 ]; then
+    if ! python3 -c "import $1" > /dev/null 2>&1; then
         if [ "$1" == "nixl" ]; then
             echo "$1 is not installed. Please refer to https://github.com/ai-dynamo/nixl for installation."
         else
@@ -102,12 +100,12 @@ main() {
     bash disagg_vllm_launcher.sh prefiller \
         > >(tee prefiller.log) 2>&1 &
     prefiller_pid=$!
-    PIDS+=($prefiller_pid)
+    PIDS+=("$prefiller_pid")
 
     bash disagg_vllm_launcher.sh decoder  \
         > >(tee decoder.log)  2>&1 &
     decoder_pid=$!
-    PIDS+=($decoder_pid)
+    PIDS+=("$decoder_pid")
 
     python3 disagg_proxy_server.py \
         --host localhost \
@@ -118,7 +116,7 @@ main() {
         --decoder-port 8200  \
         > >(tee proxy.log)    2>&1 &
     proxy_pid=$!
-    PIDS+=($proxy_pid)
+    PIDS+=("$proxy_pid")
 
     wait_for_server 8100
     wait_for_server 8200
@@ -128,7 +126,7 @@ main() {
 
     # begin benchmark
     cd ../../../../benchmarks/
-    vllm bench serve --port 9000 --seed $(date +%s) \
+    vllm bench serve --port 9000 --seed "$(date +%s)" \
         --model meta-llama/Llama-3.1-8B-Instruct \
         --dataset-name random --random-input-len 7500 --random-output-len 200 \
         --num-prompts 200 --burstiness 100 --request-rate 3.6 | tee benchmark.log

examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_vllm_launcher.sh

@@ -34,7 +34,7 @@ if [[ $1 == "prefiller" ]]; then
         VLLM_ENABLE_V1_MULTIPROCESSING=1 \
         VLLM_WORKER_MULTIPROC_METHOD=spawn \
         CUDA_VISIBLE_DEVICES=0 \
-        vllm serve $MODEL \
+        vllm serve "$MODEL" \
         --port 8100 \
         --enforce-eager \
         --kv-transfer-config \
@@ -51,7 +51,7 @@ elif [[ $1 == "decoder" ]]; then
         VLLM_ENABLE_V1_MULTIPROCESSING=1 \
         VLLM_WORKER_MULTIPROC_METHOD=spawn \
         CUDA_VISIBLE_DEVICES=1 \
-        vllm serve $MODEL \
+        vllm serve "$MODEL" \
         --port 8200 \
         --enforce-eager \
         --kv-transfer-config \

examples/pooling/embed/openai_embedding_long_text/service.sh

@@ -103,7 +103,7 @@ vllm serve "$MODEL_NAME" \
   --tensor-parallel-size "$GPU_COUNT" \
   --enforce-eager \
   --pooler-config "$POOLER_CONFIG" \
-  --served-model-name ${MODEL_CODE} \
+  --served-model-name "${MODEL_CODE}" \
   --api-key "$API_KEY" \
   --trust-remote-code \
   --port "$PORT" \

examples/pooling/plugin/prithvi_geospatial_mae_io_processor.py

@@ -20,15 +20,17 @@ def main():
     torch.set_default_dtype(torch.float16)
     image_url = "https://huggingface.co/christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM/resolve/main/valencia_example_2024-10-26.tiff"  # noqa: E501
 
-    img_prompt = dict(
+    img_data = dict(
         data=image_url,
         data_format="url",
         image_format="tiff",
         out_data_format="b64_json",
     )
 
+    prompt = dict(data=img_data)
+
     llm = LLM(
-        model="christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM",
+        model="ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11",
         skip_tokenizer_init=True,
         trust_remote_code=True,
         enforce_eager=True,
@@ -41,7 +43,7 @@ def main():
         enable_mm_embeds=True,
     )
 
-    pooler_output = llm.encode(img_prompt, pooling_task="plugin")
+    pooler_output = llm.encode(prompt, pooling_task="plugin")
     output = pooler_output[0].outputs
 
     print(output)

examples/pooling/plugin/prithvi_geospatial_mae_offline.py

@@ -391,7 +391,7 @@ if __name__ == "__main__":
     parser.add_argument(
         "--model",
         type=str,
-        default="christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM",
+        default="ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11",
         help="Path to a checkpoint file to load from.",
     )
     parser.add_argument(

examples/pooling/plugin/prithvi_geospatial_mae_online.py

@@ -14,9 +14,7 @@ import requests
 # - install TerraTorch v1.1 (or later):
 #   pip install terratorch>=v1.1
 # - start vllm in serving mode with the below args
-#   --model='christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM'
-#   --model-impl terratorch
-#   --trust-remote-code
+#   --model='ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11'
 #   --skip-tokenizer-init --enforce-eager
 #   --io-processor-plugin terratorch_segmentation
 #   --enable-mm-embeds
@@ -34,7 +32,7 @@ def main():
             "out_data_format": "b64_json",
         },
         "priority": 0,
-        "model": "christian-pinto/Prithvi-EO-2.0-300M-TL-VLLM",
+        "model": "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11",
     }
 
     ret = requests.post(server_endpoint, json=request_payload_url)

examples/pooling/score/colbert_rerank_online.py

@@ -1,15 +1,27 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
-Example of using ColBERT late interaction model for reranking.
+Example of using ColBERT late interaction models for reranking and scoring.
 
 ColBERT (Contextualized Late Interaction over BERT) uses per-token embeddings
 and MaxSim scoring for document reranking, providing better accuracy than
 single-vector models while being more efficient than cross-encoders.
 
-Start the server with:
+vLLM supports ColBERT with multiple encoder backbones. Start the server
+with one of the following:
+
+    # BERT backbone (works out of the box)
     vllm serve answerdotai/answerai-colbert-small-v1
 
+    # ModernBERT backbone
+    vllm serve lightonai/GTE-ModernColBERT-v1 \
+        --hf-overrides '{"architectures": ["ColBERTModernBertModel"]}'
+
+    # Jina XLM-RoBERTa backbone
+    vllm serve jinaai/jina-colbert-v2 \
+        --hf-overrides '{"architectures": ["ColBERTJinaRobertaModel"]}' \
+        --trust-remote-code
+
 Then run this script:
     python colbert_rerank_online.py
 """
@@ -18,39 +30,62 @@ import json
 
 import requests
 
-url = "http://127.0.0.1:8000/rerank"
+# Change this to match the model you started the server with
+MODEL = "answerdotai/answerai-colbert-small-v1"
+BASE_URL = "http://127.0.0.1:8000"
 
 headers = {"accept": "application/json", "Content-Type": "application/json"}
 
-data = {
-    "model": "answerdotai/answerai-colbert-small-v1",
-    "query": "What is machine learning?",
-    "documents": [
-        "Machine learning is a subset of artificial intelligence.",
-        "Python is a programming language.",
-        "Deep learning uses neural networks for complex tasks.",
-        "The weather today is sunny.",
-    ],
-}
+documents = [
+    "Machine learning is a subset of artificial intelligence.",
+    "Python is a programming language.",
+    "Deep learning uses neural networks for complex tasks.",
+    "The weather today is sunny.",
+]
+
+
+def rerank_example():
+    """Use the /rerank endpoint to rank documents by query relevance."""
+    print("=== Rerank Example ===")
+
+    data = {
+        "model": MODEL,
+        "query": "What is machine learning?",
+        "documents": documents,
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+    result = response.json()
+    print(json.dumps(result, indent=2))
+
+    print("\nRanked documents (most relevant first):")
+    for item in result["results"]:
+        doc_idx = item["index"]
+        score = item["relevance_score"]
+        print(f"  Score {score:.4f}: {documents[doc_idx]}")
+
+
+def score_example():
+    """Use the /score endpoint for pairwise query-document scoring."""
+    print("\n=== Score Example ===")
+
+    data = {
+        "model": MODEL,
+        "text_1": "What is machine learning?",
+        "text_2": [
+            "Machine learning is a subset of AI.",
+            "The weather is sunny.",
+        ],
+    }
+
+    response = requests.post(f"{BASE_URL}/score", headers=headers, json=data)
+    result = response.json()
+    print(json.dumps(result, indent=2))
 
 
 def main():
-    response = requests.post(url, headers=headers, json=data)
-
-    if response.status_code == 200:
-        print("ColBERT Rerank Request successful!")
-        result = response.json()
-        print(json.dumps(result, indent=2))
-
-        # Show ranked results
-        print("\nRanked documents (most relevant first):")
-        for item in result["results"]:
-            doc_idx = item["index"]
-            score = item["relevance_score"]
-            print(f"  Score {score:.4f}: {data['documents'][doc_idx]}")
-    else:
-        print(f"Request failed with status code: {response.status_code}")
-        print(response.text)
+    rerank_example()
+    score_example()
 
 
 if __name__ == "__main__":

