Update CUDA Docker base image to vLLM v0.21.0

Align Dockerfile.cuda with the v0.21.0 rebase on main and other
GPU Dockerfiles/docs. Related to #3858.

Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>

.buildkite/nightly-release-pipeline.yaml

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

.buildkite/pipeline-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: "v0.20.0"
+          VLLM_VERSION: "v0.21.0"
         priority: 100
         timeout_in_minutes: 60
         soft_fail: false

.buildkite/pipeline-npu.yaml

@@ -78,15 +78,19 @@ steps:
       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"
-  #   agents:
-  #     queue: "ascend-a2b3"
-  #     resource_class: "npu-1"
-  #   command: |
-  #     buildkite-agent pipeline upload .buildkite/test-npu-ready.yaml
+  - 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"

.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/hardware_ci/run-xpu-test.sh

@@ -52,5 +52,6 @@ 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"
+    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/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/test-merge.yml

@@ -1,5 +1,6 @@
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -172,7 +173,7 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py tests/e2e/offline_inference/test_qwen3_tts_customvoice.py -m 'advanced_model and cuda' --run-level 'advanced_model'
             "
         agents:
-          queue: "gpu_1_queue"
+          queue: "gpu_4_queue"
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -193,7 +194,7 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
             "
         agents:
-          queue: "gpu_1_queue"
+          queue: "gpu_4_queue"
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -247,6 +248,47 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
+      - label: "TTS · GLM-TTS Test"
+        timeout_in_minutes: 20
+        commands:
+          - |
+            timeout 20m bash -c "
+              export VLLM_LOGGING_LEVEL=DEBUG
+              pytest -s -v tests/e2e/online_serving/test_glm_tts.py tests/e2e/offline_inference/test_glm_tts.py -m 'advanced_model and cuda' --run-level 'advanced_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: "Omni · Qwen3-Omni Test"
         timeout_in_minutes: 30

.buildkite/test-nightly.yml

@@ -1,5 +1,6 @@
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -388,7 +389,7 @@ steps:
                   - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
                     resources:
                       limits:
-                        nvidia.com/gpu: 1
+                        nvidia.com/gpu: 2
                     volumeMounts:
                       - name: devshm
                         mountPath: /dev/shm
@@ -659,87 +660,6 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
-      - label: ":full_moon: Diffusion X2I(&A&T) · GEBench Accuracy Test"
-        timeout_in_minutes: 60
-        commands:
-          - pytest -s -v tests/e2e/accuracy/test_gebench_h100_smoke.py --run-level full_model --gebench-model Qwen/Qwen-Image-2512 --accuracy-judge-model QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ --accuracy-gpu 0 --gebench-port 8093 --accuracy-workers 1
-          - buildkite-agent artifact upload "tests/e2e/accuracy/artifacts/gebench_qwen-image-2512/summary*.json"
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 1
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
-      - label: ":full_moon: Diffusion X2I(&A&T) · GEdit-Bench Accuracy Test"
-        timeout_in_minutes: 60
-        commands:
-          - pytest -s -v tests/e2e/accuracy/test_gedit_bench_h100_smoke.py --run-level full_model --gedit-model Qwen/Qwen-Image-Edit --accuracy-judge-model QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ --accuracy-gpu 0 --gedit-port 8093 --gedit-samples-per-group 20 --accuracy-workers 1
-          - buildkite-agent artifact upload "tests/e2e/accuracy/artifacts/gedit_scores_qwen-image-edit/qwen-image-edit_all_all_vie_score_*.csv"
-          - buildkite-agent artifact upload "tests/e2e/accuracy/artifacts/gedit_scores_qwen-image-edit/qwen-image-edit_all_all_summary_*.json"
-        agents:
-          queue: "mithril-h100-pool"
-        plugins:
-          - kubernetes:
-              podSpec:
-                containers:
-                  - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-                    resources:
-                      limits:
-                        nvidia.com/gpu: 1
-                    volumeMounts:
-                      - name: devshm
-                        mountPath: /dev/shm
-                      - name: hf-cache
-                        mountPath: /root/.cache/huggingface
-                    env:
-                      - name: HF_HOME
-                        value: /root/.cache/huggingface
-                      - name: VLLM_HTTP_TIMEOUT_KEEP_ALIVE
-                        value: "120"
-                      - name: HF_TOKEN
-                        valueFrom:
-                          secretKeyRef:
-                            name: hf-token-secret
-                            key: token
-                nodeSelector:
-                  node.kubernetes.io/instance-type: gpu-h100-sxm
-                volumes:
-                  - name: devshm
-                    emptyDir:
-                      medium: Memory
-                  - name: hf-cache
-                    hostPath:
-                      path: /mnt/hf-cache
-                      type: DirectoryOrCreate
-
       - label: ":full_moon: Diffusion X2I(&A&T) · Accuracy Test"
         timeout_in_minutes: 180
         commands:
@@ -1039,7 +959,7 @@ steps:
       - label: ":full_moon: Diffusion X2V · Function Test"
         timeout_in_minutes: 90
         commands:
-          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_hunyuan_video_15_expansion.py  --run-level "full_model"
+          - pytest -s -v tests/e2e/online_serving/test_wan22_expansion.py tests/e2e/online_serving/test_wan_2_1_vace_expansion.py tests/e2e/online_serving/test_hunyuan_video_15_expansion.py tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py -m "full_model and cuda" --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -1166,9 +1086,8 @@ steps:
     if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test" || build.pull_request.labels includes "diffusion-x2iat-test"
     steps:
       - label: ":full_moon: Quantization Test with H100"
-        timeout_in_minutes: 90
         commands:
-          - timeout 30m pytest -sv tests/diffusion/quantization -m 'full_model and cuda and H100' --run-level "full_model"
+          - timeout 60m pytest -sv tests/diffusion/quantization -m 'full_model and cuda and H100' --run-level "full_model"
         agents:
           queue: "mithril-h100-pool"
         plugins:
@@ -1276,7 +1195,7 @@ steps:
       - nightly-diffusion-x2iat-performance-qwen-image-layered
       - nightly-diffusion-x2v-performance
       - nightly-testcase-statistics
-    if: build.env("NIGHTLY") == "1"
+    if: build.env("NIGHTLY") == "1" && build.env("EMAIL_DISTRIBUTION") == "1"
     commands:
       - pip install openpyxl
       - export DEFAULT_INPUT_DIR=tests/dfx/perf/results

.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

@@ -1,5 +1,6 @@
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -303,7 +304,7 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_customvoice.py -m 'core_model' --run-level 'core_model'
             "
         agents:
-          queue: "gpu_1_queue"
+          queue: "gpu_4_queue"
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
@@ -326,7 +327,7 @@ steps:
               pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py -m 'core_model' --run-level 'core_model'
             "
         agents:
-          queue: "gpu_1_queue"
+          queue: "gpu_4_queue"
         plugins:
           - docker#v5.2.0:
               image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT

.buildkite/test-weekly.yml

@@ -1,5 +1,6 @@
 env:
   VLLM_WORKER_MULTIPROC_METHOD: spawn
+  VLLM_USE_FLASHINFER_MOE_FP16: 0
   HF_HUB_DOWNLOAD_TIMEOUT: 300
   HF_HUB_ETAG_TIMEOUT: 60
 
@@ -83,3 +84,65 @@ steps:
                 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/diffusion-perf-opt/SKILL.md

@@ -0,0 +1,488 @@
+---
+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.
+
+## 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. **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.
+
+7. **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.
+
+8. **Produce an optimization plan.**
+   - Classify candidates as P0/P1/P2.
+   - For each candidate, state necessity, expected benefit, implementation path, validation plan, and quality risk.
+   - 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 quality regression. A/B means same workload and hardware before/after. Quality regression means checking generated image/video stability, artifacts, temporal flicker, and seed behavior when precision or approximate kernels change.
+
+## 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

@@ -18,6 +18,9 @@ include:
 
 - `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

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/build_dataset/seed_tts_design/en/meta.lst

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

benchmarks/build_dataset/seed_tts_smoke/en/meta.lst

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

benchmarks/diffusion/README.md

@@ -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.

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,
@@ -151,7 +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,
@@ -348,12 +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/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 (v1/videos backend)
+    # Video (/v1/videos endpoint)
     t2v:
     python3 benchmarks/diffusion/diffusion_benchmark_serving.py \
-        --backend v1/videos --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 v1/videos --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 +42,25 @@ 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)
+    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/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 +88,12 @@ 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 +140,7 @@ 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 +532,7 @@ 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."
             )
@@ -580,7 +593,7 @@ class RandomDataset(BaseDataset):
             self._sampled_requests = None
 
         # Random image generate
-        if self.args.task in ["i2v", "ti2v", "ti2i", "i2i"]:
+        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
@@ -883,13 +896,23 @@ 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"
@@ -902,18 +925,23 @@ 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":
@@ -929,6 +957,10 @@ 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)
@@ -976,15 +1008,17 @@ 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))
@@ -1047,12 +1081,17 @@ 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",
@@ -1065,7 +1104,7 @@ if __name__ == "__main__":
         "--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(
@@ -1172,9 +1211,15 @@ if __name__ == "__main__":
         default=1,
         help=(
             "Number of synthetic input images to attach for image-conditioned tasks "
-            "(i2v, ti2v, ti2i, i2i) when using random dataset."
+            "(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

@@ -58,7 +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> \
@@ -71,7 +71,7 @@ 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          |
@@ -129,7 +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 \

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 \
@@ -148,6 +158,7 @@ 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)
@@ -165,8 +176,9 @@ def _generate_image(omni, args, prompt, seed):
     elapsed = time.perf_counter() - start
     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
 
@@ -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:
@@ -223,9 +244,11 @@ 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.accelerator.empty_cache()
@@ -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/tts/README.md

@@ -40,7 +40,7 @@ vllm bench serve --omni \
     --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 \
+    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun \
     --save-result --result-dir ./results
 ```
 
@@ -58,7 +58,7 @@ vllm bench serve --omni \
     --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 \
+    --percentile-metrics ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun \
     --save-result --result-dir ./results
 ```
 
