[Bugfix] Fix NPU SDPA attention mask shape and semantics (#1031 )

Signed-off-by: gcanlin <canlinguosdu@gmail.com> Co-authored-by: muziyuhui666 <111362884+muziyuhui666@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
[FEATURE] /v1/images/edit interface (#1101 )
2026-01-31 14:14:15 +08:00 · 2026-01-31 14:13:36 +08:00 · 2026-01-31 04:26:34 +00:00 · 2026-01-31 10:42:09 +08:00 · 2026-01-31 10:15:26 +08:00 · 2026-01-31 09:46:23 +08:00
530 changed files with 99323 additions and 13280 deletions
@@ -0,0 +1,238 @@
+#!/bin/bash
+# vllm-omni customized version
+# Based on: https://github.com/vllm-project/ci-infra/blob/main/buildkite/bootstrap-amd.sh
+# Last synced: 2025-12-15
+# Modifications: Use local template file instead of downloading from ci-infra
+
+set -euo pipefail
+
+if [[ -z "${RUN_ALL:-}" ]]; then
+    RUN_ALL=0
+fi
+
+if [[ -z "${NIGHTLY:-}" ]]; then
+    NIGHTLY=0
+fi
+
+if [[ -z "${VLLM_CI_BRANCH:-}" ]]; then
+    VLLM_CI_BRANCH="main"
+fi
+
+if [[ -z "${AMD_MIRROR_HW:-}" ]]; then
+    AMD_MIRROR_HW="amdproduction"
+fi
+
+if [[ -z "${DOCS_ONLY_DISABLE:-}" ]]; then
+    DOCS_ONLY_DISABLE=0
+fi
+
+fail_fast() {
+    DISABLE_LABEL="ci-no-fail-fast"
+    # If BUILDKITE_PULL_REQUEST != "false", then we check the PR labels using curl and jq
+    if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
+        PR_LABELS=$(curl -s "https://api.github.com/repos/vllm-project/vllm-omni/pulls/$BUILDKITE_PULL_REQUEST" | jq -r '.labels[].name')
+        if [[ $PR_LABELS == *"$DISABLE_LABEL"* ]]; then
+            echo false
+        else
+            echo true
+        fi
+    else
+        echo false  # not a PR or BUILDKITE_PULL_REQUEST not set
+    fi
+}
+
+check_run_all_label() {
+    RUN_ALL_LABEL="ready-run-all-tests"
+    # If BUILDKITE_PULL_REQUEST != "false", then we check the PR labels using curl and jq
+    if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
+        PR_LABELS=$(curl -s "https://api.github.com/repos/vllm-project/vllm-omni/pulls/$BUILDKITE_PULL_REQUEST" | jq -r '.labels[].name')
+        if [[ $PR_LABELS == *"$RUN_ALL_LABEL"* ]]; then
+            echo true
+        else
+            echo false
+        fi
+    else
+        echo false  # not a PR or BUILDKITE_PULL_REQUEST not set
+    fi
+}
+
+if [[ -z "${COV_ENABLED:-}" ]]; then
+    COV_ENABLED=0
+fi
+
+upload_pipeline() {
+    echo "Uploading pipeline..."
+    # Install minijinja
+    ls .buildkite || buildkite-agent annotate --style error 'Please merge upstream main branch for buildkite CI'
+    curl -sSfL https://github.com/mitsuhiko/minijinja/releases/download/2.3.1/minijinja-cli-installer.sh | sh
+    source /var/lib/buildkite-agent/.cargo/env
+
+    if [[ $BUILDKITE_PIPELINE_SLUG == "fastcheck" ]]; then
+        AMD_MIRROR_HW="amdtentative"
+    fi
+
+    # Use local template file for vllm-omni
+    cp .buildkite/test-template-amd-omni.j2 .buildkite/test-template.j2
+
+
+    # (WIP) Use pipeline generator instead of jinja template
+    if [ -e ".buildkite/pipeline_generator/pipeline_generator.py" ]; then
+        python -m pip install click pydantic
+        python .buildkite/pipeline_generator/pipeline_generator.py --run_all=$RUN_ALL --list_file_diff="$LIST_FILE_DIFF" --nightly="$NIGHTLY" --mirror_hw="$AMD_MIRROR_HW"
+        buildkite-agent pipeline upload .buildkite/pipeline.yaml
+        exit 0
+    fi
+    echo "List file diff: $LIST_FILE_DIFF"
+    echo "Run all: $RUN_ALL"
+    echo "Nightly: $NIGHTLY"
+    echo "AMD Mirror HW: $AMD_MIRROR_HW"
+
+    FAIL_FAST=$(fail_fast)
+
+    cd .buildkite
+    (
+        set -x
+        # Output pipeline.yaml with all blank lines removed
+        minijinja-cli test-template.j2 test-amd.yaml \
+            -D branch="$BUILDKITE_BRANCH" \
+            -D list_file_diff="$LIST_FILE_DIFF" \
+            -D run_all="$RUN_ALL" \
+            -D nightly="$NIGHTLY" \
+            -D mirror_hw="$AMD_MIRROR_HW" \
+            -D fail_fast="$FAIL_FAST" \
+            -D vllm_use_precompiled="$VLLM_USE_PRECOMPILED" \
+            -D vllm_merge_base_commit="$(git merge-base origin/main HEAD)" \
+            -D cov_enabled="$COV_ENABLED" \
+            -D vllm_ci_branch="$VLLM_CI_BRANCH" \
+            | sed '/^[[:space:]]*$/d' \
+            > pipeline.yaml
+    )
+    cat pipeline.yaml
+    buildkite-agent artifact upload pipeline.yaml
+    buildkite-agent pipeline upload pipeline.yaml
+    exit 0
+}
+
+get_diff() {
+    $(git add .)
+    echo $(git diff --name-only --diff-filter=ACMDR $(git merge-base origin/main HEAD))
+}
+
+get_diff_main() {
+    $(git add .)
+    echo $(git diff --name-only --diff-filter=ACMDR HEAD~1)
+}
+
+file_diff=$(get_diff)
+if [[ $BUILDKITE_BRANCH == "main" ]]; then
+    file_diff=$(get_diff_main)
+fi
+
+# ----------------------------------------------------------------------
+# Early exit start: skip pipeline if conditions are met
+# ----------------------------------------------------------------------
+
+# skip pipeline if all changed files are under docs/
+if [[ "${DOCS_ONLY_DISABLE}" != "1" ]]; then
+  if [[ -n "${file_diff:-}" ]]; then
+    docs_only=1
+    # Robust iteration over newline-separated file_diff
+    while IFS= read -r f; do
+      [[ -z "$f" ]] && continue
+      # **Policy:** only skip if *every* path starts with docs/
+      if [[ "$f" != docs/* ]]; then
+        docs_only=0
+        break
+      fi
+    done < <(printf '%s\n' "$file_diff" | tr ' ' '\n' | tr -d '\r')
+
+    if [[ "$docs_only" -eq 1 ]]; then
+      buildkite-agent annotate ":memo: CI skipped — docs/** only changes detected
+
+\`\`\`
+${file_diff}
+\`\`\`" --style "info" || true
+      echo "[docs-only] All changes are under docs/. Exiting before pipeline upload."
+      exit 0
+    fi
+  fi
+fi
+
+# ----------------------------------------------------------------------
+# Early exit end
+# ----------------------------------------------------------------------
+
+patterns=(
+    "docker/Dockerfile"
+    "CMakeLists.txt"
+    "requirements/common.txt"
+    "requirements/cuda.txt"
+    "requirements/build.txt"
+    "requirements/test.txt"
+    "setup.py"
+    "csrc/"
+    "cmake/"
+)
+
+ignore_patterns=(
+    "docker/Dockerfile."
+    "csrc/cpu"
+    "csrc/rocm"
+    "cmake/hipify.py"
+    "cmake/cpu_extension.cmake"
+)
+
+for file in $file_diff; do
+    # First check if file matches any pattern
+    matches_pattern=0
+    for pattern in "${patterns[@]}"; do
+        if [[ $file == $pattern* ]] || [[ $file == $pattern ]]; then
+            matches_pattern=1
+            break
+        fi
+    done
+
+    # If file matches pattern, check it's not in ignore patterns
+    if [[ $matches_pattern -eq 1 ]]; then
+        matches_ignore=0
+        for ignore in "${ignore_patterns[@]}"; do
+            if [[ $file == $ignore* ]] || [[ $file == $ignore ]]; then
+                matches_ignore=1
+                break
+            fi
+        done
+
+        if [[ $matches_ignore -eq 0 ]]; then
+            RUN_ALL=1
+            echo "Found changes: $file. Run all tests"
+            break
+        fi
+    fi
+done
+
+# Check for ready-run-all-tests label
+LABEL_RUN_ALL=$(check_run_all_label)
+if [[ $LABEL_RUN_ALL == true ]]; then
+    RUN_ALL=1
+    NIGHTLY=1
+    echo "Found 'ready-run-all-tests' label. Running all tests including optional tests."
+fi
+
+# Decide whether to use precompiled wheels
+# Relies on existing patterns array as a basis.
+if [[ -n "${VLLM_USE_PRECOMPILED:-}" ]]; then
+    echo "VLLM_USE_PRECOMPILED is already set to: $VLLM_USE_PRECOMPILED"
+elif [[ $RUN_ALL -eq 1 ]]; then
+    export VLLM_USE_PRECOMPILED=0
+    echo "Detected critical changes, building wheels from source"
+else
+    export VLLM_USE_PRECOMPILED=1
+    echo "No critical changes, using precompiled wheels"
+fi
+
+
+LIST_FILE_DIFF=$(get_diff | tr ' ' '|')
+if [[ $BUILDKITE_BRANCH == "main" ]]; then
+    LIST_FILE_DIFF=$(get_diff_main | tr ' ' '|')
+fi
+upload_pipeline
@@ -1,6 +1,307 @@
 steps:
-  - label: "Simple Unit Test"
+  - label: ":docker: Build image"
+    key: image-build
    commands:
-      - ".buildkite/scripts/simple_test.sh"
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "docker build --file docker/Dockerfile.ci -t vllm-omni-ci ."
+      - "docker tag vllm-omni-ci public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT"
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT"
    agents:
      queue: "cpu_queue_premerge"
+
+  # - label: "Test on NPU"
+  #   depends_on: ~
+  #   key: npu-test
+  #   commands:
+  #     - ".buildkite/scripts/hardware_ci/run_npu_test.sh"
+  #   agents:
+  #     queue: "ascend"
+
+  - label: "Simple Unit Test"
+    depends_on: image-build
+    commands:
+      - pytest -v -s tests/entrypoints/
+      - pytest -v -s tests/diffusion/cache/
+      - pytest -v -s tests/diffusion/lora/
+      - pytest -v -s tests/model_executor/models/qwen2_5_omni/test_audio_length.py
+      - pytest -v -s tests/worker/
+      - pytest -v -s tests/distributed/omni_connectors/test_kv_flow.py
+    agents:
+      queue: "gpu_1_queue"
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Model Test"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_t2i_model.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Images API LoRA E2E"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/e2e/online_serving/test_images_generations_lora.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Model CPU offloading Test"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
+      - pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Audio Generation Model Test"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_stable_audio_model.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Cache Backend Test"
+    timeout_in_minutes: 15
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_cache_dit.py tests/e2e/offline_inference/test_teacache.py
+    agents:
+      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Sequence Parallelism Test"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion Tensor Parallelism Test"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/e2e/offline_inference/test_zimage_tensor_parallel.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Diffusion GPU Worker Test"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - pytest -s -v tests/diffusion/test_diffusion_worker.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+      - docker#v5.2.0:
+          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+          always-pull: true
+          propagate-environment: true
+          shm-size: "8gb"
+          environment:
+            - "HF_HOME=/fsx/hf_cache"
+          volumes:
+            - "/fsx/hf_cache:/fsx/hf_cache"
+
+
+  - label: "Benchmark Test"
+    timeout_in_minutes: 15
+    depends_on: image-build
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/benchmarks/test_serve_cli.py
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
+  - label: "Omni Model Test"
+    timeout_in_minutes: 15
+    depends_on: image-build
+    commands:
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+    agents:
+      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+    plugins:
+    - docker#v5.2.0:
+        image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+        always-pull: true
+        propagate-environment: true
+        environment:
+        - "HF_HOME=/fsx/hf_cache"
+        volumes:
+        - "/fsx/hf_cache:/fsx/hf_cache"
+
+  - label: "Omni Model Test with H100"
+    timeout_in_minutes: 30
+    depends_on: image-build
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
+      - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
+      - pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py
+      - pytest -s -v tests/e2e/online_serving/test_async_omni.py
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
+  - label: "Diffusion Image Edit Test with H100 (1 GPU)"
+    timeout_in_minutes: 20
+    depends_on: image-build
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 1
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
@@ -0,0 +1,73 @@
+#!/bin/bash
+# Helper function to safely login to ECR Public with per-job config isolation
+# Uses DOCKER_CONFIG environment variable to prevent race conditions
+#
+# This script prevents the "device or resource busy" error by giving each
+# Buildkite job its own isolated Docker config directory.
+#
+# Usage:
+#   source docker_login_ecr_public.sh && safe_docker_login_ecr_public
+
+set -euo pipefail
+
+# Configuration
+ECR_REGISTRY="public.ecr.aws"
+
+setup_isolated_docker_config() {
+    # Use BUILDKITE_JOB_ID for job-specific isolation
+    # Fallback to PID if running outside Buildkite
+    local job_id="${BUILDKITE_JOB_ID:-$$}"
+
+    # Set Docker config to job-specific directory
+    export DOCKER_CONFIG="/tmp/docker-config-${job_id}"
+
+    # Create directory if it doesn't exist
+    mkdir -p "$DOCKER_CONFIG"
+
+    echo "[docker-config] Using isolated Docker config: $DOCKER_CONFIG"
+}
+
+check_docker_auth() {
+    # Check if already authenticated to the given registry
+    # Returns 0 if authenticated, 1 if not
+    local registry="$1"
+
+    # Check if credentials exist in the isolated config
+    if [[ -f "$DOCKER_CONFIG/config.json" ]]; then
+        # Check if registry is present in config
+        if grep -q "$registry" "$DOCKER_CONFIG/config.json" 2>/dev/null; then
+            return 0
+        fi
+    fi
+
+    return 1
+}
+
+safe_docker_login_ecr_public() {
+    # Setup isolated config first
+    setup_isolated_docker_config
+
+    local registry="$ECR_REGISTRY"
+
+    # Check if already authenticated (within this job)
+    if check_docker_auth "$registry"; then
+        echo "[docker-login] Already authenticated to $registry in this job"
+        return 0
+    fi
+
+    # Perform login to isolated config directory
+    echo "[docker-login] Logging in to $ECR_REGISTRY (isolated config)..."
+    if aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin "$ECR_REGISTRY"; then
+        echo "[docker-login] Login successful (config: $DOCKER_CONFIG)"
+        return 0
+    else
+        local exit_code=$?
+        echo "[docker-login] ERROR: Login failed with exit code $exit_code" >&2
+        return $exit_code
+    fi
+}
+
+# Execute if run as script (not sourced)
+if [[ "${BASH_SOURCE[0]}" == "${0}" ]]; then
+    safe_docker_login_ecr_public
+fi
@@ -0,0 +1,166 @@
+#!/bin/bash
+# vllm-omni customized version
+# Based on: vllm/.buildkite/scripts/hardware_ci/run-amd-test.sh
+# Last synced: 2025-12-15
+# Modifications: docker image name for vllm-omni
+
+# This script runs test inside the corresponding ROCm docker container.
+set -o pipefail
+
+# Export Python path
+export PYTHONPATH=".."
+
+# Print ROCm version
+echo "--- Confirming Clean Initial State"
+while true; do
+        sleep 3
+        if grep -q clean /opt/amdgpu/etc/gpu_state; then
+                echo "GPUs state is \"clean\""
+                break
+        fi
+done
+
+echo "--- ROCm info"
+rocminfo
+
+# cleanup older docker images
+cleanup_docker() {
+  # Get Docker's root directory
+  docker_root=$(docker info -f '{{.DockerRootDir}}')
+  if [ -z "$docker_root" ]; then
+    echo "Failed to determine Docker root directory."
+    exit 1
+  fi
+  echo "Docker root directory: $docker_root"
+  # Check disk usage of the filesystem where Docker's root directory is located
+  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+  # Define the threshold
+  threshold=70
+  if [ "$disk_usage" -gt "$threshold" ]; then
+    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    # Remove dangling images (those that are not tagged and not used by any container)
+    docker image prune -f
+    # Remove unused volumes / force the system prune for old images as well.
+    docker volume prune -f && docker system prune --force --filter "until=72h" --all
+    echo "Docker images and volumes cleanup completed."
+  else
+    echo "Disk usage is below $threshold%. No cleanup needed."
+  fi
+}
+
+# Call the cleanup docker function
+cleanup_docker
+
+echo "--- Resetting GPUs"
+
+echo "reset" > /opt/amdgpu/etc/gpu_state
+
+while true; do
+        sleep 3
+        if grep -q clean /opt/amdgpu/etc/gpu_state; then
+                echo "GPUs state is \"clean\""
+                break
+        fi
+done
+
+echo "--- Pulling container"
+image_name="public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:${BUILDKITE_COMMIT}-rocm-omni"
+container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
+# Install AWS CLI to authenticate to ECR Public Gallery to get higher rate limit for pulling images
+sudo apt-get update && sudo apt-get install -y awscli
+# Use safe docker login helper to prevent race conditions
+source "$(dirname "${BASH_SOURCE[0]}")/../docker_login_ecr_public.sh"
+safe_docker_login_ecr_public
+# Pull the container from ECR Public Gallery
+
+docker pull "${image_name}"
+
+remove_docker_container() {
+   docker rm -f "${container_name}" || docker image rm -f "${image_name}" || true
+}
+trap remove_docker_container EXIT
+
+echo "--- Running container"
+
+HF_CACHE="$(realpath ~)/huggingface"
+mkdir -p "${HF_CACHE}"
+HF_MOUNT="/root/.cache/huggingface"
+
+commands=$@
+echo "Commands:$commands"
+
+PARALLEL_JOB_COUNT=8
+MYPYTHONPATH=".."
+
+# Test that we're launching on the machine that has
+# proper access to GPUs
+render_gid=$(getent group render | cut -d: -f3)
+if [[ -z "$render_gid" ]]; then
+  echo "Error: 'render' group not found. This is required for GPU access." >&2
+  exit 1
+fi
+
+# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
+if [[ $commands == *"--shard-id="* ]]; then
+  # assign job count as the number of shards used
+  commands=$(echo "$commands" | sed -E "s/--num-shards[[:blank:]]*=[[:blank:]]*[0-9]*/--num-shards=${PARALLEL_JOB_COUNT} /g" | sed 's/ \\ / /g')
+  for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
+    # assign shard-id for each shard
+    commands_gpu=$(echo "$commands" | sed -E "s/--shard-id[[:blank:]]*=[[:blank:]]*[0-9]*/--shard-id=${GPU} /g" | sed 's/ \\ / /g')
+    echo "Shard ${GPU} commands:$commands_gpu"
+    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
+    docker run \
+        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+        --network=host \
+        --shm-size=16gb \
+        --group-add "$render_gid" \
+        --rm \
+        -e MIOPEN_DEBUG_CONV_DIRECT=0 \
+        -e MIOPEN_DEBUG_CONV_GEMM=0 \
+        -e VLLM_ROCM_USE_AITER=1 \
+        -e HIP_VISIBLE_DEVICES="${GPU}" \
+        -e HF_TOKEN \
+        -e AWS_ACCESS_KEY_ID \
+        -e AWS_SECRET_ACCESS_KEY \
+        -v "${HF_CACHE}:${HF_MOUNT}" \
+        -e "HF_HOME=${HF_MOUNT}" \
+        -e "PYTHONPATH=${MYPYTHONPATH}" \
+        --name "${container_name}_${GPU}" \
+        "${image_name}" \
+        /bin/bash -c "${commands_gpu}" \
+        |& while read -r line; do echo ">>Shard $GPU: $line"; done &
+    PIDS+=($!)
+  done
+  #wait for all processes to finish and collect exit codes
+  for pid in "${PIDS[@]}"; do
+    wait "${pid}"
+    STATUS+=($?)
+  done
+  for st in "${STATUS[@]}"; do
+    if [[ ${st} -ne 0 ]]; then
+      echo "One of the processes failed with $st"
+      exit "${st}"
+    fi
+  done
+else
+  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
+  docker run \
+          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+          --network=host \
+          --shm-size=16gb \
+          --group-add "$render_gid" \
+          --rm \
+          -e MIOPEN_DEBUG_CONV_DIRECT=0 \
+          -e MIOPEN_DEBUG_CONV_GEMM=0 \
+          -e VLLM_ROCM_USE_AITER=1 \
+          -e HF_TOKEN \
+          -e AWS_ACCESS_KEY_ID \
+          -e AWS_SECRET_ACCESS_KEY \
+          -v "${HF_CACHE}:${HF_MOUNT}" \
+          -e "HF_HOME=${HF_MOUNT}" \
+          -e "PYTHONPATH=${MYPYTHONPATH}" \
+          --name "${container_name}" \
+          "${image_name}" \
+          /bin/bash -c "${commands}"
+fi
@@ -0,0 +1,145 @@
+#!/bin/bash
+
+# This script build the Ascend NPU docker image and run the offline inference inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+# Base ubuntu image with basic ascend development libraries and python installed
+VLLM_OMNI_REPO="https://github.com/vllm-project/vllm-omni.git"
+BASE_IMAGE_NAME="quay.nju.edu.cn/ascend/vllm-ascend:v0.11.0rc2"
+image_name="npu/vllm-omni-ci:${BUILDKITE_COMMIT}_${EPOCHSECONDS}"
+# image_name="npu/vllm-ci:${BUILDKITE_COMMIT}_${EPOCHSECONDS}"
+container_name="npu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
+# BUILDKITE_AGENT_NAME format is {hostname}-{agent_idx}-{npu_card_num}cards
+agent_idx=$(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 ${builder_cache_dir}
+
+# Try building the docker image
+cat <<EOF | DOCKER_BUILDKIT=1 docker build \
+    --add-host pypi-cache:${PYPI_CACHE_HOST} \
+    --builder ${builder_name} --cache-from type=local,src=${builder_cache_dir} \
+                           --cache-to type=local,dest=${builder_cache_dir},mode=max \
+    --build-arg BUILDKITE_PULL_REQUEST="${BUILDKITE_PULL_REQUEST}" \
+    --build-arg BUILDKITE_PULL_REQUEST_REPO="${BUILDKITE_PULL_REQUEST_REPO}" \
+    --progress=plain --load -t ${image_name} -f - .
+FROM ${BASE_IMAGE_NAME}
+
+# Define environments
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN pip config set global.index-url http://pypi-cache:${PYPI_CACHE_PORT}/pypi/simple && \
+    pip config set global.trusted-host pypi-cache && \
+    apt-get update -y && \
+    apt-get install -y python3-pip git vim wget net-tools gcc g++ cmake libnuma-dev && \
+    rm -rf /var/cache/apt/* && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install for pytest to make the docker build cache layer always valid
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install pytest>=6.0  pytest-cov modelscope
+
+COPY . .
+
+# Install vllm-omni
+WORKDIR /workspace
+ARG VLLM_OMNI_REPO=https://github.com/vllm-project/vllm-omni.git
+ARG VLLM_OMNI_TAG=main
+ARG BUILDKITE_PULL_REQUEST
+ARG BUILDKITE_PULL_REQUEST_REPO
+RUN git config --global url."https://gh-proxy.test.osinfra.cn/https://github.com/".insteadOf "https://github.com/" && \
+    if [ "\$BUILDKITE_PULL_REQUEST" != "false" ] && [ -n "\$BUILDKITE_PULL_REQUEST" ]; then \
+        echo "Cloning and checking out PR #\$BUILDKITE_PULL_REQUEST..." && \
+        git clone \$VLLM_OMNI_REPO /workspace/vllm-omni && \
+        cd /workspace/vllm-omni && \
+        git fetch origin pull/\$BUILDKITE_PULL_REQUEST/head:pr-\$BUILDKITE_PULL_REQUEST && \
+        git checkout pr-\$BUILDKITE_PULL_REQUEST; \
+    else \
+        echo "Not a PR build, using main branch" && \
+        git clone --depth 1 \$VLLM_OMNI_REPO /workspace/vllm-omni; \
+    fi
+
+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 && \
+    python3 -m pip install -v -e /workspace/vllm-omni/
+
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+WORKDIR /workspace/vllm-omni
+CMD ["/bin/bash"]
+
+EOF
+
+# Setup cleanup
+remove_docker_container() {
+  docker rm -f "${container_name}" || true;
+  docker image rm -f "${image_name}" || true;
+  docker system prune -f || true;
+}
+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
+parse_and_gen_devices() {
+    local input="$1"
+    local index cards_num
+    if [[ "$input" =~ ([0-9]+)-([0-9]+)cards$ ]]; then
+        index="${BASH_REMATCH[1]}"
+        cards_num="${BASH_REMATCH[2]}"
+    else
+        echo "parse error" >&2
+        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${dev_idx}"
+        ((i++))
+    done
+
+    # trim leading space
+    devices="${devices#"${devices%%[![:space:]]*}"}"
+    # Output devices: assigned to the caller variable
+    printf '%s' "$devices"
+}
+
+devices=$(parse_and_gen_devices "${BUILDKITE_AGENT_NAME}") || exit 1
+
+# 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.
+hf_model_cache_dir=/mnt/hf_cache${agent_idx}
+ms_model_cache_dir=/mnt/modelscope${agent_idx}
+mkdir -p ${hf_model_cache_dir}
+mkdir -p ${ms_model_cache_dir}
+docker run \
+    --init \
+    ${devices} \
+    --device /dev/davinci_manager \
+    --device /dev/devmm_svm \
+    --device /dev/hisi_hdc \
+    -v /usr/local/dcmi:/usr/local/dcmi \
+    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+    -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 ${hf_model_cache_dir}:/root/.cache/huggingface \
+    -v ${ms_model_cache_dir}:/root/.cache/modelscope \
+    --network host \
+    --entrypoint="" \
+    --name "${container_name}" \
+    "${image_name}" \
+    bash -c '
+    set -e
+    VLLM_USE_MODELSCOPE=True pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+'
@@ -1,52 +0,0 @@
-#!/usr/bin/env bash
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd -- "${SCRIPT_DIR}/../.." && pwd)"
-UV_BIN="uv"
-PYTHON_BOOTSTRAP="${PYTHON:-python3}"
-VENV_DIR="${VENV_DIR:-${REPO_ROOT}/.venv-simple-test}"
-TARGET_PY_VERSION="3.12"
-
-if ! command -v "${PYTHON_BOOTSTRAP}" >/dev/null 2>&1; then
-  PYTHON_BOOTSTRAP="python"
-fi
-
-cd "${REPO_ROOT}"
-
-# Ensure pip-installed scripts (including uv) are on PATH
-PYTHON_SCRIPTS="$(${PYTHON_BOOTSTRAP} -c 'import sysconfig; print(sysconfig.get_path("scripts"))')"
-USER_BASE="$(${PYTHON_BOOTSTRAP} -m site --user-base 2>/dev/null || true)"
-PATH="${PYTHON_SCRIPTS}:${PATH}"
-if [[ -n "${USER_BASE}" ]]; then
-  PATH="${USER_BASE}/bin:${PATH}"
-fi
-
-if ! command -v "${UV_BIN}" >/dev/null 2>&1; then
-  "${PYTHON_BOOTSTRAP}" -m pip install --upgrade pip
-  "${PYTHON_BOOTSTRAP}" -m pip install uv
-  hash -r
-fi
-
-TARGET_PY_PATH="$(${UV_BIN} python find "${TARGET_PY_VERSION}" 2>/dev/null | head -n1 || true)"
-if [[ -z "${TARGET_PY_PATH}" ]]; then
-  "${UV_BIN}" python install "${TARGET_PY_VERSION}"
-  TARGET_PY_PATH="$(${UV_BIN} python find "${TARGET_PY_VERSION}" 2>/dev/null | head -n1 || true)"
-fi
-
-if [[ -z "${TARGET_PY_PATH}" ]]; then
-  echo "Failed to locate python ${TARGET_PY_VERSION}" >&2
-  exit 1
-fi
-
-if [[ ! -d "${VENV_DIR}" ]]; then
-  "${UV_BIN}" venv --python "${TARGET_PY_PATH}" "${VENV_DIR}"
-fi
-
-VENV_PYTHON="${VENV_DIR}/bin/python"
-[[ -x "${VENV_PYTHON}" ]] || { echo "Python not found in ${VENV_DIR}"; exit 1; }
-
-"${UV_BIN}" pip install --python "${VENV_PYTHON}" vllm==0.11.0
-"${UV_BIN}" pip install --python "${VENV_PYTHON}" -e ".[dev]"
-"${VENV_PYTHON}" -m pytest -v -s tests/test_omni_llm.py
@@ -0,0 +1,116 @@
+steps:
+
+- label: "Diffusion Model Test"
+  timeout_in_minutes: 20
+  agent_pool: mi325_2
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export GPU_ARCHS=gfx942
+    - pytest -s -v tests/e2e/offline_inference/test_t2i_model.py
+
+- label: "Diffusion Images API LoRA E2E"
+  timeout_in_minutes: 20
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export GPU_ARCHS=gfx942
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/online_serving/test_images_generations_lora.py
+
+- label: "Diffusion Model CPU offloading Test"
+  timeout_in_minutes: 20
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export GPU_ARCHS=gfx942
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
+
+- label: "Diffusion Cache Backend Test"
+  timeout_in_minutes: 15
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export GPU_ARCHS=gfx942
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/offline_inference/test_cache_dit.py tests/e2e/offline_inference/test_teacache.py
+
+- label: "Diffusion Sequence Parallelism Test"
+  timeout_in_minutes: 20
+  agent_pool: mi325_2
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export GPU_ARCHS=gfx942
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
+
+- label: "Diffusion Tensor Parallelism Test"
+  timeout_in_minutes: 20
+  agent_pool: mi325_2
+  depends_on: amd-build
+  commands:
+    - export GPU_ARCHS=gfx942
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/offline_inference/test_zimage_tensor_parallel.py
+
+- label: "Diffusion GPU Worker Test"
+  timeout_in_minutes: 20
+  agent_pool: mi325_2
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - pytest -s -v tests/diffusion/test_diffusion_worker.py
+
+- label: "Omni Model Test Qwen2-5-Omni"
+  timeout_in_minutes: 15
+  agent_pool: mi325_2
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export GPU_ARCHS=gfx942
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+
+- label: "Omni Model Test Qwen3-Omni"
+  timeout_in_minutes: 15
+  agent_pool: mi325_2
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
+    - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
+    - pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py
+    - pytest -s -v tests/e2e/online_serving/test_async_omni.py
+
+- label: "Diffusion Image Edit Test"
+  timeout_in_minutes: 15
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - export GPU_ARCHS=gfx942
+    - export VLLM_LOGGING_LEVEL=DEBUG
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
@@ -0,0 +1,53 @@
+{# vllm-omni customized version
+   Based on: https://github.com/vllm-project/ci-infra/blob/main/buildkite/test-template-amd.j2
+   Last synced: 2025-12-15
+   Modifications: Removed unused CUDA/NVIDIA logic, keeping only AMD tests
+#}
+{% set docker_image_amd = "public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT-rocm-omni" %}
+{% set default_working_dir = "/app/vllm-omni" %}
+
+  - group: "AMD Tests"
+    depends_on: ~
+    steps:
+      - label: "AMD: :docker: build image"
+        depends_on: ~
+        soft_fail: false
+        commands:
+          - "source .buildkite/scripts/docker_login_ecr_public.sh && safe_docker_login_ecr_public"
+          - "docker build -f docker/Dockerfile.rocm -t {{ docker_image_amd }} --target final --progress plain ."
+          - "docker push {{ docker_image_amd }}"
+        key: "amd-build"
+        env:
+          DOCKER_BUILDKIT: "1"
+        retry:
+          automatic:
+            - exit_status: -1  # Agent was lost
+              limit: 1
+            - exit_status: -10  # Agent was lost
+              limit: 1
+            - exit_status: 1  # Machine occasionally fail
+              limit: 1
+        agents:
+          queue: cpu_queue_premerge
+
+    {% for step in steps %}
+    {% if step.mirror_hardwares and mirror_hw in step.mirror_hardwares %}
+      - label: "{{ step.agent_pool }}: {{ step.label }}"
+        depends_on: amd-build
+        agents:
+          {% if step.agent_pool %}
+          queue: amd_{{ step.agent_pool }}
+          {% else %}
+          queue: amd_mi325_1
+          {% endif %}
+        command: bash .buildkite/scripts/hardware_ci/run-amd-test.sh "(command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" && ")) | safe }}"
+        env:
+          DOCKER_BUILDKIT: "1"
+        priority: 100
+        {% if step.grade and step.grade == "Blocking" %}
+        soft_fail: false
+        {% else %}
+        soft_fail: true
+        {% endif%}
+    {% endif %}
+    {% endfor %}
@@ -24,7 +24,7 @@ body:
    description: |
      Please run the following and paste the output below.
      ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm-omni/main/vllm_omni/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm-omni/main/collect_env.py
      # For security purposes, please feel free to check the contents of collect_env.py before running it.
      python collect_env.py
      ```
@@ -40,6 +40,31 @@ body:
      </details>
  validations:
    required: true
+
+- type: textarea
+  attributes:
+    label: Your code version
+    description: |
+      Please provide a description of which commit id or version the code is.
+
+    value: |
+      <details>
+      <summary>The commit id or version of vllm</summary>
+
+      ```text
+
+      ```
+      </details>
+      <details>
+      <summary>The commit id or version of vllm-omni</summary>
+
+      ```text
+
+      ```
+      </details>
+  validations:
+    required: true
+
 - type: textarea
  attributes:
    label: 🐛 Describe the bug
@@ -20,6 +20,8 @@ body:
    label: Proposed Change.
    description: >
      The proposed change of the RFC.
+    value: |
+      Please provide the detailed design document of the RFC using the [template](https://docs.google.com/document/d/12YxSsVeD1jvL-InClkeAEnZyWFDndz_65JmXvsamuV4/edit?tab=t.0#heading=h.4ef4szrsgspp).
  validations:
    required: true
 - type: textarea
@@ -1,5 +1,5 @@
 blank_issues_enabled: false
 contact_links:
  - name: Questions
-    url: https://github.com/vllm-project/vllm-omni/discussions
-    about: Ask questions and discuss with other vLLM-Omni community members
+    url: https://slack.vllm.ai
+    about: Ask questions and discuss with other vLLM-Omni community members in our #sig-omni slack channel.
@@ -0,0 +1,61 @@
+name: Build Wheel
+
+on:
+  push:
+    branches:
+      - main
+    paths-ignore:
+      - 'docs/**'
+      - 'examples/**'
+      - '.github/**'
+      - '.buildkite/**'
+      - '**.md'
+      - '*.yml'
+      - '.gitignore'
+      - '.gitattributes'
+      - 'LICENSE'
+  pull_request:
+    branches:
+      - main
+    paths-ignore:
+      - 'docs/**'
+      - 'examples/**'
+      - '.github/**'
+      - '.buildkite/**'
+      - '*.yml'
+      - '**.md'
+      - '.gitignore'
+      - '.gitattributes'
+      - 'LICENSE'
+  workflow_dispatch:
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.11", "3.12"]
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install uv
+        run: |
+          pip install uv
+
+      - name: Build Wheel
+        env:
+          UV_SYSTEM_PYTHON: 1
+        run: |
+          bash scripts/build_wheel.sh --python python
+
+      - name: Upload Wheel Artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: vllm-omni-wheel-py${{ matrix.python-version }}
+          path: dist/*.whl
@@ -156,9 +156,6 @@ cython_debug/
 # cursor
 .cursor/

-# docker
-docker/
-
 # scripts
 /scripts/

@@ -182,6 +179,7 @@ checkpoints/

 # Cache directories
 cache/
+!vllm_omni/diffusion/cache/
 .cache/
 diffusion_cache/
 kv_cache/
@@ -223,6 +221,7 @@ datasets/
 outputs/
 results/
 generated/
+output_*/

 # Configuration overrides
 configs/local.yaml
@@ -232,3 +231,5 @@ configs/development.yaml
 # Docker
 .dockerignore
 Dockerfile.dev
+discussion
+tmp_test
@@ -3,7 +3,7 @@ default_install_hook_types:
  - commit-msg
 repos:
  - repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.6.0
+    rev: v6.0.0
    hooks:
      # list of supported hooks: https://pre-commit.com/hooks.html
      - id: check-yaml
@@ -16,19 +16,22 @@ repos:
        args: ["--markdown-linebreak-ext=md"]

  - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.14.0
+    rev: v0.14.10
    hooks:
    - id: ruff-check
      args: [--output-format, github, --fix]
    - id: ruff-format

  - repo: https://github.com/crate-ci/typos
-    rev: v1.35.5
+    rev: typos-dict-v0.13.13
    hooks:
      - id: typos
        # only for staged files

  - repo: https://github.com/rhysd/actionlint
+    # v1.7.8+ sets `go 1.24.0` in go.mod, which older Go toolchains (and most
+    # current CI images) cannot parse. Pin to v1.7.7 until actionlint fixes the
+    # go.mod directive.
    rev: v1.7.7
    hooks:
      - id: actionlint
@@ -58,3 +61,10 @@ repos:
      verbose: true
      pass_filenames: false
    # Insert new entries above the `suggestion` entry
+
+    - id: check-pickle-imports
+      name: Prevent new pickle/cloudpickle imports
+      entry: python tools/pre_commit/check_pickle_imports.py
+      language: python
+      types: [python]
+      additional_dependencies: [regex]
@@ -9,13 +9,13 @@ Easy, fast, and cheap omni-modality model serving for everyone
 </h3>

 <p align="center">
-| <a href="https://vllm-omni.readthedocs.io/en/latest/"><b>Documentation</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://vllm-omni.readthedocs.io/en/latest/"><b>Documentation</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> | <a href="docs/assets/WeChat.jpg"><b>WeChat</b></a> |
 </p>

 ---

 *Latest News* 🔥
-
+- [2026/01] We released [0.12.0rc1](https://github.com/vllm-project/vllm-omni/releases/tag/v0.12.0rc1) - a major RC milestone focused on maturing the diffusion stack, strengthening OpenAI-compatible serving, expanding omni-model coverage, and improving stability across platforms (GPU/NPU/ROCm), please check our latest [design](https://docs.google.com/presentation/d/1qv4qMW1rKAqDREMXiUDLIgqqHQe7TDPj/edit?usp=sharing&ouid=110473603432222024453&rtpof=true&sd=true).
 - [2025/11] vLLM community officially released [vllm-project/vllm-omni](https://github.com/vllm-project/vllm-omni) in order to support omni-modality models serving.

 ---
@@ -69,6 +69,10 @@ Please check out [Contributing to vLLM-Omni](https://vllm-omni.readthedocs.io/en
 ## Join the Community
 Feel free to ask questions, provide feedbacks and discuss with fellow users of vLLM-Omni in `#sig-omni` slack channel at [slack.vllm.ai](https://slack.vllm.ai) or vLLM user forum at [discuss.vllm.ai](https://discuss.vllm.ai).

+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=vllm-project/vllm-omni&type=date&legend=top-left)](https://www.star-history.com/#vllm-project/vllm-omni&type=date&legend=top-left)
+
 ## License