examples/pooling/score/colqwen3_rerank_online.py

@@ -0,0 +1,130 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Example of using ColQwen3 late interaction model for reranking.
+
+ColQwen3 is a multi-modal ColBERT-style model based on Qwen3-VL.
+It produces per-token embeddings and uses MaxSim scoring for retrieval
+and reranking. Supports both text and image inputs.
+
+Start the server with:
+    vllm serve TomoroAI/tomoro-colqwen3-embed-4b --max-model-len 50000
+
+Then run this script:
+    python colqwen3_rerank_online.py
+"""
+
+import requests
+
+MODEL = "TomoroAI/tomoro-colqwen3-embed-4b"
+BASE_URL = "http://127.0.0.1:8000"
+
+headers = {"accept": "application/json", "Content-Type": "application/json"}
+
+
+def rerank_text():
+    """Text-only reranking via /rerank endpoint."""
+    print("=" * 60)
+    print("1. Text reranking (/rerank)")
+    print("=" * 60)
+
+    data = {
+        "model": MODEL,
+        "query": "What is machine learning?",
+        "documents": [
+            "Machine learning is a subset of artificial intelligence.",
+            "Python is a programming language.",
+            "Deep learning uses neural networks for complex tasks.",
+            "The weather today is sunny.",
+        ],
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print("\n  Ranked documents (most relevant first):")
+        for item in result["results"]:
+            doc_idx = item["index"]
+            score = item["relevance_score"]
+            print(f"    [{score:.4f}] {data['documents'][doc_idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def score_text():
+    """Text-only scoring via /score endpoint."""
+    print()
+    print("=" * 60)
+    print("2. Text scoring (/score)")
+    print("=" * 60)
+
+    query = "What is the capital of France?"
+    documents = [
+        "The capital of France is Paris.",
+        "Berlin is the capital of Germany.",
+        "Python is a programming language.",
+    ]
+
+    data = {
+        "model": MODEL,
+        "text_1": query,
+        "text_2": documents,
+    }
+
+    response = requests.post(f"{BASE_URL}/score", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print(f"\n  Query: {query}\n")
+        for item in result["data"]:
+            idx = item["index"]
+            score = item["score"]
+            print(f"    Doc {idx} (score={score:.4f}): {documents[idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def score_text_top_n():
+    """Text reranking with top_n filtering via /rerank endpoint."""
+    print()
+    print("=" * 60)
+    print("3. Text reranking with top_n=2 (/rerank)")
+    print("=" * 60)
+
+    data = {
+        "model": MODEL,
+        "query": "What is the capital of France?",
+        "documents": [
+            "The capital of France is Paris.",
+            "Berlin is the capital of Germany.",
+            "Python is a programming language.",
+            "The Eiffel Tower is in Paris.",
+        ],
+        "top_n": 2,
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print(f"\n  Top {data['top_n']} results:")
+        for item in result["results"]:
+            doc_idx = item["index"]
+            score = item["relevance_score"]
+            print(f"    [{score:.4f}] {data['documents'][doc_idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def main():
+    rerank_text()
+    score_text()
+    score_text_top_n()
+
+
+if __name__ == "__main__":
+    main()

examples/pooling/token_embed/colqwen3_token_embed_online.py

@@ -0,0 +1,198 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# ruff: noqa: E501
+
+"""
+Example online usage of Pooling API for ColQwen3 multi-vector retrieval.
+
+ColQwen3 is a multi-modal late interaction model based on Qwen3-VL that
+produces per-token embeddings (320-dim, L2-normalized) for both text and
+image inputs. Similarity is computed via MaxSim scoring.
+
+This example mirrors the official TomoroAI inference code
+(https://huggingface.co/TomoroAI/tomoro-colqwen3-embed-4b) but uses the
+vLLM serving API instead of local HuggingFace model loading.
+
+Start the server with:
+    vllm serve TomoroAI/tomoro-colqwen3-embed-4b --max-model-len 4096
+
+Then run this script:
+    python colqwen3_token_embed_online.py
+"""
+
+import argparse
+import base64
+from io import BytesIO
+
+import numpy as np
+import requests
+from PIL import Image
+
+# ── Helpers ─────────────────────────────────────────────────
+
+
+def post_http_request(payload: dict, api_url: str) -> requests.Response:
+    headers = {"User-Agent": "Test Client"}
+    return requests.post(api_url, headers=headers, json=payload)
+
+
+def load_image(url: str) -> Image.Image:
+    """Download an image from URL (handles Wikimedia 403)."""
+    for hdrs in ({}, {"User-Agent": "Mozilla/5.0 (compatible; ColQwen3-demo/1.0)"}):
+        resp = requests.get(url, headers=hdrs, timeout=10)
+        if resp.status_code == 403:
+            continue
+        resp.raise_for_status()
+        return Image.open(BytesIO(resp.content)).convert("RGB")
+    raise RuntimeError(f"Could not fetch image from {url}")
+
+
+def encode_image_base64(image: Image.Image) -> str:
+    """Encode a PIL image to a base64 data URI."""
+    buf = BytesIO()
+    image.save(buf, format="PNG")
+    return "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode()
+
+
+def compute_maxsim(q_emb: np.ndarray, d_emb: np.ndarray) -> float:
+    """Compute ColBERT-style MaxSim score between query and document."""
+    sim = q_emb @ d_emb.T
+    return float(sim.max(axis=-1).sum())
+
+
+# ── Encode functions ────────────────────────────────────────
+
+
+def encode_queries(texts: list[str], model: str, api_url: str) -> list[np.ndarray]:
+    """Encode text queries → list of multi-vector embeddings."""
+    resp = post_http_request({"model": model, "input": texts}, api_url)
+    return [np.array(item["data"]) for item in resp.json()["data"]]
+
+
+def encode_images(image_urls: list[str], model: str, api_url: str) -> list[np.ndarray]:
+    """Encode image documents → list of multi-vector embeddings.
+
+    Images are sent via the chat-style `messages` field so that the
+    vLLM multimodal processor handles them correctly.
+    """
+    embeddings = []
+    for url in image_urls:
+        print(f"  Loading: {url.split('/')[-1]}...")
+        image = load_image(url)
+        image_uri = encode_image_base64(image)
+        resp = post_http_request(
+            {
+                "model": model,
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": [
+                            {"type": "image_url", "image_url": {"url": image_uri}},
+                            {"type": "text", "text": "Describe the image."},
+                        ],
+                    }
+                ],
+            },
+            api_url,
+        )
+        result = resp.json()
+        if resp.status_code != 200 or "data" not in result:
+            print(f"    Error ({resp.status_code}): {str(result)[:200]}")
+            continue
+        embeddings.append(np.array(result["data"][0]["data"]))
+    return embeddings
+
+
+# ── Main ────────────────────────────────────────────────────
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="TomoroAI/tomoro-colqwen3-embed-4b",
+    )
+    return parser.parse_args()
+
+
+def main(args):
+    pooling_url = f"http://{args.host}:{args.port}/pooling"
+    score_url = f"http://{args.host}:{args.port}/score"
+    model = args.model
+
+    # Same sample data as the official TomoroAI example
+    queries = [
+        "Retrieve the city of Singapore",
+        "Retrieve the city of Beijing",
+        "Retrieve the city of London",
+    ]
+    image_urls = [
+        "https://upload.wikimedia.org/wikipedia/commons/2/27/Singapore_skyline_2022.jpg",
+        "https://upload.wikimedia.org/wikipedia/commons/6/61/Beijing_skyline_at_night.JPG",
+        "https://upload.wikimedia.org/wikipedia/commons/4/49/London_skyline.jpg",
+    ]
+
+    # ── 1) Text query embeddings ────────────────────────────
+    print("=" * 60)
+    print("1. Encode text queries (multi-vector)")
+    print("=" * 60)
+    query_embeddings = encode_queries(queries, model, pooling_url)
+    for i, emb in enumerate(query_embeddings):
+        norm = float(np.linalg.norm(emb[0]))
+        print(f'  Query {i}: {emb.shape}  (L2 norm: {norm:.4f})  "{queries[i]}"')
+
+    # ── 2) Image document embeddings ────────────────────────
+    print()
+    print("=" * 60)
+    print("2. Encode image documents (multi-vector)")
+    print("=" * 60)
+    doc_embeddings = encode_images(image_urls, model, pooling_url)
+    for i, emb in enumerate(doc_embeddings):
+        print(f"  Doc {i}:   {emb.shape}  {image_urls[i].split('/')[-1]}")
+
+    # ── 3) Cross-modal MaxSim scoring ───────────────────────
+    if doc_embeddings:
+        print()
+        print("=" * 60)
+        print("3. Cross-modal MaxSim scores (text queries × image docs)")
+        print("=" * 60)
+        # Header
+        print(f"{'':>35s}", end="")
+        for j in range(len(doc_embeddings)):
+            print(f"  Doc {j:>2d}", end="")
+        print()
+        # Score matrix
+        for i, q_emb in enumerate(query_embeddings):
+            print(f"  {queries[i]:<33s}", end="")
+            for j, d_emb in enumerate(doc_embeddings):
+                score = compute_maxsim(q_emb, d_emb)
+                print(f"  {score:6.2f}", end="")
+            print()
+
+    # ── 4) Text-only /score endpoint ────────────────────────
+    print()
+    print("=" * 60)
+    print("4. Text-only late interaction scoring (/score endpoint)")
+    print("=" * 60)
+    text_query = "What is the capital of France?"
+    text_docs = [
+        "The capital of France is Paris.",
+        "Berlin is the capital of Germany.",
+        "Python is a programming language.",
+    ]
+    resp = post_http_request(
+        {"model": model, "text_1": text_query, "text_2": text_docs},
+        score_url,
+    )
+    print(f'  Query: "{text_query}"\n')
+    for item in resp.json()["data"]:
+        idx = item["index"]
+        print(f"  Doc {idx} (score={item['score']:.4f}): {text_docs[idx]}")
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

requirements/nightly_torch_test.txt

@@ -27,7 +27,7 @@ mistral_common[image,audio] >= 1.9.1 # required for voxtral test
 num2words # required for smolvlm test
 opencv-python-headless >= 4.13.0 # required for video test
 datamodel_code_generator # required for minicpm3 test
-lm-eval[api]>=0.4.9.2 # required for model evaluation test
+lm-eval[api]>=0.4.11 # required for model evaluation test
 mteb>=1.38.11, <2 # required for mteb test
 transformers==4.57.5
 tokenizers==0.22.0

requirements/rocm-test.txt

@@ -58,7 +58,7 @@ schemathesis==3.39.15
     # OpenAI schema test
 
 # Evaluation and benchmarking
-lm-eval[api]==0.4.9.2
+lm-eval[api]==0.4.11
 jiwer==4.0.0
 
 # Required for multiprocessed tests that use spawn method, Datasets and Evaluate Test
@@ -67,8 +67,6 @@ multiprocess==0.70.16
 # Required for v1/metrics/test_engine_logger_apis.py
 ray[cgraph,default]>=2.48.0
 
