chore: complete v0.21.0 → v0.22.0 version bump across all files

Update remaining Dockerfiles (cuda, rocm, xpu, ci) and docs
(rocm.inc.md, cuda.inc.md, quickstart.md, gpu.md) that were
missed in the initial pass — pip install lines, git checkout tags,
Docker image tags, wheel URLs, and prose references.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.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/pipeline-intel.yaml

@@ -10,7 +10,7 @@ steps:
           DOCKER_BUILDKIT: "1"
           # Buildkite will automatically replace this with the actual commit hash
           VLLM_IMAGE_TAG: "${BUILDKITE_COMMIT}"
-          VLLM_VERSION: "v0.18.0"
+          VLLM_VERSION: "v0.21.0"
         priority: 100
         timeout_in_minutes: 60
-        soft_fail: true
+        soft_fail: false

.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/pipeline.yml

@@ -1,6 +1,22 @@
+# Document 1: Buildkite loads only this block on first parse. The next step resolves docs-only skip-ci
+# from git diff, then uploads document 2. When docs-only skip applies, image-build still runs for the nightly
+# exception in upload_pipeline_with_skip_ci.sh (nightly-test label or main NIGHTLY=1); otherwise upload-nightly
+# can be skipped together with test-ready/test-merge. Weekly uploads also depend on image-build.
+
+# Document 2: appended after `---`; same file, read by upload_pipeline_with_skip_ci.sh (not evaluated as a second pipeline by Buildkite). Child uploads include test-ready.yml, test-merge.yml, test-nightly.yml, and test-weekly.yml.
+steps:
+  - label: ":github: Resolve skip-ci & upload pipeline"
+    key: upload-ci-pipeline
+    commands:
+      - "bash .buildkite/scripts/upload_pipeline_with_skip_ci.sh"
+    agents:
+      queue: "cpu_queue_premerge"
+
+---
 steps:
   - label: ":docker: Build image"
     key: image-build
+    if: __IMAGE_BUILD_IF__
     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 ."
@@ -13,7 +29,7 @@ steps:
   - label: "Upload Ready Pipeline"
     depends_on: image-build
     key: upload-ready-pipeline
-    if: build.branch != "main" && build.pull_request.labels includes "ready"
+    if: __UPLOAD_READY_IF__
     commands:
       - buildkite-agent pipeline upload .buildkite/test-ready.yml
     agents:
@@ -23,18 +39,38 @@ steps:
   - label: "Upload Merge Pipeline"
     depends_on: image-build
     key: upload-merge-pipeline
-    if: build.branch == "main" && build.env("NIGHTLY") != "1"
+    if: __UPLOAD_MERGE_IF__
     commands:
       - buildkite-agent pipeline upload .buildkite/test-merge.yml
     agents:
       queue: "cpu_queue_premerge"
 
-  # L4 Test — main+NIGHTLY=1 (scheduled), or PR with label nightly-test (e.g. add label then Rebuild)
+  # 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")'
+    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-nightly.yml
     agents:
       queue: "cpu_queue_premerge"
+
+  # L5 Test — main+WEEKLY=1 (scheduled), or PR with weekly-test label
+  - label: "Upload Weekly Pipeline"
+    depends_on: image-build
+    key: upload-weekly-pipeline
+    if: >-
+      (build.branch == "main" && build.env("WEEKLY") == "1") ||
+      (build.branch != "main" && build.pull_request.labels includes "weekly-test")
+    commands:
+      - buildkite-agent pipeline upload .buildkite/test-weekly.yml
+    agents:
+      queue: "cpu_queue_premerge"

.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/generate-and-upload-nightly-index.sh

@@ -0,0 +1,88 @@
+#!/usr/bin/env bash
+
+set -ex
+
+# Generate and upload wheel indices for all vllm-omni wheels in the commit directory.
+# This script should run once after all wheels have been built and uploaded.
+# All paths are under the omni/ prefix in the vllm-wheels S3 bucket.
+
+# ======== setup ========
+
+BUCKET="vllm-wheels"
+INDICES_OUTPUT_DIR="indices"
+PYTHON="${PYTHON_PROG:-python3}"
+SUBPATH="omni/$BUILDKITE_COMMIT"
+S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
+
+# detect if python3.12+ is available
+has_new_python=$($PYTHON -c "print(1 if __import__('sys').version_info >= (3,12) else 0)")
+if [[ "$has_new_python" -eq 0 ]]; then
+    # use new python from docker
+    docker pull python:3-slim
+    PYTHON="docker run --rm --user $(id -u):$(id -g) -v $(pwd):/app -w /app python:3-slim python3"
+fi
+
+echo "Using python interpreter: $PYTHON"
+echo "Python version: $($PYTHON --version)"
+
+# ======== generate and upload indices ========
+
+# list all wheels in the commit directory
+echo "Existing wheels on S3:"
+aws s3 ls "$S3_COMMIT_PREFIX" || echo "(no objects found)"
+obj_json="objects.json"
+aws s3api list-objects-v2 --bucket "$BUCKET" --prefix "$SUBPATH/" --delimiter / --output json > "$obj_json"
+mkdir -p "$INDICES_OUTPUT_DIR"
+
+# HACK: we do not need regex module here, but it is required by pre-commit hook
+# To avoid any external dependency, we simply replace it back to the stdlib re module
+sed -i.bak 's/import regex as re/import re/g' .buildkite/scripts/generate-nightly-index.py && rm -f .buildkite/scripts/generate-nightly-index.py.bak
+
+# Generate indices -- the version is just the commit hash (not omni/{commit})
+# because relative paths are computed between the index and wheel directories,
+# both of which live under the omni/ prefix in S3.
+$PYTHON .buildkite/scripts/generate-nightly-index.py \
+    --version "$BUILDKITE_COMMIT" \
+    --current-objects "$obj_json" \
+    --output-dir "$INDICES_OUTPUT_DIR" \
+    --comment "commit $BUILDKITE_COMMIT"
+
+# copy indices to /omni/{commit}/ unconditionally
+echo "Uploading indices to $S3_COMMIT_PREFIX"
+aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
+
+# copy to /omni/nightly/ when NIGHTLY=1
+if [[ "${NIGHTLY:-}" == "1" ]]; then
+    echo "Uploading indices to overwrite /omni/nightly/"
+    aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/omni/nightly/"
+fi
+
+# detect version from any wheel in the commit directory
+first_wheel_key=$($PYTHON -c "import json; obj=json.load(open('$obj_json')); print(next((c['Key'] for c in obj.get('Contents', []) if c['Key'].endswith('.whl')), ''))")
+if [[ -z "$first_wheel_key" ]]; then
+    echo "Error: No wheels found in $S3_COMMIT_PREFIX"
+    exit 1
+fi
+first_wheel=$(basename "$first_wheel_key")
+aws s3 cp "s3://$BUCKET/${first_wheel_key}" "/tmp/${first_wheel}"
+version=$(unzip -p "/tmp/${first_wheel}" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
+rm -f "/tmp/${first_wheel}"
+echo "Version in wheel: $version"
+pure_version="${version%%+*}"
+echo "Pure version (without variant): $pure_version"
+
+# re-generate and copy to /omni/{version}/ only if it does not have "dev" in the version
+if [[ "$version" != *"dev"* ]]; then
+    s3_version="v$pure_version"
+    echo "Re-generating indices for /omni/$s3_version/"
+    rm -rf "${INDICES_OUTPUT_DIR:?}"
+    mkdir -p "$INDICES_OUTPUT_DIR"
+    # wheel-dir is overridden to be the commit directory, so that the indices point to the correct wheel path
+    $PYTHON .buildkite/scripts/generate-nightly-index.py \
+        --version "$s3_version" \
+        --wheel-dir "$BUILDKITE_COMMIT" \
+        --current-objects "$obj_json" \
+        --output-dir "$INDICES_OUTPUT_DIR" \
+        --comment "version $pure_version"
+    aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/omni/$s3_version/"
+fi

.buildkite/scripts/generate-nightly-index.py

@@ -0,0 +1,192 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import argparse
+import json
+import re
+import sys
+from dataclasses import asdict, dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+from urllib.parse import quote
+
+
+def normalize_package_name(name: str) -> str:
+    """Normalize package name per PEP 503."""
+    return re.sub(r"[-_.]+", "-", name).lower()
+
+
+if not sys.version_info >= (3, 12):
+    raise RuntimeError("This script requires Python 3.12 or higher.")
+
+INDEX_HTML_TEMPLATE = """<!DOCTYPE html>
+<html>
+  <!-- {comment} -->
+  <meta name="pypi:repository-version" content="1.0">
+  <body>
+{items}
+  </body>
+</html>
+"""
+
+
+@dataclass
+class WheelFileInfo:
+    package_name: str
+    version: str
+    build_tag: str | None
+    python_tag: str
+    abi_tag: str
+    platform_tag: str
+    filename: str
+
+
+def parse_from_filename(file: str) -> WheelFileInfo:
+    """
+    Parse wheel filename per PEP 427:
+        {package_name}-{version}(-{build_tag})?-{python_tag}-{abi_tag}-{platform_tag}.whl
+    """
+    wheel_file_re = re.compile(
+        r"^(?P<package_name>.+)-(?P<version>[^-]+?)(-(?P<build_tag>[^-]+))?-(?P<python_tag>[^-]+)-(?P<abi_tag>[^-]+)-(?P<platform_tag>[^-]+)\.whl$"
+    )
+    match = wheel_file_re.match(file)
+    if not match:
+        raise ValueError(f"Invalid wheel file name: {file}")
+
+    return WheelFileInfo(
+        package_name=match.group("package_name"),
+        version=match.group("version"),
+        build_tag=match.group("build_tag"),
+        python_tag=match.group("python_tag"),
+        abi_tag=match.group("abi_tag"),
+        platform_tag=match.group("platform_tag"),
+        filename=file,
+    )
+
+
+def generate_project_list(package_names: list[str], comment: str = "") -> str:
+    """Generate top-level PEP 503 project list HTML."""
+    href_tags = []
+    for name in sorted(package_names):
+        href_tags.append(f'    <a href="{name}/">{name}/</a><br/>')
+    return INDEX_HTML_TEMPLATE.format(items="\n".join(href_tags), comment=comment)
+
+
+def generate_package_index(
+    wheel_files: list[WheelFileInfo],
+    wheel_base_dir: Path,
+    index_base_dir: Path,
+    comment: str = "",
+) -> tuple[str, str]:
+    """Generate package index HTML and metadata JSON linking to wheel files."""
+    href_tags = []
+    metadata = []
+    for file in sorted(wheel_files, key=lambda x: x.filename):
+        relative_path = wheel_base_dir.relative_to(index_base_dir, walk_up=True) / file.filename
+        # handle '+' in URL; avoid double-encoding '/' and '%2B' (AWS S3 behavior)
+        file_path_quoted = quote(relative_path.as_posix(), safe=":%/")
+        href_tags.append(f'    <a href="{file_path_quoted}">{file.filename}</a><br/>')
+        file_meta = asdict(file)
+        file_meta["path"] = file_path_quoted
+        metadata.append(file_meta)
+    index_str = INDEX_HTML_TEMPLATE.format(items="\n".join(href_tags), comment=comment)
+    metadata_str = json.dumps(metadata, indent=2)
+    return index_str, metadata_str
+
+
+def generate_index(
+    whl_files: list[str],
+    wheel_base_dir: Path,
+    index_base_dir: Path,
+    comment: str = "",
+):
+    """
+    Generate PEP 503 index for all wheel files.
+
+    Output structure:
+        index_base_dir/
+            index.html          # project list linking to vllm-omni/
+            vllm-omni/
+                index.html      # package index linking to wheel files
+                metadata.json   # machine-readable metadata
+    """
+    parsed_files = [parse_from_filename(f) for f in whl_files]
+
+    if not parsed_files:
+        print("No wheel files found, skipping index generation.")
+        return
+
+    comment_str = f" ({comment})" if comment else ""
+    comment_tmpl = f"Generated on {datetime.now().isoformat()}{comment_str}"
+
+    # Group by normalized package name
+    packages: dict[str, list[WheelFileInfo]] = {}
+    for file in parsed_files:
+        name = normalize_package_name(file.package_name)
+        packages.setdefault(name, []).append(file)
+
+    print(f"Found packages: {list(packages.keys())}")
+
+    # Generate per-package index
+    for package, files in packages.items():
+        package_dir = index_base_dir / package
+        package_dir.mkdir(parents=True, exist_ok=True)
+        index_str, metadata_str = generate_package_index(files, wheel_base_dir, package_dir, comment)
+        with open(package_dir / "index.html", "w") as f:
+            f.write(index_str)
+        with open(package_dir / "metadata.json", "w") as f:
+            f.write(metadata_str)
+
+    # Generate top-level project list
+    project_list_str = generate_project_list(sorted(packages.keys()), comment_tmpl)
+    with open(index_base_dir / "index.html", "w") as f:
+        f.write(project_list_str)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Generate PEP 503 wheel index from S3 object listing.")
+    parser.add_argument("--version", type=str, required=True, help="Version string (e.g., commit hash)")
+    parser.add_argument("--current-objects", type=str, required=True, help="Path to JSON from S3 list-objects-v2")
+    parser.add_argument("--output-dir", type=str, required=True, help="Directory to write index files")
+    parser.add_argument("--wheel-dir", type=str, default=None, help="Wheel directory (defaults to --version)")
+    parser.add_argument("--comment", type=str, default="", help="Comment for generated HTML")
+
+    args = parser.parse_args()
+
+    version = args.version
+    if "\\" in version or "/" in version:
+        raise ValueError("Version string must not contain slashes or backslashes.")
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    with open(args.current_objects) as f:
+        current_objects: dict[str, list[dict[str, Any]]] = json.load(f)
+
+    wheel_files = [
+        item["Key"].split("/")[-1] for item in current_objects.get("Contents", []) if item["Key"].endswith(".whl")
+    ]
+
+    print(f"Found {len(wheel_files)} wheel files for version {version}: {wheel_files}")
+
+    # For release versions, filter to only matching non-dev wheels
+    PY_VERSION_RE = re.compile(r"^\d+\.\d+\.\d+([a-zA-Z0-9.+-]*)?$")
+    if PY_VERSION_RE.match(version):
+        wheel_files = [f for f in wheel_files if version in f and "dev" not in f]
+        print(f"Non-nightly version detected, wheel files used: {wheel_files}")
+    else:
+        print("Nightly version detected, keeping all wheel files.")
+
+    wheel_dir = (args.wheel_dir or version).strip().rstrip("/")
+    wheel_base_dir = Path(output_dir).parent / wheel_dir
+    index_base_dir = Path(output_dir)
+
+    generate_index(
+        whl_files=wheel_files,
+        wheel_base_dir=wheel_base_dir,
+        index_base_dir=index_base_dir,
+        comment=args.comment.strip(),
+    )
+    print(f"Successfully generated index in {output_dir}")

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

@@ -11,15 +11,6 @@ set -o pipefail
 export PYTHONPATH=".."
 
 # Print ROCm version
-echo "--- Confirming Clean Initial State"
-while true; do
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
-
 echo "--- ROCm info"
 rocminfo
 
@@ -51,28 +42,24 @@ cleanup_docker() {
 # Call the cleanup docker function
 cleanup_docker
 
-echo "--- Resetting GPUs"
-
-echo "reset" > /opt/amdgpu/etc/gpu_state
-
-while true; do
-        sleep 3
-        if grep -q clean /opt/amdgpu/etc/gpu_state; then
-                echo "GPUs state is \"clean\""
-                break
-        fi
-done
-
 echo "--- Pulling container"
-image_name="public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:${BUILDKITE_COMMIT}-rocm-omni"
+## Temporary change to use AMD Docker Hub to store the vllm-omni image
+# to bypass the rate limit issue with ECR Public Gallery.
+# Images are now stored in a separate repository for vllm-omni, instead of vllm-ci.
+# TODO: @tjtanaa point back to ECR Public Gallery
+# once the amd agents are configured to use ECR Public Gallery.
+# image_name="public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:${BUILDKITE_COMMIT}-rocm-omni"
+image_name="rocm/vllm-omni:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 
-# Install AWS CLI to authenticate to ECR Public Gallery to get higher rate limit for pulling images
-sudo apt-get update && sudo apt-get install -y awscli
-# Use safe docker login helper to prevent race conditions
-source "$(dirname "${BASH_SOURCE[0]}")/../docker_login_ecr_public.sh"
-safe_docker_login_ecr_public
-# Pull the container from ECR Public Gallery
+# TODO: @tjtanaa uncomment this once the amd agents are configured to use ECR Public Gallery.
+# # Install AWS CLI to authenticate to ECR Public Gallery to get higher rate limit for pulling images
+# sudo apt-get update && sudo apt-get install -y awscli
+##  Use safe docker login helper to prevent race conditions
+# source "$(dirname "${BASH_SOURCE[0]}")/../docker_login_ecr_public.sh"
+# safe_docker_login_ecr_public
+
+## Pull the container from AMD Docker Hub
 
 docker pull "${image_name}"
 
@@ -126,8 +113,6 @@ if [[ $commands == *"--shard-id="* ]]; then
         -e VLLM_ROCM_USE_AITER=0 \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
         -e HF_TOKEN \
-        -e AWS_ACCESS_KEY_ID \
-        -e AWS_SECRET_ACCESS_KEY \
         -v "${HF_CACHE}:${HF_MOUNT}" \
         -e "HF_HOME=${HF_MOUNT}" \
         -e "PYTHONPATH=${MYPYTHONPATH}" \
@@ -160,8 +145,6 @@ else
           -e MIOPEN_DEBUG_CONV_GEMM=0 \
           -e VLLM_ROCM_USE_AITER=0 \
           -e HF_TOKEN \
-          -e AWS_ACCESS_KEY_ID \
-          -e AWS_SECRET_ACCESS_KEY \
           -v "${HF_CACHE}:${HF_MOUNT}" \
           -e "HF_HOME=${HF_MOUNT}" \
           -e "PYTHONPATH=${MYPYTHONPATH}" \

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

@@ -52,5 +52,7 @@ time timeout -k 30 30m docker run \
     pip install tblib==3.1.0
     cd /workspace/vllm-omni
     pytest -v -s -m "core_model and xpu and B60"
+    pytest -v -s tests/diffusion/quantization/test_mxfp8_config.py
+    export VLLM_XPU_USE_SAMPLER_KERNEL=0    # NOTE: Remove this after vLLM v0.21.1 is merged. Fixes Qwen2-5 omni-expansion tests.
     pytest -v -s -m "advanced_model and xpu and B60"
 '

.buildkite/scripts/hardware_ci/run_npu_test.sh

@@ -141,5 +141,5 @@ docker run \
     "${image_name}" \
     bash -c '
     set -e
-    VLLM_USE_MODELSCOPE=True pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+    VLLM_USE_MODELSCOPE=True pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni_expansion.py
 '

.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-nightly-wheels.sh

@@ -0,0 +1,33 @@
+#!/usr/bin/env bash
+
+set -ex
+
+# Upload a single wheel to S3 under the omni/ prefix.
+# Index generation is handled separately by generate-and-upload-nightly-index.sh.
+
+BUCKET="vllm-wheels"
+SUBPATH="omni/$BUILDKITE_COMMIT"
+S3_COMMIT_PREFIX="s3://$BUCKET/$SUBPATH/"
+
+# ========= collect & upload the wheel ==========
+
+# python3 -m build outputs to dist/ by default
+wheel_files=(dist/*.whl)
+
+# Check that exactly one wheel is found
+if [[ ${#wheel_files[@]} -ne 1 ]]; then
+  echo "Error: Expected exactly one wheel file in dist/, but found ${#wheel_files[@]}"
+  exit 1
+fi
+wheel="${wheel_files[0]}"
+
+echo "Uploading wheel: $wheel"
+
+# Extract the version from the wheel
+version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
+echo "Version in wheel: $version"
+
+# Upload wheel to S3
+aws s3 cp "$wheel" "$S3_COMMIT_PREFIX"
+
+echo "Wheel uploaded to $S3_COMMIT_PREFIX. Index generation is handled by a separate step."

.buildkite/scripts/upload_pipeline_with_skip_ci.sh

@@ -0,0 +1,146 @@
+#!/usr/bin/env bash
+# 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 "${BASH_SOURCE[0]}")/../.." && pwd)"
+PIPELINE_ARG="${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() {
+  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
+}
+
+SKIP_CI="$(resolve_skip_ci)"
+
+if [[ ! -f "${PIPELINE_YML}" ]]; then
+  echo "upload_pipeline_with_skip_ci: missing ${PIPELINE_YML}" >&2
+  exit 1
+fi
+
+export ROOT SKIP_CI PIPELINE_YML
+python3 <<'PY' | buildkite-agent pipeline upload
+import os
+import pathlib
+
+path = pathlib.Path(os.environ["PIPELINE_YML"])
+text = path.read_text(encoding="utf-8")
+sep = "\n---\n"
+# 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
+# main NIGHTLY=1), otherwise upload-nightly (depends_on image-build) would be skipped too.
+nightly_only = (
+    '(build.pull_request.labels includes "nightly-test") '
+    '|| (build.branch == "main" && build.env("NIGHTLY") == "1")'
+)
+# Placeholder in pipeline.yml is `if: __IMAGE_BUILD_IF__` (valid YAML); replace value only.
+if skip:
+    rep = f"'{nightly_only}'"
+    ready_rep = "'false'"
+    merge_rep = "'false'"
+else:
+    rep = "'true'"
+    ready_rep = "'build.branch != \"main\" && build.pull_request.labels includes \"ready\"'"
+    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)
+    .replace("__UPLOAD_READY_IF__", ready_rep)
+    .replace("__UPLOAD_MERGE_IF__", merge_rep)
+)
+print(rendered, end="")
+PY

.buildkite/test-amd-merge.yml

@@ -7,16 +7,9 @@ steps:
   grade: Blocking
   commands:
     - export VLLM_ROCM_USE_AITER=0
-    - "timeout 20m pytest -v -s -m 'core_model and cpu' --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"
+    # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
+    # that is not supported on CPU
+    - "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
@@ -24,7 +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
@@ -32,15 +25,14 @@ steps:
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - |
       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))
       '
@@ -55,7 +47,7 @@ steps:
 #     - export GPU_ARCHS=gfx942
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-#     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_model.py
+#     - 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: "Diffusion Cache Backend Test"
   agent_pool: mi325_1
@@ -63,20 +55,20 @@ steps:
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - timeout 15m pytest -s -v -m "core_model and cache and diffusion and not distributed_cuda and L4"
 
-- label: "Diffusion Sequence Parallelism Test"
-  agent_pool: mi325_2
+- label: "Diffusion Sequence Parallelism Test (Need 4 GPUs)"
+  agent_pool: mi325_4
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
   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
 - label: "Diffusion Tensor Parallelism Test"
@@ -85,7 +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
@@ -95,32 +95,26 @@ steps:
   commands:
     - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
 
-- label: "Benchmark & Engine Test"
-  agent_pool: mi325_2
+- label: "Engine Test"
+  agent_pool: mi325_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - |
-      timeout 20m bash -c '
-        set +e
-        pytest -s -v tests/benchmarks/test_serve_cli.py
-        EXIT1=\$?
-        pytest -s -v tests/engine/test_async_omni_engine_abort.py
-        EXIT2=\$?
-        exit \$((EXIT1 | EXIT2))
-      '
+    - timeout 20m pytest -s -v tests/engine/test_async_omni_engine_abort.py
 
+# This is from test-nightly.yml.
+# This model seems to be important, so I am trying to run it here.
+# Since this is from test-nightly.yml,
+# I keep this non-blocking.
 - label: "Omni Model Test Qwen2-5-Omni"
-  agent_pool: mi325_2
+  agent_pool: mi325_4
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
-  grade: Blocking
   commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 20m pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+    - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+    - pytest -s -v tests/e2e/*/test_qwen2_5_omni_expansion.py --run-level "core_model"
 
 - label: "Omni Model Test Qwen3-Omni"
   agent_pool: mi325_2
@@ -128,70 +122,98 @@ steps:
   mirror_hardwares: [amdproduction]
   grade: Blocking
   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 "advanced_model" --run-level "advanced_model"
+    - 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 E2E Test"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 20m pytest -s -v tests/e2e/online_serving/test_qwen3_tts.py
-
-- label: "Diffusion Image Edit Test"
+- label: "Qwen3-TTS CustomVoice E2E 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
+    - |
+      timeout 20m bash -c '
+        export VLLM_LOGGING_LEVEL=DEBUG
+        export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+        pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m "advanced_model" --run-level "advanced_model"
+      '
 
-# split Bagel Model Test with H100 (Real Weights) into three tests
-- label: "Bagel Text2Img Model Test"
+- label: "Qwen3-TTS Base E2E Test"
   agent_pool: mi325_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   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 bash -c '
+        export VLLM_LOGGING_LEVEL=DEBUG
+        export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+        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: "Bagel Img2Img Model Test"
+# 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
+# - label: "Bagel Text2Img Model Test (1/3)"
+#   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
+#     - export VLLM_ROCM_USE_AITER_RMSNORM=0
+#     - 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
+#   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
+#     - export VLLM_ROCM_USE_AITER_RMSNORM=0
+#     - 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
+#   depends_on: amd-build
+#   mirror_hardwares: [amdproduction]
+#   grade: Blocking
+#   commands:
+#     - export GPU_ARCHS=gfx942
+#     - 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.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:
-    - 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 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: "Bagel Online Serving Test"
+- label: "Voxtral-TTS E2E Test"
   agent_pool: mi325_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   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 20m bash -c '
+        export VLLM_LOGGING_LEVEL=DEBUG
+        export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+      '

.buildkite/test-amd-ready.yaml

@@ -7,15 +7,33 @@ steps:
   grade: Blocking
   commands:
     - export VLLM_ROCM_USE_AITER=0
-    - "timeout 20m pytest -v -s -m 'core_model and cpu' --cov=vllm_omni --cov-branch --cov-report=term-missing --cov-report=html --cov-report=xml"
+    # ignore test_teacache_extractors.py because it use rocm gemm kernel from vLLM
+    # that is not supported on CPU
+    - "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 Model Test"
-  agent_pool: mi325_2
+- label: "Voxtral TTS CUDA Unit 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_t2i_model.py -m "core_model and diffusion" --run-level "core_model"
+    - timeout 10m pytest -s -v tests/model_executor/models/voxtral_tts/test_cuda_graph_acoustic_transformer.py
+
+- label: "Diffusion Batching Test"
+  agent_pool: mi325_1
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - 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
+  depends_on: amd-build
+  mirror_hardwares: [amdproduction]
+  grade: Blocking
+  commands:
+    - timeout 20m pytest -s -v tests/e2e/offline_inference/custom_pipeline/ -m "core_model"
 
 - label: "Diffusion Model CPU offloading Test"
   agent_pool: mi325_1
@@ -23,15 +41,14 @@ steps:
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - |
       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))
       '
@@ -46,7 +63,7 @@ steps:
 #     - export GPU_ARCHS=gfx942
 #     - export VLLM_LOGGING_LEVEL=DEBUG
 #     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-#     - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_model.py
+#     - 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: "Diffusion Cache Backend Test"
   agent_pool: mi325_1
@@ -67,7 +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
@@ -77,102 +94,134 @@ steps:
   commands:
     - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
 
-- label: "Benchmark & Engine Test"
-  agent_pool: mi325_2
+- 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
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - |
-      timeout 30m bash -c '
-        set +e
-        pytest -s -v tests/benchmarks/test_serve_cli.py
-        EXIT1=\$?
-        pytest -s -v tests/engine/test_async_omni_engine_abort.py
-        EXIT2=\$?
-        exit \$((EXIT1 | EXIT2))
+      timeout 15m bash -c '
+              pytest -s -v tests/engine/test_async_omni_engine_abort.py
       '
 
-- label: "Omni Model Test Qwen2-5-Omni"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 17m pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+# - label: "Omni Model Test Qwen3-Omni"
+#   agent_pool: mi325_2
+#   depends_on: amd-build
+#   mirror_hardwares: [amdproduction]
+#   grade: Blocking
+#   commands:
+#     - export VLLM_LOGGING_LEVEL=DEBUG
+#     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+#     - 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"
 
-- label: "Omni Model Test Qwen3-Omni"
-  agent_pool: mi325_2
+- label: "MiMo-Audio E2E Test with H100"
+  agent_pool: mi325_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
   commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - export VLLM_TEST_CLEAN_GPU_MEMORY=1
-    - timeout 10m pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
-    - timeout 10m pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_model"
+    - |
+      timeout 30m bash -c '
+        export VLLM_LOGGING_LEVEL=DEBUG
+        export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m "core_model" --run-level "core_model"
+      '
 
 - label: "Qwen3-TTS E2E Test"
-  agent_pool: mi325_2
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  grade: Blocking
-  commands:
-    - export VLLM_LOGGING_LEVEL=DEBUG
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - timeout 20m pytest -s -v tests/e2e/online_serving/test_qwen3_tts.py
-
-- 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
+    - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+    - timeout 30m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m "core_model" --run-level "core_model"
 
-- label: "Bagel Text2Img Model Test"
+- label: "Voxtral-TTS E2E Test"
   agent_pool: mi325_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   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 20m bash -c '
+        export VLLM_LOGGING_LEVEL=DEBUG
+        export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+        pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
+      '
 
-- label: "Bagel Img2Img Model Test"
+# TODO: Bagel test on ROCm is very unstable. @tjtanaa
+# Need to debug before reneable numerical changes across large PRs
+# - label: "Bagel Text2Img Model 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
+#     - export VLLM_ROCM_USE_AITER_RMSNORM=0
+#     - 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
+#   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
+#     - export VLLM_ROCM_USE_AITER_RMSNORM=0
+#     - 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
+#   depends_on: amd-build
+#   mirror_hardwares: [amdproduction]
+#   grade: Blocking
+#   commands:
+#     - export GPU_ARCHS=gfx942
+#     - 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.py -m "core_model" --run-level "core_model" -k "rocm"
+
+- label: "CosyVoice3-TTS E2E Test"
   agent_pool: mi325_1
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   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"
-
-- label: "Bagel Online Serving Test"
-  agent_pool: mi325_1
-  depends_on: amd-build
-  mirror_hardwares: [amdproduction]
-  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 20m bash -c '
+        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
@@ -88,7 +88,7 @@ steps:
     - export GPU_ARCHS=gfx942
     - export VLLM_LOGGING_LEVEL=DEBUG
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni.py
+    - pytest -s -v tests/e2e/offline_inference/test_qwen2_5_omni_expansion.py
     - pytest -s -v tests/engine/test_async_omni_engine_abort.py
 
 - label: "Omni Model Test Qwen3-Omni"
@@ -100,15 +100,14 @@ 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
+- label: "Omni Sleep Mode Test"
+  timeout_in_minutes: 40
+  agent_pool: mi325_2
   depends_on: amd-build
   mirror_hardwares: [amdproduction]
   grade: Blocking
@@ -116,4 +115,4 @@ steps:
     - 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
+    - pytest -s -v tests/entrypoints/test_omni_sleep_mode.py -m "advanced_model and omni and MI325" --run-level "advanced_model"

.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-ready.yml

@@ -1,71 +1,44 @@
+env:
+  VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
+  HF_HUB_DOWNLOAD_TIMEOUT: 300
+  HF_HUB_ETAG_TIMEOUT: 60
+
 steps:
-  - label: "Simple Unit Test"
+  - group: ":card_index_dividers: Simple Test"
     depends_on: upload-ready-pipeline
-    commands:
-      - "timeout 20m pytest -v -s -m 'core_model and cpu' --cov=vllm_omni --cov-branch --cov-report=term-missing --cov-report=html --cov-report=xml"
-    agents:
-      queue: "gpu_1_queue"
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+    steps:
+      - label: "Simple · Diffusion & Model Executor Test"
+        commands:
+          - "timeout 20m pytest -sv tests/diffusion tests/model_executor -m 'core_model and cpu' --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - label: "Voxtral TTS CUDA Unit Test"
-    timeout_in_minutes: 10
-    depends_on: upload-ready-pipeline
-    commands:
-      - "timeout 10m pytest -s -v tests/model_executor/models/voxtral_tts/test_cuda_graph_acoustic_transformer.py"
-    agents:
-      queue: "gpu_1_queue"
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Diffusion Model Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - timeout 30m pytest -s -v tests/e2e/offline_inference/test_t2i_model.py -m "core_model and diffusion" --run-level "core_model"
-    agents:
-      queue: "gpu_1_queue"
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Diffusion Batching Test"
-    depends_on: upload-ready-pipeline
-    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"
-    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:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+      - label: "Simple · Other Test"
+        commands:
+          - "timeout 20m pytest -sv -m 'core_model and cpu' --ignore=tests/diffusion --ignore=tests/model_executor --cov=vllm_omni --cov-report=term-missing:skip-covered --cov-report=xml"
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
   - label: "Custom Pipeline Test"
     depends_on: upload-ready-pipeline
@@ -83,81 +56,116 @@ steps:
           volumes:
             - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - label: "Diffusion Model CPU offloading Test"
+  - group: ":card_index_dividers: Diffusion Test"
     depends_on: upload-ready-pipeline
-    commands:
-      - timeout 10m pytest -s -v tests/e2e/offline_inference/test_diffusion_cpu_offload.py
-    agents:
-      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+    steps:
+      - label: "Diffusion · Offloader Test"
+        commands:
+          - timeout 10m pytest -s -v tests/diffusion/offloader -m "core_model and cuda" --run-level "core_model"
+        agents:
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - label: "Audio Generation Model Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - timeout 20m pytest -s -v tests/e2e/offline_inference/test_stable_audio_model.py
-    agents:
-      queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+      - label: "Diffusion · Cache Test"
+        commands:
+          - timeout 15m pytest -sv tests/diffusion/cache -m 'core_model and cuda' --run-level "core_model"
+        agents:
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - label: "Diffusion Cache Backend Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - timeout 15m pytest -s -v -m 'core_model and cache and diffusion and not distributed_cuda and L4'
-    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:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+      - label: "Diffusion · Batch Test"
+        commands:
+          - timeout 15m pytest -sv tests/diffusion/batching -m 'core_model and cuda' --run-level "core_model"
+        agents:
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - label: "Diffusion Sequence Parallelism Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - timeout 20m pytest -s -v tests/e2e/offline_inference/test_sequence_parallel.py -m core_model
-    agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
+      - label: "Diffusion · Distributed Test"
+        commands:
+          - timeout 20m pytest -sv tests/diffusion/distributed -m 'core_model and cuda' --run-level "core_model"
+        agents:
+          queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - label: "Diffusion GPU Worker Test"
+      - label: "Diffusion · Model Test"
+        commands:
+          - timeout 15m pytest -sv tests/diffusion/models/ -m 'core_model and cuda' --run-level "core_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:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "Diffusion · Other Test"
+        commands:
+          - timeout 15m pytest -sv tests/diffusion/*.py -m 'core_model and cuda' --run-level "core_model"
+        agents:
+          queue: "gpu_1_queue" # g6.4xlarge instance on AWS, has 1 L4 GPU
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+
+  - label: "Engine&Model Executor Test"
     depends_on: upload-ready-pipeline
     commands:
-      - timeout 20m pytest -s -v tests/diffusion/test_diffusion_worker.py
+      - |
+        timeout 15m bash -c "
+                  pytest -sv tests/engine/ tests/model_executor/ -m 'core_model and cuda' --run-level 'core_model'
+        "
     agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
+      queue: "gpu_1_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
     plugins:
       - docker#v5.2.0:
           image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -171,168 +179,13 @@ steps:
             - "/fsx/hf_cache:/fsx/hf_cache"
 
 
-  - label: "Engine Test"
+  - label: "Distributed Test"
     depends_on: upload-ready-pipeline
     commands:
       - |
-        timeout 15m bash -c '
-                export VLLM_WORKER_MULTIPROC_METHOD=spawn
-                pytest -s -v tests/engine/test_async_omni_engine_abort.py
-        '
-    agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-
-  - label: "Omni Model Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 17m bash -c '
-          export VLLM_LOGGING_LEVEL=DEBUG
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          pytest -s -v tests/e2e/online_serving/test_qwen2_5_omni.py -m "core_model" --run-level "core_model"
-        '
-    agents:
-      queue: "gpu_4_queue" # g6.12xlarge instance on AWS, has 4 L4 GPU
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  # - label: "Omni Model Test with H100"
-  #   depends_on: upload-ready-pipeline
-  #   commands:
-  #     - |
-  #       timeout 20m bash -c '
-  #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  #         export VLLM_TEST_CLEAN_GPU_MEMORY="1"
-  #         # - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
-  #         pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_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:$BUILDKITE_COMMIT
-  #               resources:
-  #                 limits:
-  #                   nvidia.com/gpu: 2
-  #               volumeMounts:
-  #                 - name: devshm
-  #                   mountPath: /dev/shm
-  #                 - name: hf-cache
-  #                   mountPath: /root/.cache/huggingface
-  #               env:
-  #                 - name: HF_HOME
-  #                   value: /root/.cache/huggingface
-  #           nodeSelector:
-  #             node.kubernetes.io/instance-type: gpu-h100-sxm
-  #           volumes:
-  #             - name: devshm
-  #               emptyDir:
-  #                 medium: Memory
-  #             - name: hf-cache
-  #               hostPath:
-  #                 path: /mnt/hf-cache
-  #                 type: DirectoryOrCreate
-
-  - label: "MiMo-Audio E2E Test with H100"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 30m bash -c '
-          export VLLM_LOGGING_LEVEL=DEBUG
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          pytest -s -v tests/e2e/online_serving/test_mimo_audio.py -m "core_model" --run-level "core_model"
-        '
-    agents:
-      queue: "mithril-h100-pool"
-    plugins:
-      - kubernetes:
-          podSpec:
-            containers:
-              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$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: "Qwen3-TTS E2E Test"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 20m bash -c '
-          huggingface-cli download Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
-          export VLLM_LOGGING_LEVEL=DEBUG
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          pytest -s -v tests/e2e/online_serving/test_qwen3_tts.py -k "not NoAsyncChunk"
-        '
-    agents:
-      queue: "gpu_4_queue"
-    plugins:
-      - docker#v5.2.0:
-          image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-          always-pull: true
-          propagate-environment: true
-          shm-size: "8gb"
-          environment:
-            - "HF_HOME=/fsx/hf_cache"
-            - "HF_TOKEN"
-          volumes:
-            - "/fsx/hf_cache:/fsx/hf_cache"
-
-  - label: "Voxtral-TTS E2E Test"
-    timeout_in_minutes: 20
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 20m bash -c '
-          export VLLM_LOGGING_LEVEL=DEBUG
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          pytest -s -v tests/e2e/online_serving/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-          pytest -s -v tests/e2e/offline_inference/test_voxtral_tts.py -m "advanced_model" --run-level "advanced_model"
-        '
+        timeout 15m bash -c "
+                  pytest -sv tests/distributed/ -m 'core_model and cuda' --run-level 'core_model'
+        "
     agents:
       queue: "mithril-h100-pool"
     plugins:
@@ -362,169 +215,327 @@ steps:
                   path: /mnt/hf-cache
                   type: DirectoryOrCreate
 
-  # - label: "Diffusion Image Edit Test with H100 (1 GPU)"
-  #   depends_on: upload-ready-pipeline
-  #   commands:
-  #     - |
-  #       timeout 20m bash -c '
-  #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  #         pytest -s -v tests/e2e/online_serving/test_image_gen_edit.py
-  #       '
-  #   agents:
-  #     queue: "mithril-h100-pool"
-  #   plugins:
-  #     - kubernetes:
-  #         podSpec:
-  #           containers:
-  #             - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-  #               resources:
-  #                 limits:
-  #                   nvidia.com/gpu: 1
-  #               volumeMounts:
-  #                 - name: devshm
-  #                   mountPath: /dev/shm
-  #                 - name: hf-cache
-  #                   mountPath: /root/.cache/huggingface
-  #               env:
-  #                 - name: HF_HOME
-  #                   value: /root/.cache/huggingface
-  #           nodeSelector:
-  #             node.kubernetes.io/instance-type: gpu-h100-sxm
-  #           volumes:
-  #             - name: devshm
-  #               emptyDir:
-  #                 medium: Memory
-  #             - name: hf-cache
-  #               hostPath:
-  #                 path: /mnt/hf-cache
-  #                 type: DirectoryOrCreate
-
-  - label: "Bagel Text2Img Model Test with H100"
+  - group: ":card_index_dividers: E2E Test"
     depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 30m bash -c '
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          export VLLM_TEST_CLEAN_GPU_MEMORY=1
-          pytest -s -v tests/e2e/offline_inference/test_bagel_text2img.py -m "core_model" --run-level "core_model"
-        '
-    agents:
-      queue: "mithril-h100-pool"
-    plugins:
-      - kubernetes:
-          podSpec:
-            containers:
-              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$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
+    steps:
+      - label: "RLHF · VeRL-Omni E2E Test"
+        commands:
+          - |
+            timeout 20m bash -c "
+              pytest -s -v tests/e2e/offline_inference/rlhf_test/test_verl_omni_e2e.py
+            "
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
 
-  - label: "Bagel Img2Img Model Test with H100"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 30m bash -c '
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          export VLLM_TEST_CLEAN_GPU_MEMORY=1
-          pytest -s -v tests/e2e/offline_inference/test_bagel_img2img.py -m "core_model" --run-level "core_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:$BUILDKITE_COMMIT
-                resources:
-                  limits:
-                    nvidia.com/gpu: 1
-                volumeMounts:
+      - label: "Omni · Qwen3-Omni Test"
+        commands:
+          - |
+            timeout 20m bash -c "
+              pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m 'core_model' --run-level 'core_model'
+            "
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$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
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
                   - name: devshm
-                    mountPath: /dev/shm
+                    emptyDir:
+                      medium: Memory
                   - 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
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
 
-  - label: "Bagel Online Serving Test with H100"
-    depends_on: upload-ready-pipeline
-    commands:
-      - |
-        timeout 40m bash -c '
-          export VLLM_WORKER_MULTIPROC_METHOD=spawn
-          export VLLM_TEST_CLEAN_GPU_MEMORY=1
-          export VLLM_IMAGE_FETCH_TIMEOUT=60
-          pytest -s -v tests/e2e/online_serving/test_bagel_online.py -m "core_model" --run-level "core_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:$BUILDKITE_COMMIT
-                resources:
-                  limits:
-                    nvidia.com/gpu: 1
-                volumeMounts:
+      - label: "Omni · VoxCPM2 Test"
+        timeout_in_minutes: 20
+        commands:
+          - |
+            timeout 20m bash -c "
+              export VLLM_LOGGING_LEVEL=DEBUG
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              pytest -s -v tests/e2e/offline_inference/test_voxcpm2.py -m 'core_model' --run-level 'core_model'
+            "
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "TTS · Qwen3-TTS CustomVoice Test"
+        commands:
+          - |
+            timeout 20m bash -c "
+              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'
+            "
+        agents:
+          queue: "gpu_4_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "TTS · Qwen3-TTS Base Test"
+        commands:
+          - |
+            timeout 20m bash -c "
+              export VLLM_WORKER_MULTIPROC_METHOD=spawn
+              export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
+              export VLLM_OMNI_USE_V2_RUNNER=1
+              pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m 'core_model' --run-level 'core_model'
+            "
+        agents:
+          queue: "gpu_4_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"
+
+      - label: "Diffusion · Qwen Image Edit Test"
+        commands:
+          - |
+            timeout 20m bash -c "
+              pytest -s -v tests/e2e/online_serving/test_qwen_image_edit.py -m 'core_model' --run-level 'core_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
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
                   - name: devshm
-                    mountPath: /dev/shm
+                    emptyDir:
+                      medium: Memory
                   - 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
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      - label: "Diffusion · Bagel Test"
+        commands:
+          - |
+            timeout 40m bash -c "
+              export VLLM_IMAGE_FETCH_TIMEOUT=60
+              pytest -s -v tests/e2e/online_serving/test_bagel.py -m 'core_model' --run-level 'core_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: "Diffusion · Wan22 Test"
+        commands:
+          - |
+            timeout 40m bash -c "
+              export VLLM_IMAGE_FETCH_TIMEOUT=60
+              pytest -s -v tests/e2e/online_serving/test_wan22_t2v.py -m 'core_model' --run-level 'core_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:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 1
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+
+      - label: "Diffusion · Qwen Image Test"
+        commands:
+          - |
+            timeout 40m bash -c "
+              export VLLM_IMAGE_FETCH_TIMEOUT=60
+              pytest -s -v tests/e2e/online_serving/test_qwen_image.py -m 'core_model' --run-level 'core_model'
+            "
+        agents:
+          queue: "mithril-h100-pool"
+        plugins:
+          - kubernetes:
+              podSpec:
+                containers:
+                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$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: "Diffusion · Qwen Image Layered Test"
+        commands:
+          - |
+            timeout 40m bash -c "
+              export VLLM_IMAGE_FETCH_TIMEOUT=60
+              pytest -s -v tests/e2e/online_serving/test_qwen_image_layered.py -m 'core_model' --run-level 'core_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

.buildkite/test-template-amd-omni.j2

@@ -3,7 +3,7 @@
    Last synced: 2025-12-15
    Modifications: Removed unused CUDA/NVIDIA logic, keeping only AMD tests
 #}
-{% set docker_image_amd = "public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT-rocm-omni" %}
+{% set docker_image_amd = "rocm/vllm-omni:$BUILDKITE_COMMIT" %}
 {% set default_working_dir = "/app/vllm-omni" %}
 
   - group: "AMD Tests"
@@ -13,7 +13,6 @@
         depends_on: ~
         soft_fail: false
         commands:
-          - "source .buildkite/scripts/docker_login_ecr_public.sh && safe_docker_login_ecr_public"
           - "docker build -f docker/Dockerfile.rocm -t {{ docker_image_amd }} --target test --build-arg ARG_PYTORCH_ROCM_ARCH=gfx942 --progress plain ."
           - "docker push {{ docker_image_amd }}"
         key: "amd-build"
@@ -22,13 +21,15 @@
         retry:
           automatic:
             - exit_status: -1  # Agent was lost
-              limit: 1
+              limit: 2
             - exit_status: -10  # Agent was lost
-              limit: 1
+              limit: 2
+            - exit_status: 128  # Git connectivity issues
+              limit: 2
             - exit_status: 1  # Machine occasionally fail
               limit: 1
         agents:
-          queue: cpu_queue_premerge
+          queue: amd-cpu
 
     {% for step in steps %}
     {% if step.mirror_hardwares and mirror_hw in step.mirror_hardwares %}
@@ -47,6 +48,9 @@
           DOCKER_BUILDKIT: "1"
           TEST_COMMAND: |-
             (command rocm-smi || true) && cd {{ (step.working_dir or default_working_dir) | safe }}
+{% if "mi250" in step.agent_pool %}
+            python3 -m pip uninstall -y amd-aiter
+{% endif %}
             {{ indented_cmd | safe }}
         priority: 100
         {% if step.grade and step.grade == "Blocking" %}

.buildkite/test-weekly.yml

@@ -0,0 +1,148 @@
+env:
+  VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
+  HF_HUB_DOWNLOAD_TIMEOUT: 300
+  HF_HUB_ETAG_TIMEOUT: 60
+
+steps:
+  - label: "Reliability Test - qwen3-omni"
+    timeout_in_minutes: 180
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    commands:
+      - pytest -s -v tests/dfx/reliability/test_reliability_qwen3_omni.py -m "slow"
+    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:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+                  - name: HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
+  - label: "Reliability Test - wan22"
+    timeout_in_minutes: 180
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    commands:
+      - pytest -s -v tests/dfx/reliability/test_reliability_wan22.py -m "slow"
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$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
+
+  # Invalid HTTP / WS payloads vs live Omni servers (Issue #3649 tracks known skips).
+  - group: "Reliability Test - Invalid parameters Test"
+    depends_on: upload-weekly-pipeline
+    if: build.env("WEEKLY") == "1" || build.pull_request.labels includes "weekly-test"
+    steps:
+      - label: "Invalid parameters Test · H100"
+        timeout_in_minutes: 120
+        commands:
+          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and H100"
+        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:$BUILDKITE_COMMIT
+                    resources:
+                      limits:
+                        nvidia.com/gpu: 2
+                    volumeMounts:
+                      - name: devshm
+                        mountPath: /dev/shm
+                      - name: hf-cache
+                        mountPath: /root/.cache/huggingface
+                    env:
+                      - name: HF_HOME
+                        value: /root/.cache/huggingface
+                      - name: HF_TOKEN
+                        valueFrom:
+                          secretKeyRef:
+                            name: hf-token-secret
+                            key: token
+                nodeSelector:
+                  node.kubernetes.io/instance-type: gpu-h100-sxm
+                volumes:
+                  - name: devshm
+                    emptyDir:
+                      medium: Memory
+                  - name: hf-cache
+                    hostPath:
+                      path: /mnt/hf-cache
+                      type: DirectoryOrCreate
+
+      # Stable Audio Open, Qwen3-TTS speech/batch/voices paths marked ``L4``.
+      - label: "Invalid parameters Test · L4"
+        timeout_in_minutes: 120
+        commands:
+          - pytest -s -v tests/dfx/reliability/invalid_param_test/ -m "slow and L4"
+        agents:
+          queue: "gpu_1_queue"
+        plugins:
+          - docker#v5.2.0:
+              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+              always-pull: true
+              propagate-environment: true
+              shm-size: "8gb"
+              environment:
+                - "HF_HOME=/fsx/hf_cache"
+                - "HF_TOKEN"
+              volumes:
+                - "/fsx/hf_cache:/fsx/hf_cache"

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

@@ -0,0 +1,575 @@
+---
+name: add-diffusion-model
+description: Add a new diffusion model (text-to-image, text-to-video, image-to-video, text-to-audio, image editing) to vLLM-Omni, including Cache-DiT acceleration and parallelism support (TP, SP/USP, CFG-Parallel, HSDP). Use when integrating a new diffusion model, porting a diffusers pipeline or a custom model repo to vllm-omni, creating a new DiT transformer adapter, adding diffusion model support, or enabling multi-GPU parallelism and cache acceleration for an existing model.
+---
+
+# Adding a Diffusion Model to vLLM-Omni
+
+## Overview
+
+This skill guides you through adding a new diffusion model to vLLM-Omni. The model may come from HuggingFace Diffusers (structured pipeline) or from a private/custom repo. The workflow differs significantly depending on the source.
+
+## Prerequisites
+
+Before starting, determine:
+
+1. **Model category**: Text-to-Image, Text-to-Video, Image-to-Video, Image Editing, Text-to-Audio, or Omni
+2. **Reference source**: Diffusers pipeline, custom repo, or a combination
+3. **Model HuggingFace ID** or local checkpoint path
+4. **Architecture**: Scheduler, text encoder, VAE, transformer/backbone
+
+## Step 0: Classify the Migration Path
+
+Check the model's HF repo for `model_index.json`. This determines your path:
+
+| Scenario | How to identify | Migration path |
+|----------|----------------|----------------|
+| **Already supported** | `_class_name` in `model_index.json` matches a key in `_DIFFUSION_MODELS` in `registry.py` | Skip to Step 5 (test) and Step 7 (docs) |
+| **Diffusers-based** | Has standard `model_index.json` with `_diffusers_version`, subfolders for `transformer/`, `vae/`, etc. | Follow **Path A** below |
+| **Custom/private repo** | No diffusers `model_index.json`, weights in non-standard format, custom model code in a separate git repo | Follow **Path B** below |
+| **Hybrid** | Has some diffusers components (VAE) but custom transformer/fusion | Mix of Path A and Path B |
+
+## Path A: Diffusers-Based Model
+
+For models with a standard diffusers layout. See [references/transformer-adaptation.md](references/transformer-adaptation.md) for detailed code patterns.
+
+### A1. Analyze `model_index.json`
+
+Identify components: `transformer`, `scheduler`, `vae`, `text_encoder`, `tokenizer`.
+
+### A2. Create model directory
+
+```
+vllm_omni/diffusion/models/your_model_name/
+├── __init__.py
+├── pipeline_your_model.py
+└── your_model_transformer.py
+```
+
+### A3. Adapt transformer
+
+1. Copy from diffusers source. Remove mixins (`ModelMixin`, `ConfigMixin`, `AttentionModuleMixin`).
+2. Replace attention with `vllm_omni.diffusion.attention.layer.Attention` (QKV shape: `[B, seq, heads, head_dim]`).
+3. Add `od_config: OmniDiffusionConfig | None = None` to `__init__`.
+4. Add `load_weights()` method mapping diffusers weight names to vllm-omni names.
+5. Add class attributes: `_repeated_blocks`, `_layerwise_offload_blocks_attr`.
+
+### A4. Adapt pipeline
+
+Inherit from `nn.Module`. The key contract:
+
+```python
+class YourPipeline(nn.Module):
+    def __init__(self, *, od_config: OmniDiffusionConfig, prefix: str = ""):
+        # Load VAE, text encoder, tokenizer via from_pretrained()
+        # Instantiate transformer (weights loaded later via weights_sources)
+        self.weights_sources = [
+            DiffusersPipelineLoader.ComponentSource(
+                model_or_path=od_config.model, subfolder="transformer",
+                prefix="transformer.", fall_back_to_pt=True)]
+
+    def forward(self, req: OmniDiffusionRequest) -> DiffusionOutput:
+        # Encode prompt → prepare latents → denoise loop → VAE decode
+        return DiffusionOutput(output=output)
+
+    def load_weights(self, weights):
+        return AutoWeightsLoader(self).load_weights(weights)
+```
+
+Add post/pre-process functions in the same pipeline file. Register them in `registry.py`.
+
+### A5. Register, test, docs → continue at Step 4 below.
+
+---
+
+## Path B: Custom/Private Repo Model
+
+For models without a diffusers pipeline — weights in custom format, model code in a private repo. Real examples: DreamID-Omni, BAGEL, HunyuanImage3.
+
+### B1. Understand the reference repo
+
+Study the original model's code to identify:
+- **Model architecture files** (transformers, fusion modules, embeddings)
+- **Weight format** (safetensors, `.pth`, custom checkpoint structure)
+- **Weight loading helpers** (custom init functions, checkpoint loaders)
+- **Pre/post-processing** (image/audio transforms, tokenization, VAE encode/decode)
+- **External dependencies** (packages not on PyPI)
+- **Config format** (JSON config files, hardcoded dicts)
+
+### B2. Decide what lives WHERE
+
+This is the key design decision for custom models. Follow these placement rules:
+
+| Code type | Where to place | Example |
+|-----------|---------------|---------|
+| **Pipeline orchestration** (init, forward, denoise loop) | `vllm_omni/diffusion/models/<name>/pipeline_<name>.py` | Always required |
+| **Custom transformer/backbone** (ported and adapted to vllm-omni) | `vllm_omni/diffusion/models/<name>/<name>_transformer.py` or similar | `wan2_2.py`, `fusion.py`, `bagel_transformer.py` |
+| **Custom sub-models** (VAE, fusion, autoencoder) | `vllm_omni/diffusion/models/<name>/` as separate files | `autoencoder.py`, `fusion.py` |
+| **External dependency code** (original repo utilities) | **External repo**, installed via download script or pip | `dreamid_omni` package via git clone |
+| **Hardcoded model configs** | Module-level dicts in pipeline file | `VIDEO_CONFIG`, `AUDIO_CONFIG` dicts |
+| **Download/setup script** | `examples/offline_inference/<name>/download_<name>.py` | `download_dreamid_omni.py` |
+| **Custom `model_index.json`** | Generated by download script, placed at model root | Minimal: `{"_class_name": "YourPipeline", ...}` |
+
+### B3. Handle external dependencies
+
+If the model's code lives in a separate git repo:
+
+**Option 1: Import with graceful fallback** (recommended for models with external utils)
+
+```python
+try:
+    from external_model.utils import init_vae, load_checkpoint
+except ImportError:
+    raise ImportError(
+        "Failed to import from dependency 'external_model'. "
+        "Please run the download script first."
+    )
+```
+
+**Option 2: Port the code directly** (preferred when feasible)
+
+Copy the essential model files into `vllm_omni/diffusion/models/<name>/` and adapt them. This avoids external dependencies. BAGEL does this — `autoencoder.py` and `bagel_transformer.py` are ported directly.
+
+**Decision criteria**: Port if the code is self-contained and won't diverge. Use external deps if the model repo is actively maintained and the code is complex.
+
+### B4. Handle custom weight loading
+
+Custom models have two patterns for weight loading:
+
+**Pattern 1: Bypass standard loader** (DreamID-Omni style)
+
+When the original model has complex custom init functions that load weights in `__init__`:
+
+```python
+class CustomPipeline(nn.Module):
+    def __init__(self, *, od_config, prefix=""):
+        super().__init__()
+        model = od_config.model
+        # Load everything eagerly in __init__ using custom helpers
+        self.vae = custom_init_vae(model, device=self.device)
+        self.text_encoder = custom_init_text_encoder(model, device=self.device)
+        self.transformer = CustomFusionModel(CONFIG)
+        load_custom_checkpoint(self.transformer,
+            checkpoint_path=os.path.join(model, "model.safetensors"))
+        # NO weights_sources defined — bypasses standard loader
+
+    def load_weights(self, weights):
+        pass  # No-op — all weights loaded in __init__
+```
+
+**Pattern 2: Use standard loader with custom `load_weights`** (BAGEL style)
+
+When weights are in safetensors format but need name remapping:
+
+```python
+class CustomPipeline(nn.Module):
+    def __init__(self, *, od_config, prefix=""):
+        super().__init__()
+        # Instantiate model architecture without weights
+        self.bagel = BagelModel(config)
+        self.vae = AutoEncoder(ae_params)
+
+        # Point loader at the safetensors in the model root
+        self.weights_sources = [
+            DiffusersPipelineLoader.ComponentSource(
+                model_or_path=od_config.model,
+                subfolder=None,  # weights at root, not in subfolder
+                prefix="",
+                fall_back_to_pt=False,
+            )
+        ]
+
+    def load_weights(self, weights):
+        # Custom name remapping for non-diffusers weight names
+        params = dict(self.named_parameters())
+        loaded = set()
+        for name, tensor in weights:
+            # Remap original weight names to vllm-omni module names
+            name = self._remap_weight_name(name)
+            if name in params:
+                default_weight_loader(params[name], tensor)
+                loaded.add(name)
+        return loaded
+```
+
+### B5. Create the `model_index.json`
+
+Custom models need a `model_index.json` at the model root for vllm-omni to discover them. For custom models, this is minimal:
+
+```json
+{
+    "_class_name": "YourModelPipeline",
+    "custom_key": "path/to/custom_weights.safetensors"
+}
+```
+
+The `_class_name` must match a key in `_DIFFUSION_MODELS` in `registry.py`. Additional keys are model-specific (accessed via `od_config.model_config`).
+
+If the model's weights come from multiple HF repos, write a **download script** that:
+1. Downloads from each repo
+2. Assembles into a single directory
+3. Generates `model_index.json`
+4. Installs any external dependencies (git clone + `.pth` file)
+
+Place at: `examples/offline_inference/<name>/download_<name>.py`
+
+### B6. Handle multi-modal inputs
+
+If the model accepts images, audio, or other multi-modal inputs, implement the protocol classes from `vllm_omni/diffusion/models/interface.py`:
+
+```python
+from vllm_omni.diffusion.models.interface import SupportImageInput, SupportAudioInput
+
+class MyPipeline(nn.Module, SupportImageInput, SupportAudioInput):
+    # Protocol markers — the engine uses these to enable proper input routing
+    pass
+```
+
+Preprocessing for custom models is typically done **inside `forward()`** rather than via registered pre-process functions, since the logic is often tightly coupled to the model.
+
+### B7. Continue at Step 4 below.
+
+---
+
+## Common Steps (Both Paths)
+
+### Step 4: Register Model in registry.py
+
+Edit `vllm_omni/diffusion/registry.py`:
+
+```python
+_DIFFUSION_MODELS = {
+    "YourModelPipeline": ("your_model_name", "pipeline_your_model", "YourModelPipeline"),
+}
+_DIFFUSION_POST_PROCESS_FUNCS = {
+    "YourModelPipeline": "get_your_model_post_process_func",  # if applicable
+}
+_DIFFUSION_PRE_PROCESS_FUNCS = {
+    "YourModelPipeline": "get_your_model_pre_process_func",  # if applicable
+}
+```
+
+The registry key is the `_class_name` from `model_index.json`. The tuple is `(folder_name, module_file, class_name)`.
+
+Create `__init__.py` exporting the pipeline class and any factory functions.
+
+### Step 5: Run, Test, Debug
+
+Use the appropriate existing example script:
+
+| Category | Script |
+|----------|--------|
+| Text-to-Image | `examples/offline_inference/text_to_image/text_to_image.py` |
+| Text-to-Video | `examples/offline_inference/text_to_video/text_to_video.py` |
+| Image-to-Video | `examples/offline_inference/image_to_video/image_to_video.py` |
+| Image-to-Image | `examples/offline_inference/image_to_image/image_edit.py` |
+| Text-to-Audio | `examples/offline_inference/text_to_audio/text_to_audio.py` |
+
+For custom/Omni models that don't fit these categories, create a dedicated example script.
+
+**Validation**: No errors, output is meaningful, quality matches reference implementation.
+
+See [references/troubleshooting.md](references/troubleshooting.md) for common errors.
+
+### Step 6: Add Example Scripts
+
+For Omni or custom models, create:
+- `examples/offline_inference/your_model_name/` — offline script + README
+- `examples/online_serving/your_model_name/` — server script + client
+- Download script if weights require assembly from multiple sources
+
+### Step 7: Update Documentation
+
+Required updates:
+1. `docs/user_guide/diffusion/parallelism_acceleration.md` — parallelism support table
+2. `docs/user_guide/diffusion/cpu_offload_diffusion.md` — if CPU offload supported (add to supported models table)
+3. `docs/user_guide/diffusion/teacache.md` — if TeaCache supported
+4. `docs/user_guide/diffusion/cache_dit_acceleration.md` — if Cache-DiT supported
+5. `examples/offline_inference/xxx/README.md` — offline example docs
+6. `examples/online_serve/xxx/README.md` — online serve docs
+
+### Step 8: Add E2E Tests (Recommended)
+
+Create `tests/e2e/online_serving/test_your_model_expansion.py`.
+
+### Step 9: Add Cache-DiT Acceleration
+
+Cache-DiT accelerates inference by caching intermediate computation results across denoising steps. After your model is working correctly on a single GPU, add cache-dit support.
+
+See [references/cache-dit-patterns.md](references/cache-dit-patterns.md) for detailed code patterns.
+
+#### 9a. Determine your model type
+
+| Model Type | Description | Action |
+|------------|-------------|--------|
+| **Standard single-transformer** | One transformer with one `ModuleList` of blocks | No code needed — `CacheDiTBackend` auto-detects via `enable_cache_for_dit()` |
+| **Multi-block-list** | One transformer with multiple block lists (e.g., `transformer_blocks` + `single_transformer_blocks`) | Write custom enabler with `BlockAdapter` |
+| **Dual-transformer** | Two transformers (e.g., high-noise + low-noise) | Write custom enabler with `BlockAdapter` wrapping both |
+
+#### 9b. Standard models — verify automatic support
+
+For standard single-transformer models, test directly:
+
+```python
+omni = Omni(
+    model="your-model-name",
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4,
+    }
+)
+```
+
+Check logs for "Cache-dit enabled successfully on xxx". If it works, skip to Step 9e.
+
+#### 9c. Custom architectures — write a custom enabler
+
+For multi-block-list or dual-transformer models, write a custom enabler function:
+
+```python
+from cache_dit import BlockAdapter, ForwardPattern, ParamsModifier, DBCacheConfig
+
+def enable_cache_for_your_model(pipeline, cache_config):
+    db_cache_config = DBCacheConfig(
+        num_inference_steps=None,
+        Fn_compute_blocks=cache_config.Fn_compute_blocks,
+        Bn_compute_blocks=cache_config.Bn_compute_blocks,
+        max_warmup_steps=cache_config.max_warmup_steps,
+        max_cached_steps=cache_config.max_cached_steps,
+        max_continuous_cached_steps=cache_config.max_continuous_cached_steps,
+        residual_diff_threshold=cache_config.residual_diff_threshold,
+    )
+
+    cache_dit.enable_cache(
+        BlockAdapter(
+            transformer=pipeline.transformer,
+            blocks=[
+                pipeline.transformer.transformer_blocks,
+                pipeline.transformer.single_transformer_blocks,
+            ],
+            forward_pattern=[ForwardPattern.Pattern_1, ForwardPattern.Pattern_1],
+            params_modifiers=[ParamsModifier(...)],
+        ),
+        cache_config=db_cache_config,
+    )
+
+    def refresh_cache_context(pipeline, num_inference_steps, verbose=True):
+        cache_dit.refresh_context(
+            pipeline.transformer, num_inference_steps=num_inference_steps, verbose=verbose
+        )
+    return refresh_cache_context
+```
+
+#### 9d. Register the custom enabler
+
+Add your enabler to `CUSTOM_DIT_ENABLERS` in `vllm_omni/diffusion/cache/cache_dit_backend.py`:
+
+```python
+CUSTOM_DIT_ENABLERS = {
+    "Wan22Pipeline": enable_cache_for_wan22,
+    "LongCatImagePipeline": enable_cache_for_longcat_image,
+    "YourModelPipeline": enable_cache_for_your_model,  # Add here
+}
+```
+
+#### 9e. Test Cache-DiT
+
+```python
+omni = Omni(
+    model="your-model-name",
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1, "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4, "residual_diff_threshold": 0.24,
+    }
+)
+images = omni.generate("a beautiful landscape",
+    OmniDiffusionSamplingParams(num_inference_steps=50))
+```
+
+**Verify**: 1) logs show cache enabled, 2) 1.5-2x speedup, 3) output quality acceptable vs baseline.
+
+If quality degrades, lower `residual_diff_threshold` (try 0.12-0.18) or increase `max_warmup_steps` (try 6-8).
+
+---
+
+### Step 10: Add Parallelism Support
+
+After the model works on a single GPU, add multi-GPU parallelism. Add each type incrementally, testing after each addition.
+
+See [references/parallelism-patterns.md](references/parallelism-patterns.md) for detailed code patterns and API reference.
+
+**Recommended order**: TP → SP/USP → CFG Parallel → HSDP
+
+#### 10a. Tensor Parallelism (TP)
+
+Shards DiT linear layers across GPUs. Requires code changes in the transformer.
+
+**What to change in the transformer**:
+1. Replace `nn.Linear` with `ColumnParallelLinear` / `RowParallelLinear` / `QKVParallelLinear`
+2. Update `load_weights()` to handle QKV fusion with `stacked_params_mapping`
+3. Use `self.to_qkv.num_heads` (local heads) instead of total heads for split sizes
+
+```python
+from vllm.model_executor.layers.linear import (
+    QKVParallelLinear, RowParallelLinear, ColumnParallelLinear,
+)
+
+# Attention: QKV → RowParallel output
+self.to_qkv = QKVParallelLinear(dim, head_dim, num_heads, num_kv_heads)
+self.to_out = RowParallelLinear(dim, dim, input_is_parallel=True)
+
+# FFN: ColumnParallel → RowParallel
+self.w1 = ColumnParallelLinear(dim, ffn_dim)
+self.w2 = RowParallelLinear(ffn_dim, dim, input_is_parallel=True)
+```
+
+**Constraints**: `num_heads % tp_size == 0` and `num_kv_heads % tp_size == 0`.
+
+**Test**: `--tensor-parallel-size 2`
+
+#### 10b. Sequence Parallelism (SP / USP)
+
+Splits sequence tokens across GPUs. Non-intrusive via `_sp_plan` on the transformer class — no changes to `forward()`.
+
+**What to change in the transformer**:
+
+Add `_sp_plan` class attribute:
+
+```python
+from vllm_omni.diffusion.distributed.sp_plan import (
+    SequenceParallelInput, SequenceParallelOutput,
+)
+
+class YourTransformer(nn.Module):
+    _sp_plan = {
+        "blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+If inline tensor ops (e.g., `torch.cat`) exist between shard/gather points, extract them into `nn.Module` submodules so hooks can intercept them.
+
+For RoPE that needs splitting, add an entry for the RoPE module with `split_output=True`.
+
+**Test**: `--ulysses-degree 2` (offline) or `--usp 2` (online serving)
+
+#### 10c. CFG Parallel
+
+Distributes positive/negative CFG branches across 2 GPUs. Requires the pipeline to inherit `CFGParallelMixin`.
+
+**What to change in the pipeline**:
+
+```python
+from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
+
+class YourPipeline(nn.Module, CFGParallelMixin):
+    def diffuse(self, ...) -> torch.Tensor:
+        for i, t in enumerate(timesteps):
+            positive_kwargs = {...}
+            negative_kwargs = {...} if do_true_cfg else None
+            noise_pred = self.predict_noise_maybe_with_cfg(
+                do_true_cfg=do_true_cfg, true_cfg_scale=cfg_scale,
+                positive_kwargs=positive_kwargs, negative_kwargs=negative_kwargs,
+            )
+            latents = self.scheduler_step_maybe_with_cfg(
+                noise_pred, t, latents, do_true_cfg
+            )
+        return latents
+```
+
+Override `predict_noise()` if your transformer call is non-standard. Override `combine_cfg_noise()` for multi-output models (e.g., video + audio).
+
+**Constraint**: Exactly 2 GPUs. Only for models using classifier-free guidance.
+
+**Test**: `--cfg-parallel-size 2`
+
+#### 10d. HSDP (Hybrid Sharded Data Parallel)
+
+Shards transformer weights via PyTorch FSDP2 to reduce per-GPU VRAM. No code changes to the forward pass — just add a class attribute.
+
+**What to change in the transformer**:
+
+```python
+class YourTransformer(nn.Module):
+    @staticmethod
+    def _is_transformer_block(name: str, module) -> bool:
+        return "blocks" in name and name.split(".")[-1].isdigit()
+
+    _hsdp_shard_conditions = [_is_transformer_block]
+```
+
+**Constraint**: Cannot combine with TP. For standalone HSDP, set `hsdp_shard_size` explicitly.
+
+**Test**: `--use-hsdp` or `DiffusionParallelConfig(use_hsdp=True)`
+
+#### 10e. Update parallelism documentation
+
+After adding parallelism support, update:
+1. `docs/user_guide/diffusion/parallelism_acceleration.md` — add your model to the support table
+2. Record which parallelism methods are supported (USP, Ring, CFG, TP, HSDP, VAE-Patch)
+
+### Step 11: Add CPU Offload Support
+
+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 SupportsComponentDiscovery
+
+class YourPipeline(nn.Module, SupportsComponentDiscovery):
+    _dit_modules: ClassVar[list[str]] = ["transformer"]
+    _encoder_modules: ClassVar[list[str]] = ["text_encoder"]
+    _vae_modules: ClassVar[list[str]] = ["vae"]
+    _resident_modules: ClassVar[list[str]] = []  # optional
+```
+
+- `_dit_modules`: denoising submodules (kept on GPU during diffusion loop)
+- `_encoder_modules`: encoder/vision submodules (offloaded to CPU during diffusion loop)
+- `_vae_modules`: VAE(s) (handled by both sequential and layerwise backends)
+- `_resident_modules`: additional modules to pin on GPU during layerwise
+  offloading (e.g. embedders, connectors). Only used by the layerwise
+  backend. Optional — defaults to `[]`.
+
+All attribute names support dotted paths for nested submodules
+(e.g. `"pipe.transformer"`, `"bagel.time_embedder"`).
+
+Pipelines without `SupportsComponentDiscovery` fall back to scanning
+well-known attribute names (`transformer`, `text_encoder`, `vae`,
+etc.), which fails for non-standard names.
+
+### Step 12: Performance Profiling
+
+After verifying correctness and implementing parallelism/caching, profile the model's performance to identify bottlenecks and ensure optimal execution.
+
+See the [Profiling Single-Stage Diffusion](../../../docs/contributing/profiling.md#3-profiling-single-stage-diffusion) guide for detailed instructions on:
+1. Using the PyTorch profiler (`profiler: "torch"`) to capture detailed CPU/CUDA traces.
+2. Using Nsight Systems (`nsys`) with `profiler: "cuda"` for low-overhead CUDA traces.
+3. Controlling profiling via `omni.start_profile()` and `omni.stop_profile()`.
+
+---
+
+## Iterative Development Tips
+
+1. **Start minimal**: Basic generation first, no parallelism/caching
+2. **Use `--enforce-eager`**: Disable torch.compile during debugging
+3. **Use small models**: Test with smaller variants first
+4. **Check tensor shapes**: Most errors are reshape mismatches in attention
+5. **Add features incrementally**: Single GPU → TP → SP → CFG → HSDP → Cache-DiT
+6. **For custom models**: Get the model running with the original code first, then progressively replace components with vllm-omni equivalents
+7. **Cache-DiT before parallelism tuning**: Cache-DiT is lossy — verify quality at baseline before combining with parallelism
+8. **Combine lossless + lossy**: e.g., TP + SP + Cache-DiT for maximum throughput
+
+## Reference Files
+
+- [Transformer Adaptation](references/transformer-adaptation.md) — porting transformers from diffusers
+- [Custom Model Patterns](references/custom-model-patterns.md) — patterns for non-diffusers models
+- [Parallelism Patterns](references/parallelism-patterns.md) — TP, SP/USP, CFG parallel, HSDP implementation details
+- [Cache-DiT Patterns](references/cache-dit-patterns.md) — cache-dit acceleration for standard and custom architectures
+- [Troubleshooting](references/troubleshooting.md) — common errors and fixes