@@ -76,10 +76,28 @@ vllm bench serve --omni \
     --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 \
+    --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`
@@ -215,7 +233,6 @@ sentinel for regressions in this area.
 benchmarks/tts/
 ├── README.md                  (this file)
 ├── bench_tts.py               CLI — serve-mode benchmark driver
-├── bench_voxcpm_offline.py    CLI — offline VoxCPM benchmark (sync + streaming)
 ├── plot_results.py            Generate per-task / per-concurrency curves
 └── model_configs.yaml         Model registry (supported tasks + extra body)
 ```

benchmarks/tts/bench_tts.py

@@ -118,7 +118,7 @@ def build_bench_args(
         "--num-warmups",
         "2",
         "--percentile-metrics",
-        "ttft,e2el,audio_rtf,audio_ttfp,audio_duration",
+        "ttft,e2el,audio_rtf,audio_ttfp,audio_duration,audio_underrun",
     ]
 
     if resolved_dataset_path:

benchmarks/tts/bench_voxcpm_offline.py

@@ -1,922 +0,0 @@
-"""Offline VoxCPM benchmark for vLLM Omni.
-
-Supports both:
-- sync one-shot (Omni.generate)
-- streaming (AsyncOmni.generate with async_chunk config)
-- text-only synthesis
-- voice cloning
-- text/clone batch inputs from txt or jsonl
-
-Usage::
-
-    # Sync (default voice)
-    python benchmarks/tts/bench_voxcpm_offline.py \\
-        --model /path/to/VoxCPM \\
-        --text "Hello world" \\
-        --output-dir results/audio/
-
-    # Streaming (async_chunk)
-    python benchmarks/tts/bench_voxcpm_offline.py \\
-        --model /path/to/VoxCPM \\
-        --stage-configs-path vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml \\
-        --txt-prompts prompts.txt \\
-        --output-dir results/audio/
-
-    # Voice cloning batch via JSONL
-    python benchmarks/tts/bench_voxcpm_offline.py \\
-        --model /path/to/VoxCPM \\
-        --jsonl-prompts prompts.jsonl \\
-        --output-dir results/audio/
-"""
-
-from __future__ import annotations
-
-import asyncio
-import json
-import logging
-import os
-import tempfile
-import time
-import uuid
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Any
-
-import torch
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-
-from vllm_omni import AsyncOmni, Omni
-
-
-def _find_repo_root(start: Path) -> Path:
-    """Walk up from ``start`` until a repo marker is found.
-
-    Falls back to ``parents[2]`` for backwards compatibility if no marker hits
-    (which can only happen in unusual checkouts — the tree should always have
-    pyproject.toml + vllm_omni/ at the top level).
-    """
-    for candidate in [start, *start.parents]:
-        if (candidate / "pyproject.toml").is_file() and (candidate / "vllm_omni").is_dir():
-            return candidate
-    return start.parents[2]
-
-
-REPO_ROOT = _find_repo_root(Path(__file__).resolve())
-DEFAULT_STAGE_ASYNC = REPO_ROOT / "vllm_omni" / "model_executor" / "stage_configs" / "voxcpm_async_chunk.yaml"
-DEFAULT_STAGE_SYNC = REPO_ROOT / "vllm_omni" / "deploy" / "voxcpm.yaml"
-
-logger = logging.getLogger(__name__)
-
-
-@dataclass(frozen=True, slots=True)
-class PromptSpec:
-    text: str
-    label: str
-    ref_audio: str | None = None
-    ref_text: str | None = None
-
-
-def _require_soundfile():
-    try:
-        import soundfile as sf  # type: ignore
-    except ModuleNotFoundError as exc:
-        raise RuntimeError(
-            "soundfile is required to write VoxCPM benchmark WAV outputs. Install it with: pip install soundfile"
-        ) from exc
-    return sf
-
-
-def _build_prompt(
-    args,
-    *,
-    text: str,
-    ref_audio: str | None = None,
-    ref_text: str | None = None,
-    global_request_id: str | None = None,
-) -> dict[str, Any]:
-    additional_information: dict[str, list[Any]] = {
-        "text": [text],
-        "cfg_value": [args.cfg_value],
-        "inference_timesteps": [args.inference_timesteps],
-        "min_len": [args.min_len],
-        "max_new_tokens": [args.max_new_tokens],
-    }
-    if args.streaming_prefix_len is not None:
-        additional_information["streaming_prefix_len"] = [args.streaming_prefix_len]
-
-    if ref_audio:
-        additional_information["ref_audio"] = [ref_audio]
-    if ref_text:
-        additional_information["ref_text"] = [ref_text]
-    if global_request_id is not None:
-        additional_information["global_request_id"] = [global_request_id]
-
-    return {
-        "prompt_token_ids": [1],
-        "additional_information": additional_information,
-    }
-
-
-def _extract_audio_tensor(mm: dict[str, Any]) -> torch.Tensor:
-    audio = mm.get("audio", mm.get("model_outputs"))
-    if audio is None:
-        raise ValueError("No audio output found in multimodal output.")
-    if isinstance(audio, list):
-        parts = [torch.as_tensor(a).float().cpu().reshape(-1) for a in audio]
-        audio = torch.cat(parts, dim=-1) if parts else torch.zeros(0)
-    if not isinstance(audio, torch.Tensor):
-        audio = torch.as_tensor(audio)
-    return audio.float().cpu().reshape(-1)
-
-
-def _extract_sample_rate(mm: dict[str, Any]) -> int:
-    sr_raw = mm.get("sr", 24000)
-    if isinstance(sr_raw, list) and sr_raw:
-        sr_raw = sr_raw[-1]
-    if hasattr(sr_raw, "item"):
-        return int(sr_raw.item())
-    return int(sr_raw)
-
-
-def _emit_offline_metrics(
-    *,
-    request_id: str,
-    elapsed_s: float,
-    first_audio_elapsed: float | None,
-    audio_duration_s: float,
-) -> None:
-    metrics = {
-        "request_id": request_id,
-        "ttfp_ms": round(first_audio_elapsed * 1000.0, 3) if first_audio_elapsed is not None else None,
-        "audio_duration_s": round(audio_duration_s, 6),
-        "rtf": round(elapsed_s / audio_duration_s, 6) if audio_duration_s > 0 else None,
-    }
-    print(f"[OfflineMetrics] {metrics}")
-
-
-def _write_audio_tensor(output_path: Path, audio_tensor: Any, sample_rate: int) -> None:
-    sf = _require_soundfile()
-    if isinstance(audio_tensor, torch.Tensor):
-        audio_np = audio_tensor.float().cpu().clamp(-1.0, 1.0).numpy()
-    else:
-        audio_np = torch.as_tensor(audio_tensor).float().cpu().clamp(-1.0, 1.0).numpy()
-    sf.write(
-        output_path,
-        audio_np,
-        sample_rate,
-        format="WAV",
-        subtype="PCM_16",
-    )
-
-
-def _save_wav(mm: dict[str, Any], output_dir: Path, request_id: str) -> Path:
-    output_dir.mkdir(parents=True, exist_ok=True)
-    output_path = output_dir / f"output_{request_id}.wav"
-    _write_audio_tensor(output_path, _extract_audio_tensor(mm), _extract_sample_rate(mm))
-    return output_path
-
-
-def _iter_request_multimodal_outputs(request_output: Any):
-    outputs = getattr(request_output, "outputs", None)
-    if outputs:
-        for output in outputs:
-            mm = getattr(output, "multimodal_output", None)
-            if isinstance(mm, dict):
-                yield mm
-
-    mm = getattr(request_output, "multimodal_output", None)
-    if isinstance(mm, dict):
-        yield mm
-
-
-def _read_non_empty_lines(path: str) -> list[str]:
-    with open(path, encoding="utf-8") as f:
-        return [line.strip() for line in f if line.strip()]
-
-
-def _load_prompt_specs(args) -> list[PromptSpec]:
-    specs: list[PromptSpec] = []
-
-    if args.txt_prompts is not None:
-        texts = _read_non_empty_lines(args.txt_prompts)
-        if not texts:
-            raise ValueError(f"No prompts found in {args.txt_prompts}")
-        for idx, text in enumerate(texts, start=1):
-            specs.append(
-                PromptSpec(
-                    text=text,
-                    label=f"item{idx:03d}",
-                    ref_audio=args.ref_audio,
-                    ref_text=args.ref_text,
-                )
-            )
-        return specs
-
-    if args.jsonl_prompts is not None:
-        with open(args.jsonl_prompts, encoding="utf-8") as f:
-            for line_no, raw_line in enumerate(f, start=1):
-                line = raw_line.strip()
-                if not line:
-                    continue
-                try:
-                    item = json.loads(line)
-                except json.JSONDecodeError as exc:
-                    raise ValueError(f"{args.jsonl_prompts}:{line_no} is not valid JSON: {exc}") from exc
-                if not isinstance(item, dict):
-                    raise ValueError(f"{args.jsonl_prompts}:{line_no} must be a JSON object")
-
-                text = item.get("text")
-                if not isinstance(text, str) or not text.strip():
-                    raise ValueError(f"{args.jsonl_prompts}:{line_no} requires non-empty string field 'text'")
-
-                ref_audio = item.get("ref_audio", args.ref_audio)
-                ref_text = item.get("ref_text", args.ref_text)
-                if (ref_audio is None) != (ref_text is None):
-                    raise ValueError(
-                        f"{args.jsonl_prompts}:{line_no} must provide both 'ref_audio' and 'ref_text' together"
-                    )
-
-                specs.append(
-                    PromptSpec(
-                        text=text.strip(),
-                        label=f"item{len(specs) + 1:03d}",
-                        ref_audio=ref_audio,
-                        ref_text=ref_text,
-                    )
-                )
-
-        if not specs:
-            raise ValueError(f"No prompts found in {args.jsonl_prompts}")
-        return specs
-
-    specs.append(
-        PromptSpec(
-            text=args.text,
-            label="item001",
-            ref_audio=args.ref_audio,
-            ref_text=args.ref_text,
-        )
-    )
-    return specs
-
-
-def _build_prompt_for_spec(args, spec: PromptSpec, *, global_request_id: str | None = None) -> dict[str, Any]:
-    return _build_prompt(
-        args,
-        text=spec.text,
-        ref_audio=spec.ref_audio,
-        ref_text=spec.ref_text,
-        global_request_id=global_request_id,
-    )
-
-
-def _count_voice_clone_prompts(prompt_specs: list[PromptSpec]) -> int:
-    return sum(1 for spec in prompt_specs if spec.ref_audio is not None)
-
-
-def _get_warmup_specs(prompt_specs: list[PromptSpec]) -> list[PromptSpec]:
-    return prompt_specs[:1]
-
-
-def _extract_stream_finished(stage_output: Any) -> bool:
-    request_output = getattr(stage_output, "request_output", None)
-    request_finished = getattr(request_output, "finished", None)
-    if request_finished is not None:
-        return bool(request_finished)
-    return bool(getattr(stage_output, "finished", False))
-
-
-def _build_profiled_stage_config(
-    stage_configs_path: str,
-    profiler_dir: str,
-) -> str:
-    stage_config_path = Path(stage_configs_path)
-    yaml_text = stage_config_path.read_text(encoding="utf-8")
-    injected_lines: list[str] = []
-    injected_count = 0
-
-    for line in yaml_text.splitlines():
-        injected_lines.append(line)
-        if line.strip() != "engine_args:":
-            continue
-        indent = line[: len(line) - len(line.lstrip())]
-        child_indent = indent + "  "
-        grandchild_indent = child_indent + "  "
-        injected_lines.extend(
-            [
-                f"{child_indent}profiler_config:",
-                f'{grandchild_indent}profiler: "torch"',
-                f'{grandchild_indent}torch_profiler_dir: "{profiler_dir}"',
-                f"{grandchild_indent}torch_profiler_with_stack: true",
-            ]
-        )
-        injected_count += 1
-
-    if injected_count == 0:
-        raise ValueError(f"No engine_args block found in stage config: {stage_configs_path}")
-
-    tmp = tempfile.NamedTemporaryFile(
-        mode="w",
-        encoding="utf-8",
-        delete=False,
-        suffix=".yaml",
-        prefix=f"{stage_config_path.stem}_profile_",
-    )
-    tmp.write("\n".join(injected_lines) + "\n")
-    tmp.close()
-    return tmp.name
-
-
-def parse_args():
-    parser = FlexibleArgumentParser(
-        description="Offline split-stage VoxCPM inference with vLLM Omni (auto sync/streaming by stage config)"
-    )
-    parser.add_argument(
-        "--model",
-        type=str,
-        default=os.environ.get("VOXCPM_MODEL"),
-        help="Local VoxCPM model directory. Defaults to $VOXCPM_MODEL.",
-    )
-    parser.add_argument(
-        "--text",
-        type=str,
-        default="This is a split-stage VoxCPM synthesis example running on vLLM Omni.",
-        help="Text to synthesize. Ignored when --txt-prompts or --jsonl-prompts is used.",
-    )
-    parser.add_argument(
-        "--txt-prompts",
-        type=str,
-        default=None,
-        help="Path to a .txt file with one synthesis text per line.",
-    )
-    parser.add_argument(
-        "--jsonl-prompts",
-        type=str,
-        default=None,
-        help=(
-            "Path to a .jsonl file. Each line must contain at least {'text': ...}; "
-            "clone rows can also set ref_audio/ref_text, and ref_text must be the "
-            "real transcript of ref_audio."
-        ),
-    )
-    parser.add_argument(
-        "--ref-audio",
-        type=str,
-        default=None,
-        help=(
-            "Optional reference audio path for voice cloning. With --txt-prompts, "
-            "the same reference is applied to every line."
-        ),
-    )
-    parser.add_argument(
-        "--ref-text",
-        type=str,
-        default=None,
-        help=(
-            "Real transcript of the reference audio. Placeholder text or mismatched "
-            "text will usually produce noisy/electronic clone audio."
-        ),
-    )
-    parser.add_argument(
-        "--stage-configs-path",
-        type=str,
-        default=str(DEFAULT_STAGE_SYNC),
-        help="Stage config YAML path. Routing is selected only from this path.",
-    )
-    parser.add_argument(