-# Plugins test
-terratorch @ git+https://github.com/IBM/terratorch.git@07184fcf91a1324f831ff521dd238d97fe350e3e
 torchgeo==0.7.0
     # via terratorch
 # MTEB Benchmark Test
@@ -96,3 +94,11 @@ albumentations==1.4.6
 transformers==4.57.3
 # Pin HF Hub version
 huggingface-hub==0.36.2
+# Pin Mistral Common
+mistral-common[image,audio]==1.9.1
+# Required for Prithvi tests
+terratorch==1.2.2
+# Required for Prithvi tests
+segmentation-models-pytorch==0.5.0
+# Required for Prithvi tests
+imagehash==4.3.2

requirements/test.in

@@ -35,7 +35,7 @@ num2words # required for smolvlm test
 open_clip_torch==2.32.0 # Required for nemotron_vl test, Nemotron Parse in test_common.py
 opencv-python-headless >= 4.13.0 # required for video test
 datamodel_code_generator # required for minicpm3 test
-lm-eval[api]>=0.4.9.2 # required for model evaluation test
+lm-eval[api]>=0.4.11 # required for model evaluation test
 mteb[bm25s]>=2, <3 # required for mteb test
 transformers==4.57.5
 tokenizers==0.22.0
@@ -56,7 +56,14 @@ runai-model-streamer[s3,gcs]==0.15.3
 fastsafetensors>=0.2.2 # 0.2.2 contains important fixes for multi-GPU mem usage
 pydantic>=2.12 # 2.11 leads to error on python 3.13
 decord==0.6.0
-terratorch @ git+https://github.com/IBM/terratorch.git@1.1.rc3 # required for PrithviMAE test
+terratorch >= 1.2.2 # Required for Prithvi tests
+imagehash # Required for Prithvi tests
+segmentation-models-pytorch > 0.4.0 # Required for Prithvi tests
+
 gpt-oss >= 0.0.7; python_version > '3.11'
 
 perceptron # required for isaac test
+
+# Newer versions of datasets require torchcoded, that makes the tests fail in CI because of a missing library.
+# Older versions are in conflict with teerratorch requirements.
+datasets>=3.3.0,<=3.6.0