.claude/skills/add-diffusion-model/references/cache-dit-patterns.md

@@ -0,0 +1,254 @@
+# Cache-DiT Patterns Reference
+
+## Overview
+
+Cache-DiT accelerates Diffusion Transformers by caching intermediate computation results across denoising steps. Adjacent steps produce similar features, so redundant computations can be skipped.
+
+Three caching strategies:
+- **DBCache**: Dynamic block-level caching — selectively computes or caches transformer blocks based on residual differences
+- **TaylorSeer**: Calibration-based prediction using Taylor expansion to estimate block outputs
+- **SCM** (Step Computation Masking): Dynamic step skipping based on configurable policies
+
+**Typical speedup**: 1.5-2.5x depending on model and configuration.
+
+**Official docs**: https://docs.vllm.ai/projects/vllm-omni/en/latest/design/feature/cache_dit
+
+## Architecture
+
+vLLM-Omni integrates cache-dit through `CacheDiTBackend`:
+
+| Component | Purpose |
+|-----------|---------|
+| `CacheDiTBackend` | Unified backend — auto-selects enabler (standard or custom) |
+| `enable_cache_for_dit()` | Default enabler for standard single-transformer models |
+| `CUSTOM_DIT_ENABLERS` dict | Registry of custom enablers keyed by pipeline class name |
+| `BlockAdapter` | Wraps complex architectures (multi-block-list or multi-transformer) |
+| `ForwardPattern` | Specifies block forward signature: `Pattern_0`, `Pattern_1`, `Pattern_2` |
+| `ParamsModifier` | Per-transformer or per-block-list config customization |
+| `DBCacheConfig` | Configuration for DBCache parameters |
+| `cache_dit.refresh_context()` | Updates cache context when `num_inference_steps` changes |
+
+**Source files**:
+- `vllm_omni/diffusion/cache/cache_dit_backend.py` — `CacheDiTBackend`, enablers, `CUSTOM_DIT_ENABLERS`
+- `vllm_omni/diffusion/cache/` — cache backend implementations
+
+## Standard Models: Automatic Support
+
+Most DiT models follow this pattern:
+- Single transformer with one `nn.ModuleList` of blocks
+- Standard forward signature
+- Compatible with cache-dit's automatic detection
+
+**Examples**: Qwen-Image, Z-Image, FLUX
+
+No code changes needed. `CacheDiTBackend` automatically uses `enable_cache_for_dit()`:
+
+```python
+from vllm_omni import Omni
+
+omni = Omni(
+    model="Qwen/Qwen-Image",
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4,
+    }
+)
+```
+
+What happens automatically:
+
+```python
+def enable_cache_for_dit(pipeline, cache_config):
+    db_cache_config = DBCacheConfig(
+        num_inference_steps=None,
+        Fn_compute_blocks=cache_config.Fn_compute_blocks,
+        Bn_compute_blocks=cache_config.Bn_compute_blocks,
+        max_warmup_steps=cache_config.max_warmup_steps,
+        max_cached_steps=cache_config.max_cached_steps,
+        max_continuous_cached_steps=cache_config.max_continuous_cached_steps,
+        residual_diff_threshold=cache_config.residual_diff_threshold,
+    )
+
+    cache_dit.enable_cache(pipeline.transformer, cache_config=db_cache_config)
+
+    def refresh_cache_context(pipeline, num_inference_steps, verbose=True):
+        cache_dit.refresh_context(
+            pipeline.transformer, num_inference_steps=num_inference_steps, verbose=verbose
+        )
+    return refresh_cache_context
+```
+
+## Custom Architectures: Writing Custom Enablers
+
+### When you need a custom enabler
+
+- Model has multiple block lists in one transformer (e.g., `transformer_blocks` + `single_transformer_blocks`)
+- Model has two transformers (e.g., high-noise + low-noise like Wan2.2)
+- Model uses non-standard block forward signature
+
+### Pattern 1: Multi-Block-List (LongCat-Image style)
+
+Single transformer with two block lists:
+
+```python
+import cache_dit
+from cache_dit import BlockAdapter, ForwardPattern, ParamsModifier, DBCacheConfig
+
+def enable_cache_for_your_model(pipeline, cache_config):
+    db_cache_config = DBCacheConfig(
+        num_inference_steps=None,
+        Fn_compute_blocks=cache_config.Fn_compute_blocks,
+        Bn_compute_blocks=cache_config.Bn_compute_blocks,
+        max_warmup_steps=cache_config.max_warmup_steps,
+        max_cached_steps=cache_config.max_cached_steps,
+        max_continuous_cached_steps=cache_config.max_continuous_cached_steps,
+        residual_diff_threshold=cache_config.residual_diff_threshold,
+    )
+
+    cache_dit.enable_cache(
+        BlockAdapter(
+            transformer=pipeline.transformer,
+            blocks=[
+                pipeline.transformer.transformer_blocks,
+                pipeline.transformer.single_transformer_blocks,
+            ],
+            forward_pattern=[ForwardPattern.Pattern_1, ForwardPattern.Pattern_1],
+            params_modifiers=[ParamsModifier(...)],
+        ),
+        cache_config=db_cache_config,
+    )
+
+    def refresh_cache_context(pipeline, num_inference_steps, verbose=True):
+        cache_dit.refresh_context(
+            pipeline.transformer, num_inference_steps=num_inference_steps, verbose=verbose
+        )
+    return refresh_cache_context
+```
+
+For single transformer with multiple block lists, `refresh_context` works the same as standard models — call it once on the transformer.
+
+### Pattern 2: Dual-Transformer (Wan2.2 style)
+
+Two transformers with separate configs:
+
+```python
+def enable_cache_for_dual_transformer(pipeline, cache_config):
+    db_cache_config = DBCacheConfig(...)
+
+    cache_dit.enable_cache(
+        BlockAdapter(
+            transformer=[pipeline.transformer, pipeline.transformer_2],
+            blocks=[pipeline.transformer.blocks, pipeline.transformer_2.blocks],
+            forward_pattern=[ForwardPattern.Pattern_2, ForwardPattern.Pattern_2],
+            params_modifiers=[
+                ParamsModifier(...),  # Config for transformer 1
+                ParamsModifier(...),  # Config for transformer 2
+            ],
+        ),
+        cache_config=db_cache_config,
+    )
+
+    def refresh_cache_context(pipeline, num_inference_steps, verbose=True):
+        high_steps, low_steps = _split_inference_steps(num_inference_steps)
+        cache_dit.refresh_context(
+            pipeline.transformer, num_inference_steps=high_steps, verbose=verbose
+        )
+        cache_dit.refresh_context(
+            pipeline.transformer_2, num_inference_steps=low_steps, verbose=verbose
+        )
+    return refresh_cache_context
+```
+
+Key difference: `refresh_context` must be called on **each transformer separately** with its own step count.
+
+### Choosing the ForwardPattern
+
+| Pattern | Block forward signature | Example models |
+|---------|------------------------|----------------|
+| `Pattern_0` | `block(hidden_states, **kwargs)` → residual added inside block | Default |
+| `Pattern_1` | `block(hidden_states, **kwargs)` → returns `(hidden_states, ...)` tuple | FLUX-style single blocks |
+| `Pattern_2` | `block(hidden_states, **kwargs)` → `(hidden_states, ...)` with different residual pattern | Wan2.2 blocks |
+
+Inspect your block's `forward()` return type and residual connection pattern to choose the right one. See [Cache-DiT API Reference](https://cache-dit.readthedocs.io/en/latest/user_guide/CACHE_API/) for details.
+
+## Registering Custom Enablers
+
+Add your enabler to `CUSTOM_DIT_ENABLERS` in `vllm_omni/diffusion/cache/cache_dit_backend.py`:
+
+```python
+CUSTOM_DIT_ENABLERS = {
+    "Wan22Pipeline": enable_cache_for_wan22,
+    "LongCatImagePipeline": enable_cache_for_longcat_image,
+    "YourModelPipeline": enable_cache_for_your_model,
+}
+```
+
+The key must match `pipeline.__class__.__name__`.
+
+## Configuration Parameters
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `Fn_compute_blocks` | 1 | Number of blocks to always compute at the front |
+| `Bn_compute_blocks` | 0 | Number of blocks to always compute at the back |
+| `max_warmup_steps` | 4 | Steps to run without caching at the beginning |
+| `max_cached_steps` | — | Max total cached steps |
+| `max_continuous_cached_steps` | — | Max consecutive cached steps |
+| `residual_diff_threshold` | 0.24 | Threshold for deciding whether to cache a block |
+
+### Tuning for quality vs speed
+
+| Goal | Adjustments |
+|------|-------------|
+| **More speed, acceptable quality loss** | Higher `residual_diff_threshold` (0.24-0.4), lower `max_warmup_steps` (2-4) |
+| **Better quality, less speed** | Lower `residual_diff_threshold` (0.12-0.18), higher `max_warmup_steps` (6-8), lower `max_continuous_cached_steps` (2) |
+
+## Testing
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(
+    model="your-model-name",
+    cache_backend="cache_dit",
+    cache_config={
+        "Fn_compute_blocks": 1,
+        "Bn_compute_blocks": 0,
+        "max_warmup_steps": 4,
+        "residual_diff_threshold": 0.24,
+    }
+)
+images = omni.generate(
+    "a beautiful landscape",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+CLI (online serving):
+
+```bash
+vllm serve your-model --omni --port 8098 \
+  --cache-backend cache_dit \
+  --cache-config '{"Fn_compute_blocks": 1, "Bn_compute_blocks": 0, "max_warmup_steps": 4}'
+```
+
+**Verification checklist**:
+1. Logs show "Cache-dit enabled successfully on xxx"
+2. Performance: 1.5-2x speedup vs no cache
+3. Quality: compare output with `cache_backend=None`
+
+## Excluded Models
+
+Models listed in `_NO_CACHE_ACCELERATION` in `vllm_omni/diffusion/registry.py` do not support cache-dit (e.g., `NextStep11Pipeline`, `StableDiffusionPipeline`). Check this set before attempting to enable cache-dit.
+
+## Reference Implementations
+
+| Model | Path | Notes |
+|-------|------|-------|
+| Standard DiT | `cache_dit_backend.py::enable_cache_for_dit` | Default enabler, automatic |
+| Wan2.2 | `cache_dit_backend.py::enable_cache_for_wan22` | Dual-transformer, auto-detects mode |
+| LongCat | `cache_dit_backend.py::enable_cache_for_longcat_image` | Multi-block-list |
+| BAGEL | `cache_dit_backend.py::enable_cache_for_bagel` | Complex omni model |

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

@@ -0,0 +1,273 @@
+# Custom Model Patterns Reference
+
+Patterns for adding models that don't come from the standard diffusers pipeline format.
+
+## Directory Structure Comparison
+
+### Diffusers-based model (e.g., Wan2.2)
+
+```
+vllm_omni/diffusion/models/wan2_2/
+├── __init__.py                    # Exports pipeline + transformer + helpers
+├── pipeline_wan2_2.py             # Pipeline: loads components via from_pretrained()
+├── pipeline_wan2_2_i2v.py         # Variant pipeline for image-to-video
+└── wan2_2_transformer.py          # Transformer: ported from diffusers, uses Attention layer
+```
+
+The transformer is loaded separately via `weights_sources` + `load_weights()`. Non-transformer components (VAE, text encoder) are loaded in `__init__` via `from_pretrained()`.
+
+### Custom model with external deps (e.g., DreamID-Omni)
+
+```
+vllm_omni/diffusion/models/dreamid_omni/
+├── __init__.py                    # Exports pipeline only
+├── pipeline_dreamid_omni.py       # Pipeline: loads ALL weights in __init__ via custom helpers
+├── fusion.py                      # Custom fusion architecture (video + audio cross-attention)
+└── wan2_2.py                      # Re-implemented Wan backbone with split API
+
+examples/offline_inference/x_to_video_audio/
+└── download_dreamid_omni.py       # Downloads weights from 3 HF repos + clones code repo
+```
+
+All weights loaded eagerly in `__init__`. `load_weights()` is a no-op. External dependency (`dreamid_omni` package) imported with try/except.
+
+### Custom model with ported code (e.g., BAGEL)
+
+```
+vllm_omni/diffusion/models/bagel/
+├── __init__.py
+├── pipeline_bagel.py              # Pipeline: instantiates models, uses weights_sources
+├── bagel_transformer.py           # Full LLM backbone (Qwen2-MoT) ported into vllm-omni
+└── autoencoder.py                 # Custom VAE ported from original repo
+```
+
+Model code is fully ported (no external dependency). Uses `weights_sources` and `load_weights()` with custom name remapping to handle non-diffusers safetensors format.
+
+## Weight Loading Patterns
+
+### Pattern 1: Standard diffusers flow (Wan2.2, Z-Image, FLUX)
+
+```
+init → create transformer (empty) → set weights_sources → [loader calls load_weights()]
+```
+
+- `weights_sources` points to safetensors in HF subfolder (e.g., `transformer/`)
+- `load_weights()` receives `(name, tensor)` pairs from the loader
+- Name remapping handles diffusers→vllm-omni differences (QKV fusion, Sequential index removal)
+
+### Pattern 2: Custom safetensors at root (BAGEL)
+
+```
+init → create all models (empty) → set weights_sources(subfolder=None) → [loader calls load_weights()]
+```
+
+- `weights_sources` points to **root** of model directory, not a subfolder
+- Weights have non-diffusers names (e.g., `bagel.language_model.model.layers.0.self_attn.q_proj.weight`)
+- `load_weights()` does heavy name normalization
+
+```python
+self.weights_sources = [
+    DiffusersPipelineLoader.ComponentSource(
+        model_or_path=od_config.model,
+        subfolder=None,      # root directory
+        prefix="",           # no prefix stripping
+        fall_back_to_pt=False,
+    )
+]
+```
+
+### Pattern 3: Fully custom loading (DreamID-Omni)
+
+```
+init → load ALL weights eagerly via custom helpers → load_weights() = no-op
+```
+
+- No `weights_sources` attribute — standard loader finds nothing to iterate
+- Custom init functions (e.g., `init_wan_vae_2_2()`, `load_fusion_checkpoint()`) handle downloading and loading
+- `load_weights()` is `pass`
+- Weights may come from multiple HF repos in different formats (`.pth`, `.safetensors`)
+
+Use this when:
+- The original model has complex, well-tested loading code you don't want to rewrite
+- Weights span multiple HF repos
+- Weight format is non-standard (e.g., a single `.pth` file, not sharded safetensors)
+
+## model_index.json for Custom Models
+
+Standard diffusers `model_index.json`:
+```json
+{
+    "_class_name": "WanPipeline",
+    "_diffusers_version": "0.35.0.dev0",
+    "scheduler": ["diffusers", "UniPCMultistepScheduler"],
+    "transformer": ["diffusers", "WanTransformer3DModel"],
+    "vae": ["diffusers", "AutoencoderKLWan"]
+}
+```
+
+Custom model `model_index.json` (minimal):
+```json
+{
+    "_class_name": "DreamIDOmniPipeline",
+    "fusion": "DreamID-Omni/dreamid_omni.safetensors"
+}
+```
+
+The only **required** field is `_class_name` — it must match a key in `_DIFFUSION_MODELS` in `registry.py`. Other fields are model-specific and accessible via `od_config.model_config` dict.
+
+## External Dependency Management
+
+### Git clone + .pth injection (DreamID-Omni pattern)
+
+```python
+def download_dependency():
+    CACHE_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOCK_FILE, "w") as f:
+        fcntl.flock(f, fcntl.LOCK_EX)
+        if not DEPENDENCY_DIR.exists():
+            subprocess.run([
+                "git", "clone", "--depth", "1",
+                REPO_URL, "--branch", BRANCH,
+                str(DEPENDENCY_DIR)
+            ], check=True)
+        fcntl.flock(f, fcntl.LOCK_UN)
+
+    # Add to Python path via .pth file
+    site_packages = Path(site.getsitepackages()[0])
+    pth_file = site_packages / "vllm_omni_dependency.pth"
+    pth_file.write_text(str(DEPENDENCY_DIR))
+```
+
+### Direct port (BAGEL pattern)
+
+Copy essential files from the original repo into `vllm_omni/diffusion/models/<name>/`. Adapt imports to use vllm-omni utilities. Benefits: no external dependency, no git clone step. Drawback: must maintain the ported code.
+
+## Multi-Modal Input/Output Protocols
+
+Custom models that handle images, audio, or video I/O should implement protocol classes:
+
+```python
+from vllm_omni.diffusion.models.interface import (
+    SupportImageInput,    # Model accepts image input
+    SupportAudioInput,    # Model accepts audio input
+    SupportAudioOutput,   # Model produces audio output
+)
+
+class MyPipeline(nn.Module, SupportImageInput, SupportAudioInput, SupportAudioOutput):
+    pass  # Protocol markers enable proper engine routing
+```
+
+The engine checks `isinstance(pipeline, SupportImageInput)` at startup to configure input validation and warmup behavior.
+
+## Hardcoded Config vs Config Files
+
+Diffusers models use `config.json` in each subfolder. Custom models often use:
+
+**Module-level config dicts** (DreamID-Omni):
+```python
+VIDEO_CONFIG = {
+    "patch_size": [1, 2, 2], "model_type": "ti2v",
+    "dim": 3072, "ffn_dim": 14336, "num_heads": 24, "num_layers": 30, ...
+}
+```
+
+**Loaded from custom JSON** (BAGEL):
+```python
+cfg_path = os.path.join(model_path, "config.json")
+with open(cfg_path) as f:
+    bagel_cfg = json.load(f)
+vae_cfg = bagel_cfg.get("vae_config", {})
+```
+
+## Custom Architecture Patterns
+
+### Split forward API (DreamID-Omni)
+
+When a fusion model needs to interleave blocks from two backbones:
+
+```python
+class WanModel(nn.Module):
+    def prepare_transformer_block_kwargs(self, x, t, context, ...):
+        # Patch embed, time embed, text embed, RoPE
+        return x, e, kwargs
+
+    def post_transformer_block_out(self, x, grid_sizes, e):
+        # Output projection, unpatchify
+        return output
+
+    def forward(self, *args, **kwargs):
+        raise NotImplementedError  # Fusion model handles block iteration
+```
+
+The `FusionModel` then iterates blocks in lock-step:
+```python
+for video_block, audio_block in zip(self.video_model.blocks, self.audio_model.blocks):
+    video_out = video_block(video_hidden, ...)
+    audio_out = audio_block(audio_hidden, ...)
+    # Cross-attend between modalities
+    video_out = cross_attention(video_out, audio_out)
+    audio_out = cross_attention(audio_out, video_out)
+```
+
+### LLM-as-denoiser (BAGEL)
+
+When the backbone is a language model that also does diffusion:
+
+```python
+class BagelModel(nn.Module):
+    def __init__(self):
+        self.language_model = Qwen2MoTForCausalLM(config)
+        self.vit_model = SiglipVisionModel(vit_config)
+```
+
+The LLM processes both text tokens and latent image tokens in a single forward pass, using KV caching for the text portion.
+
+## Pre/Post Processing for Custom Models
+
+Custom models typically handle pre/post processing **inside `forward()`** rather than via registered functions, because the logic is tightly coupled:
+
+```python
+def forward(self, req: OmniDiffusionRequest) -> DiffusionOutput:
+    # Inline preprocessing
+    image = self._load_and_resize_image(req.prompts[0].get("multi_modal_data", {}).get("image"))
+    image_latent = self._vae_encode(image)
+
+    # ... denoising loop ...
+
+    # Inline postprocessing
+    pil_image = self._decode_to_pil(latents)
+    return DiffusionOutput(output=[pil_image])
+```
+
+If pre/post functions are not registered in `_DIFFUSION_PRE_PROCESS_FUNCS` / `_DIFFUSION_POST_PROCESS_FUNCS`, the engine simply skips those steps.
+
+## Download Script Template
+
+```python
+# examples/offline_inference/<name>/download_<name>.py
+from huggingface_hub import snapshot_download
+import json, os
+
+def main(output_dir):
+    # Download model weights from HF
+    snapshot_download(repo_id="org/model-weights", local_dir=os.path.join(output_dir, "weights"))
+
+    # Download additional components if from separate repos
+    snapshot_download(repo_id="org/vae-weights", local_dir=os.path.join(output_dir, "vae"),
+        allow_patterns=["*.safetensors"])
+
+    # Generate model_index.json
+    config = {"_class_name": "YourPipeline", "custom_key": "weights/model.safetensors"}
+    with open(os.path.join(output_dir, "model_index.json"), "w") as f:
+        json.dump(config, f, indent=2)
+
+    # Install external code dependency (if needed)
+    download_dependency()
+
+if __name__ == "__main__":
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--output-dir", default="./your_model")
+    args = parser.parse_args()
+    main(args.output_dir)
+```

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

@@ -0,0 +1,572 @@
+# Parallelism Patterns Reference
+
+## Overview
+
+vLLM-Omni supports multiple parallelism strategies for diffusion models. Each targets a different bottleneck:
+
+| Strategy | Splits | Best For | Constraint |
+|----------|--------|----------|------------|
+| Tensor Parallel (TP) | Model layers across GPUs | Latency reduction, large models | Requires fast GPU interconnect, `num_heads % tp == 0` |
+| Sequence Parallel (SP/USP) | Sequence tokens across GPUs | Long sequences (video, high-res) | Near-linear scaling |
+| CFG Parallel | Positive/negative CFG branches | Models using classifier-free guidance | Exactly 2 GPUs |
+| HSDP | Weight shards via FSDP2 | VRAM reduction | Cannot combine with TP |
+| VAE Patch Parallel | VAE decode spatial tiles | Large VAE outputs | Auto-enables tiling |
+
+**Recommended integration order**: TP → SP → CFG Parallel → HSDP
+
+**Official design docs**:
+- TP: https://docs.vllm.ai/projects/vllm-omni/en/latest/design/feature/tensor_parallel
+- SP: https://docs.vllm.ai/projects/vllm-omni/en/latest/design/feature/sequence_parallel
+- CFG: https://docs.vllm.ai/projects/vllm-omni/en/latest/design/feature/cfg_parallel
+- HSDP: https://docs.vllm.ai/projects/vllm-omni/en/latest/design/feature/hsdp
+
+---
+
+## Tensor Parallelism (TP)
+
+Replace standard `nn.Linear` with vLLM's parallel linear layers. This is the most invasive change but provides direct VRAM savings and compute speedup.
+
+### Layer replacement rules
+
+| Pattern | vLLM Layer | When to Use |
+|---------|-----------|-------------|
+| Fan-out (first in FFN) | `ColumnParallelLinear` | Projection that splits output across ranks |
+| Fan-in (second in FFN) | `RowParallelLinear` | Projection that gathers across ranks |
+| QKV projection | `QKVParallelLinear` | Fused Q/K/V for self-attention |
+| Single Q or K or V | `ColumnParallelLinear` | Separate projections (cross-attention) |
+| Attention output | `RowParallelLinear` | Output projection after attention |
+| Must not shard | `ReplicatedLinear` | Layers that must stay replicated |
+
+### MLP Block (Up-Down Pattern)
+
+```python
+from vllm.model_executor.layers.linear import (
+    ColumnParallelLinear, RowParallelLinear,
+)
+
+class TPFeedForward(nn.Module):
+    def __init__(self, dim, ffn_dim):
+        super().__init__()
+        self.fc1 = ColumnParallelLinear(dim, ffn_dim, bias=False, return_bias=False)
+        self.fc2 = RowParallelLinear(
+            ffn_dim, dim, bias=False,
+            input_is_parallel=True,  # Input already sharded from fc1
+            return_bias=False,
+        )
+
+    def forward(self, x):
+        x, _ = self.fc1(x)
+        x = torch.nn.functional.gelu(x)
+        x, _ = self.fc2(x)
+        return x
+```
+
+### Attention Block (QKV-Out Pattern)
+
+```python
+from vllm.model_executor.layers.linear import QKVParallelLinear, RowParallelLinear
+from vllm_omni.diffusion.attention.layer import Attention
+
+class TPSelfAttention(nn.Module):
+    def __init__(self, dim, num_heads, num_kv_heads=None):
+        super().__init__()
+        num_kv_heads = num_kv_heads or num_heads
+        self.head_dim = dim // num_heads
+
+        self.to_qkv = QKVParallelLinear(
+            hidden_size=dim,
+            head_size=self.head_dim,
+            total_num_heads=num_heads,
+            total_num_kv_heads=num_kv_heads,
+            bias=False,
+            return_bias=False,
+        )
+        self.to_out = RowParallelLinear(
+            dim, dim, bias=False,
+            input_is_parallel=True,
+            return_bias=False,
+        )
+        self.attn = Attention(
+            num_heads=self.to_qkv.num_heads,      # Local heads per GPU
+            head_size=self.head_dim,
+            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):
+        qkv, _ = self.to_qkv(x)
+        q, k, v = qkv.split(
+            [self.to_qkv.num_heads * self.head_dim,
+             self.to_qkv.num_kv_heads * self.head_dim,
+             self.to_qkv.num_kv_heads * self.head_dim],
+            dim=-1,
+        )
+        B, S, _ = x.shape
+        q = q.view(B, S, self.to_qkv.num_heads, self.head_dim)
+        k = k.view(B, S, self.to_qkv.num_kv_heads, self.head_dim)
+        v = v.view(B, S, self.to_qkv.num_kv_heads, self.head_dim)
+        out = self.attn(q, k, v)
+        out = out.reshape(B, S, -1)
+        out, _ = self.to_out(out)
+        return out
+```
+
+### QKV Fusion in load_weights
+
+When you fuse separate Q/K/V into `QKVParallelLinear`, map diffusers' separate weight names:
+
+```python
+stacked_params_mapping = [
+    ("to_qkv", "to_q", "q"),
+    ("to_qkv", "to_k", "k"),
+    ("to_qkv", "to_v", "v"),
+]
+
+def load_weights(self, weights):
+    params = dict(self.named_parameters())
+    loaded = set()
+    for name, tensor in weights:
+        for fused_name, orig_name, shard_id in stacked_params_mapping:
+            if orig_name in name:
+                name = name.replace(orig_name, fused_name)
+                param = params[name]
+                param.weight_loader(param, tensor, shard_id)
+                loaded.add(name)
+                break
+        else:
+            if name in params:
+                param = params[name]
+                if hasattr(param, "weight_loader"):
+                    param.weight_loader(param, tensor)
+                else:
+                    default_weight_loader(param, tensor)
+                loaded.add(name)
+    return loaded
+```
+
+### RMSNorm with TP
+
+When RMSNorm sits between TP-sharded dimensions, use `DistributedRMSNorm` — it computes global RMS via all-reduce across TP ranks. See the Wan2.2 implementation for the pattern.
+
+### TP Constraints
+
+- `num_heads % tp_size == 0`
+- `num_kv_heads % tp_size == 0`
+- Use `self.to_qkv.num_heads` (local per-GPU count), not total heads, for split sizes
+
+### Testing TP
+
+```bash
+python text_to_image.py --model Your-org/your-model \
+  --tensor-parallel-size 2 --output "tp_test.png"
+```
+
+**Verify**: speedup, memory reduction proportional to TP size, quality matches single-GPU.
+
+### Reference implementations
+
+| Model | Path |
+|-------|------|
+| Z-Image | `vllm_omni/diffusion/models/z_image/z_image_transformer.py` |
+| FLUX | `vllm_omni/diffusion/models/flux/flux_transformer.py` |
+| Qwen-Image | `vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py` |
+
+---
+
+## Sequence Parallelism (SP / USP)
+
+SP splits sequence tokens across GPUs using Ulysses (all-to-all) or Ring (P2P) communication. It is applied non-intrusively via the `_sp_plan` dict — no changes to `forward()` logic.
+
+### Approach 1: Non-Intrusive `_sp_plan` (Recommended)
+
+The framework automatically registers hooks to shard inputs and gather outputs at `nn.Module` boundaries.
+
+#### Step 1: Identify module boundaries
+
+Find where tensors need sharding/gathering:
+
+```python
+class MyTransformer(nn.Module):
+    def __init__(self):
+        self.patch_embed = PatchEmbed()    # Before blocks
+        self.pos_embed = RoPE()            # RoPE may need splitting
+        self.blocks = nn.ModuleList([...]) # Blocks process sharded x
+        self.norm_out = LayerNorm()
+        self.proj_out = Linear()           # Gather after this
+
+    def forward(self, x):
+        x = self.patch_embed(x)
+        pos = self.pos_embed(x)
+        for block in self.blocks:
+            x = block(x, pos)
+        x = self.norm_out(x)
+        return self.proj_out(x)
+```
+
+#### Step 2: Handle inline operations
+
+`_sp_plan` hooks only work at `nn.Module` boundaries. Inline ops like `torch.cat()` must be extracted into submodules:
+
+```python
+# BAD: Inline — hooks can't intercept
+unified = torch.cat([x, cap_feats], dim=1)
+
+# GOOD: Extract into submodule
+class UnifiedPrepare(nn.Module):
+    def forward(self, x, cap_feats):
+        return torch.cat([x, cap_feats], dim=1)
+
+self.unified_prepare = UnifiedPrepare()
+unified = self.unified_prepare(x, cap_feats)
+```
+
+Common cases: `torch.cat()`, `pad_sequence()`, `tensor.reshape()`, complex preprocessing.
+
+#### Step 3: Write `_sp_plan`
+
+**Pattern 1: Shard at first block, gather at output** (most common)
+
+```python
+from vllm_omni.diffusion.distributed.sp_plan import (
+    SequenceParallelInput, SequenceParallelOutput,
+)
+
+class StandardTransformer(nn.Module):
+    _sp_plan = {
+        "blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+**Pattern 2: Shard RoPE outputs separately**
+
+```python
+class TransformerWithRoPE(nn.Module):
+    _sp_plan = {
+        "rope": {
+            0: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),
+            1: SequenceParallelInput(split_dim=1, expected_dims=4, split_output=True),
+        },
+        "blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+**Pattern 3: Dual-stream (shard image, replicate text)**
+
+```python
+class DualStreamTransformer(nn.Module):
+    _sp_plan = {
+        "rope_preparer": {
+            2: SequenceParallelInput(split_dim=0, expected_dims=2, split_output=True),
+            3: SequenceParallelInput(split_dim=0, expected_dims=2, split_output=True),
+        },
+        "transformer_blocks.0": {
+            "hidden_states": SequenceParallelInput(split_dim=1, expected_dims=3),
+        },
+        "proj_out": SequenceParallelOutput(gather_dim=1, expected_dims=3),
+    }
+```
+
+### API Reference
+
+**SequenceParallelInput**:
+
+| Parameter | Type | Description |
+|-----------|------|-------------|
+| `split_dim` | int | Dimension to split (usually 1 for sequence) |
+| `expected_dims` | int/None | Expected tensor rank for validation |
+| `split_output` | bool | `False`: shard input params; `True`: shard output tensors |
+| `auto_pad` | bool | Auto-pad if sequence not divisible by world_size |
+
+**SequenceParallelOutput**:
+
+| Parameter | Type | Description |
+|-----------|------|-------------|
+| `gather_dim` | int | Dimension to gather (usually 1 for sequence) |
+| `expected_dims` | int/None | Expected tensor rank for validation |
+
+**Module naming**:
+
+| Key | Meaning |
+|-----|---------|
+| `"blocks.0"` | First element of ModuleList |
+| `"blocks.*"` | All elements of ModuleList |
+| `"rope"` | Named submodule |
+
+**Dictionary value types**:
+
+| Key type | split_output | Description |
+|----------|-------------|-------------|
+| `"param_name"` (str) | False | Shard input parameter by name |
+| `0, 1, ...` (int) | True | Shard output tuple by index |
+
+### Approach 2: Intrusive Modification (Complex Cases)
+
+For dynamic sharding logic that can't be expressed via `_sp_plan`:
+
+```python
+from vllm_omni.diffusion.distributed.sp_sharding import sp_shard, sp_gather
+
+def forward(self, hidden_states, ...):
+    if self.parallel_config.sequence_parallel_size > 1:
+        hidden_states = sp_shard(hidden_states, dim=1)
+    for block in self.blocks:
+        hidden_states = block(hidden_states)
+    if self.parallel_config.sequence_parallel_size > 1:
+        hidden_states = sp_gather(hidden_states, dim=1)
+    return hidden_states
+```
+
+Use intrusive modification as a last resort — `_sp_plan` is preferred for maintainability.
+
+### UAA Mode (Experimental)
+
+`ulysses_mode="advanced_uaa"` handles arbitrary sequence lengths and head counts that aren't divisible by `ulysses_degree`. Uses variable all-to-all split sizes and temporary head padding.
+
+### Combining SP methods
+
+Ulysses and Ring can be combined: `ulysses_degree × ring_degree = total SP GPUs`.
+
+```python
+DiffusionParallelConfig(ulysses_degree=2, ring_degree=2)  # 4 GPUs total
+```
+
+### Testing SP
+
+```bash
+# Offline
+python text_to_image.py --model Your-model --ulysses-degree 2
+
+# Online serving
+vllm serve Your-model --omni --usp 2
+```
+
+### Reference implementations
+
+| Model | Path |
+|-------|------|
+| Qwen-Image | `vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py` |
+| Wan2.2 | `vllm_omni/diffusion/models/wan2_2/wan2_2_transformer.py` |
+| Z-Image | `vllm_omni/diffusion/models/z_image/z_image_transformer.py` |
+
+---
+
+## CFG Parallelism
+
+Distributes positive/negative Classifier-Free Guidance branches across 2 GPUs.
+
+### Implementation
+
+Inherit `CFGParallelMixin` and implement `diffuse()`:
+
+```python
+from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
+
+class YourPipeline(nn.Module, CFGParallelMixin):
+    def diffuse(self, latents, timesteps, prompt_embeds, negative_embeds,
+                do_true_cfg, true_cfg_scale, **kwargs):
+        for i, t in enumerate(timesteps):
+            positive_kwargs = {
+                "hidden_states": latents,
+                "encoder_hidden_states": prompt_embeds,
+                "timestep": t,
+            }
+            negative_kwargs = {
+                "hidden_states": latents,
+                "encoder_hidden_states": negative_embeds,
+                "timestep": t,
+            } if do_true_cfg else None
+
+            noise_pred = self.predict_noise_maybe_with_cfg(
+                do_true_cfg=do_true_cfg,
+                true_cfg_scale=true_cfg_scale,
+                positive_kwargs=positive_kwargs,
+                negative_kwargs=negative_kwargs,
+            )
+            latents = self.scheduler_step_maybe_with_cfg(
+                noise_pred, t, latents, do_true_cfg
+            )
+        return latents
+```
+
+### Customization hooks
+
+| Method | Override when |
+|--------|-------------|
+| `predict_noise()` | Non-standard transformer call (e.g., dual-transformer like Wan2.2) |
+| `cfg_normalize_function()` | Custom normalization (e.g., LongCat with clamping) |
+| `combine_cfg_noise()` | Multi-output models (e.g., video + audio: CFG on video, positive-only on audio) |
+
+**Custom predict_noise** (Wan2.2 — selects active transformer):
+
+```python
+def predict_noise(self, current_model=None, **kwargs):
+    if current_model is None:
+        current_model = self.transformer
+    return current_model(**kwargs)[0]
+```
+
+**Custom combine_cfg_noise** (multi-output):
+
+```python
+def combine_cfg_noise(self, positive_pred, negative_pred, scale, normalize):
+    video_pos, audio_pos = positive_pred
+    video_neg, audio_neg = negative_pred
+    video_combined = super().combine_cfg_noise(video_pos, video_neg, scale, normalize)
+    return (video_combined, audio_pos)
+```
+
+### Composite scheduler for multi-output
+
+When each output has its own schedule:
+
+```python
+class VideoAudioScheduler:
+    def __init__(self, video_scheduler, audio_scheduler):
+        self.video_scheduler = video_scheduler
+        self.audio_scheduler = audio_scheduler
+
+    def step(self, noise_pred, t, latents, return_dict=False, generator=None):
+        video_out = self.video_scheduler.step(
+            noise_pred[0], t[0], latents[0], return_dict=False, generator=generator
+        )[0]
+        audio_out = self.audio_scheduler.step(
+            noise_pred[1], t[1], latents[1], return_dict=False, generator=generator
+        )[0]
+        return ((video_out, audio_out),)
+```
+
+### Testing CFG Parallel
+
+```bash
+python text_to_image.py --model Your-model \
+  --cfg-parallel-size 2 --cfg-scale 4.0 \
+  --negative-prompt "ugly, unclear"
+```
+
+**Constraint**: `guidance_scale > 1.0` and negative prompt must be provided.
+
+### Reference implementations
+
+| Model | Path |
+|-------|------|
+| Qwen-Image | `vllm_omni/diffusion/models/qwen_image/cfg_parallel.py` |
+| Wan2.2 | `vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py` |
+| Mixin base | `vllm_omni/diffusion/distributed/cfg_parallel.py` |
+
+---
+
+## HSDP (Hybrid Sharded Data Parallel)
+
+Shards model weights across GPUs using PyTorch FSDP2. Reduces per-GPU VRAM without changing computation.
+
+### Implementation
+
+Add `_hsdp_shard_conditions` to the transformer class:
+
+```python
+class YourTransformer(nn.Module):
+    @staticmethod
+    def _is_transformer_block(name: str, module) -> bool:
+        return "blocks" in name and name.split(".")[-1].isdigit()
+
+    _hsdp_shard_conditions = [_is_transformer_block]
+```
+
+For MoE models, add additional conditions:
+
+```python
+class MoETransformer(nn.Module):
+    @staticmethod
+    def _is_transformer_block(name, module):
+        return "blocks" in name and name.split(".")[-1].isdigit()
+
+    @staticmethod
+    def _is_moe_expert(name, module):
+        return "experts" in name and name.split(".")[-1].isdigit()
+
+    _hsdp_shard_conditions = [_is_transformer_block, _is_moe_expert]
+```
+
+A module is sharded if **any** condition returns `True`.
+
+### Constraints
+
+- Cannot combine with Tensor Parallelism
+- For standalone HSDP (no other parallelism), `hsdp_shard_size` must be specified explicitly
+- Can combine with SP: HSDP reduces memory while SP distributes sequence
+
+### Testing HSDP
+
+```python
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+
+parallel_config = DiffusionParallelConfig(use_hsdp=True, hsdp_shard_size=8)
+omni = Omni(model="your-model", parallel_config=parallel_config)
+```
+
+Or CLI:
+
+```bash
+vllm serve Your-model --omni --use-hsdp
+```
+
+**Verify**: logs show "HSDP Inference: replicate_size=..., shard_size=..." and "Sharded N modules + root". Check VRAM reduction.
+
+### Reference implementations
+
+| Model | Path |
+|-------|------|
+| Wan2.2 | `vllm_omni/diffusion/models/wan2_2/wan2_2_transformer.py` |
+| HSDP Core | `vllm_omni/diffusion/distributed/hsdp.py` |
+
+---
+
+## VAE Patch Parallelism
+
+Shards VAE decode spatially across ranks using tiling:
+
+```bash
+python text_to_image.py --model Your-model --vae-patch-parallel-size 4
+```
+
+Auto-enables `--vae-use-tiling`. Uses `DistributedAutoencoderKLWan` or similar distributed VAE. Set `vae_patch_parallel_size` in `DiffusionParallelConfig`.
+
+---
+
+## Combining Parallelism Methods
+
+Common multi-GPU recipes:
+
+```bash
+# 4 GPUs: CFG (2) × Ulysses (2)
+python text_to_image.py --model Qwen/Qwen-Image \
+  --cfg-parallel-size 2 --ulysses-degree 2
+
+# 8 GPUs: Ulysses (4) × Ring (2) + VAE patch (8)
+python text_to_video.py --model Wan-AI/Wan2.2-T2V-A14B-Diffusers \
+  --ulysses-degree 4 --ring-degree 2 --vae-patch-parallel-size 8
+
+# 2 GPUs: HSDP + Ulysses (cannot combine HSDP with TP)
+vllm serve Your-model --omni --use-hsdp --usp 2
+```
+
+## Discovering Parallelism Support
+
+Check which parallelism methods a model supports:
+
+| Check | How |
+|-------|-----|
+| **Ulysses / Ring SP** | Transformer defines `_sp_plan`. Search: `grep -r '_sp_plan' vllm_omni/diffusion/models/` |
+| **CFG Parallel** | Pipeline inherits `CFGParallelMixin`. Search: `grep -r 'CFGParallelMixin' vllm_omni/diffusion/models/` |
+| **TP** | Uses `ColumnParallelLinear` / `QKVParallelLinear`. Search: `grep -r 'ParallelLinear\|QKVParallel' vllm_omni/diffusion/models/<model>/` |
+| **HSDP** | Transformer defines `_hsdp_shard_conditions`. Search: `grep -r '_hsdp_shard_conditions' vllm_omni/diffusion/models/` |
+
+The canonical per-model support table is in `docs/user_guide/diffusion/parallelism_acceleration.md`.

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

@@ -0,0 +1,238 @@
+# Transformer Adaptation Reference
+
+## Adapting a Diffusers Transformer to vLLM-Omni
+
+### Step-by-step Checklist
+
+1. Copy the transformer class from diffusers source
+2. Remove all mixin classes — inherit only from `nn.Module`
+3. Replace attention dispatch with `vllm_omni.diffusion.attention.layer.Attention`
+4. Replace logger with `vllm.logger.init_logger`
+5. Add `od_config: OmniDiffusionConfig | None = None` to `__init__`
+6. Remove training-only code (gradient checkpointing, dropout)
+7. Add `load_weights()` method for weight loading from safetensors
+8. Add class-level attributes for acceleration features
+
+### Mixin Removal
+
+Remove these diffusers mixins (and their imports):
+
+```python
+# Remove all of these:
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import AttentionModuleMixin
+from diffusers.loaders import PeftAdapterMixin, FromOriginalModelMixin
+
+# Replace:
+class MyTransformer(ModelMixin, ConfigMixin, AttentionModuleMixin):
+# With:
+class MyTransformer(nn.Module):
+```
+
+Also remove `@register_to_config` decorators from `__init__`.
+
+### Attention Replacement
+
+The vLLM-Omni `Attention` layer wraps backend selection (FlashAttention, SDPA, SageAttn, etc.) and supports sequence parallelism hooks.
+
+**QKV tensor shape must be `[batch, seq_len, num_heads, head_dim]`.**
+
+#### Self-Attention Pattern
+
+```python
+from vllm_omni.diffusion.attention.layer import Attention
+from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata
+
+class SelfAttentionBlock(nn.Module):
+    def __init__(self, dim, num_heads):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.to_q = nn.Linear(dim, dim)
+        self.to_k = nn.Linear(dim, dim)
+        self.to_v = nn.Linear(dim, dim)
+        self.to_out = nn.Linear(dim, dim)
+
+        self.attn = Attention(
+            num_heads=num_heads,
+            head_size=self.head_dim,
+            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):
+        B, S, _ = x.shape
+        q = self.to_q(x).view(B, S, self.num_heads, self.head_dim)
+        k = self.to_k(x).view(B, S, self.num_heads, self.head_dim)
+        v = self.to_v(x).view(B, S, self.num_heads, self.head_dim)
+
+        attn_metadata = AttentionMetadata(attn_mask=attn_mask)
+        out = self.attn(q, k, v, attn_metadata=attn_metadata)
+        out = out.reshape(B, S, -1)
+        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:
+
+```python
+from vllm.model_executor.layers.linear import (
+    QKVParallelLinear, RowParallelLinear
+)
+
+class TPSelfAttention(nn.Module):
+    def __init__(self, dim, num_heads):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.to_qkv = QKVParallelLinear(
+            hidden_size=dim,
+            head_size=self.head_dim,
+            total_num_heads=num_heads,
+            total_num_kv_heads=num_heads,
+        )
+        self.to_out = RowParallelLinear(dim, dim)
+
+        self.attn = Attention(
+            num_heads=num_heads,
+            head_size=self.head_dim,
+            softmax_scale=1.0 / (self.head_dim ** 0.5),
+            causal=False,
+            num_kv_heads=num_heads,
+            role="self",
+        )
+```
+
+### Logger Replacement
+
+```python
+# Replace:
+from diffusers.utils import logging
+logger = logging.get_logger(__name__)
+
+# With:
+from vllm.logger import init_logger
+logger = init_logger(__name__)
+```
+
+### Custom Layers from vLLM-Omni
+
+Available utility layers:
+
+```python
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm_omni.diffusion.layers.rope import RotaryEmbedding
+from vllm_omni.diffusion.layers.adalayernorm import AdaLayerNorm
+```
+
+### Config Support
+
+```python
+from vllm_omni.diffusion.data import OmniDiffusionConfig
+
+class MyTransformer(nn.Module):
+    def __init__(self, *, od_config=None, num_layers=28, hidden_size=3072, **kwargs):
+        super().__init__()
+        self.od_config = od_config
+        self.parallel_config = od_config.parallel_config if od_config else None
+        # ... build layers
+```
+
+The transformer config values come from `model_index.json` → `config.json` in the transformer subfolder. The pipeline uses `get_transformer_config_kwargs(od_config.tf_model_config, TransformerClass)` to filter config keys to match the `__init__` signature.
+
+### Weight Loading
+
+The `load_weights` method receives an iterable of `(name, tensor)` from safetensors files, with the prefix (e.g., `"transformer."`) already stripped by the loader.
+
+```python
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+class MyTransformer(nn.Module):
+    def load_weights(self, weights):
+        params = dict(self.named_parameters())
+        loaded = set()
+        for name, tensor in weights:
+            # Optional: remap names from diffusers to vllm-omni naming
+            # e.g., "ff.net.0.proj" -> "ff.net_0.proj"
+
+            if name in params:
+                param = params[name]
+                if hasattr(param, "weight_loader"):
+                    param.weight_loader(param, tensor)
+                else:
+                    default_weight_loader(param, tensor)
+                loaded.add(name)
+        return loaded
+```
+
+#### QKV Fusion in load_weights
+
+If you fused separate Q/K/V into a `QKVParallelLinear`, you need to map diffusers' separate weight names:
+
+```python
+stacked_params_mapping = [
+    ("to_qkv", "to_q", "q"),
+    ("to_qkv", "to_k", "k"),
+    ("to_qkv", "to_v", "v"),
+]
+
+def load_weights(self, weights):
+    params = dict(self.named_parameters())
+    loaded = set()
+    for name, tensor in weights:
+        for fused_name, orig_name, shard_id in stacked_params_mapping:
+            if orig_name in name:
+                name = name.replace(orig_name, fused_name)
+                param = params[name]
+                param.weight_loader(param, tensor, shard_id)
+                loaded.add(name)
+                break
+        else:
+            # Normal loading
+            ...
+    return loaded
+```
+
+### Class-Level Attributes for Features
+
+```python
+class MyTransformer(nn.Module):
+    # torch.compile: list block class names that repeat and can be compiled
+    _repeated_blocks = ["MyTransformerBlock"]
+
+    # CPU offload: attribute name of the nn.ModuleList containing blocks
+    _layerwise_offload_blocks_attr = "blocks"
+
+    # LoRA: mapping of fused param names to original param names
+    packed_modules_mapping = {"to_qkv": ["to_q", "to_k", "to_v"]}
+
+    # Sequence parallelism plan (advanced — add after basic impl works)
+    _sp_plan = {
+        "blocks.0": SequenceParallelInput(split_dim=1),
+        "proj_out": SequenceParallelOutput(gather_dim=1),
+    }
+```

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

@@ -0,0 +1,178 @@
+# Troubleshooting Reference
+
+## Common Errors When Adding a Diffusion Model
+
+### ImportError / ModuleNotFoundError
+
+**Cause**: Missing or incorrect registration.
+
+**Fix checklist**:
+1. Model registered in `vllm_omni/diffusion/registry.py` `_DIFFUSION_MODELS` dict
+2. `__init__.py` exports the pipeline class
+3. Pipeline file exists at the correct path: `vllm_omni/diffusion/models/{folder}/{file}.py`
+4. Class name in registry matches the actual class name in the file
+
+### Shape Mismatch in Attention
+
+**Symptom**: `RuntimeError: shape mismatch` or `expected 4D tensor`
+
+**Cause**: QKV tensors not reshaped to `[batch, seq_len, num_heads, head_dim]`.
+
+**Fix**: Before calling `self.attn(q, k, v, ...)`, ensure:
+```python
+q = q.view(batch, seq_len, self.num_heads, self.head_dim)
+k = k.view(batch, kv_seq_len, self.num_kv_heads, self.head_dim)
+v = v.view(batch, kv_seq_len, self.num_kv_heads, self.head_dim)
+```
+
+After attention, reshape back:
+```python
+out = out.reshape(batch, seq_len, -1)
+```
+
+### Weight Loading Failures
+
+**Symptom**: `RuntimeError: size mismatch for parameter ...` or missing keys
+
+**Debugging**:
+1. Print diffusers weight names: `safetensors.safe_open(path, "pt").keys()`
+2. Print model parameter names: `dict(model.named_parameters()).keys()`
+3. Compare and add name remappings in `load_weights()`
+
+**Common remappings needed**:
+- `ff.net.0.proj` → `ff.net_0.proj` (PyTorch Sequential indexing)
+- `.to_out.0.` → `.to_out.` (Sequential unwrapping)
+- `scale_shift_table` → moved to a wrapper module
+
+### Black/Blank/Noisy Output
+
+**Possible causes**:
+1. **Wrong latent normalization**: Check VAE expects latents scaled by `vae.config.scaling_factor`
+2. **Wrong scheduler**: Using the wrong scheduler class or wrong `flow_shift`
+3. **Missing CFG**: Some models require `guidance_scale > 1.0` with negative prompt
+4. **Wrong timestep format**: Some schedulers expect float, others expect int/long
+5. **Missing post-processing**: Raw VAE output may need denormalization
+
+**Quick test**: Run with diffusers directly using the same seed and compare latents at each step.
+
+### OOM (Out of Memory)
+
+**Solutions** (in order of preference):
+1. `--enforce-eager` to disable torch.compile (saves compile memory)
+2. `--enable-cpu-offload` for model-level offload
+3. `--enable-layerwise-offload` for block-level offload (better for large models)
+4. `--vae-use-slicing --vae-use-tiling` for VAE memory reduction
+5. Reduce resolution: `--height 480 --width 832`
+6. Use TP: `--tensor-parallel-size 2`
+
+### Different Output vs Diffusers Reference
+
+**Common causes**:
+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
+
+### Tensor Parallel Errors
+
+**Symptom**: `AssertionError: not divisible` or incorrect output with TP>1
+
+**Fix**:
+1. Verify `num_heads % tp_size == 0` and `num_kv_heads % tp_size == 0`
+2. Ensure `ColumnParallelLinear` / `RowParallelLinear` are used correctly
+3. Check that norms between parallel layers use distributed norm if needed
+4. Verify `load_weights` handles TP sharding for norm weights
+5. Use `self.to_qkv.num_heads` (local heads per GPU) for QKV split sizes, not total heads
+
+**Missing `input_is_parallel=True`**:
+
+`RowParallelLinear` expects sharded input from `ColumnParallelLinear`:
+```python
+self.w1 = ColumnParallelLinear(dim, hidden_dim, return_bias=False)
+self.w2 = RowParallelLinear(hidden_dim, dim, input_is_parallel=True, return_bias=False)
+```
+
+### Sequence Parallel Errors
+
+**Symptom**: Incorrect output or crashes with `--ulysses-degree N` or `--usp N`
+
+**Possible causes**:
+1. **Inline operations between shard/gather points**: `torch.cat()`, `pad_sequence()` etc. not at `nn.Module` boundaries. Fix: extract into submodule.
+2. **Wrong `split_dim`**: Check the tensor shape at the shard point. Sequence dimension is typically `dim=1` for `[B, S, D]` tensors.
+3. **RoPE not sharded**: If RoPE is computed separately, add it to `_sp_plan` with `split_output=True`.
+4. **Sequence not divisible by SP degree**: Use `auto_pad=True` in `SequenceParallelInput` or switch to `ulysses_mode="advanced_uaa"`.
+
+**Debugging**: Add `expected_dims=N` to `SequenceParallelInput`/`Output` for shape validation at runtime.
+
+### CFG Parallel Errors
+
+**Symptom**: CFG parallel not activating, no speedup
+
+**Fix checklist**:
+1. Pipeline inherits `CFGParallelMixin`
+2. `guidance_scale > 1.0`
+3. Negative prompt provided (even if empty string)
+4. `--cfg-parallel-size 2` specified
+5. `diffuse()` method calls `predict_noise_maybe_with_cfg()` and `scheduler_step_maybe_with_cfg()`
+
+**Symptom**: Different output with CFG parallel vs sequential
+
+**Possible cause**: Non-deterministic scheduler. Fix: pass `generator=torch.Generator(device).manual_seed(seed)` to `scheduler_step_maybe_with_cfg()`.
+
+### HSDP Errors
+
+**Symptom**: HSDP not activating or errors during weight loading
+
+**Fix checklist**:
+1. Transformer defines `_hsdp_shard_conditions` class attribute
+2. Shard condition functions return `True` for correct modules (test with `model.named_modules()`)
+3. Not combining with TP (HSDP and TP are incompatible)
+4. For standalone HSDP, `hsdp_shard_size` is specified explicitly
+
+**Verify**: Check logs for "HSDP Inference: replicate_size=..., shard_size=..." and "Sharded N modules + root".
+
+### Cache-DiT Not Applied
+
+**Symptom**: No speedup, no cache-related log messages
+
+**Fix checklist**:
+1. Model not in `_NO_CACHE_ACCELERATION` in `registry.py`
+2. Pipeline class name matches `CUSTOM_DIT_ENABLERS` key (if using custom enabler)
+3. `cache_backend="cache_dit"` specified
+4. Check logs for "Cache-dit enabled successfully on xxx"
+
+**Verify pipeline name**: `print(pipeline.__class__.__name__)` — must match registry key.
+
+### Cache-DiT Quality Degradation
+
+**Symptom**: Artifacts or lower quality with cache-dit
+
+**Fix**: Reduce aggressiveness:
+```python
+cache_config={
+    "residual_diff_threshold": 0.12,      # Lower from 0.24
+    "max_warmup_steps": 6,                # Increase from 4
+    "max_continuous_cached_steps": 2,      # Reduce if higher
+}
+```
+
+If quality is still poor, the model may need a custom enabler with per-block-list `ParamsModifier` tuning.
+
+### Model Not Detected / Wrong Pipeline Class
+
+**Symptom**: `ValueError: Model class ... not found in diffusion model registry`
+
+**Cause**: The model's `model_index.json` has a `_class_name` for the pipeline that doesn't match registry keys.
+
+**Fix**: The registry key must match the diffusers pipeline class name from `model_index.json`. If using a different name, map it in the registry:
+```python
+"DiffusersPipelineClassName": ("your_folder", "your_file", "YourVllmClassName"),
+```
+
+## Debugging Workflow
+
+1. **Add verbose logging**: Use `logger.info()` to print tensor shapes at each stage
+2. **Compare step-by-step**: Run diffusers and vllm-omni side by side, comparing tensors after each major operation
+3. **Use small configs**: Reduce `num_inference_steps=2`, small resolution for fast iteration
+4. **Test transformer isolation**: Feed the same input to both diffusers and vllm-omni transformers, compare outputs
+5. **Binary search for bugs**: Comment out blocks/layers to isolate where divergence starts

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

@@ -0,0 +1,504 @@
+---
+name: add-tts-model
+description: "Integrate a new text-to-speech model into vLLM-Omni from HuggingFace reference implementation through production-ready serving with streaming and CUDA graph acceleration. Use when adding a new TTS model, wiring stage separation for speech synthesis, enabling online voice generation serving, debugging TTS integration behavior, or building audio output pipelines."
+---
+
+# TTS Model Integration Workflow
+
+## Overview
+
+```
+HF Reference -> Stage Separation -> Online Serving -> Async Chunk -> CUDA Graph -> Pre-commit/DCO
+   (Phase 1)      (Phase 2)          (Phase 3)        (Phase 4)     (Phase 5)      (Phase 6)
+```
+
+Three architecture patterns are supported:
+
+- **Two-stage pipeline** (e.g. Qwen3-TTS, Fish Speech, CosyVoice3): AR
+  code-predictor → audio decoder, connected via async_chunk for low-latency
+  streaming. Use this for maximum performance.
+- **Single-stage AR via generator** (e.g. MOSS-TTS-Nano): entire model runs
+  inside one AR worker, streaming audio chunks through a per-request
+  `inference_stream()` generator. Use this when the upstream model bundles AR
+  + codec inseparably. See [references/single-stage-ar.md](references/single-stage-ar.md).
+- **Single-stage, vLLM-native base LM + side computation** (e.g. VoxCPM2):
+  the base language model runs under vLLM's PagedAttention as a normal AR
+  model; diffusion / VAE / side computations run outside vLLM and are
+  attached via the runner post-processing hook. This is a distinct pattern
+  from the generator approach above — do not confuse the two.
+
+The single-stage variants skip Phase 4 (async_chunk) but Phase 5 (CUDA graph)
+is still encouraged for the inner AR loop.
+
+## Cross-Cutting Invariants
+
+These rules apply to every TTS model regardless of architecture (AR vs AR+diffusion, single-stage vs two-stage, codec-based vs VAE-based). They surface repeatedly across PRs — check them at the end of every phase.
+
+### I1. Streaming output contract
+
+Pick exactly one per-step semantics for `forward()` and document it in the docstring:
+
+- **Delta**: yield only new audio samples produced this step. Preferred — linear cost, low memory.
+- **Cumulative**: re-decode from step 0 every call. O(N²); only acceptable if the codec has no streaming decode path.
+
+If you choose **delta**, verify the full emit→consolidate→consume chain:
+
+1. `forward()` returns `{"model_outputs": <new_chunk_only>, ...}`
+2. `_consolidate_multimodal_tensors()` in `vllm_omni/engine/output_processor.py` concatenates the audio key into one tensor at finish. If it skips the key (`continue`), offline consumers receive only the final chunk. See `output_processor.py` for the concrete list of handled modality keys.
+3. Streaming consumers (SSE, Gradio) receive per-step deltas; offline consumers (`engine.generate()`) receive a single concatenated tensor.
+
+Cumulative-vs-delta mismatch is the most common silent bug — offline RTF benchmarks pass, but users hear replays or truncation.
+
+### I2. Multimodal output consumer hygiene
+
+`outputs[0].outputs[0].multimodal_output[<key>]` can be any of `Tensor`, `list[Tensor]` (pre-consolidation snapshot), `np.ndarray`, or scalar. When writing tests, examples, and benchmarks:
+
+- **Never** use `dict.get("a") or dict.get("b")` on tensor values — Python evaluates the tensor's boolean, raising `RuntimeError: Boolean value of Tensor with more than one value is ambiguous`. Use explicit `if x is None` chains.
+- Always defensively handle the list form: `if isinstance(x, list): x = torch.cat([t.reshape(-1) for t in x], dim=0)`.
+- Assert `shape` / `dtype` / `duration` explicitly; do not rely on truthiness for presence checks.
+
+### I3. Hot-loop GPU discipline
+
+Inside any per-step model loop (AR decode, diffusion solver, CFM Euler, vocoder block loop):
+
+- No `tensor.item()`, `.cpu()`, or `.tolist()` — each triggers a GPU→CPU sync; at 10 steps × 60 frames × 4 ops that is 2400 syncs per request.
+- Prefer `dst.copy_(src)` over `dst.fill_(src.item())` when writing a scalar tensor into a buffer.
+- Prefer `torch.compile(Model.forward, fullgraph=False)` on the whole forward over per-submodule compile — fewer dispatch boundaries, larger fusion regions. Measure before choosing granularity.
+- No Python-side control flow that depends on tensor values; use `torch.where` / masking instead.
+
+Profile first, optimize second. See the profiling docs / project memory for the trace-analysis workflow.
+
+### I4. Validation pyramid
+
+Offline RTF alone is necessary but not sufficient. Every new TTS model must pass all three:
+
+| Layer | Catches | Tool |
+|-------|---------|------|
+| Offline RTF / duration check | Throughput regressions, missing audio, wrong sample rate | `end2end.py`, pytest e2e |
+| Browser streaming playback | Delta/cumulative bugs, chunk boundary glitches, TTFP regressions | Gradio demo over `/v1/audio/speech?stream=true` |
+| Concurrent requests | Per-request state leaks, codec window round-robin gaps | `max_num_seqs>1` smoke test with 4+ parallel prompts |
+
+Declaring a model "done" without all three has shipped regressions more than once.
+
+### I5. Per-request state is owned by the request, not the model
+
+If the model caches *anything* across `forward()` calls (streaming generators, codec buffers, sliding-window pads, CUDA graph state), key it by request ID:
+
+```python
+self._state: dict[str, YourState] = {}    # request_key → state
+# fetch: request_key = str(info.get("_omni_req_id", "0"))
+# free on finish: del self._state[request_key]
+```
+
+A shared buffer silently corrupts audio across concurrent requests — the symptom is crosstalk or truncation only under load.
+
+## Phase 1: HuggingFace Reference
+
+**Goal**: Understand the reference implementation and verify it produces correct audio.
+
+### Steps
+
+1. **Run the reference model** end-to-end using the official HuggingFace / GitHub code
+2. **Document the architecture**:
+   - What are the sub-models? (AR decoder, codec decoder, vocoder, etc.)
+   - What is the token vocabulary? (semantic codes, RVQ codebooks, special tokens)
+   - What is the output format? (sample rate, channels, codec type)
+3. **Capture reference outputs** for comparison during integration
+4. **Identify the config structure**: `config.json` fields, `model_type`, sub-model configs
+
+### Key Questions
+
+- How many codebooks? What are the codebook sizes?
+- What special tokens exist? (`<|voice|>`, `<|audio_start|>`, `<|im_end|>`, etc.)
+- What is the token-to-ID mapping for codec codes?
+- What is the hop length / frame rate of the codec?
+- Does the model support voice cloning? How? (reference audio encoding, speaker embeddings, etc.)
+
+### Deliverables
+
+- Working reference script that produces audio
+- Architecture diagram / notes
+- Token vocabulary mapping
+- Reference audio samples for regression testing
+
+## Phase 2: Stage Separation (Offline Inference)
+
+**Goal**: Split the model into vLLM-Omni stages and get offline inference working.
+
+### Steps
+
+1. **Register the model** in `vllm_omni/model_executor/models/registry.py`
+2. **Create config classes** (`configuration_<model>.py`) with `model_type` registration
+3. **Implement Stage 0** (AR model):
+   - Subclass appropriate base (e.g., wrap Qwen3 decoder layers)
+   - Implement `forward()` for autoregressive token generation
+   - Handle special token logic (start/stop tokens, codec token mapping)
+   - If dual-AR (like Fish Speech), implement Fast AR as a nested module
+4. **Implement Stage 1** (Decoder):
+   - Load codec weights (may need lazy loading from separate checkpoint)
+   - Implement `forward()`: codec codes -> audio waveform
+   - Return `OmniOutput` with `multimodal_outputs`
+5. **Create stage config YAML** defining both stages, memory allocation, and model paths
+6. **Create stage input processor** for prompt building
+7. **Write end2end.py** test script
+
+### Critical Parameters to Get Right
+
+| Parameter | Impact if Wrong |
+|-----------|----------------|
+| Hop length | Audio duration wrong, streaming noise |
+| Token ID mapping | Garbage codes -> noise output |
+| Codebook count/size | Shape mismatch crashes |
+| Stop token | Generation never stops or stops too early |
+| dtype / autocast | Numerical issues, silent quality degradation |
+| Repetition penalty | Must match reference (often 1.0 for TTS) |
+
+### Debugging Priority (from experience)
+
+When audio output is wrong, check in this order:
+
+1. **RoPE / attention**: Are position encodings correct? Is the attention mask right?
+2. **Normalization**: RMSNorm epsilon, layer norm placement (pre vs post)
+3. **Hop length**: Product of all upsample rates in the codec decoder
+4. **Token mapping**: Are codec IDs correctly offset from the vocabulary base?
+5. **Sampling parameters**: Temperature, top_k, top_p, repetition_penalty
+6. **Tensor layout**: Codebook-major vs frame-major ordering
+7. **dtype**: Float32 for codec decoders (autocast can corrupt audio)
+
+### Streaming Correctness Rules (single-stage and two-stage)
+
+These bugs appear in almost every new TTS PR. Check all before the first push. See also the cross-cutting invariants I1 (output contract) and I5 (per-request state) above — the rules below are the Phase 2-specific instances of those invariants:
+
+- **Accumulate codes across AR steps** — each `forward()` appends new codes; do not reset between steps or audio will be truncated (fish speech: `fix: accumulate audio_codes across steps`)
+- **Emit delta audio, not full waveform** — in streaming mode yield only the new chunk per step, not the re-decoded full waveform from step 0 (fish speech: `fix: emit delta audio not full waveform`)
+- **All return paths must emit `model_outputs`** — if any early-return branch skips setting `model_outputs`, the serving layer silently drops that step's audio (fish speech: `fix: ensure ALL return paths emit model_outputs`)
+- **Per-request state isolation** — for batched concurrent requests, key all state by request ID; a shared buffer corrupts audio across requests (fish speech: `fix: per-request vocode + delta emission`)
+- **Codec tensor device** — move codec codes to the codec decoder's device before calling decode; mismatches cause silent CPU fallback or crashes (fish speech: `fix: use model device for CUDA stream`)
+- **AR stage `max_num_seqs`** — set to at least 4 in production configs; for single-stage models this is the only stage. For two-stage models, Stage 0 (AR) needs `max_num_seqs ≥ 4` to pipeline concurrent requests; Stage 1 (codec decoder) typically uses `max_num_seqs: 1` intentionally. Default of 1 everywhere causes audio gaps under concurrency because the codec window round-robins across requests (RFC #2568)
+
+### Optional Dependency Handling
+
+Patch optional dependencies (`torchaudio` / `torchcodec` / `soundfile`) at
+the top of `load_weights()`, not at module import. Failures to do so cause
+cryptic errors only on environments missing the optional package — after
+the model is already deployed. See
+[references/optional-deps.md](references/optional-deps.md) for the full
+pattern, signature constraints, and MOSS-TTS-Nano reference.
+
+### Single-Stage AR Pattern (alternative to two-stage)
+
+When the upstream model cannot be cleanly split into an AR stage and a
+separate decoder, run the full pipeline inside a single AR worker and
+stream audio through a per-request `inference_stream()` generator keyed by
+`_omni_req_id`. Stage config must set `worker_type: ar`,
+`engine_output_type: audio`, `final_output: true`, `is_comprehension: true`,
+and `async_chunk: false` at the top level. Only extract params from
+`additional_information` that you actually forward, or pre-commit fails
+`ruff F841`.
+
+Full walkthrough with the complete `forward()` / `_create_stream_gen()`
+skeleton and stage-config fields:
+[references/single-stage-ar.md](references/single-stage-ar.md). For an
+in-tree reference, look for any single-stage AR model under
+`vllm_omni/model_executor/models/` — e.g. the MOSS-TTS-Nano integration when
+it lands.
+
+**VoxCPM2 is a different pattern** and should not reuse this skeleton — it
+runs the base LM under vLLM PagedAttention with external side-computation.
+See `plan/voxcpm2_native_ar_design.md`.
+
+### Deliverables
+
+- Model files in `vllm_omni/model_executor/models/<model_name>/`
+- Stage config YAML
+- Working `end2end.py` at `examples/offline_inference/text_to_speech/<model>/end2end.py`
+- New section in `examples/offline_inference/text_to_speech/README.md` (table row + per-model section). Do **not** create a top-level `examples/offline_inference/<model>/` dir or a per-model `README.md` inside `text_to_speech/<model>/` — the hub README is the documented surface and the mkdocs `generate_examples` hook only descends one level into `examples/<category>/`.
+
+## Phase 3: Online Serving
+
+**Goal**: Expose the model via `/v1/audio/speech` API endpoint.
+
+### Steps
+
+1. **Register in `serving_speech.py`** — add all 5 points in a **single commit**;
+   partial integration causes hard-to-debug failures. This file is modified by every
+   model PR and is the most common source of rebase conflicts — see conflict note below.
+
+   **Point 1** — stage constant (near the top, alongside the other `_*_TTS_MODEL_STAGES` sets):
+   ```python
+   _YOUR_MODEL_TTS_MODEL_STAGES = {"your_stage_key"}
+   ```
+
+   **Point 2** — union into `_TTS_MODEL_STAGES`:
+   ```python
+   _TTS_MODEL_STAGES: set[str] = (
+       ...
+       | _YOUR_MODEL_TTS_MODEL_STAGES
+   )
+   ```
+
+   **Point 3** — model type detection in `_detect_tts_model_type()`:
+   ```python
+   if model_stage in _YOUR_MODEL_TTS_MODEL_STAGES:
+       return "your_model"
+   ```
+
+   **Point 4** — validation dispatch in `_validate_tts_request()`:
+   ```python
+   if self._tts_model_type == "your_model":
+       return self._validate_your_model_request(request)
+   ```
+
+   **Point 5** — validation + parameter-builder methods:
+   ```python
+   def _validate_your_model_request(self, request) -> str | None:
+       if not request.input or not request.input.strip():
+           return "Input text cannot be empty"
+       return None
+
+   def _build_your_model_params(self, request) -> dict:
+       params = {"text": [request.input]}
+       if request.voice is not None:
+           params["voice"] = [request.voice]
+       return params
+   ```
+   Wire `_build_your_model_params` into `_create_tts_request()` alongside the other
+   model-specific param builders.
+
+   > **Two dispatch patterns coexist**: Fish Speech uses a `self._is_fish_speech` boolean
+   > instance attribute checked before `elif self._is_tts`, while all newer models
+   > (CosyVoice3, MOSS-TTS-Nano) use the `_tts_model_type` string returned by
+   > `_detect_tts_model_type()`. For new models, always use the `_tts_model_type` string
+   > pattern — do not add new `_is_*` flags.
+
+   > **Unused variable rule**: only extract fields in `_build_your_model_params` that
+   > are actually forwarded to the model. Unused extractions fail `ruff F841`.
+   > For voice-cloning fields (`ref_audio` → `prompt_audio_path`, `ref_text` →
+   > `prompt_text`), add them to the param builder and verify they reach the model call.
+
+   **Rebase conflict note**: when rebasing onto `main` after another model was merged,
+   `serving_speech.py` will conflict. Resolution: always keep *both* the upstream
+   model's additions and your own — never discard either side.
+
+2. **Handle model-specific parameters**:
+   - Voice cloning: `ref_audio` encoding and prompt injection
+   - `max_new_tokens` override in sampling params
+   - Model-specific default values
+3. **Create client scripts**: `speech_client.py`, `run_server.sh`
+4. **Test all response formats**: wav, mp3, flac, pcm
+5. **Add Gradio demo**: Interactive web UI with streaming support
+
+### Voice Cloning Pattern
+
+```python
+import base64
+from pathlib import Path
+
+def build_voice_clone_prompt(ref_audio_path: str, text: str, codec) -> list:
+    """Build prompt with reference audio for voice cloning in serving_speech.py."""
+    audio_bytes = Path(ref_audio_path).read_bytes()
+    codes = codec.encode(audio_bytes)  # Encode on CPU using model's codec (e.g., DAC)
+    token_ids = [code + codec.vocab_offset for code in codes.flatten().tolist()]
+    return [
+        {"role": "system", "content": f"<|voice|>{''.join(chr(t) for t in token_ids)}"},
+        {"role": "user", "content": text},
+    ]
+```
+
+### Deliverables
+
+- Updated `serving_speech.py` with all 5 integration points (single commit)
+- Client scripts and server launcher under `examples/online_serving/text_to_speech/<model>/`
+- Gradio demo with streaming and voice cloning UI in the same dir
+- E2E online serving test (`tests/e2e/online_serving/test_<model>.py`)
+- Buildkite CI entry in `.buildkite/test-merge.yml`
+- New section in `examples/online_serving/text_to_speech/README.md` (table row + per-model section). Do **not** create a top-level `examples/online_serving/<model>/` dir or a per-model `README.md` inside `text_to_speech/<model>/`.
+
+### E2E test pitfalls to avoid
+
+- **One `OmniServerParams` set per file.** `omni_server` is module-scoped; a second
+  id in the same file forces mid-module teardown/restart and exposes startup
+  races (`APIConnectionError` on the first request post-restart). Split variants
+  into separate files instead.
+- **No external URL fetches from the server.** CI and some dev hosts can't
+  reach `raw.githubusercontent.com` over TLS. Inline ref audio as
+  `data:audio/wav;base64,...`; the serving layer accepts both URL and data URL.
+- **Use the harness readiness gate.** The fixture waits for HTTP 200 on
+  `/health`; don't add `time.sleep` in tests. If warmup is incomplete, make
+  `/health` return non-200 until you're actually ready.
+- **Mark with `@pytest.mark.core_model` + `hardware_test(res={"cuda": "H100"})`**
+  so the test lands in `test-ready.yml` (triggered by the `ready` label) rather
+  than only nightly.
+
+## Phase 4: Async Chunk (Streaming)
+
+**Goal**: Enable inter-stage streaming so audio chunks are produced while AR generation continues.
+
+### Steps
+
+1. **Update stage config YAML**:
+   ```yaml
+   async_chunk: true
+   codec_chunk_frames: 25      # frames per chunk
+   codec_left_context_frames: 25  # overlap for smooth boundaries
+   ```
+2. **Implement chunk handling in Stage 1**:
+   - Accept partial input (chunk of codec codes)
+   - Handle left context for smooth audio boundaries
+   - Return partial audio in `OmniOutput`
+3. **Test streaming**:
+   - Verify audio quality matches non-streaming output
+   - Check for artifacts at chunk boundaries
+   - Measure TTFA (time to first audio)
+4. **Update online serving** to support `stream=true` with PCM output
+
+### Streaming Architecture
+
+```
+Stage 0 (AR)                    Stage 1 (Decoder)
+  |                                |
+  |-- chunk 0 (25 frames) ------> decode -> audio chunk 0 -> client
+  |-- chunk 1 (25 frames) ------> decode -> audio chunk 1 -> client
+  |-- chunk 2 (25 frames) ------> decode -> audio chunk 2 -> client
+  ...
+```
+
+### Key Considerations
+
+- **Left context overlap**: Prevents audible artifacts at chunk boundaries
+- **Hop length matters**: `context_audio_samples = context_frames * hop_length`
+- **First chunk latency**: Can use larger initial chunk for better quality, then smaller chunks
+
+### Deliverables
+
+- Updated stage config with async_chunk enabled
+- Smooth streaming audio without boundary artifacts
+- TTFA metrics
+
+## Phase 5: CUDA Graph Acceleration
+
+**Goal**: Capture the AR loop as a CUDA graph for significant speedup.
+
+### Steps
+
+1. **Identify the hot loop**: The AR decoding loop that runs N steps per token
+2. **Create static buffers**:
+   - KV caches with fixed max sequence length
+   - Pre-built causal masks and position tensors per step
+   - Static input/output tensors
+3. **Implement graph capture**:
+   - Warm up with real data
+   - Capture the forward pass
+   - Replay with updated inputs
+4. **Handle constraints**:
+   - Use `torch.argmax` instead of `torch.multinomial` (graph-safe)
+   - Fixed batch size (fall back to eager for other sizes)
+   - No dynamic control flow inside the graph
+
+See [references/cuda-graph-example.md](references/cuda-graph-example.md) for
+a worked skeleton (Qwen3-TTS code predictor, 16-step AR loop), performance
+expectations (3–5× on the graphed component for fixed batch_size=1), and the
+graph-safety constraints you must honor inside the captured region.
+
+### Deliverables
+
+- CUDA graph implementation for the AR hot loop
+- Benchmark script comparing eager vs graph performance
+- Documentation of constraints and fallback behavior
+
+## Phase 6: Pre-commit and DCO
+
+**Goal**: Every commit passes `pre-commit` lint and carries a DCO
+`Signed-off-by` line that matches the author email.
+
+- Install hooks once: `pre-commit install`.
+- Run `pre-commit run --files <changed-files>` before every push; accept any
+  auto-fixes, stage, re-commit.
+- Sign every commit with `git commit -s`. DCO checks that author email and
+  `Signed-off-by` email match — `git config user.email` must match your
+  GitHub account email.
+
+Common pre-commit failures, recovery commands for missing sign-off, and the
+full `pre-commit run` invocation for a TTS model:
+[references/precommit-dco.md](references/precommit-dco.md).
+
+## Integration Checklist
+
+Use this checklist when integrating a new TTS model:
+
+### Cross-Cutting Invariants (verify at end of every phase)
+- [ ] I1: `forward()` docstring states cumulative vs delta; consolidation path audited end-to-end
+- [ ] I2: Tests / examples / benchmarks never use `dict.get(a) or dict.get(b)` on tensor values; list form handled
+- [ ] I3: No `.item()` / `.cpu()` / Python branch on tensor values inside per-step loops
+- [ ] I4: Offline RTF, browser streaming playback, and concurrent-request smoke test all pass
+- [ ] I5: Any cross-step cache keyed by `_omni_req_id`; entries freed when the request finishes
+
+### Phase 1: HF Reference
+- [ ] Reference model runs and produces correct audio
+- [ ] Architecture documented (stages, codebooks, tokens, sample rate)
+- [ ] Reference audio samples saved for comparison
+
+### Phase 2: Stage Separation
+- [ ] Model registered in `registry.py`
+- [ ] Config classes created with `model_type` registration
+- [ ] Stage 0 (AR) implemented and generates correct tokens
+- [ ] Stage 1 (Decoder) produces correct audio from tokens — dtype float32 for codec decoder
+- [ ] Stage 1 `max_num_seqs` ≥ 4 in production config (default 1 causes gaps under concurrency)
+- [ ] Optional dependency fallbacks handled at `load_weights()` time (torchaudio/soundfile/etc.)
+- [ ] Streaming: codec codes accumulated across AR steps (not reset per step)
+- [ ] Streaming: delta audio emitted per chunk, not full re-decoded waveform
+- [ ] Streaming: all `forward()` return paths emit `model_outputs`
+- [ ] Streaming: per-request state keyed by request ID (not shared across requests)
+- [ ] Streaming: codec tensors moved to codec decoder device before decode
+- [ ] Stage config YAML created
+- [ ] `end2end.py` produces audio matching reference quality
+- [ ] README.md written
+
+### Phase 3: Online Serving
+- [ ] All 5 `serving_speech.py` integration points added in one commit
+- [ ] Only extract params in `_build_*_params` that are forwarded to the model call (ruff F841)
+- [ ] Prompt builder handles text input correctly
+- [ ] Voice cloning works (if supported)
+- [ ] All response formats work (wav, mp3, flac, pcm)
+- [ ] Client scripts and server launcher created
+- [ ] E2E online serving test written (`tests/e2e/online_serving/test_<model>.py`)
+- [ ] Buildkite CI entry added to `.buildkite/test-merge.yml`
+- [ ] Gradio demo working
+- [ ] Documentation added (offline + online docs, nav, supported models)
+
+### Phase 4: Async Chunk
+- [ ] Stage config updated with `async_chunk: true`
+- [ ] Stage 1 handles partial chunks correctly
+- [ ] No audio artifacts at chunk boundaries
+- [ ] Streaming via API (`stream=true`) works
+- [ ] TTFA measured and acceptable
+
+### Phase 5: CUDA Graph
+- [ ] Hot loop identified and profiled
+- [ ] Static buffers allocated
+- [ ] Graph captured and replays correctly
+- [ ] Benchmark shows meaningful speedup
+- [ ] Fallback to eager works for unsupported configs
+
+### Phase 6: Pre-commit and DCO
+- [ ] `pre-commit run --files <changed>` passes before every push
+- [ ] Every commit has `Signed-off-by` matching the author email (`git commit -s`)
+- [ ] `git config user.email` matches the email registered on your GitHub account
+- [ ] Details and failure-recovery commands: [references/precommit-dco.md](references/precommit-dco.md)
+
+## References
+
+In-skill references (details split out of the main body):
+
+- [references/single-stage-ar.md](references/single-stage-ar.md) — full `forward()` / generator skeleton for the MOSS-TTS-Nano-style pattern
+- [references/optional-deps.md](references/optional-deps.md) — torchaudio / torchcodec fallback pattern
+- [references/cuda-graph-example.md](references/cuda-graph-example.md) — Qwen3-TTS code-predictor CUDA graph skeleton
+- [references/precommit-dco.md](references/precommit-dco.md) — full pre-commit invocation, failure table, DCO recovery
+
+Project docs and adjacent skills:
+
+- [TTS audio skill](../vllm-omni-audio-tts/SKILL.md) — supported models and usage
+- [Fish Speech integration](../vllm-omni-audio-tts/references/fish-speech.md) — complete example of Phases 1–3
+- [Qwen3-TTS reference](../vllm-omni-audio-tts/references/qwen-tts.md) — complete example of all 5 phases
+- [Adding a TTS model (developer guide)](https://github.com/vllm-project/vllm-omni/blob/main/docs/contributing/model/adding_tts_model.md)
+- `plan/voxcpm2_native_ar_design.md` — VoxCPM2's vLLM-native AR + side-computation pattern (distinct from the generator-based single-stage described above)

.claude/skills/add-tts-model/references/cuda-graph-example.md

@@ -0,0 +1,42 @@
+# CUDA Graph Example: Qwen3-TTS Code Predictor
+
+Reference sketch for capturing the 16-step code-predictor AR loop as a single
+CUDA graph. Adapt the shapes, number of steps, and KV-head layout to your
+model.
+
+```python
+import torch
+
+class CodePredictorGraph:
+    """Captures the 16-step code predictor AR loop as a single CUDA graph."""
+
+    def setup_graph(self, device: torch.device, kv_heads: int = 4, head_dim: int = 64):
+        self.num_steps = 16
+        self.kv_cache = torch.zeros(1, kv_heads, self.num_steps, head_dim, device=device)
+        self.positions = torch.arange(self.num_steps, device=device)
+        self.causal_mask = torch.tril(torch.ones(self.num_steps, self.num_steps, device=device))
+        self.input_buf = torch.zeros(1, 1, kv_heads * head_dim, device=device)
+        self.output_buf = torch.zeros(1, self.num_steps, device=device, dtype=torch.long)
+        # Warm up, then: self.graph = torch.cuda.CUDAGraph(); self.graph.capture(...)
+
+    def run_graph(self, initial_input: torch.Tensor) -> torch.Tensor:
+        self.input_buf.copy_(initial_input)
+        self.graph.replay()
+        return self.output_buf.clone()
+```
+
+## Performance expectations (Qwen3-TTS code predictor)
+
+- **3–5× speedup** on the graphed component.
+- Effective only for fixed batch sizes (typically `batch_size=1`).
+- Fall back to eager for any shape/config that wasn't captured — do not try
+  to recapture per request.
+
+## Graph-safety constraints
+
+- `torch.argmax` instead of `torch.multinomial`.
+- Fixed batch size.
+- No Python control flow that branches on tensor values inside the captured
+  region (use `torch.where` / masks).
+- No `.item()`, `.cpu()`, `.tolist()` — each would break the capture or
+  cause a GPU→CPU sync during replay.