-        "--cfg-value",
-        type=float,
-        default=2.0,
-        help="Classifier-free guidance value for VoxCPM.",
-    )
-    parser.add_argument(
-        "--inference-timesteps",
-        type=int,
-        default=10,
-        help="Number of inference timesteps.",
-    )
-    parser.add_argument(
-        "--min-len",
-        type=int,
-        default=2,
-        help="Minimum generated token length.",
-    )
-    parser.add_argument(
-        "--max-new-tokens",
-        type=int,
-        default=4096,
-        help="Maximum generated token length.",
-    )
-    parser.add_argument(
-        "--streaming-prefix-len",
-        type=int,
-        default=None,
-        help="VoxCPM streaming window (optional, streaming mode only).",
-    )
-    parser.add_argument(
-        "--output-dir",
-        type=str,
-        default=None,
-        help="Directory for output WAV files.",
-    )
-    parser.add_argument(
-        "--stage-init-timeout",
-        type=int,
-        default=600,
-        help="Stage initialization timeout in seconds.",
-    )
-    parser.add_argument(
-        "--log-stats",
-        dest="log_stats",
-        action="store_true",
-        help="Enable vLLM Omni stats logging.",
-    )
-    parser.add_argument(
-        "--no-log-stats",
-        dest="log_stats",
-        action="store_false",
-        help="Disable vLLM Omni stats logging.",
-    )
-    parser.set_defaults(log_stats=True)
-    parser.add_argument(
-        "--num-runs",
-        type=int,
-        default=1,
-        help="Number of full inference runs (same prompt each time). Default 1.",
-    )
-    parser.add_argument(
-        "--warmup-runs",
-        type=int,
-        default=0,
-        help=(
-            "Optional number of warmup passes before measured runs. Warmup uses only "
-            "the first prompt and does not save outputs."
-        ),
-    )
-    parser.add_argument(
-        "--enable-profiler",
-        action="store_true",
-        help=(
-            "Enable torch profiler for the configured stages. A temporary profiled "
-            "stage config is generated automatically."
-        ),
-    )
-    parser.add_argument(
-        "--profiler-dir",
-        type=str,
-        default=None,
-        help="Directory for profiler traces. Defaults to <output-dir>/profiler when profiling is enabled.",
-    )
-    parser.add_argument(
-        "--profiler-stages",
-        type=int,
-        nargs="*",
-        default=None,
-        help="Optional stage ids to profile. Defaults to all stages that have profiler_config.",
-    )
-    parser.add_argument(
-        "--profiler-wait-seconds",
-        type=float,
-        default=30.0,
-        help="Seconds to wait after stop_profile for trace files to flush.",
-    )
-    args = parser.parse_args()
-
-    if not args.model:
-        parser.error("--model is required unless $VOXCPM_MODEL is set")
-    if args.txt_prompts is not None and args.jsonl_prompts is not None:
-        parser.error("--txt-prompts and --jsonl-prompts are mutually exclusive")
-    if (args.ref_audio is None) != (args.ref_text is None):
-        parser.error("--ref-audio and --ref-text must be provided together")
-    if args.num_runs < 1:
-        parser.error("--num-runs must be >= 1")
-    if args.warmup_runs < 0:
-        parser.error("--warmup-runs must be >= 0")
-    if args.output_dir is None:
-        args.output_dir = (
-            "output_audio_streaming" if _is_streaming_stage_config(args.stage_configs_path) else "output_audio"
-        )
-    if args.enable_profiler and args.profiler_dir is None:
-        args.profiler_dir = str(Path(args.output_dir) / "profiler")
-    try:
-        args.prompt_specs = _load_prompt_specs(args)
-    except ValueError as exc:
-        parser.error(str(exc))
-
-    return args
-
-
-def _is_streaming_stage_config(stage_configs_path: str) -> bool:
-    cfg_name = Path(stage_configs_path).name.lower()
-    return "async_chunk" in cfg_name
-
-
-async def _collect_streaming_audio(
-    omni: AsyncOmni,
-    args: Any,
-    spec: PromptSpec,
-    request_id: str,
-    *,
-    phase_label: str,
-    prompt_index: int,
-    prompt_count: int,
-    print_prompt: bool = False,
-) -> tuple[torch.Tensor, int, float, float | None]:
-    prompt = _build_prompt_for_spec(args, spec, global_request_id=request_id)
-    delta_chunks: list[torch.Tensor] = []
-    sample_rate = 24000
-    chunk_i = 0
-    prev_total_samples = 0
-    t_start = time.perf_counter()
-    first_audio_elapsed: float | None = None
-
-    if print_prompt:
-        print(f"---prompt---:{prompt}")
-
-    async for stage_output in omni.generate(prompt, request_id=request_id):
-        mm = getattr(stage_output, "multimodal_output", None)
-        if not isinstance(mm, dict):
-            ro = getattr(stage_output, "request_output", None)
-            if ro is None:
-                continue
-            mm = getattr(ro, "multimodal_output", None)
-            if not isinstance(mm, dict) and getattr(ro, "outputs", None):
-                seq = ro.outputs[0]
-                mm = getattr(seq, "multimodal_output", None)
-        if not isinstance(mm, dict):
-            continue
-        sample_rate = _extract_sample_rate(mm)
-        try:
-            w = _extract_audio_tensor(mm)
-            n = int(w.numel())
-            if n == 0:
-                continue
-            finished = _extract_stream_finished(stage_output)
-            if n > prev_total_samples:
-                delta = w.reshape(-1)[prev_total_samples:]
-                prev_total_samples = n
-            elif finished and n == prev_total_samples:
-                delta = w.reshape(-1)[:0]
-            else:
-                delta = w.reshape(-1)
-                prev_total_samples += int(delta.numel())
-            if int(delta.numel()) > 0:
-                delta_chunks.append(delta)
-            if first_audio_elapsed is None and int(delta.numel()) > 0:
-                first_audio_elapsed = time.perf_counter() - t_start
-            logger.info(
-                "%s prompt=%d/%d chunk=%d delta_samples=%d buf_len=%d finished=%s",
-                phase_label,
-                prompt_index + 1,
-                prompt_count,
-                chunk_i,
-                int(delta.numel()),
-                n,
-                finished,
-            )
-            chunk_i += 1
-        except ValueError:
-            if not _extract_stream_finished(stage_output):
-                logger.debug("skip non-audio partial output chunk=%d", chunk_i)
-
-    if not delta_chunks:
-        raise RuntimeError("No audio chunks received; check stage config and logs.")
-
-    audio_cat = torch.cat([c.reshape(-1) for c in delta_chunks], dim=0)
-    elapsed = time.perf_counter() - t_start
-    return audio_cat, sample_rate, elapsed, first_audio_elapsed
-
-
-async def _abort_streaming_residual_work(
-    omni: AsyncOmni,
-    request_id: str,
-    *,
-    settle_seconds: float = 0.1,
-) -> None:
-    """Stop any late stage-0 work once the final audio has been collected."""
-    await omni.engine.abort_async([request_id])
-    if settle_seconds > 0:
-        await asyncio.sleep(settle_seconds)
-
-
-async def _run_streaming_single(
-    omni: AsyncOmni,
-    args: Any,
-    spec: PromptSpec,
-    output_dir: Path,
-    request_id: str,
-    *,
-    run_index: int,
-    num_runs: int,
-    prompt_index: int,
-    prompt_count: int,
-) -> Path:
-    audio_cat, sample_rate, elapsed, first_audio_elapsed = await _collect_streaming_audio(
-        omni,
-        args,
-        spec,
-        request_id,
-        phase_label=f"run={run_index + 1}/{num_runs}",
-        prompt_index=prompt_index,
-        prompt_count=prompt_count,
-        print_prompt=(run_index == 0 and prompt_index == 0),
-    )
-    await _abort_streaming_residual_work(omni, request_id)
-    output_path = output_dir / f"output_run{run_index + 1}_{spec.label}.wav"
-    _write_audio_tensor(output_path, audio_cat, sample_rate)
-    audio_duration_s = float(audio_cat.numel()) / float(sample_rate) if sample_rate > 0 else 0.0
-    ttfp_text = f", ttfp={first_audio_elapsed:.2f}s" if first_audio_elapsed is not None else ""
-    rtf_text = f", rtf={elapsed / audio_duration_s:.3f}" if audio_duration_s > 0 else ""
-    print(
-        f"Saved (streaming) run {run_index + 1}/{num_runs}, "
-        f"prompt {prompt_index + 1}/{prompt_count}: {output_path} ({elapsed:.2f}s{ttfp_text}{rtf_text})"
-    )
-    _emit_offline_metrics(
-        request_id=request_id,
-        elapsed_s=elapsed,
-        first_audio_elapsed=first_audio_elapsed,
-        audio_duration_s=audio_duration_s,
-    )
-    return output_path
-
-
-async def _run_streaming_warmup(args, omni: AsyncOmni) -> None:
-    if args.warmup_runs == 0:
-        return
-
-    warmup_specs = _get_warmup_specs(args.prompt_specs)
-    print(
-        f"Warmup: {args.warmup_runs} run(s) using the first prompt "
-        f"({len(warmup_specs)} prompt(s)); outputs will be discarded."
-    )
-    for warmup_index in range(args.warmup_runs):
-        t_warmup = time.perf_counter()
-        tasks = []
-        request_ids: list[str] = []
-        for prompt_index, spec in enumerate(warmup_specs):
-            request_id = f"warmup_stream_{warmup_index + 1}_{spec.label}_{uuid.uuid4().hex[:8]}"
-            request_ids.append(request_id)
-            tasks.append(
-                _collect_streaming_audio(
-                    omni,
-                    args,
-                    spec,
-                    request_id,
-                    phase_label=f"warmup={warmup_index + 1}/{args.warmup_runs}",
-                    prompt_index=prompt_index,
-                    prompt_count=len(warmup_specs),
-                )
-            )
-        results = await asyncio.gather(*tasks)
-        for request_id in request_ids:
-            await _abort_streaming_residual_work(omni, request_id)
-        total_samples = sum(int(audio.numel()) for audio, _, _, _ in results)
-        warmup_ttfps = [ttfp for _, _, _, ttfp in results if ttfp is not None]
-        ttfp_text = f", ttfp={min(warmup_ttfps):.2f}s" if warmup_ttfps else ""
-        print(
-            f"Warmup (streaming) {warmup_index + 1}/{args.warmup_runs} finished: "
-            f"{len(results)} prompt(s), {total_samples} sample(s) "
-            f"({time.perf_counter() - t_warmup:.2f}s{ttfp_text})"
-        )
-
-
-async def _run_streaming(args) -> list[Path]:
-    output_dir = Path(args.output_dir)
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    omni = AsyncOmni(
-        model=args.model,
-        stage_configs_path=args.stage_configs_path,
-        log_stats=args.log_stats,
-        stage_init_timeout=args.stage_init_timeout,
-    )
-
-    await _run_streaming_warmup(args, omni)
-    profiler_started = False
-    if args.enable_profiler:
-        profile_prefix = f"voxcpm_streaming_{int(time.time())}"
-        stages_text = args.profiler_stages if args.profiler_stages is not None else "all-configured"
-        print(f"Starting profiler (streaming): stages={stages_text}, dir={args.profiler_dir}")
-        await omni.start_profile(profile_prefix=profile_prefix, stages=args.profiler_stages)
-        profiler_started = True
-    t_total = time.perf_counter()
-    total_elapsed = 0.0
-    paths: list[Path] = []
-    prompt_specs: list[PromptSpec] = args.prompt_specs
-    try:
-        for run in range(args.num_runs):
-            for prompt_index, spec in enumerate(prompt_specs):
-                request_id = f"stream_{run + 1}_{spec.label}_{uuid.uuid4().hex[:8]}"
-                paths.append(
-                    await _run_streaming_single(
-                        omni,
-                        args,
-                        spec,
-                        output_dir,
-                        request_id,
-                        run_index=run,
-                        num_runs=args.num_runs,
-                        prompt_index=prompt_index,
-                        prompt_count=len(prompt_specs),
-                    )
-                )
-        total_elapsed = time.perf_counter() - t_total
-    finally:
-        if profiler_started:
-            print("Stopping profiler (streaming)...")
-            await omni.stop_profile(stages=args.profiler_stages)
-            if args.profiler_wait_seconds > 0:
-                print(f"Waiting {args.profiler_wait_seconds:.1f}s for profiler traces to flush...")
-                await asyncio.sleep(args.profiler_wait_seconds)
-
-    print(
-        f"All streaming runs finished: {args.num_runs} run(s), "
-        f"{len(prompt_specs)} prompt(s), {len(paths)} file(s) in {total_elapsed:.2f}s total"
-    )
-    return paths
-
-
-def _run_sync(args) -> list[Path]:
-    output_dir = Path(args.output_dir)
-
-    omni = Omni(
-        model=args.model,
-        stage_configs_path=args.stage_configs_path,
-        log_stats=args.log_stats,
-        stage_init_timeout=args.stage_init_timeout,
-    )
-
-    def _run_sync_single(
-        spec: PromptSpec,
-        *,
-        request_prefix: str,
-        save_outputs: bool,
-        run_index: int | None = None,
-    ) -> tuple[list[Path], int, float | None, float, float, str]:
-        global_request_id = f"{request_prefix}_{spec.label}"
-        prompt = _build_prompt_for_spec(args, spec, global_request_id=global_request_id)
-        if save_outputs and run_index == 0 and spec.label == "item001":
-            print(f"---prompt---:{prompt}")
-
-        saved_paths: list[Path] = []
-        output_count = 0
-        first_audio_elapsed: float | None = None
-        total_audio_duration_s = 0.0
-        metrics_request_id = global_request_id
-        t_start = time.perf_counter()
-        for stage_outputs in omni.generate(prompt):