 Apache License 2.0, as found in the [LICENSE](./LICENSE) file.
@@ -0,0 +1,42 @@
+# Benchmarks Overview and Architecture
+
+This document explains the benchmark architecture across all benchmark assets in this repo. It describes what we measure, and where to find or plug in new scenarios. Per-task details remain in subfolder READMEs (e.g., `benchmarks/<model>/README.md`).
+
+## Scope and goals
+- Establish repeatable latency/throughput measurements for multimodal LLM pipelines.
+- Provide both HF Transformers (offline) and vLLM-Omni (multi-stage/pipeline) baselines.
+- Make it easy to plug in new datasets and models with minimal changes to the runner scripts.
+
+## Dataset and inputs
+- Default example: SeedTTS top-100 prompts (`benchmarks/build_dataset/top100.txt`) via `benchmarks/build_dataset/`.
+- Extensible: drop in new prompt files or modality-aligned payloads; keep the expected format for the consuming scripts (e.g., one prompt per line).
+- If you add a new dataset, document it under `benchmarks/<model>/README.md` and point scripts to your data path.
+
+## Directory layout
+- `benchmarks/build_dataset/` — dataset prep utilities (e.g., SeedTTS top100).
+- `benchmarks/<model>/vllm_omni/` — vLLM-Omni pipeline benchmarks, logs, outputs.
+- Add new tasks under `benchmarks/<model>/...` with the same pattern: `transformers/`, `vllm_omni/`, task-specific README, and (optionally) dataset prep notes.
+
+## Reference workflows
+- **HF Transformers (offline, single process)**  
+  Script (example): `benchmarks/<model>/transformers/eval_qwen3_moe_omni_transformers.sh`  
+  Outputs: `benchmark_results/perf_stats.json`, `benchmark_results/results.json`, `benchmark_results/audio/` (if audio is produced).
+
+- **vLLM-Omni end-to-end pipeline**  
+  Script (example): `benchmarks/<model>/vllm_omni/eval_qwen3_moe_omni.sh`  
+  Outputs: `vllm_omni/logs/*.stats.jsonl` (per-stage/overall latency & TPS), `vllm_omni/logs/stage*.log`, `vllm_omni/outputs/` (text/audio artifacts).
+
+- **Adding a new task/model**  
+  1) Create `benchmarks/<model>/transformers/` and/or `benchmarks/<model>/vllm_omni/` with scripts referencing your model and dataset.  
+  2) Add a task README describing dataset, configs, and expected outputs.  
+  3) Keep the output/log structure similar for easy comparison (perf_stats/results/audio or text outputs; stats.jsonl/logs for pipeline).
+
+## Metrics to watch
+- **Throughput**: `overall_tps`, `*_tps_avg` per stage.
+- **Latency distribution**: look for long tails in `*.stats.jsonl`.
+- **Quality/completeness**: missing outputs or errors in stage logs indicate pipeline failures or misconfigurations.
+
+## Troubleshooting
+- Verify GPU/driver/FlashAttention2 requirements for your chosen model/config.
+- Ensure network access for dataset/model downloads (Google Drive, Hugging Face, etc.).
+- If outputs are missing or slow, inspect per-stage logs and `*.stats.jsonl` for errors, stragglers, or contention.
@@ -0,0 +1,82 @@
+# Benchmark Dataset Preparation Guide
+
+This guide describes how to download and prepare the SeedTTS test dataset for benchmarking Qwen-Omni models.
+
+## Prerequisites
+
+- Python 3.8+
+- `gdown` for downloading from Google Drive
+- Access to the benchmark scripts
+
+## Steps
+
+### 1. Navigate to the Dataset Directory
+
+```bash
+cd benchmarks/build_dataset
+```
+
+### 2. Install Dependencies
+
+```bash
+pip install gdown
+```
+
+### 3. Download the SeedTTS Test Dataset
+
+Download the dataset from Google Drive:
+
+```bash
+gdown --id 1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
+```
+
+### 4. Extract the Dataset
+
+```bash
+tar -xf seedtts_testset.tar
+```
+
+### 5. Prepare the Metadata File
+
+Copy the English metadata file to the working directory:
+
+```bash
+cp seedtts_testset/en/meta.lst meta.lst
+```
+
+### 6. Extract Prompts
+
+Extract the first N prompts from the metadata file:
+
+```bash
+# Extract top 100 prompts (adjust -n for different amounts)
+python extract_tts_prompts.py -i meta.lst -o top100.txt -n 100
+```
+
+**Options:**
+- `-i, --input`: Input metadata file (default: `meta.lst`)
+- `-o, --output`: Output prompts file (default: `prompts.txt`)
+- `-n, --num_lines`: Number of prompts to extract (required)
+
+### 7. Clean Up (Optional)
+
+Remove temporary files to save disk space:
+
+```bash
+rm -rf seedtts_testset
+rm seedtts_testset.tar
+rm meta.lst
+```
+
+## Quick Start (All-in-One)
+
+```bash
+# Full setup and benchmark
+cd benchmarks/build_dataset
+pip install gdown
+gdown --id 1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
+tar -xf seedtts_testset.tar
+cp seedtts_testset/en/meta.lst meta.lst
+python extract_tts_prompts.py -i meta.lst -o top100.txt -n 100
+rm -rf seedtts_testset seedtts_testset.tar meta.lst
+```
@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+"""
+Extract prompts from meta.lst and save them to a txt file.
+
+Each line in meta.lst has the format:
+ID|prompt_text|audio_path|target_text
+
+This script extracts the prompt_text (second field) from the first N lines.
+"""
+
+import argparse
+from pathlib import Path
+
+
+def extract_prompts(input_file: str, output_file: str, num_lines: int) -> None:
+    """
+    Extract prompts from meta.lst and save to output file.
+
+    Args:
+        input_file: Path to the meta.lst file
+        output_file: Path to the output txt file
+        num_lines: Number of lines to process
+    """
+    prompts = []
+
+    with open(input_file, encoding="utf-8") as f:
+        for i, line in enumerate(f):
+            if i >= num_lines:
+                break
+
+            line = line.strip()
+            if not line:  # Skip empty lines
+                continue
+
+            parts = line.split("|")
+            if len(parts) >= 2:
+                prompt = parts[1]  # The prompt is the second field
+                prompts.append(prompt)
+
+    # Write prompts to output file
+    with open(output_file, "w", encoding="utf-8") as f:
+        for prompt in prompts:
+            f.write(prompt + "\n")
+
+    # Print result stats
+    print(f"Extracted {len(prompts)} prompts from first {num_lines} lines")
+    print(f"Saved to: {output_file}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Extract prompts from meta.lst file")
+    parser.add_argument(
+        "-i", "--input", type=str, default="meta.lst", help="Input meta.lst file path (default: meta.lst)"
+    )
+    parser.add_argument(
+        "-o", "--output", type=str, default="prompts.txt", help="Output txt file path (default: prompts.txt)"
+    )
+    parser.add_argument(
+        "-n", "--num_lines", type=int, required=True, help="Number of lines to extract from the beginning"
+    )
+
+    args = parser.parse_args()
+
+    # Check if input file exists
+    if not Path(args.input).exists():
+        print(f"Error: Input file '{args.input}' not found")
+        return
+
+    extract_prompts(args.input, args.output, args.num_lines)
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,117 @@
+
+# Diffusion Serving Benchmark (Image/Video)
+
+This folder contains an online-serving benchmark script for diffusion models.
+It sends requests to a vLLM OpenAI-compatible endpoint and reports throughput,
+latency percentiles, and optional SLO attainment.
+
+The main entrypoint is:
+
+- `benchmarks/diffusion/diffusion_benchmark_serving.py`
+
+## 1. Quick Start
+
+1. Start the server:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni --port 8099
+```
+
+2. Run a minimal benchmark:
+
+```bash
+python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
+	--base-url http://localhost:8099 \
+	--model Qwen/Qwen-Image \
+	--task t2i \
+	--dataset vbench \
+	--num-prompts 5
+```
+
+**Notes**
+
+- The benchmark talks to `http://<host>:<port>/v1/chat/completions`.
+- If you run the server on another host or port, pass `--base-url` accordingly.
+
+## 2. Supported Datasets
+
+The benchmark supports three dataset modes via `--dataset`:
+
+- `vbench`: Built-in prompt/data loader.
+- `trace`: Heterogeneous request traces (each request can have different resolution/frames/steps).
+- `random`: Synthetic prompts for quick smoke tests.
+
+### VBench dataset
+
+If you use i2v/i2i bench datasets and need auto-download support, you may need:
+
+```bash
+uv pip install gdown
+```
+
+### Trace dataset
+
+Use `--dataset trace` to replay a trace file. The trace can specify per-request fields such as:
+
+- `width`, `height`
+- `num_frames` (video)
+- `num_inference_steps`
+- `seed`, `fps`
+- optional `slo_ms` (per-request SLO target)
+
+By default (when `--dataset-path` is not provided), the script downloads a default trace from
+the HuggingFace dataset repo `asukaqaqzz/Dit_Trace`. The default filename can depend on `--task`
+(e.g., `t2v` uses a video trace).
+
+Current defaults:
+
+- `--task t2i` -> `sd3_trace.txt`
+- `--task t2v` -> `cogvideox_trace.txt`
+
+You can point to your own trace using `--dataset-path`.
+
+## 3. Benchmark Parameters
+
+### Basic flags
+
+- `--base-url`: Server address (the script calls `.../v1/chat/completions`).
+- `--model`: The OpenAI-compatible `model` field.
+- `--task`: Task type (e.g., `t2i`, `t2v`, `i2i`, `i2v`).
+- `--dataset`: Dataset mode (`vbench` / `trace` / `random`).
+- `--num-prompts`: Number of requests to send.
+
+Common optional flags:
+
+- `--output-file`: Write metrics to a JSON file.
+- `--disable-tqdm`: Disable the progress bar.
+
+### Resolution / frames / steps: CLI defaults vs dataset fields
+
+Related flags: `--width`, `--height`, `--num-frames`, `--fps`, `--num-inference-steps`.
+
+- For `vbench` / `random`: these CLI flags act as global defaults for all generated requests.
+- For `trace`: each request can carry its own fields (e.g., `width/height/num_frames/num_inference_steps`).
+
+Precedence rules for `trace` (i.e., what actually gets sent):
+
+- `width/height`: if either `--width` or `--height` is explicitly set, it overrides per-request values from the trace; otherwise per-request values are used when present.
+- `num_frames`: per-request `num_frames` takes precedence; otherwise fall back to `--num-frames`.
+- `num_inference_steps`: per-request `num_inference_steps` takes precedence; otherwise fall back to `--num-inference-steps`.
+
+### SLO, warmup, and max concurrency
+
+Enable SLO evaluation with `--slo`.
+
+- If a request in the trace already has `slo_ms`, that value is used.
+- Otherwise, the script runs warmup requests to infer a base unit time, estimates `expected_ms` by linearly scaling with area/frames/steps, and then sets `slo_ms = expected_ms * --slo-scale`.
+
+Warmup flags:
+
+- `--warmup-requests`: Number of warmup requests.
+- `--warmup-num-inference-steps`: Steps used during warmup.
+- For `--task t2v`: warmup requests are forced to use `num_frames=1` to make warmup faster and less noisy.
+
+Traffic / concurrency flags:
+
+- `--request-rate`: Target request rate (requests/second). If set to `inf`, the script sends all requests immediately.
+- `--max-concurrency`: Max number of in-flight requests (default: `1`). This can hard-cap the achieved QPS: if it is too small, requests will queue behind the semaphore, and both achieved throughput and observed SLO attainment can be skewed.
@@ -0,0 +1,86 @@
+# Benchmarks Guide
+
+This README explains how to (1) prepare benchmark datasets and (2) run the provided Qwen3-Omni benchmarks.
+
+## 1) Prepare the dataset (SeedTTS top100)
+
+```bash
+cd benchmarks/build_dataset
+pip install gdown
+
+# Download SeedTTS test set from Google Drive
+gdown --id 1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
+
+# Extract
+tar -xf seedtts_testset.tar
+
+# Copy metadata and extract top-100 prompts
+cp seedtts_testset/en/meta.lst meta.lst
+python extract_prompts.py -i meta.lst -o top100.txt -n 100
+
+# (Optional) clean up to save space
+rm -rf seedtts_testset seedtts_testset.tar meta.lst
+```
+
+Artifacts:
+- `benchmarks/build_dataset/top100.txt` — 100 text prompts (one per line).
+
+## 2) Run benchmarks
+
+All commands assume repo root (`vllm-omni`).
+
+### A. Transformers benchmark (offline, HF Transformers)
+
+```
+bash benchmarks/qwen3-omni/transformers/eval_qwen3_moe_omni_transformers.sh
+```
+
+What it does:
+- Runs `qwen3_omni_moe_transformers.py` over `top100.txt` with `--num_prompts 100`.
+- Outputs to `benchmarks/qwen3-omni/transformers/benchmark_results/`:
+  - `perf_stats.json` — aggregated & per-prompt TPS/latency (thinker/talker/code2wav/overall).
+  - `results.json` — per-prompt outputs and audio paths.
+  - `audio/` — ~100 generated `.wav` files.
+
+Key checks:
+- `overall_tps` and `*_tps_avg` should be non-zero and reasonably stable.
+- Investigate any 0/NaN or unusually low TPS / long-tail latency.
+
+### B. vLLM Omni end-to-end benchmark (pipeline)
+
+```
+bash benchmarks/qwen3-omni/vllm_omni/eval_qwen3_moe_omni.sh
+```
+
+What it does:
+- Runs `examples/offline_inference/qwen3_omni/end2end.py` with `--enable-stats`.
+- Uses `benchmarks/build_dataset/top100.txt` and writes to:
+  - Logs: `benchmarks/qwen3-omni/vllm_omni/logs/`
+    - `omni_llm_pipeline_text.orchestrator.stats.jsonl` — per-stage latency stats.
+    - `omni_llm_pipeline_text.overall.stats.jsonl` — end-to-end latency/TPS.
+    - `omni_llm_pipeline_text.stage{0,1,2}.log` — per-stage detailed logs/errors.
+  - Outputs: `benchmarks/qwen3-omni/vllm_omni/outputs/` — ~100 text and `.wav` files.
+
+Key checks:
+- Overall stats: end-to-end latency/TPS should be reasonable.
+- Orchestrator stats: per-stage latency should be stable; investigate long tails.
+- Stage logs: ensure no errors and no unusually slow stages.
+
+
+## Performance snapshot
+
+The chart below summarizes our measured Qwen3-Omni MoE end-to-end benchmark, comparing vLLM-Omni against HF Transformers. It shows the overall throughput advantage for vLLM-Omni. These are actual experiment results—please refer to this performance when evaluating or reproducing the benchmark.
+
+![vLLM-Omni vs HF](./vllm-omni-vs-hf.png)
+
+## Directory layout
+- `benchmarks/build_dataset/` — dataset prep utilities (e.g., SeedTTS top100).
+- `benchmarks/<model>/vllm_omni/` — vLLM-Omni pipeline benchmarks, logs, outputs.
+- Add new tasks under `benchmarks/<model>/...` with the same pattern: `transformers/`, `vllm_omni/`, task-specific README, and (optionally) dataset prep notes.
+- `benchmarks/<model>/vllm-omni-vs-hf.png` — current performance snapshot (overall throughput comparison).
+- `benchmarks/<model>/transformers/` — HF Transformers benchmarks (offline reference).
+
+## Troubleshooting
+- Make sure GPU/driver/FlashAttention2 requirements are met for the chosen model.
+- If downloads fail, confirm network access to Google Drive (`gdown`) and Hugging Face.
+- If audio files are missing, check for errors in stage logs or model generation.***
@@ -0,0 +1,26 @@
+#!/bin/bash
+# Qwen3-Omni Transformers Benchmark Evaluation Script
+# This script must be run from the vllm-omni root directory
+
+# Get the directory where this script is located
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+# Navigate to vllm-omni root directory (4 levels up from script location)
+VLLM_OMNI_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
+cd "$VLLM_OMNI_ROOT" || { echo "Error: Failed to navigate to vllm-omni directory"; exit 1; }
+
+echo "Working directory: $(pwd)"
+# Verify we're in the correct directory and run benchmark
+if [[ ! -f "benchmarks/qwen3-omni/transformers/qwen3_omni_moe_transformers.py" ]]; then
+    echo "Error: Not in vllm-omni root directory. Please run from vllm-omni folder."
+else
+    cd benchmarks/qwen3-omni/transformers
+
+    python qwen3_omni_moe_transformers.py --prompts_file ../../build_dataset/top100.txt --num_prompts 100
+
+    echo "Logs and outputs are saved to $(pwd)/benchmark_results:"
+    echo "  - perf_stats.json    Aggregated/per-prompt TPS and latency (thinker/talker/code2wav/overall)"
+    echo "  - results.json       Per-prompt outputs and audio paths"
+    echo "  - audio/             Generated wav files, there should be 100 wav file generated"
+    echo "Key checks: overall_tps and *_tps_avg should be non-zero and stable; investigate 0/NaN or unusually low TPS/long-tail latency."
+fi
@@ -0,0 +1,265 @@
+import time
+
+import torch
+from transformers import Qwen3OmniMoeForConditionalGeneration
+
+
+class Qwen3OmniMoeForConditionalGenerationWithLogging(Qwen3OmniMoeForConditionalGeneration):
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids: torch.Tensor | None = None,
+        speaker: str = "Ethan",
+        use_audio_in_video: bool = False,
+        return_audio: bool | None = None,
+        thinker_max_new_tokens: int = 1024,
+        thinker_eos_token_id: int = 151645,
+        talker_max_new_tokens: int = 4096,
+        talker_do_sample: bool = True,
+        talker_top_k: int = 50,
+        talker_top_p: float = 1.0,
+        talker_temperature: float = 0.9,
+        talker_repetition_penalty: float = 1.05,
+        **kwargs,
+    ):
+        total_t0 = time.time()
+        perf_stats = {
+            "thinker_tokens": 0,
+            "thinker_time_s": 0.0,
+            "thinker_tps": 0.0,
+            "talker_tokens": 0,
+            "talker_time_s": 0.0,
+            "talker_tps": 0.0,
+            "code2wav_tokens": 0,
+            "code2wav_time_s": 0.0,
+            "code2wav_tps": 0.0,
+            "total_tokens": 0,
+            "total_time_s": 0.0,
+            "total_tps": 0.0,
+        }
+        if return_audio and not self.has_talker:
+            raise ValueError(
+                "Cannot use talker when talker module not initialized. "
+                "Use `enable_talker` method or set enable_talker in config "
+                "to enable talker."
+            )
+        if return_audio is None:
+            return_audio = self.has_talker
+
+        shared_kwargs = {"use_audio_in_video": use_audio_in_video}
+        thinker_kwargs = {
+            "max_new_tokens": thinker_max_new_tokens,
+            "eos_token_id": thinker_eos_token_id,
+        }
+
+        talker_kwargs = {}
+        token2wav_kwargs = {}
+        if return_audio:
+            speaker_id = self.config.talker_config.speaker_id.get(speaker.lower())
+            if speaker_id is None:
+                raise NotImplementedError(f"Speaker {speaker} not implemented")
+            if input_ids.shape[0] != 1:
+                raise NotImplementedError("Qwen3-Omni currently does not support batched inference with audio output")
+            talker_suppressed_tokens = [
+                i
+                for i in range(
+                    self.config.talker_config.text_config.vocab_size - 1024,
+                    self.config.talker_config.text_config.vocab_size,
+                )
+                if i != self.config.talker_config.codec_eos_token_id
+            ]  # Suppress additional special tokens, should not be predicted
+            talker_kwargs = {
+                "max_new_tokens": talker_max_new_tokens,
+                "do_sample": talker_do_sample,
+                "top_k": talker_top_k,
+                "top_p": talker_top_p,
+                "temperature": talker_temperature,
+                "eos_token_id": self.config.talker_config.codec_eos_token_id,
+                "repetition_penalty": talker_repetition_penalty,
+                "suppress_tokens": talker_suppressed_tokens,
+                "output_hidden_states": True,
+                "return_dict_in_generate": True,
+            }
+            token2wav_kwargs = {}
+
+        for key, value in kwargs.items():
+            if key.startswith("thinker_"):
+                thinker_kwargs[key[len("thinker_") :]] = value
+            elif key.startswith("talker_"):
+                talker_kwargs[key[len("talker_") :]] = value
+            elif key.startswith("token2wav_"):
+                token2wav_kwargs[key[len("token2wav_") :]] = value
+            # Process special input values
+            elif key == "feature_attention_mask":
+                thinker_kwargs[key] = value
+                talker_kwargs["audio_feature_lengths"] = torch.sum(value, dim=1)
+            elif key in ("input_features", "attention_mask"):
+                thinker_kwargs[key] = value
+            # Put other key to shared kwargs
+            else:
+                shared_kwargs[key] = value
+
+        # Merge kwargs
+        for key, value in shared_kwargs.items():
+            if key not in thinker_kwargs:
+                thinker_kwargs[key] = value
+            if key not in talker_kwargs and key in ["image_grid_thw", "video_grid_thw", "video_second_per_grid"]:
+                talker_kwargs[key] = value
+            if key not in token2wav_kwargs:
+                token2wav_kwargs[key] = value
+
+        # 1. Generate from thinker module
+        generate_audio = return_audio and self.has_talker
+        if generate_audio:
+            thinker_kwargs["output_hidden_states"] = True
+            thinker_kwargs["return_dict_in_generate"] = True
+
+        t0 = time.time()
+        thinker_result = self.thinker.generate(input_ids=input_ids, **thinker_kwargs)
+        t1 = time.time()
+        perf_stats["thinker_time_s"] = max(0.0, t1 - t0)
+        try:
+            prompt_len = int(input_ids.shape[1]) if input_ids is not None else 0
+            total_len = int(thinker_result.sequences.shape[-1])
+            thinker_out_len = max(0, total_len - prompt_len)
+        except Exception:
+            thinker_out_len = 0
+        perf_stats["thinker_tokens"] = thinker_out_len
+        perf_stats["thinker_tps"] = (
+            (thinker_out_len / perf_stats["thinker_time_s"]) if perf_stats["thinker_time_s"] > 0 else 0.0
+        )
+
+        if not generate_audio:
+            perf_stats["total_tokens"] = perf_stats["thinker_tokens"]
+            perf_stats["total_time_s"] = time.time() - total_t0
+            perf_stats["total_tps"] = (
+                (perf_stats["total_tokens"] / perf_stats["total_time_s"]) if perf_stats["total_time_s"] > 0 else 0.0
+            )
+            # attach stats to self
+            setattr(self, "_perf_stats_last", perf_stats)
+            if not hasattr(self, "_perf_stats_history"):
+                setattr(self, "_perf_stats_history", [])
+            self._perf_stats_history.append(perf_stats)
+            return thinker_result, None
+
+        # 2. Prepare talker input
+        thinker_embed = torch.cat([hidden_states[0] for hidden_states in thinker_result.hidden_states], dim=1).to(
+            self.talker.device
+        )  # [1 t d]
+        thinker_hidden = torch.cat(
+            [
+                hidden_states[self.config.talker_config.accept_hidden_layer]
+                for hidden_states in thinker_result.hidden_states
+            ],
+            dim=1,
+        ).to(self.talker.device)  # [1 t d]
+
+        im_start_indexes = torch.cat(
+            (
+                torch.nonzero(input_ids[0] == self.config.im_start_token_id).squeeze(),
+                torch.tensor([thinker_result.sequences.shape[-1]], device=input_ids.device, dtype=input_ids.dtype),
+            ),
+            dim=-1,
+        ).to(self.talker.device)  # Shape [n_starts + 1]; Take batch 0 since batched inference is not supported here.
+        multimodal_mask = (
+            (thinker_result.sequences == self.config.thinker_config.audio_token_id) |
+            (thinker_result.sequences == self.config.thinker_config.image_token_id) |
+            (thinker_result.sequences == self.config.thinker_config.video_token_id)
+        ).to(self.talker.device)  # [1 t] # fmt: skip
+
+        talker_special_tokens = torch.tensor(
+            [[self.config.tts_bos_token_id, self.config.tts_eos_token_id, self.config.tts_pad_token_id]],
+            device=self.thinker.device,
+            dtype=input_ids.dtype,
+        )
+        tts_bos_embed, tts_eos_embed, tts_pad_embed = (
+            self.talker.text_projection(self.thinker.get_input_embeddings()(talker_special_tokens))
+            .to(self.talker.device)
+            .chunk(3, dim=1)
+        )  # 3 * [1 1 d]
+
+        talker_input_embeds = []  # [1 t d]
+        talker_input_ids = []
+        # For every chatml parts
+        for i in range(len(im_start_indexes) - 1):
+            im_start_index = im_start_indexes[i]
+            segment_end_index = im_start_indexes[i + 1]
+            role_token = input_ids[0][im_start_index + 1]
+            # Talker should ignore thinker system prompt
+            if role_token == self.config.system_token_id:
+                continue
+            # Talker takes word embeddings for tokens and hidden state from `accept_hidden_layer` for multimodal inputs
+            elif role_token == self.config.user_token_id:
+                talker_user_part = self._get_talker_user_parts(
+                    im_start_index, segment_end_index, multimodal_mask, thinker_hidden, thinker_embed
+                )
+                talker_input_embeds.append(talker_user_part)
+                talker_input_ids.append(thinker_result.sequences[:, im_start_index:segment_end_index])
+            # Take assistant output (for now)
+            elif role_token == self.config.assistant_token_id and i == len(im_start_indexes) - 2:
+                talker_assistant_embeds, talker_assistant_ids, trailing_text_hidden = self._get_talker_assistant_parts(
+                    im_start_index,
+                    segment_end_index,
+                    speaker_id,
+                    thinker_embed,
+                    tts_pad_embed,
+                    tts_bos_embed,
+                    tts_eos_embed,
+                )
+                talker_input_embeds.append(talker_assistant_embeds)
+                talker_input_ids.append(talker_assistant_ids)
+            # History assistant output (ignore for now)
+            elif role_token == self.config.assistant_token_id and i != len(im_start_indexes) - 2:
+                continue
+            else:
+                raise AssertionError("Expect role id after <|im_start|> (assistant, user, system)")
+        talker_input_embed = torch.cat([embed.to(self.talker.device) for embed in talker_input_embeds], dim=1)
+        talker_input_id = torch.cat([embed.to(self.talker.device) for embed in talker_input_ids], dim=1)
+        t2 = time.time()
+        talker_result = self.talker.generate(
+            inputs_embeds=talker_input_embed,
+            trailing_text_hidden=trailing_text_hidden,
+            tts_pad_embed=tts_pad_embed,
+            talker_input_ids=talker_input_id,  # Not use input_ids to prevent repetition penalty out of bound
+            **talker_kwargs,
+        )
+        t3 = time.time()
+        perf_stats["talker_time_s"] = max(0.0, t3 - t2)
+        talker_codes = (
+            torch.stack([hid[-1] for hid in talker_result.hidden_states if hid[-1] is not None], dim=1)
+            .transpose(1, 2)
+            .to(self.code2wav.device)
+        )
+        try:
+            # codes shape: (B, num_quantizers, T). We log T as token length.
+            perf_stats["talker_tokens"] = int(talker_codes.shape[-1])
+        except Exception:
+            perf_stats["talker_tokens"] = 0
+        perf_stats["talker_tps"] = (
+            (perf_stats["talker_tokens"] / perf_stats["talker_time_s"]) if perf_stats["talker_time_s"] > 0 else 0.0
+        )
+        t4 = time.time()
+        talker_wavs = self.code2wav.chunked_decode(talker_codes, chunk_size=300, left_context_size=25).float()
+        t5 = time.time()
+        perf_stats["code2wav_time_s"] = max(0.0, t5 - t4)
+        perf_stats["code2wav_tokens"] = perf_stats["talker_tokens"]  # same T, not times 16
+        perf_stats["code2wav_tps"] = (
+            (perf_stats["code2wav_tokens"] / perf_stats["code2wav_time_s"])
+            if perf_stats["code2wav_time_s"] > 0
+            else 0.0
+        )
+        perf_stats["total_tokens"] = perf_stats["thinker_tokens"] + perf_stats["talker_tokens"]
+        perf_stats["total_time_s"] = time.time() - total_t0
+        perf_stats["total_tps"] = (
+            (perf_stats["total_tokens"] / perf_stats["total_time_s"]) if perf_stats["total_time_s"] > 0 else 0.0
+        )
+        setattr(self, "_perf_stats_last", perf_stats)
+        if not hasattr(self, "_perf_stats_history"):
+            setattr(self, "_perf_stats_history", [])
+        self._perf_stats_history.append(perf_stats)
+        return thinker_result, talker_wavs.float()
+
+
+__all__ = [
+    "Qwen3OmniMoeForConditionalGenerationWithLogging",
+]
@@ -0,0 +1,275 @@
+import argparse
+import json
+import os
+
+import soundfile as sf
+from qwen3_omni_moe_model import Qwen3OmniMoeForConditionalGenerationWithLogging
+from qwen_omni_utils import process_mm_info
+from tqdm import tqdm
+from transformers import Qwen3OmniMoeProcessor
+
+MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+# MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Thinking"
+
+
+def load_prompts(prompts_file: str) -> list[str]:
+    """Load prompts from a text file, one prompt per line."""
+    prompts = []
+    with open(prompts_file, encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                prompts.append(line)
+    return prompts
+
+
+def run_benchmark(
+    model,
+    processor,
+    prompts: list[str],
+    output_dir: str = "benchmark_results",
+    speaker: str = "Ethan",
+    use_audio_in_video: bool = True,
+):
+    """
+    Run benchmark on a list of prompts and collect performance stats.
+
+    Args:
+        model: The Qwen3OmniMoe model
+        processor: The Qwen3OmniMoe processor
+        prompts: List of text prompts to process
+        output_dir: Directory to save results
+        speaker: Speaker voice for audio output
+        use_audio_in_video: Whether to use audio in video
+
+    Returns:
+        tuple: (aggregated_stats, results, audio_outputs)
+            - aggregated_stats: dict with aggregated performance statistics
+            - results: list of dicts with per-prompt results
+            - audio_outputs: list of audio tensors/arrays (or None if no audio)
+    """
+    os.makedirs(output_dir, exist_ok=True)
+    audio_dir = os.path.join(output_dir, "audio")
+    os.makedirs(audio_dir, exist_ok=True)
+
+    all_stats = []
+    results = []
+    audio_outputs = []
+
+    for idx, prompt in enumerate(tqdm(prompts, desc="Processing prompts")):
+        conversation = [
+            {
+                "role": "user",
+                "content": [{"type": "text", "text": prompt}],
+            },
+        ]
+
+        # Preparation for inference
+        text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
+        audios, images, videos = process_mm_info(conversation, use_audio_in_video=use_audio_in_video)
+        inputs = processor(
+            text=text,
+            audio=audios,
+            images=images,
+            videos=videos,
+            return_tensors="pt",
+            padding=True,
+            use_audio_in_video=use_audio_in_video,
+        )
+        inputs = inputs.to(model.device).to(model.dtype)
+
+        # Inference: Generation of the output text and audio
+        text_ids, audio = model.generate(
+            **inputs, speaker=speaker, thinker_return_dict_in_generate=True, use_audio_in_video=use_audio_in_video
+        )
+
+        # Decode output text
+        output_text = processor.batch_decode(
+            text_ids.sequences[:, inputs["input_ids"].shape[1] :],
+            skip_special_tokens=True,
+            clean_up_tokenization_spaces=False,
+        )[0]
+
+        # Collect performance stats
+        perf_stats = None
+        if hasattr(model, "_perf_stats_last"):
+            perf_stats = model._perf_stats_last.copy()
+            perf_stats["prompt_idx"] = idx
+            perf_stats["prompt"] = prompt
+            all_stats.append(perf_stats)
+
+        # Save audio and collect audio output
+        audio_path = None
+        audio_data = None
+        if audio is not None:
+            audio_data = audio.reshape(-1).detach().cpu().numpy()
+            audio_path = os.path.join(audio_dir, f"output_{idx:04d}.wav")
+            sf.write(
+                audio_path,
+                audio_data,
+                samplerate=24000,
+            )
+            audio_outputs.append(audio_data)
+        else:
+            audio_outputs.append(None)
+
+        # Save result
+        result = {
+            "idx": idx,
+            "prompt": prompt,
+            "output": output_text,
+            "audio_path": audio_path,
+            "perf_stats": perf_stats,
+        }
+        results.append(result)
+
+    # Aggregate statistics
+    aggregated_stats = aggregate_stats(all_stats)
+
+    # Save all results
+    results_path = os.path.join(output_dir, "results.json")
+    with open(results_path, "w", encoding="utf-8") as f:
+        json.dump(results, f, ensure_ascii=False, indent=2)
+
+    # Save aggregated stats
+    stats_path = os.path.join(output_dir, "perf_stats.json")
+    with open(stats_path, "w", encoding="utf-8") as f:
+        json.dump({"aggregated": aggregated_stats, "per_prompt": all_stats}, f, ensure_ascii=False, indent=2)
+
+    # Count saved audio files
+    num_audio_saved = sum(1 for a in audio_outputs if a is not None)
+    print(f"\nSaved {num_audio_saved} audio files to {audio_dir}/")
+
+    return aggregated_stats, results, audio_outputs
+
+
+def aggregate_stats(all_stats: list[dict]) -> dict:
+    """Aggregate performance statistics from multiple runs."""
+    if not all_stats:
+        return {}
+
+    keys = [
+        "thinker_tokens",
+        "thinker_time_s",
+        "thinker_tps",
+        "talker_tokens",
+        "talker_time_s",
+        "talker_tps",
+        "code2wav_tokens",
+        "code2wav_time_s",
+        "code2wav_tps",
+        "total_tokens",
+        "total_time_s",
+        "total_tps",
+    ]
+
+    aggregated = {
+        "num_samples": len(all_stats),
+    }
+
+    for key in keys:
+        values = [s.get(key, 0) for s in all_stats if key in s]
+        if values:
+            aggregated[f"{key}_sum"] = sum(values)
+            aggregated[f"{key}_avg"] = sum(values) / len(values)
+            aggregated[f"{key}_min"] = min(values)
+            aggregated[f"{key}_max"] = max(values)
+
+    # Calculate overall throughput
+    total_tokens = aggregated.get("total_tokens_sum", 0)
+    total_time = aggregated.get("total_time_s_sum", 0)
+    if total_time > 0:
+        aggregated["overall_tps"] = total_tokens / total_time
+
+    return aggregated
+
+
+def print_stats(stats: dict):
+    """Print performance statistics in a formatted way."""
+    print("\n" + "=" * 60)
+    print("Performance Statistics Summary")
+    print("=" * 60)
+
+    print(f"\nNumber of samples: {stats.get('num_samples', 0)}")
+
+    print("\n--- Thinker ---")
+    print(f"  Total tokens:  {stats.get('thinker_tokens_sum', 0):.0f}")
+    print(f"  Total time:    {stats.get('thinker_time_s_sum', 0):.2f}s")
+    print(f"  Avg TPS:       {stats.get('thinker_tps_avg', 0):.2f}")
+    print(f"  Min TPS:       {stats.get('thinker_tps_min', 0):.2f}")
+    print(f"  Max TPS:       {stats.get('thinker_tps_max', 0):.2f}")
+
+    print("\n--- Talker ---")
+    print(f"  Total tokens:  {stats.get('talker_tokens_sum', 0):.0f}")
+    print(f"  Total time:    {stats.get('talker_time_s_sum', 0):.2f}s")
+    print(f"  Avg TPS:       {stats.get('talker_tps_avg', 0):.2f}")
+    print(f"  Min TPS:       {stats.get('talker_tps_min', 0):.2f}")
+    print(f"  Max TPS:       {stats.get('talker_tps_max', 0):.2f}")
+
+    print("\n--- Code2Wav ---")
+    print(f"  Total tokens:  {stats.get('code2wav_tokens_sum', 0):.0f}")
+    print(f"  Total time:    {stats.get('code2wav_time_s_sum', 0):.2f}s")
+    print(f"  Avg TPS:       {stats.get('code2wav_tps_avg', 0):.2f}")
+    print(f"  Min TPS:       {stats.get('code2wav_tps_min', 0):.2f}")
+    print(f"  Max TPS:       {stats.get('code2wav_tps_max', 0):.2f}")
+
+    print("\n--- Overall ---")
+    print(f"  Total tokens:  {stats.get('total_tokens_sum', 0):.0f}")
+    print(f"  Total time:    {stats.get('total_time_s_sum', 0):.2f}s")
+    print(f"  Overall TPS:   {stats.get('overall_tps', 0):.2f}")
+    print(f"  Avg TPS:       {stats.get('total_tps_avg', 0):.2f}")
+    print(f"  Min TPS:       {stats.get('total_tps_min', 0):.2f}")
+    print(f"  Max TPS:       {stats.get('total_tps_max', 0):.2f}")
+
+    print("=" * 60 + "\n")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Qwen3-Omni Benchmark Script")
+    parser.add_argument(
+        "--prompts_file",
+        type=str,
+        default="benchmark/build_dataset/top100.txt",
+        help="Path to the prompts file (one prompt per line)",
+    )
+    parser.add_argument(
+        "--output_dir", type=str, default="benchmark_results", help="Directory to save benchmark results"
+    )
+    parser.add_argument("--model_path", type=str, default=MODEL_PATH, help="Path to the model")
+    parser.add_argument("--speaker", type=str, default="Ethan", help="Speaker voice for audio output")
+    parser.add_argument("--num_prompts", type=int, default=None, help="Number of prompts to process (default: all)")
+    args = parser.parse_args()
+
+    # Load model and processor
+    print(f"Loading model from {args.model_path}...")
+    model = Qwen3OmniMoeForConditionalGenerationWithLogging.from_pretrained(
+        args.model_path,
+        dtype="auto",
+        device_map="auto",
+        attn_implementation="flash_attention_2",
+    )
+    processor = Qwen3OmniMoeProcessor.from_pretrained(args.model_path)
+
+    # Benchmark mode
+    print(f"Loading prompts from {args.prompts_file}...")
+    prompts = load_prompts(args.prompts_file)
+
+    if args.num_prompts:
+        prompts = prompts[: args.num_prompts]
+
+    print(f"Running benchmark on {len(prompts)} prompts...")
+
+    aggregated_stats, results, audio_outputs = run_benchmark(
+        model=model,
+        processor=processor,
+        prompts=prompts,
+        output_dir=args.output_dir,
+        speaker=args.speaker,
+    )
+
+    print_stats(aggregated_stats)
+    print(f"\nResults saved to {args.output_dir}/")
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,37 @@
+#!/bin/bash
+# Qwen3-Omni Benchmark Evaluation Script
+# This script must be run from the vllm-omni root directory
+
+# Get the directory where this script is located
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+# Navigate to vllm-omni root directory (4 levels up from script location)
+VLLM_OMNI_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
+cd "$VLLM_OMNI_ROOT" || { echo "Error: Failed to navigate to vllm-omni directory"; exit 1; }
+
+echo "Working directory: $(pwd)"
+
+# Verify we're in the correct directory and run benchmark
+if [[ ! -d "benchmarks/qwen3-omni/vllm_omni" ]]; then
+    echo "Error: Not in vllm-omni root directory. Please run from vllm-omni folder."
+else
+    log_dir=benchmarks/qwen3-omni/vllm_omni/logs
+    outputs_dir=benchmarks/qwen3-omni/vllm_omni/outputs
+    end2end_script_path=examples/offline_inference/qwen3_omni/end2end.py
+    build_dataset_path=benchmarks/build_dataset/top100.txt
+
+    python $end2end_script_path --output-wav $outputs_dir \
+                      --query-type text \
+                      --txt-prompts $build_dataset_path \
+                      --enable-stats \
+                      --log-dir $log_dir
+    echo "Logs and outputs are saved in ${log_dir} and ${outputs_dir} respectively:"
+    echo "  - omni_llm_pipeline_text                       run dir/base name"
+    echo "  - omni_llm_pipeline_text.orchestrator.stats.jsonl  orchestrator-stage latency stats"
+    echo "  - omni_llm_pipeline_text.overall.stats.jsonl       overall latency/TPS stats"
+    echo "  - omni_llm_pipeline_text.stage0.log                per-stage detailed logs"
+    echo "  - omni_llm_pipeline_text.stage1.log"
+    echo "  - omni_llm_pipeline_text.stage2.log"
+    echo "Key checks: overall.stats.jsonl for end-to-end latency/TPS; orchestrator.stats.jsonl for stable per-stage latency; stage*.log for errors or long tails."
+    echo "  - outputs/             Generated txt and wav files, there should be 100 text and wav files generated respectively"
+fi
@@ -0,0 +1,760 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# ruff: noqa
+# code borrowed from https://github.com/pytorch/pytorch/blob/main/torch/utils/collect_env.py
+
+import datetime
+import locale
+import os
+import subprocess
+import sys
+
+# Unlike the rest of the PyTorch this file must be python2 compliant.
+# This script outputs relevant system environment info
+# Run it with `python collect_env.py` or `python -m torch.utils.collect_env`
+from collections import namedtuple
+
+import regex as re
+
+from vllm.envs import environment_variables
+
+try:
+    import torch
+
+    TORCH_AVAILABLE = True
+except (ImportError, NameError, AttributeError, OSError):
+    TORCH_AVAILABLE = False
+
+# System Environment Information
+SystemEnv = namedtuple(
+    "SystemEnv",
+    [
+        "torch_version",
+        "is_debug_build",
+        "cuda_compiled_version",
+        "gcc_version",
+        "clang_version",
+        "cmake_version",
+        "os",
+        "libc_version",
+        "python_version",
+        "python_platform",
+        "is_cuda_available",
+        "cuda_runtime_version",
+        "cuda_module_loading",
+        "nvidia_driver_version",
+        "nvidia_gpu_models",
+        "cudnn_version",
+        "pip_version",  # 'pip' or 'pip3'
+        "pip_packages",
+        "conda_packages",
+        "hip_compiled_version",
+        "hip_runtime_version",
+        "miopen_runtime_version",
+        "caching_allocator_config",
+        "is_xnnpack_available",
+        "cpu_info",
+        "rocm_version",  # vllm specific field
+        "vllm_version",  # vllm specific field
+        "vllm_omni_version",  # vllm-omni specific field
+        "vllm_build_flags",  # vllm specific field
+        "gpu_topo",  # vllm specific field
+        "env_vars",
+    ],
+)
+
+DEFAULT_CONDA_PATTERNS = {
+    "torch",
+    "numpy",
+    "cudatoolkit",
+    "soumith",
+    "mkl",
+    "magma",
+    "triton",
+    "optree",
+    "nccl",
+    "transformers",
+    "zmq",
+    "nvidia",
+    "pynvml",
+    "flashinfer-python",
+}
+
+DEFAULT_PIP_PATTERNS = {
+    "torch",
+    "numpy",
+    "mypy",
+    "flake8",
+    "triton",
+    "optree",
+    "onnx",
+    "nccl",
+    "transformers",
+    "zmq",
+    "nvidia",
+    "pynvml",
+    "flashinfer-python",
+}
+
+
+def run(command):
+    """Return (return-code, stdout, stderr)."""
+    shell = True if type(command) is str else False
+    try:
+        p = subprocess.Popen(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, shell=shell)
+        raw_output, raw_err = p.communicate()
+        rc = p.returncode
+        if get_platform() == "win32":
+            enc = "oem"
+        else:
+            enc = locale.getpreferredencoding()
+        output = raw_output.decode(enc)
+        if command == "nvidia-smi topo -m":
+            # don't remove the leading whitespace of `nvidia-smi topo -m`
+            #   because they are meaningful
+            output = output.rstrip()
+        else:
+            output = output.strip()
+        err = raw_err.decode(enc)
+        return rc, output, err.strip()
+
+    except FileNotFoundError:
+        cmd_str = command if isinstance(command, str) else command[0]
+        return 127, "", f"Command not found: {cmd_str}"
+
+
+def run_and_read_all(run_lambda, command):
+    """Run command using run_lambda; reads and returns entire output if rc is 0."""
+    rc, out, _ = run_lambda(command)
+    if rc != 0:
+        return None
+    return out
+
+
+def run_and_parse_first_match(run_lambda, command, regex):
+    """Run command using run_lambda, returns the first regex match if it exists."""
+    rc, out, _ = run_lambda(command)
+    if rc != 0:
+        return None
+    match = re.search(regex, out)
+    if match is None:
+        return None
+    return match.group(1)
+
+
+def run_and_return_first_line(run_lambda, command):
+    """Run command using run_lambda and returns first line if output is not empty."""
+    rc, out, _ = run_lambda(command)
+    if rc != 0:
+        return None
+    return out.split("\n")[0]
+
+
+def get_conda_packages(run_lambda, patterns=None):
+    if patterns is None:
+        patterns = DEFAULT_CONDA_PATTERNS
+    conda = os.environ.get("CONDA_EXE", "conda")
+    out = run_and_read_all(run_lambda, [conda, "list"])
+    if out is None:
+        return out
+
+    return "\n".join(
+        line for line in out.splitlines() if not line.startswith("#") and any(name in line for name in patterns)
+    )
+
+
+def get_gcc_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, "gcc --version", r"gcc (.*)")
+
+
+def get_clang_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, "clang --version", r"clang version (.*)")
+
+
+def get_cmake_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, "cmake --version", r"cmake (.*)")
+
+
+def get_nvidia_driver_version(run_lambda):
+    if get_platform() == "darwin":
+        cmd = "kextstat | grep -i cuda"
+        return run_and_parse_first_match(run_lambda, cmd, r"com[.]nvidia[.]CUDA [(](.*?)[)]")
+    smi = get_nvidia_smi()
+    return run_and_parse_first_match(run_lambda, smi, r"Driver Version: (.*?) ")
+
+
+def get_gpu_info(run_lambda):
+    if get_platform() == "darwin" or (
+        TORCH_AVAILABLE and hasattr(torch.version, "hip") and torch.version.hip is not None
+    ):
+        if TORCH_AVAILABLE and torch.cuda.is_available():
+            if torch.version.hip is not None:
+                prop = torch.cuda.get_device_properties(0)
+                if hasattr(prop, "gcnArchName"):
+                    gcnArch = " ({})".format(prop.gcnArchName)
+                else:
+                    gcnArch = "NoGCNArchNameOnOldPyTorch"
+            else:
+                gcnArch = ""
+            return torch.cuda.get_device_name(None) + gcnArch
+        return None
+    smi = get_nvidia_smi()
+    uuid_regex = re.compile(r" \(UUID: .+?\)")
+    rc, out, _ = run_lambda(smi + " -L")
+    if rc != 0:
+        return None
+    # Anonymize GPUs by removing their UUID
+    return re.sub(uuid_regex, "", out)
+
+
+def get_running_cuda_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, "nvcc --version", r"release .+ V(.*)")
+
+
+def get_cudnn_version(run_lambda):
+    """Return a list of libcudnn.so; it's hard to tell which one is being used."""
+    if get_platform() == "win32":
+        system_root = os.environ.get("SYSTEMROOT", "C:\\Windows")
+        cuda_path = os.environ.get("CUDA_PATH", "%CUDA_PATH%")
+        where_cmd = os.path.join(system_root, "System32", "where")
+        cudnn_cmd = '{} /R "{}\\bin" cudnn*.dll'.format(where_cmd, cuda_path)
+    elif get_platform() == "darwin":
+        # CUDA libraries and drivers can be found in /usr/local/cuda/. See
+        # https://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html#install
+        # https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#installmac
+        # Use CUDNN_LIBRARY when cudnn library is installed elsewhere.
+        cudnn_cmd = "ls /usr/local/cuda/lib/libcudnn*"
+    else:
+        cudnn_cmd = 'ldconfig -p | grep libcudnn | rev | cut -d" " -f1 | rev'
+    rc, out, _ = run_lambda(cudnn_cmd)
+    # find will return 1 if there are permission errors or if not found
+    if len(out) == 0 or (rc != 1 and rc != 0):
+        l = os.environ.get("CUDNN_LIBRARY")
+        if l is not None and os.path.isfile(l):
+            return os.path.realpath(l)
+        return None
+    files_set = set()
+    for fn in out.split("\n"):
+        fn = os.path.realpath(fn)  # eliminate symbolic links
+        if os.path.isfile(fn):
+            files_set.add(fn)
+    if not files_set:
+        return None
+    # Alphabetize the result because the order is non-deterministic otherwise
+    files = sorted(files_set)
+    if len(files) == 1:
+        return files[0]
+    result = "\n".join(files)
+    return "Probably one of the following:\n{}".format(result)
+
+
+def get_nvidia_smi():
+    # Note: nvidia-smi is currently available only on Windows and Linux
+    smi = "nvidia-smi"
+    if get_platform() == "win32":
+        system_root = os.environ.get("SYSTEMROOT", "C:\\Windows")
+        program_files_root = os.environ.get("PROGRAMFILES", "C:\\Program Files")
+        legacy_path = os.path.join(program_files_root, "NVIDIA Corporation", "NVSMI", smi)
+        new_path = os.path.join(system_root, "System32", smi)
+        smis = [new_path, legacy_path]
+        for candidate_smi in smis:
+            if os.path.exists(candidate_smi):
+                smi = '"{}"'.format(candidate_smi)
+                break
+    return smi
+
+
+def get_rocm_version(run_lambda):
+    """Returns the ROCm version if available, otherwise 'N/A'."""
+    return run_and_parse_first_match(run_lambda, "hipcc --version", r"HIP version: (\S+)")
+
+
+def get_vllm_version():
+    from vllm import __version__, __version_tuple__
+
+    if __version__ == "dev":
+        return "N/A (dev)"
+    version_str = __version_tuple__[-1]
+    if isinstance(version_str, str) and version_str.startswith("g"):
+        # it's a dev build
+        if "." in version_str:
+            # it's a dev build containing local changes
+            git_sha = version_str.split(".")[0][1:]
+            date = version_str.split(".")[-1][1:]
+            return f"{__version__} (git sha: {git_sha}, date: {date})"
+        else:
+            # it's a dev build without local changes
+            git_sha = version_str[1:]  # type: ignore
+            return f"{__version__} (git sha: {git_sha})"
+    return __version__
+
+
+def get_vllm_omni_version(run_lambda):
+    try:
+        import vllm_omni
+        from vllm_omni import __version__, __version_tuple__
+
+        version_str = __version_tuple__[-1]
+        if isinstance(version_str, str) and version_str.startswith("g"):
+            if "." in version_str:
+                git_sha = version_str.split(".")[0][1:]
+                date = version_str.split(".")[-1][1:]
+                return f"{__version__} (git sha: {git_sha}, date: {date})"
+            else:
+                git_sha = version_str[1:]
+                return f"{__version__} (git sha: {git_sha})"
+
+        package_dir = os.path.dirname(os.path.abspath(vllm_omni.__file__))
+        git_sha = run_and_read_all(run_lambda, f"git -C {package_dir} rev-parse --short HEAD")
+        if git_sha:
+            return f"{__version__} (git sha: {git_sha})"
+
+        return __version__
+    except ImportError:
+        return "N/A (vllm_omni not installed)"
+
+
+def summarize_vllm_build_flags():
+    # This could be a static method if the flags are constant, or dynamic if you need to check environment variables, etc.
+    return "CUDA Archs: {}; ROCm: {}".format(
+        os.environ.get("TORCH_CUDA_ARCH_LIST", "Not Set"),
+        "Enabled" if os.environ.get("ROCM_HOME") else "Disabled",
+    )
+
+
+def get_gpu_topo(run_lambda):
+    output = None
+
+    if get_platform() == "linux":
+        output = run_and_read_all(run_lambda, "nvidia-smi topo -m")
+        if output is None:
+            output = run_and_read_all(run_lambda, "rocm-smi --showtopo")
+
+    return output
+
+
+def get_cpu_info(run_lambda):
+    rc, out, err = 0, "", ""
+    if get_platform() == "linux":
+        rc, out, err = run_lambda("lscpu")
+    elif get_platform() == "win32":
+        rc, out, err = run_lambda(
+            "wmic cpu get Name,Manufacturer,Family,Architecture,ProcessorType,DeviceID, \
+        CurrentClockSpeed,MaxClockSpeed,L2CacheSize,L2CacheSpeed,Revision /VALUE"
+        )
+    elif get_platform() == "darwin":
+        rc, out, err = run_lambda("sysctl -n machdep.cpu.brand_string")
+    cpu_info = "None"
+    if rc == 0:
+        cpu_info = out
+    else:
+        cpu_info = err
+    return cpu_info
+
+
+def get_platform():
+    if sys.platform.startswith("linux"):
+        return "linux"
+    elif sys.platform.startswith("win32"):
+        return "win32"
+    elif sys.platform.startswith("cygwin"):
+        return "cygwin"
+    elif sys.platform.startswith("darwin"):
+        return "darwin"
+    else:
+        return sys.platform
+
+
+def get_mac_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, "sw_vers -productVersion", r"(.*)")
+
+
+def get_windows_version(run_lambda):
+    system_root = os.environ.get("SYSTEMROOT", "C:\\Windows")
+    wmic_cmd = os.path.join(system_root, "System32", "Wbem", "wmic")
+    findstr_cmd = os.path.join(system_root, "System32", "findstr")
+    return run_and_read_all(run_lambda, "{} os get Caption | {} /v Caption".format(wmic_cmd, findstr_cmd))
+
+
+def get_lsb_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, "lsb_release -a", r"Description:\t(.*)")
+
+
+def check_release_file(run_lambda):
+    return run_and_parse_first_match(run_lambda, "cat /etc/*-release", r'PRETTY_NAME="(.*)"')
+
+
+def get_os(run_lambda):
+    from platform import machine
+
+    platform = get_platform()
+
+    if platform == "win32" or platform == "cygwin":
+        return get_windows_version(run_lambda)
+
+    if platform == "darwin":
+        version = get_mac_version(run_lambda)
+        if version is None:
+            return None
+        return "macOS {} ({})".format(version, machine())
+
+    if platform == "linux":
+        # Ubuntu/Debian based
+        desc = get_lsb_version(run_lambda)
+        if desc is not None:
+            return "{} ({})".format(desc, machine())
+
+        # Try reading /etc/*-release
+        desc = check_release_file(run_lambda)
+        if desc is not None:
+            return "{} ({})".format(desc, machine())
+
+        return "{} ({})".format(platform, machine())
+
+    # Unknown platform
+    return platform
+
+
+def get_python_platform():
+    import platform
+
+    return platform.platform()
+
+
+def get_libc_version():
+    import platform
+
+    if get_platform() != "linux":
+        return "N/A"
+    return "-".join(platform.libc_ver())
+
+
+def is_uv_venv():
+    if os.environ.get("UV"):
+        return True
+    pyvenv_cfg_path = os.path.join(sys.prefix, "pyvenv.cfg")
+    if os.path.exists(pyvenv_cfg_path):
+        with open(pyvenv_cfg_path, "r") as f:
+            return any(line.startswith("uv = ") for line in f)
+    return False
+
+
+def get_pip_packages(run_lambda, patterns=None):
+    """Return `pip list` output. Note: will also find conda-installed pytorch and numpy packages."""
+    if patterns is None:
+        patterns = DEFAULT_PIP_PATTERNS
+
+    def run_with_pip():
+        try:
+            import importlib.util
+
+            pip_spec = importlib.util.find_spec("pip")
+            pip_available = pip_spec is not None
+        except ImportError:
+            pip_available = False
+
+        if pip_available:
+            cmd = [sys.executable, "-mpip", "list", "--format=freeze"]
+        elif is_uv_venv():
+            print("uv is set")
+            cmd = ["uv", "pip", "list", "--format=freeze"]
+        else:
+            raise RuntimeError("Could not collect pip list output (pip or uv module not available)")
+
+        out = run_and_read_all(run_lambda, cmd)
+        return "\n".join(line for line in out.splitlines() if any(name in line for name in patterns))
+
+    pip_version = "pip3" if sys.version[0] == "3" else "pip"
+    out = run_with_pip()
+    return pip_version, out
+
+
+def get_cachingallocator_config():
+    ca_config = os.environ.get("PYTORCH_CUDA_ALLOC_CONF", "")
+    return ca_config
+
+
+def get_cuda_module_loading_config():
+    if TORCH_AVAILABLE and torch.cuda.is_available():
+        torch.cuda.init()
+        config = os.environ.get("CUDA_MODULE_LOADING", "")
+        return config
+    else:
+        return "N/A"
+
+
+def is_xnnpack_available():
+    if TORCH_AVAILABLE:
+        import torch.backends.xnnpack
+
+        return str(torch.backends.xnnpack.enabled)  # type: ignore[attr-defined]
+    else:
+        return "N/A"
+
+
+def get_env_vars():
+    env_vars = ""
+    secret_terms = ("secret", "token", "api", "access", "password")
+    report_prefix = (
+        "TORCH",
+        "NCCL",
+        "PYTORCH",
+        "CUDA",
+        "CUBLAS",
+        "CUDNN",
+        "OMP_",
+        "MKL_",
+        "NVIDIA",
+    )
+    for k, v in os.environ.items():
+        if any(term in k.lower() for term in secret_terms):
+            continue
+        if k in environment_variables:
+            env_vars = env_vars + "{}={}".format(k, v) + "\n"
+        if k.startswith(report_prefix):
+            env_vars = env_vars + "{}={}".format(k, v) + "\n"
+
+    return env_vars
+
+
+def get_env_info():
+    run_lambda = run
+    pip_version, pip_list_output = get_pip_packages(run_lambda)
+
+    if TORCH_AVAILABLE:
+        version_str = torch.__version__
+        debug_mode_str = str(torch.version.debug)
+        cuda_available_str = str(torch.cuda.is_available())
+        cuda_version_str = torch.version.cuda
+        if not hasattr(torch.version, "hip") or torch.version.hip is None:  # cuda version
+            hip_compiled_version = hip_runtime_version = miopen_runtime_version = "N/A"
+        else:  # HIP version
+
+            def get_version_or_na(cfg, prefix):
+                _lst = [s.rsplit(None, 1)[-1] for s in cfg if prefix in s]
+                return _lst[0] if _lst else "N/A"
+
+            cfg = torch._C._show_config().split("\n")
+            hip_runtime_version = get_version_or_na(cfg, "HIP Runtime")
+            miopen_runtime_version = get_version_or_na(cfg, "MIOpen")
+            cuda_version_str = "N/A"
+            hip_compiled_version = torch.version.hip
+    else:
+        version_str = debug_mode_str = cuda_available_str = cuda_version_str = "N/A"
+        hip_compiled_version = hip_runtime_version = miopen_runtime_version = "N/A"
+
+    sys_version = sys.version.replace("\n", " ")
+
+    conda_packages = get_conda_packages(run_lambda)
+
+    rocm_version = get_rocm_version(run_lambda)
+    vllm_version = get_vllm_version()
+    vllm_omni_version = get_vllm_omni_version(run_lambda)
+    vllm_build_flags = summarize_vllm_build_flags()
+    gpu_topo = get_gpu_topo(run_lambda)
+
+    return SystemEnv(
+        torch_version=version_str,
+        is_debug_build=debug_mode_str,
+        python_version="{} ({}-bit runtime)".format(sys_version, sys.maxsize.bit_length() + 1),
+        python_platform=get_python_platform(),
+        is_cuda_available=cuda_available_str,
+        cuda_compiled_version=cuda_version_str,
+        cuda_runtime_version=get_running_cuda_version(run_lambda),
+        cuda_module_loading=get_cuda_module_loading_config(),
+        nvidia_gpu_models=get_gpu_info(run_lambda),
+        nvidia_driver_version=get_nvidia_driver_version(run_lambda),
+        cudnn_version=get_cudnn_version(run_lambda),
+        hip_compiled_version=hip_compiled_version,
+        hip_runtime_version=hip_runtime_version,
+        miopen_runtime_version=miopen_runtime_version,
+        pip_version=pip_version,
+        pip_packages=pip_list_output,
+        conda_packages=conda_packages,
+        os=get_os(run_lambda),
+        libc_version=get_libc_version(),
+        gcc_version=get_gcc_version(run_lambda),
+        clang_version=get_clang_version(run_lambda),
+        cmake_version=get_cmake_version(run_lambda),
+        caching_allocator_config=get_cachingallocator_config(),
+        is_xnnpack_available=is_xnnpack_available(),
+        cpu_info=get_cpu_info(run_lambda),
+        rocm_version=rocm_version,
+        vllm_version=vllm_version,
+        vllm_omni_version=vllm_omni_version,
+        vllm_build_flags=vllm_build_flags,
+        gpu_topo=gpu_topo,
+        env_vars=get_env_vars(),
+    )
+
+
+env_info_fmt = """
+==============================
+        System Info
+==============================
+OS                           : {os}
+GCC version                  : {gcc_version}
+Clang version                : {clang_version}
+CMake version                : {cmake_version}
+Libc version                 : {libc_version}
+
+==============================
+       PyTorch Info
+==============================
+PyTorch version              : {torch_version}
+Is debug build               : {is_debug_build}
+CUDA used to build PyTorch   : {cuda_compiled_version}
+ROCM used to build PyTorch   : {hip_compiled_version}
+
+==============================
+      Python Environment
+==============================
+Python version               : {python_version}
+Python platform              : {python_platform}
+
+==============================
+       CUDA / GPU Info
+==============================
+Is CUDA available            : {is_cuda_available}
+CUDA runtime version         : {cuda_runtime_version}
+CUDA_MODULE_LOADING set to   : {cuda_module_loading}
+GPU models and configuration : {nvidia_gpu_models}
+Nvidia driver version        : {nvidia_driver_version}
+cuDNN version                : {cudnn_version}
+HIP runtime version          : {hip_runtime_version}
+MIOpen runtime version       : {miopen_runtime_version}
+Is XNNPACK available         : {is_xnnpack_available}
+
+==============================
+          CPU Info
+==============================
+{cpu_info}
+
+==============================
+Versions of relevant libraries
+==============================
+{pip_packages}
+{conda_packages}
+""".strip()
+
+# both the above code and the following code use `strip()` to
+# remove leading/trailing whitespaces, so we need to add a newline
+# in between to separate the two sections
+env_info_fmt += "\n\n"
+
+env_info_fmt += """
+==============================
+         vLLM Info
+==============================
+ROCM Version                 : {rocm_version}
+vLLM Version                 : {vllm_version}
+vLLM-Omni Version            : {vllm_omni_version}
+vLLM Build Flags:
+  {vllm_build_flags}
+GPU Topology:
+  {gpu_topo}
+
+==============================
+     Environment Variables
+==============================
+{env_vars}
+""".strip()
+
+
+def pretty_str(envinfo):
+    def replace_nones(dct, replacement="Could not collect"):
+        for key in dct.keys():
+            if dct[key] is not None:
+                continue
+            dct[key] = replacement
+        return dct
+
+    def replace_bools(dct, true="Yes", false="No"):
+        for key in dct.keys():
+            if dct[key] is True:
+                dct[key] = true
+            elif dct[key] is False:
+                dct[key] = false
+        return dct
+
+    def prepend(text, tag="[prepend]"):
+        lines = text.split("\n")
+        updated_lines = [tag + line for line in lines]
+        return "\n".join(updated_lines)
+
+    def replace_if_empty(text, replacement="No relevant packages"):
+        if text is not None and len(text) == 0:
+            return replacement
+        return text
+
+    def maybe_start_on_next_line(string):
+        # If `string` is multiline, prepend a \n to it.
+        if string is not None and len(string.split("\n")) > 1:
+            return "\n{}\n".format(string)
+        return string
+
+    mutable_dict = envinfo._asdict()
+
+    # If nvidia_gpu_models is multiline, start on the next line
+    mutable_dict["nvidia_gpu_models"] = maybe_start_on_next_line(envinfo.nvidia_gpu_models)
+
+    # If the machine doesn't have CUDA, report some fields as 'No CUDA'
+    dynamic_cuda_fields = [
+        "cuda_runtime_version",
+        "nvidia_gpu_models",
+        "nvidia_driver_version",
+    ]
+    all_cuda_fields = dynamic_cuda_fields + ["cudnn_version"]
+    all_dynamic_cuda_fields_missing = all(mutable_dict[field] is None for field in dynamic_cuda_fields)
+    if TORCH_AVAILABLE and not torch.cuda.is_available() and all_dynamic_cuda_fields_missing:
+        for field in all_cuda_fields:
+            mutable_dict[field] = "No CUDA"
+        if envinfo.cuda_compiled_version is None:
+            mutable_dict["cuda_compiled_version"] = "None"
+
+    # Replace True with Yes, False with No
+    mutable_dict = replace_bools(mutable_dict)
+
+    # Replace all None objects with 'Could not collect'
+    mutable_dict = replace_nones(mutable_dict)
+
+    # If either of these are '', replace with 'No relevant packages'
+    mutable_dict["pip_packages"] = replace_if_empty(mutable_dict["pip_packages"])
+    mutable_dict["conda_packages"] = replace_if_empty(mutable_dict["conda_packages"])
+
+    # Tag conda and pip packages with a prefix
+    # If they were previously None, they'll show up as ie '[conda] Could not collect'
+    if mutable_dict["pip_packages"]:
+        mutable_dict["pip_packages"] = prepend(mutable_dict["pip_packages"], "[{}] ".format(envinfo.pip_version))
+    if mutable_dict["conda_packages"]:
+        mutable_dict["conda_packages"] = prepend(mutable_dict["conda_packages"], "[conda] ")
+    mutable_dict["cpu_info"] = envinfo.cpu_info
+    return env_info_fmt.format(**mutable_dict)
+
+
+def get_pretty_env_info():
+    return pretty_str(get_env_info())
+
+
+def main():
+    print("Collecting environment information...")
+    output = get_pretty_env_info()
+    print(output)
+
+    if TORCH_AVAILABLE and hasattr(torch, "utils") and hasattr(torch.utils, "_crash_handler"):
+        minidump_dir = torch.utils._crash_handler.DEFAULT_MINIDUMP_DIR
+        if sys.platform == "linux" and os.path.exists(minidump_dir):
+            dumps = [os.path.join(minidump_dir, dump) for dump in os.listdir(minidump_dir)]
+            latest = max(dumps, key=os.path.getctime)
+            ctime = os.path.getctime(latest)
+            creation_time = datetime.datetime.fromtimestamp(ctime).strftime("%Y-%m-%d %H:%M:%S")
+            msg = (
+                "\n*** Detected a minidump at {} created on {}, ".format(latest, creation_time)
+                + "if this is related to your bug please include it when you file a report ***"
+            )
+            print(msg, file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,20 @@
+ARG VLLM_BASE_IMAGE=vllm/vllm-openai
+ARG VLLM_BASE_TAG=v0.14.0
+FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
+ARG APP_DIR=/workspace/vllm-omni
+WORKDIR ${APP_DIR}
+
+COPY . .
+
+# Install system dependencies
+RUN apt-get update && \
+    apt-get install -y ffmpeg && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install vllm-omni into the same uv-managed Python environment used by the base image.
+RUN uv pip install --python "$(python3 -c 'import sys; print(sys.executable)')" --no-cache-dir ".[dev]"
+
+RUN ln -sf /usr/bin/python3 /usr/bin/python
+
+ENTRYPOINT []
@@ -0,0 +1,15 @@
+ARG VLLM_ASCEND_IMAGE=quay.nju.edu.cn/ascend/vllm-ascend
+ARG VLLM_ASCEND_TAG=v0.11.0rc2
+FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
+
+ARG APP_DIR=/vllm-workspace/vllm-omni
+WORKDIR ${APP_DIR}
+
+COPY . .
+
+# Install vllm-omni with dev dependencies
+RUN pip install --no-cache-dir -e ".[dev]"
+
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+ENTRYPOINT []
@@ -0,0 +1,32 @@
+ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.14.0
+FROM ${BASE_IMAGE} AS final
+
+ARG COMMON_WORKDIR=/app
+
+WORKDIR ${COMMON_WORKDIR}
+
+# Step 1: Setup - Install system dependencies
+RUN apt-get update && \
+    apt-get install -y ffmpeg && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN mkdir -p ${COMMON_WORKDIR}/vllm-omni
+
+# Step 2: Copy vllm-omni code and install without uv
+COPY . ${COMMON_WORKDIR}/vllm-omni
+RUN cd ${COMMON_WORKDIR}/vllm-omni && uv pip install --python "$(python3 -c 'import sys; print(sys.executable)')" --no-cache-dir ".[dev]"
+
+# When we are installing onnxruntime-rocm, we need to uninstall the system-installed onnxruntime first.
+# These are the dependencies of Qwen3-TTS.
+RUN uv pip uninstall onnxruntime --system && uv pip install --no-cache-dir onnxruntime-rocm sox --system
+
+RUN ln -sf /usr/bin/python3 /usr/bin/python
+
+CMD ["/bin/bash"]
+
+ENTRYPOINT []
+
+#Set entrypoint for vllm-openai official images
+FROM final AS vllm-openai
+ENTRYPOINT ["vllm", "serve", "--omni"]
@@ -3,16 +3,26 @@ nav:
 - User Guide:
  - Getting Started:
    - getting_started/quickstart.md
-    - getting_started/installation
+    - getting_started/installation/*
+  - Serving:
+    - OpenAI-Compatible API:
+      - Image Generation: serving/image_generation_api.md
+      - Image Edit: serving/image_edit_api.md
  - Examples:
    - examples/README.md
    - Offline Inference:
-      - Offline Example of vLLM-Omni for Qwen2.5-Omni: user_guide/examples/offline_inference/qwen2_5_omni.md
-      - Offline Example of vLLM-Omni for Qwen3-Omni: user_guide/examples/offline_inference/qwen3_omni.md
-      - Qwen-Image Offline Inference: user_guide/examples/offline_inference/qwen_image.md
+      - Image-To-Image: user_guide/examples/offline_inference/image_to_image.md
+      - Image-To-Video: user_guide/examples/offline_inference/image_to_video.md
+      - Qwen2.5-Omni: user_guide/examples/offline_inference/qwen2_5_omni.md
+      - Qwen3-Omni: user_guide/examples/offline_inference/qwen3_omni.md
+      - Qwen3-TTS Offline Inference: user_guide/examples/offline_inference/qwen3_tts.md
+      - Text-To-Image: user_guide/examples/offline_inference/text_to_image.md
+      - Text-To-Video: user_guide/examples/offline_inference/text_to_video.md
    - Online Serving:
-      - Online serving Example of vLLM-Omni for Qwen2.5-Omni: user_guide/examples/online_serving/qwen2_5_omni.md
-      - Online serving Example of vLLM-Omni for Qwen3-omni: user_guide/examples/online_serving/qwen3_omni.md
+      - Image-To-Image: user_guide/examples/online_serving/image_to_image.md
+      - Qwen2.5-Omni: user_guide/examples/online_serving/qwen2_5_omni.md
+      - Qwen3-Omni: user_guide/examples/online_serving/qwen3_omni.md
+      - Text-To-Image: user_guide/examples/online_serving/text_to_image.md
  - General:
    - usage/*
  - Configuration:
@@ -20,6 +30,14 @@ nav:
    - configuration/*
  - Models:
    - models/supported_models.md
+  - Features:
+    - Sleep Mode: features/sleep_mode.md
+    - Diffusion Features:
+      - Overview: user_guide/diffusion_acceleration.md
+      - TeaCache: user_guide/diffusion/teacache.md
+      - Cache-DiT: user_guide/diffusion/cache_dit_acceleration.md
+      - Parallelism Acceleration: user_guide/diffusion/parallelism_acceleration.md
+      - CPU Offloading: user_guide/diffusion/cpu_offload_diffusion.md
 - Developer Guide:
  - General:
    - contributing/README.md
@@ -27,12 +45,19 @@ nav:
      flatten_single_child_sections: true
  - Model Implementation:
    - contributing/model/README.md
+    - contributing/model/adding_omni_model.md
+    - contributing/model/adding_diffusion_model.md
  - CI: contributing/ci
  - Design Documents:
    - design/index.md
    - design/architecture_overview.md
-    - design/vllm_omni_design.md
-    - design/mrs_design.md
+    - Feature Design:
+      - design/feature/disaggregated_inference.md
+      - design/feature/ray_based_execution.md
+    - Module Design:
+      - design/module/ar_module.md
+      - design/module/dit_module.md
+      - design/module/entrypoint_module.md
  - Docs Guide: contributing/DOCS_GUIDE.md
 - API Reference:
  - api/README.md
@@ -59,4 +59,6 @@ vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, in
 - Multi-modality generation models (e.g. Qwen-Image)

 For more information, checkout the following:
+
+- [vllm-omni architecture design and recent roadmaps](https://docs.google.com/presentation/d/1qv4qMW1rKAqDREMXiUDLIgqqHQe7TDPj/edit?usp=sharing&ouid=110473603432222024453&rtpof=true&sd=true)
 - [vllm-omni announcement blogpost](https://blog.vllm.ai/2025/11/30/vllm-omni.html)
@@ -5,17 +5,22 @@
 Main entry points for vLLM-Omni inference and serving.

 - [vllm_omni.entrypoints.async_omni.AsyncOmni][]
- [vllm_omni.entrypoints.async_omni.AsyncOmniStageLLM][]
+- [vllm_omni.entrypoints.async_omni_diffusion.AsyncOmniDiffusion][]
+- [vllm_omni.entrypoints.async_omni_llm.AsyncOmniLLM][]
 - [vllm_omni.entrypoints.chat_utils.OmniAsyncMultiModalContentParser][]
 - [vllm_omni.entrypoints.chat_utils.OmniAsyncMultiModalItemTracker][]
 - [vllm_omni.entrypoints.chat_utils.parse_chat_messages_futures][]
 - [vllm_omni.entrypoints.cli.serve.OmniServeCommand][]
+- [vllm_omni.entrypoints.client_request_state.ClientRequestState][]
 - [vllm_omni.entrypoints.log_utils.OrchestratorMetrics][]
+- [vllm_omni.entrypoints.log_utils.StageRequestMetrics][]
+- [vllm_omni.entrypoints.log_utils.StageStats][]
 - [vllm_omni.entrypoints.omni.Omni][]
+- [vllm_omni.entrypoints.omni.OmniBase][]
+- [vllm_omni.entrypoints.omni_diffusion.OmniDiffusion][]
 - [vllm_omni.entrypoints.omni_llm.OmniLLM][]
- [vllm_omni.entrypoints.omni_llm.OmniStageLLM][]
 - [vllm_omni.entrypoints.omni_stage.OmniStage][]
- [vllm_omni.entrypoints.openai.serving_chat.OmniOpenAIServingChat][]
+- [vllm_omni.entrypoints.stage_utils.OmniStageTaskType][]

 ## Inputs

@@ -40,81 +45,63 @@ Engine classes for offline and online inference.
 - [vllm_omni.engine.PromptEmbedsPayload][]
 - [vllm_omni.engine.arg_utils.AsyncOmniEngineArgs][]
 - [vllm_omni.engine.arg_utils.OmniEngineArgs][]
+- [vllm_omni.engine.input_processor.OmniInputProcessor][]
 - [vllm_omni.engine.output_processor.MultimodalOutputProcessor][]
 - [vllm_omni.engine.output_processor.OmniRequestState][]
- [vllm_omni.engine.processor.OmniProcessor][]

 ## Core

 Core scheduling and caching components.

