Address review feedback on precheck-pr skill

- Fix base-branch detection: use BASE_SHA, fallback to origin/main ref
- Add union instruction for multi-category PRs in Step 3
- Enforce [Model] prefix includes model identifier
- Scope Cache-DiT check: blocking for top-level pipelines, warning for subclass
- Recognize deferred accuracy/perf claims in PR body
- Add import hygiene carve-out for closures/factory functions

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Signed-off-by: hsliu <liuhongsheng4@huawei.com>

.buildkite/bootstrap-amd-omni.sh

@@ -56,6 +56,88 @@ check_run_all_label() {
     fi
 }
 
+is_docs_only_change() {
+    local file_path
+    local has_any=0
+
+    while IFS= read -r file_path; do
+        [[ -z "${file_path}" ]] && continue
+        has_any=1
+
+        if [[ "${file_path}" == docs/* ]]; then
+            continue
+        fi
+        if [[ "${file_path}" == *.md ]]; then
+            continue
+        fi
+        if [[ "${file_path}" == "mkdocs.yml" ]]; then
+            continue
+        fi
+        return 1
+    done
+
+    [[ "${has_any}" -eq 1 ]]
+}
+
+resolve_skip_ci() {
+    local is_pr_build=0
+    local files
+    local base_branch base_ref
+
+    if [[ "${BUILDKITE_PULL_REQUEST:-false}" != "false" && -n "${BUILDKITE_PULL_REQUEST:-}" ]]; then
+        is_pr_build=1
+    fi
+
+    if [[ "${is_pr_build}" -eq 1 ]]; then
+        base_branch="${BUILDKITE_PULL_REQUEST_BASE_BRANCH:-main}"
+        if ! git rev-parse --verify "origin/${base_branch}" >/dev/null 2>&1; then
+            echo "resolve_skip_ci: origin/${base_branch} not found locally; trying fetch" >&2
+            git fetch --depth=200 origin "${base_branch}" >/dev/null 2>&1 || true
+        fi
+
+        base_ref=""
+        if git rev-parse --verify "origin/${base_branch}" >/dev/null 2>&1; then
+            base_ref="origin/${base_branch}"
+        elif git rev-parse --verify "${base_branch}" >/dev/null 2>&1; then
+            base_ref="${base_branch}"
+        else
+            echo "resolve_skip_ci: cannot resolve PR base ${base_branch}; skip-ci=0" >&2
+            echo -n 0
+            return 0
+        fi
+
+        if ! files="$(git diff --name-only "${base_ref}...${BUILDKITE_COMMIT}" 2>/dev/null)"; then
+            echo "resolve_skip_ci: failed to compute PR changed files; skip-ci=0" >&2
+            echo -n 0
+            return 0
+        fi
+    elif [[ "${BUILDKITE_BRANCH:-}" == "main" ]]; then
+        if ! git rev-parse --verify "${BUILDKITE_COMMIT}^" >/dev/null 2>&1; then
+            echo "resolve_skip_ci: commit has no parent on main; skip-ci=0" >&2
+            echo -n 0
+            return 0
+        fi
+        if ! files="$(git diff --name-only "${BUILDKITE_COMMIT}^..${BUILDKITE_COMMIT}" 2>/dev/null)"; then
+            echo "resolve_skip_ci: failed to compute main changed files; skip-ci=0" >&2
+            echo -n 0
+            return 0
+        fi
+    else
+        echo "resolve_skip_ci: not PR/main build; skip-ci=0" >&2
+        echo -n 0
+        return 0
+    fi
+
+    if is_docs_only_change <<< "${files}"; then
+        echo "resolve_skip_ci: docs-only change detected; skip-ci=1" >&2
+        echo -n 1
+        return 0
+    fi
+
+    echo "resolve_skip_ci: non-doc changes detected; skip-ci=0" >&2
+    echo -n 0
+}
+
 if [[ -z "${COV_ENABLED:-}" ]]; then
     COV_ENABLED=0
 fi
@@ -91,8 +173,24 @@ upload_pipeline() {
 
     cd .buildkite
 
-    # Select test definition file: merge suite for main, ready suite for PRs
-    if [[ $BUILDKITE_BRANCH == "main" ]]; then
+    # Select test definition file: merge suite for main, ready suite for PRs.
+    # For debugging, DEBUG_TEST_YAML can override the selection — accepts
+    # "merge" or "ready" (case-insensitive).
+    if [[ -n "${DEBUG_TEST_YAML:-}" ]]; then
+        case "${DEBUG_TEST_YAML,,}" in
+            merge)
+                TEST_YAML="test-amd-merge.yml"
+                ;;
+            ready)
+                TEST_YAML="test-amd-ready.yaml"
+                ;;
+            *)
+                echo "ERROR: DEBUG_TEST_YAML must be 'merge' or 'ready', got '$DEBUG_TEST_YAML'" >&2
+                exit 1
+                ;;
+        esac
+        echo "DEBUG_TEST_YAML override: using $TEST_YAML"
+    elif [[ $BUILDKITE_BRANCH == "main" ]]; then
         TEST_YAML="test-amd-merge.yml"
     else
         TEST_YAML="test-amd-ready.yaml"
@@ -140,30 +238,13 @@ fi
 # Early exit start: skip pipeline if conditions are met
 # ----------------------------------------------------------------------
 
-# skip pipeline if all changed files are under docs/
+# Match CUDA Buildkite skip-ci behavior for docs/markdown-only changes.
 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
-
-\`\`\`
-${file_diff}
-\`\`\`" --style "info" || true
-      echo "[docs-only] All changes are under docs/. Exiting before pipeline upload."
-      exit 0
+    SKIP_CI="$(resolve_skip_ci)"
+    if [[ "${SKIP_CI}" == "1" ]]; then
+        echo "[docs-only] Docs/markdown-only changes detected. Exiting before pipeline upload."
+        exit 0
     fi
-  fi
 fi
 
 # ----------------------------------------------------------------------

.buildkite/nightly-release-pipeline.yaml

@@ -1,20 +0,0 @@
-steps:
-  - label: "Build and upload wheel"
-    key: "build-wheel"
-    agents:
-      queue: cpu_queue_release
-    commands:
-      - "curl -LsSf https://astral.sh/uv/install.sh | sh"
-      - 'export PATH="$HOME/.local/bin:$PATH"'
-      - "uv venv --python=3.12 && source .venv/bin/activate"
-      - "uv pip install --upgrade build"
-      - "python3 -m build"
-      - "bash .buildkite/scripts/upload-nightly-wheels.sh"
-
-  - label: "Generate and upload wheel indices"
-    depends_on: "build-wheel"
-    allow_dependency_failure: true
-    agents:
-      queue: small_cpu_queue_release
-    commands:
-      - "bash .buildkite/scripts/generate-and-upload-nightly-index.sh"

.buildkite/pipeline-npu-a3.yaml

@@ -0,0 +1,110 @@
+# steps:
+#   - label: ":buildkit: Build and Push NPU Test Image - A3"
+#     key: image-build-a3
+#     agents:
+#       queue: "ascend-a3"
+#       resource_class: "npu-2"
+#     plugins:
+#       - kubernetes:
+#           metadata:
+#             annotations:
+#               vault.hashicorp.com/agent-init-first: "true"
+#               vault.hashicorp.com/agent-inject: "true"
+#               vault.hashicorp.com/agent-inject-perms-ca.pem: "0400"
+#               vault.hashicorp.com/agent-inject-perms-cert.pem: "0400"
+#               vault.hashicorp.com/agent-inject-perms-config.json: "0400"
+#               vault.hashicorp.com/agent-inject-perms-key.pem: "0400"
+#               vault.hashicorp.com/agent-inject-secret-ca.pem: internal/data/ascend/buildkitd
+#               vault.hashicorp.com/agent-inject-secret-cert.pem: internal/data/ascend/buildkitd
+#               vault.hashicorp.com/agent-inject-secret-config.json: internal/data/ascend/buildkitd
+#               vault.hashicorp.com/agent-inject-secret-key.pem: internal/data/ascend/buildkitd
+#               vault.hashicorp.com/agent-inject-template-ca.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.RootCA }}\n{{- end }}"
+#               vault.hashicorp.com/agent-inject-template-cert.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaCert }}\n{{- end }}"
+#               vault.hashicorp.com/agent-inject-template-config.json: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.dockerConfig }}\n{{- end }}"
+#               vault.hashicorp.com/agent-inject-template-key.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaKey }}\n{{- end }}"
+#               vault.hashicorp.com/agent-pre-populate-only: "true"
+#               vault.hashicorp.com/agent-run-as-group: "1000"
+#               vault.hashicorp.com/agent-run-as-user: "1000"
+#               vault.hashicorp.com/agent-service-account-token-volume-name: token-vol
+#               vault.hashicorp.com/role: ascend-gha-runners
+#               vault.hashicorp.com/secret-volume-path: /home/user/.docker/
+#               vault.hashicorp.com/tls-skip-verify: "true"
+#           podSpecPatch:
+#             volumes:
+#             - name: token-vol
+#               projected:
+#                 defaultMode: 420
+#                 sources:
+#                 - serviceAccountToken:
+#                     audience: api
+#                     expirationSeconds: 600
+#                     path: token
+#     env:
+#       VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
+#       IMAGE_NAME: "vllm-omni-ci-npu"
+#       IMAGE_REGISTRY: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry"
+#       BUILDKITD_ADDR: "tcp://buildkitd-service.buildkitd:1234"
+#     command: |
+#       set -ex
+
+#       echo "--- Building and pushing NPU Test Image"
+#       echo "Image: $${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+#       echo "buildkitd address: $${BUILDKITD_ADDR}"
+
+#       git config --global url."https://gh-proxy.test.osinfra.cn/https://github.com/".insteadOf "https://github.com/"
+
+#       if ! command -v buildctl &> /dev/null; then
+#         echo "Installing buildctl..."
+#         mkdir -p /tmp/buildkit
+#         BUILDKIT_VERSION="v0.29.0"
+#         wget -q "https://gh-proxy.test.osinfra.cn/https://github.com/moby/buildkit/releases/download/$${BUILDKIT_VERSION}/buildkit-$${BUILDKIT_VERSION}.linux-arm64.tar.gz" -O /tmp/buildkit.tar.gz
+#         tar -xzf /tmp/buildkit.tar.gz -C /tmp/buildkit
+#         cp /tmp/buildkit/bin/buildctl /usr/local/bin/
+#       fi
+
+#       export DOCKER_CONFIG=/home/user/.docker
+#       buildctl \
+#         --addr="$${BUILDKITD_ADDR}" \
+#         --tlscacert=/home/user/.docker/ca.pem \
+#         --tlscert=/home/user/.docker/cert.pem \
+#         --tlskey=/home/user/.docker/key.pem \
+#         build \
+#         --frontend dockerfile.v0 \
+#         --local context=. \
+#         --local dockerfile=./docker \
+#         --opt filename=Dockerfile.npu.ci.a3 \
+#         --secret id=dockerconfig,src=/home/user/.docker/config.json \
+#         --output type=image,name=$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG},push=true \
+#         --progress=plain
+
+#       echo "--- Image pushed successfully"
+#       echo "$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+
+#   # - label: "Upload Ready Pipeline - A3"
+#   #   depends_on: image-build-a3
+#   #   key: upload-ready-pipeline-a3
+#   #   if: build.branch != "main" && build.pull_request.labels includes "npu-test"
+#   #   agents:
+#   #     queue: "ascend-a3"
+#   #     resource_class: "npu-2"
+#   #   command: |
+#   #     buildkite-agent pipeline upload .buildkite/test-npu-ready.yaml
+
+  # # L4 Test — main+NIGHTLY=1 (scheduled), or PR with specific label (e.g. add label then Rebuild)
+  # - label: "Upload Nightly Pipeline"
+  #   depends_on: image-build
+  #   key: upload-nightly-pipeline
+  #   if: >-
+  #     (build.branch == "main" && build.env("NIGHTLY") == "1") ||
+  #     (build.branch != "main" && (
+  #       build.pull_request.labels includes "nightly-test" ||
+  #       build.pull_request.labels includes "omni-test" ||
+  #       build.pull_request.labels includes "tts-test" ||
+  #       build.pull_request.labels includes "diffusion-x2iat-test" ||
+  #       build.pull_request.labels includes "diffusion-x2v-test"
+  #     ))
+  #   commands:
+  #     - buildkite-agent pipeline upload .buildkite/test-npu-nightly.yaml
+  #   agents:
+  #     queue: "ascend-a3"
+  #     resource_class: "npu-2"

.buildkite/pipeline-npu.yaml

@@ -0,0 +1,112 @@
+steps:
+  - label: ":buildkit: Build and Push NPU Test Image - A2"
+    key: image-build
+    agents:
+      queue: "ascend-a2b3"
+      resource_class: "npu-1"
+    plugins:
+      - kubernetes:
+          metadata:
+            annotations:
+              vault.hashicorp.com/agent-init-first: "true"
+              vault.hashicorp.com/agent-inject: "true"
+              vault.hashicorp.com/agent-inject-perms-ca.pem: "0400"
+              vault.hashicorp.com/agent-inject-perms-cert.pem: "0400"
+              vault.hashicorp.com/agent-inject-perms-config.json: "0400"
+              vault.hashicorp.com/agent-inject-perms-key.pem: "0400"
+              vault.hashicorp.com/agent-inject-secret-ca.pem: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-secret-cert.pem: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-secret-config.json: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-secret-key.pem: internal/data/ascend/buildkitd
+              vault.hashicorp.com/agent-inject-template-ca.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.RootCA }}\n{{- end }}"
+              vault.hashicorp.com/agent-inject-template-cert.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaCert }}\n{{- end }}"
+              vault.hashicorp.com/agent-inject-template-config.json: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.dockerConfig }}\n{{- end }}"
+              vault.hashicorp.com/agent-inject-template-key.pem: "{{- with secret \"internal/data/ascend/buildkitd\" -}}\n{{ .Data.data.ClientCaKey }}\n{{- end }}"
+              vault.hashicorp.com/agent-pre-populate-only: "true"
+              vault.hashicorp.com/agent-run-as-group: "1000"
+              vault.hashicorp.com/agent-run-as-user: "1000"
+              vault.hashicorp.com/agent-service-account-token-volume-name: token-vol
+              vault.hashicorp.com/role: ascend-gha-runners
+              vault.hashicorp.com/secret-volume-path: /home/user/.docker/
+              vault.hashicorp.com/tls-skip-verify: "true"
+          podSpecPatch:
+            volumes:
+            - name: token-vol
+              projected:
+                defaultMode: 420
+                sources:
+                - serviceAccountToken:
+                    audience: api
+                    expirationSeconds: 600
+                    path: token
+    env:
+      VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
+      IMAGE_NAME: "vllm-omni-ci-npu"
+      IMAGE_REGISTRY: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry"
+      BUILDKITD_ADDR: "tcp://buildkitd-service.buildkitd:1234"
+    command: |
+      set -ex
+
+      echo "--- Building and pushing NPU Test Image"
+      echo "Image: $${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+      echo "buildkitd address: $${BUILDKITD_ADDR}"
+
+      if ! command -v buildctl &> /dev/null; then
+        echo "Installing buildctl..."
+        mkdir -p /tmp/buildkit
+        BUILDKIT_VERSION="v0.29.0"
+        wget -q "https://github.com/moby/buildkit/releases/download/$${BUILDKIT_VERSION}/buildkit-$${BUILDKIT_VERSION}.linux-arm64.tar.gz" -O /tmp/buildkit.tar.gz
+        tar -xzf /tmp/buildkit.tar.gz -C /tmp/buildkit
+        cp /tmp/buildkit/bin/buildctl /usr/local/bin/
+      fi
+
+      export DOCKER_CONFIG=/home/user/.docker
+      buildctl \
+        --addr="$${BUILDKITD_ADDR}" \
+        --tlscacert=/home/user/.docker/ca.pem \
+        --tlscert=/home/user/.docker/cert.pem \
+        --tlskey=/home/user/.docker/key.pem \
+        build \
+        --frontend dockerfile.v0 \
+        --local context=. \
+        --local dockerfile=./docker \
+        --opt filename=Dockerfile.npu.ci \
+        --secret id=dockerconfig,src=/home/user/.docker/config.json \
+        --output type=image,name=$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG},push=true \
+        --progress=plain
+
+      echo "--- Image pushed successfully"
+      echo "$${IMAGE_REGISTRY}/$${IMAGE_NAME}:$${VLLM_IMAGE_TAG}"
+
+  - label: "Upload Ready Pipeline"
+    depends_on: image-build
+    key: upload-ready-pipeline
+    if: >-
+      build.branch != "main" && (
+        build.pull_request.labels includes "npu-test" ||
+        build.pull_request.labels includes "ready"
+      )
+    agents:
+      queue: "ascend-a2b3"
+      resource_class: "npu-1"
+    command: |
+      buildkite-agent pipeline upload .buildkite/test-npu-ready.yaml
+
+  # L4 Test — main+NIGHTLY=1 (scheduled), or PR with specific label (e.g. add label then Rebuild)
+  - label: "Upload Nightly Pipeline"
+    depends_on: image-build
+    key: upload-nightly-pipeline
+    if: >-
+      (build.branch == "main" && build.env("NIGHTLY") == "1") ||
+      (build.branch != "main" && (
+        build.pull_request.labels includes "nightly-test" ||
+        build.pull_request.labels includes "omni-test" ||
+        build.pull_request.labels includes "tts-test" ||
+        build.pull_request.labels includes "diffusion-x2iat-test" ||
+        build.pull_request.labels includes "diffusion-x2v-test"
+      ))
+    commands:
+      - buildkite-agent pipeline upload .buildkite/test-npu-nightly.yaml
+    agents:
+      queue: "ascend-a2b3"
+      resource_class: "npu-1"

.buildkite/rebase-pipeline.yaml

@@ -0,0 +1,38 @@
+steps:
+  - label: ":docker: Build image"
+    key: image-build
+    commands:
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "docker build --progress=plain --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"
+
+  # L2 Test
+  - label: "Upload Ready Pipeline"
+    depends_on: image-build
+    key: upload-ready-pipeline
+    commands:
+      - buildkite-agent pipeline upload .buildkite/test-ready.yml
+    agents:
+      queue: "cpu_queue_premerge"
+
+  # L3 Test
+  - label: "Upload Merge Pipeline"
+    depends_on: image-build
+    key: upload-merge-pipeline
+    commands:
+      - buildkite-agent pipeline upload .buildkite/test-merge.yml
+    agents:
+      queue: "cpu_queue_premerge"
+
+  # L4 Test
+  - label: "Upload Nightly Pipeline"
+    depends_on: image-build
+    key: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1"
+    commands:
+      - buildkite-agent pipeline upload .buildkite/test-nightly.yml
+    agents:
+      queue: "cpu_queue_premerge"

.buildkite/release-pipeline.yaml

@@ -1,38 +1,110 @@
 steps:
-  - label: ":docker: Build image"
-    key: image-build
-    commands:
-      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "docker build --progress=plain --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"
+  - input: "Provide Release version here"
+    id: input-release-version
+    fields:
+      - text: "What is the release version?"
+        key: release-version
 
-  # L2 Test
-  - label: "Upload Ready Pipeline"
-    depends_on: image-build
-    key: upload-ready-pipeline
-    commands:
-      - buildkite-agent pipeline upload .buildkite/test-ready.yml
+  - label: "Build and upload wheel"
+    depends_on: ~
+    key: "build-wheel"
     agents:
-      queue: "cpu_queue_premerge"
+      queue: cpu_queue_release
+    commands:
+      - "curl -LsSf https://astral.sh/uv/install.sh | sh"
+      - 'export PATH="$HOME/.local/bin:$PATH"'
+      - "uv venv --python=3.12 && source .venv/bin/activate"
+      - "uv pip install --upgrade build"
+      - "python3 -m build"
+      - "bash .buildkite/scripts/upload-nightly-wheels.sh"
 
-  # L3 Test
-  - label: "Upload Merge Pipeline"
-    depends_on: image-build
-    key: upload-merge-pipeline
-    commands:
-      - buildkite-agent pipeline upload .buildkite/test-merge.yml
+  - label: "Generate and upload wheel indices"
+    depends_on: "build-wheel"
+    allow_dependency_failure: true
     agents:
-      queue: "cpu_queue_premerge"
+      queue: small_cpu_queue_release
+    commands:
+      - "bash .buildkite/scripts/generate-and-upload-nightly-index.sh"
 
-  # L4 Test
-  - label: "Upload Nightly Pipeline"
-    depends_on: image-build
-    key: upload-nightly-pipeline
-    if: build.env("NIGHTLY") == "1"
-    commands:
-      - buildkite-agent pipeline upload .buildkite/test-nightly.yml
+  - block: "Upload release wheels to PyPI"
+    depends_on: "build-wheel"
+    key: block-upload-release-wheels
+
+  - label: "Upload release wheels to PyPI"
+    depends_on:
+      - block-upload-release-wheels
+      - input-release-version
+    key: upload-release-wheels
     agents:
-      queue: "cpu_queue_premerge"
+      queue: small_cpu_queue_release
+    commands:
+      - "bash .buildkite/scripts/publish-release-wheels-pypi.sh"
+
+  - block: "Unblock to build release Docker images"
+    depends_on: ~
+    key: block-build-release-images
+
+  - group: "Build release Docker images"
+    key: "build-release-images"
+    depends_on: block-build-release-images
+    allow_dependency_failure: true
+    steps:
+      - label: "Build release image - x86_64"
+        key: build-release-image-x86
+        depends_on: ~
+        agents:
+          queue: cpu_queue_release
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              --build-arg max_jobs=16 \
+              --tag public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT-$(uname -m) \
+              --progress plain \
+              -f docker/Dockerfile.cuda .
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
+          # re-tag to default image tag and push, just in case arm64 build fails
+          - "docker tag public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT-$(uname -m) public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT"
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+      - label: "Build release image - aarch64"
+        key: build-release-image-arm64
+        depends_on: ~
+        agents:
+          queue: arm64_cpu_queue_release
+        commands:
+          - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+          - |
+            DOCKER_BUILDKIT=1 docker build \
+              --build-arg max_jobs=16 \
+              --tag public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT-$(uname -m) \
+              --progress plain \
+              -f docker/Dockerfile.cuda .
+          - "docker push public.ecr.aws/q9t5s3a7/vllm-omni-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
+        env:
+          DOCKER_BUILDKIT: "1"
+
+  - block: "Publish release images to DockerHub"
+    key: block-publish-release-images
+    depends_on:
+      - build-release-image-x86
+      - build-release-image-arm64
+      - input-release-version
+
+  - label: "Publish release images to DockerHub"
+    depends_on:
+      - block-publish-release-images
+    key: publish-release-images-dockerhub
+    agents:
+      queue: small_cpu_queue_release
+    commands:
+      - "bash .buildkite/scripts/publish-release-images.sh"
+    plugins:
+      - docker-login#v3.0.0:
+          username: vllmbot
+          password-env: DOCKERHUB_TOKEN
+    env:
+      DOCKER_BUILDKIT: "1"
+      DOCKERHUB_USERNAME: "vllmbot"

.buildkite/scripts/publish-release-images.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+#
+# Publish release Docker images from ECR to DockerHub.
+# Pulls per-arch images, tags with `latest` and versioned tags, pushes them,
+# then creates and pushes a multi-arch manifest.
+
+set -euo pipefail
+
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version --default "" | sed 's/^v//')
+if [ -z "${RELEASE_VERSION}" ]; then
+  echo "ERROR: release-version metadata not set"
+  exit 1
+fi
+
+COMMIT="$BUILDKITE_COMMIT"
+DOCKERHUB_REPO="vllm/vllm-omni"
+ECR_REPO="public.ecr.aws/q9t5s3a7/vllm-omni-release-repo"
+
+echo "========================================"
+echo "Publishing release images v${RELEASE_VERSION}"
+echo "  Commit: ${COMMIT}"
+echo "========================================"
+
+# Login to ECR to pull staging images
+aws ecr-public get-login-password --region us-east-1 | \
+  docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7
+
+docker pull "${ECR_REPO}:${COMMIT}-x86_64"
+docker pull "${ECR_REPO}:${COMMIT}-aarch64"
+
+docker tag "${ECR_REPO}:${COMMIT}-x86_64" "${DOCKERHUB_REPO}:latest-x86_64"
+docker tag "${ECR_REPO}:${COMMIT}-x86_64" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-x86_64"
+docker push "${DOCKERHUB_REPO}:latest-x86_64"
+docker push "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-x86_64"
+
+docker tag "${ECR_REPO}:${COMMIT}-aarch64" "${DOCKERHUB_REPO}:latest-aarch64"
+docker tag "${ECR_REPO}:${COMMIT}-aarch64" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-aarch64"
+docker push "${DOCKERHUB_REPO}:latest-aarch64"
+docker push "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-aarch64"
+
+docker manifest rm "${DOCKERHUB_REPO}:latest" || true
+docker manifest rm "${DOCKERHUB_REPO}:v${RELEASE_VERSION}" || true
+docker manifest create "${DOCKERHUB_REPO}:latest" "${DOCKERHUB_REPO}:latest-x86_64" "${DOCKERHUB_REPO}:latest-aarch64"
+docker manifest create "${DOCKERHUB_REPO}:v${RELEASE_VERSION}" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-x86_64" "${DOCKERHUB_REPO}:v${RELEASE_VERSION}-aarch64"
+docker manifest push "${DOCKERHUB_REPO}:latest"
+docker manifest push "${DOCKERHUB_REPO}:v${RELEASE_VERSION}"
+
+echo ""
+echo "Successfully published release images for v${RELEASE_VERSION}"

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

@@ -0,0 +1,52 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+
+set -e
+
+BUCKET="vllm-wheels"
+SUBPATH="omni/$BUILDKITE_COMMIT"
+S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
+
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version | sed 's/^v//')
+if [[ -z "$RELEASE_VERSION" ]]; then
+  echo "[FATAL] release-version metadata not set."
+  exit 1
+fi
+echo "Release version: $RELEASE_VERSION"
+
+if [[ -z "$PYPI_TOKEN" ]]; then
+  echo "[FATAL] PYPI_TOKEN is not set."
+  exit 1
+else
+  export TWINE_USERNAME="__token__"
+  export TWINE_PASSWORD="$PYPI_TOKEN"
+fi
+
+set -x
+
+if ! command -v uv &> /dev/null; then
+  curl -LsSf https://astral.sh/uv/install.sh | sh
+  export PATH="$HOME/.local/bin:$PATH"
+fi
+
+uv venv --python 3.12 /tmp/vllm-omni-release-env
+source /tmp/vllm-omni-release-env/bin/activate
+uv pip install twine
+
+DIST_DIR=/tmp/vllm-omni-release-dist
+mkdir -p "$DIST_DIR"
+
+echo "Downloading wheels from S3:"
+aws s3 ls "$S3_COMMIT_PREFIX"
+aws s3 cp --recursive --exclude "*" --include "vllm_omni-${RELEASE_VERSION}*.whl" --exclude "*dev*" "$S3_COMMIT_PREFIX" "$DIST_DIR"
+ls -la "$DIST_DIR"
+
+PYPI_WHEEL_FILES=$(find "$DIST_DIR" -name "vllm_omni-${RELEASE_VERSION}*.whl" -not -name "*+*")
+if [[ -z "$PYPI_WHEEL_FILES" ]]; then
+  echo "[FATAL] No wheels found for version ${RELEASE_VERSION}"
+  exit 1
+fi
+
+python3 -m twine check $PYPI_WHEEL_FILES
+python3 -m twine upload --non-interactive --verbose $PYPI_WHEEL_FILES
+echo "Wheels uploaded to PyPI"

.buildkite/scripts/upload_pipeline_with_skip_ci.sh

@@ -1,10 +1,18 @@
 #!/usr/bin/env bash
-# Evaluate docs-only skip-ci and upload continuation steps from the same `.buildkite/pipeline.yml`
+# Evaluate docs-only skip-ci and upload continuation steps from the given pipeline YAML
 # (YAML document after the first `---`). Buildkite `if` is evaluated at upload time.
+#
+# Usage: upload_pipeline_with_skip_ci.sh [pipeline_yaml]
+#   pipeline_yaml: path relative to repo root or absolute (default: .buildkite/pipeline.yml)
 set -euo pipefail
 
 ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
-PIPELINE_YML="${ROOT}/.buildkite/pipeline.yml"
+PIPELINE_ARG="${1:-.buildkite/pipeline.yml}"
+if [[ "${PIPELINE_ARG}" = /* ]]; then
+  PIPELINE_YML="${PIPELINE_ARG}"
+else
+  PIPELINE_YML="${ROOT}/${PIPELINE_ARG}"
+fi
 
 # Prints a single digit to stdout: 1 = skip image CI, 0 = run. Logs go to stderr.
 is_docs_only_change() {
@@ -21,7 +29,7 @@ is_docs_only_change() {
     if [[ "${file_path}" == *.md ]]; then
       continue
     fi
-    if [[ "${file_path}" == "mkdocs.yaml" ]]; then
+    if [[ "${file_path}" == "mkdocs.yml" ]]; then
       continue
     fi
     return 1
@@ -104,12 +112,13 @@ import pathlib
 path = pathlib.Path(os.environ["PIPELINE_YML"])
 text = path.read_text(encoding="utf-8")
 sep = "\n---\n"
-if sep not in text:
-    raise SystemExit(
-        "upload_pipeline_with_skip_ci: .buildkite/pipeline.yml must contain a '\\n---\\n' separator "
-        "(document 1 = bootstrap, document 2 = uploaded steps)"
-    )
-_, continuation = text.split(sep, 1)
+# Two supported layouts:
+#   - multi-doc: doc 1 = bootstrap (loaded by Buildkite), doc 2 (after `---`) = uploaded steps.
+#   - single-doc: whole file is the uploaded steps; caller (UI init step) is the bootstrap.
+if sep in text:
+    _, continuation = text.split(sep, 1)
+else:
+    continuation = text
 
 skip = os.environ.get("SKIP_CI") == "1"
 # When docs-only skip-ci: skip default CI image, but still build for L4 nightly (PR label nightly-test or
@@ -126,7 +135,7 @@ if skip:
 else:
     rep = "'true'"
     ready_rep = "'build.branch != \"main\" && build.pull_request.labels includes \"ready\"'"
-    merge_rep = "'(build.branch == \"main\" && build.env(\"NIGHTLY\") != \"1\") || (build.branch != \"main\" && build.pull_request.labels includes \"merge-test\")'"
+    merge_rep = "'(build.branch == \"main\" && build.env(\"NIGHTLY\") != \"1\" && build.env(\"WEEKLY\") != \"1\") || (build.branch != \"main\" && build.pull_request.labels includes \"merge-test\")'"
 rendered = (
     continuation
     .replace("__IMAGE_BUILD_IF__", rep)

.buildkite/test-amd-merge.yml

@@ -9,16 +9,7 @@ steps:
     - export VLLM_ROCM_USE_AITER=0
     # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
     # that is not supported on CPU
-    - "timeout 20m pytest -v -s -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-branch --cov-report=term-missing --cov-report=html --cov-report=xml"
-
-- label: "Diffusion Model Test"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export GPU_ARCHS=gfx942
-    - timeout 20m pytest -s -v tests/e2e/offline_inference/test_t2i_model.py -m "advanced_model and diffusion" --run-level "advanced_model"
+    - "timeout 60m pytest -v -s -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
 
 - label: "Diffusion Images API LoRA E2E"
   agent_pool: mi325_1
@@ -26,7 +17,7 @@ steps:
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - timeout 20m pytest -s -v tests/e2e/online_serving/test_images_generations_lora.py
+    - timeout 20m pytest -s -v tests/diffusion/lora/test_images_generations_lora.py
 
 - label: "Diffusion Model CPU offloading Test"
   agent_pool: mi325_1
@@ -39,9 +30,9 @@ steps:
     - |
       timeout 20m bash -c '
         set +e
-        pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
+        pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
         EXIT1=\$?
-        pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
+        pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
         EXIT2=\$?
         exit \$((EXIT1 | EXIT2))
       '
@@ -76,7 +67,7 @@ steps:
   commands:
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 20m pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py
+    - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py
     - timeout 20m pytest -s -v tests/diffusion/distributed/test_ulysses_uaa_perf.py
 
 # merge-only tests
@@ -86,7 +77,15 @@ steps:
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - pytest -s -v tests/e2e/offline_inference/test_zimage_parallelism.py
+    - pytest -s -v tests/diffusion/distributed/test_zimage_parallelism.py
+
+- label: "Diffusion Model Test"
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - timeout 30m pytest -s -v tests/e2e/offline_inference/test_z_image.py
 
 - label: "Diffusion GPU Worker Test"
   agent_pool: mi325_2
@@ -124,7 +123,6 @@ steps:
   grade: Blocking
   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 tests/e2e/online_serving/test_qwen3_omni.py tests/e2e/online_serving/test_mimo_audio.py -m "advanced_model" --run-level "advanced_model"
 
 - label: "Qwen3-TTS CustomVoice E2E Test"
@@ -155,17 +153,6 @@ steps:
         pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model" --run-level "advanced_model"
       '
 
-- label: "Diffusion Image Edit Test"
-  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
-    - timeout 20m pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
-
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
 # # split Bagel Model Test with H100 (Real Weights) into three tests
@@ -176,11 +163,10 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
-#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel_text2img.py -m "advanced_model" --run-level "advanced_model" -k "shared_memory" -k "rocm"
+#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "shared_memory" -k "rocm"
 
 # - label: "Bagel Img2Img Model Test (2/3)"
 #   agent_pool: mi325_1
@@ -189,11 +175,10 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
-#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel_img2img.py -m "advanced_model" --run-level "advanced_model" -k "rocm"
+#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "rocm"
 
 # - label: "Bagel Online Serving Test (3/3)"
 #   agent_pool: mi325_1
@@ -202,12 +187,23 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_IMAGE_FETCH_TIMEOUT=60
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
-#     - timeout 40m pytest -s -v tests/e2e/online_serving/test_bagel_online.py -m "advanced_model" --run-level "advanced_model" -k "rocm"
+#     - timeout 40m pytest -s -v tests/e2e/online_serving/test_bagel.py -m "advanced_model" --run-level "advanced_model" -k "rocm"
+
+- label: "Diffusion · Wan22 Test"
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - |
+      timeout 40m bash -c "
+        export VLLM_IMAGE_FETCH_TIMEOUT=60
+        pytest -s -v tests/e2e/offline_inference/test_wan22_t2v.py tests/e2e/online_serving/test_wan22_t2v.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+      "
 
 - label: "Voxtral-TTS E2E Test"
   agent_pool: mi325_1

.buildkite/test-amd-ready.yaml

@@ -9,7 +9,7 @@ steps:
     - export VLLM_ROCM_USE_AITER=0
     # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
     # that is not supported on CPU
-    - "timeout 20m pytest -vvvv -s -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-branch --cov-report=term-missing --cov-report=html --cov-report=xml"
+    - "timeout 60m pytest -v -s -m 'core_model and cpu' --ignore=tests/diffusion/cache/test_teacache_extractors.py --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
 
 - label: "Voxtral TTS CUDA Unit Test"
   agent_pool: mi325_1
@@ -19,21 +19,13 @@ steps:
   commands:
     - timeout 10m pytest -s -v tests/model_executor/models/voxtral_tts/test_cuda_graph_acoustic_transformer.py
 
-- label: "Diffusion Model Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - timeout 30m pytest -s -v tests/e2e/offline_inference/test_t2i_model.py -m "core_model and diffusion" --run-level "core_model"
-
 - label: "Diffusion Batching Test"
   agent_pool: mi325_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - timeout 20m pytest -s -v tests/e2e/offline_inference/test_qwen_image_diffusion_batching.py -m "core_model and diffusion" --run-level "core_model"
+    - timeout 20m pytest -s -v tests/diffusion/batching/test_diffusion_batching.py -m "core_model and diffusion" --run-level "core_model"
 
 - label: "Custom Pipeline Test"
   agent_pool: mi325_1
@@ -54,9 +46,9 @@ steps:
     - |
       timeout 20m bash -c '
         set +e
-        pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
+        pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
         EXIT1=\$?
-        pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
+        pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
         EXIT2=\$?
         exit \$((EXIT1 | EXIT2))
       '
@@ -73,17 +65,6 @@ steps:
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_expansion.py -m "advanced_model and diffusion and L4" --run-level advanced_model
 
-- label: "AudioX Online Test"
-  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
-    - timeout 20m pytest -s -v tests/e2e/online_serving/test_audiox_online.py
-
 - label: "Diffusion Cache Backend Test"
   agent_pool: mi325_1
   depends_on: amd-build
@@ -103,7 +84,7 @@ steps:
   commands:
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 20m pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py -m core_model
+    - timeout 20m pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py -m core_model
 
 - label: "Diffusion GPU Worker Test"
   agent_pool: mi325_2
@@ -113,6 +94,27 @@ steps:
   commands:
     - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
 