-            request_output = stage_outputs.request_output
-            if request_output is None:
-                continue
-            request_output_id = getattr(request_output, "request_id", None)
-            if isinstance(request_output_id, str) and request_output_id:
-                metrics_request_id = request_output_id
-            for j, mm in enumerate(_iter_request_multimodal_outputs(request_output)):
-                output_count += 1
-                if first_audio_elapsed is None:
-                    try:
-                        audio_tensor = _extract_audio_tensor(mm)
-                        if int(audio_tensor.numel()) > 0:
-                            first_audio_elapsed = time.perf_counter() - t_start
-                        total_audio_duration_s += float(audio_tensor.numel()) / float(_extract_sample_rate(mm))
-                    except ValueError:
-                        pass
-                else:
-                    try:
-                        audio_tensor = _extract_audio_tensor(mm)
-                        total_audio_duration_s += float(audio_tensor.numel()) / float(_extract_sample_rate(mm))
-                    except ValueError:
-                        pass
-                if not save_outputs:
-                    continue
-                save_stem = f"run{run_index + 1}_{spec.label}" if j == 0 else f"run{run_index + 1}_{spec.label}_{j}"
-                saved_paths.append(_save_wav(mm, output_dir, save_stem))
-
-        if output_count == 0:
-            raise RuntimeError("No output from Omni.generate")
-        elapsed_s = time.perf_counter() - t_start
-        return saved_paths, output_count, first_audio_elapsed, elapsed_s, total_audio_duration_s, metrics_request_id
-
-    if args.warmup_runs:
-        warmup_specs = _get_warmup_specs(args.prompt_specs)
-        print(
-            f"Warmup: {args.warmup_runs} run(s) using the first prompt "
-            f"({len(warmup_specs)} prompt(s)); outputs will be discarded."
-        )
-        for warmup_index in range(args.warmup_runs):
-            t_warmup = time.perf_counter()
-            _, output_count, first_audio_elapsed, elapsed_s, audio_duration_s, _ = _run_sync_single(
-                warmup_specs[0],
-                request_prefix=f"warmup_sync{warmup_index + 1}",
-                save_outputs=False,
-            )
-            ttfp_text = f", ttfp={first_audio_elapsed:.2f}s" if first_audio_elapsed is not None else ""
-            rtf_text = f", rtf={elapsed_s / audio_duration_s:.3f}" if audio_duration_s > 0 else ""
-            print(
-                f"Warmup (sync) {warmup_index + 1}/{args.warmup_runs} finished: "
-                f"{output_count} output(s) ({time.perf_counter() - t_warmup:.2f}s{ttfp_text}{rtf_text})"
-            )
-
-    profiler_started = False
-    if args.enable_profiler:
-        profile_prefix = f"voxcpm_sync_{int(time.time())}"
-        stages_text = args.profiler_stages if args.profiler_stages is not None else "all-configured"
-        print(f"Starting profiler (sync): stages={stages_text}, dir={args.profiler_dir}")
-        omni.start_profile(profile_prefix=profile_prefix, stages=args.profiler_stages)
-        profiler_started = True
-
-    t_total = time.perf_counter()
-    total_elapsed = 0.0
-    saved_paths: list[Path] = []
-    prompt_specs: list[PromptSpec] = args.prompt_specs
-    try:
-        for run in range(args.num_runs):
-            t_run = time.perf_counter()
-            run_paths: list[Path] = []
-            for prompt_index, spec in enumerate(prompt_specs):
-                prompt_paths, _, first_audio_elapsed, elapsed_s, audio_duration_s, metrics_request_id = (
-                    _run_sync_single(
-                        spec,
-                        request_prefix=f"sync_run{run + 1}_{prompt_index + 1:03d}",
-                        save_outputs=True,
-                        run_index=run,
-                    )
-                )
-                run_paths.extend(prompt_paths)
-                ttfp_text = f", ttfp={first_audio_elapsed:.2f}s" if first_audio_elapsed is not None else ""
-                rtf_text = f", rtf={elapsed_s / audio_duration_s:.3f}" if audio_duration_s > 0 else ""
-                print(
-                    f"Saved (sync) run {run + 1}/{args.num_runs}, "
-                    f"prompt {prompt_index + 1}/{len(prompt_specs)}: {len(prompt_paths)} file(s){ttfp_text}{rtf_text}"
-                )
-                _emit_offline_metrics(
-                    request_id=metrics_request_id,
-                    elapsed_s=elapsed_s,
-                    first_audio_elapsed=first_audio_elapsed,
-                    audio_duration_s=audio_duration_s,
-                )
-
-            saved_paths.extend(run_paths)
-            print(
-                f"Run {run + 1}/{args.num_runs} finished: {len(run_paths)} file(s) ({time.perf_counter() - t_run:.2f}s)"
-            )
-            for path in run_paths:
-                print(f"  {path}")
-
-        total_elapsed = time.perf_counter() - t_total
-    finally:
-        if profiler_started:
-            print("Stopping profiler (sync)...")
-            omni.stop_profile(stages=args.profiler_stages)
-            if args.profiler_wait_seconds > 0:
-                print(f"Waiting {args.profiler_wait_seconds:.1f}s for profiler traces to flush...")
-                time.sleep(args.profiler_wait_seconds)
-
-    print(
-        f"All sync runs finished: {args.num_runs} run(s), "
-        f"{len(prompt_specs)} prompt(s), {len(saved_paths)} file(s) in {total_elapsed:.2f}s total"
-    )
-    return saved_paths
-
-
-def main(args) -> int:
-    logging.basicConfig(level=logging.INFO)
-    profiled_stage_config_path: str | None = None
-    original_stage_config_path = args.stage_configs_path
-    if args.enable_profiler:
-        Path(args.profiler_dir).mkdir(parents=True, exist_ok=True)
-        profiled_stage_config_path = _build_profiled_stage_config(
-            args.stage_configs_path,
-            str(Path(args.profiler_dir).resolve()),
-        )
-        args.stage_configs_path = profiled_stage_config_path
-
-    is_streaming = _is_streaming_stage_config(args.stage_configs_path)
-    voice_clone_count = _count_voice_clone_prompts(args.prompt_specs)
-    print(f"Model: {args.model}")
-    print(f"Stage config: {original_stage_config_path}")
-    print(f"Route: {'streaming' if is_streaming else 'sync'} (from stage-configs-path)")
-    print(f"Prompt count: {len(args.prompt_specs)}")
-    print("Batch mode: sequential (aligned with native VoxCPM)")
-    print(f"Warmup runs: {args.warmup_runs}")
-    print(f"Voice cloning prompts: {voice_clone_count}/{len(args.prompt_specs)}")
-    if args.enable_profiler:
-        print(f"Profiler: enabled (dir={args.profiler_dir}, stages={args.profiler_stages or 'all-configured'})")
-        print(f"Profiled stage config: {args.stage_configs_path}")
-    if voice_clone_count:
-        print("Voice cloning note: --ref-text/ref_text must match the spoken content of the reference audio.")
-    print(f"Num runs: {args.num_runs}")
-    try:
-        if is_streaming:
-            asyncio.run(_run_streaming(args))
-        else:
-            _run_sync(args)
-    finally:
-        if profiled_stage_config_path is not None and os.path.exists(profiled_stage_config_path):
-            os.unlink(profiled_stage_config_path)
-    return 0
-
-
-if __name__ == "__main__":
-    os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-    raise SystemExit(main(parse_args()))

docker/Dockerfile.ci

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

docker/Dockerfile.cuda

@@ -1,4 +1,4 @@
-ARG BASE_IMAGE=vllm/vllm-openai:v0.20.0
+ARG BASE_IMAGE=vllm/vllm-openai:v0.21.0
 FROM ${BASE_IMAGE}
 
 ARG COMMON_WORKDIR=/app

docker/Dockerfile.npu.ci

@@ -15,7 +15,8 @@ RUN apt-get update && \
         libsm6 \
         libxext6 \
         libxrender1 \
-        libxcb1 && \
+        libxcb1 \
+        espeak-ng && \
     rm -rf /var/cache/apt/* && \
     rm -rf /var/lib/apt/lists/*
 
@@ -29,6 +30,11 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "pytest-cov>=4.0.0" \
       "pytest-mock>=3.10.0" \
       "openpyxl>=3.0.0" \
+      "psutil>=7.2.0" \
+      "soundfile>=0.13.1" \
+      "pyttsx3>=2.99" \
+      "opencv-python>=4.12.0.88" \
+      "imageio[ffmpeg]>=0.6.0" \
       --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
       --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
       --default-timeout=60 && \

docker/Dockerfile.npu.ci.a3

@@ -15,7 +15,8 @@ RUN apt-get update && \
         libsm6 \
         libxext6 \
         libxrender1 \
-        libxcb1 && \
+        libxcb1 \
+        espeak-ng && \
     rm -rf /var/cache/apt/* && \
     rm -rf /var/lib/apt/lists/*
 
@@ -29,6 +30,11 @@ RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
       "pytest-cov>=4.0.0" \
       "pytest-mock>=3.10.0" \
       "openpyxl>=3.0.0" \
+      "psutil>=7.2.0" \
+      "soundfile>=0.13.1" \
+      "pyttsx3>=2.99" \
+      "opencv-python>=4.12.0.88" \
+      "imageio[ffmpeg]>=0.6.0" \
       --index-url http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
       --trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
       --default-timeout=60 && \

docker/Dockerfile.rocm

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

docker/Dockerfile.xpu

@@ -88,7 +88,7 @@ ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE="copy"
 
-ARG VLLM_VERSION=v0.20.0
+ARG VLLM_VERSION=v0.21.0
 RUN git clone -b ${VLLM_VERSION} https://github.com/vllm-project/vllm
 WORKDIR /workspace/vllm
 
@@ -133,6 +133,11 @@ ENTRYPOINT []
 
 FROM ${VLLM_BASE} AS vllm-omni
 
+RUN apt-get update && \
+    apt-get install -y espeak-ng ffmpeg git jq && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
 WORKDIR /workspace/vllm-omni
 COPY . .
 

docs/.nav.yml

@@ -6,10 +6,11 @@ nav:
     - getting_started/installation/*
   - Serving:
     - OpenAI-Compatible API:
-      - Diffusion Chat API: serving/diffusion_chat_api.md
+      - Chat Completions API: serving/chat_completions_api.md
       - Image Generation: serving/image_generation_api.md
       - Image Edit: serving/image_edit_api.md
       - Text to Speech: serving/speech_api.md
+      - Videos: serving/videos_api.md
       - Audio Generation: serving/audio_generate_api.md
       - Streaming Video Input: serving/video_stream_api.md
   - Examples:
@@ -79,6 +80,8 @@ nav:
       - Quantized KV Cache: user_guide/quantization/quantized_kvcache.md
       - FP8 W8A8: user_guide/quantization/fp8.md
       - Int8 W8A8: user_guide/quantization/int8.md
+      - MXFP8 W8A8: user_guide/quantization/mxfp8.md
+      - MXFP4 W4A4: user_guide/quantization/mxfp4.md
       - ModelOpt: user_guide/quantization/modelopt.md
       - GGUF: user_guide/quantization/gguf.md
       - AutoRound: user_guide/quantization/autoround.md

docs/api/README.md

@@ -99,7 +99,6 @@ Configuration classes.
 - [vllm_omni.diffusion.cache.teacache.config.TeaCacheConfig][]
 - [vllm_omni.distributed.omni_connectors.utils.config.ConnectorSpec][]
 - [vllm_omni.distributed.omni_connectors.utils.config.OmniTransferConfig][]
-- [vllm_omni.model_executor.models.cosyvoice3.config.CosyVoice3Config][]
 - [vllm_omni.model_executor.models.fish_speech.configuration_fish_speech.FishSpeechConfig][]
 - [vllm_omni.model_executor.models.fish_speech.configuration_fish_speech.FishSpeechFastARConfig][]
 - [vllm_omni.model_executor.models.fish_speech.configuration_fish_speech.FishSpeechSlowARConfig][]
@@ -116,10 +115,13 @@ Configuration classes.
 - [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1DecoderConfig][]
 - [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1DecoderDiTConfig][]
 - [vllm_omni.model_executor.models.qwen3_tts.tokenizer_25hz.configuration_qwen3_tts_tokenizer_v1.Qwen3TTSTokenizerV1EncoderConfig][]
+- [vllm_omni.transformers_utils.configs.cosyvoice3.CosyVoice3Config][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Config][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLConfig][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLTextConfig][]
 - [vllm_omni.transformers_utils.configs.mammoth_moda2.Mammothmoda2Qwen2_5_VLVisionConfig][]
+- [vllm_omni.transformers_utils.configs.omnivoice.OmniVoiceConfig][]
+- [vllm_omni.transformers_utils.configs.voxcpm2.VoxCPM2Config][]
 
 ## Workers
 

docs/community/governance.md

@@ -39,11 +39,34 @@ Committers have write access and merge rights. They typically have deep expertis
 - [@linyueqian](https://github.com/linyueqian): TTS and Omni Support
 - [@lishunyang12](https://github.com/lishunyang12): Quantization and Configuration
 - [@princepride](https://github.com/princepride): Diffusion and Omni Support
-- [@SamitHuang](https://github.com/SamitHuang): RL and Diffusion
+- [@RuixiangMa](https://github.com/RuixiangMa): Diffusion models, parallel, cache, and docs
+- [@SamitHuang](https://github.com/SamitHuang): RL, Diffusion, and cache
 - [@tzhouam](https://github.com/tzhouam): Engine and New Model Support
-- [@wtomin](https://github.com/wtomin): Diffusion and Parallelism
+- [@wtomin](https://github.com/wtomin): Diffusion models, parallel, and docs
 - [@ZeldaHuang](https://github.com/ZeldaHuang): Omni Support
-- [@ZJY0516](https://github.com/ZJY0516): Diffusion and CustomOp
+- [@ZJY0516](https://github.com/ZJY0516): Diffusion attention backend, kernel fusion, and CustomOp
+- [@yuanheng-zhao](https://github.com/yuanheng-zhao): Diffusion cache, offload, and Omni Support
+
+## Diffusion Workload Division
+
+This section breaks down diffusion-related responsibilities. If you have PRs touching these areas, please ping the listed owners for review.
+
+### Runtime-related
+
+- Diffusion models: @RuixiangMa, @wtomin
+
+### Optimization
+