- [vllm_omni.core.dit_cache_manager.DiTCacheManager][]
- [vllm_omni.core.sched.diffusion_scheduler.DiffusionScheduler][]
- [vllm_omni.core.sched.generation_scheduler.GenerationScheduler][]
+- [vllm_omni.core.sched.omni_ar_scheduler.KVCacheTransferData][]
+- [vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler][]
+- [vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler][]
+- [vllm_omni.core.sched.output.OmniCachedRequestData][]
 - [vllm_omni.core.sched.output.OmniNewRequestData][]
- [vllm_omni.core.sched.scheduler.OmniScheduler][]
-
-## Model Executor
-
-Model execution components.
-
- [vllm_omni.model_executor.models.output_templates.OmniOutput][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni.Qwen2_5OmniForConditionalGeneration][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_talker.Qwen2_5OmniTalkerForConditionalGeneration][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker.Qwen2_5OmniAudioFeatureInputs][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker.Qwen2_5OmniConditionalGenerationMixin][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker.Qwen2_5OmniThinkerDummyInputsBuilder][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker.Qwen2_5OmniThinkerForConditionalGeneration][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker.Qwen2_5OmniThinkerMultiModalDataParser][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker.Qwen2_5OmniThinkerMultiModalProcessor][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_thinker.Qwen2_5OmniThinkerProcessingInfo][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_token2wav.Qwen2_5OmniToken2WavBigVGANModel][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_token2wav.Qwen2_5OmniToken2WavDiTModel][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_token2wav.Qwen2_5OmniToken2WavForConditionalGenerationVLLM][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_5_omni_token2wav.Qwen2_5OmniToken2WavModel][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_old.Qwen2EmbeddingModel][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_old.Qwen2ForCausalLM][]
- [vllm_omni.model_executor.models.qwen2_5_omni.qwen2_old.Qwen2Model][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_moe.Qwen3MoeForCausalLM][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni.Qwen3OmniMoeForConditionalGeneration][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_code2wav.Qwen3OmniMoeCode2Wav][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_code_predictor_mtp.Qwen3OmniCodePredictorBaseModel][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_code_predictor_mtp.Qwen3OmniMoeTalkerCodePredictor][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_talker.Qwen3OmniMoeModel][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_talker.Qwen3OmniMoeTalkerForConditionalGeneration][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3MoeLLMForCausalLM][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3MoeLLMModel][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3OmniMoeConditionalGenerationMixin][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3OmniMoeThinkerForConditionalGeneration][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3OmniMoeThinkerMultiModalProcessor][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3OmniMoeThinkerProcessingInfo][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3Omni_VisionTransformer][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3_VisionPatchEmbed][]
- [vllm_omni.model_executor.models.qwen3_omni.qwen3_omni_moe_thinker.Qwen3_VisionPatchMerger][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.vq.core_vq.DistributedGroupResidualVectorQuantization][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.vq.core_vq.DistributedResidualVectorQuantization][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.vq.core_vq.EuclideanCodebook][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.vq.core_vq.VectorQuantization][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.vq.core_vq.preprocess][]

 ## Configuration

 Configuration classes.

 - [vllm_omni.config.model.OmniModelConfig][]
+- [vllm_omni.diffusion.cache.teacache.config.TeaCacheConfig][]
+- [vllm_omni.distributed.omni_connectors.utils.config.ConnectorSpec][]
+- [vllm_omni.distributed.omni_connectors.utils.config.OmniTransferConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.configuration_qwen3_tts.Qwen3TTSConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.configuration_qwen3_tts.Qwen3TTSSpeakerEncoderConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.configuration_qwen3_tts.Qwen3TTSTalkerCodePredictorConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.configuration_qwen3_tts.Qwen3TTSTalkerConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_12hz.configuration_qwen3_tts_tokenizer_v2.Qwen3TTSTokenizerV2Config][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_12hz.configuration_qwen3_tts_tokenizer_v2.Qwen3TTSTokenizerV2DecoderConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1Config][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1DecoderBigVGANConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1DecoderConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1DecoderDiTConfig][]
+- [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1EncoderConfig][]

 ## Workers

 Worker classes and model runners for distributed inference.

- [vllm_omni.diffusion.worker.gpu_worker.GPUWorker][]
- [vllm_omni.diffusion.worker.gpu_worker.WorkerProc][]
+- [vllm_omni.diffusion.worker.gpu_diffusion_model_runner.GPUDiffusionModelRunner][]
+- [vllm_omni.diffusion.worker.gpu_diffusion_worker.GPUDiffusionWorker][]
+- [vllm_omni.diffusion.worker.gpu_diffusion_worker.WorkerProc][]
 - [vllm_omni.diffusion.worker.npu.npu_worker.NPUWorker][]
 - [vllm_omni.diffusion.worker.npu.npu_worker.NPUWorkerProc][]
+- [vllm_omni.worker.gpu_ar_model_runner.ExecuteModelState][]
 - [vllm_omni.worker.gpu_ar_model_runner.GPUARModelRunner][]
 - [vllm_omni.worker.gpu_ar_worker.GPUARWorker][]
- [vllm_omni.worker.gpu_diffusion_model_runner.GPUDiffusionModelRunner][]
- [vllm_omni.worker.gpu_diffusion_worker.GPUDiffusionWorker][]
 - [vllm_omni.worker.gpu_generation_model_runner.GPUGenerationModelRunner][]
 - [vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker][]
 - [vllm_omni.worker.gpu_model_runner.OmniGPUModelRunner][]
+- [vllm_omni.worker.npu.npu_ar_model_runner.ExecuteModelState][]
 - [vllm_omni.worker.npu.npu_ar_model_runner.NPUARModelRunner][]
 - [vllm_omni.worker.npu.npu_ar_worker.NPUARWorker][]
- [vllm_omni.worker.npu.npu_diffusion_model_runner.NPUDiffusionModelRunner][]
- [vllm_omni.worker.npu.npu_diffusion_worker.NPUDiffusionWorker][]
+- [vllm_omni.worker.npu.npu_generation_model_runner.NPUGenerationModelRunner][]
+- [vllm_omni.worker.npu.npu_generation_worker.NPUGenerationWorker][]
 - [vllm_omni.worker.npu.npu_model_runner.OmniNPUModelRunner][]
@@ -22,3 +22,21 @@ If you have custom stage configs file, launch the server with command below
 ```bash
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --stage-configs-path /path/to/stage_configs_file
 ```
+
+
+## bench
+
+Run benchmark tests for online serving throughput.
+Available Commands:
+
+```bash
+vllm bench serve --omni \
+    --model Qwen/Qwen2.5-Omni-7B \
+    --host server-host \
+    --port server-port \
+    --random-input-len 32 \
+    --random-output-len 4  \
+    --num-prompts  5
+```
+
+See [vllm bench serve](./bench/serve.md) for the full reference of all available arguments.
@@ -0,0 +1,359 @@
+# vLLM-Omni Benchmark CLI Guide
+The vllm bench command launches the vLLM-Omni benchmark to evaluate the performance of multimodal models.
+
+## Notes
+We currently only support using the "openai-chat-omni" backend.
+
+## Basic Parameter Description
+You can use `vllm bench serve --omni --help=all` to get descriptions of all parameters. The commonly used parameters are described below:
+- `--omni`  
+  Enable Omni (multimodal) mode, supporting multimodal inputs and outputs such as images, videos, and audio.
+
+- `--backend`  
+  Specify the backend adapter as openai-chat-omni, using OpenAI Chat compatible API behavior as the protocol. Currently only openai-chat-omni is supported.
+
+- `--model`  
+  The model identifier to load, filled according to the models supported by vLLM-Omni.
+
+- `--endpoint`  
+  The API endpoint exposed externally, to which clients send their requests.
+
+- `--dataset-name`  
+  The name of the dataset used; random-mm indicates generating random multimodal inputs (images, videos, audio).
+
+- `--num-prompts`  
+  The total number of requests to send, an integer.
+
+- `--max-concurrency`  
+  "Maximum number of concurrent requests. This can be used "
+        "to help simulate an environment where a higher level component "
+        "is enforcing a maximum number of concurrent requests. While the "
+        "--request-rate argument controls the rate at which requests are "
+        "initiated, this argument will control how many are actually allowed "
+        "to execute at a time. This means that when used in combination, the "
+        "actual request rate may be lower than specified with --request-rate, "
+        "if the server is not processing requests fast enough to keep up."
+
+- `--request-rate`  
+  "Number of requests per second. If this is inf, "
+        "then all the requests are sent at time 0. "
+        "Otherwise, we use Poisson process or gamma distribution "
+        "to synthesize the request arrival times."
+
+- `--ignore-eos`  
+  "Set ignore_eos flag when sending the benchmark request."
+
+- `--metric-percentiles`  
+  Comma-separated list of percentiles for selected metrics. "
+        "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
+        "Default value is \"99\"."
+        "Use \"--percentile-metrics\" to select metrics.
+
+- `--percentile-metrics`  
+        "Comma-separated list of selected metrics to report percentiles."
+                    "This argument specifies the metrics to report percentiles."
+                    'Allowed metric names are "ttft", "tpot", "itl", "e2el", "audio_ttfp", "audio_rtf". '
+
+- `--save-result`  
+Specify to save benchmark results to a json file
+
+- `--save-detailed`  
+"When saving the results, whether to include per request "
+        "information such as response, error, ttfs, tpots, etc."
+
+- `--result-dir`  
+ "Specify directory to save benchmark json results."
+        "If not specified, results are saved in the current directory."
+
+- `--result-filename`  
+"Specify the filename to save benchmark json results."
+        "If not specified, results will be saved in "
+        "{label}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
+
+- `--random-prefix-len`  
+  Number of fixed prefix tokens before the random context in a request.
+  The total input length is the sum of random-prefix-len and a random
+  context length sampled from [input_len * (1 - range_ratio),
+  input_len * (1 + range_ratio)].Only the random and random-mm modes
+  support this parameter.
+
+- `--random-input-len`  
+  Number of input tokens per request.Only the random and random-mm modes support this parameter.
+
+- `--random-output-len`  
+  Number of output tokens per request.Only the random and random-mm modes support this parameter.
+
+- `--random-range-ratio`  
+  Range ratio for sampling input/output length,
+  used only for random sampling. Must be in the range [0, 1) to define
+  a symmetric sampling range
+  [length * (1 - range_ratio), length * (1 + range_ratio)].
+  Only the random and random-mm modes support this parameter.
+
+- `--random-mm-base-items-per-request`  
+  Base number of multimodal items per request for random-mm.
+  Actual per-request count is sampled around this base using
+  --random-mm-num-mm-items-range-ratio.
+  Only the random-mm mode supports this parameter.
+
+- `--random-mm-limit-mm-per-prompt`  
+  Per-modality hard caps for items attached per request, e.g.
+  '{"image": 3, "video": 1, "audio": 1}'. The sampled per-request item
+  count is clamped to the sum of these limits. When a modality
+  reaches its cap, its buckets are excluded and probabilities are
+  renormalized.
+  Only the random-mm mode supports this parameter.
+
+- `--random-mm-num-mm-items-range-ratio`  
+  Range ratio r in [0, 1] for sampling items per request.
+  We sample uniformly from the closed integer range
+  [floor(n*(1-r)), ceil(n*(1+r))]
+  where n is the base items per request.
+  r=0 keeps it fixed; r=1 allows 0 items. The maximum is clamped
+  to the sum of per-modality limits from
+  --random-mm-limit-mm-per-prompt.
+  An error is raised if the computed min exceeds the max.
+  Only the random-mm mode supports this parameter.
+
+- `--random-mm-bucket-config`  
+  The bucket config is a dictionary mapping a multimodal item
+  sampling configuration to a probability.
+  Currently allows for 3 modalities: audio, images and videos.
+  A bucket key is a tuple of (height, width, num_frames)
+  The value is the probability of sampling that specific item.
+  Example:
+  --random-mm-bucket-config
+  "{(256, 256, 1): 0.5, (720, 1280, 16): 0.4, (0, 1, 5): 0.10}"
+  First item: images with resolution 256x256 w.p. 0.5
+  Second item: videos with resolution 720x1280 and 16 frames
+  Third item: audios with 1s duration and 5 channels w.p. 0.1
+  OBS.: If the probabilities do not sum to 1, they are normalized.
+  Only the random-mm mode supports this parameter
+
+## Usage Examples
+
+### Online Benchmark
+<details class="admonition abstract" markdown="1">
+<summary>Show more</summary>
+
+First start serving your model:
+
+```bash
+vllm serve Qwen/Qwen2.5-Omni-7B --omni
+```
+
+Then run the benchmarking for sharegpt:
+
+```bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+vllm bench serve \
+  --omni \
+  --port 43845 \
+  --model /home/models/Qwen/Qwen3-Omni-30B-A3B-Instruct \
+  --endpoint /v1/chat/completions \
+  --backend openai-chat-omni \
+  --num-prompts 2 \
+  --dataset-name sharegpt \
+  --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json \
+  --percentile-metrics ttft,tpot,itl,e2el
+```
+If successful, you will see the following output:
+```text
+============ Serving Benchmark Result ============
+Successful requests:                     2
+Failed requests:                         0
+Benchmark duration (s):                  81.63
+Request throughput (req/s):              0.02
+Peak concurrent requests:                2.00
+----------------End-to-end Latency----------------
+Mean E2EL (ms):                          56966.13
+Median E2EL (ms):                        56966.13
+P99 E2EL (ms):                           81016.80
+================== Text Result ===================
+Total input tokens:                      36
+Total generated tokens:                  5926
+Output token throughput (tok/s):         72.60
+Peak output token throughput (tok/s):    103.00
+Peak concurrent requests:                2.00
+Total Token throughput (tok/s):          73.04
+---------------Time to First Token----------------
+Mean TTFT (ms):                          124.76
+Median TTFT (ms):                        124.76
+P99 TTFT (ms):                           156.10
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          481.30
+Median TPOT (ms):                        481.30
+P99 TPOT (ms):                           947.55
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           25.11
+Median ITL (ms):                         0.33
+P99 ITL (ms):                            25.17
+================== Audio Result ==================
+Total audio duration generated(s):       3.95
+Total audio frames generated:            94890
+Audio throughput(audio duration/s):      0.05
+==================================================
+```
+
+Or run the benchmarking for random:
+
+```bash
+vllm bench serve \
+  --omni \
+  --port 43845 \
+  --endpoint /v1/chat/completions \
+  --backend openai-chat-omni \
+  --model /home/models/Qwen/Qwen3-Omni-30B-A3B-Instruct \
+  --dataset-name random \
+  --num-prompts 2 \
+  --random-prefix-len 5 \
+  --random-input-len 10 \
+  --random-output-len 100 \
+  --percentile-metrics ttft,tpot,itl,e2el,audio_ttfp,audio_rtf \
+  --ignore-eos
+```
+
+If successful, you will see the following output:
+
+```text
+============ Serving Benchmark Result ============
+Successful requests:                     2
+Failed requests:                         0
+Benchmark duration (s):                  24.35
+Request throughput (req/s):              0.08
+Peak concurrent requests:                2.00
+----------------End-to-end Latency----------------
+Mean E2EL (ms):                          22576.23
+Median E2EL (ms):                        22576.23
+P99 E2EL (ms):                           24205.72
+================== Text Result ===================
+Total input tokens:                      30
+Total generated tokens:                  8973
+Output token throughput (tok/s):         368.52
+Peak output token throughput (tok/s):    81.00
+Peak concurrent requests:                2.00
+Total Token throughput (tok/s):          369.76
+---------------Time to First Token----------------
+Mean TTFT (ms):                          125.16
+Median TTFT (ms):                        125.16
+P99 TTFT (ms):                           155.88
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          5.01
+Median TPOT (ms):                        5.01
+P99 TPOT (ms):                           5.42
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           34.15
+Median ITL (ms):                         0.01
+P99 ITL (ms):                            376.19
+================== Audio Result ==================
+Total audio duration generated(s):       3.95
+Total audio frames generated:            94890
+Audio throughput(audio duration/s):      0.16
+---------------Time to First Packet---------------
+Mean AUDIO_TTFP (ms):                    11756.89
+Median AUDIO_TTFP (ms):                  11756.89
+P99 AUDIO_TTFP (ms):                     20854.25
+-----------------Real Time Factor-----------------
+Mean AUDIO_RTF:                          3.75
+Median AUDIO_RTF:                        3.75
+P99 AUDIO_RTF:                           7.39
+==================================================
+```
+Notes:
+We use (audio generation time - first packet latency) / audio duration to calculate RTF.
+
+</details>
+
+### Multi-Modal Benchmark
+
+<details class="admonition abstract" markdown="1">
+<summary>Show more</summary>
+
+Benchmark the performance of multi-modal requests in vLLM-Omni.
+
+Generate synthetic image、video、audio inputs alongside random text prompts to stress-test vision models without external datasets.
+
+Notes:
+
+- Works only with online benchmark via the OpenAI backend (`--backend openai-chat-omni`) and endpoint `/v1/chat/completions`.
+
+Start the server (example):
+
+```bash
+vllm serve Qwen/Qwen2.5-Omni-7B --omni
+```
+
+It is recommended to use the flag `--ignore-eos` to simulate real responses. You can set the size of the output via the arg `random-output-len`.
+
+Then run the benchmarking script:
+```bash
+vllm bench serve \
+    --omni \
+  --dataset-name random-mm \
+  --port 40849 \
+  --model /home/models/Qwen/Qwen3-Omni-30B-A3B-Instruct \
+  --endpoint /v1/chat/completions \
+  --backend openai-chat-omni \
+  --request-rate 1 \
+  --num-prompts 1 \
+  --random-input-len 10 \
+  --random-range-ratio 0.0 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-num-mm-items-range-ratio 0 \
+  --random-mm-limit-mm-per-prompt '{"image":1,"video":1, "audio": 1}' \
+  --random-mm-bucket-config '{"(32, 32, 1)": 0.5, "(0, 1, 1)": 0.1, "(32, 32, 2)":0.4}' \
+  --ignore-eos \
+  --percentile-metrics ttft,tpot,itl \
+  --random-output-len 2 \
+  --extra_body '{"modalities": ["text"]}'
+```
+
+If successful, you will see the following output:
+
+```text
+============ Serving Benchmark Result ============
+Successful requests:                     1
+Failed requests:                         0
+Request rate configured (RPS):           1.00
+Benchmark duration (s):                  1.21
+Request throughput (req/s):              0.83
+Peak concurrent requests:                1.00
+================== Text Result ===================
+Total input tokens:                      10
+Total generated tokens:                  3
+Output token throughput (tok/s):         2.49
+Peak output token throughput (tok/s):    3.00
+Peak concurrent requests:                1.00
+Total Token throughput (tok/s):          10.77
+---------------Time to First Token----------------
+Mean TTFT (ms):                          179.74
+Median TTFT (ms):                        179.74
+P99 TTFT (ms):                           179.74
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          12.76
+Median TPOT (ms):                        12.76
+P99 TPOT (ms):                           12.76
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           12.76
+Median ITL (ms):                         12.76
+P99 ITL (ms):                            25.24
+================== Audio Result ==================
+Total audio duration generated(s):       0.00
+Total audio frames generated:            0
+Audio throughput(audio duration/s):      0.00
+==================================================
+```
+
+Behavioral notes:
+
+- If the requested base item count cannot be satisfied under the provided per-prompt limits, the tool raises an error rather than silently clamping.
+
+How sampling works:
+
+- Determine per-request item count k by sampling uniformly from the integer range defined by `--random-mm-base-items-per-request` and `--random-mm-num-mm-items-range-ratio`, then clamp k to at most the sum of per-modality limits.
+- For each of the k items, sample a bucket (H, W, T) according to the normalized probabilities in `--random-mm-bucket-config`, while tracking how many items of each modality have been added.
+- If a modality (e.g., image) reaches its limit from `--random-mm-limit-mm-per-prompt`, all buckets of that modality are excluded and the remaining bucket probabilities are renormalized before continuing.
+This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`.
+- The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`.
+</details>
@@ -2,4 +2,4 @@

 - For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm-omni/issues)
 - For coordinating contributions and development and discussing with other users and developers, please join `sig-omni` channel in our [Slack](https://slack.vllm.ai/) or use the [vLLM Forum](https://discuss.vllm.ai/)
- For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
+- For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm-omni/security/advisories) feature
@@ -0,0 +1,12 @@
+# Volunteers for Bugfix and CI
+
+We encourage you to check current docs and [issues](https://github.com/vllm-project/vllm-omni/issues) to find possible solutions for your questions. If non of these can solve it, please propose an issue to describe your questions about bug or CI problems for developing.
+
+If you have urgent need for locating and solving bugfix or CI problems, please find community volunteers below.
+
+| Dec 4-Dec 12 | Dec 15-Dec 19 | Dec 22-Dec 26 | Dec 29- Jan 2, 2026| Jan 5-Jan 9 | Jan 12-Jan 16 |
+|----------|----------|----------|----------|----------|----------|
+| <a href="https://github.com/congw729">Conw729</a> | <a href="https://github.com/yinpeiqi">yinpeiqi</a> | <a href="https://github.com/tzhouam">tzhouam</a> | <a href="https://github.com/SamitHuang">SamitHuang</a> | <a href="https://github.com/gcanlin">gcanlin</a> | <a href="https://github.com/natureofnature">natureofnature</a> |
+| <a href="https://github.com/david6666666">david6666666</a> | <a href="https://github.com/R2-Y">R2-Y</a> | <a href="https://github.com/hsliuustc0106">hsliuustc0106</a> | <a href="https://github.com/Gaohan123">Gaohan123</a> | <a href="https://github.com/ZJY0516">ZJY0516</a> | <a href="https://github.com/qibaoyuan">qibaoyuan</a> |
+
+We kindly welcome more contributors to fix bugs and contribute new features!
@@ -2,10 +2,20 @@

 This section lists the most common options for running vLLM-Omni.

-For options within a vLLM Engine. Please refer to [vLLM Configuration](https://docs.vllm.ai/en/v0.11.0/configuration/index.html)
+For options within a vLLM Engine. Please refer to [vLLM Configuration](https://docs.vllm.ai/en/v0.14.0/configuration/index.html)

 Currently, the main options are maintained by stage configs for each model.

 For specific example, please refer to [Qwen2.5-omni stage config](stage_configs/qwen2_5_omni.yaml)

 For introduction, please check [Introduction for stage config](./stage_configs.md)
+
+## Memory Configuration
+
+- **[GPU Memory Calculation and Configuration](./gpu_memory_utilization.md)** - Guide on how to calculate memory requirements and set up `gpu_memory_utilization` for optimal performance
+
+## Optimization Features
+
+- **[TeaCache Configuration](../user_guide/diffusion/teacache.md)** - Enable TeaCache adaptive caching for DiT models to achieve 1.5x-2.0x speedup with minimal quality loss
+- **[Cache-DiT Configuration](../user_guide/diffusion/cache_dit_acceleration.md)** - Enable Cache-DiT as cache acceleration backends for DiT models
+- **[Parallelism Configuration](../user_guide/diffusion/parallelism_acceleration.md)** - Enable parallelism (e.g., sequence parallelism) for for DiT models
@@ -0,0 +1,207 @@
+# GPU Memory Calculation and Configuration
+
+This guide explains how to calculate GPU memory requirements and properly configure `gpu_memory_utilization` for vLLM-Omni stages.
+
+## Overview
+
+`gpu_memory_utilization` is a critical parameter that controls how much GPU memory each stage can use. It's specified as a fraction between 0.0 and 1.0, where:
+- `0.8` means 80% of the GPU's total memory
+- `1.0` means 100% of the GPU's total memory (not recommended, leaves no buffer)
+
+## How Memory is Calculated
+
+### Memory Allocation Formula
+
+For each stage, vLLM-Omni calculates the requested memory as:
+
+```
+requested_memory = total_gpu_memory × gpu_memory_utilization
+```
+
+The system checks that:
+```
+free_memory ≥ requested_memory
+```
+
+If this condition is not met, the stage will fail to initialize with an error message showing the memory requirements.
+
+### Memory Components
+
+The total memory used by a stage includes:
+
+1. **Model Weights**: The size of the model parameters loaded on the GPU
+2. **KV Cache**: Memory for storing key-value cache during generation
+3. **Activation Memory**: Temporary memory for intermediate computations
+4. **System Overhead**: Memory used by CUDA, PyTorch, and other system components
+5. **Non-Torch Memory**: Memory allocated outside of PyTorch (e.g., CUDA graphs)
+
+### Example Calculation
+
+For a GPU with 80GB total memory:
+- `gpu_memory_utilization: 0.8` → 64GB available for the stage
+- `gpu_memory_utilization: 0.6` → 48GB available for the stage
+- `gpu_memory_utilization: 0.15` → 12GB available for the stage
+
+## Setting Up `gpu_memory_utilization`
+
+### Step 1: Determine GPU Memory
+
+First, check your GPU's total memory:
+
+```bash
+# Using nvidia-smi
+nvidia-smi --query-gpu=memory.total --format=csv
+
+# Or using Python
+python -c "import torch; print(f'{torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB')"
+```
+
+### Step 2: Estimate Model Memory Requirements
+
+#### For Autoregressive (AR) Stages
+
+AR stages typically need more memory due to:
+- Large model weights
+- KV cache for attention
+- Activation buffers
+
+#### For Diffusion/Generation Stages
+
+Diffusion stages (like code2wav) typically need less memory:
+- Smaller model components
+- Different memory access patterns
+
+**Typical values:**
+- `0.1 - 0.3` for most diffusion stages
+
+### Step 3: Consider Multi-Stage Scenarios
+
+When multiple stages share the same GPU, you must ensure the sum of their `gpu_memory_utilization` values doesn't exceed 1.0.
+
+**Example: Two stages on GPU 0**
+```yaml
+stage_args:
+  - stage_id: 0
+    runtime:
+      devices: "0"
+    engine_args:
+      gpu_memory_utilization: 0.6  # Uses 60% of GPU 0
+
+  - stage_id: 1
+    runtime:
+      devices: "0"
+    engine_args:
+      gpu_memory_utilization: 0.3  # Uses 30% of GPU 0
+      # Total: 90% of GPU 0 (safe, leaves 10% buffer)
+```
+
+**Important:** If stages run on different GPUs, each can use up to 1.0 independently.
+
+### Step 4: Account for Tensor Parallelism
+
+When using `tensor_parallel_size > 1`, the model is split across multiple GPUs, so each GPU needs less memory.
+
+**Example: 2-way tensor parallelism**
+```yaml
+stage_args:
+  - stage_id: 0
+    runtime:
+      devices: "0,1"  # Uses both GPUs
+    engine_args:
+      tensor_parallel_size: 2
+      gpu_memory_utilization: 0.6  # 60% per GPU
+      # Model is split, so each GPU uses ~30% of model memory
+```
+
+## Examples
+
+### Qwen3-Omni-MoE on 2x H100-80GB
+
+```yaml
+stage_args:
+  - stage_id: 0  # Thinker stage with TP=2
+    runtime:
+      devices: "0,1"
+    engine_args:
+      tensor_parallel_size: 2
+      gpu_memory_utilization: 0.6  # 48GB per GPU
+
+  - stage_id: 1  # Talker stage
+    runtime:
+      devices: "1"
+    engine_args:
+      gpu_memory_utilization: 0.3  # 24GB on GPU 1
+
+  - stage_id: 2  # Code2Wav stage
+    runtime:
+      devices: "0"
+    engine_args:
+      gpu_memory_utilization: 0.1  # 8GB on GPU 0
+```
+**Note:** In this configuration, stages 0 and 2 share GPU 0, but they run at different times in the pipeline, so their memory usage doesn't overlap.
+
+## Troubleshooting
+
+### Error: "Free memory is less than desired GPU memory utilization"
+
+This means the GPU doesn't have enough free memory when the stage starts.
+
+**Solutions:**
+1. Free up memory by closing other processes
+2. Reduce `gpu_memory_utilization` for this stage
+3. Use a GPU with more memory
+4. Move the stage to a different GPU
+
+### Error: OOM during inference
+
+The stage initialized but ran out of memory during processing.
+
+**Solutions:**
+1. Reduce `max_num_batched_tokens`
+2. Reduce `max_batch_size` in runtime config
+3. Lower `gpu_memory_utilization` slightly
+4. Enable quantization if supported
+
+### Memory Not Fully Utilized
+
+If you see low memory usage, you can:
+1. Increase `gpu_memory_utilization` to allow larger KV cache
+2. Increase `max_num_batched_tokens` for better batching
+3. Check if other stages are limiting throughput
+
+## Useful formula for Memory Calculation
+
+### KV Cache Memory
+
+The KV cache size depends on:
+- Number of sequences in batch
+- Sequence length (prompt + generation)
+- Model hidden size
+- Number of attention heads
+- Number of layers
+
+approximate Formula:
+```
+kv_cache_memory ≈ batch_size × seq_len × hidden_size × num_layers × 2 × dtype_size
+```
+2 for k & v
+
+### Model Weight Memory
+
+```
+model_memory ≈ num_parameters × dtype_size
+```
+
+For example:
+- 7B parameters in FP16: ~14GB
+- 7B parameters in FP32: ~28GB
+- 7B parameters in INT8: ~7GB
+
+### Activation Memory
+
+Activation memory is typically smaller but varies with:
+- Batch size
+- Sequence length
+- Model architecture
+
+It's usually 10-30% of model weight memory during inference.
@@ -40,8 +40,8 @@ stage_args:
    engine_args: # Engine arguments for a certain engine
      model_stage: thinker
      model_arch: Qwen2_5OmniForConditionalGeneration # The model implementation registered in model_executor/models/registry.py
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker # The specific worker used
-      scheduler_cls: vllm_omni.core.sched.scheduler.OmniScheduler # The specific scehduler used
+      worker_type: ar # The specific worker used
+      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler # The specific scehduler used
      gpu_memory_utilization: 0.8 # The gpu memory allocation for the stage within a single chip
      enforce_eager: true  # Now we only support eager mode
      trust_remote_code: true # Needed by huggingface config parsing
@@ -66,8 +66,8 @@ stage_args:
    engine_args:
      model_stage: talker
      model_arch: Qwen2_5OmniForConditionalGeneration
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.scheduler.OmniScheduler
+      worker_type: ar
+      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
      gpu_memory_utilization: 0.8
      enforce_eager: true
      trust_remote_code: true
@@ -92,8 +92,8 @@ stage_args:
    engine_args:
      model_stage: code2wav
      model_arch: Qwen2_5OmniForConditionalGeneration
-      worker_cls: vllm_omni.worker.gpu_diffusion_worker.GPUDiffusionWorker
-      scheduler_cls: vllm_omni.core.sched.diffusion_scheduler.DiffusionScheduler
+      worker_type: generation
+      scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
      gpu_memory_utilization: 0.15
      enforce_eager: true
      trust_remote_code: true
@@ -171,11 +171,11 @@ The model architecture class name that is registered in `model_executor/models/r

 #### `engine_args.worker_cls`

-The specific worker class to use for this stage. This determines how the model computations are executed. Examples include `vllm_omni.worker.gpu_ar_worker.GPUARWorker` for autoregressive stages and `vllm_omni.worker.gpu_diffusion_worker.GPUDiffusionWorker` for diffusion-based stages.
+The specific worker class to use for this stage. This determines how the model computations are executed. Examples include `vllm_omni.worker.gpu_ar_worker.GPUARWorker` for autoregressive stages and `vllm_omni.worker.gpu_generation_worker.GPUGenerationWorker` for diffusion-based stages.

 #### `engine_args.scheduler_cls`

-The scheduler class to use for this stage. The scheduler manages request queuing, batching, and execution order. Examples include `vllm_omni.core.sched.scheduler.OmniScheduler` for standard stages and `vllm_omni.core.sched.diffusion_scheduler.DiffusionScheduler` for diffusion stages.
+The scheduler class to use for this stage. The scheduler manages request queuing, batching, and execution order. Examples include `vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler` for standard stages and `vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler` for diffusion stages.

 #### `engine_args.gpu_memory_utilization`

@@ -183,6 +183,9 @@ The fraction of GPU memory to allocate for this stage within a single GPU chip.

 Default: `0.8`

+!!! tip "Memory Configuration Guide"
+    For detailed information on how to calculate memory requirements and properly configure `gpu_memory_utilization`, see the [GPU Memory Calculation and Configuration Guide](./gpu_memory_utilization.md).
+
 #### `engine_args.enforce_eager`