+- label: "Diffusion · Model Test"
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_model"
+
+- label: "Diffusion · Wan22 Test"
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - |
+      timeout 40m bash -c "
+        export VLLM_TEST_CLEAN_GPU_MEMORY=1
+        export VLLM_IMAGE_FETCH_TIMEOUT=60
+        pytest -s -v tests/e2e/online_serving/test_wan22_t2v.py -m 'core_model' --run-level 'core_model'
+      "
+
 - label: "Engine Test"
   agent_pool: mi325_1
   depends_on: amd-build
@@ -133,7 +135,6 @@ steps:
 #   commands:
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - timeout 10m pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
 #     - timeout 20m pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_model"
 
@@ -175,17 +176,6 @@ steps:
         pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
       '
 
-- label: "Diffusion Image Edit Test"
-  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
-    - timeout 20m pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
-
 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
 # - label: "Bagel Text2Img Model Test"
@@ -195,11 +185,10 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
-#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel_text2img.py -m "core_model" --run-level "core_model" -k "rocm"
+#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"
 
 # - label: "Bagel Img2Img Model Test"
 #   agent_pool: mi325_1
@@ -208,11 +197,10 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
-#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel_img2img.py -m "core_model" --run-level "core_model" -k "rocm"
+#     - timeout 30m pytest -s -v tests/e2e/offline_inference/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"
 
 # - label: "Bagel Online Serving Test"
 #   agent_pool: mi325_1
@@ -221,12 +209,11 @@ steps:
 #   grade: Blocking
 #   commands:
 #     - export GPU_ARCHS=gfx942
-#     - export VLLM_TEST_CLEAN_GPU_MEMORY=1
 #     - export VLLM_IMAGE_FETCH_TIMEOUT=60
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
 #     - export VLLM_ROCM_USE_AITER_RMSNORM=0
-#     - timeout 40m pytest -s -v tests/e2e/online_serving/test_bagel_online.py -m "core_model" --run-level "core_model" -k "rocm"
+#     - timeout 40m pytest -s -v tests/e2e/online_serving/test_bagel.py -m "core_model" --run-level "core_model" -k "rocm"
 
 - label: "CosyVoice3-TTS E2E Test"
   agent_pool: mi325_1
@@ -236,5 +223,5 @@ steps:
   commands:
     - |
       timeout 20m bash -c '
-        pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts.py -m "core_model" --run-level "core_model"
+        pytest -s -v tests/e2e/online_serving/test_cosyvoice3_tts_expansion.py -m "full_model" --run-level "core_model"
       '

.buildkite/test-amd.yaml

@@ -8,7 +8,7 @@ steps:
   grade: Blocking
   commands:
     - export GPU_ARCHS=gfx942
-    - pytest -s -v tests/e2e/offline_inference/test_t2i_model.py
+    - pytest -s -v tests/e2e/offline_inference/test_z_image.py
 
 - label: "Diffusion Images API LoRA E2E"
   timeout_in_minutes: 30
@@ -20,7 +20,7 @@ steps:
     - 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
+    - pytest -s -v tests/diffusion/lora/test_images_generations_lora.py
 
 - label: "Diffusion Model CPU offloading Test"
   timeout_in_minutes: 20
@@ -32,8 +32,8 @@ steps:
     - 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
-    - pytest -s -v tests/e2e/offline_inference/test_diffusion_layerwise_offload.py
+    - pytest -s -v tests/diffusion/offloader/test_diffusion_cpu_offload.py
+    - pytest -s -v tests/diffusion/offloader/test_diffusion_layerwise_offload.py
 
 - label: "Diffusion Cache Backend Test"
   timeout_in_minutes: 15
@@ -57,7 +57,7 @@ steps:
     - 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 -m core_model
+    - pytest -s -v tests/diffusion/distributed/test_sequence_parallel.py -m core_model
 
 - label: "Diffusion Tensor Parallelism Test"
   timeout_in_minutes: 20
@@ -67,7 +67,7 @@ steps:
     - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -s -v tests/e2e/offline_inference/test_zimage_parallelism.py
+    - pytest -s -v tests/diffusion/distributed/test_zimage_parallelism.py
 
 - label: "Diffusion GPU Worker Test"
   timeout_in_minutes: 20
@@ -100,25 +100,11 @@ steps:
   commands:
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
     - export VLLM_ROCM_USE_AITER=0
     - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
     - pytest -s -v tests/e2e/online_serving/test_qwen3_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
-
-
 - label: "Omni Sleep Mode Test"
   timeout_in_minutes: 40
   agent_pool: mi325_2
@@ -129,5 +115,4 @@ steps:
     - export GPU_ARCHS=gfx942
     - 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_omni_sleep_mode.py -m "advanced_model and omni and MI325" --run-level "advanced_model"
+    - pytest -s -v tests/entrypoints/test_omni_sleep_mode.py -m "advanced_model and omni and MI325" --run-level "advanced_model"

.buildkite/test-nightly.yml

@@ -1,5 +1,6 @@
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -13,7 +14,7 @@ steps:
       - label: ":full_moon: Omni · Function Test with H100"
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/ -m "full_model and H100 and omni" --run-level "full_model" --ignore=tests/e2e/accuracy
+          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -169,14 +170,104 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Omni · Perf Test"
-        key: nightly-omni-performance
+      - label: ":full_moon: Omni · Perf Test · No Async Chunk"
+        key: nightly-omni-performance-no-async-chunk
         timeout_in_minutes: 180
         commands:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_omni.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/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
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                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: ":full_moon: Omni · Perf Test · Async Chunk"
+        key: nightly-omni-performance-async-chunk
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/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
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                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: ":full_moon: Omni · Perf Test · vLLM Text"
+        key: nightly-omni-performance-vllm-text
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
@@ -218,7 +309,7 @@ steps:
       - label: ":full_moon: Omni · Multi-Replica Startup Test with 4x H100"
         timeout_in_minutes: 45
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py -m "core_model" --run-level "core_model"
+          - pytest -s -v tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py -m "full_model" --run-level "core_model"
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -262,7 +353,7 @@ steps:
         timeout_in_minutes: 90
         commands:
           - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-          - pytest -s -v tests/e2e/ -m "full_model and L4 and omni" -k "not test_qwen2_5" --run-level "full_model" --ignore=tests/e2e/accuracy
+          - pytest -s -v tests/e2e/ -m "full_model and L4 and tts" --run-level "full_model" --ignore=tests/e2e/accuracy
         agents:
           queue: "gpu_1_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
         plugins:
@@ -298,7 +389,7 @@ steps:
                   - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
                     resources:
                       limits:
-                        nvidia.com/gpu: 1
+                        nvidia.com/gpu: 2
                     volumeMounts:
                       - name: devshm
                         mountPath: /dev/shm
@@ -332,13 +423,65 @@ steps:
       build.pull_request.labels includes "nightly-test" ||
       build.pull_request.labels includes "diffusion-x2iat-test"
     steps:
-      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100"
+      # Use explicit file shards so single-card cases do not reserve 2+ GPUs,
+      # and Bagel feature coverage is not duplicated via a broad tests/e2e sweep.
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Single-GPU"
         timeout_in_minutes: 120
         commands:
-          - pytest -sv tests/e2e/ -k "not test_wan and not test_bagel_expansion and not hunyuan" -m "full_model and diffusion and H100" --run-level "full_model" --ignore=tests/e2e/accuracy
-          # revoke after bagel optimization finished
-          # Keep Bagel expansion and multi-replica tests in their dedicated H100 jobs below.
-          # - pytest -sv tests/e2e/ -k "not test_wan and not test_bagel_expansion and not hunyuan" -m "full_model and diffusion and H100" --run-level "full_model" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_bagel_multi_replicas.py
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_qwen_image_expansion.py
+            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
+            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
+            tests/e2e/online_serving/test_longcat_image_expansion.py
+            tests/e2e/online_serving/test_longcat_image_edit_expansion.py
+            tests/e2e/online_serving/test_flux_2_dev_expansion.py
+            tests/e2e/online_serving/test_bagel_expansion.py
+            -m "full_model and diffusion and H100 and not distributed_cuda"
+            --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/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
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                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: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Qwen-Image"
+        timeout_in_minutes: 120
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_qwen_image_expansion.py
+            tests/e2e/online_serving/test_qwen_image_edit_expansion.py
+            -m "full_model and diffusion and H100 and distributed_cuda"
+            --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -373,10 +516,59 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · Bagel Function Test with H100"
+      - label: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Layered/LongCat/FLUX"
+        timeout_in_minutes: 90
+        commands:
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_qwen_image_layered_expansion.py
+            tests/e2e/online_serving/test_longcat_image_expansion.py
+            tests/e2e/online_serving/test_longcat_image_edit_expansion.py
+            tests/e2e/online_serving/test_flux_2_dev_expansion.py
+            -m "full_model and diffusion and H100 and distributed_cuda"
+            --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/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
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                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: ":full_moon: Diffusion X2I(&A&T) · Function Test with H100 · Multi-GPU Bagel"
         timeout_in_minutes: 120
         commands:
-          - pytest -sv tests/e2e/online_serving/test_bagel_expansion.py -m "full_model and diffusion and H100" --run-level "full_model"
+          - >-
+            pytest -sv
+            tests/e2e/online_serving/test_bagel_expansion.py
+            -m "full_model and diffusion and H100 and distributed_cuda"
+            --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -411,44 +603,6 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      # revoke after bagel optimization finished
-      # - label: ":full_moon: Diffusion X2I(&A&T) · BAGEL Multi-Replica with H100"
-      #   timeout_in_minutes: 120
-      #   commands:
-      #     - pytest -s -v tests/e2e/online_serving/test_bagel_multi_replicas.py -m "full_model and diffusion and H100" --run-level "full_model"
-      #   agents:
-      #     queue: "mithril-h100-pool"
-      #   plugins:
-      #     - kubernetes:
-      #         podSpec:
-      #           containers:
-      #             - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-      #               resources:
-      #                 limits:
-      #                   nvidia.com/gpu: 4
-      #               volumeMounts:
-      #                 - name: devshm
-      #                   mountPath: /dev/shm
-      #                 - name: hf-cache
-      #                   mountPath: /root/.cache/huggingface
-      #               env:
-      #                 - name: HF_HOME
-      #                   value: /root/.cache/huggingface
-      #                 - name: HF_TOKEN
-      #                   valueFrom:
-      #                     secretKeyRef:
-      #                       name: hf-token-secret
-      #                       key: token
-      #           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: ":full_moon: Diffusion X2I(&A&T) · Function Test with L4"
         timeout_in_minutes: 60
@@ -471,7 +625,6 @@ steps:
       - label: ":full_moon: Diffusion X2I(&A&T) · Doc Test"
         timeout_in_minutes: 60
         commands:
-          - export VLLM_TEST_CLEAN_GPU_MEMORY="1"
           - pytest -s -v tests/examples/*/test_text_to_image.py -m "full_model and example and H100" --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
@@ -507,87 +660,6 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · GEBench Accuracy Test"
-        timeout_in_minutes: 60
-        commands:
-          - pytest -s -v tests/e2e/accuracy/test_gebench_h100_smoke.py --run-level full_model --gebench-model Qwen/Qwen-Image-2512 --accuracy-judge-model QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ --accuracy-gpu 0 --gebench-port 8093 --accuracy-workers 1
-          - buildkite-agent artifact upload "tests/e2e/accuracy/artifacts/gebench_qwen-image-2512/summary*.json"
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/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
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                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: ":full_moon: Diffusion X2I(&A&T) · GEdit-Bench Accuracy Test"
-        timeout_in_minutes: 60
-        commands:
-          - pytest -s -v tests/e2e/accuracy/test_gedit_bench_h100_smoke.py --run-level full_model --gedit-model Qwen/Qwen-Image-Edit --accuracy-judge-model QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ --accuracy-gpu 0 --gedit-port 8093 --gedit-samples-per-group 20 --accuracy-workers 1
-          - buildkite-agent artifact upload "tests/e2e/accuracy/artifacts/gedit_scores_qwen-image-edit/qwen-image-edit_all_all_vie_score_*.csv"
-          - buildkite-agent artifact upload "tests/e2e/accuracy/artifacts/gedit_scores_qwen-image-edit/qwen-image-edit_all_all_summary_*.json"
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/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
-                      - name: VLLM_HTTP_TIMEOUT_KEEP_ALIVE
-                        value: "120"
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                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: ":full_moon: Diffusion X2I(&A&T) · Accuracy Test"
         timeout_in_minutes: 180
         commands:
@@ -629,28 +701,20 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test"
-        key: nightly-diffusion-x2iat-performance
-        timeout_in_minutes: 360
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image"
+        key: nightly-diffusion-x2iat-performance-qwen-image
+        timeout_in_minutes: 180
         commands:
           - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
           - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
           - export CACHE_DIT_VERSION=1.3.0
-          # [HACK]: run upload in the same command block as pytest.
-          # Because `exit` aborts the entire commands list.
           - |
             set +e
             pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_vllm_omni.json
-            EXIT1=$$?
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_vllm_omni.json
-            EXIT2=$$?
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_2509_vllm_omni.json
-            EXIT3=$$?
-            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_layered_vllm_omni.json
-            EXIT4=$$?
+            EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
             buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
-            exit $$((EXIT1 | EXIT2 | EXIT3 | EXIT4))
+            exit $$EXIT
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -685,6 +749,150 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
+      - label: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit"
+        key: nightly-diffusion-x2iat-performance-qwen-image-edit
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - export CACHE_DIT_VERSION=1.3.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 4
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                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: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Edit-2509"
+        key: nightly-diffusion-x2iat-performance-qwen-image-edit-2509
+        timeout_in_minutes: 180
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - export CACHE_DIT_VERSION=1.3.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_edit_2509_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 4
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                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: ":full_moon: Diffusion X2I(&A&T) · Perf Test · Qwen-Image-Layered"
+        key: nightly-diffusion-x2iat-performance-qwen-image-layered
+        timeout_in_minutes: 120
+        commands:
+          - export DIFFUSION_BENCHMARK_DIR=tests/dfx/perf/results
+          - export DIFFUSION_ATTENTION_BACKEND=FLASH_ATTN
+          - export CACHE_DIT_VERSION=1.3.0
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen_image_layered_vllm_omni.json
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/diffusion_result_*.json"
+            buildkite-agent artifact upload "tests/dfx/perf/results/logs/*.log"
+            exit $$EXIT
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/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
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                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: ":full_moon: Diffusion X2I(&A&T) · HunyuanImage3 · DiT Perf Test"
         key: nightly-hunyuan-image3-performance
         soft_fail: true
@@ -751,7 +959,7 @@ steps:
       - label: ":full_moon: Diffusion X2V · Function Test"
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_hunyuan_video_15_expansion.py  --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_hunyuan_video_15_expansion.py  -m "full_model and cuda" --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -790,7 +998,7 @@ steps:
         timeout_in_minutes: 180
         commands:
           - pytest -s -v tests/e2e/accuracy/wan22_i2v/test_wan22_i2v_video_similarity.py -m full_model --run-level full_model
-          - pytest -s -v tests/e2e/accuracy/test_ltx2_3_video_similarity.py -m advanced_model --run-level advanced_model
+          - pytest -s -v tests/e2e/accuracy/test_ltx2_3_video_similarity.py -m full_model --run-level full_model
           - pytest -s -v tests/e2e/accuracy/test_diffusers_backend_similarity.py -k '2v_matches_diffusers' -m full_model --run-level full_model
         agents:
           queue: "mithril-h100-pool"
@@ -873,6 +1081,65 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
+  - group: ":card_index_dividers: Diffusion · Quantization Test"
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "diffusion-x2iat-test"
+    steps:
+      - label: ":full_moon: Quantization Test with H100"
+        timeout_in_minutes: 90
+        commands:
+          - timeout 30m pytest -sv tests/diffusion/quantization -m 'full_model and cuda and H100' --run-level "full_model"
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 1
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: ":full_moon: Quantization Test with L4"
+        timeout_in_minutes: 60
+        commands:
+          - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and L4' --run-level "full_model"
+        agents:
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
   - label: ":bar_chart: Testcase Statistics"
     key: nightly-testcase-statistics
     timeout_in_minutes: 120
@@ -890,7 +1157,7 @@ steps:
               - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
                 resources:
                   limits:
-                    nvidia.com/gpu: 2
+                    nvidia.com/gpu: 1
                 volumeMounts:
                   - name: devshm
                     mountPath: /dev/shm
@@ -919,9 +1186,14 @@ steps:
   - label: ":email: Nightly Collection & Email"
     key: nightly-perf-distribution
     depends_on:
-      - nightly-omni-performance
+      - nightly-omni-performance-no-async-chunk
+      - nightly-omni-performance-async-chunk
+      - nightly-omni-performance-vllm-text
       - nightly-tts-performance
-      - nightly-diffusion-x2iat-performance
+      - nightly-diffusion-x2iat-performance-qwen-image
+      - nightly-diffusion-x2iat-performance-qwen-image-edit
+      - nightly-diffusion-x2iat-performance-qwen-image-edit-2509
+      - nightly-diffusion-x2iat-performance-qwen-image-layered
       - nightly-diffusion-x2v-performance
       - nightly-testcase-statistics
     if: build.env("NIGHTLY") == "1"
@@ -930,8 +1202,13 @@ steps:
       - export DEFAULT_INPUT_DIR=tests/dfx/perf/results
       - export DEFAULT_OUTPUT_DIR=tests/dfx/perf/results
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-tts-performance
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance
-      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-no-async-chunk
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-async-chunk
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-omni-performance-vllm-text
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-edit
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-edit-2509
+      - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2iat-performance-qwen-image-layered
       - buildkite-agent artifact download "tests/dfx/perf/results/*.json" . --step nightly-diffusion-x2v-performance
       - buildkite-agent artifact download "tests/dfx/perf/results/*.html" . --step nightly-testcase-statistics
       - python tools/nightly/generate_nightly_perf_excel.py

.buildkite/test-npu-nightly.yaml

@@ -0,0 +1,24 @@
+steps:
+  - group: ":card_index_dividers: Diffusion X2V Model Test"
+    key: nightly-diffusion-x2v-group
+    depends_on: upload-nightly-pipeline
+    if: >-
+      build.env("NIGHTLY") == "1" ||
+      build.pull_request.labels includes "nightly-test" ||
+      build.pull_request.labels includes "diffusion-x2v-test"
+    steps:
+      - label: ":full_moon: Diffusion X2V · Function Test"
+        agents:
+          queue: "ascend-a2b3"
+          resource_class: "npu-4"
+        image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}"
+        plugins:
+          - kubernetes:
+              podSpecPatch:
+                imagePullSecrets:
+                  - name: swr-secret
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py -k npu_i2v --run-level "full_model"

.buildkite/test-npu-ready.yaml

@@ -0,0 +1,21 @@
+steps:
+  - group: ":card_index_dividers: TTS Model Test"
+    key: ready-tts-group
+    depends_on: upload-ready-pipeline
+    steps:
+      - label: ":sound: TTS · Qwen3-TTS CustomVoice Test"
+        agents:
+          queue: "ascend-a2b3"
+          resource_class: "npu-1"
+        image: "swr.cn-southwest-2.myhuaweicloud.com/modelfoundry/vllm-omni-ci-npu:${BUILDKITE_COMMIT}"
+        plugins:
+          - kubernetes:
+              podSpecPatch:
+                imagePullSecrets:
+                  - name: swr-secret
+        env:
+          VLLM_WORKER_MULTIPROC_METHOD: spawn
+          VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+          HF_TOKEN: "${HF_TOKEN}"
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'

.buildkite/test-weekly.yml

@@ -1,5 +1,6 @@
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 

.claude/skills/add-diffusion-model/SKILL.md

@@ -515,15 +515,15 @@ After adding parallelism support, update:
 
 ### Step 11: Add CPU Offload Support
 
-Implement `SupportsModuleOffload` on your pipeline class to enable
+Implement `SupportsComponentDiscovery` on your pipeline class to enable
 `--enable-cpu-offload` and `--enable-layerwise-offload`. The protocol
 declares which submodules the offloader should manage:
 
 ```python
 from typing import ClassVar
-from vllm_omni.diffusion.models.interface import SupportsModuleOffload
+from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
 
-class YourPipeline(nn.Module, SupportsModuleOffload):
+class YourPipeline(nn.Module, SupportsComponentDiscovery):
     _dit_modules: ClassVar[list[str]] = ["transformer"]
     _encoder_modules: ClassVar[list[str]] = ["text_encoder"]
     _vae_modules: ClassVar[list[str]] = ["vae"]
@@ -540,7 +540,7 @@ class YourPipeline(nn.Module, SupportsModuleOffload):
 All attribute names support dotted paths for nested submodules
 (e.g. `"pipe.transformer"`, `"bagel.time_embedder"`).
 
-Pipelines without `SupportsModuleOffload` fall back to scanning
+Pipelines without `SupportsComponentDiscovery` fall back to scanning
 well-known attribute names (`transformer`, `text_encoder`, `vae`,
 etc.), which fails for non-standard names.
 

.claude/skills/add-diffusion-model/references/parallelism-patterns.md

@@ -92,6 +92,7 @@ class TPSelfAttention(nn.Module):
             softmax_scale=1.0 / (self.head_dim ** 0.5),
             causal=False,
             num_kv_heads=self.to_qkv.num_kv_heads,  # Local KV heads per GPU
+            role="self",
         )
 
     def forward(self, x):

.claude/skills/add-diffusion-model/references/transformer-adaptation.md

@@ -61,6 +61,7 @@ class SelfAttentionBlock(nn.Module):
             softmax_scale=1.0 / (self.head_dim ** 0.5),
             causal=False,
             num_kv_heads=num_heads,
+            role="self",
         )
 
     def forward(self, x, attn_mask=None):
@@ -75,6 +76,24 @@ class SelfAttentionBlock(nn.Module):
         return self.to_out(out)
 ```
 
+**Declaring the attention role.** Always pass a `role` string to `Attention(...)`. It is the lookup key users target with `--diffusion-attention-config.per_role.<role>.backend` to swap kernels per site without touching model code.
+
+| Convention | When to use |
+|---|---|
+| `role="self"` | Q/K/V come from the same hidden state |
+| `role="cross"` | K/V come from a separate `encoder_hidden_states`; pair with `skip_sequence_parallel=True` if K/V is replicated across SP ranks |
+
+For multi-modal sites that don't fit `self` / `cross`, use a dot-namespaced role and provide `role_category` so user config can fall back to the generic category:
+
+```python
+self.audio_to_video_attn = Attention(
+    num_heads=num_heads, head_size=self.head_dim,
+    softmax_scale=1.0 / (self.head_dim ** 0.5), causal=False,
+    role="mymodel.audio_to_video",
+    role_category="cross",
+)
+```
+
 #### Fused QKV with TP (Advanced)
 