requirements/test.txt

@@ -1,11 +1,11 @@
 # This file was autogenerated by uv via the following command:
-#    uv pip compile requirements/test.in -o requirements/test.txt --index-strategy unsafe-best-match --torch-backend cu129 --python-platform x86_64-manylinux_2_28 --python-version 3.12
+#    uv pip compile requirements/test.in -o requirements/test.txt --index-strategy unsafe-best-match --torch-backend cu128 --python-platform x86_64-manylinux_2_28 --python-version 3.12
 absl-py==2.1.0
-    # via rouge-score
-accelerate==1.0.1
     # via
-    #   lm-eval
-    #   peft
+    #   rouge-score
+    #   tensorboard
+accelerate==1.0.1
+    # via peft
 aenum==3.1.16
     # via lightly
 affine==2.4.0
@@ -31,9 +31,7 @@ albumentations==1.4.6
     #   -r requirements/test.in
     #   terratorch
 alembic==1.16.4
-    # via
-    #   mlflow
-    #   optuna
+    # via optuna
 annotated-doc==0.0.4
     # via fastapi
 annotated-types==0.7.0
@@ -74,8 +72,6 @@ bitsandbytes==0.46.1
     #   lightning
 black==24.10.0
     # via datamodel-code-generator
-blinker==1.9.0
-    # via flask
 blobfile==3.0.0
     # via -r requirements/test.in
 bm25s==0.2.13
@@ -93,9 +89,7 @@ bounded-pool-executor==0.0.3
 buildkite-test-collector==0.1.9
     # via -r requirements/test.in
 cachetools==5.5.2
-    # via
-    #   google-auth
-    #   mlflow-skinny
+    # via google-auth
 certifi==2024.8.30
     # via
     #   fiona
@@ -106,6 +100,7 @@ certifi==2024.8.30
     #   pyproj
     #   rasterio
     #   requests
+    #   sentry-sdk
 cffi==1.17.1
     # via soundfile
 chardet==5.2.0
@@ -120,15 +115,14 @@ click==8.1.7
     #   click-plugins
     #   cligj
     #   fiona
-    #   flask
     #   jiwer
-    #   mlflow-skinny
     #   nltk
     #   rasterio
     #   ray
     #   schemathesis
     #   typer
     #   uvicorn
+    #   wandb
 click-plugins==1.1.1.2
     # via
     #   fiona
@@ -137,14 +131,11 @@ cligj==0.7.2
     # via
     #   fiona
     #   rasterio
-cloudpickle==3.1.1
-    # via mlflow-skinny
 colorama==0.4.6
     # via
     #   perceptron
     #   sacrebleu
     #   schemathesis
-    #   tqdm-multiprocess
 colorful==0.5.6
     # via ray
 colorlog==6.10.1
@@ -163,16 +154,15 @@ cupy-cuda12x==13.6.0
     # via ray
 cycler==0.12.1
     # via matplotlib
-databricks-sdk==0.59.0
-    # via mlflow-skinny
 datamodel-code-generator==0.26.3
     # via -r requirements/test.in
 dataproperty==1.0.1
     # via
     #   pytablewriter
     #   tabledata
-datasets==3.0.2
+datasets==3.3.0
     # via
+    #   -r requirements/test.in
     #   evaluate
     #   lm-eval
     #   mteb
@@ -180,6 +170,8 @@ decorator==5.1.1
     # via librosa
 decord==0.6.0
     # via -r requirements/test.in
+diffusers==0.36.0
+    # via terratorch
 dill==0.3.8
     # via
     #   datasets
@@ -191,15 +183,11 @@ distlib==0.3.9
 dnspython==2.7.0
     # via email-validator
 docker==7.1.0
-    # via
-    #   gpt-oss
-    #   mlflow
+    # via gpt-oss
 docopt==0.6.2
     # via num2words
 docstring-parser==0.17.0
     # via jsonargparse
-efficientnet-pytorch==0.7.1
-    # via segmentation-models-pytorch
 einops==0.8.1
     # via
     #   -r requirements/test.in
@@ -217,9 +205,7 @@ encodec==0.1.1
 evaluate==0.4.3
     # via lm-eval
 fastapi==0.128.0
-    # via
-    #   gpt-oss
-    #   mlflow-skinny
+    # via gpt-oss
 fastparquet==2024.11.0
     # via genai-perf
 fastrlock==0.8.2
@@ -230,6 +216,7 @@ filelock==3.16.1
     # via
     #   blobfile
     #   datasets
+    #   diffusers
     #   huggingface-hub
     #   ray
     #   torch
@@ -237,8 +224,6 @@ filelock==3.16.1
     #   virtualenv
 fiona==1.10.1
     # via torchgeo
-flask==3.1.1
-    # via mlflow
 fonttools==4.55.0
     # via matplotlib
 fqdn==1.5.1
@@ -249,7 +234,7 @@ frozenlist==1.5.0
     # via
     #   aiohttp
     #   aiosignal
-fsspec==2024.9.0
+fsspec==2024.12.0
     # via
     #   datasets
     #   evaluate
@@ -257,6 +242,7 @@ fsspec==2024.9.0
     #   huggingface-hub
     #   lightning
     #   pytorch-lightning
+    #   tacoreader
     #   torch
 ftfy==6.3.1
     # via open-clip-torch
@@ -269,7 +255,7 @@ geopandas==1.0.1
 gitdb==4.0.12
     # via gitpython
 gitpython==3.1.44
-    # via mlflow-skinny
+    # via wandb
 google-api-core==2.24.2
     # via
     #   google-cloud-core
@@ -277,7 +263,6 @@ google-api-core==2.24.2
     #   opencensus
 google-auth==2.40.2
     # via
-    #   databricks-sdk
     #   google-api-core
     #   google-cloud-core
     #   google-cloud-storage
@@ -296,25 +281,17 @@ googleapis-common-protos==1.70.0
     # via google-api-core
 gpt-oss==0.0.8
     # via -r requirements/test.in
-graphene==3.4.3
-    # via mlflow
 graphql-core==3.2.6
-    # via
-    #   graphene
-    #   graphql-relay
-    #   hypothesis-graphql
-graphql-relay==3.2.0
-    # via graphene
+    # via hypothesis-graphql
 greenlet==3.2.3
     # via sqlalchemy
 grpcio==1.78.0
     # via
     #   grpcio-tools
     #   ray
+    #   tensorboard
 grpcio-tools==1.78.0
     # via -r requirements/test.in
-gunicorn==23.0.0
-    # via mlflow
 h11==0.14.0
     # via
     #   httpcore
@@ -338,12 +315,14 @@ httpcore==1.0.6
 httpx==0.27.2
     # via
     #   -r requirements/test.in
+    #   diffusers
     #   perceptron
     #   schemathesis
 huggingface-hub==0.36.2
     # via
     #   accelerate
     #   datasets
+    #   diffusers
     #   evaluate
     #   open-clip-torch
     #   peft
@@ -379,11 +358,13 @@ idna==3.10
     #   jsonschema
     #   requests
     #   yarl
+imagehash==4.3.2
+    # via -r requirements/test.in
 imageio==2.37.0
     # via scikit-image
 importlib-metadata==8.7.0
     # via
-    #   mlflow-skinny
+    #   diffusers
     #   opentelemetry-api
 importlib-resources==6.5.2
     # via typeshed-client
@@ -395,14 +376,11 @@ isoduration==20.11.0
     # via jsonschema
 isort==5.13.2
     # via datamodel-code-generator
-itsdangerous==2.2.0
-    # via flask
 jinja2==3.1.6
     # via
     #   datamodel-code-generator
-    #   flask
     #   genai-perf
-    #   mlflow
+    #   lm-eval
     #   torch
 jiwer==3.0.5
     # via -r requirements/test.in
@@ -415,12 +393,14 @@ joblib==1.4.2
     #   librosa
     #   nltk
     #   scikit-learn
-jsonargparse==4.35.0
+jsonargparse==4.46.0
     # via
     #   lightning
     #   terratorch
 jsonlines==4.0.0
     # via lm-eval
+jsonnet==0.21.0
+    # via jsonargparse
 jsonpointer==3.0.0
     # via jsonschema
 jsonschema==4.23.0
@@ -449,13 +429,13 @@ libnacl==2.1.0
     # via tensorizer
 librosa==0.10.2.post1
     # via -r requirements/test.in
-lightly==1.5.20
+lightly==1.5.22
     # via
     #   terratorch
     #   torchgeo
 lightly-utils==0.0.2
     # via lightly
-lightning==2.5.1.post0
+lightning==2.6.1
     # via
     #   terratorch
     #   torchgeo
@@ -466,7 +446,7 @@ lightning-utilities==0.14.3
     #   torchmetrics
 llvmlite==0.44.0
     # via numba
-lm-eval==0.4.9.2
+lm-eval==0.4.11
     # via -r requirements/test.in
 lxml==5.3.0
     # via
@@ -476,12 +456,11 @@ lxml==5.3.0
 mako==1.3.10
     # via alembic
 markdown==3.8.2
-    # via mlflow
+    # via tensorboard
 markdown-it-py==3.0.0
     # via rich
 markupsafe==3.0.1
     # via
-    #   flask
     #   jinja2
     #   mako
     #   werkzeug
@@ -489,7 +468,6 @@ matplotlib==3.9.2
     # via
     #   -r requirements/test.in
     #   lightning
-    #   mlflow
     #   pycocotools
     #   torchgeo
 mbstrdecoder==1.1.3
@@ -501,10 +479,6 @@ mdurl==0.1.2
     # via markdown-it-py
 mistral-common==1.9.1
     # via -r requirements/test.in
-mlflow==2.22.0
-    # via terratorch
-mlflow-skinny==2.22.0
-    # via mlflow
 more-itertools==10.5.0
     # via lm-eval
 mpmath==1.3.0
@@ -523,8 +497,6 @@ multiprocess==0.70.16
     # via
     #   datasets
     #   evaluate
-munch==4.0.0
-    # via pretrainedmodels
 mypy-extensions==1.0.0
     # via black
 networkx==3.2.1
@@ -539,8 +511,6 @@ numba==0.61.2
     # via
     #   -r requirements/test.in
     #   librosa
-numexpr==2.10.1
-    # via lm-eval
 numpy==2.2.6
     # via
     #   -r requirements/test.in
@@ -553,6 +523,7 @@ numpy==2.2.6
     #   cupy-cuda12x
     #   datasets
     #   decord
+    #   diffusers
     #   einx
     #   encodec
     #   evaluate
@@ -560,16 +531,16 @@ numpy==2.2.6
     #   genai-perf
     #   geopandas
     #   h5py
+    #   imagehash
     #   imageio
     #   librosa
     #   lightly
     #   lightly-utils
+    #   lm-eval
     #   matplotlib
     #   mistral-common
-    #   mlflow
     #   mteb
     #   numba
-    #   numexpr
     #   opencv-python-headless
     #   optuna
     #   pandas
@@ -578,6 +549,7 @@ numpy==2.2.6
     #   perceptron
     #   pycocotools
     #   pyogrio
+    #   pywavelets
     #   rasterio
     #   rioxarray
     #   rouge-score
@@ -590,8 +562,10 @@ numpy==2.2.6
     #   shapely
     #   soxr
     #   statsmodels
+    #   tensorboard
     #   tensorboardx
     #   tensorizer
+    #   terratorch
     #   tifffile
     #   torchgeo
     #   torchmetrics
@@ -600,28 +574,28 @@ numpy==2.2.6
     #   tritonclient
     #   vocos
     #   xarray
-nvidia-cublas-cu12==12.9.1.4
+nvidia-cublas-cu12==12.8.4.1
     # via
     #   nvidia-cudnn-cu12
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cuda-cupti-cu12==12.9.79
+nvidia-cuda-cupti-cu12==12.8.90
     # via torch
-nvidia-cuda-nvrtc-cu12==12.9.86
+nvidia-cuda-nvrtc-cu12==12.8.93
     # via torch
-nvidia-cuda-runtime-cu12==12.9.79
+nvidia-cuda-runtime-cu12==12.8.90
     # via torch
 nvidia-cudnn-cu12==9.10.2.21
     # via torch
-nvidia-cufft-cu12==11.4.1.4
+nvidia-cufft-cu12==11.3.3.83
     # via torch
-nvidia-cufile-cu12==1.14.1.1
+nvidia-cufile-cu12==1.13.1.3
     # via torch
-nvidia-curand-cu12==10.3.10.19
+nvidia-curand-cu12==10.3.9.90
     # via torch
-nvidia-cusolver-cu12==11.7.5.82
+nvidia-cusolver-cu12==11.7.3.90
     # via torch
-nvidia-cusparse-cu12==12.5.10.65
+nvidia-cusparse-cu12==12.5.8.93
     # via
     #   nvidia-cusolver-cu12
     #   torch
@@ -629,7 +603,7 @@ nvidia-cusparselt-cu12==0.7.1
     # via torch
 nvidia-nccl-cu12==2.27.5
     # via torch
-nvidia-nvjitlink-cu12==12.9.86
+nvidia-nvjitlink-cu12==12.8.93
     # via
     #   nvidia-cufft-cu12
     #   nvidia-cusolver-cu12
@@ -637,7 +611,7 @@ nvidia-nvjitlink-cu12==12.9.86
     #   torch
 nvidia-nvshmem-cu12==3.4.5
     # via torch
-nvidia-nvtx-cu12==12.9.79
+nvidia-nvtx-cu12==12.8.90
     # via torch
 omegaconf==2.3.0
     # via
@@ -659,7 +633,6 @@ opencv-python-headless==4.13.0.90
     #   mistral-common
 opentelemetry-api==1.35.0