 Whether to enforce eager execution mode. When set to `true`, the engine will run in eager mode without using CUDA graphs or other compilation optimizations. Currently, vLLM-Omni only supports eager mode.
@@ -8,8 +8,8 @@ stage_args:
    engine_args:
      model_stage: thinker
      model_arch: Qwen2_5OmniForConditionalGeneration
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.scheduler.OmniScheduler
+      worker_type: ar
+      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
      gpu_memory_utilization: 0.8
      enforce_eager: true  # Now we only support eager mode
      trust_remote_code: true
@@ -34,8 +34,8 @@ stage_args:
    engine_args:
      model_stage: talker
      model_arch: Qwen2_5OmniForConditionalGeneration
-      worker_cls: vllm_omni.worker.gpu_ar_worker.GPUARWorker
-      scheduler_cls: vllm_omni.core.sched.scheduler.OmniScheduler
+      worker_type: ar
+      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
      gpu_memory_utilization: 0.8
      enforce_eager: true
      trust_remote_code: true
@@ -60,8 +60,8 @@ stage_args:
    engine_args:
      model_stage: code2wav
      model_arch: Qwen2_5OmniForConditionalGeneration
-      worker_cls: vllm_omni.worker.gpu_diffusion_worker.GPUDiffusionWorker
-      scheduler_cls: vllm_omni.core.sched.diffusion_scheduler.DiffusionScheduler
+      worker_type: generation
+      scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
      gpu_memory_utilization: 0.15
      enforce_eager: true
      trust_remote_code: true
@@ -13,23 +13,18 @@ vLLM-Omni uses `uv` as the environment manager, to create and manage Python envi
 uv venv --python 3.12 --seed
 source .venv/bin/activate
 ```
-### Development Environment for vLLM

-vLLM-Omni is built on top of the stable release 0.11.0 of vLLM. You may install `vllm==0.11.0` directly or install the library from source if you need to check, modify or debug with source code of vLLM.
+### Development Environment for vLLM and vLLM-Omni

-### Development Environment for vLLM-Omni
-
-Install vLLM-Omni from source with development dependencies:
-
-```bash
-git clone https://github.com/vllm-project/vllm-omni.git
-cd vllm_omni
-uv pip install -e ".[dev]"
-```
+vLLM-Omni is quickly evolving, please see the [installation guide](../getting_started/installation/README.md) for details. It's recommended to build from source to provide the latest development environment.

 !!! tip
    vLLM-Omni is compatible with Python versions 3.10 to 3.12. However, we recommend developing with Python 3.12 to minimize the chance of your local environment clashing with our CI environment.

+### Adding a new model to vLLM-Omni
+
+Please check [model implementation](model/README.md) for how to add diffusion and omni-modality models to vLLM-Omni.
+
 ### Linting

 vLLM-Omni uses `pre-commit` to lint and format the codebase. See [pre-commit documentation](https://pre-commit.com/#usage) if `pre-commit` is new to you. Setting up `pre-commit` is as easy as:
@@ -121,7 +116,7 @@ Only specific types of PRs will be reviewed. The PR title is prefixed appropriat
 - `[Model]` for adding a new model or improving an existing model. Model name should appear in the title.
 - `[Frontend]` For changes on the vLLM-Omni frontend (e.g., OpenAI API server, `OmniLLM` class, etc.)
 - `[Kernel]` for changes affecting CUDA kernels or other compute kernels.
- `[Core]` for changes in the core vLLM-Omni logic (e.g., `OmniProcessor`, `OmniScheduler`, etc.)
+- `[Core]` for changes in the core vLLM-Omni logic (e.g., `OmniProcessor`, `OmniARScheduler`, etc.)
 - `[Hardware][Vendor]` for hardware-specific changes. Vendor name should appear in the prefix, such as [Ascend] for Ascend NPUs.
 - `[Misc]` for PRs that do not fit the above categories. Please use this sparingly.

@@ -0,0 +1,160 @@
+# Markers for Tests
+
+By adding markers before test functions, tests can later be executed uniformly by simply declaring the corresponding marker type.
+
+## Current Markers
+Defined in `pyproject.toml`:
+
+| Marker             | Description                                             |
+| ------------------ | ------------------------------------------------------- |
+| `core_model`       | Core model tests (run in each PR)                       |
+| `diffusion`        | Diffusion model tests                                   |
+| `omni`             | Omni model tests                                        |
+| `cache`            | Cache backend tests                                     |
+| `parallel`         | Parallelism/distributed tests                           |
+| `cpu`              | Tests that run on CPU                                   |
+| `gpu`              | Tests that run on GPU (auto-added)                      |
+| `cuda`             | Tests that run on CUDA (auto-added)                     |
+| `rocm`             | Tests that run on AMD/ROCm (auto-added)                 |
+| `npu`              | Tests that run on NPU/Ascend (auto-added)               |
+| `H100`             | Tests that require H100 GPU                             |
+| `L4`               | Tests that require L4 GPU                               |
+| `MI325`            | Tests that require MI325 GPU (AMD/ROCm)                 |
+| `A2`               | Tests that require A2 NPU                               |
+| `A3`               | Tests that require A3 NPU                               |
+| `distributed_cuda` | Tests that require multi cards on CUDA platform         |
+| `distributed_rocm` | Tests that require multi cards on ROCm platform         |
+| `distributed_npu`  | Tests that require multi cards on NPU platform          |
+| `skipif_cuda`      | Skip if the num of CUDA cards is less than the required |
+| `skipif_rocm`      | Skip if the num of ROCm cards is less than the required |
+| `skipif_npu`       | Skip if the num of NPU cards is less than the required  |
+| `slow`             | Slow tests (may skip in quick CI)                       |
+| `benchmark`        | Benchmark tests                                         |
+
+For those markers shown as auto-added, they will be added by the `@hardware_test` decorator.
+
+### Example usage for markers
+
+```python
+from tests.utils import hardware_test
+
+@pytest.mark.core_model
+@pytest.mark.omni
+@hardware_test(
+   res={"cuda": "L4", "rocm": "MI325", "npu": "A2"},
+   num_cards=2,
+)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_video_to_audio()
+    ...
+```
+### Decorator: `@hardware_test`
+
+This decorator is intended to make hardware-aware, cross-platform test authoring easier and more robust for CI/CD environments. The `hardware_test` decorator in `vllm-omni/tests/utils.py` performs the following actions:
+
+1. **Applies platform and resource markers**  
+   Adds the appropriate pytest markers for each specified hardware platform (e.g., `cuda`, `rocm`, `npu`) and resource type (e.g., `L4`, `H100`, `MI325`, `A2`, `A3`).
+   ```
+   @pytest.mark.cuda
+   @pytest.mark.L4
+   ```
+2. **Handles multi-card (distributed) scenarios**  
+   For tests requiring multiple cards, it automatically adds distributed markers such as `distributed_cuda`, `distributed_rocm`, or `distributed_npu`.
+   ```
+   @pytest.mark.distributed_cuda(num_cards=num_cards)
+   ```
+3. **Supports flexible card requirements**  
+   Accepts `num_cards` as either a single integer for all platforms or as a dictionary with per-platform values. If not specified, defaults to 1 card per platform.
+
+4. **Integrates resource validation**  
+   On CUDA, adds a skip marker (`skipif_cuda`) if the system does not have the required number of devices.
+   Support for `skipif_rocm` and `skipif_npu` will be implemented later.
+
+
+5. **Runs each test in a new process**  
+   Automatically wraps the distributed test with a decorator (`@create_new_process_for_each_test`) to ensure isolation and compatibility with multi-process hardware backends.
+
+6. **Works with pytest filtering**  
+   Allows tests to be filtered and selected at runtime using standard pytest marker expressions (e.g., `-m "distributed_cuda and L4"`).
+
+#### Example usage for decorator
+- Single call for multiple platforms:
+    ```python
+    @hardware_test(
+        res={"cuda": "L4", "rocm": "MI325", "npu": "A2"},
+        num_cards={"cuda": 2, "rocm": 2, "npu": 2},
+    )
+    ```
+    or
+    ```python
+    @hardware_test(
+        res={"cuda": "L4", "rocm": "MI325", "npu": "A2"},
+        num_cards=2,
+    )
+    ```
+- `res` must be a dict; supported resources: CUDA (L4/H100), ROCm (MI325), NPU (A2/A3)
+- `num_cards` can be int (all platforms) or dict (per platform); defaults to 1 when missing
+- `hardware_test` automatically applies `@create_new_process_for_each_test` for distributed tests.
+- Distributed markers (`distributed_cuda`, `distributed_rocm`, `distributed_npu`) are auto-added for multi-card cases
+- Filtering examples:
+    - CUDA only: `pytest -m "distributed_cuda and L4"`
+    - ROCm only: `pytest -m "distributed_rocm and MI325"`
+    - NPU only: `pytest -m "distributed_npu"`
+
+## Add Support for a New Platform
+
+If you want to add support for a new platform (e.g., "tpu" for a new accelerator), follow these steps:
+
+1. **Extend the marker list in your pytest config** so that platform/resource markers are defined:
+   ```toml
+   # In pyproject.toml or pytest.ini
+   [tool.pytest.ini_options]
+   markers = [
+       # ... existing markers ...
+       "tpu: Tests that require TPU device",
+       "TPU_V3: Tests that require TPU v3 hardware",
+       "distributed_tpu: Tests that require multiple TPU devices",
+   ]
+   ```
+2. **Implement a marker construction function for your platform** in `vllm-omni/tests/utils.py`:
+   ```python
+   # In vllm-omni/tests/utils.py
+
+   def tpu_marks(*, res: str, num_cards: int):
+       test_platform = pytest.mark.tpu
+       if res == "TPU_V3":
+           test_resource = pytest.mark.TPU_V3
+       else:
+           raise ValueError(
+               f"Invalid TPU resource type: {res}. Supported: TPU_V3")
+
+       if num_cards == 1:
+           return [test_platform, test_resource]
+       else:
+           test_distributed = pytest.mark.distributed_tpu(num_cards=num_cards)
+           # Optionally: add skipif_tpu when implemented
+           return [test_platform, test_resource, test_distributed]
+   ```
+3. **Update `hardware_test` to recognize your new platform**:
+    In the relevant place (see the `hardware_test` implementation), add:
+    ```python
+    if platform == "tpu":
+        marks = tpu_marks(res=resource, num_cards=cards)
+    ```
+4. **(Recommended) Add a test using your new markers**:
+   ```python
+   @hardware_test(
+       res={"tpu": "TPU_V3"},
+       num_cards=2,
+   )
+   def test_my_tpu_feature():
+       ...
+   ```
+
+**Summary**:  
+- Add pytest markers for your new platform/resources  
+- Implement a marker function (`xxx_marks`)  
+- Plug into `hardware_test`  
+- You're done: tests decorated with `@hardware_test` using your platform now automatically get the correct markers, distribution, and isolation!
+
+See code in `vllm-omni/tests/utils.py` for existing examples (`cuda_marks`, `rocm_marks`, `npu_marks`).
@@ -0,0 +1,280 @@
+# Test File Structure and Style Guide
+
+To ensure project maintainability and sustainable development, we encourage contributors to submit test code (unit tests, system tests, or end-to-end tests) alongside their code changes. This document outlines the guidelines for organizing and naming test files.
+
+## Test Types
+
+### Unit Tests and System Tests
+For unit tests and system tests, we strongly recommend placing test files in the same directory structure as the source code being tested, using the naming convention `test_*.py`.
+
+### End-to-End (E2E) Tests for Models
+End-to-end tests verify the complete functionality of a system or component. For our project, the E2E tests for different omni models are organized into two subdirectories:
+
+- **`tests/e2e/offline_inference/`**: Tests for offline inference modes (e.g., Qwen3Omni offline inference)
+
+- **`tests/e2e/online_serving/`**: Tests for online serving scenarios (e.g., API server tests)
+
+**Example:** The test file for `vllm_omni/entrypoints/omni_llm.py` should be located at `tests/entrypoints/test_omni_llm.py`.
+
+## Test Directory Structure
+
+The ideal directory structure mirrors the source code organization:
+
+```
+vllm_omni/                          tests/
+├── config/                    →    ├── config/
+│   └── model.py                    │   └── test_model.py
+│
+├── core/                      →    ├── core/
+│   └── sched/                      │   └── sched/                    # Maps to core/sched/
+│       ├── omni_ar_scheduler.py    │       ├── test_omni_ar_scheduler.py
+│       ├── omni_generation_scheduler.py │  ├── test_omni_generation_scheduler.py
+│       └── output.py               │       └── test_output.py
+│
+├── diffusion/                 →    ├── diffusion/
+│   ├── diffusion_engine.py         │   ├── test_diffusion_engine.py
+│   ├── omni_diffusion.py           │   ├── test_omni_diffusion.py
+│   ├── attention/                  │   ├── attention/                # Maps to diffusion/attention/
+│   │   └── backends/               │   │   └── test_*.py
+│   ├── models/                     │   ├── models/                   # Maps to diffusion/models/
+│   │   ├── qwen_image/             │   │   ├── qwen_image/
+│   │   │   └── ...                 │   │   │   └── test_*.py
+│   │   └── z_image/                │   │   └── z_image/
+│   │       └── ...                 │   │       └── test_*.py
+│   └── worker/                     │   └── worker/                   # Maps to diffusion/worker/
+│       └── ...                     │       └── test_*.py
+│
+├── distributed/               →    ├── distributed/
+│   └── ...                         │   └── test_*.py
+│
+├── engine/                    →    ├── engine/
+│   ├── processor.py                │   ├── test_processor.py
+│   └── output_processor.py         │   └── test_output_processor.py
+│
+├── entrypoints/               →    ├── entrypoints/
+│   ├── omni_llm.py                 │   ├── test_omni_llm.py          # UT: OmniLLM core logic (mocked)
+│   ├── omni_stage.py               │   ├── test_omni_stage.py         # UT: OmniStage logic
+│   ├── omni.py                     │   ├── test_omni.py               # E2E: Omni class (offline inference)
+│   ├── async_omni.py               │   ├── test_async_omni.py         # E2E: AsyncOmni class
+│   ├── cli/                        │   ├── cli/                       # Maps to entrypoints/cli/
+│   │   └── ...                     │   │   └── test_*.py
+│   └── openai/                     │   └── openai/                     # Maps to entrypoints/openai/
+│       ├── api_server.py           │       ├── test_api_server.py     # E2E: API server (online serving)
+│       └── serving_chat.py         │       └── test_serving_chat.py
+│
+├── inputs/                    →    ├── inputs/
+│   ├── data.py                     │   ├── test_data.py
+│   ├── parse.py                    │   ├── test_parse.py
+│   └── preprocess.py               │   └── test_preprocess.py
+│
+├── model_executor/            →    ├── model_executor/
+│   ├── layers/                     │   ├── layers/
+│   │   └── mrope.py                │   │   └── test_mrope.py
+│   ├── model_loader/               │   ├── model_loader/
+│   │   └── weight_utils.py         │   │   └── test_weight_utils.py
+│   ├── models/                     │   ├── models/
+│   │   ├── qwen2_5_omni/           │   │   ├── qwen2_5_omni/
+│   │   │   ├── qwen2_5_omni_thinker.py │ │   │   ├── test_qwen2_5_omni_thinker.py  # UT
+│   │   │   ├── qwen2_5_omni_talker.py │ │   │   ├── test_qwen2_5_omni_talker.py  # UT
+│   │   │   └── qwen2_5_omni_token2wav.py │ │   │   └── test_qwen2_5_omni_token2wav.py  # UT
+│   │   └── qwen3_omni/             │   │   └── qwen3_omni/
+│   │       └── ...                 │   │       └── test_*.py
+│   ├── stage_configs/              │   └── stage_configs/             # Configuration tests (if needed)
+│   │   └── ...                     │       └── test_*.py
+│   └── stage_input_processors/     │   └── stage_input_processors/
+│       └── ...                     │       └── test_*.py
+│
+├── sample/                    →    ├── sample/
+│   └── ...                         │   └── test_*.py
+│
+├── utils/                     →    ├── utils/
+│   └── platform_utils.py           │   └── test_platform_utils.py
+│
+├── worker/                    →    ├── worker/
+    ├── gpu_ar_worker.py            │   ├── test_gpu_ar_worker.py
+    ├── gpu_generation_worker.py    │   ├── test_gpu_generation_worker.py
+    ├── gpu_model_runner.py         │   ├── test_gpu_model_runner.py
+    └── npu/                        │   └── npu/                       # Maps to worker/npu/
+        └── ...                     │       └── test_*.py
+│
+└── e2e/                       →    ├── e2e/                # End-to-end scenarios (no 1:1 source mirror)
+                                    ├── online_serving/       # Full-stack online serving flows
+                                    │   └── (empty for now)
+                                    └── offline_inference/    # Full offline inference flows
+                                        ├── test_qwen2_5_omni.py     # Moved from multi_stages/
+                                        ├── test_qwen3_omni.py       # Moved from multi_stages_h100/
+                                        ├── test_t2i_model.py  # Moved from single_stage/
+                                        └── stage_configs/           # Shared stage configs
+                                            ├── qwen2_5_omni_ci.yaml
+                                            └── qwen3_omni_ci.yaml
+```
+
+
+
+### Naming Conventions
+
+- **Unit Tests**: Use `test_<module_name>.py` format. Example: `omni_llm.py` → `test_omni_llm.py`
+
+- **E2E Tests**: Place in `tests/e2e/offline_inference/` or `tests/e2e/online_serving/` with descriptive names. Example: `tests/e2e/offline_inference/test_qwen3_omni.py`, `tests/e2e/offline_inference/test_diffusion_model.py`
+
+### Best Practices
+
+1. **Mirror Source Structure**: Test directories should mirror the source code structure
+2. **Test Type Indicators**: Use comments to indicate test types (UT for unit tests, E2E for end-to-end tests)
+3. **Shared Resources**: Place shared test configurations (e.g., CI configs) in appropriate subdirectories
+4. **Consistent Naming**: Follow the `test_*.py` naming convention consistently across all test files
+
+
+## Test codes requirements
+
+### Coding style
+
+1. **File header**: Add SPDX license header to all test files
+2. **Imports**: Pls don't use manual `sys.path` modifications, use standard imports instead.
+3. **Test type differentiation**:
+
+      - Unit tests: Maintain mock style
+      - E2E tests for models: Consider using OmniRunner uniformly, avoid decorators
+
+4. **Documentation**: Add docstrings to all test functions
+5. **Environment variables**: Set uniformly in `conftest.py` or at the top of files
+6. **Type annotations**: Add type annotations to all test function parameters
+7. **Pytest Markers**: Add necessary markers like `@pytest.mark.core_model` and use `@hardware_test` to declare hardware requirements (check detailed in [Markers for Tests](../ci/tests_markers.md)).
+
+### Template
+#### E2E - Online serving
+
+```python
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Online E2E smoke test for an omni model (video,text,audio → audio).
+"""
+from pathlib import Path
+
+import pytest
+import openai
+
+from tests.utils import hardware_test
+
+# Optional: set process start method for workers
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+models = ["{your model name}"] #Edit here to load your model
+stage_configs = [str(Path(__file__).parent / "stage_configs" / {your model yaml})] #Edit here to load your model yaml
+test_params = [(model, stage_config) for model in models for stage_config in stage_configs]
+
+#OmniServer，Used to start the vllm-omni server
+class OmniServer:
+    xxx
+
+
+@pytest.fixture
+def omni_server(request):
+    model, stage_config_path = request.param
+    with OmniServer(model, ["--stage-configs-path", stage_config_path]) as server:
+        yield server
+
+
+#handle request message
+@pytest.fixture(scope="session")
+def base64_encoded_video() -> str:
+    xxx
+
+@pytest.fixture(scope="session")
+def dummy_messages_from_video_data(video_data_url: str, content_text: str) -> str:
+    xxx
+
+@pytest.mark.core_model
+@pytest.mark.omni
+@hardware_test(
+    res={"cuda": "L4", "rocm": "MI325", "npu": "A2"},
+    num_cards={"cuda": 2, "rocm": 2, "npu": 4},
+)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_video_to_audio(
+    client: openai.OpenAI,
+    omni_server,
+    base64_encoded_video: str,
+) -> None:
+    #set message
+    video_data_url = f"data:video/mp4;base64, {base64_encoded_video}"
+    messages = dummy_messages_from_video_data(video_data_url)
+
+    #send request
+    chat_completion = client.chat.completions.create(
+        model=omni_server.model,
+        messages=messages,
+    )
+
+    #verify text output
+    text_choice = chat_completion.choices[0]
+    assert text_choice.finish_reason == "length"
+
+    #verify audio output
+    audio_choice = chat_completion.choices[1]
+    audio_message = audio_choice.message
+    if hasattr(audio_message, "audio") and audio_message.audio:
+        assert audio_message.audio.data is not None
+        assert len(audio_message.audio.data) > 0
+```
+
+#### E2E - Offline inference
+```python
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Offline E2E smoke test for an omni model (video → audio).
+"""
+
+import os
+from pathlib import Path
+
+import pytest
+from vllm.assets.video import VideoAsset
+
+from tests.utils import hardware_test
+from ..multi_stages.conftest import OmniRunner
+
+# Optional: set process start method for workers
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+models = ["{your model name}"] #Edit here to load your model
+stage_configs = [str(Path(__file__).parent / "stage_configs" / {your model yaml})] #Edit here to load your model yaml
+
+# Create parameter combinations for model and stage config
+test_params = [(model, stage_config) for model in models for stage_config in stage_configs]
+
+# function name: test_{input_modality}_to_{output_modality}
+# modality candidate: text, image, audio, video, mixed_modalities
+@pytest.mark.core_model
+@pytest.mark.omni
+@hardware_test(
+    res={"cuda": "L4", "rocm": "MI325", "npu": "A2"},
+    num_cards=2,
+)
+@pytest.mark.parametrize("test_config", test_params)
+def test_video_to_audio(omni_runner: type[OmniRunner], model: str) -> None:
+    """Offline inference: video input, audio output."""
+    model, stage_config_path = test_config
+    with omni_runner(model, seed=42, stage_configs_path=stage_config_path) as runner:
+        # Prepare inputs
+        video = VideoAsset(name="sample", num_frames=4).np_ndarrays
+
+        outputs = runner.generate_multimodal(
+            prompts="Describe this video briefly.",
+            videos=video,
+        )
+
+        # Minimal assertions: got outputs and at least one audio result
+        assert outputs
+        has_audio = any(o.final_output_type == "audio" for o in outputs)
+        assert has_audio
+```
+
+## Checklist before submitting your test files
+
+1. The file is saved in an appropriate place and the file name is clear.
+2. The coding style follows the requirements outlined above.
+3. **All test functions have appropriate pytest markers**
+4. For tests that need run in CI, please ensure the test is configured under the `./buildkite/` folder.
@@ -1,2 +1,15 @@
 # Adding a New Model
-Stay tuned for guidance on how to add a new model on vLLM-Omni!
+
+This section provides comprehensive guidance on how to add a new model to vLLM-Omni.
+
+## Documentation
+
+- **[Adding an Omni-Modality Model](adding_omni_model.md)**: Complete step-by-step guide using Qwen3-Omni as an example.
+
+- **[Adding a Diffusion Model](adding_diffusion_model.md)**: Complete step-by-step guide using Qwen/Qwen-Image-Edit as an example.
+
+
+
+## Quick Start
+
+For a quick reference, see the [Adding a New Multi-Stage Model Guide](adding_omni_model.md) and [Adding a New Diffusion Model Guide](adding_diffusion_model.md).
@@ -0,0 +1,227 @@
+# Adding a Diffusion Model
+This guide walks through the process of adding a new Diffusion model to vLLM-Omni, using Qwen/Qwen-Image-Edit as a comprehensive example.
+
+# Table of Contents
+1. [Overview](#overview)
+2. [Directory Structure](#directory-structure)
+3. [Step-by-Step Implementation](#step-by-step-implementation)
+4. [Testing](#testing)
+5. [Adding a Model Recipe](#adding-a-model-recipe)
+
+
+# Overview
+When add a new diffusion model into vLLM-Omni, additional adaptation work is required due to the following reasons:
+
+ New model must follow the framework’s parameter passing mechanisms and inference flow.
+
+ Replacing the model’s default implementations with optimized modules, which is necessary to achieve the better performance.
+
+The diffusion execution flow as follow:
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-diffusion-flow.png">
+    <img alt="Diffusion Flow" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-diffusion-flow.png" width=55%>
+  </picture>
+</p>
+
+
+# Directory Structure
+File Structure for Adding a New Diffusion Model
+
+```
+vllm_omni/
+└── examples/
+    └──offline_inference
+        └── example script                # reuse existing if possible (e.g., image_edit.py)
+    └──online_serving
+        └── example script
+└── diffusion/
+    └── registry.py                       # Registry work
+    ├── request.py                        # Request Info
+    └── models/your_model_name/           # Model directory (e.g., qwen_image)
+        └── pipeline_xxx.py               # Model implementation (e.g., pipeline_qwen_image_edit.py)
+```
+
+# Step-by-step-implementation
+## Step 1: Model Implementation
+The diffusion pipeline’s implementation follows **HuggingFace Diffusers**.
+### 1.1 Define the Pipeline Class
+Define the pipeline class, e.g., `QwenImageEditPipeline`, and initialize all required submodules, either from HuggingFace `diffusers` or custom implementations. In `QwenImageEditPipeline`, only `QwenImageTransformer2DModel` is re-implemented to support optimizations such as Ulysses-SP. When adding new models in the future, you can either reuse this re-implemented `QwenImageTransformer2DModel` or extend it as needed.
+
+### 1.2 Pre-Processing and Post-Processing Extraction
+Extract the pre-processing and post-processing logic from the pipeline class to follow vLLM-Omni’s execution flow. For Qwen-Image-Edit:
+```python
+def get_qwen_image_edit_pre_process_func(
+    od_config: OmniDiffusionConfig,
+):
+    """
+    Define a pre-processing function that resizes input images and
+    pre-process for subsequent inference.
+    """
+```
+
+```python
+def get_qwen_image_edit_post_process_func(
+    od_config: OmniDiffusionConfig,
+):
+    """
+    Defines a post-processing function that post-process images.
+    """
+```
+
+### 1.3 Define the forward function
+The forward function of `QwenImageEditPipeline` follows the HuggingFace `diffusers` design for the most part. The key differences are:
+ As described in the overview, arguments are passed through `OnniDiffusionRequest`, so we need to get user parameters from it accordingly.
+```python
+prompt = req.prompt
+```
+ pre/post-processing are handled by the framework elsewhere, so skip them.
+
+### 1.4 Replace some ops or layers in DiT component
+
+vLLM-Omni provides a set of optimized operators with better performance and built-in support for parallelism, including attention, rotary embeddings (RoPE), and linear layers.
+
+Below is an example showing how to replace standard Transformer attention and FFN layers with vLLM-Omni implementations:
+
+```python
+from vllm_omni.diffusion.attention.layer import Attention
+from vllm_omni.diffusion.layers.rope import RotaryEmbedding
+
+class MyAttention(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.attn = Attention()
+        self.to_qkv = QKVParallelLinear()
+        self.to_out = RowParallelLinear()
+        self.rope = RotaryEmbedding(is_neox_style=False)
+
+    def forward(self, hidden_states):
+        qkv, _ = self.to_qkv(hidden_states)
+        q, k, v = qkv.split(...)
+        q, k = self.rope(...)
+        attn_output = self.attn(q, k, v)
+        output = self.to_out(attn_output)
+
+class MyFFN(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.fc1 = ColumnParallelLinear()
+        self.fc2 = RowParallelLinear()
+        self.act = F.gelu
+
+    def forward(self, hidden_states):
+        hidden, _ = self.fc1(hidden_states)
+        hidden = self.act(hidden)
+        output = self.fc2(hidden)
+        return output
+```
+
+In this example:
+
+ Attention uses vLLM-Omni’s optimized attention kernel together with parallel QKV projection and RoPE.
+
+ Linear layers are replaced with column- and row-parallel variants to enable tensor parallelism.
+
+ The FFN follows a standard two-layer structure and can be further optimized (e.g., using fused or merged projections) if needed.
+
+
+### 1.5 Provide a `_repeated_blocks` in DiT model
+`_repeated_blocks` is the small and frequently-repeated block(s) of a model -- typically a transformer layer.
+
+It's used for torch compile optimizations.
+```python
+_repeated_blocks = ["QwenImageTransformerBlock"]
+```
+
+
+### 1.6 (Optional) implement sequence parallelism
+vLLM-Omni has a non-intrusive `_sp_plan` that enable sequence parallel without modifying `forward()` logic.
+You can refer to [How to parallelize a new model](../../user_guide/diffusion/parallelism_acceleration.md)
+
+
+### 1.7 (Optional) integrate with Cache-Dit
+vLLM-Omni supports acceleration via [Cache-Dit](../../user_guide/diffusion/cache_dit_acceleration.md). Most models compatible with Diffusers can use Cache-Dit seamlessly. For new models, you can extend support by modifying`cache_dit_backend.py`
+
+## Step 2: Extend OmniDiffusionRequest Fields
+User-provided inputs are ultimately passed to the model’s forward method through OmniDiffusionRequest, so we add the required fields here to support the new model.
+```python
+prompt: str | list[str] | None = None
+negative_prompt: str | list[str] | None = None
+...
+```
+
+## Step 3: Registry
+ registry diffusion model in registry.py
+```python
+_DIFFUSION_MODELS = {
+    # arch:(mod_folder, mod_relname, cls_name)
+    ...
+    "QwenImageEditPipeline": (
+        "qwen_image",
+        "pipeline_qwen_image_edit",
+        "QwenImageEditPipeline",
+    ),
+    ...
+}
+```
+ registry pre-process get function
+```python
+_DIFFUSION_PRE_PROCESS_FUNCS = {
+    # arch: pre_process_func
+    ...
+    "QwenImageEditPipeline": "get_qwen_image_edit_pre_process_func",
+    ...
+}
+```
+
+ registry post-process get function
+```python
+_DIFFUSION_POST_PROCESS_FUNCS = {
+    # arch: post_process_func
+    ...
+    "QwenImageEditPipeline": "get_qwen_image_edit_post_process_func",
+    ...
+}
+```
+
+## Step 4: Add an Example Script
+For each newly integrated model, we need to provide examples script under the examples/ to demonstrate how to initialize the pipeline with Omni, pass in user inputs, and generate outputs.
+Key point for writing the example:
+
+ Use the Omni entrypoint to load the model and construct the pipeline.
+
+ Show how to format user inputs and pass them via omni.generate(...).
+
+ Demonstrate the common runtime arguments, such as:
+
+    + model path or model name
+
+    + input image(s) or prompt text
+
+    + key diffusion parameters (e.g., inference steps, guidance scale)
+
+    + optional acceleration backends (e.g., Cache-DiT, TeaCache)
+
+ Save or display the generated results so users can validate the integration.
+
+## Step 5: Open a Pull Request
+
+When submitting a pull request to add support for a new model, please include the following information in the PR description:
+
+ Output verification: provide generation outputs to verify correctness and model behavior.
+
+ Inference speed: provide a comparison with the corresponding implementation in Diffusers.
+
+ Parallelism support: specify the supported parallel sizes and any relevant limitations.
+
+ Cache acceleration: check whether the model can be accelerated using Cache-Dit or not.
+
+
+Providing these details helps reviewers evaluate correctness, performance improvements, and parallel scalability of the new model integration.
+
+# Testing
+For comprehensive testing guidelines, please refer to the [Test File Structure and Style Guide](../ci/tests_style.md).
+
+
+## Adding a Model Recipe
+After implementing and testing your model, please add a model recipe to the [vllm-project/recipes](https://github.com/vllm-project/recipes) repository. This helps other users understand how to use your model with vLLM-Omni.
@@ -0,0 +1,633 @@
+# Adding an Omni-Modality Model
+
+This guide walks through the process of adding a new multi-stage model to vLLM-Omni, using **Qwen3-Omni** as a comprehensive example. Qwen3-Omni is a multi-stage omni-modality model that demonstrates the full capabilities of vLLM-Omni's architecture.
+
+## Table of Contents
+
+1. [Overview](#overview)
+2. [Directory Structure](#directory-structure)
+3. [Step-by-Step Implementation](#step-by-step-implementation)
+4. [Key Components](#key-components)
+5. [Model Registration](#model-registration)
+6. [Stage Configuration](#stage-configuration)
+7. [Stage Input Processors](#stage-input-processors)
+8. [Testing](#testing)
+9. [Adding a Model Recipe](#adding-a-model-recipe)
+10. [Summary](#summary)
+
+## Overview
+
+vLLM-Omni supports multi-stage model architectures where different stages can run on different devices and process different modalities. The Qwen3-Omni model exemplifies this with three stages:
+
+1. **Thinker Stage**: Multimodal understanding (text + audio + video) → text generation
+2. **Talker Stage**: Text embeddings → RVQ codec codes
+3. **Code2Wav Stage**: RVQ codes → audio waveform
+
+Each stage is implemented as a separate model class that can be configured independently.
+
+## Directory Structure
+
+When adding a new model, you'll need to create the following structure:
+
+```
+vllm_omni/model_executor/models/
+└── your_model_name/              # Model directory (e.g., qwen3_omni)
+    ├── __init__.py               # Exports main model class
+    ├── your_model.py             # Main unified model class
+    ├── your_model_stage1_implementation.py      # Stage 1 implementation (e.g., thinker)
+    ├── your_model_stage2_implementation.py      # Stage 2 implementation (e.g., talker)
+    └── your_model_stage3_implementation.py      # Stage 3 implementation (e.g., code2wav)
+    └── ... maybe other stage implementations
+
+vllm_omni/model_executor/stage_input_processors/
+└── your_model_name.py            # Stage transition processors
+
+vllm_omni/model_executor/stage_configs/
+└── your_model_name.yaml          # Stage configuration file
+```
+
+## Step-by-Step Implementation
+
+### Step 1: Create the Model Directory
+
+Create a new directory under `vllm_omni/model_executor/models/`
+
+### Step 2: Implement Stage Components
+
+For Qwen3-Omni, we have three stage components:
+
+#### 2.1 Thinker Stage (`qwen3_omni_moe_thinker.py`)
+
+The thinker stage handles multimodal understanding. Key features:
+
+- Inherits from base Qwen3 MoE model in vLLM, using vLLM fused ops & page attn to accelerate
+- Implements multimodal processing interfaces
+- Handles audio, video, and image inputs
+- Generates text outputs
+
+```python
+from vllm.model_executor.models.interfaces import SupportsMultiModal, SupportsPP
+from vllm.model_executor.models.qwen3_moe import Qwen3MoeForCausalLM
+
+class Qwen3OmniMoeThinkerForConditionalGeneration(
+    Qwen3MoeForCausalLM,
+    SupportsMultiModal,
+    SupportsPP
+):
+    """Thinker stage: multimodal understanding → text generation."""
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        # Initialize base model
+        # Set up multimodal processors
+        # Configure audio/video/image encoders
+        pass
+```
+
+#### 2.2 Talker Stage (`qwen3_omni_moe_talker.py`)
+
+The talker stage converts text embeddings to codec codes:
+
+```python
+class Qwen3OmniMoeTalkerForConditionalGeneration(
+    Qwen3MoeForCausalLM,
+    SupportsPP
+):
+    """Talker stage: text embeddings → RVQ codec codes."""
+
+    def __init__(self, vllm_config, talker_config, prefix):
+        # Initialize base model
+        # Replace LM head with codec head
+        # Set up text projection from thinker
+        pass
+```
+
+#### 2.3 Code2Wav Stage (`qwen3_omni_code2wav.py`)
+
+The code2wav stage generates audio waveforms:
+
+```python
+class Qwen3OmniMoeCode2Wav(nn.Module):
+    """Code2Wav stage: RVQ codes → audio waveform."""
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        # Initialize audio decoder
+        # Set up codec processing
+        pass
+```
+
+### Step 3: Implement the Unified Model Class
+
+The main model class (`qwen3_omni.py`) orchestrates all stages:
+
+```python
+@MULTIMODAL_REGISTRY.register_processor(
+    Qwen3OmniMoeThinkerMultiModalProcessor,
+    info=Qwen3OmniMoeThinkerProcessingInfo,
+    dummy_inputs=Qwen3OmniMoeThinkerDummyInputsBuilder,
+)
+class Qwen3OmniMoeForConditionalGeneration(
+    nn.Module, SupportsMultiModal, SupportsPP, Qwen3OmniMoeConditionalGenerationMixin
+):
+    """
+    Unified Qwen3 Omni MoE model combining thinker, talker, and code2wav.
+
+    Architecture:
+    - Thinker: Multimodal understanding (text + audio + video) → text generation
+    - Talker: Text embeddings → RVQ codec codes
+    - Code2Wav: RVQ codes → audio waveform
+
+    Usage:
+        Set `model_stage` in vllm_config to one of: "thinker", "talker", "code2wav"
+    """
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        self.have_multimodal_outputs = True
+        config: Qwen3OmniMoeConfig = vllm_config.model_config.hf_config
+
+        # Determine which stage to initialize
+        self.model_stage = vllm_config.model_config.model_stage
+
+        if self.model_stage == "thinker":
+            # Initialize thinker model
+            thinker_vllm_config = vllm_config.with_hf_config(
+                config.thinker_config,
+                architectures=["Qwen3OmniMoeThinkerForConditionalGeneration"]
+            )
+            self.thinker = init_vllm_registered_model(
+                vllm_config=thinker_vllm_config,
+                prefix=maybe_prefix(prefix, "thinker"),
+                hf_config=config.thinker_config,
+                architectures=["Qwen3OmniMoeThinkerForConditionalGeneration"],
+            )
+            self.model = self.thinker
+
+        elif self.model_stage == "talker":
+            # Initialize talker model
+            talker_vllm_config = vllm_config.with_hf_config(
+                config.talker_config,
+                architectures=["Qwen3OmniMoeTalkerForConditionalGeneration"]
+            )
+            self.talker = init_vllm_registered_model(
+                vllm_config=talker_vllm_config,
+                prefix=maybe_prefix(prefix, "talker"),
+                hf_config=config.talker_config,
+                architectures=["Qwen3OmniMoeTalkerForConditionalGeneration"],
+            )
+            self.model = self.talker
+
+        elif self.model_stage == "code2wav":
+            # Initialize code2wav model
+            code2wav_vllm_config = vllm_config.with_hf_config(
+                config.code2wav_config,
+                architectures=["Qwen3OmniMoeCode2Wav"]
+            )
+            self.code2wav = init_vllm_registered_model(
+                vllm_config=code2wav_vllm_config,
+                prefix=maybe_prefix(prefix, "code2wav"),
+                hf_config=config.code2wav_config,
+                architectures=["Qwen3OmniMoeCode2Wav"],
+            )
+            self.model = self.code2wav
+        else:
+            raise ValueError(
+                f"Invalid model_stage: {self.model_stage}. "
+                f"Must be one of: 'thinker', 'talker', 'code2wav'"
+            )
+```
+
+#### Key Methods to Implement
+
+1. **`forward()`**: Handles the forward pass for each stage
+2. **`embed_input_ids()`**: Embeds input token IDs
+3. **`embed_multimodal()`**: Processes multimodal inputs (if applicable)
+4. **`compute_logits()`**: Computes logits from hidden states
+5. **`load_weights()`**: Loads model weights with proper prefixing of different stages
+
+### Step 4: Create `__init__.py`
+
+Export the main model class:
+
+```python
+# vllm_omni/model_executor/models/qwen3_omni/__init__.py
+from .qwen3_omni import Qwen3OmniMoeForConditionalGeneration
+
+__all__ = ["Qwen3OmniMoeForConditionalGeneration"]
+```
+
+## Key Components
+
+### 1. Model Interfaces
+
+Your model should implement the appropriate interfaces:
+
+- **`SupportsMultiModal`**: For models that process multimodal inputs
+- **`SupportsPP`**: For models that support pipeline parallelism
+- **`SupportsMRoPE`**: For models using multi-dimensional RoPE (if applicable)
+
+### 2. Multimodal Registration
+
+If your model processes multimodal inputs, register it with the multimodal registry:
+
+```python
+@MULTIMODAL_REGISTRY.register_processor(
+    YourMultiModalProcessor,
+    info=YourProcessingInfo,
+    dummy_inputs=YourDummyInputsBuilder,
+)
+class YourModel(nn.Module, SupportsMultiModal):
+    pass
+```
+
+### 3. Weight Loading
+
+Implement `load_weights()` to handle weight loading with proper prefixing:
+
+```python
+def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+    """Load weights for all components of the omni model."""
+    loaded_weights = set()
+    thinker_weights = []
+    talker_weights = []
+    code2wav_weights = []
+
+    # Separate weights by component
+    for k, v in weights:
+        if k.startswith("thinker."):
+            thinker_weights.append((k, v))
+        elif k.startswith("talker."):
+            talker_weights.append((k, v))
+        elif k.startswith("code2wav."):
+            code2wav_weights.append((k, v))
+
+    # Load each component's weights
+    if self.thinker and thinker_weights:
+        thinker_loaded = self.thinker.load_weights(thinker_weights)
+        thinker_loaded = add_prefix_to_loaded_weights(thinker_loaded, "thinker")
+        loaded_weights.update(thinker_loaded)
+
+    # Similar for talker and code2wav...
+
+    return loaded_weights
+```
+
+### 4. Output Format
+
+Use `OmniOutput` for stage outputs:
+
+```python
+from vllm_omni.model_executor.models.output_templates import OmniOutput
+
+# In forward method
+return OmniOutput(
+    text_hidden_states=hidden_states,
+    multimodal_outputs={"additional_data": data},
+    next_token_id=next_token_id,
+)
+```
+
+## Model Registration
+
+Register your model in `vllm_omni/model_executor/models/registry.py`:
+
+```python
+_OMNI_MODELS = {
+    # ... existing models ...
+
+    # Your new model
+    "YourModelForConditionalGeneration": (
+        "your_model_name",        # Module folder name
+        "your_model",             # Module file name (without .py)
+        "YourModelForConditionalGeneration",  # Class name
+    ),
+    "YourModelThinkerForConditionalGeneration": (
+        "your_model_name",
+        "your_model_thinker",
+        "YourModelThinkerForConditionalGeneration",
+    ),
+    # ... other stages ...
+}
+```
+
+The registry uses lazy loading, so the model class is imported only when needed.
+
+## Stage Configuration
+
+Create a YAML configuration file in `vllm_omni/model_executor/stage_configs/`. For a complete example, see the [Qwen3-Omni configuration file](gh-file:vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml).
+
+### Key Configuration Fields
+
+- **`model_stage`**: Which stage to run ("thinker", "talker", "code2wav", etc.)
+- **`model_arch`**: The model architecture name (must match registry)
+- **`engine_input_source`**: List of stage IDs that provide input to this stage
+- **`custom_process_input_func`**: Function to process inputs from previous stages
+- **`final_output`**: Whether this stage produces the final output (True/False)
+- **`final_output_type`**: Type of final output ("text", "audio", "image", etc.)
+
+## Stage Input Processors
+
+Stage transitions are the mechanism by which outputs from one stage are converted into inputs for the next stage. This section explains where and how stage transitions occur.
+
+### Where Stage Transitions Are Called
+
+Stage transitions happen automatically in the orchestrator (`OmniLLM` class) during the generation loop. Here's the detailed flow:
+
+1. **Location**: `vllm_omni/entrypoints/omni_llm.py` in the `_run_generation()` method
+2. **Trigger**: When a stage completes processing and produces outputs
+3. **Execution Flow**:
+   ```python
+   # In omni_llm.py, _run_generation() method (around line 345-460)
+
+   # Main orchestrator loop polls each stage for completed requests
+   for stage_id, stage in enumerate(self.stage_list):
+       result = stage.try_collect()  # Get completed request
+       if result is None:
+           continue
+
+       # Store outputs from this stage
+       engine_outputs = _load(result, obj_key="engine_outputs", shm_key="engine_outputs_shm")
+       stage.set_engine_outputs(engine_outputs)
+
+       # Check if there's a next stage to forward to
+       next_stage_id = stage_id + 1
+       if next_stage_id < len(self.stage_list):
+           next_stage: OmniStage = self.stage_list[next_stage_id]
+
+           # THIS IS WHERE STAGE TRANSITION HAPPENS
+           next_inputs = next_stage.process_engine_inputs(
+               self.stage_list,
+               [request_id_to_prompt[req_id]]
+           )
+
+           # Submit to next stage
+           task = {
+               "type": OmniStageTaskType.GENERATE,
+               "request_id": req_id,
+               "engine_inputs": next_inputs[0],
+               "sampling_params": sampling_params_list[next_stage_id],
+           }
+           next_stage.submit(task)
+   ```
+
+### How Stage Transitions Work
+
+The stage transition process follows these steps:
+
+1. **Stage Completion**: When a stage finishes processing a request, it stores outputs via `stage.set_engine_outputs(engine_outputs)`
+
+2. **Transition Detection**: The orchestrator checks if there's a next stage and calls `process_engine_inputs()` on it
+
+3. **Input Processing**: The `process_engine_inputs()` method in `OmniStage` (`omni_stage.py`) handles the transition:
+   ```python
+   def process_engine_inputs(
+       self, stage_list: list[Any], prompt: OmniTokensPrompt | TextPrompt = None
+   ) -> list[OmniTokensPrompt | TextPrompt]:
+       """Process engine inputs for this stage from upstream stage outputs."""
+
+       if self.custom_process_input_func is None:
+           # Default behavior: pass token IDs directly
+           # Extract outputs from source stage
+           source_stage_id = self.engine_input_source[0]
+           source_outputs = stage_list[source_stage_id].engine_outputs
+           # ... create OmniTokensPrompt from token_ids ...
+       else:
+           # Custom transition function (YOUR CODE HERE)
+           return self.custom_process_input_func(
+               stage_list,
+               self.engine_input_source,
+               prompt,
+               self.requires_multimodal_data
+           )
+   ```
+   - If `custom_process_input_func` is configured, it calls that function
+   - Otherwise, it uses default behavior (passing token IDs directly)
+
+4. **Custom Function Execution**: Your custom function receives:
+   - `stage_list`: List of all stage objects (to access upstream stage outputs)
+   - `engine_input_source`: List of source stage IDs (e.g., `[0]` for stage 0)
+   - `prompt`: Original prompt data (for preserving multimodal data)
+   - `requires_multimodal_data`: Whether multimodal data is required
+
+5. **Output Format**: The function must return a list of `OmniTokensPrompt` objects ready for the next stage
+
+### Data Structures in Stage Transitions
+
+Understanding the data structures is crucial for implementing stage transitions:
+
+**Input to your function:**
+- `stage_list[source_stage_id].engine_outputs`: List of `EngineCoreOutput` objects
+  - Each contains `outputs`: List of `RequestOutput` objects
+  - Each `RequestOutput` has:
+    - `token_ids`: Generated token IDs
+    - `multimodal_output`: Dict with keys like `"code_predictor_codes"`, etc.
+      - These are the hidden states or intermediate outputs from the model's forward pass
+    - `prompt_token_ids`: Original prompt token IDs
+
+**Output from your function:**
+- Must return `list[OmniTokensPrompt]` where each `OmniTokensPrompt` contains:
+  - `prompt_token_ids`: List[int] - Token IDs for the next stage
+  - `additional_information`: Dict[str, Any] - Optional metadata (e.g., embeddings, hidden states)
+  - `multi_modal_data`: Optional multimodal data if needed
+
+### How Model Outputs Are Stored
+
+The model's `forward()` method returns an `OmniOutput` object that contains:
+- `text_hidden_states`: Final hidden states for text generation
+- `multimodal_outputs`: Dict containing intermediate outputs
+
+These outputs are captured during the forward pass and stored in `multimodal_output` with specific keys:
+
+```python
+# In your model's forward() method (e.g., qwen3_omni.py)
+def forward(self, ...):
+    # ... processing ...
+
+    # For thinker stage: capture embeddings and hidden states
+    multimodal_outputs = {
+        "0": captured_embeddings,      # Layer 0 embeddings
+        "24": captured_hidden_states,  # Layer 24 hidden states
+        "tts_bos_embed": tts_bos_embed,
+        "tts_eos_embed": tts_eos_embed,
+        # ... other intermediate outputs ...
+    }
+
+    return OmniOutput(
+        text_hidden_states=hidden_states,
+        multimodal_outputs=multimodal_outputs,
+    )
+```
+
+These keys are then accessible in your stage transition function:
+```python
+# In stage_input_processors/qwen3_omni.py
+thinker_embeddings = output.multimodal_output["0"]  # Access by key
+thinker_hidden_states = output.multimodal_output["24"]
+```
+
+### Key Points
+
+1. **Accessing Upstream Outputs**: Use `stage_list[source_stage_id].engine_outputs` to get outputs from the source stage
+2. **Extracting Data**: Access `output.multimodal_output[key]` to get specific hidden states or intermediate results
+   - Keys are defined by your model's `forward()` method when it creates `multimodal_outputs`
+3. **Device Management**: Move tensors to appropriate devices (CPU for serialization, GPU for processing)
+4. **Shape Transformations**: Reshape tensors as needed for the next stage (e.g., flattening codec codes)
+5. **Batch Handling**: Process each request in the batch separately and return a list
+
+### Complete Flow Diagram
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-dataflow-between-stages.png">
+    <img alt="Data Flow between stages" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-dataflow-between-stages.png" width=55%>
+  </picture>
+</p>
+
+### Implementation Example
+
+Create stage transition processors in `vllm_omni/model_executor/stage_input_processors/your_model_name.py`:
+
+```python
+# qwen3_omni.py
+
+def thinker2talker(
+    stage_list: list[Any],
+    engine_input_source: list[int],
+    prompt: OmniTokensPrompt | TextPrompt | None = None,
+    requires_multimodal_data: bool = False,
+) -> list[OmniTokensPrompt]:
+    """
+    Process thinker outputs to create talker inputs.
+
+    Args:
+        stage_list: List of stage objects
+        engine_input_source: Source stage IDs (typically [0] for thinker)
+        prompt: Original prompt data
+
+    Returns:
+        List of OmniTokensPrompt for talker stage
+    """
+    source_stage_id = engine_input_source[0]
+    thinker_outputs = stage_list[source_stage_id].engine_outputs
+    talker_inputs = []
+
+    for thinker_output in thinker_outputs:
+        output = thinker_output.outputs[0]
+        # Extract thinker embeddings and hidden states
+        thinker_embeddings = output.multimodal_output["0"].float().clone().detach().cuda()
+        thinker_hidden_states = output.multimodal_output["24"].float().clone().detach().cuda()
+
+        info = {
+            "thinker_embeddings": thinker_embeddings,
+            "thinker_hidden_states": thinker_hidden_states,
+            "thinker_sequences": thinker_output.prompt_token_ids + output.token_ids,
+            "thinker_input_ids": thinker_output.prompt_token_ids,
+        }
+
+        talker_inputs.append(
+            OmniTokensPrompt(
+                prompt_token_ids=[0] * computed_length,
+                additional_information=info,
+                multi_modal_data=None,
+            )
+        )
+
+    return talker_inputs
+
+
+def talker2code2wav(
+    stage_list: list[Any],
+    engine_input_source: list[int],
+    prompt: OmniTokensPrompt | TextPrompt | None = None,
+    requires_multimodal_data: bool = False,
+) -> list[OmniTokensPrompt]:
+    """
+    Process talker outputs to create code2wav inputs.
+    """
+    source_stage_id = engine_input_source[0]
+    talker_outputs = stage_list[source_stage_id].engine_outputs
+    code2wav_inputs = []
+
+    for talker_output in talker_outputs:
+        output = talker_output.outputs[0]
+        # Extract codec codes
+        codec_codes = (
+            output.multimodal_output["code_predictor_codes"]
+            .to(torch.long)
+            .transpose(0, 1)
+            .cpu()
+            .to(torch.long)
+            .reshape(-1)
+            .tolist()
+        )
+
+        code2wav_inputs.append(
+            OmniTokensPrompt(
+                prompt_token_ids=codec_codes,
+                multi_modal_data=None,
+            )
+        )
+
+    return code2wav_inputs
+```
+
+## Testing
+
+For comprehensive testing guidelines, please refer to the [Test File Structure and Style Guide](../ci/tests_style.md).
+
+## Adding a Model Recipe
+
+After implementing and testing your model, please add a model recipe to the [vllm-project/recipes](https://github.com/vllm-project/recipes) repository. This helps other users understand how to use your model with vLLM-Omni.
+
+### What to Include
+
+Your recipe should include:
+
+1. **Model Overview**: Brief description of the model and its capabilities
+2. **Installation Instructions**: Step-by-step setup instructions including:
+   - Installing vllm-omni and dependencies
+   - Installing any additional required packages (e.g., xformers, diffusers)
+   - Any version requirements
+3. **Usage Examples**: Command-line examples demonstrating how to run the model
+4. **Configuration Details**: Important configuration parameters and their meanings
+
+### Example
+
+For reference, see the [LongCat recipe example](https://github.com/vllm-project/recipes/pull/179) which demonstrates the expected format and structure.
+
+### Recipe Location
+
+Create your recipe file in the appropriate directory structure:
+- For organization-specific models: `OrganizationName/ModelName.md`
+- For general models: `ModelName.md`
+
+The recipe should be a Markdown file that provides clear, reproducible instructions for users to get started with your model.
+
+## Summary
+
+Adding a new model to vLLM-Omni involves:
+
+1. **Create model directory structure** with stage implementations
+2. **Implement unified model class** that orchestrates stages
+3. **Register model** in `registry.py`
+4. **Create stage configuration** YAML file
+5. **Implement stage input processors** for stage transitions
+6. **Write tests** to verify functionality
+7. **Add model recipe** to the [vllm-project/recipes](https://github.com/vllm-project/recipes) repository (see [Adding a Model Recipe](#adding-a-model-recipe) section)
+
+### Qwen3-Omni Reference Files
+
+For a complete reference implementation, see:
+
+- **Main model**: `vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py`
+- **Thinker**: `vllm_omni/model_executor/models/qwen3_omni/qwen3_omni_moe_thinker.py`
+- **Talker**: `vllm_omni/model_executor/models/qwen3_omni/qwen3_omni_moe_talker.py`
+- **Code2Wav**: `vllm_omni/model_executor/models/qwen3_omni/qwen3_omni_code2wav.py`
+- **Stage config**: `vllm_omni/model_executor/stage_configs/qwen3_omni_moe.yaml`
+- **Input processors**: `vllm_omni/model_executor/stage_input_processors/qwen3_omni.py`
+- **Registry**: `vllm_omni/model_executor/models/registry.py`
+- **Testing**: `vllm_omni/tests/e2e/offline_inference/test_qwen3_omni.py`
+
+For more information, see:
+- [Architecture Overview](../../design/architecture_overview.md)
+- [Supported Models](../../models/supported_models.md)
+- [Stage Configuration Guide](../../configuration/stage_configs.md)
@@ -0,0 +1,149 @@
+# Profiling vLLM-Omni
+
+> **Warning:** Profiling incurs significant overhead. Use only for development and debugging, never in production.
+
+vLLM-Omni uses the PyTorch Profiler to analyze performance across both **multi-stage omni-modality models** and **diffusion models**.
+
+### 1. Set the Output Directory
+Before running any script, set this environment variable. The system detects this and automatically saves traces here.
+
+```bash
+export VLLM_TORCH_PROFILER_DIR=./profiles
+```
+
+### 2. Profiling Omni-Modality Models
+
+It is best to limit profiling to one iteration to keep trace files manageable.
+
+```bash
+export VLLM_PROFILER_MAX_ITERS=1
+```
+
+**Selective Stage Profiling**
+The profiler is default to function across all stages. But It is highly recommended to profile specific stages by passing the stages list, preventing from producing too large trace files:
+```python
+# Profile all stages
+omni_llm.start_profile()
+
+# Only profile Stage 1
+omni_llm.start_profile(stages=[1])
+```
+
+```python
+# Stage 0 (Thinker) and Stage 2 (Audio Decoder) for qwen omni
+omni_llm.start_profile(stages=[0, 2])
+```
+
+**Python Usage**: Wrap your generation logic with `start_profile()` and `stop_profile()`.
+
+```python
+from vllm_omni import omni_llm
+
+profiler_enabled = bool(os.getenv("VLLM_TORCH_PROFILER_DIR"))
+
+# 1. Start profiling if enabled
+if profiler_enabled:
+    omni_llm.start_profile(stages=[0])
+
+# Initialize generator
+omni_generator = omni_llm.generate(prompts, sampling_params_list, py_generator=args.py_generator)
+
+total_requests = len(prompts)
+processed_count = 0
+
+# Main Processing Loop
+for stage_outputs in omni_generator:
+
+    # ... [Output processing logic for text/audio would go here] ...
+
+    # Update count to track when to stop profiling
+    processed_count += len(stage_outputs.request_output)
+
+    # 2. Check if all requests are done to stop the profiler safely
+    if profiler_enabled and processed_count >= total_requests:
+        print(f"[Info] Processed {processed_count}/{total_requests}. Stopping profiler inside active loop...")
+
+        # Stop the profiler while workers are still active
+        omni_llm.stop_profile()
+
+        # Wait for traces to flush to disk
+        print("[Info] Waiting 30s for workers to write trace files to disk...")
+        time.sleep(30)
+        print("[Info] Trace export wait time finished.")
+
+omni_llm.close()
+```
+
+
+**Examples**:
+
+1. **Qwen2.5-Omni**:  [https://github.com/vllm-project/vllm-omni/blob/main/examples/offline_inference/qwen2_5_omni/end2end.py](https://github.com/vllm-project/vllm-omni/blob/main/examples/offline_inference/qwen2_5_omni/end2end.py)
+
+2. **Qwen3-Omni**:   [https://github.com/vllm-project/vllm-omni/blob/main/examples/offline_inference/qwen3_omni/end2end.py](https://github.com/vllm-project/vllm-omni/blob/main/examples/offline_inference/qwen3_omni/end2end.py)
+
+
+### 3. Profiling diffusion models
+
+Diffusion profiling is End-to-End, capturing encoding, denoising loops, and decoding.
+
+**CLI Usage:**
+```python
+
+python image_to_video.py \
+    --model Wan-AI/Wan2.2-I2V-A14B-Diffusers \
+    --image qwen-bear.png \
+    --prompt "A cat playing with yarn, smooth motion" \
+    \
+    # Minimize Spatial Dimensions (Optional but helpful):
+    #    Drastically reduces memory usage so the profiler doesn't
+    #    crash due to overhead, though for accurate performance
+    #    tuning you often want target resolutions.
+    --height 48 \
+    --width 64 \
+    \
+    # Minimize Temporal Dimension (Frames):
+    #    Video models process 3D tensors (Time, Height, Width).
+    #    Reducing frames to the absolute minimum (2) keeps the
+    #    tensor size small, ensuring the trace file doesn't become
+    #    multi-gigabytes in size.
+    --num_frames 2 \
+    \
+    # Minimize Iteration Loop (Steps):
+    #    This is the most critical setting for profiling.
+    #    Diffusion models run the same loop X times.
+    #    Profiling 2 steps gives you the exact same performance
+    #    data as 50 steps, but saves minutes of runtime and
+    #    prevents the trace viewer from freezing.
+    --num_inference_steps 2 \
+    \
+    --guidance_scale 5.0 \
+    --guidance_scale_high 6.0 \
+    --boundary_ratio 0.875 \
+    --flow_shift 12.0 \
+    --fps 16 \
+    --output i2v_output.mp4
+
+```
+
+**Examples**:
+
+1. **Qwen image edit**:  [https://github.com/vllm-project/vllm-omni/blob/main/examples/offline_inference/image_to_image/image_edit.py](https://github.com/vllm-project/vllm-omni/blob/main/examples/offline_inference/image_to_image/image_edit.py)
+
+2. **Wan-AI/Wan2.2-I2V-A14B-Diffusers**:   [https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/image_to_video](https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/image_to_video)
+
+> **Note:**
+As of now, asynchronous (online) profiling is not fully supported in vLLM-Omni. While start_profile() and stop_profile() methods exist, they are only reliable in offline inference scripts (e.g., the provided end2end.py examples). Do not use them in server-mode or streaming scenarios—traces may be incomplete or fail to flush.
+
+### 4. Analyzing Omni Traces
+
+Output files are saved to your configured ```VLLM_TORCH_PROFILER_DIR```.
+
+**Output**
+**Chrome Trace** (```.json.gz```): Visual timeline of kernels and stages. Open in Perfetto UI.
+
+**Viewing Tools:**
+
+- [Perfetto](https://ui.perfetto.dev/)(recommended)
+- ```chrome://tracing```(Chrome only)
+
+**Note**: vLLM-Omni reuses the PyTorch Profiler infrastructure from vLLM. See the official vLLM profiler documentation:  [vLLM Profiling Guide](https://docs.vllm.ai/en/stable/contributing/profiling/)
@@ -1,113 +1,195 @@
 # Architecture Overview

-# Introduction
+This document outlines the architectural design for vLLM-Omni.

-This document outlines the architectural design for vLLM-Omni, a new repository developed under the vLLM project. vLLM-Omni supports multi-modality models with non-autoregressive structures and non-textual outputs, extending vLLM beyond traditional text-based, autoregressive generation.
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/omni-modality-model-architecture.png">
+    <img alt="Omni-Modality Model Architecture" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/omni-modality-model-architecture.png" width=55%>
+  </picture>
+</p>

 # Goals

-The primary goals for vLLM-Omni are:
+The primary goal of the vLLM-Omni project is to build the fastest and easiest-to-use open-source Omni-Modality model inference & serving engine. vLLM-Omni extends the original vLLM, which was created to support large language models for text-based autoregressive (AR) generation tasks. vLLM-Omni is designed to support:

-* **Support for Non-textual Output:** Enables the integration, efficient processing and output of various data types, including but not limited to, images, audio, and video, alongside text.  
-* **Non-Autoregressive Structure:** Support model structure beyond autoregressive, especially Diffusion Transformer (DiT), which is widely used in visual and audio generation.  
-* **Integration with vLLM Core:** Maintain compatibility and leverage existing vLLM key modules and optimizations where applicable.  
+* **Non-textual Output:** Enables the integration, efficient processing and output of various data types, including but not limited to, images, audio, and video, alongside text.
+* **Non-Autoregressive Structure:** Support model structure beyond autoregressive, especially Diffusion Transformer (DiT), which is widely used in visual and audio generation.
+* **Integration with vLLM Core:** Maintain compatibility and leverage existing vLLM key modules and optimizations where applicable.
 * **Extensibility:** Design a modular and flexible architecture that can easily accommodate new modalities, model architectures, and output formats.

-# Expected Gains

-* It breaks through the limitation of vLLM that only supports autoregressive text generation models, and provides architectural support for popular model structures, especially multimodal generation.  
-* By supporting new output formats and new model architectures, developers can easily connect to mature ecosystems such as Higress Gateway and Ray, improving the ease of deployment and efficiency of corresponding models.  
-* Integrating efficient cache management, multiple parallel strategies, disaggregation from vLLM, to improve inference efficiency of corresponding models.
-
-# Non-Autoregressive Structure
-
-* **Diffusion Transformer:** initialized with a random latent noise, the transformer model iteratively predicts latent output for multiple steps, under control of sampling timestamps. Finally, it is decoded by the decoder of Variational AutoEncoder (VAE). It is not supported by vLLM now.  
-* **Autoregressive:** dominant generation paradigm for text, now supported by vLLM. It generates tokens conditioned on previous tokens one by one. The efficient KV Cache management provided by vLLM can accelerate inference effectively.  
-  ![][image1]                       ![][image2]
-
-
-|  | Autoregressive | DiT |
-| :---- | ----- | ----- |
-| Scenarios | Text generation | Multimodal generation |
-| Generation | Token-by-token, left to right | Continuous vector, iteratively refine |
-| Key advantages | Smooth, good local consistency | Global consistency, diversity and controlability |
-| Efficiency | More efficient with KV cache | Need generation repeatedly, slow |
-| Error accumulation | Easy to accumulate | Can fix possibly |
-| Key challenges | Limited global planning, long context generation | Large global computational cost |
-| Sequence length | Varied length | Fixed length |
-| Attention mask | Lower triangular | Full bi-directional matrix |
-| Parallel strategy | TP, DP, PP, .etc | DP, CP, .etc |
-| vLLM support | Supported | Unsupported |
-
-# Targeted models to support
+# Representative omni-modality models

 According to analysis for current popular open-source models, most of them have the combination of AR+DiT. Specifically, they can be further categorized into 3 types below:

-* **DiT as a main structure, with AR as text encoder (e.g.: Qwen-Image)**  
-  A popular model for image generation and editing. Main visual generation models are similar to the structure, such as Flux.  
-  ![][image3]  
-* **AR as a main structure, with DiT as multi-modal generator (e.g. BAGEL)**  
-  A unified multimodal comprehension and generation model. Visual generation can utilize CoT text generation output. Main unified multimodal models are similar to the structure.  
-      ![][image4]  
-* **Extension, multiple AR+DiT (e.g. Qwen-omni)**  
-  A new model for multimodal input and output. It designs the Thinker-Talker-Codec structure, which is a double-AR+DiT format.  
-  ![][image5]  
+**DiT as a main structure, with AR as text encoder (e.g.: Qwen-Image)**
+  A powerful image generation foundation model capable of complex text rendering and precise image editing.

+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/dit-main-architecture.png">
+    <img alt="Qwen-Image" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/dit-main-architecture.png" width=30%>
+  </picture>
+</p>

+**AR as a main structure, with DiT as multi-modal generator (e.g. BAGEL)**
+  A unified multimodal comprehension and generation model, with cot text output and visual generation.

-# Architectural Overview
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/ar-main-architecture.png">
+    <img alt="Bagel" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/ar-main-architecture.png" width=30%>
+  </picture>
+</p>

-vLLM-Omni will import core modules from existing vLLM framework and introduce several new components and modifications to achieve its goals. The core architecture will be based on a modular design, allowing for independent development and integration of different modality handlers and output generators.  
-![][image6]
+**AR+DiT (e.g. Qwen-Omni)**
+  A natively end-to-end omni-modal LLM for multimodal inputs (text/image/audio/video...) and outputs (text/audio...).

-## General ideas
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/ar-dit-main-architecture.png">
+    <img alt="Qwen-Omni" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/ar-dit-main-architecture.png" width=30%>
+  </picture>
+</p>

-* It divides models in vLLM-Omni into multiple stages, which are processed by different Engine Cores. Each stage can choose AR Engine Core for AR model part, and Diffusion Engine Core for diffusion model part.  
-* Designed a specialized diffusion engine core for DiT or dLLM.  
-* Plug-in with external DiT acceleration modules such as xDiT, Cache-DiT.  
-* Output linked with Higress API, Gradio for online demo, ComfyUI for less-code process.
+# vLLM-Omni main architecture
+
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-main-architecture.png">
+    <img alt="vLLM-Omni Main Architecture" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-main-architecture.png" width=55%>
+  </picture>
+</p>

 ## Key Components

-* **Processor:** Input processing, tokenization or encoding  
-* **Scheduler:** for order and queue of incoming requests  
-* **Executor:**  Prepare model input and run  
-* **Multimodal output processor:** support multimodal outputs  
-* **Hidden State Output processor:** support intermediate hidden vectors as request output  
-* **Engine Core:** accelerate AR modules, imported from vLLM main repo  
-* **Diffusion Engine Core:** Wrap and accelerate DiT modules.  
-* **Diffusion model:** Wrap up diffusion models with vLLM format
+| Component         | Description                                                                                                                              |
+| ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
+| **OmniRouter**    | provide an intelligent router for Omni-modality requests dispatch                                                                        |
+| **EntryPoints**   | define the APIs for offline/online serving (APIServer, Omni/AsyncOmni) and provide the OmniStage abstraction for different AR/DiT stages |
+| **AR**            | adapted for omni-modality models while inheriting efficient features from vLLM, such as cache management                                 |
+| **Diffusion**     | natively implemented and optimized using acceleration components                                                                         |
+| **OmniConnector** | supports fully disaggregation based on E/P/D/G (Encoding/Processing/Decoding/Generation) disaggregation across stages                    |

-# Examples data flow of specific models
+Disaggregated stages are managed through configuration, such as in the Qwen3-Omni example, where stages like Thinker, Talker, and Code2wav are defined as separate OmniStage instances with specific resources and input/output type.

-1. Qwen-omni  
-   ![][image7]![][image8]  
-2. Qwen-Image  
-   ![][image9]      ![][image10]
+## Main features

-# Overview of function call dependency
+vLLM-Omni aims to be fast, flexible, and easy to use with the following features:

-![][image11]
+### Performance and Acceleration

+The framework achieves high performance through several optimization techniques:

-[image1]: <data:image/png;base64,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 truncated
+* **Efficient AR Support:** Leverages efficient KV cache management inherited from vLLM.
+* **Pipelined Execution:** Uses pipelined stage execution overlapping to ensure high throughput.
+* **Full Disaggregation:** Relies on the OmniConnector and dynamic resource allocation across stages.
+* **Diffusion Acceleration:** Includes integrated support for diffusion acceleration. This is managed by the acceleration layer, which handles:
+    * **Cache:** Includes DBCache, TeaCache and third-party integration(e.g., [cache-dit](https://github.com/vipshop/cache-dit)).
+    * **Parallelism:** Supports TP, CP, USP, and CFG.
+    * **Attention:** Provides an interface for third-party integration (e.g., FA3, SAGE, MindIE-SD).
+    * **Quantization:** Supports various quantization implementations including FP8 and AWQ.
+    * **FusedOps:** Allows for custom and third-party integration.

-[image2]: <data:image/png;base64,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 truncated
+### Flexibility and Usability

-[image3]: <data:image/png;base64,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 truncated
+vLLM-Omni is designed to be flexible and straightforward for users:

-[image4]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAO0AAAFdCAYAAADv8OpOAABCCElEQVR4Xu2dB3gUxRvGlwRCh9BTCb2KVAWkKiDSiwKiCPxV7IACKlbABohBESk2pAkC0kR6SWjSW+i9Kr0TapL3P7N3c+x911I2l1z4fs/zPrmd/XZmd2fendkaTWMYhmEYhmEYhmEYhmEYhmEYhmEYhmEYhmEYhmEYJoUA8OtVqnqTj8rUfappcMmmOfyzNhXJKVJw1hxNWwSXbdo9vFK9IdUb59UYhjGHryrUfapFSIlnS2XPs6NOgdA9Imm3WSqcNfvuanmLrK0WGNxLYxgmZcyr1SL020qNhvlrmdCocDHMr9MR0fWfx+qGL7jVGqEV9Z/TtbJBF/1vVAPny8nYn6s1R818IagaWKhnEa1ITroeDMMkko/K1u1QPTDo4FslquNEszeBZz5AwtP9cbplT6FeLvVfi5641eZd3GjVD9fb9EVs63643fY9nG3Z2yFW6lLrdzC71tMQRUZny6zVoevBMEwi6BRSrmWpHHn3BmfLiXhh1GvCfBVzF5TGSpSWDWoDzOuL7Hn8sPLL9ujfqZpDDFWmTPrf2IjQnG9Pi6yVX2MYJvGUyJ63XW7/zIdr5w8Vpv1ADI076Mbq1KIEpg5vhCmRT1g03FGThz2Oazu7Aadfw8JfmuH6ru44uvJZTPu2kUOsUVPF/O8/fgz+fpkOly0V2puuE8Mw7mkrdLBx4QjsaPwSauYNwdLfWgBnegGHXsPtnS/jg1eroGeXiuhlot7u/hAqlgpEZj8NT7Ro+3vjlq3/R1eMYRgn+BtM+8/jXVExRyFsmN4WOPUmTq3qghfalHYY3pqtXLnzxGXLnn1f7wHfBmoMw7jHaNp1j3fTTbt+mjTtW3i1U3ndVN3efAf9Ph+GPoOGmqp+n32Nd78Yhi6vvomQsHCEFS/1xoABA9i4DOMOl6Y92wuBuTOjRr0ncPQOsP9aQop18AZ00fTj94C1B/9DoSJBl1s/+z++j8sw7vBk2gZPtdGNtufyvRTp0E1g75U47Lsaj/3X7fM7GAusP3YOgQUKXmrS8mk2LcO4w5Np6zdtnSLTHr4FLNq6Dy06dLadw/YdNATbTl+3M+26o2flvIuBgfl72q8hwzB2pMS0+8TQ9thd6L0onae08fgFNGvX0XJu3L40nn6yGDL5Z8ankaP0XpeaVohNyzDuSI5pd1+6hw3HzqPPwK9QskRxNG7eCgfEkPfA9QRi2jj0GfClbtj+r1bWbyHh8OsILxyArNlz68vsEUNmfXh8/LyMOyX0hv0aMgxjR3JMK3vXds93swx3C4Xpf4f9PFE3HzXt+18M0+e/9Exp/TYSTr6JIvkzI3uuQJtpZU+9at/xe8FhRd8SseXtVpBhGHuSYtoj4vx0+op1qFKjJrRSlaENnQNtTTy0OSd1YwYFh+gGlMNmW698ORa9P/5MzPfHIw8VQMVSeVAkJAwT/l6mX5jS870NLIs5dGPBgQNZ6foxDENIrGnlFd+f/pyPEqWsD1uMXgVtRSy0hRegLb8BrVUPPX3k73/aneMeEr3vplOXMOi70fDLkhu5Agvjt3lL7a4gy4tVy2IOxu5JSChAVo9hGEpiTLv3yj30/yrSYtZmYlg87QC0RRcthlWSxpXpIqZ4iZK6WdWFJil10Urek5U97K4Lt7FX3v6R92+tt3zEsleFmtutIMMw9ngy7RMt2uON9z5BtoAAaJ37iqHwKWHQ6/aG1XXRYtxGz+rGHTd3MQ6Tq8qyd5Xq/kZvVKxUCQOHj8bW/67i0C3b1ePrQi3s15BhGDvcmTZ/YFb4+flBKxgCrf/P0JZccWJWouViyPz7Lt245R96WDeu7EnlkPmN9z5G3pw5oFVpCG3A79AKhOhxE/5ernraK4ULF25tv4YMw1DaCR1q5My0eQMsQ+LBs6AtvuxoUKe6aDnXfayFvuy0Zf+Ic+K7mLlyoyWvnHmhTT8EbZnorUcshVa5HoIKFcJb73+MHNmyXQwLC3vTfvUYhrEjOCB72xz+/ocq5SmEnU1eRiWjacXwWCtUTBjsmjDtJScGdaMJ26HVbIpHHquLdz79AqVKl4UWUR7ax+OhLb1qOSdeIYbTU/dCy5YLosdHYGDgjR49ekyh68gwjAEMGOD3aGBwvWyZMs8pn7sA8vtntzftI80sJqOmTIyib0Gr2tDSwxYIhrY2wbn5ZZzozbOKofjSpUuRkJBQiq4nwzCECrkK1dUyaZeFEjbMaGeOaaV+3gDt4ToW4/60ztJr0xg5VJ59HFlFzOzZs6Vp36brxzCMEyZ81ahA+WKBH3dvVwZXdr6MgnmypNy0UtG3oY1ZDS0gK7TIBY7zlwojTz+InH7arSlTpqwUpq1K141hGBe0bhxeI1s2/7stGxRFzqx+0Go1T7lppeTFqS9nQPvziOM8q2kLZvW/2LZ1C/5mFMMklbLFC1fKlyvLy+LnXa1mM8vwVV44Si3J+77CtFrhkMs5m/D7tAyTXLJoWXNt1IKLWe69Lr5iue2TGlp5B9rwhdCCwo/6PdbieboiDMMkkor1Hn9MCys1TcscIM5Hs0PLki115JcZWbLm+Dc8vEAIXQeGYZJIoT7fNG3+ZBM8Xq8uGtatkypq0qBefL0XekTRshmGSSL37t1rGR8fvxxeQpT1XUJCQkm6HgyT4RENv49V7em8xCI8lEWYKIr4KlUR5V0UGkHXhWEyPAYTXBbG/VT8HCj/CtUSv/2E/KXEdFZnEvNyiV62lcrHm8TFxV0U61Dauh4B1r/+dBsZJkNBjeAKYeqrwiSXqcSsKzTWm4j1ui3XQ/y9JHTNmnZEGDiMbivDZAioCdI1wpPnb9DE+1w7fwZnrt6Vpo0Rpo2g28owGQLa8G2cmQGtwPP4S76Jng7YNPYl1KzUBDfpDAMbv2yCoGLNdNNOB4rSbWWYDAFt+IrYqCEo+MxP2HMpgc5KA+LQrHxpNB8UTWfYcWHtp6gZVAL32LRMRoY2fAvnMLNfQ7w5Zy9ui6mXH6uCRzoNRO0yRRFcpAiCitbFxpVj0aNhBRQuVBghD3cRtrIw5IV6KF00DEFBwQgNL4n6Lw615bpqTC80figcwSFhKN+gC16p/yhavT8HuHMKS0e+hvDQMISFh4rlyiLmYrx1qVvY/ctz6D7hMOThI+HKbvz+YRuEFyuGYkVD0azvROy8ZIm9c+Ef9K1REPPuxscMOMymZTIoNkcZubseg9sWx09bzuuTFXNp0HKVwLglG7Bu7rdoV8wPDz/eCj0+m4yZn7REsMjmrHTUjc3Im7cMnhk4GYdiVmDM241RwL8A/pOZJFxFxby5UbH9R4hZ/Qc+eroycmYLQqtv9yDqm2cQmisPJq3ciyNHtmHqhy3Q+puVOCu8GHdlHfrXCULUVcuqbR7fA6VzhWLnieM4tnspcmQqgI5jtuBigozdh8/r58PPV+JjfrzJpmUyKMqnRs4tGISGBTTs0y/6XBZGLI93l57T513a+Qc6BflhtfVU95/IjiiWORyx+lQCDh/Yi107tiFq8SyM7N8RZTLnwRVcx9/9G6L3zL24Y1kMuL0XBct1xZwT/0LzD0erj//Eju3bsV1oR8xi+Id2wsy9N7C036MoUKiyWgo7ZryLGnkz4/XIGdhw8LRIice9e3FIEKa9d3k7PqmdD7/cYNMyGRibGwysG/kCimsB0PvZu6uRv+LrWHHpnpi4g4OL3kc5/8LQ760IfnulBnLkeRLSsJd2z0eVMsUQHhaGsuXKoGRwLmgBFUUXuBlftQzB9Jj7l37jrq5GyZaRiPl3BzIF5EZIibIoWbKkrtJlyyLsiQ8QffwWvm4ahoIlnrUth4QziBrdE/kDNGQNfgT7r6iBuRhIn5mPl0rnxyIeHjMZmftuuM+g5kWgBXXUf1/88w00HroV1+Vp47WD+L1rBAJK3X+WokNpP4R3+Q1357+JMvlz3O9JD89Gt5IactV8TRh/A4a0K4rx2y39Me4dwYTuFdB16jncPbcdOav3wJwLasEE7Jo5BH/tkbeAgedLi3Pmpt/ov+PP7kSn93/FdlXIv7OghXbF3H2WsfOeMW1RJF8JxMXH71jA92mZjIq1+dvRpWwmZHt8EOTQc9WnT+DtJaeh97Mn1+DdatkR3KCvLbZGzmxoMiwatxb2Qrm8WXFKBgq++V9lZNH8UO7pwZBD7DWRHVD9peFYsikGP/ZriQjRUw7blSCvHiGrfwhaD1qAy6f3Y9XkfqhTJDPmH7EYUZo2tPlw/XfChS3Qcj+C92bs16dvn5qFLBV7Yckxy42gb5sXRb6yHeUtn50H2LRMRkVv7YSxg4di+vpj4tcprJwyFidiLU68emIVfo0chT9WHbFGXkXk8D+w+aTlBPfSrvlo07QJmrXvivUxe7Doh4/x1bgl1lhgxe8j8eWgz7A26js0L14IJ+9a0mNPrMPU7z5B15d7YsD307Dh5P27sUObRyC46lu26cOrfsfgPi+iYaMWeO71j3BJjY7jT6CgMOzEffJ6t/5U1E66rQyTIbC5wSSuXbqEaxbf2Di1diI+6NrdNn10THuE5iqCe4m4Bbykb20EFWtBk3H+0nXbbSbJqXn98NyoTbaHL+RjjXRbGSZDYGj3qcaFpV+gjhgOf/LzdPw1bwYeC82OQo/1oWFOOT2vJyqHVqTJ9pxZgj71y2IXecSRbivDZAjsm3kqEheLzevWYNWq1Th+4Rad65qEWzi6YbHLxxcTYk9g3cIFWHvQcuHKCN1WhskQ0IaeXnE5kpY3aF1At5VhmBQQpWmBUjSdYZh0iDRrtKZtk2LjMkw6x2pYGMXGZZh0iqGHtTOtTPtb0/IZYxmGSWOc9bBUbFyGSSe46WGpuMdlmLQmMT2sE23TF2YYxvsI07YVGigVrWlHox0NqnRUxUnZMmAYJu2IFh6OdjSrUtT9SIZh0gXRbFqG8S2i2bQM41Line truncated
+* **Heterogeneous Pipeline Abstraction:** Manages complex model workflows effectively.
+* **Hugging Face Integration:** Offers seamless integration with popular Hugging Face models.
+* **Distributed Inference:** Supports tensor, pipeline, data, and expert parallelism.
+* **Streaming Outputs:** Supports streaming outputs.
+* **Unified API:** Provides a consistent and unified API interface compatible with vLLM.
+* **OpenAI-compatible API Server:** Includes a FastAPI-based server for online serving that is compatible with the OpenAI API.

-[image5]: <data:image/png;base64,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 truncated
+# Interface design

-[image6]: <data:image/png;base64,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 truncated
+If you use vLLM, then you know how to use vLLM-Omni from Day 0:

-[image7]: <data:image/png;base64,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 truncated
+<p align="center">
+  <picture>
+    <source media="(prefers-color-scheme: dark)" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-user-interface.png">
+    <img alt="vLLM-Omni interface design" src="https://raw.githubusercontent.com/vllm-project/vllm-omni/refs/heads/main/docs/source/architecture/vllm-omni-user-interface.png" width=55%>
+  </picture>
+</p>

-[image8]: <data:image/png;base64,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 truncated
+Taking **Qwen3-Omni** as an example:

-[image9]: <data:image/png;base64,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 truncated
+## Offline Inference
+The **Omni** class provides a Python interface for offline batched inference. Users initialize the Omni class with a Hugging Face model name and use the generate method, passing inputs that include both text prompts and multi-modal data:

-[image10]: <data:image/png;base64,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 truncated
+```
+# Create an omni_lm with HF model name.
+from vllm_omni.entrypoints.omni import Omni

-[image11]: <data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAABM8AAARlCAYAAACukOBGAACAAElEQVR4Xuy9B7hV13WuHSd5ch3Zia17rcQl13+K88dyXOLYlmPLRb13CVUkUEMIEB3Rkeiig+iiI3rvvffeO4gmeu+dM+56xz4TbRZ7IxDt7M33Ps98zll97rXGnGvOb4055l+YEEIIIYQQQgghhBAiJX8RXyGEEEIIIYQQQgghhEgg8UwIIYQQQgghhBBCiDRIPBNCCCGEEEIIIYQQIg0Sz4QQQgghhBBCCCGESIPEMyGEEEIIIYQQQggh0iDxTAghhBBCCCGEEEKINEg8E0IIIYQQQgghhBAiDRLPhBBCCCGEEEIIIYRIg8QzIYQQQgghhBBCCCHSIPFMCCGEEEIIIYQQQog0SDwTQgghhBBCCCGEECINEs+EuAhnz5jlKCnF09k0Kb49/B9fH98/1bazSfaXYhvr025PPm98e3xd8nJ826VcK36eVNeN/e/nih+Xajl+vaT1550jvu+lbIv/TfV/fDn5nPHz5yYhhBBCCCFE9iHxTIh05JgdP5Bjx3YrKcXSntTpeG5KXk61/lzafZFtsXPE07n1KfJz7pwp8nvB9ZL2uWBb8j7ptuWm+DWT973geqmOS943zfW+7Den+j++HD9Hym25/8eXk895wfmj8506mhOvRYQQQgghhBBZgMQzIdKBeLY/x45sz7GjO5SUvkhH0qT49vB/fH18/1TbLnbclx2bvC2+Pb4ueTm+Lb5Pqm2p9onvG///Upfj10ten25b/P/4cvzcqbalukb8nBecf7vEMyGEEEIIIbIViWdCXIQT+y3ROd6ppKSklD5RT0g8E0IIIYQQIjuReCZEOnIQz3ISHigpOstKSkpKIUk8E0IIIYQQInuReCZEOsKwTYlnSkpKX5IkngkhhBBCCJG9SDwTIh3yPFNSUrrEJPFMCCGEEEKI7EXimRDpkOeZkpLSJSaJZ0IIIYQQQmQvEs+ESIfEMyUlpUtMEs+EEEIIIYTIXiSeCZEOiWdKSkqXmCSeCSGEEEIIkb1IPBPiIpzYbxLPbmA6vjt6CEej57Dni+dgh83OHkgsu2CxL9rnpPlflk9wzIlon4Nf7HNyb2IfO56bonMe22UXXC+euD7nuZR9SX79KK/k8cuOYd/TkX3lHJKNZUOSeCaEEEIIIUT2IvFMiHRowoAbmhC8tq3bby0afWKfLd3mQhPrenToZ2MGT3ax7PRBs3lTllmJImVsxbz1dvpQ4pjypSrb8H7jfDv7LZu91t59s5gVzk1Vy9ew3RtQuOycoIZQhuCFsBYEtu1r91vfbkNs/+fsYC7K2ZGEKOYiGeuOJY4NQt6GKK9d2va0PZuOfnGu4wkB0MXAIOBFl58zabGNHDDBBTpSOD/58GXyd+ILYZB74PvkCm5nDnyxHL9/Stc3STwTQgghhBAie5F4JkQ6NGzzhiaEorWLt9hD9z9mcycvdVGJde++Ucwa1m7m4hMi1IAew+373/uBNW/Y1h/bkN6jfblRneYuPuHZNbDHCLvvrgds5viFtmDaUlsya5Xt2XjE5kxaZD079bcB3YfZwa2nXOBCdOvZaUC03zL7fNVuGzFgvK1bssUmjphhndv0sCmjZrvnG/v36NjPBvYc4fk9sSfHBa/5U5ZZwfxv2cblO23RjOXWu8tA69NlkO3//IQd3n7Gl7u172N7N+2zSmWq2aMPPm57Nh2wxTNXW6fW3WzCsOkuni2ZtSbK93Dr3qFvlI89djr6vRuWbY/y0N1mTljowtyaRZs9T3OnLPU8xe+h0vVLEs+EEEIIIYTIXiSeCXERNGzzxiWEMjzO8j39ki2YtsLFMtaVKVbeWjRs66IYAtKwvmPtyUeftVJFy9mhHcetecM2lu+pF619y09dcEPQGtpnjL30XH4XoXatP2CHo/3WR+f+nzvudE+0P995t/XpOtjGD5saXe9Fe+O1QvbgfY/YzAkLrGzxCtamWQe741f/YyXeLWP3/Pk+Wz5njbVq3N5KFStnBV9509o07eBeYORp0fSV5zzh7r/nQXvztXeiY+73/HRq3d3ee6ekVS73gQtolcpWs8ceftKFOfLfpF5Le+bxfDZ93Hx75YUC9nz025945Gmr8n4N2711X/S3ur1fspIVi84xacxUK/1eeav9QT17KV9+mzNxkZ2WgHbDksQzIYQQQgghsheJZ0JcBIlnNy5dqng2uNcoK1mkjL1fqrIN6D7CalerbzUq17HWTTu4eMZ+YwZNstv/4z8t/wsFrMDLr1vLxm1s7aLP7bmnXrAd6w5Yp1bdrH7NJlauRAUfErp1zT4r9EYRmzZmrpUrWTHa/xN7r3ApP1fZaJ9P2/exJx992lo1aW8fVKhlTz/+nB3YctLzE8SzZXPWuaj12dKt1rfrYKsV5avtx52sepS3kQNG277NB21g9+HWqM7HdmLP2SjvQ21Qr5H2wrMvW99Ph1iRQiVsanT9GePmW8WyVaPtw/28eNMd2XPUOrfuYX+68y4b2meYPfLA4+5pR/7i91Hp+iSJZ0IIIYQQQmQvEs+ESIeGbd7QhFC2bulWe+HZV1zock4kxLMOLbuee0yDeo6wmlU/cq8uvLVqVatnrZu0TwzjzPU8Q2B7u0DhRKyxXH1j3eIt7nW2c/2h6HyfutcXItzEETNt+7oDVrRQiYR4VqKifdK8k4tfDKdkqGW39n3tuadfsJpVPrKPG7T0oZOHtp2xnMNfiGcMu+TvphU7rVfHAVbnw4Z2cNux6Nge7s3W9ZNePuyzWb1WNnviYnv5+VetWf1W9uC9D/tQ0MrlPrSF01fYhOHTo99X14eSFn+39Lk4bd2jPNz9p/tcCKxfs7FNGjnTTu298D4qXZ8k8UwIIYQQQojsReKZEOmQeHZDE+LZ+mXb7Q+/v8uFq95detnMCXOtyFvF7eV8r1qvzr1sxMARLnzh/TV2yBS77Tu32bC+Y6xFo7ZWv1ZT9wTDGwvx7Df//VsX2Lq1722Deg6xuZOXWIGX37Bta/db8wZtrHHdFja490h7q0Bhq/NBA3vmiXw2bew8K/p2cR8KWrFMNc8XotqwfmPdY6xhnWY+bJLjT+xNzARKzLO3Cha2hdNX+vDPz5ZstS5te9iHlWpbuxado3O19mM/ad7Z+n06xAq88oa1b9HFr9e/21B75IHHXIwrWbSsTygwatBEK1uivK2cv87PhyhYqlhZ69Wlr4tzvToPsBLvlrbJo2b5PYvfR6XrkySeCSGEEEIIkb1IPBMiHRLPbmg6vjvH9m85YZ+262N1PmxgtarWs8G9R3nA/no1Gvsyccfw2po1YaHt3njExadd6w/ZwmkrfBbOk/uYtTPHA+03rdfSaler555pzeq3ttULNrlXF9fAW2zBtOW2d9Mxa1i7oRV45U2PZcZxU0fPskUzVvrwSfJFfLItq/b49Zi4oEGtprZpxS47tS8xacC2Nfv8vMzUOX7YNNuz8agtn7POxbrN0X51qzf0vByIrrt55S6Pp8YQT2boRFQbPWiSX2/K6Nm2dfVej82GMMaEALMmLvJhqUP6jLacI4nZOvk9vToNiOz0rN+z+H1Uuj5J4pkQQgghhBDZi8QzIdIh8eyGJ4ZJMlTTTuemI5aYZTMsn0gMy8yJ0ond5sMZmTGT4P2ITTw70sm9SceQTpqd2pc4F9dgf87F7JYli5axd94o4gIY52afMBkAeWKZ83EdO5VILIdr+frDudujv8d3J2K1cS7PB8ecTKz35ZOJ7f47o//5jZ6fQ4lznGKfQ4nze5y33PtwlGXyx/KxxO+I3z+l65ckngkhhBBCCJG9SDwTIh0Sz266hACFCIewpiGQSpeTJJ4JIYQQQgiRvUg8EyIdEs9uyoSAJi8upctNEs+EEEIIIYTIXiSeCZEOiWdKSkqXmCSeCSGEEEIIkb1IPBMiHRLPlJSULjFJPBNCCCGEECJ7kXgmRDoknikpKV1ikngmhBBCCCFE9iLxTIh0SDxTUlK6xCTxTAghhBBCiOxF4pkQ6ZB4pqSkdIlJ4pkQQgghhBDZi8QzIdIh8UxJKU8kZj89nsdnQJV4JoQQQgghRPYi8UyIdCCe7Ys6xdsTHWMlJaXrn45G6cRus5N77YJteSpF9cSpIxLPhBBCCCGEyEYkngmRhpyoH4wnCR1iJaXz0lG7cDms+7JtydtTLce3pTrvxY5LtRzfFj82+RrxY+P7xrelOzbVvun+j58nadvZE2Zd2/ewMsUr2IlDZ+x0/Nj4NeLnSZe/VMelOybVtvj/UTpzQuKZEEIIIYQQ2YjEMyGEEHmacePHWdu2bS0n52x8kxBCCCGEEEJccySeCSGEEEIIIYQQQgiRBolnQgghhBBCCCGEEEKkQeKZEEIIIYQQQgghhBBpkHgmhBAiTzNixAj76KOP7OxZxTwTQgghhBBCXH8kngkhhMjTFC9e3G699VY7ffp0fJMQQgghhBBCXHMkngkhhMjTVKxY0f7t3/7NTp06Fd8khBBCCCGEENcciWdCCCHyNJUqVbJ//dd/lXgmhBBCCCGEuCFIPBNCCJGnqVy5sv3Lv/yLxDMhhBBCCCHEDUHimRBCiDwNwzZ/+MMfSjwTQgghhBBC3BAkngkhhMjT4Hkm8UwIIYQQQghxo5B4JoQQIk+za9cuW7NmjeXk5MQ3CSGEEEIIIcQ1R+KZEEIIIYQQQgghhBBpkHgmhBBCCCGEEEIIIUQaJJ4JIYQQQgghhBBCCJEGiWdCCCHyNK1bt7Z77rlHEwYIIYQQQgghbggSz4QQQuRpqlSpYn//939vJ0+ejG8SQgghhBBCiGuOxDMhhBB5mpo1a9p3v/tdiWdCCCGEEEKIG4LEMyGEEHmaWrVq2T/+4z9KPBNCCCGEEELcECSeCSGEyNMgnv3DP/yDxDMhhBBCCCHEDUHimRBCiDxNtWrV7JZbbpF4JoQQQgghhLghSLine truncated
+omni_lm = Omni(model="Qwen/Qwen3-Omni-30B-A3B-Instruct")
+
+# Example prompts.
+om_inputs = {"prompt": prompt,
+             "multi_modal_data": {
+                 "video": video_frames,
+                 "audio": audio_signal,
+             }}
+
+# Generate texts and audio from the multi-modality inputs.
+outputs = omni_lm.generate(om_inputs, sampling_params_list)
+```
+
+## Online Serving
+Similar to vLLM, vLLM-Omni also provides a FastAPI-based server for online serving. Users can launch the server using the vllm serve command with the `--omni` flag:
+
+```
+vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
+```
+
+Users can send requests to the server using curl:
+
+```
+# prepare user content
+user_content='[
+        {
+          "type": "video_url",
+          "video_url": {
+            "url": "'"$SAMPLE_VIDEO_URL"'"
+          }
+        },
+        {
+          "type": "text",
+          "text": "Why is this video funny?"
+        }
+      ]'
+    sampling_params_list='[
+      '"$thinker_sampling_params"',
+      '"$talker_sampling_params"',
+      '"$code2wav_sampling_params"'
+    ]'
+    mm_processor_kwargs="{}"
+
+# send the request
+curl -sS -X POST http://localhost:8091/v1/chat/completions \
+    -H "Content-Type: application/json" \
+    -d @- <<EOF
+{
+  "model": "Qwen/Qwen3-Omni-30B-A3B-Instruct",
+  "sampling_params_list": $sampling_params_list,
+  "mm_processor_kwargs": $mm_processor_kwargs,
+  "messages": [
+    {
+      "role": "system",
+      "content": [
+        {
+          "type": "text",
+          "text": "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech."
+        }
+      ]
+    },
+    {
+      "role": "user",
+      "content": $user_content
+    }
+  ]
+}
+```
+
+For more usages, please refer to [examples](https://github.com/vllm-project/vllm-omni/tree/main/examples).
@@ -0,0 +1,106 @@
+# Disaggregated Inference for Omni-Modality Models
+
+This guide explains how to configure and use distributed connectors
+(`vllm_omni/distributed/omni_connectors`) in vllm-omni for multi-stage pipelines.
+
+Backend-specific setup lives in separate docs:
+
+- [SharedMemoryConnector](omni_connectors/shared_memory_connector.md)
+- [MooncakeConnector](omni_connectors/mooncake_connector.md)
+- [YuanrongConnector](omni_connectors/yuanrong_connector.md)
+
+## Overview
+
+Connectors enable data transfer between pipeline stages (e.g., Thinker -> Talker).
+Current connectors operate in D2H2D (device to host to device) mode.
+
+## Connector Choices
+
+| Use Case | Recommended Connector | Notes |
+| :--- | :--- | :--- |
+| Single node | SharedMemoryConnector | Auto-configured if no connector is specified. |
+| Multi node (Mooncake) | MooncakeConnector | Requires Mooncake Master + metadata server. |
+| Multi node (Yuanrong) | YuanrongConnector | Requires Yuanrong Datasystem + etcd. |
+
+## Core API
+
+The connector system is built around `OmniConnectorBase`.
+
+```python
+class OmniConnectorBase(ABC):
+    @abstractmethod
+    def put(self, from_stage: str, to_stage: str, put_key: str, data: Any) -> tuple[bool, int, Optional[dict]]:
+        """
+        Store data.
+        Returns: (success, serialized_size, metadata)
+        """
+        pass
+
+    @abstractmethod
+    def get(self, from_stage: str, to_stage: str, get_key: str, metadata: Optional[dict] = None) -> Optional[tuple[Any, int]]:
+        """
+        Retrieve data.
+        Args: metadata - transport-specific handles returned by put() (e.g., SHM name).
+        Returns: (object, serialized_size)
+        """
+        pass
+```
+
+### Metadata Passing
+
+Some connectors (e.g., SharedMemoryConnector) generate transient resources during `put()`.
+This `metadata` must be passed through the control plane so `get()` can locate the data.
+
+## Configuration Model
+
+Define connectors in runtime:
+
+```yaml
+runtime:
+  connectors:
+    connector_of_shared_memory:
+      name: SharedMemoryConnector
+      extra:
+        shm_threshold_bytes: 65536
+```
+
+Wire stages to connectors:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    output_connectors:
+      to_stage_1: connector_of_shared_memory
+
+  - stage_id: 1
+    input_connectors:
+      from_stage_0: connector_of_shared_memory
+```
+
+If a pipeline edge has no explicit connector, the system auto-creates a
+SharedMemoryConnector for that edge.
+
+## Relationship with vLLM
+
+vLLM provides specialized distributed mechanisms for specific artifacts:
+
+- KV Transfer (`vllm.distributed.kv_transfer`): optimized for KV caches.
+- EC Transfer (`vllm.distributed.ec_transfer`): optimized for encoder embeddings.
+- Device Communicators (`vllm.distributed.device_communicators`): low-level primitives (NCCL, SHM).
+
+vllm-omni complements this with a generalized connector abstraction:
+
+1. Unifies transport via a single `put`/`get` API for any stage artifact.
+2. Enables DAG-style pipelines across processes or nodes with per-edge transports.