 For tensor parallelism, use vLLM's parallel linear layers:
@@ -104,6 +123,7 @@ class TPSelfAttention(nn.Module):
             softmax_scale=1.0 / (self.head_dim ** 0.5),
             causal=False,
             num_kv_heads=num_heads,
+            role="self",
         )
 ```
 

.claude/skills/add-diffusion-model/references/troubleshooting.md

@@ -68,7 +68,7 @@ out = out.reshape(batch, seq_len, -1)
 ### Different Output vs Diffusers Reference
 
 **Common causes**:
-1. **Attention backend difference**: FlashAttention vs SDPA may produce slightly different results. Set `DIFFUSION_ATTENTION_BACKEND=TORCH_SDPA` to match diffusers
+1. **Attention backend difference**: FlashAttention vs SDPA may produce slightly different results. Force SDPA via `--diffusion-attention-backend TORCH_SDPA` (or `DIFFUSION_ATTENTION_BACKEND=TORCH_SDPA`) to match diffusers; for finer-grained per-role overrides see `--diffusion-attention-config.per_role.<role>.backend`
 2. **Float precision**: vLLM-Omni may use bfloat16 where diffusers uses float32 for some operations
 3. **Missing normalization**: Check all LayerNorm/RMSNorm are preserved
 4. **Scheduler rounding**: Some schedulers have numerical sensitivity

.claude/skills/precheck-pr/SKILL.md

@@ -0,0 +1,97 @@
+---
+name: precheck-pr
+description: Self-check your branch before creating a PR — catch dead code, verify accuracy/perf claims, validate PR title format, and confirm merge readiness. Use when the user says "precheck", "self review", "pre-submit check", or "check my PR before I open it." Never posts to GitHub.
+---
+
+# PR Pre-Check
+
+Self-review your branch before creating a PR against `vllm-project/vllm-omni`. Two modes: **quick** catches showstoppers, **full** does a thorough maintainer-grade review. Never posts to GitHub; the report is for the contributor's terminal only.
+
+## Mode Selection
+
+| Mode | When | Time |
+|------|------|------|
+| **Quick** | About to push, final sanity check | ~3 min |
+| **Full** | Ready for review, want maintainer-level scan | ~10 min |
+
+Default to quick if unsure. Run full before marking a PR "ready for review."
+
+## Workflow
+
+### Step 1: Detect Base Branch
+
+```bash
+BASE_SHA=$(git merge-base HEAD origin/main 2>/dev/null \
+         || git merge-base HEAD main 2>/dev/null \
+         || echo origin/main)
+echo "diffing against ${BASE_SHA}"
+git diff --name-only ${BASE_SHA}...HEAD
+```
+
+### Step 2: Validate PR Title
+
+Check the most recent commit message (or branch name if no commit yet) against the project convention. Valid prefixes:
+
+| Prefix | Applies to |
+|--------|-----------|
+| `[Bugfix]` | Bug fixes |
+| `[CI/Build]` | Build or CI improvements |
+| `[Doc]` | Documentation changes |
+| `[Model]` | New/improved models (include model name) |
+| `[Frontend]` | Frontend changes (API server, OmniLLM class, etc.) |
+| `[Kernel]` | CUDA/kernel changes |
+| `[Core]` | Core logic changes (OmniProcessor, OmniARScheduler, etc.) |
+| `[Hardware][Vendor]` | Hardware-specific (e.g., `[Hardware][Ascend]`) |
+| `[Misc]` | Other changes (use sparingly) |
+
+✗ if: missing prefix, wrong case (`[bugfix]`), or WIP/Draft in title.
+⚠ if: `[Model]` prefix without the model identifier (e.g., `[Model] Add new model` — should be `[Model] Add <ModelName> ...`).
+
+### Step 3: Categorize the PR
+
+| Diff contains | PR type |
+|---------------|---------|
+| New files under `vllm_omni/model_executor/models/<name>/` | **New Model** |
+| Changes to `vllm_omni/diffusion/` | **Diffusion Model** |
+| `[Bugfix]` prefix or single-file fix | **Bug Fix** |
+| Perf/benchmark/throughput claims in commit msg or diff | **Performance** |
+| Everything else | **General** |
+
+If multiple rows apply (e.g., a diffusion model is also a new model), union the checklists.
+
+### Step 4: Run Checklist
+
+Ask: "Quick mode or full mode?" Then walk the checklist for the detected PR type from [references/checklists.md](references/checklists.md). Each item produces ✓, ✗, or ⚠.
+
+### Step 5: Print Report
+
+```
+Pre-check report for <branch>
+
+  Mode: quick | full
+  Type: <new-model | diffusion-model | bug-fix | perf | general>
+
+  Dimension          Result
+  ─────────────────  ──────
+  PR title format    ✓
+  PR desc integrity  ✓
+  Registry/config    ✓
+  Dead code          ⚠ 2 warnings
+  Accuracy           ✓
+  Benchmark          ✗ missing software versions
+
+  Verdict: 1 blocking | 2 warnings | recommend fixing ✗ before PR
+```
+
+**Severity:**
+
+| Mark | Meaning |
+|------|---------|
+| ✗ | Blocking — fix before opening PR |
+| ⚠ | Warning — consider fixing |
+| ✓ | Pass |
+| — | Skipped (not applicable) |
+
+### Stop Here
+
+Do not post comments, open PRs, or modify files. The report is for the contributor's terminal only.

.claude/skills/precheck-pr/references/checklists.md

@@ -0,0 +1,135 @@
+# Pre-Check Checklists
+
+Each item is a concrete check to run against the diff. ✗ = fix before PR. ⚠ = consider fixing.
+
+---
+
+## All PRs: Title Format (check first)
+
+PR title must follow the project convention documented in `docs/contributing/README.md`:
+
+| Prefix | Applies to |
+|--------|-----------|
+| `[Bugfix]` | Bug fixes |
+| `[CI/Build]` | Build or CI improvements |
+| `[Doc]` | Documentation changes |
+| `[Model]` | New/improved models (include model name) |
+| `[Frontend]` | Frontend changes (API server, OmniLLM class, etc.) |
+| `[Kernel]` | CUDA/kernel changes |
+| `[Core]` | Core logic changes (OmniProcessor, OmniARScheduler, etc.) |
+| `[Hardware][Vendor]` | Hardware-specific (e.g., `[Hardware][Ascend]`) |
+| `[Misc]` | Other changes (use sparingly) |
+
+**Examples:**
+```
+✓ [Model] Add MiniCPM-o 4.5 support
+✓ [Bugfix] Fix Ovis image text encoder dtype
+✓ [Core] Qwen3-Omni performance optimization
+✓ [Hardware][Ascend] Fix VoxCPM2 on Ascend
+✗ Fix Ovis image text encoder dtype              ← missing prefix
+✗ [bugfix] Fix Ovis image text encoder           ← wrong case
+✗ [WIP] Add model                                ← WIP in title
+```
+
+- [ ] **Title has valid prefix** — ✗ if missing prefix, wrong case, or WIP/Draft in title. For `[Model]` PRs, the title MUST also include the model identifier (e.g., `[Model] Add <ModelName> ...`). ⚠ if missing the model name.
+
+---
+
+## Bug Fix PRs
+
+### Quick
+
+- [ ] **PR body has 4 sections:** what broke (error message), repro steps, root cause, fix description
+- [ ] **Fix is minimal:** diff touches only files directly related to the bug. No unrelated refactors or style changes.
+- [ ] **Repro is runnable:** the PR body includes a copy-paste command that reproduces the bug
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+
+### Full (adds)
+
+- [ ] **Regression test exists:** `git diff --name-only` includes at least one `tests/` file
+- [ ] **Regression test reproduces the original error:** the test would fail on main, pass on this branch
+- [ ] **Fix matches root cause exactly:** no "fix the symptom + something else" — if the root cause is dtype, the fix is dtype, not dtype + reformatting
+- [ ] **No silent failure risk:** no bare `except: pass`, no `try: ... except: return None`, no empty fallback added
+- [ ] **Upstream pattern match:** the fix follows the same pattern as existing code for similar cases (grep for analogous `from_pretrained` calls, etc.)
+- [ ] **Environment documented:** torch/transformers/vllm versions listed if the bug is version-dependent
+
+---
+
+## Performance PRs
+
+### Quick
+
+- [ ] **Before/after numbers in PR body:** at minimum, one metric (RTF, RPS, latency, VRAM) with main vs PR values
+- [ ] **Hardware specified:** GPU model, count, VRAM. "L4" or "H20" is sufficient. "tested on GPU" is not.
+- [ ] **Concurrency scaling:** if the optimization claims better throughput, numbers at 2+ concurrency levels
+- [ ] **No unexplained regressions:** if any metric goes down at any concurrency level, explain why
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+
+### Full (adds)
+
+- [ ] **Software versions:** torch, CUDA, vllm, vllm-omni versions in PR body
+- [ ] **Warmup stated:** "1 warmup + 3 measured" or equivalent
+- [ ] **Stddev or range:** not just a single best run. "average of 2 rounds" is minimum.
+- [ ] **VRAM measured:** peak GPU memory before and after
+- [ ] **Benchmark script checked in:** under `tests/` or `examples/`, runnable by reviewer
+- [ ] **Exact command line in PR:** not pseudocode, not "run the benchmark"
+- [ ] **Pytest summary included:** `36 passed, 3 skipped, 17 warnings in 2897.66s` — proves the whole suite passes
+- [ ] **Config changes documented:** if `max_num_seqs`, `gpu_memory_utilization`, or deploy YAML changed, old and new values stated with reasoning
+
+---
+
+## New Model PRs
+
+### Quick
+
+- [ ] **PR body file list matches `git diff --name-only`:** every file path in the description exists in the diff
+- [ ] **Registry entries match claimed count:** count `_OMNI_MODELS` or `_DIFFUSION_MODELS` entries added vs claimed in body
+- [ ] **Pipeline config `custom_process_input_func` paths resolve:** each referenced module.function exists
+- [ ] **`__init__.py` exports match registry:** each registered class appears in `__init__.py` and `__all__`
+- [ ] **Shell script modes match README:** `run_curl.sh` supports every mode the README documents
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+
+### Full (adds)
+
+- [ ] **Dead code scan (5 patterns):**
+  1. Dead `forward()` — defined but never called through the pipeline path
+  2. Dead factory/builder functions — instantiated nowhere
+  3. Dead wrappers — thin wrappers with no callers
+  4. Dead branch guards — `if version > X` where X is always satisfied
+  5. Unused parameters — `__init__` args never accessed
+- [ ] **Copy-paste detection:** no string constant defined in 3+ files, no cross-module validation duplication, no near-identical shape coercion fns
+- [ ] **Import hygiene:** no re-export-only imports, no module-level side effects without docs, no `import os` in function body when already at top. Function-body imports are acceptable when the import is only reachable from a closure or factory return value (i.e., `import` would otherwise execute at module load even though the user may never trigger that path). Either lift to top-of-file or annotate with `# noqa: PLC0415`.
+- [ ] **Accuracy:** at least one test compares output against reference; audio = valid WAV at correct sample rate; determinism verified if seed claimed. ⚠ if missing and not explicitly deferred to a follow-up PR or RFC section; — if explicitly deferred (the deferral target SHOULD be linked in the PR body).
+- [ ] **Performance:** hardware, software versions, warmup, and metrics (RTF, VRAM, RPS, latency) stated. ⚠ if missing and not explicitly deferred; — if explicitly deferred.
+- [ ] **Benchmark settings:** model config (TP, PP, max_model_len, enforce_eager, quant), runtime config (batch, input/output spec), environment versions
+- [ ] **Benchmark script checked in + exact command line + pytest summary**
+
+---
+
+## Diffusion Model PRs
+
+Diffusion models live under `vllm_omni/diffusion/`. In addition to the New Model checklist above, verify:
+
+- [ ] **Transformer adapter follows contract:** inherits from `nn.Module`, implements `load_weights()`, uses `vllm_omni.diffusion.attention.layer.Attention`
+- [ ] **Pipeline has both offline and online paths:** `forward()` for batch generation, streaming path for online serving
+- [ ] **Cache-DiT integration:** at least one acceleration (TeaCache, FBCache, etc.) is wired up. ✗ for new **top-level** pipelines. For **subclass** pipelines that inherit `__call__` from a parent where Cache-DiT is already absent, downgrade to ⚠ and recommend filing a follow-up issue against the parent pipeline family.
+- [ ] **Parallelism config:** TP/SP/USP/CFG-Parallel options exposed in pipeline config
+- [ ] **Docs table updated:** model added to `docs/models/supported_models.md` with correct metadata
+- [ ] **E2E test exists:** at least one test exercises the full generate path
+
+---
+
+## General PRs
+
+### Quick
+
+- [ ] **PR body matches diff:** description doesn't claim changes not in the diff
+- [ ] **Branch is rebased:** `git merge-base HEAD origin/main` is recent, no merge conflicts
+- [ ] **CI gates passing:** DCO, pre-commit, build — check `gh pr view --json statusCheckRollup` if a PR already exists
+
+### Full (adds)
+
+- [ ] **No new dead code in changed files:** scan for unreachable paths, unused imports
+- [ ] **Test coverage:** if core code changed (`engine/`, `stages/`, `connectors/`), tests are added or existing tests cover the path
+- [ ] **Import hygiene:** no redundant imports added
+- [ ] **No unrelated changes:** diff scope matches PR description

.claude/skills/readme.md

@@ -22,7 +22,12 @@ include:
 - `add-tts-model`: covers integration of new TTS models and related serving
   workflows
 - `generate-release-note`: helps prepare release notes for repository changes
+- `precheck-pr`: self-check a branch before creating a PR — validates PR title
+  format, catches dead code, verifies accuracy/perf claims, and confirms merge
+  readiness
 - `review-pr`: provides a structured workflow for reviewing pull requests
+- `vllm-omni-npu-model-runner-upgrade`: upgrades NPU model runners to align with the
+  latest vllm-ascend NPUModelRunner
 
 ## Maintenance Guidelines
 

.github/CODEOWNERS

@@ -3,15 +3,64 @@
 # More info: https://docs.github.com/en/repositories/managing-your-repositorys-settings-and-features/customizing-your-repository/about-code-owners
 
 # Default owners (fallback reviewers for all files)
-# later, we can add more reviewers here
-* @hsliuustc0106
+* @Gaohan123 @hsliuustc0106 @ywang96
 
-# Assign reviewers by directory
-/.github/ @owner1
-/tests/ @owner1
-/docs/ @owner1
+# Repository operations
+/.github/ @hsliuustc0106
 
-# Example: reviewers for specific directories
-# /examples/offline_inference/ @offline-inference-team
-# /vllm_omni/core/ @core-team
-# /vllm_omni/model_executor/ @model-team
+# Documentation and Recipes
+/docs/ @hsliuustc0106 @Gaohan123 @david6666666
+/recipes @hsliuustc0106 @Gaohan123 @david6666666
+
+# Skills
+/.claude/ @hsliuustc0106
+
+# Tests
+/tests/ @yenuo26
+
+# CI
+/.buildkite/ @yenuo26 @congw729
+
+# Engine and model_executor
+/vllm_omni/core/ @tzhouam @yinpeiqi @fake0fan
+/vllm_omni/engine/ @tzhouam @yinpeiqi @fake0fan
+/vllm_omni/entrypoints/ @tzhouam @yinpeiqi @fake0fan
+/vllm_omni/model_executor/ @tzhouam @gcanlin
+/vllm_omni/worker/ @tzhouam @gcanlin
+
+# Diffusion
+/vllm_omni/diffusion/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @david6666666
+/vllm_omni/lora/ @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @yuanheng-zhao
+
+# Diffusion CPU offloading
+/vllm_omni/diffusion/offloader/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa @yuanheng-zhao
+
+# Omni and TTS support
+/vllm_omni/transformers_utils/ @tzhouam @yuanheng-zhao
+/vllm_omni/model_executor/models/ @linyueqian @ZeldaHuang @princepride @yuanheng-zhao @amy-why-3459 @Bounty-hunter
+
+# Benchmarks
+/benchmarks/accuracy/ @david6666666
+/benchmarks/build_dataset/ @linyueqian @yuanheng-zhao
+/benchmarks/diffusion/ @Isotr0py @princepride @SamitHuang @wtomin @ZJY0516 @RuixiangMa
+/benchmarks/distributed/ @princepride
+/benchmarks/glm_image/ @princepride @JaredforReal
+/benchmarks/tts/ @linyueqian @ZeldaHuang @princepride
+
+# Usability and Observability
+/vllm_omni/deploy/ @tzhouam @lishunyang12
+/vllm_omni/metrics/ @lishunyang12
+/vllm_omni/profiler/ @gcanlin
+
+# ComfyUI
+/apps/ @fhfuih
+
+# Large Scale Serving
+/vllm_omni/distributed/ @princepride @yuanheng-zhao @xuechendi @natureofnature @fake0fan
+
+# Quantization and configuration
+/vllm_omni/config/ @lishunyang12
+/vllm_omni/quantization/ @david6666666 @Isotr0py @lishunyang12
+
+# Hardware plugins and accelerators
+/vllm_omni/platforms/ @gcanlin @tjtanaa @xuechendi

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,4 +1,6 @@
 <!-- markdownlint-disable -->
+**Before submitting:** run the precheck-pr skill with code agent for a self-check against project conventions — catches dead code, missing benchmarks, and title format issues.
+
 PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.
 
 ## Purpose
@@ -18,4 +20,5 @@ PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTT
 - [ ] (Optional) Release notes update. If your change is user-facing, please update the release notes draft.
 </details>
 
-**BEFORE SUBMITTING, PLEASE READ <https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md>** (anything written below this line will be removed by GitHub Actions)
+**BEFORE SUBMITTING,** please read <https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md>
+(anything written below this line will be removed by GitHub Actions)

.gitignore

@@ -239,6 +239,7 @@ data/
 datasets/
 *.csv
 *.json
+!tests/dfx/perf/tests/*.json
 !apps/ComfyUI-vLLM-Omni/example_workflows/*.json
 *.jsonl
 *.parquet

README.md

@@ -9,19 +9,19 @@ 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="docs/assets/WeChat.jpg"><b>WeChat</b></a> | <a href="https://arxiv.org/abs/2602.02204"><b>Paper</b></a> | <a href="https://docs.google.com/presentation/d/1XJWgv79lORl8rbaVvp2d5Sqs6ZEBgAgj/edit?slide=id.p1#slide=id.p1"><b>Slides</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> | <a href="https://arxiv.org/abs/2602.02204"><b>Paper</b></a> | <a href="https://docs.google.com/presentation/d/111-L8zF7A1j_YI_cR8JsblofdScdRr2f/edit?usp=sharing&ouid=110473603432222024453&rtpof=true&sd=true"><b>Slides</b></a> |
 </p>
 
 
 ---
 
 *Latest News* 🔥
+- [2026/05] We released [0.20.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.20.0) - refreshes the serving/runtime stack for large-scale omni workloads, and improves diffusion model performance, quantization, and hardware readiness across CUDA, ROCm, MUSA, NPU, and XPU backends.
 - [2026/03] We released [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0) - strengthens the core runtime through a large entrypoint refactor and scheduler/runtime cleanups, expands unified quantization and diffusion execution, broadens multimodal model coverage, and improves production readiness across audio, omni, image, video, RL, and multi-platform deployments.
 - [2026/03] Check out our first public [project deepdive](https://youtu.be/sgwNfsNnR9I) at the vLLM Hong Kong Meetup!
 - [2026/03] **[vllm-omni-skills](https://github.com/hsliuustc0106/vllm-omni-skills)** is a community-driven collection of AI assistant skills that help developers work with vLLM-Omni more effectively. These skills can be used with popular agentic AI coding assistants like **Cursor IDE**, **Claude**, **Codex**, and more.
 - [2026/02] We released [0.16.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.16.0) - A major alignment + capability release that rebases onto **upstream vLLM v0.16.0** and significantly expands performance, distributed execution, and production readiness across **Qwen3-Omni / Qwen3-TTS**, **Bagel**, **MiMo-Audio**, **GLM-Image** and the **Diffusion (DiT) image/video stack**—while also improving platform coverage (CUDA / ROCm / NPU / XPU), CI quality, and documentation.
 - [2026/02] We released [0.14.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.14.0) - This is the first **stable release** of vLLM-Omni that expands Omni’s diffusion / image-video generation and audio / TTS stack, improves distributed execution and memory efficiency, and broadens platform/backend coverage (GPU/ROCm/NPU/XPU). It also brings meaningful upgrades to serving APIs, profiling & benchmarking, and overall stability. Please check our latest [paper](https://arxiv.org/abs/2602.02204) for architecture design and performance results.
-- [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).
 - [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.
 
 ---

benchmarks/build_dataset/seed_tts_design/en/meta.lst

@@ -18,3 +18,13 @@ vd017|||The quarterly results exceed expectations across all major metrics.|A sh
 vd018|||Chapter one. The morning sun filtered gently through the forest canopy.|A smooth, rich male audiobook narrator voice, expressive and engaging.
 vd019|||To be or not to be, that is the question.|A theatrical female voice, dramatic and expressive, stage projection quality.
 vd020|||And that is all for tonight. Stay well out there, everyone.|A warm, velvety male late-night radio DJ voice, smooth and intimate.
+vd021|||Please fasten your seatbelts; we will be experiencing some turbulence shortly.|A calm, professional female flight attendant voice, reassuring and clear.
+vd022|||The fossil record suggests these creatures lived more than sixty million years ago.|A measured male documentary narrator voice, scientific and curious.
+vd023|||Hey kiddo, did you finish your homework before dinner today?|A gentle, caring father voice, warm and slightly tired after a long day.
+vd024|||Live from courtside, the home team has just taken a commanding lead.|A high-energy male sports broadcaster, urgent and engaged.
+vd025|||Please remain on the line while I transfer you to the next available agent.|A polite, neutral female call-center voice, professional and patient.
+vd026|||I have a dream that one day this nation will rise up and live out the true meaning of its creed.|A powerful, resonant male orator voice, slow and impassioned.
+vd027|||Mama always said life was like a box of chocolates; you never know what you're gonna get.|A folksy male voice with a Southern American accent, warm and reflective.
+vd028|||Ahoy there, matey! Hoist the colors and prepare to set sail.|A theatrical pirate male voice, gruff and boisterous.
+vd029|||Ladies and gentlemen, please welcome to the stage our keynote speaker.|A polished female event MC voice, gracious and articulate.
+vd030|||Once you've completed the form, click submit and we'll review it within two business days.|A friendly female product-onboarding voice, encouraging and clear.

benchmarks/build_dataset/seed_tts_smoke/en/meta.lst

@@ -18,3 +18,13 @@ smoke017|||In the distance, the mountains were shrouded in thick morning fog.
 smoke018|||Our company reported record revenue for the fourth quarter of the fiscal year.
 smoke019|||She explained the new policy in detail during the staff meeting this morning.
 smoke020|||The children laughed and played in the garden until the sun began to set.
+smoke021|||A long line of cars stretched along the coastal highway after the lighthouse opened.
+smoke022|||The chef recommended the seared scallops with a lemon butter sauce as the special.
+smoke023|||Our quarterly engineering review begins promptly at two in the afternoon on Thursday.
+smoke024|||Despite the heavy rain, the marathon runners kept a steady pace through downtown.
+smoke025|||She finished the crossword puzzle in under fifteen minutes while sipping her tea.
+smoke026|||The new release adds streaming output, faster startup, and better error reporting.
+smoke027|||His grandfather taught him how to whittle small wooden birds from cedar offcuts.
+smoke028|||Travelers were advised to check the timetable since trains were running on a holiday schedule.
+smoke029|||The startup launched a public beta after eighteen months of closed development.
+smoke030|||At night the desert sky filled with stars so bright the Milky Way was clearly visible.

benchmarks/diffusion/README.md

@@ -32,8 +32,6 @@ python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
 
 - The benchmark talks to `http://<host>:<port>/v1/chat/completions`.
 - If you run the server on another host or port, pass `--base-url` accordingly.
-- To compare diffusion multi-replica serving, restart the server with different
-  `runtime.num_replicas` values and keep the benchmark parameters unchanged.
 
 ## 2. Supported Datasets
 
@@ -95,35 +93,6 @@ You can point to your own trace using `--dataset-path`.
 
 ## 3. Benchmark Parameters
 
-### Multi-replica sweep
-
-For replica scaling, use the same model, prompt count, image/video size, and
-step count for every run. Change only the serving stage config:
-
-```yaml
-runtime:
-  devices: "0,1,2,3"
-  num_replicas: 4
-```
-
-Then run a fixed benchmark and write one JSON file per replica count:
-
-```bash
-python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-  --base-url http://localhost:8099 \
-  --model Qwen/Qwen-Image \
-  --task t2i \
-  --dataset random \
-  --num-prompts 32 \
-  --max-concurrency 8 \
-  --request-rate inf \
-  --num-inference-steps 10 \
-  --width 512 --height 512 \
-  --output-file diffusion_replica_4.json
-```
-
-Repeat for `num_replicas=1`, `2`, `3`, and `4`, using matching `runtime.devices`.
-
 ### Basic flags
 
 - `--base-url`: Server address (the script calls `.../v1/chat/completions`).

benchmarks/diffusion/quantization_quality.py

@@ -151,7 +151,7 @@ def _generate_image(omni, args, prompt, seed):
     from vllm_omni.platforms import current_omni_platform
 
     generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(seed)
-    torch.cuda.reset_peak_memory_stats()
+    torch.accelerator.reset_peak_memory_stats()
     start = time.perf_counter()
     outputs = omni.generate(
         {"prompt": prompt},
@@ -163,7 +163,7 @@ def _generate_image(omni, args, prompt, seed):
         ),
     )
     elapsed = time.perf_counter() - start
-    peak_mem = torch.cuda.max_memory_allocated() / (1024**3)
+    peak_mem = torch.accelerator.max_memory_allocated() / (1024**3)
 
     first = outputs[0]
     req_out = first.request_output[0] if hasattr(first, "request_output") else first
@@ -178,7 +178,7 @@ def _generate_video(omni, args, prompt, seed):
     from vllm_omni.platforms import current_omni_platform
 
     generator = torch.Generator(device=current_omni_platform.device_type).manual_seed(seed)
-    torch.cuda.reset_peak_memory_stats()
+    torch.accelerator.reset_peak_memory_stats()
     start = time.perf_counter()
     outputs = omni.generate(
         {"prompt": prompt, "negative_prompt": ""},
@@ -192,7 +192,7 @@ def _generate_video(omni, args, prompt, seed):
         ),
     )
     elapsed = time.perf_counter() - start
-    peak_mem = torch.cuda.max_memory_allocated() / (1024**3)
+    peak_mem = torch.accelerator.max_memory_allocated() / (1024**3)
 
     first = outputs[0]
     if hasattr(first, "request_output") and isinstance(first.request_output, list):
@@ -228,8 +228,8 @@ def _unload_omni(omni):
     del omni
     gc.collect()
     if torch.cuda.is_available():
-        torch.cuda.empty_cache()
-        torch.cuda.synchronize()
+        torch.accelerator.empty_cache()
+        torch.accelerator.synchronize()
 
 
 def run_benchmark(args):

benchmarks/distributed/omni_connectors/cross_node_mooncake_transfer_engine.py

@@ -614,8 +614,8 @@ Examples:
         if not torch.cuda.is_available():
             print("[ERROR] CUDA is not available but GPU mode was requested.")
             sys.exit(1)
-        if args.gpu_id >= torch.cuda.device_count():
-            print(f"[ERROR] GPU {args.gpu_id} not available. Found {torch.cuda.device_count()} GPUs.")
+        if args.gpu_id >= torch.accelerator.device_count():
+            print(f"[ERROR] GPU {args.gpu_id} not available. Found {torch.accelerator.device_count()} GPUs.")
             sys.exit(1)
         print(f"[INFO] Using GPU {args.gpu_id}: {torch.cuda.get_device_name(args.gpu_id)}")
 

benchmarks/tts/bench_voxcpm_offline.py

@@ -63,7 +63,7 @@ def _find_repo_root(start: Path) -> Path:
 
 REPO_ROOT = _find_repo_root(Path(__file__).resolve())
 DEFAULT_STAGE_ASYNC = REPO_ROOT / "vllm_omni" / "model_executor" / "stage_configs" / "voxcpm_async_chunk.yaml"
-DEFAULT_STAGE_SYNC = REPO_ROOT / "vllm_omni" / "model_executor" / "stage_configs" / "voxcpm.yaml"
+DEFAULT_STAGE_SYNC = REPO_ROOT / "vllm_omni" / "deploy" / "voxcpm.yaml"
 
 logger = logging.getLogger(__name__)
 

docker/Dockerfile.ci

@@ -1,5 +1,5 @@
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.20.0
+ARG VLLM_BASE_TAG=v0.21.0
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
 ARG APP_DIR=/workspace/vllm-omni
 WORKDIR ${APP_DIR}

docker/Dockerfile.npu.ci

@@ -0,0 +1,48 @@
+ARG VLLM_ASCEND_IMAGE=quay.io/ascend/vllm-ascend
+ARG VLLM_ASCEND_TAG=nightly-main
+FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
+
+ARG APP_DIR=/vllm-workspace/vllm-omni
+WORKDIR ${APP_DIR}
+COPY . .
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        ffmpeg \
+        libgl1 \
+        libglib2.0-0 \
+        libgomp1 \
+        libsm6 \
+        libxext6 \
+        libxrender1 \
+        libxcb1 \
+        espeak-ng && \
+    rm -rf /var/cache/apt/* && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+    source /usr/local/Ascend/nnal/atb/set_env.sh && \
+    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
+    python3 -m pip install -U pip && \
+    python3 -m pip install -v \
+      "pytest>=7.0.0" \
+      "pytest-asyncio>=0.21.0" \
+      "pytest-cov>=4.0.0" \
+      "pytest-mock>=3.10.0" \
+      "openpyxl>=3.0.0" \
+      "psutil>=7.2.0" \
+      "soundfile>=0.13.1" \
+      "pyttsx3>=2.99" \
+      "opencv-python>=4.12.0.88" \
+      "imageio[ffmpeg]>=0.6.0" \
+      --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
+      --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
+      --default-timeout=60 && \
+    python3 -m pip install -v -e /vllm-workspace/vllm-omni/ --no-build-isolation \
+      --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
+      --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
+      --default-timeout=60
+
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+ENTRYPOINT []

docker/Dockerfile.npu.ci.a3

@@ -0,0 +1,48 @@
+ARG VLLM_ASCEND_IMAGE=quay.io/ascend/vllm-ascend
+ARG VLLM_ASCEND_TAG=nightly-main-a3
+FROM ${VLLM_ASCEND_IMAGE}:${VLLM_ASCEND_TAG}
+
+ARG APP_DIR=/vllm-workspace/vllm-omni
+WORKDIR ${APP_DIR}
+COPY . .
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        ffmpeg \
+        libgl1 \
+        libglib2.0-0 \
+        libgomp1 \
+        libsm6 \
+        libxext6 \
+        libxrender1 \
+        libxcb1 \
+        espeak-ng && \
+    rm -rf /var/cache/apt/* && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
+    source /usr/local/Ascend/nnal/atb/set_env.sh && \
+    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
+    python3 -m pip install -U pip && \
+    python3 -m pip install -v \
+      "pytest>=7.0.0" \
+      "pytest-asyncio>=0.21.0" \
+      "pytest-cov>=4.0.0" \
+      "pytest-mock>=3.10.0" \
+      "openpyxl>=3.0.0" \
+      "psutil>=7.2.0" \
+      "soundfile>=0.13.1" \
+      "pyttsx3>=2.99" \
+      "opencv-python>=4.12.0.88" \
+      "imageio[ffmpeg]>=0.6.0" \
+      --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
+      --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
+      --default-timeout=60 && \
+    python3 -m pip install -v -e /vllm-workspace/vllm-omni/ --no-build-isolation \
+      --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
+      --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
+      --default-timeout=60
+
+ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
+
+ENTRYPOINT []

docker/Dockerfile.rocm

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.20.0
+ARG BASE_IMAGE=vllm/vllm-openai-rocm:v0.21.0
 FROM ${BASE_IMAGE} AS base
 
 # Declare a variable to know if we want to use the nightly build or the stable build.

docs/.nav.yml

@@ -76,8 +76,10 @@ nav:
     - Quantization:
       - Overview: user_guide/quantization/overview.md
       - Online Quantization: user_guide/quantization/online.md
+      - Quantized KV Cache: user_guide/quantization/quantized_kvcache.md
       - FP8 W8A8: user_guide/quantization/fp8.md
       - Int8 W8A8: user_guide/quantization/int8.md
+      - ModelOpt: user_guide/quantization/modelopt.md
       - GGUF: user_guide/quantization/gguf.md
       - AutoRound: user_guide/quantization/autoround.md
       - msModelSlim: user_guide/quantization/msmodelslim.md

docs/api/README.md

@@ -99,7 +99,6 @@ Configuration classes.
 - [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.cosyvoice3.config.CosyVoice3Config][]
 - [vllm_omni.model_executor.models.fish_speech.configuration_fish_speech.FishSpeechConfig][]
 - [vllm_omni.model_executor.models.fish_speech.configuration_fish_speech.FishSpeechFastARConfig][]
 - [vllm_omni.model_executor.models.fish_speech.configuration_fish_speech.FishSpeechSlowARConfig][]
@@ -116,10 +115,14 @@ Configuration classes.
 - [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][]
+- [vllm_omni.transformers_utils.configs.cosyvoice3.CosyVoice3Config][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Config][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLConfig][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLTextConfig][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLVisionConfig][]
+- [vllm_omni.transformers_utils.configs.omnivoice.OmniVoiceConfig][]
+- [vllm_omni.transformers_utils.configs.voxcpm.VoxCPMConfig][]
+- [vllm_omni.transformers_utils.configs.voxcpm2.VoxCPM2Config][]
 
 ## Workers
 

docs/contributing/README.md

@@ -116,6 +116,20 @@ Only specific types of PRs will be reviewed. The PR title is prefixed appropriat
 !!! note
     If the PR spans more than one category, please include all relevant prefixes.
 
+### Pre-Check Before Submitting
+
+Before submitting a PR, run the precheck-pr skill with code agent for a self-review against project conventions:
+
+```
+/precheck-pr
+```
+
+The skill offers two modes:
+- **Quick (~3 min):** catches showstoppers — PR title format, missing benchmark claims, rebase status
+- **Full (~10 min):** thorough maintainer-grade review — dead code scan, copy-paste detection, import hygiene
+
+The precheck covers five PR types: Bug Fix, Performance, New Model, Diffusion Model, and General. Each type has a tailored checklist that validates evidence quality (repro steps, A/B benchmarks, registry entries, etc.). See the skill in `.claude/skills/precheck-pr/` for the full checklist.
+
 ### Local Test
 Please run the L1 and L2 test cases locally first and attach the results before contacting us to add the "ready" label. Please refer to the [test instructions](./ci/test_guide.md) for running the test cases.
 

docs/contributing/ci/CI_5levels.md

@@ -152,6 +152,8 @@ vllm_omni/                                    tests/
 │   │                                            │   │   └── test_lora_manager.py        ✅
 │   ├── models/                                 │   ├── models/
 │   │   ├── qwen_image/                         │   │   ├── qwen_image/                 (e2e coverage)
+│   │   ├── ovis_image/                         │   │   ├── ovis_image/
+│   │   │   └── ...                             │   │   │   └── test_ovis_image.py     ✅
 │   │   ├── z_image/                            │   │   └── z_image/
 │   │   └── ...                                 │   │       └── test_zimage_tp_constraints.py  ✅
 │   └── worker/                                 │   └── worker/
@@ -232,15 +234,13 @@ vllm_omni/                                    tests/
                                                │   ├── test_qwen3_omni.py
                                                │   ├── test_qwen3_omni_expansion.py
                                                │   ├── test_mimo_audio.py
-                                               │   ├── test_image_gen_edit.py
                                                │   └── test_images_generations_lora.py
                                                └── offline_inference/                  ✅
                                                    ├── test_qwen2_5_omni.py
                                                    ├── test_qwen3_omni.py
                                                    ├── test_bagel_text2img.py
-                                                   ├── test_t2i_model.py
-                                                   ├── test_t2v_model.py
-                                                   ├── test_ovis_image.py
+                                                   ├── test_z_image.py
+                                                   ├── test_wan22.py
                                                    ├── test_zimage_tensor_parallel.py
                                                    ├── test_cache_dit.py
                                                    ├── test_teacache.py

docs/contributing/ci/test_guide.md

@@ -80,7 +80,7 @@ Our test scripts use the pytest framework. First, please use `git clone https://
 
     ```bash
     cd tests
-    pytest -s -v -m "full_model and omni and H100" --run-level=full_model
+    pytest -s -v -m "full_model and (omni or tts) and H100" --run-level=full_model
     ```
 
 === "L5 level"
@@ -91,10 +91,16 @@ Our test scripts use the pytest framework. First, please use `git clone https://
     cd tests
 
     # Stability: Qwen3-Omni
-    pytest -s -v dfx/stability/scripts/test_stability_qwen3_omni.py
+    pytest -s -v dfx/stability/scripts/test_stability_qwen3_omni.py -m slow
 
     # Stability: Wan2.2 (v1/videos diffusion benchmark loop)
-    pytest -s -v dfx/stability/scripts/test_stability_wan22.py
+    pytest -s -v dfx/stability/scripts/test_stability_wan22.py -m slow
+
+    # Reliability: Qwen3-Omni
+    pytest -s -v dfx/reliability/test_reliability_qwen3_omni.py -m slow
+
+    # Reliability: Wan2.2
+    pytest -s -v dfx/reliability/test_reliability_wan22.py -m slow
 
     ```
 

docs/contributing/ci/tests_markers.md

@@ -11,7 +11,8 @@ Defined in `pyproject.toml`:
 | `advanced_model`   | L3 tests (run on each merge to main)                 |
 | `full_model`       | L4 tests (run nightly) |
 | `diffusion`        | Diffusion model tests                                     |
-| `omni`             | Omni model tests                                          |
+| `omni`             | Omni multimodal model tests |
+| `tts`              | TTS model tests     |
 | `cache`            | Cache backend tests                                       |
 | `parallel`         | Parallelism/distributed tests                             |
 | `cpu`              | Tests that run on CPU                                     |

docs/contributing/ci/tests_style.md

@@ -54,6 +54,8 @@ vllm_omni/                                    tests/
 │   │                                            │   │   └── test_lora_manager.py        ✅
 │   ├── models/                                 │   ├── models/
 │   │   ├── qwen_image/                         │   │   ├── qwen_image/                 (e2e coverage)
+│   │   ├── ovis_image/                         │   │   ├── ovis_image/
+│   │   │   └── ...                             │   │   │   └── test_ovis_image.py     ✅
 │   │   ├── z_image/                            │   │   └── z_image/
 │   │   └── ...                                 │   │       └── test_zimage_tp_constraints.py  ✅
 │   └── worker/                                 │   └── worker/
@@ -134,15 +136,13 @@ vllm_omni/                                    tests/
                                                │   ├── test_qwen3_omni.py
                                                │   ├── test_qwen3_omni_expansion.py
                                                │   ├── test_mimo_audio.py
-                                               │   ├── test_image_gen_edit.py
                                                │   └── test_images_generations_lora.py
                                                └── offline_inference/                  ✅
                                                    ├── test_qwen2_5_omni.py
                                                    ├── test_qwen3_omni.py
                                                    ├── test_bagel_text2img.py
-                                                   ├── test_t2i_model.py
-                                                   ├── test_t2v_model.py
-                                                   ├── test_ovis_image.py
+                                                   ├── test_z_image.py
+                                                   ├── test_wan22.py
                                                    ├── test_zimage_tensor_parallel.py
                                                    ├── test_cache_dit.py
                                                    ├── test_teacache.py
@@ -213,7 +213,6 @@ E2E Online tests for Qwen3-Omni model with mix input and audio+text output.
 import os
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
 
 import threading
 from pathlib import Path

docs/contributing/model/adding_diffusion_model.md

@@ -115,17 +115,20 @@ class YourAttentionBlock(nn.Module):
 from vllm_omni.diffusion.attention.layer import Attention
 from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata
 
-class YourAttentionBlock(nn.Module):
+class YourSelfAttentionBlock(nn.Module):
     def __init__(self, ...):
         super().__init__()
 
-        # Initialize vLLM-Omni's Attention layer
+        # Initialize vLLM-Omni's Attention layer.
+        # `role` lets users target this site with --diffusion-attention-config
+        # (e.g. --diffusion-attention-config.per_role.self.backend SAGE_ATTN).
         self.attn = Attention(
             num_heads=self.num_heads,
             head_size=self.head_dim,
             softmax_scale=1.0 / (self.head_dim ** 0.5),
             causal=False,  # Diffusion models typically use bidirectional attention
             num_kv_heads=self.num_kv_heads,
+            role="self",
         )
 
     def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, ...):
@@ -133,7 +136,6 @@ class YourAttentionBlock(nn.Module):
         # Create attention metadata
         attn_metadata = AttentionMetadata(attn_mask=attention_mask)
         hidden_states = self.attn(query, key, value, attn_metadata=attn_metadata)
-
 ```
 
 **Key Points:**
@@ -141,8 +143,35 @@ class YourAttentionBlock(nn.Module):
 - **Attention layer initialization:** Done in `__init__`, not per-forward
 - **Tensor shapes:** vLLM-Omni `Attention` expects QKV to have `[B, seq, num_heads, head_dim]` shape
 - **AttentionMetadata:** Wraps attention mask and other metadata
+- **Role:** Tag every `Attention` site with a `role` string so users can configure backends per role (see below)
 
-**Attention backends:** vLLM-Omni automatically selects the attention backend given the environmental variable `DIFFUSION_ATTENTION_BACKEND`. The default attention backend is `FLASH_ATTN` for diffusion models.
+**Declaring attention roles**
+
+The `role` argument is a free-form string that identifies this attention site. Users can match it from `--diffusion-attention-config.per_role.<role>.*` to swap backends without touching model code. Two conventions cover the common cases:
+
+| Convention | When to use | Example |
+|---|---|---|
+| `"self"` | Q/K/V come from the same hidden state | DiT self-attention block |
+| `"cross"` | K/V come from a separate `encoder_hidden_states` | Text-conditioned cross-attention |
+
+For multi-modal or unusual sites, use a dot-namespaced role and pair it with `role_category` so it can fall back to the generic config when nothing model-specific is set:
+
+```python
+# A model-specific cross-attention site that user config can target
+# either as 'mymodel.audio_to_video' (exact) or as 'cross' (category fallback).
+self.audio_to_video_attn = Attention(
+    num_heads=self.num_heads,
+    head_size=self.head_dim,
+    softmax_scale=1.0 / (self.head_dim ** 0.5),
+    causal=False,
+    role="mymodel.audio_to_video",
+    role_category="cross",
+)
+```
+
+For cross-attention sites whose K/V are replicated across ranks (e.g. text encoder output), pass `skip_sequence_parallel=True` to opt this layer out of sequence-parallel sharding.
+
+**Attention backends:** When the user does not configure a backend, vLLM-Omni asks the current platform for its default (typically `FLASH_ATTN` on CUDA when available). Users override the default via `--diffusion-attention-backend`, the `DIFFUSION_ATTENTION_BACKEND` env var, or finer-grained `--diffusion-attention-config.per_role.*` flags. See [Diffusion Attention Backends](../../user_guide/diffusion/attention_backends.md) for the full configuration surface.
 
 #### 1.3: Replace Imports and Utilities
 

docs/design/module/dit_module.md

@@ -455,7 +455,7 @@ To learn more about the diffusion pipeline and how to add a new diffusion pipeli
 
 #### Architecture
 
-The attention system uses a **backend selector pattern** that automatically chooses the optimal attention implementation based on hardware and model configuration.
+The attention system uses a **role-aware backend selector pattern**. Each `Attention` site declares a semantic `role` (`"self"`, `"cross"`, or a model-specific string). The selector consults the user's `AttentionConfig` to pick a backend per role and falls back to the platform default when nothing is configured.
 
 #### Backend Selection
 
@@ -463,92 +463,90 @@ The attention system uses a **backend selector pattern** that automatically choo
 
 ```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(...)
+    def __init__(self, num_heads, head_size, causal, softmax_scale, *,
+                 role="self", role_category=None, ...):
+        # Resolve backend for this role from the active OmniDiffusionConfig
+        config = get_current_diffusion_config_or_none()
+        attention_config = config.attention_config if config is not None else None
+        attn_backend_cls, spec = get_attn_backend_for_role(
+            role=role,
+            head_size=head_size,
+            attention_config=attention_config,
+            role_category=role_category,
+        )
+        self.attn_impl_cls = attn_backend_cls.get_impl_cls()
+        self.attention = self.attn_impl_cls(..., backend_kwargs=spec.extra if spec else None)
 ```
 
 **Available Backends**:
 
-- **FlashAttention**: Optimized CUDA kernel (FA2/FA3) - memory efficient via tiling
+- **FlashAttention** (`FLASH_ATTN`): Memory-efficient kernel that dispatches to FA2/FA3 on CUDA, AITER on ROCm, and `mindiesd` on Ascend NPU.
 
-- **SDPA**: PyTorch's scaled dot-product attention - default, cross-platform
+- **SDPA** (`TORCH_SDPA`): PyTorch's scaled dot-product attention — cross-platform fallback.
 
-- **SageAttention**: Sparse attention implementation from SageAttention library
-
-- **AscendAttention**: NPU-optimized attention for Ascend hardware
+- **SageAttention** (`SAGE_ATTN`): Quantized attention implementation from the SageAttention library.
 
 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
 
+Selection is driven by `AttentionConfig` on `OmniDiffusionConfig`, which carries a global `default` plus a `per_role` map of `AttentionSpec` entries. A role string is resolved in this order:
+
 ```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
+def get_attn_backend_for_role(role, head_size, attention_config=None, role_category=None):
+    # 1. attention_config.per_role[role]            — exact match
+    # 2. attention_config.per_role[role_category]   — category fallback
+    # 3. attention_config.default                   — global default
+    # 4. Platform default                           — hardware-specific
+    if attention_config is not None:
+        spec, source = attention_config.resolve_with_source(
+            role=role, role_category=role_category,
+        )
+        if spec is not None:
+            return load_backend(spec.backend), spec
+    return current_omni_platform.get_diffusion_attn_backend_cls(...), None
 ```
 
 **Selection Priority**:
 
-1. **Environment Variable**: `DIFFUSION_ATTENTION_BACKEND` for manual override
+1. **Per-role override** (`--diffusion-attention-config.per_role.<role>.backend`) — finest control; matched on the layer's exact `role` string.
 
-    - Valid values: `FLASH_ATTN`, `TORCH_SDPA`, `SAGE_ATTN`, `ASCEND`
+2. **Role category fallback** (`--diffusion-attention-config.per_role.<category>.backend`) — used when an exact match is missing and the layer declared `role_category` (e.g. a `"mymodel.audio_to_video"` site can fall back to `"cross"`).
 
-    - Example: `export DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN`
+3. **Global default** (`--diffusion-attention-backend FLASH_ATTN`, or `--diffusion-attention-config.default.backend`, or `DIFFUSION_ATTENTION_BACKEND` env var).
 
-2. **Automatic Fallback**: Falls back to SDPA if selected backend unavailable
+4. **Platform default** — `current_omni_platform.get_diffusion_attn_backend_cls(...)` picks the best-available kernel for the hardware.
 
-3. **Hardware Detection**: Can select based on device type (NPU, CUDA, etc.)
+`AttentionSpec.extra` is forwarded to the backend constructor as `backend_kwargs`, so backend-specific parameters (e.g. SparseBlock `block_size`) can be set without changing model code.
+
+For the user-facing CLI surface, see [Diffusion Attention Backends](../../user_guide/diffusion/attention_backends.md). For the role declaration contract on the model side, see [Adding a Diffusion Model](../../contributing/model/adding_diffusion_model.md).
 
 **Backend Availability**:
 
 - **SDPA**: Always available (PyTorch built-in)
 
-- **FlashAttention**: Requires `flash-attn` package installed
+- **FlashAttention**: Requires `flash-attn` on CUDA, `aiter` on ROCm, or `mindiesd` on Ascend NPU
 
 - **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`
+**Location**: `vllm_omni/diffusion/attention/selector.py` + `vllm_omni/platforms/<device>/platform.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**:
+Backend resolution is delegated to the active platform: `current_omni_platform.get_diffusion_attn_backend_cls(selected_backend, head_size)` returns a fully-qualified class path that `_load_backend_cls` imports lazily, with the result cached by `(backend_name, head_size)`.
 
 ```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",
-    },
-}
+@cache
+def _cached_get_backend_cls(backend_name: str | None, head_size: int) -> type[AttentionBackend]:
+    backend_cls_path = current_omni_platform.get_diffusion_attn_backend_cls(
+        selected_backend=backend_name,
+        head_size=head_size,
+    )
+    return _load_backend_cls(backend_cls_path)
 ```
 
+Each platform (`cuda`, `rocm`, `xpu`, `musa`, `npu`) maps backend names like `"FLASH_ATTN"`, `"TORCH_SDPA"`, `"SAGE_ATTN"` to the right backend class path for that hardware, including head-size compatibility checks. Passing `selected_backend=None` lets the platform pick its own default.
+
 #### Attention Backend Integration
 
 The `Attention` layer integrates backends through a unified interface. Here's how **FlashAttentionBackend** is integrated as an example:

docs/getting_started/installation/gpu.md

@@ -34,7 +34,7 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr
 
 ### Pre-built wheels
 
-Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, and 0.18.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
+Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, and 0.21.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
 
 === "NVIDIA CUDA"
 

docs/getting_started/installation/gpu/cuda.inc.md

@@ -20,7 +20,7 @@ Therefore, it is recommended to install vLLM and vLLM-Omni with a **fresh new**
 
 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm --torch-backend=auto
+uv pip install vllm==0.21.0 --torch-backend=auto
 ```
 
 #### Installation of vLLM-Omni
@@ -39,13 +39,13 @@ uv pip install 'vllm-omni[demo]'
 # --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 0.20.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.21.0 release version of the library
 
 ```bash
-uv pip install vllm==0.20.0 --torch-backend=auto
+uv pip install vllm==0.21.0 --torch-backend=auto
 ```
 
-The 0.20.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The 0.21.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
 
 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -66,12 +66,12 @@ 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 v0.20.0
+git checkout v0.21.0
 ```
 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
-#For CUDA 13.0 (the default for v0.20.0; the wheel filename has no `+cu130` suffix)
-export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.20.0/vllm-0.20.0-cp38-abi3-manylinux_2_35_x86_64.whl
+#For CUDA 13.0 (the default for v0.21.0; the wheel filename has no `+cu130` suffix)
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.21.0/vllm-0.21.0-cp38-abi3-manylinux_2_35_x86_64.whl
 ```
 Install vllm with command below (If you have no existing PyTorch).
 ```bash
@@ -102,7 +102,7 @@ docker run --runtime nvidia --gpus 2 \
     --env "HF_TOKEN=$HF_TOKEN" \
     -p 8091:8091 \
     --ipc=host \
-    vllm/vllm-omni:v0.18.0 \
+    vllm/vllm-omni:v0.21.0 \
     --model Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8091
 ```
 

docs/getting_started/installation/gpu/rocm.inc.md

@@ -13,7 +13,7 @@ vLLM-Omni current recommends the steps in under setup through Docker Images.
 
 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm==0.20.0+rocm721 --extra-index-url https://wheels.vllm.ai/rocm/0.20.0/rocm721
+uv pip install vllm==0.21.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.21.0/rocm722
 ```
 
 #### Installation of vLLM-Omni
@@ -34,13 +34,13 @@ uv pip install onnxruntime-rocm
 # --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 0.20.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.21.0 release version of the library
 
 ```bash
-uv pip install vllm==0.20.0+rocm721 --extra-index-url https://wheels.vllm.ai/rocm/0.20.0/rocm721
+uv pip install vllm==0.21.0+rocm722 --extra-index-url https://wheels.vllm.ai/rocm/0.21.0/rocm722
 ```
 
-The pre-built 0.20.0 vLLM wheel targets ROCm 7.2.1. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The pre-built 0.21.0 vLLM wheel targets ROCm 7.2.2. If you need a different ROCm stack or want to reuse an existing PyTorch installation, build vLLM from source instead.
 