     # via
-    #   mlflow-skinny
     #   opentelemetry-exporter-prometheus
     #   opentelemetry-sdk
     #   opentelemetry-semantic-conventions
@@ -669,7 +642,6 @@ opentelemetry-proto==1.36.0
     # via ray
 opentelemetry-sdk==1.35.0
     # via
-    #   mlflow-skinny
     #   opentelemetry-exporter-prometheus
     #   ray
 opentelemetry-semantic-conventions==0.56b0
@@ -687,7 +659,6 @@ packaging==24.2
     #   evaluate
     #   fastparquet
     #   geopandas
-    #   gunicorn
     #   huggingface-hub
     #   hydra-core
     #   kornia
@@ -695,7 +666,6 @@ packaging==24.2
     #   lightning
     #   lightning-utilities
     #   matplotlib
-    #   mlflow-skinny
     #   optuna
     #   peft
     #   plotly
@@ -708,10 +678,12 @@ packaging==24.2
     #   rioxarray
     #   scikit-image
     #   statsmodels
+    #   tensorboard
     #   tensorboardx
     #   torchmetrics
     #   transformers
     #   typepy
+    #   wandb
     #   xarray
 pandas==2.2.3
     # via
@@ -720,8 +692,8 @@ pandas==2.2.3
     #   fastparquet
     #   genai-perf
     #   geopandas
-    #   mlflow
     #   statsmodels
+    #   tacoreader
     #   torchgeo
     #   xarray
 pathspec==0.12.1
@@ -731,16 +703,16 @@ pathvalidate==3.2.1
 patsy==1.0.1
     # via statsmodels
 peft==0.16.0
-    # via
-    #   -r requirements/test.in
-    #   lm-eval
+    # via -r requirements/test.in
 perceptron==0.1.4
     # via -r requirements/test.in
 perf-analyzer==0.1.0
     # via genai-perf
 pillow==10.4.0
     # via
+    #   diffusers
     #   genai-perf
+    #   imagehash
     #   imageio
     #   lightly-utils
     #   matplotlib
@@ -748,6 +720,7 @@ pillow==10.4.0
     #   perceptron
     #   scikit-image
     #   segmentation-models-pytorch
+    #   tensorboard
     #   torchgeo
     #   torchvision
 platformdirs==4.3.6
@@ -755,6 +728,7 @@ platformdirs==4.3.6
     #   black
     #   pooch
     #   virtualenv
+    #   wandb
 plotly==5.24.1
     # via genai-perf
 pluggy==1.5.0
@@ -769,8 +743,6 @@ portalocker==2.10.1
     # via sacrebleu
 pqdm==0.2.0
     # via -r requirements/test.in
-pretrainedmodels==0.7.4
-    # via segmentation-models-pytorch
 prometheus-client==0.22.0
     # via
     #   opentelemetry-exporter-prometheus
@@ -786,12 +758,13 @@ protobuf==6.33.2
     #   google-api-core
     #   googleapis-common-protos
     #   grpcio-tools
-    #   mlflow-skinny
     #   opentelemetry-proto
     #   proto-plus
     #   ray
+    #   tensorboard
     #   tensorboardx
     #   tensorizer
+    #   wandb
 psutil==6.1.0
     # via
     #   accelerate
@@ -801,19 +774,18 @@ py==1.11.0
     # via pytest-forked
 py-spy==0.4.0
     # via ray
-pyarrow==18.0.0
+pyarrow==23.0.0
     # via
     #   datasets
     #   genai-perf
-    #   mlflow
+    #   tacoreader
+    #   terratorch
 pyasn1==0.6.1
     # via
     #   pyasn1-modules
     #   rsa
 pyasn1-modules==0.4.2
     # via google-auth
-pybind11==2.13.6
-    # via lm-eval
 pycocotools==2.0.8
     # via terratorch
 pycountry==24.6.1
@@ -831,11 +803,11 @@ pydantic==2.12.0
     #   gpt-oss
     #   lightly
     #   mistral-common
-    #   mlflow-skinny
     #   mteb
     #   openai-harmony
     #   pydantic-extra-types
     #   ray
+    #   wandb
 pydantic-core==2.41.1
     # via pydantic
 pydantic-extra-types==2.10.5
@@ -873,7 +845,6 @@ pytest==8.3.5
     #   pytest-subtests
     #   pytest-timeout
     #   schemathesis
-    #   terratorch
 pytest-asyncio==0.24.0
     # via -r requirements/test.in
 pytest-cov==6.3.0
@@ -896,7 +867,6 @@ python-dateutil==2.9.0.post0
     # via
     #   arrow
     #   botocore
-    #   graphene
     #   lightly
     #   matplotlib
     #   pandas
@@ -913,6 +883,8 @@ pytz==2024.2
     # via
     #   pandas
     #   typepy
+pywavelets==1.9.0
+    # via imagehash
 pyyaml==6.0.2
     # via
     #   accelerate
@@ -923,7 +895,6 @@ pyyaml==6.0.2
     #   huggingface-hub
     #   jsonargparse
     #   lightning
-    #   mlflow-skinny
     #   omegaconf
     #   optuna
     #   peft
@@ -934,6 +905,7 @@ pyyaml==6.0.2
     #   timm
     #   transformers
     #   vocos
+    #   wandb
 rapidfuzz==3.12.1
     # via jiwer
 rasterio==1.4.3
@@ -951,6 +923,7 @@ referencing==0.35.1
     #   jsonschema-specifications
 regex==2024.9.11
     # via
+    #   diffusers
     #   nltk
     #   open-clip-torch
     #   sacrebleu
@@ -959,8 +932,8 @@ regex==2024.9.11
 requests==2.32.3
     # via
     #   buildkite-test-collector
-    #   databricks-sdk
     #   datasets
+    #   diffusers
     #   docker
     #   evaluate
     #   google-api-core
@@ -970,15 +943,16 @@ requests==2.32.3
     #   lightly
     #   lm-eval
     #   mistral-common
-    #   mlflow-skinny
     #   mteb
     #   pooch
     #   ray
     #   responses
     #   schemathesis
     #   starlette-testclient
+    #   tacoreader
     #   tiktoken
     #   transformers
+    #   wandb
 responses==0.25.3
     # via genai-perf
 rfc3339-validator==0.1.4
@@ -991,6 +965,7 @@ rich==13.9.4
     #   lightning
     #   mteb
     #   perceptron
+    #   terratorch
     #   typer
 rioxarray==0.19.0
     # via terratorch
@@ -1017,47 +992,55 @@ sacrebleu==2.4.3
 safetensors==0.4.5
     # via
     #   accelerate
+    #   diffusers
     #   open-clip-torch
     #   peft
+    #   segmentation-models-pytorch
     #   timm
     #   transformers
 schemathesis==3.39.15
     # via -r requirements/test.in
 scikit-image==0.25.2
-    # via albumentations
+    # via
+    #   albumentations
+    #   terratorch
 scikit-learn==1.5.2
     # via
     #   albumentations
     #   librosa
     #   lm-eval
-    #   mlflow
     #   mteb
     #   sentence-transformers
+    #   terratorch
 scipy==1.13.1
     # via
     #   albumentations
     #   bm25s
+    #   imagehash
     #   librosa
-    #   mlflow
     #   mteb
     #   scikit-image
     #   scikit-learn
     #   sentence-transformers
     #   statsmodels
     #   vocos
-segmentation-models-pytorch==0.4.0
+segmentation-models-pytorch==0.5.0
     # via
+    #   -r requirements/test.in
     #   terratorch
     #   torchgeo
 sentence-transformers==5.2.0
     # via
     #   -r requirements/test.in
     #   mteb
+sentry-sdk==2.52.0
+    # via wandb
 setuptools==77.0.3
     # via
     #   grpcio-tools
     #   lightning-utilities
     #   pytablewriter
+    #   tensorboard
     #   torch
 shapely==2.1.1
     # via
@@ -1075,7 +1058,6 @@ six==1.16.0
     #   python-dateutil
     #   rfc3339-validator
     #   rouge-score
-    #   segmentation-models-pytorch
 smart-open==7.1.0
     # via ray
 smmap==5.0.2
@@ -1099,12 +1081,9 @@ soxr==0.5.0.post1
 sqlalchemy==2.0.41
     # via
     #   alembic
-    #   mlflow
     #   optuna
 sqlitedict==2.1.0
     # via lm-eval
-sqlparse==0.5.3
-    # via mlflow-skinny
 starlette==0.50.0
     # via
     #   fastapi
@@ -1124,6 +1103,8 @@ tabledata==1.3.3
     # via pytablewriter
 tabulate==0.9.0
     # via sacrebleu
+tacoreader==0.5.6
+    # via terratorch
 tblib==3.1.0
     # via -r requirements/test.in
 tcolorpy==0.1.6
@@ -1133,13 +1114,19 @@ tenacity==9.1.2
     #   gpt-oss
     #   lm-eval
     #   plotly
+tensorboard==2.20.0
+    # via terratorch
+tensorboard-data-server==0.7.2
+    # via tensorboard
 tensorboardx==2.6.4
     # via lightning
 tensorizer==2.10.1
     # via -r requirements/test.in
 termcolor==3.1.0
-    # via gpt-oss
-terratorch @ git+https://github.com/IBM/terratorch.git@07184fcf91a1324f831ff521dd238d97fe350e3e
+    # via
+    #   gpt-oss
+    #   terratorch
+terratorch==1.2.2
     # via -r requirements/test.in
 threadpoolctl==3.5.0
     # via scikit-learn
@@ -1167,21 +1154,18 @@ tomli==2.2.1
     # via schemathesis
 tomli-w==1.2.0
     # via schemathesis
-torch==2.10.0+cu129
+torch==2.10.0+cu128
     # via
     #   -r requirements/test.in
     #   accelerate
     #   bitsandbytes
-    #   efficientnet-pytorch
     #   encodec
     #   kornia
     #   lightly
     #   lightning
-    #   lm-eval
     #   mteb
     #   open-clip-torch
     #   peft
-    #   pretrainedmodels
     #   pytorch-lightning
     #   runai-model-streamer
     #   segmentation-models-pytorch
@@ -1195,7 +1179,7 @@ torch==2.10.0+cu129
     #   torchvision
     #   vector-quantize-pytorch
     #   vocos
-torchaudio==2.10.0+cu129
+torchaudio==2.10.0+cu128
     # via
     #   -r requirements/test.in
     #   encodec
@@ -1208,17 +1192,16 @@ torchmetrics==1.7.4
     #   pytorch-lightning
     #   terratorch
     #   torchgeo
-torchvision==0.25.0+cu129
+torchvision==0.25.0+cu128
     # via
     #   -r requirements/test.in
     #   lightly
     #   open-clip-torch
-    #   pretrainedmodels
     #   segmentation-models-pytorch
     #   terratorch
     #   timm
     #   torchgeo
-tqdm==4.66.6
+tqdm==4.67.3
     # via
     #   datasets
     #   evaluate
@@ -1232,19 +1215,16 @@ tqdm==4.66.6
     #   optuna
     #   peft
     #   pqdm
-    #   pretrainedmodels
     #   pytorch-lightning
     #   segmentation-models-pytorch
     #   sentence-transformers
-    #   tqdm-multiprocess
+    #   tacoreader
+    #   terratorch
     #   transformers
-tqdm-multiprocess==0.0.11
-    # via lm-eval
 transformers==4.57.5
     # via
     #   -r requirements/test.in
     #   genai-perf
-    #   lm-eval
     #   peft
     #   sentence-transformers
     #   transformers-stream-generator
@@ -1274,14 +1254,13 @@ typing-extensions==4.15.0
     #   alembic
     #   chz
     #   fastapi
-    #   graphene
     #   grpcio
     #   huggingface-hub
     #   librosa
     #   lightning
     #   lightning-utilities
+    #   lm-eval
     #   mistral-common
-    #   mlflow-skinny
     #   mteb
     #   opentelemetry-api
     #   opentelemetry-sdk
@@ -1299,6 +1278,7 @@ typing-extensions==4.15.0
     #   typer
     #   typeshed-client
     #   typing-inspection
+    #   wandb
 typing-inspection==0.4.2
     # via pydantic
 tzdata==2024.2
@@ -1313,25 +1293,26 @@ urllib3==2.2.3
     #   lightly
     #   requests
     #   responses
+    #   sentry-sdk
     #   tritonclient
 uvicorn==0.35.0
-    # via
-    #   gpt-oss
-    #   mlflow-skinny
+    # via gpt-oss
 vector-quantize-pytorch==1.21.2
     # via -r requirements/test.in
 virtualenv==20.31.2
     # via ray
 vocos==0.1.0
     # via -r requirements/test.in
+wandb==0.24.2
+    # via terratorch
 wcwidth==0.2.13
     # via ftfy
 webcolors==24.11.1
     # via jsonschema
 werkzeug==3.1.3
     # via
-    #   flask
     #   schemathesis
+    #   tensorboard
 word2number==1.1
     # via lm-eval
 wrapt==1.17.2