.claude/skills/add-tts-model/references/optional-deps.md

@@ -0,0 +1,47 @@
+# Optional Dependency Handling
+
+Models that rely on `torchaudio`, `torchcodec`, `soundfile`, or other optional
+packages must handle the missing-package case at import time, not at call
+time. Failing to do this causes cryptic errors only on environments without
+the optional package — after the model is already deployed.
+
+## Pattern (used in MOSS-TTS-Nano)
+
+```python
+def _patch_torchaudio_load() -> None:
+    """Fallback torchaudio.load/save to soundfile if torchcodec is unavailable."""
+    try:
+        import torchcodec  # noqa: F401
+        return  # torchcodec present, torchaudio works as-is
+    except ImportError:
+        pass
+
+    import soundfile as sf
+    import torchaudio
+
+    def _sf_load(path, **kwargs):
+        data, sr = sf.read(str(path), dtype="float32", always_2d=True)
+        return torch.from_numpy(data).T, sr
+
+    torchaudio.load = _sf_load
+    # patch .save similarly if needed
+```
+
+## Rules
+
+- Mirror the full signature of the replaced function. `torchaudio.load`
+  accepts `frame_offset`, `num_frames`, `normalize`, `channels_first`,
+  `format` — missing any of them causes `TypeError` from calling code.
+- Catch `except Exception`, not just `ImportError`. `import torchaudio`
+  itself can fail with non-`ImportError` errors on broken installs. Log the
+  exception type and message (`logger.warning("torchaudio probe failed: %s: %s",
+  type(exc).__name__, exc)`) before falling back, so unrelated errors are not
+  silently swallowed.
+- Call the patch function at the top of `load_weights()` before loading any
+  audio assets. Do not call it at module import time.
+
+## Reference implementation
+
+Any in-tree model that patches `torchaudio.load` in its `load_weights()` —
+e.g. MOSS-TTS-Nano's `modeling_moss_tts_nano.py` once that integration
+lands.

.claude/skills/add-tts-model/references/precommit-dco.md

@@ -0,0 +1,54 @@
+# Pre-commit and DCO
+
+Every commit must pass `pre-commit` lint and carry a `Signed-off-by` line
+that matches the commit author email.
+
+## Pre-commit
+
+Install hooks once:
+
+```bash
+pre-commit install
+```
+
+Run before every push on the files you changed:
+
+```bash
+pre-commit run --files \
+  vllm_omni/model_executor/models/<model_name>/*.py \
+  vllm_omni/entrypoints/openai/serving_speech.py \
+  vllm_omni/model_executor/models/registry.py \
+  tests/e2e/offline_inference/test_<model_name>.py \
+  tests/e2e/online_serving/test_<model_name>.py
+```
+
+When pre-commit **modifies files** (ruff format auto-fix), it exits non-zero
+but the changes are correct — stage the modified files and re-commit.
+
+| Failure | Root cause | Fix |
+|---------|-----------|-----|
+| `ruff F841` | Variable extracted but never forwarded to model call | Remove the extraction or wire it through |
+| `ruff E402` | Import added below function definitions | Move to top-level import block |
+| `ruff format` | Line length, spacing, quote style | Accept auto-fix, stage, re-commit |
+
+## DCO sign-off
+
+Every commit must carry `Signed-off-by: Your Name <your@email.com>`. Use
+`-s`:
+
+```bash
+git commit -s -m "feat(<model>): add <Model> TTS support"
+```
+
+Or set it permanently: `git config format.signOff true`.
+
+The DCO check verifies that the commit author email matches the
+`Signed-off-by` line. Confirm `git config user.email` matches your GitHub
+account email before committing.
+
+Fix a missing or mismatched sign-off on the latest commit:
+
+```bash
+git commit --amend -s --no-edit
+git push origin <branch> --force-with-lease
+```

.claude/skills/add-tts-model/references/single-stage-ar.md

@@ -0,0 +1,108 @@
+# Single-Stage AR Pattern
+
+When the upstream model cannot be cleanly split into an AR stage and a separate
+decoder (e.g. MOSS-TTS-Nano, or any model that bundles AR + codec via an
+`inference_stream()` generator), run the whole pipeline inside a single AR
+worker that yields audio chunks per request.
+
+This is distinct from VoxCPM2's pattern, which also runs in a single stage but
+uses vLLM's native PagedAttention on the base language model with diffusion /
+VAE side-computation outside vLLM — see
+`plan/voxcpm2_native_ar_design.md` for that variant.
+
+## Implementation
+
+1. **Single model file** — load both AR LM and codec inside
+   `modeling_<model>.py`.
+2. **Load weights in `load_weights()`**, not `__init__()` — vLLM initializes
+   distributed state before any CUDA allocations.
+3. **Stream via a per-request generator** stored in `self._stream_gens`:
+
+```python
+class YourModelForCausalLM(nn.Module):
+    def __init__(self, *, vllm_config, prefix=""):
+        super().__init__()
+        self._lm = None                   # populated in load_weights()
+        self._stream_gens: dict = {}      # request_key → generator
+
+    def load_weights(self, weights):
+        # Load self._lm here, after vLLM distributed init
+        ...
+
+    def forward(
+        self,
+        input_ids,
+        positions,
+        intermediate_tensors=None,
+        inputs_embeds=None,
+        runtime_additional_information: list[dict] | None = None,  # one dict per request
+        **kwargs,
+    ) -> OmniOutput:
+        infos = runtime_additional_information or [{}]
+        # Skip dummy/profiling calls
+        if not runtime_additional_information or all(i.get("_is_dummy") for i in infos):
+            self._ar_emit_stop_token = True
+            return OmniOutput(...)  # return empty outputs
+
+        outputs, last_flags = [], []
+        for info in infos:
+            request_key = str(info.get("_omni_req_id", "0"))  # per-request ID from vLLM
+            if request_key not in self._stream_gens:
+                self._stream_gens[request_key] = self._create_stream_gen(info)
+            try:
+                chunk, is_last = next(self._stream_gens[request_key])
+            except StopIteration:
+                chunk, is_last = torch.zeros(0), True
+            if is_last:
+                del self._stream_gens[request_key]
+            outputs.append(chunk)
+            last_flags.append(is_last)
+
+        self._ar_emit_stop_token = all(last_flags)
+        return OmniOutput(multimodal_outputs={"model_outputs": outputs, ...})
+
+    def _create_stream_gen(self, info: dict):
+        """Yield (waveform_tensor, is_last) tuples from inference_stream()."""
+        for event in self._lm.inference_stream(...):
+            if event["type"] == "audio":
+                yield event["waveform"], False
+            elif event["type"] == "result":
+                # Fallback: some models emit a single "result" event instead of
+                # incremental "audio" events — handle both paths
+                yield event.get("waveform", torch.zeros(0)), True
+                return
+        yield torch.zeros(0), True
+
+    def compute_logits(self, hidden_states, sampling_metadata):
+        # Emit EOS only after the last chunk so the AR scheduler ends the request
+        ...
+```
+
+## Key points
+
+- `runtime_additional_information` is the correct parameter name (not
+  `**kwargs`) — it carries one dict per request in the batch.
+- The request ID is `info.get("_omni_req_id")` — set by vLLM, not by user code.
+- Handle both `"audio"` (incremental) and `"result"` (final combined) event
+  types from upstream models.
+
+## Stage config
+
+Single stage with `worker_type: ar`, `engine_output_type: audio`,
+`final_output: true`, `is_comprehension: true`, and `async_chunk: false` at
+the top level. Omitting any of these causes silent misclassification in the
+serving layer.
+
+## Lint discipline
+
+Only extract variables from `additional_information` that you actually
+forward to the model call — unused extractions trip `ruff F841` in
+pre-commit.
+
+## Reference implementation
+
+Look for any single-stage AR model under
+`vllm_omni/model_executor/models/` — e.g. `moss_tts_nano/` when its
+integration lands. If none is in tree yet, follow the skeleton above and
+cross-check against the `is_comprehension: true` / `async_chunk: false`
+dispatch in `vllm_omni/entrypoints/openai/serving_speech.py`.

.claude/skills/diffusion-perf-opt/SKILL.md

@@ -0,0 +1,510 @@
+---
+name: diffusion-perf-opt
+description: Diagnose and optimize vLLM Omni diffusion workloads, especially Wan/Qwen/Flux-style image and video generation. Use when Codex is asked to analyze profiling traces, choose parallel strategies, inspect torch profiler trace.json or trace.json.gz timelines, estimate optimization ROI, investigate GPU idle/free bubbles, compare USP/CFG/HSDP/VAE parallelism, or design operator/host/quantization optimizations for vLLM Omni.
+---
+
+# vLLM Omni Performance Optimization
+
+Use this skill to run a disciplined optimization loop for vLLM Omni diffusion workloads. Keep two ideas separate: real performance baselines are collected with low overhead, while torch profiler traces are diagnostic artifacts and may distort latency.
+
+## First Questions
+
+Before proposing changes, ask for the optimization scene if it is not already known:
+
+- GPU model, card count, topology, and whether NVLink is present.
+- Model and pipeline, for example Wan2.2 I2V A14B.
+- User workload: resolution, frames, steps, batch/concurrency, CFG scales, prompt/image inputs.
+- Three runnable commands for each model/workload, or scripts that generate them:
+  - **Server startup command**: exact `vllm serve` command, environment variables, model path, port, parallelism flags, profiler flags, and precision/compile settings.
+  - **User request command**: exact single-request client command (for example curl or Python client) with fixed prompt/media/seed/size/steps. Use this to validate correctness and collect per-request stage timings.
+  - **User benchmark command**: exact repeatable benchmark command or script with warmup count, measured iteration count, concurrency/batch policy, output directory, and summary format.
+- Current enabled strategies: USP/SP, CFG parallel, HSDP/FSDP, VAE patch parallel, torch.compile, profiler options.
+- Optimization target: latency, throughput, memory, cost, or quality-preserving speed.
+- Precision/quality tolerance: bf16/fp8/quantization/sparsity/approximate attention allowed or not.
+- Quality validation method: exact-output comparison when feasible, or metrics
+  such as SSIM, PSNR, LPIPS, cosine similarity, MAE, MSE, temporal flicker
+  checks, baseline self-run variance, and reviewable output artifacts.
+
+## Workflow
+
+1. **Freeze the measurement protocol and commands.**
+   - Before analyzing parallel strategies or traces, establish the three commands for the target model/workload: server startup, single user request, and user benchmark.
+   - Prefer checked-in scripts or one-off shell scripts over free-form commands in chat. The scripts should make workload variables explicit, including model path, port/base URL, prompt/media inputs, resolution, frames, steps, seed, warmup, measured iterations, concurrency, and output path.
+   - Measurement reports must preserve the concrete process, not just the final numbers. For each tested configuration, write down the server startup command, the user request command, the user benchmark command or polling loop, the final server response metrics, and the client-side timing/HTTP result.
+   - Present all duration and latency values in **milliseconds (ms)** in tables and summaries. If an API returns seconds, convert to ms before reporting. Keep raw units only inside quoted raw logs or source snippets.
+   - Keep profiler modes separate in the command set:
+     - Baseline/benchmark commands should avoid torch profiler and stack collection.
+     - Diagnostic commands may enable torch profiler after a baseline identifies a bottleneck.
+   - If any of the three commands is missing, ask for it or create a proposed script before proceeding.
+   - Make clear which command is authoritative for correctness validation and which command is authoritative for performance numbers.
+
+2. **Collect a real baseline.**
+   - Disable PyTorch profiler and stack collection.
+   - Disable or fix torch.compile state so A/B is fair.
+   - Avoid `--enforce-eager` for production-speed baselines unless eager is the target.
+   - Prefer `--log-stats` and `--enable-diffusion-pipeline-profiler` for low-overhead stage timing. PR 3069 has the relevant metrics/log-stats changes; if local code does not include them, fetch or cherry-pick the minimal metrics changes rather than merging unrelated PR drift.
+   - Run warmup requests and exclude first-request lazy init.
+
+3. **Model the parallel strategy before testing.**
+   - Estimate compute and communication for self-attention, cross-attention, FFN, CFG, VAE encode/decode, and HSDP.
+   - Select a small candidate matrix rather than exhaustively testing everything.
+   - Typical 2-card candidates: `USP=2`, `CFG=2`, `USP=1/HSDP on-off`, VAE parallel on-off if memory allows.
+   - Typical 4-card candidates: `USP=4`, `CFG=2 x USP=2`, `USP=2 x HSDP`, VAE parallel on-off.
+   - Typical 8-card candidates for official CFG-enabled video diffusion models:
+     - Primary candidate: `CFG=2 x USP=4` with VAE patch parallel across all 8 ranks.
+     - Compare against: `CFG=1 x USP=8` to test whether larger sequence parallel groups beat CFG branch parallelism.
+     - Isolate VAE parallelism: keep the DiT strategy fixed and compare VAE patch parallel on/off or different VAE patch sizes.
+     - If long-video `diffuse` remains dominant and Ulysses all-to-all is suspected, test a hybrid sequence strategy such as `CFG=2 x USP=2 x Ring=2`.
+     - Test HSDP off only if the model fits without it; treat HSDP primarily as a memory strategy until A/B proves otherwise.
+   - Prefer USP/SP for long video token sequences; prefer CFG parallel when CFG doubles transformer forwards and sequence length is modest.
+   - Convert workload shape to latent/patch token counts before choosing candidates. For video models, record frames, latent frames, latent HxW, patch size, approximate token count, and which stages should scale with the token count.
+
+4. **Search for the best parallel configuration.**
+   - Start with a small matrix that answers one question per comparison:
+     - `CFG parallel vs larger USP`: compare `CFG=2 x USP=world/2` against `CFG=1 x USP=world` for CFG-enabled workloads.
+     - `VAE patch parallel value`: compare the best DiT strategy with VAE patch parallel enabled and disabled.
+     - `HSDP cost`: compare HSDP on/off only when both configurations fit memory.
+     - `Ulysses vs Ring`: test a Ring/Ulysses hybrid only after long-sequence `diffuse` is confirmed dominant.
+   - For each configuration, create or record a stable config id, for example `A_cfg2_usp4_vaepp8_hsdp_tiling`.
+   - For each config id, capture:
+     - exact server command and environment variables,
+     - observed distributed setup from logs, such as SP groups, CFG groups, HSDP shard/replicate sizes, VAE patch size,
+     - exact request command for every scenario,
+     - final server response metrics and client-side elapsed time,
+     - output artifact paths and any failed/empty responses.
+   - Use one warmup request per scenario, then at least three measured repeats for the shortlisted config. For exploratory matrix pruning, one measured request is acceptable only if the margin is large; label it as one-shot.
+   - Compare configurations by stage, not only end-to-end latency. A configuration can improve `diffuse` while hurting `vae.decode`; record both effects.
+   - Select the best config only after checking the target metric, dominant stages, memory headroom, and output correctness.
+
+5. **Run targeted A/B tests.**
+   - Change one variable per test.
+   - Keep model, input, seed, request parameters, GPU placement, and warmup policy fixed.
+   - Record latency, stage timings, memory, output quality, and logs.
+   - Report comparison tables in ms. Include at least end-to-end client time, server `inference_time`, server stage generation time if available, `vae.encode`, `diffuse`, `vae.decode`, and peak memory.
+
+6. **Enforce the quality and precision gate for every optimization.**
+   - Do not mark an optimization ready until quality is checked on the same
+     workload: model, prompt/media, seed, shape, steps, scheduler, dtype,
+     backend, parallelism, and output encoding.
+   - For math-preserving changes, require exact/near-exact agreement or show
+     that differences are within baseline self-run variance.
+   - For precision, quantization, approximate attention, sparse/custom kernels,
+     or backend changes, include reviewable artifacts and metrics such as SSIM,
+     PSNR, LPIPS, cosine similarity, MAE, MSE, and temporal flicker checks.
+   - If quality tolerance is not stated, default to quality-preserving behavior.
+     Failed or inconclusive quality validation blocks a ready-to-merge claim.
+
+7. **Collect diagnostic trace only after narrowing hypotheses.**
+   - Use torch profiler for a small number of requests.
+   - Run two separate diagnostic traces instead of mixing concerns:
+     - **Operator/shape trace**: enable `torch_profiler_record_shapes=True` and keep stack collection disabled. Use this to rank CUDA kernels, NCCL collectives, attention/MLP/norm/RoPE work, and shape-specific hot operators.
+     - **Host-stack trace**: enable `torch_profiler_with_stack=True` and normally keep shape collection disabled. Use this to map CPU/Python host gaps, synchronization points, scheduler paths, and request handling overhead.
+   - Keep both trace commands and reports separate from baseline/benchmark commands. Torch profiler latency is diagnostic only and must not be used as the final latency claim.
+   - Prefer profiling only the narrowed dominant scenario, for example the highest-resolution/video-length case where `diffuse` dominates.
+   - If profiler endpoints are available, run one warmup request first, then call `/start_profile`, run one profiled request, and call `/stop_profile`. This keeps model initialization and warmup out of the diagnostic trace.
+   - Start by analyzing rank 0 only. Expand to more ranks only if rank 0 suggests imbalance, unclear GPU idle/free bubbles, high NCCL wait, server timing mismatch, CFG branch imbalance, or USP group stragglers.
+   - Diagnostic reports must be written to disk and preserve: server command, profiler config, warmup/request/polling commands, trace artifact paths, rank analyzed, analyzer output or summary, and the decision about whether additional ranks are necessary.
+   - Analyze both rank-level balance and device-level free bubbles when additional ranks are opened.
+
+8. **Analyze host, communication, and operators.**
+   - Find GPU idle/free intervals and map each large gap to the enclosing CPU/Python code.
+   - Separate real GPU idle from profiler overhead such as CUPTI `Command Buffer Full`.
+   - Compare NCCL kernel time to user annotations; annotations can overcount nested intervals.
+   - Rank operator work by total CUDA time and by repeated small-kernel launch count.
+
+9. **Produce an optimization plan.**
+   - Classify candidates as P0/P1/P2.
+   - For each candidate, state necessity, expected benefit, implementation path, validation plan, and quality risk.
+   - Include a concrete quality/precision validation gate for each candidate.
+   - Do not implement high-risk operator rewrites before proving the operator is a bottleneck for the target shapes.
+   - End the plan with a user-facing candidate selection table. The assistant
+     should not automatically choose a risky optimization just because it is
+     technically possible. Present the options clearly and let the user decide
+     which item is worth implementing based on latency target, engineering
+     budget, memory headroom, and quality tolerance.
+   - Organize the plan by optimization layer, not by a flat list of ideas:
+     - host/runtime optimization,
+     - measurement/benchmark reliability,
+     - parallelism and communication,
+     - VAE encode/decode and media pre/post processing,
+     - operator fusion and layout cleanup,
+     - attention main-path optimization,
+     - algorithmic, precision, or approximation changes.
+   - For each layer, tie every candidate back to evidence from baseline
+     metrics, diagnostic trace, source code, or output quality requirements.
+     Do not list generic optimizations without a trace or workload reason.
+
+## Priority Rules
+
+- **P0:** low risk, likely useful, or required for trustworthy measurement. Examples: real baseline, warmup, targeted parallel A/B, disabling avoidable `empty_cache`, scheduler coefficient caching.
+- **P1:** meaningful code changes with contained risk. Examples: cross-attention KV caching, VAE gather/broadcast reduction, AdaLayerNorm/RMSNorm/RoPE fusion after trace evidence.
+- **P2:** high implementation or quality risk. Examples: FA to LA replacement, custom Triton/CUDA fused kernels, FP8/quantization, sparsity/Rainfusion-style acceleration.
+
+Every implemented optimization needs A/B validation and a passing
+quality/precision gate, including math-preserving changes such as padding trim,
+layout cleanup, cache reuse, or host/runtime cleanup. Use objective metrics
+when available, keep reviewable artifacts, and compare against baseline
+self-run variance for precision, quantization, approximate kernels, or backend
+changes. If quality validation fails or is inconclusive, do not present the
+optimization as ready to merge.
+
+## Optimization Layers
+
+After baseline, parallel-search, and diagnostic traces, summarize optimization
+opportunities by layer. This is the core of the performance analysis: the goal
+is to connect evidence to a scoped implementation and a validation plan.
+
+### Host and Runtime Optimization
+
+Purpose: remove CPU/Python stalls, synchronization points, allocator overhead,
+and request-path overhead that leave GPU lanes empty.
+
+Evidence to look for:
+
+- High `idle_pct` or large `GAP` blocks in `trace_analyzer.py`.
+- Host-stack trace lines such as `torch.cuda.empty_cache`,
+  `cudaStreamSynchronize`, `cudaDeviceSynchronize`, Python locks, scheduler
+  waits, image/video preprocessing, or repeated small allocation paths.
+- Difference between client wall-clock time and server `inference_time_s`.
+
+Typical candidates:
+
+- Make avoidable `torch.cuda.empty_cache()` optional or guard it by memory
+  headroom.
+- Cache scheduler coefficients, timesteps, masks, or other tiny repeated CPU
+  computations when the request shape/steps are fixed.
+- Remove avoidable host-device synchronizations and blocking logging/stat calls.
+- Move expensive preprocessing out of the critical path or cache fixed prompt,
+  image, and transform work for benchmark scenarios.
+- Ensure benchmark scripts record client-side elapsed time, HTTP status, output
+  path, and server response metrics.
+
+Priority guidance:
+
+- Usually P0 when the change is measurement reliability or an obvious removable
+  synchronization.
+- Usually P1 when it changes scheduling, memory lifetime, or request execution
+  order.
+
+Validation:
+
+- Re-run non-profiler baseline with same workload and seed.
+- Confirm peak memory headroom if disabling cache cleanup.
+- Confirm generated output exists and quality/seed behavior is unchanged.
+
+### Parallelism and Communication Optimization
+
+Purpose: choose the right decomposition for CFG branches, sequence tokens, model
+weights, VAE tiles, and rank topology.
+
+Evidence to look for:
+
+- Baseline A/B across `CFG`, `USP/SP`, `Ring`, `HSDP/FSDP`, and VAE patch
+  parallelism.
+- Stage timing shifts: `diffuse`, `vae.encode`, `vae.decode`, and server
+  end-to-end.
+- NCCL kernel time from trace, not only `user_annotation` time.
+- Rank imbalance across SP group ranks or CFG branch ranks.
+- Memory headroom and OOM risk.
+
+Typical candidates:
+
+- `CFG=2 x USP=world/2` versus `CFG=1 x USP=world` for CFG-enabled models.
+- VAE patch parallel on/off or patch size tuning.
+- HSDP/FSDP on/off only if both configurations fit memory.
+- Ulysses versus Ulysses+Ring only after long-sequence `diffuse` is confirmed
+  dominant and all-to-all is suspected.
+- Rank mapping/topology changes if all-rank traces show stragglers or NCCL wait.
+- Buffer reuse or preallocation for FSDP/HSDP all-gather paths.
+
+Priority guidance:
+
+- P0/P1 for configuration-only changes with strong measured wins.
+- P1 for buffer reuse or rank mapping changes.
+- P2 for invasive distributed algorithm changes.
+
+Validation:
+
+- Measure by stage and memory, not only end-to-end.
+- Use one-variable A/B with identical prompt/media/seed/shape/steps.
+- When communication is suspected, compare rank0-3 in one USP group and rank0
+  versus rank4 across CFG branches for `CFG=2 x USP=4`.
+
+### VAE and Media Pipeline Optimization
+
+Purpose: reduce encode/decode, tiling, split/gather, and media conversion time.
+
+Evidence to look for:
+
+- Large `vae.encode` or `vae.decode` in low-overhead stage timings.
+- Host-stack gaps in VAE tile split, gather, merge, broadcast, or image/video
+  transforms.
+- VAE kernels or cuDNN convolution in operator trace.
+- Whether every rank needs the final decoded tensor.
+
+Typical candidates:
+
+- Keep VAE patch parallel enabled when it has clear measured benefit.
+- Reduce VAE gather/broadcast to only ranks that need the final media output.
+- Reuse tile metadata, split buffers, or gather buffers.
+- Evaluate bf16/autocast behavior for VAE only with visual quality checks.
+- Avoid redundant image conversion, resize, or tensor construction in repeated
+  benchmark runs.
+
+Priority guidance:
+
+- P0/P1 if VAE is a large share of the target workload or if a host gap is
+  obvious and low risk.
+- Lower priority when `diffuse` dominates and VAE is already patch-parallelized.
+
+Validation:
+
+- Compare `vae.encode`, `vae.decode`, server end-to-end, and peak memory.
+- Check output video integrity, artifacts, flicker, and seed stability.
+
+### Operator Fusion and Layout Cleanup
+
+Purpose: reduce high-frequency small kernels, memory bandwidth pressure, layout
+conversions, and launch overhead in transformer and VAE blocks.
+
+Evidence to look for:
+
+- Top operator tables showing many `aten::copy_`, `aten::cat`,
+  `split_with_sizes_copy`, `aten::add`, `aten::mul`, `aten::div`, norm,
+  activation, RoPE, or reshape/layout kernels.
+- `ops_rankN.xlsx` `by_shape` sheet showing repeated small shapes inside the
+  same block path.
+- Trace lanes showing many short kernels between larger GEMM/attention kernels.
+- Source code patterns with repeated elementwise chains or layout conversions.
+
+Typical fusion targets:
+
+- AdaLayerNorm / RMSNorm / LayerNorm plus scale/shift fusion.
+- RoPE fusion with Q/K layout preparation when shapes are stable.
+- Residual add, scale, gate, and elementwise chains.
+- MLP gate/up/down cleanup, such as fusing activation and multiply around
+  GEMM outputs when feasible.
+- QKV projection and reshape/split/cat path cleanup.
+- Attention pre/post layout cleanup to avoid unnecessary copies, cats, and
+  splits around sequence parallel all-to-all.
+
+Priority guidance:
+
+- P1 when implemented with existing PyTorch/Triton/local helper patterns and
+  validated against exact outputs or tolerances.
+- P2 when it requires custom CUDA/Triton kernels, changes numerics, or touches
+  attention math directly.
+
+Validation:
+
+- First prove the operator family is material for the target shape.
+- Use non-profiler A/B for latency and stage timing.
+- Use quality regression checks for generated video stability.
+- Check compile behavior and graph breaks if using `torch.compile`.
+
+### Attention Main-Path Optimization
+
+Purpose: address the dominant self-attention cost when FlashAttention or other
+attention kernels dominate CUDA time.
+
+Evidence to look for:
+
+- Operator trace where FlashAttention/SDPA kernels dominate total CUDA time.
+- Attention shape from `ops_rankN.xlsx` `by_shape`, model code, or trace
+  metadata.
+- Whether attention cost scales with latent frames, latent H/W, patch size, or
+  CFG duplication.
+- Layout/copy/all-to-all work around attention.
+
+Typical candidates:
+
+- Verify the attention backend and shape are on the intended fast kernel path.
+- Compare supported attention backends only with identical workload and quality
+  settings.
+- Reduce attention input size by safe model/config choices when allowed:
+  latent resolution, frame count, patching, boundary ratio, or windowing.
+- Remove avoidable layout conversions before/after attention.
+- Reuse condition-side KV or other static inputs if the model structure allows.
+- Consider custom kernels, sparse/window/linear attention, or approximation only
+  after quality risk is accepted.
+
+Priority guidance:
+
+- P1 for backend/config/layout changes with preserved math.
+- P2 for approximate attention, sparsity, custom kernels, or any change that can
+  alter quality/temporal consistency.
+
+Validation:
+
+- Always include output quality and seed behavior checks.
+- Compare `diffuse`, server end-to-end, peak memory, and attention kernel time
+  in diagnostic traces if needed.
+
+### Algorithmic, Precision, and Approximation Optimization
+
+Purpose: reduce mathematical work or precision cost beyond local code cleanup.
+
+Evidence to look for:
+
+- A single operator family dominates even after low-risk runtime, parallel, and
+  fusion work.
+- Memory bandwidth or compute utilization suggests precision or quantization
+  could matter.
+- The user explicitly allows quality-preserving or approximate methods.
+
+Typical candidates:
+
+- FP8/quantization for transformer or selected projections.
+- Sparsity or Rainfusion-style acceleration.
+- Reduced steps, scheduler changes, distillation, or caching across frames.
+- Approximate attention or linear attention.
+
+Priority guidance:
+
+- Usually P2 because quality, numerics, and implementation risk are high.
+
+Validation:
+
+- Requires strict A/B, visual quality review, temporal flicker checks, seed
+  stability, and possibly human evaluation.
+
+### Interpolation, Super-Resolution, and E2E Pipeline Optimization
+
+Purpose: optimize the whole user-visible video product, not only the base
+diffusion invocation. Some deployments trade base-model latency against
+post-processing, interpolation, or super-resolution stages.
+
+Evidence to look for:
+
+- E2E latency breakdown across base generation, interpolation, super-resolution,
+  encoding, storage, and response streaming.
+- Fast/slow GPU or fast/slow stage analysis across multiple cards and pipeline
+  stages.
+- User quality target: resolution, FPS, temporal smoothness, and acceptable
+  post-processing artifacts.
+
+Typical candidates:
+
+- Add or optimize a frame interpolation stage when it reduces required base
+  model frames for the same perceived FPS.
+- Add or optimize a super-resolution model when generating lower base
+  resolution plus SR is faster for the target quality.
+- Analyze E2E2 pipeline behavior: client request, service scheduling, diffusion,
+  VAE/media, post-process, file write, and response.
+- Identify fast/slow cards or stages and rebalance pipeline placement.
+
+Priority guidance:
+
+- P1 when using proven interpolation/SR components without changing diffusion
+  math.
+- P2 when quality risk is high or the pipeline adds significant operational
+  complexity.
+
+Validation:
+
+- Measure E2E wall-clock, per-stage server timings, output FPS/resolution,
+  artifacts, flicker, and user-visible quality.
+
+### Optimization Candidate Library
+
+Use this table as a compact menu, not as automatic recommendations. Pick items
+only when baseline metrics, trace evidence, source inspection, or quality
+tolerance supports them.
+
+| Layer | Candidate | Evidence | Priority | Validation focus |
+|---|---|---|---|---|
+| Measurement | Freeze server/request/benchmark commands | Missing or drifting commands | P0 | Repeatable non-profiler A/B |
+| Measurement | Separate baseline and diagnostic profiler runs | Profiler used for latency claims | P0 | Low-overhead stage timings |
+| Host/runtime | Guard or remove avoidable `empty_cache` | Host-stack gaps or sync stalls | P0/P1 | Latency, peak memory, OOM safety |
+| Host/runtime | Cache scheduler coefficients/timesteps | Repeated tiny CPU/GPU work | P0/P1 | Same seed/output, stage timing |
+| Host/runtime | Reduce framework scheduling overhead | Client time exceeds server time | P1 | E2E latency, throughput |
+| Parallel | `CFG=2 x USP=world/2` vs `USP=world` | CFG doubles forward work | P0/P1 | `diffuse`, NCCL, memory |
+| Parallel | Tune VAE patch parallelism | VAE encode/decode is material | P0/P1 | VAE time, output correctness |
+| Parallel | HSDP on/off or buffer reuse | HSDP affects memory/all-gather | P1 | Memory, latency, OOM risk |
+| Parallel | Ulysses vs Ulysses+Ring | Long sequence all-to-all suspected | P1/P2 | Rank balance, NCCL kernels |
+| Cross-attn | Disable SP for short condition tokens | Cross-attn comm exceeds compute | P1 | `diffuse`, correctness |
+| VAE/media | Reduce VAE gather/broadcast | Rank traces show VAE wait | P1 | Rank balance, output file |
+| VAE/media | Reuse tile metadata/buffers | Tile split/merge host gaps | P1 | `vae.encode/decode`, memory |
+| VAE/media | VAE bf16/autocast | VAE float kernels are slow | P1/P2 | Artifacts, flicker, seed stability |
+| Operator fusion | AdaLayerNorm/LayerNorm fusion | Norm plus scale/shift kernels | P1 | Numeric tolerance, latency |
+| Operator fusion | RMSNorm fusion | Many small RMSNorm kernels | P1 | Numeric tolerance, latency |
+| Operator fusion | RoPE cache/fuse/layout cleanup | RoPE copy/reshape kernels | P1/P2 | Kernel count, correctness |
+| Operator layout | QKV or attention layout cleanup | Copy/cat/split around attention | P1 | Copy kernels, compile behavior |
+| Attention | Verify backend fast path | FA/SDPA dominates trace | P1 | `diffuse`, attention kernels |
+| Attention | FA to LA or selected-head LA | Attention remains dominant | P2 | Quality, temporal stability |
+| Precision | Transformer FP8/quantization | Compute/bandwidth bound and allowed | P2 | Quality, speed, stability |
+| Sparsity | Rainfusion-style acceleration | DiT compute remains dominant | P2 | Prompt diversity, quality |
+| Pipeline | Frame interpolation | Fewer base frames can meet FPS | P1/P2 | E2E latency, motion artifacts |
+| Pipeline | Super-resolution | Lower base res plus SR may win | P1/P2 | Detail quality, artifacts |
+| E2E | Fast/slow-card analysis | Multi-card stragglers | P0/P1 | Per-rank/stage wall-clock |
+
+### Optimization Plan Template
+
+Use this table shape when reporting the next work items:
+
+| Priority | Layer | Candidate | Evidence | Expected benefit | Implementation path | Validation | Quality risk |
+|---|---|---|---|---|---|---|---|
+| P0 | Host/runtime | Guard `empty_cache` | Host-stack gap points to `torch.cuda.empty_cache` | Small latency reduction, less idle | Add config/env guard | Non-profiler A/B, memory check | Low |
+| P1 | Operator fusion | RMSNorm/AdaLayerNorm fusion | High-frequency norm/elementwise kernels | Lower launch/bandwidth overhead | Use existing fusion helper or targeted Triton | A/B + output check | Medium |
+| P1/P2 | Attention | Attention layout/backend investigation | FA kernel dominates CUDA time | Potentially large | Inspect shapes/backend and remove layout copies | A/B + trace + quality | Medium/high |
+
+Then present a short selection prompt using the same rows:
+
+```text
+Which candidate should we implement next?
+
+1. P0 Host/runtime: guard empty_cache
+   - Expected benefit: small but low-risk latency reduction.
+   - Risk: possible memory increase/OOM if memory headroom is insufficient.
+
+2. P1 Operator fusion: inspect by_shape and implement first norm/RoPE/layout fusion
+   - Expected benefit: medium if high-frequency small kernels are confirmed.
+   - Risk: numerical/compile/quality validation needed.
+
+3. P1/P2 Attention: FA/LA/backend/layout investigation
+   - Expected benefit: potentially large.
+   - Risk: high quality and implementation risk.
+```
+
+If the user has not chosen an item, default to explaining tradeoffs and asking
+which candidate to execute. Only proceed autonomously on low-risk P0 measurement
+or instrumentation fixes.
+
+## Analysis Helpers
+
+PyTorch profiler traces are Chrome/Perfetto-compatible JSON files, usually
+`trace_rankN.json` or `trace_rankN.json.gz`. They normally contain a top-level
+`traceEvents` list, though some exporters emit the raw event list directly.
+
+Use the checked-in analyzer from the repository root:
+
+```bash
+python3 .claude/skills/diffusion-perf-opt/scripts/trace_analyzer.py \
+  vllm_profile/.../trace_rank0.json.gz \
+  --min-gap-ms 5 \
+  --topn 20
+```
+
+For rank imbalance or communication questions, pass all relevant rank traces in
+one command. For host gaps, lower `--min-gap-ms` to `1` and use a host-stack
+trace. Read `gpu_span_s`, `busy_union_s`, `idle_union_s`, `idle_pct`, `GAP`
+blocks, `Top GPU/operator events by total duration`, and `Top NCCL-like events
+by category`. Treat `cat=user_annotation` NCCL ranges as enclosing annotations;
+prefer `cat=kernel` or `cat=gpu_user_annotation` for real device work.
+
+The analyzer summarizes timing only. It does not parse tensor shapes, attribute
+overlap to individual streams, prove quality, or provide final latency claims.
+Use `ops_rankN.xlsx` or PyTorch key averages for shape analysis, and re-test any
+optimization with non-profiler baseline commands.
+
+Read `references/optimization-playbook.md` when drafting the optimization table or comparing candidate techniques.
+
+## vLLM Omni Heuristics
+
+- Cross-attention usually should not use USP/SP when text/image condition token count is much smaller than latent video tokens. Confirm via trace; in Wan2.2 I2V, self-attention dominates cross-attention.
+- VAE bf16/autocast is often worthwhile but requires visual quality checks.
+- VAE patch parallel can help decode/encode but may add gather/merge/broadcast overhead. Check whether all ranks need the final decoded tensor.
+- HSDP/FSDP is primarily a memory strategy. If the model fits without it, run an on/off latency comparison.
+- Scheduler work can create small host/device gaps; cache tiny solve coefficients when timesteps/order are known.
+- `torch.cuda.empty_cache()` can prevent OOM but creates synchronization/idle. Make it optional if memory headroom is sufficient.
+- `Command Buffer Full` in profiler output is profiler overhead, not a model optimization target.

.claude/skills/diffusion-perf-opt/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "vLLM Omni Perf Opt"
+  short_description: "Profile and optimize vLLM Omni diffusion performance."
+  default_prompt: "Analyze and optimize a vLLM Omni diffusion workload using baselines, traces, parallel strategy modeling, and prioritized optimization plans."

.claude/skills/diffusion-perf-opt/references/optimization-playbook.md

@@ -0,0 +1,82 @@
+# Optimization Playbook
+
+Use this reference when turning trace evidence into prioritized vLLM Omni optimizations.
+
+## Parallel Strategy
+
+Start from workload shape:
+
+- Long video tokens, high resolution, many frames: self-attention and FFN dominate. USP/SP is usually valuable.
+- CFG enabled with positive and negative prompts: transformer forward count doubles. CFG parallel may beat USP on small/moderate sequence lengths.
+- Cross-attention: skip SP/USP when condition tokens are much smaller than latent tokens; extra all-to-all can exceed compute saved.
+- HSDP/FSDP: use to fit memory first. If memory allows, compare without HSDP because all-gathers and lazy init add overhead.
+- VAE parallel: useful when encode/decode is material. Check gather/merge/broadcast and whether every rank needs final decoded output.
+
+Candidate matrices:
+
+- 2 cards: `USP=2`, `CFG=2`, HSDP on/off, VAE PP on/off.
+- 4 cards: `USP=4`, `CFG=2 x USP=2`, `USP=2 x HSDP`, VAE PP on/off.
+- 8 cards: add pipeline/stage parallel only after measuring orchestration and inter-stage transfer overhead.
+
+## Host And Free Bubble Analysis
+
+Look for device intervals where no kernel, memcpy, or memset is active.
+
+Common causes:
+
+- First request FSDP/HSDP lazy init.
+- Python preprocessing: image load, resize, `to_tensor`, tokenizer, request marshaling.
+- `torch.cuda.empty_cache()` or explicit synchronizations.
+- Scheduler small linear algebra such as tiny `torch.linalg.solve`.
+- VAE tile split/merge Python loops.
+- Profiler overhead such as CUPTI command buffer flush/full.
+
+Also inspect trace-external time:
+
+- API request parsing and multipart upload.
+- Engine/orchestrator dispatch.
+- Worker RPC and queue wait.
+- Output serialization, video write, base64, HTTP response.
+
+## Operator Optimization Candidates
+
+Evaluate necessity before implementation:
+
+- **FA to LA / LA for head:** high potential only if attention is the main CUDA hotspot and target shapes fit LA well.
+- **Cross-attention BSND/no transpose:** do only if transpose/copy around cross-attn is measurable.
+- **AdaLayerNorm fusion:** fuse norm, scale/shift, gate, residual, and dtype conversion if repeated elementwise kernels are significant.
+- **RMSNorm fusion:** useful for Q/K norms and VAE norms when small-kernel launch count is high.
+- **FP32 LayerNorm to bf16:** quality-sensitive; run visual regression.
+- **LA preprocess:** integrate only if preprocess overhead is cached or amortized.
+- **RoPE move/cache/fuse:** useful when RoPE small kernels are visible around every self-attn block.
+- **VAE RMSNorm replacement:** only after VAE internal trace shows norm as a real component of decode/encode.
+- **Scheduler fusion/cache:** cache coefficient solves and fuse CFG combine with scheduler arithmetic when possible.
+
+Before writing a custom kernel, check:
+
+- Local vLLM Omni implementations.
+- Upstream vLLM kernels and attention backends.
+- SGLang, FlashInfer, FlashAttention, and Triton examples.
+- Shape constraints and layout expectations.
+
+## Quantization And Sparsity
+
+Treat these as high-risk optimizations:
+
+- Rainfusion/sparse diffusion can be high reward but needs model-specific quality validation.
+- VAE decode quantization can reduce decode time but usually has smaller end-to-end upside than transformer attention/FFN work.
+- FP8 or approximate kernels require seed-stable visual regression, temporal flicker checks, and prompt diversity.
+
+## Output Format For Recommendations
+
+Return a table with:
+
+- Priority: P0/P1/P2.
+- Optimization.
+- Evidence from trace or code.
+- Expected benefit.
+- Risk/quality impact.
+- Implementation path.
+- Validation command or A/B test.
+
+State clearly when a recommendation is a hypothesis requiring new trace with shapes.

.claude/skills/diffusion-perf-opt/scripts/trace_analyzer.py

@@ -0,0 +1,144 @@
+#!/usr/bin/env python3
+"""Summarize torch profiler Chrome traces for vLLM Omni diffusion optimization."""
+
+from __future__ import annotations
+
+import argparse
+import collections
+import gzip
+import json
+from pathlib import Path
+from typing import Any
+
+GPU_CATS = {"kernel", "gpu_memcpy", "gpu_memset"}
+CPU_CATS = {"python_function", "user_annotation", "cpu_op", "cuda_runtime", "cuda_driver"}
+
+
+def open_trace(path: Path):
+    if path.suffix == ".gz":
+        return gzip.open(path, "rt")
+    return path.open("rt")
+
+
+def event_name(event: dict[str, Any]) -> str:
+    return str(event.get("name", ""))
+
+
+def summarize_trace(path: Path, min_gap_us: float, topn: int) -> None:
+    with open_trace(path) as f:
+        data = json.load(f)
+
+    events = data if isinstance(data, list) else data.get("traceEvents", [])
+    gpu: list[tuple[float, float, float, str, str, Any, Any]] = []
+    cpu: list[tuple[float, float, float, str, str, Any, Any]] = []
+    by_gpu_name: dict[str, list[float]] = collections.defaultdict(lambda: [0, 0.0, 0.0])
+    by_name: dict[str, list[float]] = collections.defaultdict(lambda: [0, 0.0, 0.0])
+    nccl: dict[tuple[str, str], list[float]] = collections.defaultdict(lambda: [0, 0.0, 0.0])
+
+    for event in events:
+        dur = event.get("dur")
+        ts = event.get("ts")
+        if dur is None or ts is None or dur <= 0:
+            continue
+        cat = str(event.get("cat", ""))
+        name = event_name(event)
+        row = (float(ts), float(ts + dur), float(dur), name, cat, event.get("pid"), event.get("tid"))
+        stat = by_name[name]
+        stat[0] += 1
+        stat[1] += dur
+        stat[2] = max(stat[2], dur)
+        if cat in GPU_CATS:
+            gpu.append(row)
+            gstat = by_gpu_name[name]
+            gstat[0] += 1
+            gstat[1] += dur
+            gstat[2] = max(gstat[2], dur)
+        elif cat in CPU_CATS:
+            cpu.append(row)
+        if "nccl" in name.lower():
+            nstat = nccl[(cat, name)]
+            nstat[0] += 1
+            nstat[1] += dur
+            nstat[2] = max(nstat[2], dur)
+
+    print(f"\n== {path}")
+    print(f"events={len(events)} gpu_events={len(gpu)} cpu_events={len(cpu)}")
+    if not gpu:
+        print("No GPU events found.")
+        return
+
+    gpu.sort()
+    merged: list[list[Any]] = []
+    for start, end, dur, name, cat, pid, tid in gpu:
+        if not merged or start > merged[-1][1]:
+            merged.append([start, end, [(start, end, dur, name, cat, pid, tid)]])
+        else:
+            merged[-1][1] = max(merged[-1][1], end)
+            merged[-1][2].append((start, end, dur, name, cat, pid, tid))
+
+    span = max(end for _, end, *_ in gpu) - min(start for start, *_ in gpu)
+    busy = sum(end - start for start, end, _ in merged)
+    idle = span - busy
+    print(
+        f"gpu_span_s={span / 1e6:.3f} "
+        f"busy_union_s={busy / 1e6:.3f} "
+        f"idle_union_s={idle / 1e6:.3f} "
+        f"idle_pct={idle / span * 100:.2f}"
+    )
+
+    interesting_cpu = [
+        row
+        for row in cpu
+        if row[2] >= 1000
+        and (
+            row[4] in {"python_function", "user_annotation"}
+            or "cudaStreamSynchronize" in row[3]
+            or "cudaDeviceSynchronize" in row[3]
+            or "cudaLaunch" in row[3]
+            or "cudaMemcpy" in row[3]
+        )
+    ]
+
+    gaps = []
+    for idx in range(1, len(merged)):
+        gap_start = merged[idx - 1][1]
+        gap_end = merged[idx][0]
+        gap_dur = gap_end - gap_start
+        if gap_dur >= min_gap_us:
+            prev_event = max(merged[idx - 1][2], key=lambda x: x[1])
+            next_event = min(merged[idx][2], key=lambda x: x[0])
+            mid = (gap_start + gap_end) / 2
+            containers = [row for row in interesting_cpu if row[0] <= mid <= row[1]]
+            containers = sorted(containers, key=lambda x: x[2])[:8]
+            gaps.append((gap_dur, gap_start, gap_end, prev_event, next_event, containers))
+
+    print(f"gaps_ge_{min_gap_us / 1000:.3f}ms count={len(gaps)} sum_s={sum(g[0] for g in gaps) / 1e6:.3f}")
+    for gap_dur, gap_start, gap_end, prev_event, next_event, containers in sorted(gaps, reverse=True)[:topn]:
+        print(f"\nGAP {gap_dur / 1000:.3f} ms ts={gap_start:.0f}->{gap_end:.0f}")
+        print(f"  prev {prev_event[4]} {prev_event[2] / 1000:.3f} ms {prev_event[3][:160]}")
+        print(f"  next {next_event[4]} {next_event[2] / 1000:.3f} ms {next_event[3][:160]}")
+        for row in containers:
+            print(f"  in   {row[4]} {row[2] / 1000:.3f} ms {row[3][:180]}")
+
+    print("\nTop GPU/operator events by total duration:")
+    for name, (count, total, max_dur) in sorted(by_gpu_name.items(), key=lambda kv: kv[1][1], reverse=True)[:topn]:
+        print(f"  {int(count):8d} total={total / 1e6:9.3f}s max={max_dur / 1000:9.3f}ms {name[:180]}")
+
+    print("\nTop NCCL-like events by category:")
+    for (cat, name), (count, total, max_dur) in sorted(nccl.items(), key=lambda kv: kv[1][1], reverse=True)[:topn]:
+        print(f"  {int(count):8d} total={total / 1e6:9.3f}s max={max_dur / 1000:9.3f}ms cat={cat} {name[:160]}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("traces", nargs="+", type=Path, help="trace.json or trace.json.gz files")
+    parser.add_argument("--min-gap-ms", type=float, default=5.0, help="minimum GPU idle gap to print")
+    parser.add_argument("--topn", type=int, default=20, help="number of gaps/hotspots to print")
+    args = parser.parse_args()
+
+    for trace in args.traces:
+        summarize_trace(trace, args.min_gap_ms * 1000.0, args.topn)
+
+
+if __name__ == "__main__":
+    main()

.claude/skills/readme.md

@@ -0,0 +1,37 @@
+# Claude Skills for vLLM-Omni
+
+This directory contains Claude Code skills maintained for the `vllm-omni`
+repository. These skills capture repeatable workflows for common contributor
+tasks such as model integration, pull request review, and release note
+generation.
+
+## Directory Structure
+
+Each skill lives in its own directory under `.claude/skills/`. A skill may
+include:
+
+- `SKILL.md`: the main workflow and operating instructions
+- `references/`: focused reference material used by the skill
+- `scripts/`: small helper scripts used by the skill
+
+## Available Skills
+
+- `add-diffusion-model`: guides integration of a new diffusion model into
+  `vllm-omni`
+- `diffusion-perf-opt`: guides diffusion model performance optimization,
+  including profiling traces, parallel strategies, stage timing analysis, and
+  benchmark-driven tuning
+- `add-omni-model`: covers addition of new omni-modality model support
+- `add-tts-model`: covers integration of new TTS models and related serving
+  workflows
+- `generate-release-note`: helps prepare release notes for repository changes
+- `review-pr`: provides a structured workflow for reviewing pull requests
+
+## Maintenance Guidelines
+
+- Keep skill names short and task-oriented.
+- Prefer repository-local paths, commands, and examples.
+- Avoid hardcoding fast-changing support matrices unless the skill is actively
+  maintained alongside those changes.
+- Treat skills as contributor tooling: optimize for clarity, actionability, and
+  low maintenance overhead.

.claude/skills/vllm-omni-npu-upgrade/SKILL.md

@@ -0,0 +1,300 @@
+---
+name: vllm-omni-npu-model-runner-upgrade
+description: "Upgrade vllm-omni NPU model runners (OmniNPUModelRunner, NPUARModelRunner, NPUGenerationModelRunner) to align with the latest vllm-ascend NPUModelRunner while preserving omni-specific logic."
+---
+
+# vLLM-Omni NPU Model Runner Upgrade Skill
+
+## Overview
+
+This skill guides the process of upgrading vllm-omni's NPU model runners to align with the latest vllm-ascend codebase while preserving omni-specific enhancements. The NPU runners are designed to run omni multimodal models (like Qwen3-Omni, Bagel, MiMoAudio) on Ascend NPUs.
+
+## File Structure
+
+### NPU Model Runner Files
+```
+vllm-omni/vllm_omni/platforms/npu/worker/
+├── __init__.py
+├── npu_model_runner.py           # OmniNPUModelRunner (base class)
+├── npu_ar_model_runner.py        # NPUARModelRunner (autoregressive)
+├── npu_ar_worker.py              # AR worker
+├── npu_generation_model_runner.py # NPUGenerationModelRunner (diffusion/non-AR)
+└── npu_generation_worker.py      # Generation worker
+```
+
+### GPU Reference Files (for omni-specific logic sync)
+```
+vllm-omni/vllm_omni/worker/
+├── __init__.py
+├── gpu_model_runner.py           # OmniGPUModelRunner
+├── gpu_ar_model_runner.py        # GPUARModelRunner
+├── gpu_ar_worker.py
+├── gpu_generation_model_runner.py
+├── gpu_generation_worker.py
+├── mixins.py
+├── base.py
+└── gpu_memory_utils.py
+```
+
+### vllm-ascend Reference Files
+```
+vllm-ascend/vllm_ascend/worker/
+├── model_runner_v1.py            # NPUModelRunner (base class to copy from)
+├── npu_input_batch.py
+├── block_table.py
+├── pcp_utils.py
+└── worker.py
+```
+
+## Inheritance Hierarchy
+
+```
+                    GPUModelRunner (vllm)
+                         |
+        +----------------+----------------+
+        |                                 |
+  OmniGPUModelRunner              NPUModelRunner (vllm-ascend)
+  (vllm_omni/worker)              (vllm_ascend/worker)
+        |                                 |
+        +----------- OmniNPUModelRunner --+
+                     (multiple inheritance)
+                            |
+            +---------------+---------------+
+            |                               |
+    NPUARModelRunner            NPUGenerationModelRunner
+    (autoregressive)            (non-autoregressive/diffusion)
+```
+
+## Omni-Specific Comment Markers
+
+Omni-specific logic is marked with comment blocks:
+```python
+# -------------------------------------- Omni-new -------------------------------------------------
+# ... omni-specific code ...
+# -------------------------------------- Omni-new -------------------------------------------------
+```
+
+Or simpler variations:
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+#  ------------------------------------------------------------------------------------------------
+```
+
+**Important**:
+- Always preserve and add these markers when modifying code.
+- **The reference documents (`references/omni-specific-blocks.md`) may not be up-to-date.** Always grep for `Omni-new` in the GPU implementations to find the authoritative list of omni-specific blocks.
+- When you discover new omni-specific code that is not documented in the references, please update the reference files.
+
+## Key Methods Requiring Attention
+
+### OmniNPUModelRunner (npu_model_runner.py)
+
+| Method | Description | Omni-Specific Logic |
+|--------|-------------|---------------------|
+| `load_model` | Load model and initialize talker_mtp | Uses `ACLGraphWrapper` instead of `CUDAGraphWrapper`, initializes talker buffers |
+| `_dummy_run` | Warmup/profiling run | talker_mtp dummy forward, `extract_multimodal_outputs` |
+| `_model_forward` | Forward pass wrapper | Injects `model_kwargs_extra`, wraps with `OmniOutput`, NPU-specific graph updates |
+| `_talker_mtp_forward` | Talker MTP forward for Qwen3-Omni | Uses `set_ascend_forward_context` |
+
+### NPUARModelRunner (npu_ar_model_runner.py)
+
+| Method | Description | Omni-Specific Logic |
+|--------|-------------|---------------------|
+| `__init__` | Initialize with KV transfer manager | `OmniKVTransferManager` setup |
+| `execute_model` | Main inference entry | KV transfer handling, `_update_states` override, `extract_multimodal_outputs` |
+| `sample_tokens` | Token sampling | Hidden states extraction, multimodal outputs processing, `OmniModelRunnerOutput` |
+| `_resolve_global_request_id` | Request ID resolution | For disaggregated inference |
+
+### NPUGenerationModelRunner (npu_generation_model_runner.py)
+
+| Method | Description | Omni-Specific Logic |
+|--------|-------------|---------------------|
+| `_update_request_states` | Update request states for async chunk | async_chunk handling |
+| `execute_model` | Generation forward | async_chunk, `seq_token_counts`, `_run_generation_model` |
+| `sample_tokens` | Output processing | multimodal output packaging to `OmniModelRunnerOutput` |
+| `_dummy_run` | Dummy run override | model_kwargs initialization, multimodal extraction |
+| `_run_generation_model` | Run generation model | Calls `_model_forward` with sampler |
+
+## Upgrade Workflow
+
+### Step 1: Preparation
+
+1. **Identify target versions**(Use gh cli to check):
+   - We're using vllm-omni main branch
+   - Check the last release of vllm-omni
+   - Target vllm-ascend version(Just directly use the local latest vllm-ascend code)
+
+2. **Check GPU-side changes** (since last release):
+   ```bash
+   cd /root/vllm-workspace/vllm-omni
+   git log --oneline --since="<last-release-date>" -- vllm_omni/worker/
+   ```
+
+3. **Read latest vllm-ascend code**:
+   - We don't track vllm-ascend changes - just directly use the latest code from `/root/vllm-workspace/vllm-ascend/vllm_ascend/worker/model_runner_v1.py`
+   - Copy the relevant methods and re-insert omni-specific blocks
+
+### Step 2: Analyze Omni-Specific Logic
+
+For each NPU model runner file:
+
+1. **Extract existing omni-specific blocks**:
+   ```bash
+   grep -n "Omni-new" vllm_omni/platforms/npu/worker/npu_model_runner.py
+   ```
+
+2. **Document each omni block**:
+   - Which method it belongs to
+   - What functionality it provides
+   - Dependencies on other omni code
+
+### Step 3: Update Base Class (OmniNPUModelRunner)
+
+**Note**: Always check the GPU implementation `gpu_model_runner.py` for any new omni logic not yet documented in references.
+
+1. **Read the latest vllm-ascend `NPUModelRunner.load_model`**
+2. **Copy the method, keeping the structure**
+3. **Re-insert omni-specific logic** (check GPU `gpu_model_runner.py` for authoritative list):
+   - Replace `CUDAGraphWrapper` with `ACLGraphWrapper`
+   - Keep talker_mtp initialization
+   - Preserve buffer allocations for talker
+   - Check for any new omni blocks added since last sync
+
+4. **Update `_dummy_run`**:
+   - Copy from vllm-ascend
+   - Compare with GPU `_dummy_run` for omni-specific blocks
+   - Re-insert all `Omni-new` marked code from GPU version
+
+5. **Update `_model_forward`**:
+   - Keep the omni wrapper logic
+   - Update NPU-specific parts (graph params, SP all-gather)
+   - Check GPU version for any new omni logic
+
+### Step 4: Update AR Model Runner
+
+1. **Compare with GPU `gpu_ar_model_runner.py`** for any new omni features
+2. **Copy `execute_model` from vllm-ascend**
+3. **Re-insert omni blocks** (reference `references/omni-specific-blocks.md`, but note it may be incomplete):
+   - **IMPORTANT**: Always check the GPU implementation `gpu_ar_model_runner.py` for all `Omni-new` marked code blocks
+   - The reference doc may not include newly added omni logic - treat it as a starting point, not exhaustive
+   - When discovering new omni code blocks, please update `references/omni-specific-blocks.md`
+   - Common omni blocks include but are not limited to: KV transfer, multimodal outputs, sampling_metadata handling, etc.
+
+4. **Update `sample_tokens`** (also compare with GPU implementation):
+   - Compare with `gpu_ar_model_runner.py`'s `sample_tokens` method
+   - Identify all `Omni-new` marked code blocks
+   - Ensure NPU version includes all omni-specific logic
+
+### Step 5: Update Generation Model Runner
+
+**Note**: Generation model runner may have unique omni logic for diffusion/non-AR models.
+
+1. **Compare with GPU `gpu_generation_model_runner.py`** - grep for all `Omni-new` blocks
+2. **Update `execute_model`**:
+   - Check GPU version for all omni-specific blocks
+   - Keep async_chunk handling
+   - Keep `seq_token_counts` injection
+   - Update forward/context setup from vllm-ascend
+   - Look for any new omni logic not documented in references
+
+3. **Update `_dummy_run`**:
+   - Copy from vllm-ascend base
+   - Compare with GPU `_dummy_run` if exists
+   - Re-insert all omni-specific logic
+
+### Step 6: Update Imports
+
+Check and update imports at the top of each file:
+
+```python
+# Common vllm-ascend imports
+from vllm_ascend.ascend_forward_context import get_forward_context, set_ascend_forward_context
+from vllm_ascend.attention.attention_v1 import AscendAttentionState
+from vllm_ascend.attention.utils import using_paged_attention
+from vllm_ascend.compilation.acl_graph import ACLGraphWrapper, update_full_graph_params
+from vllm_ascend.ops.rotary_embedding import update_cos_sin
+from vllm_ascend.utils import enable_sp, lmhead_tp_enable
+from vllm_ascend.worker.model_runner_v1 import SEQ_LEN_WITH_MAX_PA_WORKSPACE, NPUModelRunner
+
+# Omni-specific imports
+from vllm_omni.model_executor.models.output_templates import OmniOutput
+from vllm_omni.worker.gpu_model_runner import OmniGPUModelRunner
+from vllm_omni.outputs import OmniModelRunnerOutput
+from vllm_omni.distributed.omni_connectors.kv_transfer_manager import OmniKVTransferManager
+```
+
+### Step 7: Sync GPU-Side Omni Changes
+
+1. **Check recent GPU worker changes**:
+   ```bash
+   git diff <from-tag>..<to-tag> -- vllm_omni/worker/gpu_model_runner.py
+   git diff <from-tag>..<to-tag> -- vllm_omni/worker/gpu_ar_model_runner.py
+   ```
+
+2. **Identify new omni features** that need to be ported to NPU
+
+3. **Apply corresponding changes** to NPU runners
+
+### Step 8: Validation
+
+1. **Run type checking**:
+   ```bash
+   cd /root/vllm-workspace/vllm-omni
+   python -m py_compile vllm_omni/platforms/npu/worker/npu_model_runner.py
+   python -m py_compile vllm_omni/platforms/npu/worker/npu_ar_model_runner.py
+   python -m py_compile vllm_omni/platforms/npu/worker/npu_generation_model_runner.py
+   ```
+
+2. **Run import test**:
+   ```bash
+   python -c "from vllm_omni.platforms.npu.worker import *"
+   ```
+
+3. **Run model serving test** (if hardware available):
+   ```bash
+   vllm serve <model-path> --trust-remote-code
+   ```
+
+## Common Pitfalls
+
+### 1. Forward Context Differences
+- GPU uses `set_forward_context`
+- NPU uses `set_ascend_forward_context`
+- Parameters may differ slightly
+
+### 2. Graph Wrapper Differences
+- GPU: `CUDAGraphWrapper`
+- NPU: `ACLGraphWrapper`
+- Constructor parameters may differ
+
+### 3. Buffer Creation
+- GPU: `_make_buffer` returns different structure
+- NPU: May need numpy=True/False parameter
+
+### 4. Attention Metadata
+- GPU: Uses vllm attention metadata builders
+- NPU: Uses `AscendCommonAttentionMetadata`
+
+### 5. Sampling
+- GPU: Uses vllm sampler
+- NPU: Uses `AscendSampler`
+
+## Checklist Before Commit
+
+- [ ] All omni-specific comment markers preserved
+- [ ] New omni logic from GPU side synced
+- [ ] Imports updated to latest vllm-ascend
+- [ ] No `CUDAGraphWrapper` references in NPU code
+- [ ] `set_ascend_forward_context` used instead of `set_forward_context`
+- [ ] `ACLGraphWrapper` used for talker_mtp wrapping
+- [ ] Type hints match vllm-ascend signatures
+- [ ] No duplicate code blocks
+- [ ] Python syntax valid (py_compile passes)
+
+## Reference Files for Comparison
+
+When upgrading, keep these files open for reference:
+
+1. **vllm-ascend NPUModelRunner**: `/root/vllm-workspace/vllm-ascend/vllm_ascend/worker/model_runner_v1.py`
+2. **vllm GPUModelRunner**: `/root/vllm-workspace/vllm/vllm/v1/worker/gpu_model_runner.py`
+3. **vllm-omni OmniGPUModelRunner**: `/root/vllm-workspace/vllm-omni/vllm_omni/worker/gpu_model_runner.py`