 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -58,7 +58,7 @@ 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 v0.20.0
+git checkout v0.21.0
 python3 -m pip install -r requirements/rocm.txt
 python3 setup.py develop
 ```
@@ -130,7 +130,7 @@ docker run --rm \
   -v ~/.cache/huggingface:/root/.cache/huggingface \
   --env "HF_TOKEN=$HF_TOKEN" \
   -p 8091:8091 \
-  vllm/vllm-omni-rocm:v0.20.0 \
+  vllm/vllm-omni-rocm:v0.21.0 \
   --model Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 
@@ -149,7 +149,7 @@ docker run --rm -it \
   -v ~/.cache/huggingface:/root/.cache/huggingface \
   --env "HF_TOKEN=$HF_TOKEN" \
   --entrypoint bash \
-  vllm/vllm-omni-rocm:v0.20.0
+  vllm/vllm-omni-rocm:v0.21.0
 ```
 
 # --8<-- [end:pre-built-images]

docs/getting_started/quickstart.md

@@ -19,10 +19,10 @@ uv venv --python 3.12 --seed
 source .venv/bin/activate
 
 # On CUDA
-uv pip install vllm==0.20.0 --torch-backend=auto
+uv pip install vllm==0.21.0 --torch-backend=auto
 
 # On ROCm
-uv pip install vllm==0.20.0+rocm721 --extra-index-url https://wheels.vllm.ai/rocm/0.20.0/rocm721
+uv pip install vllm==0.21.0+rocm721 --extra-index-url https://wheels.vllm.ai/rocm/0.21.0/rocm721
 
 git clone https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
@@ -35,7 +35,7 @@ For additional installation methods — please see the [installation guide](inst
 !!! note
     It is important to install the same major & minor version of vLLM and vLLM Omni, otherwise things may not work as expected. If the versions are misaligned, you will see a warning when you import vLLM Omni.
 
-    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.20.0` and vLLM Omni `0.20.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
+    If you are seeing strange behavior with the `vllm` command not handling the `--omni` flag correctly, you most likely have a version mismatch with vLLM < `0.21.0` and vLLM Omni `0.21.0`, as vLLM Omni no longer hijacks the vLLM entrypoint. Updating vLLM should resolve this issue.
 
 ## Offline Inference
 

docs/models/supported_models.md

@@ -64,9 +64,11 @@ th {
 | `Flux2Pipeline` | FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | ✅︎ | ✅︎ | | |
 | `FishSpeechSlowARForConditionalGeneration` | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✅︎ | ✅︎ | | |
 | `DreamIDOmniPipeline` | DreamID-Omni | `XuGuo699/DreamID-Omni` | ✅︎ | ✅︎ | | |
+| `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only) | `SenseNova/SenseNova-U1-8B-MoT` | ✅︎ | | | |
 | `HunyuanVideo15Pipeline` | HunyuanVideo-1.5-T2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | ✅︎ | ✅︎ | | |
 | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | |
 | `VoxtralTTSForConditionalGeneration` | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✅︎ | ✅︎ | | |
+| `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | |
 |`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | |
 | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 ✅︎ indicates the model is supported on that backend. Empty cells mean not listed as supported on that backend.

docs/serving/speech_api.md

@@ -531,16 +531,18 @@ for result in response.json()["results"]:
 
 All items are fanned out to `generate()` concurrently. The engine's stage worker automatically batches them up to the configured `max_batch_size` and queues the rest — no client-side throttling needed.
 
-For best throughput, set both stages' `max_num_seqs` to ≥4 via `--stage-overrides`:
+For best throughput, set both stages' `max_num_seqs` above 1 via `--stage-overrides`. On the current Qwen3-TTS CustomVoice benchmark, stage 1 performed best at `max_num_seqs: 10`:
 
 ```bash
 vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
     --omni --port 8091 --trust-remote-code --enforce-eager \
-    --stage-overrides '{"0":{"max_num_seqs":4,"gpu_memory_utilization":0.2},
-                        "1":{"max_num_seqs":4,"gpu_memory_utilization":0.2}}'
+    --stage-overrides '{"0":{"max_num_seqs":10,"gpu_memory_utilization":0.2},
+                        "1":{"max_num_seqs":10,"gpu_memory_utilization":0.2}}'
 ```
 
-The bundled `qwen3_tts.yaml` uses `max_num_seqs: 1` (single request) on both stages. Bumping to 4 yields roughly 4× throughput on the talker and lets stage 1 batch chunks across in-flight requests.
+The bundled `qwen3_tts.yaml` uses a multi-request default and lets stage 1 batch chunks across in-flight requests. For latency-sensitive deployments, avoid forcing stage 1 back to `max_num_seqs: 1`; benchmark before reducing it below `10`.
+
+The bundled config also sets `initial_codec_chunk_frames: 1`. This emits only the first audio chunk early for lower TTFA, then returns to the normal `codec_chunk_frames` window so Code2Wav does not repeatedly decode tiny overlapping chunks.
 
 ## Supported Models
 

docs/user_guide/diffusion/attention_backends.md

@@ -1,10 +1,10 @@
 # Diffusion Attention Backends
 
-This document describes the diffusion attention backends available in vLLM-Omni, how to select them, and how to use SageAttention.
+This document describes the diffusion attention backends available in vLLM-Omni, how to select them globally and per-role, and how to use SageAttention.
 
 ## Overview
 
-Diffusion attention backend selection is controlled by the `DIFFUSION_ATTENTION_BACKEND` environment variable and resolved in `vllm_omni.diffusion.attention.selector`.
+Diffusion attention backend selection is resolved in `vllm_omni.diffusion.attention.selector`. It looks up the backend from a structured `AttentionConfig` carried on `OmniDiffusionConfig` and falls back to the platform default when nothing is configured.
 
 This backend is used by diffusion attention layers such as the DiT attention in video and image generation models.
 
@@ -14,7 +14,7 @@ On CUDA, the practical choices today are:
 - `TORCH_SDPA`: PyTorch `scaled_dot_product_attention`.
 - `SAGE_ATTN`: SageAttention backend, if `sageattention` is installed.
 
-If `DIFFUSION_ATTENTION_BACKEND` is unset, vLLM-Omni asks the current platform to choose the default backend. On CUDA, that normally means `FLASH_ATTN` when available, otherwise `TORCH_SDPA`.
+If no attention backend is configured, vLLM-Omni asks the current platform to choose the default. On CUDA, that normally means `FLASH_ATTN` when available, otherwise `TORCH_SDPA`.
 
 ## Backend Options
 
@@ -24,19 +24,77 @@ If `DIFFUSION_ATTENTION_BACKEND` is unset, vLLM-Omni asks the current platform t
 | `TORCH_SDPA` | Most conservative fallback. Useful for debugging or compatibility. |
 | `SAGE_ATTN` | Requires `sageattention`. Can improve performance on some workloads, but output quality must be validated model-by-model. |
 
-## Selection Priority
+## Configuration
 
-Diffusion attention backend selection follows this order:
+Diffusion attention backends can be configured three ways, in priority order:
 
-1. `DIFFUSION_ATTENTION_BACKEND`
-2. Platform default
+1. **`--diffusion-attention-config`** — structured per-role config (highest priority).
+2. **`--diffusion-attention-backend` / `DIFFUSION_ATTENTION_BACKEND` env var** — global shorthand that sets the default backend.
+3. **Platform default** — used when nothing is configured.
 
-Example:
+`--diffusion-attention-backend` is shorthand for `--diffusion-attention-config.default.backend`. It may be combined with `--diffusion-attention-config.per_role.*` overrides, but is mutually exclusive with `--diffusion-attention-config.default.backend`.
+
+### Global default
+
+Set the default backend for every diffusion attention layer:
 
 ```bash
+# CLI flag
+vllm-omni serve <model> --diffusion-attention-backend SAGE_ATTN
+
+# Environment variable (also recognized for backwards compatibility)
 export DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN
 ```
 
+### Per-role configuration
+
+Roles are free-form strings declared by each diffusion model. The two common categories are `"self"` and `"cross"`; model-specific roles (e.g. `"ltx2.audio_to_video"`) may also be declared. A role string is matched in this order:
+
+1. Exact `per_role[role]` match
+2. `per_role[role_category]` fallback (e.g. `"ltx2.audio_to_video"` → `"cross"`)
+3. `default`
+4. Platform default
+
+Use vLLM-style dotted flags or one JSON blob:
+
+```bash
+# Dotted flags
+vllm-omni serve <model> \
+    --diffusion-attention-config.default.backend FLASH_ATTN \
+    --diffusion-attention-config.per_role.cross.backend TORCH_SDPA
+
+# JSON
+vllm-omni serve <model> \
+    --diffusion-attention-config '{"default":{"backend":"FLASH_ATTN"},"per_role":{"cross":{"backend":"TORCH_SDPA"}}}'
+```
+
+Backends may also accept backend-specific parameters via `extra`:
+
+```bash
+--diffusion-attention-config.per_role.self.backend SPARSE_BLOCK \
+--diffusion-attention-config.per_role.self.extra.block_size 128
+```
+
+### Programmatic API
+
+When constructing `OmniDiffusionConfig` directly:
+
+```python
+from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, OmniDiffusionConfig
+
+config = OmniDiffusionConfig(
+    attention_config=AttentionConfig(
+        default=AttentionSpec(backend="FLASH_ATTN"),
+        per_role={
+            "cross": AttentionSpec(backend="TORCH_SDPA"),
+        },
+    ),
+    ...,
+)
+```
+
+A plain dict is also accepted and normalized to `AttentionConfig`.
+
 ## SageAttention Installation
 
 vLLM-Omni expects SageAttention to be installed into the same Python environment as vLLM-Omni.
@@ -85,7 +143,21 @@ DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN python examples/offline_inference/text_to_
     --output outputs/wan22_sage.mp4
 ```
 
-### Compare Against FlashAttention
+### Mixed backends across roles
+
+Use `FLASH_ATTN` for self-attention and `TORCH_SDPA` for cross-attention:
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+    --model Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+    --prompt "A dog running across a field of golden wheat." \
+    --diffusion-attention-config.per_role.self.backend FLASH_ATTN \
+    --diffusion-attention-config.per_role.cross.backend TORCH_SDPA \
+    --tensor-parallel-size 2 \
+    --output outputs/wan22_mixed.mp4
+```
+
+### Compare against FlashAttention
 
 Unset the backend override, or explicitly use `FLASH_ATTN`:
 

docs/user_guide/diffusion/cpu_offload_diffusion.md

@@ -38,14 +38,15 @@ vllm-omni serve diffusion Wan-AI/Wan2.2-T2V-A14B-Diffusers --enable-cpu-offload
 
 ### To Support a Model
 
-Implement the `SupportsModuleOffload` protocol to declare which
-submodules participate in offloading:
+Implement the `SupportsComponentDiscovery` protocol to declare which
+submodules serve as pipeline components (used by offloading, HSDP
+sharding, and other framework features):
 
 ```python
 from typing import ClassVar
-from vllm_omni.diffusion.models.interface import SupportsModuleOffload
+from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
 
-class MyPipeline(nn.Module, SupportsModuleOffload):
+class MyPipeline(nn.Module, SupportsComponentDiscovery):
     _dit_modules: ClassVar[list[str]] = ["transformer"]
     _encoder_modules: ClassVar[list[str]] = ["text_encoder", "vision_model"]
     _vae_modules: ClassVar[list[str]] = ["vae"]
@@ -158,7 +159,7 @@ class Flux2Transformer2DModel(nn.Module):
 The offloader discovers pipeline components in two ways:
 
 1. **Protocol-based** (preferred): If the pipeline implements
-    `SupportsModuleOffload`, its `_dit_modules`, `_encoder_modules`,
+    `SupportsComponentDiscovery`, its `_dit_modules`, `_encoder_modules`,
     `_vae_modules`, and `_resident_modules` class variables are used
     directly.  All attribute names support dotted paths (e.g.
     `"pipe.transformer"`, `"bagel.time_embedder"`) for nested submodules.

docs/user_guide/diffusion/multi_replica_serving.md

@@ -1,80 +0,0 @@
-# Diffusion Multi-Replica Serving
-
-Diffusion stages can be replicated in online serving by setting `runtime.num_replicas` on a `stage_type: diffusion` stage. Each replica owns an independent diffusion engine process and receives requests through round-robin routing with request affinity.
-
-## When to Use
-
-Use multi-replica serving when one diffusion process does not saturate the available GPUs, or when you want higher request throughput for many independent image/video generation requests. It improves concurrency throughput; it does not make a single request faster.
-
-## Configuration
-
-A replica consumes the devices required by that diffusion stage. For single-device diffusion pipelines, set `runtime.devices` to one device per replica:
-
-```yaml
-stages:
-  - stage_id: 0
-    stage_type: diffusion
-    runtime:
-      devices: "0,1,2,3"
-      num_replicas: 4
-    engine_args:
-      model: Qwen/Qwen-Image
-      parallel_config:
-        world_size: 1
-```
-
-For multi-device diffusion pipelines, `parallel_config.world_size` is treated as the number of devices per replica. For example, four 2-GPU replicas need eight devices:
-
-```yaml
-stages:
-  - stage_id: 0
-    stage_type: diffusion
-    runtime:
-      devices: "0,1,2,3,4,5,6,7"
-      num_replicas: 4
-    engine_args:
-      model: Wan-AI/Wan2.2-T2V-A14B-Diffusers
-      parallel_config:
-        world_size: 2
-```
-
-The runtime splits the device list into contiguous chunks per replica. With the second example, replica 0 uses `0,1`, replica 1 uses `2,3`, replica 2 uses `4,5`, and replica 3 uses `6,7`.
-
-## Serving
-
-Start the server with the stage config:
-
-```bash
-vllm serve Qwen/Qwen-Image \
-  --omni \
-  --port 8099 \
-  --stage-configs-path /path/to/diffusion_multi_replica.yaml
-```
-
-Then send normal OpenAI-compatible diffusion chat requests. The orchestrator routes new requests across replicas and keeps updates/abort calls on the same replica for each request.
-
-## Benchmarking
-
-Benchmark each replica count by changing `runtime.num_replicas` and `runtime.devices`, restarting the server, then running:
-
-```bash
-python benchmarks/diffusion/diffusion_benchmark_serving.py \
-  --base-url http://localhost:8099 \
-  --model Qwen/Qwen-Image \
-  --task t2i \
-  --dataset random \
-  --num-prompts 8 \
-  --request-rate inf \
-  --num-inference-steps 2 \
-  --width 256 --height 256 \
-  --output-file diffusion_replica_4.json
-```
-
-For a quick scaling sweep, test `num_replicas` values `1`, `2`, `3`, and `4` with the same prompt count and generation parameters. Compare throughput and end-to-end latency from the JSON outputs.
-
-## Notes and Limitations
-
-- Multi-replica serving is currently supported for diffusion stages. Single-stage mode still rejects non-diffusion replicas.
-- `runtime.devices` must provide exactly `num_replicas * devices_per_replica` devices.
-- Request-level latency may stay similar or slightly increase because each request still runs on one replica.
-- Throughput improves only when enough concurrent requests are issued and each replica has enough GPU memory.

docs/user_guide/diffusion_features.md

@@ -15,7 +15,7 @@ vLLM-Omni supports various advanced features for diffusion models:
 - Acceleration: **cache methods**, **parallelism methods**, **startup optimizations**
 - Memory optimization: **cpu offloading**, **quantization**
 - Extensions: **LoRA inference**, **frame interpolation**
-- Execution modes: **step execution**, **multi-replica serving**
+- Execution modes: **step execution**
 
 ## Supported Features
 
@@ -43,7 +43,6 @@ Parallelism methods distribute computation across GPUs without quality loss (mat
 | **[Tensor Parallelism](diffusion/parallelism/tensor_parallel.md)** | Shards model weights across devices | Large models that don't fit in single GPU, with 2+ GPUs |
 | **[HSDP](diffusion/parallelism/hsdp.md)** | Weight sharding via FSDP2, redistributed on-demand at runtime | Very large models (14B+) on limited VRAM, combinable with SP |
 | **[Expert Parallelism](diffusion/parallelism/expert_parallel.md)** | Shards MoE expert MLP blocks across devices | MoE diffusion models (e.g., HunyuanImage3.0) |
-| **[Multi-Replica Serving](diffusion/multi_replica_serving.md)** | Runs multiple independent diffusion stage replicas and routes requests across them | Online serving throughput with many concurrent requests |
 
 #### Startup Optimization
 
@@ -108,7 +107,7 @@ The following tables show which models support each feature:
 
 | Model | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
 |-------|:----------:|:-----------:|:---------------------:|:--------------:|:-----------------:|:------:|:------------------------:|:--------------------:|:--------------:|:----------------:|
-| **Bagel** | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅  | ❌ | ❌ | ❌ |
+| **Bagel** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅  | ❌ | ❌ | ❌ |
 | **FLUX.1-dev** | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
 | **FLUX.1-schnell** | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
 | **FLUX.2-klein** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
@@ -128,6 +127,7 @@ The following tables show which models support each feature:
 | **Qwen-Image-Edit** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ |
 | **Qwen-Image-Edit-2509** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ✅ | ❌ | ❌ |
 | **Qwen-Image-Layered** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ |
+| **SenseNova-U1** | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
 | **Stable-Diffusion3.5** | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ (decode) | ❌ | ❌ |
 | **Z-Image** | ✅ | ✅ | ✅ | ❓ | ✅ (TP=2 only) | ✅ | ❌ | ✅ (decode) | ✅ | ❌ |
 | **ERNIE-Image** | ❌ | ✅ | ✅ | ❓ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -144,7 +144,7 @@ The following tables show which models support each feature:
 | **Wan2.1-VACE** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ |
 | **LTX-2** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 | **LTX-2.3** | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-| **Helios** | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+| **Helios** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 | **HunyuanVideo-1.5 T2V I2V** | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ✅ | ❌ |
 | **DreamID-Omni** | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 

docs/user_guide/quantization/fp8.md

@@ -2,11 +2,11 @@
 
 ## Overview
 
-FP8 quantization converts BF16/FP16 weights to FP8 at model load time, or loads
-a checkpoint whose target stage already declares an FP8 quantization config.
-Online activation scaling is the default and does not require calibration.
-Static activation scaling is supported when calibrated scale information is
-available.
+FP8 quantization converts BF16/FP16 weights to FP8 at model load time. Online
+activation scaling is the default and does not require calibration. Static
+activation scaling is supported when calibrated scale information is available.
+For ModelOpt-produced pre-quantized checkpoints, see
+[ModelOpt Quantization](modelopt.md).
 
 Some architectures can quantize all linear layers. Others have
 quality-sensitive layers that should stay in BF16 through `ignored_layers`.
@@ -46,7 +46,7 @@ guide. FP8 on Ampere may use a weight-only path where available.
 
 | Model | Scope | Format | Status |
 |-------|-------|--------|--------|
-| Qwen3-Omni | Thinker language-model stage | ModelOpt `quant_algo=FP8` | Tested for thinker memory reduction |
+| Qwen3-Omni | Thinker language-model stage | [ModelOpt](modelopt.md) `quant_algo=FP8` | Tested for thinker memory reduction |
 | Qwen3-TTS | TTS language-model stage | Checkpoint config | Not validated |
 
 Audio encoder, vision encoder, talker, and code2wav stay in BF16 unless a

docs/user_guide/quantization/modelopt.md

@@ -0,0 +1,138 @@
+# ModelOpt Quantization
+
+## Overview
+
+ModelOpt quantization loads checkpoints produced by NVIDIA ModelOpt. The
+quantized weights and scale tensors are generated before serving, so inference
+does not run online calibration or convert a BF16 checkpoint at startup.
+
+vLLM-Omni currently validates the ModelOpt FP8 checkpoint path for diffusion
+transformers. The loader auto-detects supported ModelOpt FP8 checkpoint configs
+and keeps non-transformer components, such as the tokenizer, scheduler, text
+encoder, and VAE, on the base checkpoint unless a model-specific guide says
+otherwise.
+
+!!! note
+    `--force-cutlass-fp8` is an explicit runtime override for diffusion
+    checkpoints that already carry a supported ModelOpt FP8 config. It does not
+    quantize BF16 checkpoints and it does not apply to online `--quantization
+    fp8`. The flag only takes effect for ModelOpt FP8 diffusion stages on CUDA
+    SM89+ devices; other platforms and non-ModelOpt FP8 paths fall back to the
+    normal vLLM kernel selection.
+
+## Supported ModelOpt Checkpoint Formats
+
+vLLM-Omni treats ModelOpt checkpoints as pre-quantized checkpoints. The
+checkpoint config must identify ModelOpt as the quantization method or producer,
+and the quantization algorithm must be one of the validated FP8 algorithms.
+
+| Checkpoint field | Supported value |
+|------------------|-----------------|
+| `method` / `quant_method` | `modelopt` |
+| `producer.name` | `modelopt` |
+| `quant_algo` | `FP8`, `FP8_PER_CHANNEL_PER_TOKEN` |
+
+Other ModelOpt algorithms, such as NVFP4, are not enabled by this diffusion
+FP8 path until they have separate model and quality validation.
+
+## Hardware Support
+
+| Device | Support |
+|--------|---------|
+| NVIDIA Blackwell GPU (SM 100+) | ✅ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ |
+| NVIDIA Ampere GPU (SM 80+) | ⭕ |
+| AMD ROCm | ⭕ |
+| Intel XPU | ⭕ |
+| Ascend NPU | ❌ |
+
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this guide.
+The optional CUTLASS FP8 runtime override requires CUDA SM89+.
+
+## Model Type Support
+
+### Diffusion Model
+
+| Model | HF checkpoint | Scope | Status |
+|-------|---------------|-------|--------|
+| Qwen-Image 2512 | `feizhai123/qwen-image-2512-modelopt-fp8-dynamic-all` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
+| Z-Image | `feizhai123/z-image-modelopt-fp8-conservative` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
+| FLUX.2-dev | `feizhai123/flux2-dev-modelopt-fp8` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
+| FLUX.2-klein 4B | `feizhai123/flux2-klein-4b-modelopt-fp8` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
+| HunyuanImage-3.0 | `feizhai123/hunyuan-image3-modelopt-fp8` | MoE diffusion transformer | Validated for ModelOpt FP8 checkpoints |
+| Wan2.2 | Not available | Diffusion transformer | Not validated |
+
+### Multi-Stage Omni/TTS Model
+
+| Model | Scope | Status |
+|-------|-------|--------|
+| Qwen3-Omni | Thinker language-model stage | ModelOpt FP8 checkpoint path |
+| Qwen3-TTS | TTS language-model stage | Not validated |
+
+Audio encoder, vision encoder, talker, and code2wav stages stay in BF16 unless
+a model-specific guide documents otherwise.
+
+### Multi-Stage Diffusion Model
+
+ModelOpt checkpoints must be routed to the stage whose checkpoint contains the
+ModelOpt `quantization_config`. BAGEL and GLM-Image are not listed as validated
+ModelOpt targets yet.
+
+## Configuration
+
+For pre-quantized ModelOpt FP8 checkpoints, no `--quantization fp8` flag is
+needed. The checkpoint config selects the ModelOpt path.
+
+Online serving:
+
+```bash
+vllm serve <modelopt-fp8-checkpoint> \
+  --omni \
+  --tensor-parallel-size <N> \
+  --force-cutlass-fp8
+```
+
+Offline inference:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model <modelopt-fp8-checkpoint> \
+  --tensor-parallel-size <N> \
+  --prompt "a red ceramic teapot on a wooden table" \
+  --height 1024 \
+  --width 1024 \
+  --num-inference-steps 20 \
+  --seed 42 \
+  --output outputs/modelopt_fp8.png
+```
+
+Python API:
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="<modelopt-fp8-checkpoint>",
+    tensor_parallel_size=2,
+    force_cutlass_fp8=True,
+)
+```
+
+## Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `force_cutlass_fp8` / `--force-cutlass-fp8` | bool | `False` | Force CUTLASS FP8 linear kernels for supported ModelOpt FP8 diffusion stages on CUDA SM89+ |
+
+## Validation and Notes
+
+1. Compare the ModelOpt FP8 checkpoint against the BF16 baseline with the same
+   prompt, resolution, seed, and inference steps.
+2. Use `tests/diffusion/quantization/test_quantization_quality.py` with
+   `VLLM_OMNI_QUALITY_CONFIGS` to validate local baseline and quantized model
+   paths.
+3. Report LPIPS, PSNR, MAE, throughput, latency, and peak memory when adding a
+   new validated ModelOpt diffusion checkpoint.
+4. Keep `--quantization fp8` for online FP8 from BF16 checkpoints; use this
+   ModelOpt path only when the checkpoint already contains ModelOpt FP8 weights
+   and scales.

docs/user_guide/quantization/mxfp8.md

@@ -0,0 +1,187 @@
+# W8A8 MXFP8 Quantization
+
+## Overview
+
+W8A8 MXFP8 (Microscaling FP8) quantizes both weights and activations to FP8
+using the OCP MX format: groups of 32 K-dimension elements share a single
+`float8_e8m0fnu` exponent scale. This gives better accuracy than channel-wise
+FP8 while keeping the same 8-bit weight footprint.
+
+This method supports two modes:
+
+| Mode | Description |
+|------|-------------|
+| **Online** | BF16 weights are quantized to MXFP8 at load time — no pre-processing needed |
+| **Offline** | msModelSlim-exported MXFP8 weights converted to diffusers format via `merge_mxfp8_checkpoint.py` — weights and scales are loaded directly from the preprocessed checkpoint |
+
+## Hardware Support
+
+| Device | Support |
+|--------|---------|
+| NVIDIA Blackwell GPU (SM 100+) | ⭕ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ⭕ |
+| NVIDIA Ampere GPU (SM 80+) | ⭕ |
+| AMD ROCm | ⭕ |
+| Intel XPU | ⭕ |
+| Ascend NPU (Atlas 950 A5) | ✅ |
+
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
+guide.
+
+## Model Type Support
+
+### Diffusion Model (Wan2.2)
+
+| Model | Mode | Notes |
+|-------|------|-------|
+| Wan2.2-T2V-A14B | Online + Offline | MoE cascade; quantizes two transformers (`transformer` + `transformer_2`) |
+| Wan2.2-I2V-A14B | Online + Offline | MoE cascade; quantizes two transformers (`transformer` + `transformer_2`) |
+| Wan2.2-TI2V-5B | Online + Offline | Single transformer |
+
+### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
+
+| Model | Status | Notes |
+|-------|--------|-------|
+| Qwen3-Omni | Not validated | — |
+| Qwen3-TTS | Not validated | — |
+
+### Multi-Stage Diffusion Model (BAGEL, GLM-Image)
+
+| Model | Status | Notes |
+|-------|--------|-------|
+| BAGEL | Not validated | — |
+| GLM-Image | Not validated | — |
+
+## Configuration
+
+### Online Mode
+
+Online mode requires no pre-processing. vLLM-Omni quantizes BF16 weights to
+MXFP8 at load time.
+
+Python API:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(model="<your-model>", quantization="mxfp8")
+
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+CLI:
+
+```bash
+python text_to_video.py --model <your-model> --quantization mxfp8
+
+# Online serving
+vllm serve <your-model> --omni --quantization mxfp8
+```
+
+### Offline Mode
+
+Offline mode loads a pre-quantized checkpoint from msModelSlim. A preprocessing
+step converts the raw quantized output to the diffusers format expected by
+vLLM-Omni and injects the quantization config into `transformer/config.json` so
+that vLLM-Omni auto-detects the offline path without a `--quantization` flag.
+
+#### Step 1 — Quantize with msModelSlim
+
+```bash
+msmodelslim quant \
+  --model_path /path/to/Wan2.2-TI2V-5B-Diffusers \
+  --save_path  /path/to/wan2_2_ti2v_quantized_raw \
+  --device npu \
+  --model_type Wan2_2 \
+  --config_path /path/to/wan2_2_w8a8f8_mxfp.yaml \
+  --trust_remote_code True
+```
+
+After this step, `--save_path` contains the raw quantized safetensors files and
+a metadata JSON (`quant_model_description*.json`).
+
+For cascade MoE models (T2V-A14B, I2V-A14B), msModelSlim outputs two
+subdirectories: `high_noise_model/` and `low_noise_model/`.
+
+#### Step 2 — Preprocess with merge_mxfp8_checkpoint.py
+
+The script (`vllm_omni/quantization/tools/merge_mxfp8_checkpoint.py`):
+
+1. Copies the original diffusers model to `--output-path` (VAE, text encoder,
+   scheduler, etc. are preserved).
+2. Remaps tensor names from msModelSlim convention to diffusers convention.
+3. Saves the converted weights as `diffusion_pytorch_model.safetensors`.
+4. Copies the original `transformer/config.json` and injects
+   `quantization_config` so that vLLM-Omni auto-detects offline MXFP8.
+
+For cascade MoE models, steps 2–4 run separately for `high_noise_model/` →
+`transformer/` and `low_noise_model/` → `transformer_2/`.
+
+```bash
+python vllm_omni/quantization/tools/merge_mxfp8_checkpoint.py \
+  --model-type     Wan2.2-TI2V-5B \
+  --original-model /path/to/Wan2.2-TI2V-5B-Diffusers \
+  --quant-path     /path/to/wan2_2_ti2v_quantized_raw \
+  --output-path    /path/to/Wan2.2-TI2V-5B-MXFP8
+```
+
+| Argument | Description |
+|----------|-------------|
+| `--model-type` | Model variant: `Wan2.2-T2V-A14B`, `Wan2.2-I2V-A14B`, or `Wan2.2-TI2V-5B` |
+| `--original-model` | Root directory of the original BF16 diffusers model |
+| `--quant-path` | Root directory of the msModelSlim quantized output |
+| `--output-path` | Output directory for the merged model (created by the script) |
+
+The script outputs a complete diffusers model directory at `--output-path`,
+with each transformer subfolder containing:
+
+- `diffusion_pytorch_model.safetensors` — converted FP8 weights
+- `config.json` — original transformer config with `quantization_config` injected
+- `quant_model_description.json` — renamed quantization metadata (reference only)
+
+#### Step 3 — Serve
+
+```bash
+python text_to_video.py --model /path/to/Wan2.2-TI2V-5B-MXFP8
+
+# Online serving
+vllm serve /path/to/Wan2.2-TI2V-5B-MXFP8 --omni
+```
+
+Python API:
+
+```python
+omni = Omni(model="/path/to/Wan2.2-TI2V-5B-MXFP8")
+```
+
+!!! note
+    No `--quantization` flag is needed for offline mode. The preprocessing
+    script injects `quantization_config` into each `transformer/config.json`,
+    which vLLM-Omni reads automatically to activate the offline MXFP8 method.
+
+## Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `method` | str | — | Must be `"mxfp8"` |
+| `is_checkpoint_mxfp8_serialized` | bool | `False` | `True` for offline pre-quantized checkpoints; auto-set from `config.json` when using the preprocessing script |
+| `ignored_layers` | list[str] | `[]` | Layer name substrings to keep in BF16 (e.g. `"to_out"` matches `blocks.0.attn1.to_out.0`) |
+
+## Validation and Notes
+
+1. Online mode quantizes BF16 weights at load time using
+   `npu_dynamic_mx_quant`. This adds a one-time overhead on the first load
+   but requires no checkpoint preparation.
+2. Offline mode loads FP8 weights directly from the checkpoint. Scales are
+   stored as `uint8` bytes in safetensors (same bit layout as
+   `float8_e8m0fnu`) and are reinterpreted at load time without a dtype
+   conversion.
+3. If the offline checkpoint was produced with the old `merge_mxfp8_checkpoint.py`
+   interface (arguments `--quant-dir`, `--orig-dir`, `--meta-json`,
+   `--output-dir`), regenerate it with the current script. The old script
+   wrote a separate `quantization_config.json` that is not read by vLLM-Omni;
+   the current script injects the config directly into `transformer/config.json`.

docs/user_guide/quantization/overview.md

@@ -1,4 +1,4 @@
-# Quantization
+# Quantization
 
 vLLM-Omni exposes quantization through the unified `quantization_config`
 path. The same configuration entrypoint is used across diffusion-only models,
@@ -9,19 +9,20 @@ type has a different quantization scope.
 
 | Mode | Guide | Description | Methods |
 |------|-------|-------------|---------|
-| Online quantization | [Online Quantization](online.md) | vLLM-Omni computes quantized weights and scales while loading the model. | FP8 W8A8, Int8 W8A8 |
-| Pre-quantized checkpoints | Method-specific guides | The checkpoint or an offline quantizer provides quantized weights and scales before serving. | GGUF, AutoRound, msModelSlim, serialized Int8 |
+| Online quantization | [Online Quantization](online.md) | vLLM-Omni computes quantized weights and scales while loading the model. | FP8 W8A8, Int8 W8A8, MXFP8 W8A8 |
+| Runtime attention quantization | [Quantized KV Cache](quantized_kvcache.md) | vLLM-Omni dynamically quantizes eligible diffusion Flash Attention tensors during inference. | FP8 FA |
+| Pre-quantized checkpoints | Method-specific guides | The checkpoint or an offline quantizer provides quantized weights and scales before serving. | ModelOpt, GGUF, AutoRound, msModelSlim, serialized Int8, offline MXFP8 |
 
 ## Hardware Support
 
-| Device | FP8 W8A8 | Int8 W8A8 | GGUF | AutoRound | msModelSlim |
-|--------|----------|-----------|------|-----------|-------------|
-| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ✅ | ❌ |
-| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ✅ | ❌ |
-| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ✅ | ✅ | ❌ |
-| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ | ❌ |
-| Intel XPU | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
-| Ascend NPU | ❌ | ✅ | ❌ | ❌ | ✅ |
+| Device | FP8 W8A8 | Int8 W8A8 | ModelOpt | MXFP8 W8A8 | GGUF | AutoRound | msModelSlim |
+|--------|----------|-----------|----------|------------|------|-----------|-------------|
+| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ⭕ | ✅ | ✅ | ❌ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ⭕ | ✅ | ✅ | ❌ |
+| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
+| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ❌ |
+| Intel XPU | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
+| Ascend NPU | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ |
 
 Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
 guide. FP8 on Ampere may use a weight-only path where available.
@@ -39,6 +40,8 @@ otherwise.
 |--------|-------|------|----------------|--------|
 | FP8 W8A8 | [FP8](fp8.md) | Online W8A8 or checkpoint FP8 | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image family and other DiT models |
 | Int8 W8A8 | [Int8](int8.md) | Online or serialized W8A8 | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image and Z-Image |
+| ModelOpt | [ModelOpt](modelopt.md) | Pre-quantized FP8 checkpoints | Qwen-Image, Z-Image, FLUX.2, HunyuanImage-3.0 | Validated for ModelOpt FP8 diffusion checkpoints |
+| MXFP8 W8A8 | [MXFP8](mxfp8.md) | Online W8A8 or offline pre-quantized | Wan2.2-T2V-A14B, I2V-A14B, TI2V-5B | Ascend NPU only; validated for Wan2.2 |
 | GGUF | [GGUF](gguf.md) | Pre-quantized transformer weights | Qwen-Image | Validated where a model-specific GGUF adapter exists |
 | AutoRound | [AutoRound](autoround.md) | Pre-quantized W4A16 checkpoints | FLUX.1-dev; Qwen-Image/Wan2.2 not validated | Checkpoint-driven |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Pre-quantized Ascend checkpoints | Wan2.2 recipe; HunyuanImage-3.0 inference target | Ascend/NPU path |
@@ -52,8 +55,9 @@ in BF16 unless the model guide explicitly adds support.
 