scripts/autotune_helion_kernels.py

@@ -0,0 +1,430 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Autotune registered Helion kernels for optimal configurations.
+
+Usage:
+    # Autotune all registered kernels
+    python scripts/autotune_helion_kernels.py
+
+    # Autotune specific kernel
+    python scripts/autotune_helion_kernels.py --kernels silu_mul_fp8
+
+    # Autotune multiple kernels
+    python scripts/autotune_helion_kernels.py --kernels silu_mul_fp8 rms_norm_fp8
+
+    # Force re-autotuning
+    python scripts/autotune_helion_kernels.py --force
+
+    # List available kernels
+    python scripts/autotune_helion_kernels.py --list
+"""
+
+import argparse
+import sys
+import time
+from dataclasses import dataclass
+
+import torch
+
+try:
+    import helion
+
+    from vllm.kernels.helion import (
+        ConfigManager,
+        get_kernel_by_name,
+        get_registered_kernels,
+    )
+    from vllm.kernels.helion.utils import get_canonical_gpu_name
+    from vllm.logger import init_logger
+    from vllm.utils.import_utils import has_helion
+except ImportError as e:
+    print(f"Error importing vLLM: {e}")
+    print("Please ensure vLLM is installed and in your Python path")
+    sys.exit(1)
+
+logger = init_logger("vllm.scripts.autotune_helion_kernels")
+
+
+@dataclass
+class AutotuneResult:
+    status: str  # "success" | "partial" | "error" | "skipped"
+    successful: int
+    failed: int
+    configs: dict[str, "helion.Config"]
+    message: str = ""
+
+
+def list_kernels() -> None:
+    kernels = get_registered_kernels()
+
+    if not kernels:
+        print("No Helion kernels found in registry.")
+        return
+
+    print("Available Helion kernels:")
+    print("=" * 50)
+
+    for name in sorted(kernels.keys()):
+        print(f"  {name}")
+
+    print(f"\nTotal: {len(kernels)} kernels")
+
+
+def check_requirements() -> bool:
+    if not torch.cuda.is_available():
+        logger.error("CUDA is not available. Helion autotuning requires GPU.")
+        return False
+
+    if not has_helion():
+        logger.error("Helion is not installed. Please install Helion package.")
+        return False
+
+    return True
+
+
+def autotune_kernel(
+    kernel_name: str,
+    platform: str,
+    config_manager: ConfigManager,
+    force: bool = False,
+    autotune_effort: str = "quick",
+) -> AutotuneResult:
+    logger.debug(
+        "Starting autotune for kernel '%s' with effort='%s'",
+        kernel_name,
+        autotune_effort,
+    )
+    kernel_wrapper = get_kernel_by_name(kernel_name)
+    if kernel_wrapper is None:
+        error_msg = f"Kernel '{kernel_name}' not found in registry"
+        logger.error(error_msg)
+        return AutotuneResult(
+            status="error",
+            message=error_msg,
+            successful=0,
+            failed=0,
+            configs={},
+        )
+
+    try:
+        inputs_dict = kernel_wrapper.get_inputs()
+    except NotImplementedError:
+        error_msg = f"Kernel '{kernel_name}' has no input generator registered"
+        logger.error(error_msg)
+        return AutotuneResult(
+            status="error",
+            message=error_msg,
+            successful=0,
+            failed=0,
+            configs={},
+        )
+
+    try:
+        logger.info(
+            "Autotuning kernel '%s' for platform '%s' with %d configs",
+            kernel_name,
+            platform,
+            len(inputs_dict),
+        )
+
+        configs_to_autotune = {}
+        if not force:
+            existing_configs = config_manager.get_platform_configs(
+                kernel_name, platform
+            )
+            for config_key, inputs in inputs_dict.items():
+                if config_key in existing_configs:
+                    logger.debug(
+                        "Config '%s' already exists for platform '%s', skipping",
+                        config_key,
+                        platform,
+                    )
+                else:
+                    configs_to_autotune[config_key] = inputs
+        else:
+            logger.debug("Force mode enabled, will re-autotune all configs")
+            configs_to_autotune = inputs_dict
+
+        if not configs_to_autotune:
+            logger.info(
+                "All configs already exist for kernel '%s' on platform '%s'. "
+                "Use --force to re-autotune.",
+                kernel_name,
+                platform,
+            )
+            return AutotuneResult(
+                status="skipped",
+                message="All configs already exist",
+                successful=0,
+                failed=0,
+                configs={},
+            )
+
+        total_start_time = time.time()
+        autotuned_configs = {}
+        failed_configs = []
+
+        for config_key, inputs in configs_to_autotune.items():
+            logger.info("Autotuning config: %s", config_key)
+            logger.debug(
+                "Input shapes: %s",
+                [getattr(inp, "shape", type(inp).__name__) for inp in inputs],
+            )
+
+            try:
+                config_start_time = time.time()
+                config = kernel_wrapper.run_autotune(inputs, autotune_effort)
+                config_duration = time.time() - config_start_time
+
+                # Save immediately for checkpointing
+                config_manager.save_configs(kernel_name, platform, {config_key: config})
+
+                autotuned_configs[config_key] = config
+                logger.debug("Config details: %s", config)
+
+                logger.info(
+                    "✓ Autotuned and saved config '%s' (%.2fs)",
+                    config_key,
+                    config_duration,
+                )
+
+            except (RuntimeError, ValueError, OSError) as e:
+                logger.exception(
+                    "Failed to autotune config '%s': %s",
+                    config_key,
+                    e,
+                )
+                failed_configs.append(config_key)
+
+        total_duration = time.time() - total_start_time
+        successful = len(autotuned_configs)
+        failed = len(failed_configs)
+
+        logger.info(
+            "Completed autotuning for kernel '%s': %d successful, %d failed (%.2fs)",
+            kernel_name,
+            successful,
+            failed,
+            total_duration,
+        )
+
+        status = "success" if failed == 0 else "partial"
+        return AutotuneResult(
+            status=status,
+            successful=successful,
+            failed=failed,
+            configs=autotuned_configs,
+        )
+
+    except (KeyError, RuntimeError, ValueError, OSError) as e:
+        error_msg = f"Unexpected error: {e}"
+        logger.exception("Failed to autotune kernel '%s': %s", kernel_name, e)
+        return AutotuneResult(
+            status="error",
+            message=error_msg,
+            successful=0,
+            failed=0,
+            configs={},
+        )
+
+
+def summarize_results(results: dict[str, AutotuneResult]) -> bool:
+    logger.info("=" * 50)
+    logger.info("Autotuning Results Summary")
+    logger.info("=" * 50)
+
+    total_successful = 0
+    total_failed = 0
+    success_kernels = []
+    partial_kernels = []
+    error_kernels = []
+    skipped_kernels = []
+
+    for kernel_name, result in results.items():
+        total_successful += result.successful
+        total_failed += result.failed
+
+        if result.status == "success":
+            success_kernels.append(f"{kernel_name} ({result.successful} configs)")
+            logger.info("✓ %s: %d configs successful", kernel_name, result.successful)
+        elif result.status == "partial":
+            partial_kernels.append(
+                f"{kernel_name} ({result.successful} ok, {result.failed} failed)"
+            )
+            logger.warning(
+                "⚠ %s: %d successful, %d failed",
+                kernel_name,
+                result.successful,
+                result.failed,
+            )
+        elif result.status == "error":
+            error_kernels.append(f"{kernel_name}: {result.message or 'Unknown error'}")
+            logger.error("✗ %s: %s", kernel_name, result.message or "Unknown error")
+        elif result.status == "skipped":
+            skipped_kernels.append(f"{kernel_name}: {result.message or 'Skipped'}")
+            logger.info("- %s: %s", kernel_name, result.message or "Skipped")
+
+    logger.info("=" * 50)
+    logger.info(
+        "Summary: %d total configs (%d successful, %d failed)",
+        total_successful + total_failed,
+        total_successful,
+        total_failed,
+    )
+    logger.info(
+        "Kernels: %d success, %d partial, %d error, %d skipped",
+        len(success_kernels),
+        len(partial_kernels),
+        len(error_kernels),
+        len(skipped_kernels),
+    )
+
+    has_failures = bool(error_kernels or partial_kernels)
+
+    if not has_failures:
+        if total_successful > 0:
+            logger.info("All configs autotuned successfully!")
+        else:
+            logger.info("No new configs were generated (all may already exist)")
+
+    return not has_failures
+
+
+def get_kernels_to_autotune(requested_kernels: list[str] | None) -> list[str]:
+    all_kernels = get_registered_kernels()
+    if not all_kernels:
+        logger.error("No Helion kernels found in registry")
+        sys.exit(1)
+
+    if not requested_kernels:
+        return list(all_kernels.keys())
+
+    if len(requested_kernels) != len(set(requested_kernels)):
+        duplicates = [
+            k for k in set(requested_kernels) if requested_kernels.count(k) > 1
+        ]
+        logger.error("Duplicate kernel names in --kernels flag: %s", duplicates)
+        sys.exit(1)
+
+    kernels_to_autotune = []
+    missing_kernels = []
+
+    for kernel_name in requested_kernels:
+        if kernel_name in all_kernels:
+            kernels_to_autotune.append(kernel_name)
+        else:
+            missing_kernels.append(kernel_name)
+
+    if missing_kernels:
+        logger.error("Kernel(s) not found: %s", missing_kernels)
+        logger.error("Available kernels: %s", list(all_kernels.keys()))
+        sys.exit(1)
+
+    return kernels_to_autotune
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Autotune Helion kernels",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=__doc__.split("Usage:")[1] if "Usage:" in __doc__ else "",
+    )
+
+    parser.add_argument(
+        "--kernels",
+        nargs="+",
+        help="Kernel(s) to autotune (default: all kernels)",
+    )
+
+    parser.add_argument(
+        "--config-dir",
+        type=str,
+        help="Config directory for config files (default: vLLM helion configs dir)",
+    )
+
+    parser.add_argument(
+        "--list",
+        action="store_true",
+        help="List available Helion kernels and exit",
+    )
+
+    parser.add_argument(
+        "--force",
+        action="store_true",
+        help=(
+            "Force re-autotuning even if configs already exist for the "
+            "platform and config keys"
+        ),
+    )
+
+    parser.add_argument(
+        "--autotune-effort",
+        type=str,
+        default="quick",
+        help=(
+            "Helion autotune effort level: 'quick' (smaller search) or "
+            "'full' (full search budget) (default: quick)"
+        ),
+    )
+
+    parser.add_argument(
+        "--verbose",
+        action="store_true",
+        help="Enable verbose logging",
+    )
+
+    args = parser.parse_args()
+
+    import logging
+
+    if args.verbose:
+        logging.getLogger("vllm").setLevel(logging.DEBUG)
+        logger.debug("Verbose mode enabled")
+        logger.debug("Arguments: %s", vars(args))
+    else:
+        logging.getLogger("vllm").setLevel(logging.INFO)
+
+    if args.list:
+        list_kernels()
+        return
+
+    if not check_requirements():
+        sys.exit(1)
+
+    platform = get_canonical_gpu_name()
+    logger.info("Detected GPU platform: %s", platform)
+
+    config_manager = (
+        ConfigManager(args.config_dir) if args.config_dir else ConfigManager()
+    )
+
+    try:
+        config_manager.ensure_base_dir_writable()
+    except OSError as e:
+        logger.error("Failed to access config directory: %s", e)
+        sys.exit(1)
+
+    kernels_to_autotune = get_kernels_to_autotune(args.kernels)
+
+    logger.info(
+        "Will autotune %d kernel(s) for platform '%s': %s",
+        len(kernels_to_autotune),
+        platform,
+        kernels_to_autotune,
+    )
+
+    results = {}
+    for kernel_name in kernels_to_autotune:
+        result = autotune_kernel(
+            kernel_name, platform, config_manager, args.force, args.autotune_effort
+        )
+        results[kernel_name] = result
+
+    success = summarize_results(results)
+    sys.exit(0 if success else 1)
+
+
+if __name__ == "__main__":
+    main()

tests/basic_correctness/test_cpu_offload.py

@@ -1,10 +1,29 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import pytest
+
 from ..utils import compare_two_settings
 