+3. Can wrap vLLM-specific transfers for KV paths while keeping a consistent interface.
+
+## Operational Notes
+
+- Fail-fast config validation: missing expected edges cause startup failures.
+- Missing payloads halt stages: verify connector wiring and metadata propagation.
+
+## Future Roadmap: D2D Transport
+
+Current connectors use D2H2D paths. Future versions will introduce direct
+device-to-device connectors (NCCL, UCX, IPC) to reduce latency for large
+tensor payloads.
@@ -0,0 +1,74 @@
+# MooncakeConnector
+
+## When to Use
+
+Best for multi-node distributed inference using Mooncake.
+
+## Installation
+
+```bash
+# For CUDA-enabled systems (recommended)
+pip install mooncake-transfer-engine
+
+# For non-CUDA systems
+pip install mooncake-transfer-engine-non-cuda
+```
+
+## Start Mooncake Master
+
+```bash
+# If you use Mooncake SSD storage
+mkdir -p ./mc_storage
+
+mooncake_master \
+  --rpc_port=50051 \
+  --enable_http_metadata_server=true \
+  --http_metadata_server_host=0.0.0.0 \
+  --http_metadata_server_port=8080 \
+  --metrics_port=9003 \
+  --root_fs_dir=./mc_storage/ \
+  --cluster_id=mc-local-1 &
+```
+
+## Configuration
+
+Define the connector in runtime:
+
+```yaml
+runtime:
+  connectors:
+    connector_of_mooncake:
+      name: MooncakeConnector
+      extra:
+        host: "127.0.0.1"
+        metadata_server: "http://<MASTER_IP>:8080/metadata"
+        master: "<MASTER_IP>:50051"
+        segment: 512000000
+        localbuf: 64000000
+        proto: "tcp"
+```
+
+Wire stages to the connector:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    output_connectors:
+      to_stage_1: connector_of_mooncake
+
+  - stage_id: 1
+    input_connectors:
+      from_stage_0: connector_of_mooncake
+```
+
+Parameters:
+
+- host: local worker IP registered in the metadata server.
+- metadata_server: metadata server URL for discovery and setup.
+- master: Mooncake Master address.
+- segment: global memory segment size in bytes.
+- localbuf: local buffer size in bytes.
+- proto: transport protocol ("tcp" or "rdma").
+
+For more details, refer to the
+[Mooncake repository](https://github.com/kvcache-ai/Mooncake).
@@ -0,0 +1,27 @@
+# SharedMemoryConnector
+
+## When to Use
+
+Best for single-node deployments where stages run on the same host. It is
+auto-configured when no explicit connector is specified for an edge.
+
+## How It Works
+
+- Small payloads (< threshold): serialized and passed inline in metadata.
+- Large payloads (>= threshold): stored in shared memory; the SHM name is
+  returned in metadata.
+
+## Configuration
+
+```yaml
+runtime:
+  connectors:
+    connector_of_shared_memory:
+      name: SharedMemoryConnector
+      extra:
+        shm_threshold_bytes: 65536
+```
+
+## Notes
+
+- Auto-mode uses SharedMemoryConnector if no connector is declared for an edge.
@@ -0,0 +1,101 @@
+# YuanrongConnector
+
+## When to Use
+
+Best for multi-node distributed inference using Yuanrong Datasystem.
+
+## Mechanism
+
+Uses Yuanrong Datasystem's distributed KV store (`datasystem.kv_client`).
+
+- Data Plane: TCP or RDMA for high-bandwidth transfer.
+- Control Plane: Yuanrong Datasystem workers and etcd.
+- Keying: deterministic keys based on `put_key` (often composed as `request_id:fromStage_toStage`).
+
+## Installation
+
+```bash
+pip install openyuanrong-datasystem
+```
+
+## Start etcd
+
+```bash
+# Download and install etcd (v3.5.12 or higher)
+ETCD_VERSION="v3.5.12"
+ETCD_ARCH="linux-arm64"
+wget https://github.com/etcd-io/etcd/releases/download/${ETCD_VERSION}/etcd-${ETCD_VERSION}-${ETCD_ARCH}.tar.gz
+tar -xvf etcd-${ETCD_VERSION}-${ETCD_ARCH}.tar.gz
+cd etcd-${ETCD_VERSION}-${ETCD_ARCH}
+sudo cp etcd etcdctl /usr/local/bin/
+
+# Start etcd
+etcd \
+  --name etcd-single \
+  --data-dir /tmp/etcd-data \
+  --listen-client-urls http://0.0.0.0:2379 \
+  --advertise-client-urls http://0.0.0.0:2379 \
+  --listen-peer-urls http://0.0.0.0:2380 \
+  --initial-advertise-peer-urls http://0.0.0.0:2380 \
+  --initial-cluster etcd-single=http://0.0.0.0:2380 &
+
+# Verify etcd is running
+etcdctl --endpoints "127.0.0.1:2379" put key "value"
+etcdctl --endpoints "127.0.0.1:2379" get key
+```
+
+For production environments, refer to the
+[official etcd clustering documentation](https://etcd.io/docs/current/op-guide/clustering/).
+
+## Start Datasystem Worker
+
+```bash
+# Replace ${ETCD_IP} with etcd node IP, ${WORKER_IP} with local node IP
+dscli start -w \
+  --worker_address "${WORKER_IP}:31501" \
+  --etcd_address "${ETCD_IP}:2379" \
+  --shared_memory_size_mb 20480
+```
+
+To stop the worker:
+
+```bash
+dscli stop --worker_address "${WORKER_IP}:31501"
+```
+
+## Configuration
+
+Define the connector in runtime:
+
+```yaml
+runtime:
+  connectors:
+    connector_of_yuanrong:
+      name: YuanrongConnector
+      extra:
+        host: "127.0.0.1"
+        port: 31501
+        get_sub_timeout_ms: 1000
+```
+
+Wire stages to the connector:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    output_connectors:
+      to_stage_1: connector_of_yuanrong
+
+  - stage_id: 1
+    input_connectors:
+      from_stage_0: connector_of_yuanrong
+```
+
+Parameters:
+
+- host: datasystem worker host.
+- port: datasystem worker port.
+- get_sub_timeout_ms: get timeout in milliseconds (0 for no timeout).
+
+For more details, refer to the
+[Yuanrong Datasystem repository](https://atomgit.com/openeuler/yuanrong-datasystem).
@@ -0,0 +1,59 @@
+# Distributed utils
+
+This directory (vllm_omni/distributed/ray_utils) contains utilities for distributed execution in vllm-omni, supporting both **Ray** and **Multiprocessing** backends.
+## 1. Installation
+```bash
+pip install "ray[default]"
+```
+## 2. Ray Utils
+
+The `ray_utils` module provides helper functions for managing Ray clusters and actors, which is essential for:
+*   **Multi-node deployment**: Running pipeline stages across different physical machines.
+*   **Resource management**: Efficient GPU/CPU allocation.
+
+### 2.1 Basic Usage
+
+To use the Ray backend, specify `worker_backend="ray"` when initializing the engine.
+
+**Command Line Example:**
+```bash
+vllm serve Qwen/Qwen2.5-Omni-7B \
+  --omni \
+  --port 8091 \
+  --worker-backend ray \
+  --ray-address auto
+```
+
+### 2.2 Cluster Setup
+
+**Step 1: Start Head Node**
+Run this on your primary machine:
+```bash
+ray start --head --port=6399
+```
+
+**Step 2: Connect Worker Nodes**
+Run this on each worker machine:
+```bash
+ray start --address=<HEAD_NODE_IP>:6399
+```
+
+> **Tip**: For a complete cluster setup script, refer to the vLLM example:
+> [run_cluster.sh](https://github.com/vllm-project/vllm/blob/main/examples/online_serving/run_cluster.sh)
+
+### 2.3 Distributed Connector Support
+
+When running on Ray, the system automatically adapts its communication strategy:
+
+*   **Cross-Node**: Recommended to use `MooncakeConnector` (requires separate configuration).
+*   **Same-Node**: Can still use `SharedMemoryConnector` for efficiency, or Ray's native object store (plasma).
+*   **SHM threshold default differs**: when `worker_backend="ray"`, the SharedMemoryConnector default threshold is set to `sys.maxsize`, which forces payloads to go inline (no SHM). Override `shm_threshold_bytes` in the connector config if you want SHM for Ray runs.
+
+### 2.4 Internal Helpers
+
+*   **`initialize_ray_cluster`**: Connects to an existing Ray cluster or starts a local one.
+
+## 3. Troubleshooting
+
+*   **Connection Issues**: Ensure the Ray head node is accessible and ports (default 6399 in this example) are open.
+*   **Version Mismatch**: Ensure all nodes run the same version of Ray and Python.
@@ -5,5 +5,14 @@ This section contains design documents and architecture specifications for vLLM-
 ## Architecture Documents

 - [Architecture Overview](architecture_overview.md)
- [vLLM-Omni Software Design Document](vllm_omni_design.md)
- [Multi-Request Streaming](mrs_design.md)
+
+## Feature Design Documents
+
+- [Disaggregated Inference](feature/disaggregated_inference.md)
+- [Ray-based Execution](feature/ray_based_execution.md)
+
+## Module Design Documents
+
+- [AR Module](module/ar_module.md)
+- [DIT Module](module/dit_module.md)
+- [Entrypoint Module](module/entrypoint_module.md)
@@ -0,0 +1,414 @@
+# AutoRegressive (AR) Module
+
+## 1. Overview
+
+The AutoRegressive (AR) module in vLLM-Omni handles autoregressive generation stages, primarily used for text, chain-of-thought(COT), and audio latent tokens generation stages in multi-stage models like Qwen2.5-Omni, Qwen3-Omni, BAGEL, .etc. Unlike some representative non-autoregressive generation stages (e.g., Diffusion), AR stages generate tokens sequentially, one at a time, following the standard transformer decoder pattern.
+
+The AR module of vLLM-Omni extends vLLM's core components to support:
+
+- **Multimodal inputs/outputs**: Processing images, videos, and audio alongside text
+- **Direct embedding transfer**: Passing pre-computed prompt embeddings between pipeline stages via serialized payloads
+- **Additional information flow**: Carrying per-request metadata (tensors, lists) through the pipeline
+- **Hidden state exposure**: Exposing per-request hidden representations for downstream stages
+- **Basic generator support**: Support some basic heterogeneous architecture such as Convolution, LSTM, etc.
+
+As shown in the [end2end example](../../user_guide/examples/offline_inference/qwen3_omni.md), AR module can be widely applied across multiple stages, generating text tokens in thinker(AR), audio latent tokens in talker(AR) and audio wave in code2wav(Convolution).
+
+## 2. Relationship with vLLM
+
+The AR module builds upon vLLM main framework through inheritance, extending core classes while preserving compatibility with vLLM's scheduling, batching, KV cache management, and execution mechanisms.
+
+### Inheritance Hierarchy
+- Scheduler
+
+```mermaid
+classDiagram
+    class VLLMScheduler {
+        +schedule() SchedulerOutput
+        +update_from_output() EngineCoreOutputs
+    }
+    class OmniARScheduler {
+        +schedule() SchedulerOutput
+    }
+    class OmniGenerationScheduler {
+        +schedule() SchedulerOutput
+        +update_from_output() EngineCoreOutputs
+    }
+    VLLMScheduler <|-- OmniARScheduler
+    VLLMScheduler <|-- OmniGenerationScheduler
+```
+- Worker
+
+```mermaid
+classDiagram
+    class GPUWorker {
+        +init_device()
+        +model_runner
+    }
+    class GPUARWorker {
+        +init_device()
+    }
+    class GPUGenerationWorker {
+        +init_device()
+    }
+    GPUWorker <|-- GPUARWorker
+    GPUWorker <|-- GPUGenerationWorker
+```
+- ModelRunner
+
+```mermaid
+classDiagram
+    class GPUModelRunner {
+        +execute_model()
+        +sample_tokens()
+    }
+    class OmniGPUModelRunner {
+        +_update_states()
+        +_preprocess()
+        +_model_forward()
+    }
+    class GPUARModelRunner {
+        +execute_model()
+        +sample_tokens()
+    }
+    class GPUGenerationModelRunner {
+        +execute_model()
+    }
+    GPUModelRunner <|-- OmniGPUModelRunner
+    OmniGPUModelRunner <|-- GPUARModelRunner
+    OmniGPUModelRunner <|-- GPUGenerationModelRunner
+```
+- InputProcessor/OutputProcessor
+
+```mermaid
+classDiagram
+    class InputProcessor {
+        +process_inputs() EngineCoreRequest
+    }
+    class OmniInputProcessor {
+        +process_inputs() OmniEngineCoreRequest
+    }
+
+    class VLLMOutputProcessor {
+        +process_outputs() OutputProcessorOutput
+    }
+    class MultimodalOutputProcessor {
+        +process_outputs() OutputProcessorOutput
+        +_route_and_normalize()
+    }
+    InputProcessor <|-- OmniInputProcessor
+    VLLMOutputProcessor <|-- MultimodalOutputProcessor
+```
+
+### Key Extensions
+
+- **Scheduler**: `OmniARScheduler` extends `vllm.v1.core.sched.scheduler.Scheduler` to enrich scheduled requests with omni-specific payloads
+- **Worker**: `GPUARWorker` extends `vllm.v1.worker.gpu_worker.Worker` to initialize AR-specific model runners
+- **ModelRunner**: `GPUARModelRunner` extends `OmniGPUModelRunner` → `vllm.v1.worker.gpu_model_runner.GPUModelRunner` to expose hidden states and handle multimodal outputs
+- **InputProcessor**: `OmniInputProcessor` extends `vllm.v1.engine.input_processor.InputProcessor` to serialize prompt embeddings and additional information
+- **OutputProcessor**: `MultimodalOutputProcessor` extends `vllm.v1.engine.output_processor.OutputProcessor` to route and accumulate multimodal outputs
+
+## 3. Scheduler Design
+
+The AR module provides two scheduler implementations: one for standard autoregressive generation and one for basic heterogeneous architectures.
+
+### Request Flow
+
+The following diagram illustrates the request flow through the AR module components:
+
+```mermaid
+flowchart TD
+    A[OmniInputProcessor] -->|OmniEngineCoreRequest| B[OmniARScheduler]
+    B -->|schedule: OmniNewRequestData| C[GPUARWorker]
+    C -->|SchedulerOutput| D[GPUARModelRunner]
+    D -->|execute_model: None| E[Model Forward Pass]
+    E -->|hidden_states, logits| D
+    D -->|sample_tokens: OmniModelRunnerOutput| F[OmniARScheduler]
+    F -->|update_from_output| G[MultimodalOutputProcessor]
+    G -->|RequestOutput| H[Client/Downstream Stage]
+
+    style A fill:#e1f5ff
+    style B fill:#fff4e1
+    style C fill:#e8f5e9
+    style D fill:#f3e5f5
+    style G fill:#fce4ec
+```
+
+The flow follows vLLM's standard pattern: input processing → scheduling → worker execution → output processing, with omni-specific enrichments at each stage.
+
+### OmniARScheduler
+
+`OmniARScheduler` extends the base vLLM scheduler with minimal modifications, focusing on enriching scheduled requests with omni-specific payloads.
+
+#### Modified API: `schedule()`
+
+The scheduler wraps base `NewRequestData` entries with `OmniNewRequestData` to include prompt embeddings and additional information:
+
+```python
+def schedule(self) -> SchedulerOutput:
+    scheduler_output = super().schedule()
+    # Rewrap base NewRequestData entries with OmniNewRequestData
+    new_list = []
+    for nr in scheduler_output.scheduled_new_reqs:
+        request = self.requests.get(nr.req_id)
+        omni_nr = OmniNewRequestData(
+            req_id=nr.req_id,
+            prompt_token_ids=nr.prompt_token_ids,
+            # ... other base fields ...
+            prompt_embeds=getattr(request, "prompt_embeds", None),
+            additional_information=getattr(request, "additional_information", None),
+        )
+        new_list.append(omni_nr)
+    scheduler_output.scheduled_new_reqs = new_list
+    return scheduler_output
+```
+
+The `update_from_output()` method remains unchanged, inheriting standard request lifecycle management from the base scheduler.
+
+### OmniGenerationScheduler
+
+`OmniGenerationScheduler` implements a fast-path scheduling strategy for basic heterogeneous architectures that process all input tokens in a single step.
+
+#### Modified API: `schedule()`
+
+Allocates all input tokens for a request at once (or 1 placeholder if zero), falling back to default scheduling if budget is insufficient:
+
+```python
+def schedule(self) -> SchedulerOutput:
+    # Fast path: allocate all input tokens at once
+    while self.waiting and token_budget > 0:
+        request = self.waiting.peek_request()
+        required_tokens = max(getattr(request, "num_prompt_tokens", 0), 1)
+        if required_tokens > token_budget:
+            break  # Fall back to default scheduling
+        # Allocate and schedule...
+```
+
+#### Modified API: `update_from_output()`
+
+Marks requests as finished immediately after one step, since generation models complete in a single forward pass:
+
+```python
+def update_from_output(self, ...) -> dict[int, EngineCoreOutputs]:
+    # ...
+    # Diffusion request: completes in one step
+    request.status = RequestStatus.FINISHED_STOPPED
+    kv_transfer_params = self._free_request(request)
+    # ...
+```
+
+## 4. Worker and ModelRunner Design
+
+### GPUARWorker
+
+`GPUARWorker` initializes the AR-specific model runner while maintaining standard device initialization:
+
+```python
+class GPUARWorker(GPUWorker):
+    def init_device(self):
+        # ... standard device initialization ...
+        self.model_runner = GPUARModelRunner(self.vllm_config, self.device)
+```
+
+### GPUARModelRunner
+
+`GPUARModelRunner` follows vLLM's two-phase execute/sample flow while exposing hidden states and multimodal outputs.
+
+#### Two-Phase Execution
+
+**Phase 1: `execute_model()`** - Runs forward pass and stores state:
+- Computes logits from hidden states
+- Stores `ExecuteModelState` with hidden states, logits, and multimodal outputs
+- Returns `None` to defer sampling
+
+**Phase 2: `sample_tokens()`** - Samples tokens and builds output:
+- Retrieves stored state from `execute_model()`
+- Samples tokens using logits
+- Extracts per-request hidden states and multimodal outputs
+- Builds `OmniModelRunnerOutput` with `pooler_output` containing hidden states
+
+```python
+def sample_tokens(self, grammar_output) -> OmniModelRunnerOutput:
+    # Retrieve stored state
+    hidden_states, multimodal_outputs = self.execute_model_state
+
+    # Sample tokens
+    sampler_output = self._sample(logits, spec_decode_metadata)
+
+    # Extract per-request hidden states
+    pooler_output = []
+    for rid in req_ids:
+        hidden_slice = hidden_states_cpu[start:end]
+        payload = {"hidden": hidden_slice}
+        # Add multimodal outputs if present
+        pooler_output.append(payload)
+
+    return OmniModelRunnerOutput(
+        pooler_output=pooler_output,
+        # ... other fields ...
+    )
+```
+
+### GPUGenerationModelRunner
+
+`GPUGenerationModelRunner` implements a simplified single-phase execution for basic heterogeneous architectures:
+
+- No logits computation or token sampling
+- Direct generation from forward pass in model implementation
+- Returns outputs via `pooler_output` immediately after forward pass
+
+### OmniGPUModelRunner
+
+`OmniGPUModelRunner` provides shared functionality for both AR and Generation runners:
+
+#### Prompt Embeddings Overlay
+
+During prefill, overlays custom `prompt_embeds` from request state onto `inputs_embeds`:
+
+```python
+def _collect_additional_information_for_prefill(self, num_scheduled_tokens_np):
+    for req_index, req_id in enumerate(self.input_batch.req_ids):
+        req_state = self.requests[req_id]
+        pe_cpu = getattr(req_state, "prompt_embeds_cpu", None)
+        # Overlay prompt_embeds for prefill portion
+        if pe_cpu is not None:
+            src = pe_cpu[num_computed_tokens:num_computed_tokens + overlay_len]
+            self.inputs_embeds[start_offset:start_offset + overlay_len].copy_(src)
+```
+
+#### Additional Information Processing
+
+Decodes and manages `additional_information` payloads:
+- Decodes serialized payloads → CPU tensors in request state
+- Passes runtime information to model via `runtime_additional_information` kwarg
+- Processes model-provided updates via `postprocess()` hook
+- Merges updates back into request state
+
+#### M-RoPE Position Initialization
+
+For multimodal models using M-RoPE (e.g., Qwen2-VL), computes position encodings from multimodal feature metadata (image grids, video grids, audio features).
+
+## 5. Input/Output Processing
+
+### Processing Pipeline
+
+The input/output processing pipeline handles serialization, routing, and accumulation of multimodal data:
+
+```mermaid
+sequenceDiagram
+    participant Client
+    participant OmniInputProcessor
+    participant Scheduler
+    participant ModelRunner
+    participant MultimodalOutputProcessor
+    participant Client
+
+    Client->>OmniInputProcessor: prompt + prompt_embeds + additional_info
+    OmniInputProcessor->>OmniInputProcessor: Serialize tensors to payloads
+    OmniInputProcessor->>Scheduler: OmniEngineCoreRequest (with payloads)
+    Scheduler->>ModelRunner: OmniNewRequestData (with payloads)
+    ModelRunner->>ModelRunner: Decode payloads → CPU tensors
+    ModelRunner->>ModelRunner: Overlay prompt_embeds on inputs_embeds
+    ModelRunner->>ModelRunner: Forward pass with runtime_additional_information
+    ModelRunner->>ModelRunner: Extract hidden states + multimodal outputs
+    ModelRunner->>MultimodalOutputProcessor: OmniModelRunnerOutput (pooler_output)
+    MultimodalOutputProcessor->>MultimodalOutputProcessor: Route by output_type
+    MultimodalOutputProcessor->>MultimodalOutputProcessor: Accumulate tensors in OmniRequestState
+    MultimodalOutputProcessor->>MultimodalOutputProcessor: Consolidate tensor lists
+    MultimodalOutputProcessor->>Client: RequestOutput (with multimodal_output)
+```
+
+### OmniInputProcessor
+
+`OmniInputProcessor` extends the base input processor to serialize prompt embeddings and additional information for inter-stage transfer.
+
+#### Payload Serialization
+
+Converts PyTorch tensors to serialized payloads:
+
+```python
+def process_inputs(self, ...) -> OmniEngineCoreRequest:
+    # Serialize prompt_embeds
+    if "prompt_embeds" in decoder_inputs:
+        pe_cpu = decoder_inputs["prompt_embeds"].detach().to("cpu").contiguous()
+        prompt_embeds_payload = PromptEmbedsPayload(
+            data=pe_cpu.numpy().tobytes(),
+            shape=[seq_len, hidden_size],
+            dtype=dtype_str,
+        )
+
+    # Serialize additional_information
+    if "additional_information" in decoder_inputs:
+        entries = {}
+        for key, value in raw_info.items():
+            if isinstance(value, torch.Tensor):
+                entry = AdditionalInformationEntry(
+                    tensor_data=value.numpy().tobytes(),
+                    tensor_shape=list(value.shape),
+                    tensor_dtype=dtype_str,
+                )
+            entries[key] = entry
+        additional_information_payload = AdditionalInformationPayload(entries=entries)
+
+    return OmniEngineCoreRequest(
+        # ... standard fields ...
+        prompt_embeds=prompt_embeds_payload,
+        additional_information=additional_information_payload,
+    )
+```
+
+### MultimodalOutputProcessor
+
+`MultimodalOutputProcessor` routes outputs by modality type and accumulates multimodal tensors.
+
+#### Output Routing
+
+Routes `EngineCoreOutput` by `output_type` attribute:
+- `"text"`: Standard text generation path
+- `"image"`, `"audio"`, `"latents"`: Extract from `pooling_output` or `multimodal_outputs`
+- Fallback: Heuristic based on presence of `pooling_output`
+
+#### Tensor Accumulation
+
+`OmniRequestState` accumulates multimodal tensors across multiple steps:
+
+```python
+def add_multimodal_tensor(self, payload, mm_type):
+    # Normalize payload to dict
+    incoming = {mm_type or "hidden": payload}
+
+    # Accumulate: convert tensors to lists for deferred concatenation
+    if isinstance(v, torch.Tensor) and isinstance(existing, torch.Tensor):
+        self.mm_accumulated[k] = [existing, v]  # List accumulation
+```
+
+Before final output, consolidates tensor lists via concatenation:
+
+```python
+def _consolidate_multimodal_tensors(self):
+    for k, v in self.mm_accumulated.items():
+        if isinstance(v, list) and isinstance(v[0], torch.Tensor):
+            self.mm_accumulated[k] = torch.cat(v, dim=0)  # Concatenate
+```
+
+The consolidated tensors are attached to `RequestOutput.multimodal_output` for consumption by downstream stages or clients.
+
+## 6. Summary
+
+The AR module of vLLM-Omni extends vLLM through strategic inheritance and minimal API modifications:
+
+### Key Design Patterns
+
+1. **Inheritance over composition**: Extends vLLM classes to preserve compatibility with existing scheduling, batching, and execution mechanisms
+2. **Payload serialization**: Uses `PromptEmbedsPayload` and `AdditionalInformationPayload` for efficient inter-stage data transfer
+3. **Two-phase execution**: Maintains vLLM's execute/sample separation for AR models while supporting single-phase execution for generation models
+4. **Multimodal routing**: Routes outputs by `output_type` and accumulates tensors incrementally to support streaming
+
+### Differences from vLLM
+
+- **Payload support**: Serialized prompt embeddings and additional information enable direct transfer between pipeline stages
+- **Multimodal handling**: Extended input/output processors support images, audio, and other modalities alongside text
+- **Hidden state exposure**: AR model runners expose per-request hidden states via `pooler_output` for downstream consumption
+- **Generation scheduler**: Fast-path scheduling for basic heterogeneous architectures that complete in one step
+
+The AR module seamlessly integrates with vLLM's existing infrastructure while adding the necessary extensions for multi-stage, multimodal generation pipelines.
@@ -0,0 +1,906 @@
+---
+toc_depth: 4
+---
+
+# Diffusion Module Architecture Design
+
+The vLLM-Omni diffusion module (`vllm_omni/diffusion`) is a high-performance inference engine for diffusion models, designed with a modular architecture that separates concerns across multiple components. It provides efficient execution for non-autoregressive generation tasks such as image and video generation.
+
+This document describes the architecture design of the diffusion module, including the diffusion engine, scheduler, worker, diffusion pipeline, and acceleration components.
+
+<p align="center">
+   <img src="https://github.com/user-attachments/assets/cde1ebea-4006-4d30-8283-15a5721a09d2" alt="vLLM-Omni Diffusion Module Components" width="80%">
+</p>
+<p align="center">
+  <em> Main Components of the Diffusion Module </em>
+</p>
+
+
+**Table of Content:**
+
+- [Architecture Overview](#architecture-overview)
+- [Diffusion Engine](#1-diffusion-engine)
+- [Scheduler](#2-scheduler)
+- [Worker](#3-worker)
+- [Diffusion Pipeline](#4-diffusion-pipeline)
+- [Acceleration Components](#5-acceleration-components)
+    - [Attention Backends](#51-attention-backends)
+    - [Parallel Attention](#52-parallel-attention)
+    - [Cache Backends](#53-cache-backends)
+    - [Parallel Strategies](#54-parallel-strategies)
+- [Data Flow](#6-data-flow)
+
+---
+
+## Architecture Overview
+
+The diffusion module follows a **multi-process, distributed architecture** with clear separation of concerns:
+
+<p align="center">
+  <img src="https://github.com/user-attachments/assets/78c4d446-d238-4406-a057-eb17d2ccc8e0" alt="vLLM-Omni Diffusion Module Architecture" width="100%">
+</p>
+<p align="center">
+  <em> Diffusion Architecture Overview </em>
+</p>
+
+
+---
+
+## 1. Diffusion Engine
+
+**Location**: `vllm_omni/diffusion/diffusion_engine.py`
+
+### Responsibilities
+
+The `DiffusionEngine` is the **orchestrator** of the diffusion inference system. It manages the lifecycle of worker processes and coordinates the execution flow.
+
+### Key Components
+
+#### 1.1 Initialization
+
+```python
+class DiffusionEngine:
+    def __init__(self, od_config: OmniDiffusionConfig):
+        self.od_config = od_config
+        self.post_process_func = get_diffusion_post_process_func(od_config)
+        self.pre_process_func = get_diffusion_pre_process_func(od_config)
+        self._processes: list[mp.Process] = []
+        self._make_client()
+```
+
+**Key Features**:
+
+- **Pre/Post Processing**: Registers model-specific pre-processing and post-processing functions via registry pattern
+
+- **Worker Management**: Launches and manages multiple worker processes (one per GPU)
+
+- **Process Isolation**: Uses multiprocessing for true parallelism
+
+#### 1.2 Worker Launch Process
+
+The engine launches workers using a **spawn** method:
+
+```python
+def _launch_workers(self, broadcast_handle):
+    # Creates one process per GPU
+    for i in range(num_gpus):
+        process = mp.Process(
+            target=worker_proc.worker_main,
+            args=(i, od_config, writer, broadcast_handle),
+            name=f"DiffusionWorker-{i}",
+        )
+        process.start()
+```
+
+**Design Decisions**:
+
+- **Spawn Method**: Ensures clean state for each worker (no shared memory issues)
+
+- **Pipe Communication**: Uses `mp.Pipe` for initialization handshake
+
+- **Device Selection**: Each worker is assigned a specific GPU (`cuda:{rank}`)
+
+#### 1.3 Request Processing Flow
+
+```python
+def step(self, requests: list[OmniDiffusionRequest]):
+    # 1. Pre-process requests
+    requests = self.pre_process_func(requests)
+
+    # 2. Send to scheduler and wait for response
+    output = self.add_req_and_wait_for_response(requests)
+
+    # 3. Post-process results
+    result = self.post_process_func(output.output)
+    return result
+```
+
+**Flow**:
+
+1. **Pre-processing**: Applies model-specific transformations
+
+2. **Scheduling**: Delegates to scheduler for distribution
+
+3. **Post-processing**: Converts raw outputs to final format (e.g., PIL images)
+
+---
+
+## 2. Scheduler
+
+**Location**: `vllm_omni/diffusion/scheduler.py`
+
+### Architecture
+
+The `Scheduler` is implemented as a **Singleton** pattern to ensure a single coordination point across the system, i.e., only one scheduler instance exists for coordination.
+
+### Key Components
+
+#### 2.1 Message Queue System
+
+```python
+class Scheduler:
+    def initialize(self, od_config: OmniDiffusionConfig):
+        # Broadcast queue: scheduler -> all workers
+        self.mq = MessageQueue(
+            n_reader=od_config.num_gpus,
+            n_local_reader=od_config.num_gpus,
+            local_reader_ranks=list(range(od_config.num_gpus)),
+        )
+
+        # Result queue: rank 0 worker -> scheduler
+        self.result_mq = None  # Initialized later
+```
+
+**Communication Pattern**:
+
+- **Broadcast Queue**: One-to-many communication (scheduler → all workers)
+
+- **Result Queue**: One-to-one communication (rank 0 → scheduler)
+
+- **Shared Memory**: Uses `MessageQueue` (ZMQ-based) for efficient IPC
+
+#### 2.2 Request Distribution
+
+```python
+def add_req(self, requests: list[OmniDiffusionRequest]) -> DiffusionOutput:
+    # Broadcast request to all workers
+    self.mq.enqueue(requests)
+
+    # Wait for result from Rank 0
+    output = self.result_mq.dequeue()
+    return output
+```
+
+**Design Features**:
+
+- **Broadcast Model**: All workers receive the same request (for tensor parallelism)
+
+- **Single Response**: Only rank 0 sends results back (avoids duplicate outputs)
+
+- **Synchronous**: Blocks until result is received (can be made async)
+
+#### 2.3 Singleton Pattern
+
+```python
+class Scheduler:
+    _instance = None
+
+    def __new__(cls, *args, **kwargs):
+        if not cls._instance:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+# Global singleton instance
+scheduler = Scheduler()
+```
+
+**Benefits**:
+
+- **Single Point of Control**: Ensures consistent state
+
+- **Easy Access**: Global `scheduler` instance accessible everywhere
+
+- **Resource Management**: Centralized queue management
+
+---
+
+## 3. Worker
+
+**Location**: `vllm_omni/diffusion/worker/gpu_worker.py`
+
+### Architecture
+
+Workers are **independent processes** that execute the actual model inference. Each worker runs on a dedicated GPU and participates in distributed inference.
+
+### Key Components
+
+#### 3.1 Worker Process Structure
+
+```python
+class WorkerProc:
+    def __init__(self, od_config, gpu_id, broadcast_handle):
+        # Initialize ZMQ context for IPC
+        self.context = zmq.Context(io_threads=2)
+
+        # Connect to broadcast queue (receive requests)
+        self.mq = MessageQueue.create_from_handle(broadcast_handle, gpu_id)
+
+        # Create result queue (only rank 0)
+        if gpu_id == 0:
+            self.result_mq = MessageQueue(n_reader=1, ...)
+
+        # Initialize GPU worker
+        self.worker = GPUWorker(local_rank=gpu_id, rank=gpu_id, od_config=od_config)
+```
+
+**Initialization Steps**:
+
+1. **IPC Setup**: Creates ZMQ context and message queues
+
+2. **Distributed Environment Setup**: Initializes PyTorch distributed communication
+
+    - For CUDA GPUs: Uses NCCL (fast GPU communication)
+
+    - For NPU: Uses HCCL (Huawei Collective Communications Library)
+
+    - For other devices: Uses appropriate backend (GLOO, MCCL, etc.)
+
+3. **Model Loading**: Loads diffusion pipeline on assigned GPU
+
+4. **Cache Setup**: Enables cache backend if configured.
+
+#### 3.2 GPU Worker
+
+```python
+class GPUWorker:
+    def init_device_and_model(self):
+        # Set distributed environment variables
+        os.environ["RANK"] = str(rank)
+        os.environ["WORLD_SIZE"] = str(world_size)
+
+        # Initialize PyTorch distributed
+        init_distributed_environment(world_size, rank)
+        parallel_config = self.od_config.parallel_config
+        initialize_model_parallel(
+            data_parallel_size=parallel_config.data_parallel_size,
+            cfg_parallel_size=parallel_config.cfg_parallel_size,
+            sequence_parallel_size=parallel_config.sequence_parallel_size,
+            tensor_parallel_size=parallel_config.tensor_parallel_size,
+            pipeline_parallel_size=parallel_config.pipeline_parallel_size,
+        )
+
+        # Load model
+        model_loader = DiffusersPipelineLoader(load_config)
+        self.pipeline = model_loader.load_model(od_config, load_device=f"cuda:{rank}")
+
+        # Setup cache backend
+        from vllm_omni.diffusion.cache.selector import get_cache_backend
+        self.cache_backend = get_cache_backend(od_config.cache_backend, od_config.cache_config)
+
+        if self.cache_backend is not None:
+            self.cache_backend.enable(self.pipeline)
+```
+
+**Key Features**:
+
+- **Tensor Parallelism**: Supports multi-GPU tensor parallelism via PyTorch distributed
+
+- **Model Loading**: Uses `DiffusersPipelineLoader` for efficient weight loading
+
+- **Cache Integration**: Enables cache backends (TeaCache, cache-dit, etc.) transparently
+
+#### 3.3 Worker Busy Loop
+
+```python
+def worker_busy_loop(self):
+    while self._running:
+        # 1. Receive unified message (generation request, RPC request, or shutdown)
+        msg = self.recv_message()
+
+        # 2. Route message based on type
+        if isinstance(msg, dict) and msg.get("type") == "rpc":
+            # Handle RPC request
+            result, should_reply = self.execute_rpc(msg)
+            if should_reply:
+                self.return_result(result)
+
+        elif isinstance(msg, dict) and msg.get("type") == "shutdown":
+            # Handle shutdown message
+            self._running = False
+
+        else:
+            # Handle generation request (OmniDiffusionRequest list)
+            output = self.worker.execute_model(msg, self.od_config)
+            self.return_result(output)
+```
+
+**Execution Flow**:
+
+1. **Receive**: Dequeues unified messages from shared memory queue
+
+2. **Route**: Handles different message types (generation, RPC, shutdown)
+
+3. **Execute**: Runs forward pass through pipeline for generation requests
+
+4. **Respond**: Sends results back (rank 0 for generation, specified rank for RPC)
+
+#### 3.4 Model Execution
+
+```python
+@torch.inference_mode()
+def execute_model(self, reqs: list[OmniDiffusionRequest], od_config):
+    req = reqs[0]  # TODO: support batching
+
+    # Refresh cache backend if enabled
+    if self.cache_backend is not None and self.cache_backend.is_enabled():
+        self.cache_backend.refresh(self.pipeline, req.num_inference_steps)
+
+    # Set forward context for parallelism
+    with set_forward_context(
+        vllm_config=self.vllm_config,
+        omni_diffusion_config=self.od_config
+    ):
+        output = self.pipeline.forward(req)
+    return output
+```
+
+The model execution leverages multiple parallelism strategies that are transparently applied during the forward pass. The `set_forward_context()` context manager makes parallel group information available throughout the forward pass:
+
+```python
+# Inside transformer layers, parallel groups are accessed via:
+from vllm_omni.diffusion.distributed.parallel_state import (
+    get_sp_group, get_dp_group, get_cfg_group, get_pp_group
+)
+```
+
+**Optimizations**:
+
+- **Cache Refresh**: Clears cache state before each generation for clean state
+
+- **Context Management**: Forward context ensures parallel groups are available during execution
+
+- **Single Request**: Currently processes one request at a time (batching TODO)
+
+---
+
+## 4. Diffusion Pipeline
+
+**Location**: `vllm_omni/diffusion/models/*/pipeline_*.py`
+
+The pipeline is the **model-specific implementation** that orchestrates the diffusion process. Different models (QwenImage, Wan2.2, Z-Image) have their own pipeline implementations.
+
+Most pipeline implementation are referred from `diffusers`. The multi-step diffusion loop is usually the most time-consuming part during the overall inference process, which is defined by the `diffuse` function in the pipeline class. An example is as follows:
+
+```python
+def diffuse(self, ...):
+    for i, t in enumerate(timesteps):
+        # Forward pass for positive prompt
+        transformer_kwargs = {
+            "hidden_states": latents,
+            "timestep": timestep / 1000,
+            "encoder_hidden_states": prompt_embeds,
+        }
+        noise_pred = self.transformer(**transformer_kwargs)[0]
+
+        # Forward pass for negative prompt (CFG)
+        if do_true_cfg:
+            neg_transformer_kwargs = {...}
+            neg_transformer_kwargs["cache_branch"] = "negative"
+            neg_noise_pred = self.transformer(**neg_transformer_kwargs)[0]
+
+            # Combine predictions
+            comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
+            noise_pred = comb_pred * (cond_norm / noise_norm)
+
+        # Scheduler step
+        latents = self.scheduler.step(noise_pred, t, latents)[0]
+
+    return latents
+```
+
+**Key Features**:
+
+- **CFG Support**: Handles classifier-free guidance with separate forward passes
+
+- **Cache Branching**: Uses `cache_branch` parameter for cache-aware execution
+
+- **True CFG**: Implements advanced CFG with norm preservation
+
+To learn more about the diffusion pipeline and how to add a new diffusion pipeline, please view [Adding Diffusion Model](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/model/adding_diffusion_model)
+
+---
+
+## 5. Acceleration Components
+
+### 5.1 Attention Backends
+
+**Location**: `vllm_omni/diffusion/attention/`
+
+#### Architecture
+
+The attention system uses a **backend selector pattern** that automatically chooses the optimal attention implementation based on hardware and model configuration.
+
+#### Backend Selection
+
+**Location**: `vllm_omni/diffusion/attention/selector.py`
+
+```python
+class Attention(nn.Module):
+    def __init__(self, num_heads, head_size, causal, softmax_scale, ...):
+        # Auto-select backend
+        self.attn_backend = get_attn_backend(-1)
+        self.attn_impl_cls = self.attn_backend.get_impl_cls()
+        self.attention = self.attn_impl_cls(...)
+```
+
+**Available Backends**:
+
+- **FlashAttention**: Optimized CUDA kernel (FA2/FA3) - memory efficient via tiling
+
+- **SDPA**: PyTorch's scaled dot-product attention - default, cross-platform
+
+- **SageAttention**: Sparse attention implementation from SageAttention library
+
+- **AscendAttention**: NPU-optimized attention for Ascend hardware
+
+These backends provide the **kernel implementations** for attention computation. For attention-level sequence parallelism strategies (Ring Attention, Ulysses), see [Parallel Attention](#52-parallel-attention).
+
+#### Backend Selection Mechanism
+
+```python
+def get_attn_backend(head_size: int) -> type[AttentionBackend]:
+    # Check environment variable
+    backend_name = os.environ.get("DIFFUSION_ATTENTION_BACKEND")
+
+    if backend_name:
+        return load_backend(backend_name.upper())
+
+    # Default to SDPA
+    return SDPABackend
+```
+
+**Selection Priority**:
+
+1. **Environment Variable**: `DIFFUSION_ATTENTION_BACKEND` for manual override
+
+    - Valid values: `FLASH_ATTN`, `TORCH_SDPA`, `SAGE_ATTN`, `ASCEND`
+
+    - Example: `export DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN`
+
+2. **Automatic Fallback**: Falls back to SDPA if selected backend unavailable
+
+3. **Hardware Detection**: Can select based on device type (NPU, CUDA, etc.)
+
+**Backend Availability**:
+
+- **SDPA**: Always available (PyTorch built-in)
+
+- **FlashAttention**: Requires `flash-attn` package installed
+
+- **SageAttention**: Requires `sage-attention` package (from THU-ML GitHub)
+
+- **AscendAttention**: Only available on Ascend NPU hardware
+
+#### Attention Backend Registry
+
+**Location**: `vllm_omni/diffusion/attention/selector.py`
+
+The attention system uses a **registry pattern** to manage and dynamically load attention backends. This allows for easy extension and runtime selection of backends.
+
+
+**Registry Structure**:
+
+```python
+# Registry mapping backend names to their module paths and class names
+_BACKEND_CONFIG = {
+    "FLASH_ATTN": {
+        "module": "vllm_omni.diffusion.attention.backends.flash_attn",
+        "class": "FlashAttentionBackend",
+    },
+    "TORCH_SDPA": {
+        "module": "vllm_omni.diffusion.attention.backends.sdpa",
+        "class": "SDPABackend",
+    },
+    "SAGE_ATTN": {
+        "module": "vllm_omni.diffusion.attention.backends.sage_attn",
+        "class": "SageAttentionBackend",
+    },
+    "ASCEND": {
+        "module": "vllm_omni.diffusion.attention.backends.ascend_attn",
+        "class": "AscendAttentionBackend",
+    },
+}
+```
+
+#### Attention Backend Integration
+
+The `Attention` layer integrates backends through a unified interface. Here's how **FlashAttentionBackend** is integrated as an example:
+
+```python
+# attention/backends/flash_attn.py
+
+class FlashAttentionBackend(AttentionBackend):
+    @staticmethod
+    def get_name() -> str:
+        return "FLASH_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> type["FlashAttentionImpl"]:
+        return FlashAttentionImpl
+
+    @staticmethod
+    def get_supported_head_sizes() -> list[int]:
+        return [64, 96, 128, 192, 256]  # FlashAttention supports these head sizes
+
+
+class FlashAttentionImpl(AttentionImpl):
+    def __init__(self, num_heads, head_size, softmax_scale, causal, ...):
+        self.num_heads = num_heads
+        self.causal = causal
+        self.softmax_scale = softmax_scale
+
+    def forward(self, query, key, value, attn_metadata=None):
+        # Call FlashAttention kernel
+        out = flash_attn_func(
+            query, key, value,
+            causal=self.causal,
+            softmax_scale=self.softmax_scale,
+        )
+        return out
+```
+
+---
+
+### 5.2 Parallel Attention
+
+**Location**: `vllm_omni/diffusion/attention/parallel/`
+
+#### Architecture
+
+Parallel attention strategies implement **Sequence Parallelism (SP) at the attention layer level**. These strategies distribute attention computation across multiple GPUs by splitting the sequence dimension, using different communication patterns. They work **on top of** AttentionBackend implementations (FlashAttention, SDPA, etc.), handling the parallelization/communication while the backends handle the actual attention computation.
+
+**Key Distinction**: Unlike AttentionBackend (which provides kernel implementations), ParallelAttentionStrategy provides communication patterns for multi-GPU attention parallelism. These strategies implement the `ParallelAttentionStrategy` interface and use AttentionBackend implementations internally.
+
+Both Ring Attention and Ulysses are forms of Sequence Parallelism (SP) that:
+
+- Split the sequence dimension across GPUs
+
+- Contribute to `sequence_parallel_size` (via `ring_degree` and `ulysses_degree`)
+
+- Work at the attention layer level (not model/pipeline level)
+
+#### Ulysses Sequence Parallelism (USP)
+
+**Location**: `vllm_omni/diffusion/attention/parallel/ulysses.py`
+
+USP is a sequence-parallel attention strategy that splits attention computation across multiple GPUs by distributing both the sequence dimension and attention heads. It uses **all-to-all communication** to efficiently parallelize attention for very long sequences. Specifically, it uses **all-to-all** collective operations to redistribute Q/K/V tensors before attention computation and gather results afterward.
+
+Ulysses splits attention computation in two dimensions:
+
+1. **Sequence Dimension**: Splits the sequence length across GPUs
+
+2. **Head Dimension**: Splits attention heads across GPUs
+
+**Configuration**: `ulysses_degree` contributes to `sequence_parallel_size`
+
+#### Ring Sequence Parallelism
+
+**Location**: `vllm_omni/diffusion/attention/parallel/ring.py`
+
+Ring Attention is a **parallel attention strategy** that implements sequence parallelism using ring-based point-to-point (P2P) communication. Unlike attention backends that provide the attention kernel implementation, Ring Attention is a **communication pattern** that works on top of attention backends (FlashAttention or SDPA).
+
+Ring Attention splits sequence dimension across GPUs in a ring topology, implemented via the `ParallelAttentionStrategy` interface, instead of `AttentionBackend`. P2P ring communication is applied to circulate Key/Value blocks across GPUs. Internally, `ring_flash_attn_func` or `ring_pytorch_attn_func` is used depending on available backends.
+
+**Architecture**:
+```python
+class RingParallelAttention:
+    """Ring sequence-parallel strategy."""
+
+    def run_attention(self, query, key, value, attn_metadata, ...):
+        # Selects underlying attention kernel (FlashAttention or SDPA)
+        if backend_pref == "sdpa":
+            return ring_pytorch_attn_func(...)  # Uses SDPA kernel
+        else:
+            return ring_flash_attn_func(...)    # Uses FlashAttention kernel
+```
+
+**Integration**:
+
+- Ring Attention is activated when `ring_degree > 1` in parallel config
+
+- It's selected by `build_parallel_attention_strategy()` in the attention layer
+
+- The `Attention` layer routes to `_run_ring_attention()` when Ring is enabled
+
+- Works alongside attention backends: Ring handles communication, backends handle computation
+
+**Configuration**: `ring_degree` contributes to `sequence_parallel_size`
+
+#### Relationship with AttentionBackend
+
+Parallel attention strategies (Ring, Ulysses) work **on top of** AttentionBackend implementations:
+
+- They use AttentionBackend for the actual attention computation (FlashAttention, SDPA, etc.)
+
+- They handle the multi-GPU communication/parallelization layer
+
+- They implement `ParallelAttentionStrategy` interface (not `AttentionBackend`)
+
+For general parallelism strategies (Data Parallelism, Tensor Parallelism, Pipeline Parallelism), see [Parallel Strategies](#54-parallel-strategies).
+
+---
+
+### 5.3 Cache Backends
+
+**Location**: `vllm_omni/diffusion/cache/`
+
+#### Architecture
+
+Cache backends provide a **unified interface** for applying different caching strategies to accelerate diffusion inference. The system supports multiple backends (TeaCache, cache-dit) with a consistent API for enabling and refreshing cache state.
+
+#### Cache Backend Interface
+
+```python
+class CacheBackend(ABC):
+    def __init__(self, config: DiffusionCacheConfig):
+        self.config = config
+        self.enabled = False
+
+    @abstractmethod
+    def enable(self, pipeline: Any) -> None:
+        """Enable cache on the pipeline."""
+        raise NotImplementedError
+
+    @abstractmethod
+    def refresh(self, pipeline: Any, num_inference_steps: int, verbose: bool = True) -> None:
+        """Refresh cache state for new generation."""
+        raise NotImplementedError
+
+    def is_enabled(self) -> bool:
+        """Check if cache is enabled."""
+        return self.enabled
+```
+
+**Design Pattern**:
+
+- **Abstract Base Class**: Defines contract for all cache backends
+
+- **Pipeline-based**: Works with pipeline instances (not just transformers)
+
+- **State Management**: Provides refresh mechanism for clean state between generations
+
+#### Available Backends
+
+**1. TeaCache Backend**
+
+**Location**: `vllm_omni/diffusion/cache/teacache/backend.py`
+
+```python
+class TeaCacheBackend(CacheBackend):
+    def enable(self, pipeline: Any):
+        # Extract transformer from pipeline
+        transformer = pipeline.transformer
+        transformer_type = transformer.__class__.__name__
+
+        # Create TeaCacheConfig from DiffusionCacheConfig
+        teacache_config = TeaCacheConfig(
+            transformer_type=transformer_type,
+            rel_l1_thresh=self.config.rel_l1_thresh,
+            coefficients=self.config.coefficients,
+        )
+
+        # Apply hooks to transformer
+        apply_teacache_hook(transformer, teacache_config)
+        self.enabled = True
+
+    def refresh(self, pipeline: Any, num_inference_steps: int, verbose: bool = True):
+        transformer = pipeline.transformer
+        if hasattr(transformer, "_hook_registry"):
+            transformer._hook_registry.reset_hook(TeaCacheHook._HOOK_NAME)
+```
+
+**TeaCache Features**:
+
+- **Timestep-aware**: Caches based on timestep embedding similarity
+
+- **Adaptive**: Dynamically decides when to reuse cached computations
+
+- **CFG-aware**: Handles positive/negative branches separately
+
+- **Custom Hook System**: Uses a custom forward interception mechanism (via `HookRegistry`) that wraps the module's `forward` method, allowing transparent integration without modifying model code
+
+**2. Cache-DiT Backend**
+
+**Location**: `vllm_omni/diffusion/cache/cache_dit_backend.py`
+
+```python
+class CacheDiTBackend(CacheBackend):
+    def enable(self, pipeline: Any):
+        # Uses cache-dit library for acceleration
+        # Supports DBCache, SCM (Step Computation Masking), TaylorSeer
+        # Works with single and dual-transformer architectures
+        ...
+        self.enabled = True
+
+    def refresh(self, pipeline: Any, num_inference_steps: int, verbose: bool = True):
+        # Updates cache context with new num_inference_steps
+        ...
+```
+
+**Cache-DiT Features**:
+
+- **DBCache**: Dynamic block caching with configurable compute blocks
+
+- **SCM**: Step Computation Masking for additional speedup
+
+- **TaylorSeer**: Advanced calibration for cache accuracy
+
+- **Dual-transformer Support**: Handles models like Wan2.2 with two transformers
+
+#### Cache Backend Selector
+
+**Location**: `vllm_omni/diffusion/cache/selector.py`
+
+```python
+def get_cache_backend(
+    cache_backend: str | None,
+    cache_config: dict | DiffusionCacheConfig
+) -> CacheBackend | None:
+    """Get cache backend instance based on cache_backend string.
+
+    Args:
+        cache_backend: Cache backend name ("cache_dit", "tea_cache", or None)
+        cache_config: Cache configuration (dict or DiffusionCacheConfig)
+
+    Returns:
+        Cache backend instance or None if cache_backend is "none"
+    """
+    if cache_backend is None or cache_backend == "none":
+        return None
+
+    if isinstance(cache_config, dict):
+        cache_config = DiffusionCacheConfig.from_dict(cache_config)
+
+    if cache_backend == "cache_dit":
+        return CacheDiTBackend(cache_config)
+    elif cache_backend == "tea_cache":
+        return TeaCacheBackend(cache_config)
+    else:
+        raise ValueError(f"Unsupported cache backend: {cache_backend}")
+```
+
+**Usage Flow**:
+
+1. **Selection**: `get_cache_backend()` returns appropriate backend instance
+
+2. **Enable**: `backend.enable(pipeline)` called during worker initialization
+
+3. **Refresh**: `backend.refresh(pipeline, num_inference_steps)` called before each generation
+
+4. **Check**: `backend.is_enabled()` verifies cache is active
+
+### 5.4 Parallel Strategies
+
+**Location**: `vllm_omni/diffusion/distributed/parallel_state.py`
+
+#### Parallelism Types
+
+The system supports multiple orthogonal parallelism strategies:
+
+**Sequence Parallelism (SP)**
+
+- **Purpose**: Split sequence dimension across GPUs
+
+- **Attention-level SP**: Ring Attention and Ulysses (USP) implement SP at the attention layer level
+
+    - See [Parallel Attention](#52-parallel-attention) for details
+
+    - Configuration: `ulysses_degree` × `ring_degree` = `sequence_parallel_size`
+
+- **Use Case**: Very long sequences (e.g., high-resolution images)
+
+**Data Parallelism (DP)**
+
+- **Purpose**: Replicate model across GPUs, split batch
+
+- **Use Case**: Batch processing, throughput optimization
+
+**Tensor Parallelism (TP)** (Experimental)
+
+- **Purpose**: Split model weights across GPUs
+
+- **Implementation**: Uses vLLM's tensor parallel groups
+
+- **Use Case**: Large models that don't fit on single GPU
+
+**CFG Parallelism**  (under development)
+
+- **Purpose**: Parallelize Classifier-Free Guidance (positive/negative prompts)
+
+- **Infrastructure**: CFG parallel groups are initialized and available via `get_cfg_group()`
+
+#### Parallel Group Management
+
+```python
+def initialize_model_parallel(
+    data_parallel_size: int = 1,
+    cfg_parallel_size: int = 1,
+    sequence_parallel_size: int | None = None,
+    ulysses_degree: int = 1,
+    ring_degree: int = 1,
+    tensor_parallel_size: int = 1,
+    pipeline_parallel_size: int = 1,
+    vae_parallel_size: int = 0,
+):
+    # Generate orthogonal parallel groups
+    rank_generator = RankGenerator(
+        tensor_parallel_size,
+        sequence_parallel_size,
+        pipeline_parallel_size,
+        cfg_parallel_size,
+        data_parallel_size,
+        "tp-sp-pp-cfg-dp",
+    )
+
+    # Initialize each parallel group
+    _DP = init_model_parallel_group(rank_generator.get_ranks("dp"), ...)
+    _CFG = init_model_parallel_group(rank_generator.get_ranks("cfg"), ...)
+    _SP = init_model_parallel_group(rank_generator.get_ranks("sp"), ...)
+    _PP = init_model_parallel_group(rank_generator.get_ranks("pp"), ...)
+    _TP = init_model_parallel_group(rank_generator.get_ranks("tp"), ...)
+```
+
+**Rank Order**: `tp-sp-pp-cfg-dp` (tensor → sequence → pipeline → cfg → data)
+
+**Note**: For attention-level Sequence Parallelism implementations (Ring Attention and Ulysses), see [Parallel Attention](#52-parallel-attention). This section covers higher-level parallelism strategies.
+
+
+---
+
+## 6. Data Flow
+
+### Complete Request Flow
+
+<p align="center">
+   <img src="https://github.com/user-attachments/assets/6e093a0d-29c0-4efd-85d5-747002bd2fed" alt="vLLM-Omni Diffusion Module Components" width="100%">
+</p>
+<p align="center">
+  <em> End-to-end Data Flow in the vLLM-Omni Diffusion Module </em>
+</p>
+
+
+```
+1. User Request
+   └─> OmniDiffusion.generate(prompt)
+       └─> Prepare OmniDiffusionRequest
+           └─> DiffusionEngine.step(requests)
+
+2. Pre-processing
+   └─> pre_process_func(requests)
+       └─> Model-specific transformations
+
+3. Scheduling
+   └─> scheduler.add_req(requests)
+       └─> Broadcast via MessageQueue to all workers
+
+4. Worker Execution
+   └─> WorkerProc.worker_busy_loop()
+       └─> GPUWorker.execute_model(reqs)
+           └─> Pipeline.forward(req)
+               ├─> encode_prompt()
+               ├─> prepare_latents()
+               ├─> diffuse() [loop]
+               │   ├─> transformer.forward() [with cache backend hooks]
+               │   └─> scheduler.step()
+               └─> vae.decode()
+
+5. Result Collection
+   └─> Rank 0 sends DiffusionOutput via result queue
+       └─> Scheduler receives and returns
+
+6. Post-processing
+   └─> post_process_func(output)
+       └─> Convert to PIL images / final format
+```
+
+---
@@ -0,0 +1 @@
+Architecture design of the entrypoint (update soon)
@@ -1,188 +0,0 @@
-## Multi-Request Streaming (MRS) on a Single Machine
-
-### 1. Background & Scope
- All processing runs on a single physical machine with multi-process, per-stage workers. No proxy or network transport involved.
- Current alignment with vllm-omni: `OmniLLM` supports multiple stages (`OmniStage`). GPU runners already expose streamable steps (prefill/decoding/diffusion), but the entry layer still collects lists and lacks intra-stage streaming and window scheduling.
- Goal: implement multi-stage, multi-request streaming (MRS) locally. Each stage outputs segments; downstream stages stitch and trigger compute based on configured windows. Shared memory and zero-copy strategies reduce data movement overhead.
-
-### 2. Key Constraints
- Multi-process per stage: each stage is an independent process with a while loop; device visibility can be configured (`CUDA_VISIBLE_DEVICES`/`torch.cuda.set_device`).
- Simple IPC (copy-based): use `multiprocessing.Queue`/Pipe for inter-process communication with CPU copies/serialization; do not rely on CUDA IPC/SHM zero-copy in this version.
- Cross-stage pipeline: different stages can process different requests concurrently (e.g., stage A handles request 1 while stage B handles request 0).
-
-### 3. Architecture Overview
- Processes & IPC queues
-  - Each "sub-stage" is an OS process (worker). The loop: take from input_queue → compute → put to output_queue.
-  - Inter-stage connection via IPC: copy-based `multiprocessing.Queue` passing dict payloads; use shared memory for large objects.
-  - Each link is SPSC (single-producer/single-consumer): the upstream is the orchestrator and the downstream is a single stage process; queues are unbounded (maxsize=0) on the orchestrator side.
- Device visibility
-  - Each stage sets `CUDA_VISIBLE_DEVICES` or calls `torch.cuda.set_device` to bind to GPU sets.
-  - A stage may use multiple GPUs internally (TP/PP/DP) but presents as a single stage unit.
- Simplified IPC: copy-based queues/pipes for data transfer; zero-copy is future work.
- Pipeline progression: when a stage finishes a request, it enqueues outputs to the downstream stage; if downstream is idle, it starts immediately.
- Scheduling
-  - A downstream stage triggers only after the upstream completes the request.
-  - Windowed segmentation/stitched triggering is not implemented; intra-stage streaming is not provided.
-
-### 7. IPC Implementation (simplified: copy-based)
- Use `multiprocessing.Queue`/Pipe for inter-process communication (control + data).
- Data is serialized/copied via CPU; no CUDA IPC/SHM zero-copy in this version.
- Backpressure: queues are unbounded; pressure manifests as compute-rate differences. Optional SHM reduces large-object transfer cost; RX/decoding overhead is recorded for observability.
-
-### 8. Scheduling & Cancellation (simplified)
- Pipeline: when a stage finishes a request, it enqueues to the next stage; that stage immediately pulls the next request from its input queue, enabling cross-stage concurrency.
- Cancellation/timeout: explicit cancellation/timeouts are not provided; graceful shutdown uses a `None` sentinel sent to each stage input queue.
-
-#### Short sequence example (req0/req1, stage A→B)
-1) t0: stage A handles req0
-2) t1: req0 completes on A → enters B; A immediately starts req1
-3) t2: B handles req0 while A handles req1 (parallel across stages)
-
-### 9. Integration Points (by file)
- `vllm_omni/entrypoints/omni_llm.py` (Orchestrator)
-  - Class `OmniLLM` orchestrates multi-process stages; constructs `OmniStage` instances in parallel and spawns per-stage workers.
-  - Spawns stage processes per config (set `CUDA_VISIBLE_DEVICES`/`torch.cuda.set_device`), creates control/data channels, builds simple full-trigger flow.
-  - Stats/logging are disabled by default; per-stage and orchestrator stats are only written when explicitly enabled.
-  - Manages process lifecycle: start/wait for readiness, graceful shutdown; forwards results between stages using copy-based IPC and optional SHM.
-  - Stage readiness: each stage emits `{"type": "stage_ready"}` after initialization; the orchestrator waits for all stages or times out and logs diagnostic suggestions.
-
-### 9.1 Process Device Visibility
- Device binding: set `CUDA_VISIBLE_DEVICES` before process start, or call `torch.cuda.set_device` early in init; configs can be single/multi-GPU.
- Cross-device transfer (simplified): via CPU copies; zero-copy paths are out of scope for now.
-
-### 10. Worker Template (pseudo-code)
-```python
-def stage_worker(input_q, output_q, runtime, shm_threshold):
-    # Device binding
-    set_stage_gpu_devices(runtime.devices)
-    engine = OmniStageLLM(model, **engine_args)
-    batch_seq = 0
-    while True:
-        first = input_q.get()
-        if first is None:
-            break
-        # Batch intake (up to runtime.max_batch_size)
-        batch = [first]
-        while len(batch) < int(runtime.max_batch_size or 1) and not input_q.empty():
-            nxt = input_q.get()
-            if nxt is None:
-                input_q.put(None)
-                break
-            batch.append(nxt)
-        # Decode IPC payload and sampling params
-        request_ids = []
-        engine_inputs = []
-        rx_bytes = {}
-        rx_decode_ms = {}
-        for t in batch:
-            ein, rx = maybe_load_from_ipc_with_metrics(t, "engine_inputs", "engine_inputs_shm")
-            request_ids.append(t["request_id"])
-            engine_inputs.extend(ein if isinstance(ein, list) else [ein])
-            rx_bytes[t["request_id"]] = rx["rx_transfer_bytes"]
-            rx_decode_ms[t["request_id"]] = rx["rx_decode_time_ms"]
-        # Generate and dispatch (grouped by request_id)
-        batch_seq += 1
-        outputs = list(engine.generate(engine_inputs, batch[0]["sampling_params"], use_tqdm=False))
-        grouped = group_by_request_id(outputs, request_ids)
-        for rid in request_ids:
-            r_out = grouped.get(rid, [])
-            use_shm, payload = maybe_dump_to_shm(r_out, shm_threshold)
-            msg = {
-                "request_id": rid,
-                "metrics": {
-                    "batch_id": batch_seq,
-                    "rx_transfer_bytes": int(rx_bytes.get(rid, 0)),
-                    "rx_decode_time_ms": float(rx_decode_ms.get(rid, 0.0)),
-                    # omitted: generation latency and token stats
-                },
-            }
-            if use_shm:
-                msg["engine_outputs_shm"] = payload
-            else:
-                msg["engine_outputs"] = payload
-            output_q.put(msg)
-```
-
-### 12. Observability
- Metrics (as implemented):
-  - per-request (emitted by stages): `num_tokens_out`, `stage_gen_time_ms`, `rx_transfer_bytes`, `rx_decode_time_ms`, `rx_in_flight_time_ms`, `batch_id`
-  - orchestrator aggregates: E2E latency, tokens/s (written only when stats are enabled)
-  - optional per-stage JSONL: `{log_file}.stage{stage_id}.stats.jsonl`
- Logs/tracing:
-  - optional per-stage log files: `{log_file}.stage{stage_id}.log`
-  - the orchestrator can log readiness, forward size/time, and summary information
-
-### 15. Risks & Mitigations
-— End-to-end latency: full-trigger sacrifices interactivity; start simple/correct, add windowed streaming later.
-— CPU copy overhead: copy-based IPC for maintainability; add SHM/zero-copy later as an optimization path.
-
-### 16. Refactor: Sink LLM init and process into Stage (diagram)
-
-Objective: without changing external call patterns, encapsulate per-stage LLM init, process creation, and worker logic inside `Stage`, so `PipelinedOmniLLM` focuses on orchestration (seeding, polling, forwarding, collecting) while keeping multi-process, shared memory, and device mapping capabilities.
-
-Key changes (vs. current):
- Before: Orchestrator created processes directly, passing `_stage_worker`; device setup, LLM init, batching, and SHM were implemented there.
- After: enhanced `OmniStage` owns its subprocess (`init_stage_worker`), with `_stage_worker` as the entry; device setup, LLM init, batching, and SHM live in `OmniStage`.
-
-Classes & responsibilities
- OmniLLM (orchestrator)
-  - Build `OmniStage` list in parallel (preserve `process_engine_inputs` wiring)
-  - Connect adjacent `Stage` input/output queues; seed requests to stage 0
-  - Poll stage outputs, decode results, call `process_engine_inputs`, then encode and forward to next stage
-  - Termination & cleanup: distribute shutdown signals; join/terminate subprocesses
- Stage (stage unit)
-  - Members: `stage_config` (with `engine_args` and `runtime`), `in_q/out_q`, subprocess handle, stats
-  - API:
-    - `init_stage_worker(...)`: spawn subprocess
-    - `stop_stage_worker()`: graceful exit
-    - `submit(ipc_payload)`: submit to input queue (may use shared memory)
-    - `try_collect() -> Optional[payload]`: non-blocking get from output queue
-    - `process_engine_inputs(stages, prompts)`: reuse/delegate existing logic
-  - Subprocess entry `worker_main()`:
-    - `set_stage_gpu_devices` (`CUDA_VISIBLE_DEVICES` & logical index mapping)
-    - build `OmniStageLLM(model, **engine_args)`
-    - while-loop:
-      - take first task + up to `runtime.max_batch_size-1` non-blocking; support `max_batch_size>=1`
-      - `maybe_load_from_ipc_with_metrics` per `engine_inputs` in batch
-      - `stage_engine.generate(batched_inputs, sampling_params)` (window `-1`)
-      - write per-request aggregated outputs to output queue after `maybe_dump_to_shm`
-      - log start/finish; report `error` per request on exceptions
-
-Data flow (sketch)
-```
-OmniLLM
-  ├─ Stage[0].start_process(in0, out0)
-  ├─ Stage[1].start_process(in1, out1)
-  ├─ ...
-  └─ Orchestrator loop:
-       - stage0.submit(encode(engine_inputs))
-       - for each stage:
-           res = stage.try_collect()
-           outputs = decode(res)
-           stage.set_engine_outputs(outputs)
-           if !final: next_inputs = next_stage.process_engine_inputs(stages, original_prompt)
-                    next_stage.submit(encode(next_inputs))
-```
-
-Config (preserve & extend)
- `runtime.devices`:
-  - Comma-separated visibility list (e.g., "2,5,7"): set visible set; default logical index 0 (first GPU)
-  - Number/string: treat as logical index (0/1-based per implementation) mapping to physical GPUs
- `runtime.max_batch_size`: max requests taken per intake (default 1)
- `shm_threshold_bytes`: objects larger than this go via shared memory (default 64KB; injected into stages)
- Stage link and inputs:
-  - `engine_input_source`: list of upstream stage_id(s) from which the downstream stage takes inputs (typically a single source)
-  - `custom_process_input_func`: function to construct downstream engine_inputs from upstream outputs
- Output flags:
-  - `final_output`, `final_output_type`: marks that a stage produces the final external output and its type
-
-Kept & strengthened
- Multi-process: each stage in its own process; parallel across stages; serial/batching inside a stage controlled by `max_batch_size`
- Shared memory: large objects via SHM with strict `memoryview -> close -> unlink`
- Device mapping: unified `set_stage_gpu_devices`, support comma lists and logical indices
-
-Migration (execution order)
-1) Enhanced `OmniStage` with `init_stage_worker/stop_stage_worker/submit/try_collect/_stage_worker`.
-2) `OmniLLM` holds `OmniStage` instances, replacing the older pipelined class; uses Stage APIs.
-3) Reuse `entrypoints/stage_utils`: device/SHM/encode-decode utilities live here.
-4) Regression: single/multi-request, various `max_batch_size`, SHM threshold, device mapping, multi-stage parallelism, error paths, and shutdown flows.
@@ -1,509 +0,0 @@
-# Architecture Design Principles
-
-## Overview
-
-vLLM-Omni is a multi-modality extension for vLLM that supports non-autoregressive structures and non-textual outputs. This document outlines the key software abstractions, APIs, and dependencies for the system, designed to maximize reuse of vLLM's proven architecture.
-
-## Architecture Principles
-
-1. **vLLM V1 Compatibility**: Built on vLLM's Engine V1 architecture with AsyncLLM and EngineCore patterns
-2. **Stage-based Processing**: Models are divided into multiple stages, each processed by different Engine Cores
-3. **Multiple Engine Core Support**: Each stage can use either AR Engine Core (reusing vLLM) or Diffusion Engine Core (new DiT support), or other new Cores
-4. **Worker Process Pattern**: Follows vLLM's multiprocess worker architecture for scalability
-5. **Extensibility**: Easy integration of new modalities, model architectures, and output formats
-
-## Key Data Flow
-
-API Server --> OmniLLM/AsyncOmni (New, including multi engines) --> LLMEngine/AsyncLLM --> Engine Core
- --> Scheduler (New one for DiT) --> Executor (New one for diffusers) --> Worker (New one for DiT)
- --> ModelRunner (New one for AR hiddenstate, New one for DiT) --> RequestState --> OutputProcessoer (New one for final multimodal output)
-
-## Core Components (Ordered by Data Flow)
-
-### 1. Installation
-```bash
-pip install vllm
-pip install vllm-omni
-```
-
-### 2. Online inference launch (Entry Point)
-
-Keep consistency with the vllm main branch
-```bash
-vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8000
-```
-
-#### Design Architecture
-```mermaid
-graph TD
-    A[vllm serve model --omni] --> B[vLLM-Omni CLI Wrapper]
-    B --> C{Detect --omni flag}
-    C -->|Yes| D[Parse OmniConfig]
-    C -->|No| E[Forward to vLLM CLI]
-    D --> F[Initialize AsyncOmni]
-    F --> G[Start omni Server]
-    G --> H[Multi-stage Processing]
-    E --> I[Standard vLLM Pipeline]
-
-    subgraph "vLLM-Omni Components"
-        F
-        G
-        H
-    end
-
-    subgraph "vLLM Components"
-        I
-    end
-```
-
-#### omni Serve Command Implementation
-File: vllm_omni/entrypoints/cli/main.py
-```python
-import sys
-import argparse
-from typing import List, Optional
-from vllm_omni.entrypoints.omni import OmniServeCommand
-
-def main():
-    """Main CLI entry point that intercepts vLLM commands"""
-    # Check if --omni flag is present
-    if "--omni" in sys.argv:
-        # Remove --omni flag and process with vLLM-Omni
-        omni_args = [arg for arg in sys.argv[1:] if arg != "--omni"]
-        omni_serve = OmniServeCommand()
-        omni_serve.run(omni_args)
-    else:
-        # Forward to original vLLM CLI
-        from vllm.entrypoints.cli.main import main as vllm_main
-        vllm_main()
-
-if __name__ == "__main__":
-    main()
-```
-
-#### vLLM Plugin Integration
-File: vllm_omni/pyproject.toml
-```python
-[project.scripts]
-# Override vLLM's CLI entry point
-vllm = "vllm_omni.cli:main"
-
-# Add entry point for vLLM integration
-[project.entry-points."vllm.plugins"]
-omni = "vllm_omni.plugin:OmniPlugin"
-```
-
-### 3. Offline inference launch (Entry Point)
-Design an upper level class to incorporate multi Engines, each engine has a engine call.
-```python
-from vllm.entrypoints.llm import LLM
-from vllm.v1.engine.llm_engine import LLMEngine
-
-class OmniLLM(LLM):
-    """Extended LLM supporting multiple engines and stage-based processing"""
-
-    def __init__(self, stage_configs: List[StageConfig]):
-        super().__init__()
-        self.stage_configs = stage_configs
-        self.engine_list = []  # List of AsyncLLM instances for each stage
-        self.output_processor = MultimodalOutputProcessor()
-        self._initialize_stage_engines()
-
-    def _initialize_stage_engines(self) -> None:
-        """Initialize LLMEngine instances for each stage"""
-        for stage_config in self.stage_configs:
-            stage_llm = LLMEngine.from_vllm_config(
-                    vllm_config=stage_config.vllm_config,
-                    executor_class=self.executor_class,
-                    log_stats=self.log_stats
-                )
-            self.engine_list.append(stage_llm)
-
-    def generate(
-        self,
-        stage_args_list: List[stage_args],
-        use_tqdm: Union[bool, Callable[..., tqdm]] = True,
-        lora_request: Optional[Union[list[LoRARequest], LoRARequest]] = None,
-        priority: Optional[list[int]] = None,
-    ) -> list[RequestOutput]:
-        """Main generation interface - orchestrates multi-stage processing"""
-
-        # Process through each stage sequentially
-        for i, stage_config in enumerate(self.stage_configs):
-            stage_engine = self.engine_list[i]
-
-            # Prepare input for this stage
-            stage_args = stage_args_list[i]
-            prompt_str, engine_request, tokenization_kwargs = self._process_stage_inputs(stage_config, **stage_args)
-
-            # Add inputs to Engine
-            stage_engine.add_request(requesy_id, prompt_str, tokenization_kwargs)
-            # Run Engine
-            stage_output = stage_engine.step()
-
-            # Update stage input and output of each stage for later usage
-            self._update_stage_io(stage_output, stage_config)
-
-
-        # Process final output
-        final_output = self.output_processor.process_output(
-            self.stage_configs[-1].stage_output
-        )
-        return final_output
-
-    def _process_stage_inputs(stage_config, **stage_args) -> Any:
-        """Prepare input for specific stage"""
-        if stage_config.engine_type == "AR":
-            return self._process_ar_inputs(**stage_args)
-        elif stage_config.engine_type == "DiT":
-            return self._process_dit_inputs(**stage_args)
-        else
-            raise NotImplementedError
-
-    def _process_dit_inputs(**stage_args)-> Any:
-        image_latent = self.vae.encode(**stage_args)
-
-        return image_latent
-```
-
-### 4. Stage config for setting of different model stages
-```python
-@dataclass
-class StageConfig:
-    """Configuration for a processing stage"""
-    stage_id: int
-    engine_type: str  # "AR" or "DiT"
-    model_path: str
-    input_modalities: List[str]
-    output_modalities: List[str]
-    vllm_config: Optional[VllmConfig] = None  # For engine config of corresponding stage
-    executor_class: type[Executor]  # For execute class config of corresponding stage
-    dit_config: Optional[DiTConfig] = None  # For diffusion stages
-    cache_config: Optional[DiTCacheConfig] = None
-```
-For AR stage, the setting is:
-```python
-from vllm.v1.core.sched.scheduler import Scheduler
-from vllm.v1.executor.multiproc_executor import MultiprocExecutor
-ar_stage_config = StageConfig()
-ar_stage_config.vllm_config.scheduler_config.scheduler_cls = Scheduler # original vllm scheduler for AR
-ar_stage_config.executor_class = MultiprocExecutor # original vllm executor for AR
-```
-For DiT stage, the setting is:
-```python
-from vllm.v1.executor.multiproc_executor import MultiprocExecutor
-dit_stage_config = StageConfig()
-dit_stage_config.vllm_config.scheduler_config.scheduler_cls = DiffusionScheduler # New scheduler for DiT
-
-# For diffusion models without using diffusers
-dit_stage_config.executor_class = MultiprocExecutor
-
-# For diffusion models using diffusers
-dit_stage_config.executor_class = DiffusersPipelineExecutor
-```
-
-### 5. Online inference main class
-
-Similar to OmniLLM in offline inference, add some asynchronous processing, referring to AsyncLLM
-```python
-from vllm.v1.engine.async_llm import AsyncLLM
-
-
-class AsyncOmni(AsyncLLM):
-    """Extended AsyncLLM supporting multiple engines and stage-based processing"""
-
-    def __init__(self, stage_configs: List[StageConfig]):
-        super().__init__()
-```
-
-### 6. Engine Core
-No need to change. The specific executor, scheduler and output type can be transferred into it with new configs.
-```python
-# class: EngineCore
-# EngineCore.step (simplified)
-class EngineCore:
-    def step(self):
-        scheduler_output = scheduler.schedule()
-        model_output = executor.execute_model(scheduler_output)
-        engine_outputs = scheduler.update_from_output(
-            scheduler_output, model_output
-        )
-        return engine_outputs
-```
-
-### 7. Scheduler
-Create a new Diffusion Scheduler for DiT, which is inherited from original vllm AR scheduler. At first, no complicated strategy.
-Just process the request first in first out. Then create a class of DiT Cache Manager for optimization in the future.
-```python
-from vllm.v1.core.sched.scheduler import Scheduler
-class DiffusionScheduler(Scheduler):
-
-    def __init__(
-        self,
-        vllm_config: VllmConfig,
-        kv_cache_config: KVCacheConfig,
-        dit_cache_config: DiTCacheConfig,
-        structured_output_manager: StructuredOutputManager,
-        mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
-        include_finished_set: bool = False,
-        log_stats: bool = False,
-    ) -> None:
-
-    self.dit_cache_manager = DiTCacheManager(dit_cache_config)
-```
-For DiT Cache Manager (Can refer to xDiT):
-```python
-class DiTCacheManager:
-    """Manages DiT-specific caching"""
-
-    def __init__(self, config: DiTCacheConfig):
-        self.cache_tensors: Dict[str, torch.Tensor] = {}
-        self.cache_groups: List[DiTCacheTensor] = config.dit_cache_tensors
-
-    def allocate_cache(self, request_id: str, size: int) -> torch.Tensor
-    def get_cache(self, request_id: str) -> Optional[torch.Tensor]
-    def release_cache(self, request_id: str) -> None
-    def clear_expired_cache(self) -> None
-```
-
-### 8. Executor for DiT without diffusers
-
-No need to change. Just import original AR executor.
-
-### 9. Diffusers Pipeline Executor (no worker)
-
-A single-process executor that directly runs the Diffusers pipeline without spawning workers or using RPC. Interfaces remain identical to `Executor` so the EngineCore loop is unchanged.
-
-#### Function map（Pipeline Executor）
-
-##### 1) Inherited and overridden
-```python
-from concurrent.futures import Future
-from typing import Optional, Union, Callable, Any
-import torch
-import torch.nn as nn
-from vllm.v1.executor.abstract import Executor
-from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheSpec
-from vllm.v1.outputs import ModelRunnerOutput
-from vllm.tasks import SupportedTask
-from diffusers import DiffusionPipeline
-
-class DiffusersPipelineExecutor(Executor):
-    supports_pp: bool = False  # Single-process, no TP/PP/DP
-
-    def _init_executor(self) -> None:
-        # Called by ExecutorBase.__init__
-        self._failure_callback: Optional[Callable[[], None]] = None
-        self._device = self._resolve_device()
-        self._dtype = self._resolve_dtype()
-        self._pipeline = self._build_pipeline(device=self._device, dtype=self._dtype)
-        self._profiler = None
-        self._is_failed = False
-        self.is_sleeping = False
-        self.sleeping_tags: set[str] = set()
-
-    # major functions to build/run diffusers pipeline
-    def _build_pipeline(self, device:torch.device, dtype:torch.dtype)->DiffusionPipeline:
-        model_name = "Qwen/Qwen-Image"
-
-        self.pipe = DiffusionPipeline.from_pretrained(model_name, torch_dtype=dtype)
-        self.pipe = pipe.to(device)
-
-    def _run_pipeline(self, scheduler_output) -> ModelRunnerOutput: ...
-        positive_magic = {
-            "en": ", Ultra HD, 4K, cinematic composition.", # for english prompt
-            "zh": ", 超清，4K，电影级构图." # for chinese prompt
-        }
-
-        # Generate image
-        prompt_embeds = self._get_and_process_prompt_embeds(scheduler_output, positive_magic)
-        negative_prompt_embeds = self.pipe.embed_prompt(" ")
-
-        width, height = aspect_ratios["16:9"]
-
-        image = pipe(
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            width=width,
-            height=height,
-            num_inference_steps=50,
-            true_cfg_scale=4.0,
-            generator=torch.Generator(device="cuda").manual_seed(42)
-        ).images[0]
-
-        output = self.wrap_image_as_ModelRunnerOutput(image)
-        return output
-
-    # ---- Internal helpers (implementation-specific, not public API) ----
-    def _resolve_device(self): ...
-    def _resolve_dtype(self): ...
-    def _get_model(self) -> nn.Module: ...
-    def _get_and_process_prompt_embeds(self, scheduler_output, positive_magic):
-        ...
-        #append the positive_magic to prompt and embed them to prompt embed tensors
-
-    # Functions related to workers should either raise NotImplementedError
-    # or return empty defaults in this no-worker executor.
-
-    def collective_rpc(self, method, timeout=None, args=(), kwargs=None) -> list[Any]:
-        # No workers in pipeline executor
-        raise NotImplementedError("No workers in DiffusersPipelineExecutor")
-
-    def initialize_from_config(self, kv_cache_configs: list[KVCacheConfig]) -> None:
-        return  # no-op (pipeline already built in _init_executor)
-
-    def register_failure_callback(self, callback):
-        self._failure_callback = callback
-
-    def determine_available_memory(self) -> list[int]:  # bytes
-        # Single device; return [available_bytes]. If CPU-only, return [0].
-        return [self._determine_available_bytes(self._device)]
-
-```
-
-### 10. Worker
-
-#### Inheritance strategy
-Prefer reusing the mature GPU Worker end-to-end. Worker class is selected by configuration (vllm_config). Do not add a new executor-specific worker binding. If customization is needed, override only `init_device` to construct the `DiffusionModelRunner`; all other behaviors (device init details, profiling, sleep/wake, PP/TP comms, execute path) remain from `vllm/v1/worker/gpu_worker.py::Worker`.
-
-
-Inherited and overridden
-```python
-# Optional: only if you need to plug a custom DiffusionModelRunner.
-from vllm.v1.worker.gpu_worker import Worker as GPUWorker
-from vllm.v1.worker.diffusion_model_runner import DiffusionModelRunner
-
-class Worker(GPUWorker):
-    def init_device(self) -> None:
-        #those related to device check and init
-        ...
-
-        self.model_runner = DiffusionModelRunner(self.vllm_config, ...)
-```
-
-### 11. Model Runner
-
-### Function map (Model Runner)
-#### 1) Inherited and overridden
-Those parts relied to the KV Cache will be omitted if we do not register the model to the vllm config. The engine core will view it as do not require KV Cache, and handle it properly
-
-Reuse `vllm/v1/outputs.py::ModelRunnerOutput`：
- DiffusionModelRunner: Use the `pooler_output=[Tensor,...]` to return multi modal tensors
- ARModelRunner: Use the `pooler_output=[Tensor,...]` to return hidden states.
-```python
-from typing import Optional, Union
-import torch
-from vllm.v1.worker.gpu_model_runner import GPUModelRunner
-from vllm.v1.outputs import ModelRunnerOutput
-
-
-class DiffusionModelRunner(GPUModelRunner):
-    @torch.inference_mode()
-    def execute_model(
-        self,
-        scheduler_output: "SchedulerOutput",
-        intermediate_tensors: Optional[IntermediateTensors] = None,
-    ) -> Union[ModelRunnerOutput, IntermediateTensors]:
-        ...
-        return ModelRunnerOutput(
-            req_ids=[...],
-            req_id_to_index={...},
-            sampled_token_ids=[],
-            spec_token_ids=None,
-            logprobs=None,
-            prompt_logprobs_dict={},
-            pooler_output=[Tensor,...],         # Return Hidden states
-            kv_connector_output=None,
-            num_nans_in_logits=None,
-        )# return multi modal tensors via pooler_output=[Tensor,...]
-
-
-class ARModelRunner(GPUModelRunner):
-    @torch.inference_mode()
-    def execute_model(
-        self,
-        scheduler_output: "SchedulerOutput",
-        intermediate_tensors: Optional[IntermediateTensors] = None,
-    ) -> Union[ModelRunnerOutput, IntermediateTensors]:
-        ...
-        return ModelRunnerOutput(
-            req_ids=[...],
-            req_id_to_index={...},
-            sampled_token_ids=[],
-            spec_token_ids=None,
-            logprobs=None,
-            prompt_logprobs_dict={},
-            pooler_output=[Tensor,...],         # Return Hidden states
-            kv_connector_output=None,
-            num_nans_in_logits=None,
-        )# return hidden states via pooler_output=[Tensor,...]
-```
-
-### 12. RequestState Role and Usage
-```bash
-# RequestState serves as the per-request state tracker in OutputProcessor:
-# - Maintains request-specific state (tokens, logprobs, detokenizer, etc.)
-# - Converts EngineCoreOutput → RequestOutput/PoolingRequestOutput
-# - Manages request lifecycle from registration to completion
-AsyncLLM.add_request()
-↓
-OutputProcessor.add_request()
-↓
-RequestState.from_new_request() → create request state
-
-AsyncLLM.__init__() / AsyncLLM.generate() / AsyncLLM.encode() →  create a background loop that continuously pulls output from EngineCore
-↓
-OutputProcessor.process_outputs() → update state and process outputs
-↓
-RequestState.make_request_output() → convert to the final output, formatted as RequestOutput or PoolingRequestOutput
-↓
-RequestOutputCollector.put() → push to the queue (AsyncLLM)
-```
-
-Need to add implementation for an existing method
-```python
-class RequestState:
-    def __init__(self, request_id: str, parent_req: Optional[ParentRequest],
-                 request_index: int, lora_name: Optional[str],
-                 output_kind: RequestOutputKind, prompt: Optional[str],
-                 prompt_token_ids: list[int], logprobs_processor: Optional[LogprobsProcessor],
-                 detokenizer: Optional[IncrementalDetokenizer],
-                 max_tokens_param: Optional[int], arrival_time: float,
-                 queue: Optional[RequestOutputCollector], log_stats: bool):
-
-    def _new_pooling_output(self, pooling_output: torch.Tensor) -> PoolingOutput:
-        """Create PoolingOutput for multimodal/pooling requests"""
-        return PoolingOutput(data=pooling_output)
-```
-
-
-### 13. OutputProcessor
-
-For hidden state output, original OutputProcessor already support.
-
-We only need to add new one for final multimodal output.
-```python
-class MultimodalOutputProcessor(OutputProcessor):
-    """Handles multimodal output processing"""
-
-    def __init__(self):
-        self.output_handlers: Dict[str, OutputProcessor] = {}
-
-    def process_outputs(self, engine_core_outputs: list[EngineCoreOutput], ...):
-        for engine_core_output in engine_core_outputs:
-            # Option 1: Use output_type field (if EngineCoreOutput is extended)
-            if engine_core_output.output_type == "image":
-                self._process_image_output(engine_core_output)
-            elif engine_core_output.output_type == "text+image":
-                self._process_text_image_output(engine_core_output)
-            elif engine_core_output.output_type == "latents":
-                self._process_latents_output(engine_core_output)
-            elif engine_core_output.output_type == "text":
-                self._process_text_output(engine_core_output)
-            else:
-                # Fallback: use existing pooling_output logic
-                if engine_core_output.pooling_output is not None:
-                    self._process_pooling_output(engine_core_output)
-                else:
-                    self._process_text_output(engine_core_output)
-```
@@ -0,0 +1,39 @@
+# Sleep Mode
+
+vLLM-Omni’s **Sleep Mode** allows you to temporarily release most GPU memory used by a model—such as model weights and key-value (KV) caches (for autoregressive models)—**without stopping the server or unloading the Docker container**.
+
+This feature is inherited from [vLLM’s Sleep Mode](https://blog.vllm.ai/2025/10/26/sleep-mode.html), which provides zero-reload model switching for multi-model serving.  
+
+It is especially useful in **RLHF**, **training**, or **cost-saving scenarios**, where GPU resources must be freed between inference workloads.
+
+---
+
+## Omni Model
+
+Omni model inherit the feature from vLLM' Sleep Mode
+
+This means:
+
+- Support both Level 1 and Level 2 sleep, allow to release and reset both model weights and KV Cache
+
+## Diffusion Model Extension
+
+We added Sleep Mode support for **diffusion models**, which previously lacked this functionality.  
+In diffusion pipelines, this currently only offloads **model weight memory**, as these models typically do not use KV caches.
+
+This means:
+
+- Diffusion models can now enter Level 1 sleep.
+- Pipeline states (e.g., noise schedulers, buffers) remain intact after waking.
+- Useful for releasing VRAM between image generation or training cycles.
+
+---
+
+## Enable sleep mode
+To enable sleep mode, set the `enable_sleep_mode` in `engine_args` to `True`
+
+
+Example:
+```python
+omni = Omni(model=...,enable_sleep_mode=True)
+```
@@ -1,7 +1,8 @@
 # Installation

-vLLM supports the following hardware platforms:
+vLLM-Omni supports the following hardware platforms:

 - [GPU](gpu.md)
    - [NVIDIA CUDA](gpu.md)
+    - [AMD ROCm](gpu.md)
 - [NPU](npu.md)
@@ -14,6 +14,10 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr

    --8<-- "docs/getting_started/installation/gpu/cuda.inc.md:requirements"

+=== "AMD ROCm"
+
+    --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:requirements"
+
 ## Set up using Python

 ### Create a new Python environment
@@ -26,6 +30,11 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr

    --8<-- "docs/getting_started/installation/gpu/cuda.inc.md:pre-built-wheels"

+
+=== "AMD ROCm"
+
+    --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:pre-built-wheels"
+
 [](){ #build-from-source }

 ### Build wheel from source
@@ -34,7 +43,24 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr

    --8<-- "docs/getting_started/installation/gpu/cuda.inc.md:build-wheel-from-source"

+=== "AMD ROCm"
+
+    --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:build-wheel-from-source"
+
 ## Set up using Docker

 ### Pre-built images
-To be released... Stay tuned!
+
+=== "NVIDIA CUDA"
+
+    --8<-- "docs/getting_started/installation/gpu/cuda.inc.md:pre-built-images"
+
+=== "AMD ROCm"
+
+    --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:pre-built-images"
+
+### Build your own docker image
+
+=== "AMD ROCm"
+
+    --8<-- "docs/getting_started/installation/gpu/rocm.inc.md:build-docker"
@@ -17,10 +17,12 @@ Therefore, it is recommended to install vLLM and vLLM-Omni with a **fresh new**
 # --8<-- [start:pre-built-wheels]

 #### Installation of vLLM
+Note: Pre-built wheels are currently only available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1. For the latest version, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).

-vLLM-Omni is built based on vLLM v0.11.0. Please install it with command below.
+
+vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.11.0 --torch-backend=auto
+uv pip install vllm==0.14.0 --torch-backend=auto
 ```