.claude/skills/vllm-omni-npu-upgrade/references/gpu-to-npu-translation.md

@@ -0,0 +1,335 @@
+# GPU to NPU Translation Patterns
+
+This document provides a quick reference for translating GPU code patterns to NPU equivalents when porting omni-specific logic.
+
+## Import Translations
+
+### Forward Context
+```python
+# GPU
+from vllm.forward_context import set_forward_context
+
+# NPU
+from vllm_ascend.ascend_forward_context import set_ascend_forward_context
+```
+
+### Graph Wrapper
+```python
+# GPU
+from vllm.compilation.cuda_graph import CUDAGraphWrapper
+
+# NPU
+from vllm_ascend.compilation.acl_graph import ACLGraphWrapper
+```
+
+### Attention State
+```python
+# GPU (no equivalent - uses FlashAttention states directly)
+
+# NPU
+from vllm_ascend.attention.attention_v1 import AscendAttentionState
+```
+
+### Utilities
+```python
+# GPU
+# (directly use torch.cuda functions)
+
+# NPU
+from vllm_ascend.utils import enable_sp, lmhead_tp_enable
+from vllm_ascend.ops.rotary_embedding import update_cos_sin
+```
+
+## Context Manager Translations
+
+### Forward Context Setup
+```python
+# GPU
+with set_forward_context(
+    attn_metadata,
+    self.vllm_config,
+    num_tokens=num_tokens_padded,
+    num_tokens_across_dp=num_tokens_across_dp,
+    cudagraph_runtime_mode=cudagraph_mode,
+    batch_descriptor=batch_desc,
+):
+    # forward pass
+
+# NPU
+with set_ascend_forward_context(
+    attn_metadata,
+    self.vllm_config,
+    num_tokens=num_tokens_padded,
+    num_tokens_across_dp=num_tokens_across_dp,
+    aclgraph_runtime_mode=cudagraph_mode,  # Note: 'aclgraph' not 'cudagraph'
+    batch_descriptor=batch_desc,
+    num_actual_tokens=scheduler_output.total_num_scheduled_tokens,
+    model_instance=self.model,
+):
+    # forward pass
+```
+
+### Graph Capture Context
+```python
+# GPU
+from vllm.compilation.cuda_graph import graph_capture as cuda_graph_capture
+with cuda_graph_capture(self.device):
+    # capture
+
+# NPU
+from vllm_ascend.worker.model_runner_v1 import graph_capture
+with graph_capture(self.device):
+    # capture
+```
+
+## Graph Wrapper Usage
+
+### Creating Graph Wrapper
+```python
+# GPU
+if cudagraph_mode.has_full_cudagraphs() and has_separate_talker:
+    self.talker_mtp = CUDAGraphWrapper(
+        talker_mtp,
+        self.vllm_config,
+        runtime_mode=CUDAGraphMode.FULL
+    )
+
+# NPU
+if cudagraph_mode.has_full_cudagraphs() and has_separate_talker:
+    self.talker_mtp = ACLGraphWrapper(
+        talker_mtp,
+        self.vllm_config,
+        runtime_mode=CUDAGraphMode.FULL
+    )
+```
+
+### Checking Graph Wrapper Type
+```python
+# GPU
+if not isinstance(self.talker_mtp, CUDAGraphWrapper):
+    _cudagraph_mode = CUDAGraphMode.NONE
+
+# NPU
+if not isinstance(self.talker_mtp, ACLGraphWrapper):
+    _cudagraph_mode = CUDAGraphMode.NONE
+```
+
+## Device Operations
+
+### Synchronization
+```python
+# GPU
+torch.cuda.synchronize()
+
+# NPU
+torch.npu.synchronize()
+```
+
+### Stream Operations
+```python
+# GPU
+stream = torch.cuda.Stream(device=device)
+torch.cuda.current_stream()
+
+# NPU
+stream = torch.npu.Stream(device=device)
+torch.npu.current_stream()
+```
+
+## Attention Metadata
+
+### State Setting (NPU-specific)
+```python
+# GPU - handled internally by attention backends
+
+# NPU - explicit state setting required
+self.attn_state = AscendAttentionState.DecodeOnly
+if self.speculative_config and self.speculative_config.method == "mtp":
+    if self.vllm_config.model_config.use_mla:
+        self.attn_state = AscendAttentionState.SpecDecoding
+    else:
+        self.attn_state = AscendAttentionState.ChunkedPrefill
+```
+
+### Building Attention Metadata
+```python
+# GPU - uses vllm attention builders
+
+# NPU - may need additional parameters
+(attn_metadata, spec_decode_common_attn_metadata) = self._build_attention_metadata(
+    num_tokens=num_tokens_unpadded,
+    num_tokens_padded=num_tokens_padded,
+    num_reqs=num_reqs,
+    num_reqs_padded=num_reqs_padded,
+    max_query_len=max_num_scheduled_tokens,
+    ubatch_slices=ubatch_slices_attn,
+    logits_indices=logits_indices,
+    use_spec_decode=use_spec_decode,
+    num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
+    num_scheduled_tokens_np=num_scheduled_tokens_np,
+    cascade_attn_prefix_lens=cascade_attn_prefix_lens,
+)
+```
+
+## Rotary Embedding
+
+### Update Cos/Sin Cache
+```python
+# GPU - typically handled inside attention
+
+# NPU - explicit update required before forward
+from vllm_ascend.ops.rotary_embedding import update_cos_sin
+update_cos_sin(positions)
+```
+
+## Sequence Parallelism
+
+### Enable SP Check
+```python
+# GPU - use vllm distributed utilities
+
+# NPU - use vllm-ascend wrapper
+from vllm_ascend.utils import enable_sp
+
+if enable_sp():
+    # sequence parallelism enabled
+```
+
+## Sampler
+
+### Sampler Type
+```python
+# GPU - uses vllm sampler
+self.sampler = Sampler()
+
+# NPU - uses AscendSampler
+from vllm_ascend.sample.sampler import AscendSampler
+self.sampler = AscendSampler()
+```
+
+## Input Batch
+
+### Batch Class
+```python
+# GPU
+from vllm.v1.worker.gpu_input_batch import InputBatch
+
+# NPU
+from vllm_ascend.worker.npu_input_batch import NPUInputBatch
+```
+
+## Graph Parameter Updates
+
+### Full Graph Params Update (NPU-specific)
+```python
+# GPU - not needed
+
+# NPU - required for FULL graph mode
+from vllm_ascend.compilation.acl_graph import update_full_graph_params
+
+forward_context = get_forward_context()
+if (
+    forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL
+    and not forward_context.capturing
+    and not self.use_sparse
+):
+    update_full_graph_params(
+        self.attn_backend,
+        self.update_stream,
+        forward_context,
+        num_tokens_padded,
+        self.vllm_config,
+        self.speculative_config,
+        positions.shape[0],
+    )
+```
+
+## Paged Attention Check
+
+```python
+# GPU - not typically needed
+
+# NPU
+from vllm_ascend.attention.utils import using_paged_attention
+
+if is_graph_capturing and using_paged_attention(num_tokens, self.vllm_config):
+    seq_lens = SEQ_LEN_WITH_MAX_PA_WORKSPACE
+```
+
+## Common Method Signature Differences
+
+### _dummy_run Parameters
+```python
+# GPU (v0.17.0)
+def _dummy_run(
+    self,
+    num_tokens: int,
+    cudagraph_runtime_mode: CUDAGraphMode | None = None,
+    force_attention: bool = False,
+    uniform_decode: bool = False,
+    allow_microbatching: bool = True,
+    skip_eplb: bool = False,
+    is_profile: bool = False,
+    create_mixed_batch: bool = False,
+    remove_lora: bool = True,
+    is_graph_capturing: bool = False,
+    num_active_loras: int = 0,
+) -> tuple[torch.Tensor, torch.Tensor]:
+
+# NPU (v0.17.0) - adds with_prefill, activate_lora
+def _dummy_run(
+    self,
+    num_tokens: int,
+    with_prefill: bool = False,
+    cudagraph_runtime_mode: CUDAGraphMode | None = None,
+    force_attention: bool = False,
+    uniform_decode: bool = False,
+    is_profile: bool = False,
+    create_mixed_batch: bool = False,
+    allow_microbatching: bool = True,
+    skip_eplb: bool = False,
+    remove_lora: bool = True,
+    activate_lora: bool = False,
+    is_graph_capturing: bool = False,
+    num_active_loras: int = 0,
+) -> tuple[torch.Tensor, torch.Tensor]:
+```
+
+### _model_forward Parameters
+```python
+# GPU - no num_tokens_padded
+def _model_forward(
+    self,
+    input_ids: torch.Tensor | None = None,
+    positions: torch.Tensor | None = None,
+    intermediate_tensors: IntermediateTensors | None = None,
+    inputs_embeds: torch.Tensor | None = None,
+    **model_kwargs: dict[str, Any],
+):
+
+# NPU - has num_tokens_padded as first parameter
+def _model_forward(
+    self,
+    num_tokens_padded: int,
+    input_ids: torch.Tensor | None = None,
+    positions: torch.Tensor | None = None,
+    intermediate_tensors: IntermediateTensors | None = None,
+    inputs_embeds: torch.Tensor | None = None,
+    **model_kwargs: dict[str, Any],
+):
+```
+
+## Quick Reference Table
+
+| Feature | GPU | NPU |
+|---------|-----|-----|
+| Graph wrapper | `CUDAGraphWrapper` | `ACLGraphWrapper` |
+| Forward context | `set_forward_context` | `set_ascend_forward_context` |
+| Runtime mode param | `cudagraph_runtime_mode` | `aclgraph_runtime_mode` |
+| Device sync | `torch.cuda.synchronize()` | `torch.npu.synchronize()` |
+| Stream | `torch.cuda.Stream` | `torch.npu.Stream` |
+| Current stream | `torch.cuda.current_stream()` | `torch.npu.current_stream()` |
+| Input batch | `InputBatch` | `NPUInputBatch` |
+| Sampler | `Sampler` | `AscendSampler` |
+| Attention state | N/A | `AscendAttentionState` |
+| RoPE update | N/A | `update_cos_sin()` |

.claude/skills/vllm-omni-npu-upgrade/references/omni-specific-blocks.md

@@ -0,0 +1,374 @@
+# Omni-Specific Code Blocks Reference
+
+This document catalogs omni-specific code blocks in the NPU model runners, making it easier to identify what needs to be preserved during upgrades.
+
+> **IMPORTANT**: This document may not be complete or up-to-date!
+>
+> - Always grep for `Omni-new` in the GPU implementations (`vllm_omni/worker/`) to find the authoritative list
+> - New omni features may be added that are not yet documented here
+> - When you discover new omni-specific blocks during an upgrade, please update this document
+> - Last verified: Check git history for this file
+
+## OmniNPUModelRunner (npu_model_runner.py)
+
+### load_model - Talker MTP Initialization
+
+```python
+def load_model(self, *args, **kwargs) -> None:
+    NPUModelRunner.load_model(self, *args, **kwargs)
+    # Initialize enable_sp cache to avoid get_current_vllm_config() error
+    # in _pad_for_sequence_parallelism during execute_model.
+    # This is a workaround for vllm-ascend not passing vllm_config to enable_sp().
+    enable_sp(self.vllm_config)
+    # TODO move this model specific logic to a separate class
+    # TTS model IS the talker (no .talker sub-attr); use getattr to support both Omni and TTS.
+    talker_mtp = getattr(self.model, "talker_mtp", None)
+    if talker_mtp is not None:
+        self.talker_mtp = talker_mtp  # type: ignore[assignment]
+        cudagraph_mode = self.compilation_config.cudagraph_mode
+        assert cudagraph_mode is not None
+        # Only wrap talker_mtp in CUDAGraphWrapper for Omni models that
+        # have a separate .talker sub-module.  TTS models' code predictor
+        # has internal AR loops / torch.multinomial — not graph-safe.
+        has_separate_talker = getattr(self.model, "talker", None) is not None
+        if cudagraph_mode.has_full_cudagraphs() and has_separate_talker:
+            # NOTE: Use ACLGraphWrapper on NPU, not CUDAGraphWrapper
+            self.talker_mtp = ACLGraphWrapper(talker_mtp, self.vllm_config, runtime_mode=CUDAGraphMode.FULL)
+        # TTS exposes mtp_hidden_size; Omni uses hf_text_config.hidden_size.
+        hidden_size = int(
+            getattr(self.model, "mtp_hidden_size", 0) or getattr(self.model_config.hf_text_config, "hidden_size")
+        )
+        max_batch_size = max(self.max_num_reqs, self.compilation_config.max_cudagraph_capture_size)
+        self.talker_mtp_input_ids = self._make_buffer(max_batch_size, dtype=torch.int32)
+        self.talker_mtp_inputs_embeds = self._make_buffer(
+            max_batch_size, hidden_size, dtype=self.dtype, numpy=False
+        )
+        self.last_talker_hidden = self._make_buffer(max_batch_size, hidden_size, dtype=self.dtype, numpy=False)
+        self.text_step = self._make_buffer(max_batch_size, hidden_size, dtype=self.dtype, numpy=False)
+```
+
+### _dummy_run - Talker MTP Dummy Forward
+
+Location: Inside `set_ascend_forward_context` block, before main model forward
+
+```python
+# ---------------------------------------Omni-new----------------------------------------------
+if getattr(self.model, "talker", None) is not None and hasattr(self.model, "talker_mtp"):
+    num_tokens_padded_talker_mtp = num_tokens_padded
+    if num_tokens_padded_talker_mtp == self.max_num_tokens:
+        num_tokens_padded_talker_mtp = self.talker_mtp_input_ids.gpu.shape[0]
+    outputs = self.talker_mtp(
+        self.talker_mtp_input_ids.gpu[:num_tokens_padded_talker_mtp],
+        self.talker_mtp_inputs_embeds.gpu[:num_tokens_padded_talker_mtp],
+        self.last_talker_hidden.gpu[:num_tokens_padded_talker_mtp],
+        self.text_step.gpu[:num_tokens_padded_talker_mtp],
+    )
+    self.compilation_config.cache_dir = None
+# ---------------------------------------Omni-new----------------------------------------------
+```
+
+### _dummy_run - Extract Multimodal Outputs
+
+Location: After model forward, before dummy_compute_logits
+
+```python
+# ---------------------------------------Omni-new----------------------------------------------
+hidden_states, multimodal_outputs = self.extract_multimodal_outputs(hidden_states)
+# ---------------------------------------Omni-new----------------------------------------------
+```
+
+### _model_forward - Omni Output Wrapping
+
+```python
+def _model_forward(
+    self,
+    num_tokens_padded: int,
+    input_ids: torch.Tensor | None = None,
+    positions: torch.Tensor | None = None,
+    intermediate_tensors: IntermediateTensors | None = None,
+    inputs_embeds: torch.Tensor | None = None,
+    **model_kwargs: dict[str, Any],
+):
+    """Override to combine NPUModelRunner's signature with OmniGPUModelRunner's logic."""
+    # Omni-specific: build and inject extra model kwargs
+    model_kwargs_extra = self._build_model_kwargs_extra()
+
+    # Call the model forward (same as NPUModelRunner)
+    assert self.model is not None
+    model_output = self.model(
+        input_ids=input_ids,
+        positions=positions,
+        intermediate_tensors=intermediate_tensors,
+        inputs_embeds=inputs_embeds,
+        **model_kwargs,
+        **model_kwargs_extra,
+    )
+
+    # Omni-specific: wrap output if needed
+    if not isinstance(model_output, OmniOutput) and hasattr(self.model, "make_omni_output"):
+        model_output = self.model.make_omni_output(model_output, **model_kwargs_extra)
+
+    # Omni-specific: cache model output for later sample_tokens
+    self._omni_last_model_output = model_output
+
+    # NPU-specific: update full graph params (keep from vllm-ascend)
+    forward_context = get_forward_context()
+    # ... NPU graph update logic ...
+
+    # NPU-specific: all-gather for sequence parallelism (keep from vllm-ascend)
+    if get_forward_context().sp_enabled and not isinstance(model_output, IntermediateTensors):
+        model_output = self._all_gather_hidden_states_and_aux(model_output)
+
+    return model_output
+```
+
+---
+
+## NPUARModelRunner (npu_ar_model_runner.py)
+
+### __init__ - KV Transfer Manager
+
+```python
+def __init__(self, *args, **kwargs):
+    super().__init__(*args, **kwargs)
+    self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
+    # each model stage has their own hidden size
+    self.hidden_size = self.model_config.hf_text_config.hidden_size
+    self.inputs_embeds = self._make_buffer(self.max_num_tokens, self.hidden_size, dtype=self.dtype, numpy=False)
+    # Initialize KV cache manager (preserve vllm_config fallback behavior)
+    self.kv_transfer_manager = OmniKVTransferManager.from_vllm_config(self.vllm_config, self.model_config)
+```
+
+### execute_model - KV Transfer Before Update States
+
+Location: At the very beginning of execute_model
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+# [Omni] Handle KV transfer BEFORE updating states (which removes finished requests)
+self.kv_extracted_req_ids = self.kv_transfer_manager.handle_finished_requests_kv_transfer(
+    finished_reqs=getattr(scheduler_output, "finished_requests_needing_kv_transfer", {}),
+    kv_caches=self.kv_caches,
+    block_size=self.cache_config.block_size,
+    cache_dtype=str(self.cache_config.cache_dtype),
+    request_id_resolver=self._resolve_global_request_id,
+)
+#  -------------------------------------- Omni-new -------------------------------------------------
+```
+
+### execute_model - Custom _update_states Call
+
+Location: Inside synchronize_input_prep context
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+self._update_states(scheduler_output)
+#  ------------------------------------------------------------------------------------------------
+```
+
+### execute_model - Extract Multimodal Outputs
+
+Location: In post process section, after hidden_states assignment
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+hidden_states, multimodal_outputs = self.extract_multimodal_outputs(hidden_states)
+
+if multimodal_outputs is not None:
+    keys_or_type = (
+        list(multimodal_outputs.keys())
+        if isinstance(multimodal_outputs, dict)
+        else type(multimodal_outputs)
+    )
+    logger.debug(f"[AR] execute_model: multimodal_outputs keys = {keys_or_type}")
+else:
+    logger.debug("[AR] execute_model: multimodal_outputs is None")
+#  -------------------------------------- Omni-new -------------------------------------------------
+```
+
+### execute_model - Compute Logits with sampling_metadata
+
+Location: In both broadcast_pp_output True and False branches
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+# Try with sampling_metadata first; fall back to without for models that don't support it
+try:
+    logits = self.model.compute_logits(
+        sample_hidden_states, sampling_metadata=self.input_batch.sampling_metadata
+    )
+except TypeError:
+    logits = self.model.compute_logits(sample_hidden_states)
+#  -------------------------------------- Omni-new -------------------------------------------------
+```
+
+### sample_tokens - KV Extracted Req IDs
+
+Location: At the beginning of sample_tokens
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+kv_extracted_req_ids = getattr(self, "kv_extracted_req_ids", None)
+self.kv_extracted_req_ids = None
+#  -------------------------------------- Omni-new -------------------------------------------------
+```
+
+### sample_tokens - Process Additional Information and Build Output
+
+Location: After bookkeeping sync, replacing the original output construction
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+hidden_states_cpu = hidden_states.detach().to("cpu").contiguous()
+num_scheduled_tokens_np = getattr(self, "_omni_num_scheduled_tokens_np", None)
+if num_scheduled_tokens_np is None:
+    req_ids = self.input_batch.req_ids
+    num_scheduled_tokens_np = np.array(
+        [scheduler_output.num_scheduled_tokens[rid] for rid in req_ids],
+        dtype=np.int32,
+    )
+
+self._process_additional_information_updates(
+    hidden_states, multimodal_outputs, num_scheduled_tokens_np, scheduler_output
+)
+
+pooler_output: list[dict[str, object]] = []
+for rid in req_ids_output_copy:
+    idx = req_id_to_index_output_copy[rid]
+    start = int(self.query_start_loc.cpu[idx])
+    sched = int(num_scheduled_tokens_np[idx])
+    end = start + sched
+    hidden_slice = hidden_states_cpu[start:end]
+    payload: dict[str, object] = {"hidden": hidden_slice}
+    if isinstance(multimodal_outputs, dict) and multimodal_outputs:
+        # ... multimodal output slicing logic ...
+    pooler_output.append(payload)
+
+model_runner_output = OmniModelRunnerOutput(
+    req_ids=req_ids_output_copy,
+    req_id_to_index=req_id_to_index_output_copy,
+    sampled_token_ids=valid_sampled_token_ids,
+    logprobs=logprobs_lists,
+    prompt_logprobs_dict=prompt_logprobs_dict,
+    pooler_output=(pooler_output if self.vllm_config.model_config.engine_output_type != "text" else None),
+    kv_connector_output=kv_connector_output,
+)
+model_runner_output.kv_extracted_req_ids = kv_extracted_req_ids
+#  -------------------------------------- Omni-new -------------------------------------------------
+```
+
+---
+
+## NPUGenerationModelRunner (npu_generation_model_runner.py)
+
+### execute_model - Async Chunk Update
+
+Location: Inside prepare input section, before synchronize_input_prep
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+if self.model_config.async_chunk and num_scheduled_tokens:
+    self._update_request_states(scheduler_output)
+#  -------------------------------------- Omni-new -------------------------------------------------
+```
+
+### execute_model - Seq Token Counts
+
+Location: After _preprocess call
+
+```python
+# [Omni] Pass token counts per request for code2wav output slicing
+model_kwargs["seq_token_counts"] = tokens
+```
+
+### execute_model - Run Generation Model
+
+Location: Inside forward context
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+outputs = self._run_generation_model(
+    num_tokens_padded=num_tokens_padded,
+    input_ids=input_ids,
+    positions=positions,
+    intermediate_tensors=intermediate_tensors,
+    inputs_embeds=inputs_embeds,
+    model_kwargs=model_kwargs,
+    logits_indices=logits_indices,
+)
+_, multimodal_outputs = self.extract_multimodal_outputs(outputs)
+# -------------------------------------- Omni-new -------------------------------------------------
+```
+
+### sample_tokens - Multimodal Output Processing
+
+The entire sample_tokens method body is omni-specific for generation models:
+
+```python
+#  -------------------------------------- Omni-new -------------------------------------------------
+pooler_output: list[object] = []
+if isinstance(multimodal_outputs, torch.Tensor):
+    # ... tensor handling ...
+elif isinstance(multimodal_outputs, list):
+    # ... list handling ...
+elif isinstance(multimodal_outputs, dict):
+    # ... dict handling per request ...
+else:
+    raise RuntimeError("Unsupported diffusion output type")
+# [Omni] Copy req_id mappings to avoid async scheduling mutation.
+req_ids_output_copy = self.input_batch.req_ids.copy()
+req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
+output = OmniModelRunnerOutput(
+    req_ids=req_ids_output_copy,
+    req_id_to_index=req_id_to_index_output_copy,
+    sampled_token_ids=[],
+    logprobs=None,
+    prompt_logprobs_dict={},
+    pooler_output=pooler_output,
+    kv_connector_output=kv_connector_output,
+    num_nans_in_logits={},
+    ec_connector_output=ec_connector_output if self.supports_mm_inputs else None,
+)
+#  -------------------------------------- Omni-new -------------------------------------------------
+```
+
+### _dummy_run - Model Kwargs Init and Multimodal Extract
+
+Location: Before model forward and after
+
+```python
+model_kwargs = self._init_model_kwargs()  # Before forward
+
+# ... forward ...
+
+# -------------------------------------- Omni-new -------------------------------------------------
+hidden_states, _ = self.extract_multimodal_outputs(hidden_states)
+# -------------------------------------------------------------------------------------------------
+```
+
+---
+
+## ExecuteModelState Extension
+
+The `ExecuteModelState` NamedTuple is extended for omni:
+
+```python
+class ExecuteModelState(NamedTuple):
+    """Ephemeral cached state transferred between execute_model() and
+    sample_tokens(), after execute_model() returns None."""
+
+    scheduler_output: SchedulerOutput
+    logits: torch.Tensor
+    spec_decode_metadata: SpecDecodeMetadata | None
+    spec_decode_common_attn_metadata: AscendCommonAttentionMetadata | None
+    hidden_states: torch.Tensor
+    sample_hidden_states: torch.Tensor
+    aux_hidden_states: list[torch.Tensor] | None
+    attn_metadata: PerLayerAttnMetadata
+    positions: torch.Tensor
+    ec_connector_output: ECConnectorOutput | None
+    cudagraph_stats: CUDAGraphStat | None
+    multimodal_outputs: Any  # <-- Omni extension
+```
+
+This extended state must be imported from `npu_ar_model_runner` in `npu_generation_model_runner`.

.claude/skills/vllm-omni-npu-upgrade/references/workflow-checklist.md

@@ -0,0 +1,222 @@
+# NPU Model Runner Upgrade Workflow Checklist
+
+> **Note**: Reference documents (`omni-specific-blocks.md`) may not be complete. Always grep for `Omni-new` in GPU implementations to find all omni-specific code blocks. Update the reference docs when discovering new blocks.
+
+## Pre-Upgrade Preparation
+
+### 1. Version Information
+- [ ] Identify current vllm-omni version: `_________`
+- [ ] Identify target vllm-ascend version: `_________`
+- [ ] Identify target vllm version: `_________`
+- [ ] Last release date for GPU worker changes: `_________`
+
+### 2. Gather Git History
+```bash
+# GPU-side omni changes since last release
+cd /root/vllm-workspace/vllm-omni
+git log --oneline --since="YYYY-MM-DD" -- vllm_omni/worker/
+
+# vllm-ascend NPUModelRunner changes
+cd /root/vllm-workspace/vllm-ascend
+git log --oneline <from-tag>..<to-tag> -- vllm_ascend/worker/model_runner_v1.py
+```
+
+### 3. Backup Current Files
+- [ ] Create backup of current NPU runners:
+  ```bash
+  cp -r vllm_omni/platforms/npu/worker vllm_omni/platforms/npu/worker.backup
+  ```
+
+---
+
+## OmniNPUModelRunner (npu_model_runner.py)
+
+### Read and Understand
+- [ ] Read current `npu_model_runner.py`
+- [ ] Read latest `vllm_ascend/worker/model_runner_v1.py`
+- [ ] Read latest `vllm_omni/worker/gpu_model_runner.py`
+
+### Method: load_model
+- [ ] Document existing omni-specific logic
+- [ ] Copy latest NPUModelRunner.load_model structure
+- [ ] Re-insert: `enable_sp(self.vllm_config)` call
+- [ ] Re-insert: talker_mtp detection and setup
+- [ ] Replace: `CUDAGraphWrapper` → `ACLGraphWrapper`
+- [ ] Re-insert: Buffer allocations (talker_mtp_input_ids, etc.)
+
+### Method: _dummy_run
+- [ ] Document existing omni-specific logic locations
+- [ ] Copy latest NPUModelRunner._dummy_run
+- [ ] Re-insert: talker_mtp dummy forward block (inside context)
+- [ ] Re-insert: `extract_multimodal_outputs` call
+- [ ] Verify: Comment markers are present
+
+### Method: _model_forward
+- [ ] Copy latest NPUModelRunner._model_forward structure
+- [ ] Re-insert: `_build_model_kwargs_extra()` call
+- [ ] Re-insert: OmniOutput wrapping logic
+- [ ] Re-insert: `_omni_last_model_output` caching
+- [ ] Keep: NPU graph params update
+- [ ] Keep: SP all-gather logic
+
+### Method: _talker_mtp_forward
+- [ ] Verify: Uses `set_ascend_forward_context`
+- [ ] Verify: Uses `ACLGraphWrapper` check
+- [ ] Sync any changes from GPU `_talker_mtp_forward`
+
+### Imports
+- [ ] Update vllm-ascend imports to latest paths
+- [ ] Verify all omni imports are present
+- [ ] Remove any deprecated imports
+
+---
+
+## NPUARModelRunner (npu_ar_model_runner.py)
+
+### Read and Understand
+- [ ] Read current `npu_ar_model_runner.py`
+- [ ] Read latest `vllm_ascend/worker/model_runner_v1.py` execute_model
+- [ ] Read latest `vllm_omni/worker/gpu_ar_model_runner.py`
+
+### Method: __init__
+- [ ] Sync any new initialization from GPU side
+- [ ] Keep: `OmniKVTransferManager` setup
+- [ ] Keep: Custom buffer allocations
+
+### Method: execute_model
+- [ ] Document all omni blocks with line numbers
+- [ ] Copy latest NPUModelRunner.execute_model structure
+- [ ] Re-insert: KV transfer handling (beginning)
+- [ ] Re-insert: Custom `_update_states` call
+- [ ] Re-insert: `extract_multimodal_outputs`
+- [ ] Re-insert: `compute_logits` with sampling_metadata try/except
+- [ ] Update: ExecuteModelState to include multimodal_outputs
+
+### Method: sample_tokens
+- [ ] Document all omni blocks
+- [ ] Copy latest NPUModelRunner.sample_tokens structure
+- [ ] Re-insert: `kv_extracted_req_ids` handling
+- [ ] Re-insert: Hidden states CPU copy
+- [ ] Re-insert: `_process_additional_information_updates`
+- [ ] Re-insert: `OmniModelRunnerOutput` construction
+
+### ExecuteModelState
+- [ ] Verify: `multimodal_outputs` field is present
+- [ ] Verify: Imported/used correctly in execute_model
+
+### Imports
+- [ ] Update all vllm-ascend imports
+- [ ] Keep omni-specific imports
+
+---
+
+## NPUGenerationModelRunner (npu_generation_model_runner.py)
+
+### Read and Understand
+- [ ] Read current `npu_generation_model_runner.py`
+- [ ] Read latest GPU `gpu_generation_model_runner.py`
+
+### Method: _update_request_states
+- [ ] Verify: async_chunk handling is correct
+- [ ] Sync any changes from GPU side
+
+### Method: execute_model
+- [ ] Document all omni blocks
+- [ ] Copy latest NPUModelRunner.execute_model base structure
+- [ ] Re-insert: async_chunk update logic
+- [ ] Re-insert: `seq_token_counts` injection
+- [ ] Re-insert: `_run_generation_model` call
+- [ ] Re-insert: `extract_multimodal_outputs`
+- [ ] Use: ExecuteModelState from npu_ar_model_runner
+
+### Method: sample_tokens
+- [ ] Keep: Entire omni multimodal output processing
+- [ ] Update: Any new output fields needed
+- [ ] Keep: `OmniModelRunnerOutput` construction
+
+### Method: _run_generation_model
+- [ ] Sync any changes from GPU side
+- [ ] Keep: `_model_forward` call with sampler
+
+### Method: _dummy_run
+- [ ] Copy latest NPUModelRunner._dummy_run
+- [ ] Re-insert: `model_kwargs = self._init_model_kwargs()`
+- [ ] Re-insert: `extract_multimodal_outputs` at end
+
+### Imports
+- [ ] Import ExecuteModelState from npu_ar_model_runner
+- [ ] Update vllm-ascend imports
+
+---
+
+## Post-Upgrade Validation
+
+### Syntax Validation
+- [ ] `python -m py_compile vllm_omni/platforms/npu/worker/npu_model_runner.py`
+- [ ] `python -m py_compile vllm_omni/platforms/npu/worker/npu_ar_model_runner.py`
+- [ ] `python -m py_compile vllm_omni/platforms/npu/worker/npu_generation_model_runner.py`
+
+### Import Validation
+- [ ] `python -c "from vllm_omni.platforms.npu.worker.npu_model_runner import OmniNPUModelRunner"`
+- [ ] `python -c "from vllm_omni.platforms.npu.worker.npu_ar_model_runner import NPUARModelRunner"`
+- [ ] `python -c "from vllm_omni.platforms.npu.worker.npu_generation_model_runner import NPUGenerationModelRunner"`
+
+### Comment Markers
+- [ ] Grep for "Omni-new" in all three files
+- [ ] Verify all omni blocks have closing markers
+
+### Code Review
+- [ ] No `CUDAGraphWrapper` references
+- [ ] All `set_forward_context` replaced with `set_ascend_forward_context`
+- [ ] Parameter names correct (`aclgraph_runtime_mode` not `cudagraph_runtime_mode`)
+- [ ] No duplicate code blocks
+- [ ] No missing imports
+
+---
+
+## Git Commit
+
+### Commit Message Template
+```
+[NPU] Upgrade model runners to align with vllm-ascend vX.Y.Z
+
+- Update OmniNPUModelRunner with latest NPUModelRunner base
+- Update NPUARModelRunner execute_model and sample_tokens
+- Update NPUGenerationModelRunner for async_chunk changes
+- Sync GPU-side omni changes from vX.Y.Z release
+- Preserve all omni-specific logic (marked with Omni-new comments)
+
+Changes from vllm-ascend:
+- <list key changes>
+
+Changes synced from GPU:
+- <list key GPU-side omni changes>
+```
+
+### Files to Stage
+- [ ] `vllm_omni/platforms/npu/worker/npu_model_runner.py`
+- [ ] `vllm_omni/platforms/npu/worker/npu_ar_model_runner.py`
+- [ ] `vllm_omni/platforms/npu/worker/npu_generation_model_runner.py`
+- [ ] Any other modified files
+
+---
+
+## Troubleshooting
+
+### Import Errors
+- Check if vllm-ascend module paths have changed
+- Verify PYTHONPATH includes both vllm-ascend and vllm-omni
+
+### Type Errors
+- Check method signatures match between GPU and NPU
+- Verify NamedTuple fields match expected structure
+
+### Runtime Errors
+- Enable debug logging: `export VLLM_LOGGING_LEVEL=DEBUG`
+- Check graph capture issues: try `--enforce-eager`
+- Check attention issues: verify AscendAttentionState usage
+
+### Performance Regression
+- Compare with previous version on same model
+- Check if graph capture is working: look for ACLGraph logs
+- Verify SP/EP configurations are correct

.gitattributes

@@ -27,6 +27,7 @@
 *.mov binary
 *.mp4 binary
 *.mp3 binary
+*.wav binary
 *.flv binary
 *.fla binary
 *.swf binary

.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/ISSUE_TEMPLATE/750-RFC.yml

@@ -7,7 +7,7 @@ body:
 - type: markdown
   attributes:
     value: >
-      #### Please take a look at previous [RFCs](https://github.com/vllm-project/vllm-omni/issues?q=label%3ARFC+sort%3Aupdated-desc) for reference.
+      #### Please take a look at previous [RFCs](https://github.com/vllm-project/vllm-omni/issues?q=in%3Atitle%20RFC%20sort%3Aupdated-desc) for reference.
 - type: textarea
   attributes:
     label: Motivation.
@@ -21,7 +21,7 @@ body:
     description: >
       The proposed change of the RFC.
     value: |
-      Please provide the detailed design document of the RFC using the [template](https://docs.google.com/document/d/12YxSsVeD1jvL-InClkeAEnZyWFDndz_65JmXvsamuV4/edit?tab=t.0#heading=h.4ef4szrsgspp).
+      Please provide the detailed design document of the RFC using the [template](https://docs.google.com/document/d/1jcgR3cDaUQH3VczD4ZcKaJAoYWHjCmnYzHkCNyz-9fk/edit?usp=sharing).
   validations:
     required: true
 - type: textarea

.github/mergify.yml

@@ -0,0 +1,223 @@
+pull_request_rules:
+- name: label-documentation
+  description: Automatically apply documentation label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^[^/]+\.md$
+      - files~=^docs/
+      - files~=^examples/
+  actions:
+    label:
+      add:
+        - documentation
+    comment:
+      message: "Documentation preview: https://vllm--{{number}}.org.readthedocs.build/en/{{number}}/"
+
+- name: comment-pre-commit-failure
+  description: Comment on PR when pre-commit check fails
+  conditions:
+    - status-failure=pre-commit
+    - -closed
+    - -draft
+  actions:
+    comment:
+      message: |
+        Hi @{{author}}, the pre-commit checks have failed. Please run:
+
+        ```bash
+        uv pip install pre-commit
+        pre-commit install
+        pre-commit run --all-files
+        ```
+
+        Then, commit the changes and push to your branch.
+
+        For future commits, `pre-commit` will run automatically on changed files before each commit.
+
+        > [!TIP]
+        > <details>
+        > <summary>Is <code>mypy</code> or <code>markdownlint</code> failing?</summary>
+        > <br/>
+        > <code>mypy</code> and <code>markdownlint</code> are run differently in CI. If the failure is related to either of these checks, please use the following commands to run them locally:
+        >
+        > ```bash
+        > # For mypy (substitute "3.10" with the failing version if needed)
+        > pre-commit run --hook-stage manual mypy-3.10
+        > # For markdownlint
+        > pre-commit run --hook-stage manual markdownlint
+        > ```
+        > </details>
+
+- name: comment-dco-failure
+  description: Comment on PR when DCO check fails
+  conditions:
+    - status-failure=dco
+    - -closed
+    - -draft
+  actions:
+    comment:
+      message: |
+        Hi @{{author}}, the DCO check has failed. Please click on DCO in the Checks section for instructions on how to resolve this.
+
+- name: label-ci-build
+  description: Automatically apply ci/build label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^\.github/
+      - files~=\.buildkite/
+      - files~=^cmake/
+      - files=CMakeLists.txt
+      - files~=^docker/Dockerfile
+      - files~=^requirements.*\.txt
+      - files=setup.py
+  actions:
+    label:
+      add:
+        - ci/build
+
+- name: label-frontend
+  description: Automatically apply frontend label
+  conditions:
+    - label != stale
+    - files~=^vllm_omni/entrypoints/
+  actions:
+    label:
+      add:
+        - frontend
+
+- name: label-diffusion
+  description: Automatically apply diffusion label
+  conditions:
+    - label != stale
+    - files~=^vllm_omni/diffusion/
+  actions:
+    label:
+      add:
+        - diffusion
+
+- name: label-new-model
+  description: Automatically apply new-model label
+  conditions:
+    - label != stale
+    - or:
+        - files=vllm_omni/model_executor/models/registry.py
+        - files=vllm_omni/diffusion/registry.py
+  actions:
+    label:
+      add:
+        - new-model
+
+- name: label-npu
+  description: Automatically apply npu label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^vllm_omni/platforms/npu/
+      - title~=(?i)NPU
+      - title~=(?i)Ascend
+  actions:
+    label:
+      add:
+        - npu
+    assign:
+      users:
+        - "gcanlin"
+
+- name: label-rocm
+  description: Automatically apply rocm label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^vllm_omni/platforms/rocm/
+      - title~=(?i)AMD
+      - title~=(?i)ROCm
+  actions:
+    label:
+      add:
+        - rocm
+    assign:
+      users:
+        - "tjtanaa"
+
+- name: label-xpu
+  description: Automatically apply xpu label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^vllm_omni/platforms/xpu/
+      - title~=(?i)Intel
+      - title~=(?i)XPU
+  actions:
+    label:
+      add:
+        - xpu
+    assign:
+      users:
+        - "yma11"
+
+- name: label-modelrunner
+  description: Automatically apply modelrunner label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^vllm_omni/worker/
+      - title~=(?i)ModelRunner
+      - title~=(?i)worker
+  actions:
+    label:
+      add:
+        - ModelRunner
+    assign:
+      users:
+        - "gcanlin" # To check whether need to sync to NPU
+
+- name: auto-rebase if approved, ready, and 30 commits behind main
+  conditions:
+    - base = main
+    - label=ready
+    - "#approved-reviews-by >= 1"
+    - "#commits-behind >= 30"
+    - -closed
+    - -draft
+    - -conflict
+  actions:
+    rebase: {}
+
+- name: ping author on conflicts and add 'needs-rebase' label
+  conditions:
+    - label != stale
+    - conflict
+    - -closed
+  actions:
+    label:
+      add:
+        - needs-rebase
+    comment:
+      message: |
+       This pull request has merge conflicts that must be resolved before it can be
+       merged. Please rebase the PR, @{{author}}.
+
+       https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork
+
+- name: remove 'needs-rebase' label when conflict is resolved
+  conditions:
+    - -conflict
+    - -closed
+  actions:
+    label:
+      remove:
+        - needs-rebase
+
+- name: label-bug
+  description: Automatically apply bug label
+  conditions:
+    - label != stale
+    - or:
+      - title~=(?i)\bbug\b
+      - title~=(?i)\bbugfix\b
+  actions:
+    label:
+      add:
+        - bug

.github/scripts/pr_reviewer.py

@@ -1,629 +0,0 @@
-#!/usr/bin/env python3
-"""
-PR Reviewer using GLM API for vllm-omni project.
-"""
-
-import json
-import logging
-import os
-import sys
-import time
-from dataclasses import dataclass
-from typing import Any, TypedDict
-
-import requests
-
-
-# Type definitions for API responses
-class PRDetails(TypedDict):
-    """Type definition for GitHub PR details response."""
-
-    title: str
-    body: str
-    number: int
-    state: str
-    user: dict[str, Any]
-
-
-class GLMMessage(TypedDict):
-    """Type definition for GLM API message."""
-
-    role: str
-    content: str
-
-
-class GLMChoice(TypedDict):
-    """Type definition for GLM API choice."""
-
-    message: GLMMessage
-    finish_reason: str
-
-
-class GLMResponse(TypedDict):
-    """Type definition for GLM API response."""
-
-    choices: list[GLMChoice]
-    usage: dict[str, int] | None
-
-
-class GitHubComment(TypedDict):
-    """Type definition for GitHub comment."""
-
-    id: int
-    body: str
-    created_at: str
-    user: dict[str, Any]
-
-
-# Configuration
-TRIGGER_PHRASE: str = "@vllm-omni-reviewer"
-DEFAULT_GLM_API_URL: str = "https://open.bigmodel.cn/api/paas/v4/chat/completions"  # noqa: E501
-DEFAULT_GLM_MODEL: str = "glm-5"
-DEFAULT_COOLDOWN_MINUTES: int = 5
-DEFAULT_MAX_RETRIES: int = 3
-DEFAULT_RETRY_DELAY: float = 1.0
-MAX_DIFF_SIZE: int = 100_000  # Maximum diff size in characters
-
-
-@dataclass
-class Config:
-    """Configuration for the PR reviewer."""
-
-    glm_api_url: str
-    glm_model: str
-    cooldown_minutes: int
-    max_retries: int
-    retry_delay: float
-    max_diff_size: int
-
-
-# Setup logging
-logging.basicConfig(
-    level=logging.INFO,
-    format="[PR Reviewer] %(asctime)s - %(levelname)s - %(message)s",
-    datefmt="%Y-%m-%d %H:%M:%S",
-)
-logger: logging.Logger = logging.getLogger(__name__)
-
-
-def get_config() -> Config:
-    """Load configuration from environment variables with defaults."""
-    return Config(
-        glm_api_url=os.getenv("GLM_API_URL", DEFAULT_GLM_API_URL),
-        glm_model=os.getenv("GLM_MODEL", DEFAULT_GLM_MODEL),
-        cooldown_minutes=int(
-            os.getenv(
-                "PR_REVIEWER_COOLDOWN_MINUTES",
-                str(DEFAULT_COOLDOWN_MINUTES),
-            )
-        ),
-        max_retries=int(
-            os.getenv(
-                "PR_REVIEWER_MAX_RETRIES",
-                str(DEFAULT_MAX_RETRIES),
-            )
-        ),
-        retry_delay=float(os.getenv("PR_REVIEWER_RETRY_DELAY", str(DEFAULT_RETRY_DELAY))),
-        max_diff_size=int(os.getenv("PR_REVIEWER_MAX_DIFF_SIZE", str(MAX_DIFF_SIZE))),  # noqa: E501
-    )
-
-
-def get_env_var(name: str) -> str:
-    """
-    Get an environment variable or raise an error.
-
-    Args:
-        name: Name of the environment variable.
-
-    Returns:
-        The value of the environment variable.
-
-    Raises:
-        SystemExit: If the environment variable is not set.
-    """
-    value = os.environ.get(name)
-    if not value:
-        logger.error(f"Environment variable {name} is not set")
-        sys.exit(1)
-    return value
-
-
-def check_trigger(comment_body: str) -> bool:
-    """
-    Check if the comment contains the trigger phrase.
-
-    Args:
-        comment_body: The body of the comment to check.
-
-    Returns:
-        True if the trigger phrase is found, False otherwise.
-    """
-    return TRIGGER_PHRASE in comment_body
-
-
-def fetch_pr_diff(
-    repo_name: str,
-    pr_number: int,
-    token: str,
-    max_size: int = MAX_DIFF_SIZE,
-) -> str | None:
-    """
-    Fetch the diff for a pull request.
-
-    Args:
-        repo_name: The repository name in format "owner/repo".
-        pr_number: The pull request number.
-        token: GitHub authentication token.
-        max_size: Maximum diff size in characters.
-
-    Returns:
-        The diff content as a string, or None if fetching failed.
-        Returns empty string if diff is larger than max_size.
-    """
-    url: str = f"https://api.github.com/repos/{repo_name}/pulls/{pr_number}"
-    headers: dict[str, str] = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/vnd.github.v3.diff",
-    }
-
-    logger.info(f"Fetching PR diff from {url}")
-    response = requests.get(url, headers=headers, timeout=30)
-
-    if response.status_code == 200:
-        diff: str = response.text
-        if len(diff) > max_size:
-            logger.warning(
-                f"Diff size ({len(diff)} bytes) exceeds maximum "
-                f"({max_size} bytes), truncating to first "
-                f"{max_size} bytes"
-            )
-            return diff[:max_size] + "\n\n... [Diff truncated due to size] ..."
-        logger.info(f"Successfully fetched diff ({len(diff)} bytes)")
-        return diff
-    else:
-        logger.error(f"Failed to fetch PR diff: {response.status_code}")
-        logger.error(f"Response: {response.text}")
-        return None
-
-
-def fetch_pr_details(
-    repo_name: str,
-    pr_number: int,
-    token: str,
-) -> PRDetails | None:
-    """
-    Fetch PR details including title and description.
-
-    Args:
-        repo_name: The repository name in format "owner/repo".
-        pr_number: The pull request number.
-        token: GitHub authentication token.
-
-    Returns:
-        A dictionary containing PR details, or None if fetching failed.
-    """
-    url: str = f"https://api.github.com/repos/{repo_name}/pulls/{pr_number}"
-    headers: dict[str, str] = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/vnd.github.v3+json",
-    }
-
-    logger.info(f"Fetching PR details from {url}")
-    response = requests.get(url, headers=headers, timeout=30)
-
-    if response.status_code == 200:
-        return response.json()
-    else:
-        logger.error(f"Failed to fetch PR details: {response.status_code}")
-        return None
-
-
-def build_review_prompt(pr_title: str, pr_description: str, diff: str) -> str:
-    """
-    Build the prompt for the GLM-4.7 API.
-
-    Args:
-        pr_title: The title of the pull request.
-        pr_description: The description/body of the pull request.
-        diff: The diff content of the pull request.
-
-    Returns:
-        The formatted prompt string for the API.
-    """
-    return f"""You are an expert code reviewer for the VLLM-Omni project. \
-Please review the following pull request:
-
-## Pull Request Details
-**Title:** {pr_title}
-
-**Description:**
-{pr_description if pr_description else "No description provided."}
-
-## Code Changes (Diff)
-{diff}
-
-## Review Guidelines
-
-Please provide a comprehensive code review with the following sections:
-
-### 1. Overview
-- Brief summary of the changes
-- Overall assessment (positive, neutral, or concerns)
-
-### 2. Code Quality
-- Code style and consistency
-- Potential bugs or edge cases
-- Performance considerations
-- Error handling
-
-### 3. Architecture & Design
-- Integration with existing codebase
-- Design patterns and best practices
-- Potential improvements
-
-### 4. Security & Safety
-- Security concerns (if any)
-- Resource management
-- Input validation
-
-### 5. Testing & Documentation
-- Test coverage considerations
-- Documentation completeness
-- Examples and usage clarity
-
-### 6. Specific Suggestions
-- Line-by-line specific feedback (use `file:line` format)
-- Concrete actionable suggestions
-- Code examples for improvements (if applicable)
-
-### 7. Approval Status
-- **LGTM** (Looks Good To Me) if the PR is ready to merge
-- **LGTM with suggestions** if the PR is good but has minor suggestions
-- **Changes requested** if significant changes are needed
-
-## Important Notes
-- Be constructive and helpful
-- Focus on objective technical feedback
-- Acknowledge good practices when you see them
-- Prioritize critical issues over nitpicks
-- If the diff is empty or minimal, acknowledge this and provide
-  any relevant context-specific guidance
-
-Please format your response in Markdown with clear section headers.
-"""
-
-
-def validate_glm_response(data: dict[str, Any]) -> str | None:
-    """
-    Validate and extract content from GLM API response.
-
-    Args:
-        data: The response data from GLM API.
-
-    Returns:
-        The review content string if valid, None otherwise.
-    """
-    # Check if choices exists and is a non-empty list
-    if "choices" not in data:
-        logger.error("GLM API response missing 'choices' field")
-        logger.error(f"Response structure: {json.dumps(data, indent=2)}")
-        return None
-
-    choices = data["choices"]
-    if not isinstance(choices, list):
-        logger.error(f"GLM API 'choices' is not a list: {type(choices)}")
-        return None
-
-    if len(choices) == 0:
-        logger.error("GLM API 'choices' is an empty list")
-        return None
-
-    # Check if first choice has message
-    try:
-        first_choice = choices[0]
-        if not isinstance(first_choice, dict):
-            logger.error(f"GLM API choice is not a dict: {type(first_choice)}")
-            return None
-
-        if "message" not in first_choice:
-            logger.error("GLM API choice missing 'message' field")
-            logger.error(f"Choice structure: {json.dumps(first_choice, indent=2)}")  # noqa: E501
-            return None
-
-        message = first_choice["message"]
-        if not isinstance(message, dict):
-            logger.error(f"GLM API message is not a dict: {type(message)}")
-            return None
-
-        if "content" not in message:
-            logger.error("GLM API message missing 'content' field")
-            logger.error(f"Message structure: {json.dumps(message, indent=2)}")
-            return None
-
-        content = message["content"]
-        if not isinstance(content, str):
-            logger.error(f"GLM API content is not a string: {type(content)}")
-            return None
-
-        return content
-
-    except (KeyError, IndexError, TypeError) as e:
-        logger.error(f"Failed to parse GLM API response: {e}")
-        logger.error(f"Response: {json.dumps(data, indent=2)}")
-        return None
-
-
-def call_glm_api(prompt: str, api_key: str, config: Config) -> str | None:
-    """
-    Call the GLM-4.7 API to get code review with retry logic.
-
-    Args:
-        prompt: The prompt to send to the API.
-        api_key: The GLM API key.
-        config: Configuration object.
-
-    Returns:
-        The review content as a string, or None if all retries failed.
-    """
-    headers: dict[str, str] = {
-        "Authorization": f"Bearer {api_key}",
-        "Content-Type": "application/json",
-    }
-
-    payload: dict[str, Any] = {
-        "model": config.glm_model,
-        "messages": [{"role": "user", "content": prompt}],
-        "temperature": 0.3,
-        "max_tokens": 32000,
-        "top_p": 0.9,
-    }
-
-    last_error: str | None = None
-
-    for attempt in range(config.max_retries):
-        try:
-            logger.info(f"Calling GLM API ({config.glm_model}) - Attempt {attempt + 1}/{config.max_retries}")
-            response = requests.post(
-                config.glm_api_url,
-                headers=headers,
-                json=payload,
-                timeout=120,
-            )
-
-            if response.status_code == 200:
-                data = response.json()
-                review = validate_glm_response(data)
-                if review:
-                    logger.info(f"Successfully received review ({len(review)} chars)")  # noqa: E501
-                    return review
-                else:
-                    last_error = "Failed to validate API response structure"
-                    logger.error(last_error)
-            else:
-                last_error = f"GLM API request failed: {response.status_code} - {response.text}"
-                logger.error(last_error)
-
-        except requests.exceptions.Timeout:
-            last_error = f"GLM API request timed out (attempt {attempt + 1})"
-            logger.error(last_error)
-        except requests.exceptions.RequestException as e:
-            last_error = f"GLM API request exception: {e}"
-            logger.error(last_error)
-        except json.JSONDecodeError as e:
-            last_error = f"Failed to decode GLM API response as JSON: {e}"
-            logger.error(last_error)
-
-        # Exponential backoff before retry
-        if attempt < config.max_retries - 1:
-            wait_time: float = config.retry_delay * (2**attempt)
-            logger.info(f"Waiting {wait_time}s before retry...")  # noqa: E501
-            time.sleep(wait_time)
-
-    logger.error(
-        f"All {config.max_retries} attempts failed. Last error: {last_error}"  # noqa: E501
-    )
-    return None
-
-
-def check_cooldown(  # noqa: E501
-    repo_name: str,
-    pr_number: int,
-    token: str,
-    cooldown_minutes: int,
-) -> bool:
-    """
-    Check if the PR is within the cooldown period.
-
-    Args:
-        repo_name: The repository name in format "owner/repo".
-        pr_number: The pull request number.
-        token: GitHub authentication token.
-        cooldown_minutes: Cooldown period in minutes.
-
-    Returns:
-        True if within cooldown period (should skip), False otherwise.
-    """
-    from datetime import datetime, timedelta
-
-    url: str = (
-        f"https://api.github.com/repos/{repo_name}/issues/"
-        f"{pr_number}/comments"  # noqa: E501
-    )
-    headers: dict[str, str] = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/vnd.github.v3+json",
-    }
-
-    logger.info(f"Checking cooldown period ({cooldown_minutes} minutes)")
-    response = requests.get(url, headers=headers, timeout=30)
-
-    if response.status_code != 200:
-        logger.warning(f"Failed to check cooldown: {response.status_code}, proceeding with review")
-        return False
-
-    comments: list[dict[str, Any]] = response.json()
-    cutoff_time: datetime = datetime.utcnow() - timedelta(minutes=cooldown_minutes)  # noqa: E501
-
-    for comment in reversed(comments):
-        # Check if this is a bot comment
-        body: str = comment.get("body", "")
-        if "VLLM-Omni PR Review" in body or "PR Reviewer Bot" in body:
-            created_at_str: str = comment.get("created_at", "")
-            try:
-                # Parse GitHub timestamp format
-                created_at = datetime.fromisoformat(created_at_str.replace("Z", "+00:00"))
-                created_at = created_at.replace(tzinfo=None)
-                if created_at > cutoff_time:
-                    logger.info(f"PR is within cooldown period (last review: {created_at_str})")
-                    return True
-            except ValueError:
-                logger.warning(f"Failed to parse comment timestamp: {created_at_str}")  # noqa: E501
-                continue
-
-    logger.info("PR is outside cooldown period, proceeding with review")
-    return False
-
-
-def post_review_comment(  # noqa: E501
-    repo_name: str,
-    pr_number: int,
-    token: str,
-    review: str,
-) -> bool:
-    """
-    Post the review as a comment on the PR.
-
-    Args:
-        repo_name: The repository name in format "owner/repo".
-        pr_number: The pull request number.
-        token: GitHub authentication token.
-        review: The review content to post.
-
-    Returns:
-        True if posting succeeded, False otherwise.
-    """
-    url: str = (
-        f"https://api.github.com/repos/{repo_name}/issues/"
-        f"{pr_number}/comments"  # noqa: E501
-    )
-    headers: dict[str, str] = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/vnd.github.v3+json",
-    }
-
-    # Format the review comment
-    comment_body: str = f"""## 🤖 VLLM-Omni PR Review
-
-{review}
-
----
-*This review was generated automatically by the VLLM-Omni PR Reviewer Bot
-using {os.getenv("GLM_MODEL", DEFAULT_GLM_MODEL)}.*
-"""
-
-    payload: dict[str, str] = {"body": comment_body}
-
-    logger.info(f"Posting review comment to PR #{pr_number}")
-    response = requests.post(url, headers=headers, json=payload, timeout=30)
-
-    if response.status_code == 201:
-        logger.info("Successfully posted review comment")
-        return True
-    else:
-        logger.error(f"Failed to post comment: {response.status_code}")
-        logger.error(f"Response: {response.text}")
-        return False
-
-
-def main() -> int:
-    """
-    Main entry point for the PR reviewer bot.
-
-    Returns:
-        0 on success, 1 on error.
-    """
-    logger.info("VLLM-Omni PR Reviewer Bot starting...")
-
-    # Load configuration
-    config: Config = get_config()
-    logger.info(
-        f"Configuration: model={config.glm_model}, "
-        f"cooldown={config.cooldown_minutes}min, "
-        f"max_retries={config.max_retries}"
-    )
-
-    # Get environment variables
-    token: str = get_env_var("GITHUB_TOKEN")
-    api_key: str = get_env_var("GLM_API_KEY")
-    repo_name: str = get_env_var("REPO_NAME")
-    pr_number_str: str = get_env_var("PR_NUMBER")
-    comment_body: str = get_env_var("COMMENT_BODY")
-
-    try:
-        pr_number: int = int(pr_number_str)
-    except ValueError:
-        logger.error(f"Invalid PR number: {pr_number_str}")
-        return 1
-
-    logger.info(f"Repository: {repo_name}")
-    logger.info(f"PR Number: {pr_number}")
-
-    # Check if the comment contains the trigger phrase
-    if not check_trigger(comment_body):
-        logger.info(
-            f"Comment does not contain trigger phrase '{TRIGGER_PHRASE}', exiting"  # noqa: E501
-        )
-        return 0
-
-    logger.info("Trigger phrase detected! Starting review process...")
-
-    # Check cooldown period
-    if check_cooldown(repo_name, pr_number, token, config.cooldown_minutes):
-        logger.info("Skipping review due to cooldown period")
-        return 0
-
-    # Fetch PR details
-    logger.info("Step 1/4: Fetching PR details...")
-    pr_details: PRDetails | None = fetch_pr_details(repo_name, pr_number, token)  # noqa: E501
-    if not pr_details:
-        logger.error("Failed to fetch PR details")
-        return 1
-
-    pr_title: str = pr_details.get("title", "Unknown")
-    pr_description: str = pr_details.get("body", "")
-
-    logger.info(f"PR Title: {pr_title}")
-
-    # Fetch PR diff
-    logger.info("Step 2/4: Fetching PR diff...")
-    diff: str | None = fetch_pr_diff(repo_name, pr_number, token, config.max_diff_size)
-    if diff is None:
-        logger.error("Failed to fetch PR diff")
-        return 1
-
-    if not diff:
-        logger.warning("Warning: Empty diff - this might be a draft PR or no code changes")
-
-    # Build prompt
-    logger.info("Step 3/4: Building review prompt...")
-    prompt: str = build_review_prompt(pr_title, pr_description, diff)
-
-    # Call GLM API
-    logger.info("Step 4/4: Calling GLM API...")
-    review: str | None = call_glm_api(prompt, api_key, config)
-    if not review:
-        logger.error("Failed to get review from GLM API")
-        return 1
-
-    # Post review comment
-    logger.info("Posting review comment...")
-    if not post_review_comment(repo_name, pr_number, token, review):
-        logger.error("Failed to post review comment")
-        return 1
-
-    logger.info("PR review completed successfully!")
-    return 0
-
-
-if __name__ == "__main__":
-    sys.exit(main())

.github/workflows/pr-reviewer.yml

@@ -1,62 +0,0 @@
-name: VLLM-Omni PR Reviewer
-
-on:
-  issue_comment:
-    types: [created]
-
-permissions:
-  contents: read
-  pull-requests: write
-  issues: write
-
-jobs:
-  pr-reviewer:
-    name: Review Pull Request
-    runs-on: ubuntu-latest
-    timeout-minutes: 10
-    # Only run when the comment is from a collaborator/owner/member
-    if: |
-      github.event_name == 'issue_comment' &&
-      github.event.issue.pull_request != null &&
-      contains(github.event.comment.body, '@vllm-omni-reviewer') &&
-      (github.event.comment.author_association == 'MEMBER' ||
-       github.event.comment.author_association == 'COLLABORATOR' ||
-       github.event.comment.author_association == 'OWNER')
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v6.0.2
-        with:
-          fetch-depth: 0
-
-      - name: Set up Python
-        uses: actions/setup-python@v6.2.0
-        with:
-          python-version: '3.11'
-          cache: 'pip'
-
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip==24.0
-          pip install requests==2.31.0 pyyaml==6.0.1
-
-      - name: Run PR Reviewer
-        id: reviewer
-        env:
-          GLM_API_KEY: ${{ secrets.GLM_API_KEY }}
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          PR_NUMBER: ${{ github.event.issue.number || github.event.pull_request.number }}
-          COMMENT_BODY: ${{ github.event.comment.body }}
-          REPO_NAME: ${{ github.repository }}
-          PR_REVIEWER_COOLDOWN_MINUTES: 5
-          PR_REVIEWER_MAX_RETRIES: 3
-        run: |
-          python .github/scripts/pr_reviewer.py 2>&1 | tee "/tmp/pr_review_${PR_NUMBER}.log"
-
-      - name: Upload review logs
-        if: always()
-        uses: actions/upload-artifact@v7.0.0
-        with:
-          name: pr-review-logs-${{ github.event.issue.number || github.event.pull_request.number }}
-          path: /tmp/pr_review_*.log
-          retention-days: 7
-          if-no-files-found: ignore

.gitignore

@@ -83,6 +83,9 @@ target/
 profile_default/
 ipython_config.py
 
+# uv
+uv.lock
+
 # pyenv
 #   For a library or package, you might want to ignore these files since the code is
 #   intended to run in multiple environments; otherwise, check them in:
@@ -155,10 +158,23 @@ cython_debug/
 
 # Claude
 CLAUDE.md
-.claude/
+/.claude/*
+!.claude/skills/
+!.claude/skills/readme.md
+!.claude/skills/add-diffusion-model/
+!.claude/skills/add-diffusion-model/SKILL.md
+!.claude/skills/add-diffusion-model/references/
+!.claude/skills/add-diffusion-model/references/*.md
+!.claude/skills/add-tts-model/
+!.claude/skills/add-tts-model/SKILL.md
+!.claude/skills/review-pr/
+!.claude/skills/review-pr/SKILL.md
+!.claude/skills/review-pr/references/
+!.claude/skills/review-pr/references/*.md
 
 # Codex
 AGENTS.md
+.codex
 .codex/
 
 # cursor
@@ -188,6 +204,7 @@ checkpoints/
 # Cache directories
 cache/
 !vllm_omni/diffusion/cache/
+!tests/diffusion/cache/
 .cache/
 diffusion_cache/
 kv_cache/
@@ -222,6 +239,7 @@ data/
 datasets/
 *.csv
 *.json
+!tests/dfx/perf/tests/*.json
 !apps/ComfyUI-vLLM-Omni/example_workflows/*.json
 *.jsonl
 *.parquet
@@ -247,3 +265,8 @@ tmp_test
 vllm_omni/_version.py
 # output files
 *.wav
+# Vendored test reference audio (small, intentionally checked in)
+!tests/assets/**/*.wav
+.worktrees/
+# CI overlay yamls materialized from tests/utils.py:_CI_OVERLAYS at test time
+tests/.ci_generated/

.readthedocs.yml

@@ -10,6 +10,11 @@ build:
   jobs:
     post_checkout:
       - git fetch --unshallow || true
+    pre_install:
+      # Pre-install CPU torch so the subsequent `pip install .[docs]` does not
+      # pull ~2 GB of CUDA 13 libs (cuBLAS/cuDNN/NCCL/...) that exceed RTD's
+      # 1 GB memory cap and time out the mkdocs build.
+      - python -m pip install --index-url https://download.pytorch.org/whl/cpu torch
 
 mkdocs:
   configuration: mkdocs.yml

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://deepwiki.com/vllm-project/vllm-omni"><b>DeepWiki</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/README.md

@@ -1,44 +1,60 @@
-# Benchmarks Overview and Architecture
+# Benchmarks
 
-This document explains the benchmark architecture across all benchmark assets in this repo. It describes what we measure, and where to find or plug in new scenarios. Per-task details remain in subfolder READMEs (e.g., `benchmarks/<model>/README.md`).
+This directory contains benchmark suites for evaluating different model families and infrastructure components in vLLM-Omni. Each subfolder targets a different benchmark family with its own scripts, configs, and metrics. See the per-subfolder READMEs for detailed usage.
 
-## Scope and goals
-- Establish repeatable latency/throughput measurements for multimodal LLM pipelines.
-- Provide both HF Transformers (offline) and vLLM-Omni (multi-stage/pipeline) baselines.
-- Make it easy to plug in new datasets and models with minimal changes to the runner scripts.
+## Benchmark families
 
-## Dataset and inputs
-- Default example: SeedTTS top-100 prompts (`benchmarks/build_dataset/top100.txt`) via `benchmarks/build_dataset/`.
-- Extensible: drop in new prompt files or modality-aligned payloads; keep the expected format for the consuming scripts (e.g., one prompt per line).
-- If you add a new dataset, document it under `benchmarks/<model>/README.md` and point scripts to your data path.
+### [TTS](tts/README.md) — Text-to-Speech
 
-## Directory layout
-- `benchmarks/build_dataset/` — dataset prep utilities (e.g., SeedTTS top100).
-- `benchmarks/<model>/vllm_omni/` — vLLM-Omni pipeline benchmarks, logs, outputs.
-- `benchmarks/accuracy/` — accuracy benchmark integrations that adapt external
-  benchmark suites to vLLM-Omni serving and evaluation flows.
-- Add new tasks under `benchmarks/<model>/...` with the same pattern: `transformers/`, `vllm_omni/`, task-specific README, and (optionally) dataset prep notes.
+Model-agnostic serving benchmarks for TTS models, including Qwen3-TTS and VoxCPM2.
 
-## Reference workflows
-- **HF Transformers (offline, single process)**  
-  Script (example): `benchmarks/<model>/transformers/eval_qwen3_moe_omni_transformers.sh`  
-  Outputs: `benchmark_results/perf_stats.json`, `benchmark_results/results.json`, `benchmark_results/audio/` (if audio is produced).
+- **Layout**: `tts/bench_tts.py` (serving benchmark driver), `tts/model_configs.yaml` (model registry), `tts/plot_results.py` (result plotting)
+- **Dataset**: Seed-TTS full or text-only datasets, plus bundled smoke/design prompts under `build_dataset/`
+- **Key metrics**: TTFP (time to first audio packet), E2E latency, RTF (real-time factor), throughput (audio seconds / wall-clock second)
 
-- **vLLM-Omni end-to-end pipeline**  
-  Script (example): `benchmarks/<model>/vllm_omni/eval_qwen3_moe_omni.sh`  
-  Outputs: `vllm_omni/logs/*.stats.jsonl` (per-stage/overall latency & TPS), `vllm_omni/logs/stage*.log`, `vllm_omni/outputs/` (text/audio artifacts).
+### [Diffusion](diffusion/README.md) — Image and Video Generation
 
-- **Adding a new task/model**  
-  1) Create `benchmarks/<model>/transformers/` and/or `benchmarks/<model>/vllm_omni/` with scripts referencing your model and dataset.  
-  2) Add a task README describing dataset, configs, and expected outputs.  
-  3) Keep the output/log structure similar for easy comparison (perf_stats/results/audio or text outputs; stats.jsonl/logs for pipeline).
+Online-serving benchmark for diffusion image/video models, sending requests to the configured vLLM serving endpoint (`/v1/chat/completions`, `/v1/images/generations`, or `/v1/videos`, depending on backend/task).
 
-## Metrics to watch
-- **Throughput**: `overall_tps`, `*_tps_avg` per stage.
-- **Latency distribution**: look for long tails in `*.stats.jsonl`.
-- **Quality/completeness**: missing outputs or errors in stage logs indicate pipeline failures or misconfigurations.
+- **Tasks**: text-to-image, text-to-video, image-to-image, image-to-video, text+image-to-image, text+image-to-video
+- **Datasets**: `vbench`, `trace`, `random`
+- **Key metrics**: request throughput, latency percentiles, optional SLO attainment
 
-## Troubleshooting
-- Verify GPU/driver/FlashAttention2 requirements for your chosen model/config.
-- Ensure network access for dataset/model downloads (Google Drive, Hugging Face, etc.).
-- If outputs are missing or slow, inspect per-stage logs and `*.stats.jsonl` for errors, stragglers, or contention.
+### [GLM-Image](glm_image/README.md) — Text-to-Image and Image-to-Image
+
+Benchmarks for GLM-Image performance across HuggingFace baseline, vLLM-Omni offline inference, and vLLM-Omni online serving.
+
+- **Layout**: `glm_image/huggingface/` (HF baseline), `glm_image/vllm-omni/` (offline inference), `glm_image/benchmark_glm_image.py` (online serving)
+- **Tasks**: text-to-image and image-to-image
+- **Key metrics**: request/image throughput, latency percentiles, optional per-stage pipeline timings
+
+### [Distributed](distributed/omni_connectors/README.md) — RDMA Connector Testing
+
+RDMA environment setup and transfer tests for `MooncakeTransferEngineConnector`, including pytest-based single-node checks and manual cross-node benchmarks.
+
+- **Transfer modes**: `copy`, `zerocopy`, `gpu` (GPUDirect)
+- **Supports**: single-node pytest suites and manual multi-node/cross-node transfer testing
+
+### [Accuracy](accuracy/README.md) — Image Generation and Editing Quality
+
+Accuracy benchmarks for image generation/editing models, adapting external suites to vLLM-Omni serving and local judge-evaluation flows.
+
+- **Layout**: `accuracy/text_to_image/` (GEBench), `accuracy/image_to_image/` (GEdit-Bench)
+- **Method**: generation and judge scoring both run through local `vllm-omni serve` endpoints
+
+### Common serving metrics framework
+
+`vllm_omni/benchmarks/` extends `vllm bench serve --omni` with Omni-specific datasets, backends, and multimodal metrics. Key metrics include:
+
+- **Text output**: TTFT (time to first token), TPOT (time per output token), ITL (inter-token latency)
+- **Audio output**: TTFP (time to first audio packet), E2E latency, RTF (real-time factor)
+- **Throughput**: request throughput, output token throughput, total token throughput, audio throughput
+
+See `vllm_omni/benchmarks/serve.py` for the `vllm bench serve --omni` runner wrapper and `vllm_omni/benchmarks/metrics/` for Omni metric definitions.
+
+## Adding a new benchmark
+
+1. Create a subfolder under `benchmarks/<name>/` with scripts, configs if needed, and a `README.md`.
+2. If comparing against another runtime, use clear backend subfolders where applicable, such as `huggingface/` and `vllm-omni/`, or follow the shared TTS serving benchmark pattern in `tts/`.
+3. Add reusable dataset or prompt-building utilities to `build_dataset/` if applicable.
+4. Update this README with a link to the new benchmark family.

benchmarks/accuracy/README.md

@@ -23,5 +23,5 @@ Test guidance:
 - Local static/self-checks live in `tests/benchmarks/test_accuracy_bench_utils.py`.
 - End-to-end generation/evaluation should be validated in a remote GPU
   environment. In the current repo marker system there is `L4` but no `L5`
-  marker, so benchmark smoke tests should be wired as `advanced_model +
-  benchmark + L4` when GPU capacity is available.
+  marker, so benchmark smoke tests should be wired as `full_model +
+  benchmark + L4` for nightly when GPU capacity is available.

benchmarks/accuracy/image_to_image/README.md

@@ -99,5 +99,5 @@ Notes:
 - This flow requires the optional Hugging Face `datasets` package.
 - `generate` writes `generation_manifest.json` with local output coverage.
 - The current repo marker set exposes `L4` but not `L5`, so if you promote an
-  end-to-end smoke test into CI, use the existing `advanced_model`, `benchmark`,
-  and `L4` markers or introduce a new repo-wide marker explicitly first.
+  end-to-end smoke test into CI, use the `full_model`, `benchmark`,
+  and `L4` markers for nightly (or `advanced_model` for merge) or introduce a new repo-wide marker explicitly first.

benchmarks/build_dataset/download_process_data_seedtts.md

@@ -27,7 +27,7 @@ pip install gdown
 Download the dataset from Google Drive:
 
 ```bash
-gdown --id 1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
+gdown 1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
 ```
 
 ### 4. Extract the Dataset
@@ -74,7 +74,7 @@ rm meta.lst
 # Full setup and benchmark
 cd benchmarks/build_dataset
 pip install gdown
-gdown --id 1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
+gdown  1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
 tar -xf seedtts_testset.tar
 cp seedtts_testset/en/meta.lst meta.lst
 python extract_tts_prompts.py -i meta.lst -o top100.txt -n 100

benchmarks/build_dataset/seed_tts_design/en/meta.lst

@@ -0,0 +1,30 @@
+vd001|||The quick brown fox jumps over the lazy dog.|A warm, friendly female voice with a slight American Midwest accent, speaking at a moderate pace with natural inflection.
+vd002|||Welcome to the future of text-to-speech synthesis.|A deep, authoritative male news anchor voice, clear and professional with a measured cadence.
+vd003|||The sunset painted the sky in brilliant shades of orange and pink.|A gentle elderly female voice, soft and wise, with a slight Southern American accent.
+vd004|||Scientists have discovered a new species of deep-sea creature.|A young male voice with an Australian accent, curious and enthusiastic.
+vd005|||Breaking news: a major climate summit opens today in Geneva.|A crisp female newsreader voice, neutral accent, confident and precise.
+vd006|||In the beginning, there was darkness and silence across the universe.|A rich, dramatic bass male narrator voice, slow and deeply resonant.
+vd007|||Come closer, I have something important to tell you.|A soft, intimate female voice, slightly whispery, warm and gentle.
+vd008|||And they're off! The horses race toward the first turn at incredible speed.|An energetic male sports commentator, fast-paced and excited.
+vd009|||Once upon a time, in a land far away, lived a very clever fox.|A light, playful voice with childlike enthusiasm, bright and clear.
+vd010|||The ancient manuscript reveals secrets hidden for a thousand years.|A wise, measured elderly male voice, slow and deliberate, British English accent.
+vd011|||Good evening, ladies and gentlemen, and welcome to our show.|A sophisticated female voice with a slight French accent speaking English, elegant and refined.
+vd012|||System initialized. Running diagnostics. All systems nominal.|A clear, precise robotic-sounding voice, neutral and monotone with slight synthetic quality.
+vd013|||I hear what you are saying, and it is completely understandable to feel that way.|A warm, empathetic female therapist voice, calm and reassuring, unhurried pace.
+vd014|||Attention all units: proceed to grid reference seven-seven-alpha.|A firm, authoritative military male voice, clipped and commanding.
+vd015|||Oh my goodness, you have to try this amazing new recipe I just found!|An enthusiastic, bubbly female voice, high energy and friendly.
+vd016|||Dude, the waves were totally amazing out there today. Super happy about it!|A relaxed male voice with a California accent, casual and laid-back.
+vd017|||The quarterly results exceed expectations across all major metrics.|A sharp, businesslike female voice, confident and efficient, fast-paced delivery.
+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

@@ -0,0 +1,30 @@
+smoke001|||The quick brown fox jumps over the lazy dog near the riverbank at sunset.
+smoke002|||Welcome to the future of text-to-speech synthesis in production systems.
+smoke003|||Yesterday the team finished rolling out the new authentication flow.
+smoke004|||She walked carefully across the wet cobblestones, careful not to slip.
+smoke005|||The conference call is scheduled for nine in the morning, Pacific time.
+smoke006|||Please remember to save your work before closing the editor.
+smoke007|||Two plus two equals four, but five hundred and forty three digits is long.
+smoke008|||I would like a coffee with oat milk and a chocolate croissant please.
+smoke009|||The library closes at eight on weekdays and six on Saturdays.
+smoke010|||During the Renaissance, art and science flourished in European cities.
+smoke011|||He whispered the secret word so quietly that no one else could hear.
+smoke012|||Our flight departs from gate twenty three at eleven fifteen.
+smoke013|||The storm knocked out power for six hours, but the backup generator kicked in.
+smoke014|||Reading a good book on a rainy afternoon is one of life's great pleasures.
+smoke015|||When the kettle whistled, she poured the hot water over the fresh tea leaves.
+smoke016|||The algorithm runs in linear time, which is a big improvement over the previous approach.
+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

@@ -30,7 +30,7 @@ python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
 
 **Notes**
 
-- The benchmark talks to `http://<host>:<port>/v1/chat/completions`.
+- By default, image tasks talk to `http://<host>:<port>/v1/chat/completions`; video tasks talk to `/v1/videos`.
 - If you run the server on another host or port, pass `--base-url` accordingly.
 
 ## 2. Supported Datasets
@@ -95,8 +95,9 @@ You can point to your own trace using `--dataset-path`.
 
 ### Basic flags
 
-- `--base-url`: Server address (the script calls `.../v1/chat/completions`).
+- `--base-url`: Server address; `--endpoint` selects the path appended to this base URL.
 - `--model`: The OpenAI-compatible `model` field.
+- `--endpoint`: API endpoint path. Leading `/` is optional, e.g. `/v1/videos` or `v1/videos`.
 - `--task`: Task type (e.g., `t2i`, `t2v`, `i2i`, `i2v`).
 - `--dataset`: Dataset mode (`vbench` / `trace` / `random`).
 - `--num-prompts`: Number of requests to send.
@@ -130,9 +131,21 @@ Warmup flags:
 
 - `--warmup-requests`: Number of warmup requests.
 - `--warmup-num-inference-steps`: Steps used during warmup.
+- `--warmup-concurrency`: Maximum concurrent warmup requests. Use this to warm
+  the same batch shape as the measured run instead of warming only batch=`1`.
 - For `--task t2v`: warmup requests are forced to use `num_frames=1` to make warmup faster and less noisy.
 