 
-def test_cpu_offload():
+@pytest.mark.parametrize("disable_pin_memory", [False, True])
+@pytest.mark.parametrize("disable_uva", [False, True])
+def test_cpu_offload(disable_pin_memory, disable_uva):
+    env_vars = {
+        "VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY": str(int(disable_pin_memory)),
+        "VLLM_WEIGHT_OFFLOADING_DISABLE_UVA": str(int(disable_uva)),
+    }
+
+    args = ["--cpu-offload-gb", "1"]
+
+    # cuda graph only works with UVA offloading
+    if disable_uva:
+        args.append("--enforce-eager")
+
     compare_two_settings(
-        "hmellor/tiny-random-LlamaForCausalLM", [], ["--cpu-offload-gb", "1"]
+        model="hmellor/tiny-random-LlamaForCausalLM",
+        arg1=[],
+        arg2=args,
+        env1=None,
+        env2=env_vars,
     )

tests/compile/correctness_e2e/test_sequence_parallel.py

@@ -229,7 +229,7 @@ def _compare_sp(
     if chunked_prefill:
         common_args.append("--enable-chunked-prefill")
     if eager_mode:
-        common_args.append("--enforce-eager")
+        common_args.append("-cc.cudagraph_mode=none")
     if runner != "auto":
         common_args.extend(["--runner", runner])
     if trust_remote_code:

tests/config/test_multimodal_config.py

@@ -3,6 +3,7 @@
 
 import pytest
 
+from vllm.config.model import ModelConfig
 from vllm.config.multimodal import MultiModalConfig
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
@@ -23,3 +24,20 @@ def test_mm_encoder_attn_backend_hash_updates():
         mm_encoder_attn_backend=AttentionBackendEnum.FLASH_ATTN
     ).compute_hash()
     assert base_hash != overridden_hash
+
+
+def test_language_model_only_does_not_affect_mm_hash():
+    """language_model_only does not affect the ViT computation graph,
+    so it should not change the multimodal config hash."""
+    base_hash = MultiModalConfig().compute_hash()
+    lm_only_hash = MultiModalConfig(language_model_only=True).compute_hash()
+    assert base_hash == lm_only_hash
+
+
+def test_language_model_only_affects_model_hash():
+    """language_model_only affects the LM computation graph,
+    so it should change the model config hash."""
+    model = "llava-hf/llava-1.5-7b-hf"
+    base_hash = ModelConfig(model).compute_hash()
+    lm_only_hash = ModelConfig(model, language_model_only=True).compute_hash()
+    assert base_hash != lm_only_hash

tests/conftest.py

@@ -419,7 +419,6 @@ class HfRunner:
             self.tokenizer: "PreTrainedTokenizer | PreTrainedTokenizerFast" = (
                 AutoTokenizer.from_pretrained(
                     model_name,
-                    dtype=dtype,
                     trust_remote_code=trust_remote_code,
                 )
             )
@@ -430,7 +429,6 @@ class HfRunner:
 
         self.processor = AutoProcessor.from_pretrained(
             model_name,
-            dtype=dtype,
             trust_remote_code=trust_remote_code,
         )
         if skip_tokenizer_init:

tests/detokenizer/test_min_tokens.py

@@ -39,7 +39,6 @@ def test_min_tokens_with_stop(min_tokens: int, stop: str, truth: str):
         mm_features=None,
         sampling_params=params,
         pooling_params=None,
-        eos_token_id=None,
         arrival_time=0.0,
         lora_request=None,
         cache_salt=None,

tests/detokenizer/test_stop_string_while_stop_model_terminates.py

@@ -35,7 +35,6 @@ def _make_request(stop, include_stop_str_in_output: bool, min_tokens: int = 0):
         mm_features=None,
         sampling_params=params,
         pooling_params=None,
-        eos_token_id=None,
         arrival_time=0.0,
         lora_request=None,
         cache_salt=None,

tests/distributed/test_comm_ops.py

@@ -19,6 +19,8 @@ from vllm.distributed import (
     tensor_model_parallel_all_reduce,
     tensor_model_parallel_reduce_scatter,
 )
+from vllm.distributed.parallel_state import GroupCoordinator, TensorMetadata
+from vllm.v1.worker.gpu_worker import AsyncIntermediateTensors
 
 from ..utils import (
     init_test_distributed_environment,
@@ -200,6 +202,111 @@ def send_recv_tensor_dict_test_worker(
         torch.testing.assert_close(recv_dict["f"], test_dict["f"])
 
 
+class _DummyWork:
+    def __init__(self) -> None:
+        self.wait_calls = 0
+
+    def wait(self) -> None:
+        self.wait_calls += 1
+
+
+class _DummyAllGatherGroup:
+    def __init__(self, world_size: int, rank_in_group: int) -> None:
+        self.world_size = world_size
+        self.rank_in_group = rank_in_group
+
+    def all_gather(self, t: torch.Tensor, dim: int = 0) -> torch.Tensor:
+        # duplicate local slice across ranks.
+        assert dim == 0
+        return torch.cat([t for _ in range(self.world_size)], dim=0)
+
+
+def _make_group_for_unit_test(
+    rank_in_group: int = 0, world_size: int = 2
+) -> GroupCoordinator:
+    # avoid running GroupCoordinator.__init__ (it wires up real process groups).
+    g = GroupCoordinator.__new__(GroupCoordinator)
+    g.world_size = world_size
+    g.rank_in_group = rank_in_group
+    g.ranks = list(range(world_size))
+    g.use_cpu_custom_send_recv = False
+    g.device_group = None
+    g.cpu_group = None
+    return g
+
+
+def test_irecv_tensor_dict_send_allgather_postprocess_binds_keys(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    def fake_irecv(t: torch.Tensor, *args: Any, **kwargs: Any) -> _DummyWork:
+        t.fill_(1)
+        return _DummyWork()
+
+    monkeypatch.setattr(torch.distributed, "is_initialized", lambda: True)
+    monkeypatch.setattr(torch.distributed, "irecv", fake_irecv)
+
+    g = _make_group_for_unit_test(rank_in_group=0, world_size=2)
+    # 2 tensors so we can catch late-binding bugs in postprocess closures.
+    metadata_list = [
+        ("a", TensorMetadata("cpu", torch.int32, torch.Size([4]))),
+        ("b", TensorMetadata("cpu", torch.int32, torch.Size([4]))),
+    ]
+    g.recv_object = lambda src=None: metadata_list  # type: ignore[method-assign]
+
+    ag = _DummyAllGatherGroup(world_size=2, rank_in_group=0)
+    td, handles, postprocess = g.irecv_tensor_dict(all_gather_group=ag)
+
+    assert td is not None
+    assert len(handles) == 2
+    assert len(postprocess) == 2
+
+    # before postprocess, dict holds the TP slice (shape 2).
+    assert td["a"].shape == torch.Size([2])
+    assert td["b"].shape == torch.Size([2])
+
+    # simulate worker-side "defer wait": wait + postprocess later.
+    for handle in handles:
+        handle.wait()
+    for fn in postprocess:
+        fn()
+
+    # after postprocess, dict values are reconstructed to full shape (shape 4),
+    # and each key should be updated independently
+    assert td["a"].shape == torch.Size([4])
+    assert td["b"].shape == torch.Size([4])
+    torch.testing.assert_close(td["a"], torch.ones(4, dtype=torch.int32))
+    torch.testing.assert_close(td["b"], torch.ones(4, dtype=torch.int32))
+
+
+def test_async_intermediate_tensors_lazy_wait() -> None:
+    work = _DummyWork()
+    post_calls = {"n": 0}
+
+    def post() -> None:
+        post_calls["n"] += 1
+
+    it = AsyncIntermediateTensors(
+        {"x": torch.tensor([1])},
+        comm_handles=[work],
+        comm_postprocess=[post],
+    )
+
+    # accessing non-tensor attributes should not trigger wait.
+    assert it.kv_connector_output is None
+    assert work.wait_calls == 0
+    assert post_calls["n"] == 0
+
+    # first access of `.tensors` triggers wait + postprocess.
+    _ = it.tensors
+    assert work.wait_calls == 1
+    assert post_calls["n"] == 1
+
+    # subsequent access should not re-wait.
+    _ = it.tensors
+    assert work.wait_calls == 1
+    assert post_calls["n"] == 1
+
+
 @ray.remote(num_gpus=1, max_calls=1)
 def send_recv_test_worker(
     monkeypatch: pytest.MonkeyPatch,

tests/entrypoints/instrumentator/test_basic.py

@@ -145,6 +145,7 @@ async def test_request_cancellation(server: RemoteOpenAIServer):
                 model=MODEL_NAME,
                 max_tokens=10000,
                 extra_body={"min_tokens": 10000},
+                temperature=0.0,
             )
         )
         tasks.append(task)
@@ -163,7 +164,7 @@ async def test_request_cancellation(server: RemoteOpenAIServer):
     # be able to respond to this one within the timeout
     client = server.get_async_client(timeout=5)
     response = await client.chat.completions.create(
-        messages=chat_input, model=MODEL_NAME, max_tokens=10
+        messages=chat_input, model=MODEL_NAME, max_tokens=10, temperature=0.0
     )
 
     assert len(response.choices) == 1

tests/entrypoints/instrumentator/test_metrics.py

@@ -17,6 +17,7 @@ from transformers import AutoTokenizer
 from tests.conftest import LocalAssetServer
 from tests.utils import RemoteOpenAIServer
 from vllm import version
+from vllm.utils.network_utils import get_open_port
 
 MODELS = {
     "text": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
@@ -315,14 +316,26 @@ async def test_abort_metrics_reset(
             client.completions.create(
                 model=model_name,
                 prompt=prompt_ids,
-                max_tokens=100,  # Long generation to give time to abort
+                max_tokens=500,  # Long generation to give time to abort
                 temperature=0.0,
             )
         )
         tasks.append(task)
 
-    # Wait a bit for requests to start processing
-    await asyncio.sleep(0.5)
+    # Poll until we see running requests rather than using a fixed sleep,
+    # since generation speed varies across hardware.
+    try:
+        await _poll_until(
+            lambda: _get_running_metrics_from_api(server)[0] > 0,
+            timeout=10.0,
+            interval=0.1,
+            description="running_requests > 0",
+        )
+    except TimeoutError:
+        for task in tasks:
+            task.cancel()
+        await asyncio.gather(*tasks, return_exceptions=True)
+        pytest.fail("Requests never appeared as running in metrics")
 
     # Check that we have running requests
     running_requests, waiting_requests, kv_cache_usage = _get_running_metrics_from_api(
@@ -336,13 +349,15 @@ async def test_abort_metrics_reset(
     # Cancel all tasks to abort the requests
     for task in tasks:
         task.cancel()
+    await asyncio.gather(*tasks, return_exceptions=True)
 