 | Method | Guide | Scope | Example models | Status |
 |--------|-------|-------|----------------|--------|
-| FP8 | [FP8](fp8.md) | Thinker or language-model checkpoint config | Qwen3-Omni thinker | ModelOpt checkpoint path |
+| ModelOpt | [ModelOpt](modelopt.md) | Thinker or language-model checkpoint config | Qwen3-Omni thinker | ModelOpt checkpoint path |
 | Int8 | [Int8](int8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
+| MXFP8 | [MXFP8](mxfp8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | GGUF | [GGUF](gguf.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | AutoRound | [AutoRound](autoround.md) | Thinker or language-model checkpoint config | Qwen2.5-Omni, Qwen3-Omni | Supported through AutoRound checkpoints |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
@@ -67,6 +71,8 @@ attached to the intended stage rather than applied globally.
 |--------|-------|-------|----------------|--------|
 | FP8 | [FP8](fp8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Requires model-specific validation |
 | Int8 | [Int8](int8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Requires model-specific validation |
+| ModelOpt | [ModelOpt](modelopt.md) | Checkpoint-defined diffusion stage | BAGEL, GLM-Image | Requires model-specific validation |
+| MXFP8 | [MXFP8](mxfp8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Not validated |
 | GGUF | [GGUF](gguf.md) | Stage-specific transformer weights | BAGEL, GLM-Image | No validated adapter listed |
 | AutoRound | [AutoRound](autoround.md) | Checkpoint-defined stage | BAGEL, GLM-Image | No validated checkpoint listed |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Ascend-generated stage weights | GLM-Image | Requires model-specific adaptation |
@@ -94,7 +100,7 @@ config = build_quant_config({
 
 | Component | Default quantized? | Notes |
 |-----------|--------------------|-------|
-| Diffusion transformer | Yes | Primary target for FP8, Int8, GGUF, AutoRound, and msModelSlim |
+| Diffusion transformer | Yes | Primary target for FP8, Int8, ModelOpt, GGUF, AutoRound, and msModelSlim |
 | Text encoder | No | Keep BF16 unless a method-specific guide documents support |
 | VAE | No | Keep BF16; storage-only paths are method-specific |
 | Scheduler/tokenizer | No | Loaded from the base model repository |

docs/user_guide/quantization/quantized_kvcache.md

@@ -0,0 +1,125 @@
+# Quantized KV Cache
+
+## Overview
+
+In DiT-based image and video generation, Flash Attention can take a large share
+of denoising time, especially for high-resolution or long-frame workloads.
+vLLM-Omni supports online FP8 quantization for eligible diffusion Flash
+Attention (FA) to reduce FA latency while keeping model weights in their
+original dtype.
+
+This feature is configured through `diffusion_kv_cache_dtype` on
+`OmniDiffusionConfig` (CLI: `--diffusion-kv-cache-dtype`). It is intentionally
+**not** the same as vLLM's `--kv-cache-dtype`, which controls autoregressive
+language-model KV cache storage and defaults to `"auto"`. Diffusion FA
+quantization uses the dedicated diffusion flags so omni serve does not inherit
+that default.
+
+In vLLM-Omni diffusion pipelines, this is a runtime FA path: Q/K/V tensors are
+dynamically quantized before the attention operator. It does not quantize model
+weights and is separate from [FP8 W8A8](fp8.md), [Int8 W8A8](int8.md), or
+pre-quantized checkpoint formats.
+
+If `diffusion_kv_cache_dtype` is not set, behavior is unchanged and attention
+runs in the native dtype.
+
+## Hardware Support
+
+| Device | FP8 FA |
+|--------|--------|
+| Ascend NPU | ✅ |
+| NVIDIA GPU | ❌ |
+| AMD ROCm | ❌ |
+| Intel XPU | ❌ |
+
+Legend: `✅` supported, `❌` unsupported.
+
+FP8 FA is currently implemented only for the NPU Flash Attention backend. Other
+backends do not support `diffusion_kv_cache_dtype="fp8"` for diffusion attention
+and fall back to native dtype execution.
+
+## Model Type Support
+
+### Diffusion Model
+
+| Model | Scope | Status | Notes |
+|-------|-------|--------|-------|
+| Wan2.2 | Eligible DiT full-attention FA on Ascend NPU | Tested | Compare quality and latency against a BF16 baseline before production use |
+| Other diffusion models | Eligible DiT full-attention FA on Ascend NPU | Not tested | You can try `diffusion_kv_cache_dtype="fp8"`; tune `diffusion_kv_cache_skip_steps` and `diffusion_kv_cache_skip_layers` when higher precision is needed |
+
+### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
+
+Not tested for FP8 FA. Treat any use as experimental unless a model-specific
+guide documents support.
+
+### Multi-Stage Diffusion Model (BAGEL, GLM-Image)
+
+Not tested. If the diffusion stage uses the same NPU Flash Attention backend,
+`diffusion_kv_cache_dtype` may apply in theory; validate quality and latency for
+each stage and model.
+
+## Configuration
+
+Offline diffusion example:
+
+```bash
+python examples/offline_inference/image_to_video/image_to_video.py \
+    --model <your-wan2.2-model> \
+    --prompt "A cat sitting on a surfboard at the beach" \
+    --height 1280 \
+    --width 720 \
+    --num-frames 61 \
+    --num-inference-steps 4 \
+    --ulysses-degree 4 \
+    --vae-patch-parallel-size 4 \
+    --diffusion-kv-cache-dtype fp8 \
+    --diffusion-kv-cache-skip-steps "0,1" \
+    --diffusion-kv-cache-skip-layers "0-2"
+```
+
+Online serving:
+
+```bash
+vllm serve <your-model> --omni --diffusion-kv-cache-dtype fp8
+```
+
+Stage config:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    stage_type: diffusion
+    engine_args:
+      model_stage: dit
+      diffusion_kv_cache_dtype: "fp8"
+      diffusion_kv_cache_skip_steps: "0,1"
+      diffusion_kv_cache_skip_layers: "0-2"
+```
+
+Legacy YAML keys `kv_cache_dtype`, `kv_cache_skip_steps`, and
+`kv_cache_skip_layers` are still accepted when constructing
+`OmniDiffusionConfig` (for example via `from_kwargs`); prefer the `diffusion_*`
+names for new configs.
+
+## Parameters
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `diffusion_kv_cache_dtype` | str \| None | `None` | Set to `"fp8"` to enable dynamic FP8 FA on supported attention backends |
+| `diffusion_kv_cache_skip_steps` | str \| None | `None` | Denoising step selector to keep in native dtype, for example `"0,1,4-6"` |
+| `diffusion_kv_cache_skip_layers` | str \| None | `None` | Transformer layer selector to keep in native dtype, for example `"0-2,10"` |
+
+Selectors use comma-separated integers and inclusive ranges. Listed steps or
+layers skip FP8 FA; all other eligible full-attention forwards use the FP8 path.
+
+## Validation and Notes
+
+1. Compare generated images or videos against a BF16 baseline with the same
+   seed, prompt, resolution, frame count, and denoising steps.
+2. Use `diffusion_kv_cache_skip_steps` for denoising steps where quality is more
+   sensitive.
+3. Use `diffusion_kv_cache_skip_layers` for transformer layers that show visible quality
+   regressions.
+4. Report both latency and quality results when enabling this option for a new
+   model. For image or video models, include visual comparison and quantitative
+   metrics when available, such as PSNR or SSIM.

examples/offline_inference/bagel/README.md

@@ -195,6 +195,25 @@ stages:
     devices: "0,1"
 ```
 
+#### Hybrid Sharded Data Parallel (HSDP)
+
+For larger Bagel deployments on multiple GPUs, you can enable HSDP (Hybrid Sharded Data Parallel) by modifying the stage configuration (for example, [`bagel.yaml`](../../../vllm_omni/deploy/bagel.yaml)). HSDP shards transformer weights across GPUs to reduce per-GPU memory usage.
+
+1. **Enable HSDP**: Set `use_hsdp: true`.
+2. **Set shard size**: Set `hsdp_shard_size` to the number of GPUs used for sharding (for example, `4`).
+3. **Set replicate size**: Usually keep `hsdp_replicate_size: 1` unless you want replicated HSDP groups.
+4. **Set devices**: Specify the comma-separated GPU IDs used by the diffusion stage (for example, `"0,1,2,3"`).
+
+Example configuration for HSDP across 4 GPUs:
+```yaml
+  - stage_id: 1
+    devices: "0,1,2,3"
+    parallel_config:
+      use_hsdp: true
+      hsdp_shard_size: 4
+      hsdp_replicate_size: 1
+```
+
 Then pass the custom deploy YAML:
 
 ```bash

examples/offline_inference/covo_audio/README.md

@@ -0,0 +1,76 @@
+# Covo-Audio-Chat (Offline Inference)
+
+## 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.
+
+> **Note**
+> Covo-Audio code2wav requires `torchdiffeq`. Install it with: `pip install torchdiffeq`
+
+## Run examples
+
+Get into the example folder:
+```bash
+cd examples/offline_inference/covo_audio
+```
+
+### Audio input chat
+
+Using the default audio asset:
+```bash
+python end2end.py
+```
+
+Using a custom audio file:
+```bash
+python end2end.py --audio-path /path/to/audio.wav
+```
+
+Using a local model:
+```bash
+python end2end.py -m /path/to/Covo-Audio-Chat --output-dir ./my_output
+```
+
+### Command-line Arguments
+
+| Argument | Short | Default | Description |
+|----------|-------|---------|-------------|
+| `--model-name` | `-m` | `tencent/Covo-Audio-Chat` | Model path or HuggingFace model ID |
+| `--text` | `-t` | `请回答这段音频里的问题。` | Text prompt / question for the audio |
+| `--audio-path` | `-a` | default audio asset | Path to local audio file |
+| `--sampling-rate` | | `16000` | Sampling rate for audio loading (Hz) |
+| `--output-dir` | | `./output_audio` | Output directory for generated files |
+| `--num-prompts` | | `1` | Number of prompts to generate |
+| `--stage-configs-path` | | (auto) | Path to stage configs YAML file |
+| `--log-stats` | | `false` | Enable detailed statistics logging |
+| `--stage-init-timeout` | | `300` | Stage initialization timeout (seconds) |
+| `--batch-timeout` | | `5` | Batching timeout (seconds) |
+| `--init-timeout` | | `300` | Overall initialization timeout (seconds) |
+| `--shm-threshold-bytes` | | `65536` | Shared memory threshold (bytes) |
+
+## Pipeline
+
+Covo-Audio-Chat uses a 2-stage pipeline:
+
+- **Stage 0 (fused_thinker_talker):** The 7B LLM generates interleaved text and audio tokens in a single autoregressive pass.
+- **Stage 1 (code2wav):** A BigVGAN-based vocoder converts the extracted audio codes into a 24kHz WAV waveform.
+
+## Output
+
+The script generates two files per request in the output directory:
+
+- `{request_id}.txt` -- prompt and generated text
+- `{request_id}.wav` -- generated audio (24kHz WAV)
+
+## FAQ
+
+If you encounter `ModuleNotFoundError: No module named 'librosa'`, install it with:
+```bash
+pip install librosa
+```
+
+## Environment
+
+- GPU: 1x A100 (80 GiB)
+- Stage 0 (7B LLM): ~16 GiB VRAM
+- Stage 1 (BigVGAN vocoder): ~2 GiB VRAM

examples/offline_inference/covo_audio/end2end.py

@@ -0,0 +1,222 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+This example shows how to use vLLM-Omni for running offline inference
+with the correct prompt format on Covo-Audio-Chat.
+
+Usage:
+    python end2end.py --audio-path /path/to/audio.wav
+"""
+
+import os
+
+import soundfile as sf
+from vllm.assets.audio import AudioAsset
+from vllm.multimodal.media.audio import load_audio
+from vllm.sampling_params import SamplingParams
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.model_executor.models.covo_audio.prompt_utils import (
+    COVO_AUDIO_INPUT_PREFIX,
+    build_covo_audio_chat_prompt,
+)
+
+SEED = 42
+
+
+def get_audio_query(
+    question: str | None = None,
+    audio_path: str | None = None,
+    sampling_rate: int = 16000,
+) -> dict:
+    if question is None:
+        question = "请回答这段音频里的问题。"
+    user_content = COVO_AUDIO_INPUT_PREFIX + question
+    prompt = build_covo_audio_chat_prompt(user_content)
+
+    if audio_path is None:
+        audio_data = AudioAsset("mary_had_lamb").audio_and_sample_rate
+    else:
+        import numpy as np
+
+        audio_signal, sr = load_audio(audio_path, sr=sampling_rate)
+        audio_data = (audio_signal.astype(np.float32), sr)
+
+    return {
+        "prompt": prompt,
+        "multi_modal_data": {"audio": audio_data},
+        "modalities": ["audio"],
+    }
+
+
+def main(args):
+    query_result = get_audio_query(
+        question=args.text,
+        audio_path=args.audio_path,
+        sampling_rate=args.sampling_rate,
+    )
+
+    omni = Omni(
+        model=args.model_name,
+        stage_configs_path=args.stage_configs_path,
+        log_stats=args.log_stats,
+        stage_init_timeout=args.stage_init_timeout,
+        batch_timeout=args.batch_timeout,
+        init_timeout=args.init_timeout,
+        shm_threshold_bytes=args.shm_threshold_bytes,
+    )
+
+    # Stage 0: fused_thinker_talker
+    # stop_token_ids=[151645] (<|im_end|>) and ignore_eos=True are required
+    # so the model generates interleaved text+audio tokens before stopping.
+    thinker_sampling_params = SamplingParams(
+        temperature=0.0,
+        top_p=1.0,
+        top_k=-1,
+        max_tokens=2048,
+        seed=SEED,
+        detokenize=True,
+        repetition_penalty=1.05,
+        stop_token_ids=[151645],
+        ignore_eos=True,
+    )
+    # Stage 1: code2wav (audio codes, not real token IDs — skip detokenize)
+    code2wav_sampling_params = SamplingParams(
+        temperature=0.0,
+        top_p=1.0,
+        top_k=-1,
+        max_tokens=2048,
+        seed=SEED,
+        detokenize=False,
+        repetition_penalty=1.1,
+    )
+
+    sampling_params_list = [
+        thinker_sampling_params,
+        code2wav_sampling_params,
+    ]
+
+    prompts = [query_result for _ in range(args.num_prompts)]
+
+    omni_outputs = omni.generate(prompts, sampling_params_list)
+
+    output_dir = args.output_dir
+    os.makedirs(output_dir, exist_ok=True)
+
+    for stage_outputs in omni_outputs:
+        output = stage_outputs.request_output
+        if stage_outputs.final_output_type == "text":
+            request_id = output.request_id
+            text_output = output.outputs[0].text
+            prompt_text = output.prompt
+            out_txt = os.path.join(output_dir, f"{request_id}.txt")
+            lines = [
+                "Prompt:\n",
+                str(prompt_text) + "\n",
+                "vllm_text_output:\n",
+                str(text_output).strip() + "\n",
+            ]
+            try:
+                with open(out_txt, "w", encoding="utf-8") as f:
+                    f.writelines(lines)
+            except Exception as e:
+                print(f"[Warn] Failed writing text file {out_txt}: {e}")
+            print(f"Request ID: {request_id}, Text saved to {out_txt}")
+        elif stage_outputs.final_output_type == "audio":
+            request_id = output.request_id
+            audio_tensor = output.outputs[0].multimodal_output.get("audio")
+            if audio_tensor is None:
+                continue
+            output_wav = os.path.join(output_dir, f"{request_id}.wav")
+            audio_numpy = audio_tensor.float().detach().cpu().numpy()
+            if audio_numpy.ndim > 1:
+                audio_numpy = audio_numpy.flatten()
+            sf.write(output_wav, audio_numpy, samplerate=24000, format="WAV")
+            print(f"Request ID: {request_id}, Audio saved to {output_wav}")
+
+    omni.close()
+
+
+def parse_args():
+    parser = FlexibleArgumentParser(description="Offline inference demo for Covo-Audio-Chat")
+    parser.add_argument(
+        "--model-name",
+        "-m",
+        type=str,
+        default="tencent/Covo-Audio-Chat",
+        help="Model path or HuggingFace model ID.",
+    )
+    parser.add_argument(
+        "--text",
+        "-t",
+        type=str,
+        default=None,
+        help="Text prompt / question for the audio.",
+    )
+    parser.add_argument(
+        "--audio-path",
+        "-a",
+        type=str,
+        default=None,
+        help="Path to local audio file. Uses default asset if not provided.",
+    )
+    parser.add_argument(
+        "--sampling-rate",
+        type=int,
+        default=16000,
+        help="Sampling rate for audio loading (default: 16000).",
+    )
+    parser.add_argument(
+        "--stage-configs-path",
+        type=str,
+        default=None,
+        help="Path to stage configs YAML file.",
+    )
+    parser.add_argument(
+        "--log-stats",
+        action="store_true",
+        default=False,
+        help="Enable writing detailed statistics.",
+    )
+    parser.add_argument(
+        "--stage-init-timeout",
+        type=int,
+        default=300,
+        help="Timeout for initializing a single stage in seconds.",
+    )
+    parser.add_argument(
+        "--batch-timeout",
+        type=int,
+        default=5,
+        help="Timeout for batching in seconds.",
+    )
+    parser.add_argument(
+        "--init-timeout",
+        type=int,
+        default=300,
+        help="Timeout for initializing stages in seconds.",
+    )
+    parser.add_argument(
+        "--shm-threshold-bytes",
+        type=int,
+        default=65536,
+        help="Threshold for using shared memory in bytes.",
+    )
+    parser.add_argument(
+        "--output-dir",
+        default="./output_audio",
+        help="Output directory for generated files.",
+    )
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        default=1,
+        help="Number of prompts to generate.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

examples/offline_inference/hunyuan_image3/README.md

@@ -1,172 +1,155 @@
 # HunyuanImage-3.0-Instruct
 
-## Set up
+This example runs HunyuanImage-3.0-Instruct offline with the unified deploy
+YAMLs under `vllm_omni/deploy/`.
 
-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.
+## Deploy Configs
 
-## Run examples
+| File | Topology | Default use |
+| :--- | :--- | :--- |
+| `vllm_omni/deploy/hunyuan_image3.yaml` | AR + DiT | Default for `text2img` and `img2img`. |
+| `vllm_omni/deploy/hunyuan_image3_ar.yaml` | AR only | Default for `img2text` and `text2text`. |
+| `vllm_omni/deploy/hunyuan_image3_dit.yaml` | DiT only | Standalone diffusion stage. Pass it explicitly with `--deploy-config`. |
 
-**Note**: These examples work with the default configuration on **8x NVIDIA L40S (48GB)**. For different GPU setups, modify the stage configuration to adjust device allocation and memory utilization.
+The example chooses a deploy config automatically when `--deploy-config` and
+`--stage-configs-path` are both omitted:
 
-Get into the hunyuan_image3 folder:
+| `--modality` | `mode` passed to Omni | Default deploy |
+| :--- | :--- | :--- |
+| `text2img` | `text-to-image` | `hunyuan_image3.yaml` |
+| `img2img` | `image-editing` | `hunyuan_image3.yaml` |
+| `img2text` | `image-to-text` | `hunyuan_image3_ar.yaml` |
+| `text2text` | `text-to-text` | `hunyuan_image3_ar.yaml` |
+
+`--modality` is an offline example convenience flag. It maps to the internal
+`mode` argument passed to `Omni(...)` by this script. HunyuanImage3 uses
+separate deploy YAMLs for AR + DiT, AR-only, and DiT-only topologies, so the
+stage topology is selected by the deploy file rather than by YAML mode
+overrides.
+
+Online serving does not expose a `--modality` flag or accept `mode` as an API
+request field. Choose the deploy topology when starting the server with
+`--deploy-config`, then use the OpenAI-compatible endpoint and request shape for
+the scenario. The `modalities` request field is used by the chat completions
+path; the image endpoints infer the image task from the endpoint and payload.
+
+| Online scenario | Server deploy | Request |
+| :--- | :--- | :--- |
+| Text to image | `--deploy-config vllm_omni/deploy/hunyuan_image3.yaml` | `POST /v1/images/generations`, or `POST /v1/chat/completions` with `"modalities": ["image"]`. |
+| Image editing | `--deploy-config vllm_omni/deploy/hunyuan_image3.yaml` | `POST /v1/images/edits`. |
+| Image/text to text | `--deploy-config vllm_omni/deploy/hunyuan_image3_ar.yaml` | `POST /v1/chat/completions` for text output, for example with `"modalities": ["text"]`. |
+| DiT-only image generation | `--deploy-config vllm_omni/deploy/hunyuan_image3_dit.yaml` | `POST /v1/images/generations`. |
+
+## Run Examples
+
+Text to image, using the default AR + DiT deploy:
 
 ```bash
-cd examples/offline_inference/hunyuan_image3
+python examples/offline_inference/hunyuan_image3/end2end.py \
+  --model tencent/HunyuanImage-3.0-Instruct \
+  --modality text2img \
+  --prompts "A cute cat sitting on a windowsill watching the sunset"
 ```
 
-### Modality Control
-
-HunyuanImage-3.0-Instruct supports multiple modality modes. You can control the mode using the `--modality` argument:
-
-#### Text to Image (text2img)
-
-- **Pipeline**: Text → AR (CoT + latent tokens) → DiT (denoise) → VAE Decode → Image
-- **Stages Used**: Stage 0 (AR) + Stage 1 (DiT)
-- **KV Transfer**: AR sends KV cache to DiT for conditioned generation
-- **Default Config**: `hunyuan_image3_t2i.yaml`
+Image editing, using the default AR + DiT deploy:
 
 ```bash
-python end2end.py --model tencent/HunyuanImage-3.0-Instruct \
-                  --modality text2img \
-                  --prompts "A cute cat sitting on a windowsill watching the sunset"
+python examples/offline_inference/hunyuan_image3/end2end.py \
+  --model tencent/HunyuanImage-3.0-Instruct \
+  --modality img2img \
+  --image-path /path/to/image.png \
+  --prompts "Make the petals neon pink"
 ```
 
-**With VAE tiling (required on A100 GPUs):**
-```bash
-python end2end.py --model tencent/HunyuanImage-3.0-Instruct \
-                  --modality text2img \
-                  --prompts "A cute cat sitting on a windowsill watching the sunset" \
-                  --vae-use-tiling
-```
-
-#### Image to Image (img2img)
-
-- **Pipeline**: Image + Text → AR (CoT + recaption + latent) → DiT → Edited Image
-- **Stages Used**: Stage 0 (AR) + Stage 1 (DiT)
-- **KV Transfer**: AR sends KV cache to DiT
-- **Default Config**: `hunyuan_image3_it2i.yaml`
+Image to text, using the AR-only deploy:
 
 ```bash
-python end2end.py --model tencent/HunyuanImage-3.0-Instruct \
-                  --modality img2img \
-                  --image-path /path/to/image.png \
-                  --prompts "Make the petals neon pink"
+python examples/offline_inference/hunyuan_image3/end2end.py \
+  --model tencent/HunyuanImage-3.0-Instruct \
+  --modality img2text \
+  --image-path /path/to/image.jpg \
+  --prompts "Describe the content of the picture."
 ```
 
-#### Image to Text (img2text)
-
-- **Pipeline**: Image + Question → AR → Text description
-- **Stages Used**: Stage 0 (AR) only
-- **Default Config**: `hunyuan_image3_i2t.yaml`
+Text to text, using the AR-only deploy:
 
 ```bash
-python end2end.py --model tencent/HunyuanImage-3.0-Instruct \
-                  --modality img2text \
-                  --image-path /path/to/image.jpg \
-                  --prompts "Describe the content of the picture."
+python examples/offline_inference/hunyuan_image3/end2end.py \
+  --model tencent/HunyuanImage-3.0-Instruct \
+  --modality text2text \
+  --prompts "What is the capital of France?"
 ```
 
-#### Text to Text (text2text)
-
-- **Pipeline**: Text → AR → Text
-- **Stages Used**: Stage 0 (AR) only
-- **Default Config**: `hunyuan_image3_t2t.yaml`
+Standalone DiT, using the DiT-only deploy explicitly:
 
 ```bash
-python end2end.py --model tencent/HunyuanImage-3.0-Instruct \
-                  --modality text2text \
-                  --prompts "What is the capital of France?"
+python examples/offline_inference/hunyuan_image3/end2end.py \
+  --model tencent/HunyuanImage-3.0-Instruct \
+  --modality text2img \
+  --deploy-config vllm_omni/deploy/hunyuan_image3_dit.yaml \
+  --prompts "A cinematic portrait of an astronaut in a greenhouse"
 ```
 
-### Inference Steps & Guidance
+Override the default full AR + DiT deploy explicitly:
 
-Control generation quality for image modalities:
+```bash
+python examples/offline_inference/hunyuan_image3/end2end.py \
+  --model tencent/HunyuanImage-3.0-Instruct \
+  --modality text2img \
+  --deploy-config vllm_omni/deploy/hunyuan_image3.yaml \
+  --prompts "A cute cat"
+```
+
+## Additional Config
+
+You can pass diffusion worker `additional_config` from the offline example as a JSON object.
+This maps to the upstream vLLM `VllmConfig.additional_config` platform extension field:
+https://docs.vllm.ai/en/stable/api/vllm/config/#vllm.config.VllmConfig.additional_config
 
 ```bash
 python end2end.py --modality text2img \
-                  --steps 50 \
-                  --guidance-scale 5.0 \
-                  --height 1024 --width 1024 \
-                  --prompts "A photo-realistic sunset over the ocean"
+                  --prompts "A cute cat" \
+                  --additional-config '{"torchair_graph_config":{"enabled":true}}'
 ```
 
-### Key Arguments
 
-#### 📌 Command Line Arguments (end2end.py)
+## Key Arguments
 
-| Argument               | Type   | Default                              | Description                                                  |
-| :--------------------- | :----- | :----------------------------------- | :----------------------------------------------------------- |
-| `--model`              | string | `tencent/HunyuanImage-3.0-Instruct` | Model path or name                                           |
-| `--modality`           | choice | `text2img`                           | Modality: `text2img`, `img2img`, `img2text`, `text2text`     |
-| `--prompts`            | list   | `None`                               | Input text prompts                                           |
-| `--image-path`         | string | `None`                               | Input image path (for `img2img`/`img2text`)                  |
-| `--output`             | string | `.`                                  | Output directory for saved images                            |
-| `--steps`              | int    | `50`                                 | Number of inference steps                                    |
-| `--guidance-scale`     | float  | `5.0`                                | Classifier-free guidance scale                               |
-| `--seed`               | int    | `42`                                 | Random seed                                                  |
-| `--height`             | int    | `1024`                               | Output image height                                          |
-| `--width`              | int    | `1024`                               | Output image width                                           |
-| `--bot-task`           | string | auto                                 | Override prompt task (e.g. `it2i_think`, `t2i_recaption`)    |
-| `--sys-type`           | string | auto                                 | Override system prompt type (e.g. `en_unified`, `en_vanilla`) |
-| `--stage-configs-path` | string | auto                                 | Custom stage config YAML path                                |
-| `--enforce-eager`      | flag   | `False`                              | Disable torch.compile                                        |
-| `--init-timeout`       | int    | `300`                                | Initialization timeout (seconds)                             |
-| `--vae-use-tiling`     | flag   | `False`                              | Enable VAE tiling for memory optimization (required to avoid OOM on A100) |
+| Argument | Description |
+| :--- | :--- |
+| `--deploy-config` | Preferred config path for unified deploy YAMLs. |
+| `--stage-configs-path` | Legacy stage config path, kept only for compatibility. Prefer `--deploy-config`. |
+| `--additional-config` | JSON object forwarded to diffusion worker `additional_config`. |
+| `--modality` | Offline-only convenience flag. One of `text2img`, `img2img`, `img2text`, `text2text`. It selects prompt formatting, internal `mode`, and default deploy config for this script. Online serving uses `--deploy-config` plus the endpoint and, for chat completions, request `modalities` instead. |
+| `--steps` | Number of diffusion inference steps for image generation. |
+| `--guidance-scale` | Classifier-free guidance scale for image generation. |
+| `--height`, `--width` | Output image size for `text2img`. |
+| `--bot-task` | Override prompt mode. `none`, `think`, `recaption`, `think_recaption`, or `vanilla`. |
+| `--sys-type` | Override the system prompt type, for example `en_unified` or `en_vanilla`. |
+| `--vae-use-tiling` | Enable VAE tiling for memory reduction. |
 
-------
+## Notes
 
-#### ⚙️ Stage Configurations
-
-| Config YAML                         | Modality  | Stages | GPUs   | Description                           |
-| :---------------------------------- | :-------- | :----- | :----- | :------------------------------------ |
-| `hunyuan_image3_t2i.yaml`           | text2img  | 2      | 8      | T2I with AR→DiT, 4 GPU each          |
-| `hunyuan_image3_it2i.yaml`          | img2img   | 2      | 8      | IT2I with AR→DiT, 4 GPU each         |
-| `hunyuan_image3_i2t.yaml`           | img2text  | 1      | 4      | I2T (AR only)                         |
-| `hunyuan_image3_t2t.yaml`           | text2text | 1      | 4      | T2T (AR only)                         |
-| `hunyuan_image3_t2i_2gpu.yaml`      | text2img  | 2      | 2      | T2I for 2-GPU setups                  |
-| `hunyuan_image3_moe.yaml`           | text2img  | 2      | 8      | T2I with MoE AR→DiT KV reuse          |
-| `hunyuan_image3_moe_dit_2gpu_fp8.yaml` | text2img | 2   | 2      | T2I with FP8 quantization             |
-
-------
-
-## Using MoE Config
-
-The `hunyuan_image3_moe.yaml` config enables AR→DiT KV cache reuse with 8 GPUs (4 for AR + 4 for DiT).
-
-```bash
-python end2end.py --model tencent/HunyuanImage-3.0-Instruct \
-                  --modality text2img \
-                  --stage-configs-path hunyuan_image3_moe.yaml \
-                  --prompts "A cute cat"
-```
-
-------
+- `hunyuan_image3_ar.yaml` is a 4-card AR-only text/comprehension deploy.
+- `hunyuan_image3_dit.yaml` is a single-stage DiT deploy with `stage_id: 0`.
+- The old HunyuanImage3 YAMLs under `model_executor/stage_configs/` and `platforms/*/stage_configs/` have been folded into the deploy YAMLs.
 
 ## Prompt Format
 
 HunyuanImage-3.0-Instruct uses an instruct chat template:
 