 Traffic / concurrency flags:
 
 - `--request-rate`: Target request rate (requests/second). If set to `inf`, the script sends all requests immediately.
 - `--max-concurrency`: Max number of in-flight requests (default: `1`). This can hard-cap the achieved QPS: if it is too small, requests will queue behind the semaphore, and both achieved throughput and observed SLO attainment can be skewed.
+
+### Batched warmup note
+
+For batched serving runs, warm the same in-flight shape you plan to measure.
+For example, a run with `--max-concurrency 8` should usually also use
+`--warmup-requests 8 --warmup-concurrency 8`; otherwise the first measured
+batch may still pay compile or CUDA-graph capture cost.
+
+For a Qwen-Image continuous-batching replay example, see
+[`performance_dashboard/qwen_image_serving_performance.md`](./performance_dashboard/qwen_image_serving_performance.md).

benchmarks/diffusion/backends.py

@@ -10,6 +10,8 @@ from typing import Any
 import aiohttp
 from tqdm import tqdm
 
+DEFAULT_EDITS_BOT_TASK = "think"
+
 
 @dataclass
 class RequestFuncInput:
@@ -27,6 +29,7 @@ class RequestFuncInput:
     extra_body: dict[str, Any] = field(default_factory=dict)
     image_paths: list[str] | None = None
     request_id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    default_bot_task: str | None = DEFAULT_EDITS_BOT_TASK
 
 
 @dataclass
@@ -53,6 +56,95 @@ def _encode_image_as_data_url(path: str) -> str:
     return f"data:{mime};base64,{encoded}"
 
 
+async def async_request_image_edits(
+    input: RequestFuncInput,
+    session: aiohttp.ClientSession,
+    pbar: tqdm | None = None,
+    enable_diffusion_pipeline_profiler: bool = False,
+) -> RequestFuncOutput:
+    """POST /v1/images/edits (multipart)."""
+    del enable_diffusion_pipeline_profiler
+    output = RequestFuncOutput()
+    output.start_time = time.perf_counter()
+
+    extra_body = dict(input.extra_body)
+    width = input.width or extra_body.get("width") or 1024
+    height = input.height or extra_body.get("height") or 1024
+    edits_url = input.api_url
+
+    form = aiohttp.FormData()
+    form.add_field("model", input.model)
+    form.add_field("prompt", input.prompt)
+    form.add_field("size", f"{width}x{height}")
+    form.add_field("response_format", "b64_json")
+
+    if input.num_inference_steps is not None:
+        form.add_field("num_inference_steps", str(input.num_inference_steps))
+    elif extra_body.get("num_inference_steps") is not None:
+        form.add_field("num_inference_steps", str(extra_body["num_inference_steps"]))
+
+    if input.seed is not None:
+        form.add_field("seed", str(input.seed))
+    elif extra_body.get("seed") is not None:
+        form.add_field("seed", str(extra_body["seed"]))
+
+    if extra_body.get("guidance_scale") is not None:
+        form.add_field("guidance_scale", str(extra_body["guidance_scale"]))
+    if extra_body.get("negative_prompt") is not None:
+        form.add_field("negative_prompt", str(extra_body["negative_prompt"]))
+    if extra_body.get("true_cfg_scale") is not None:
+        form.add_field("true_cfg_scale", str(extra_body["true_cfg_scale"]))
+    if extra_body.get("sys_type") is not None:
+        form.add_field("sys_type", str(extra_body["sys_type"]))
+    if extra_body.get("system_prompt") is not None:
+        form.add_field("system_prompt", str(extra_body["system_prompt"]))
+
+    bot_task = extra_body.get("bot_task")
+    if bot_task is None and input.default_bot_task is not None:
+        bot_task = input.default_bot_task
+    if bot_task is not None:
+        form.add_field("bot_task", str(bot_task))
+
+    assert input.image_paths is not None
+    for img_path in input.image_paths:
+        if not os.path.exists(img_path):
+            output.error = f"Image file not found: {img_path}"
+            output.success = False
+            if pbar:
+                pbar.update(1)
+            return output
+        with open(img_path, "rb") as img_f:
+            image_bytes = img_f.read()
+        form.add_field(
+            "image",
+            image_bytes,
+            filename=os.path.basename(img_path),
+            content_type=_guess_mime_type(img_path),
+        )
+
+    try:
+        async with session.post(edits_url, data=form) as response:
+            if response.status == 200:
+                resp_json = await response.json()
+                output.response_body = resp_json
+                output.success = True
+            else:
+                output.error = f"HTTP {response.status}: {await response.text()}"
+                output.success = False
+    except Exception as e:
+        output.error = str(e)
+        output.success = False
+
+    output.latency = time.perf_counter() - output.start_time
+
+    if output.success and input.slo_ms is not None:
+        output.slo_achieved = (output.latency * 1000.0) <= float(input.slo_ms)
+
+    if pbar:
+        pbar.update(1)
+    return output
+
+
 async def async_request_chat_completions(
     input: RequestFuncInput,
     session: aiohttp.ClientSession,
@@ -122,6 +214,18 @@ async def async_request_chat_completions(
                                     output.peak_memory_mb = first_item.get("peak_memory_mb", 0.0)
                 except (IndexError, TypeError, AttributeError):
                     pass
+
+                if (not output.stage_durations or output.peak_memory_mb == 0.0) and isinstance(
+                    resp_json.get("metrics"), dict
+                ):
+                    m = resp_json["metrics"]
+                    if not output.stage_durations and isinstance(m.get("stage_durations"), dict):
+                        output.stage_durations = m.get("stage_durations") or {}
+                    if output.peak_memory_mb == 0.0 and m.get("peak_memory_mb") is not None:
+                        try:
+                            output.peak_memory_mb = float(m.get("peak_memory_mb") or 0.0)
+                        except (TypeError, ValueError):
+                            pass
             else:
                 output.error = f"HTTP {response.status}: {await response.text()}"
                 output.success = False
@@ -139,7 +243,7 @@ async def async_request_chat_completions(
     return output
 
 
-async def async_request_openai_images(
+async def async_request_openai_image_generations(
     input: RequestFuncInput,
     session: aiohttp.ClientSession,
     pbar: tqdm | None = None,
@@ -306,6 +410,8 @@ async def async_request_v1_videos(
             video_bytes = await content_response.read()
             output.response_body = video_bytes
             output.success = True
+            if "stage_durations" in poll_json:
+                output.stage_durations = poll_json["stage_durations"] or {}
             if "peak_memory_mb" in poll_json:
                 output.peak_memory_mb = poll_json["peak_memory_mb"]
             elif "peak_memory_mb" in resp_json:
@@ -334,12 +440,39 @@ async def async_request_v1_videos(
     return output
 
 
+LEGACY_BACKEND_ENDPOINT_ALIASES = {
+    "vllm-omni": "/v1/chat/completions",
+    "openai": "/v1/images/generations",
+}
+
+
+def normalize_endpoint(value: str) -> str:
+    endpoint = str(value).strip()
+    if not endpoint:
+        raise ValueError("endpoint must not be empty.")
+    endpoint = LEGACY_BACKEND_ENDPOINT_ALIASES.get(
+        endpoint,
+        LEGACY_BACKEND_ENDPOINT_ALIASES.get(endpoint.lstrip("/"), endpoint),
+    )
+    if not endpoint.startswith("/"):
+        endpoint = f"/{endpoint}"
+    return endpoint
+
+
+def endpoint_filename_token(value: str) -> str:
+    token = normalize_endpoint(value).lstrip("/")
+    for bad in ("/", "\\", ":", "*", "?", '"', "<", ">", "|"):
+        token = token.replace(bad, "_")
+    return token or "endpoint"
+
+
 backends_function_mapping = {
     "2i": {
-        "vllm-omni": (async_request_chat_completions, "/v1/chat/completions"),
-        "openai": (async_request_openai_images, "/v1/images/generations"),
+        "/v1/chat/completions": (async_request_chat_completions, "/v1/chat/completions"),
+        "/v1/images/generations": (async_request_openai_image_generations, "/v1/images/generations"),
+        "/v1/images/edits": (async_request_image_edits, "/v1/images/edits"),
     },
     "2v": {
-        "v1/videos": (async_request_v1_videos, "/v1/videos"),
+        "/v1/videos": (async_request_v1_videos, "/v1/videos"),
     },
 }

benchmarks/diffusion/bench_attention_backends.py

@@ -0,0 +1,328 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Standalone diagnostic for the diffusion attention backends added in #3079.
+
+Exercises the same synthetic attention shape through each backend and each
+torch SDPA sub-kernel, with and without an attn_mask, so we can pinpoint why
+CUDNN_ATTN underperforms on some SKUs. Two typical causes:
+
+  (a) cuDNN lacks a tuned kernel for this (SM, head_dim, seq) combination
+      and silently falls back to MATH.
+  (b) attn_mask is non-None and the cuDNN SDPA dispatch rejects it,
+      walking CUDNN -> FLASH -> MATH.
+
+Run:
+    python benchmarks/diffusion/bench_attention_backends.py --preset hv15
+    python benchmarks/diffusion/bench_attention_backends.py --preset wan22
+    python benchmarks/diffusion/bench_attention_backends.py \
+        --batch 1 --heads 24 --seq 14336 --head-dim 128
+
+Optional extras:
+    pip install --pre flash-attn-4        # FA4 is currently pre-release only
+    pip install -U flashinfer              # latest FlashInfer (0.6.9)
+
+The table at the end is the data we want on the PR — surface the row where a
+backend is >1.5x the SDPA baseline, that's the one to gate off the auto-route.
+"""
+
+from __future__ import annotations
+
+import argparse
+import time
+
+import torch
+from torch.nn.attention import SDPBackend, sdpa_kernel
+
+# Presets mirror the DiT hot-path attention call (single-stream block) at the
+# shapes we run in the PR's validation configs.
+_PRESETS = {
+    # HunyuanVideo-1.5 480p/33f: (30 * 52 * 9) latent + ~256 text tokens.
+    "hv15": {"batch": 1, "heads": 24, "seq": 14336, "head_dim": 128},
+    # Wan 2.2 480p/33f rough estimate; adjust when we measure real shapes.
+    "wan22": {"batch": 1, "heads": 40, "seq": 16384, "head_dim": 128},
+    # Smaller image-gen shape for quick smoke tests.
+    "flux": {"batch": 1, "heads": 24, "seq": 4096, "head_dim": 128},
+}
+
+_SDPA_BACKENDS = [
+    ("CUDNN_ATTENTION", [SDPBackend.CUDNN_ATTENTION]),
+    ("FLASH_ATTENTION", [SDPBackend.FLASH_ATTENTION]),
+    ("EFFICIENT_ATTENTION", [SDPBackend.EFFICIENT_ATTENTION]),
+    ("MATH", [SDPBackend.MATH]),
+    # The PR's CUDNN_ATTN impl uses this priority chain; keep it here to show
+    # what torch actually picks when multiple are allowed.
+    ("CUDNN_ATTN_CHAIN", [SDPBackend.CUDNN_ATTENTION, SDPBackend.FLASH_ATTENTION, SDPBackend.MATH]),
+]
+
+
+def _env_report() -> None:
+    if not torch.cuda.is_available():
+        raise SystemExit("CUDA not available; this diagnostic needs a GPU.")
+    cc = torch.cuda.get_device_capability()
+    print("=" * 72)
+    print(f"GPU           : {torch.cuda.get_device_name()}")
+    print(f"SM capability : sm_{cc[0]}{cc[1]}")
+    print(f"torch         : {torch.__version__}")
+    print(f"cuDNN version : {torch.backends.cudnn.version()}")
+    print("flashinfer    : ", end="")
+    try:
+        import flashinfer
+
+        print(getattr(flashinfer, "__version__", "present"))
+    except Exception as e:
+        print(f"not installed ({type(e).__name__})")
+    print("=" * 72)
+
+
+def _make_qkv(batch: int, heads: int, seq: int, head_dim: int, device: str, dtype: torch.dtype):
+    # Layout: (B, S, H, D) — what the PR's backends accept as input.
+    q = torch.randn(batch, seq, heads, head_dim, device=device, dtype=dtype)
+    k = torch.randn(batch, seq, heads, head_dim, device=device, dtype=dtype)
+    v = torch.randn(batch, seq, heads, head_dim, device=device, dtype=dtype)
+    return q, k, v
+
+
+def _make_mask(batch: int, seq: int, device: str, dtype: torch.dtype, pad_tokens: int = 128) -> torch.Tensor:
+    # Mirrors a text-encoder padding mask: last `pad_tokens` positions masked.
+    mask = torch.zeros(batch, 1, seq, seq, device=device, dtype=dtype)
+    if pad_tokens > 0:
+        mask[..., -pad_tokens:] = float("-inf")
+    return mask
+
+
+def _time_call(fn, *args, warmup: int = 3, iters: int = 10) -> tuple[float, str]:
+    """Return (median ms, error string). Skips on any exception so that a
+    backend that rejects our args (wrong dtype, missing JIT module, unsupported
+    kwarg value) doesn't abort the whole sweep."""
+    try:
+        for _ in range(warmup):
+            fn(*args)
+        torch.accelerator.synchronize()
+        times = []
+        for _ in range(iters):
+            torch.accelerator.synchronize()
+            t0 = time.perf_counter()
+            fn(*args)
+            torch.accelerator.synchronize()
+            times.append((time.perf_counter() - t0) * 1000.0)
+        times.sort()
+        return times[len(times) // 2], ""
+    except Exception as e:  # noqa: BLE001 — probe script, keep the whole table going
+        msg = str(e).split("\n", 1)[0][:60]
+        return float("nan"), f"{type(e).__name__}: {msg}" if msg else type(e).__name__
+
+
+def _run_sdpa_variants(q, k, v, attn_mask, scale: float) -> list[tuple[str, float, str]]:
+    # q/k/v come in (B, S, H, D); F.sdpa wants (B, H, S, D).
+    q_t = q.permute(0, 2, 1, 3)
+    k_t = k.permute(0, 2, 1, 3)
+    v_t = v.permute(0, 2, 1, 3)
+    rows: list[tuple[str, float, str]] = []
+    for name, backends in _SDPA_BACKENDS:
+
+        def _call(mask=attn_mask):
+            with sdpa_kernel(backends):
+                return torch.nn.functional.scaled_dot_product_attention(
+                    q_t, k_t, v_t, attn_mask=mask, dropout_p=0.0, is_causal=False, scale=scale
+                )
+
+        ms, err = _time_call(_call)
+        rows.append((name, ms, err))
+    return rows
+
+
+def _run_flashinfer(q, k, v, scale: float, backend: str | None = None, attn_mask=None) -> tuple[float, str]:
+    """Call FlashInfer's dense single-prefill.
+
+    ``backend`` hints at cutlass/fa3/trtllm if the installed version exposes
+    that kwarg. ``attn_mask`` is the same additive 4D mask we pass to SDPA —
+    we convert it to the 2D boolean ``custom_mask`` form FlashInfer accepts.
+    Per FlashInfer docs, ``custom_mask`` only applies when ``causal=False``.
+    """
+    try:
+        import inspect
+
+        from flashinfer.prefill import single_prefill_with_kv_cache
+    except Exception as e:
+        return float("nan"), f"import-{type(e).__name__}"
+
+    kwargs: dict = {"sm_scale": scale, "causal": False, "return_lse": False}
+    if backend is not None:
+        sig = inspect.signature(single_prefill_with_kv_cache)
+        if "backend" not in sig.parameters:
+            return float("nan"), "no-backend-kwarg"
+        kwargs["backend"] = backend
+
+    if attn_mask is not None:
+        # Collapse (B, 1, S, S) additive float mask to (S, S) boolean.
+        # FlashInfer expects True = keep, False = masked out.
+        mask_2d = attn_mask[0, 0]
+        kwargs["custom_mask"] = mask_2d != float("-inf")
+
+    def _call():
+        out = single_prefill_with_kv_cache(q[0], k[0], v[0], **kwargs)
+        return out.unsqueeze(0)
+
+    return _time_call(_call)
+
+
+def _run_fa4(q, k, v, scale: float) -> tuple[float, str]:
+    """Call FlashAttention-4 directly (``pip install flash-attn-4``). FA4
+    ships a Blackwell-native kernel via CuTe-DSL; on sm_120 it should beat
+    cuDNN by ~20%. API lives under ``flash_attn.cute``; the older
+    ``flash_attn.flash_attn_func`` path is FA2/FA3 only."""
+    try:
+        from flash_attn.cute import flash_attn_func
+    except Exception as e:
+        return float("nan"), f"import-{type(e).__name__}"
+
+    def _call():
+        # FA4 accepts (B, S, H, D) directly — same layout the PR's backends use.
+        return flash_attn_func(q, k, v, softmax_scale=scale, causal=False)
+
+    return _time_call(_call)
+
+
+def _run_flashinfer_cudnn_batch(q, k, v, scale: float) -> tuple[float, str]:
+    """Call FlashInfer's direct cuDNN wrapper. Bypasses PyTorch SDPA dispatch
+    (which has a few hundred ns of overhead per call) and talks to cuDNN FMHA
+    straight. Useful as a ceiling for 'pure cuDNN, no SDPA' on Blackwell."""
+    try:
+        from flashinfer.prefill import cudnn_batch_prefill_with_kv_cache
+    except Exception as e:
+        return float("nan"), f"import-{type(e).__name__}"
+
+    b, s, h, d = q.shape
+    qo_indptr = torch.tensor([0, s], dtype=torch.int32, device=q.device)
+    kv_indptr = torch.tensor([0, s], dtype=torch.int32, device=q.device)
+
+    def _call():
+        return cudnn_batch_prefill_with_kv_cache(
+            q.reshape(b * s, h, d),
+            k.reshape(b * s, h, d),
+            v.reshape(b * s, h, d),
+            qo_indptr=qo_indptr,
+            kv_indptr=kv_indptr,
+            max_qo_len=s,
+            max_kv_len=s,
+            sm_scale=scale,
+            causal=False,
+        )
+
+    return _time_call(_call)
+
+
+def _print_table(title: str, rows: list[tuple[str, float, str]], baseline_name: str | None = None) -> None:
+    print(f"\n{title}")
+    print("-" * 88)
+    header = f"{'backend':<24} {'median (ms)':>14}    {'vs baseline':>12}    status"
+    print(header)
+    print("-" * 88)
+    baseline_ms = None
+    if baseline_name is not None:
+        for name, ms, _ in rows:
+            if name == baseline_name and ms == ms:  # not NaN
+                baseline_ms = ms
+                break
+    for name, ms, err in rows:
+        ms_str = f"{ms:>14.3f}" if ms == ms else f"{'n/a':>14}"
+        if baseline_ms is not None and ms == ms:
+            ratio = baseline_ms / ms
+            ratio_str = f"{ratio:>11.2f}x"
+        else:
+            ratio_str = f"{'—':>12}"
+        status = "ok" if not err else f"FAILED ({err})"
+        print(f"{name:<24} {ms_str}    {ratio_str}    {status}")
+
+
+def _pick_winner(rows: list[tuple[str, float, str]]) -> tuple[str, float] | None:
+    """Return (backend, ms) of the fastest non-failing row, or None."""
+    ok_rows = [(n, m) for n, m, err in rows if not err and m == m]
+    if not ok_rows:
+        return None
+    return min(ok_rows, key=lambda x: x[1])
+
+
+def _bench_one_shape(shape: dict, dtype: torch.dtype, device: str) -> tuple[list, list]:
+    torch.manual_seed(0)
+    q, k, v = _make_qkv(**shape, device=device, dtype=dtype)
+    scale = 1.0 / (shape["head_dim"] ** 0.5)
+
+    rows_nomask = _run_sdpa_variants(q, k, v, attn_mask=None, scale=scale)
+    rows_nomask.append(("FLASHINFER (default)", *_run_flashinfer(q, k, v, scale)))
+    # `trtllm-gen` has no sm_120 cubins (NVIDIA/TensorRT-LLM#11799) — skipped.
+    for fi_backend in ("fa2", "fa3", "cutlass", "auto"):
+        rows_nomask.append((f"FLASHINFER ({fi_backend})", *_run_flashinfer(q, k, v, scale, backend=fi_backend)))
+    rows_nomask.append(("FLASHINFER (cudnn-batch)", *_run_flashinfer_cudnn_batch(q, k, v, scale)))
+    rows_nomask.append(("FA4 (direct)", *_run_fa4(q, k, v, scale)))
+
+    mask = _make_mask(shape["batch"], shape["seq"], device=device, dtype=dtype, pad_tokens=256)
+    rows_mask = _run_sdpa_variants(q, k, v, attn_mask=mask, scale=scale)
+    rows_mask.append(("FLASHINFER (dense)", *_run_flashinfer(q, k, v, scale, attn_mask=mask)))
+    rows_mask.append(("FA4 (direct)", float("nan"), "mask-not-supported"))
+
+    return rows_nomask, rows_mask
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("--preset", choices=list(_PRESETS.keys()), default="hv15")
+    parser.add_argument("--sweep", action="store_true", help="Run all presets and print a ranking")
+    parser.add_argument("--batch", type=int, default=None)
+    parser.add_argument("--heads", type=int, default=None)
+    parser.add_argument("--seq", type=int, default=None)
+    parser.add_argument("--head-dim", type=int, default=None)
+    parser.add_argument("--dtype", choices=["bf16", "fp16"], default="bf16")
+    parser.add_argument("--device", default="cuda:0")
+    args = parser.parse_args()
+
+    dtype = torch.bfloat16 if args.dtype == "bf16" else torch.float16
+    _env_report()
+    print(f"dtype         : {args.dtype}")
+
+    presets = list(_PRESETS.keys()) if args.sweep else [args.preset]
+    summary: dict[str, tuple[str, float] | None] = {}
+
+    for preset_name in presets:
+        shape = dict(_PRESETS[preset_name])
+        if not args.sweep:
+            for k in ("batch", "heads", "seq", "head_dim"):
+                v = getattr(args, k if k != "head_dim" else "head_dim")
+                if v is not None:
+                    shape[k] = v
+
+        print("\n" + "=" * 88)
+        print(
+            f"Preset: {preset_name}  |  batch={shape['batch']} heads={shape['heads']} "
+            f"seq={shape['seq']} head_dim={shape['head_dim']}"
+        )
+        print("=" * 88)
+
+        rows_nomask, rows_mask = _bench_one_shape(shape, dtype, args.device)
+        _print_table(f"[{preset_name}] No attention mask", rows_nomask, baseline_name="CUDNN_ATTENTION")
+        _print_table(
+            f"[{preset_name}] With attention mask (pad 256 tokens)", rows_mask, baseline_name="CUDNN_ATTENTION"
+        )
+
+        summary[preset_name] = _pick_winner(rows_nomask)
+
+    if args.sweep or len(presets) > 1:
+        print("\n" + "=" * 88)
+        print("Winners per preset (no-mask path)")
+        print("=" * 88)
+        for preset_name, winner in summary.items():
+            if winner is None:
+                print(f"  {preset_name:<10} — no successful backend")
+            else:
+                name, ms = winner
+                print(f"  {preset_name:<10} {name:<24} {ms:>8.3f} ms")
+
+    print("\nNotes:")
+    print("  * Ratios are relative to CUDNN_ATTENTION. >1.0x means faster than cuDNN.")
+    print("  * Mask-path winner inherits CUDNN_ATTN's fallback in the PR's backends.")
+    print("  * `trtllm-gen` and FA4 4.0.0b10 are known-broken on sm_120 as of Apr 2026.")
+    print("  * For e2e timings, run text_to_video.py / text_to_image.py while varying DIFFUSION_ATTENTION_BACKEND.")
+
+
+if __name__ == "__main__":
+    main()

benchmarks/diffusion/bench_sage_attn3_matrix.sh

@@ -0,0 +1,108 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# 2x2 timing matrix (backend x compile mode) for SAGE_ATTN_3 vs TORCH_SDPA,
+# on HunyuanVideo-1.5 480p T2V and Wan 2.2 TI2V-5B. Run from the repo root:
+#
+#     bash benchmarks/diffusion/bench_sage_attn3_matrix.sh
+#
+# Requires: sageattn3 installed (for SAGE_ATTN_3 rows) and a Blackwell GPU.
+
+set -u
+
+STEPS=${STEPS:-50}
+SEED=${SEED:-42}
+PROMPT=${PROMPT:-"A dog running across a field of golden wheat."}
+OUT_DIR=${OUT_DIR:-/tmp/bench_sage3}
+mkdir -p "$OUT_DIR"
+
+declare -A TOTAL PER_STEP TEXT_ENC TRANSFORMER VAE
+
+run() {
+    local name="$1"; shift
+    local model="$1"; shift
+    local h="$1"; shift
+    local w="$1"; shift
+    local f="$1"; shift
+    local g="$1"; shift
+    local log="$OUT_DIR/${name}.log"
+    echo "[$(date +%T)] Running $name ..."
+    env "$@" python examples/offline_inference/text_to_video/text_to_video.py \
+        --model "$model" \
+        --prompt "$PROMPT" \
+        --height "$h" --width "$w" --num-frames "$f" \
+        --num-inference-steps "$STEPS" --seed "$SEED" --guidance-scale "$g" \
+        --enable-diffusion-pipeline-profiler \
+        --output "$OUT_DIR/${name}.mp4" > "$log" 2>&1 || {
+            echo "  FAILED — see $log"
+            TOTAL[$name]="FAIL"; PER_STEP[$name]="FAIL"
+            TEXT_ENC[$name]="-"; TRANSFORMER[$name]="-"; VAE[$name]="-"
+            return
+        }
+
+    local wait_ms
+    wait_ms=$(grep "add_req_and_wait=" "$log" | tail -1 | sed -nE 's/.*add_req_and_wait=([0-9.]+) ms.*/\1/p')
+    local total_s
+    total_s=$(grep -oE "Total generation time: [0-9.]+ seconds" "$log" | tail -1 | sed -nE 's/.*: ([0-9.]+) seconds.*/\1/p')
+    local text_enc_s
+    text_enc_s=$(grep -oE "text_encoder.forward took [0-9.]+s" "$log" | awk '{sub(/s$/,"",$NF); sum+=$NF} END {printf "%.3f", sum+0}')
+    local vae_s
+    vae_s=$(grep -oE "vae.decode took [0-9.]+s" "$log" | awk '{sub(/s$/,"",$NF); sum+=$NF} END {printf "%.3f", sum+0}')
+    local pipeline_s
+    pipeline_s=$(grep -oE "Pipeline\.forward took [0-9.]+s" "$log" | tail -1 | sed -nE 's/.*took ([0-9.]+)s.*/\1/p')
+    local transformer_s
+    transformer_s=$(awk "BEGIN {printf \"%.3f\", ${pipeline_s:-0} - ${vae_s:-0} - ${text_enc_s:-0}}")
+
+    TOTAL[$name]="${total_s:-0}"
+    PER_STEP[$name]=$(awk "BEGIN {printf \"%.3f\", ${wait_ms:-0} / 1000 / $STEPS}")
+    TEXT_ENC[$name]="$text_enc_s"
+    TRANSFORMER[$name]="$transformer_s"
+    VAE[$name]="$vae_s"
+
+    echo "  total=${TOTAL[$name]}s  per-step=${PER_STEP[$name]}s/it  transformer=${TRANSFORMER[$name]}s  vae=${VAE[$name]}s"
+}
+
+# ========== HunyuanVideo-1.5 480p T2V (480x832x33, CFG=6.0) ==========
+HV_MODEL=hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v
+run hv15_sdpa_eager     "$HV_MODEL" 480 832 33 6.0 TORCH_COMPILE_DISABLE=1
+run hv15_sdpa_compiled  "$HV_MODEL" 480 832 33 6.0
+run hv15_sage3_eager    "$HV_MODEL" 480 832 33 6.0 TORCH_COMPILE_DISABLE=1 DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN_3
+run hv15_sage3_compiled "$HV_MODEL" 480 832 33 6.0                         DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN_3
+
+# ========== Wan 2.2 TI2V-5B (704x1280x33, CFG=5.0) ==========
+WAN_MODEL=Wan-AI/Wan2.2-TI2V-5B-Diffusers
+run wan22_sdpa_eager     "$WAN_MODEL" 704 1280 33 5.0 TORCH_COMPILE_DISABLE=1
+run wan22_sdpa_compiled  "$WAN_MODEL" 704 1280 33 5.0
+run wan22_sage3_eager    "$WAN_MODEL" 704 1280 33 5.0 TORCH_COMPILE_DISABLE=1 DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN_3
+run wan22_sage3_compiled "$WAN_MODEL" 704 1280 33 5.0                         DIFFUSION_ATTENTION_BACKEND=SAGE_ATTN_3
+
+print_table() {
+    local title="$1"; shift
+    echo
+    echo "$title (${STEPS} steps, seed=${SEED})"
+    echo
+    printf "| %-23s | %-9s | %-15s | %-12s | %-15s | %-14s |\n" \
+        "Config" "Total (s)" "Per-step (s/it)" "Text Enc (s)" "Transformer (s)" "VAE Decode (s)"
+    printf "| %-23s | %-9s | %-15s | %-12s | %-15s | %-14s |\n" \
+        "-----------------------" "---------" "---------------" "------------" "---------------" "--------------"
+    for row in "$@"; do
+        case "$row" in
+            *_sdpa_eager)     label="SDPA + Eager" ;;
+            *_sdpa_compiled)  label="SDPA + Compiled" ;;
+            *_sage3_eager)    label="SAGE_ATTN_3 + Eager" ;;
+            *_sage3_compiled) label="SAGE_ATTN_3 + Compiled" ;;
+        esac
+        printf "| %-23s | %-9s | %-15s | %-12s | %-15s | %-14s |\n" \
+            "$label" "${TOTAL[$row]}" "${PER_STEP[$row]}" "${TEXT_ENC[$row]}" "${TRANSFORMER[$row]}" "${VAE[$row]}"
+    done
+}
+
+print_table "HunyuanVideo-1.5 480p T2V, 832x480x33" \
+    hv15_sdpa_eager hv15_sdpa_compiled hv15_sage3_eager hv15_sage3_compiled
+
+print_table "Wan 2.2 TI2V-5B, 1280x704x33" \
+    wan22_sdpa_eager wan22_sdpa_compiled wan22_sage3_eager wan22_sage3_compiled
+
+echo
+echo "Logs: $OUT_DIR/*.log"

benchmarks/diffusion/diffusion_benchmark_serving.py

@@ -6,31 +6,34 @@
 Benchmark online serving for diffusion models (Image/Video Generation).
 If you want to use i2v, i2i dataset, you should `uv pip install gdown` first
 
-Supports multiple backends:
-    - vllm-omni: Uses /v1/chat/completions endpoint (default)
-    - openai: Uses /v1/images/generations endpoint
-    - v1/videos: Use /v1/videos endpoint
+Supports multiple endpoints:
+    - /v1/chat/completions: OpenAI chat-compatible image requests (e.g. t2i, Qwen i2i)
+    - /v1/images/edits: OpenAI image edit / IT2I (multipart; e.g. Hunyuan --bot-task think)
+    - /v1/images/generations: OpenAI image generation requests
+    - /v1/videos: Async video jobs
+
+Legacy --backend vllm-omni and openai are aliases for chat/completions and images/generations.
 
 Usage:
-    # Video (vllm-omni backend)
+    # Video (/v1/videos endpoint)
     t2v:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend vllm-omni --dataset vbench --task t2v --num-prompts 10 \
+        --endpoint /v1/videos --dataset vbench --task t2v --num-prompts 10 \
         --height 480 --width 640 --fps 16 --num-frames 80
 
     i2v:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend vllm-omni --dataset vbench --task i2v --num-prompts 10
+        --endpoint /v1/videos --dataset vbench --task i2v --num-prompts 10
 
 
-    # Image (vllm-omni backend)
+    # Image (/v1/chat/completions endpoint)
     t2i:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend vllm-omni --dataset vbench --task t2i --num-prompts 10 \
+        --endpoint /v1/chat/completions --dataset vbench --task t2i --num-prompts 10 \
         --height 1024 --width 1024
 
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend vllm-omni --dataset random --task t2i --num-prompts 1 \
+        --endpoint /v1/chat/completions --dataset random --task t2i --num-prompts 1 \
         --max-concurrency 1 --enable-negative-prompt \
         --random-request-config '[
             {"width":512,"height":512,"num_inference_steps":20,"weight":0.15},
@@ -39,20 +42,32 @@ Usage:
             {"width":1536,"height":1536,"num_inference_steps":35,"weight":0.15}
         ]'
 
-    i2i:
+    ti2i (Hunyuan / OpenAI image edit API):
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend vllm-omni --dataset vbench --task i2i --num-prompts 10
+        --endpoint /v1/images/edits --dataset random --task ti2i --num-prompts 10 \
+        --bot-task think
 
-    # Image (openai backend)
+    python benchmarks/diffusion/diffusion_benchmark_serving.py \
+      --endpoint /v1/images/edits \
+      --dataset custom \
+      --dataset-path custom_requests.jsonl \
+      --task ti2i \
+      --bot-task think
+
+    i2i (chat-based models such as Qwen-Image-Edit):
+    python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
+        --endpoint /v1/chat/completions --dataset vbench --task i2i --num-prompts 10
+
+    # Image (/v1/images/generations endpoint)
     t2i:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend openai --dataset vbench --task t2i --num-prompts 10 \
+        --endpoint /v1/images/generations --dataset vbench --task t2i --num-prompts 10 \
         --height 1024 --width 1024 --port 3000
 
-    # Video (v1/vedeos)
+    # Video (v1/videos)
     t2v:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend v1/videos --dataset random --task t2v --num-prompts 1 \
+        --endpoint /v1/videos --dataset random --task t2v --num-prompts 1 \
         --max-concurrency 1 --enable-negative-prompt \
         --random-request-config '[
             {"width":854,"height":480,"num_inference_steps":18,"num_frames":120,"fps":24,"weight":1}
@@ -80,7 +95,12 @@ from typing import Any
 import aiohttp
 import numpy as np
 import requests
-from backends import RequestFuncInput, RequestFuncOutput, backends_function_mapping
+from backends import (
+    RequestFuncInput,
+    RequestFuncOutput,
+    backends_function_mapping,
+    normalize_endpoint,
+)
 from PIL import Image
 from tqdm.asyncio import tqdm
 
@@ -127,7 +147,7 @@ class VBenchDataset(BaseDataset):
     def _load_data(self) -> list[dict[str, Any]]:
         if self.args.task == "t2v":
             return self._load_t2v_prompts()
-        elif self.args.task in ["i2v", "ti2v", "ti2i", "i2i"]:
+        elif self.args.task in ["i2v", "ti2v", "ti2i", "i2i", "it2i"]:
             return self._load_i2v_data()
         else:
             return self._load_t2v_prompts()
@@ -519,7 +539,7 @@ class TraceDataset(BaseDataset):
             single = row.get("image_path")
             image_paths = [single] if single else None
 
-        if not image_paths and self.args.task in ["i2v", "i2i", "ti2v", "ti2i"]:
+        if not image_paths and self.args.task in ["i2v", "i2i", "ti2v", "ti2i", "it2i"]:
             raise ValueError(
                 f"Task {self.args.task} requires image input, but no image_path or image_paths found in trace row."
             )
@@ -553,11 +573,141 @@ class TraceDataset(BaseDataset):
         return [self[i] for i in range(len(self))]
 
 
+class CustomDataset(BaseDataset):
+    """
+    Custom dataset that loads requests from a JSONL file.
+
+    Each line in the JSONL file should be a JSON object with the following fields:
+    - prompt (required): The text prompt for the request
+    - width (optional): Image/video width
+    - height (optional): Image height
+    - num_inference_steps (optional): Number of diffusion steps
+    - seed (optional): Random seed
+    - image_paths (optional): List of input image paths for i2i/i2v/ti2i tasks
+    - image_urls (optional): List of input image URLs (alternative to image_paths)
+
+    Example JSONL for ti2i:
+    {"prompt": "Add sunset lighting", "width": 1024, "height": 1024, "image_urls": ["https://example.com/image.jpg"]}
+    """
+
+    def __init__(self, args, api_url: str, model: str):
+        super().__init__(args, api_url, model)
+        self.dataset_path = args.dataset_path
+        if not self.dataset_path:
+            raise ValueError("--dataset-path must be provided when using 'custom' dataset")
+        self.load_data()
+
+    def load_data(self) -> None:
+        """Load data from JSONL file."""
+        import pandas as pd
+
+        if not self.dataset_path.endswith(".jsonl"):
+            raise ValueError("Custom dataset must be a JSONL file")
+
+        try:
+            df = pd.read_json(path_or_buf=self.dataset_path, lines=True)
+        except Exception as e:
+            raise ValueError(f"Failed to load JSONL file {self.dataset_path}: {e}")
+
+        if "prompt" not in df.columns:
+            raise ValueError("JSONL file must contain a 'prompt' column")
+
+        self.data = df.to_dict("records")
+        print(f"Loaded {len(self.data)} requests from {self.dataset_path}")
+
+    def _resolve_image_paths(self, item: dict) -> list[str] | None:
+        # Handle local file paths
+        if "image_paths" in item and item["image_paths"]:
+            paths = item["image_paths"]
+            if isinstance(paths, str):
+                paths = [paths]
+
+            # Verify all paths exist
+            valid_paths = []
+            for path in paths:
+                if os.path.exists(path):
+                    valid_paths.append(path)
+                else:
+                    raise ValueError(f"Image file not found: {path}")
+
+            return valid_paths if valid_paths else None
+
+        # Handle URLs - download to temp files
+        if "image_urls" in item and item["image_urls"]:
+            urls = item["image_urls"]
+            if isinstance(urls, str):
+                urls = [urls]
+
+            downloaded_paths = []
+            for url in urls:
+                try:
+                    response = requests.get(url, timeout=30)
+                    response.raise_for_status()
+
+                    # Create temp file with appropriate extension
+                    suffix = os.path.splitext(url)[1] or ".png"
+                    with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
+                        tmp.write(response.content)
+                        downloaded_paths.append(tmp.name)
+                except Exception as e:
+                    raise ValueError(f"Failed to download image from {url}: {e}")
+
+            return downloaded_paths if downloaded_paths else None
+
+        return None
+
+    def __len__(self) -> int:
+        return len(self.data)
+
+    def __getitem__(self, idx: int) -> RequestFuncInput:
+        item = self.data[idx]
+
+        # Extract fields with fallback to args defaults
+        prompt = item["prompt"]
+        width = item.get("width", self.args.width)
+        height = item.get("height", self.args.height)
+        num_inference_steps = item.get("num_inference_steps", self.args.num_inference_steps)
+        seed = item.get("seed", self.args.seed)
+        reserved_keys = {"prompt", "width", "height", "num_inference_steps", "seed", "image_paths", "image_urls"}
+        extra_body = {k: v for k, v in item.items() if k not in reserved_keys}
+
+        # Handle image paths/URLs
+        image_paths = self._resolve_image_paths(item)
+
+        return RequestFuncInput(
+            prompt=prompt,
+            api_url=self.api_url,
+            model=self.model,
+            seed=seed,
+            image_paths=image_paths,
+            extra_body=extra_body,
+            width=width,
+            height=height,
+            num_inference_steps=num_inference_steps,
+        )
+
+    def get_requests(self) -> list[RequestFuncInput]:
+        """Get all requests, cycling through data if num_prompts > dataset size."""
+        num_requests = getattr(self.args, "num_prompts", len(self.data))
+
+        if num_requests <= len(self.data):
+            return [self[i] for i in range(num_requests)]
+
+        # Cycle through the dataset to reach num_prompts
+        requests = []
+        for i in range(num_requests):
+            idx = i % len(self.data)
+            requests.append(self[idx])
+
+        return requests
+
+
 class RandomDataset(BaseDataset):
     def __init__(self, args, api_url: str, model: str, enable_negative_prompt: bool = False):
         super().__init__(args, api_url, model)
         self.num_prompts = args.num_prompts
         self.enable_negative_prompt = enable_negative_prompt
+        self.num_input_images = max(1, args.num_input_images)
         self.random_request_config = getattr(args, "random_request_config", None)
         if self.random_request_config:
             self.random_request_config = json.loads(self.random_request_config)
@@ -579,12 +729,8 @@ class RandomDataset(BaseDataset):
             self._sampled_requests = None
 
         # Random image generate
-        if self.args.task in ["i2v", "ti2v", "ti2i", "i2i"]:
-            img = Image.new("RGB", (512, 512), (255, 255, 255))
-
-            image_path = os.path.join(tempfile.gettempdir(), "diffusion_benchmark_random_image.png")
-            self._random_image_path = [image_path]
-            img.save(image_path)
+        if self.args.task in ["i2v", "ti2v", "ti2i", "i2i", "it2i"]:
+            self._random_image_path = self._generate_random_image_paths()
         else:
             self._random_image_path = None
 
@@ -619,6 +765,18 @@ class RandomDataset(BaseDataset):
     def get_requests(self) -> list[RequestFuncInput]:
         return [self[i] for i in range(len(self))]
 
+    def _generate_random_image_paths(self) -> list[str]:
+        image_paths: list[str] = []
+        for image_idx in range(self.num_input_images):
+            img = Image.new("RGB", (512, 512), (255, 255, 255))
+            image_path = os.path.join(
+                tempfile.gettempdir(),
+                f"diffusion_benchmark_random_image_{image_idx}.png",
+            )
+            img.save(image_path)
+            image_paths.append(image_path)
+        return image_paths
+
 
 def _compute_expected_latency_ms_from_base(req: RequestFuncInput, args, base_time_ms: float | None) -> float | None:
     """Compute expected execution time (ms) based on a base per-step-per-frame unit time.
@@ -742,6 +900,52 @@ async def iter_requests(
         yield req
 
 
+def _make_warmup_request(
+    requests_list: list[RequestFuncInput],
+    index: int,
+    args,
+) -> RequestFuncInput:
+    warm_req = requests_list[index % len(requests_list)]
+    if args.warmup_num_inference_steps is not None:
+        warm_req = replace(
+            warm_req,
+            num_inference_steps=args.warmup_num_inference_steps,
+        )
+    if args.task == "t2v":
+        warm_req = replace(warm_req, num_frames=1)
+    return warm_req
+
+
+async def _run_warmups(
+    requests_list: list[RequestFuncInput],
+    args,
+    session: aiohttp.ClientSession,
+    request_func,
+) -> list[tuple[RequestFuncInput, RequestFuncOutput]]:
+    if not args.warmup_requests or not requests_list:
+        return []
+
+    warmup_requests = [_make_warmup_request(requests_list, i, args) for i in range(args.warmup_requests)]
+    warmup_concurrency = min(int(args.warmup_concurrency), len(warmup_requests))
+    warmup_semaphore = asyncio.Semaphore(warmup_concurrency)
+
+    print(
+        f"Running {len(warmup_requests)} warmup request(s) "
+        f"with num_inference_steps={args.warmup_num_inference_steps} "
+        f"and warmup_concurrency={warmup_concurrency}..."
+    )
+
+    async def limited_warmup_request_func(
+        req: RequestFuncInput,
+    ) -> RequestFuncOutput:
+        async with warmup_semaphore:
+            return await request_func(req, session, None)
+
+    warmup_tasks = [asyncio.create_task(limited_warmup_request_func(req)) for req in warmup_requests]
+    warmup_outputs = await asyncio.gather(*warmup_tasks)
+    return list(zip(warmup_requests, warmup_outputs))
+
+
 def calculate_metrics(
     outputs: list[RequestFuncOutput],
     total_duration: float,
@@ -828,13 +1032,23 @@ def wait_for_service(base_url: str, timeout: int = 120) -> None:
         time.sleep(1)
 
 
+def _default_endpoint_for_task(task: str) -> str:
+    if task in {"t2v", "i2v", "ti2v"}:
+        return "/v1/videos"
+    if task in {"i2i", "ti2i", "it2i"}:
+        return "/v1/images/edits"
+    if task == "t2i":
+        return "/v1/chat/completions"
+    raise ValueError(f"Unsupported task for endpoint resolution: {task}")
+
+
 async def benchmark(args):
     # Construct base_url if not provided
     if args.base_url is None:
         args.base_url = f"http://{args.host}:{args.port}"
 
     VIDEO_TASKS = {"t2v", "i2v", "ti2v"}
-    IMAGE_TASKS = {"t2i", "i2i", "ti2i"}
+    IMAGE_TASKS = {"t2i", "i2i", "ti2i", "it2i"}
 
     if args.task in VIDEO_TASKS:
         task_type = "2v"
@@ -847,18 +1061,23 @@ async def benchmark(args):
             f"Valid image tasks: {sorted(IMAGE_TASKS)}"
         )
 
-    valid_backends = sorted(backends_function_mapping[task_type].keys())
+    raw_endpoint = args.endpoint if args.endpoint is not None else args.backend
+    if raw_endpoint is None:
+        raw_endpoint = _default_endpoint_for_task(args.task)
+    args.endpoint = normalize_endpoint(raw_endpoint)
 
-    if args.backend not in valid_backends:
+    valid_endpoints = sorted(backends_function_mapping[task_type].keys())
+
+    if args.endpoint not in valid_endpoints:
         logger.error(
-            f"Invalid backend '{args.backend}' for task '{args.task}' (task type: '{task_type}').\n"
-            f"Valid backends for this task type: {valid_backends}\n"
-            f"Example usage: --task {args.task} --backend {valid_backends[0]}"
+            f"Invalid endpoint '{args.endpoint}' for task '{args.task}' (task type: '{task_type}').\n"
+            f"Valid endpoints for this task type: {valid_endpoints}\n"
+            f"Example usage: --task {args.task} --endpoint {valid_endpoints[0]}"
         )
-        raise ValueError("Backend validation failed. See log above for valid options.")
+        raise ValueError("Endpoint validation failed. See log above for valid options.")
 
-    # Setup API URL and request function based on backend
-    request_func, api_url = backends_function_mapping[task_type][args.backend]
+    # Setup API URL and request function based on endpoint.
+    request_func, api_url = backends_function_mapping[task_type][args.endpoint]
     api_url = f"{args.base_url}{api_url}"
 
     if args.dataset == "vbench":
@@ -867,6 +1086,8 @@ async def benchmark(args):
         dataset = TraceDataset(args, api_url, args.model)
     elif args.dataset == "random":
         dataset = RandomDataset(args, api_url, args.model, args.enable_negative_prompt)
+    elif args.dataset == "custom":
+        dataset = CustomDataset(args, api_url, args.model)
     else:
         raise ValueError(f"Unknown dataset: {args.dataset}")
 
@@ -874,6 +1095,10 @@ async def benchmark(args):
     requests_list = dataset.get_requests()
     print(f"Prepared {len(requests_list)} requests from {args.dataset} dataset.")
 
+    if args.endpoint == "/v1/images/edits":
+        for req in requests_list:
+            req.default_bot_task = args.bot_task
+
     # Limit concurrency
     if args.max_concurrency is not None:
         semaphore = asyncio.Semaphore(args.max_concurrency)
@@ -891,23 +1116,12 @@ async def benchmark(args):
     pbar = tqdm(total=len(requests_list), disable=args.disable_tqdm)
 
     async with aiohttp.ClientSession() as session:
-        warmup_pairs: list[tuple[RequestFuncInput, RequestFuncOutput]] = []
-        if args.warmup_requests and requests_list:
-            print(
-                f"Running {args.warmup_requests} warmup request(s) \
-                with num_inference_steps={args.warmup_num_inference_steps}..."
-            )
-            for i in range(args.warmup_requests):
-                warm_req = requests_list[i % len(requests_list)]
-                if args.warmup_num_inference_steps is not None:
-                    warm_req = replace(
-                        warm_req,
-                        num_inference_steps=args.warmup_num_inference_steps,
-                    )
-                if args.task == "t2v":
-                    warm_req = replace(warm_req, num_frames=1)
-                warm_out = await limited_request_func(warm_req, session, None)
-                warmup_pairs.append((warm_req, warm_out))
+        warmup_pairs = await _run_warmups(
+            requests_list=requests_list,
+            args=args,
+            session=session,
+            request_func=request_func,
+        )
 
         if args.slo:
             # Prefer trace-provided per-request slo_ms. Only populate when missing.
@@ -932,15 +1146,17 @@ async def benchmark(args):
     metrics = calculate_metrics(outputs, total_duration, requests_list, args, args.slo)
 
     # Add configuration info to metrics for JSON output
-    metrics["backend"] = args.backend
+    metrics["endpoint"] = args.endpoint
     metrics["model"] = args.model
     metrics["dataset"] = args.dataset
     metrics["task"] = args.task
+    if args.endpoint == "/v1/images/edits":
+        metrics["bot_task"] = args.bot_task
 
-    print("\n{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=60, c="="))
+    print("\n{s:{c}^{n}}".format(s=" Serving Benchmark Result ", n=50, c="="))
 
     # Section 1: Configuration
-    print("{:<40} {:<15}".format("Backend:", args.backend))
+    print("{:<40} {:<15}".format("Endpoint:", args.endpoint))
     print("{:<40} {:<15}".format("Model:", args.model))
     print("{:<40} {:<15}".format("Dataset:", args.dataset))
     print("{:<40} {:<15}".format("Task:", args.task))
@@ -1003,25 +1219,30 @@ if __name__ == "__main__":
     parser.add_argument("--host", type=str, default="localhost", help="Server host.")
     parser.add_argument("--port", type=int, default=8091, help="Server port.")
     parser.add_argument("--model", type=str, default="default", help="Model name.")
+    parser.add_argument(
+        "--endpoint",
+        type=str,
+        default=None,
+        help=("Endpoint path to target. Leading '/' is optional, e.g. /v1/videos or v1/videos."),
+    )
     parser.add_argument(
         "--backend",
         type=str,
-        default="vllm-omni",
-        choices=["vllm-omni", "openai", "v1/videos"],
-        help="Backend to target the benchmark to.",
+        default=None,
+        help=argparse.SUPPRESS,
     )
     parser.add_argument(
         "--dataset",
         type=str,
         default="vbench",
-        choices=["vbench", "trace", "random"],
+        choices=["vbench", "trace", "random", "custom"],
         help="Dataset to use.",
     )
     parser.add_argument(
         "--task",
         type=str,
         default="t2v",
-        choices=["t2v", "i2v", "ti2v", "ti2i", "i2i", "t2i"],
+        choices=["t2v", "i2v", "ti2v", "ti2i", "i2i", "it2i", "t2i"],
         help="Task type.",
     )
     parser.add_argument(
@@ -1063,6 +1284,13 @@ if __name__ == "__main__":
         default=1,
         help="num_inference_steps used for warmup requests.",
     )
+    parser.add_argument(
+        "--warmup-concurrency",
+        type=int,
+        default=1,
+        help="Maximum number of warmup requests to run concurrently. "
+        "Set this to match the real batch shape when warming up torch.compile or CUDA graphs.",
+    )
     parser.add_argument("--width", type=int, default=None, help="Image/Video width.")
     parser.add_argument("--height", type=int, default=None, help="Image/Video height.")
     parser.add_argument("--num-frames", type=int, default=None, help="Number of frames (for video).")
@@ -1115,7 +1343,21 @@ if __name__ == "__main__":
             '{"width":768,"height":768,"num_inference_steps":20,"weight":0.85}]'
         ),
     )
+    parser.add_argument(
+        "--num-input-images",
+        type=int,
+        default=1,
+        help=(
+            "Number of synthetic input images to attach for image-conditioned tasks "
+            "(i2v, ti2v, ti2i, i2i, it2i) when using random dataset."
+        ),
+    )
+    parser.add_argument(
+        "--bot-task",
+        type=str,
+        default="think",
+        help=("bot_task form field for --endpoint /v1/images/edits (think, recaption, think_recaption, vanilla)."),
+    )
 
     args = parser.parse_args()
-
     asyncio.run(benchmark(args))

benchmarks/diffusion/performance_dashboard/qwen_image_serving_performance.md

@@ -33,6 +33,12 @@ vllm serve Qwen/Qwen-Image --omni \
     --port 8091
 ```
 
+To replay step-wise continuous batching, compare
+`vllm serve Qwen/Qwen-Image --omni --port 8089 --step-execution --max-num-seqs 1`
+against the same command with `--max-num-seqs 8`. `--step-execution` is the
+feature gate, and increasing `--max-num-seqs` above `1` lets the scheduler
+keep more compatible requests active at once.
+
 ## 3.2 Key Parameters
 
 | Parameter             | Description              |
@@ -52,7 +58,7 @@ Record these parameters when reporting performance results.
 
 ```bash
 python benchmarks/diffusion/diffusion_benchmark_serving.py \
-    --backend vllm-omni \
+    --endpoint /v1/chat/completions \
     --dataset <DATASET_NAME> \
     --task t2i \
     --num-prompts <N> \
@@ -65,11 +71,12 @@ python benchmarks/diffusion/diffusion_benchmark_serving.py \
 
 | Parameter              | Description                       |
 | ---------------------- | --------------------------------- |
-| `--backend`            | Serving backend (use `vllm-omni`) |
+| `--endpoint`           | API endpoint (use `/v1/chat/completions`; leading `/` is optional) |
 | `--dataset`            | Dataset name (`random` or custom) |
 | `--task`               | Task type (e.g., `t2i`)           |
 | `--num-prompts`        | Total number of requests          |
 | `--max-concurrency`    | Client-side concurrency           |
+| `--warmup-concurrency` | Warmup concurrency for batch-shape prewarming |
 | `--random-request-config`| JSON string defining random request |
 
 ---
@@ -122,7 +129,7 @@ python benchmarks/diffusion/diffusion_benchmark_serving.py \
 
 ```bash
 python benchmarks/diffusion/diffusion_benchmark_serving.py \
-    --backend vllm-omni \
+    --endpoint /v1/chat/completions \
     --dataset random \
     --task t2i \
     --num-prompts 1 \
@@ -133,6 +140,12 @@ python benchmarks/diffusion/diffusion_benchmark_serving.py \
     ]'
 ```
 
+For a continuous-batching A/B replay, run the same benchmark command twice
+against the two serving configs above and keep traffic plus warmup settings
+identical across both runs. If the measured run uses `--max-concurrency 8`,
+warm with `--warmup-requests 8 --warmup-concurrency 8` so the first measured
+batch does not include compile or CUDA-graph capture overhead.
+
 ---
 
 # 6. Performance Metrics
@@ -163,6 +176,7 @@ To ensure consistent and comparable benchmark results:
 * Record parallel configuration
 * Record benchmark parameters (resolution, concurrency, number of prompts)
 * Ensure no background workload on GPUs during testing
+* When comparing continuous batching, keep warmup settings the same across runs
 
 ---
 

benchmarks/diffusion/performance_dashboard/wan_2_2_serving_performance.md

@@ -53,7 +53,7 @@ Record these parameters when reporting performance results.
 
 ```bash
 python benchmarks/diffusion/diffusion_benchmark_serving.py \
-    --backend v1/videos \
+    --endpoint /v1/videos \
     --dataset <DATASET_NAME> \
     --task t2v \
     --num-prompts <N> \
@@ -66,7 +66,7 @@ python benchmarks/diffusion/diffusion_benchmark_serving.py \
 
 | Parameter              | Description                       |
 | ---------------------- | --------------------------------- |
-| `--backend`            | Serving backend (use `v1/videos`) |
+| `--endpoint`           | API endpoint (use `/v1/videos`; leading `/` is optional) |
 | `--dataset`            | Dataset name (`random` or custom) |
 | `--task`               | Task type (e.g., `t2v`)           |
 | `--num-prompts`        | Total number of requests          |
@@ -120,7 +120,7 @@ python benchmarks/diffusion/diffusion_benchmark_serving.py \
 
 ```bash
 python benchmarks/diffusion/diffusion_benchmark_serving.py \
-    --backend v1/videos \
+    --endpoint /v1/videos \
     --dataset random \
     --task t2v \
     --num-prompts 1 \

benchmarks/diffusion/quantization_quality.py

@@ -34,6 +34,16 @@ Video example (text-to-video):
         --height 720 --width 1280 \
         --num-frames 81 --num-inference-steps 40 --seed 42
 
+LTX-2 example (text-to-video; audio output is generated but not scored by LPIPS):
+    python benchmarks/diffusion/quantization_quality.py \
+        --model Lightricks/LTX-Video-2 \
+        --task t2v \
+        --quantization fp8 int8 \
+        --prompts \
+            "A serene lakeside sunrise with mist over the water" \
+        --height 704 --width 1216 \
+        --num-frames 121 --num-inference-steps 40 --seed 42
+
 Multiple quantization methods:
     python benchmarks/diffusion/quantization_quality.py \
         --model Tongyi-MAI/Z-Image-Turbo \
@@ -148,10 +158,11 @@ def _build_omni_kwargs(args, quantization=None):
 def _generate_image(omni, args, prompt, seed):
     """Generate a single image and return (PIL.Image, time_seconds, memory_gib)."""
     from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+    from vllm_omni.outputs import OmniRequestOutput
     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},
@@ -163,10 +174,11 @@ 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
+    req_out = OmniRequestOutput.unwrap_result(outputs)
+    if not req_out.images:
+        raise ValueError("Could not extract image output from result.")
     img = req_out.images[0]
     return img, elapsed, peak_mem
 
@@ -178,7 +190,7 @@ 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 +204,7 @@ 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):
@@ -202,10 +214,19 @@ def _generate_video(omni, args, prompt, seed):
         else:
             frames = inner
     elif hasattr(first, "images") and first.images:
-        frames = first.images
+        frames = first.images[0]
     else:
         raise ValueError("Could not extract video frames from output.")
 
+    # LTX-2 (and similar audio+video models) may surface a dict or (video, audio) tuple
+    if isinstance(frames, dict):
+        frames = frames.get("video") or frames.get("frames")
+    elif isinstance(frames, tuple) and len(frames) == 2:
+        frames = frames[0]
+
+    if frames is None:
+        raise ValueError("Could not extract video frames from output.")
+
     if isinstance(frames, torch.Tensor):
         video = frames.detach().cpu()
         if video.dim() == 5:
@@ -223,13 +244,15 @@ def _generate_video(omni, args, prompt, seed):
     return frames_array, elapsed, peak_mem
 
 
-def _unload_omni(omni):
-    """Delete Omni instance and free GPU memory."""
-    del omni
+def _free_gpu_memory():
+    """Force GC and release cached GPU memory.
+
+    Must be called AFTER the caller has dropped (i.e., via `del`)
+    every reference to the Omni instance"""
     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):
@@ -265,7 +288,9 @@ def run_benchmark(args):
 
     bl_avg_time = np.mean([v[1] for v in baseline_outputs.values()])
     bl_mem = baseline_outputs[prompts[0]][2]  # use first prompt's memory
-    _unload_omni(omni_bl)
+    omni_bl.shutdown()
+    del omni_bl
+    _free_gpu_memory()
 
     # Save baseline outputs
     bl_dir = output_dir / "baseline"
@@ -305,7 +330,9 @@ def run_benchmark(args):
 
         qt_avg_time = np.mean([v[1] for v in qt_outputs.values()])
         qt_mem = qt_outputs[prompts[0]][2]
-        _unload_omni(omni_qt)
+        omni_qt.shutdown()
+        del omni_qt
+        _free_gpu_memory()
 
         # Save quantized outputs
         qt_dir = output_dir / config_label.replace(" ", "_")

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/fish-speech/bench_speaker_cache.py

@@ -0,0 +1,290 @@
+"""Benchmark Fish Speech voice cache: inline ref_audio vs uploaded voice.
+
+Measures TTFP improvement from DAC-code caching when using uploaded voices.
+
+Setup:
+  1. Start vllm-omni with Fish Speech S2 Pro
+  2. Provide a reference audio file for voice cloning
+
+Usage:
+    python bench_speaker_cache.py \
+        --ref-audio /path/to/reference.wav \
+        --ref-text "Transcript of the reference audio." \
+        --num-prompts 20 \
+        --port 8091
+
+The script runs two rounds:
+  A) Inline ref_audio: every request sends base64 audio (no cache)
+  B) Uploaded voice: upload once, then use voice name (cache hits after 1st)
+"""
+
+import argparse
+import asyncio
+import base64
+import json
+import os
+import sys
+import time
+from pathlib import Path
+
+import aiohttp
+
+# Allow imports from benchmarks/fish-speech/
+sys.path.insert(0, str(Path(__file__).resolve().parent))
+
+from fish_bench_utils import (  # noqa: E402
+    BenchmarkResult,
+    RequestResult,
+    compute_stats,
+    print_benchmark_results,
+    send_streaming_request,
+)
+
+SAMPLE_RATE = 44100
+SAMPLE_WIDTH = 2
+
+PROMPTS = [
+    "Hello, welcome to the voice synthesis benchmark test.",
+    "She said she would be here by noon, but nobody showed up.",
+    "The quick brown fox jumps over the lazy dog near the riverbank.",
+    "I can't believe how beautiful the sunset looks from up here.",
+    "Please remember to bring your identification documents tomorrow morning.",
+    "Have you ever wondered what it would be like to travel through time?",
+    "The restaurant on the corner serves the best pasta I have ever tasted.",
+    "After the meeting, we should discuss the quarterly results.",
+    "Learning a new language takes patience and genuine curiosity.",
+    "The train leaves at half past seven, so we need to arrive early.",
+    "Could you please turn down the music, I'm trying to concentrate.",
+    "It was a dark and stormy night when the keeper heard a knock.",
+]
+
+
+def encode_audio_to_base64(audio_path: str) -> str:
+    """Encode a local audio file to base64 data URL."""
+    ext = audio_path.lower().rsplit(".", 1)[-1]
+    mime_map = {"wav": "audio/wav", "mp3": "audio/mpeg", "flac": "audio/flac"}
+    mime_type = mime_map.get(ext, "audio/wav")
+    with open(audio_path, "rb") as f:
+        audio_b64 = base64.b64encode(f.read()).decode("utf-8")
+    return f"data:{mime_type};base64,{audio_b64}"
+
+
+async def upload_voice(
+    host: str,
+    port: int,
+    audio_path: str,
+    ref_text: str,
+    voice_name: str = "bench_voice",
+) -> dict:
+    """Upload a voice via POST /v1/audio/voices."""
+    url = f"http://{host}:{port}/v1/audio/voices"
+    data = aiohttp.FormData()
+    data.add_field("name", voice_name)
+    data.add_field("consent", "true")
+    if ref_text:
+        data.add_field("ref_text", ref_text)
+    data.add_field(
+        "audio_sample",
+        open(audio_path, "rb"),
+        filename=os.path.basename(audio_path),
+        content_type="audio/wav",
+    )
+
+    async with aiohttp.ClientSession() as session:
+        async with session.post(url, data=data) as resp:
+            result = await resp.json()
+            print(f"  Upload response ({resp.status}): {json.dumps(result, indent=2)}")
+            return result
+
+
+async def delete_voice(host: str, port: int, voice_name: str) -> None:
+    """Delete an uploaded voice."""
+    url = f"http://{host}:{port}/v1/audio/voices/{voice_name}"
+    async with aiohttp.ClientSession() as session:
+        async with session.delete(url) as resp:
+            if resp.status == 200:
+                print(f"  Deleted voice '{voice_name}'")
+
+
+async def run_round(
+    host: str,
+    port: int,
+    num_prompts: int,
+    create_payload_fn,
+    label: str,
+    num_warmups: int = 2,
+    timeout_s: float = 120.0,
+) -> BenchmarkResult:
+    """Run one benchmark round and return results."""
+    api_url = f"http://{host}:{port}/v1/audio/speech"
+    connector = aiohttp.TCPConnector(limit=1, limit_per_host=1)
+    session = aiohttp.ClientSession(
+        connector=connector,
+        timeout=aiohttp.ClientTimeout(total=timeout_s),
+    )
+
+    try:
+        # Warmup.
+        if num_warmups > 0:
+            print(f"  [{label}] Warming up ({num_warmups} requests)...")
+            for i in range(num_warmups):
+                payload = create_payload_fn(PROMPTS[i % len(PROMPTS)])
+                r = await send_streaming_request(
+                    session,
+                    api_url,
+                    payload,
+                    SAMPLE_RATE,
+                    SAMPLE_WIDTH,
+                )
+                status = "OK" if r.success else f"FAIL: {r.error[:80]}"
+                print(f"    warmup {i + 1}: ttfp={r.ttfp * 1000:.0f}ms  {status}")
+
+        # Benchmark.
+        print(f"  [{label}] Running {num_prompts} requests (concurrency=1)...")
+        results: list[RequestResult] = []
+        start = time.perf_counter()
+        for i in range(num_prompts):
+            prompt = PROMPTS[i % len(PROMPTS)]
+            payload = create_payload_fn(prompt)
+            r = await send_streaming_request(
+                session,
+                api_url,
+                payload,
+                SAMPLE_RATE,
+                SAMPLE_WIDTH,
+            )
+            results.append(r)
+            tag = "HIT" if i > 0 and label == "uploaded_voice" else ""
+            print(
+                f"    req {i + 1:3d}: ttfp={r.ttfp * 1000:7.1f}ms  "
+                f"e2e={r.e2e * 1000:7.1f}ms  "
+                f"{'OK' if r.success else 'FAIL'} {tag}"
+            )
+        wall_time = time.perf_counter() - start
+    finally:
+        await session.close()
+
+    bench = compute_stats(results, wall_time)
+    bench.concurrency = 1
+    bench.num_prompts = num_prompts
+    bench.config_name = label
+    return bench
+
+
+async def main():
+    parser = argparse.ArgumentParser(
+        description="Benchmark Fish Speech voice cache (inline vs uploaded)",
+    )
+    parser.add_argument("--host", default="127.0.0.1")
+    parser.add_argument("--port", type=int, default=8091)
+    parser.add_argument("--ref-audio", required=True, help="Path to reference audio file")
+    parser.add_argument("--ref-text", required=True, help="Transcript of reference audio")
+    parser.add_argument("--num-prompts", type=int, default=20)
+    parser.add_argument("--num-warmups", type=int, default=2)
+    parser.add_argument("--voice-name", default="bench_voice")
+    args = parser.parse_args()
+
+    if not os.path.exists(args.ref_audio):
+        print(f"Error: ref_audio not found: {args.ref_audio}")
+        sys.exit(1)
+
+    ref_audio_b64 = encode_audio_to_base64(args.ref_audio)
+    print(f"Reference audio: {args.ref_audio} ({len(ref_audio_b64) // 1024}KB base64)")
+
+    # ---- Round A: Inline ref_audio (no cache) ----
+    print(f"\n{'=' * 60}")
+    print("Round A: INLINE ref_audio (every request sends full audio)")
+    print(f"{'=' * 60}")
+
+    def make_inline_payload(prompt: str) -> dict:
+        return {
+            "input": prompt,
+            "voice": "default",
+            "stream": True,
+            "response_format": "pcm",
+            "ref_audio": ref_audio_b64,
+            "ref_text": args.ref_text,
+            "max_new_tokens": 2048,
+        }
+
+    bench_inline = await run_round(
+        args.host,
+        args.port,
+        args.num_prompts,
+        make_inline_payload,
+        "inline_ref_audio",
+        num_warmups=args.num_warmups,
+    )
+    print_benchmark_results(bench_inline)
+
+    # ---- Upload voice ----
+    print(f"\n{'=' * 60}")
+    print("Uploading voice for cache test...")
+    print(f"{'=' * 60}")
+    await delete_voice(args.host, args.port, args.voice_name)
+    await upload_voice(
+        args.host,
+        args.port,
+        args.ref_audio,
+        args.ref_text,
+        args.voice_name,
+    )
+
+    # ---- Round B: Uploaded voice (cache hits after 1st request) ----
+    print(f"\n{'=' * 60}")
+    print("Round B: UPLOADED VOICE (cache hits after 1st request)")
+    print(f"{'=' * 60}")
+
+    def make_uploaded_payload(prompt: str) -> dict:
+        return {
+            "input": prompt,
+            "voice": args.voice_name,
+            "stream": True,
+            "response_format": "pcm",
+            "ref_text": args.ref_text,
+            "max_new_tokens": 2048,
+        }
+
+    bench_cached = await run_round(
+        args.host,
+        args.port,
+        args.num_prompts,
+        make_uploaded_payload,
+        "uploaded_voice",
+        num_warmups=args.num_warmups,
+    )
+    print_benchmark_results(bench_cached)
+
+    # ---- Comparison ----
+    print(f"\n{'=' * 60}")
+    print("COMPARISON: Inline ref_audio vs Uploaded voice (cached)")
+    print(f"{'=' * 60}")
+    print(f"{'Metric':<30} {'Inline':>12} {'Cached':>12} {'Speedup':>10}")
+    print(f"{'-' * 64}")
+
+    def fmt_speedup(inline_val: float, cached_val: float) -> str:
+        if cached_val > 0 and inline_val > 0:
+            ratio = inline_val / cached_val
+            return f"{ratio:.2f}x"
+        return "N/A"
+
+    rows = [
+        ("Mean TTFP (ms)", bench_inline.mean_ttfp_ms, bench_cached.mean_ttfp_ms),
+        ("Median TTFP (ms)", bench_inline.median_ttfp_ms, bench_cached.median_ttfp_ms),
+        ("P99 TTFP (ms)", bench_inline.p99_ttfp_ms, bench_cached.p99_ttfp_ms),
+        ("Mean E2E (ms)", bench_inline.mean_e2e_ms, bench_cached.mean_e2e_ms),
+        ("Median E2E (ms)", bench_inline.median_e2e_ms, bench_cached.median_e2e_ms),
+        ("Mean RTF", bench_inline.mean_rtf, bench_cached.mean_rtf),
+    ]
+    for label, a, b in rows:
+        print(f"{label:<30} {a:>12.1f} {b:>12.1f} {fmt_speedup(a, b):>10}")
+
+    print("\nNote: Round B request #1 is a cache MISS (cold start).")
+    print("      Requests #2+ are cache HITs (skip DAC encoding).")
+
+    # Cleanup.
+    await delete_voice(args.host, args.port, args.voice_name)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