-    # Wait for cancellations to be processed
-    await asyncio.sleep(1.0)
-
-    # Check that metrics have reset to zero
-    response = requests.get(server.url_for("metrics"))
-    assert response.status_code == HTTPStatus.OK
+    # Poll until metrics reset rather than using a fixed sleep
+    await _poll_until(
+        lambda: _get_running_metrics_from_api(server) == (0, 0, 0),
+        timeout=10.0,
+        interval=0.2,
+        description="gauge metrics back to zero",
+    )
 
     # Verify running and waiting requests counts and KV cache usage are zero
     running_requests_after, waiting_requests_after, kv_cache_usage_after = (
@@ -360,6 +375,18 @@ async def test_abort_metrics_reset(
     )
 
 
+async def _poll_until(
+    predicate, *, timeout: float, interval: float = 0.5, description: str = "condition"
+):
+    """Poll until predicate() returns True, or raise TimeoutError."""
+    start = time.time()
+    while time.time() - start < timeout:
+        if predicate():
+            return
+        await asyncio.sleep(interval)
+    raise TimeoutError(f"Timed out after {timeout}s waiting for: {description}")
+
+
 def _get_running_metrics_from_api(server: RemoteOpenAIServer):
     """Return (running_count, waiting_count, kv_cache_usage)"""
 
@@ -399,7 +426,7 @@ def test_metrics_exist_run_batch():
     input_batch = """{"custom_id": "request-0", "method": "POST", "url": "/v1/embeddings", "body": {"model": "intfloat/multilingual-e5-small", "input": "You are a helpful assistant."}}"""  # noqa: E501
 
     base_url = "0.0.0.0"
-    port = "8001"
+    port = str(get_open_port())
     server_url = f"http://{base_url}:{port}"
 
     with (
@@ -427,17 +454,32 @@ def test_metrics_exist_run_batch():
             ],
         )
 
-        def is_server_up(url):
+        try:
+
+            def is_server_up(url):
+                try:
+                    response = requests.get(url)
+                    return response.status_code == 200
+                except requests.ConnectionError:
+                    return False
+
+            start = time.time()
+            timeout = 120
+            while not is_server_up(server_url):
+                if proc.poll() is not None:
+                    pytest.fail(
+                        f"Batch process exited early with returncode={proc.returncode}"
+                    )
+                if time.time() - start > timeout:
+                    pytest.fail("Batch server did not start within timeout")
+                time.sleep(1)
+
+            response = requests.get(server_url + "/metrics")
+            assert response.status_code == HTTPStatus.OK
+        finally:
+            proc.terminate()
             try:
-                response = requests.get(url)
-                return response.status_code == 200
-            except requests.ConnectionError:
-                return False
-
-        while not is_server_up(server_url):
-            time.sleep(1)
-
-        response = requests.get(server_url + "/metrics")
-        assert response.status_code == HTTPStatus.OK
-
-        proc.wait()
+                proc.wait(timeout=15)
+            except subprocess.TimeoutExpired:
+                proc.kill()
+                proc.wait(timeout=5)

tests/entrypoints/llm/test_chat.py

@@ -195,18 +195,15 @@ def test_chat_batch_failure_cleanup(llm_for_failure_test):
     valid_msg = [{"role": "user", "content": "Hello"}]
     long_text = "This is a very long text to test the error " * 50
     invalid_msg = [{"role": "user", "content": long_text}]
-    batch_1 = [
-        valid_msg,
-        valid_msg,
-        invalid_msg,
-    ]
-    batch_2 = [
-        valid_msg,
-        valid_msg,
-    ]
+
+    batch_1 = [valid_msg, valid_msg, invalid_msg]
+    batch_2 = [valid_msg, valid_msg]
     sampling_params = SamplingParams(temperature=0, max_tokens=10)
+
     with pytest.raises(ValueError, match="context length is only"):
         llm.chat(batch_1, sampling_params=sampling_params)
+    assert llm.llm_engine.get_num_unfinished_requests() == 0
+
     outputs_2 = llm.chat(batch_2, sampling_params=sampling_params)
     assert len(outputs_2) == len(batch_2)
     assert llm.llm_engine.get_num_unfinished_requests() == 0

tests/entrypoints/openai/responses/test_harmony.py

@@ -6,7 +6,6 @@ import time
 
 import pytest
 import pytest_asyncio
-import requests
 from openai import BadRequestError, NotFoundError, OpenAI
 from openai_harmony import (
     Message,
@@ -513,11 +512,9 @@ async def test_code_interpreter(client: OpenAI, model_name: str):
 
 
 def get_weather(latitude, longitude):
-    response = requests.get(
-        f"https://api.open-meteo.com/v1/forecast?latitude={latitude}&longitude={longitude}&current=temperature_2m,wind_speed_10m&hourly=temperature_2m,relative_humidity_2m,wind_speed_10m"  # noqa
-    )
-    data = response.json()
-    return data["current"]["temperature_2m"]
+    # Return a static temperature value to avoid flaky SSL/network errors
+    # from calling the external api.open-meteo.com API in CI.
+    return 15.0
 
 
 def get_place_to_travel():

tests/entrypoints/openai/responses/test_parsable_context.py

@@ -172,19 +172,26 @@ async def test_mcp_tool_call(client: OpenAI, model_name: str):
 
     assert response is not None
     assert response.status == "completed"
-    assert response.output[0].type == "reasoning"
-    assert response.output[1].type == "mcp_call"
-    assert type(response.output[1].arguments) is str
-    assert type(response.output[1].output) is str
-    assert response.output[2].type == "reasoning"
-    # make sure the correct math is in the final output
-    assert response.output[3].type == "message"
-    assert any(s in response.output[3].content[0].text for s in ("56088", "56,088"))
 
-    # test raw input_messages / output_messages
+    # The model may produce multiple reasoning/mcp_call rounds before the
+    # final message, so validate structurally rather than by exact index.
+    output_types = [o.type for o in response.output]
+    assert "reasoning" in output_types
+    mcp_calls = [o for o in response.output if o.type == "mcp_call"]
+    assert len(mcp_calls) >= 1
+    assert type(mcp_calls[0].arguments) is str
+    assert type(mcp_calls[0].output) is str
+
+    # The final output should be a message containing the correct answer
+    assert response.output[-1].type == "message"
+    assert any(s in response.output[-1].content[0].text for s in ("56088", "56,088"))
+
+    # Test raw input_messages / output_messages
     assert len(response.input_messages) == 1
-    assert len(response.output_messages) == 3
-    assert any(s in response.output_messages[2]["message"] for s in ("56088", "56,088"))
+    assert len(response.output_messages) >= 3
+    assert any(
+        s in response.output_messages[-1]["message"] for s in ("56088", "56,088")
+    )
 
 
 @pytest.mark.asyncio

tests/entrypoints/openai/responses/test_sampling_params.py

@@ -4,8 +4,17 @@
 """Unit tests for ResponsesRequest.to_sampling_params() parameter mapping."""
 
 import pytest
+import torch
+from openai.types.responses.response_format_text_json_schema_config import (
+    ResponseFormatTextJSONSchemaConfig,
+)
+from pydantic import ValidationError
 
-from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
+from vllm.entrypoints.openai.responses.protocol import (
+    ResponsesRequest,
+    ResponseTextConfig,
+)
+from vllm.sampling_params import StructuredOutputsParams
 
 
 class TestResponsesRequestSamplingParams:
@@ -76,9 +85,6 @@ class TestResponsesRequestSamplingParams:
 
     def test_seed_bounds_validation(self):
         """Test that seed values outside torch.long bounds are rejected."""
-        import torch
-        from pydantic import ValidationError
-
         # Test seed below minimum
         with pytest.raises(ValidationError) as exc_info:
             ResponsesRequest(
@@ -111,3 +117,40 @@ class TestResponsesRequestSamplingParams:
             seed=torch.iinfo(torch.long).max,
         )
         assert request_max.seed == torch.iinfo(torch.long).max
+
+    def test_structured_outputs_passed_through(self):
+        """Test that structured_outputs field is passed to SamplingParams."""
+        structured_outputs = StructuredOutputsParams(grammar="root ::= 'hello'")
+        request = ResponsesRequest(
+            model="test-model",
+            input="test input",
+            structured_outputs=structured_outputs,
+        )
+
+        sampling_params = request.to_sampling_params(default_max_tokens=1000)
+
+        assert sampling_params.structured_outputs is not None
+        assert sampling_params.structured_outputs.grammar == "root ::= 'hello'"
+
+    def test_structured_outputs_and_json_schema_conflict(self):
+        """Test that specifying both structured_outputs and json_schema raises."""
+        structured_outputs = StructuredOutputsParams(grammar="root ::= 'hello'")
+        text_config = ResponseTextConfig()
+        text_config.format = ResponseFormatTextJSONSchemaConfig(
+            type="json_schema",
+            name="test",
+            schema={"type": "object"},
+        )
+        request = ResponsesRequest(
+            model="test-model",
+            input="test input",
+            structured_outputs=structured_outputs,
+            text=text_config,
+        )
+
+        with pytest.raises(ValueError) as exc_info:
+            request.to_sampling_params(default_max_tokens=1000)
+
+        assert "Cannot specify both structured_outputs and text.format" in str(
+            exc_info.value
+        )

tests/entrypoints/openai/test_chat_error.py

@@ -54,16 +54,22 @@ class MockModelConfig:
     media_io_kwargs: dict[str, dict[str, Any]] = field(default_factory=dict)
     skip_tokenizer_init = False
     is_encoder_decoder: bool = False
+    is_multimodal_model: bool = False
 
     def get_diff_sampling_param(self):
         return self.diff_sampling_param or {}
 
 
+@dataclass
+class MockVllmConfig:
+    model_config: MockModelConfig
+
+
 def _build_renderer(model_config: MockModelConfig):
     _, tokenizer_name, _, kwargs = tokenizer_args_from_config(model_config)
 
-    return HfRenderer(
-        model_config,
+    return HfRenderer.from_config(
+        MockVllmConfig(model_config),
         tokenizer_kwargs={**kwargs, "tokenizer_name": tokenizer_name},
     )