-```
-<|startoftext|>{system_prompt}\n\nUser: {<img>?}{user_prompt}\n\nAssistant: {trigger_tag?}
+```text
+<|startoftext|>{system_prompt}
+
+User: {<img>?}{user_prompt}
+
+Assistant: {trigger_tag?}
 ```
 
-- `<img>`: Placeholder for each input image (single token; expanded by the multimodal pipeline)
-- Trigger tags: `<think>` (CoT), `<recaption>` (recaptioning) — placed AFTER `Assistant: `
-- System prompt: Auto-selected based on task
-- `t2i_vanilla` is the only task that uses the bare pretrain template (no chat structure)
+- `<img>`: Placeholder for each input image (single token; expanded by the multimodal pipeline).
+- Trigger tags: `<think>` for CoT and `<recaption>` for recaptioning, placed after `Assistant: `.
+- System prompt: Auto-selected from `task` and `bot_task`.
+- `bot_task='vanilla'` with `task='t2i'` uses the bare pretrain template.
 
 The shared `vllm_omni.diffusion.models.hunyuan_image3.prompt_utils.build_prompt_tokens()`
-helper handles segment-by-segment tokenization (matches HF `apply_chat_template` byte-for-byte).
-
-------
-
-## FAQ
-
-- **OOM errors**: Decrease `gpu_memory_utilization` in the YAML stage config, use a smaller `max_num_batched_tokens`, or enable VAE tiling with `--vae-use-tiling` (required on A100 GPUs).
-- **Custom image sizes**: Use `--height` and `--width` flags (multiples of 16 recommended).
-
-| Stage             | VRAM (approx)        |
-| :---------------- | :------------------- |
-| Stage 0 (AR)      | ~15 GiB + KV Cache   |
-| Stage 1 (DiT)     | ~30 GiB              |
-| Total (8-GPU)     | ~45 GiB + KV Cache   |
+helper handles segment-by-segment tokenization and matches HF `apply_chat_template`.

examples/offline_inference/hunyuan_image3/end2end.py

@@ -1,141 +1,93 @@
 """
 HunyuanImage-3.0-Instruct unified end-to-end inference script.
-
-Supports all modalities through a single entry point:
-  - text2img:  Text → AR → DiT → Image
-  - img2img:   Text+Image → AR → DiT → Edited Image (IT2I)
-  - img2text:  Image+Text → AR → Text description (I2T)
-  - text2text: Text → AR → Text (comprehension, no image)
-
-Usage:
-    python end2end.py --modality text2img --prompts "A cute cat"
-    python end2end.py --modality img2img --image-path input.png --prompts "Make it snowy"
-    python end2end.py --modality img2text --image-path input.png --prompts "Describe this image"
 """
 
 import argparse
+import json
 import os
+from pathlib import Path
 
 from vllm_omni.diffusion.models.hunyuan_image3.prompt_utils import (
+    MAX_IMAGES_PER_REQUEST,
     build_prompt_tokens,
+    resolve_stop_token_ids,
+    resolve_sys_type,
 )
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniPromptType
-from vllm_omni.platforms import current_omni_platform
 
-# task -> (sys_type, bot_task, trigger_tag)
-_TASK_PRESETS: dict[str, tuple[str, str | None, str | None]] = {
-    "t2t": ("en_unified", None, None),
-    "i2t": ("en_unified", None, None),
-    "it2i_think": ("en_unified", "think", "<think>"),
-    "it2i_recaption": ("en_unified", "recaption", "<recaption>"),
-    "t2i_think": ("en_unified", "think", "<think>"),
-    "t2i_recaption": ("en_unified", "recaption", "<recaption>"),
-    "t2i_vanilla": ("en_vanilla", "image", None),
+_REPO_ROOT = Path(__file__).resolve().parents[3]
+_DEFAULT_DEPLOY_CONFIG = str(_REPO_ROOT / "vllm_omni" / "deploy" / "hunyuan_image3.yaml")
+_DEFAULT_AR_DEPLOY_CONFIG = str(_REPO_ROOT / "vllm_omni" / "deploy" / "hunyuan_image3_ar.yaml")
+
+_MODALITY_TASK_MAP: dict[str, tuple[str, str | None]] = {
+    "text2img": ("t2i", "think"),
+    "img2img": ("it2i", "think"),
+    "img2text": ("i2t", None),
+    "text2text": ("t2t", None),
 }
 
-# Modality → prompt_utils task mapping
-_MODALITY_TASK_MAP = {
-    "text2img": "t2i_think",
-    "img2img": "it2i_think",
-    "img2text": "i2t",
-    "text2text": "t2t",
+_MODALITY_DEFAULT_DEPLOY_CONFIG = {
+    "text2img": _DEFAULT_DEPLOY_CONFIG,
+    "img2img": _DEFAULT_DEPLOY_CONFIG,
+    "img2text": _DEFAULT_AR_DEPLOY_CONFIG,
+    "text2text": _DEFAULT_AR_DEPLOY_CONFIG,
 }
 
-
-# Modality → default stage config
-_MODALITY_DEFAULT_CONFIG = {
-    "text2img": "hunyuan_image3_t2i.yaml",
-    "img2img": "hunyuan_image3_it2i.yaml",
-    "img2text": "hunyuan_image3_i2t.yaml",
-    "text2text": "hunyuan_image3_t2t.yaml",
+_MODALITY_MODE = {
+    "text2img": "text-to-image",
+    "img2img": "image-editing",
+    "img2text": "image-to-text",
+    "text2text": "text-to-text",
 }
 
 
 def parse_args():
     parser = argparse.ArgumentParser(description="HunyuanImage-3.0-Instruct end-to-end inference.")
-    parser.add_argument(
-        "--model",
-        default="tencent/HunyuanImage-3.0-Instruct",
-        help="Model name or local path.",
-    )
+    parser.add_argument("--model", default="tencent/HunyuanImage-3.0-Instruct", help="Model name or local path.")
     parser.add_argument(
         "--modality",
         default="text2img",
-        choices=["text2img", "img2img", "img2text", "text2text"],
-        help="Modality mode to control stage execution.",
+        choices=list(_MODALITY_TASK_MAP),
     )
     parser.add_argument("--prompts", nargs="+", default=None, help="Input text prompts.")
     parser.add_argument(
         "--image-path",
         type=str,
         default=None,
-        help="Path to input image (for img2img/img2text).",
+        help="Input image path(s) for img2img/img2text. Comma-separated for multi-image (up to 3).",
     )
-    parser.add_argument(
-        "--output",
-        type=str,
-        default=".",
-        help="Output directory to save results.",
-    )
-
-    # Generation parameters
+    parser.add_argument("--output", type=str, default=".", help="Output directory to save results.")
     parser.add_argument("--steps", type=int, default=50, help="Number of inference steps.")
     parser.add_argument("--guidance-scale", type=float, default=5.0, help="Classifier-free guidance scale.")
     parser.add_argument("--seed", type=int, default=42, help="Random seed.")
     parser.add_argument("--height", type=int, default=1024, help="Output image height.")
     parser.add_argument("--width", type=int, default=1024, help="Output image width.")
-    parser.add_argument(
-        "--vae-use-tiling",
-        action="store_true",
-        help="Enable VAE tiling for memory optimization.",
-    )
-
-    # Prompt configuration
+    parser.add_argument("--vae-use-tiling", action="store_true", help="Enable VAE tiling.")
     parser.add_argument(
         "--bot-task",
         type=str,
         default=None,
-        help="Override prompt task (e.g. it2i_think, t2i_recaption). Default: auto from modality.",
+        choices=["none", "think", "recaption", "think_recaption", "vanilla"],
+        help="Override prompt mode. Default: auto from --modality.",
     )
-    parser.add_argument(
-        "--sys-type",
-        type=str,
-        default=None,
-        help="Override system prompt type (e.g. en_unified, en_vanilla).",
-    )
-
-    # Omni init args
-    parser.add_argument("--stage-configs-path", type=str, default=None, help="Custom stage config YAML path.")
+    parser.add_argument("--sys-type", type=str, default=None, help="Override system prompt type.")
+    parser.add_argument("--deploy-config", type=str, default=None, help="Custom deploy YAML path.")
+    parser.add_argument("--stage-configs-path", type=str, default=None, help="Custom legacy stage config YAML path.")
     parser.add_argument("--log-stats", action="store_true", default=False)
     parser.add_argument("--init-timeout", type=int, default=300, help="Initialization timeout in seconds.")
     parser.add_argument("--enforce-eager", action="store_true", help="Disable torch.compile.")
     parser.add_argument(
-        "--quantization",
+        "--additional-config",
         type=str,
         default=None,
-        choices=["fp8", "int8", "gguf"],
-        help="Quantization method for the transformer. "
-        "Options: 'fp8' (FP8 W8A8 on Ada/Hopper, weight-only on older GPUs), 'int8' (Int8 W8A8), 'gguf' (GGUF quantized weights). "
-        "Default: None (no quantization, uses BF16).",
+        help=(
+            "JSON object forwarded to Omni/additional_config, for example "
+            '\'{"torchair_graph_config":{"enabled":true}}\'. Different platforms may support different '
+            "configs. Make sure the configs are valid for the platform you are using. "
+            "Contents must be hashable."
+        ),
     )
-    parser.add_argument(
-        "--gguf-model",
-        type=str,
-        default=None,
-        help=("GGUF file path or HF reference for transformer weights. Required when --quantization gguf is set."),
-    )
-    parser.add_argument(
-        "--ignored-layers",
-        type=str,
-        default=None,
-        help="Comma-separated list of layer name patterns to skip quantization. "
-        "Only used when --quantization is set. "
-        "Available layers: to_qkv, to_out, add_kv_proj, to_add_out, img_mlp, txt_mlp, proj_out. "
-        "Example: --ignored-layers 'add_kv_proj,to_add_out'",
-    )
-
-    current_omni_platform.pre_register_and_update(parser)
 
     from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 
@@ -143,124 +95,143 @@ def parse_args():
     return parser.parse_args()
 
 
+def parse_additional_config(raw_value: str | None) -> dict | None:
+    """Parse a JSON string into an additional_config mapping."""
+    if raw_value is None:
+        return None
+
+    try:
+        additional_config = json.loads(raw_value)
+    except json.JSONDecodeError as exc:
+        raise ValueError(f"Invalid --additional-config JSON: {exc}") from exc
+
+    if additional_config is None:
+        return None
+    if not isinstance(additional_config, dict):
+        raise ValueError(f"--additional-config must decode to a JSON object, got {type(additional_config).__name__}")
+    return additional_config
+
+
 def main():
     args = parse_args()
     os.makedirs(args.output, exist_ok=True)
+    additional_config = parse_additional_config(args.additional_config)
 
-    # Determine task for prompt formatting
-    task = args.bot_task or _MODALITY_TASK_MAP[args.modality]
+    task, default_bot_task = _MODALITY_TASK_MAP[args.modality]
+    if args.bot_task is None:
+        bot_task: str | None = default_bot_task
+    elif args.bot_task == "none":
+        bot_task = None
+    else:
+        bot_task = args.bot_task
 
-    # Determine stage config
-    stage_configs_path = args.stage_configs_path or _MODALITY_DEFAULT_CONFIG[args.modality]
-    # Build quantization kwargs: use quantization_config dict when
-    # ignored_layers is specified so the list flows through OmniDiffusionConfig
-    quant_kwargs: dict[str, Any] = {}
-    ignored_layers = [s.strip() for s in args.ignored_layers.split(",") if s.strip()] if args.ignored_layers else None
-    if args.quantization == "gguf":
-        if not args.gguf_model:
-            raise ValueError("--gguf-model is required when --quantization gguf is set.")
-        quant_kwargs["quantization_config"] = {
-            "method": "gguf",
-            "gguf_model": args.gguf_model,
-        }
-    elif args.quantization and ignored_layers:
-        quant_kwargs["quantization_config"] = {
-            "method": args.quantization,
-            "ignored_layers": ignored_layers,
-        }
-    elif args.quantization:
-        quant_kwargs["quantization"] = args.quantization
+    if args.deploy_config is not None and args.stage_configs_path is not None:
+        raise ValueError("--deploy-config and --stage-configs-path are mutually exclusive.")
+
+    deploy_config = args.deploy_config
+    stage_configs_path = args.stage_configs_path
+    if deploy_config is None and stage_configs_path is None:
+        deploy_config = _MODALITY_DEFAULT_DEPLOY_CONFIG[args.modality]
 
-    # Build Omni
     omni_kwargs = {
         "model": args.model,
         "vae_use_tiling": args.vae_use_tiling,
-        "stage_configs_path": stage_configs_path,
         "log_stats": args.log_stats,
         "init_timeout": args.init_timeout,
         "enforce_eager": args.enforce_eager,
-        **quant_kwargs,
+        "mode": _MODALITY_MODE[args.modality],
     }
-    if args.modality in ("text2img", "img2img"):
-        omni_kwargs["mode"] = "text-to-image"
+
+    if additional_config is not None:
+        omni_kwargs["additional_config"] = additional_config
+    if deploy_config is not None:
+        omni_kwargs["deploy_config"] = deploy_config
+    else:
+        omni_kwargs["stage_configs_path"] = stage_configs_path
 
     omni = Omni(**omni_kwargs)
 
-    # Prepare prompts
     prompts = args.prompts or ["A cute cat"]
-    if not prompts:
-        print("[Info] No prompts provided, using default.")
-        prompts = ["A cute cat"]
-
-    # Load image if needed
-    input_image = None
+    input_images: list = []
     if args.modality in ("img2img", "img2text"):
-        if not args.image_path or not os.path.exists(args.image_path):
+        if not args.image_path:
             raise ValueError(f"--image-path required for {args.modality}, got: {args.image_path}")
         from PIL import Image
 
-        input_image = Image.open(args.image_path).convert("RGB")
+        image_paths = [p.strip() for p in args.image_path.split(",") if p.strip()]
+        if len(image_paths) > MAX_IMAGES_PER_REQUEST:
+            raise ValueError(
+                f"--image-path accepts at most {MAX_IMAGES_PER_REQUEST} images for "
+                f"HunyuanImage-3.0 IT2I, got {len(image_paths)}: {args.image_path}"
+            )
+        for image_path in image_paths:
+            if not os.path.exists(image_path):
+                raise ValueError(f"Image path does not exist: {image_path}")
+            input_images.append(Image.open(image_path).convert("RGB"))
+        if not input_images:
+            raise ValueError(f"--image-path produced no usable paths: {args.image_path!r}")
 
-    # Load tokenizer for segment-wise prompt tokenization (matches HF
-    # apply_chat_template byte-for-byte; see build_prompt_tokens docstring).
     from transformers import AutoTokenizer
 
     tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+    mm_image_payload = (input_images[0] if len(input_images) == 1 else input_images) if input_images else None
 
-    # Format prompts
     formatted_prompts: list[OmniPromptType] = []
-    for p in prompts:
-        token_ids = build_prompt_tokens(p, tokenizer, task=task, sys_type=args.sys_type)
-        preset_sys_type, _, _ = _TASK_PRESETS[task]
-        effective_sys_type = args.sys_type or preset_sys_type
+    for prompt in prompts:
+        build_kwargs: dict = {"task": task, "bot_task": bot_task, "sys_type": args.sys_type}
+        if input_images:
+            build_kwargs["num_images"] = len(input_images)
+        result = build_prompt_tokens(prompt, tokenizer, **build_kwargs)
+        token_ids = result.token_ids
+        effective_sys_type = args.sys_type or resolve_sys_type(bot_task)
 
-        # `prompt_token_ids` drives the AR stage (matches HF byte-for-byte).
-        # `prompt` and `use_system_prompt` are forwarded by ar2diffusion to
-        # the DiT stage so the diffusion pipeline can rebuild the same
-        # system prefix when constructing its model inputs.
         prompt_dict: dict = {
             "prompt_token_ids": token_ids,
-            "prompt": p,
+            "prompt": prompt,
             "use_system_prompt": effective_sys_type,
         }
-
         if args.modality == "text2img":
             prompt_dict["modalities"] = ["image"]
         elif args.modality == "img2img":
             prompt_dict["modalities"] = ["image"]
-            prompt_dict["multi_modal_data"] = {"image": input_image}
-            prompt_dict["height"] = input_image.height
-            prompt_dict["width"] = input_image.width
+            prompt_dict["multi_modal_data"] = {"image": mm_image_payload}
+            prompt_dict["height"] = input_images[0].height
+            prompt_dict["width"] = input_images[0].width
         elif args.modality == "img2text":
             prompt_dict["modalities"] = ["text"]
-            prompt_dict["multi_modal_data"] = {"image": input_image}
-        elif args.modality == "text2text":
+            prompt_dict["multi_modal_data"] = {"image": mm_image_payload}
+        else:
             prompt_dict["modalities"] = ["text"]
-
         formatted_prompts.append(prompt_dict)
 
-    # Build sampling params from defaults
     params_list = list(omni.default_sampling_params_list)
 
-    # Override diffusion params if applicable
     from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 
-    for i, sp in enumerate(params_list):
+    ar_stop_token_ids = resolve_stop_token_ids(task=task, bot_task=bot_task, tokenizer=tokenizer)
+    for sp in params_list:
         if isinstance(sp, OmniDiffusionSamplingParams):
             sp.num_inference_steps = args.steps
             sp.guidance_scale = args.guidance_scale
+            sp.guidance_scale_provided = True
             if args.seed is not None:
                 sp.seed = args.seed
-            if args.modality in ("text2img",):
+            if args.modality == "text2img":
                 sp.height = args.height
                 sp.width = args.width
+        elif hasattr(sp, "stop_token_ids"):
+            sp.stop_token_ids = ar_stop_token_ids
 
-    # Print configuration
     print(f"\n{'=' * 60}")
     print("HunyuanImage-3.0 Generation Configuration:")
     print(f"  Model: {args.model}")
     print(f"  Modality: {args.modality}")
-    print(f"  Stage config: {stage_configs_path}")
+    print(f"  Prompt task: {task}")
+    print(f"  Bot task: {bot_task}")
+    if deploy_config is not None:
+        print(f"  Deploy config: {deploy_config}")
+    else:
+        print(f"  Stage config: {stage_configs_path}")
     print(f"  Num stages: {omni.num_stages}")
     if args.modality in ("text2img", "img2img"):
         print(f"  Inference steps: {args.steps}")
@@ -270,27 +241,30 @@ def main():
         print(f"  Output size: {args.width}x{args.height}")
     if args.image_path:
         print(f"  Input image: {args.image_path}")
+    if additional_config is not None:
+        print(f"  Additional config: {additional_config}")
     print(f"  Prompts: {prompts}")
     print(f"{'=' * 60}\n")
 
-    # Generate
     omni_outputs = list(omni.generate(prompts=formatted_prompts, sampling_params_list=params_list))
-
-    # Process outputs
     img_idx = 0
     for req_output in omni_outputs:
-        # Text output (AR stage or text-only)
         ro = getattr(req_output, "request_output", None)
+        txt = ""
         if ro and getattr(ro, "outputs", None):
             txt = "".join(getattr(o, "text", "") or "" for o in ro.outputs)
-            if txt:
-                print(f"[Output] Text:\n{txt}")
+        if not txt:
+            ar_text = getattr(req_output, "custom_output", {}).get("ar_generated_text")
+            if isinstance(ar_text, list):
+                txt = "\n".join(text for text in ar_text if text)
+            else:
+                txt = ar_text or ""
+        if txt:
+            print(f"[Output] Text:\n{txt}")
 
-        # Image output (DiT stage)
         images = getattr(req_output, "images", None)
         if not images and ro and hasattr(ro, "images"):
             images = ro.images
-
         if images:
             for j, img in enumerate(images):
                 save_path = os.path.join(args.output, f"output_{img_idx}_{j}.png")

examples/offline_inference/image_to_video/image_to_video.py

@@ -102,6 +102,24 @@ def parse_args() -> argparse.Namespace:
         choices=["unipc", "euler"],
         help="Sampling solver for Wan2.2 pipelines. Use 'euler' for Lightning/Distill setups.",
     )
+    parser.add_argument(
+        "--diffusion-kv-cache-dtype",
+        type=str,
+        default=None,
+        help="Diffusion attention KV cache dtype (e.g. float8_e4m3fn). Separate from vLLM --kv-cache-dtype.",
+    )
+    parser.add_argument(
+        "--diffusion-kv-cache-skip-steps",
+        type=str,
+        default=None,
+        help="Diffusion KV-cache quantization skip-step selector, e.g. '0-9,20,25-30'.",
+    )
+    parser.add_argument(
+        "--diffusion-kv-cache-skip-layers",
+        type=str,
+        default=None,
+        help="Diffusion KV-cache quantization skip-layer selector, e.g. '0,1,4-8'.",
+    )
     parser.add_argument("--output", type=str, default="i2v_output.mp4", help="Path to save the video (mp4).")
     parser.add_argument("--fps", type=int, default=None, help="Frames per second for the output video.")
     parser.add_argument(
@@ -200,6 +218,13 @@ def parse_args() -> argparse.Namespace:
             "Default 1 means pure sharding (no replication). "
         ),
     )
+    parser.add_argument(
+        "--quantization",
+        type=str,
+        default=None,
+        choices=["fp8", "mxfp8", "int8", "gguf"],
+        help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 online quant (NPU). fp8: online FP8 (GPU).",
+    )
     parser.add_argument(
         "--enable-diffusion-pipeline-profiler",
         action="store_true",
@@ -302,13 +327,16 @@ def main():
         hsdp_shard_size=args.hsdp_shard_size,
         hsdp_replicate_size=args.hsdp_replicate_size,
     )
-    omni = Omni(
+    omni_kwargs = dict(
         model=args.model,
         enable_layerwise_offload=args.enable_layerwise_offload,
         vae_use_slicing=args.vae_use_slicing,
         vae_use_tiling=args.vae_use_tiling,
         boundary_ratio=args.boundary_ratio,
         flow_shift=args.flow_shift,
+        diffusion_kv_cache_dtype=args.diffusion_kv_cache_dtype,
+        diffusion_kv_cache_skip_steps=args.diffusion_kv_cache_skip_steps,
+        diffusion_kv_cache_skip_layers=args.diffusion_kv_cache_skip_layers,
         enable_cpu_offload=args.enable_cpu_offload,
         parallel_config=parallel_config,
         enforce_eager=args.enforce_eager,
@@ -318,6 +346,9 @@ def main():
         enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler,
         profiler_config=args.profiler_config,
     )
+    if args.quantization is not None:
+        omni_kwargs["quantization"] = args.quantization
+    omni = Omni(**omni_kwargs)
 
     if profiler_enabled:
         print("[Profiler] Starting profiling...")
@@ -330,6 +361,9 @@ def main():
     print(f"  Inference steps: {args.num_inference_steps}")
     print(f"  Frames: {args.num_frames}")
     print(f"  Solver: {args.sample_solver}")
+    print(f"  diffusion_kv_cache_dtype(config): {args.diffusion_kv_cache_dtype}")
+    print(f"  diffusion_kv_cache_skip_steps(config): {args.diffusion_kv_cache_skip_steps}")
+    print(f"  diffusion_kv_cache_skip_layers(config): {args.diffusion_kv_cache_skip_layers}")
     print(
         f"  Parallel configuration: cfg_parallel_size={args.cfg_parallel_size},"
         f" tensor_parallel_size={args.tensor_parallel_size}, vae_patch_parallel_size={args.vae_patch_parallel_size}"

examples/offline_inference/internvla_a1/end2end.py

@@ -172,7 +172,7 @@ def run_pipeline_forward(
 
 def _synchronize(device: str) -> None:
     if device.startswith("cuda") and torch.cuda.is_available():
-        torch.cuda.synchronize()
+        torch.accelerator.synchronize()
 
 
 def _latency_summary(values_ms: list[float]) -> dict[str, float]:

examples/offline_inference/sensenova_u1/README.md

@@ -0,0 +1,208 @@
+# SenseNova-U1-8B-MoT
+
+## Architecture
+
+SenseNova-U1 is a unified Qwen3-based LLM with Mixture-of-Tokenizers (MoT) attention. Unlike two-stage pipelines (e.g., BAGEL), it handles text encoding, optional reasoning ("think mode"), and flow-matching-based image denoising entirely within a single diffusion stage.
+
+| Feature | Description |
+| :------ | :---------- |
+| **Model type** | Unified LLM (single-stage diffusion pipeline) |
+| **Base LLM** | Qwen3 with MoT attention and 3D RoPE |
+| **Image generation** | Flow-matching Euler sampler, no separate VAE |
+| **Think mode** | Optional chain-of-thought reasoning before image generation |
+| **Parallelism** | Tensor Parallelism (TP) with fused QKV and fused gate/up projections |
+| **Modalities** | text2img, img2img, img2text (understanding), text2text |
+
+## Quick Start
+
+```bash
+cd examples/offline_inference/sensenova_u1
+
+# Text-to-image
+python end2end.py --prompt "A cute cat" --think
+
+# Image-to-image editing
+python end2end.py --prompt "Turn this into an oil painting" \
+                  --image input.png --think
+
+# Image understanding (img2text)
+python end2end.py --modality img2text \
+                  --prompt "Describe this image in detail" \
+                  --image photo.jpg
+
+# Text-to-text (pure chat)
+python end2end.py --modality text2text \
+                  --prompt "What is the capital of France?"
+
+# Custom resolution (image generation)
+python end2end.py --prompt "A futuristic cityscape at sunset" \
+                  --width 2048 --height 1024 --think
+```
+
+> **Note**: Default configuration works on a single **NVIDIA A100 (80GB)** or **H100** GPU.
+
+## Text-to-Image (t2i)
+
+Standard text-to-image with optional think mode:
+
+```bash
+python end2end.py \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, textured skin, gentle smile, warm natural light" \
+    --width 1536 --height 2720 \
+    --think --print-think
+```
+
+## Image-to-Image Editing (img2img)
+
+Pass one or more `--image` paths to trigger img2img mode. The model uses the
+input image(s) as visual context for editing:
+
+```bash
+# Single input image
+python end2end.py \
+    --prompt "Add a sunset sky in the background" \
+    --image photo.jpg \
+    --width 2048 --height 2048 \
+    --think
+
+# Multiple reference images
+python end2end.py \
+    --prompt "Combine the style of Image-1 with the content of Image-2" \
+    --image style_ref.png content_ref.png \
+    --width 2048 --height 2048 \
+    --think
+```
+
+### Dual CFG for img2img
+
+img2img supports dual classifier-free guidance (text CFG + image CFG):
+
+```bash
+python end2end.py \
+    --prompt "Make the person smile" \
+    --image portrait.jpg \
+    --cfg-scale 4.0 \
+    --img-cfg-scale 2.0 \
+    --think
+```
+
+| `img_cfg_scale` | Behavior |
+| :-------------- | :------- |
+| `1.0` (default) | Text-only CFG: guidance = image-condition → full-condition |
+| `== cfg_scale` | Standard CFG: guidance = unconditional → full-condition |
+| Other value | Dual CFG: separate text and image guidance strengths |
+
+## Full Parameter Reference
+
+### General Parameters
+
+| Parameter | Default | Description |
+| :-------- | :------ | :---------- |
+| `--modality` | `auto` | Task modality: `auto`, `text2img`, `img2img`, `img2text`, `text2text` |
+| `--prompt` | "A cute cat..." | Text prompt / editing instruction / question |
+| `--image` | None | Input image path(s) for img2img or img2text |
+
+### Image Generation Parameters (text2img / img2img)
+
+| Parameter | Default | Description |
+| :-------- | :------ | :---------- |
+| `--height` | 2048 | Height of generated image (pixels) |
+| `--width` | 2048 | Width of generated image (pixels) |
+| `--seed` | 42 | Random seed for reproducibility |
+| `--num-steps` | 50 | Number of denoising steps |
+| `--cfg-scale` | 4.0 | Text classifier-free guidance scale |
+| `--img-cfg-scale` | 1.0 | Image CFG scale (img2img only, 1.0 = disabled) |
+| `--cfg-norm` | "none" | CFG normalization: none, global, channel, cfg_zero_star (t2i only) |
+| `--timestep-shift` | 3.0 | Timestep shift for flow-matching schedule |
+| `--t-eps` | 0.02 | Epsilon for timestep schedule |
+| `--think` | False | Enable think mode |
+| `--print-think` | False | Print think text to stdout |
+
+### Text Generation Parameters (text2text / img2text)
+
+| Parameter | Default | Description |
+| :-------- | :------ | :---------- |
+| `--max-tokens` | 512 | Maximum number of tokens to generate |
+| `--do-sample` | False | Use sampling instead of greedy decoding |
+| `--temperature` | 0.7 | Sampling temperature (higher = more diverse) |
+
+### Infrastructure Parameters
+
+| Parameter | Default | Description |
+| :-------- | :------ | :---------- |
+| `--model` | `SenseNova/SenseNova-U1-8B-MoT` | HuggingFace model ID or local path |
+| `--output` | `.` | Output directory for saved images |
+| `--tensor-parallel-size` | 1 | Number of GPUs for tensor parallelism |
+| `--enforce-eager` | False | Disable torch.compile |
+| `--enable-cpu-offload` | False | Enable module-wise (sequential) CPU offload to reduce peak VRAM |
+
+## Reducing GPU Memory Usage
+
+For hardware with limited VRAM, enable **module-wise CPU offload** with
+`--enable-cpu-offload`. The pipeline implements `SupportsModuleOffload`, so the
+vision encoder (`vision_model`) and the Qwen3 LLM (`language_model`) are
+swapped between CPU and GPU on demand:
+
+- During text/vision encoding, the LLM is on CPU.
+- During the diffusion loop, the vision encoder is on CPU.
+- Lightweight FM modules stay resident on GPU.
+
+This lowers peak VRAM at the cost of extra CPU<->GPU transfers (use pinned
+memory) — useful for running SenseNova-U1 on consumer-grade GPUs.
+
+```bash
+# Text-to-image with CPU offload
+python end2end.py \
+    --prompt "A cute cat sitting on a windowsill" \
+    --width 2048 --height 2048 \
+    --enable-cpu-offload --think
+
+# Image-to-image editing with CPU offload
+python end2end.py \
+    --prompt "Turn this into an oil painting" \
+    --image input.png \
+    --width 2048 --height 2048 \
+    --enable-cpu-offload --think
+
+# Image understanding (img2text) with CPU offload
+python end2end.py \
+    --modality img2text \
+    --prompt "Describe this image in detail" \
+    --image photo.jpg \
+    --enable-cpu-offload
+```
+
+> **Notes**
+> - CPU offload is **single-GPU only** (incompatible with `--tensor-parallel-size > 1`).
+> - First-step latency is higher because of the cold-start CPU<->GPU transfers.
+> - For more details and other offloading strategies, see
+>   [CPU Offloading for Diffusion Models](../../../docs/user_guide/diffusion/cpu_offload_diffusion.md).
+
+## Reproducing the E2E Test
+
+The following command reproduces the pixel-validated CI test case:
+
+```bash
+python end2end.py \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, textured skin, gentle smile, warm natural light, emotional documentary look. The portrait should feel polished and natural, with sharp eyes, realistic skin texture, accurate facial anatomy, and premium lighting that keeps the face as the main focus." \
+    --width 1536 --height 2720 \
+    --seed 42 --num-steps 50 \
+    --cfg-scale 4.0 --timestep-shift 3.0 --cfg-norm none \
+    --think --print-think \
+    --output outputs
+```
+
+The corresponding pytest:
+
+```bash
+pytest -s -v tests/e2e/offline_inference/test_sensenova_u1_text2img.py \
+    -m "advanced_model" --run-level "advanced_model"
+```
+
+The img2img counterpart lives at
+`tests/e2e/offline_inference/test_sensenova_u1_img2img.py`.
+
+## Online Serving
+
+For OpenAI-compatible API serving, see
+[`examples/online_serving/sensenova_u1/`](../../examples/online_serving/sensenova_u1/).

examples/offline_inference/sensenova_u1/end2end.py

@@ -0,0 +1,281 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+"""
+End-to-end offline inference for SenseNova-U1-8B-MoT.
+
+Supports four modalities: text2img, img2img, img2text, text2text.
+
+Text-to-image:
+    python end2end.py --prompt "A cute cat" --think
+
+Image-to-image (editing):
+    python end2end.py --prompt "Turn this into an oil painting" \
+        --image input.png --think
+
+Image understanding:
+    python end2end.py --modality img2text \
+        --prompt "Describe this image" --image photo.jpg
+
+Text chat:
+    python end2end.py --modality text2text \
+        --prompt "What is the capital of France?"
+
+See README.md for more examples.
+"""
+
+import argparse
+import os
+
+from PIL import Image
+
+from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="SenseNova-U1 text-to-image / image-to-image / text / understanding via vLLM-Omni.",
+    )
+    parser.add_argument(
+        "--model",
+        default="SenseNova/SenseNova-U1-8B-MoT",
+        help="HuggingFace model ID or local path.",
+    )
+    parser.add_argument(
+        "--modality",
+        default="auto",
+        choices=["auto", "text2img", "img2img", "img2text", "text2text"],
+        help="Task modality. 'auto' infers from --image (img2img/img2text if images, else t2i/t2t).",
+    )
+    parser.add_argument(
+        "--prompt",
+        default="A cute cat sitting on a windowsill, soft natural light",
+        help="Text prompt for generation or editing instruction.",
+    )
+    parser.add_argument(
+        "--image",
+        nargs="+",
+        metavar="PATH",
+        default=None,
+        help="Input image path(s) for img2img or img2text.",
+    )
+    parser.add_argument(
+        "--output",
+        type=str,
+        default=".",
+        help="Output directory to save generated images.",
+    )
+    # Image dimensions
+    parser.add_argument("--height", type=int, default=2048, help="Height of generated image.")
+    parser.add_argument("--width", type=int, default=2048, help="Width of generated image.")
+
+    # Generation parameters
+    parser.add_argument("--seed", type=int, default=42, help="Random seed for deterministic results.")
+    parser.add_argument("--num-steps", type=int, default=50, help="Number of denoising steps.")
+    parser.add_argument("--cfg-scale", type=float, default=4.0, help="Text classifier-free guidance scale.")
+    parser.add_argument(
+        "--img-cfg-scale",
+        type=float,
+        default=1.0,
+        help="Image CFG scale for img2img (1.0 = disabled). "
+        "When both --cfg-scale and --img-cfg-scale differ, dual CFG is used.",
+    )
+    parser.add_argument(
+        "--cfg-norm",
+        type=str,
+        default="none",
+        choices=["none", "global", "channel", "cfg_zero_star"],
+        help="CFG normalization mode. cfg_zero_star is only for t2i.",
+    )
+    parser.add_argument(
+        "--timestep-shift",
+        type=float,
+        default=3.0,
+        help="Timestep shift for flow-matching schedule.",
+    )
+    parser.add_argument(
+        "--t-eps",
+        type=float,
+        default=0.02,
+        help="Epsilon for flow-matching timestep schedule.",
+    )
+
+    # Text generation parameters (for text2text / img2text)
+    parser.add_argument(
+        "--max-tokens",
+        type=int,
+        default=2048,
+        help="Max tokens for text generation (text2text / img2text).",
+    )
+    parser.add_argument(
+        "--do-sample",
+        action="store_true",
+        help="Use sampling instead of greedy decoding for text generation.",
+    )
+    parser.add_argument(
+        "--temperature",
+        type=float,
+        default=0.7,
+        help="Sampling temperature for text generation.",
+    )
+
+    # Think mode
+    parser.add_argument(
+        "--think",
+        action="store_true",
+        help="Enable think mode: the model reasons about the prompt before generating the image.",
+    )
+    parser.add_argument(
+        "--print-think",
+        action="store_true",
+        help="Print the generated think text to stdout.",
+    )
+
+    # Advanced
+    parser.add_argument(
+        "--tensor-parallel-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for tensor parallelism.",
+    )
+    parser.add_argument(
+        "--enforce-eager",
+        action="store_true",
+        help="Disable torch.compile and force eager execution.",
+    )
+    parser.add_argument(
+        "--enable-cpu-offload",
+        action="store_true",
+        help=(
+            "Enable model-level (sequential) CPU offloading. "
+            "Mutually exclusive between vision_model and language_model: "
+            "only one is on GPU at a time, reducing peak VRAM at the cost "
+            "of extra CPU<->GPU transfers."
+        ),
+    )
+
+    nullify_stage_engine_defaults(parser)
+    return parser.parse_args()
+
+
+def _resolve_modality(args):
+    """Determine the effective modality from CLI args."""
+    if args.modality != "auto":
+        return args.modality
+    has_images = args.image is not None
+    # Default: image-producing tasks for backward compatibility
+    return "img2img" if has_images else "text2img"
+
+
+def main():
+    args = parse_args()
+    os.makedirs(args.output, exist_ok=True)
+
+    modality = _resolve_modality(args)
+    is_text_output = modality in ("text2text", "img2text")
+    has_images = modality in ("img2img", "img2text")
+
+    omni = Omni(
+        model=args.model,
+        tensor_parallel_size=args.tensor_parallel_size,
+        enforce_eager=args.enforce_eager,
+        enable_cpu_offload=args.enable_cpu_offload,
+    )
+
+    extra_args = {
+        "cfg_scale": args.cfg_scale,
+        "cfg_norm": args.cfg_norm,
+        "timestep_shift": args.timestep_shift,
+        "cfg_interval": (0.0, 1.0),
+        "batch_size": 1,
+        "think": args.think,
+        "t_eps": args.t_eps,
+    }
+    if modality == "img2img":
+        extra_args["img_cfg_scale"] = args.img_cfg_scale
+    if is_text_output:
+        extra_args["max_tokens"] = args.max_tokens
+        extra_args["do_sample"] = args.do_sample
+        extra_args["temperature"] = args.temperature
+
+    sampling_params = OmniDiffusionSamplingParams(
+        height=args.height,
+        width=args.width,
+        seed=args.seed,
+        num_inference_steps=args.num_steps,
+        extra_args=extra_args,
+    )
+
+    print(f"\n{'=' * 60}")
+    print(f"SenseNova-U1 Configuration ({modality}):")
+    print(f"  Model          : {args.model}")
+    if not is_text_output:
+        print(f"  Image size     : {args.width}x{args.height}")
+        print(f"  Steps          : {args.num_steps}")
+        print(f"  CFG scale      : {args.cfg_scale}")
+    if modality == "img2img":
+        print(f"  Img CFG scale  : {args.img_cfg_scale}")
+    if has_images:
+        print(f"  Input images   : {args.image}")
+    if is_text_output:
+        print(f"  Max tokens     : {args.max_tokens}")
+        print(f"  Temperature    : {args.temperature}")
+    print(f"  Seed           : {args.seed}")
+    print(f"  Think mode     : {args.think}")
+    print(f"  TP size        : {args.tensor_parallel_size}")
+    print(f"{'=' * 60}\n")
+
+    # Build prompt dict
+    if has_images:
+        if not args.image:
+            raise ValueError(f"{modality} requires --image.")
+        input_images = [Image.open(p).convert("RGB") for p in args.image]
+        if is_text_output:
+            prompt_dict = {
+                "prompt": args.prompt,
+                "multi_modal_data": {"image": input_images},
+                "modalities": ["text"],
+            }
+        else:
+            prompt_dict = {
+                "prompt": args.prompt,
+                "multi_modal_data": {"image": input_images},
+                "modalities": ["img2img"],
+            }
+    elif is_text_output:
+        prompt_dict = {"prompt": args.prompt, "modalities": ["text"]}
+    else:
+        prompt_dict = {"prompt": args.prompt, "modalities": ["image"]}
+
+    outputs = list(
+        omni.generate(
+            prompts=prompt_dict,
+            sampling_params_list=sampling_params,
+        )
+    )
+
+    for req_output in outputs:
+        custom = getattr(req_output, "_custom_output", {}) or {}
+        if args.print_think and custom.get("think_text"):
+            print(f"[Think]\n{custom['think_text']}\n")
+
+        if is_text_output:
+            text = custom.get("text_output", "")
+            if not text:
+                text = getattr(req_output, "text", "") or ""
+            print(f"[Response]\n{text}")
+        else:
+            images = getattr(req_output, "images", None) or []
+            if not images:
+                print("[Warning] No images generated.")
+                continue
+            for j, img in enumerate(images):
+                save_path = os.path.join(args.output, f"sensenova_u1_output_{j}.png")
+                img.save(save_path)
+                print(f"[Output] Saved {img.size[0]}x{img.size[1]} image to {save_path}")
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/text_to_speech/README.md

@@ -73,13 +73,15 @@ python examples/offline_inference/text_to_speech/cosyvoice3/end2end.py \
 ```
 
 ### Voice cloning
-Pass a reference audio. Note that CosyVoice3's `--prompt-text` is a system-style prompt for the GPT stage, not a reference transcript:
+If `--ref-audio` is omitted, the script downloads the upstream
+[`zero_shot_prompt.wav`](https://github.com/FunAudioLLM/CosyVoice/blob/main/asset/zero_shot_prompt.wav) from the CosyVoice repo into the current directory.
+To use your own clip, pass `--ref-audio /path/to/reference.wav`, and modify `--prompt-text` correspondingly.
 ```bash
 python examples/offline_inference/text_to_speech/cosyvoice3/end2end.py \
     --model pretrained_models/Fun-CosyVoice3-0.5B \
     --tokenizer pretrained_models/Fun-CosyVoice3-0.5B/CosyVoice-BlankEN \
-    --ref-audio prompt.wav \
-    --prompt-text "You are a helpful assistant.<|endofprompt|>Testing my voices. Why should I not?"
+    --ref-audio /path/to/reference.wav \
+    --prompt-text "You are a helpful assistant.<|endofprompt|>Trascript in your ref audio clip"
 ```
 