benchmarks/fish-speech/fish_bench_utils.py

@@ -0,0 +1,501 @@
+"""Shared benchmark infrastructure for Fish Speech serving benchmarks.
+
+Provides common dataclasses, metrics computation, streaming HTTP client,
+and result formatting used by model-specific benchmark scripts.
+
+Model-specific scripts supply a ``create_payload_fn(prompt) -> dict``
+callback and audio parameters; everything else is handled here.
+"""
+
+import asyncio
+import base64
+import json
+import time
+from collections.abc import Callable
+from dataclasses import asdict, dataclass, field
+from datetime import datetime
+from pathlib import Path
+
+import aiohttp
+import numpy as np
+from tqdm.asyncio import tqdm
+
+# ---------------------------------------------------------------------------
+# Shared test prompts (varying length for realistic workload)
+# ---------------------------------------------------------------------------
+PROMPTS = [
+    "Hello, welcome to the voice synthesis benchmark test.",
+    "She said she would be here by noon, but nobody showed up.",
+    "The quick brown fox jumps over the lazy dog near the riverbank.",
+    "I can't believe how beautiful the sunset looks from up here on the mountain.",
+    "Please remember to bring your identification documents to the appointment tomorrow morning.",
+    "Have you ever wondered what it would be like to travel through time and visit ancient civilizations?",
+    "The restaurant on the corner serves the best pasta I have ever tasted in my entire life.",
+    "After the meeting, we should discuss the quarterly results and plan for the next phase.",
+    "Learning a new language takes patience, practice, and a genuine curiosity about other cultures.",
+    "The train leaves at half past seven, so we need to arrive at the station before then.",
+    "Could you please turn down the music a little bit, I'm trying to concentrate on my work.",
+    "It was a dark and stormy night when the old lighthouse keeper heard a knock at the door.",
+]
+
+
+# ---------------------------------------------------------------------------
+# Dataclasses
+# ---------------------------------------------------------------------------
+@dataclass
+class RequestResult:
+    success: bool = False
+    ttfp: float = 0.0  # Time to first audio packet (seconds)
+    e2e: float = 0.0  # End-to-end latency (seconds)
+    audio_bytes: int = 0  # Total audio bytes received
+    audio_duration: float = 0.0  # Audio duration in seconds
+    rtf: float = 0.0  # Real-time factor = e2e / audio_duration
+    prompt: str = ""
+    error: str = ""
+
+
+@dataclass
+class BenchmarkResult:
+    config_name: str = ""
+    concurrency: int = 0
+    num_prompts: int = 0
+    completed: int = 0
+    failed: int = 0
+    duration_s: float = 0.0
+    # TTFP stats (ms)
+    mean_ttfp_ms: float = 0.0
+    median_ttfp_ms: float = 0.0
+    std_ttfp_ms: float = 0.0
+    p90_ttfp_ms: float = 0.0
+    p95_ttfp_ms: float = 0.0
+    p99_ttfp_ms: float = 0.0
+    # E2E stats (ms)
+    mean_e2e_ms: float = 0.0
+    median_e2e_ms: float = 0.0
+    std_e2e_ms: float = 0.0
+    p90_e2e_ms: float = 0.0
+    p95_e2e_ms: float = 0.0
+    p99_e2e_ms: float = 0.0
+    # RTF stats
+    mean_rtf: float = 0.0
+    median_rtf: float = 0.0
+    std_rtf: float = 0.0
+    p99_rtf: float = 0.0
+    # Audio stats
+    mean_audio_duration_s: float = 0.0
+    total_audio_duration_s: float = 0.0
+    audio_throughput: float = 0.0  # audio_duration / wall_time
+    request_throughput: float = 0.0  # requests / second
+    # Per-request details
+    per_request: list = field(default_factory=list)
+
+
+# ---------------------------------------------------------------------------
+# Audio helpers
+# ---------------------------------------------------------------------------
+def pcm_bytes_to_duration(
+    num_bytes: int,
+    sample_rate: int = 24000,
+    sample_width: int = 2,
+) -> float:
+    """Convert raw PCM byte count to duration in seconds."""
+    return num_bytes / sample_width / sample_rate
+
+
+def _is_sse_response(response: aiohttp.ClientResponse) -> bool:
+    content_type = (response.headers.get("Content-Type") or "").lower()
+    return "text/event-stream" in content_type
+
+
+async def _read_raw_audio_stream(
+    response: aiohttp.ClientResponse,
+    *,
+    start_time: float,
+) -> tuple[int, float]:
+    first_audio_at = 0.0
+    total_bytes = 0
+
+    async for chunk in response.content.iter_any():
+        if chunk and first_audio_at <= 0:
+            first_audio_at = time.perf_counter() - start_time
+        total_bytes += len(chunk)
+
+    return total_bytes, first_audio_at
+
+
+def _extract_sse_payload(raw_event: bytes) -> bytes | None:
+    data_lines: list[bytes] = []
+    for raw_line in raw_event.splitlines():
+        line = raw_line.rstrip(b"\r")
+        if line.startswith(b"data: "):
+            data_lines.append(line[6:])
+        elif line.startswith(b"data:"):
+            data_lines.append(line[5:].lstrip())
+
+    if not data_lines:
+        return None
+    return b"\n".join(data_lines).strip()
+
+
+async def _read_sse_audio_stream(
+    response: aiohttp.ClientResponse,
+    *,
+    start_time: float,
+) -> tuple[int, float]:
+    """Decode SSE events and count raw audio bytes from base64 payloads."""
+    first_audio_at = 0.0
+    total_bytes = 0
+    pending = b""
+
+    async for chunk in response.content.iter_any():
+        if not chunk:
+            continue
+        pending += chunk
+        pending = pending.replace(b"\r\n", b"\n")
+
+        while b"\n\n" in pending:
+            raw_event, pending = pending.split(b"\n\n", 1)
+            payload_bytes = _extract_sse_payload(raw_event)
+            if payload_bytes is None:
+                continue
+            if payload_bytes == b"[DONE]":
+                return total_bytes, first_audio_at
+
+            try:
+                payload = json.loads(payload_bytes)
+            except json.JSONDecodeError as exc:
+                raise ValueError(f"Invalid SSE JSON payload: {exc}") from exc
+
+            audio = payload.get("audio")
+            if not isinstance(audio, dict):
+                continue
+
+            audio_b64 = audio.get("data")
+            if not audio_b64:
+                continue
+
+            try:
+                audio_bytes = base64.b64decode(audio_b64)
+            except Exception as exc:
+                raise ValueError(f"Invalid base64 audio chunk: {exc}") from exc
+
+            if audio_bytes and first_audio_at <= 0:
+                first_audio_at = time.perf_counter() - start_time
+            total_bytes += len(audio_bytes)
+
+    return total_bytes, first_audio_at
+
+
+# ---------------------------------------------------------------------------
+# Metrics
+# ---------------------------------------------------------------------------
+def compute_stats(
+    results: list[RequestResult],
+    wall_time: float,
+) -> BenchmarkResult:
+    """Compute aggregate statistics from per-request results."""
+    successful = [r for r in results if r.success]
+    failed = [r for r in results if not r.success]
+
+    bench = BenchmarkResult(
+        completed=len(successful),
+        failed=len(failed),
+        duration_s=wall_time,
+    )
+
+    if not successful:
+        return bench
+
+    ttfps = [r.ttfp * 1000 for r in successful]
+    e2es = [r.e2e * 1000 for r in successful]
+    rtfs = [r.rtf for r in successful]
+    audio_durs = [r.audio_duration for r in successful]
+
+    bench.mean_ttfp_ms = float(np.mean(ttfps))
+    bench.median_ttfp_ms = float(np.median(ttfps))
+    bench.std_ttfp_ms = float(np.std(ttfps))
+    bench.p90_ttfp_ms = float(np.percentile(ttfps, 90))
+    bench.p95_ttfp_ms = float(np.percentile(ttfps, 95))
+    bench.p99_ttfp_ms = float(np.percentile(ttfps, 99))
+
+    bench.mean_e2e_ms = float(np.mean(e2es))
+    bench.median_e2e_ms = float(np.median(e2es))
+    bench.std_e2e_ms = float(np.std(e2es))
+    bench.p90_e2e_ms = float(np.percentile(e2es, 90))
+    bench.p95_e2e_ms = float(np.percentile(e2es, 95))
+    bench.p99_e2e_ms = float(np.percentile(e2es, 99))
+
+    bench.mean_rtf = float(np.mean(rtfs))
+    bench.median_rtf = float(np.median(rtfs))
+    bench.std_rtf = float(np.std(rtfs))
+    bench.p99_rtf = float(np.percentile(rtfs, 99))
+
+    bench.mean_audio_duration_s = float(np.mean(audio_durs))
+    bench.total_audio_duration_s = float(np.sum(audio_durs))
+    bench.audio_throughput = bench.total_audio_duration_s / wall_time
+    bench.request_throughput = len(successful) / wall_time
+
+    bench.per_request = [
+        {
+            "ttfp_ms": r.ttfp * 1000,
+            "e2e_ms": r.e2e * 1000,
+            "rtf": r.rtf,
+            "audio_duration_s": r.audio_duration,
+            "prompt": r.prompt,
+        }
+        for r in successful
+    ]
+
+    return bench
+
+
+# ---------------------------------------------------------------------------
+# Output formatting
+# ---------------------------------------------------------------------------
+def print_benchmark_results(bench: BenchmarkResult) -> None:
+    """Print benchmark results in standardized format."""
+    W = 50
+    print("")
+    print(f"{'=' * W}")
+    print(f"{'Serving Benchmark Result':^{W}}")
+    print(f"{'=' * W}")
+    print(f"{'Successful requests:':<40}{bench.completed:<10}")
+    print(f"{'Failed requests:':<40}{bench.failed:<10}")
+    print(f"{'Maximum request concurrency:':<40}{bench.concurrency:<10}")
+    print(f"{'Benchmark duration (s):':<40}{bench.duration_s:<10.2f}")
+    print(f"{'Request throughput (req/s):':<40}{bench.request_throughput:<10.2f}")
+    print(f"{'-' * W}")
+    print(f"{'End-to-end Latency':^{W}}")
+    print(f"{'-' * W}")
+    print(f"{'Mean E2EL (ms):':<40}{bench.mean_e2e_ms:<10.2f}")
+    print(f"{'Median E2EL (ms):':<40}{bench.median_e2e_ms:<10.2f}")
+    print(f"{'P99 E2EL (ms):':<40}{bench.p99_e2e_ms:<10.2f}")
+    print(f"{'=' * W}")
+    print(f"{'Audio Result':^{W}}")
+    print(f"{'=' * W}")
+    print(f"{'Total audio duration generated (s):':<40}{bench.total_audio_duration_s:<10.2f}")
+    print(f"{'Audio throughput (audio duration/s):':<40}{bench.audio_throughput:<10.2f}")
+    print(f"{'-' * W}")
+    print(f"{'Time to First Packet':^{W}}")
+    print(f"{'-' * W}")
+    print(f"{'Mean AUDIO_TTFP (ms):':<40}{bench.mean_ttfp_ms:<10.2f}")
+    print(f"{'Median AUDIO_TTFP (ms):':<40}{bench.median_ttfp_ms:<10.2f}")
+    print(f"{'P99 AUDIO_TTFP (ms):':<40}{bench.p99_ttfp_ms:<10.2f}")
+    print(f"{'-' * W}")
+    print(f"{'Real Time Factor':^{W}}")
+    print(f"{'-' * W}")
+    print(f"{'Mean AUDIO_RTF:':<40}{bench.mean_rtf:<10.3f}")
+    print(f"{'Median AUDIO_RTF:':<40}{bench.median_rtf:<10.3f}")
+    print(f"{'P99 AUDIO_RTF:':<40}{bench.p99_rtf:<10.3f}")
+    print(f"{'=' * W}")
+    print("")
+
+
+def save_results(
+    all_results: list[dict],
+    result_dir: str,
+    config_name: str,
+) -> Path:
+    """Save benchmark results as JSON and return the file path."""
+    out = Path(result_dir)
+    out.mkdir(parents=True, exist_ok=True)
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    result_file = out / f"bench_{config_name}_{timestamp}.json"
+
+    with open(result_file, "w") as f:
+        json.dump(all_results, f, indent=2)
+    print(f"Results saved to {result_file}")
+    return result_file
+
+
+# ---------------------------------------------------------------------------
+# Streaming HTTP client
+# ---------------------------------------------------------------------------
+async def send_streaming_request(
+    session: aiohttp.ClientSession,
+    api_url: str,
+    payload: dict,
+    sample_rate: int,
+    sample_width: int,
+    pbar: tqdm | None = None,
+) -> RequestResult:
+    """Send a streaming TTS request and measure latency metrics."""
+    result = RequestResult(prompt=payload.get("input", ""))
+    st = time.perf_counter()
+
+    try:
+        async with session.post(api_url, json=payload) as response:
+            if response.status != 200:
+                result.error = f"HTTP {response.status}: {await response.text()}"
+            else:
+                if _is_sse_response(response):
+                    total_bytes, result.ttfp = await _read_sse_audio_stream(
+                        response,
+                        start_time=st,
+                    )
+                else:
+                    total_bytes, result.ttfp = await _read_raw_audio_stream(
+                        response,
+                        start_time=st,
+                    )
+
+                result.e2e = time.perf_counter() - st
+                result.audio_bytes = total_bytes
+                result.audio_duration = pcm_bytes_to_duration(total_bytes, sample_rate, sample_width)
+
+                if total_bytes <= 0 or result.ttfp <= 0:
+                    result.error = "HTTP 200 but no audio bytes were received"
+                else:
+                    if result.audio_duration > 0:
+                        result.rtf = result.e2e / result.audio_duration
+                    result.success = True
+
+    except Exception as e:
+        result.error = str(e)
+        result.e2e = time.perf_counter() - st
+
+    finally:
+        if pbar:
+            pbar.update(1)
+    return result
+
+
+# ---------------------------------------------------------------------------
+# Benchmark runner
+# ---------------------------------------------------------------------------
+async def run_benchmark(
+    host: str,
+    port: int,
+    num_prompts: int,
+    max_concurrency: int,
+    create_payload_fn: Callable[[str], dict],
+    sample_rate: int,
+    sample_width: int = 2,
+    num_warmups: int = 3,
+    request_timeout_s: float = 120.0,
+) -> BenchmarkResult:
+    """Run a TTS streaming benchmark at a given concurrency level.
+
+    Args:
+        create_payload_fn: Model-specific function that takes a prompt string
+            and returns the request JSON payload dict.
+        sample_rate: PCM sample rate for audio duration calculation.
+        sample_width: PCM sample width in bytes (default 2 for 16-bit).
+    """
+    api_url = f"http://{host}:{port}/v1/audio/speech"
+
+    connector = aiohttp.TCPConnector(
+        limit=max_concurrency,
+        limit_per_host=max_concurrency,
+        keepalive_timeout=60,
+    )
+    session = aiohttp.ClientSession(
+        connector=connector,
+        timeout=aiohttp.ClientTimeout(
+            total=request_timeout_s,
+            connect=min(10.0, request_timeout_s),
+            sock_connect=min(10.0, request_timeout_s),
+            sock_read=request_timeout_s,
+        ),
+    )
+
+    try:
+        # Warmup
+        if num_warmups > 0:
+            print(f"  Warming up with {num_warmups} requests...")
+            warmup_tasks = [
+                send_streaming_request(
+                    session,
+                    api_url,
+                    create_payload_fn(PROMPTS[i % len(PROMPTS)]),
+                    sample_rate,
+                    sample_width,
+                )
+                for i in range(num_warmups)
+            ]
+            warmup_results = await asyncio.gather(*warmup_tasks)
+            warmup_ok = sum(1 for r in warmup_results if r.success)
+            if warmup_ok == 0:
+                print("  WARNING: All warmup requests failed!")
+                for r in warmup_results:
+                    if r.error:
+                        print(f"    {r.error[:200]}")
+            print(f"  Warmup done ({warmup_ok}/{num_warmups} succeeded).")
+
+        # Build request list
+        request_prompts = [PROMPTS[i % len(PROMPTS)] for i in range(num_prompts)]
+
+        # Run
+        print(f"  Running {num_prompts} requests with concurrency={max_concurrency}...")
+        semaphore = asyncio.Semaphore(max_concurrency)
+        pbar = tqdm(total=num_prompts, desc=f"  concurrency={max_concurrency}")
+
+        async def limited_request(prompt: str) -> RequestResult:
+            async with semaphore:
+                return await send_streaming_request(
+                    session,
+                    api_url,
+                    create_payload_fn(prompt),
+                    sample_rate,
+                    sample_width,
+                    pbar,
+                )
+
+        start_time = time.perf_counter()
+        tasks = [asyncio.create_task(limited_request(p)) for p in request_prompts]
+        results: list[RequestResult] = await asyncio.gather(*tasks)
+        wall_time = time.perf_counter() - start_time
+        pbar.close()
+
+    finally:
+        await session.close()
+
+    # Compute stats
+    bench = compute_stats(results, wall_time)
+    bench.concurrency = max_concurrency
+    bench.num_prompts = num_prompts
+
+    print_benchmark_results(bench)
+
+    # Print sample errors
+    failed = [r for r in results if not r.success]
+    if failed:
+        for r in failed[:3]:
+            print(f"  [ERROR] {r.error[:200]}")
+
+    return bench
+
+
+async def run_benchmark_sweep(
+    host: str,
+    port: int,
+    num_prompts: int,
+    concurrency_levels: list[int],
+    create_payload_fn: Callable[[str], dict],
+    sample_rate: int,
+    sample_width: int = 2,
+    num_warmups: int = 3,
+    request_timeout_s: float = 120.0,
+    config_name: str = "benchmark",
+    result_dir: str = "results",
+) -> list[dict]:
+    """Run benchmarks across multiple concurrency levels and save results."""
+    all_results = []
+
+    for concurrency in concurrency_levels:
+        result = await run_benchmark(
+            host=host,
+            port=port,
+            num_prompts=num_prompts,
+            max_concurrency=concurrency,
+            create_payload_fn=create_payload_fn,
+            sample_rate=sample_rate,
+            sample_width=sample_width,
+            num_warmups=num_warmups,
+            request_timeout_s=request_timeout_s,
+        )
+        result.config_name = config_name
+        all_results.append(asdict(result))
+
+    save_results(all_results, result_dir, config_name)
+    return all_results

benchmarks/glm_image/README.md

@@ -0,0 +1,157 @@
+# GLM-Image Benchmarks
+
+Benchmark GLM-Image T2I (text-to-image) and I2I (image-to-image) performance across three backends: HuggingFace baseline, vLLM-Omni offline, and vLLM-Omni online serving.
+
+## Benchmarks
+
+| Benchmark | Script | Description |
+|-----------|--------|-------------|
+| HuggingFace Baseline | `huggingface/inference.py` | Single-GPU transformers + diffusers pipeline |
+| vLLM-Omni Offline | `vllm-omni/inference.py` | Offline inference with continuous batching |
+| vLLM-Omni Online | `benchmark_glm_image.py` | Online serving via `/v1/chat/completions` |
+
+## HuggingFace Baseline
+
+Single-request sequential inference using the reference HuggingFace pipeline.
+
+```bash
+# T2I
+CUDA_VISIBLE_DEVICES=0 python benchmarks/glm_image/huggingface/inference.py \
+    --model-path /path/to/GLM-Image --mode t2i --num-prompts 10
+
+# I2I
+CUDA_VISIBLE_DEVICES=0 python benchmarks/glm_image/huggingface/inference.py \
+    --model-path /path/to/GLM-Image --mode i2i --num-prompts 10
+```
+
+### Options
+
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--model-path` | `zai-org/GLM-Image` | Model path |
+| `--mode` | `t2i` | `t2i` or `i2i` |
+| `--dataset-path` | `prompt/prompt.json` | Path to prompt.json |
+| `--num-prompts` | `10` | Number of images to generate |
+| `--width` / `--height` | `1024` | Output image size |
+| `--num-inference-steps` | `50` | Diffusion denoising steps |
+| `--output-dir` | `benchmarks/glm_image/huggingface/outputs` | Output directory |
+| `--output-file` | - | JSON file for metrics |
+
+## vLLM-Omni Offline
+
+Multi-GPU offline inference with pipeline parallelism and continuous batching.
+
+```bash
+# T2I
+CUDA_VISIBLE_DEVICES=0,1 python benchmarks/glm_image/vllm-omni/inference.py \
+    --model-path /path/to/GLM-Image --mode t2i --num-prompts 10
+
+# I2I
+CUDA_VISIBLE_DEVICES=0,1 python benchmarks/glm_image/vllm-omni/inference.py \
+    --model-path /path/to/GLM-Image --mode i2i --num-prompts 10
+```
+
+### Options
+
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--model-path` | `zai-org/GLM-Image` | Model path |
+| `--deploy-config` | - | Deploy config YAML |
+| `--mode` | `t2i` | `t2i` or `i2i` |
+| `--dataset-path` | `prompt/prompt.json` | Path to prompt.json |
+| `--num-prompts` | `10` | Number of images to generate |
+| `--width` / `--height` | `1024` | Output image size |
+| `--num-inference-steps` | `50` | Diffusion denoising steps |
+| `--output-dir` | `benchmarks/glm_image/vllm-omni/outputs` | Output directory |
+| `--output-file` | - | JSON file for metrics |
+| `--stage-init-timeout` | `600` | Stage initialization timeout (s) |
+
+### Latency Computation
+
+In offline mode all requests are submitted simultaneously and processed with continuous batching. The per-request latency is computed by summing the actual per-stage times (with `stage_0_gen_ms` diffed against the previous request to remove accumulated queue/scheduling wait).
+
+## vLLM-Omni Online Serving
+
+### Start the server
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 vllm serve /path/to/GLM-Image \
+    --omni --port 8091 --host 0.0.0.0 \
+    --served-model-name glm-image
+```
+
+### Run the benchmark
+
+```bash
+# T2I
+python benchmarks/glm_image/benchmark_glm_image.py \
+    --mode t2i --num-prompts 10 --model glm-image
+
+# I2I
+python benchmarks/glm_image/benchmark_glm_image.py \
+    --mode i2i --num-prompts 10 --model glm-image
+
+# Custom dataset
+python benchmarks/glm_image/benchmark_glm_image.py \
+    --mode i2i --dataset custom \
+    --dataset-path prompts.json --num-prompts 5
+```
+
+### Options
+
+| Flag | Default | Description |
+|------|---------|-------------|
+| `--mode` | `t2i` | `t2i` or `i2i` |
+| `--dataset` | `prompt` | `prompt`, `random`, or `custom` |
+| `--dataset-path` | - | JSON file path (required for `custom`) |
+| `--num-prompts` | `10` | Number of benchmark requests |
+| `--max-concurrency` | `1` | Max concurrent requests |
+| `--request-rate` | `inf` | Requests per second (Poisson arrival) |
+| `--warmup-requests` | `1` | Warmup requests before measurement |
+| `--width` / `--height` | `1024` | Output image size |
+| `--num-inference-steps` | `50` | Diffusion denoising steps |
+| `--seed` | - | Random seed |
+| `--model` | `default` | Model name (must match `--served-model-name`) |
+| `--host` | `localhost` | Server host |
+| `--port` | `8091` | Server port |
+| `--output-file` | - | JSON output file for metrics |
+| `--num-input-images` | `1` | Number of input images for random I2I |
+
+## Dataset
+
+The default dataset is hosted on [HuggingFace](https://huggingface.co/datasets/JaredforReal/glm-image-bench) (`prompt.json`). It is automatically downloaded and cached to `prompt/prompt.json` on first run. No manual setup needed.
+
+Each entry contains:
+
+- `t2i_prompt`: Text prompt for text-to-image generation
+- `i2i_prompt`: Text prompt for image-to-image editing
+- `image_url`: Source image URL for I2I (downloaded and cached on first use)
+
+Custom datasets use the same JSON format and can be provided via `--dataset-path`.
+
+## Pipeline Timings
+
+All three benchmarks report per-stage pipeline timings (in milliseconds):
+
+| Key | Description |
+|-----|-------------|
+| `preprocess_ms` | Input preprocessing (tokenization, multimodal encoding) |
+| `stage_0_gen_ms` | AR (autoregressive) model generation time |
+| `ar2diffusion_ms` | AR output to diffusion input conversion |
+| `stage_1_gen_ms` | Diffusion model denoising time |
+| `queue_wait_ms` | Queue wait time before processing |
+
+The stages are ordered by execution: `preprocess → stage_0 (AR) → ar2diffusion → stage_1 (Diffusion)`.
+
+## Sample Results
+
+Tested on 2x GPU with 10 prompts, 1024x1024, 50 denoising steps:
+
+| Backend | Mode | Latency Mean (s) | Throughput (img/s) |
+|---------|------|-------------------|--------------------|
+| HuggingFace | T2I | 72.6 | 0.014 |
+| HuggingFace | I2I | 70.9 | 0.014 |
+| vLLM-Omni Offline | T2I | 35.0 | 0.044 |
+| vLLM-Omni Offline | I2I | 31.0 | 0.053 |
+| vLLM-Omni Online | T2I | 38.8 | 0.026 |
+| vLLM-Omni Online | I2I | 34.7 | 0.029 |

benchmarks/glm_image/benchmark_glm_image.py

@@ -0,0 +1,464 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Online serving benchmark for GLM-Image (T2I and I2I modes).
+
+Sends requests to the /v1/chat/completions endpoint and reports end-to-end
+latency, throughput, and per-stage durations (when the server is started with
+--enable-diffusion-pipeline-profiler and/or --enable-ar-profiler).
+
+Supports three dataset types:
+  - prompt:   Use prompt.json (default). T2I uses t2i_prompt, I2I uses i2i_prompt
+              and sends source images from image_url.
+  - random:   Generate synthetic prompts (and random images for I2I).
+  - custom:   Load from a user-specified JSON file.
+
+Usage:
+    # T2I with prompt.json (default)
+    python benchmarks/glm_image/benchmark_glm_image.py \
+        --mode t2i --num-prompts 10
+
+    # I2I with prompt.json (downloads source images automatically)
+    python benchmarks/glm_image/benchmark_glm_image.py \
+        --mode i2i --num-prompts 10
+
+    # Random dataset
+    python benchmarks/glm_image/benchmark_glm_image.py \
+        --mode t2i --dataset random --num-prompts 20
+
+    # Custom dataset
+    python benchmarks/glm_image/benchmark_glm_image.py \
+        --mode i2i --dataset custom \
+        --dataset-path my_prompts.json --num-prompts 5
+"""
+
+import argparse
+import asyncio
+import base64
+import json
+import os
+import sys
+import tempfile
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+import aiohttp
+import numpy as np
+import requests as sync_requests
+from PIL import Image
+from tqdm.asyncio import tqdm
+
+# Import backends from the diffusion benchmark (add parent dirs to path)
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "diffusion"))
+from backends import RequestFuncOutput
+
+BENCHMARK_DIR = Path(__file__).resolve().parent
+DEFAULT_PROMPT_JSON = BENCHMARK_DIR / "prompt" / "prompt.json"
+IMAGE_CACHE_DIR = BENCHMARK_DIR / "prompt" / "images"
+
+DATASET_REPO = "JaredforReal/glm-image-bench"
+DATASET_FILE = "prompt.json"
+
+
+def _ensure_prompt_json(dataset_path: str | None) -> str:
+    """Return path to prompt.json, downloading from HuggingFace if needed."""
+    if dataset_path:
+        return dataset_path
+    local = DEFAULT_PROMPT_JSON
+    if local.exists():
+        return str(local)
+    print(f"Downloading {DATASET_FILE} from {DATASET_REPO} ...")
+    try:
+        from huggingface_hub import hf_hub_download
+
+        downloaded = hf_hub_download(
+            repo_id=DATASET_REPO,
+            filename=DATASET_FILE,
+            repo_type="dataset",
+        )
+        local.parent.mkdir(parents=True, exist_ok=True)
+        import shutil
+
+        shutil.copy2(downloaded, local)
+        print(f"Saved to {local}")
+    except ImportError:
+        url = f"https://huggingface.co/datasets/{DATASET_REPO}/resolve/main/{DATASET_FILE}"
+        import urllib.request
+
+        local.parent.mkdir(parents=True, exist_ok=True)
+        urllib.request.urlretrieve(url, local)
+        print(f"Saved to {local}")
+    return str(local)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+@dataclass
+class GLMImageRequest:
+    prompt: str
+    image_path: str | None = None  # Only for I2I mode
+
+
+def download_image(url: str) -> str:
+    """Download an image to cache and return the local path."""
+    IMAGE_CACHE_DIR.mkdir(parents=True, exist_ok=True)
+    fname = url.rsplit("/", 1)[-1]
+    local_path = IMAGE_CACHE_DIR / fname
+    if local_path.exists():
+        return str(local_path)
+    resp = sync_requests.get(url, timeout=30)
+    resp.raise_for_status()
+    local_path.write_bytes(resp.content)
+    return str(local_path)
+
+
+def encode_image_as_data_url(path: str) -> str:
+    """Encode a local image file as a base64 data URL."""
+    with open(path, "rb") as f:
+        encoded = base64.b64encode(f.read()).decode("utf-8")
+    ext = Path(path).suffix.lower()
+    mime = {"png": "image/png", ".jpg": "image/jpeg", ".jpeg": "image/jpeg"}.get(ext, "image/png")
+    return f"data:{mime};base64,{encoded}"
+
+
+# ---------------------------------------------------------------------------
+# Datasets
+# ---------------------------------------------------------------------------
+
+
+class PromptDataset:
+    """Load from prompt.json. T2I uses t2i_prompt, I2I uses i2i_prompt + image_url."""
+
+    def __init__(self, args: argparse.Namespace):
+        path = _ensure_prompt_json(args.dataset_path)
+        with open(path, encoding="utf-8") as f:
+            raw = json.load(f)
+
+        prompt_key = "t2i_prompt" if args.mode == "t2i" else "i2i_prompt"
+        self.items: list[GLMImageRequest] = []
+
+        for entry in raw:
+            prompt = entry.get(prompt_key, "").strip()
+            if not prompt:
+                continue
+            image_path = None
+            if args.mode == "i2i":
+                url = entry.get("image_url", "")
+                if url:
+                    image_path = download_image(url)
+            self.items.append(GLMImageRequest(prompt=prompt, image_path=image_path))
+
+        if args.num_prompts and len(self.items) > args.num_prompts:
+            self.items = self.items[: args.num_prompts]
+
+    def __len__(self) -> int:
+        return len(self.items)
+
+    def __getitem__(self, idx: int) -> GLMImageRequest:
+        return self.items[idx]
+
+    def get_requests(self) -> list[GLMImageRequest]:
+        return list(self.items)
+
+
+class RandomDataset:
+    """Generate synthetic prompts (and optional random images for I2I)."""
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+        self.num_prompts = args.num_prompts
+        self._random_image_paths: list[str] | None = None
+        if args.mode == "i2i":
+            self._random_image_paths = self._generate_random_images()
+
+    def _generate_random_images(self) -> list[str]:
+        paths: list[str] = []
+        for i in range(self.args.num_input_images):
+            img = Image.new("RGB", (512, 512), (128 + i * 30 % 128, 64, 192))
+            path = os.path.join(tempfile.gettempdir(), f"glm_image_bench_input_{i}.png")
+            img.save(path)
+            paths.append(path)
+        return paths
+
+    def __len__(self) -> int:
+        return self.num_prompts
+
+    def __getitem__(self, idx: int) -> GLMImageRequest:
+        image_path = None
+        if self._random_image_paths is not None:
+            image_path = self._random_image_paths[idx % len(self._random_image_paths)]
+        return GLMImageRequest(
+            prompt=f"A beautiful scene with vivid colors and intricate details, prompt {idx}",
+            image_path=image_path,
+        )
+
+    def get_requests(self) -> list[GLMImageRequest]:
+        return [self[i] for i in range(len(self))]
+
+
+class CustomDataset:
+    """Load from a user-specified JSON file.
+
+    Expected format:
+    [
+        {"prompt": "A cat sitting on a windowsill"},
+        {"prompt": "Make it look like winter", "image_path": "/path/to/img.png"}
+    ]
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        if not args.dataset_path:
+            raise ValueError("--dataset-path is required for custom dataset")
+        with open(args.dataset_path, encoding="utf-8") as f:
+            raw = json.load(f)
+        self.items: list[GLMImageRequest] = []
+        for item in raw:
+            self.items.append(
+                GLMImageRequest(
+                    prompt=item.get("prompt", ""),
+                    image_path=item.get("image_path"),
+                )
+            )
+        if args.num_prompts and len(self.items) > args.num_prompts:
+            self.items = self.items[: args.num_prompts]
+
+    def __len__(self) -> int:
+        return len(self.items)
+
+    def __getitem__(self, idx: int) -> GLMImageRequest:
+        return self.items[idx]
+
+    def get_requests(self) -> list[GLMImageRequest]:
+        return list(self.items)
+
+
+# ---------------------------------------------------------------------------
+# Async request for GLM-Image (chat completions with image support)
+# ---------------------------------------------------------------------------
+
+
+async def async_glm_image_request(
+    req: GLMImageRequest,
+    api_url: str,
+    model: str,
+    session: aiohttp.ClientSession,
+    pbar: Any,
+    args: argparse.Namespace,
+) -> RequestFuncOutput:
+    """Send a single T2I or I2I request via chat completions endpoint."""
+    output = RequestFuncOutput()
+    output.start_time = time.perf_counter()
+
+    # Build messages
+    if req.image_path and args.mode == "i2i":
+        data_url = encode_image_as_data_url(req.image_path)
+        content = [
+            {"type": "text", "text": req.prompt},
+            {"type": "image_url", "image_url": {"url": data_url}},
+        ]
+    else:
+        content = req.prompt
+
+    messages = [{"role": "user", "content": content}]
+
+    extra_body: dict[str, Any] = {}
+    if args.height:
+        extra_body["height"] = args.height
+    if args.width:
+        extra_body["width"] = args.width
+    if args.num_inference_steps:
+        extra_body["num_inference_steps"] = args.num_inference_steps
+    if args.seed is not None:
+        extra_body["seed"] = args.seed
+
+    payload: dict[str, Any] = {
+        "model": model,
+        "messages": messages,
+    }
+    if extra_body:
+        payload["extra_body"] = extra_body
+
+    try:
+        async with session.post(api_url, json=payload) as response:
+            if response.status == 200:
+                resp_json = await response.json()
+                output.response_body = resp_json
+                output.success = True
+                try:
+                    choices = resp_json.get("choices", [])
+                    if choices and isinstance(choices, list):
+                        msg = choices[0].get("message", {})
+                        if isinstance(msg, dict):
+                            resp_content = msg.get("content", [])
+                            if resp_content and isinstance(resp_content, list) and len(resp_content) > 0:
+                                first_item = resp_content[0]
+                                if isinstance(first_item, dict):
+                                    output.stage_durations = first_item.get("stage_durations") or {}
+                                    output.peak_memory_mb = first_item.get("peak_memory_mb", 0.0)
+                except (IndexError, TypeError, AttributeError):
+                    pass
+            else:
+                output.error = f"HTTP {response.status}: {await response.text()}"
+                output.success = False
+    except Exception as e:
+        output.error = str(e)
+        output.success = False
+
+    output.latency = time.perf_counter() - output.start_time
+    if pbar:
+        pbar.update(1)
+    return output
+
+
+# ---------------------------------------------------------------------------
+# Benchmark
+# ---------------------------------------------------------------------------
+
+
+async def iter_requests(n: int, request_rate: float) -> Any:
+    import random as _random
+
+    for i in range(n):
+        if request_rate != float("inf") and i > 0:
+            await asyncio.sleep(_random.expovariate(request_rate))
+        yield i
+
+
+def calculate_metrics(outputs: list[RequestFuncOutput], total_duration: float) -> dict[str, Any]:
+    success = [o for o in outputs if o.success]
+    errors = [o for o in outputs if not o.success]
+    latencies = [o.latency for o in success]
+    peak_mems = [o.peak_memory_mb for o in success if o.peak_memory_mb > 0]
+
+    stage_duration_lists: dict[str, list[float]] = {}
+    for o in success:
+        for stage, dur in (o.stage_durations or {}).items():
+            stage_duration_lists.setdefault(stage, []).append(dur)
+
+    return {
+        "duration": total_duration,
+        "completed_requests": len(success),
+        "failed_requests": len(errors),
+        "throughput_qps": len(success) / total_duration if total_duration > 0 else 0,
+        "latency_mean": float(np.mean(latencies)) if latencies else 0,
+        "latency_median": float(np.median(latencies)) if latencies else 0,
+        "latency_p99": float(np.percentile(latencies, 99)) if latencies else 0,
+        "latency_p95": float(np.percentile(latencies, 95)) if latencies else 0,
+        "peak_memory_mb_max": max(peak_mems) if peak_mems else 0,
+        "stage_durations_mean": {s: float(np.mean(v)) for s, v in stage_duration_lists.items()},
+        "stage_durations_p50": {s: float(np.percentile(v, 50)) for s, v in stage_duration_lists.items()},
+    }
+
+
+async def benchmark(args: argparse.Namespace) -> None:
+    api_url = f"http://{args.host}:{args.port}/v1/chat/completions"
+
+    # Load dataset
+    if args.dataset == "prompt":
+        dataset = PromptDataset(args)
+    elif args.dataset == "random":
+        dataset = RandomDataset(args)
+    elif args.dataset == "custom":
+        dataset = CustomDataset(args)
+    else:
+        raise ValueError(f"Unknown dataset: {args.dataset}")
+
+    glm_requests = dataset.get_requests()
+    print(f"Prepared {len(glm_requests)} requests (mode={args.mode}, dataset={args.dataset})")
+
+    semaphore = asyncio.Semaphore(args.max_concurrency) if args.max_concurrency else None
+
+    async def limited_request(idx: int, req: GLMImageRequest, session: aiohttp.ClientSession, pbar: Any):
+        if semaphore:
+            async with semaphore:
+                return await async_glm_image_request(req, api_url, args.model, session, pbar, args)
+        return await async_glm_image_request(req, api_url, args.model, session, pbar, args)
+
+    async with aiohttp.ClientSession() as session:
+        # Warmup
+        if args.warmup_requests and glm_requests:
+            print(f"Running {args.warmup_requests} warmup request(s)...")
+            for i in range(args.warmup_requests):
+                await limited_request(i, glm_requests[i % len(glm_requests)], session, None)
+
+        # Main benchmark
+        pbar = tqdm(total=len(glm_requests), disable=args.disable_tqdm)
+        start_time = time.perf_counter()
+        tasks = []
+        async for idx in iter_requests(len(glm_requests), args.request_rate):
+            tasks.append(asyncio.create_task(limited_request(idx, glm_requests[idx], session, pbar)))
+        outputs = await asyncio.gather(*tasks)
+        total_duration = time.perf_counter() - start_time
+        pbar.close()
+
+    # Metrics
+    metrics = calculate_metrics(outputs, total_duration)
+    metrics["mode"] = args.mode
+    metrics["model"] = args.model
+    metrics["dataset"] = args.dataset
+
+    print(f"\n{' GLM-Image Online Benchmark Result ':=^60}")
+    print(f"{'Mode:':<40} {args.mode}")
+    print(f"{'Model:':<40} {args.model}")
+    print(f"{'Dataset:':<40} {args.dataset}")
+    print("-" * 50)
+    print(f"{'Benchmark duration (s):':<40} {metrics['duration']:.2f}")
+    print(f"{'Request rate:':<40} {args.request_rate}")
+    print(f"{'Max concurrency:':<40} {args.max_concurrency}")
+    print(f"{'Successful requests:':<40} {metrics['completed_requests']}/{len(glm_requests)}")
+    print("-" * 50)
+    print(f"{'Throughput (req/s):':<40} {metrics['throughput_qps']:.2f}")
+    print(f"{'Latency Mean (s):':<40} {metrics['latency_mean']:.4f}")
+    print(f"{'Latency Median (s):':<40} {metrics['latency_median']:.4f}")
+    print(f"{'Latency P95 (s):':<40} {metrics['latency_p95']:.4f}")
+    print(f"{'Latency P99 (s):':<40} {metrics['latency_p99']:.4f}")
+
+    if metrics["peak_memory_mb_max"] > 0:
+        print("-" * 50)
+        print(f"{'Peak Memory Max (MB):':<40} {metrics['peak_memory_mb_max']:.2f}")
+
+    if metrics["stage_durations_mean"]:
+        print("-" * 50)
+        print("Stage Durations Mean:")
+        for stage, val in sorted(metrics["stage_durations_mean"].items()):
+            unit = "ms" if stage.endswith("_ms") else "s"
+            print(f"  {stage + ':':<38} {val:.4f} ({unit})")
+
+    print("=" * 60)
+
+    if args.output_file:
+        with open(args.output_file, "w") as f:
+            json.dump(metrics, f, indent=2)
+        print(f"Metrics saved to {args.output_file}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Benchmark GLM-Image T2I/I2I online serving.")
+    parser.add_argument("--mode", type=str, default="t2i", choices=["t2i", "i2i"])
+    parser.add_argument("--dataset", type=str, default="prompt", choices=["prompt", "random", "custom"])
+    parser.add_argument("--dataset-path", type=str, default=None)
+    parser.add_argument("--num-prompts", type=int, default=10)
+    parser.add_argument("--max-concurrency", type=int, default=1)
+    parser.add_argument("--request-rate", type=float, default=float("inf"))
+    parser.add_argument("--warmup-requests", type=int, default=1)
+    parser.add_argument("--width", type=int, default=1024)
+    parser.add_argument("--height", type=int, default=1024)
+    parser.add_argument("--num-inference-steps", type=int, default=50)
+    parser.add_argument("--seed", type=int, default=None)
+    parser.add_argument("--model", type=str, default="default")
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8091)
+    parser.add_argument("--output-file", type=str, default=None)
+    parser.add_argument("--disable-tqdm", action="store_true")
+    parser.add_argument("--num-input-images", type=int, default=1, help="For random I2I dataset.")
+    args = parser.parse_args()
+    asyncio.run(benchmark(args))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/glm_image/huggingface/inference.py

@@ -0,0 +1,291 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+HuggingFace (transformers + diffusers) baseline benchmark for GLM-Image.
+
+Supports T2I and I2I modes with the prompt.json dataset.
+Downloads source images for I2I from image_url on first run and caches locally.
+
+Usage:
+    # T2I mode (text-to-image, no source images needed)
+    python benchmarks/glm_image/huggingface/inference.py \
+        --model-path zai-org/GLM-Image \
+        --mode t2i --num-prompts 10
+
+    # I2I mode (image-to-image, downloads source images)
+    python benchmarks/glm_image/huggingface/inference.py \
+        --model-path zai-org/GLM-Image \
+        --mode i2i --num-prompts 10
+
+    # With custom prompt.json
+    python benchmarks/glm_image/huggingface/inference.py \
+        --model-path zai-org/GLM-Image \
+        --mode i2i --dataset-path prompts.json --num-prompts 5
+"""
+
+import argparse
+import json
+import os
+import time
+from pathlib import Path
+
+import numpy as np
+import requests
+import torch
+from PIL import Image
+
+BENCHMARK_DIR = Path(__file__).resolve().parent.parent
+DEFAULT_PROMPT_JSON = BENCHMARK_DIR / "prompt" / "prompt.json"
+IMAGE_CACHE_DIR = BENCHMARK_DIR / "prompt" / "images"
+
+DATASET_REPO = "JaredforReal/glm-image-bench"
+DATASET_FILE = "prompt.json"
+
+
+def _ensure_prompt_json(dataset_path: str | None) -> str:
+    """Return path to prompt.json, downloading from HuggingFace if needed."""
+    if dataset_path:
+        return dataset_path
+    local = DEFAULT_PROMPT_JSON
+    if local.exists():
+        return str(local)
+    print(f"Downloading {DATASET_FILE} from {DATASET_REPO} ...")
+    try:
+        from huggingface_hub import hf_hub_download
+
+        downloaded = hf_hub_download(
+            repo_id=DATASET_REPO,
+            filename=DATASET_FILE,
+            repo_type="dataset",
+        )
+        local.parent.mkdir(parents=True, exist_ok=True)
+        import shutil
+
+        shutil.copy2(downloaded, local)
+        print(f"Saved to {local}")
+    except ImportError:
+        url = f"https://huggingface.co/datasets/{DATASET_REPO}/resolve/main/{DATASET_FILE}"
+        import urllib.request
+
+        local.parent.mkdir(parents=True, exist_ok=True)
+        urllib.request.urlretrieve(url, local)
+        print(f"Saved to {local}")
+    return str(local)
+
+
+HEIGHT = 1024
+WIDTH = 1024
+SEED = 42
+NUM_INFERENCE_STEPS = 50
+GUIDANCE_SCALE = 1.5
+
+
+# ---------------------------------------------------------------------------
+# Dataset
+# ---------------------------------------------------------------------------
+
+
+def load_dataset(
+    dataset_path: str | None,
+    mode: str,
+    num_prompts: int,
+) -> list[dict]:
+    """Load prompts from prompt.json and prepare per-request data."""
+    path = _ensure_prompt_json(dataset_path)
+    with open(path, encoding="utf-8") as f:
+        raw = json.load(f)
+
+    items = []
+    for entry in raw:
+        if mode == "t2i":
+            prompt_key = "t2i_prompt"
+        else:
+            prompt_key = "i2i_prompt"
+
+        prompt_text = entry.get(prompt_key, "").strip()
+        if not prompt_text:
+            continue
+
+        item = {"prompt": prompt_text}
+        if mode == "i2i":
+            item["image_url"] = entry.get("image_url", "")
+        items.append(item)
+
+    if num_prompts and len(items) > num_prompts:
+        items = items[:num_prompts]
+    return items
+
+
+def download_image(url: str, cache_dir: Path) -> str:
+    """Download an image to cache_dir and return the local path."""
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    fname = url.rsplit("/", 1)[-1]
+    local_path = cache_dir / fname
+    if local_path.exists():
+        return str(local_path)
+    print(f"  Downloading {url} ...")
+    resp = requests.get(url, timeout=30)
+    resp.raise_for_status()
+    local_path.write_bytes(resp.content)
+    return str(local_path)
+
+
+# ---------------------------------------------------------------------------
+# Benchmark
+# ---------------------------------------------------------------------------
+
+
+def benchmark(args: argparse.Namespace) -> None:
+    from diffusers.pipelines.glm_image import GlmImagePipeline
+
+    print("=" * 60)
+    print("GLM-Image HuggingFace Baseline Benchmark")
+    print(f"Mode: {args.mode}  |  Model: {args.model_path}")
+    print(f"Size: {args.height}x{args.width}  |  Steps: {args.num_inference_steps}")
+    print("=" * 60)
+
+    # Load dataset
+    items = load_dataset(args.dataset_path, args.mode, args.num_prompts)
+    if not items:
+        print("No prompts loaded. Exiting.")
+        return
+    print(f"Loaded {len(items)} prompts for {args.mode} mode")
+
+    # Download I2I source images
+    if args.mode == "i2i":
+        print("Preparing source images...")
+        for item in items:
+            url = item.get("image_url", "")
+            if url:
+                item["image_path"] = download_image(url, IMAGE_CACHE_DIR)
+            else:
+                item["image_path"] = None
+
+    # Load pipeline
+    print(f"\nLoading pipeline from {args.model_path} ...")
+    t0 = time.perf_counter()
+    pipe = GlmImagePipeline.from_pretrained(
+        args.model_path,
+        torch_dtype=torch.bfloat16,
+        device_map="cuda",
+    )
+    init_time = time.perf_counter() - t0
+    print(f"Pipeline loaded in {init_time:.2f}s")
+
+    # Create output dir
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    # Run benchmark
+    generator = torch.Generator(device="cuda").manual_seed(args.seed)
+    latencies = []
+    success = 0
+    failed = 0
+
+    print(f"\nRunning {len(items)} requests sequentially...")
+    print("-" * 60)
+
+    for i, item in enumerate(items):
+        prompt = item["prompt"]
+        gen_kwargs: dict = {
+            "prompt": prompt,
+            "height": args.height,
+            "width": args.width,
+            "num_inference_steps": args.num_inference_steps,
+            "guidance_scale": args.guidance_scale,
+            "generator": generator,
+        }
+
+        if args.mode == "i2i":
+            img_path = item.get("image_path")
+            if img_path and os.path.exists(img_path):
+                gen_kwargs["image"] = [Image.open(img_path).convert("RGB")]
+            else:
+                print(f"  [{i + 1}] SKIP: no source image")
+                failed += 1
+                continue
+
+        t_start = time.perf_counter()
+        try:
+            result = pipe(**gen_kwargs)
+            image = result.images[0]
+            elapsed = time.perf_counter() - t_start
+            latencies.append(elapsed)
+            success += 1
+
+            out_path = os.path.join(args.output_dir, f"{i:04d}.png")
+            image.save(out_path)
+            print(f"  [{i + 1}/{len(items)}] {elapsed:.3f}s -> {out_path}")
+        except Exception as e:
+            elapsed = time.perf_counter() - t_start
+            failed += 1
+            print(f"  [{i + 1}/{len(items)}] FAILED ({elapsed:.3f}s): {e}")
+
+    # Report
+    total_gen_time = sum(latencies) if latencies else 0
+    print("\n" + "=" * 60)
+    print("HuggingFace Baseline Results")
+    print("=" * 60)
+    print(f"{'Mode:':<40} {args.mode}")
+    print(f"{'Model:':<40} {args.model_path}")
+    print(f"{'Image size:':<40} {args.height}x{args.width}")
+    print(f"{'Num inference steps:':<40} {args.num_inference_steps}")
+    print("-" * 50)
+    print(f"{'Pipeline init time (s):':<40} {init_time:.2f}")
+    print(f"{'Successful:':<40} {success}/{len(items)}")
+    print(f"{'Failed:':<40} {failed}")
+    print("-" * 50)
+    if latencies:
+        arr = np.array(latencies)
+        print(f"{'Total generation time (s):':<40} {total_gen_time:.2f}")
+        print(f"{'Throughput (img/s):':<40} {success / total_gen_time:.4f}")
+        print(f"{'Latency Mean (s):':<40} {arr.mean():.4f}")
+        print(f"{'Latency Median (s):':<40} {np.median(arr):.4f}")
+        print(f"{'Latency P95 (s):':<40} {np.percentile(arr, 95):.4f}")
+        print(f"{'Latency P99 (s):':<40} {np.percentile(arr, 99):.4f}")
+
+    print(f"\n{'Output dir:':<40} {args.output_dir}")
+    print("=" * 60)
+
+    # Save metrics JSON
+    metrics = {
+        "backend": "huggingface",
+        "mode": args.mode,
+        "model": args.model_path,
+        "height": args.height,
+        "width": args.width,
+        "num_inference_steps": args.num_inference_steps,
+        "init_time_s": init_time,
+        "completed_requests": success,
+        "failed_requests": failed,
+        "total_gen_time_s": total_gen_time,
+        "throughput_qps": success / total_gen_time if total_gen_time > 0 else 0,
+        "latency_mean": float(np.mean(latencies)) if latencies else 0,
+        "latency_median": float(np.median(latencies)) if latencies else 0,
+        "latency_p95": float(np.percentile(latencies, 95)) if latencies else 0,
+        "latency_p99": float(np.percentile(latencies, 99)) if latencies else 0,
+    }
+    if args.output_file:
+        with open(args.output_file, "w") as f:
+            json.dump(metrics, f, indent=2)
+        print(f"Metrics saved to {args.output_file}")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="GLM-Image HuggingFace baseline benchmark")
+    parser.add_argument("--model-path", type=str, default="zai-org/GLM-Image")
+    parser.add_argument("--mode", type=str, default="t2i", choices=["t2i", "i2i"])
+    parser.add_argument("--dataset-path", type=str, default=None, help="Path to prompt.json")
+    parser.add_argument("--num-prompts", type=int, default=10)
+    parser.add_argument("--height", type=int, default=HEIGHT)
+    parser.add_argument("--width", type=int, default=WIDTH)
+    parser.add_argument("--num-inference-steps", type=int, default=NUM_INFERENCE_STEPS)
+    parser.add_argument("--guidance-scale", type=float, default=GUIDANCE_SCALE)
+    parser.add_argument("--seed", type=int, default=SEED)
+    parser.add_argument("--output-dir", type=str, default="benchmarks/glm_image/huggingface/outputs")
+    parser.add_argument("--output-file", type=str, default=None, help="JSON file for metrics")
+    args = parser.parse_args()
+    benchmark(args)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/glm_image/vllm-omni/inference.py

@@ -0,0 +1,505 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+vLLM-Omni offline benchmark for GLM-Image.
+
+Supports T2I and I2I modes with the prompt.json dataset.
+Downloads source images for I2I from image_url on first run and caches locally.
+
+Usage:
+    # T2I mode
+    python benchmarks/glm_image/vllm-omni/inference.py \
+        --model-path zai-org/GLM-Image \
+        --mode t2i --num-prompts 10
+
+    # I2I mode (downloads source images)
+    python benchmarks/glm_image/vllm-omni/inference.py \
+        --model-path zai-org/GLM-Image \
+        --mode i2i --num-prompts 10
+"""
+
+import argparse
+import json
+import math
+import os
+import time
+from pathlib import Path
+
+import numpy as np
+import requests
+from PIL import Image
+from vllm import SamplingParams
+
+from vllm_omni.entrypoints.omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+BENCHMARK_DIR = Path(__file__).resolve().parent.parent
+DEFAULT_PROMPT_JSON = BENCHMARK_DIR / "prompt" / "prompt.json"
+IMAGE_CACHE_DIR = BENCHMARK_DIR / "prompt" / "images"
+DEFAULT_DEPLOY_CONFIG = "vllm_omni/deploy/glm_image.yaml"
+
+DATASET_REPO = "JaredforReal/glm-image-bench"
+DATASET_FILE = "prompt.json"
+
+
+def _ensure_prompt_json(dataset_path: str | None) -> str:
+    """Return path to prompt.json, downloading from HuggingFace if needed."""
+    if dataset_path:
+        return dataset_path
+    local = DEFAULT_PROMPT_JSON
+    if local.exists():
+        return str(local)
+    print(f"Downloading {DATASET_FILE} from {DATASET_REPO} ...")
+    try:
+        from huggingface_hub import hf_hub_download
+
+        downloaded = hf_hub_download(
+            repo_id=DATASET_REPO,
+            filename=DATASET_FILE,
+            repo_type="dataset",
+        )
+        local.parent.mkdir(parents=True, exist_ok=True)
+        import shutil
+
+        shutil.copy2(downloaded, local)
+        print(f"Saved to {local}")
+    except ImportError:
+        url = f"https://huggingface.co/datasets/{DATASET_REPO}/resolve/main/{DATASET_FILE}"
+        import urllib.request
+
+        local.parent.mkdir(parents=True, exist_ok=True)
+        urllib.request.urlretrieve(url, local)
+        print(f"Saved to {local}")
+    return str(local)
+
+
+SEED = 42
+HEIGHT = 1024
+WIDTH = 1024
+NUM_INFERENCE_STEPS = 50
+GUIDANCE_SCALE = 1.5
+
+GLM_IMAGE_EOS_TOKEN_ID = 16385
+GLM_IMAGE_VISION_VOCAB_SIZE = 16512
+
+
+# ---------------------------------------------------------------------------
+# Dataset
+# ---------------------------------------------------------------------------
+
+
+def load_dataset(
+    dataset_path: str | None,
+    mode: str,
+    num_prompts: int,
+) -> list[dict]:
+    path = _ensure_prompt_json(dataset_path)
+    with open(path, encoding="utf-8") as f:
+        raw = json.load(f)
+
+    items = []
+    for entry in raw:
+        prompt_key = "t2i_prompt" if mode == "t2i" else "i2i_prompt"
+        prompt_text = entry.get(prompt_key, "").strip()
+        if not prompt_text:
+            continue
+
+        item = {"prompt": prompt_text}
+        if mode == "i2i":
+            item["image_url"] = entry.get("image_url", "")
+        items.append(item)
+
+    if num_prompts and len(items) > num_prompts:
+        items = items[:num_prompts]
+    return items
+
+
+def download_image(url: str, cache_dir: Path) -> str:
+    cache_dir.mkdir(parents=True, exist_ok=True)
+    fname = url.rsplit("/", 1)[-1]
+    local_path = cache_dir / fname
+    if local_path.exists():
+        return str(local_path)
+    print(f"  Downloading {url} ...")
+    resp = requests.get(url, timeout=30)
+    resp.raise_for_status()
+    local_path.write_bytes(resp.content)
+    return str(local_path)
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+
+def compute_max_tokens(height: int, width: int, is_i2i: bool = False) -> int:
+    factor = 32
+    token_h = height // factor
+    token_w = width // factor
+    large_tokens = token_h * token_w
+
+    # Small preview tokens (half resolution in each dimension)
+
+    ratio = token_h / token_w if token_w > 0 else 1.0
+    small_token_h = max(1, int(math.sqrt(ratio) * (factor // 2)))
+    small_token_w = max(1, int(math.sqrt(1 / ratio) * (factor // 2)))
+    small_tokens = small_token_h * small_token_w
+
+    # Mode-dependent totals:
+    # - t2i: small + large + EOS
+    # - i2i: large + EOS
+    if is_i2i:
+        return large_tokens + 1
+    return small_tokens + large_tokens + 1
+
+
+def build_prompt_t2i(prompt: str, height: int, width: int, **gen_kw) -> dict:
+    return {
+        "prompt": prompt,
+        "height": height,
+        "width": width,
+        "mm_processor_kwargs": {"target_h": height, "target_w": width},
+        **gen_kw,
+    }
+
+
+def build_prompt_i2i(prompt: str, image_path: str, height: int, width: int, **gen_kw) -> dict:
+    return {
+        "prompt": prompt,
+        "height": height,
+        "width": width,
+        "mm_processor_kwargs": {"target_h": height, "target_w": width},
+        "multi_modal_data": {"image": Image.open(image_path).convert("RGB")},
+        **gen_kw,
+    }
+
+
+def resolve_deploy_config(args: argparse.Namespace) -> str:
+    if args.deploy_config:
+        return args.deploy_config
+    if os.path.exists(DEFAULT_DEPLOY_CONFIG):
+        return DEFAULT_DEPLOY_CONFIG
+    fallback = Path(__file__).resolve().parents[3] / DEFAULT_DEPLOY_CONFIG
+    if fallback.exists():
+        return str(fallback)
+    raise FileNotFoundError("Deploy config not found. Specify --deploy-config.")
+
+
+# ---------------------------------------------------------------------------
+# Benchmark
+# ---------------------------------------------------------------------------
+
+
+def benchmark(args: argparse.Namespace) -> None:
+    is_i2i = args.mode == "i2i"
+
+    print("=" * 60)
+    print("GLM-Image vLLM-Omni Benchmark")
+    print(f"Mode: {args.mode}  |  Model: {args.model_path}")
+    print(f"Size: {args.height}x{args.width}  |  Steps: {args.num_inference_steps}")
+    print("=" * 60)
+
+    # Load dataset
+    items = load_dataset(args.dataset_path, args.mode, args.num_prompts)
+    if not items:
+        print("No prompts loaded. Exiting.")
+        return
+    print(f"Loaded {len(items)} prompts for {args.mode} mode")
+
+    # Download I2I source images
+    if is_i2i:
+        print("Preparing source images...")
+        for item in items:
+            url = item.get("image_url", "")
+            if url:
+                item["image_path"] = download_image(url, IMAGE_CACHE_DIR)
+            else:
+                item["image_path"] = None
+
+    # Init Omni
+    deploy_config = resolve_deploy_config(args)
+    print(f"\nInitializing vLLM-Omni (deploy config: {deploy_config}) ...")
+    t0 = time.perf_counter()
+
+    omni = Omni(
+        model=args.model_path,
+        deploy_config=deploy_config,
+        log_stats=args.log_stats,
+        stage_init_timeout=args.stage_init_timeout,
+        enable_diffusion_pipeline_profiler=args.enable_diffusion_pipeline_profiler,
+        enable_ar_profiler=args.enable_ar_profiler,
+    )
+
+    init_time = time.perf_counter() - t0
+    print(f"Initialized in {init_time:.2f}s")
+
+    # Sampling params
+    max_tokens = compute_max_tokens(args.height, args.width, is_i2i=is_i2i)
+    ar_params = SamplingParams(
+        temperature=0.9,
+        top_p=0.75,
+        top_k=GLM_IMAGE_VISION_VOCAB_SIZE,
+        max_tokens=max_tokens,
+        stop_token_ids=[GLM_IMAGE_EOS_TOKEN_ID],
+        seed=args.seed,
+        detokenize=False,
+        extra_args={"target_h": args.height, "target_w": args.width},
+    )
+    diff_params = OmniDiffusionSamplingParams(
+        num_inference_steps=args.num_inference_steps,
+        guidance_scale=args.guidance_scale,
+        height=args.height,
+        width=args.width,
+        seed=args.seed,
+    )
+    sampling_params_list = [ar_params, diff_params]
+
+    # Build all prompts
+    gen_kw = {
+        "seed": args.seed,
+        "num_inference_steps": args.num_inference_steps,
+        "guidance_scale": args.guidance_scale,
+    }
+    all_prompts = []
+    for item in items:
+        if is_i2i:
+            img_path = item.get("image_path")
+            if not img_path or not os.path.exists(img_path):
+                continue
+            all_prompts.append(build_prompt_i2i(item["prompt"], img_path, args.height, args.width, **gen_kw))
+        else:
+            all_prompts.append(build_prompt_t2i(item["prompt"], args.height, args.width, **gen_kw))
+
+    valid = len(all_prompts)
+    print(f"Valid prompts: {valid}")
+
+    # Create output dir
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    # Warmup: run 1 request to prime caches, CUDA graphs, etc.
+    if all_prompts:
+        print("Running warmup request...")
+        try:
+            warmup_prompt = [all_prompts[0]]
+            omni.generate(warmup_prompt, sampling_params_list, py_generator=False)
+            print("Warmup done.\n")
+        except Exception as e:
+            print(f"Warmup failed (continuing): {e}")
+
+    # Run
+    print(f"\nRunning {valid} requests...")
+    print("-" * 60)
+
+    latencies = []
+    all_stage_durations: list[dict[str, float]] = []
+    success = 0
+    failed = 0
+    wall_start = time.perf_counter()
+
+    try:
+        output_idx = 0
+        for stage_outputs in omni.generate(all_prompts, sampling_params_list, py_generator=True):
+            if stage_outputs.final_output_type == "image":
+                request_output = stage_outputs.request_output
+                request_id = getattr(request_output, "request_id", "")
+
+                images = getattr(request_output, "images", [])
+                if not images and hasattr(request_output, "multimodal_output"):
+                    mm = request_output.multimodal_output
+                    if isinstance(mm, dict):
+                        images = mm.get("images", [])
+
+                elapsed = time.perf_counter() - wall_start
+                if images:
+                    for img in images:
+                        if isinstance(img, Image.Image):
+                            out_path = os.path.join(args.output_dir, f"{output_idx:04d}.png")
+                            img.save(out_path)
+                    success += 1
+                    latencies.append(elapsed)
+                    stage_durations = getattr(stage_outputs, "stage_durations", {})
+                    if stage_durations:
+                        all_stage_durations.append(stage_durations)
+                    # Show wall-clock elapsed and pipeline breakdown if available
+                    preprocess_str = ""
+                    if "preprocess_ms" in stage_durations:
+                        preprocess_str = f" preprocess={stage_durations['preprocess_ms'] / 1000.0:.2f}s"
+                    print(f"  [{success}/{valid}] id={request_id[:8]} {elapsed:.2f}s{preprocess_str}")
+                    output_idx += 1
+                else:
+                    failed += 1
+    except Exception as e:
+        print(f"Error: {e}")
+        failed = valid - success
+
+    total_gen_time = time.perf_counter() - wall_start
+
+    # Diff stage_0_gen_ms with previous request to remove accumulated wait time.
+    # stage_0_gen_ms is measured from submit_ts (same for all requests submitted
+    # at once), so it accumulates queue/scheduling overhead across requests.
+    # Other stages and pipeline timings are per-request already.
+    _TIMING_ORDER = [
+        "preprocess_ms",
+        "stage_0_gen_ms",
+        "ar2diffusion_ms",
+        "stage_1_gen_ms",
+        "queue_wait_ms",
+    ]
+
+    per_request_actual: list[dict[str, float]] = []
+    prev_stage_0_ms = 0.0
+    for sd in all_stage_durations:
+        actual = dict(sd)
+        s0 = sd.get("stage_0_gen_ms", 0.0)
+        actual["stage_0_gen_ms"] = s0 - prev_stage_0_ms
+        prev_stage_0_ms = s0
+        per_request_actual.append(actual)
+
+    per_request_e2e_ms: list[float] = []
+    for actual in per_request_actual:
+        e2e_ms = sum(v for k, v in actual.items() if k in _TIMING_ORDER)
+        if e2e_ms > 0:
+            per_request_e2e_ms.append(e2e_ms)
+
+    # Report
+    print("\n" + "=" * 60)
+    print("vLLM-Omni Benchmark Results")
+    print("=" * 60)
+    print(f"{'Mode:':<40} {args.mode}")
+    print(f"{'Model:':<40} {args.model_path}")
+    print(f"{'Image size:':<40} {args.height}x{args.width}")
+    print(f"{'Num inference steps:':<40} {args.num_inference_steps}")
+    print("-" * 50)
+    print(f"{'Init time (s):':<40} {init_time:.2f}")
+    print(f"{'Successful:':<40} {success}/{valid}")
+    print(f"{'Failed:':<40} {failed}")
+    print("-" * 50)
+
+    if per_request_e2e_ms:
+        per_request_s = np.array(per_request_e2e_ms) / 1000.0
+        print(f"{'Total generation time (s):':<40} {total_gen_time:.2f}")
+        print(f"{'Throughput (img/s):':<40} {success / total_gen_time:.4f}")
+        print(f"{'Latency Mean (s):':<40} {per_request_s.mean():.4f}")
+        print(f"{'Latency Median (s):':<40} {np.median(per_request_s):.4f}")
+        print(f"{'Latency P95 (s):':<40} {np.percentile(per_request_s, 95):.4f}")
+        print(f"{'Latency P99 (s):':<40} {np.percentile(per_request_s, 99):.4f}")
+        print(f"{'Latency Min (s):':<40} {per_request_s.min():.4f}")
+        print(f"{'Latency Max (s):':<40} {per_request_s.max():.4f}")
+    elif latencies:
+        per_request = np.diff([0.0] + list(latencies))
+        print(f"{'Total generation time (s):':<40} {total_gen_time:.2f}")
+        print(f"{'Throughput (img/s):':<40} {success / total_gen_time:.4f}")
+        print(f"{'Latency Mean (s) [wall-clock]:':<40} {per_request.mean():.4f}")
+        print(f"{'Latency Median (s) [wall-clock]:':<40} {np.median(per_request):.4f}")
+        print(f"{'Latency P95 (s) [wall-clock]:':<40} {np.percentile(per_request, 95):.4f}")
+        print(f"{'Latency P99 (s) [wall-clock]:':<40} {np.percentile(per_request, 99):.4f}")
+        print(f"{'Latency Min (s) [wall-clock]:':<40} {per_request.min():.4f}")
+        print(f"{'Latency Max (s) [wall-clock]:':<40} {per_request.max():.4f}")
+
+    if per_request_actual:
+        print("-" * 50)
+        print("Pipeline Timings Mean:")
+        for key in _TIMING_ORDER:
+            vals = [d.get(key, 0.0) for d in per_request_actual]
+            if any(v != 0 for v in vals):
+                unit = "ms" if key.endswith("_ms") else "s"
+                print(f"  {key + ':':<38} {np.mean(vals):.4f} ({unit})")
+        # Show any extra keys not in the ordered list
+        ordered_set = set(_TIMING_ORDER)
+        extra_keys = sorted(k for k in per_request_actual[0].keys() if k not in ordered_set)
+        for key in extra_keys:
+            vals = [d.get(key, 0.0) for d in per_request_actual]
+            if any(v != 0 for v in vals):
+                unit = "ms" if key.endswith("_ms") else "s"
+                print(f"  {key + ':':<38} {np.mean(vals):.4f} ({unit})")
+
+    print(f"\n{'Output dir:':<40} {args.output_dir}")
+    print("=" * 60)
+
+    # Metrics JSON
+    metrics = {
+        "backend": "vllm-omni",
+        "mode": args.mode,
+        "model": args.model_path,
+        "height": args.height,
+        "width": args.width,
+        "num_inference_steps": args.num_inference_steps,
+        "init_time_s": init_time,
+        "completed_requests": success,
+        "failed_requests": failed,
+        "total_gen_time_s": total_gen_time,
+        "throughput_qps": success / total_gen_time if total_gen_time > 0 else 0,
+    }
+    if per_request_e2e_ms:
+        per_request_s = np.array(per_request_e2e_ms) / 1000.0
+        metrics["latency_mean"] = float(per_request_s.mean())
+        metrics["latency_median"] = float(np.median(per_request_s))
+        metrics["latency_p95"] = float(np.percentile(per_request_s, 95))
+        metrics["latency_p99"] = float(np.percentile(per_request_s, 99))
+    elif latencies:
+        per_request = np.diff([0.0] + list(latencies))
+        metrics["latency_mean"] = float(per_request.mean())
+        metrics["latency_median"] = float(np.median(per_request))
+        metrics["latency_p95"] = float(np.percentile(per_request, 95))
+        metrics["latency_p99"] = float(np.percentile(per_request, 99))
+    else:
+        metrics["latency_mean"] = 0
+        metrics["latency_median"] = 0
+        metrics["latency_p95"] = 0
+        metrics["latency_p99"] = 0
+    if per_request_actual:
+        all_keys = list(_TIMING_ORDER) + sorted(k for k in per_request_actual[0].keys() if k not in set(_TIMING_ORDER))
+        stage_metrics = {}
+        for key in all_keys:
+            vals = [d.get(key, 0.0) for d in per_request_actual]
+            stage_metrics[key] = {
+                "mean": float(np.mean(vals)),
+                "median": float(np.median(vals)),
+                "p95": float(np.percentile(vals, 95)),
+            }
+        metrics["stage_durations"] = stage_metrics
+    if args.output_file:
+        with open(args.output_file, "w") as f:
+            json.dump(metrics, f, indent=2)
+        print(f"Metrics saved to {args.output_file}")
+
+    omni.close()
+    print("Done!")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="GLM-Image vLLM-Omni offline benchmark")
+    parser.add_argument("--model-path", type=str, default="zai-org/GLM-Image")
+    parser.add_argument("--deploy-config", type=str, default=None, help="Deploy config YAML")
+    parser.add_argument("--mode", type=str, default="t2i", choices=["t2i", "i2i"])
+    parser.add_argument("--dataset-path", type=str, default=None, help="Path to prompt.json")
+    parser.add_argument("--num-prompts", type=int, default=10)
+    parser.add_argument("--height", type=int, default=HEIGHT)
+    parser.add_argument("--width", type=int, default=WIDTH)
+    parser.add_argument("--num-inference-steps", type=int, default=NUM_INFERENCE_STEPS)
+    parser.add_argument("--guidance-scale", type=float, default=GUIDANCE_SCALE)
+    parser.add_argument("--seed", type=int, default=SEED)
+    parser.add_argument("--output-dir", type=str, default="benchmarks/glm_image/vllm-omni/outputs")
+    parser.add_argument("--output-file", type=str, default=None, help="JSON file for metrics")
+    parser.add_argument("--stage-init-timeout", type=int, default=600)
+    parser.add_argument(
+        "--enable-diffusion-pipeline-profiler",
+        action="store_true",
+        help="Enable diffusion pipeline profiler for stage-level timing",
+    )
+    parser.add_argument(
+        "--enable-ar-profiler",
+        action="store_true",
+        help="Enable AR stage profiler to include AR timing in stage_durations",
+    )
+    parser.add_argument(
+        "--log-stats",
+        action="store_true",
+        help="Enable detailed per-request pipeline stats logging",
+    )
+    args = parser.parse_args()
+    benchmark(args)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/qwen3-omni/README.md