+- Parallel: @RuixiangMa, @wtomin
+- Attention backend: @ZJY0516
+- Cache: @yuanheng-zhao, @RuixiangMa, @SamitHuang
+- Offload: @yuanheng-zhao
+- Quantization: @lishunyang12, @david6666666
+- Kernel fusion / communication-computation: @ZJY0516
+
+### Docs & Test
+
+- Diffusion docs: @RuixiangMa, @wtomin
 
 ## Meetings
 

docs/design/feature/diffusion_continuous_batching.md

@@ -83,7 +83,10 @@ There are two important details:
 
 Today that compatibility rule is still shape-sensitive. `height`, `width`,
 `num_frames`, and CFG-related fields remain part of the key, so different
-resolutions or incompatible guidance settings do **not** co-batch yet.
+resolutions or incompatible guidance settings do **not** co-batch yet. The
+key also covers LoRA identity (`lora_int_id`, `lora_scale`), so requests
+targeting different adapters or scales run in separate batches and the
+worker can activate exactly one adapter per step.
 
 The current batching unit is one `OmniDiffusionRequest`. Requests with
 multiple prompts do not participate in batching today.

docs/design/feature/disaggregated_inference.md

@@ -9,6 +9,7 @@ Backend-specific setup lives in separate docs:
 - [MooncakeStoreConnector](omni_connectors/mooncake_store_connector.md)
 - [MooncakeTransferEngineConnector](omni_connectors/mooncake_transfer_engine_connector.md)
 - [YuanrongConnector](omni_connectors/yuanrong_connector.md)
+- [YuanrongTransferEngineConnector](omni_connectors/yuanrong_transfer_engine_connector.md)
 
 ## Overview
 
@@ -23,6 +24,7 @@ Current connectors operate in D2H2D (device to host to device) mode.
 | Multi node (Mooncake Store) | MooncakeStoreConnector | TCP-based, requires Mooncake Master + metadata server. |
 | Multi node (Mooncake RDMA) | MooncakeTransferEngineConnector | RDMA/TCP direct transfer with managed memory pool. Fastest. |
 | Multi node (Yuanrong) | YuanrongConnector | Requires Yuanrong Datasystem + etcd. |
+| Ascend NPU P2P (Yuanrong TE) | YuanrongTransferEngineConnector | Uses Yuanrong TransferEngine directly. Configure NPU device IPv4 and `memory_pool_device: "npu"`. |
 
 ## Core API
 

docs/design/feature/omni_connectors/yuanrong_transfer_engine_connector.md

@@ -0,0 +1,211 @@
+# YuanrongTransferEngineConnector
+
+## When to Use
+
+Use `YuanrongTransferEngineConnector` for high-performance peer-to-peer KV cache
+transfer on Ascend NPU. It is different from `YuanrongConnector`:
+
+- `YuanrongConnector` uses Yuanrong Datasystem as a distributed KV store and
+  requires Datasystem workers plus etcd.
+- `YuanrongTransferEngineConnector` uses Yuanrong TransferEngine directly. It
+  registers local memory in each stage worker and lets the receiver pull data
+  from the sender with TransferEngine.
+
+The implementation currently lives under
+`vllm_omni/platforms/npu/omni_connectors/` because the supported backend is
+Ascend NPU only. The connector is still registered with the generic
+OmniConnector factory under the name `YuanrongTransferEngineConnector`.
+
+For Ascend P2P transfer, the connector should use NPU memory as its transfer
+pool. The Mooncake-style CPU pool is not supported for this connector.
+
+## Prerequisites
+
+Install Yuanrong Datasystem/TransferEngine Python bindings in the runtime
+environment. The packaged install is:
+
+```bash
+pip install openyuanrong-datasystem
+```
+
+To build from source, follow the upstream project instructions:
+
+- Source repository: [openeuler/yuanrong-datasystem](https://atomgit.com/openeuler/yuanrong-datasystem)
+- Installation guide: [Yuanrong Datasystem Linux installation](https://pages.openeuler.openatom.cn/openyuanrong-datasystem/docs/zh-cn/latest/installation/installation_linux.html)
+
+The Ascend driver, CANN runtime, HCCL/HCCP, and NPU device network must also be
+configured on every machine participating in the transfer.
+
+## Check NPU Device IPv4
+
+Yuanrong TransferEngine with `protocol: "ascend"` uses the NPU device network,
+not only the host management IP. Each participating NPU must have an IPv4 HCCN
+device IP.
+
+Check the device IP, link, and health state:
+
+```bash
+for i in {0..7}; do
+  echo "===== NPU $i ====="
+  hccn_tool -i "$i" -ip -g
+  hccn_tool -i "$i" -link -g
+  hccn_tool -i "$i" -net_health -g
+done
+```
+
+Expected output for each NPU should include an IPv4 address and healthy link:
+
+```text
+ipaddr:10.86.1.80
+netmask:255.255.255.0
+link status: UP
+net health status: Success
+```
+
+If the output says `Get ipconf failed, because no ip was preset there!`, the NPU
+device IP is not configured and Ascend P2P will fail. If the address is IPv6,
+configure IPv4 for this connector path or validate that the underlying Yuanrong
+TransferEngine build supports the IPv6 HCCN path before using it.
+
+## Check Device-to-Device Connectivity
+
+After collecting the target-stage NPU IPv4 addresses, ping them from each
+source-stage NPU:
+
+```bash
+# Example: AR uses NPU 0..3 and DiT uses NPU 4..7.
+# Replace the destination IPs with the values reported by hccn_tool -i 4..7 -ip -g.
+for src in 0 1 2 3; do
+  for dst_ip in 10.86.1.78 10.86.1.79 10.86.1.80 10.86.1.81; do
+    echo "===== NPU $src -> $dst_ip ====="
+    hccn_tool -i "$src" -ping -g address "$dst_ip" || true
+  done
+done
+```
+
+The expected result is `0.00% packet loss` for every required source/target pair.
+This verifies basic HCCN IPv4 connectivity. TransferEngine can still fail later
+if RDMA/QP/HCCL setup is broken, but failed ping means the device network must be
+fixed first.
+
+## YAML Configuration
+
+For Ascend P2P, configure the connector with `protocol: "ascend"` and an NPU
+memory pool:
+
+```yaml
+runtime:
+  enabled: true
+  defaults:
+    window_size: -1
+    max_inflight: 1
+  connectors:
+    yuanrong_te_connector:
+      name: YuanrongTransferEngineConnector
+      extra:
+        host: "auto"
+        zmq_port: 50051
+        rpc_port: "auto"
+        protocol: "ascend"
+        device_name: "auto"
+        memory_pool_size: 1073741824
+        memory_pool_device: "npu"
+  edges:
+    - from: 0
+      to: 1
+      window_size: -1
+```
+
+Important fields:
+
+| Parameter | Recommended Value | Notes |
+|---|---|---|
+| `host` | `"auto"` | Advertises a routable local host IP for ZMQ metadata exchange. |
+| `zmq_port` | `50051` | Base ZMQ port. The runtime applies stage/rank offsets. |
+| `rpc_port` | `"auto"` | Lets TransferEngine choose an available RPC port. |
+| `protocol` | `"ascend"` | Uses the Ascend TransferEngine backend. |
+| `device_name` | `"auto"` | Each TP worker resolves to its local logical NPU, for example `npu:0`, `npu:1`, etc. |
+| `memory_pool_device` | `"npu"` | Required for the Ascend fast path. This registers NPU memory with TransferEngine. |
+| `memory_pool_size` | `1073741824` | 1 GiB per worker is a practical starting point for KV-cache transfer validation. Increase if the transfer pool is exhausted. |
+
+`memory_pool_device: "cpu"` is rejected for
+`YuanrongTransferEngineConnector` with `protocol: "ascend"`. CPU pool registration
+can lead to Ascend P2P failures such as HCCL RA init errors, QP timeouts, or
+FFTS/SDMA runtime errors. `memory_pool_device: "cpu"` is common for
+`MooncakeTransferEngineConnector`, but Yuanrong TE on Ascend only supports
+`"npu"`.
+
+## 4-Card AR-to-DiT TP Example
+
+For AR 4-way TP sending KV cache to DiT 4-way TP, set the stage TP topology in
+both stage `omni_kv_config` blocks:
+
+```yaml
+stage_args:
+  - stage_id: 0
+    runtime:
+      devices: "0,1,2,3"
+    engine_args:
+      tensor_parallel_size: 4
+      omni_kv_config:
+        need_send_cache: true
+        rank_mapping:
+          from_tp: 4
+          to_tp: 4
+    output_connectors:
+      to_stage_1: yuanrong_te_connector
+
+  - stage_id: 1
+    runtime:
+      devices: "4,5,6,7"
+    engine_args:
+      parallel_config:
+        tensor_parallel_size: 4
+      omni_kv_config:
+        need_recv_cache: true
+        rank_mapping:
+          from_tp: 4
+          to_tp: 4
+    input_connectors:
+      from_stage_0: yuanrong_te_connector
+```
+
+In this topology each DiT rank receives the corresponding AR rank shard:
+
+```text
+DiT rank 0 <- AR rank 0
+DiT rank 1 <- AR rank 1
+DiT rank 2 <- AR rank 2
+DiT rank 3 <- AR rank 3
+```
+
+## Validate a Successful Run
+
+Search the inference log for Yuanrong TE receive-side messages:
+
+```bash
+grep -E "\[YR GET\]|Successfully received KV cache|transfer_engine get failed" inference.log
+```
+
+A successful Ascend fast-path transfer contains lines like:
+
+```text
+[YR GET] ... fast_path, zero-copy
+Successfully received KV cache ... across 1 key(s)
+Applied CFG KV caches
+[KV Reuse] ...
+```
+
+Sender-side `KV transfer OK` only means that the sender registered the payload
+and metadata. The actual P2P data transfer happens on the receiver side during
+`[YR GET]`.
+
+## Troubleshooting
+
+| Symptom | Likely Cause | Action |
+|---|---|---|
+| `IPv4 device IP not found` | NPU HCCN IPv4 is missing. | Run `hccn_tool -i <id> -ip -g` and configure IPv4 for every participating NPU. |
+| `P2PCommInitRootInfo failed`, `ra init failed`, `RA_QP_STATUS_TIMEOUT` | Ascend P2P/RDMA/QP setup failed. | First verify device IPv4, link health, and device-to-device ping. Then test a minimal Yuanrong TE pair outside vLLM. |
+| `fftsplus sdma error`, `context is abort` | Runtime SDMA/FFTS failure, often after an invalid or failed P2P transfer. | Use `memory_pool_device: "npu"` for Ascend TE, reduce `memory_pool_size` if OOM occurs, and isolate the failing NPU pair with a Yuanrong TE demo. |
+| Sender logs `KV transfer OK` but receiver never logs `[YR GET]` | Sender `put()` succeeded, receiver P2P read failed or timed out. | Check receiver logs and search for `transfer_engine get failed`. |
+| Pool allocation fails or OOM | Transfer pool is too small or too large for available NPU memory. | Start with 1 GiB (`1073741824`), reduce to 512 MiB for validation, or increase for larger KV payloads. |

docs/design/feature/pipeline_parallel.md

@@ -0,0 +1,356 @@
+# Pipeline Parallel
+
+This section describes how to add Pipeline Parallelism (PP) to a diffusion pipeline. We use the Wan2.2 text-to-video and
+image-to-video pipelines as the reference implementations.
+
+## Implementation Checklist
+
+Adding Pipeline Parallel support requires:
+
+1. ✅ **Inherit the mixin** - add `PipelineParallelMixin` before `CFGParallelMixin` in the pipeline class
+2. ✅ **Make stage forward resumable** - support `intermediate_tensors` in the transformer forward path
+3. ✅ **Return the right object type** - `IntermediateTensors` on non-last PP ranks, final model output on the last rank
+4. ✅ **Use the standard denoising loop** - call `predict_noise_maybe_with_cfg()` and `scheduler_step_maybe_with_cfg()`
+5. ✅ **Test parity** - compare PP results against the single-GPU baseline
+
+---
+
+## Table of Contents
+
+- [Overview](#overview)
+- [Step-by-Step Implementation](#step-by-step-implementation)
+- [Testing](#testing)
+- [Troubleshooting](#troubleshooting)
+- [Reference Implementations](#reference-implementations)
+
+---
+
+## Overview
+
+### What is Pipeline Parallelism?
+
+Pipeline Parallelism splits the denoising transformer into multiple sequential stages and places each stage on a
+different rank. Instead of every rank holding the full DiT, each PP rank owns only a slice of the layers.
+
+For each denoising step:
+
+1. Rank 0 starts the forward pass using the current latents.
+2. Each intermediate rank receives hidden states from the previous rank, runs its local layer slice, and forwards the
+   intermediate tensors downstream.
+3. The last PP rank produces the final noise prediction.
+4. The last PP rank applies the scheduler step and sends the updated latents back to rank 0 for the next timestep.
+
+This reduces per-rank model memory and enables larger diffusion transformers to run across multiple GPUs. It can also be
+combined with CFG-Parallel, where each PP pipeline carries one CFG branch.
+
+### Architecture
+
+vLLM-Omni provides `PipelineParallelMixin` to encapsulate the PP communication pattern for diffusion pipelines.
+
+| Method                            | Purpose                       | Automatic Behavior                                                              |
+|-----------------------------------|-------------------------------|---------------------------------------------------------------------------------|
+| `diffuse()`                       | Denoising loop boundary       | Wrapped by `PipelineParallelMixin` to flush pending async PP sends on exit      |
+| `predict_noise_maybe_with_cfg()`  | Predict noise with PP support | Runs partial forwards on non-last PP ranks, combines with CFG logic when needed |
+| `scheduler_step_maybe_with_cfg()` | Step scheduler with PP sync   | Runs scheduler on the last PP rank, returns updated latents to rank 0           |
+| `_sync_pp_send()`                 | Flush pending async sends     | Waits for outstanding `isend` handles before later collectives or decode        |
+
+`PipelineParallelMixin` is intentionally a pipeline-level abstraction. Your model-specific `predict_noise()` still
+defines how a local stage executes.
+
+### How It Works
+
+`predict_noise_maybe_with_cfg()` automatically switches between these modes:
+
+- **PP disabled** (`pipeline_parallel_size == 1`):
+    - Falls back to `CFGParallelMixin.predict_noise_maybe_with_cfg()`
+- **PP only** (`pipeline_parallel_size > 1`, `cfg_parallel_size == 1`):
+    - Rank 0 starts with the input latents
+    - Middle ranks receive `intermediate_tensors`, run their local layer range, and asynchronously send downstream
+    - The last rank returns the final noise prediction
+    - When CFG is enabled in this mode (sequential CFG), both the positive and negative branches run through the same