 ### Streaming
@@ -317,44 +319,39 @@ pip install voxcpm soundfile
 export VLLM_OMNI_VOXCPM_CODE_PATH=/path/to/VoxCPM/src
 ```
 
-If the native VoxCPM `config.json` does not contain HF metadata such as `model_type`, prepare a persistent HF-compatible config directory and point the stage configs to it via `VLLM_OMNI_VOXCPM_HF_CONFIG_PATH`:
-
-```bash
-export VOXCPM_MODEL=/path/to/voxcpm-model
-export VLLM_OMNI_VOXCPM_HF_CONFIG_PATH=/tmp/voxcpm_hf_config
-mkdir -p "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"
-cp "$VOXCPM_MODEL/config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/config.json"
-cp "$VOXCPM_MODEL/generation_config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/generation_config.json" 2>/dev/null || true
-python3 -c 'import json, os; p=os.path.join(os.environ["VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"], "config.json"); cfg=json.load(open(p, "r", encoding="utf-8")); cfg["model_type"]="voxcpm"; cfg.setdefault("architectures", ["VoxCPMForConditionalGeneration"]); json.dump(cfg, open(p, "w", encoding="utf-8"), indent=2, ensure_ascii=False)'
-```
-
 ### Quick start
 ```bash
 python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
+    --model openbmb/VoxCPM-0.5B \
     --text "This is a split-stage VoxCPM synthesis example running on vLLM Omni."
 ```
 
 ### Voice cloning
 ```bash
 python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
+    --model openbmb/VoxCPM-0.5B \
     --text "This sentence is synthesized with a cloned voice." \
     --ref-audio /path/to/reference.wav \
     --ref-text  "The exact transcript spoken in reference.wav."
 ```
 
 ### Streaming
-Pass the async-chunk stage config:
+Pass `--streaming` together with the legacy async-chunk stage config:
 ```bash
 python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
+    --model openbmb/VoxCPM-0.5B \
+    --streaming \
     --stage-configs-path vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml \
     --text "This is a split-stage VoxCPM streaming example running on vLLM Omni."
 ```
 
+### Persistent HF config (optional)
+The engine auto-patches a missing `model_type` in the checkpoint's `config.json` in-memory at startup (see `vllm_omni/engine/arg_utils.py`), so no manual setup is required.
+
+If you'd rather maintain a persistent patched config dir (e.g. to share across runs or to add fields the auto-patcher doesn't set), export `VLLM_OMNI_VOXCPM_HF_CONFIG_PATH` to its location — both the bundled deploy and async-chunk yamls read it via `${oc.env:VLLM_OMNI_VOXCPM_HF_CONFIG_PATH,}`.
+
 ### Notes
-- `voxcpm.yaml` is the default non-streaming stage config; `voxcpm_async_chunk.yaml` enables streaming.
+- Non-streaming auto-loads `vllm_omni/deploy/voxcpm.yaml` from the model directory name / HF `model_type`. Streaming uses the legacy `voxcpm_async_chunk.yaml` passed via `--stage-configs-path`.
 - Streaming is currently single-request oriented.
 - `--ref-text` must be the real transcript of `--ref-audio`; mismatched text degrades quality.
 - For online serving, see the [VoxCPM section in the online hub](../../online_serving/text_to_speech/README.md#voxcpm). For benchmark reporting, see [`benchmarks/voxcpm`](../../../benchmarks/voxcpm/README.md).

examples/offline_inference/text_to_speech/cosyvoice3/end2end.py

@@ -2,18 +2,32 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import argparse
 import os
+import urllib.request
+from pathlib import Path
 
 import numpy as np
 import soundfile as sf
 from vllm import SamplingParams
-from vllm.assets.audio import AudioAsset
 from vllm.multimodal.media.audio import load_audio
 
 from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 from vllm_omni.entrypoints.omni import Omni
-from vllm_omni.model_executor.models.cosyvoice3.config import CosyVoice3Config
 from vllm_omni.model_executor.models.cosyvoice3.tokenizer import get_qwen_tokenizer
 from vllm_omni.model_executor.models.cosyvoice3.utils import extract_text_token
+from vllm_omni.transformers_utils.configs.cosyvoice3 import CosyVoice3Config
+
+# Upstream zero-shot reference clip
+ZERO_SHOT_PROMPT_URL = "https://raw.githubusercontent.com/FunAudioLLM/CosyVoice/main/asset/zero_shot_prompt.wav"
+
+
+def _default_ref_audio() -> str:
+    # Download the upstream zero_shot_prompt.wav into the current dir
+    dest = Path("zero_shot_prompt.wav")
+    if not dest.exists() or dest.stat().st_size == 0:
+        print(f"Downloading default reference audio to {dest}")
+        urllib.request.urlretrieve(ZERO_SHOT_PROMPT_URL, dest)
+
+    return str(dest)
 
 
 def run_e2e():
@@ -36,9 +50,15 @@ def run_e2e():
     parser.add_argument(
         "--prompt-text",
         type=str,
-        default="You are a helpful assistant.<|endofprompt|>Testing my voices. Why should I not?",
+        default="You are a helpful assistant.<|endofprompt|>希望你以后，能够做的比我还好呦!",
+    )
+    parser.add_argument(
+        "--ref-audio",
+        type=str,
+        default=None,
+        help="Path to reference audio for voice cloning. "
+        "If unset, downloads the upstream CosyVoice3 zero-shot prompt audio clip",
     )
-    parser.add_argument("--ref-audio", type=str, default="prompt.wav")
     parser.add_argument(
         "--tokenizer",
         type=str,
@@ -66,24 +86,22 @@ def run_e2e():
     sampling_cfg = {"top_p": 0.8, "top_k": 25, "eos_token_id": 6561 + 1}
 
     print("Model initialized. Preparing inputs...")
-    if args.ref_audio:
-        if not os.path.exists(args.ref_audio):
-            raise FileNotFoundError(f"Audio file not found: {args.ref_audio}")
-        # Load at native sample rate
-        audio_signal, sr = load_audio(args.ref_audio, sr=None)
+    ref_audio_path = args.ref_audio or _default_ref_audio()
+    if not os.path.exists(ref_audio_path):
+        raise FileNotFoundError(f"Audio file not found: {ref_audio_path}")
+    # Load at native sample rate
+    audio_signal, sr = load_audio(ref_audio_path, sr=None)
 
-        # Validate sample rate before processing (similar to original CosyVoice)
-        min_sr = 16000
-        if sr < min_sr:
-            raise ValueError(
-                f"Audio sample rate {sr} Hz is too low. "
-                f"Minimum required: {min_sr} Hz. "
-                f"Please provide audio with sample rate >= {min_sr} Hz."
-            )
+    # Validate sample rate before processing (similar to original CosyVoice)
+    min_sr = 16000
+    if sr < min_sr:
+        raise ValueError(
+            f"Audio sample rate {sr} Hz is too low. "
+            f"Minimum required: {min_sr} Hz. "
+            f"Please provide audio with sample rate >= {min_sr} Hz."
+        )
 
-        audio_data = (audio_signal.astype(np.float32), sr)
-    else:
-        audio_data = AudioAsset("mary_had_lamb").audio_and_sample_rate
+    audio_data = (audio_signal.astype(np.float32), sr)
 
     prompts = {
         "prompt": args.text,

examples/offline_inference/text_to_speech/omnivoice/end2end.py

@@ -37,7 +37,7 @@ def run_e2e():
     parser.add_argument(
         "--stage-config",
         type=str,
-        default="vllm_omni/model_executor/stage_configs/omnivoice.yaml",
+        default="vllm_omni/deploy/omnivoice.yaml",
     )
     parser.add_argument(
         "--text",

examples/offline_inference/text_to_speech/voxcpm/end2end.py

@@ -14,9 +14,6 @@ from vllm.utils.argparse_utils import FlexibleArgumentParser
 from vllm_omni import AsyncOmni, Omni
 from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 
-REPO_ROOT = Path(__file__).resolve().parents[4]
-DEFAULT_SYNC_STAGE_CONFIG = REPO_ROOT / "vllm_omni" / "model_executor" / "stage_configs" / "voxcpm.yaml"
-
 
 def _build_prompt(args) -> dict[str, Any]:
     additional_information: dict[str, list[Any]] = {
@@ -59,10 +56,6 @@ def _extract_sample_rate(mm: dict[str, Any]) -> int:
     return int(sr_raw)
 
 
-def _is_streaming_stage_config(stage_config_path: str) -> bool:
-    return "async_chunk" in Path(stage_config_path).stem
-
-
 def _save_audio(audio: torch.Tensor, sample_rate: int, output_dir: Path, request_id: str) -> Path:
     output_dir.mkdir(parents=True, exist_ok=True)
     output_path = output_dir / f"output_{request_id}.wav"
@@ -87,6 +80,7 @@ async def _run_streaming(args) -> Path:
     omni = AsyncOmni(
         model=args.model,
         stage_configs_path=args.stage_configs_path,
+        deploy_config=args.deploy_config,
         log_stats=args.log_stats,
         stage_init_timeout=args.stage_init_timeout,
     )
@@ -133,6 +127,7 @@ def _run_sync(args) -> Path:
     omni = Omni(
         model=args.model,
         stage_configs_path=args.stage_configs_path,
+        deploy_config=args.deploy_config,
         log_stats=args.log_stats,
         stage_init_timeout=args.stage_init_timeout,
     )
@@ -164,11 +159,27 @@ def _run_sync(args) -> Path:
 def parse_args():
     parser = FlexibleArgumentParser(description="Minimal offline VoxCPM example for vLLM Omni.")
     parser.add_argument("--model", type=str, required=True, help="Local VoxCPM model directory.")
+    parser.add_argument(
+        "--deploy-config",
+        type=str,
+        default=None,
+        help=("Override the deploy config path."),
+    )
     parser.add_argument(
         "--stage-configs-path",
         type=str,
-        default=str(DEFAULT_SYNC_STAGE_CONFIG),
-        help=("Stage config path. Use voxcpm.yaml for non-streaming or voxcpm_async_chunk.yaml for streaming."),
+        default=None,
+        help=(
+            "Legacy stage_args yaml path. Required for streaming "
+            "(vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml); "
+            "leave unset for non-streaming to use the auto-resolved deploy config."
+        ),
+    )
+    parser.add_argument(
+        "--streaming",
+        action="store_true",
+        default=False,
+        help="Stream audio chunks as they arrive via AsyncOmni.",
     )
     parser.add_argument("--text", type=str, required=True, help="Input text for synthesis.")
     parser.add_argument("--ref-audio", type=str, default=None, help="Reference audio path for voice cloning.")
@@ -194,9 +205,10 @@ def parse_args():
 
 
 def main(args) -> None:
-    route = "streaming" if _is_streaming_stage_config(args.stage_configs_path) else "sync"
+    route = "streaming" if args.streaming else "sync"
     print(f"Model: {args.model}")
-    print(f"Stage config: {args.stage_configs_path}")
+    print(f"Deploy config: {args.deploy_config or '<auto from HF model_type>'}")
+    print(f"Stage configs path: {args.stage_configs_path or '<unused>'}")
     print(f"Route: {route}")
     if route == "streaming":
         asyncio.run(_run_streaming(args))

examples/offline_inference/text_to_speech/voxcpm2/end2end.py

@@ -18,7 +18,6 @@ from vllm_omni import Omni
 from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 
 REPO_ROOT = Path(__file__).resolve().parents[4]
-DEFAULT_STAGE_CONFIGS_PATH = str(REPO_ROOT / "vllm_omni" / "model_executor" / "stage_configs" / "voxcpm2.yaml")
 SAMPLE_RATE = 48_000
 
 
@@ -43,10 +42,11 @@ def parse_args():
         help="Directory for output WAV files.",
     )
     parser.add_argument(
-        "--stage-configs-path",
+        "--deploy-config",
         type=str,
-        default=DEFAULT_STAGE_CONFIGS_PATH,
-        help="Path to the stage config YAML file.",
+        default=None,
+        help="Override the deploy config path. If unset, auto-loads "
+        "vllm_omni/deploy/voxcpm2.yaml based on the HF model_type.",
     )
     parser.add_argument(
         "--ref-audio",
@@ -96,7 +96,7 @@ def main():
 
     engine = Omni(
         model=args.model,
-        stage_configs_path=args.stage_configs_path,
+        deploy_config=args.deploy_config,
     )
 
     from transformers import AutoTokenizer

examples/offline_inference/text_to_speech/voxtral_tts/end2end.py

@@ -42,6 +42,7 @@ async def run_streaming(inputs, sampling_params_list, model_name, args, output_d
         model=model_name,
         deploy_config=args.deploy_config,
         log_stats=args.log_stats,
+        quantization=args.quantization,
     )
 
     # Normalize to a list so batch and single-input share the same code path
@@ -184,7 +185,7 @@ async def run_streaming(inputs, sampling_params_list, model_name, args, output_d
     )
 
     async_omni.shutdown()
-    torch.cuda.empty_cache()
+    torch.accelerator.empty_cache()
     gc.collect()
 
 
@@ -194,6 +195,7 @@ def run_non_streaming(inputs, sampling_params_list, model_name, args, output_dir
         model=model_name,
         log_stats=args.log_stats,
         deploy_config=args.deploy_config,
+        quantization=args.quantization,
     )
 
     if args.profiling_mode:
@@ -223,7 +225,7 @@ def run_non_streaming(inputs, sampling_params_list, model_name, args, output_dir
     print(f"RTF: {output_audio_dur / vllm_elapsed:.4f}")
 
     del llm
-    torch.cuda.empty_cache()
+    torch.accelerator.empty_cache()
     gc.collect()
 
 
@@ -305,6 +307,12 @@ def parse_args() -> Namespace:
         default=None,
         help="CFG alpha for flow-matching guidance (default: use value from stage config, typically 1.2).",
     )
+    parser.add_argument(
+        "--quantization",
+        type=str,
+        default=None,
+        help="Quantization method (e.g. 'fp8'). Applied to the language model only.",
+    )
     nullify_stage_engine_defaults(parser)
     return parser.parse_args()
 
@@ -379,7 +387,7 @@ def main(args: Any) -> None:
             f"--num-prompts ({args.num_prompts}) must be divisible by --concurrency ({args.concurrency})"
         )
 
-    torch.cuda.empty_cache()
+    torch.accelerator.empty_cache()
     gc.collect()
 
     if args.streaming:

examples/offline_inference/text_to_video/text_to_video.py

@@ -206,8 +206,8 @@ def parse_args() -> argparse.Namespace:
         "--quantization",
         type=str,
         default=None,
-        choices=["fp8", "gguf"],
-        help="Quantization method for the transformer (fp8 for online FP8 quantization).",
+        choices=["fp8", "mxfp8", "int8", "gguf"],
+        help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 online quant (NPU). fp8: online FP8 (GPU).",
     )
     parser.add_argument(
         "--use-hsdp",

examples/online_serving/bagel/README.md

@@ -68,6 +68,25 @@ vllm serve ByteDance-Seed/BAGEL-7B-MoT --omni --port 8091 --tensor-parallel-size
 
 Or set `tensor_parallel_size` per stage in a custom deploy YAML.
 
+#### Hybrid Sharded Data Parallel (HSDP)
+
+For larger Bagel deployments on multiple GPUs, you can enable HSDP (Hybrid Sharded Data Parallel) by modifying the stage configuration (for example, [`bagel.yaml`](../../../vllm_omni/deploy/bagel.yaml)). HSDP shards transformer weights across GPUs to reduce per-GPU memory usage.
+
+1. **Enable HSDP**: Set `use_hsdp: true`.
+2. **Set shard size**: Set `hsdp_shard_size` to the number of GPUs used for sharding (for example, `4`).
+3. **Set replicate size**: Usually keep `hsdp_replicate_size: 1` unless you want replicated HSDP groups.
+4. **Set devices**: Specify the comma-separated GPU IDs used by the diffusion stage (for example, `"0,1,2,3"`).
+
+Example configuration for HSDP across 4 GPUs:
+```yaml
+  - stage_id: 1
+    devices: "0,1,2,3"
+    parallel_config:
+      use_hsdp: true
+      hsdp_shard_size: 4
+      hsdp_replicate_size: 1
+```
+
 ### Multi-Node Deployment
 
 Deploy each stage on a **separate node** for better resource utilization. Replace `<ORCHESTRATOR_IP>` with the actual IP address of your orchestrator node.

examples/online_serving/covo_audio/README.md

@@ -0,0 +1,93 @@
+# Covo-Audio-Chat
+
+## Installation
+
+Please refer to [README.md](../../../README.md)
+
+> **Note**
+> Covo-Audio code2wav requires `torchdiffeq`. Install it with: `pip install torchdiffeq`
+
+## Run examples (Covo-Audio-Chat)
+
+### Launch the Server
+
+```bash
+CUDA_VISIBLE_DEVICES=0 vllm serve tencent/Covo-Audio-Chat --omni \
+    --trust-remote-code --port 18091
+```
+
+> The default deploy config is located at `vllm_omni/deploy/covo_audio.yaml` and is loaded automatically by the model registry — no `--deploy-config` flag needed for default use.
+
+### Send Request
+
+Get into the example folder:
+```bash
+cd examples/online_serving/covo_audio
+```
+
+#### Audio input chat
+
+Using the default audio asset:
+```bash
+python openai_chat_completion_client.py
+```
+
+Using a custom audio file:
+```bash
+python openai_chat_completion_client.py --audio-path /path/to/audio.wav
+```
+
+#### Streaming mode
+
+```bash
+python openai_chat_completion_client.py --audio-path /path/to/audio.wav --stream
+```
+
+### Command-line Arguments
+
+The Python client supports the following arguments:
+
+- `--prompt` (or `-p`): Text prompt (default: `"请回答这段音频里的问题。"`)
+- `--audio-path`: Path to audio file. Supports local file paths or HTTP/HTTPS URLs. Common audio formats: WAV, MP3, OGG, FLAC, M4A. Uses the default audio asset if omitted
+- `--output-audio-path` (or `-o`): Output path for generated audio (default: `./audio_0.wav`)
+- `--model`: Model name for the API. Auto-detected from server if omitted
+- `--stream`: Enable streaming mode
+
+## Pipeline
+
+Covo-Audio-Chat uses a 2-stage pipeline:
+
+- **Stage 0 (fused_thinker_talker):** The 7B LLM generates interleaved text and audio tokens in a single autoregressive pass.
+- **Stage 1 (code2wav):** A BigVGAN-based vocoder converts the extracted audio codes into a 24kHz WAV waveform.
+
+## Environment
+
+- GPU: 1x H100 (80 GiB)
+- Stage 0 (7B LLM): ~16 GiB VRAM
+- Stage 1 (BigVGAN vocoder): ~2 GiB VRAM
+
+## Important: System Prompt
+
+Covo-Audio-Chat requires a specific system prompt to enable interleaved text+audio generation. Without it, Stage 0 produces text-only tokens, and the output audio will be **silent**.
+
+The system prompt must include the instruction:
+
+```
+请用文本和音频进行对话，交替生成5个文本token和15个音频token，音频部分使用发音人：default_female
+```
+
+The full system prompt is defined in `vllm_omni/model_executor/models/covo_audio/prompt_utils.py` (`COVO_AUDIO_SYSTEM_PROMPT`).
+
+## FAQ
+
+If you encounter `ModuleNotFoundError: No module named 'librosa'`, install it with:
+```bash
+pip install librosa
+```
+
+### Audio output is silent (0.1s, all zeros)
+
+This means Stage 0 did not generate any audio tokens. Most likely causes:
+
+1. **Missing system prompt** -- see the "Important: System Prompt" section above. The model needs the interleaved generation instruction in the system message.
+2. **Missing `ignore_eos: true` / `stop_token_ids: [151645]`** in Stage 0 sampling params -- without these, the model stops at `<|endoftext|>` before generating audio tokens.

examples/online_serving/covo_audio/openai_chat_completion_client.py

@@ -0,0 +1,216 @@
+"""
+Client example for Covo-Audio-Chat with vllm-omni.
+
+Usage
+-----
+# Start the server first:
+#   CUDA_VISIBLE_DEVICES=0 vllm serve tencent/Covo-Audio-Chat --omni --trust-remote-code
+
+# Audio input chat (uses default audio asset if --audio-path not provided):
+python openai_chat_completion_client.py
+python openai_chat_completion_client.py --audio-path /path/to/audio.wav
+
+# Streaming mode:
+python openai_chat_completion_client.py --stream
+"""
+
+import base64
+import os
+import sys
+import time
+
+from openai import OpenAI
+from vllm.assets.audio import AudioAsset
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+from vllm_omni.model_executor.models.covo_audio.prompt_utils import (
+    COVO_AUDIO_SYSTEM_PROMPT,
+)
+
+openai_api_key = "EMPTY"
+openai_api_base = "http://localhost:18091/v1"
+
+client = OpenAI(api_key=openai_api_key, base_url=openai_api_base)
+
+SEED = 42
+
+
+def encode_base64_content_from_file(file_path: str) -> str:
+    """Encode a local file to base64 format."""
+    with open(file_path, "rb") as f:
+        return base64.b64encode(f.read()).decode("utf-8")
+
+
+def get_audio_url(audio_path: str | None) -> str:
+    """Convert an audio path to a data URL for the API."""
+    if not audio_path:
+        return AudioAsset("mary_had_lamb").url
+
+    if audio_path.startswith(("data:", "http://", "https://")):
+        return audio_path
+
+    if not os.path.exists(audio_path):
+        raise FileNotFoundError(f"Audio file not found: {audio_path}")
+
+    ext = audio_path.rsplit(".", 1)[-1].lower() if "." in audio_path else "wav"
+    mime_map = {
+        "wav": "audio/wav",
+        "mp3": "audio/mpeg",
+        "mpeg": "audio/mpeg",
+        "ogg": "audio/ogg",
+        "flac": "audio/flac",
+        "m4a": "audio/mp4",
+    }
+    mime = mime_map.get(ext, "audio/wav")
+    b64 = encode_base64_content_from_file(audio_path)
+    return f"data:{mime};base64,{b64}"
+
+
+def build_audio_messages(audio_path: str | None, prompt: str) -> list[dict]:
+    """Build messages for an audio chat request."""
+    audio_url = get_audio_url(audio_path)
+    return [
+        {"role": "system", "content": COVO_AUDIO_SYSTEM_PROMPT},
+        {
+            "role": "user",
+            "content": [
+                {"type": "audio_url", "audio_url": {"url": audio_url}},
+                {"type": "text", "text": prompt},
+            ],
+        },
+    ]
+
+
+def get_model_name(override: str | None = None) -> str:
+    """Return the model name to use in API calls."""
+    if override:
+        return override
+    # Auto-detect from the server
+    models = client.models.list()
+    return models.data[0].id
+
+
+def run(args) -> None:
+    model_name = get_model_name(args.model)
+    thinker_params = {
+        "temperature": 0.0,
+        "top_p": 1.0,
+        "top_k": -1,
+        "max_tokens": 2048,
+        "seed": SEED,
+        "detokenize": True,
+        "repetition_penalty": 1.05,
+        "stop_token_ids": [151645],
+        "ignore_eos": True,
+    }
+    code2wav_params = {
+        "temperature": 0.0,
+        "top_p": 1.0,
+        "top_k": -1,
+        "max_tokens": 2048,
+        "seed": SEED,
+        "detokenize": False,
+        "repetition_penalty": 1.05,
+    }
+
+    extra_body = {
+        "sampling_params_list": [thinker_params, code2wav_params],
+    }
+
+    messages = build_audio_messages(args.audio_path, args.prompt)
+
+    start = time.perf_counter()
+
+    if not args.stream:
+        completion = client.chat.completions.create(
+            messages=messages,
+            model=model_name,
+            extra_body=extra_body,
+            stream=False,
+        )
+        elapsed = time.perf_counter() - start
+
+        for choice in completion.choices:
+            if choice.message.audio:
+                audio_bytes = base64.b64decode(choice.message.audio.data)
+                out_path = args.output_audio_path
+                os.makedirs(os.path.dirname(out_path) or ".", exist_ok=True)
+                with open(out_path, "wb") as f:
+                    f.write(audio_bytes)
+                print(f"Audio saved to {out_path} ({len(audio_bytes)} bytes)")
+            elif choice.message.content:
+                print("Chat completion output from text:", choice.message.content)
+
+        print(f"Elapsed: {elapsed:.2f}s", file=sys.stderr)
+
+    else:
+        stream = client.chat.completions.create(
+            messages=messages,
+            model=model_name,
+            extra_body=extra_body,
+            stream=True,
+        )
+        audio_idx = 0
+        for chunk in stream:
+            modality = getattr(chunk, "modality", None)
+            elapsed = time.perf_counter() - start
+            for choice in chunk.choices:
+                content = getattr(choice.delta, "content", None) if hasattr(choice, "delta") else None
+                if not content:
+                    continue
+                if modality == "audio":
+                    audio_bytes = base64.b64decode(content)
+                    out_dir = os.path.dirname(args.output_audio_path) or "."
+                    out_path = os.path.join(out_dir, f"audio_{audio_idx}.wav")
+                    os.makedirs(out_dir, exist_ok=True)
+                    with open(out_path, "wb") as f:
+                        f.write(audio_bytes)
+                    print(f"[{elapsed:.2f}s] Audio chunk saved to {out_path}", file=sys.stderr)
+                    audio_idx += 1
+                elif modality == "text":
+                    print(content, end="", flush=True)
+
+        print()  # trailing newline
+        elapsed = time.perf_counter() - start
+        print(f"Elapsed: {elapsed:.2f}s", file=sys.stderr)
+
+
+def parse_args():
+    parser = FlexibleArgumentParser(description="Covo-Audio-Chat client example")
+    parser.add_argument(
+        "--prompt",
+        "-p",
+        type=str,
+        default="请回答这段音频里的问题。",
+        help="Text prompt.",
+    )
+    parser.add_argument(
+        "--audio-path",
+        type=str,
+        default=None,
+        help="Path to audio file. Uses default audio asset if omitted.",
+    )
+    parser.add_argument(
+        "--output-audio-path",
+        "-o",
+        type=str,
+        default="./audio_0.wav",
+        help="Output path for generated audio.",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default=None,
+        help="Model name for the API. Auto-detected from server if omitted.",
+    )
+    parser.add_argument(
+        "--stream",
+        action="store_true",
+        help="Enable streaming mode.",
+    )
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    run(args)

examples/online_serving/sensenova_u1/README.md

@@ -0,0 +1,163 @@
+# SenseNova-U1 Online Serving
+
+## Launch the Server
+
+```bash
+vllm serve SenseNova/SenseNova-U1-8B-MoT --omni --port 8091
+```
+
+Or use the convenience script:
+
+```bash
+cd examples/online_serving/sensenova_u1
+bash run_server.sh
+```
+
+### Tensor Parallelism (TP)
+
+```bash
+vllm serve SenseNova/SenseNova-U1-8B-MoT --omni --port 8091 \
+    --tensor-parallel-size 2
+```
+
+## Send Requests
+
+```bash
+cd examples/online_serving/sensenova_u1
+```
+
+### Text to Image (text2img)
+
+**Python client:**
+
+```bash
+python openai_chat_client.py \
+    --prompt "A beautiful sunset over mountains" \
+    --modality text2img \
+    --height 2048 --width 2048 --num-steps 50
+```
+
+**curl:**
+
+```bash
+curl http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [{"role": "user", "content": [{"type": "text", "text": "A beautiful sunset over mountains"}]}],
+    "modalities": ["image"],
+    "height": 2048,
+    "width": 2048,
+    "num_inference_steps": 50,
+    "seed": 42
+  }'
+```
+
+### Image to Image (img2img)
+
+**Python client:**
+
+```bash
+python openai_chat_client.py \
+    --prompt "Turn this into an oil painting" \
+    --modality img2img \
+    --image-url /path/to/input.jpg \
+    --height 2048 --width 2048
+```
+
+**curl:**
+
+```bash
+IMAGE_BASE64=$(base64 -w 0 input.jpg)
+
+cat <<EOF > payload.json
+{
+    "messages": [{
+      "role": "user",
+      "content": [
+        {"type": "text", "text": "Turn this into an oil painting"},
+        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,${IMAGE_BASE64}"}}
+      ]
+    }],
+    "modalities": ["image"],
+    "height": 2048,
+    "width": 2048,
+    "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)
+
+**Python client:**
+
+```bash
+python openai_chat_client.py \
+    --prompt "Describe this image in detail" \
+    --modality img2text \
+    --image-url /path/to/image.jpg
+```
+
+**curl:**
+
+```bash
+IMAGE_BASE64=$(base64 -w 0 image.jpg)
+
+cat <<EOF > payload.json
+{
+  "messages": [{
+    "role": "user",
+    "content": [
+      {"type": "text", "text": "Describe this image in detail"},
+      {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,${IMAGE_BASE64}"}}
+    ]
+  }],
+  "modalities": ["text"]
+}
+EOF
+
+curl http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d @payload.json
+```
+
+### Text to Text (text2text)
+
+**Python client:**
+
+```bash
+python openai_chat_client.py \
+    --prompt "What is the capital of France?" \
+    --modality text2text
+```
+
+**curl:**
+
+```bash
+curl http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [{"role": "user", "content": [{"type": "text", "text": "What is the capital of France?"}]}],
+    "modalities": ["text"]
+  }'
+```
+
+## Python Client Arguments
+
+| Argument | Default | Description |
+| :------- | :------ | :---------- |
+| `--prompt` / `-p` | `A cute cat` | Text prompt |
+| `--output` / `-o` | `sensenova_u1_output.png` | Output file path |
+| `--server` / `-s` | `http://localhost:8091` | Server URL |
+| `--image-url` / `-i` | `None` | Input image URL or local path (img2img/img2text) |
+| `--modality` / `-m` | `text2img` | `text2img`, `img2img`, `img2text`, `text2text` |
+| `--height` | `2048` | Image height (image generation only) |
+| `--width` | `2048` | Image width (image generation only) |
+| `--num-steps` | `50` | Number of inference steps (image generation only) |
+| `--seed` | `42` | Random seed |
+| `--cfg-scale` | `4.0` | CFG scale (image generation only) |
+| `--think` | `False` | Enable think mode |

examples/online_serving/sensenova_u1/openai_chat_client.py

@@ -0,0 +1,188 @@
+#!/usr/bin/env python3
+"""
+SenseNova-U1 OpenAI-compatible chat client for all four task modalities.
+
+Usage:
+    # Text-to-image
+    python openai_chat_client.py --prompt "A cute cat" --modality text2img
+
+    # Image-to-image editing
+    python openai_chat_client.py --prompt "Turn this into an oil painting" \
+        --modality img2img --image-url /path/to/image.jpg
+
+    # Image understanding (img2text)
+    python openai_chat_client.py --prompt "Describe this image in detail" \
+        --modality img2text --image-url /path/to/image.jpg
+
+    # Text-to-text (chat)
+    python openai_chat_client.py --prompt "What is the capital of France?" \
+        --modality text2text
+"""
+
+import argparse
+import base64
+from pathlib import Path
+
+import requests
+
+
+def _encode_image(image_url: str) -> str:
+    """Encode a local file or URL to a base64 data URI."""
+    if Path(image_url).exists():
+        with open(image_url, "rb") as f:
+            b64_data = base64.b64encode(f.read()).decode("utf-8")
+        return f"data:image/jpeg;base64,{b64_data}"
+    return image_url
+
+
+def generate(
+    prompt: str,
+    server_url: str = "http://localhost:8091",
+    image_url: str | None = None,
+    modality: str = "text2img",
+    **kwargs: object,
+) -> bytes | str | None:
+    """Send a request to the SenseNova-U1 server.
+
+    All keyword arguments (height, width, seed, cfg_scale, think, …) are
+    forwarded as top-level fields in the request payload.  The serving layer
+    maps standard keys (height, width, seed, num_inference_steps) to sampling
+    params and forwards model-specific keys (think, cfg_norm, img_cfg_scale,
+    …) to ``extra_args`` based on the pipeline's ``EXTRA_BODY_PARAMS``.
+
+    Returns:
+        Image bytes (for image outputs) or text string (for text outputs).
+    """
+    content = [{"type": "text", "text": prompt}]
+
+    if image_url:
+        content.append({"type": "image_url", "image_url": {"url": _encode_image(image_url)}})
+
+    messages = [{"role": "user", "content": content}]
+    payload: dict = {"messages": messages}
+
+    if modality in ("text2img", "img2img"):
+        payload["modalities"] = ["image"]
+    else:
+        payload["modalities"] = ["text"]
+
+    for key, val in kwargs.items():
+        if val is not None and val is not False:
+            payload[key] = val
+
+    try:
+        print(f"Sending {modality} request to {server_url}...")
+        response = requests.post(
+            f"{server_url}/v1/chat/completions",
+            headers={"Content-Type": "application/json"},
+            json=payload,
+            timeout=600,
+        )
+        response.raise_for_status()
+        data = response.json()
+
+        # Print think text if present in response metrics
+        metrics = data.get("metrics") or {}
+        think_text = metrics.get("think_text")
+        if think_text:
+            print(f"\n[Think]\n{think_text}\n")
+
+        choices = data.get("choices", [])
+
+        for choice in choices:
+            choice_content = choice.get("message", {}).get("content")
+
+            if isinstance(choice_content, list) and len(choice_content) > 0:
+                first_item = choice_content[0]
+                if isinstance(first_item, dict) and "image_url" in first_item:
+                    img_url_str = first_item["image_url"].get("url", "")
+                    if img_url_str.startswith("data:image"):
+                        _, b64_data = img_url_str.split(",", 1)
+                        return base64.b64decode(b64_data)
+
+        for choice in choices:
+            choice_content = choice.get("message", {}).get("content")
+            if isinstance(choice_content, str) and choice_content:
+                return choice_content
+
+        print(f"Unexpected response format: {choices}")
+        return None
+
+    except Exception as e:
+        print(f"Error: {e}")
+        return None
+
+
+def main():
+    parser = argparse.ArgumentParser(description="SenseNova-U1 multimodal chat client")
+    parser.add_argument("--prompt", "-p", default="A cute cat", help="Text prompt")
+    parser.add_argument("--output", "-o", default="sensenova_u1_output.png", help="Output file (for image results)")
+    parser.add_argument("--server", "-s", default="http://localhost:8091", help="Server URL")
+    parser.add_argument("--image-url", "-i", type=str, help="Input image URL or local path")
+    parser.add_argument(
+        "--modality",
+        "-m",
+        default="text2img",
+        choices=["text2img", "img2img", "img2text", "text2text"],
+        help="Task modality",
+    )
+    # Standard generation parameters
+    parser.add_argument("--height", type=int, default=2048, help="Image height")
+    parser.add_argument("--width", type=int, default=2048, help="Image width")
+    parser.add_argument("--num-steps", type=int, default=50, help="Inference steps")
+    parser.add_argument("--seed", type=int, default=42, help="Random seed")
+    # Model-specific parameters (forwarded via EXTRA_BODY_PARAMS)
+    parser.add_argument("--cfg-scale", type=float, default=4.0, help="CFG scale")
+    parser.add_argument("--img-cfg-scale", type=float, help="Image CFG scale (img2img)")
+    parser.add_argument("--cfg-norm", type=str, help="CFG normalization mode")
+    parser.add_argument("--timestep-shift", type=float, help="Timestep shift")
+    parser.add_argument("--think", action="store_true", help="Enable think mode")
+
+    args = parser.parse_args()
+    is_image = args.modality in ("text2img", "img2img")
+
+    print(f"Mode: {args.modality}")
+    if args.image_url:
+        print(f"Input Image: {args.image_url}")
+
+    extra: dict[str, object] = {
+        "seed": args.seed,
+        "think": args.think or None,
+    }
+    if is_image:
+        extra.update(
+            height=args.height,
+            width=args.width,
+            num_inference_steps=args.num_steps,
+            cfg_scale=args.cfg_scale,
+        )
+        if args.img_cfg_scale is not None:
+            extra["img_cfg_scale"] = args.img_cfg_scale
+        if args.cfg_norm is not None:
+            extra["cfg_norm"] = args.cfg_norm
+        if args.timestep_shift is not None:
+            extra["timestep_shift"] = args.timestep_shift
+
+    result = generate(
+        prompt=args.prompt,
+        server_url=args.server,
+        image_url=args.image_url,
+        modality=args.modality,
+        **extra,
+    )
+
+    if result:
+        if isinstance(result, bytes):
+            output_path = Path(args.output)
+            output_path.write_bytes(result)
+            print(f"Image saved to: {output_path}")
+            print(f"Size: {len(result) / 1024:.1f} KB")
+        elif isinstance(result, str):
+            print(f"[Response]\n{result}")
+    else:
+        print("Failed to generate response")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()

examples/online_serving/sensenova_u1/run_server.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+# SenseNova-U1 online serving startup script
+
+MODEL="${MODEL:-SenseNova/SenseNova-U1-8B-MoT}"
+PORT="${PORT:-8091}"
+
+echo "Starting SenseNova-U1 server..."
+echo "Model: $MODEL"
+echo "Port: $PORT"
+
+vllm serve "$MODEL" --omni \
+    --port "$PORT"

examples/online_serving/text_to_speech/qwen3_tts/streaming_speech_client.py

@@ -7,6 +7,12 @@ Usage:
     # Send full text at once
     python streaming_speech_client.py --text "Hello world. How are you? I am fine."
 
+    # Pick a built-in speaker for CustomVoice models
+    python streaming_speech_client.py \
+        --text "打开电子枪，关闭扫描，切换到低倍模式。" \
+        --speaker "Serena" \
+        --language "Chinese"
+
     # Simulate STT: send text word-by-word with delay
     python streaming_speech_client.py \
         --text "Hello world. How are you? I am fine." \
@@ -211,7 +217,9 @@ def main():
 
     args = parser.parse_args()
 
-    # Build session config (only include non-None values)
+    # Build session config (only include non-None values).
+    # Server canonical field is `voice`; `speaker` is accepted as an alias
+    # (see StreamingSpeechSessionConfig.voice in protocol/audio.py).
     config = {}
     for key in [
         "model",

examples/online_serving/text_to_speech/voxcpm/run_server.sh

@@ -17,7 +17,7 @@ case "$MODE" in
         STAGE_CONFIG="vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml"
         ;;
     sync)
-        STAGE_CONFIG="vllm_omni/model_executor/stage_configs/voxcpm.yaml"
+        STAGE_CONFIG="vllm_omni/deploy/voxcpm.yaml"
         ;;
     *)