@@ -1,86 +0,0 @@
-# Benchmarks Guide
-
-This README explains how to (1) prepare benchmark datasets and (2) run the provided Qwen3-Omni benchmarks.
-
-## 1) Prepare the dataset (SeedTTS top100)
-
-```bash
-cd benchmarks/build_dataset
-pip install gdown
-
-# Download SeedTTS test set from Google Drive
-gdown --id 1GlSjVfSHkW3-leKKBlfrjuuTGqQ_xaLP
-
-# Extract
-tar -xf seedtts_testset.tar
-
-# Copy metadata and extract top-100 prompts
-cp seedtts_testset/en/meta.lst meta.lst
-python extract_prompts.py -i meta.lst -o top100.txt -n 100
-
-# (Optional) clean up to save space
-rm -rf seedtts_testset seedtts_testset.tar meta.lst
-```
-
-Artifacts:
-- `benchmarks/build_dataset/top100.txt` — 100 text prompts (one per line).
-
-## 2) Run benchmarks
-
-All commands assume repo root (`vllm-omni`).
-
-### A. Transformers benchmark (offline, HF Transformers)
-
-```
-bash benchmarks/qwen3-omni/transformers/eval_qwen3_moe_omni_transformers.sh
-```
-
-What it does:
-- Runs `qwen3_omni_moe_transformers.py` over `top100.txt` with `--num-prompts 100`.
-- Outputs to `benchmarks/qwen3-omni/transformers/benchmark_results/`:
-  - `perf_stats.json` — aggregated & per-prompt TPS/latency (thinker/talker/code2wav/overall).
-  - `results.json` — per-prompt outputs and audio paths.
-  - `audio/` — ~100 generated `.wav` files.
-
-Key checks:
-- `overall_tps` and `*_tps_avg` should be non-zero and reasonably stable.
-- Investigate any 0/NaN or unusually low TPS / long-tail latency.
-
-### B. vLLM Omni end-to-end benchmark (pipeline)
-
-```
-bash benchmarks/qwen3-omni/vllm_omni/eval_qwen3_moe_omni.sh
-```
-
-What it does:
-- Runs `examples/offline_inference/qwen3_omni/end2end.py` with `--log-stats`.
-- Uses `benchmarks/build_dataset/top100.txt` and writes to:
-  - Logs: `benchmarks/qwen3-omni/vllm_omni/logs/`
-    - `omni_pipeline_text.orchestrator.stats.jsonl` — per-stage latency stats.
-    - `omni_pipeline_text.overall.stats.jsonl` — end-to-end latency/TPS.
-    - `omni_pipeline_text.stage{0,1,2}.log` — per-stage detailed logs/errors.
-  - Outputs: `benchmarks/qwen3-omni/vllm_omni/outputs/` — ~100 text and `.wav` files.
-
-Key checks:
-- Overall stats: end-to-end latency/TPS should be reasonable.
-- Orchestrator stats: per-stage latency should be stable; investigate long tails.
-- Stage logs: ensure no errors and no unusually slow stages.
-
-
-## Performance snapshot
-
-The chart below summarizes our measured Qwen3-Omni MoE end-to-end benchmark, comparing vLLM-Omni against HF Transformers. It shows the overall throughput advantage for vLLM-Omni. These are actual experiment results—please refer to this performance when evaluating or reproducing the benchmark.
-
-![vLLM-Omni vs HF](./vllm-omni-vs-hf.png)
-
-## Directory layout
-- `benchmarks/build_dataset/` — dataset prep utilities (e.g., SeedTTS top100).
-- `benchmarks/<model>/vllm_omni/` — vLLM-Omni pipeline benchmarks, logs, outputs.
-- Add new tasks under `benchmarks/<model>/...` with the same pattern: `transformers/`, `vllm_omni/`, task-specific README, and (optionally) dataset prep notes.
-- `benchmarks/<model>/vllm-omni-vs-hf.png` — current performance snapshot (overall throughput comparison).
-- `benchmarks/<model>/transformers/` — HF Transformers benchmarks (offline reference).
-
-## Troubleshooting
-- Make sure GPU/driver/FlashAttention2 requirements are met for the chosen model.
-- If downloads fail, confirm network access to Google Drive (`gdown`) and Hugging Face.
-- If audio files are missing, check for errors in stage logs or model generation.***

benchmarks/qwen3-omni/transformers/eval_qwen3_moe_omni_transformers.sh

@@ -1,26 +0,0 @@
-#!/bin/bash
-# Qwen3-Omni Transformers Benchmark Evaluation Script
-# This script must be run from the vllm-omni root directory
-
-# Get the directory where this script is located
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-
-# Navigate to vllm-omni root directory (4 levels up from script location)
-VLLM_OMNI_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
-cd "$VLLM_OMNI_ROOT" || { echo "Error: Failed to navigate to vllm-omni directory"; exit 1; }
-
-echo "Working directory: $(pwd)"
-# Verify we're in the correct directory and run benchmark
-if [[ ! -f "benchmarks/qwen3-omni/transformers/qwen3_omni_moe_transformers.py" ]]; then
-    echo "Error: Not in vllm-omni root directory. Please run from vllm-omni folder."
-else
-    cd benchmarks/qwen3-omni/transformers
-
-    python qwen3_omni_moe_transformers.py --prompts-file ../../build_dataset/top100.txt --num-prompts 100
-
-    echo "Logs and outputs are saved to $(pwd)/benchmark_results:"
-    echo "  - perf_stats.json    Aggregated/per-prompt TPS and latency (thinker/talker/code2wav/overall)"
-    echo "  - results.json       Per-prompt outputs and audio paths"
-    echo "  - audio/             Generated wav files, there should be 100 wav file generated"
-    echo "Key checks: overall_tps and *_tps_avg should be non-zero and stable; investigate 0/NaN or unusually low TPS/long-tail latency."
-fi

benchmarks/qwen3-omni/transformers/qwen3_omni_moe_model.py

@@ -1,265 +0,0 @@
-import time
-
-import torch
-from transformers import Qwen3OmniMoeForConditionalGeneration
-
-
-class Qwen3OmniMoeForConditionalGenerationWithLogging(Qwen3OmniMoeForConditionalGeneration):
-    @torch.no_grad()
-    def generate(
-        self,
-        input_ids: torch.Tensor | None = None,
-        speaker: str = "Ethan",
-        use_audio_in_video: bool = False,
-        return_audio: bool | None = None,
-        thinker_max_new_tokens: int = 1024,
-        thinker_eos_token_id: int = 151645,
-        talker_max_new_tokens: int = 4096,
-        talker_do_sample: bool = True,
-        talker_top_k: int = 50,
-        talker_top_p: float = 1.0,
-        talker_temperature: float = 0.9,
-        talker_repetition_penalty: float = 1.05,
-        **kwargs,
-    ):
-        total_t0 = time.time()
-        perf_stats = {
-            "thinker_tokens": 0,
-            "thinker_time_s": 0.0,
-            "thinker_tps": 0.0,
-            "talker_tokens": 0,
-            "talker_time_s": 0.0,
-            "talker_tps": 0.0,
-            "code2wav_tokens": 0,
-            "code2wav_time_s": 0.0,
-            "code2wav_tps": 0.0,
-            "total_tokens": 0,
-            "total_time_s": 0.0,
-            "total_tps": 0.0,
-        }
-        if return_audio and not self.has_talker:
-            raise ValueError(
-                "Cannot use talker when talker module not initialized. "
-                "Use `enable_talker` method or set enable_talker in config "
-                "to enable talker."
-            )
-        if return_audio is None:
-            return_audio = self.has_talker
-
-        shared_kwargs = {"use_audio_in_video": use_audio_in_video}
-        thinker_kwargs = {
-            "max_new_tokens": thinker_max_new_tokens,
-            "eos_token_id": thinker_eos_token_id,
-        }
-
-        talker_kwargs = {}
-        token2wav_kwargs = {}
-        if return_audio:
-            speaker_id = self.config.talker_config.speaker_id.get(speaker.lower())
-            if speaker_id is None:
-                raise NotImplementedError(f"Speaker {speaker} not implemented")
-            if input_ids.shape[0] != 1:
-                raise NotImplementedError("Qwen3-Omni currently does not support batched inference with audio output")
-            talker_suppressed_tokens = [
-                i
-                for i in range(
-                    self.config.talker_config.text_config.vocab_size - 1024,
-                    self.config.talker_config.text_config.vocab_size,
-                )
-                if i != self.config.talker_config.codec_eos_token_id
-            ]  # Suppress additional special tokens, should not be predicted
-            talker_kwargs = {
-                "max_new_tokens": talker_max_new_tokens,
-                "do_sample": talker_do_sample,
-                "top_k": talker_top_k,
-                "top_p": talker_top_p,
-                "temperature": talker_temperature,
-                "eos_token_id": self.config.talker_config.codec_eos_token_id,
-                "repetition_penalty": talker_repetition_penalty,
-                "suppress_tokens": talker_suppressed_tokens,
-                "output_hidden_states": True,
-                "return_dict_in_generate": True,
-            }
-            token2wav_kwargs = {}
-
-        for key, value in kwargs.items():
-            if key.startswith("thinker_"):
-                thinker_kwargs[key[len("thinker_") :]] = value
-            elif key.startswith("talker_"):
-                talker_kwargs[key[len("talker_") :]] = value
-            elif key.startswith("token2wav_"):
-                token2wav_kwargs[key[len("token2wav_") :]] = value
-            # Process special input values
-            elif key == "feature_attention_mask":
-                thinker_kwargs[key] = value
-                talker_kwargs["audio_feature_lengths"] = torch.sum(value, dim=1)
-            elif key in ("input_features", "attention_mask"):
-                thinker_kwargs[key] = value
-            # Put other key to shared kwargs
-            else:
-                shared_kwargs[key] = value
-
-        # Merge kwargs
-        for key, value in shared_kwargs.items():
-            if key not in thinker_kwargs:
-                thinker_kwargs[key] = value
-            if key not in talker_kwargs and key in ["image_grid_thw", "video_grid_thw", "video_second_per_grid"]:
-                talker_kwargs[key] = value
-            if key not in token2wav_kwargs:
-                token2wav_kwargs[key] = value
-
-        # 1. Generate from thinker module
-        generate_audio = return_audio and self.has_talker
-        if generate_audio:
-            thinker_kwargs["output_hidden_states"] = True
-            thinker_kwargs["return_dict_in_generate"] = True
-
-        t0 = time.time()
-        thinker_result = self.thinker.generate(input_ids=input_ids, **thinker_kwargs)
-        t1 = time.time()
-        perf_stats["thinker_time_s"] = max(0.0, t1 - t0)
-        try:
-            prompt_len = int(input_ids.shape[1]) if input_ids is not None else 0
-            total_len = int(thinker_result.sequences.shape[-1])
-            thinker_out_len = max(0, total_len - prompt_len)
-        except Exception:
-            thinker_out_len = 0
-        perf_stats["thinker_tokens"] = thinker_out_len
-        perf_stats["thinker_tps"] = (
-            (thinker_out_len / perf_stats["thinker_time_s"]) if perf_stats["thinker_time_s"] > 0 else 0.0
-        )
-
-        if not generate_audio:
-            perf_stats["total_tokens"] = perf_stats["thinker_tokens"]
-            perf_stats["total_time_s"] = time.time() - total_t0
-            perf_stats["total_tps"] = (
-                (perf_stats["total_tokens"] / perf_stats["total_time_s"]) if perf_stats["total_time_s"] > 0 else 0.0
-            )
-            # attach stats to self
-            setattr(self, "_perf_stats_last", perf_stats)
-            if not hasattr(self, "_perf_stats_history"):
-                setattr(self, "_perf_stats_history", [])
-            self._perf_stats_history.append(perf_stats)
-            return thinker_result, None
-
-        # 2. Prepare talker input
-        thinker_embed = torch.cat([hidden_states[0] for hidden_states in thinker_result.hidden_states], dim=1).to(
-            self.talker.device
-        )  # [1 t d]
-        thinker_hidden = torch.cat(
-            [
-                hidden_states[self.config.talker_config.accept_hidden_layer]
-                for hidden_states in thinker_result.hidden_states
-            ],
-            dim=1,
-        ).to(self.talker.device)  # [1 t d]
-
-        im_start_indexes = torch.cat(
-            (
-                torch.nonzero(input_ids[0] == self.config.im_start_token_id).squeeze(),
-                torch.tensor([thinker_result.sequences.shape[-1]], device=input_ids.device, dtype=input_ids.dtype),
-            ),
-            dim=-1,
-        ).to(self.talker.device)  # Shape [n_starts + 1]; Take batch 0 since batched inference is not supported here.
-        multimodal_mask = (
-            (thinker_result.sequences == self.config.thinker_config.audio_token_id) |
-            (thinker_result.sequences == self.config.thinker_config.image_token_id) |
-            (thinker_result.sequences == self.config.thinker_config.video_token_id)
-        ).to(self.talker.device)  # [1 t] # fmt: skip
-
-        talker_special_tokens = torch.tensor(
-            [[self.config.tts_bos_token_id, self.config.tts_eos_token_id, self.config.tts_pad_token_id]],
-            device=self.thinker.device,
-            dtype=input_ids.dtype,
-        )
-        tts_bos_embed, tts_eos_embed, tts_pad_embed = (
-            self.talker.text_projection(self.thinker.get_input_embeddings()(talker_special_tokens))
-            .to(self.talker.device)
-            .chunk(3, dim=1)
-        )  # 3 * [1 1 d]
-
-        talker_input_embeds = []  # [1 t d]
-        talker_input_ids = []
-        # For every chatml parts
-        for i in range(len(im_start_indexes) - 1):
-            im_start_index = im_start_indexes[i]
-            segment_end_index = im_start_indexes[i + 1]
-            role_token = input_ids[0][im_start_index + 1]
-            # Talker should ignore thinker system prompt
-            if role_token == self.config.system_token_id:
-                continue
-            # Talker takes word embeddings for tokens and hidden state from `accept_hidden_layer` for multimodal inputs
-            elif role_token == self.config.user_token_id:
-                talker_user_part = self._get_talker_user_parts(
-                    im_start_index, segment_end_index, multimodal_mask, thinker_hidden, thinker_embed
-                )
-                talker_input_embeds.append(talker_user_part)
-                talker_input_ids.append(thinker_result.sequences[:, im_start_index:segment_end_index])
-            # Take assistant output (for now)
-            elif role_token == self.config.assistant_token_id and i == len(im_start_indexes) - 2:
-                talker_assistant_embeds, talker_assistant_ids, trailing_text_hidden = self._get_talker_assistant_parts(
-                    im_start_index,
-                    segment_end_index,
-                    speaker_id,
-                    thinker_embed,
-                    tts_pad_embed,
-                    tts_bos_embed,
-                    tts_eos_embed,
-                )
-                talker_input_embeds.append(talker_assistant_embeds)
-                talker_input_ids.append(talker_assistant_ids)
-            # History assistant output (ignore for now)
-            elif role_token == self.config.assistant_token_id and i != len(im_start_indexes) - 2:
-                continue
-            else:
-                raise AssertionError("Expect role id after <|im_start|> (assistant, user, system)")
-        talker_input_embed = torch.cat([embed.to(self.talker.device) for embed in talker_input_embeds], dim=1)
-        talker_input_id = torch.cat([embed.to(self.talker.device) for embed in talker_input_ids], dim=1)
-        t2 = time.time()
-        talker_result = self.talker.generate(
-            inputs_embeds=talker_input_embed,
-            trailing_text_hidden=trailing_text_hidden,
-            tts_pad_embed=tts_pad_embed,
-            talker_input_ids=talker_input_id,  # Not use input_ids to prevent repetition penalty out of bound
-            **talker_kwargs,
-        )
-        t3 = time.time()
-        perf_stats["talker_time_s"] = max(0.0, t3 - t2)
-        talker_codes = (
-            torch.stack([hid[-1] for hid in talker_result.hidden_states if hid[-1] is not None], dim=1)
-            .transpose(1, 2)
-            .to(self.code2wav.device)
-        )
-        try:
-            # codes shape: (B, num_quantizers, T). We log T as token length.
-            perf_stats["talker_tokens"] = int(talker_codes.shape[-1])
-        except Exception:
-            perf_stats["talker_tokens"] = 0
-        perf_stats["talker_tps"] = (
-            (perf_stats["talker_tokens"] / perf_stats["talker_time_s"]) if perf_stats["talker_time_s"] > 0 else 0.0
-        )
-        t4 = time.time()
-        talker_wavs = self.code2wav.chunked_decode(talker_codes, chunk_size=300, left_context_size=25).float()
-        t5 = time.time()
-        perf_stats["code2wav_time_s"] = max(0.0, t5 - t4)
-        perf_stats["code2wav_tokens"] = perf_stats["talker_tokens"]  # same T, not times 16
-        perf_stats["code2wav_tps"] = (
-            (perf_stats["code2wav_tokens"] / perf_stats["code2wav_time_s"])
-            if perf_stats["code2wav_time_s"] > 0
-            else 0.0
-        )
-        perf_stats["total_tokens"] = perf_stats["thinker_tokens"] + perf_stats["talker_tokens"]
-        perf_stats["total_time_s"] = time.time() - total_t0
-        perf_stats["total_tps"] = (
-            (perf_stats["total_tokens"] / perf_stats["total_time_s"]) if perf_stats["total_time_s"] > 0 else 0.0
-        )
-        setattr(self, "_perf_stats_last", perf_stats)
-        if not hasattr(self, "_perf_stats_history"):
-            setattr(self, "_perf_stats_history", [])
-        self._perf_stats_history.append(perf_stats)
-        return thinker_result, talker_wavs.float()
-
-
-__all__ = [
-    "Qwen3OmniMoeForConditionalGenerationWithLogging",
-]

benchmarks/qwen3-omni/transformers/qwen3_omni_moe_transformers.py

@@ -1,275 +0,0 @@
-import argparse
-import json
-import os
-
-import soundfile as sf
-from qwen3_omni_moe_model import Qwen3OmniMoeForConditionalGenerationWithLogging
-from qwen_omni_utils import process_mm_info
-from tqdm import tqdm
-from transformers import Qwen3OmniMoeProcessor
-
-MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
-# MODEL_PATH = "Qwen/Qwen3-Omni-30B-A3B-Thinking"
-
-
-def load_prompts(prompts_file: str) -> list[str]:
-    """Load prompts from a text file, one prompt per line."""
-    prompts = []
-    with open(prompts_file, encoding="utf-8") as f:
-        for line in f:
-            line = line.strip()
-            if line:
-                prompts.append(line)
-    return prompts
-
-
-def run_benchmark(
-    model,
-    processor,
-    prompts: list[str],
-    output_dir: str = "benchmark_results",
-    speaker: str = "Ethan",
-    use_audio_in_video: bool = True,
-):
-    """
-    Run benchmark on a list of prompts and collect performance stats.
-
-    Args:
-        model: The Qwen3OmniMoe model
-        processor: The Qwen3OmniMoe processor
-        prompts: List of text prompts to process
-        output_dir: Directory to save results
-        speaker: Speaker voice for audio output
-        use_audio_in_video: Whether to use audio in video
-
-    Returns:
-        tuple: (aggregated_stats, results, audio_outputs)
-            - aggregated_stats: dict with aggregated performance statistics
-            - results: list of dicts with per-prompt results
-            - audio_outputs: list of audio tensors/arrays (or None if no audio)
-    """
-    os.makedirs(output_dir, exist_ok=True)
-    audio_dir = os.path.join(output_dir, "audio")
-    os.makedirs(audio_dir, exist_ok=True)
-
-    all_stats = []
-    results = []
-    audio_outputs = []
-
-    for idx, prompt in enumerate(tqdm(prompts, desc="Processing prompts")):
-        conversation = [
-            {
-                "role": "user",
-                "content": [{"type": "text", "text": prompt}],
-            },
-        ]
-
-        # Preparation for inference
-        text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
-        audios, images, videos = process_mm_info(conversation, use_audio_in_video=use_audio_in_video)
-        inputs = processor(
-            text=text,
-            audio=audios,
-            images=images,
-            videos=videos,
-            return_tensors="pt",
-            padding=True,
-            use_audio_in_video=use_audio_in_video,
-        )
-        inputs = inputs.to(model.device).to(model.dtype)
-
-        # Inference: Generation of the output text and audio
-        text_ids, audio = model.generate(
-            **inputs, speaker=speaker, thinker_return_dict_in_generate=True, use_audio_in_video=use_audio_in_video
-        )
-
-        # Decode output text
-        output_text = processor.batch_decode(
-            text_ids.sequences[:, inputs["input_ids"].shape[1] :],
-            skip_special_tokens=True,
-            clean_up_tokenization_spaces=False,
-        )[0]
-
-        # Collect performance stats
-        perf_stats = None
-        if hasattr(model, "_perf_stats_last"):
-            perf_stats = model._perf_stats_last.copy()
-            perf_stats["prompt_idx"] = idx
-            perf_stats["prompt"] = prompt
-            all_stats.append(perf_stats)
-
-        # Save audio and collect audio output
-        audio_path = None
-        audio_data = None
-        if audio is not None:
-            audio_data = audio.reshape(-1).detach().cpu().numpy()
-            audio_path = os.path.join(audio_dir, f"output_{idx:04d}.wav")
-            sf.write(
-                audio_path,
-                audio_data,
-                samplerate=24000,
-            )
-            audio_outputs.append(audio_data)
-        else:
-            audio_outputs.append(None)
-
-        # Save result
-        result = {
-            "idx": idx,
-            "prompt": prompt,
-            "output": output_text,
-            "audio_path": audio_path,
-            "perf_stats": perf_stats,
-        }
-        results.append(result)
-
-    # Aggregate statistics
-    aggregated_stats = aggregate_stats(all_stats)
-
-    # Save all results
-    results_path = os.path.join(output_dir, "results.json")
-    with open(results_path, "w", encoding="utf-8") as f:
-        json.dump(results, f, ensure_ascii=False, indent=2)
-
-    # Save aggregated stats
-    stats_path = os.path.join(output_dir, "perf_stats.json")
-    with open(stats_path, "w", encoding="utf-8") as f:
-        json.dump({"aggregated": aggregated_stats, "per_prompt": all_stats}, f, ensure_ascii=False, indent=2)
-
-    # Count saved audio files
-    num_audio_saved = sum(1 for a in audio_outputs if a is not None)
-    print(f"\nSaved {num_audio_saved} audio files to {audio_dir}/")
-
-    return aggregated_stats, results, audio_outputs
-
-
-def aggregate_stats(all_stats: list[dict]) -> dict:
-    """Aggregate performance statistics from multiple runs."""
-    if not all_stats:
-        return {}
-
-    keys = [
-        "thinker_tokens",
-        "thinker_time_s",
-        "thinker_tps",
-        "talker_tokens",
-        "talker_time_s",
-        "talker_tps",
-        "code2wav_tokens",
-        "code2wav_time_s",
-        "code2wav_tps",
-        "total_tokens",
-        "total_time_s",
-        "total_tps",
-    ]
-
-    aggregated = {
-        "num_samples": len(all_stats),
-    }
-
-    for key in keys:
-        values = [s.get(key, 0) for s in all_stats if key in s]
-        if values:
-            aggregated[f"{key}_sum"] = sum(values)
-            aggregated[f"{key}_avg"] = sum(values) / len(values)
-            aggregated[f"{key}_min"] = min(values)
-            aggregated[f"{key}_max"] = max(values)
-
-    # Calculate overall throughput
-    total_tokens = aggregated.get("total_tokens_sum", 0)
-    total_time = aggregated.get("total_time_s_sum", 0)
-    if total_time > 0:
-        aggregated["overall_tps"] = total_tokens / total_time
-
-    return aggregated
-
-
-def print_stats(stats: dict):
-    """Print performance statistics in a formatted way."""
-    print("\n" + "=" * 60)
-    print("Performance Statistics Summary")
-    print("=" * 60)
-
-    print(f"\nNumber of samples: {stats.get('num_samples', 0)}")
-
-    print("\n--- Thinker ---")
-    print(f"  Total tokens:  {stats.get('thinker_tokens_sum', 0):.0f}")
-    print(f"  Total time:    {stats.get('thinker_time_s_sum', 0):.2f}s")
-    print(f"  Avg TPS:       {stats.get('thinker_tps_avg', 0):.2f}")
-    print(f"  Min TPS:       {stats.get('thinker_tps_min', 0):.2f}")
-    print(f"  Max TPS:       {stats.get('thinker_tps_max', 0):.2f}")
-
-    print("\n--- Talker ---")
-    print(f"  Total tokens:  {stats.get('talker_tokens_sum', 0):.0f}")
-    print(f"  Total time:    {stats.get('talker_time_s_sum', 0):.2f}s")
-    print(f"  Avg TPS:       {stats.get('talker_tps_avg', 0):.2f}")
-    print(f"  Min TPS:       {stats.get('talker_tps_min', 0):.2f}")
-    print(f"  Max TPS:       {stats.get('talker_tps_max', 0):.2f}")
-
-    print("\n--- Code2Wav ---")
-    print(f"  Total tokens:  {stats.get('code2wav_tokens_sum', 0):.0f}")
-    print(f"  Total time:    {stats.get('code2wav_time_s_sum', 0):.2f}s")
-    print(f"  Avg TPS:       {stats.get('code2wav_tps_avg', 0):.2f}")
-    print(f"  Min TPS:       {stats.get('code2wav_tps_min', 0):.2f}")
-    print(f"  Max TPS:       {stats.get('code2wav_tps_max', 0):.2f}")
-
-    print("\n--- Overall ---")
-    print(f"  Total tokens:  {stats.get('total_tokens_sum', 0):.0f}")
-    print(f"  Total time:    {stats.get('total_time_s_sum', 0):.2f}s")
-    print(f"  Overall TPS:   {stats.get('overall_tps', 0):.2f}")
-    print(f"  Avg TPS:       {stats.get('total_tps_avg', 0):.2f}")
-    print(f"  Min TPS:       {stats.get('total_tps_min', 0):.2f}")
-    print(f"  Max TPS:       {stats.get('total_tps_max', 0):.2f}")
-
-    print("=" * 60 + "\n")
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Qwen3-Omni Benchmark Script")
-    parser.add_argument(
-        "--prompts-file",
-        type=str,
-        default="benchmark/build_dataset/top100.txt",
-        help="Path to the prompts file (one prompt per line)",
-    )
-    parser.add_argument(
-        "--output-dir", type=str, default="benchmark_results", help="Directory to save benchmark results"
-    )
-    parser.add_argument("--model-path", type=str, default=MODEL_PATH, help="Path to the model")
-    parser.add_argument("--speaker", type=str, default="Ethan", help="Speaker voice for audio output")
-    parser.add_argument("--num-prompts", type=int, default=None, help="Number of prompts to process (default: all)")
-    args = parser.parse_args()
-
-    # Load model and processor
-    print(f"Loading model from {args.model_path}...")
-    model = Qwen3OmniMoeForConditionalGenerationWithLogging.from_pretrained(
-        args.model_path,
-        dtype="auto",
-        device_map="auto",
-        attn_implementation="flash_attention_2",
-    )
-    processor = Qwen3OmniMoeProcessor.from_pretrained(args.model_path)
-
-    # Benchmark mode
-    print(f"Loading prompts from {args.prompts_file}...")
-    prompts = load_prompts(args.prompts_file)
-
-    if args.num_prompts:
-        prompts = prompts[: args.num_prompts]
-
-    print(f"Running benchmark on {len(prompts)} prompts...")
-
-    aggregated_stats, results, audio_outputs = run_benchmark(
-        model=model,
-        processor=processor,
-        prompts=prompts,
-        output_dir=args.output_dir,
-        speaker=args.speaker,
-    )
-
-    print_stats(aggregated_stats)
-    print(f"\nResults saved to {args.output_dir}/")
-
-
-if __name__ == "__main__":
-    main()

benchmarks/qwen3-omni/vllm_omni/eval_qwen3_moe_omni.sh

@@ -1,37 +0,0 @@
-#!/bin/bash
-# Qwen3-Omni Benchmark Evaluation Script
-# This script must be run from the vllm-omni root directory
-
-# Get the directory where this script is located
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-
-# Navigate to vllm-omni root directory (4 levels up from script location)
-VLLM_OMNI_ROOT="$(cd "$SCRIPT_DIR/../../.." && pwd)"
-cd "$VLLM_OMNI_ROOT" || { echo "Error: Failed to navigate to vllm-omni directory"; exit 1; }
-
-echo "Working directory: $(pwd)"
-
-# Verify we're in the correct directory and run benchmark
-if [[ ! -d "benchmarks/qwen3-omni/vllm_omni" ]]; then
-    echo "Error: Not in vllm-omni root directory. Please run from vllm-omni folder."
-else
-    log_dir=benchmarks/qwen3-omni/vllm_omni/logs
-    outputs_dir=benchmarks/qwen3-omni/vllm_omni/outputs
-    end2end_script_path=examples/offline_inference/qwen3_omni/end2end.py
-    build_dataset_path=benchmarks/build_dataset/top100.txt
-
-    python $end2end_script_path --output-wav $outputs_dir \
-                      --query-type text \
-                      --txt-prompts $build_dataset_path \
-                      --log-stats \
-                      --log-dir $log_dir
-    echo "Logs and outputs are saved in ${log_dir} and ${outputs_dir} respectively:"
-    echo "  - omni_pipeline_text                       run dir/base name"
-    echo "  - omni_pipeline_text.orchestrator.stats.jsonl  orchestrator-stage latency stats"
-    echo "  - omni_pipeline_text.overall.stats.jsonl       overall latency/TPS stats"
-    echo "  - omni_pipeline_text.stage0.log                per-stage detailed logs"
-    echo "  - omni_pipeline_text.stage1.log"
-    echo "  - omni_pipeline_text.stage2.log"
-    echo "Key checks: overall.stats.jsonl for end-to-end latency/TPS; orchestrator.stats.jsonl for stable per-stage latency; stage*.log for errors or long tails."
-    echo "  - outputs/             Generated txt and wav files, there should be 100 text and wav files generated respectively"
-fi

benchmarks/qwen3-tts/README.md

@@ -1,100 +0,0 @@
-# Qwen3-TTS Benchmark
-
-Benchmarks for Qwen3-TTS text-to-speech models, comparing vLLM-Omni streaming serving against HuggingFace Transformers offline inference.
-
-## Prerequisites
-
-```bash
-pip install matplotlib aiohttp soundfile numpy tqdm
-pip install qwen_tts  # for HF baseline
-```
-
-## Quick Start
-
-Run the full benchmark (vllm-omni + HF baseline) with a single command:
-
-```bash
-cd benchmarks/qwen3-tts
-bash run_benchmark.sh
-```
-
-Results (JSON + PNG plots) are saved to `results/`.
-
-### Common options
-
-```bash
-# Only vllm-omni (skip HF baseline)
-bash run_benchmark.sh --async-only
-
-# Only HF baseline
-bash run_benchmark.sh --hf-only
-
-# Use a different model (e.g. 1.7B)
-MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice bash run_benchmark.sh --async-only
-
-# Use bs16 config for higher throughput
-STAGE_CONFIG=vllm_omni/configs/qwen3_tts_bs16.yaml bash run_benchmark.sh --async-only
-
-# Custom GPU, prompt count, concurrency levels
-GPU_DEVICE=1 NUM_PROMPTS=20 CONCURRENCY="1 4" bash run_benchmark.sh
-```
-
-## Manual Steps
-
-### 1) Start the vLLM-Omni server
-
-```bash
-CUDA_VISIBLE_DEVICES=0 python -m vllm_omni.entrypoints.cli.main serve \
-    "Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice" \
-    --omni --host 127.0.0.1 --port 8000 \
-    --stage-configs-path benchmarks/qwen3-tts/vllm_omni/configs/qwen3_tts_bs1.yaml \
-    --trust-remote-code
-```
-
-### 2) Run online serving benchmark
-
-```bash
-python benchmarks/qwen3-tts/vllm_omni/bench_tts_serve.py \
-    --port 8000 \
-    --num-prompts 50 \
-    --max-concurrency 1 4 10 \
-    --config-name "async_chunk" \
-    --result-dir results/
-```
-
-### 3) Run HuggingFace baseline
-
-```bash
-python benchmarks/qwen3-tts/transformers/bench_tts_hf.py \
-    --model "Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice" \
-    --num-prompts 50 \
-    --gpu-device 0 \
-    --result-dir results/
-```
-
-### 4) Generate comparison plots
-
-```bash
-python benchmarks/qwen3-tts/plot_results.py \
-    --results results/bench_async_chunk_*.json results/bench_hf_transformers_*.json \
-    --labels "vllm-omni" "hf_transformers" \
-    --output results/comparison.png
-```
-
-## Stage Configs
-
-| Config | max_num_seqs | Description |
-|--------|:------------:|-------------|
-| `vllm_omni/configs/qwen3_tts_bs1.yaml` | 1 | Single-request processing (lowest latency) |
-| `vllm_omni/configs/qwen3_tts_bs16.yaml` | 16 | High-throughput concurrent processing |
-
-All configs use a 2-stage pipeline (Talker -> Code2Wav) with `async_chunk` streaming enabled. The `SharedMemoryConnector` streams codec frames (25-frame chunks with 25-frame context overlap) between stages.
-
-The model is specified via the CLI `--model` flag (or `MODEL` env var), so the same configs work for both the 0.6B and 1.7B model variants.
-
-## Metrics
-
-- **TTFP (Time to First Audio Packet)**: Time from request to first audio chunk (streaming latency)
-- **E2E (End-to-End Latency)**: Total time from request to complete audio response
-- **RTF (Real-Time Factor)**: E2E latency / audio duration. RTF < 1.0 means faster-than-real-time synthesis
-- **Throughput**: Total audio seconds generated per wall-clock second

benchmarks/qwen3-tts/plot_results.py

@@ -1,254 +0,0 @@
-"""Plot Qwen3-TTS benchmark results.
-
-Generates comparison bar charts similar to the async_chunk design doc:
-- TTFP (Time-to-First-Packet) across concurrency levels
-- E2E latency across concurrency levels
-- RTF (Real-Time Factor) across concurrency levels
-
-Usage:
-    # Compare two configs (async_chunk vs no_async_chunk):
-    python plot_results.py \
-        --results results/bench_async_chunk_*.json results/bench_no_async_chunk_*.json \
-        --labels "async_chunk" "no_async_chunk" \
-        --output results/qwen3_tts_benchmark.png
-
-    # Single config:
-    python plot_results.py \
-        --results results/bench_async_chunk_*.json \
-        --labels "async_chunk" \
-        --output results/qwen3_tts_benchmark.png
-"""
-
-import argparse
-import json
-from pathlib import Path
-
-import matplotlib.pyplot as plt
-import numpy as np
-
-
-def load_results(result_files: list[str]) -> list[list[dict]]:
-    """Load benchmark results from JSON files."""
-    all_results = []
-    for f in result_files:
-        with open(f) as fh:
-            data = json.load(fh)
-        all_results.append(data)
-    return all_results
-
-
-def plot_comparison(
-    all_results: list[list[dict]],
-    labels: list[str],
-    output_path: str,
-    title_prefix: str = "Qwen3-TTS",
-):
-    """Generate comparison bar charts."""
-    n_configs = len(all_results)
-
-    # Collect concurrency levels present in ALL configs (skip missing data)
-    all_concurrencies = [set(r["concurrency"] for r in results) for results in all_results]
-    concurrencies = sorted(set.union(*all_concurrencies))
-
-    # Build data arrays, using None for missing concurrency levels
-    ttfp_data = {label: [] for label in labels}
-    e2e_data = {label: [] for label in labels}
-    rtf_data = {label: [] for label in labels}
-    throughput_data = {label: [] for label in labels}
-
-    for results, label in zip(all_results, labels):
-        conc_map = {r["concurrency"]: r for r in results}
-        for c in concurrencies:
-            r = conc_map.get(c)
-            ttfp_data[label].append(r["mean_ttfp_ms"] if r else None)
-            e2e_data[label].append(r["mean_e2e_ms"] if r else None)
-            rtf_data[label].append(r["mean_rtf"] if r else None)
-            throughput_data[label].append(r["audio_throughput"] if r else None)
-
-    fig, axes = plt.subplots(2, 2, figsize=(14, 10))
-    fig.suptitle(f"{title_prefix} Performance Benchmark", fontsize=16, fontweight="bold")
-
-    x = np.arange(len(concurrencies))
-    width = 0.35 if n_configs == 2 else 0.5
-    if n_configs > 1:
-        offsets = np.linspace(-width / 2 * (n_configs - 1), width / 2 * (n_configs - 1), n_configs)
-    else:
-        offsets = [0]
-
-    colors = ["#2196F3", "#FF5722", "#4CAF50", "#FFC107"]
-
-    def plot_metric(ax, data_dict, ylabel, title, fmt=".1f"):
-        bars = []
-        for i, (label, values) in enumerate(data_dict.items()):
-            # Replace None with 0 for plotting, but track which are missing
-            plot_values = [v if v is not None else 0 for v in values]
-            color = colors[i % len(colors)]
-            bar = ax.bar(x + offsets[i], plot_values, width, label=label, color=color, alpha=0.85)
-            bars.append(bar)
-            # Add value labels on bars (skip None/missing data)
-            max_val = max((v for v in values if v is not None), default=1)
-            for rect, val in zip(bar, values):
-                if val is not None and val > 0:
-                    ax.text(
-                        rect.get_x() + rect.get_width() / 2,
-                        rect.get_height() + max_val * 0.02,
-                        f"{val:{fmt}}",
-                        ha="center",
-                        va="bottom",
-                        fontsize=9,
-                        fontweight="bold",
-                    )
-        ax.set_xlabel("Concurrency", fontsize=12)
-        ax.set_ylabel(ylabel, fontsize=12)
-        ax.set_title(title, fontsize=13, fontweight="bold")
-        ax.set_xticks(x)
-        ax.set_xticklabels([str(c) for c in concurrencies])
-        ax.legend(fontsize=10)
-        ax.grid(axis="y", alpha=0.3)
-        ax.set_axisbelow(True)
-
-    plot_metric(axes[0, 0], ttfp_data, "TTFP (ms)", "Time to First Audio Packet (TTFP)")
-    plot_metric(axes[0, 1], e2e_data, "E2E Latency (ms)", "End-to-End Latency (E2E)")
-    plot_metric(axes[1, 0], rtf_data, "RTF", "Real-Time Factor (RTF)", fmt=".3f")
-    plot_metric(axes[1, 1], throughput_data, "Audio-sec / Wall-sec", "Audio Throughput", fmt=".2f")
-
-    plt.tight_layout()
-    plt.savefig(output_path, dpi=150, bbox_inches="tight")
-    print(f"Plot saved to {output_path}")
-    plt.close()
-
-
-def plot_single_summary(results: list[dict], label: str, output_path: str):
-    """Generate a single-config summary with percentile breakdown."""
-    concurrencies = [r["concurrency"] for r in results]
-
-    fig, axes = plt.subplots(1, 3, figsize=(16, 5))
-    fig.suptitle(f"Qwen3-TTS Benchmark - {label}", fontsize=15, fontweight="bold")
-
-    # TTFP breakdown
-    ax = axes[0]
-    means = [r["mean_ttfp_ms"] for r in results]
-    medians = [r["median_ttfp_ms"] for r in results]
-    p90s = [r["p90_ttfp_ms"] for r in results]
-    p99s = [r["p99_ttfp_ms"] for r in results]
-    x = np.arange(len(concurrencies))
-    w = 0.2
-    ax.bar(x - 1.5 * w, means, w, label="mean", color="#2196F3")
-    ax.bar(x - 0.5 * w, medians, w, label="median", color="#4CAF50")
-    ax.bar(x + 0.5 * w, p90s, w, label="p90", color="#FF9800")
-    ax.bar(x + 1.5 * w, p99s, w, label="p99", color="#F44336")
-    ax.set_xticks(x)
-    ax.set_xticklabels([str(c) for c in concurrencies])
-    ax.set_xlabel("Concurrency")
-    ax.set_ylabel("TTFP (ms)")
-    ax.set_title("Time to First Audio Packet")
-    ax.legend(fontsize=9)
-    ax.grid(axis="y", alpha=0.3)
-
-    # E2E breakdown
-    ax = axes[1]
-    means = [r["mean_e2e_ms"] for r in results]
-    medians = [r["median_e2e_ms"] for r in results]
-    p90s = [r["p90_e2e_ms"] for r in results]
-    p99s = [r["p99_e2e_ms"] for r in results]
-    ax.bar(x - 1.5 * w, means, w, label="mean", color="#2196F3")
-    ax.bar(x - 0.5 * w, medians, w, label="median", color="#4CAF50")
-    ax.bar(x + 0.5 * w, p90s, w, label="p90", color="#FF9800")
-    ax.bar(x + 1.5 * w, p99s, w, label="p99", color="#F44336")
-    ax.set_xticks(x)
-    ax.set_xticklabels([str(c) for c in concurrencies])
-    ax.set_xlabel("Concurrency")
-    ax.set_ylabel("E2E Latency (ms)")
-    ax.set_title("End-to-End Latency")
-    ax.legend(fontsize=9)
-    ax.grid(axis="y", alpha=0.3)
-
-    # RTF
-    ax = axes[2]
-    means = [r["mean_rtf"] for r in results]
-    medians = [r["median_rtf"] for r in results]
-    ax.bar(x - 0.15, means, 0.3, label="mean", color="#2196F3")
-    ax.bar(x + 0.15, medians, 0.3, label="median", color="#4CAF50")
-    ax.set_xticks(x)
-    ax.set_xticklabels([str(c) for c in concurrencies])
-    ax.set_xlabel("Concurrency")
-    ax.set_ylabel("RTF")
-    ax.set_title("Real-Time Factor")
-    ax.legend(fontsize=9)
-    ax.grid(axis="y", alpha=0.3)
-
-    plt.tight_layout()
-    plt.savefig(output_path, dpi=150, bbox_inches="tight")
-    print(f"Plot saved to {output_path}")
-    plt.close()
-
-
-def print_comparison_table(all_results: list[list[dict]], labels: list[str]):
-    """Print a markdown-formatted comparison table."""
-    concurrencies = sorted(set(r["concurrency"] for r in all_results[0]))
-
-    print("\n## Benchmark Results\n")
-    header = "| Metric | Concurrency |"
-    sep = "| --- | --- |"
-    for label in labels:
-        header += f" {label} |"
-        sep += " --- |"
-    print(header)
-    print(sep)
-
-    for metric, key, fmt in [
-        ("TTFP (ms)", "mean_ttfp_ms", ".1f"),
-        ("E2E (ms)", "mean_e2e_ms", ".1f"),
-        ("RTF", "mean_rtf", ".3f"),
-        ("Throughput (audio-s/s)", "audio_throughput", ".2f"),
-    ]:
-        for c in concurrencies:
-            row = f"| {metric} | {c} |"
-            for results in all_results:
-                conc_map = {r["concurrency"]: r for r in results}
-                val = conc_map.get(c, {}).get(key, 0)
-                row += f" {val:{fmt}} |"
-            print(row)
-
-    # Improvement calculation (only if 2 configs)
-    if len(all_results) == 2:
-        print(f"\n## Improvement ({labels[0]} vs {labels[1]})\n")
-        print("| Metric | Concurrency | Improvement |")
-        print("| --- | --- | --- |")
-        for metric, key in [("TTFP", "mean_ttfp_ms"), ("E2E", "mean_e2e_ms"), ("RTF", "mean_rtf")]:
-            for c in concurrencies:
-                m0 = {r["concurrency"]: r for r in all_results[0]}
-                m1 = {r["concurrency"]: r for r in all_results[1]}
-                v0 = m0.get(c, {}).get(key, 0)
-                v1 = m1.get(c, {}).get(key, 0)
-                if v1 > 0:
-                    pct = (v1 - v0) / v1 * 100
-                    print(f"| {metric} | {c} | {pct:+.1f}% |")
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Plot Qwen3-TTS benchmark results")
-    parser.add_argument(
-        "--results", type=str, nargs="+", required=True, help="Path(s) to result JSON files (one per config)"
-    )
-    parser.add_argument(
-        "--labels", type=str, nargs="+", required=True, help="Labels for each config (must match --results count)"
-    )
-    parser.add_argument("--output", type=str, default="results/qwen3_tts_benchmark.png", help="Output image path")
-    parser.add_argument("--title", type=str, default="Qwen3-TTS", help="Title prefix for the plot")
-    return parser.parse_args()
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    assert len(args.results) == len(args.labels), "--results and --labels must have the same count"
-
-    all_results = load_results(args.results)
-    print_comparison_table(all_results, args.labels)
-
-    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
-
-    if len(all_results) == 1:
-        plot_single_summary(all_results[0], args.labels[0], args.output)
-    else:
-        plot_comparison(all_results, args.labels, args.output, title_prefix=args.title)

benchmarks/qwen3-tts/results/.gitignore

@@ -1,3 +0,0 @@
-# Benchmark results are machine-specific - do not commit
-*
-!.gitignore

benchmarks/qwen3-tts/run_benchmark.sh

@@ -1,272 +0,0 @@
-#!/bin/bash
-# Qwen3-TTS Benchmark Runner
-#
-# Compares vllm-omni streaming serving vs HuggingFace transformers offline inference.
-# Produces JSON results and comparison plots.
-#
-# Usage:
-#   # Full comparison (vllm-omni + HF):
-#   bash run_benchmark.sh
-#
-#   # Only vllm-omni async_chunk config:
-#   bash run_benchmark.sh --async-only
-#
-#   # Only HuggingFace baseline:
-#   bash run_benchmark.sh --hf-only
-#
-#   # vllm-omni only (skip HF):
-#   bash run_benchmark.sh --skip-hf
-#
-#   # Custom settings:
-#   GPU_DEVICE=1 NUM_PROMPTS=20 CONCURRENCY="1 4" bash run_benchmark.sh
-#
-#   # Use 1.7B model:
-#   MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice bash run_benchmark.sh --async-only
-#
-#   # Use batch_size=4 config:
-#   STAGE_CONFIG=vllm_omni/configs/qwen3_tts_bs4.yaml bash run_benchmark.sh --async-only
-#
-# Environment variables:
-#   GPU_DEVICE       - GPU index to use (default: 0)
-#   NUM_PROMPTS      - Number of prompts per concurrency level (default: 50)
-#   CONCURRENCY      - Space-separated concurrency levels (default: "1 4 10")
-#   MODEL            - Model name (default: Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice)
-#   PORT             - Server port (default: 8000)
-#   GPU_MEM_TALKER   - gpu_memory_utilization for talker stage (default: 0.3)
-#   GPU_MEM_CODE2WAV - gpu_memory_utilization for code2wav stage (default: 0.2)
-#   STAGE_CONFIG     - Path to stage config YAML (default: configs/qwen3_tts_bs1.yaml)
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-PROJECT_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
-
-# Defaults
-GPU_DEVICE="${GPU_DEVICE:-0}"
-NUM_PROMPTS="${NUM_PROMPTS:-50}"
-CONCURRENCY="${CONCURRENCY:-1 4 10}"
-MODEL="${MODEL:-Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice}"
-PORT="${PORT:-8000}"
-GPU_MEM_TALKER="${GPU_MEM_TALKER:-0.3}"
-GPU_MEM_CODE2WAV="${GPU_MEM_CODE2WAV:-0.2}"
-NUM_WARMUPS="${NUM_WARMUPS:-3}"
-STAGE_CONFIG="${STAGE_CONFIG:-vllm_omni/configs/qwen3_tts_bs1.yaml}"
-RESULT_DIR="${SCRIPT_DIR}/results"
-TIMESTAMP="$(date +%Y%m%d_%H%M%S)"
-
-# Parse args
-RUN_ASYNC=true
-RUN_HF=true
-for arg in "$@"; do
-    case "$arg" in
-        --async-only) RUN_HF=false ;;
-        --hf-only) RUN_ASYNC=false ;;
-        --skip-hf) RUN_HF=false ;;
-    esac
-done
-
-mkdir -p "${RESULT_DIR}"
-
-echo "============================================================"
-echo " Qwen3-TTS Benchmark"
-echo "============================================================"
-echo " GPU:          ${GPU_DEVICE}"
-echo " Model:        ${MODEL}"
-echo " Prompts:      ${NUM_PROMPTS}"
-echo " Concurrency:  ${CONCURRENCY}"
-echo " Port:         ${PORT}"
-echo " Stage config: ${STAGE_CONFIG}"
-echo " Results:      ${RESULT_DIR}"
-echo "============================================================"
-
-# Prepare stage config with correct GPU device and memory settings
-prepare_config() {
-    local config_template="$1"
-    local config_name="$2"
-    local output_path="${RESULT_DIR}/${config_name}_stage_config.yaml"
-
-    # Use sed to patch GPU device and memory utilization
-    sed \
-        -e "s/devices: \"0\"/devices: \"${GPU_DEVICE}\"/g" \
-        -e "s/gpu_memory_utilization: 0.3/gpu_memory_utilization: ${GPU_MEM_TALKER}/g" \
-        -e "s/gpu_memory_utilization: 0.2/gpu_memory_utilization: ${GPU_MEM_CODE2WAV}/g" \
-        "${config_template}" > "${output_path}"
-
-    echo "${output_path}"
-}
-
-# Start server and wait for it to be ready
-start_server() {
-    local stage_config="$1"
-    local config_name="$2"
-    local log_file="${RESULT_DIR}/server_${config_name}_${TIMESTAMP}.log"
-
-    echo ""
-    echo "Starting server with config: ${config_name}"
-    echo "  Stage config: ${stage_config}"
-    echo "  Log file: ${log_file}"
-
-    VLLM_WORKER_MULTIPROC_METHOD=spawn \
-    CUDA_VISIBLE_DEVICES="${GPU_DEVICE}" \
-    python -m vllm_omni.entrypoints.cli.main serve "${MODEL}" \
-        --omni \
-        --host 127.0.0.1 \
-        --port "${PORT}" \
-        --stage-configs-path "${stage_config}" \
-        --stage-init-timeout 120 \
-        --trust-remote-code \
-        --disable-log-stats \
-        > "${log_file}" 2>&1 &
-
-    SERVER_PID=$!
-    echo "  Server PID: ${SERVER_PID}"
-
-    # Wait for server to be ready
-    echo "  Waiting for server to be ready..."
-    local max_wait=300
-    local waited=0
-    while [ ${waited} -lt ${max_wait} ]; do
-        if curl -sf "http://127.0.0.1:${PORT}/v1/models" > /dev/null 2>&1; then
-            echo "  Server is ready! (waited ${waited}s)"
-            return 0
-        fi
-        # Check if process is still alive
-        if ! kill -0 ${SERVER_PID} 2>/dev/null; then
-            echo "  ERROR: Server process died. Check log: ${log_file}"
-            tail -20 "${log_file}"
-            return 1
-        fi
-        sleep 2
-        waited=$((waited + 2))
-    done
-
-    echo "  ERROR: Server did not start within ${max_wait}s. Check log: ${log_file}"
-    kill ${SERVER_PID} 2>/dev/null || true
-    return 1
-}
-
-# Stop the server
-stop_server() {
-    if [ -n "${SERVER_PID:-}" ]; then
-        echo "  Stopping server (PID: ${SERVER_PID})..."
-        kill ${SERVER_PID} 2>/dev/null || true
-        wait ${SERVER_PID} 2>/dev/null || true
-        # Kill any remaining child processes on the port
-        local pids
-        pids=$(lsof -ti:${PORT} 2>/dev/null || true)
-        if [ -n "${pids}" ]; then
-            echo "  Cleaning up remaining processes on port ${PORT}..."
-            echo "${pids}" | xargs kill -9 2>/dev/null || true
-        fi
-        echo "  Server stopped."
-        SERVER_PID=""
-    fi
-}
-
-# Cleanup on exit
-trap 'stop_server' EXIT
-
-# Run benchmark for a given config
-run_bench() {
-    local config_name="$1"
-    local config_template="$2"
-
-    echo ""
-    echo "============================================================"
-    echo " Benchmarking: ${config_name}"
-    echo "============================================================"
-
-    local stage_config
-    stage_config=$(prepare_config "${config_template}" "${config_name}")
-
-    start_server "${stage_config}" "${config_name}"
-
-    # Convert concurrency string to args
-    local conc_args=""
-    for c in ${CONCURRENCY}; do
-        conc_args="${conc_args} ${c}"
-    done
-
-    cd "${PROJECT_ROOT}"
-    python "${SCRIPT_DIR}/vllm_omni/bench_tts_serve.py" \
-        --host 127.0.0.1 \
-        --port "${PORT}" \
-        --num-prompts "${NUM_PROMPTS}" \
-        --max-concurrency ${conc_args} \
-        --num-warmups "${NUM_WARMUPS}" \
-        --config-name "${config_name}" \
-        --result-dir "${RESULT_DIR}"
-
-    stop_server
-
-    # Allow GPU memory to settle
-    sleep 5
-}
-
-# Run vllm-omni benchmark
-if [ "${RUN_ASYNC}" = true ]; then
-    run_bench "async_chunk" "${SCRIPT_DIR}/${STAGE_CONFIG}"
-fi
-
-# Run HuggingFace baseline benchmark
-if [ "${RUN_HF}" = true ]; then
-    echo ""
-    echo "============================================================"
-    echo " Benchmarking: HuggingFace transformers (offline)"
-    echo "============================================================"
-
-    cd "${PROJECT_ROOT}"
-    python "${SCRIPT_DIR}/transformers/bench_tts_hf.py" \
-        --model "${MODEL}" \
-        --num-prompts "${NUM_PROMPTS}" \
-        --num-warmups "${NUM_WARMUPS}" \
-        --gpu-device "${GPU_DEVICE}" \
-        --config-name "hf_transformers" \
-        --result-dir "${RESULT_DIR}"
-
-    # Allow GPU memory to settle
-    sleep 5
-fi
-
-# Plot results
-echo ""
-echo "============================================================"
-echo " Generating plots..."
-echo "============================================================"
-
-RESULT_FILES=""
-LABELS=""
-
-if [ "${RUN_ASYNC}" = true ]; then
-    ASYNC_FILE=$(ls -t "${RESULT_DIR}"/bench_async_chunk_*.json 2>/dev/null | head -1)
-    if [ -n "${ASYNC_FILE}" ]; then
-        RESULT_FILES="${ASYNC_FILE}"
-        LABELS="async_chunk"
-    fi
-fi
-
-if [ "${RUN_HF}" = true ]; then
-    HF_FILE=$(ls -t "${RESULT_DIR}"/bench_hf_transformers_*.json 2>/dev/null | head -1)
-    if [ -n "${HF_FILE}" ]; then
-        if [ -n "${RESULT_FILES}" ]; then
-            RESULT_FILES="${RESULT_FILES} ${HF_FILE}"
-            LABELS="${LABELS} hf_transformers"
-        else
-            RESULT_FILES="${HF_FILE}"
-            LABELS="hf_transformers"
-        fi
-    fi
-fi
-
-if [ -n "${RESULT_FILES}" ]; then
-    python "${SCRIPT_DIR}/plot_results.py" \
-        --results ${RESULT_FILES} \
-        --labels ${LABELS} \
-        --output "${RESULT_DIR}/qwen3_tts_benchmark_${TIMESTAMP}.png"
-fi
-
-echo ""
-echo "============================================================"
-echo " Benchmark complete!"
-echo " Results: ${RESULT_DIR}"
-echo "============================================================"

benchmarks/qwen3-tts/transformers/bench_tts_hf.py

@@ -1,281 +0,0 @@
-"""Benchmark Qwen3-TTS using HuggingFace transformers (qwen_tts library).
-
-Measures E2E latency, RTF, and audio duration for offline (non-serving) inference.
-Results are saved in the same JSON format as bench_tts_serve.py for unified plotting.
-
-Usage:
-    python bench_tts_hf.py \
-        --model Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice \
-        --num-prompts 50 \
-        --num-warmups 3 \
-        --gpu-device 0 \
-        --result-dir results/
-"""
-
-import argparse
-import json
-import time
-from dataclasses import asdict, dataclass, field
-from datetime import datetime
-from pathlib import Path
-
-import numpy as np
-import soundfile as sf
-import torch
-
-PROMPTS = [
-    "Hello, welcome to the voice synthesis benchmark test.",
-    "She said she would be here by noon, but nobody showed up.",
-    "The quick brown fox jumps over the lazy dog near the riverbank.",
-    "I can't believe how beautiful the sunset looks from up here on the mountain.",
-    "Please remember to bring your identification documents to the appointment tomorrow morning.",
-    "Have you ever wondered what it would be like to travel through time and visit ancient civilizations?",
-    "The restaurant on the corner serves the best pasta I have ever tasted in my entire life.",
-    "After the meeting, we should discuss the quarterly results and plan for the next phase.",
-    "Learning a new language takes patience, practice, and a genuine curiosity about other cultures.",
-    "The train leaves at half past seven, so we need to arrive at the station before then.",
-    "Could you please turn down the music a little bit, I'm trying to concentrate on my work.",
-    "It was a dark and stormy night when the old lighthouse keeper heard a knock at the door.",
-]
-
-
-@dataclass
-class BenchmarkResult:
-    config_name: str = ""
-    concurrency: int = 1  # always 1 for offline
-    num_prompts: int = 0
-    completed: int = 0
-    failed: int = 0
-    duration_s: float = 0.0
-    # TTFP stats - not applicable for HF offline, set to E2E for compatibility
-    mean_ttfp_ms: float = 0.0
-    median_ttfp_ms: float = 0.0
-    std_ttfp_ms: float = 0.0
-    p90_ttfp_ms: float = 0.0
-    p95_ttfp_ms: float = 0.0
-    p99_ttfp_ms: float = 0.0
-    # E2E stats (ms)
-    mean_e2e_ms: float = 0.0
-    median_e2e_ms: float = 0.0
-    std_e2e_ms: float = 0.0
-    p90_e2e_ms: float = 0.0
-    p95_e2e_ms: float = 0.0
-    p99_e2e_ms: float = 0.0
-    # RTF stats
-    mean_rtf: float = 0.0
-    median_rtf: float = 0.0
-    std_rtf: float = 0.0
-    p99_rtf: float = 0.0
-    # Audio stats
-    mean_audio_duration_s: float = 0.0
-    total_audio_duration_s: float = 0.0
-    audio_throughput: float = 0.0
-    request_throughput: float = 0.0
-    # Per-request details
-    per_request: list = field(default_factory=list)
-
-
-def run_benchmark(args):
-    from qwen_tts import Qwen3TTSModel
-
-    device = f"cuda:{args.gpu_device}"
-    print(f"Loading model: {args.model} on {device}")
-    model = Qwen3TTSModel.from_pretrained(
-        args.model,
-        device_map=device,
-        dtype=torch.bfloat16,
-    )
-    print("Model loaded.")
-
-    # Build prompt list
-    prompts = [PROMPTS[i % len(PROMPTS)] for i in range(args.num_prompts)]
-
-    # Warmup
-    if args.num_warmups > 0:
-        print(f"Warming up with {args.num_warmups} requests...")
-        for i in range(args.num_warmups):
-            p = PROMPTS[i % len(PROMPTS)]
-            wavs, sr = model.generate_custom_voice(
-                text=p,
-                language=args.language,
-                speaker=args.voice,
-            )
-        # Sync GPU
-        torch.cuda.synchronize(device)
-        print("Warmup done.")
-
-    # Benchmark
-    print(f"Running {args.num_prompts} requests sequentially...")
-    e2e_times = []
-    rtfs = []
-    audio_durations = []
-    per_request = []
-    failed = 0
-
-    audio_dir = None
-    if args.save_audio:
-        audio_dir = Path(args.result_dir) / "audio_hf"
-        audio_dir.mkdir(parents=True, exist_ok=True)
-
-    total_start = time.perf_counter()
-
-    for i, prompt in enumerate(prompts):
-        try:
-            torch.cuda.synchronize(device)
-            st = time.perf_counter()
-
-            wavs, sr = model.generate_custom_voice(
-                text=prompt,
-                language=args.language,
-                speaker=args.voice,
-            )
-
-            torch.cuda.synchronize(device)
-            elapsed = time.perf_counter() - st
-
-            # Compute audio duration
-            audio_samples = wavs[0]
-            if isinstance(audio_samples, torch.Tensor):
-                audio_samples = audio_samples.cpu().numpy()
-            audio_dur = len(audio_samples) / sr
-
-            rtf = elapsed / audio_dur if audio_dur > 0 else 0.0
-
-            e2e_times.append(elapsed)
-            rtfs.append(rtf)
-            audio_durations.append(audio_dur)
-            per_request.append(
-                {
-                    "e2e_ms": elapsed * 1000,
-                    "ttfp_ms": elapsed * 1000,  # no streaming, TTFP = E2E
-                    "rtf": rtf,
-                    "audio_duration_s": audio_dur,
-                    "prompt": prompt,
-                }
-            )
-
-            if audio_dir:
-                sf.write(str(audio_dir / f"output_{i:04d}.wav"), audio_samples, sr)
-
-            if (i + 1) % 10 == 0 or i == 0:
-                print(
-                    f"  [{i + 1}/{args.num_prompts}] e2e={elapsed * 1000:.0f}ms  rtf={rtf:.3f}  audio={audio_dur:.2f}s"
-                )
-
-        except Exception as e:
-            print(f"  [{i + 1}/{args.num_prompts}] FAILED: {e}")
-            failed += 1
-
-    total_duration = time.perf_counter() - total_start
-    completed = len(e2e_times)
-
-    # Compute stats
-    result = BenchmarkResult(
-        config_name=args.config_name,
-        concurrency=1,
-        num_prompts=args.num_prompts,
-        completed=completed,
-        failed=failed,
-        duration_s=total_duration,
-    )
-
-    if e2e_times:
-        e2e_ms = [t * 1000 for t in e2e_times]
-
-        result.mean_e2e_ms = float(np.mean(e2e_ms))
-        result.median_e2e_ms = float(np.median(e2e_ms))
-        result.std_e2e_ms = float(np.std(e2e_ms))
-        result.p90_e2e_ms = float(np.percentile(e2e_ms, 90))
-        result.p95_e2e_ms = float(np.percentile(e2e_ms, 95))
-        result.p99_e2e_ms = float(np.percentile(e2e_ms, 99))
-
-        # For HF offline, TTFP = E2E (no streaming)
-        result.mean_ttfp_ms = result.mean_e2e_ms
-        result.median_ttfp_ms = result.median_e2e_ms
-        result.std_ttfp_ms = result.std_e2e_ms
-        result.p90_ttfp_ms = result.p90_e2e_ms
-        result.p95_ttfp_ms = result.p95_e2e_ms
-        result.p99_ttfp_ms = result.p99_e2e_ms
-
-        result.mean_rtf = float(np.mean(rtfs))
-        result.median_rtf = float(np.median(rtfs))
-        result.std_rtf = float(np.std(rtfs))
-        result.p99_rtf = float(np.percentile(rtfs, 99))
-
-        result.mean_audio_duration_s = float(np.mean(audio_durations))
-        result.total_audio_duration_s = float(np.sum(audio_durations))
-        result.audio_throughput = result.total_audio_duration_s / total_duration
-        result.request_throughput = completed / total_duration
-        result.per_request = per_request
-
-    # Print summary in standardized performance template
-    W = 50
-    print("")
-    print(f"{'=' * W}")
-    print(f"{'Serving Benchmark Result':^{W}}")
-    print(f"{'=' * W}")
-    print(f"{'Successful requests:':<40}{completed:<10}")
-    print(f"{'Failed requests:':<40}{failed:<10}")
-    print(f"{'Maximum request concurrency:':<40}{1:<10}")
-    print(f"{'Benchmark duration (s):':<40}{total_duration:<10.2f}")
-    print(f"{'Request throughput (req/s):':<40}{result.request_throughput:<10.2f}")
-    print(f"{'-' * W}")
-    print(f"{'End-to-end Latency':^{W}}")
-    print(f"{'-' * W}")
-    print(f"{'Mean E2EL (ms):':<40}{result.mean_e2e_ms:<10.2f}")
-    print(f"{'Median E2EL (ms):':<40}{result.median_e2e_ms:<10.2f}")
-    print(f"{'P99 E2EL (ms):':<40}{result.p99_e2e_ms:<10.2f}")
-    print(f"{'=' * W}")
-    print(f"{'Audio Result':^{W}}")
-    print(f"{'=' * W}")
-    print(f"{'Total audio duration generated (s):':<40}{result.total_audio_duration_s:<10.2f}")
-    print(f"{'Audio throughput (audio duration/s):':<40}{result.audio_throughput:<10.2f}")
-    print(f"{'-' * W}")
-    print(f"{'Time to First Packet':^{W}}")
-    print(f"{'-' * W}")
-    print(f"{'Mean AUDIO_TTFP (ms):':<40}{result.mean_ttfp_ms:<10.2f}")
-    print(f"{'Median AUDIO_TTFP (ms):':<40}{result.median_ttfp_ms:<10.2f}")
-    print(f"{'P99 AUDIO_TTFP (ms):':<40}{result.p99_ttfp_ms:<10.2f}")
-    print(f"{'-' * W}")
-    print(f"{'Real Time Factor':^{W}}")
-    print(f"{'-' * W}")
-    print(f"{'Mean AUDIO_RTF:':<40}{result.mean_rtf:<10.3f}")
-    print(f"{'Median AUDIO_RTF:':<40}{result.median_rtf:<10.3f}")
-    print(f"{'P99 AUDIO_RTF:':<40}{result.p99_rtf:<10.3f}")
-    print(f"{'=' * W}")
-    print("")
-
-    # Save results (as a list with single concurrency=1 entry, matching serve format)
-    result_dir = Path(args.result_dir)
-    result_dir.mkdir(parents=True, exist_ok=True)
-    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
-    result_file = result_dir / f"bench_{args.config_name}_{timestamp}.json"
-
-    with open(result_file, "w") as f:
-        json.dump([asdict(result)], f, indent=2)
-    print(f"Results saved to {result_file}")
-
-    return result
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Qwen3-TTS HuggingFace Benchmark")
-    parser.add_argument(
-        "--model", type=str, default="Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice", help="HuggingFace model name or path"
-    )
-    parser.add_argument("--num-prompts", type=int, default=50)
-    parser.add_argument("--num-warmups", type=int, default=3)
-    parser.add_argument("--gpu-device", type=int, default=0)
-    parser.add_argument("--voice", type=str, default="Vivian")
-    parser.add_argument("--language", type=str, default="English")
-    parser.add_argument(
-        "--config-name", type=str, default="hf_transformers", help="Label for this config (used in filenames)"
-    )
-    parser.add_argument("--result-dir", type=str, default="results")
-    parser.add_argument("--save-audio", action="store_true", help="Save generated audio files")
-    return parser.parse_args()
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    run_benchmark(args)

benchmarks/qwen3-tts/vllm_omni/bench_async_chunk.py

@@ -1,301 +0,0 @@
-"""Benchmark comparing async_chunk on vs off for Qwen3-TTS.
-
-Measures TTFP (Time-to-First-Packet), E2E latency, and RTF across
-concurrency levels for both async_chunk modes. Saves results as JSON.
-
-Usage:
-    # Run against a server already serving with a given config:
-    python bench_async_chunk.py \
-        --host 127.0.0.1 --port 8000 \
-        --config-name async_chunk_on \
-        --num-prompts 50 \
-        --max-concurrency 1 10 \
-        --result-dir results/
-"""
-
-import argparse
-import asyncio
-import json
-import time
-from dataclasses import asdict, dataclass, field
-from datetime import datetime
-from pathlib import Path
-
-import aiohttp
-import numpy as np
-from tqdm.asyncio import tqdm
-
-PROMPTS = [
-    "Hello, welcome to the voice synthesis benchmark test.",
-    "She said she would be here by noon, but nobody showed up.",
-    "The quick brown fox jumps over the lazy dog near the riverbank.",
-    "I can't believe how beautiful the sunset looks from up here on the mountain.",
-    "Please remember to bring your identification documents to the appointment tomorrow morning.",
-    "Have you ever wondered what it would be like to travel through time and visit ancient civilizations?",
-    "The restaurant on the corner serves the best pasta I have ever tasted in my entire life.",
-    "After the meeting, we should discuss the quarterly results and plan for the next phase.",
-    "Learning a new language takes patience, practice, and a genuine curiosity about other cultures.",
-    "The train leaves at half past seven, so we need to arrive at the station before then.",
-    "Could you please turn down the music a little bit, I'm trying to concentrate on my work.",
-    "It was a dark and stormy night when the old lighthouse keeper heard a knock at the door.",
-]
-
-
-@dataclass
-class RequestResult:
-    success: bool = False
-    ttfp: float = 0.0
-    e2e: float = 0.0
-    audio_bytes: int = 0
-    audio_duration: float = 0.0
-    rtf: float = 0.0
-    prompt: str = ""
-    error: str = ""
-
-
-@dataclass
-class BenchmarkResult:
-    config_name: str = ""
-    concurrency: int = 0
-    num_prompts: int = 0
-    completed: int = 0
-    failed: int = 0
-    duration_s: float = 0.0
-    mean_ttfp_ms: float = 0.0
-    median_ttfp_ms: float = 0.0
-    std_ttfp_ms: float = 0.0
-    p90_ttfp_ms: float = 0.0
-    p95_ttfp_ms: float = 0.0
-    p99_ttfp_ms: float = 0.0
-    mean_e2e_ms: float = 0.0
-    median_e2e_ms: float = 0.0
-    std_e2e_ms: float = 0.0
-    p90_e2e_ms: float = 0.0
-    p95_e2e_ms: float = 0.0
-    p99_e2e_ms: float = 0.0
-    mean_rtf: float = 0.0
-    median_rtf: float = 0.0
-    std_rtf: float = 0.0
-    mean_audio_duration_s: float = 0.0
-    total_audio_duration_s: float = 0.0
-    audio_throughput: float = 0.0
-    request_throughput: float = 0.0
-    per_request: list = field(default_factory=list)
-
-
-def pcm_bytes_to_duration(num_bytes: int, sample_rate: int = 24000, sample_width: int = 2) -> float:
-    return num_bytes / sample_width / sample_rate
-
-
-async def send_tts_request(
-    session: aiohttp.ClientSession,
-    api_url: str,
-    prompt: str,
-    voice: str = "vivian",
-    language: str = "English",
-    stream: bool = True,
-    pbar: tqdm | None = None,
-) -> RequestResult:
-    payload = {
-        "input": prompt,
-        "voice": voice,
-        "language": language,
-        "stream": stream,
-        "response_format": "pcm",
-    }
-
-    result = RequestResult(prompt=prompt)
-    st = time.perf_counter()
-
-    try:
-        async with session.post(api_url, json=payload) as response:
-            if response.status != 200:
-                result.error = f"HTTP {response.status}: {await response.text()}"
-                return result
-
-            first_chunk = True
-            total_bytes = 0
-
-            async for chunk in response.content.iter_any():
-                if first_chunk and len(chunk) > 0:
-                    result.ttfp = time.perf_counter() - st
-                    first_chunk = False
-                total_bytes += len(chunk)
-
-            result.e2e = time.perf_counter() - st
-            result.audio_bytes = total_bytes
-            result.audio_duration = pcm_bytes_to_duration(total_bytes)
-            if result.audio_duration > 0:
-                result.rtf = result.e2e / result.audio_duration
-            result.success = True
-
-    except Exception as e:
-        result.error = str(e)
-        result.e2e = time.perf_counter() - st
-
-    if pbar:
-        pbar.update(1)
-    return result
-
-
-async def run_benchmark(
-    host: str,
-    port: int,
-    num_prompts: int,
-    max_concurrency: int,
-    num_warmups: int = 3,
-    voice: str = "vivian",
-    language: str = "English",
-    stream: bool = True,
-) -> BenchmarkResult:
-    api_url = f"http://{host}:{port}/v1/audio/speech"
-
-    connector = aiohttp.TCPConnector(limit=max_concurrency, limit_per_host=max_concurrency, keepalive_timeout=60)
-    session = aiohttp.ClientSession(connector=connector, timeout=aiohttp.ClientTimeout(total=600))
-
-    if num_warmups > 0:
-        print(f"  Warming up with {num_warmups} requests...")
-        warmup_tasks = [
-            send_tts_request(session, api_url, PROMPTS[i % len(PROMPTS)], voice, language, stream)
-            for i in range(num_warmups)
-        ]
-        await asyncio.gather(*warmup_tasks)
-        print("  Warmup done.")
-
-    request_prompts = [PROMPTS[i % len(PROMPTS)] for i in range(num_prompts)]
-
-    print(f"  Running {num_prompts} requests with concurrency={max_concurrency}...")
-    semaphore = asyncio.Semaphore(max_concurrency)
-    pbar = tqdm(total=num_prompts, desc=f"  concurrency={max_concurrency}")
-
-    async def limited_request(prompt):
-        async with semaphore:
-            return await send_tts_request(session, api_url, prompt, voice, language, stream, pbar)
-
-    start_time = time.perf_counter()
-    tasks = [asyncio.create_task(limited_request(p)) for p in request_prompts]
-    results: list[RequestResult] = await asyncio.gather(*tasks)
-    duration = time.perf_counter() - start_time
-    pbar.close()
-
-    await session.close()
-
-    successful = [r for r in results if r.success]
-    failed = [r for r in results if not r.success]
-
-    bench = BenchmarkResult(
-        concurrency=max_concurrency,
-        num_prompts=num_prompts,
-        completed=len(successful),
-        failed=len(failed),
-        duration_s=duration,
-    )
-
-    if successful:
-        ttfps = [r.ttfp * 1000 for r in successful]
-        e2es = [r.e2e * 1000 for r in successful]
-        rtfs = [r.rtf for r in successful]
-        audio_durs = [r.audio_duration for r in successful]
-
-        bench.mean_ttfp_ms = float(np.mean(ttfps))
-        bench.median_ttfp_ms = float(np.median(ttfps))
-        bench.std_ttfp_ms = float(np.std(ttfps))
-        bench.p90_ttfp_ms = float(np.percentile(ttfps, 90))
-        bench.p95_ttfp_ms = float(np.percentile(ttfps, 95))
-        bench.p99_ttfp_ms = float(np.percentile(ttfps, 99))
-
-        bench.mean_e2e_ms = float(np.mean(e2es))
-        bench.median_e2e_ms = float(np.median(e2es))
-        bench.std_e2e_ms = float(np.std(e2es))
-        bench.p90_e2e_ms = float(np.percentile(e2es, 90))
-        bench.p95_e2e_ms = float(np.percentile(e2es, 95))
-        bench.p99_e2e_ms = float(np.percentile(e2es, 99))
-
-        bench.mean_rtf = float(np.mean(rtfs))
-        bench.median_rtf = float(np.median(rtfs))
-        bench.std_rtf = float(np.std(rtfs))
-
-        bench.mean_audio_duration_s = float(np.mean(audio_durs))
-        bench.total_audio_duration_s = float(np.sum(audio_durs))
-        bench.audio_throughput = bench.total_audio_duration_s / duration
-        bench.request_throughput = len(successful) / duration
-
-        bench.per_request = [
-            {
-                "ttfp_ms": r.ttfp * 1000,
-                "e2e_ms": r.e2e * 1000,
-                "rtf": r.rtf,
-                "audio_duration_s": r.audio_duration,
-                "prompt": r.prompt,
-            }
-            for r in successful
-        ]
-
-    print(f"\n{'=' * 60}")
-    print(f"  Concurrency: {max_concurrency}  |  Completed: {bench.completed}  |  Failed: {bench.failed}")
-    print(f"  Duration: {duration:.2f}s  |  Throughput: {bench.request_throughput:.2f} req/s")
-    print(
-        f"  TTFP (ms):  mean={bench.mean_ttfp_ms:.1f}  median={bench.median_ttfp_ms:.1f}"
-        f"  p90={bench.p90_ttfp_ms:.1f}  p99={bench.p99_ttfp_ms:.1f}"
-    )
-    print(
-        f"  E2E (ms):   mean={bench.mean_e2e_ms:.1f}  median={bench.median_e2e_ms:.1f}"
-        f"  p90={bench.p90_e2e_ms:.1f}  p99={bench.p99_e2e_ms:.1f}"
-    )
-    print(f"  RTF:        mean={bench.mean_rtf:.3f}  median={bench.median_rtf:.3f}")
-    print(f"  Throughput: {bench.audio_throughput:.2f} audio-sec/wall-sec")
-    print(f"{'=' * 60}\n")
-
-    if failed:
-        for r in failed[:3]:
-            print(f"  [ERROR] {r.error[:200]}")
-
-    return bench
-
-
-async def main(args):
-    all_results = []
-
-    for concurrency in args.max_concurrency:
-        result = await run_benchmark(
-            host=args.host,
-            port=args.port,
-            num_prompts=args.num_prompts,
-            max_concurrency=concurrency,
-            num_warmups=args.num_warmups,
-            voice=args.voice,
-            language=args.language,
-            stream=args.stream,
-        )
-        result.config_name = args.config_name
-        all_results.append(asdict(result))
-
-    result_dir = Path(args.result_dir)
-    result_dir.mkdir(parents=True, exist_ok=True)
-    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
-    result_file = result_dir / f"bench_{args.config_name}_{timestamp}.json"
-
-    with open(result_file, "w") as f:
-        json.dump(all_results, f, indent=2)
-    print(f"Results saved to {result_file}")
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Qwen3-TTS async_chunk benchmark client")
-    parser.add_argument("--host", type=str, default="127.0.0.1")
-    parser.add_argument("--port", type=int, default=8000)
-    parser.add_argument("--num-prompts", type=int, default=50)
-    parser.add_argument("--max-concurrency", type=int, nargs="+", default=[1, 10])
-    parser.add_argument("--num-warmups", type=int, default=3)
-    parser.add_argument("--voice", type=str, default="vivian")
-    parser.add_argument("--language", type=str, default="English")
-    parser.add_argument("--stream", action="store_true", default=True)
-    parser.add_argument("--no-stream", dest="stream", action="store_false")
-    parser.add_argument("--config-name", type=str, default="async_chunk_on")
-    parser.add_argument("--result-dir", type=str, default="results")
-    return parser.parse_args()
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    asyncio.run(main(args))