 #### Installation of vLLM-Omni
@@ -34,17 +36,19 @@ uv pip install vllm-omni
 # --8<-- [start:build-wheel-from-source]

 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.11.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.14.0 release version of the library

 ```bash
-uv pip install vllm==0.11.0 --torch-backend=auto
+uv pip install vllm==0.14.0 --torch-backend=auto
 ```

+The release 0.14.0 of vLLM is based on PyTorch 2.9.0 which requires CUDA 12.9 environment.
+
 #### Installation of vLLM-Omni
-Install additional requirements for vLLM-Omni
+Since vllm-omni is rapidly evolving, it's recommended to install it from source
 ```bash
 git clone https://github.com/vllm-project/vllm-omni.git
-cd vllm_omni
+cd vllm-omni
 uv pip install -e .
 ```

@@ -54,11 +58,11 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.11.0
+git checkout v0.14.0
 ```
 Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
 ```bash
-export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.11.0/vllm-0.11.0-cp38-abi3-manylinux1_x86_64.whl
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.14.0/vllm-0.14.0-cp38-abi3-manylinux_2_31_x86_64.whl
 ```
 Install vllm with command below (If you have no existing PyTorch).
 ```bash
@@ -73,3 +77,27 @@ uv pip install --no-build-isolation --editable .
 </details>

 # --8<-- [end:build-wheel-from-source]
+
+# --8<-- [start:build-wheel-from-source-in-docker]
+
+# --8<-- [end:build-wheel-from-source-in-docker]
+
+# --8<-- [start:pre-built-images]
+
+vLLM-Omni offers an official docker image for deployment. These images are built on top of vLLM docker images and available on Docker Hub as [vllm/vllm-omni](https://hub.docker.com/r/vllm/vllm-omni/tags). The version of vLLM-Omni indicates which release of vLLM it is based on.
+
+Here's an example deployment command that has been verified on 2 x H100's:
+```bash
+docker run --runtime nvidia --gpus 2 \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    --env "HF_TOKEN=$HF_TOKEN" \
+    -p 8091:8091 \
+    --ipc=host \
+    vllm/vllm-omni:v0.14.0rc1 \
+    --model Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8091
+```
+
+!!! tip
+    You can use this docker image to serve models the same way you would with in vLLM! To do so, make sure you overwrite the default entrypoint (`vllm serve --omni`) which works only for models supported in the vLLM-Omni project.
+
+# --8<-- [end:pre-built-images]
@@ -0,0 +1,105 @@
+# --8<-- [start:requirements]
+
+- GPU: Validated on gfx942 (It should be supported on the AMD GPUs that are supported by vLLM.)
+
+# --8<-- [end:requirements]
+# --8<-- [start:set-up-using-python]
+
+vLLM-Omni current recommends the steps in under setup through Docker Images.
+
+# --8<-- [start:pre-built-wheels]
+
+# --8<-- [end:pre-built-wheels]
+
+# --8<-- [start:build-wheel-from-source]
+
+# --8<-- [end:build-wheel-from-source]
+
+# --8<-- [start:build-docker]
+
+#### Build docker image
+
+```bash
+DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.rocm -t vllm-omni-rocm .
+```
+
+#### Launch the docker image
+
+##### Launch with OpenAI API Server
+
+```
+docker run --rm \
+--group-add=video \
+--ipc=host \
+--cap-add=SYS_PTRACE \
+--security-opt seccomp=unconfined \
+--device /dev/kfd \
+--device /dev/dri \
+-v ~/.cache/huggingface:/root/.cache/huggingface \
+--env "HF_TOKEN=$HF_TOKEN" \
+-p 8091:8091 \
+--ipc=host \
+vllm-omni-rocm \
+--model Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8091
+```
+
+##### Launch with interactive session for development
+
+```
+docker run --rm -it \
+--network=host \
+--group-add=video \
+--ipc=host \
+--cap-add=SYS_PTRACE \
+--security-opt seccomp=unconfined \
+--device /dev/kfd \
+--device /dev/dri \
+-v <path/to/model>:/app/model \
+-v ~/.cache/huggingface:/root/.cache/huggingface \
+--entrypoint bash \
+vllm-omni-rocm
+```
+
+# --8<-- [end:build-docker]
+
+# --8<-- [start:pre-built-images]
+
+vLLM-Omni offers an official docker image for deployment. These images are built on top of vLLM docker images and available on Docker Hub as [vllm/vllm-omni-rocm](https://hub.docker.com/r/vllm/vllm-omni-rocm/tags). The version of vLLM-Omni indicates which release of vLLM it is based on.
+
+#### Launch vLLM-Omni Server
+Here's an example deployment command that has been verified on 2 x MI300's:
+```bash
+docker run --rm \
+  --group-add=video \
+  --ipc=host \
+  --cap-add=SYS_PTRACE \
+  --security-opt seccomp=unconfined \
+  --device /dev/kfd \
+  --device /dev/dri \
+  -v <path/to/model>:/app/model \
+  -v ~/.cache/huggingface:/root/.cache/huggingface \
+  --env "HF_TOKEN=$HF_TOKEN" \
+  -p 8091:8091 \
+  vllm/vllm-omni-rocm:v0.14.0rc1 \
+  --model Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
+```
+
+#### Launch an interactive terminal with prebuilt docker image.
+If you want to run in dev environment you can launch the docker image as follows:
+```bash
+docker run --rm -it \
+  --network=host \
+  --group-add=video \
+  --ipc=host \
+  --cap-add=SYS_PTRACE \
+  --security-opt seccomp=unconfined \
+  --device /dev/kfd \
+  --device /dev/dri \
+  -v <path/to/model>:/app/model \
+  -v ~/.cache/huggingface:/root/.cache/huggingface \
+  --env "HF_TOKEN=$HF_TOKEN" \
+  --entrypoint bash \
+  vllm/vllm-omni-rocm:v0.14.0rc1
+```
+
+# --8<-- [end:pre-built-images]
@@ -13,10 +13,10 @@ export DEVICE0=/dev/davinci0
 export DEVICE1=/dev/davinci1
 # Update the vllm-ascend image
 # Atlas A2:
-# export IMAGE=quay.io/ascend/vllm-ascend:v0.11.0rc2
+# export IMAGE=quay.io/ascend/vllm-ascend:v0.14.0
 # Atlas A3:
-# export IMAGE=quay.io/ascend/vllm-ascend:v0.11.0rc2-a3
-export IMAGE=quay.io/ascend/vllm-ascend:v0.11.0rc2
+# export IMAGE=quay.io/ascend/vllm-ascend:v0.14.0-a3
+export IMAGE=quay.io/ascend/vllm-ascend:v0.14.0
 docker run --rm \
    --name vllm-omni-npu \
    --shm-size=1g \
@@ -34,12 +34,22 @@ docker run --rm \
    -p 8000:8000 \
    -it $IMAGE bash