         echo "Unknown mode: $MODE"

pyproject.toml

@@ -55,11 +55,16 @@ dev = [
     "opencc>=1.2.0",
     "mistune>=3.2.0", # for example tests
     "torchmetrics>=1.4.0", # for accuracy similarity metrics
+    "ray[default]", # for RL test
     "jiwer>=3.0.0",
     "zhon>=2.0.0",
     "zhconv>=1.4.2",
     "scipy>=1.10.0",
     "funasr>=1.0.0",
+    "lpips>=0.1.4", # for quantization fp8
+    "orjson>=3.11.7", # to make v1/embeddings API fast
+    "voxcpm>=2.0.2", # for voxcpm model Test
+    "torchdiffeq>=0.2.5" # for covo_audio Test
 ]
 
 demo = [
@@ -142,6 +147,12 @@ ignore = [
 
 [tool.ruff.lint.flake8-tidy-imports.banned-api]
 "librosa".msg = "The librosa module is banned, use vllm.multimodal helpers instead"
+"torch.cuda.max_memory_allocated".msg = "Use torch.accelerator.max_memory_allocated"
+"torch.cuda.reset_peak_memory_stats".msg = "Use torch.accelerator.reset_peak_memory_stats"
+"torch.cuda.device_count".msg = "Use torch.accelerator.device_count"
+"torch.cuda.current_device".msg = "Use torch.accelerator.current_device_index"
+"torch.cuda.empty_cache".msg  = "Use torch.accelerator.empty_cache"
+"torch.cuda.synchronize".msg = "Use torch.accelerator.synchronize"
 
 [tool.mypy]
 python_version = "3.12, 3.13"
@@ -176,6 +187,7 @@ markers = [
     # function module markers
     "diffusion: Diffusion model tests",
     "omni: Omni model tests",
+    "tts: TTS model tests",
     "cache: Cache backend tests",
     "parallel: Parallelism/distributed tests",
     "example: Doc example code tests",

recipes/Qwen/Qwen3-TTS.md

@@ -0,0 +1,168 @@
+# Qwen3-TTS for text-to-speech
+
+## Summary
+
+- Vendor: Qwen
+- Model: `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` (and VoiceDesign / Base variants)
+- Task: Text-to-speech with predefined voices, voice design, or voice cloning
+- Mode: Online serving with the OpenAI-compatible `/v1/audio/speech` API
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe when you want a known-good starting point for serving Qwen3-TTS
+models with vLLM-Omni and validate the deployment with the existing TTS client
+examples in this repository.
+
+Qwen3-TTS supports three task types, each backed by a dedicated model checkpoint:
+
+| Task Type     | Model                                    | Description                                                  |
+| ------------- | ---------------------------------------- | ------------------------------------------------------------ |
+| `CustomVoice` | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice`  | Predefined speaker voices with optional style/emotion control |
+| `VoiceDesign` | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`  | Generate speech from a natural language voice description     |
+| `Base`        | `Qwen/Qwen3-TTS-12Hz-1.7B-Base`         | Voice cloning from reference audio + transcript              |
+
+Smaller 0.6B variants are also available for `CustomVoice` and `Base`.
+
+## References
+
+- Upstream or canonical docs:
+  [`docs/user_guide/examples/online_serving/qwen3_tts.md`](../../docs/user_guide/examples/online_serving/qwen3_tts.md)
+- Related examples under `examples/`:
+  [`examples/online_serving/qwen3_tts/`](../../examples/online_serving/qwen3_tts/),
+  [`examples/offline_inference/qwen3_tts/`](../../examples/offline_inference/qwen3_tts/)
+- Related issue or discussion:
+  [RFC: add recipes folder](https://github.com/vllm-project/vllm-omni/issues/2645)
+
+## Environment
+
+- OS: Linux
+- Python: 3.10+
+- vLLM / vLLM-Omni: use versions from your current checkout, >=0.20.0
+
+## Command
+
+Start the server from the repository root. Pick the model that matches your
+task type:
+
+```bash
+# CustomVoice (predefined speakers with optional style control)
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+    --deploy-config vllm_omni/deploy/qwen3_tts.yaml \
+    --omni --port 8091
+
+# VoiceDesign (natural language voice description)
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
+    --deploy-config vllm_omni/deploy/qwen3_tts.yaml \
+    --omni --port 8091
+
+# Base (voice cloning from reference audio)
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+    --deploy-config vllm_omni/deploy/qwen3_tts.yaml \
+    --omni --port 8091
+```
+
+Alternatively, use the convenience script:
+
+```bash
+./examples/online_serving/qwen3_tts/run_server.sh                  # Default: CustomVoice
+./examples/online_serving/qwen3_tts/run_server.sh VoiceDesign      # VoiceDesign
+./examples/online_serving/qwen3_tts/run_server.sh Base             # Base (voice clone)
+```
+
+The bundled deploy config (`vllm_omni/deploy/qwen3_tts.yaml`) enables async
+chunking for low first-audio latency. For advanced deployment tuning, pass a
+custom deploy config:
+
+```bash
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+    --deploy-config /path/to/your_qwen3_tts_overrides.yaml \
+    --omni --port 8091 --trust-remote-code
+```
+
+## Verification
+
+**Quick smoke test with curl (CustomVoice):**
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "input": "Hello, how are you?",
+        "voice": "vivian",
+        "language": "English"
+    }' --output output.wav
+```
+
+**CustomVoice with emotion instruction:**
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "input": "I am so excited!",
+        "voice": "vivian",
+        "instructions": "Speak with great enthusiasm"
+    }' --output excited.wav
+```
+
+**List available voices:**
+
+```bash
+curl http://localhost:8091/v1/audio/voices
+```
+
+**Using the Python client:**
+
+```bash
+cd examples/online_serving/qwen3_tts
+
+# CustomVoice
+python openai_speech_client.py \
+    --text "Hello, how are you?" \
+    --speaker vivian --language English
+
+# VoiceDesign (requires VoiceDesign model)
+python openai_speech_client.py \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign \
+    --task-type VoiceDesign \
+    --text "Hello world" \
+    --instructions "A warm, friendly female voice"
+
+# Base / voice clone (requires Base model)
+python openai_speech_client.py \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+    --task-type Base \
+    --text "Hello, this is a cloned voice" \
+    --ref-audio /path/to/reference.wav \
+    --ref-text "Transcript of the reference audio"
+```
+
+**Streaming audio (low latency):**
+
+```bash
+curl -X POST http://localhost:8091/v1/audio/speech \
+    -H "Content-Type: application/json" \
+    -d '{
+        "input": "Hello, how are you?",
+        "voice": "vivian",
+        "language": "English",
+        "stream": true,
+        "response_format": "pcm"
+    }' --no-buffer | play -t raw -r 24000 -e signed -b 16 -c 1 -
+```
+
+**Offline inference (no server needed):**
+
+```bash
+python examples/offline_inference/qwen3_tts/end2end.py --query-type CustomVoice
+python examples/offline_inference/qwen3_tts/end2end.py --query-type CustomVoice --streaming
+```
+
+## Notes
+
+- Memory usage: The deploy config allocates `gpu_memory_utilization: 0.3` per stage (talker + code2wav share a single GPU). For the 0.6B variants or constrained GPUs, adjust via `--gpu-memory-utilization`.
+- Key flags: `--omni` is required. `--deploy-config` points to the bundled two-stage pipeline config.
+- Async chunking: Enabled by default in `qwen3_tts.yaml` for streaming-friendly first-audio latency. Streaming requires `stream=true` with `response_format="pcm"`.
+- Task/model matching: Each task type requires its matching model checkpoint. Using a CustomVoice model for a Base (voice clone) request will fail.
+- Known limitations: The server serves one model variant at a time. To switch task types (e.g., CustomVoice to Base), restart the server with the corresponding model.

recipes/README.md

@@ -29,6 +29,10 @@ recipes/
   Qwen-Image on `1x A100 80GB`, including optional step-wise continuous batching replay
 - [`Qwen/Qwen3-Omni.md`](./Qwen/Qwen3-Omni.md): online serving recipe for
   multimodal chat on `1x A100 80GB`
+- [`Tencent/Covo-Audio-Chat.md`](./Tencent/Covo-Audio-Chat.md): online
+  serving recipe for audio chat on `1x A100 80GB`
+- [`Qwen/Qwen3-TTS.md`](./Qwen/Qwen3-TTS.md): text-to-speech serving recipe
+  for Qwen3-TTS (CustomVoice / VoiceDesign / Base) on `1x H100/A100 80GB`
 - [`LTX/LTX-2.md`](./LTX/LTX-2.md): text-to-video and image-to-video serving
   recipe for LTX-2 on `1x H200 141GB`
 - [`Wan-AI/Wan2.2-I2V.md`](./Wan-AI/Wan2.2-I2V.md): image-to-video serving

recipes/SenseNova/SenseNova-U1.md

@@ -0,0 +1,187 @@
+# SenseNova-U1 for unified image generation and understanding
+
+## Summary
+
+- Vendor: SenseNova
+- Model: `SenseNova/SenseNova-U1-8B-MoT`
+- Task: text2img, img2img, img2text (visual understanding), text2text (chat)
+- Mode: Offline inference, Online serving (OpenAI-compatible API)
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe to run SenseNova-U1-8B-MoT via vLLM-Omni. SenseNova-U1 is a
+unified Qwen3-based LLM with Mixture-of-Tokenizers (MoT) attention that handles
+text encoding, optional chain-of-thought reasoning, flow-matching image
+denoising, and visual understanding in a single pipeline — no separate text
+encoder or VAE needed. It supports four task modalities: text-to-image,
+image-to-image editing (with dual CFG), image-to-text understanding, and
+text-to-text chat.
+
+## References
+
+- Offline example:
+  [`examples/offline_inference/sensenova_u1/end2end.py`](../../examples/offline_inference/sensenova_u1/end2end.py)
+- Online serving:
+  [`examples/online_serving/sensenova_u1/`](../../examples/online_serving/sensenova_u1/)
+- E2E tests:
+  [`tests/e2e/offline_inference/test_sensenova_u1_text2img.py`](../../tests/e2e/offline_inference/test_sensenova_u1_text2img.py),
+  [`tests/e2e/offline_inference/test_sensenova_u1_img2img.py`](../../tests/e2e/offline_inference/test_sensenova_u1_img2img.py)
+- HuggingFace model page:
+  [SenseNova/SenseNova-U1-8B-MoT](https://huggingface.co/SenseNova/SenseNova-U1-8B-MoT)
+
+## Hardware Support
+
+## GPU
+
+### 1x H200 (144GB)
+
+#### Environment
+
+- OS: Linux
+- Python: 3.12
+- Driver / runtime: NVIDIA 590.48.01, CUDA 13.1
+- vLLM-Omni version: 0.18.1.dev
+
+#### Command
+
+```bash
+python examples/offline_inference/sensenova_u1/end2end.py \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, textured skin, gentle smile, warm natural light, emotional documentary look. The portrait should feel polished and natural, with sharp eyes, realistic skin texture, accurate facial anatomy, and premium lighting that keeps the face as the main focus." \
+    --width 1536 --height 2720 \
+    --seed 42 --num-steps 50 \
+    --cfg-scale 4.0 --timestep-shift 3.0 --cfg-norm none \
+    --think --print-think \
+    --output outputs
+```
+
+#### Verification
+
+```bash
+pytest -s -v tests/e2e/offline_inference/test_sensenova_u1_text2img.py \
+    -m "advanced_model" --run-level "advanced_model"
+```
+
+#### Notes
+
+- E2E latency: **32.1s** (1536×2720, 50 steps, think mode, CFG scale 4.0)
+- Peak VRAM: **35.9 GB** reserved, 35.1 GB allocated
+- Model loading: 32.8 GiB, 8.7s
+- No deploy YAML needed — the engine auto-generates a single-stage diffusion config.
+- Think mode (`--think`) is recommended for higher image quality.
+
+#### Image-to-Image Editing (img2img)
+
+```bash
+python examples/offline_inference/sensenova_u1/end2end.py \
+    --prompt "Turn this into an oil painting" \
+    --image input.png \
+    --width 2048 --height 2048 \
+    --seed 42 --num-steps 50 \
+    --cfg-scale 4.0 --img-cfg-scale 1.0 --cfg-norm none \
+    --think --print-think \
+    --output outputs
+```
+
+- img2img uses dual CFG: `--cfg-scale` controls text guidance, `--img-cfg-scale`
+  controls image guidance (1.0 = image CFG disabled).
+- Pass multiple `--image` paths for multi-reference editing.
+
+#### Image Understanding (img2text)
+
+```bash
+python examples/offline_inference/sensenova_u1/end2end.py \
+    --modality img2text \
+    --prompt "Describe this image in detail" \
+    --image photo.jpg \
+    --max-tokens 512
+```
+
+#### Text-to-Text Chat (text2text)
+
+```bash
+python examples/offline_inference/sensenova_u1/end2end.py \
+    --modality text2text \
+    --prompt "Explain the theory of relativity in simple terms" \
+    --max-tokens 256
+```
+
+- For img2text and text2text, image generation parameters (height, width,
+  num-steps, cfg-scale) are ignored.
+- Use `--do-sample --temperature 0.7` for more diverse text responses.
+
+### 2x H200 (144GB) — TP=2
+
+#### Environment
+
+- OS: Linux
+- Python: 3.12
+- Driver / runtime: NVIDIA 590.48.01, CUDA 13.1
+- vLLM-Omni version: 0.18.1.dev
+
+#### Command
+
+```bash
+python examples/offline_inference/sensenova_u1/end2end.py \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, textured skin, gentle smile, warm natural light, emotional documentary look. The portrait should feel polished and natural, with sharp eyes, realistic skin texture, accurate facial anatomy, and premium lighting that keeps the face as the main focus." \
+    --width 1536 --height 2720 \
+    --seed 42 --num-steps 50 \
+    --cfg-scale 4.0 --timestep-shift 3.0 --cfg-norm none \
+    --think --print-think \
+    --tensor-parallel-size 2 \
+    --output outputs
+```
+
+#### Verification
+
+Verify the output image is generated at `outputs/sensenova_u1_output_0.png`
+with the expected 1536×2720 resolution.
+
+#### Notes
+
+- E2E latency: **28.3s** (1536×2720, 50 steps, think mode, CFG scale 4.0)
+- Peak VRAM (per GPU): **18.2 GB** reserved, 17.9 GB allocated
+- Model loading: 16.5 GiB per GPU, 7.0s
+- TP=2 provides ~12% speedup over TP=1; limited by serial CFG dual-forward
+  and communication overhead.
+- The LLM transformer uses `QKVParallelLinear` and `MergedColumnParallelLinear`
+  for fused QKV and gate/up projections with TP support.
+
+## Online Serving
+
+SenseNova-U1 supports all four modalities via the OpenAI-compatible
+`/v1/chat/completions` API.
+
+### Launch
+
+```bash
+vllm serve SenseNova/SenseNova-U1-8B-MoT --omni --port 8091
+```
+
+### Send Requests
+
+```bash
+cd examples/online_serving/sensenova_u1
+
+# Text-to-image
+python openai_chat_client.py \
+    --prompt "A beautiful sunset" --modality text2img
+
+# Image-to-image editing
+python openai_chat_client.py \
+    --prompt "Turn this into an oil painting" \
+    --modality img2img --image-url input.jpg
+
+# Image understanding
+python openai_chat_client.py \
+    --prompt "Describe this image" \
+    --modality img2text --image-url photo.jpg
+
+# Text chat
+python openai_chat_client.py \
+    --prompt "What is the capital of France?" \
+    --modality text2text
+```
+
+For full API documentation and curl examples, see
+[`examples/online_serving/sensenova_u1/README.md`](../../examples/online_serving/sensenova_u1/README.md).

recipes/Tencent-Hunyuan/HunyuanImage-3.0-Instruct.md

@@ -186,9 +186,9 @@ Validated score summary:
 
 The CI assertion threshold is `overall_mean >= 0.45`, so the smoke result is
 comfortably above the gate. The generate server and judge server run
-sequentially through the `OmniServer` fixture, and
-`VLLM_TEST_CLEAN_GPU_MEMORY=1` is used to wait for memory cleanup between
-server lifetimes.
+sequentially through the `OmniServer` fixture, with GPU memory cleanup
+between server lifetimes (for example via the `clean_gpu_memory_between_tests`
+pytest fixture in the smoke path).
 
 The lower-cost 2-GPU Instruct setup was tried for this smoke path but did not
 fit in memory. A previous 2-GPU experiment used the base HunyuanImage-3.0

recipes/Tencent/Covo-Audio-Chat.md

@@ -0,0 +1,92 @@
+# Covo-Audio-Chat for audio chat on 1x A100 80GB
+
+## Summary
+
+- Vendor: Tencent
+- Model: `tencent/Covo-Audio-Chat`
+- Task: Audio chat — audio (and optional text) input, interleaved text + audio output
+- Mode: Online serving with the OpenAI-compatible API (offline inference also supported)
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe when you want a known-good starting point for serving
+`tencent/Covo-Audio-Chat` with vLLM-Omni on a single 80 GB-class GPU and
+validate the deployment with the bundled client example.
+
+## References
+
+- Related example under `examples/`:
+  [`examples/online_serving/covo_audio/README.md`](../../examples/online_serving/covo_audio/README.md)
+- Offline end-to-end script:
+  [`examples/offline_inference/covo_audio/end2end.py`](../../examples/offline_inference/covo_audio/end2end.py)
+- Deploy config:
+  [`vllm_omni/deploy/covo_audio.yaml`](../../vllm_omni/deploy/covo_audio.yaml)
+- Pipeline topology:
+  [`vllm_omni/model_executor/models/covo_audio/pipeline.py`](../../vllm_omni/model_executor/models/covo_audio/pipeline.py)
+
+## Hardware Support
+
+This recipe currently documents one tested reference configuration for
+CUDA GPU serving. The fused thinker+talker stage and the BigVGAN code2wav
+stage both run on the same physical device.
+
+## GPU
+
+### 1x A100 80GB
+
+#### Environment
+
+- OS: Linux
+- Python: 3.12
+- Driver / runtime: NVIDIA CUDA 12.x with an A100 80 GB GPU
+- vLLM version: Match the repository requirements for your checkout
+- vLLM-Omni version or commit: Use the commit you are deploying from
+- Extra Python deps: `torchdiffeq`, `librosa`
+
+#### Command
+
+Install the two runtime-only deps that are not pulled in by vLLM-Omni itself:
+
+```bash
+pip install torchdiffeq librosa
+```
+
+Start the server from the repository root:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 vllm serve tencent/Covo-Audio-Chat --omni \
+    --trust-remote-code --port 18091
+```
+
+The deploy config at `vllm_omni/deploy/covo_audio.yaml` is loaded
+automatically by the model registry.
+
+#### Verification
+
+With the server ready, send a test request via the bundled client:
+
+```bash
+cd examples/online_serving/covo_audio
+python openai_chat_completion_client.py
+```
+
+Expected: the client prints the generated text and writes an output WAV
+file to the current directory.
+
+#### Notes
+
+- Memory usage: Stage 0 (7B LLM) ~16 GiB + stage 1 (BigVGAN vocoder)
+  ~2 GiB, both resident on the same device.
+- Key flags: `--omni` and `--trust-remote-code` are both required.
+- System prompt: the bundled `openai_chat_completion_client.py` and
+  `end2end.py` already include `COVO_AUDIO_SYSTEM_PROMPT` (defined in
+  `vllm_omni/model_executor/models/covo_audio/prompt_utils.py`) in the
+  system message by default. When integrating your own client, make sure
+  to include it — without it, stage 0 emits text-only tokens and the
+  output audio is silent.
+- Stage 1 runs with `enforce_eager: true` and `dtype: float32` — see the
+  header of `deploy/covo_audio.yaml` for the specific reasons.
+- Known limitations: this starter recipe covers the single-GPU online-serving
+  path only; multi-GPU layouts and platform-specific overrides are not
+  documented here.

recipes/Wan-AI/Wan2.2-I2V.md

@@ -11,7 +11,9 @@
 ## When to use this recipe
 
 Use this recipe when you want to deploy the Wan2.2 14B image-to-video model
-with vLLM-Omni using multi-card parallelism. Two configurations are provided:
+with vLLM-Omni using multi-card parallelism. The recommended parallel strategy
+depends on the hardware backend and whether the checkpoint needs
+classifier-free guidance.
 
 1. **Distilled model (no negative-prompt / CFG computation)** — higher
    throughput, recommended when using a distilled checkpoint that does not
@@ -25,6 +27,66 @@ with vLLM-Omni using multi-card parallelism. Two configurations are provided:
 
 ## Hardware Support
 
+## GPU
+
+### 8x NVIDIA H20 / H100 / A100
+
+#### Environment
+
+- OS: Linux
+- Python: 3.10+
+- Driver / runtime: NVIDIA driver with CUDA runtime supported by your PyTorch
+  build
+- Recommended operator library: Triton, installed through the vLLM/vLLM-Omni
+  Python environment
+- vLLM version: Match the repository requirements for your checkout
+- vLLM-Omni version or commit: Use the commit you are deploying from
+
+#### Command
+
+**Distilled model (no CFG, recommended for distilled checkpoints):**
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+vllm serve Wan-AI/Wan2.2-I2V-A14B-Diffusers \
+  --omni \
+  --use-hsdp \
+  --usp 8 \
+  --vae-patch-parallel-size 8 \
+  --vae-use-tiling
+```
+
+**Official open-source model (with CFG):**
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
+vllm serve Wan-AI/Wan2.2-I2V-A14B-Diffusers \
+  --omni \
+  --use-hsdp \
+  --cfg-parallel-size 2 \
+  --usp 4 \
+  --vae-patch-parallel-size 8 \
+  --vae-use-tiling
+```
+
+The official model benefits from splitting the positive and negative CFG
+branches with `--cfg-parallel-size 2`, while keeping Ulysses sequence
+parallelism at `--usp 4` so the two CFG branches cover all 8 GPUs.
+
+#### Performance Notes
+
+- `--use-hsdp` is recommended for memory efficiency on the 14B DiT backbone.
+- `--vae-patch-parallel-size 8` is recommended for the standard 8-card setup;
+  disabling VAE patch parallelism can significantly increase VAE latency.
+- The CUDA Triton fused AdaLayerNorm path is not available in the current
+  release. The `VLLM_OMNI_ENABLE_TRITON_ADALN` toggle is reserved for that
+  future implementation and should not be used until the kernel lands.
+- In an 8x NVIDIA H20 test with the official model, `num_inference_steps=4`,
+  `flow_shift=12.0`, `--use-hsdp`, `--cfg-parallel-size 2`, `--usp 4`, and
+  `--vae-patch-parallel-size 8`, an internal Triton AdaLayerNorm prototype
+  reduced the measured diffusion stage by about `65 ms` at
+  `832x480, 81 frames` and about `741 ms` at `1280x720, 121 frames`.
+
 ## NPU
 
 ### 8x Ascend A2 / A3

requirements/common.txt

@@ -1,7 +1,8 @@
 # Common dependencies for all platforms
 av>=14.0.0
 omegaconf>=2.3.0
-diffusers>=0.36.0
+diffusers==0.38.0
+safetensors>=0.8.0rc0
 accelerate==1.12.0
 soundfile>=0.13.1
 cache-dit==1.3.0

tests/conftest.py

@@ -8,6 +8,14 @@ Root pytest entrypoint for the vLLM-Omni test suite.
 
 from __future__ import annotations
 
+# Before ``pytest_plugins`` and before any other test path imports vLLM, pin op
+# registration order (see :func:`tests.model_executor.helpers.bootstrap_vllm_layer_custom_op_modules`).
+# Subdir ``conftest`` hooks can run after other tests are collected/imported, which is
+# too late and can trigger duplicate ``vllm::flashinfer_rotary_embedding`` (etc.) errors.
+from tests.model_executor.helpers import bootstrap_vllm_layer_custom_op_modules
+
+bootstrap_vllm_layer_custom_op_modules()
+
 pytest_plugins = (
     "tests.helpers.fixtures.env",
     "tests.helpers.fixtures.log",

tests/core/sched/test_omni_scheduler_input_coordinator_cleanup.py

@@ -0,0 +1,94 @@
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+from vllm.v1.request import RequestStatus
+
+import vllm_omni.core.sched.omni_ar_scheduler as ar_sched_mod
+import vllm_omni.core.sched.omni_generation_scheduler as gen_sched_mod
+from vllm_omni.core.sched.omni_ar_scheduler import OmniARScheduler
+from vllm_omni.core.sched.omni_generation_scheduler import OmniGenerationScheduler
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+class FakeInputCoordinator:
+    def __init__(self) -> None:
+        self.freed: list[str] = []
+
+    def free_finished_request(self, request_id: str) -> None:
+        self.freed.append(request_id)
+
+
+class DummyRequest:
+    request_id = "req-free"
+    client_index = 0
+    additional_information = None
+
+    def is_finished(self) -> bool:
+        return True
+
+
+@pytest.mark.parametrize(
+    ("scheduler_cls", "scheduler_mod"),
+    [
+        (OmniARScheduler, ar_sched_mod),
+        (OmniGenerationScheduler, gen_sched_mod),
+    ],
+)
+def test_finish_requests_cleans_input_coordinator_for_finished_ids(
+    monkeypatch: pytest.MonkeyPatch,
+    scheduler_cls,
+    scheduler_mod,
+) -> None:
+    coordinator = FakeInputCoordinator()
+    scheduler = scheduler_cls.__new__(scheduler_cls)
+    scheduler.chunk_transfer_adapter = None
+    scheduler.input_coordinator = coordinator
+
+    def fake_finish_requests(self, request_ids, finished_status):
+        assert request_ids == ["req-a", "req-b"]
+        assert finished_status == RequestStatus.FINISHED_STOPPED
+        return [("req-a", 0), ("req-b", 1)]
+
+    monkeypatch.setattr(scheduler_mod.VLLMScheduler, "finish_requests", fake_finish_requests)
+
+    result = scheduler_cls.finish_requests(
+        scheduler,
+        ["req-a", "req-b"],
+        RequestStatus.FINISHED_STOPPED,
+    )
+
+    assert result == [("req-a", 0), ("req-b", 1)]
+    assert coordinator.freed == ["req-a", "req-b"]
+
+
+def test_ar_free_request_cleans_input_coordinator_on_normal_free() -> None:
+    coordinator = FakeInputCoordinator()
+    scheduler = OmniARScheduler.__new__(OmniARScheduler)
+    scheduler.input_coordinator = coordinator
+    scheduler._omits_kv_transfer_cache = {"req-free": True}
+    scheduler.encoder_cache_manager = SimpleNamespace(free=lambda request: None)
+    scheduler.finished_req_ids = set()
+    scheduler.finished_req_ids_dict = None
+    scheduler._new_prompt_len_snapshot = {"req-free": 3}
+    scheduler._connector_finished = lambda request: (False, None)
+    scheduler._should_transfer_kv_for_request = lambda request_id: False
+    scheduler._free_blocks = lambda request: None
+
+    result = OmniARScheduler._free_request(scheduler, DummyRequest())
+
+    assert result is None
+    assert coordinator.freed == ["req-free"]
+    assert scheduler._omits_kv_transfer_cache == {}
+    assert scheduler._new_prompt_len_snapshot == {}
+
+
+def test_generation_free_request_cleans_input_coordinator(monkeypatch: pytest.MonkeyPatch) -> None:
+    coordinator = FakeInputCoordinator()
+    scheduler = OmniGenerationScheduler.__new__(OmniGenerationScheduler)
+    scheduler.input_coordinator = coordinator
+
+    def fake_free_request(self, request, delay_free_blocks=False):
+        assert delay_free_blocks is True

tests/core/sched/test_omni_scheduling_coordinator.py

@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""Unit tests for OmniSchedulingCoordinator (formerly ChunkSchedulingCoordinator).
+"""Unit tests for OmniSchedulingCoordinator.
 
 These tests use mock request objects and mock queues.  They do not require
 GPU, vLLM runtime, or any connector.
@@ -11,10 +11,12 @@ from __future__ import annotations
 import unittest
 from types import SimpleNamespace
 
+import torch
+
 import vllm_omni.core.sched.omni_scheduling_coordinator as coord_mod
 from vllm_omni.core.sched.omni_scheduling_coordinator import (
-    ChunkSchedulingCoordinator,
     OmniSchedulingCoordinator,
+    uses_qwen3_omni_full_payload_input_coordinator,
 )
 
 # ------------------------------------------------------------------ #
@@ -89,11 +91,55 @@ class MockQueue:
 # ------------------------------------------------------------------ #
 
 
+class TestFullPayloadCoordinatorSelection(unittest.TestCase):
+    def test_qwen3_omni_talker_and_code2wav_use_full_payload_input_coordinator(self):
+        for model_stage in ("talker", "code2wav"):
+            model_config = SimpleNamespace(
+                stage_id=1,
+                async_chunk=False,
+                model_arch="Qwen3OmniMoeForConditionalGeneration",
+                model_stage=model_stage,
+            )
+
+            self.assertTrue(uses_qwen3_omni_full_payload_input_coordinator(model_config))
+
+    def test_async_chunk_and_non_qwen3_omni_do_not_use_full_payload_input_coordinator(self):
+        cases = [
+            SimpleNamespace(
+                stage_id=1,
+                async_chunk=True,
+                model_arch="Qwen3OmniMoeForConditionalGeneration",
+                model_stage="talker",
+            ),
+            SimpleNamespace(
+                stage_id=1,
+                async_chunk=False,
+                model_arch="Qwen3TTSForConditionalGeneration",
+                model_stage="code2wav",
+            ),
+            SimpleNamespace(
+                stage_id=1,
+                async_chunk=False,
+                model_arch="Qwen2_5OmniForConditionalGeneration",
+                model_stage="talker",
+            ),
+            SimpleNamespace(
+                stage_id=0,
+                async_chunk=False,
+                model_arch="Qwen3OmniMoeForConditionalGeneration",
+                model_stage="thinker",
+            ),
+        ]
+
+        for model_config in cases:
+            self.assertFalse(uses_qwen3_omni_full_payload_input_coordinator(model_config))
+
+
 class TestChunkCoordinatorStateTransition(unittest.TestCase):
     """Test 5: process_pending_chunks transitions WAITING_FOR_CHUNK → target."""
 
     def test_ready_request_transitions_to_waiting(self):
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1, async_chunk=True)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1, async_chunk=True)
 
         req = _make_request("r1", status=RequestStatus.WAITING_FOR_CHUNK)
         waiting = MockQueue([req])
@@ -110,7 +156,7 @@ class TestChunkCoordinatorStateTransition(unittest.TestCase):
         self.assertIn("r1", coord.requests_with_ready_chunks)
 
     def test_non_ready_stays_waiting_for_chunk(self):
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1, async_chunk=True)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1, async_chunk=True)
 
         req = _make_request("r1", status=RequestStatus.WAITING_FOR_CHUNK)
         waiting = MockQueue([req])
@@ -126,7 +172,7 @@ class TestChunkCoordinatorStateTransition(unittest.TestCase):
         self.assertEqual(req.status, RequestStatus.WAITING_FOR_CHUNK)
 
     def test_stage_0_is_noop(self):
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=0)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=0)
         req = _make_request("r1")
         waiting = MockQueue([req])
         running: list = []
@@ -144,7 +190,7 @@ class TestChunkCoordinatorRestoreQueues(unittest.TestCase):
     """Test 6: restore_queues returns waiting-for-chunk requests."""
 
     def test_restore(self):
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
 
         r1 = _make_request("r1")
         r2 = _make_request("r2")
@@ -166,7 +212,7 @@ class TestChunkCoordinatorFinishedSignal(unittest.TestCase):
     """Test 8: chunk_finished_req_ids → finished_requests."""
 
     def test_finished_signal(self):
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1, async_chunk=True)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1, async_chunk=True)
 
         req = _make_request("r1", status=RequestStatus.WAITING_FOR_CHUNK)
         waiting = MockQueue([req])
@@ -187,7 +233,7 @@ class TestChunkCoordinatorUpdateRequestMetadata(unittest.TestCase):
 
     def test_ar_mode_no_longer_sets_additional_information(self):
         """AR mode only processes scheduling metadata, not full payloads."""
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
 
         req = _make_request("r1")
         requests = {"r1": req}
@@ -204,10 +250,14 @@ class TestChunkCoordinatorUpdateRequestMetadata(unittest.TestCase):
         self.assertIsNone(getattr(req, "additional_information", None))
 
     def test_generation_mode(self):
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
 
         req = _make_request("r1")
         req.prompt_token_ids = [0, 0, 0]
+        req.num_prompt_tokens = 3
+        req.num_computed_tokens = 3
+        req._all_token_ids = [0, 0, 0, 99]
+        req._output_token_ids = [99]
         requests = {"r1": req}
 
         request_metadata = {
@@ -220,16 +270,61 @@ class TestChunkCoordinatorUpdateRequestMetadata(unittest.TestCase):
         coord.update_request_metadata(requests, request_metadata, model_mode="generation")
 
         self.assertEqual(req.prompt_token_ids, [10, 20, 30])
+        self.assertEqual(req.num_prompt_tokens, 3)
         self.assertEqual(req.num_computed_tokens, 0)
+        self.assertEqual(req._all_token_ids, [10, 20, 30])
+        self.assertEqual(req._output_token_ids, [])
         self.assertIsNone(req.additional_information)
-        self.assertEqual(req._omni_initial_model_buffer, {"left_context_size": 25})
+        self.assertEqual(req._omni_initial_model_buffer, {"meta": {"left_context_size": 25}})
+
+    def test_generation_mode_flattens_tensor_code_predictor_codes(self):
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
+
+        req = _make_request("r1")
+        req.prompt_token_ids = [9]
+        req.num_prompt_tokens = 1
+        req._all_token_ids = [9, 8]
+        req._output_token_ids = [8]
+        requests = {"r1": req}
+
+        coord.update_request_metadata(
+            requests,
+            {"r1": {"code_predictor_codes": torch.tensor([[1, 2, 3]], dtype=torch.long)}},
+            model_mode="generation",
+        )
+
+        self.assertEqual(req.prompt_token_ids, [1, 2, 3])
+        self.assertEqual(req.num_prompt_tokens, 3)
+        self.assertEqual(req._all_token_ids, [1, 2, 3])
+        self.assertEqual(req._output_token_ids, [])
+
+    def test_generation_mode_flattens_nested_code_predictor_codes(self):
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
+
+        req = _make_request("r1")
+        req.prompt_token_ids = [9]
+        req.num_prompt_tokens = 1
+        req._all_token_ids = [9, 8]
+        req._output_token_ids = [8]
+        requests = {"r1": req}
+
+        coord.update_request_metadata(
+            requests,
+            {"r1": {"code_predictor_codes": [[1, 2], [3, 4]]}},
+            model_mode="generation",
+        )
+
+        self.assertEqual(req.prompt_token_ids, [1, 2, 3, 4])
+        self.assertEqual(req.num_prompt_tokens, 4)
+        self.assertEqual(req._all_token_ids, [1, 2, 3, 4])
+        self.assertEqual(req._output_token_ids, [])
 
 
 class TestChunkCoordinatorPostprocess(unittest.TestCase):
     """Test postprocess_scheduler_output clears ready chunks."""
 
     def test_clear_ready(self):
-        coord = ChunkSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
+        coord = OmniSchedulingCoordinator(scheduler_max_num_seqs=10, stage_id=1)
         coord.requests_with_ready_chunks = {"r1", "r2"}
 
         new_req = SimpleNamespace(req_id="r1")
@@ -366,6 +461,33 @@ class TestWaitingForInputTransition(unittest.TestCase):
         self.assertEqual(len(coord.pending_input_registrations), 1)
         self.assertEqual(coord.pending_input_registrations[0].request_id, "r1")
 
+    def test_idle_cycles_retain_received_marker_before_request_appears(self):
+        coord = OmniSchedulingCoordinator(
+            scheduler_max_num_seqs=10,
+            stage_id=1,
+            async_chunk=False,
+        )
+        coord._full_payload_input_received.add("late")
+        coord.finished_requests.add("late")
+
+        waiting = MockQueue()
+        running: list = []
+
+        coord.process_pending_full_payload_inputs(waiting, running, stage_recv_req_ids=set())
+
+        self.assertIn("late", coord._full_payload_input_received)
+        self.assertIn("late", coord.finished_requests)
+
+        late_req = _make_request("late", status=RequestStatus.WAITING)
+        waiting.add_request(late_req)
+
+        coord.process_pending_full_payload_inputs(waiting, running, stage_recv_req_ids=set())
+
+        self.assertEqual(late_req.status, RequestStatus.WAITING)
+        self.assertEqual(coord.pending_input_registrations, [])
+        self.assertIn("late", coord._full_payload_input_received)
+        self.assertIn("late", coord.finished_requests)
+
 
 class TestTimeoutDetection(unittest.TestCase):
     """Regression tests for orphaned pending-recv timeout detection.
@@ -508,14 +630,18 @@ class TestTimeoutDetection(unittest.TestCase):
         self.assertEqual(len(coord._waiting_for_input), 0)
 
     def test_free_finished_request_clears_waiting_since(self):
-        """free_finished_request clears _waiting_since."""
+        """free_finished_request clears coordinator lifecycle markers."""
         coord = OmniSchedulingCoordinator(
             scheduler_max_num_seqs=10,
             stage_id=1,
         )
         coord._waiting_since["r1"] = 0.0
+        coord._full_payload_input_received.add("r1")
+        coord.finished_requests.add("r1")
         coord.free_finished_request("r1")
         self.assertNotIn("r1", coord._waiting_since)
+        self.assertNotIn("r1", coord._full_payload_input_received)
+        self.assertNotIn("r1", coord.finished_requests)
 
     def test_timeout_from_running_queue_full_lifecycle(self):
         """End-to-end: request from running → WAITING_FOR_CHUNK → restore →

tests/dfx/conftest.py

@@ -2,6 +2,7 @@ import json
 import os
 import re
 import subprocess
+import threading
 import time
 from datetime import datetime
 from pathlib import Path
@@ -377,11 +378,23 @@ def run_benchmark(
         command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, bufsize=1, universal_newlines=True
     )
 
-    for line in iter(process.stdout.readline, ""):
-        print(line, end=" ")
+    if process.stdout is None or process.stderr is None:
+        raise RuntimeError("Failed to capture benchmark process output streams")
 
-    for line in iter(process.stderr.readline, ""):
-        print(line, end=" ")
+    def _forward_stream(stream) -> None:
+        try:
+            for line in iter(stream.readline, ""):
+                print(line, end="")
+        finally:
+            stream.close()
+
+    stdout_thread = threading.Thread(target=_forward_stream, args=(process.stdout,))
+    stderr_thread = threading.Thread(target=_forward_stream, args=(process.stderr,))
+    stdout_thread.start()
+    stderr_thread.start()
+    stdout_thread.join()
+    stderr_thread.join()
+    process.wait()
 
     if "--result-dir" in command:
         index = command.index("--result-dir")