+      PP pipeline, doubling the communication volume per denoising step. Prefer PP + CFG-Parallel when
+      `cfg_parallel_size > 1` is available to avoid this overhead.
+- **PP + CFG-Parallel** (`pipeline_parallel_size > 1`, `cfg_parallel_size > 1`):
+    - Each PP pipeline carries one CFG branch
+    - The last PP rank all-gathers across the CFG group
+    - CFG combination happens on every last PP rank in the CFG group, matching the non-PP CFG-parallel behavior
+
+`scheduler_step_maybe_with_cfg()` keeps the denoising loop consistent:
+
+- **PP disabled**:
+    - Falls back to `scheduler_step_maybe_with_cfg()`
+- **PP enabled**:
+    - Only the last PP rank has `noise_pred` and runs the scheduler step
+    - The resulting latents are sent back to rank 0
+    - Rank 0 receives an `AsyncLatents` wrapper that resolves only when the tensor is next consumed
+
+This asynchronous design avoids unnecessary blocking between denoising steps.
+
+When a pipeline class is defined, `PipelineParallelMixin` wraps its `diffuse()` method and calls `_sync_pp_send()` in a
+`finally` block after `diffuse()` returns or raises. This keeps model code free of explicit PP cleanup while still
+ensuring that the final non-blocking PP send completes before decode or later collectives.
+
+---
+
+## Step-by-Step Implementation
+
+### Step 1: Inherit `PipelineParallelMixin` and `CFGParallelMixin`
+
+`PipelineParallelMixin` **requires** `CFGParallelMixin` and must appear before it in the class MRO. This is enforced at
+class definition time via `__init_subclass__`: defining a pipeline that inherits `PipelineParallelMixin` without
+`CFGParallelMixin`, or after `CFGParallelMixin`, raises a `TypeError` immediately on import.
+
+`PipelineParallelMixin` delegates noise prediction, CFG combination, and scheduler stepping to `CFGParallelMixin`, which
+supplies `predict_noise()`, `predict_noise_maybe_with_cfg()`, `scheduler_step_maybe_with_cfg()`, and
+`combine_cfg_noise()`.
+
+**Example:**
+
+```python
+from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
+from vllm_omni.diffusion.distributed.pipeline_parallel import PipelineParallelMixin
+import torch.nn as nn
+
+
+class YourPipeline(nn.Module, PipelineParallelMixin, CFGParallelMixin):
+    ...
+```
+
+The order matters: `PipelineParallelMixin` must be listed before `CFGParallelMixin` so calls to
+`predict_noise_maybe_with_cfg()` and `scheduler_step_maybe_with_cfg()` resolve to the PP-aware wrappers, while their
+`super()` calls delegate to the CFG implementation when PP is disabled or after the last PP stage.
+
+### Step 2: Make model forward and `predict_noise()` PP-aware
+
+The PP mixin injects `intermediate_tensors` into the normal `predict_noise()` call. Your model forward path must support
+two inputs:
+
+- Normal input from rank 0, typically passed as `hidden_states` or `x`
+- `intermediate_tensors` from upstream PP ranks
+
+The standard model forward pattern is:
+
+1. If `intermediate_tensors` is present, read the local hidden state from it.
+2. Run only this rank's layer slice.
+3. Return `IntermediateTensors(...)` on non-last PP ranks.
+4. Return the final model output on the last PP rank. `CFGParallelMixin.predict_noise()` already follows this contract
+   for common pipelines.
+
+**Minimal example:**
+
+```python
+from vllm.sequence import IntermediateTensors
+from vllm_omni.diffusion.distributed.parallel_state import get_pp_group
+
+
+def forward(self, hidden_states=None, intermediate_tensors=None, **kwargs):
+    if intermediate_tensors is not None:
+        hidden_states = intermediate_tensors["hidden_states"]
+
+    for i in range(self.start_layer, self.end_layer):
+        hidden_states = self.layers[i](hidden_states)
+
+    pp_group = get_pp_group()
+    if not pp_group.is_last_rank:
+        return IntermediateTensors({"hidden_states": hidden_states})
+    return (hidden_states,)
+```
+
+### Step 3: Partition the transformer layers
+
+The local module on each PP rank must expose only that rank's layer slice. In the reference tests and vLLM model
+implementations, this is typically done with vLLM utilities such as `make_layers(...)`, which fill missing ranges with
+`PPMissingLayer`. Besides partitioning the layers themselves, model authors should also wire
+`make_empty_intermediate_tensors_factory(...)` for intermediate tensor allocation and `is_pp_missing_parameter(...)` for
+PP-aware weight loading.
+
+To prepare a transformer for PP, implement the following pieces in order.
+
+#### 3.1 Split the transformer layers across PP ranks
+
+Each PP rank should own only its local layer range, typically exposed as `[start_layer, end_layer)`. In practice this is
+usually done with `make_layers(...)`, which constructs local layers and fills missing ranges with `PPMissingLayer`.
+
+The goal is:
+
+- every PP rank knows its `[start_layer, end_layer)` range
+- non-local layers do not execute on this rank
+- the forward can resume from incoming `intermediate_tensors`
+
+By default, layers are distributed evenly across PP ranks using `get_pp_indices()`. When the number of layers is not
+divisible by the PP size, remaining layers are assigned to middle partitions to balance compute and memory. You can
+override this with the `VLLM_PP_LAYER_PARTITION` environment variable to specify an exact per-rank layer count:
+
+```bash
+# Example: 40 layers across 4 PP ranks, assigning 8 / 12 / 12 / 8 layers
+export VLLM_PP_LAYER_PARTITION=8,12,12,8
+```
+
+The value must be a comma-separated list of integers whose length equals `pipeline_parallel_size` and whose sum equals
+the total number of transformer layers.
+
+#### 3.2 Expose `make_empty_intermediate_tensors`
+
+Your transformer module should expose `self.make_empty_intermediate_tensors`, typically created with
+`make_empty_intermediate_tensors_factory(...)`.
+
+This is important because PP ranks need a consistent way to allocate placeholder `IntermediateTensors` with the expected
+keys and hidden dimension.
+
+**Example:**
+
+```python
+from vllm.model_executor.models.utils import make_empty_intermediate_tensors_factory
+
+self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
+    ["hidden_states"],
+    inner_dim,
+)
+```
+
+For Wan2.2, the intermediate payload between PP stages is the token sequence stored under `"hidden_states"`, so the
+factory is created with that key and the transformer hidden size.
+
+#### 3.3 Return `IntermediateTensors` on non-last PP ranks
+
+Your model `forward()` or custom `predict_noise()` implementation should consume `intermediate_tensors` on non-first
+ranks and return `IntermediateTensors(...)` on non-last ranks.
+
+This lets each PP stage resume from the upstream hidden states and pass the local result to the next stage.
+
+#### 3.4 Skip non-local weights during `load_weights()`
+
+When a model is PP-partitioned, many parameters in the checkpoint belong to layers that do not exist on the current
+rank. `load_weights()` must therefore skip parameters for missing PP stages using `is_pp_missing_parameter(...)`.
+
+If this is not done, weight loading will either fail or incorrectly try to load tensors into `PPMissingLayer`
+placeholders.
+
+The Wan2.2 transformer is the best reference here: it uses `is_pp_missing_parameter(...)` before loading both remapped
+and fused parameters.
+
+If your model has multiple transformer variants, PP still works as long as each selected transformer obeys the same
+contract.
+
+### Step 4: Use the standard denoising contract
+
+vLLM-Omni diffusion pipelines already route denoising through `diffuse()`, `predict_noise_maybe_with_cfg()`, and
+`scheduler_step_maybe_with_cfg()`. `PipelineParallelMixin` overrides those standard helpers when
+`pipeline_parallel_size > 1`, so model integrations should not add separate PP-specific helper names or manual post-loop
+synchronization.
+
+In PP mode:
+
+- Non-last PP ranks return `None` from `predict_noise_maybe_with_cfg()` after sending `IntermediateTensors` downstream.
+- The last PP rank returns the final prediction, and with CFG-Parallel every last PP rank in the CFG group receives the
+  combined result.
+- Rank 0 receives `AsyncLatents` from `scheduler_step_maybe_with_cfg()`, which resolves only when the latents are
+  consumed.
+- Pending non-blocking sends are flushed automatically when `diffuse()` exits.
+
+---
+
+## Testing
+
+Use an offline inference script with `pipeline_parallel_size > 1`.
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+--model=Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+--width=1280 \
+--height=704 \
+--guidance-scale=5.0 \
+--prompt="Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage" \
+--output=t2v_5B_pp2.mp4 \
+--pipeline-parallel-size=2
+```
+
+For PP + CFG-Parallel together:
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+--model=Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+--width=1280 \
+--height=704 \
+--guidance-scale=5.0 \
+--prompt="Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage" \
+--output=t2v_5B_pp2_cfg2.mp4 \
+--pipeline-parallel-size=2 \
+--cfg-parallel-size=2
+```
+
+**Verify:**
+
+1. The run completes without hangs at the PP stage boundary
+2. Output quality matches the non-PP baseline within normal numerical variance
+3. Peak memory per GPU drops compared with the single-rank model
+4. There are no outstanding communication errors before decode
+
+---
+
+## Troubleshooting
+
+### Issue: `TypeError` on import — `CFGParallelMixin` missing
+
+**Symptoms:** Importing a pipeline that inherits `PipelineParallelMixin` raises:
+
+```
+TypeError: YourPipeline inherits PipelineParallelMixin but not CFGParallelMixin.
+```
+
+or:
+
+```
+TypeError: YourPipeline must inherit PipelineParallelMixin before CFGParallelMixin ...
+```
+
+**Cause:** `PipelineParallelMixin` enforces via `__init_subclass__` that the subclass also inherits `CFGParallelMixin`
+and lists `PipelineParallelMixin` first in the MRO.
+
+**Solution:** Add `CFGParallelMixin` to the pipeline's base classes after `PipelineParallelMixin`:
+
+```python
+from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
+from vllm_omni.diffusion.distributed.pipeline_parallel import PipelineParallelMixin
+
+
+class YourPipeline(nn.Module, PipelineParallelMixin, CFGParallelMixin):
+    ...
+```
+
+### Issue: Non-last PP ranks crash when calling `predict_noise`
+
+**Symptoms:** Shape or missing-input errors on ranks other than the first or last PP rank.
+
+**Cause:** The model forward path assumes direct input tensors and ignores `intermediate_tensors`.
+
+**Solution:** Update the transformer `forward()` or custom `predict_noise()` path to load hidden states from
+`intermediate_tensors` when present.
+
+### Issue: PP outputs differ from single-GPU baseline
+
+**Symptoms:** The PP run finishes but produces numerically inconsistent outputs.
+
+**Causes & Solutions:**
+
+- **Local layer partitioning is wrong**
+    - Verify each rank runs only its `[start_layer, end_layer)` slice
+- **Non-last ranks return a plain tensor instead of `IntermediateTensors`**
+    - Return `IntermediateTensors({...})` until the last PP stage
+- **CFG branch wiring is incorrect**
+    - When CFG is enabled, confirm positive and negative kwargs are passed exactly as in the non-PP path
+
+---
+
+## Reference Implementations
+
+Complete examples in the codebase:
+
+| Component               | Path                                                       | Notes                                                                                                                   |
+|-------------------------|------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------|
+| `PipelineParallelMixin` | `vllm_omni/diffusion/distributed/pipeline_parallel.py`     | Core PP communication and scheduler helpers                                                                             |
+| `CFGParallelMixin`      | `vllm_omni/diffusion/distributed/cfg_parallel.py`          | Default `predict_noise()` tuple normalization and CFG helper fallback                                                   |
+| Wan2.2 transformer      | `vllm_omni/diffusion/models/wan2_2/wan2_2_transformer.py`  | Reference for layer partitioning, `IntermediateTensors`, `make_empty_intermediate_tensors`, and PP-aware weight loading |
+| Wan2.2 T2V pipeline     | `vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2.py`     | Reference PP + CFG integration for text-to-video                                                                        |
+| Wan2.2 I2V pipeline     | `vllm_omni/diffusion/models/wan2_2/pipeline_wan2_2_i2v.py` | Reference PP + CFG integration for image-to-video                                                                       |
+| PP tests                | `tests/diffusion/distributed/test_pipeline_parallel.py`    | Baseline parity and async communication tests                                                                           |

docs/getting_started/installation/gpu.md

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

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

@@ -20,7 +20,7 @@ Therefore, it is recommended to install vLLM and vLLM-Omni with a **fresh new**
 
 vLLM-Omni is built based on vLLM. Please install it with command below.
 ```bash
-uv pip install vllm --torch-backend=auto
+uv pip install vllm==0.21.0 --torch-backend=auto
 ```
 
 #### Installation of vLLM-Omni
@@ -39,13 +39,13 @@ uv pip install 'vllm-omni[demo]'
 # --8<-- [start:build-wheel-from-source]
 
 #### Installation of vLLM
-If you do not need to modify source code of vLLM, you can directly install the stable 0.20.0 release version of the library
+If you do not need to modify source code of vLLM, you can directly install the stable 0.21.0 release version of the library
 