+# Install the missing dependency of mooncake in the origin image.
+apt update
+apt install libjemalloc2
+echo "export LD_PRELOAD=/usr/lib/$(uname -m)-linux-gnu/libjemalloc.so.2:$LD_PRELOAD" >> ~/.bashrc
+source ~/.bashrc
+
 # Inside the container, install vLLM-Omni from source
 cd /vllm-workspace
-git clone https://github.com/vllm-project/vllm-omni.git
+git clone -b v0.14.0 https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
 pip install -v -e .
 export VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+# (Optional) Disable mooncake for stable capability
+mv /usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake \
+   /usr/local/Ascend/ascend-toolkit/latest/python/site-packages/mooncake.disabled
 ```

 The default workdir is `/workspace`, with vLLM, vLLM-Ascend and vLLM-Omni code placed in `/vllm-workspace` installed in development mode.
@@ -12,12 +12,105 @@ This guide will help you quickly get started with vLLM-Omni to perform:

 ## Installation

-Please refer to [installation](installation/README.md)
+For installation on GPU from source:
+
+```bash
+uv venv --python 3.12 --seed
+source .venv/bin/activate
+
+# On CUDA
+uv pip install vllm==0.14.0 --torch-backend=auto
+
+# On ROCm
+uv pip install vllm==0.14.0 --extra-index-url https://wheels.vllm.ai/rocm/0.14.0/rocm700
+
+git clone https://github.com/vllm-project/vllm-omni.git
+cd vllm-omni
+uv pip install -e .
+```
+
+For additional installation methods — please see the [installation guide](installation/README.md).

 ## Offline Inference

-Please refer to [offline inference](../user_guide/examples/offline_inference/qwen2_5_omni.md)
+Text-to-image generation quickstart with vLLM-Omni:
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+
+if __name__ == "__main__":
+    omni = Omni(model="Tongyi-MAI/Z-Image-Turbo")
+    prompt = "a cup of coffee on the table"
+    outputs = omni.generate(prompt)
+    images = outputs[0].request_output[0].images
+    images[0].save("coffee.png")
+```
+
+You can pass a list of prompts and wait for them to process altogether, shown below.
+
+!!! info
+
+    However, it is not currently recommended to do so
+    because not all models support batch inference,
+    and batch requesting mostly does not provide significant performance improvement (despite the impression that it does).
+    This feature is primarily for the sake of interface compatibility with vLLM and to allow for future improvements.
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+
+if __name__ == "__main__":
+    omni = Omni(
+        model="Tongyi-MAI/Z-Image-Turbo",
+        # stage_configs_path="./stage-config.yaml",  # See below
+    )
+    prompts = [
+        "a cup of coffee on a table",
+        "a toy dinosaur on a sandy beach",
+        "a fox waking up in bed and yawning",
+    ]
+    omni_outputs = omni.generate(prompts)
+    for i_prompt, prompt_output in enumerate(omni_outputs):
+        this_request_output = prompt_output.request_output[0]
+        this_images = this_request_output.images
+        for i_image, image in enumerate(this_images):
+            image.save(f"p{i_prompt}-img{i_image}.jpg")
+            print("saved to", f"p{i_prompt}-img{i_image}.jpg")
+            # saved to p0-img0.jpg
+            # saved to p1-img0.jpg
+            # saved to p2-img0.jpg
+```
+
+!!! info
+
+    For diffusion pipelines, the stage config field `stage_args.[].runtime.max_batch_size` is 1 by default, and the input
+    list is sliced into single-item requests before feeding into the diffusion pipeline. For models that do internally support
+    batched inputs, you can [modify this configuration](../configuration/stage_configs.md) to let the model accept a longer batch of prompts.
+
+For more usages, please refer to [offline inference](../user_guide/examples/offline_inference/qwen2_5_omni.md)

 ## Online Serving with OpenAI-Completions API

-Please refer to [online serving](../user_guide/examples/online_serving/qwen2_5_omni.md)
+Text-to-image generation quickstart with vLLM-Omni:
+
+```bash
+vllm serve Tongyi-MAI/Z-Image-Turbo --omni --port 8091
+```
+
+```bash
+curl -s http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [
+      {"role": "user", "content": "a cup of coffee on the table"}
+    ],
+    "extra_body": {
+      "height": 1024,
+      "width": 1024,
+      "num_inference_steps": 50,
+      "guidance_scale": 4.0,
+      "seed": 42
+    }
+  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2 | base64 -d > coffee.png
+```
+
+For more details, please refer to [online serving](../user_guide/examples/online_serving/text_to_image.md).
@@ -34,10 +34,10 @@ CATEGORIES = {
        "name": "Core",
        "description": "Core scheduling and caching components.",
    },
-    "model_executor": {
-        "name": "Model Executor",
-        "description": "Model execution components.",
-    },
+    # "model_executor": {
+    #     "name": "Model Executor",
+    #     "description": "Model execution components.",
+    # },
    "config": {
        "name": "Configuration",
        "description": "Configuration classes.",
@@ -136,6 +136,15 @@ def scan_package(package_name: str = "vllm_omni") -> dict[str, list[str]]:
            relative_path = py_file.relative_to(ROOT_DIR)
            module_path = str(relative_path.with_suffix("")).replace("/", ".").replace("\\", ".")

+            # Skip excluded modules (avoid importing vllm during docs build)
+            excluded_prefixes = [
+                "vllm_omni.diffusion.models.qwen_image",
+                "vllm_omni.entrypoints.async_diffusion",
+                "vllm_omni.entrypoints.openai",
+            ]
+            if any(module_path.startswith(prefix) for prefix in excluded_prefixes):
+                continue
+
            # Handle __init__.py - use parent module path
            if py_file.name == "__init__.py":
                # Remove .__init__ from module path
@@ -173,7 +173,7 @@ class Example:
            gh_file = gh_file.relative_to(ROOT_DIR)

            # Make GitHub URL
-            url = "https://github.com/vllm-project/vllm/"
+            url = "https://github.com/vllm-project/vllm-omni/"
            url += "tree/main" if self.path.is_dir() else "blob/main"
            gh_url = f"{url}/{gh_file}"

@@ -183,7 +183,7 @@ class Example:

    def generate(self) -> str:
        content = f"# {self.title}\n\n"
-        url = "https://github.com/vllm-project/vllm/"
+        url = "https://github.com/vllm-project/vllm-omni/"
        url += "tree/main" if self.path.is_dir() else "blob/main"
        content += f"Source <{url}/{self.path.relative_to(ROOT_DIR)}>.\n\n"

@@ -195,7 +195,7 @@ class Example:
            main_file_rel = self.main_file.relative_to(ROOT_DIR)
            content += f'{code_fence}{self.main_file.suffix[1:]}\n--8<-- "{main_file_rel}"\n{code_fence}\n'
        else:
-            with open(self.main_file) as f:
+            with open(self.main_file, encoding="utf-8") as f:
                # Skip the title from md snippets as it's been included above
                main_content = f.readlines()[1:]
            content += self.fix_relative_links("".join(main_content))
@@ -0,0 +1,21 @@
+// Initialize Mermaid for diagram rendering
+mermaid.initialize({
+  startOnLoad: false,
+  theme: 'default',
+  securityLevel: 'loose',
+  flowchart: {
+    useMaxWidth: true,
+    htmlLabels: true
+  }
+});
+
+// Render Mermaid diagrams when page content is ready
+document$.subscribe(() => {
+  const mermaidElements = document.querySelectorAll('.mermaid');
+  if (mermaidElements.length > 0) {
+    mermaid.run({
+      querySelector: '.mermaid',
+      nodes: mermaidElements
+    });
+  }
+});
@@ -25,8 +25,11 @@ a:not(:has(svg)):not(.md-icon):not(.autorefs-external) {

 a[href*="localhost"]::after,
 a[href*="127.0.0.1"]::after,
-a[href*="org.readthedocs.build"]::after,
-a[href*="docs.vllm.ai"]::after {
+
+/* Hide external link icons for all links */
+a[href^="//"]::after,
+a[href^="http://"]::after,
+a[href^="https://"]::after {
    display: none !important;
 }

@@ -4,9 +4,9 @@ vLLM-Omni supports unified multimodal comprehension and generation models across

 ## Model Implementation

-If vLLM-Omni natively supports a model, its implementation can be found in <gh-file:vllm-omni/model_executor/models> and <gh-file:vllm_omni/diffusion/models>.
+If vLLM-Omni natively supports a model, its implementation can be found in <gh-file:vllm_omni/model_executor/models> and <gh-file:vllm_omni/diffusion/models>.

-## List of Supported Models for Nvidia GPU
+## List of Supported Models for Nvidia GPU / AMD GPU

 <style>
 th {
@@ -19,10 +19,28 @@ th {
 |--------------|--------|-------------------|
 | `Qwen3OmniMoeForConditionalGeneration` | Qwen3-Omni | `Qwen/Qwen3-Omni-30B-A3B-Instruct` |
 | `Qwen2_5OmniForConditionalGeneration` | Qwen2.5-Omni | `Qwen/Qwen2.5-Omni-7B`, `Qwen/Qwen2.5-Omni-3B` |
+| `BagelForConditionalGeneration` | BAGEL (DiT-only) | `ByteDance-Seed/BAGEL-7B-MoT` |
 | `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` |
+| `QwenImagePipeline` | Qwen-Image-2512 | `Qwen/Qwen-Image-2512` |
+| `QwenImageEditPipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image-Edit` |
+| `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` |
+| `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` |
+|`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` |
+| `WanPipeline` | Wan2.2-T2V, Wan2.2-TI2V | `Wan-AI/Wan2.2-T2V-A14B-Diffusers`, `Wan-AI/Wan2.2-TI2V-5B-Diffusers` |
+| `WanImageToVideoPipeline` | Wan2.2-I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` |
+| `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` |
+|`LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` |
+|`LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` |
+|`StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` |
+|`Flux2KleinPipeline` | FLUX.2-klein | `black-forest-labs/FLUX.2-klein-4B`, `black-forest-labs/FLUX.2-klein-9B` |
+|`FluxPipeline` | FLUX.1-dev | `black-forest-labs/FLUX.1-dev` |
+|`StableAudioPipeline` | Stable-Audio-Open | `stabilityai/stable-audio-open-1.0` |
+|`Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` |
+|`Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` |
+|`Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-0.6B-Base` |


-## List of Supported Models for Ascend NPU
+## List of Supported Models for NPU

 <style>
 th {
@@ -35,3 +53,9 @@ th {
 |--------------|--------|-------------------|
 | `Qwen2_5OmniForConditionalGeneration` | Qwen2.5-Omni | `Qwen/Qwen2.5-Omni-7B`, `Qwen/Qwen2.5-Omni-3B`|
 | `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` |
+| `QwenImagePipeline` | Qwen-Image-2512 | `Qwen/Qwen-Image-2512` |
+| `QwenImageEditPipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image-Edit` |
+| `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` |
+| `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` |
+| `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2511 | `Qwen/Qwen-Image-Edit-2511` |
+|`ZImagePipeline` | Z-Image | `Tongyi-MAI/Z-Image-Turbo` |
@@ -0,0 +1,205 @@
+# Image Edit API
+
+vLLM-Omni provides an OpenAI DALL-E compatible API for image edit using diffusion models.
+
+Each server instance runs a single model (specified at startup via `vllm serve <model> --omni`).
+
+## Quick Start
+
+### Start the Server
+
+For example...
+
+```bash
+# Qwen-Image
+vllm serve Qwen/Qwen-Image-Edit-2511 --omni --port 8000
+
+
+### Generate Images
+
+**Using curl:**
+
+```bash
+curl -s -D >(grep -i x-request-id >&2) \
+  -o >(jq -r '.data[0].b64_json' | base64 --decode > gift-basket.png) \
+  -X POST "http://localhost:8000/v1/images/edits" \
+  -F "model=xxx" \
+  -F "image=@./xx.png" \
+  -F "prompt='this bear is wearing sportwear. holding a basketball, and bending one leg.'" \
+  -F "size=1024x1024" \
+  -F "output_format=png"
+```
+
+
+**Using OpenAI SDK:**
+
+```python
+import base64
+from openai import OpenAI
+from pathlib import Path
+client = OpenAI(
+    api_key="None",
+    base_url="http://localhost:8000/v1"
+)
+
+input_image_url = "https://vllm-public-assets.s3.us-west-2.amazonaws.com/omni-assets/qwen-bear.png"
+
+result = client.images.edit(
+    image=[],
+    model="Qwen-Image-Edit-2511",
+    prompt="Change the bears in the two input images into walking together.",
+    size='512x512',
+    stream=False,
+    output_format='jpeg',
+    # url格式
+    extra_body={
+        "url": [input_image_url1,input_image_url],
+        "num_inference_steps": 50,
+        "guidance_scale": 1,
+        "seed": 777,
+    }
+)
+
+image_base64 = result.data[0].b64_json
+image_bytes = base64.b64decode(image_base64)
+
+# Save the image to a file
+with open("edit_out_http.jpeg", "wb") as f:
+    f.write(image_bytes)
+```
+
+## API Reference
+
+### Endpoint
+
+```
+POST /v1/images/edits
+Content-Type: multipart/form-data
+```
+
+### Request Parameters
+
+#### OpenAI Standard Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | string | **required** | A text description of the desired image |
+| `model` | string | server's model | Model to use (optional, should match server if specified) |
+| `image` | string or array | **required** | The image(s) to edit. |
+| `n` | integer | 1 | Number of images to generate (1-10) |
+| `size` | string | "auto" | Image dimensions in WxH format (e.g., "1024x1024", "512x512"), when set to auto, it decide size from first input image. |
+| `response_format` | string | "b64_json" | Response format (only "b64_json" supported) |
+| `user` | string | null | User identifier for tracking |
+| `output_format` | string | "png" | The format in which the generated images are returned. Must be one of "png", "jpg", "jpeg", "webp". |
+| `output_compression` | integer | 100 | The compression level (0-100%) for the generated images. |
+| `background` | string or null | "auto" | Allows to set transparency for the background of the generated image(s).
+
+#### vllm-omni Extension Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `url` | string or array | None | The image(s) to edit. |
+| `negative_prompt` | string | null | Text describing what to avoid in the image |
+| `num_inference_steps` | integer | model defaults | Number of diffusion steps |
+| `guidance_scale` | float | model defaults | Classifier-free guidance scale (typically 0.0-20.0) |
+| `true_cfg_scale` | float | model defaults | True CFG scale (model-specific parameter, may be ignored if not supported) |
+| `seed` | integer | null | Random seed for reproducibility |
+
+### Response Format
+
+```json
+{
+  "created": 1701234567,
+  "data": [
+    {
+      "b64_json": "<base64-encoded PNG>",
+      "url": null,
+      "revised_prompt": null
+    }
+  ],
+  "output_format": null,
+  "size": null,
+}
+```
+
+## Examples
+
+### Multiple Images input
+
+```bash
+curl -s -D >(grep -i x-request-id >&2) \
+  -o >(jq -r '.data[0].b64_json' | base64 --decode > gift-basket.png) \
+  -X POST "http://localhost:8000/v1/images/edits" \
+  -F "model=xxx" \
+  -F "image=@xx.png" \
+  -F "image=@xx.png"
+  -F "prompt='this bear is wearing sportwear. holding a basketball, and bending one leg.'" \
+  -F "size=1024x1024" \
+  -F "output_format=png"
+```
+
+
+## Parameter Handling
+
+The API passes parameters directly to the diffusion pipeline without model-specific transformation:
+
+- **Default values**: When parameters are not specified, the underlying model uses its own defaults
+- **Pass-through design**: User-provided values are forwarded directly to the diffusion engine
+- **Minimal validation**: Only basic type checking and range validation at the API level
+
+### Parameter Compatibility
+
+The API passes parameters directly to the diffusion pipeline without model-specific validation.
+
+- Unsupported parameters may be silently ignored by the model
+- Incompatible values will result in errors from the underlying pipeline
+- Recommended values vary by model - consult model documentation
+
+**Best Practice:** Start with the model's recommended parameters, then adjust based on your needs.
+
+## Error Responses
+
+### 400 Bad Request
+
+Invalid parameters (e.g., model mismatch):
+
+```json
+{
+  "detail": "Invalid size format: '1024x'. Expected format: 'WIDTHxHEIGHT' (e.g., '1024x1024')."
+}
+```
+
+### 422 Unprocessable Entity
+
+Validation errors (missing required fields):
+
+```json
+{
+  "detail": "Field 'image' or 'url' is required"
+}
+```
+
+## Troubleshooting
+
+### Server Not Running
+
+```bash
+# Check if server is responding
+curl -X http://localhost:8000/v1/images/edit \
+  -F "prompt='test'"
+```
+
+### Out of Memory
+
+If you encounter OOM errors:
+1. Reduce image size: `"size": "512x512"`
+2. Reduce inference steps: `"num_inference_steps": 25`
+
+## Development
+
+Enable debug logging to see prompts and generation details:
+
+```bash
+vllm serve Qwen/Qwen-Image-Edit-2511 --omni \
+  --uvicorn-log-level debug
+```
@@ -0,0 +1,249 @@
+# Image Generation API
+
+vLLM-Omni provides an OpenAI DALL-E compatible API for text-to-image generation using diffusion models.
+
+Each server instance runs a single model (specified at startup via `vllm serve <model> --omni`).
+
+## Quick Start
+
+### Start the Server
+
+For example...
+
+```bash
+# Qwen-Image
+vllm serve Qwen/Qwen-Image --omni --port 8000
+
+# Z-Image Turbo
+vllm serve Tongyi-MAI/Z-Image-Turbo --omni --port 8000
+```
+
+### Generate Images
+
+**Using curl:**
+
+```bash
+curl -X POST http://localhost:8000/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "prompt": "a dragon laying over the spine of the Green Mountains of Vermont",
+    "size": "1024x1024",
+    "seed": 42
+  }' | jq -r '.data[0].b64_json' | base64 -d > dragon.png
+```
+
+**Using Python:**
+
+```python
+import requests
+import base64
+from PIL import Image
+import io
+
+response = requests.post(
+    "http://localhost:8000/v1/images/generations",
+    json={
+        "prompt": "a black and white cat wearing a princess tiara",
+        "size": "1024x1024",
+        "num_inference_steps": 50,
+        "seed": 42,
+    }
+)
+
+# Decode and save
+img_data = response.json()["data"][0]["b64_json"]
+img_bytes = base64.b64decode(img_data)
+img = Image.open(io.BytesIO(img_bytes))
+img.save("cat.png")
+```
+
+**Using OpenAI SDK:**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")
+
+response = client.images.generate(
+    model="Qwen/Qwen-Image",
+    prompt="a horse jumping over a fence nearby a babbling brook",
+    n=1,
+    size="1024x1024",
+    response_format="b64_json"
+)
+
+# Note: Extension parameters (seed, steps, cfg) require direct HTTP requests
+```
+
+## API Reference
+
+### Endpoint
+
+```
+POST /v1/images/generations
+Content-Type: application/json
+```
+
+### Request Parameters
+
+#### OpenAI Standard Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | string | **required** | Text description of the desired image |
+| `model` | string | server's model | Model to use (optional, should match server if specified) |
+| `n` | integer | 1 | Number of images to generate (1-10) |
+| `size` | string | model defaults | Image dimensions in WxH format (e.g., "1024x1024", "512x512") |
+| `response_format` | string | "b64_json" | Response format (only "b64_json" supported) |
+| `user` | string | null | User identifier for tracking |
+
+#### vllm-omni Extension Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `negative_prompt` | string | null | Text describing what to avoid in the image |
+| `num_inference_steps` | integer | model defaults | Number of diffusion steps |
+| `guidance_scale` | float | model defaults | Classifier-free guidance scale (typically 0.0-20.0) |
+| `true_cfg_scale` | float | model defaults | True CFG scale (model-specific parameter, may be ignored if not supported) |
+| `seed` | integer | null | Random seed for reproducibility |
+
+### Response Format
+
+```json
+{
+  "created": 1701234567,
+  "data": [
+    {
+      "b64_json": "<base64-encoded PNG>",
+      "url": null,
+      "revised_prompt": null
+    }
+  ]
+}
+```
+
+## Examples
+
+### Multiple Images
+
+```bash
+curl -X POST http://localhost:8000/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "prompt": "a steampunk city set in a valley of the Adirondack mountains",
+    "n": 4,
+    "size": "1024x1024",
+    "seed": 123
+  }'
+```
+
+This generates 4 images in a single request.
+
+### With Negative Prompt
+
+```python
+response = requests.post(
+    "http://localhost:8000/v1/images/generations",
+    json={
+        "prompt": "a portrait of a skier in deep powder snow",
+        "negative_prompt": "blurry, low quality, distorted, ugly",
+        "num_inference_steps": 100,
+        "size": "1024x1024",
+    }
+)
+```
+
+## Parameter Handling
+
+The API passes parameters directly to the diffusion pipeline without model-specific transformation:
+
+- **Default values**: When parameters are not specified, the underlying model uses its own defaults
+- **Pass-through design**: User-provided values are forwarded directly to the diffusion engine
+- **Minimal validation**: Only basic type checking and range validation at the API level
+
+### Parameter Compatibility
+
+The API passes parameters directly to the diffusion pipeline without model-specific validation.
+
+- Unsupported parameters may be silently ignored by the model
+- Incompatible values will result in errors from the underlying pipeline
+- Recommended values vary by model - consult model documentation
+
+**Best Practice:** Start with the model's recommended parameters, then adjust based on your needs.
+
+## Error Responses
+
+### 400 Bad Request
+
+Invalid parameters (e.g., model mismatch):
+
+```json
+{
+  "detail": "Invalid size format: '1024x'. Expected format: 'WIDTHxHEIGHT' (e.g., '1024x1024')."
+}
+```
+
+### 422 Unprocessable Entity
+
+Validation errors (missing required fields):
+
+```json
+{
+  "detail": [
+    {
+      "loc": ["body", "prompt"],
+      "msg": "field required",
+      "type": "value_error.missing"
+    }
+  ]
+}
+```
+
+### 503 Service Unavailable
+
+Diffusion engine not initialized:
+
+```json
+{
+  "detail": "Diffusion engine not initialized. Start server with a diffusion model."
+}
+```
+
+## Troubleshooting
+
+### Server Not Running
+
+```bash
+# Check if server is responding
+curl http://localhost:8000/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "test"}'
+```
+
+### Out of Memory
+
+If you encounter OOM errors:
+1. Reduce image size: `"size": "512x512"`
+2. Reduce inference steps: `"num_inference_steps": 25`
+3. Generate fewer images: `"n": 1`
+
+## Testing
+
+Run the test suite to verify functionality:
+
+```bash
+# All image generation tests
+pytest tests/entrypoints/openai_api/test_image_server.py -v
+
+# Specific test
+pytest tests/entrypoints/openai_api/test_image_server.py::test_generate_single_image -v
+```
+
+## Development
+
+Enable debug logging to see prompts and generation details:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni \
+  --uvicorn-log-level debug
+```
@@ -2,7 +2,7 @@

 > Q: How many chips do I need to infer a model in vLLM-Omni?

-A: Now, we support natively disaggregated deployment for different model stages within a model. There is a restriction that one chip can only has one AutoRegressive model stage. This is because the unified KV cache management of vLLM. Stages of other types can coexist with in a chip. The restriction will be resolved in later version.
+A: Now, we support natively disaggregated deployment for different model stages within a model. There is a restriction that one chip can only have one AutoRegressive model stage. This is because the unified KV cache management of vLLM. Stages of other types can coexist within a chip. The restriction will be resolved in later version.

 > Q: When trying to run examples, I encounter error about backend of librosa or soundfile. How to solve it?

@@ -11,3 +11,19 @@ A: If you encounter error about backend of librosa, try to install ffmpeg with c
 sudo apt update
 sudo apt install ffmpeg
 ```
+
+> Q: I see GPU OOM or "free memory is less than desired GPU memory utilization" errors. How can I fix it?
+
+A: Refer to [GPU memory calculation and configuration](../configuration/gpu_memory_utilization.md) for guidance on tuning `gpu_memory_utilization` and related settings.
+
+> Q: I encounter some bugs or CI problems, which is urgent. How can I solve it?
+
+A: At first, you can check current [issues](https://github.com/vllm-project/vllm-omni/issues) to find possible solutions. If none of these satisfy your demand and it is urgent, please find these [volunteers](https://docs.vllm.ai/projects/vllm-omni/en/latest/community/volunteers/) for help.
+
+> Q: Does vLLM-Omni support AWQ or any other quantization?
+
+A: vLLM-Omni partitions model into several stages. For AR stages, it will reuse main logic of LLMEngine in vLLM. So current quantization supported in vLLM should be also supported in vLLM-Omni for them. But systematic verification is ongoing. For quantization for DiffusionEngine, we are working on it. Please stay tuned and welcome contribution!
+
+> Q: Does vLLM-Omni support multimodal streaming input and output?
+
+A: Not yet. We already put it on the [Roadmap](https://github.com/vllm-project/vllm-omni/issues/165). Please stay tuned!
@@ -0,0 +1,228 @@
+# Cache-DiT Acceleration Guide
+
+This guide explains how to use cache-dit acceleration in vLLM-Omni to speed up diffusion model inference.
+
+## Overview
+
+Cache-dit is a library that accelerates diffusion transformer models through intelligent caching mechanisms. It supports multiple acceleration techniques that can be combined for optimal performance:
+
+- **DBCache**: Dual Block Cache for reducing redundant computations
+- **TaylorSeer**: Taylor expansion-based forecasting for faster inference
+- **SCM**: Step Computation Masking for selective step computation
+
+## Quick Start
+
+### Basic Usage
+
+Enable cache-dit acceleration by simply setting `cache_backend="cache_dit"`. Cache-dit will use its recommended default parameters:
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+# Simplest way: just enable cache-dit with default parameters
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="cache_dit",
+)
+
+images = omni.generate(
+    "a beautiful landscape",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+**Default Parameters**: When `cache_config` is not provided, cache-dit uses optimized default values. See the [Configuration Reference](#configuration-reference) section for a complete list of all parameters and their default values.
+
+### Custom Configuration
+
+To customize cache-dit settings, provide a `cache_config` dictionary, for example:
+
+```python
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4,
+        "residual_diff_threshold": 0.12,
+    },
+)
+```
+
+## Online Serving (OpenAI-Compatible)
+
+Enable Cache-DiT for online serving by passing `--cache-backend cache_dit` when starting the server:
+
+```bash
+# Use Cache-DiT default (recommended) parameters
+vllm serve Qwen/Qwen-Image --omni --port 8091 --cache-backend cache_dit
+```
+
+To customize Cache-DiT settings for online serving, pass a JSON string via `--cache-config`:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni --port 8091 \
+  --cache-backend cache_dit \
+  --cache-config '{"Fn_compute_blocks": 1, "Bn_compute_blocks": 0, "max_warmup_steps": 4, "residual_diff_threshold": 0.12}'
+```
+
+## Acceleration Methods
+
+For comprehensive illustration, please view cache-dit [User_Guide](https://cache-dit.readthedocs.io/en/latest/user_guide/OVERVIEWS/)
+
+### 1. DBCache (Dual Block Cache)
+
+DBCache intelligently caches intermediate transformer block outputs when the residual differences between consecutive steps are small, reducing redundant computations without sacrificing quality.
+
+**Key Parameters**:
+
+- `Fn_compute_blocks` (int, default: 1): Number of **first n** transformer blocks used to compute stable feature differences. Higher values provide more accurate caching decisions but increase computation.
+- `Bn_compute_blocks` (int, default: 0): Number of **last n** transformer blocks used for additional fusion. These blocks act as an auto-scaler for approximate hidden states.
+- `max_warmup_steps` (int, default: 4): Number of initial steps where caching is disabled to ensure the model learns sufficient features before caching begins. Optimized for few-step distilled models.
+- `residual_diff_threshold` (float, default: 0.24): Threshold for residual difference. Higher values lead to faster performance but may reduce precision. Default uses a relatively higher threshold for more aggressive caching.
+- `max_cached_steps` (int, default: -1): Maximum number of cached steps. Set to -1 for unlimited caching.
+- `max_continuous_cached_steps` (int, default: 3): Maximum number of consecutive cached steps. Limits consecutive caching to prevent precision degradation.
+
+**Example Configuration**:
+
+```python
+cache_config={
+    "Fn_compute_blocks": 8,      # Use first 8 blocks for difference computation
+    "Bn_compute_blocks": 0,       # No additional fusion blocks
+    "max_warmup_steps": 8,        # Cache after 8 warmup steps
+    "residual_diff_threshold": 0.12,  # Higher threshold for faster inference
+    "max_cached_steps": -1,        # No limit on cached steps
+}
+```
+
+**Performance Tips**:
+
+- Default `Fn_compute_blocks=1` works well for most cases. Increase to 8-12 for larger models or when more accuracy is needed
+- Increase `residual_diff_threshold` (e.g., 0.12-0.15) for faster inference with slight quality trade-off, or decrease from default 0.24 for higher quality
+- Default `max_warmup_steps=4` is optimized for few-step models. Increase to 6-8 for more steps if needed
+
+### 2. TaylorSeer
+
+TaylorSeer uses Taylor expansion to forecast future hidden states, allowing the model to skip some computation steps while maintaining quality.
+
+**Key Parameters**:
+
+- `enable_taylorseer` (bool, default: False): Enable TaylorSeer acceleration
+- `taylorseer_order` (int, default: 1): Order of Taylor expansion. Higher orders provide better accuracy but require more computation.
+
+**Example Configuration**:
+
+```python
+cache_config={
+    "enable_taylorseer": True,
+    "taylorseer_order": 1,  # First-order Taylor expansion
+}
+```
+
+**Performance Tips**:
+
+- Use `taylorseer_order=1` for most cases (good balance of speed and quality)
+- Combine with DBCache for maximum acceleration
+- Higher orders (2-3) may improve quality but reduce speed gains
+
+### 3. SCM (Step Computation Masking)
+
+SCM allows you to specify which steps must be computed and which can use cached results, similar to LeMiCa/EasyCache style acceleration.
+
+**Key Parameters**:
+
+- `scm_steps_mask_policy` (str | None, default: None): Predefined mask policy. Options:
+  - `None`: SCM disabled (default)
+  - `"slow"`: More compute steps, higher quality (18 compute steps out of 28)
+  - `"medium"`: Balanced (15 compute steps out of 28)
+  - `"fast"`: More cache steps, faster inference (11 compute steps out of 28)
+  - `"ultra"`: Maximum speed (8 compute steps out of 28)
+- `scm_steps_policy` (str, default: "dynamic"): Policy for cached steps:
+  - `"dynamic"`: Use dynamic cache for masked steps (recommended)
+  - `"static"`: Use static cache for masked steps
+
+**Example Configuration**:
+
+```python
+cache_config={
+    "scm_steps_mask_policy": "medium",  # Balanced speed/quality
+    "scm_steps_policy": "dynamic",      # Use dynamic cache
+}
+```
+
+**Performance Tips**:
+
+- SCM is disabled by default (`scm_steps_mask_policy=None`). Enable it by setting a policy value if you need additional acceleration
+- Start with `"medium"` policy and adjust based on quality requirements
+- Use `"fast"` or `"ultra"` for maximum speed when quality can be slightly compromised
+- `"dynamic"` policy generally provides better quality than `"static"`
+- SCM mask is automatically regenerated when `num_inference_steps` changes during inference
+
+## Configuration Reference
+
+### DiffusionCacheConfig Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `Fn_compute_blocks` | int | 1 | First n blocks for difference computation (optimized for single-transformer models) |
+| `Bn_compute_blocks` | int | 0 | Last n blocks for fusion |
+| `max_warmup_steps` | int | 4 | Steps before caching starts (optimized for few-step distilled models) |
+| `max_cached_steps` | int | -1 | Max cached steps (-1 = unlimited) |
+| `max_continuous_cached_steps` | int | 3 | Max consecutive cached steps (prevents precision degradation) |
+| `residual_diff_threshold` | float | 0.24 | Residual difference threshold (higher for more aggressive caching) |
+| `num_inference_steps` | int \| None | None | Initial inference steps for SCM mask generation (optional, auto-refreshed during inference) |
+| `enable_taylorseer` | bool | False | Enable TaylorSeer acceleration (not suitable for few-step distilled models) |
+| `taylorseer_order` | int | 1 | Taylor expansion order |
+| `scm_steps_mask_policy` | str \| None | None | SCM mask policy (None, "slow", "medium", "fast", "ultra") |
+| `scm_steps_policy` | str | "dynamic" | SCM computation policy ("dynamic" or "static") |
+
+## Example: Accelerate Text-to-Image Generation with CacheDiT
+
+See `examples/offline_inference/text_to_image/text_to_image.py` for a complete working example with cache-dit acceleration.
+
+```bash
+# Enable cache-dit with hybrid acceleration
+cd examples/offline_inference/text_to_image
+python text_to_image.py \
+    --model Qwen/Qwen-Image \
+    --prompt "a cup of coffee on the table" \
+    --cache_backend cache_dit \
+    --num_inference_steps 50
+```
+
+
+The script uses cache-dit acceleration with a hybrid configuration combining DBCache, SCM, and TaylorSeer:
+
+```python
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="cache_dit",
+    cache_config={
+        # Scheme: Hybrid DBCache + SCM + TaylorSeer
+        # DBCache
+        "Fn_compute_blocks": 8,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4,
+        "residual_diff_threshold": 0.12,
+        # TaylorSeer
+        "enable_taylorseer": True,
+        "taylorseer_order": 1,
+        # SCM
+        "scm_steps_mask_policy": "fast",  # Set to None to disable SCM
+        "scm_steps_policy": "dynamic",
+    },
+)
+```
+
+You can customize the configuration by modifying the `cache_config` dictionary to use only specific methods (e.g., DBCache only, DBCache + SCM, etc.) based on your quality and speed requirements.
+
+To test another model, you can modify `--model` with the target model identifier like `Tongyi-MAI/Z-Image-Turbo` and update `cache_config` according the model architecture (e.g., number of transformer blocks).
+
+
+## Additional Resources
+
+- [Cache-DiT User Guide](https://cache-dit.readthedocs.io/en/latest/user_guide/OVERVIEWS/)
+- [Cache-DiT Benchmark](https://cache-dit.readthedocs.io/en/latest/benchmark/HYBRID_CACHE/)
+- [DBCache Technical Details](https://cache-dit.readthedocs.io/en/latest/user_guide/CACHE_API/)
@@ -0,0 +1,101 @@
+# CPU Offloading for Diffusion Model
+
+## Overview
+
+vLLM-Omni provides two offloading strategies to reduce GPU memory usage for diffusion models, allowing you to run larger models on GPUs with limited VRAM:
+
+1. **Model-level (Component) Offloading**: Swaps entire model components (DiT transformer, VAE, encoders) between GPU and CPU.
+2. **Layerwise (Blockwise) Offloading**: Keeps only a single or a few transformer blocks on GPU at a time, with compute - memory copy overlap.
+
+Both approaches use pinned memory for faster CPU-GPU transfers. For now, the two offloading strategies could not be used at the same time.
+
+
+## Model-level CPU Offloading
+
+### Implementation
+
+CPU offload lets the diffusion worker move large model components between GPU and CPU memory on demand. It keeps the DiT transformer resident on GPU only while it is actively running, and swaps it out when encoders modules need the device. This reduces peak VRAM usage so bigger checkpoints run on smaller GPUs, or multiple requests can share the same GPU.
+
+**Execution Flow**:
+1. Text encoders run on GPU while the DiT transformer is offloaded to CPU.
+2. Before denoising, weights are prefetched back to GPU, honoring pinned-memory copies for speed.
+3. After the diffusion step, the transformer returns to CPU and the process repeats as needed.
+
+Transfers use pinned host buffers, and the worker coordinates swaps via mutex-style hooks so components never compete for memory.
+
+### Configuration
+You can enable CPU offload in two ways:
+
+1. **Python API**: set `enable_cpu_offload=True`.
+
+```python
+from vllm_omni import Omni
+
+if __name__ == "__main__":
+
+    m = Omni(model="Qwen/Qwen-Image",enable_cpu_offload=True)
+```
+
+2. **CLI**: pass `--enable-cpu-offload` to the diffusion service entrypoint.
+
+### Limitations
+- Cold start latency increases for over one minute for some models(e.g., Qwen-Image)
+
+
+## Layerwise (Blockwise) Offloading
+
+### Implementation
+Layerwise offload operates at transformer block granularity, keeping a single transformer block, or a specified number of blocks, on GPU while others stay in CPU memory.
+
+Unlike full model-wise CPU offload which swaps entire components like DiT and encoders, layerwise offloading applies a sliding window way of loading and offloading weights between gpu and cpu: while block `i` computes, block `i+1` get prefetched asynchronously via pinned memory. In this way, only partial blocks(s) reside on GPU at any moment during inference, so that greatly decrease the memory occupancy.
+
+**Execution Flow**:
+
+1. During model initialization, all components are loaded to CPU first. Then components other than DiT model(s) in the pipeline, such as VAE and encoders, are moved to GPU. The weights of target transformer blocks are collected as contiguous tensors per layer on CPU with pinned memory; and non-block modules (embeddings, norms, etc) in the DiT model are moved to and stay on GPU.
+2. The first block(s) are transferred to GPU during initialization of `LayerwiseOffloader`, before the first denoising step of the very first request.
+3. As each block executes, the next block prefetches on a separate CUDA stream for compute - memory copy overlap. After execution, the current block is immediately freed from GPU memory.
+4. When the last block completes, the first block prefetches for the next denoising step.
+
+
+Example of hook executions of a DiT model with n layers, by default keep a single layer on GPU:
+| Layer (block) idx | forward pre-hook               | forward          | forward post-hook         |
+|-------------------|--------------------------------|------------------|---------------------------|
+| layer-0           | prefetch layer 1 (copy stream) | compute layer 0  | free layer-0 gpu weights  |
+| layer-1           | prefetch layer 2 (copy stream) | compute layer 1  | free layer-1 gpu weights  |
+| layer-2           | prefetch layer 3 (copy stream) | compute layer 2  | free layer-2 gpu weights  |
+| ...               | ...                            | ...              | ...                       |
+| layer-(n-1)       | **prefetch layer 0 (copy stream)** | compute layer (n-1) | free layer (n-1) gpu weights  |
+
+
+### Configuration
+
+1. **Python API**: set `enable_layerwise_offload=True` and optionally `layerwise_num_gpu_layers`.
+
+```python
+from vllm_omni import Omni
+
+if __name__ == "__main__":
+    m = Omni(
+        model="Wan-AI/Wan2.2-T2V-A14B-Diffusers",
+        enable_layerwise_offload=True,
+        ...
+    )
+```
+
+2. **CLI**: pass `--enable-layerwise-offload` and `--layerwise-num-gpu-layers` to the diffusion service entrypoint.
+
+### Supported Models
+
+| Architecture | Models | Example HF Models | DiT Model Cls | Blocks Attr Name |
+|--------------|--------|-------------------|----------|----------|
+| `QwenImagePipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image` | `QwenImageTransformer2DModel` | "transformer_blocks" |
+| `Wan22Pipeline` | Wan2.2 | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | `WanTransformer3DModel` | "blocks" |
+
+NOTE: Models must define `_layerwise_offload_blocks_attr` class attribute so that the layerwise offloader finds the target transformer blocks.
+
+### Limitations
+- Cold start latency increases because of
+    1) components are loaded to CPU first at the very first during initialization,  
+    2) weight consolidation and pinning
+- Performance depends on CPU <-> GPU interconnection (e.g., PCIe bandwidth).
+- Support single GPU only for now
@@ -0,0 +1,568 @@
+# Parallelism Acceleration Guide
+
+This guide includes how to use parallelism methods in vLLM-Omni to speed up diffusion model inference as well as reduce the memory requirement on each device.
+
+## Overview
+
+The following parallelism methods are currently supported in vLLM-Omni:
+
+1. DeepSpeed Ulysses Sequence Parallel (DeepSpeed Ulysses-SP) ([arxiv paper](https://arxiv.org/pdf/2309.14509)): Ulysses-SP splits the input along the sequence dimension and uses all-to-all communication to allow each device to compute only a subset of attention heads.
+
+2. [Ring-Attention](#ring-attention) - splits the input along the sequence dimension and uses ring-based P2P communication to accumulate attention results, keeping the sequence dimension sharded
+
+3. Classifier-Free-Guidance Parallel (CFG-Parallel): CFG-Parallel runs the positive/negative prompts of classifier-free guidance (CFG) on different devices, then merges on a single device to perform the scheduler step.
+
+4. [Tensor Parallelism](#tensor-parallelism): Tensor parallelism shards model weights across devices. This can reduce per-GPU memory usage. Note that for diffusion models we currently shard the majority of layers within the DiT.
+
+The following table shows which models are currently supported by parallelism method:
+
+### ImageGen
+
+| Model | Model Identifier | Ulysses-SP | Ring-SP | CFG-Parallel | Tensor-Parallel |
+|-------|------------------|------------|---------|--------------|--------------------------|
+| **LongCat-Image** | `meituan-longcat/LongCat-Image` | ✅ | ✅ | ❌ | ❌ |
+| **LongCat-Image-Edit** | `meituan-longcat/LongCat-Image-Edit` | ✅ | ✅ | ❌ | ❌ |
+| **Ovis-Image** | `OvisAI/Ovis-Image` | ❌ | ❌ | ❌ | ❌ |
+| **Qwen-Image** | `Qwen/Qwen-Image` | ✅ | ✅ | ✅ | ✅ |
+| **Qwen-Image-Edit** | `Qwen/Qwen-Image-Edit` | ✅ | ✅ | ✅ | ✅ |
+| **Qwen-Image-Edit-2509** | `Qwen/Qwen-Image-Edit-2509` | ✅ | ✅ | ✅ | ✅ |
+| **Qwen-Image-Layered** | `Qwen/Qwen-Image-Layered` | ✅ | ✅ | ✅ | ✅ |
+| **Z-Image** | `Tongyi-MAI/Z-Image-Turbo` | ✅ | ✅ | ❌ | ✅ (TP=2 only) |
+| **Stable-Diffusion3.5** | `stabilityai/stable-diffusion-3.5` | ❌ | ❌ | ❌ | ❌ |
+| **FLUX.2-klein** | `black-forest-labs/FLUX.2-klein-4B` | ❌ | ❌ | ❌ | ✅ |
+| **FLUX.1-dev** | `black-forest-labs/FLUX.1-dev` | ❌ | ❌ | ❌ | ✅ |
+
+
+!!! note "TP Limitations for Diffusion Models"
+    We currently implement Tensor Parallelism (TP) only for the DiT (Diffusion Transformer) blocks. This is because the `text_encoder` component in vLLM-Omni uses the original Transformers implementation, which does not yet support TP.
+
+    - Good news: The text_encoder typically has minimal impact on overall inference performance.
+    - Bad news: When TP is enabled, every TP process retains a full copy of the text_encoder weights, leading to significant GPU memory waste.
+
+    We are actively refactoring this design to address this. For details and progress, please refer to [Issue #771](https://github.com/vllm-project/vllm-omni/issues/771).
+
+
+!!! note "Why Z-Image is TP=2 only"
+    Z-Image Turbo is currently limited to `tensor_parallel_size` of **1 or 2** due to model shape divisibility constraints.
+    For example, the model has `n_heads=30` and a final projection out dimension of `64`, so valid TP sizes must divide both 30 and 64; the only common divisors are **1 and 2**.
+
+### VideoGen
+
+| Model | Model Identifier | Ulysses-SP | Ring-SP | Tensor-Parallel |
+|-------|------------------|------------|---------|--------------------------|
+| **Wan2.2** | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | ✅ | ✅ | ❌ |
+
+### Tensor Parallelism
+
+Tensor parallelism splits model parameters across GPUs. In vLLM-Omni, tensor parallelism is configured via `DiffusionParallelConfig.tensor_parallel_size`.
+
+#### Offline Inference
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+omni = Omni(
+    model="Tongyi-MAI/Z-Image-Turbo",
+    parallel_config=DiffusionParallelConfig(tensor_parallel_size=2),
+)
+
+outputs = omni.generate(
+    "a cat reading a book",
+    OmniDiffusionSamplingParams(
+        num_inference_steps=9,
+        width=512,
+        height=512,
+    ),
+)
+```
+
+### Sequence Parallelism
+
+#### Ulysses-SP
+
+##### Offline Inference
+
+An example of offline inference script using [Ulysses-SP](https://arxiv.org/pdf/2309.14509) is shown below:
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+ulysses_degree = 2
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    parallel_config=DiffusionParallelConfig(ulysses_degree=2)
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50, width=2048, height=2048),
+)
+```
+
+See `examples/offline_inference/text_to_image/text_to_image.py` for a complete working example.
+
+##### Online Serving
+
+You can enable Ulysses-SP in online serving for diffusion models via `--usp`:
+
+```bash
+# Text-to-image (requires >= 2 GPUs)
+vllm serve Qwen/Qwen-Image --omni --port 8091 --usp 2
+```
+
+##### Benchmarks
+!!! note "Benchmark Disclaimer"
+    These benchmarks are provided for **general reference only**. The configurations shown use default or common parameter settings and have not been exhaustively optimized for maximum performance. Actual performance may vary based on:
+
+    - Specific model and use case
+    - Hardware configuration
+    - Careful parameter tuning
+    - Different inference settings (e.g., number of steps, image resolution)
+
+
+To measure the parallelism methods, we run benchmarks with **Qwen/Qwen-Image** model generating images (**2048x2048** as long sequence input) with 50 inference steps. The hardware devices are NVIDIA H800 GPUs. `sdpa` is the attention backends.
+
+| Configuration | Ulysses degree |Generation Time | Speedup |
+|---------------|----------------|---------|---------|
+| **Baseline (diffusers)** | - | 112.5s | 1.0x |
+| Ulysses-SP  |  2  |  65.2s | 1.73x |
+| Ulysses-SP  |  4  | 39.6s | 2.84x |
+| Ulysses-SP  |  8  | 30.8s | 3.65x |
+
+#### Ring-Attention
+
+Ring-Attention ([arxiv paper](https://arxiv.org/abs/2310.01889)) splits the input along the sequence dimension and uses ring-based P2P communication to accumulate attention results. Unlike Ulysses-SP which uses all-to-all communication, Ring-Attention keeps the sequence dimension sharded throughout the computation and circulates Key/Value blocks through a ring topology.
+
+##### Offline Inference
+
+An example of offline inference script using Ring-Attention is shown below:
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+ring_degree = 2
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    parallel_config=DiffusionParallelConfig(ring_degree=2)
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50, width=2048, height=2048),
+)
+```
+
+See `examples/offline_inference/text_to_image/text_to_image.py` for a complete working example.
+
+
+##### Online Serving
+
+You can enable Ring-Attention in online serving for diffusion models via `--ring`:
+
+```bash
+# Text-to-image (requires >= 2 GPUs)
+vllm serve Qwen/Qwen-Image --omni --port 8091 --ring 2
+```
+
+##### Benchmarks
+!!! note "Benchmark Disclaimer"
+    These benchmarks are provided for **general reference only**. The configurations shown use default or common parameter settings and have not been exhaustively optimized for maximum performance. Actual performance may vary based on:
+
+    - Specific model and use case
+    - Hardware configuration
+    - Careful parameter tuning
+    - Different inference settings (e.g., number of steps, image resolution)
+
+
+To measure the parallelism methods, we run benchmarks with **Qwen/Qwen-Image** model generating images (**1024x1024** as long sequence input) with 50 inference steps. The hardware devices are NVIDIA A100 GPUs. `flash_attn` is the attention backends.
+
+| Configuration | Ring degree |Generation Time | Speedup |
+|---------------|----------------|---------|---------|
+| **Baseline (diffusers)** | - | 45.2s | 1.0x |
+| Ring-Attention  |  2  |  29.9s | 1.51x |
+| Ring-Attention  |  4  | 23.3s | 1.94x |
+
+
+#### Hybrid Ulysses + Ring
+
+You can combine both Ulysses-SP and Ring-Attention for larger scale parallelism. The total sequence parallel size equals `ulysses_degree × ring_degree`.
+
+##### Offline Inference
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+# Hybrid: 2 Ulysses × 2 Ring = 4 GPUs total
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    parallel_config=DiffusionParallelConfig(ulysses_degree=2, ring_degree=2)
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50, width=2048, height=2048),
+)
+```
+
+##### Online Serving
+
+```bash
+# Text-to-image (requires >= 4 GPUs)
+vllm serve Qwen/Qwen-Image --omni --port 8091 --usp 2 --ring 2
+```
+
+##### Benchmarks
+!!! note "Benchmark Disclaimer"
+    These benchmarks are provided for **general reference only**. The configurations shown use default or common parameter settings and have not been exhaustively optimized for maximum performance. Actual performance may vary based on:
+
+    - Specific model and use case
+    - Hardware configuration
+    - Careful parameter tuning
+    - Different inference settings (e.g., number of steps, image resolution)
+
+
+To measure the parallelism methods, we run benchmarks with **Qwen/Qwen-Image** model generating images (**1024x1024** as long sequence input) with 50 inference steps. The hardware devices are NVIDIA A100 GPUs. `flash_attn` is the attention backends.
+
+| Configuration | Ulysses degree | Ring degree | Generation Time | Speedup |
+|---------------|----------------|-------------|-----------------|---------|
+| **Baseline (diffusers)** | - | - | 45.2s | 1.0x |
+| Hybrid Ulysses + Ring  |  2  |  2  |  24.3s | 1.87x |
+
+
+##### How to parallelize a new model
+
+NOTE: "Terminology: SP vs CP"
+    Our "Sequence Parallelism" (SP) corresponds to "Context Parallelism" (CP) in the [diffusers library](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/_modeling_parallel.py).
+    We use "Sequence Parallelism" to align with vLLM-Omni's terminology.
+
+---
+
+###### Non-intrusive `_sp_plan` (Recommended)
+
+The `_sp_plan` mechanism allows SP without modifying `forward()` logic. The framework automatically registers hooks to shard inputs and gather outputs at module boundaries.
+
+**Requirements for `forward()` function:**
+
+- Tensor operations that need sharding/gathering must happen at **`nn.Module` boundaries** (not inline Python operations)
+- If your `forward()` contains inline tensor operations (e.g., `torch.cat`, `pad_sequence`) that need sharding, **extract them into a submodule**
+
+**When to create a submodule:**
+
+```python
+# ❌ BAD: Inline operations - hooks cannot intercept
+def forward(self, x, cap_feats):
+    unified = torch.cat([x, cap_feats], dim=1)  # Cannot be sharded via _sp_plan
+    ...
+
+# ✅ GOOD: Extract into a submodule
+class UnifiedPrepare(nn.Module):
+    def forward(self, x, cap_feats):
+        return torch.cat([x, cap_feats], dim=1)  # Now can be sharded via _sp_plan
+
+class MyModel(nn.Module):
+    def __init__(self):
+        self.unified_prepare = UnifiedPrepare()  # Submodule
+
+    def forward(self, x, cap_feats):
+        unified = self.unified_prepare(x, cap_feats)  # Hook can intercept here
+```
+
+---
+
+###### Defining `_sp_plan`
+
+**Type definitions** (see [diffusers `_modeling_parallel.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/_modeling_parallel.py) for reference):
+
+```python
+from vllm_omni.diffusion.distributed.sp_plan import (
+    SequenceParallelInput,   # Corresponds to diffusers' ContextParallelInput
+    SequenceParallelOutput,  # Corresponds to diffusers' ContextParallelOutput
+)
+```
+
+| Parameter | Description |
+|-----------|-------------|
+| `split_dim` | Dimension to split/gather (usually `1` for sequence) |
+| `expected_dims` | Expected tensor rank for validation (optional) |
+| `split_output` | `False`: shard **input** parameters; `True`: shard **output** tensors |
+| `auto_pad` | Auto-pad if sequence not divisible by world_size (Ulysses only) |
+
+**Key naming convention:**
+
+| Key | Meaning | Python equivalent |
+|-----|---------|-------------------|
+| `""` | Root model | `model` |
+| `"blocks.0"` | First element of ModuleList | `model.blocks[0]` |
+| `"blocks.*"` | All elements of ModuleList | `for b in model.blocks` |
+| `"outputs.main"` | ModuleDict entry | `model.outputs["main"]` |
+
+**Dictionary key types:**
+
+| Key type | `split_output` | Description |
+|----------|----------------|-------------|
+| `"param_name"` (str) | `False` | Shard **input parameter** by name |
+| `0`, `1` (int) | `True` | Shard **output tuple** by index |
+
+**Example** (similar to [diffusers `transformer_wan.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/transformer_wan.py)):
+
+```python
+class MyTransformer(nn.Module):
+    _sp_plan = {
+        # Shard rope module OUTPUTS (returns tuple)
+        "rope": {
+            0: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),  # cos
+            1: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),  # sin
+        },
+        # Shard transformer block INPUT parameter
+        "blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        # Gather at final projection
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+---
+
+###### Hook flow
+
+```
+Input → [SequenceParallelSplitHook: pre_forward] → Module.forward() → [post_forward] → ...
+                                                                              ↓
+... → [SequenceParallelGatherHook: post_forward] → Output
+```
+
+1. **SplitHook** shards tensors before/after the target module
+2. **Attention layers** handle Ulysses/Ring communication internally
+3. **GatherHook** collects sharded outputs
+
+The framework automatically applies these hooks when `sequence_parallel_size > 1`.
+
+---
+
+###### Method 2: Intrusive modification (For complex cases)
+
+For models with dynamic sharding logic that cannot be expressed via `_sp_plan`:
+
+```python
+from vllm_omni.diffusion.distributed.sp_sharding import sp_shard, sp_gather
+
+def forward(self, hidden_states, ...):
+    if self.parallel_config.sequence_parallel_size > 1:
+        hidden_states = sp_shard(hidden_states, dim=1)
+        # ... computation ...
+        output = sp_gather(output, dim=1)
+    return output
+```
+
+---
+
+###### Choosing the right approach
+
+| Scenario | Approach |
+|----------|----------|
+| Standard transformer | `_sp_plan` |
+| Inline tensor ops need sharding | Extract to submodule + `_sp_plan` |
+| Dynamic/conditional sharding | Intrusive modification |
+
+
+### CFG-Parallel
+
+#### Offline Inference
+
+CFG-Parallel is enabled through `DiffusionParallelConfig(cfg_parallel_size=2)`, which runs one rank for the positive branch and one rank for the negative branch.
+
+An example of offline inference using CFG-Parallel (image-to-image) is shown below:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+image_path = "path_to_image.png"
+omni = Omni(
+    model="Qwen/Qwen-Image-Edit",
+    parallel_config=DiffusionParallelConfig(cfg_parallel_size=2),
+)
+input_image = Image.open(image_path).convert("RGB")
+
+outputs = omni.generate(
+    {
+        "prompt": "turn this cat to a dog",
+        "negative_prompt": "low quality, blurry",
+        "multi_modal_data": {"image": input_image},
+    },
+    OmniDiffusionSamplingParams(
+        true_cfg_scale=4.0,
+        num_inference_steps=50,
+    ),
+)
+```
+
+Notes:
+
+- CFG-Parallel is only effective when a `negative_prompt` is provided AND a guidance scale (or `cfg_scale`) is greater than 1.
+
+See `examples/offline_inference/image_to_image/image_edit.py` for a complete working example.
+```bash
+cd examples/offline_inference/image_to_image/
+python image_edit.py \
+  --model "Qwen/Qwen-Image-Edit" \
+  --image "qwen_image_output.png" \
+  --prompt "turn this cat to a dog" \
+  --negative_prompt "low quality, blurry" \
+  --cfg_scale 4.0 \
+  --output "edited_image.png" \
+  --cfg_parallel_size 2
+```
+
+#### Online Serving
+
+You can enable CFG-Parallel in online serving for diffusion models via `--cfg-parallel-size`:
+
+```bash
+vllm serve Qwen/Qwen-Image-Edit --omni --port 8091 --cfg-parallel-size 2
+```
+
+#### How to parallelize a pipeline
+
+This section describes how to add CFG-Parallel to a diffusion **pipeline**. We use the Qwen-Image pipeline (`vllm_omni/diffusion/models/qwen_image/pipeline_qwen_image.py`) as the reference implementation.
+
+In `QwenImagePipeline`, each diffusion step runs two denoiser forward passes sequentially:
+
+- positive (prompt-conditioned)
+- negative (negative-prompt-conditioned)
+
+CFG-Parallel assigns these two branches to different ranks in the **CFG group** and synchronizes the results.
+
+vLLM-omni provides `CFGParallelMixin` base class that encapsulates the CFG parallel logic. By inheriting from this mixin and calling its methods, pipelines can easily implement CFG parallel without writing repetitive code.
+
+**Key Methods in CFGParallelMixin:**
+- `predict_noise_maybe_with_cfg()`: Automatically handles CFG parallel noise prediction
+- `scheduler_step_maybe_with_cfg()`: Scheduler step with automatic CFG rank synchronization
+
+**Example Implementation:**
+
+```python
+class QwenImageCFGParallelMixin(CFGParallelMixin):
+    """
+    Base Mixin class for Qwen Image pipelines providing shared CFG methods.
+    """
+
+    def diffuse(
+        self,
+        prompt_embeds: torch.Tensor,
+        prompt_embeds_mask: torch.Tensor,
+        negative_prompt_embeds: torch.Tensor,
+        negative_prompt_embeds_mask: torch.Tensor,
+        latents: torch.Tensor,
+        img_shapes: torch.Tensor,
+        txt_seq_lens: torch.Tensor,
+        negative_txt_seq_lens: torch.Tensor,
+        timesteps: torch.Tensor,
+        do_true_cfg: bool,
+        guidance: torch.Tensor,
+        true_cfg_scale: float,
+        image_latents: torch.Tensor | None = None,
+        cfg_normalize: bool = True,
+        additional_transformer_kwargs: dict[str, Any] | None = None,
+    ) -> torch.Tensor:
+        self.transformer.do_true_cfg = do_true_cfg
+
+        for i, t in enumerate(timesteps):
+            timestep = t.expand(latents.shape[0]).to(device=latents.device, dtype=latents.dtype)
+
+            # Prepare kwargs for positive (conditional) prediction
+            positive_kwargs = {
+                "hidden_states": latents,
+                "timestep": timestep / 1000,
+                "guidance": guidance,
+                "encoder_hidden_states_mask": prompt_embeds_mask,
+                "encoder_hidden_states": prompt_embeds,
+                "img_shapes": img_shapes,
+                "txt_seq_lens": txt_seq_lens,
+            }
+
+            # Prepare kwargs for negative (unconditional) prediction
+            if do_true_cfg:
+                negative_kwargs = {
+                    "hidden_states": latents,
+                    "timestep": timestep / 1000,
+                    "guidance": guidance,
+                    "encoder_hidden_states_mask": negative_prompt_embeds_mask,
+                    "encoder_hidden_states": negative_prompt_embeds,
+                    "img_shapes": img_shapes,
+                    "txt_seq_lens": negative_txt_seq_lens,
+                }
+            else:
+                negative_kwargs = None
+
+            # Predict noise with automatic CFG parallel handling
+            # - In CFG parallel mode: rank0 computes positive, rank1 computes negative
+            # - Automatically gathers results and combines them on rank0
+            noise_pred = self.predict_noise_maybe_with_cfg(
+                do_true_cfg=do_true_cfg,
+                true_cfg_scale=true_cfg_scale,
+                positive_kwargs=positive_kwargs,
+                negative_kwargs=negative_kwargs,
+                cfg_normalize=cfg_normalize,
+            )
+
+            # Step scheduler with automatic CFG synchronization
+            # - Only rank0 computes the scheduler step
+            # - Automatically broadcasts updated latents to all ranks
+            latents = self.scheduler_step_maybe_with_cfg(
+                noise_pred, t, latents, do_true_cfg
+            )
+
+        return latents
+```
+
+**How it works:**
+1. Prepare separate `positive_kwargs` and `negative_kwargs` for conditional and unconditional predictions
+2. Call `predict_noise_maybe_with_cfg()` which:
+   - Detects if CFG parallel is enabled (`get_classifier_free_guidance_world_size() > 1`)
+   - Distributes computation: rank0 processes positive, rank1 processes negative
+   - Gathers predictions and combines them using `combine_cfg_noise()` on rank0
+   - Returns combined noise prediction (only valid on rank0)
+3. Call `scheduler_step_maybe_with_cfg()` which:
+   - Only rank0 computes the scheduler step
+   - Broadcasts the updated latents to all ranks for synchronization
+
+**How to customize**
+
+Some pipelines may need to customize the following functions in `CFGParallelMixin`:
+1. You may need to edit `predict_noise` function for custom behaviors.
+```python
+def predict_noise(self, *args, **kwargs):
+    """
+    Forward pass through transformer to predict noise.
+
+    Subclasses should override this if they need custom behavior,
+    but the default implementation calls self.transformer.
+    """
+    return self.transformer(*args, **kwargs)[0]
+
+```
+2. The default normalization function after combining the noise predictions from both branches is as follows. You may need to customize it.
+```python
+def cfg_normalize_function(self, noise_pred, comb_pred):
+    """
+    Normalize the combined noise prediction.
+
+    Args:
+        noise_pred: positive noise prediction
+        comb_pred: combined noise prediction after CFG
+
+    Returns:
+        Normalized noise prediction tensor
+    """
+    cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
+    noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
+    noise_pred = comb_pred * (cond_norm / noise_norm)
+    return noise_pred
+```
@@ -0,0 +1,145 @@
+# TeaCache Configuration Guide
+
+TeaCache speeds up diffusion model inference by caching transformer computations when consecutive timesteps are similar. This typically provides **1.5x-2.0x speedup** with minimal quality loss.
+
+## Quick Start
+
+Enable TeaCache by setting `cache_backend` to `"tea_cache"`:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+# Simple configuration - model_type is automatically extracted from pipeline.__class__.__name__
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="tea_cache",
+    cache_config={
+        "rel_l1_thresh": 0.2  # Optional, defaults to 0.2
+    }
+)
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(
+        num_inference_steps=50,
+    ),
+)
+```
+
+### Using Environment Variable
+
+You can also enable TeaCache via environment variable:
+
+```bash
+export DIFFUSION_CACHE_BACKEND=tea_cache
+```
+
+Then initialize without explicitly setting `cache_backend`:
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_config={"rel_l1_thresh": 0.2}  # Optional
+)
+```
+
+## Online Serving (OpenAI-Compatible)
+
+Enable TeaCache for online serving by passing `--cache-backend tea_cache` when starting the server:
+
+```bash
+vllm serve Qwen/Qwen-Image --omni --port 8091 \
+  --cache-backend tea_cache \
+  --cache-config '{"rel_l1_thresh": 0.2}'
+```
+
+## Configuration Parameters
+
+### `rel_l1_thresh` (float, default: `0.2`)
+
+Controls the balance between speed and quality. Lower values prioritize quality, higher values prioritize speed.
+
+**Recommended values:**
+
+- `0.2` - **~1.5x speedup** with minimal quality loss (recommended)
+- `0.4` - **~1.8x speedup** with slight quality loss
+- `0.6` - **~2.0x speedup** with noticeable quality loss
+- `0.8` - **~2.25x speedup** with significant quality loss
+
+## Examples
+
+### Python API
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="tea_cache",
+    cache_config={"rel_l1_thresh": 0.2}
+)
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(
+        num_inference_steps=50,
+    ),
+)
+```
+
+## Performance Tuning
+
+Start with the default `rel_l1_thresh=0.2` and adjust based on your needs:
+
+- **Maximum quality**: Use `0.1-0.2`
+- **Balanced**: Use `0.2-0.4` (recommended)
+- **Maximum speed**: Use `0.6-0.8` (may reduce quality)
+
+## Troubleshooting
+
+### Quality Degradation
+
+If you notice quality issues, lower the threshold:
+
+```python
+cache_config={"rel_l1_thresh": 0.1}  # More conservative caching
+```
+
+## Supported Models
+
+### ImageGen
+
+<style>
+th {
+  white-space: nowrap;
+  min-width: 0 !important;
+}
+</style>
+
+| Architecture | Models | Example HF Models |
+|--------------|--------|-------------------|
+| `QwenImagePipeline` | Qwen-Image | `Qwen/Qwen-Image` |
+| `QwenImageEditPipeline` | Qwen-Image-Edit | `Qwen/Qwen-Image-Edit` |
+| `QwenImageEditPlusPipeline` | Qwen-Image-Edit-2509 | `Qwen/Qwen-Image-Edit-2509` |
+| `QwenImageLayeredPipeline` | Qwen-Image-Layered | `Qwen/Qwen-Image-Layered` |
+| `BagelForConditionalGeneration` | BAGEL (DiT-only) | `ByteDance-Seed/BAGEL-7B-MoT` |
+
+### VideoGen
+
+No VideoGen models are supported by TeaCache yet.
+
+### Coming Soon
+
+<style>
+th {
+  white-space: nowrap;
+  min-width: 0 !important;
+}
+</style>
+
+| Architecture | Models | Example HF Models |
+|--------------|--------|-------------------|
+| `FluxPipeline` | Flux | - |
+| `CogVideoXPipeline` | CogVideoX | - |
@@ -0,0 +1,236 @@
+# Diffusion Acceleration Overview
+
+vLLM-Omni supports various cache acceleration methods to speed up diffusion model inference with minimal quality degradation. These methods include **cache methods** that intelligently cache intermediate computations to avoid redundant work across diffusion timesteps, and **parallelism methods** that distribute the computation across multiple devices.
+
+## Supported Acceleration Methods
+
+vLLM-Omni currently supports two main cache acceleration backends:
+
+1. **[TeaCache](diffusion/teacache.md)** - Hook-based adaptive caching that caches transformer computations when consecutive timesteps are similar
+2. **[Cache-DiT](diffusion/cache_dit_acceleration.md)** - Library-based acceleration using multiple techniques:
+    - **DBCache** (Dual Block Cache): Caches intermediate transformer block outputs based on residual differences
+    - **TaylorSeer**: Uses Taylor expansion-based forecasting for faster inference
+    - **SCM** (Step Computation Masking): Selectively computes steps based on adaptive masking
+
+Both methods can provide significant speedups (typically **1.5x-2.0x**) while maintaining high output quality.
+
+vLLM-Omni also supports parallelism methods for diffusion models, including:
+
+1. [Ulysses-SP](diffusion/parallelism_acceleration.md#ulysses-sp) - splits the input along the sequence dimension and uses all-to-all communication to allow each device to compute only a subset of attention heads.
+
+2. [Ring-Attention](diffusion/parallelism_acceleration.md#ring-attention) - splits the input along the sequence dimension and uses ring-based P2P communication to accumulate attention results, keeping the sequence dimension sharded.
+
+3. [CFG-Parallel](diffusion/parallelism_acceleration.md#cfg-parallel) - runs the positive/negative prompts of classifier-free guidance (CFG) on different devices, then merges on a single device to perform the scheduler step.
+
+## Quick Comparison
+
+### Cache Methods
+
+| Method | Configuration | Description | Best For |
+|--------|--------------|-------------|----------|
+| **TeaCache** | `cache_backend="tea_cache"` | Simple, adaptive caching with minimal configuration | Quick setup, balanced speed/quality |
+| **Cache-DiT** | `cache_backend="cache_dit"` | Advanced caching with multiple techniques (DBCache, TaylorSeer, SCM) | Maximum acceleration, fine-grained control |
+
+## Supported Models
+
+The following table shows which models are currently supported by each acceleration method:
+
+### ImageGen
+
+| Model | Model Identifier | TeaCache | Cache-DiT | Ulysses-SP | Ring-Attention | CFG-Parallel |
+|-------|------------------|:----------:|:-----------:|:-----------:|:----------------:|:----------------:|
+| **LongCat-Image** | `meituan-longcat/LongCat-Image` | ❌ | ✅ | ❌ | ❌ | ✅ |
+| **LongCat-Image-Edit** | `meituan-longcat/LongCat-Image-Edit` | ❌ | ✅ | ❌ | ❌ | ✅ |
+| **Ovis-Image** | `OvisAI/Ovis-Image` | ❌ | ✅ | ❌ | ❌ | ✅ |
+| **Qwen-Image** | `Qwen/Qwen-Image` | ✅ | ✅ | ✅ | ✅ | ✅ |
+| **Qwen-Image-2512** | `Qwen/Qwen-Image-2512` | ✅ | ✅ | ✅ | ✅ | ✅ |
+| **Qwen-Image-Edit** | `Qwen/Qwen-Image-Edit` | ✅ | ✅ | ✅ | ✅ | ✅ |
+| **Qwen-Image-Edit-2509** | `Qwen/Qwen-Image-Edit-2509` | ✅ | ✅ | ✅ | ✅ | ✅ |
+| **Qwen-Image-Layered** | `Qwen/Qwen-Image-Layered` | ❌ | ✅ | ✅ | ✅ | ✅ |
+| **Z-Image** | `Tongyi-MAI/Z-Image-Turbo` | ✅ | ✅ | ❌ | ❌ | ❌ |
+| **Stable-Diffusion3.5** | `stabilityai/stable-diffusion-3.5` | ❌ | ✅ | ❌ | ❌ | ✅ |
+| **Bagel** | `ByteDance-Seed/BAGEL-7B-MoT` | ✅ | ✅ | ❌ | ❌ | ❌ |
+
+### VideoGen
+
+| Model | Model Identifier | TeaCache | Cache-DiT | Ulysses-SP | Ring-Attention |CFG-Parallel |
+|-------|------------------|:--------:|:---------:|:----------:|:--------------:|:----------------:|
+| **Wan2.2** | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` | ❌ | ✅ | ✅ | ✅ | ✅ |
+
+
+## Performance Benchmarks
+
+The following benchmarks were measured on **Qwen/Qwen-Image** and **Qwen/Qwen-Image-Edit** models generating 1024x1024 images with 50 inference steps:
+
+!!! note "Benchmark Disclaimer"
+    These benchmarks are provided for **general reference only**. The configurations shown use default or common parameter settings and have not been exhaustively optimized for maximum performance. Actual performance may vary based on:
+
+    - Specific model and use case
+    - Hardware configuration
+    - Careful parameter tuning
+    - Different inference settings (e.g., number of steps, image resolution)
+
+    For optimal performance in your specific scenario, we recommend experimenting with different parameter configurations as described in the detailed guides below.
+
+| Model | Cache Backend | Cache Config | Generation Time | Speedup | Notes |
+|-------|---------------|--------------|----------------|---------|-------|
+| **Qwen/Qwen-Image** | None  | None | 20.0s | 1.0x | Baseline (diffusers) |
+| **Qwen/Qwen-Image** | TeaCache | `rel_l1_thresh=0.2` | 10.47s | **1.91x** | Recommended default setting |
+| **Qwen/Qwen-Image** | Cache-DiT | DBCache + TaylorSeer (Fn=1, Bn=0, W=8, TaylorSeer order=1) | 10.8s | **1.85x** | - |
+| **Qwen/Qwen-Image** | Cache-DiT | DBCache + TaylorSeer + SCM (Fn=8, Bn=0, W=4, TaylorSeer order=1, SCM fast) | 14.0s | **1.43x** | - |
+| **Qwen/Qwen-Image-Edit** | None | No acceleration | 51.5s | 1.0x | Baseline (diffusers) |
+| **Qwen/Qwen-Image-Edit** | Cache-DiT | Default (Fn=1, Bn=0, W=4, TaylorSeer disabled, SCM disabled) | 21.6s | **2.38x** | - |
+
+To measure the parallelism methods, we run benchmarks with **Qwen/Qwen-Image** model generating images (**2048x2048** as long sequence input) with 50 inference steps. The hardware devices are NVIDIA H800 GPUs. `sdpa` is the attention backends.
+
+| Configuration | Ulysses degree |Generation Time | Speedup |
+|---------------|----------------|---------|---------|
+| **Baseline (diffusers)** | - | 112.5s | 1.0x |
+| Ulysses-SP  |  2  |  65.2s | 1.73x |
+| Ulysses-SP  |  4  | 39.6s | 2.84x |
+| Ulysses-SP  |  8  | 30.8s | 3.65x |
+
+## Quick Start
+
+### Using TeaCache
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="tea_cache",
+    cache_config={"rel_l1_thresh": 0.2}  # Optional, defaults to 0.2
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(
+        num_inference_steps=50,
+    ),
+)
+```
+
+### Using Cache-DiT
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 8,
+        "enable_taylorseer": True,
+        "taylorseer_order": 1,
+    }
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(
+        num_inference_steps=50,
+    ),
+)
+```
+
+### Using Ulysses-SP
+
+Run text-to-image:
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+ulysses_degree = 2
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    parallel_config=DiffusionParallelConfig(ulysses_degree=ulysses_degree)
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50, width=2048, height=2048),
+)
+```
+
+
+Run image-to-image:
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+ulysses_degree = 2
+
+omni = Omni(
+    model="Qwen/Qwen-Image-Edit",
+    parallel_config=DiffusionParallelConfig(ulysses_degree=ulysses_degree)
+)
+
+outputs = omni.generate(
+    {
+        "prompt": "turn this cat to a dog",
+        "multi_modal_data": {"image": input_image}
+    },
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+### Using Ring-Attention
+
+Run text-to-image:
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+ring_degree = 2
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    parallel_config=DiffusionParallelConfig(ring_degree=2)
+)
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50, width=2048, height=2048),
+)
+```
+
+### Using CFG-Parallel
+
+Run image-to-image:
+
+CFG-Parallel splits the CFG positive/negative branches across GPUs. Use it when you set a non-trivial `true_cfg_scale`.
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+cfg_parallel_size = 2
+
+omni = Omni(
+    model="Qwen/Qwen-Image-Edit",
+    parallel_config=DiffusionParallelConfig(cfg_parallel_size=cfg_parallel_size)
+)
+
+outputs = omni.generate(
+    {
+        "prompt": "turn this cat to a dog",
+        "multi_modal_data": {"image": input_image}
+    },
+    OmniDiffusionSamplingParams(num_inference_steps=50, true_cfg_scale=4.0),
+)
+```
+
+## Documentation
+
+For detailed information on each acceleration method:
+
+- **[TeaCache Guide](diffusion/teacache.md)** - Complete TeaCache documentation, configuration options, and best practices
+- **[Cache-DiT Acceleration Guide](diffusion/cache_dit_acceleration.md)** - Comprehensive Cache-DiT guide covering DBCache, TaylorSeer, SCM, and configuration parameters
+- **[Sequence Parallelism](diffusion/parallelism_acceleration.md#sequence-parallelism)** - Guidance on how to set sequence parallelism with configuration.
+- **[CFG-Parallel](diffusion/parallelism_acceleration.md#cfg-parallel)** - Guidance on how to set CFG-Parallel to run positive/negative branches across ranks.
@@ -0,0 +1,179 @@
+# BAGEL-7B-MoT
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/bagel>.
+
+## Set up
+
+Please refer to the [stage configuration documentation](https://docs.vllm.ai/projects/vllm-omni/en/latest/configuration/stage_configs/) to configure memory allocation appropriately for your hardware setup.
+
+## Run examples
+
+**Note**: These examples work with the default configuration on an **NVIDIA A100 (80GB)**. We also tested on dual **NVIDIA RTX 5000 Ada (32GB each)**. For dual-GPU setups, please modify the stage configuration to distribute the model across devices.
+
+Get into the bagel folder
+
+```bash
+cd examples/offline_inference/bagel
+```
+
+### Modality Control
+
+BAGEL-7B-MoT supports multiple modality modes. You can control the mode using the `--modality` argument:
+
+#### Text to Image (text2img)
+
+- **Pipeline**: Text → Thinker  → DiT → VAE Decode → Image
+- **Stages Used**: Stage 0 (Thinker) + Stage 1 (DiT)
+- **KV Transfer**: Thinker sends KV cache to DiT for conditioned generation
+
+Generate images from text prompts:
+
+```bash
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality text2img \
+                  --prompts "A cute cat"
+```
+
+#### Image to Image (img2img)
+
+- **Pipeline**: Image → VAE Encode → DiT → VAE Decode → New Image
+- **Stages Used**: Stage 1 (DiT) only
+- **Special**: Bypasses the Thinker stage, direct image-to-image transformation
+
+Transform images based on text prompts:
+
+```bash
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality img2img \
+                  --image-path /path/to/image.jpg \
+                  --prompts "Let the woman wear a blue dress"
+```
+
+#### Image to Text (img2text)
+
+- **Pipeline**: Image → ViT + VAE Encode → Thinker → Text Output
+- **Stages Used**: Stage 0 (Thinker) only
+- **Special**: Uses both VAE latent encoding AND ViT semantic encoding for comprehensive image understanding
+
+Generate text descriptions from images:
+
+```bash
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality img2text \
+                  --image-path /path/to/image.jpg \
+                  --prompts "Describe this image in detail"
+```
+
+#### Text to Text (text2text)
+
+- **Pipeline**: Text → Thinker → Text Output
+- **Stages Used**: Stage 0 (Thinker) only
+- **Special**: No visual components involved, operates as pure language model
+
+Pure text generation:
+
+```bash
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality text2text \
+                  --prompts "What is the capital of France?"
+
+# You can load prompts from a text file (one prompt per line):  
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality text2text \
+                  --txt-prompts /path/to/prompts.txt
+```
+
+### Inference Steps
+
+Control the number of inference steps for image generation:
+
+```bash
+# You can adjust steps to 100 to improve image quality
+python end2end.py --model ByteDance-Seed/BAGEL-7B-MoT \
+                  --modality text2img \
+                  --steps 50 \
+                  --prompts "A cute cat"
+```
+
+### Key arguments
+
+BAGEL-7B-MoT supports **multiple modality modes** for different use cases.
+
+The default yaml configuration deploys Thinker and DiT on the same GPU. You can use the default configuration file: [`bagel.yaml`](../../../vllm_omni/model_executor/stage_configs/bagel.yaml)
+
+#### 📌 Command Line Arguments (end2end.py)
+
+| Argument               | Type   | Default                       | Description                                                  |
+| :--------------------- | :----- | :---------------------------- | :----------------------------------------------------------- |
+| `--model`              | string | `ByteDance-Seed/BAGEL-7B-MoT` | Model path or name                                           |
+| `--modality`           | choice | `text2img`                    | Modality mode: `text2img`, `img2img`, `img2text`, `text2text` |
+| `--prompts`            | list   | `None`                        | Input text prompts directly                                  |
+| `--txt-prompts`        | string | `None`                        | Path to txt file with one prompt per line                    |
+| `--image-path`         | string | `None`                        | Input image path (for `img2img`/`img2text`)                  |
+| `--steps`              | int    | `50`                          | Number of inference steps                                    |
+| `--stage-configs-path` | string | `None`                        | Custom stage config file path                                |
+| `--worker-backend`     | choice | `process`                     | Worker backend: `process` or `ray`                           |
+| `--ray-address`        | string | `None`                        | Ray cluster address                                          |
+| `--enable-stats`       | flag   | `False`                       | Enable statistics logging                                    |
+| `--init-sleep-seconds` | int    | `20`                          | Initialization sleep time                                    |
+| `--batch-timeout`      | int    | `5`                           | Batch timeout                                                |
+| `--init-timeout`       | int    | `300`                         | Initialization timeout                                       |
+
+------
+
+#### ⚙️ Stage Configuration Parameters (bagel.yaml)
+
+ **Stage 0 - Thinker (LLM Stage)**
+
+| Parameter                        | Value                           | Description              |
+| :------------------------------- | :------------------------------ | :----------------------- |
+| `stage_type`                     | `llm`                           | Stage type               |
+| `devices`                        | `"0"`                           | GPU device ID            |
+| `max_batch_size`                 | `1`                             | Maximum batch size       |
+| `model_stage`                    | `thinker`                       | Model stage identifier   |
+| `model_arch`                     | `BagelForConditionalGeneration` | Model architecture       |
+| `gpu_memory_utilization`         | `0.4`                           | GPU memory utilization   |
+| `tensor_parallel_size`           | `1`                             | Tensor parallel size     |
+| `max_num_batched_tokens`         | `32768`                         | Maximum batched tokens   |
+| `omni_kv_config.need_send_cache` | `true`                          | Whether to send KV cache |
+
+------
+
+**Stage 1 - DiT (Diffusion Stage)**
+
+| Parameter                        | Value       | Description                 |
+| :------------------------------- | :---------- | :-------------------------- |
+| `stage_type`                     | `diffusion` | Stage type                  |
+| `devices`                        | `"0"`       | GPU device ID               |
+| `max_batch_size`                 | `1`         | Maximum batch size          |
+| `model_stage`                    | `dit`       | Model stage identifier      |
+| `gpu_memory_utilization`         | `0.4`       | GPU memory utilization      |
+| `omni_kv_config.need_recv_cache` | `true`      | Whether to receive KV cache |
+| `engine_input_source`            | `[0]`       | Input source from Stage 0   |
+
+------
+
+#### 🔗 Runtime Configuration
+
+| Parameter             | Value   | Description                      |
+| :-------------------- | :------ | :------------------------------- |
+| `window_size`         | `-1`    | Window size (-1 means unlimited) |
+| `max_inflight`        | `1`     | Maximum inflight requests        |
+| `shm_threshold_bytes` | `65536` | Shared memory threshold (64KB)   |
+
+## FAQ
+
+- If you encounter an error about the backend of librosa, try to install ffmpeg with the command below.
+
+```bash
+sudo apt update
+sudo apt install ffmpeg
+```
+
+- If you don’t know how much VRAM is needed for the model or encounter the OOM error, you can try to decrease the max_model_len.
+
+| Stage               | VRAM                         |
+| :------------------ | :--------------------------- |
+| Stage-0 (Thinker)   | **15.04 GiB** **+ KV Cache** |
+| Stage-1 (DiT)       | **26.50 GiB**                |
+| Total               | **~42 GiB + KV Cache**       |
@@ -0,0 +1,64 @@
+# Image-To-Image
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/image_to_image>.
+
+
+This example edits an input image with `Qwen/Qwen-Image-Edit` using the `image_edit.py` CLI.
+
+## Local CLI Usage
+
+### Single Image Editing
+
+Download the example image:
+
+```bash
+wget https://vllm-public-assets.s3.us-west-2.amazonaws.com/omni-assets/qwen-bear.png
+```
+
+Then run:
+
+```bash
+python image_edit.py \
+  --image qwen-bear.png \
+  --prompt "Let this mascot dance under the moon, surrounded by floating stars and poetic bubbles such as 'Be Kind'" \
+  --output output_image_edit.png \
+  --num_inference_steps 50 \
+  --cfg_scale 4.0
+```
+
+### Multiple Image Editing (Qwen-Image-Edit-2509)
+
+For multiple image inputs, use `Qwen/Qwen-Image-Edit-2509` or  `Qwen/Qwen-Image-Edit-2511`:
+
+```bash
+python image_edit.py \
+  --model Qwen/Qwen-Image-Edit-2509 \
+  --image img1.png img2.png \
+  --prompt "Combine these images into a single scene" \
+  --output output_image_edit.png \
+  --num_inference_steps 50 \
+  --cfg_scale 4.0 \
+  --guidance_scale 1.0
+```
+
+Key arguments:
+
+- `--model`: model name or path. Use `Qwen/Qwen-Image-Edit-2509` or later for multiple image support.
+- `--image`: path(s) to the source image(s) (PNG/JPG, converted to RGB). Can specify multiple images.
+- `--prompt` / `--negative_prompt`: text description (string).
+- `--cfg_scale`: true classifier-free guidance scale (default: 4.0). Classifier-free guidance is enabled by setting cfg_scale > 1 and providing a negative_prompt. Higher guidance scale encourages images closely linked to the text prompt, usually at the expense of lower image quality.
+- `--cfg_parallel_size`: the number of devices to run CFG Parallel. CFG Parallel is valid only if classifier-free guidance is enabled and `cfg_parallel_size` is set to 2.
+- `--guidance_scale`: guidance scale for guidance-distilled models (default: 1.0, disabled). Unlike classifier-free guidance (--cfg_scale), guidance-distilled models take the guidance scale directly as an input parameter. Enabled when guidance_scale > 1. Ignored when not using guidance-distilled models.
+- `--num_inference_steps`: diffusion sampling steps (more steps = higher quality, slower).
+- `--output`: path to save the generated PNG.
+
+## Example materials
+
+??? abstract "image_edit.py"
+    ``````py
+    --8<-- "examples/offline_inference/image_to_image/image_edit.py"
+    ``````
+??? abstract "run_qwen_image_edit_2511.sh"
+    ``````sh
+    --8<-- "examples/offline_inference/image_to_image/run_qwen_image_edit_2511.sh"
+    ``````
@@ -0,0 +1,67 @@
+# Image-To-Video
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/image_to_video>.
+
+
+This example demonstrates how to generate videos from images using Wan2.2 Image-to-Video models with vLLM-Omni's offline inference API.
+
+## Local CLI Usage
+
+### Wan2.2-I2V-A14B-Diffusers (MoE)
+```bash
+python image_to_video.py \
+  --model Wan-AI/Wan2.2-I2V-A14B-Diffusers \
+  --image input.png \
+  --prompt "A cat playing with yarn, smooth motion" \
+  --negative_prompt "<optional quality filter>" \
+  --height 480 \
+  --width 832 \
+  --num_frames 48 \
+  --guidance_scale 5.0 \
+  --guidance_scale_high 6.0 \
+  --num_inference_steps 40 \
+  --boundary_ratio 0.875 \
+  --flow_shift 12.0 \
+  --fps 16 \
+  --output i2v_output.mp4
+```
+
+### Wan2.2-TI2V-5B-Diffusers (Unified)
+```bash
+python image_to_video.py \
+  --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+  --image input.png \
+  --prompt "A cat playing with yarn, smooth motion" \
+  --negative_prompt "<optional quality filter>" \
+  --height 480 \
+  --width 832 \
+  --num_frames 48 \
+  --guidance_scale 4.0 \
+  --num_inference_steps 40 \
+  --flow_shift 12.0 \
+  --fps 16 \
+  --output i2v_output.mp4
+```
+
+Key arguments:
+
+- `--model`: Model ID (I2V-A14B for MoE, TI2V-5B for unified T2V+I2V).
+- `--image`: Path to input image (required).
+- `--prompt`: Text description of desired motion/animation.
+- `--height/--width`: Output resolution (auto-calculated from image if not set). Dimensions should be multiples of 16.
+- `--num_frames`: Number of frames (default 81).
+- `--guidance_scale` and `--guidance_scale_high`: CFG scale (applied to low/high-noise stages for MoE).
+- `--negative_prompt`: Optional list of artifacts to suppress.
+- `--cfg_parallel_size`: the number of devices to run CFG Parallel. CFG Parallel is valid only if classifier-free guidance is enabled and `cfg_parallel_size` is set to 2.
+- `--boundary_ratio`: Boundary split ratio for two-stage MoE models.
+- `--flow_shift`: Scheduler flow shift (5.0 for 720p, 12.0 for 480p).
+- `--num_inference_steps`: Number of denoising steps (default 50).
+- `--fps`: Frames per second for the saved MP4 (requires `diffusers` export_to_video).
+- `--output`: Path to save the generated video.
+
+## Example materials
+
+??? abstract "image_to_video.py"
+    ``````py
+    --8<-- "examples/offline_inference/image_to_video/image_to_video.py"
+    ``````
@@ -0,0 +1,107 @@
+# LoRA-Inference
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/lora_inference>.
+
+This contains examples for using LoRA (Low-Rank Adaptation) adapters with vLLM-omni diffusion models for offline inference.
+The example uses the  `stabilityai/stable-diffusion-3.5-medium` as the default model, but you can replace it with other models in vLLM-omni.
+
+## Overview
+
+Similar to vLLM, vLLM-omni uses a unified LoRA handling mechanism:
+
+- **Pre-loaded LoRA**: Loaded at initialization via `--lora-path` (pre-loaded into cache)
+- **Per-request LoRA**: Loaded on-demand. In the example, the LoRA is loaded via `--lora-request-path` in each request
+
+Both approaches use the same underlying mechanism - all LoRA adapters are handled uniformly through `set_active_adapter()`. If no LoRA request is provided in a request, all adapters are deactivated.
+
+## Usage
+
+### Pre-loaded LoRA (via --lora-path)
+
+Load a LoRA adapter at initialization. This adapter is pre-loaded into the cache and can be activated by requests:
+
+```bash
+python -m examples.offline_inference.lora_inference.lora_inference \
+    --prompt "A piece of cheesecake" \
+    --lora-path /path/to/lora/ \
+    --lora-scale 1.0 \
+    --num_inference_steps 50 \
+    --height 1024 \
+    --width 1024 \
+    --output output_preloaded.png
+```
+
+**Note**: When using `--lora-path`, the adapter is loaded at init time with a stable ID derived from the adapter path. This example activates it automatically for the request.
+
+### Per-request LoRA (via --lora-request-path)
+
+Load a LoRA adapter on-demand for each request:
+
+```bash
+python -m examples.offline_inference.lora_inference.lora_inference \
+    --prompt "A piece of cheesecake" \
+    --lora-request-path /path/to/lora/ \
+    --lora-scale 1.0 \
+    --num_inference_steps 50 \
+    --height 1024 \
+    --width 1024 \
+    --output output_per_request.png
+```
+
+### No LoRA
+
+If no LoRA request is provided, we will use the base model without any LoRA adapters:
+
+```bash
+python -m examples.offline_inference.lora_inference.lora_inference \
+    --prompt "A piece of cheesecake" \
+    --num_inference_steps 50 \
+    --height 1024 \
+    --width 1024 \
+    --output output_no_lora.png
+```
+
+## Parameters
+
+### LoRA Parameters
+
+- `--lora-path`: Path to LoRA adapter folder to pre-load at initialization (loads into cache with a stable ID derived from the path)
+- `--lora-request-path`: Path to LoRA adapter folder for per-request loading
+- `--lora-request-id`: Integer ID for the LoRA adapter (optional). If not provided and `--lora-request-path` is set, will derive a stable ID from the path.
+- `--lora-scale`: Scale factor for LoRA weights (default: 1.0). Higher values increase the influence of the LoRA adapter.
+
+### Standard Parameters
+
+- `--prompt`: Text prompt for image generation (required)
+- `--seed`: Random seed for reproducibility (default: 42)
+- `--height`: Image height in pixels (default: 1024)
+- `--width`: Image width in pixels (default: 1024)
+- `--num_inference_steps`: Number of denoising steps (default: 50)
+- `--output`: Output file path (default: `lora_output.png`)
+
+## How LoRA Works
+
+All LoRA adapters are handled uniformly:
+
+1. **Initialization**: If `--lora-path` is provided, the adapter is loaded into cache with a stable ID derived from the adapter path
+2. **Per-request**: If `--lora-request-path` is provided, the adapter is loaded/activated for that request
+3. **No LoRA**: If no LoRA request is provided (`req.lora_request` is None), all adapters are deactivated
+
+The system uses LRU cache management - adapters are cached and evicted when the cache is full (unless pinned).
+
+## LoRA Adapter Format
+
+LoRA adapters must be in PEFT (Parameter-Efficient Fine-Tuning) format. A typical LoRA adapter directory structure:
+
+```
+lora_adapter/
+├── adapter_config.json
+└── adapter_model.safetensors
+```
+
+## Example materials
+
+??? abstract "lora_inference.py"
+    ``````py
+    --8<-- "examples/offline_inference/lora_inference/lora_inference.py"
+    ``````
@@ -1,31 +1,23 @@
-# Offline Example of vLLM-Omni for Qwen2.5-Omni
+# Qwen2.5-Omni