benchmarks/qwen3-tts/vllm_omni/bench_tts_serve.py

@@ -1,348 +0,0 @@
-"""Benchmark client for Qwen3-TTS via /v1/audio/speech endpoint.
-
-Measures TTFP (Time-to-First-Packet), E2E latency, and RTF (Real-Time Factor)
-across configurable concurrency levels. Saves results as JSON for plotting.
-
-Usage:
-    python bench_tts_serve.py \
-        --host 127.0.0.1 --port 8000 \
-        --num-prompts 50 \
-        --max-concurrency 1 4 10 \
-        --result-dir results/
-"""
-
-import argparse
-import asyncio
-import json
-import time
-from dataclasses import asdict, dataclass, field
-from datetime import datetime
-from pathlib import Path
-
-import aiohttp
-import numpy as np
-from tqdm.asyncio import tqdm
-
-PROMPTS = [
-    "Hello, welcome to the voice synthesis benchmark test.",
-    "She said she would be here by noon, but nobody showed up.",
-    "The quick brown fox jumps over the lazy dog near the riverbank.",
-    "I can't believe how beautiful the sunset looks from up here on the mountain.",
-    "Please remember to bring your identification documents to the appointment tomorrow morning.",
-    "Have you ever wondered what it would be like to travel through time and visit ancient civilizations?",
-    "The restaurant on the corner serves the best pasta I have ever tasted in my entire life.",
-    "After the meeting, we should discuss the quarterly results and plan for the next phase.",
-    "Learning a new language takes patience, practice, and a genuine curiosity about other cultures.",
-    "The train leaves at half past seven, so we need to arrive at the station before then.",
-    "Could you please turn down the music a little bit, I'm trying to concentrate on my work.",
-    "It was a dark and stormy night when the old lighthouse keeper heard a knock at the door.",
-]
-
-
-@dataclass
-class RequestResult:
-    success: bool = False
-    ttfp: float = 0.0  # Time to first audio packet (seconds)
-    e2e: float = 0.0  # End-to-end latency (seconds)
-    audio_bytes: int = 0  # Total audio bytes received
-    audio_duration: float = 0.0  # Audio duration in seconds (estimated from PCM)
-    rtf: float = 0.0  # Real-time factor = e2e / audio_duration
-    prompt: str = ""
-    error: str = ""
-
-
-@dataclass
-class BenchmarkResult:
-    config_name: str = ""
-    concurrency: int = 0
-    num_prompts: int = 0
-    completed: int = 0
-    failed: int = 0
-    duration_s: float = 0.0
-    # TTFP stats (ms)
-    mean_ttfp_ms: float = 0.0
-    median_ttfp_ms: float = 0.0
-    std_ttfp_ms: float = 0.0
-    p90_ttfp_ms: float = 0.0
-    p95_ttfp_ms: float = 0.0
-    p99_ttfp_ms: float = 0.0
-    # E2E stats (ms)
-    mean_e2e_ms: float = 0.0
-    median_e2e_ms: float = 0.0
-    std_e2e_ms: float = 0.0
-    p90_e2e_ms: float = 0.0
-    p95_e2e_ms: float = 0.0
-    p99_e2e_ms: float = 0.0
-    # RTF stats
-    mean_rtf: float = 0.0
-    median_rtf: float = 0.0
-    std_rtf: float = 0.0
-    p99_rtf: float = 0.0
-    # Audio stats
-    mean_audio_duration_s: float = 0.0
-    total_audio_duration_s: float = 0.0
-    audio_throughput: float = 0.0  # audio_duration / wall_time
-    request_throughput: float = 0.0  # requests / second
-    # Per-request details
-    per_request: list = field(default_factory=list)
-
-
-def pcm_bytes_to_duration(num_bytes: int, sample_rate: int = 24000, sample_width: int = 2) -> float:
-    """Convert raw PCM byte count to duration in seconds."""
-    num_samples = num_bytes / sample_width
-    return num_samples / sample_rate
-
-
-async def send_tts_request(
-    session: aiohttp.ClientSession,
-    api_url: str,
-    prompt: str,
-    voice: str = "vivian",
-    language: str = "English",
-    pbar: tqdm | None = None,
-) -> RequestResult:
-    """Send a streaming TTS request and measure latency metrics."""
-    payload = {
-        "input": prompt,
-        "voice": voice,
-        "language": language,
-        "stream": True,
-        "response_format": "pcm",
-    }
-
-    result = RequestResult(prompt=prompt)
-    st = time.perf_counter()
-
-    try:
-        async with session.post(api_url, json=payload) as response:
-            if response.status != 200:
-                result.error = f"HTTP {response.status}: {await response.text()}"
-                result.success = False
-                return result
-
-            first_chunk = True
-            total_bytes = 0
-
-            async for chunk in response.content.iter_any():
-                if first_chunk and len(chunk) > 0:
-                    result.ttfp = time.perf_counter() - st
-                    first_chunk = False
-                total_bytes += len(chunk)
-
-            result.e2e = time.perf_counter() - st
-            result.audio_bytes = total_bytes
-            result.audio_duration = pcm_bytes_to_duration(total_bytes)
-
-            if result.audio_duration > 0:
-                result.rtf = result.e2e / result.audio_duration
-            result.success = True
-
-    except Exception as e:
-        result.error = str(e)
-        result.success = False
-        result.e2e = time.perf_counter() - st
-
-    if pbar:
-        pbar.update(1)
-    return result
-
-
-async def run_benchmark(
-    host: str,
-    port: int,
-    num_prompts: int,
-    max_concurrency: int,
-    num_warmups: int = 3,
-    voice: str = "vivian",
-    language: str = "English",
-) -> BenchmarkResult:
-    """Run benchmark at a given concurrency level."""
-    api_url = f"http://{host}:{port}/v1/audio/speech"
-
-    connector = aiohttp.TCPConnector(
-        limit=max_concurrency,
-        limit_per_host=max_concurrency,
-        keepalive_timeout=60,
-    )
-    session = aiohttp.ClientSession(
-        connector=connector,
-        timeout=aiohttp.ClientTimeout(total=600),
-    )
-
-    # Warmup
-    if num_warmups > 0:
-        print(f"  Warming up with {num_warmups} requests...")
-        warmup_tasks = []
-        for i in range(num_warmups):
-            prompt = PROMPTS[i % len(PROMPTS)]
-            warmup_tasks.append(send_tts_request(session, api_url, prompt, voice, language))
-        await asyncio.gather(*warmup_tasks)
-        print("  Warmup done.")
-
-    # Build request list
-    request_prompts = [PROMPTS[i % len(PROMPTS)] for i in range(num_prompts)]
-
-    # Run benchmark
-    print(f"  Running {num_prompts} requests with concurrency={max_concurrency}...")
-    semaphore = asyncio.Semaphore(max_concurrency)
-    pbar = tqdm(total=num_prompts, desc=f"  concurrency={max_concurrency}")
-
-    async def limited_request(prompt):
-        async with semaphore:
-            return await send_tts_request(session, api_url, prompt, voice, language, pbar)
-
-    start_time = time.perf_counter()
-    tasks = [asyncio.create_task(limited_request(p)) for p in request_prompts]
-    results: list[RequestResult] = await asyncio.gather(*tasks)
-    duration = time.perf_counter() - start_time
-    pbar.close()
-
-    await session.close()
-
-    # Compute stats
-    successful = [r for r in results if r.success]
-    failed = [r for r in results if not r.success]
-
-    bench = BenchmarkResult(
-        concurrency=max_concurrency,
-        num_prompts=num_prompts,
-        completed=len(successful),
-        failed=len(failed),
-        duration_s=duration,
-    )
-
-    if successful:
-        ttfps = [r.ttfp * 1000 for r in successful]  # convert to ms
-        e2es = [r.e2e * 1000 for r in successful]
-        rtfs = [r.rtf for r in successful]
-        audio_durs = [r.audio_duration for r in successful]
-
-        bench.mean_ttfp_ms = float(np.mean(ttfps))
-        bench.median_ttfp_ms = float(np.median(ttfps))
-        bench.std_ttfp_ms = float(np.std(ttfps))
-        bench.p90_ttfp_ms = float(np.percentile(ttfps, 90))
-        bench.p95_ttfp_ms = float(np.percentile(ttfps, 95))
-        bench.p99_ttfp_ms = float(np.percentile(ttfps, 99))
-
-        bench.mean_e2e_ms = float(np.mean(e2es))
-        bench.median_e2e_ms = float(np.median(e2es))
-        bench.std_e2e_ms = float(np.std(e2es))
-        bench.p90_e2e_ms = float(np.percentile(e2es, 90))
-        bench.p95_e2e_ms = float(np.percentile(e2es, 95))
-        bench.p99_e2e_ms = float(np.percentile(e2es, 99))
-
-        bench.mean_rtf = float(np.mean(rtfs))
-        bench.median_rtf = float(np.median(rtfs))
-        bench.std_rtf = float(np.std(rtfs))
-        bench.p99_rtf = float(np.percentile(rtfs, 99))
-
-        bench.mean_audio_duration_s = float(np.mean(audio_durs))
-        bench.total_audio_duration_s = float(np.sum(audio_durs))
-        bench.audio_throughput = bench.total_audio_duration_s / duration
-        bench.request_throughput = len(successful) / duration
-
-        bench.per_request = [
-            {
-                "ttfp_ms": r.ttfp * 1000,
-                "e2e_ms": r.e2e * 1000,
-                "rtf": r.rtf,
-                "audio_duration_s": r.audio_duration,
-                "prompt": r.prompt,
-            }
-            for r in successful
-        ]
-
-    # Print summary in standardized performance template
-    W = 50
-    print("")
-    print(f"{'=' * W}")
-    print(f"{'Serving Benchmark Result':^{W}}")
-    print(f"{'=' * W}")
-    print(f"{'Successful requests:':<40}{bench.completed:<10}")
-    print(f"{'Failed requests:':<40}{bench.failed:<10}")
-    print(f"{'Maximum request concurrency:':<40}{max_concurrency:<10}")
-    print(f"{'Benchmark duration (s):':<40}{duration:<10.2f}")
-    print(f"{'Request throughput (req/s):':<40}{bench.request_throughput:<10.2f}")
-    print(f"{'-' * W}")
-    print(f"{'End-to-end Latency':^{W}}")
-    print(f"{'-' * W}")
-    print(f"{'Mean E2EL (ms):':<40}{bench.mean_e2e_ms:<10.2f}")
-    print(f"{'Median E2EL (ms):':<40}{bench.median_e2e_ms:<10.2f}")
-    print(f"{'P99 E2EL (ms):':<40}{bench.p99_e2e_ms:<10.2f}")
-    print(f"{'=' * W}")
-    print(f"{'Audio Result':^{W}}")
-    print(f"{'=' * W}")
-    print(f"{'Total audio duration generated (s):':<40}{bench.total_audio_duration_s:<10.2f}")
-    print(f"{'Audio throughput (audio duration/s):':<40}{bench.audio_throughput:<10.2f}")
-    print(f"{'-' * W}")
-    print(f"{'Time to First Packet':^{W}}")
-    print(f"{'-' * W}")
-    print(f"{'Mean AUDIO_TTFP (ms):':<40}{bench.mean_ttfp_ms:<10.2f}")
-    print(f"{'Median AUDIO_TTFP (ms):':<40}{bench.median_ttfp_ms:<10.2f}")
-    print(f"{'P99 AUDIO_TTFP (ms):':<40}{bench.p99_ttfp_ms:<10.2f}")
-    print(f"{'-' * W}")
-    print(f"{'Real Time Factor':^{W}}")
-    print(f"{'-' * W}")
-    print(f"{'Mean AUDIO_RTF:':<40}{bench.mean_rtf:<10.3f}")
-    print(f"{'Median AUDIO_RTF:':<40}{bench.median_rtf:<10.3f}")
-    print(f"{'P99 AUDIO_RTF:':<40}{bench.p99_rtf:<10.3f}")
-    print(f"{'=' * W}")
-    print("")
-
-    if failed:
-        for r in failed[:3]:
-            print(f"  [ERROR] {r.error[:200]}")
-
-    return bench
-
-
-async def main(args):
-    all_results = []
-
-    for concurrency in args.max_concurrency:
-        result = await run_benchmark(
-            host=args.host,
-            port=args.port,
-            num_prompts=args.num_prompts,
-            max_concurrency=concurrency,
-            num_warmups=args.num_warmups,
-            voice=args.voice,
-            language=args.language,
-        )
-        result.config_name = args.config_name
-        all_results.append(asdict(result))
-
-    # Save results
-    result_dir = Path(args.result_dir)
-    result_dir.mkdir(parents=True, exist_ok=True)
-    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
-    result_file = result_dir / f"bench_{args.config_name}_{timestamp}.json"
-
-    with open(result_file, "w") as f:
-        json.dump(all_results, f, indent=2)
-    print(f"Results saved to {result_file}")
-
-    return all_results
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Qwen3-TTS Benchmark Client")
-    parser.add_argument("--host", type=str, default="127.0.0.1")
-    parser.add_argument("--port", type=int, default=8000)
-    parser.add_argument("--num-prompts", type=int, default=50, help="Number of prompts per concurrency level")
-    parser.add_argument(  # noqa: E501
-        "--max-concurrency", type=int, nargs="+", default=[1, 4, 10], help="Concurrency levels to test"
-    )
-    parser.add_argument("--num-warmups", type=int, default=3)
-    parser.add_argument("--voice", type=str, default="vivian")
-    parser.add_argument("--language", type=str, default="English")
-    parser.add_argument(
-        "--config-name", type=str, default="async_chunk", help="Label for this config (used in filenames)"
-    )
-    parser.add_argument("--result-dir", type=str, default="results")
-    return parser.parse_args()
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    asyncio.run(main(args))

benchmarks/qwen3-tts/vllm_omni/configs/qwen3_tts_bs1.yaml

@@ -1,93 +0,0 @@
-# Qwen3-TTS batch_size=1 config (streaming with async_chunk)
-# 2-stage pipeline: Talker -> Code2Wav
-async_chunk: true
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    is_comprehension: true
-    runtime:
-      devices: "0"
-    engine_args:
-      max_num_seqs: 1
-      model_stage: qwen3_tts
-      model_arch: Qwen3TTSTalkerForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      enforce_eager: false
-      trust_remote_code: true
-      async_scheduling: true
-      enable_prefix_caching: false
-      engine_output_type: latent
-      gpu_memory_utilization: 0.3
-      distributed_executor_backend: "mp"
-      max_num_batched_tokens: 512
-      max_model_len: 4096
-      custom_process_next_stage_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_tts.talker2code2wav_async_chunk
-    output_connectors:
-      to_stage_1: connector_of_shared_memory
-    default_sampling_params:
-      temperature: 0.9
-      top_k: 50
-      max_tokens: 4096
-      seed: 42
-      detokenize: false
-      repetition_penalty: 1.05
-      stop_token_ids: [2150]
-
-  - stage_id: 1
-    stage_type: llm
-    runtime:
-      devices: "0"
-    engine_args:
-      max_num_seqs: 1
-      model_stage: code2wav
-      model_arch: Qwen3TTSCode2Wav
-      worker_type: generation
-      scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
-      enforce_eager: true
-      trust_remote_code: true
-      async_scheduling: true
-      enable_prefix_caching: false
-      engine_output_type: audio
-      gpu_memory_utilization: 0.3
-      distributed_executor_backend: "mp"
-      max_num_batched_tokens: 8192
-      max_model_len: 32768
-    engine_input_source: [0]
-    final_output: true
-    final_output_type: audio
-    input_connectors:
-      from_stage_0: connector_of_shared_memory
-    tts_args:
-      max_instructions_length: 500
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 1.0
-      top_k: -1
-      max_tokens: 65536
-      seed: 42
-      detokenize: true
-      repetition_penalty: 1.0
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 1
-
-  connectors:
-    connector_of_shared_memory:
-      name: SharedMemoryConnector
-      extra:
-        shm_threshold_bytes: 65536
-        codec_streaming: true
-        connector_get_sleep_s: 0.01
-        connector_get_max_wait_first_chunk: 3000
-        connector_get_max_wait: 300
-        codec_chunk_frames: 25
-        codec_left_context_frames: 25
-
-  edges:
-    - from: 0
-      to: 1
-      window_size: -1

benchmarks/qwen3-tts/vllm_omni/configs/qwen3_tts_bs16.yaml

@@ -1,94 +0,0 @@
-# Qwen3-TTS max_num_seqs=16 config (streaming with async_chunk)
-# High-throughput concurrent request processing
-# 2-stage pipeline: Talker -> Code2Wav
-async_chunk: true
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    is_comprehension: true
-    runtime:
-      devices: "0"
-    engine_args:
-      max_num_seqs: 16
-      model_stage: qwen3_tts
-      model_arch: Qwen3TTSTalkerForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      enforce_eager: false
-      trust_remote_code: true
-      async_scheduling: true
-      enable_prefix_caching: false
-      engine_output_type: latent
-      gpu_memory_utilization: 0.3
-      distributed_executor_backend: "mp"
-      max_num_batched_tokens: 4096
-      max_model_len: 4096
-      custom_process_next_stage_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_tts.talker2code2wav_async_chunk
-    output_connectors:
-      to_stage_1: connector_of_shared_memory
-    default_sampling_params:
-      temperature: 0.9
-      top_k: 50
-      max_tokens: 4096
-      seed: 42
-      detokenize: false
-      repetition_penalty: 1.05
-      stop_token_ids: [2150]
-
-  - stage_id: 1
-    stage_type: llm
-    runtime:
-      devices: "0"
-    engine_args:
-      max_num_seqs: 16
-      model_stage: code2wav
-      model_arch: Qwen3TTSCode2Wav
-      worker_type: generation
-      scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
-      enforce_eager: true
-      trust_remote_code: true
-      async_scheduling: true
-      enable_prefix_caching: false
-      engine_output_type: audio
-      gpu_memory_utilization: 0.2
-      distributed_executor_backend: "mp"
-      max_num_batched_tokens: 16384
-      max_model_len: 32768
-    engine_input_source: [0]
-    final_output: true
-    final_output_type: audio
-    input_connectors:
-      from_stage_0: connector_of_shared_memory
-    tts_args:
-      max_instructions_length: 500
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 1.0
-      top_k: -1
-      max_tokens: 65536
-      seed: 42
-      detokenize: true
-      repetition_penalty: 1.0
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 16
-
-  connectors:
-    connector_of_shared_memory:
-      name: SharedMemoryConnector
-      extra:
-        shm_threshold_bytes: 65536
-        codec_streaming: true
-        connector_get_sleep_s: 0.01
-        connector_get_max_wait_first_chunk: 3000
-        connector_get_max_wait: 300
-        codec_chunk_frames: 25
-        codec_left_context_frames: 25
-
-  edges:
-    - from: 0
-      to: 1
-      window_size: -1

benchmarks/qwen3-tts/vllm_omni/configs/qwen3_tts_bs4.yaml

@@ -1,94 +0,0 @@
-# Qwen3-TTS batch_size=4 config (streaming with async_chunk)
-# Enables concurrent request processing
-# 2-stage pipeline: Talker -> Code2Wav
-async_chunk: true
-stage_args:
-  - stage_id: 0
-    stage_type: llm
-    is_comprehension: true
-    runtime:
-      devices: "0"
-    engine_args:
-      max_num_seqs: 4
-      model_stage: qwen3_tts
-      model_arch: Qwen3TTSTalkerForConditionalGeneration
-      worker_type: ar
-      scheduler_cls: vllm_omni.core.sched.omni_ar_scheduler.OmniARScheduler
-      enforce_eager: false
-      trust_remote_code: true
-      async_scheduling: true
-      enable_prefix_caching: false
-      engine_output_type: latent
-      gpu_memory_utilization: 0.3
-      distributed_executor_backend: "mp"
-      max_num_batched_tokens: 512
-      max_model_len: 4096
-      custom_process_next_stage_input_func: vllm_omni.model_executor.stage_input_processors.qwen3_tts.talker2code2wav_async_chunk
-    output_connectors:
-      to_stage_1: connector_of_shared_memory
-    default_sampling_params:
-      temperature: 0.9
-      top_k: 50
-      max_tokens: 4096
-      seed: 42
-      detokenize: false
-      repetition_penalty: 1.05
-      stop_token_ids: [2150]
-
-  - stage_id: 1
-    stage_type: llm
-    runtime:
-      devices: "0"
-    engine_args:
-      max_num_seqs: 4
-      model_stage: code2wav
-      model_arch: Qwen3TTSCode2Wav
-      worker_type: generation
-      scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler
-      enforce_eager: true
-      trust_remote_code: true
-      async_scheduling: true
-      enable_prefix_caching: false
-      engine_output_type: audio
-      gpu_memory_utilization: 0.2
-      distributed_executor_backend: "mp"
-      max_num_batched_tokens: 8192
-      max_model_len: 32768
-    engine_input_source: [0]
-    final_output: true
-    final_output_type: audio
-    input_connectors:
-      from_stage_0: connector_of_shared_memory
-    tts_args:
-      max_instructions_length: 500
-    default_sampling_params:
-      temperature: 0.0
-      top_p: 1.0
-      top_k: -1
-      max_tokens: 65536
-      seed: 42
-      detokenize: true
-      repetition_penalty: 1.0
-
-runtime:
-  enabled: true
-  defaults:
-    window_size: -1
-    max_inflight: 4
-
-  connectors:
-    connector_of_shared_memory:
-      name: SharedMemoryConnector
-      extra:
-        shm_threshold_bytes: 65536
-        codec_streaming: true
-        connector_get_sleep_s: 0.01
-        connector_get_max_wait_first_chunk: 3000
-        connector_get_max_wait: 300
-        codec_chunk_frames: 25
-        codec_left_context_frames: 25
-
-  edges:
-    - from: 0
-      to: 1
-      window_size: -1

benchmarks/qwen3-tts/vllm_omni/plot_async_chunk.py

@@ -1,249 +0,0 @@
-"""Plot TTFP comparison: async_chunk off vs on.
-
-Generates a bar chart with improvement arrows, matching the Qwen3-Omni
-async_chunk benchmark figure style.
-
-Usage:
-    python plot_async_chunk.py \
-        --off results/bench_async_chunk_off_*.json \
-        --on results/bench_async_chunk_on_*.json \
-        --output results/qwen3_tts_async_chunk_ttfp.png
-
-    # Also supports E2E and RTF metrics:
-    python plot_async_chunk.py \
-        --off results/bench_async_chunk_off_*.json \
-        --on results/bench_async_chunk_on_*.json \
-        --metric e2e \
-        --output results/qwen3_tts_async_chunk_e2e.png
-"""
-
-import argparse
-import json
-from pathlib import Path
-
-import matplotlib.pyplot as plt
-import numpy as np
-
-METRIC_CONFIG = {
-    "ttfp": {
-        "key": "mean_ttfp_ms",
-        "ylabel": "TTFP (s)",
-        "title": "TTFP (Time to First Audio Packet) - Qwen3-TTS, by concurrency",
-        "to_seconds": True,
-    },
-    "e2e": {
-        "key": "mean_e2e_ms",
-        "ylabel": "E2E (s)",
-        "title": "E2E Latency - Qwen3-TTS, by concurrency",
-        "to_seconds": True,
-    },
-    "rtf": {
-        "key": "mean_rtf",
-        "ylabel": "RTF",
-        "title": "Real-Time Factor - Qwen3-TTS, by concurrency",
-        "to_seconds": False,
-    },
-}
-
-
-def load_results(path: str) -> list[dict]:
-    with open(path) as f:
-        return json.load(f)
-
-
-def plot_ttfp_comparison(
-    off_results: list[dict],
-    on_results: list[dict],
-    metric: str,
-    output_path: str,
-    title_override: str | None = None,
-):
-    cfg = METRIC_CONFIG[metric]
-    key = cfg["key"]
-    to_seconds = cfg["to_seconds"]
-
-    off_map = {r["concurrency"]: r for r in off_results}
-    on_map = {r["concurrency"]: r for r in on_results}
-    concurrencies = sorted(set(off_map.keys()) & set(on_map.keys()))
-
-    off_vals = []
-    on_vals = []
-    for c in concurrencies:
-        v_off = off_map[c][key]
-        v_on = on_map[c][key]
-        if to_seconds:
-            v_off /= 1000.0
-            v_on /= 1000.0
-        off_vals.append(v_off)
-        on_vals.append(v_on)
-
-    fig, ax = plt.subplots(figsize=(8, 6))
-
-    x = np.arange(len(concurrencies))
-    width = 0.3
-
-    ax.bar(x - width / 2, off_vals, width, label="async_chunk off", color="#87CEEB", edgecolor="none")
-    ax.bar(x + width / 2, on_vals, width, label="async_chunk on", color="#FFF8DC", edgecolor="#DDD8B8")
-
-    # Draw improvement arrows and labels
-    for i in range(len(concurrencies)):
-        v_off = off_vals[i]
-        v_on = on_vals[i]
-        if v_on > 0:
-            improvement = v_off / v_on
-        else:
-            improvement = float("inf")
-
-        # Arrow from top of off-bar to top of on-bar
-        arrow_start_x = x[i] - width / 2
-        arrow_start_y = v_off * 0.95
-        arrow_end_x = x[i] + width / 2
-        arrow_end_y = v_on * 1.05
-
-        ax.annotate(
-            "",
-            xy=(arrow_end_x, arrow_end_y),
-            xytext=(arrow_start_x, arrow_start_y),
-            arrowprops=dict(arrowstyle="->", color="red", lw=1.5),
-        )
-
-        # Improvement label
-        label_x = (arrow_start_x + arrow_end_x) / 2
-        label_y = arrow_start_y + (v_off - v_on) * 0.15
-        ax.text(
-            label_x,
-            label_y,
-            f"{improvement:.1f}x improvement",
-            ha="center",
-            va="bottom",
-            fontsize=10,
-            color="red",
-            fontweight="bold",
-        )
-
-    title = title_override or cfg["title"]
-    ax.set_title(title, fontsize=13, fontweight="bold")
-    ax.set_ylabel(cfg["ylabel"], fontsize=12)
-    ax.set_xlabel("Max concurrency", fontsize=12)
-    ax.set_xticks(x)
-    ax.set_xticklabels([str(c) for c in concurrencies])
-    ax.set_yscale("log")
-    ax.legend(loc="upper left", fontsize=11)
-    ax.grid(axis="y", alpha=0.3, linestyle="--")
-    ax.set_axisbelow(True)
-
-    plt.tight_layout()
-    Path(output_path).parent.mkdir(parents=True, exist_ok=True)
-    plt.savefig(output_path, dpi=150, bbox_inches="tight")
-    print(f"Plot saved to {output_path}")
-    plt.close()
-
-
-def plot_all_metrics(off_results: list[dict], on_results: list[dict], output_path: str):
-    """Generate a 1x3 subplot with TTFP, E2E, and RTF comparisons."""
-    off_map = {r["concurrency"]: r for r in off_results}
-    on_map = {r["concurrency"]: r for r in on_results}
-    concurrencies = sorted(set(off_map.keys()) & set(on_map.keys()))
-
-    fig, axes = plt.subplots(1, 3, figsize=(18, 6))
-    fig.suptitle("Qwen3-TTS: async_chunk on vs off", fontsize=15, fontweight="bold")
-
-    for ax, metric in zip(axes, ["ttfp", "e2e", "rtf"]):
-        cfg = METRIC_CONFIG[metric]
-        key = cfg["key"]
-        to_seconds = cfg["to_seconds"]
-
-        off_vals = []
-        on_vals = []
-        for c in concurrencies:
-            v_off = off_map[c][key]
-            v_on = on_map[c][key]
-            if to_seconds:
-                v_off /= 1000.0
-                v_on /= 1000.0
-            off_vals.append(v_off)
-            on_vals.append(v_on)
-
-        x = np.arange(len(concurrencies))
-        width = 0.3
-        ax.bar(x - width / 2, off_vals, width, label="async_chunk off", color="#87CEEB")
-        ax.bar(x + width / 2, on_vals, width, label="async_chunk on", color="#FFF8DC", edgecolor="#DDD8B8")
-
-        for i in range(len(concurrencies)):
-            if on_vals[i] > 0:
-                improvement = off_vals[i] / on_vals[i]
-                ax.annotate(
-                    "",
-                    xy=(x[i] + width / 2, on_vals[i] * 1.05),
-                    xytext=(x[i] - width / 2, off_vals[i] * 0.95),
-                    arrowprops=dict(arrowstyle="->", color="red", lw=1.5),
-                )
-                label_y = off_vals[i] * 0.85
-                ax.text(x[i], label_y, f"{improvement:.1f}x", ha="center", fontsize=10, color="red", fontweight="bold")
-
-        ax.set_title(cfg["title"].split(" - ")[0], fontsize=12, fontweight="bold")
-        ax.set_ylabel(cfg["ylabel"], fontsize=11)
-        ax.set_xlabel("Max concurrency", fontsize=11)
-        ax.set_xticks(x)
-        ax.set_xticklabels([str(c) for c in concurrencies])
-        if metric != "rtf":
-            ax.set_yscale("log")
-        ax.legend(fontsize=9)
-        ax.grid(axis="y", alpha=0.3, linestyle="--")
-        ax.set_axisbelow(True)
-
-    plt.tight_layout()
-    Path(output_path).parent.mkdir(parents=True, exist_ok=True)
-    plt.savefig(output_path, dpi=150, bbox_inches="tight")
-    print(f"Plot saved to {output_path}")
-    plt.close()
-
-
-def print_table(off_results: list[dict], on_results: list[dict]):
-    off_map = {r["concurrency"]: r for r in off_results}
-    on_map = {r["concurrency"]: r for r in on_results}
-    concurrencies = sorted(set(off_map.keys()) & set(on_map.keys()))
-
-    print("\n## Benchmark Results: async_chunk off vs on\n")
-    print("| Metric | Concurrency | async_chunk off | async_chunk on | Improvement |")
-    print("| --- | --- | --- | --- | --- |")
-
-    for name, key, fmt in [
-        ("TTFP (ms)", "mean_ttfp_ms", ".1f"),
-        ("E2E (ms)", "mean_e2e_ms", ".1f"),
-        ("RTF", "mean_rtf", ".3f"),
-        ("Throughput", "audio_throughput", ".2f"),
-    ]:
-        for c in concurrencies:
-            v_off = off_map[c].get(key, 0)
-            v_on = on_map[c].get(key, 0)
-            if v_on > 0 and key != "audio_throughput":
-                ratio = f"{v_off / v_on:.1f}x"
-            elif v_off > 0 and key == "audio_throughput":
-                ratio = f"{v_on / v_off:.1f}x"
-            else:
-                ratio = "N/A"
-            print(f"| {name} | {c} | {v_off:{fmt}} | {v_on:{fmt}} | {ratio} |")
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Plot async_chunk comparison for Qwen3-TTS")
-    parser.add_argument("--off", type=str, required=True, help="JSON results for async_chunk off")
-    parser.add_argument("--on", type=str, required=True, help="JSON results for async_chunk on")
-    parser.add_argument("--metric", type=str, default="ttfp", choices=["ttfp", "e2e", "rtf", "all"])
-    parser.add_argument("--output", type=str, default="results/qwen3_tts_async_chunk.png")
-    parser.add_argument("--title", type=str, default=None, help="Custom title override")
-    return parser.parse_args()
-
-
-if __name__ == "__main__":
-    args = parse_args()
-    off_results = load_results(args.off)
-    on_results = load_results(args.on)
-
-    print_table(off_results, on_results)
-
-    if args.metric == "all":
-        plot_all_metrics(off_results, on_results, args.output)
-    else:
-        plot_ttfp_comparison(off_results, on_results, args.metric, args.output, args.title)

benchmarks/qwen3-tts/vllm_omni/run_async_chunk_benchmark.sh

@@ -1,167 +0,0 @@
-#!/bin/bash
-# Qwen3-TTS async_chunk on vs off Benchmark
-#
-# Starts two servers (async_chunk on and off), benchmarks both,
-# and generates comparison plots.
-#
-# Usage:
-#   bash run_async_chunk_benchmark.sh
-#
-# Environment variables:
-#   GPU_DEVICE       - GPU index (default: 0)
-#   MODEL            - Model path (default: Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice)
-#   NUM_PROMPTS      - Prompts per concurrency level (default: 50)
-#   CONCURRENCY      - Space-separated concurrency levels (default: "1 10")
-#   PORT_ON          - Port for async_chunk on server (default: 8000)
-#   PORT_OFF         - Port for async_chunk off server (default: 8001)
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-PROJECT_ROOT="$(cd "${SCRIPT_DIR}/../../.." && pwd)"
-cd "$PROJECT_ROOT"
-
-GPU_DEVICE="${GPU_DEVICE:-0}"
-MODEL="${MODEL:-Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice}"
-NUM_PROMPTS="${NUM_PROMPTS:-50}"
-CONCURRENCY="${CONCURRENCY:-1 10}"
-NUM_WARMUPS="${NUM_WARMUPS:-3}"
-PORT_ON="${PORT_ON:-8000}"
-PORT_OFF="${PORT_OFF:-8001}"
-RESULT_DIR="${SCRIPT_DIR}/results"
-TIMESTAMP="$(date +%Y%m%d_%H%M%S)"
-
-STAGE_CONFIG_ON="vllm_omni/model_executor/stage_configs/qwen3_tts.yaml"
-STAGE_CONFIG_OFF="vllm_omni/model_executor/stage_configs/qwen3_tts_no_async_chunk.yaml"
-
-mkdir -p "${RESULT_DIR}"
-
-echo "============================================================"
-echo " Qwen3-TTS async_chunk Benchmark"
-echo "============================================================"
-echo " GPU:            ${GPU_DEVICE}"
-echo " Model:          ${MODEL}"
-echo " Prompts:        ${NUM_PROMPTS}"
-echo " Concurrency:    ${CONCURRENCY}"
-echo " Port (on/off):  ${PORT_ON} / ${PORT_OFF}"
-echo " Results:        ${RESULT_DIR}"
-echo "============================================================"
-
-cleanup() {
-    echo "Cleaning up servers..."
-    kill "$PID_ON" 2>/dev/null || true
-    kill "$PID_OFF" 2>/dev/null || true
-    wait "$PID_ON" 2>/dev/null || true
-    wait "$PID_OFF" 2>/dev/null || true
-}
-trap cleanup EXIT
-
-wait_for_server() {
-    local port=$1
-    local name=$2
-    local max_wait=300
-    local elapsed=0
-    echo "Waiting for ${name} server on port ${port}..."
-    while ! curl -s "http://localhost:${port}/health" >/dev/null 2>&1; do
-        sleep 5
-        elapsed=$((elapsed + 5))
-        if [ $elapsed -ge $max_wait ]; then
-            echo "ERROR: ${name} server failed to start within ${max_wait}s"
-            exit 1
-        fi
-    done
-    echo "${name} server ready (${elapsed}s)"
-}
-
-# ---- Phase 1: Start async_chunk ON server ----
-echo ""
-echo "[Phase 1] Starting async_chunk ON server on port ${PORT_ON}..."
-CUDA_VISIBLE_DEVICES=${GPU_DEVICE} vllm-omni serve "${MODEL}" \
-    --stage-configs-path "${STAGE_CONFIG_ON}" \
-    --host 0.0.0.0 --port "${PORT_ON}" \
-    --trust-remote-code --enforce-eager --omni \
-    > "${RESULT_DIR}/server_on_${TIMESTAMP}.log" 2>&1 &
-PID_ON=$!
-
-wait_for_server "${PORT_ON}" "async_chunk_on"
-
-echo "[Phase 1] Benchmarking async_chunk ON..."
-# shellcheck disable=SC2086
-python "${SCRIPT_DIR}/bench_async_chunk.py" \
-    --host 127.0.0.1 --port "${PORT_ON}" \
-    --config-name "async_chunk_on" \
-    --num-prompts "${NUM_PROMPTS}" \
-    --max-concurrency ${CONCURRENCY} \
-    --num-warmups "${NUM_WARMUPS}" \
-    --result-dir "${RESULT_DIR}"
-
-echo "[Phase 1] Stopping async_chunk ON server..."
-kill "$PID_ON" 2>/dev/null || true
-wait "$PID_ON" 2>/dev/null || true
-sleep 5
-
-# ---- Phase 2: Start async_chunk OFF server ----
-echo ""
-echo "[Phase 2] Starting async_chunk OFF server on port ${PORT_OFF}..."
-CUDA_VISIBLE_DEVICES=${GPU_DEVICE} vllm-omni serve "${MODEL}" \
-    --stage-configs-path "${STAGE_CONFIG_OFF}" \
-    --host 0.0.0.0 --port "${PORT_OFF}" \
-    --trust-remote-code --enforce-eager --omni \
-    > "${RESULT_DIR}/server_off_${TIMESTAMP}.log" 2>&1 &
-PID_OFF=$!
-
-wait_for_server "${PORT_OFF}" "async_chunk_off"
-
-echo "[Phase 2] Benchmarking async_chunk OFF (non-streaming)..."
-# shellcheck disable=SC2086
-python "${SCRIPT_DIR}/bench_async_chunk.py" \
-    --host 127.0.0.1 --port "${PORT_OFF}" \
-    --config-name "async_chunk_off" \
-    --num-prompts "${NUM_PROMPTS}" \
-    --max-concurrency ${CONCURRENCY} \
-    --num-warmups "${NUM_WARMUPS}" \
-    --no-stream \
-    --result-dir "${RESULT_DIR}"
-
-echo "[Phase 2] Stopping async_chunk OFF server..."
-kill "$PID_OFF" 2>/dev/null || true
-wait "$PID_OFF" 2>/dev/null || true
-
-# ---- Phase 3: Plot results ----
-echo ""
-echo "[Phase 3] Generating plots..."
-
-# Find the latest result files
-RESULT_ON=$(ls -t "${RESULT_DIR}"/bench_async_chunk_on_*.json 2>/dev/null | head -1)
-RESULT_OFF=$(ls -t "${RESULT_DIR}"/bench_async_chunk_off_*.json 2>/dev/null | head -1)
-
-if [ -z "$RESULT_ON" ] || [ -z "$RESULT_OFF" ]; then
-    echo "ERROR: Could not find result files. Check logs in ${RESULT_DIR}/"
-    exit 1
-fi
-
-echo "  ON results:  ${RESULT_ON}"
-echo "  OFF results: ${RESULT_OFF}"
-
-# TTFP comparison (main figure)
-python "${SCRIPT_DIR}/plot_async_chunk.py" \
-    --off "${RESULT_OFF}" \
-    --on "${RESULT_ON}" \
-    --metric ttfp \
-    --output "${RESULT_DIR}/qwen3_tts_async_chunk_ttfp.png"
-
-# All metrics comparison
-python "${SCRIPT_DIR}/plot_async_chunk.py" \
-    --off "${RESULT_OFF}" \
-    --on "${RESULT_ON}" \
-    --metric all \
-    --output "${RESULT_DIR}/qwen3_tts_async_chunk_all.png"
-
-echo ""
-echo "============================================================"
-echo " Benchmark complete!"
-echo " Results: ${RESULT_DIR}/"
-echo " Plots:"
-echo "   - ${RESULT_DIR}/qwen3_tts_async_chunk_ttfp.png"
-echo "   - ${RESULT_DIR}/qwen3_tts_async_chunk_all.png"
-echo "============================================================"

benchmarks/tts/README.md

@@ -0,0 +1,244 @@
+# TTS Universal Benchmark
+
+A model-agnostic serving benchmark for TTS models in vllm-omni. One CLI
+(`bench_tts.py`) + one YAML registry (`model_configs.yaml`) drive perf and
+quality runs for every registered checkpoint: **Qwen3-TTS** (Base / CustomVoice)
+and **VoxCPM2** today, more to come.
+
+The same three task types — `voice_clone`, `default_voice`, `voice_design` —
+are wired into both the manual CLI and the DFX nightly CI matrix
+(`tests/dfx/perf/tests/test_tts.json`).
+
+## Quick start
+
+### 1. Start the server
+
+```bash
+vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-Base --omni --port 8000
+```
+
+The server auto-loads its Deploy YAML from `vllm_omni/deploy/qwen3_tts.yaml`
+(Pipeline + Deploy schema introduced in #2383). No `--stage-configs-path` or
+`--deploy-config` flag is needed for any registered model.
+
+### 2. Run the benchmark (`vllm bench serve --omni`)
+
+The primary, directly-controllable path. Copy-paste one of these and tweak
+any bench flag (sampling params, endpoint, extra body, warmups, etc.):
+
+#### voice_clone (Qwen3-TTS-Base, seed-tts dataset)
+
+```bash
+vllm bench serve --omni \
+    --host 127.0.0.1 --port 8000 \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+    --backend openai-audio-speech \
+    --endpoint /v1/audio/speech \
+    --dataset-name seed-tts \
+    --dataset-path /path/to/seed-tts-eval \
+    --seed-tts-locale en \
+    --num-prompts 20 --num-warmups 2 \
+    --extra-body '{"task_type":"Base"}' \
+    --max-concurrency 1 --request-rate inf \
+    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun \
+    --save-result --result-dir ./results
+```
+
+#### default_voice (Qwen3-TTS-CustomVoice, bundled seed_tts_smoke)
+
+```bash
+vllm bench serve --omni \
+    --host 127.0.0.1 --port 8000 \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+    --backend openai-audio-speech \
+    --endpoint /v1/audio/speech \
+    --dataset-name seed-tts-text \
+    --dataset-path benchmarks/build_dataset/seed_tts_smoke \
+    --seed-tts-locale en \
+    --num-prompts 20 --num-warmups 2 \
+    --extra-body '{"voice":"Vivian","language":"English","task_type":"CustomVoice"}' \
+    --max-concurrency 1 --request-rate inf \
+    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun \
+    --save-result --result-dir ./results
+```
+
+#### voice_design (Qwen3-TTS-CustomVoice, bundled seed_tts_design)
+
+```bash
+vllm bench serve --omni \
+    --host 127.0.0.1 --port 8000 \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \
+    --backend openai-audio-speech \
+    --endpoint /v1/audio/speech \
+    --dataset-name seed-tts-design \
+    --dataset-path benchmarks/build_dataset/seed_tts_design \
+    --seed-tts-locale en \
+    --num-prompts 20 --num-warmups 2 \
+    --extra-body '{"task_type":"VoiceDesign","language":"English"}' \
+    --max-concurrency 1 --request-rate inf \
+    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun \
+    --save-result --result-dir ./results
+```
+
+#### Streaming continuity (`audio_underrun`)
+
+`audio_underrun` (seconds) is the per-request worst-case buffer deficit
+under a realtime-rate player simulation. The audio-speech backend records
+chunk arrival times during the SSE stream and surfaces:
+
+- `Mean / Median / P{50,99} AUDIO_UNDERRUN (s)` per the standard percentile
+  output - any positive value means at least one inter-chunk gap was longer
+  than the chunk's audio duration.
+- `Streaming continuity OK rate` - fraction of requests whose worst-case
+  underrun stayed under the threshold (default 100 ms, the commonly cited
+  "audible gap" budget). Override the threshold with
+  `VLLM_OMNI_BENCH_AUDIO_CONTINUITY_THRESHOLD_S=<float>`.
+
+This captures the failure mode where `RTF_p50 < 1` (server keeps up in
+aggregate) but per-stream chunk arrival is bursty enough that listeners
+still hear gaps - common at high concurrency on streaming TTS pipelines.
+
+#### Add WER / SIM / UTMOS to any of the above
+
+Append `--seed-tts-wer-eval` (and optionally `SEED_TTS_EVAL_DEVICE=cuda:0`
+in the env, per PR #2558). This triggers the seed-tts-eval protocol:
+Whisper-large-v3 ASR → WER, WavLM embeddings → SIM, balacoon/utmos → UTMOS.
+
+### 3. Convenience wrapper (`bench_tts.py`)
+
+If you're running the **canonical** configuration for a registered model,
+`bench_tts.py` loads the right defaults from `model_configs.yaml` and
+emits the exact `vllm bench serve --omni` command above — useful for
+concurrency sweeps and multi-task runs:
+
+```bash
+# Smallest smoke — 5 prompts, concurrency=1
+python benchmarks/tts/bench_tts.py \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+    --task voice_clone \
+    --dataset-path /path/to/seed-tts-eval \
+    --concurrency 1 --num-prompts 5 \
+    --output-dir ./results
+
+# Full concurrency sweep
+python benchmarks/tts/bench_tts.py \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+    --task voice_clone \
+    --dataset-path /path/to/seed-tts-eval \
+    --concurrency 1 2 4 8 16 32 \
+    --num-prompts 20 \
+    --output-dir ./results
+
+# With WER / SIM / UTMOS quality eval (adds ASR + embedding compute)
+python benchmarks/tts/bench_tts.py \
+    --model Qwen/Qwen3-TTS-12Hz-1.7B-Base \
+    --task voice_clone \
+    --dataset-path /path/to/seed-tts-eval \
+    --wer-eval \
+    --concurrency 4 --num-prompts 200 \
+    --output-dir ./results
+```
+
+### 4. Plot a sweep
+
+```bash
+python benchmarks/tts/plot_results.py \
+    --results ./results/*.json \
+    --output ./results/curve.png
+```
+
+Outputs TTFP / RTF / throughput curves (and a markdown table) for every
+`(task, concurrency)` combination in the result set.
+
+## Task types
+
+| Task            | Dataset           | Request body                                        | Checkpoints that support it              |
+|-----------------|-------------------|-----------------------------------------------------|------------------------------------------|
+| `voice_clone`   | `seed-tts`        | `ref_audio` + `ref_text` + `task_type=Base`         | `Qwen3-TTS-*-Base`, `VoxCPM2`            |
+| `default_voice` | `seed-tts-text`   | `voice=Vivian` + `task_type=CustomVoice`            | `Qwen3-TTS-*-CustomVoice`                |
+| `voice_design`  | `seed-tts-design` | `instructions=<natural-language description>` + `task_type=VoiceDesign` | `Qwen3-TTS-*-CustomVoice` |
+
+**`-CustomVoice` checkpoints do NOT ship `speaker_encoder` weights**, so
+voice_clone requests raise `ValueError` at model runtime. Use `-Base` for
+voice_clone.
+
+## Adding a new TTS model
+
+Drop an entry into `model_configs.yaml` — no Python changes required:
+
+```yaml
+models:
+  <org>/<model-id>:
+    supported_tasks: [voice_clone]          # or default_voice / voice_design
+    backend: openai-audio-speech            # vllm bench serve backend
+    endpoint: /v1/audio/speech              # OpenAI-compatible endpoint
+    task_extra_body:                        # merged into every request's body
+      voice_clone:
+        task_type: Base
+```
+
+Then add the model's Deploy YAML under `vllm_omni/deploy/<model>.yaml`
+(Pipeline + Deploy schema) and it's immediately benchable.
+
+## Datasets
+
+| Dataset            | Bundled? | Format            | Source                                                         |
+|--------------------|----------|-------------------|----------------------------------------------------------------|
+| `seed-tts-design`  | ✅       | 5-field meta.lst  | `benchmarks/build_dataset/seed_tts_design/en/meta.lst` (20 prompts) |
+| `seed_tts_smoke`   | ✅       | 4-field meta.lst  | `benchmarks/build_dataset/seed_tts_smoke/en/meta.lst` (20 text-only) |
+| `seed-tts`         | ❌       | 4-field meta.lst + WAVs | Google-Drive: [BytedanceSpeech/seed-tts-eval][seedtts] (~1.2 GB) |
+| `seed-tts-text`    | ❌       | 4-field meta.lst  | Same archive as `seed-tts` (wav column unused)                 |
+
+[seedtts]: https://github.com/BytedanceSpeech/seed-tts-eval
+
+For manual voice_clone / default_voice runs against the full corpus, follow
+`benchmarks/build_dataset/download_process_data_seedtts.md` and point
+`--dataset-path` at the extracted `seedtts_testset` directory.
+
+## DFX nightly CI
+
+`tests/dfx/perf/tests/test_tts.json` wires three perf regimes plus quality:
+
+| eval_phase    | concurrency | purpose                                                 | Baseline metrics                        |
+|---------------|-------------|---------------------------------------------------------|-----------------------------------------|
+| `latency`     | 1           | Single-request TTFP / RTF SLO                           | `median_audio_ttfp_ms`, `median_audio_rtf` |
+| `throughput`  | 8           | Codec-batching cliff sentinel (PDF #272 concurrency≥8)  | `median_audio_ttfp_ms`, `median_audio_rtf` |
+| `quality`     | 4           | WER / SIM / UTMOS regression (disabled in CI by default)| `mean_audio_rtf`                        |
+
+Why `median_*` for latency/throughput and `mean_*` for quality: latency
+distributions have cold-start tails that drag the mean; quality aggregates
+over 200 prompts so single-request outliers don't matter.
+
+Quality entries are `enabled: false` in CI because seed-tts-eval is not
+staged in the Buildkite container (matches the precedent in
+PR #2558 — quality runs are manual / release-validation, not nightly).
+
+## Concurrency cliff regression sentinel
+
+Observed on H20-3e, Qwen3-TTS-1.7B (measured pre-merge on this branch):
+
+| Task          | Model         | c=1    | c=4    | **c=8**    | c=16   | c=32   |
+|---------------|---------------|--------|--------|------------|--------|--------|
+| voice_clone   | 1.7B-Base     | RTF 0.15 / TTFP 165ms | 0.28 / 412ms | **0.49 / 1701ms** | 0.72 / 3355ms | 0.77 / 3772ms |
+| voice_design  | 1.7B-CustomVoice | RTF 0.08 / TTFP 53ms  | 0.11 / 154ms | **0.21 / 872ms**  | 0.33 / 1801ms | 0.38 / 1989ms |
+
+Both models show a **4–6× TTFP jump from c=4 to c=8** while audio throughput
+saturates around c=4–8 — the codec-bs=1 bottleneck documented in
+vllm-project/vllm-omni#272. The `throughput` CI regime at c=8 is the
+sentinel for regressions in this area.
+
+## File layout
+
+```
+benchmarks/tts/
+├── README.md                  (this file)
+├── bench_tts.py               CLI — serve-mode benchmark driver
+├── plot_results.py            Generate per-task / per-concurrency curves
+└── model_configs.yaml         Model registry (supported tasks + extra body)
+```
+
+## Related
+
+- Upstream seed-tts-eval integration: vllm-project/vllm-omni#2558
+- Pipeline + Deploy schema: vllm-project/vllm-omni#2383
+- Concurrency cliff RFC: vllm-project/vllm-omni#272

benchmarks/tts/bench_tts.py

@@ -0,0 +1,311 @@
+#!/usr/bin/env python3
+"""Universal TTS benchmark CLI for vllm-omni.
+
+Runs ``vllm bench serve --omni`` with model-aware defaults loaded from
+``model_configs.yaml``.  Supports Qwen3-TTS, VoxCPM2, and any future TTS
+model registered in the config file -- no code changes needed to add models.
+
+Usage::
+
+    python benchmarks/tts/bench_tts.py \\
+        --model Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice \\
+        --task voice_clone \\
+        --locale en \\
+        --concurrency 1 4 \\
+        --num-prompts 20 \\
+        --dataset-path /path/to/seed-tts-eval \\
+        --host localhost --port 8000
+
+See ``--help`` for full option list.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import os
+import subprocess
+import sys
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+
+def _vllm_omni_bin() -> str:
+    """Return the vllm-omni (or vllm) binary co-located with the current Python."""
+    bin_dir = Path(sys.executable).parent
+    for candidate in ("vllm-omni", "vllm"):
+        p = bin_dir / candidate
+        if p.is_file():
+            return str(p)
+    return "vllm-omni"  # fall back and let the shell resolve it
+
+
+_REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+_SCRIPT_DIR = Path(__file__).resolve().parent
+_DEFAULT_MODEL_CONFIGS = _SCRIPT_DIR / "model_configs.yaml"
+
+# Maps task name to the dataset_name used with vllm bench serve
+_TASK_TO_DATASET: dict[str, str] = {
+    "voice_clone": "seed-tts",
+    "default_voice": "seed-tts-text",
+    "voice_design": "seed-tts-design",
+}
+
+# Default design dataset path (bundled with the repo)
+_DEFAULT_DESIGN_DATASET_PATH = str(_REPO_ROOT / "benchmarks" / "build_dataset" / "seed_tts_design")
+
+
+def load_model_configs(path: Path) -> dict[str, Any]:
+    """Load model registry from YAML file."""
+    with open(path, encoding="utf-8") as f:
+        data = yaml.safe_load(f)
+    return data.get("models", {})
+
+
+def build_bench_args(
+    *,
+    host: str,
+    port: int,
+    model: str,
+    task: str,
+    model_cfg: dict[str, Any],
+    locale: str,
+    num_prompts: int,
+    concurrency: int | None,
+    dataset_path: str | None,
+    wer_eval: bool,
+    output_dir: str | None,
+    result_filename: str | None,
+    extra_cli_args: list[str],
+) -> list[str]:
+    """Build the ``vllm bench serve --omni`` command for one (task, concurrency) run."""
+    dataset_name = _TASK_TO_DATASET[task]
+    backend: str = model_cfg["backend"]
+    endpoint: str = model_cfg["endpoint"]
+    task_extra_body: dict[str, Any] = (model_cfg.get("task_extra_body") or {}).get(task) or {}
+
+    # Resolve dataset path
+    if dataset_path:
+        resolved_dataset_path = dataset_path
+    elif task == "voice_design":
+        resolved_dataset_path = _DEFAULT_DESIGN_DATASET_PATH
+    else:
+        resolved_dataset_path = None
+
+    cmd = [
+        _vllm_omni_bin(),
+        "bench",
+        "serve",
+        "--omni",
+        "--host",
+        host,
+        "--port",
+        str(port),
+        "--model",
+        model,
+        "--backend",
+        backend,
+        "--endpoint",
+        endpoint,
+        "--dataset-name",
+        dataset_name,
+        "--num-prompts",
+        str(num_prompts),
+        "--num-warmups",
+        "2",
+        "--percentile-metrics",
+        "ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun",
+    ]
+
+    if resolved_dataset_path:
+        cmd += ["--dataset-path", resolved_dataset_path]
+
+    if locale:
+        cmd += ["--seed-tts-locale", locale]
+
+    if task_extra_body:
+        cmd += ["--extra-body", json.dumps(task_extra_body, separators=(",", ":"))]
+
+    if model_cfg.get("trust_remote_code"):
+        cmd.append("--trust-remote-code")
+
+    if concurrency is not None:
+        cmd += ["--max-concurrency", str(concurrency), "--request-rate", "inf"]
+
+    if wer_eval:
+        cmd.append("--seed-tts-wer-eval")
+
+    if output_dir or result_filename:
+        out_dir = output_dir or "."
+        os.makedirs(out_dir, exist_ok=True)
+        cmd += ["--save-result", "--result-dir", out_dir]
+        if result_filename:
+            cmd += ["--result-filename", result_filename]
+
+    cmd += extra_cli_args
+    return cmd
+
+
+def run_one_benchmark(cmd: list[str]) -> dict[str, Any] | None:
+    """Run a single benchmark subprocess and return parsed JSON result if available."""
+    print(f"\n{'=' * 60}")
+    print("Running:", " ".join(cmd))
+    print("=" * 60)
+    result = subprocess.run(cmd, check=False)
+    if result.returncode != 0:
+        print(f"[bench_tts] WARNING: benchmark exited with code {result.returncode}")
+        return None
+    # If --save-result was used, find the result file
+    try:
+        result_dir_idx = cmd.index("--result-dir")
+        result_dir = Path(cmd[result_dir_idx + 1])
+        if "--result-filename" in cmd:
+            fname_idx = cmd.index("--result-filename")
+            result_file = result_dir / cmd[fname_idx + 1]
+        else:
+            # find most recently modified json
+            jsons = sorted(result_dir.glob("result_*.json"), key=lambda p: p.stat().st_mtime)
+            result_file = jsons[-1] if jsons else None
+        if result_file and result_file.is_file():
+            return json.loads(result_file.read_text(encoding="utf-8"))
+    except (ValueError, IndexError, OSError):
+        pass
+    return None
+
+
+def print_summary_table(results: list[dict[str, Any]]) -> None:
+    """Print a unified metrics table across all (task, concurrency) runs."""
+    if not results:
+        return
+    header = (
+        f"{'Task':<16} {'Concurrency':>11} {'RTF mean':>10} "
+        f"{'TTFP (ms)':>10} {'Throughput':>12} {'WER':>7} {'SIM':>7} {'UTMOS':>7}"
+    )
+    print(f"\n{'=' * len(header)}")
+    print("BENCHMARK SUMMARY")
+    print("=" * len(header))
+    print(header)
+    print("-" * len(header))
+    for r in results:
+        task = r.get("_task", "?")
+        conc = r.get("_concurrency", "?")
+        rtf = r.get("mean_audio_rtf", float("nan"))
+        ttfp = r.get("mean_audio_ttfp_ms", float("nan"))
+        throughput = r.get("audio_throughput", float("nan"))
+        wer = r.get("seed_tts_mean_wer", float("nan"))
+        sim = r.get("seed_tts_mean_sim", float("nan"))
+        utmos = r.get("seed_tts_mean_utmos", float("nan"))
+
+        def fmt(v: float, digits: int = 3) -> str:
+            return f"{v:.{digits}f}" if not math.isnan(v) else "  n/a"
+
+        print(
+            f"{task:<16} {str(conc):>11} {fmt(rtf):>10} {fmt(ttfp, 0):>10} "
+            f"{fmt(throughput):>12} {fmt(wer):>7} {fmt(sim):>7} {fmt(utmos):>7}"
+        )
+    print("=" * len(header))
+
+
+def main() -> None:
+    """Entry point for the universal TTS benchmark CLI."""
+    parser = argparse.ArgumentParser(
+        description=__doc__,
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "--model", required=True, help="HuggingFace model ID (e.g. Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice)"
+    )
+    parser.add_argument("--task", default="all", help="Task type: voice_clone | default_voice | voice_design | all")
+    parser.add_argument("--locale", default="en", choices=["en", "zh"])
+    parser.add_argument("--concurrency", type=int, nargs="+", default=[1, 4], metavar="N")
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        nargs="+",
+        default=[20],
+        metavar="N",
+        help="Number of prompts per run. If one value, applied to all concurrency levels.",
+    )
+    parser.add_argument(
+        "--dataset-path", default=None, help="Root of seed-tts-eval dataset (required for voice_clone/default_voice)"
+    )
+    parser.add_argument("--wer-eval", action="store_true", help="Enable WER/SIM/UTMOS quality eval")
+    parser.add_argument("--output-dir", default=None, help="Directory to save result JSON files")
+    parser.add_argument("--host", default="localhost")
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument("--model-configs", default=str(_DEFAULT_MODEL_CONFIGS), help="Path to model_configs.yaml")
+    parser.add_argument("extra", nargs=argparse.REMAINDER, help="Extra args passed directly to vllm bench serve")
+    args = parser.parse_args()
+
+    model_configs = load_model_configs(Path(args.model_configs))
+    if args.model not in model_configs:
+        known = "\n  ".join(model_configs.keys())
+        print(f"[bench_tts] ERROR: model '{args.model}' not in model_configs.yaml.\nKnown models:\n  {known}")
+        sys.exit(1)
+
+    model_cfg = model_configs[args.model]
+    supported_tasks: list[str] = model_cfg.get("supported_tasks", [])
+
+    tasks_to_run: list[str]
+    if args.task == "all":
+        tasks_to_run = supported_tasks
+    elif args.task in supported_tasks:
+        tasks_to_run = [args.task]
+    else:
+        print(
+            f"[bench_tts] ERROR: task '{args.task}' not supported by {args.model}.\nSupported tasks: {supported_tasks}"
+        )
+        sys.exit(1)
+
+    # Align num_prompts list with concurrency list
+    num_prompts_list: list[int] = args.num_prompts
+    if len(num_prompts_list) == 1:
+        num_prompts_list = num_prompts_list * len(args.concurrency)
+    elif len(num_prompts_list) != len(args.concurrency):
+        print(
+            f"[bench_tts] ERROR: --num-prompts ({len(num_prompts_list)} values) must be "
+            f"length 1 or match --concurrency ({len(args.concurrency)} values)."
+        )
+        sys.exit(1)
+
+    all_results: list[dict[str, Any]] = []
+
+    for task in tasks_to_run:
+        for concurrency, num_prompts in zip(args.concurrency, num_prompts_list):
+            ts = datetime.now().strftime("%Y%m%d-%H%M%S")
+            result_filename = f"bench_tts_{args.model.replace('/', '_')}_{task}_c{concurrency}_{ts}.json"
+            cmd = build_bench_args(
+                host=args.host,
+                port=args.port,
+                model=args.model,
+                task=task,
+                model_cfg=model_cfg,
+                locale=args.locale,
+                num_prompts=num_prompts,
+                concurrency=concurrency,
+                dataset_path=args.dataset_path,
+                wer_eval=args.wer_eval,
+                output_dir=args.output_dir,
+                result_filename=result_filename,
+                extra_cli_args=args.extra or [],
+            )
+            result = run_one_benchmark(cmd)
+            if result is not None:
+                result["_task"] = task
+                result["_concurrency"] = concurrency
+                all_results.append(result)
+                # Persist the metadata so plot_results.py can pick it up.
+                if args.output_dir and result_filename:
+                    result_path = Path(args.output_dir) / result_filename
+                    if result_path.is_file():
+                        result_path.write_text(json.dumps(result, indent=2), encoding="utf-8")
+
+    print_summary_table(all_results)
+
+
+if __name__ == "__main__":
+    main()

benchmarks/tts/model_configs.yaml

@@ -0,0 +1,42 @@
+# Universal TTS benchmark model registry.
+# Maps HuggingFace model ID → supported tasks + per-task extra body fields.
+# To add a new TTS model: add an entry here. No code changes required.
+#
+# The server auto-loads its Deploy YAML from vllm_omni/deploy/<model>.yaml via
+# the Pipeline + Deploy schema introduced in #2383, so no stage_config path
+# is tracked here.
+
+models:
+  # -CustomVoice checkpoints lack speaker_encoder weights, so voice_clone is
+  # NOT supported (an attempt raises ValueError from _extract_speaker_embedding
+  # at model runtime).  Use -Base for voice_clone.
+  Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice:
+    supported_tasks: [default_voice, voice_design]
+    backend: openai-audio-speech
+    endpoint: /v1/audio/speech
+    task_extra_body:
+      default_voice:
+        voice: Vivian
+        language: English
+        task_type: CustomVoice
+      voice_design:
+        task_type: VoiceDesign
+        language: English
+
+  Qwen/Qwen3-TTS-12Hz-1.7B-Base:
+    supported_tasks: [voice_clone]
+    backend: openai-audio-speech
+    endpoint: /v1/audio/speech
+    task_extra_body:
+      voice_clone:
+        task_type: Base
+
+  openbmb/VoxCPM2:
+    supported_tasks: [voice_clone, default_voice]
+    trust_remote_code: true
+    backend: openai-audio-speech
+    endpoint: /v1/audio/speech
+    task_extra_body:
+      voice_clone: {}
+      default_voice:
+        voice: default

benchmarks/tts/plot_results.py

@@ -0,0 +1,324 @@
+"""Plot universal TTS benchmark results.
+
+Reads JSON files saved by ``bench_tts.py`` (via ``vllm bench serve --omni``)
+and generates comparison bar charts grouped by task type.
+
+Metrics plotted:
+- AUDIO_TTFP  (mean audio time-to-first-packet, ms)
+- E2EL        (mean end-to-end latency, ms)
+- Audio RTF   (mean real-time factor)
+- Audio throughput (audio-seconds / wall-second)
+
+Quality metrics (WER / SIM / UTMOS) are printed in a table when present.
+
+Usage::
+
+    # Single run — one JSON per task, all in results/
+    python benchmarks/tts/plot_results.py \\
+        --results results/bench_tts_*.json \\
+        --output results/tts_benchmark.png
+
+    # Compare two runs (e.g. async_chunk on vs off)
+    python benchmarks/tts/plot_results.py \\
+        --results run_a/bench_tts_*.json \\
+        --results run_b/bench_tts_*.json \\
+        --labels "async_chunk_on" "async_chunk_off" \\
+        --output results/comparison.png
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+
+# ---------------------------------------------------------------------------
+# JSON loading
+# ---------------------------------------------------------------------------
+
+
+def load_run(paths: list[str]) -> list[dict]:
+    """Load and merge all JSON files for one run into a flat list of records.
+
+    Each record is expected to have at least ``_concurrency`` (int) and
+    ``_task`` (str) keys injected by ``bench_tts.py``.  Records that come
+    from a file that contains a list are flattened.
+    """
+    records: list[dict] = []
+    for p in paths:
+        raw = json.loads(Path(p).read_text(encoding="utf-8"))
+        if isinstance(raw, list):
+            records.extend(raw)
+        elif isinstance(raw, dict):
+            records.append(raw)
+    return records
+
+
+def _get(record: dict, key: str) -> float:
+    v = record.get(key, float("nan"))
+    if v is None or (isinstance(v, float) and math.isnan(v)):
+        return float("nan")
+    try:
+        return float(v)
+    except (TypeError, ValueError):
+        return float("nan")
+
+
+# ---------------------------------------------------------------------------
+# Plotting helpers
+# ---------------------------------------------------------------------------
+
+
+def _bar_group(
+    ax: plt.Axes,
+    x: np.ndarray,
+    data_per_label: dict[str, list[float]],
+    width: float,
+    colors: list[str],
+    ylabel: str,
+    title: str,
+    concurrency_labels: list[str],
+    fmt: str = ".1f",
+) -> None:
+    n = len(data_per_label)
+    offsets = np.linspace(-(n - 1) * width / 2, (n - 1) * width / 2, n) if n > 1 else [0.0]
+
+    for i, (label, values) in enumerate(data_per_label.items()):
+        plot_vals = [0.0 if math.isnan(v) else v for v in values]
+        bar = ax.bar(x + offsets[i], plot_vals, width, label=label, color=colors[i % len(colors)], alpha=0.85)
+        max_val = max((v for v in values if not math.isnan(v)), default=1.0)
+        for rect, val in zip(bar, values):
+            if not math.isnan(val) and val > 0:
+                ax.text(
+                    rect.get_x() + rect.get_width() / 2,
+                    rect.get_height() + max_val * 0.02,
+                    f"{val:{fmt}}",
+                    ha="center",
+                    va="bottom",
+                    fontsize=8,
+                    fontweight="bold",
+                )
+
+    ax.set_xlabel("Concurrency", fontsize=11)
+    ax.set_ylabel(ylabel, fontsize=11)
+    ax.set_title(title, fontsize=12, fontweight="bold")
+    ax.set_xticks(x)
+    ax.set_xticklabels(concurrency_labels)
+    ax.legend(fontsize=9)
+    ax.grid(axis="y", alpha=0.3)
+    ax.set_axisbelow(True)
+
+
+COLORS = ["#2196F3", "#FF5722", "#4CAF50", "#FFC107", "#9C27B0"]
+
+
+# ---------------------------------------------------------------------------
+# Comparison plot (multiple labels / runs)
+# ---------------------------------------------------------------------------
+
+
+def plot_comparison(
+    all_runs: list[list[dict]],
+    labels: list[str],
+    output_path: str,
+    task_filter: str | None = None,
+    title_prefix: str = "TTS",
+) -> None:
+    """One 2×2 subplot per task found in the data."""
+    # Determine tasks to plot
+    tasks: list[str] = []
+    for run in all_runs:
+        for r in run:
+            t = r.get("_task", "unknown")
+            if t not in tasks:
+                tasks.append(t)
+    if task_filter:
+        tasks = [t for t in tasks if t == task_filter]
+
+    n_tasks = len(tasks)
+    if n_tasks == 0:
+        print("[plot_results] No tasks found in data.")
+        return
+
+    fig, axes_grid = plt.subplots(n_tasks, 4, figsize=(18, 4.5 * n_tasks))
+    fig.suptitle(f"{title_prefix} Benchmark", fontsize=15, fontweight="bold")
+
+    # Ensure axes_grid is always 2D
+    if n_tasks == 1:
+        axes_grid = [axes_grid]
+
+    for row_idx, task in enumerate(tasks):
+        # Collect concurrencies across all runs for this task
+        all_concs: set[int] = set()
+        for run in all_runs:
+            for r in run:
+                if r.get("_task") == task:
+                    c = r.get("_concurrency")
+                    if c is not None:
+                        all_concs.add(int(c))
+        concurrencies = sorted(all_concs)
+        x = np.arange(len(concurrencies))
+        conc_labels = [str(c) for c in concurrencies]
+
+        def _series(run: list[dict], metric_key: str) -> list[float]:
+            conc_map = {int(r["_concurrency"]): r for r in run if r.get("_task") == task and "_concurrency" in r}
+            return [_get(conc_map.get(c, {}), metric_key) for c in concurrencies]
+
+        metrics = [
+            ("mean_audio_ttfp_ms", "TTFP (ms)", "Time-to-First-Packet", ".0f"),
+            ("mean_e2el_ms", "E2E Latency (ms)", "End-to-End Latency", ".0f"),
+            ("mean_audio_rtf", "RTF", "Real-Time Factor (RTF)", ".3f"),
+            ("audio_throughput", "audio-s / wall-s", "Audio Throughput", ".2f"),
+        ]
+
+        axes_row = axes_grid[row_idx]
+        for col_idx, (key, ylabel, subtitle, fmt) in enumerate(metrics):
+            data_per_label = {lbl: _series(run, key) for lbl, run in zip(labels, all_runs)}
+            _bar_group(
+                axes_row[col_idx],
+                x,
+                data_per_label,
+                width=0.3 if len(labels) > 1 else 0.5,
+                colors=COLORS,
+                ylabel=ylabel,
+                title=f"{task} — {subtitle}",
+                concurrency_labels=conc_labels,
+                fmt=fmt,
+            )
+
+    plt.tight_layout()
+    Path(output_path).parent.mkdir(parents=True, exist_ok=True)
+    plt.savefig(output_path, dpi=150, bbox_inches="tight")
+    print(f"Plot saved to {output_path}")
+    plt.close()
+
+
+# ---------------------------------------------------------------------------
+# Markdown comparison table
+# ---------------------------------------------------------------------------
+
+
+def print_comparison_table(all_runs: list[list[dict]], labels: list[str]) -> None:
+    tasks: list[str] = []
+    for run in all_runs:
+        for r in run:
+            t = r.get("_task", "unknown")
+            if t not in tasks:
+                tasks.append(t)
+
+    perf_metrics = [
+        ("TTFP (ms)", "mean_audio_ttfp_ms", ".1f"),
+        ("E2E (ms)", "mean_e2el_ms", ".1f"),
+        ("RTF", "mean_audio_rtf", ".3f"),
+        ("Throughput (a-s/s)", "audio_throughput", ".2f"),
+    ]
+    quality_metrics = [
+        ("WER (%)", "seed_tts_mean_wer", ".1f"),
+        ("SIM", "seed_tts_mean_sim", ".3f"),
+        ("UTMOS", "seed_tts_mean_utmos", ".2f"),
+    ]
+
+    for task in tasks:
+        all_concs: set[int] = set()
+        for run in all_runs:
+            for r in run:
+                if r.get("_task") == task:
+                    c = r.get("_concurrency")
+                    if c is not None:
+                        all_concs.add(int(c))
+        concurrencies = sorted(all_concs)
+
+        print(f"\n## {task}\n")
+        col_header = "| Metric | Concurrency |" + "".join(f" {lbl} |" for lbl in labels)
+        sep = "| --- | --- |" + " --- |" * len(labels)
+        print(col_header)
+        print(sep)
+
+        for metric, key, fmt in perf_metrics + quality_metrics:
+            for c in concurrencies:
+                row = f"| {metric} | {c} |"
+                for run in all_runs:
+                    conc_map = {
+                        int(r["_concurrency"]): r for r in run if r.get("_task") == task and "_concurrency" in r
+                    }
+                    val = _get(conc_map.get(c, {}), key)
+                    row += f" {val:{fmt}} |" if not math.isnan(val) else " n/a |"
+                print(row)
+
+        # Improvement column (2-run comparison only)
+        if len(all_runs) == 2:
+            print(f"\n### Improvement ({labels[0]} vs {labels[1]})\n")
+            print("| Metric | Concurrency | Change |")
+            print("| --- | --- | --- |")
+            for metric, key, _ in perf_metrics:
+                for c in concurrencies:
+                    conc_map0 = {
+                        int(r["_concurrency"]): r for r in all_runs[0] if r.get("_task") == task and "_concurrency" in r
+                    }
+                    conc_map1 = {
+                        int(r["_concurrency"]): r for r in all_runs[1] if r.get("_task") == task and "_concurrency" in r
+                    }
+                    v0 = _get(conc_map0.get(c, {}), key)
+                    v1 = _get(conc_map1.get(c, {}), key)
+                    if not math.isnan(v0) and not math.isnan(v1) and v1 > 0:
+                        pct = (v1 - v0) / v1 * 100
+                        print(f"| {metric} | {c} | {pct:+.1f}% |")
+
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description=__doc__,
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument(
+        "--results",
+        type=str,
+        nargs="+",
+        action="append",
+        required=True,
+        metavar="FILE",
+        help="JSON result file(s) for one run. Repeat --results for multiple runs to compare.",
+    )
+    parser.add_argument(
+        "--labels",
+        type=str,
+        nargs="+",
+        default=None,
+        help="Label for each --results group (must match the number of --results groups).",
+    )
+    parser.add_argument("--output", type=str, default="results/tts_benchmark.png", help="Output image path.")
+    parser.add_argument("--title", type=str, default="TTS", help="Title prefix for the plot.")
+    parser.add_argument("--task", type=str, default=None, help="Filter to a single task (e.g. voice_clone).")
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+
+    # args.results is a list-of-lists due to action="append"
+    all_runs: list[list[dict]] = [load_run(group) for group in args.results]
+    n_runs = len(all_runs)
+
+    labels: list[str]
+    if args.labels:
+        if len(args.labels) != n_runs:
+            raise SystemExit(f"--labels count ({len(args.labels)}) must match --results groups ({n_runs})")
+        labels = args.labels
+    else:
+        labels = [f"run{i + 1}" for i in range(n_runs)]
+
+    print_comparison_table(all_runs, labels)
+    plot_comparison(all_runs, labels, args.output, task_filter=args.task, title_prefix=args.title)
+
+
+if __name__ == "__main__":
+    main()

docker/Dockerfile.ci

@@ -1,5 +1,5 @@
 ARG VLLM_BASE_IMAGE=vllm/vllm-openai
-ARG VLLM_BASE_TAG=v0.18.0
+ARG VLLM_BASE_TAG=v0.22.0
 FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG}
 ARG APP_DIR=/workspace/vllm-omni
 WORKDIR ${APP_DIR}
@@ -7,13 +7,11 @@ COPY . .
 
 # Install system dependencies
 RUN apt-get update && \
-    apt-get install -y espeak-ng ffmpeg git sox libsox-fmt-all jq && \
+    apt-get install -y espeak-ng git jq && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
-# Install vllm-omni into the same uv-managed Python environment used by the base image.
-# Use bash -c so that $(python3 -c ...) is expanded inside the container.
-RUN uv pip install --system --no-cache-dir ".[dev]"
+RUN uv pip install --system ".[dev]"
 
 RUN ln -sf /usr/bin/python3 /usr/bin/python