 ```bash
-uv pip install vllm==0.20.0 --torch-backend=auto
+uv pip install vllm==0.21.0 --torch-backend=auto
 ```
 
-The 0.20.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
+The 0.21.0 release of vLLM ships CUDA 13.0-compatible binaries by default. If you need a different CUDA variant or want to reuse an existing PyTorch installation, build vLLM from source instead.
 
 #### Installation of vLLM-Omni
 Since vllm-omni is rapidly evolving, it's recommended to install it from source
@@ -66,12 +66,12 @@ If you want to check, modify or debug with source code of vLLM, install the libr
 ```bash
 git clone https://github.com/vllm-project/vllm.git
 cd vllm
-git checkout v0.20.0
+git checkout v0.21.0
 ```
 Set up environment variables to get pre-built wheels. If there are internet problems, just download the whl file manually. And set `VLLM_PRECOMPILED_WHEEL_LOCATION` as your local absolute path of whl file.
 ```bash
-#For CUDA 13.0 (the default for v0.20.0; the wheel filename has no `+cu130` suffix)
-export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.20.0/vllm-0.20.0-cp38-abi3-manylinux_2_35_x86_64.whl
+#For CUDA 13.0 (the default for v0.21.0; the wheel filename has no `+cu130` suffix)
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://github.com/vllm-project/vllm/releases/download/v0.21.0/vllm-0.21.0-cp38-abi3-manylinux_2_35_x86_64.whl
 ```
 Install vllm with command below (If you have no existing PyTorch).
 ```bash
@@ -102,11 +102,11 @@ docker run --runtime nvidia --gpus 2 \
     --env "HF_TOKEN=$HF_TOKEN" \
     -p 8091:8091 \
     --ipc=host \
-    vllm/vllm-omni:v0.18.0 \
-    --model Qwen/Qwen3-Omni-30B-A3B-Instruct --port 8091
+    vllm/vllm-omni:v0.21.0 \
+    vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
 ```
 
 !!! tip
-    You can use this docker image to serve models the same way you would with in vLLM! To do so, make sure you overwrite the default entrypoint (`vllm serve --omni`) which works only for models supported in the vLLM-Omni project.
+    The CUDA image does not define a default entrypoint, so include `vllm serve ... --omni` after the image name.
 
 # --8<-- [end:pre-built-images]

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

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

docs/getting_started/quickstart.md

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

docs/models/supported_models.md

@@ -58,6 +58,7 @@ th {
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-CustomVoice | `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-VoiceDesign | `Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
 | `Qwen3TTSForConditionalGeneration` | Qwen3-TTS-12Hz-1.7B-Base | `Qwen/Qwen3-TTS-12Hz-0.6B-Base` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
+| `GLMTTSForConditionalGeneration` | GLM-TTS | `zai-org/GLM-TTS` | ✅︎ | | | |
 | `NextStep11Pipeline` | NextStep-1.1 | `stepfun-ai/NextStep-1.1` | ✅︎ | ✅︎ | | ✅︎ |
 | `MiMoAudioModel` | MiMo-Audio-7B-Instruct | `XiaomiMiMo/MiMo-Audio-7B-Instruct` | ✅︎ | ✅︎ | | |
 | `MiMoV2ASRForCausalLM` | MiMo-V2.5-ASR | `XiaomiMiMo/MiMo-V2.5-ASR` | ✅︎ | ✅︎ | | |
@@ -71,4 +72,6 @@ th {
 | `CovoAudioForConditionalGeneration` | Covo-Audio-Chat | `tencent/Covo-Audio-Chat` | ✅︎ | | | |
 |`DyninOmniForConditionalGeneration` | Dynin-Omni | `snu-aidas/Dynin-Omni` | ✅︎ | | | |
 | `ErnieImagePipeline` | ERNIE-Image | `baidu/ERNIE-Image`, `baidu/ERNIE-Image-Turbo` | ✅︎ | ✅︎ | ✅︎ | ✅︎ |
+|`HiDreamImagePipeline` | HiDream-I1-Full | `HiDream-ai/HiDream-I1-Full` | ✅︎ | ✅︎ | | |
+
 ✅︎ indicates the model is supported on that backend. Empty cells mean not listed as supported on that backend.

docs/serving/chat_completions_api.md

@@ -1,4 +1,4 @@
-# Diffusion Chat Completions API
+# Chat Completions API
 
 vLLM-Omni supports generating and editing images via the `/v1/chat/completions`
 endpoint using diffusion models. This page explains how to pass generation

docs/serving/speech_api.md

@@ -5,6 +5,9 @@ vLLM-Omni provides an OpenAI-compatible API for text-to-speech (TTS) generation.
 - **Qwen3-TTS** (`Qwen/Qwen3-TTS-12Hz-*`) -- Qwen3-based TTS with CustomVoice, VoiceDesign, and Base (voice cloning) task types. Output: 24 kHz.
 - **Fish Speech S2 Pro** (`fishaudio/s2-pro`) -- Dual-AR TTS with DAC codec. Supports text-to-speech and voice cloning via reference audio. Output: 44.1 kHz.
 - **Voxtral TTS** (`mistralai/Voxtral-4B-TTS-2603`) -- AR + FlowMatching TTS with preset voices. Output: 24 kHz.
+- **CosyVoice3** (`FunAudioLLM/Fun-CosyVoice3-0.5B-2512`) -- 2-stage talker + flow-matching code2wav. Voice cloning via `ref_audio` + `ref_text` (no presets). Output: 24 kHz.
+
+See the [Supported Models](#supported-models) section below for the full list, including OmniVoice, VoxCPM2, and MOSS-TTS-Nano.
 
 Each server instance runs a single model (specified at startup via `vllm serve <model> --omni`).
 
@@ -26,6 +29,10 @@ vllm serve fishaudio/s2-pro --omni --port 8091
 
 # Voxtral TTS
 vllm serve mistralai/Voxtral-4B-TTS-2603 --omni --port 8091
+
+# CosyVoice3 (voice cloning only — supply ref_audio + ref_text per request)
+vllm serve FunAudioLLM/Fun-CosyVoice3-0.5B-2512 \
+    --omni --port 8091 --trust-remote-code
 ```
 