-Source <https://github.com/vllm-project/vllm/tree/main/examples/offline_inference/qwen2_5_omni>.
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/qwen2_5_omni>.


-## 🛠️ Installation
+## Setup
+Please refer to the [stage configuration documentation](https://docs.vllm.ai/projects/vllm-omni/en/latest/configuration/stage_configs/) to configure memory allocation appropriately for your hardware setup.

-Please refer to [README.md](https://github.com/vllm-project/vllm/tree/main/README.md)
+## Run examples

-## Run examples (Qwen2.5-Omni)
 ### Multiple Prompts
-Download dataset from [seed_tts](https://drive.google.com/file/d/1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP/edit). To get the prompt, you can:
-```bash
-tar -xf <Your Download Path>/seedtts_testset.tar
-cp seedtts_testset/en/meta.lst examples/offline_inference/qwen2_5_omni/meta.lst
-python3 examples/offline_inference/qwen2_5_omni/extract_prompts.py \
-  --input examples/offline_inference/qwen2_5_omni/meta.lst \
-  --output examples/offline_inference/qwen2_5_omni/top100.txt \
-  --topk 100
-```
 Get into the example folder
 ```bash
 cd examples/offline_inference/qwen2_5_omni
 ```
-Then run the command below.
+Then run the command below. Note: for processing large volume data, it uses py_generator mode, which will return a python generator from Omni class.
 ```bash
 bash run_multiple_prompts.sh
 ```
+
 ### Single Prompt
 Get into the example folder
 ```bash
@@ -36,6 +28,42 @@ Then run the command below.
 bash run_single_prompt.sh
 ```

+### Modality control
+If you want to control output modalities, e.g. only output text, you can run the command below:
+```bash
+python end2end.py --output-wav output_audio \
+                  --query-type mixed_modalities \
+                  --modalities text
+```
+
+#### Using Local Media Files
+The `end2end.py` script supports local media files (audio, video, image) via CLI arguments:
+
+```bash
+# Use single local media files
+python end2end.py --query-type use_image --image-path /path/to/image.jpg
+python end2end.py --query-type use_video --video-path /path/to/video.mp4
+python end2end.py --query-type use_audio --audio-path /path/to/audio.wav
+
+# Combine multiple local media files
+python end2end.py --query-type mixed_modalities \
+    --video-path /path/to/video.mp4 \
+    --image-path /path/to/image.jpg \
+    --audio-path /path/to/audio.wav
+
+# Use audio from video file
+python end2end.py --query-type use_audio_in_video --video-path /path/to/video.mp4
+
+```
+
+If media file paths are not provided, the script will use default assets. Supported query types:
+- `use_image`: Image input only
+- `use_video`: Video input only
+- `use_audio`: Audio input only
+- `mixed_modalities`: Audio + image + video
+- `use_audio_in_video`: Extract audio from video
+- `text`: Text-only query
+
 ### FAQ

 If you encounter error about backend of librosa, try to install ffmpeg with command below.
@@ -1,20 +1,19 @@
-# Offline Example of vLLM-Omni for Qwen3-Omni
+# Qwen3-Omni

-Source <https://github.com/vllm-project/vllm/tree/main/examples/offline_inference/qwen3_omni>.
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/qwen3_omni>.


-## 🛠️ Installation
+## Setup
+Please refer to the [stage configuration documentation](https://docs.vllm.ai/projects/vllm-omni/en/latest/configuration/stage_configs/) to configure memory allocation appropriately for your hardware setup.

-Please refer to [README.md](https://github.com/vllm-project/vllm/tree/main/README.md)
+## Run examples

-## Run examples (Qwen3-Omni)
 ### Multiple Prompts
-Download dataset from [seed_tts](https://drive.google.com/file/d/1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP/edit). For processing dataset please refer to [Qwen2.5-Omni README.md](https://github.com/vllm-project/vllm/tree/main/examples/offline_inference/qwen2_5_omni/README.md)
 Get into the example folder
 ```bash
 cd examples/offline_inference/qwen3_omni
 ```
-Then run the command below.
+Then run the command below. Note: for processing large volume data, it uses py_generator mode, which will return a python generator from Omni class.
 ```bash
 bash run_multiple_prompts.sh
 ```
@@ -32,6 +31,42 @@ If you have not enough memory, you can set thinker with tensor parallel. Just ru
 bash run_single_prompt_tp.sh
 ```

+### Modality control
+If you want to control output modalities, e.g. only output text, you can run the command below:
+```bash
+python end2end.py --output-wav output_audio \
+                  --query-type use_audio \
+                  --modalities text
+```
+
+#### Using Local Media Files
+The `end2end.py` script supports local media files (audio, video, image) via command-line arguments:
+
+```bash
+# Use local video file
+python end2end.py --query-type use_video --video-path /path/to/video.mp4
+
+# Use local image file
+python end2end.py --query-type use_image --image-path /path/to/image.jpg
+
+# Use local audio file
+python end2end.py --query-type use_audio --audio-path /path/to/audio.wav
+
+# Combine multiple local media files
+python end2end.py --query-type mixed_modalities \
+    --video-path /path/to/video.mp4 \
+    --image-path /path/to/image.jpg \
+    --audio-path /path/to/audio.wav
+```
+
+If media file paths are not provided, the script will use default assets. Supported query types:
+- `use_video`: Video input
+- `use_image`: Image input
+- `use_audio`: Audio input
+- `text`: Text-only query
+- `multi_audios`: Multiple audio inputs
+- `mixed_modalities`: Combination of video, image, and audio inputs
+
 ### FAQ

 If you encounter error about backend of librosa, try to install ffmpeg with command below.
@@ -58,3 +93,7 @@ sudo apt install ffmpeg
    ``````sh
    --8<-- "examples/offline_inference/qwen3_omni/run_single_prompt_tp.sh"
    ``````
+??? abstract "text_prompts_10.txt"
+    ``````txt
+    --8<-- "examples/offline_inference/qwen3_omni/text_prompts_10.txt"
+    ``````
@@ -0,0 +1,94 @@
+# Qwen3-TTS Offline Inference
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/qwen3_tts>.
+
+
+This directory contains an offline demo for running Qwen3 TTS models with vLLM Omni. It builds task-specific inputs and generates WAV files locally.
+
+## Model Overview
+
+Qwen3 TTS provides multiple task variants for speech generation:
+
+- **CustomVoice**: Generate speech with a known speaker identity (speaker ID) and optional instruction.
+- **VoiceDesign**: Generate speech from text plus a descriptive instruction that designs a new voice.
+- **Base**: Voice cloning using a reference audio + reference transcript, with optional mode selection.
+
+## Setup
+Please refer to the [stage configuration documentation](https://docs.vllm.ai/projects/vllm-omni/en/latest/configuration/stage_configs/) to configure memory allocation appropriately for your hardware setup.
+
+## Quick Start
+
+Run a single sample for a task:
+
+```
+python end2end.py --query-type CustomVoice
+```
+
+Generated audio files are saved to `output_audio/` by default.
+
+## Task Usage
+
+### CustomVoice
+
+Single sample:
+
+```
+python end2end.py --query-type CustomVoice
+```
+
+Batch sample (multiple prompts in one run):
+
+```
+python end2end.py --query-type CustomVoice --use-batch-sample
+```
+
+### VoiceDesign
+
+Single sample:
+
+```
+python end2end.py --query-type VoiceDesign
+```
+
+Batch sample:
+
+```
+python end2end.py --query-type VoiceDesign --use-batch-sample
+```
+
+### Base (Voice Clone)
+
+Single sample:
+
+```
+python end2end.py --query-type Base
+```
+
+Batch sample:
+
+```
+python end2end.py --query-type Base --use-batch-sample
+```
+
+Mode selection for Base:
+
+- `--mode-tag icl` (default): standard mode
+- `--mode-tag xvec_only`: enable `x_vector_only_mode` in the request
+
+Examples:
+
+```
+python end2end.py --query-type Base --mode-tag icl
+```
+
+## Notes
+
+- The script uses the model paths embedded in `end2end.py`. Update them if your local cache path differs.
+- Use `--output-dir` (preferred) or `--output-wav` to change the output folder.
+
+## Example materials
+
+??? abstract "end2end.py"
+    ``````py
+    --8<-- "examples/offline_inference/qwen3_tts/end2end.py"
+    ``````
@@ -1,57 +0,0 @@
-# Qwen-Image Offline Inference
-
-Source <https://github.com/vllm-project/vllm/tree/main/examples/offline_inference/qwen_image>.
-
-
-This folder provides two simple entrypoints for experimenting with `Qwen/Qwen-Image` using vLLM-Omni:
-
- `text_to_image.py`: command-line script for single image generation.
- `web_demo.py`: lightweight Gradio UI for interactive prompt/seed/CFG exploration.
-
-
-## Local CLI Usage
-
-```bash
-python text_to_image.py \
-  --prompt "a cup of coffee on the table" \
-  --seed 42 \
-  --cfg_scale 4.0 \
-  --num_images_per_prompt 1 \
-  --num_inference_steps 50 \
-  --height 1024 \
-  --width 1024 \
-  --output outputs/coffee.png
-```
-
-Key arguments:
-
- `--prompt`: text description (string).
- `--seed`: integer seed for deterministic sampling.
- `--cfg_scale`: true CFG scale (model-specific guidance strength).
- `--num_images_per_prompt`: number of images to generate per prompt (saves as `output`, `output_1`, ...).
- `--num_inference_steps`: diffusion sampling steps (more steps = higher quality, slower).
- `--height/--width`: output resolution (defaults 1024x1024).
- `--output`: path to save the generated PNG.
-
-> ℹ️ Qwen-Image currently publishes best-effort presets at `1328x1328`, `1664x928`, `928x1664`, `1472x1140`, `1140x1472`, `1584x1056`, and `1056x1584`. Adjust `--height/--width` accordingly for the most reliable outcomes.
-
-## Web UI Demo
-
-Launch the gradio demo:
-
-```bash
-python gradio_demo.py --port 7862
-```
-
-Then open `http://localhost:7862/` on your local browser to interact with the web UI.
-
-## Example materials
-
-??? abstract "gradio_demo.py"
-    ``````py
-    --8<-- "examples/offline_inference/qwen_image/gradio_demo.py"
-    ``````
-??? abstract "text_to_image.py"
-    ``````py
-    --8<-- "examples/offline_inference/qwen_image/text_to_image.py"
-    ``````
@@ -0,0 +1,125 @@
+# Text-To-Image
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/text_to_image>.
+
+
+This folder provides several entrypoints for experimenting with `Qwen/Qwen-Image` `Qwen/Qwen-Image-2512` `Tongyi-MAI/Z-Image-Turbo` using vLLM-Omni:
+
+- `text_to_image.py`: command-line script for single image generation with advanced options.
+- `web_demo.py`: lightweight Gradio UI for interactive prompt/seed/CFG exploration.
+
+Note that when you pass in multiple independent prompts, they will be processed sequentially. Batching requests is currently not supported.
+
+## Basic Usage
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+
+if __name__ == "__main__":
+    omni = Omni(model="Qwen/Qwen-Image")
+    prompt = "a cup of coffee on the table"
+    outputs = omni.generate(prompt)
+    images = outputs[0].request_output[0].images
+    images[0].save("coffee.png")
+```
+
+Or put more than one prompt in a request.
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+
+if __name__ == "__main__":
+    omni = Omni(model="Qwen/Qwen-Image")
+    prompts = [
+      "a cup of coffee on a table",
+      "a toy dinosaur on a sandy beach",
+      "a fox waking up in bed and yawning",
+    ]
+    outputs = omni.generate(prompts)
+    for i, output in enumerate(outputs):
+      image = output.request_output[0].images[0].save(f"{i}.jpg")
+```
+
+!!! info
+
+    However, it is not currently recommended to do so
+    because not all models support batch inference,
+    and batch requesting mostly does not provide significant performance improvement (despite the impression that it does).
+    This feature is primarily for the sake of interface compatibility with vLLM and to allow for future improvements.
+
+!!! info
+
+    For diffusion pipelines, the stage config field `stage_args.[].runtime.max_batch_size` is 1 by default, and the input
+    list is sliced into single-item requests before feeding into the diffusion pipeline. For models that do internally support
+    batched inputs, you can [modify this configuration](../../../configuration/stage_configs.md) to let the model accept a longer batch of prompts.
+
+Apart from string prompt, vLLM-Omni also supports dictionary prompts in the same style as vLLM.
+This is useful for models that support negative prompts.
+
+```python
+from vllm_omni.entrypoints.omni import Omni
+
+if __name__ == "__main__":
+    omni = Omni(model="Qwen/Qwen-Image")
+    outputs = omni.generate([
+      {
+        "prompt": "a cup of coffee on a table"，
+        "negative_prompt": "low resolution"
+      },
+      {
+        "prompt": "a toy dinosaur on a sandy beach"，
+        "negative_prompt": "cinematic, realistic"
+      }
+    ])
+    for i, output in enumerate(outputs):
+      image = output.request_output[0].images[0].save(f"{i}.jpg")
+```
+
+## Local CLI Usage
+
+```bash
+python text_to_image.py \
+  --model Tongyi-MAI/Z-Image-Turbo \
+  --prompt "a cup of coffee on the table" \
+  --seed 42 \
+  --cfg_scale 4.0 \
+  --num_images_per_prompt 1 \
+  --num_inference_steps 50 \
+  --height 1024 \
+  --width 1024 \
+  --output outputs/coffee.png
+```
+
+Key arguments:
+
+- `--prompt`: text description (string).
+- `--seed`: integer seed for deterministic sampling.
+- `--cfg_scale`: true CFG scale (model-specific guidance strength).
+- `--cfg_parallel_size`: the number of devices to run CFG Parallel. CFG Parallel is valid only if classifier-free guidance is enabled and `cfg_parallel_size` is set to 2.
+- `--num_images_per_prompt`: number of images to generate per prompt (saves as `output`, `output_1`, ...).
+- `--num_inference_steps`: diffusion sampling steps (more steps = higher quality, slower).
+- `--height/--width`: output resolution (defaults 1024x1024).
+- `--output`: path to save the generated PNG.
+
+> ℹ️ Qwen-Image currently publishes best-effort presets at `1328x1328`, `1664x928`, `928x1664`, `1472x1140`, `1140x1472`, `1584x1056`, and `1056x1584`. Adjust `--height/--width` accordingly for the most reliable outcomes.
+
+## Web UI Demo
+
+Launch the gradio demo:
+
+```bash
+python gradio_demo.py --port 7862
+```
+
+Then open `http://localhost:7862/` on your local browser to interact with the web UI.
+
+## Example materials
+
+??? abstract "gradio_demo.py"
+    ``````py
+    --8<-- "examples/offline_inference/text_to_image/gradio_demo.py"
+    ``````
+??? abstract "text_to_image.py"
+    ``````py
+    --8<-- "examples/offline_inference/text_to_image/text_to_image.py"
+    ``````
@@ -0,0 +1,41 @@
+# Text-To-Video
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/text_to_video>.
+
+
+The `Wan-AI/Wan2.2-T2V-A14B-Diffusers` pipeline generates short videos from text prompts.
+
+## Local CLI Usage
+
+```bash
+python text_to_video.py \
+  --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
+  --negative_prompt "<optional quality filter>" \
+  --height 480 \
+  --width 640 \
+  --num_frames 32 \
+  --guidance_scale 4.0 \
+  --guidance_scale_high 3.0 \
+  --num_inference_steps 40 \
+  --fps 16 \
+  --output t2v_out.mp4
+```
+
+Key arguments:
+
+- `--prompt`: text description (string).
+- `--height/--width`: output resolution (defaults 720x1280). Dimensions should align with Wan VAE downsampling (multiples of 8).
+- `--num_frames`: Number of frames (Wan default is 81).
+- `--guidance_scale` and `--guidance_scale_high`: CFG scale (applied to low/high)..
+- `--negative_prompt`: optional list of artifacts to suppress (the PR demo used a long Chinese string).
+- `--cfg_parallel_size`: the number of devices to run CFG Parallel. CFG Parallel is valid only if classifier-free guidance is enabled and `cfg_parallel_size` is set to 2.
+- `--boundary_ratio`: Boundary split ratio for low/high DiT.
+- `--fps`: frames per second for the saved MP4 (requires `diffusers` export_to_video).
+- `--output`: path to save the generated video.
+
+## Example materials
+
+??? abstract "text_to_video.py"
+    ``````py
+    --8<-- "examples/offline_inference/text_to_video/text_to_video.py"
+    ``````
@@ -0,0 +1,232 @@
+# BAGEL-7B-MoT
+
+Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/bagel>.
+
+## 🛠️ Installation
+
+Please refer to [README.md](../../../README.md)
+
+## Run examples (BAGEL-7B-MoT)
+
+**Note**: These examples work with the default configuration on an **NVIDIA A100 (80GB)**. We also tested on dual **NVIDIA RTX 5000 Ada (32GB each)**. For dual-GPU setups, please modify the stage configuration to distribute the model across devices.
+
+### Launch the Server
+
+```bash
+# Use default configuration
+vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091
+```
+
+Or use the convenience script:
+
+```bash
+cd /workspace/vllm-omni/examples/online_serving/bagel
+bash run_server.sh
+```
+
+If you have a custom stage configs file, launch the server with the command below:
+
+```bash
+vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 --stage-configs-path /path/to/stage_configs_file
+```
+
+### Send Multi-modal Request
+
+Get into the bagel folder:
+
+```bash
+cd examples/online_serving/bagel
+```
+
+Send request via Python
+
+```bash
+python openai_chat_client.py --prompt "A cute cat" --modality text2img
+```
+
+The Python client supports the following command-line arguments:
+
+- `--prompt` (or `-p`): Text prompt for generation (default: `A cute cat`)
+- `--output` (or `-o`): Output file path for image results (default: `bagel_output.png`)
+- `--server` (or `-s`): Server URL (default: `http://localhost:8091`)
+- `--image-url` (or `-i`): Input image URL or local file path (for img2img/img2text modes)
+- `--modality` (or `-m`): Task modality (default: `text2img`). Options: `text2img`, `img2img`, `img2text`, `text2text`
+- `--height`: Image height in pixels (default: 512)
+- `--width`: Image width in pixels (default: 512)
+- `--steps`: Number of inference steps (default: 25)
+- `--seed`: Random seed (default: 42)
+- `--negative`: Negative prompt for image generation
+
+Example with custom parameters:
+
+```bash
+python openai_chat_client.py \
+    --prompt "A futuristic city" \
+    --modality text2img \
+    --height 768 \
+    --width 768 \
+    --steps 50 \
+    --seed 42 \
+    --negative "blurry, low quality"
+```
+
+## Modality Control
+
+BAGEL-7B-MoT supports **multiple modality modes** for different use cases.
+
+The default yaml configuration deploys Thinker and DiT on the same GPU. You can use the default configuration file: [`bagel.yaml`](../../../vllm_omni/model_executor/stage_configs/bagel.yaml)
+
+| Modality    | Input        | Output | Description                            |
+| ----------- | ------------ | ------ | -------------------------------------- |
+| `text2img`  | Text         | Image  | Generate images from text prompts      |
+| `img2img`   | Image + Text | Image  | Transform images using text guidance   |
+| `img2text`  | Image + Text | Text   | Generate text descriptions from images |
+| `text2text` | Text         | Text   | Pure text generation                   |
+
+### Text to Image (text2img)
+
+Generate images from text prompts:
+
+**Using Python client**
+
+```bash
+python openai_chat_client.py \
+    --prompt "A beautiful sunset over mountains" \
+    --modality text2img \
+    --output sunset.png \
+    --steps 50
+```
+
+**Using curl**
+
+```bash
+curl http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [{"role": "user", "content": [{"type": "text", "text": "<|im_start|>A beautiful sunset over mountains<|im_end|>"}]}],
+    "modalities": ["image"],
+    "height": 512,
+    "width": 512,
+    "num_inference_steps": 50,
+    "seed": 42
+  }'
+```
+
+
+### Image to Image (img2img)
+
+Transform images based on text prompts:
+
+**Using Python client**
+
+```bash
+python openai_chat_client.py \
+    --prompt "Make the cat stand up" \
+    --modality img2img \
+    --image-url /path/to/input.jpg \
+    --output transformed.png
+```
+
+**Using curl**
+
+```bash
+IMAGE_BASE64=$(base64 -w 0 cat.jpg)
+
+cat <<EOF > payload.json
+{
+    "messages": [{
+      "role": "user",
+      "content": [
+        {"type": "text", "text": "<|im_start|>Make the cat stand up<|im_end|>"},
+        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,${IMAGE_BASE64}"}}
+      ]
+    }],
+    "modalities": ["image"],
+    "height": 512,
+    "width": 512,
+    "num_inference_steps": 50,
+    "seed": 42
+}
+EOF
+
+curl http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d @payload.json
+
+```
+
+### Image to Text (img2text)
+
+Generate text descriptions from images:
+
+**Using Python client**
+
+```bash
+python openai_chat_client.py \
+    --prompt "Describe this image in detail" \
+    --modality img2text \
+    --image-url /path/to/image.jpg
+```
+
+**Using curl**
+
+```bash
+IMAGE_BASE64=$(base64 -w 0 cat.jpg)
+
+cat <<EOF > payload.json
+{
+  "messages": [{
+    "role": "user",
+    "content": [
+      {"type": "text", "text": "<|im_start|>user\n<|image_pad|>\nDescribe this image in detail<|im_end|>\n<|im_start|>assistant\n"},
+      {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,${IMAGE_BASE64}"}}
+    ]
+  }],
+  "modalities": ["text"]
+}
+EOF
+
+curl http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d @payload.json
+```
+
+### Text to Text (text2text)
+
+Pure text generation:
+
+**Using Python client**
+
+```bash
+python openai_chat_client.py \
+    --prompt "What is the capital of France?" \
+    --modality text2text
+```
+
+**Using curl**
+
+```bash
+curl http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [{"role": "user", "content": [{"type": "text", "text": "<|im_start|>user\nWhat is the capital of France?<|im_end|>\n<|im_start|>assistant\n"}]}]
+    "modalities": ["text"]
+  }'
+```
+
+## FAQ
+
+- If you encounter an error about the backend of librosa, try to install ffmpeg with the command below.
+
+```bash
+sudo apt update
+sudo apt install ffmpeg
+```
+
+- If you don’t know how much VRAM is needed for the model or encounter the OOM error, you can try to decrease the max_model_len.
+
+| Stage               | VRAM                         |
+| :------------------ | :--------------------------- |
+| Stage-0 (Thinker)   | **15.04 GiB** **+ KV Cache** |
+| Stage-1 (DiT)       | **26.50 GiB**                |
+| Total               | **~42 GiB + KV Cache**       |
@@ -1,7 +0,0 @@
-# Gradio Demo
-
-Source <https://github.com/vllm-project/vllm-omni/blob/main/examples/online_serving/gradio_demo.py>.
-
-``````py
--8<-- "examples/online_serving/gradio_demo.py"
-``````
				`@@ -0,0 +1 @@`
				`Architecture design of the entrypoint (update soon)`