 ### Generate Speech

docs/serving/videos_api.md

@@ -0,0 +1,152 @@
+# Videos API
+
+vLLM-Omni provides an OpenAI-compatible video generation API for diffusion
+video models. The API supports asynchronous video jobs through `/v1/videos` and
+a synchronous benchmark-oriented endpoint through `/v1/videos/sync`.
+
+Each server instance runs a single model specified at startup with
+`vllm serve <model> --omni`.
+
+## Quick Start
+
+### Start the Server
+
+```bash
+vllm serve Wan-AI/Wan2.2-T2V-A14B-Diffusers --omni --port 8091
+```
+
+### Create a Video Job
+
+```bash
+create_response=$(curl -s http://localhost:8091/v1/videos \
+  -F "prompt=A cinematic tracking shot of a mountain lake at sunrise" \
+  -F "width=1280" \
+  -F "height=720" \
+  -F "num_frames=80" \
+  -F "fps=16" \
+  -F "num_inference_steps=40")
+
+video_id=$(echo "${create_response}" | jq -r '.id')
+```
+
+### Poll and Download
+
+```bash
+curl -s "http://localhost:8091/v1/videos/${video_id}" | jq .
+curl -L "http://localhost:8091/v1/videos/${video_id}/content" -o output.mp4
+```
+
+## API Reference
+
+### Endpoints
+
+| Endpoint | Method | Description |
+|----------|--------|-------------|
+| `/v1/videos` | `POST` | Create an asynchronous video generation job |
+| `/v1/videos/sync` | `POST` | Generate a video synchronously and return raw video bytes |
+| `/v1/videos/{video_id}` | `GET` | Retrieve job status and metadata |
+| `/v1/videos` | `GET` | List stored video jobs |
+| `/v1/videos/{video_id}/content` | `GET` | Download generated video content |
+| `/v1/videos/{video_id}` | `DELETE` | Delete a video job and stored output |
+
+### Request Parameters
+
+`POST /v1/videos` and `POST /v1/videos/sync` accept `multipart/form-data`.
+
+#### OpenAI-style fields
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `prompt` | string | **required** | Text prompt for video generation |
+| `model` | string | server's model | Optional model name |
+| `seconds` | string | null | Requested clip duration in seconds |
+| `size` | string | null | Requested output size in `WIDTHxHEIGHT` format |
+| `user` | string | null | Optional user identifier |
+
+#### vLLM-Omni extension fields
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `input_reference` | file | null | Uploaded reference image for image-to-video requests |
+| `image_reference` | string | null | JSON-encoded reference image payload; do not combine with `input_reference` |
+| `width` | integer | model default | Output video width |
+| `height` | integer | model default | Output video height |
+| `num_frames` | integer | model default | Number of generated frames |
+| `fps` | integer | model default | Output frames per second |
+| `num_inference_steps` | integer | model default | Number of diffusion steps |
+| `guidance_scale` | number | null | CFG guidance scale for the low-noise stage |
+| `guidance_scale_2` | number | null | CFG guidance scale for the high-noise stage |
+| `boundary_ratio` | number | null | Boundary split ratio for multi-stage denoising |
+| `flow_shift` | number | null | Scheduler flow-shift value |
+| `true_cfg_scale` | number | null | True CFG scale when supported by the model |
+| `seed` | integer | null | Random seed for reproducibility |
+| `negative_prompt` | string | null | Text describing what to avoid in the generated video |
+| `enable_frame_interpolation` | boolean | null | Enable post-generation frame interpolation |
+| `frame_interpolation_exp` | integer | null | Interpolation exponent; `1=2x`, `2=4x`, and so on |
+| `frame_interpolation_scale` | number | null | RIFE inference scale |
+| `frame_interpolation_model_path` | string | null | Local path or Hugging Face repo for the interpolation model |
+| `lora` | string | null | JSON-encoded LoRA configuration object |
+| `extra_params` | string | null | JSON-encoded object for additional model-specific parameters |
+
+### Create Response
+
+`POST /v1/videos` returns a job record:
+
+```json
+{
+  "id": "video-123",
+  "status": "queued",
+  "created_at": 1701234567
+}
+```
+
+The final content is available from `/v1/videos/{video_id}/content` after the
+job status becomes `completed`.
+
+### Synchronous Response
+
+`POST /v1/videos/sync` blocks until generation finishes and returns raw video
+bytes. It is useful for benchmarks and simple scripts that do not need job
+storage or polling.
+
+## Examples
+
+### Image-to-Video
+
+```bash
+curl -s http://localhost:8091/v1/videos \
+  -F "prompt=animate this image with subtle camera movement" \
+  -F "input_reference=@input.png" \
+  -F "width=1280" \
+  -F "height=720" \
+  -F "num_frames=80" \
+  -F "fps=16"
+```
+
+### Synchronous Generation
+
+```bash
+curl -X POST http://localhost:8091/v1/videos/sync \
+  -F "prompt=A small robot walking through a neon city" \
+  -F "width=854" \
+  -F "height=480" \
+  -F "num_frames=80" \
+  -F "fps=16" \
+  -o output.mp4
+```
+
+## Storage
+
+Set `VLLM_OMNI_STORAGE_PATH` to control where asynchronous video outputs are
+stored:
+
+```bash
+export VLLM_OMNI_STORAGE_PATH=/var/tmp/vllm-omni-videos
+```
+
+## Model-Specific Examples
+
+For complete text-to-video and image-to-video walkthroughs, see:
+
+- [Text-to-Video](../user_guide/examples/online_serving/text_to_video.md)
+- [Image-to-Video](../user_guide/examples/online_serving/image_to_video.md)

docs/user_guide/diffusion/frame_interpolation.md

@@ -28,7 +28,6 @@ Frame interpolation is currently supported for:
 
 - `WanPipeline` (Wan2.2 text-to-video)
 - `WanImageToVideoPipeline`
-- `Wan22TI2VPipeline`
 
 ## Request Parameters
 

docs/user_guide/diffusion/parallelism/overview.md

@@ -4,13 +4,14 @@ This guide covers the parallelism methods in vLLM-Omni for speeding up diffusion
 
 ## Supported Methods
 
-| Method | Description |
-|--------|-------------|
-| **[Tensor Parallelism](tensor_parallel.md)** | Shards DiT weights across GPUs to reduce per-GPU memory |
-| **[Sequence Parallelism](sequence_parallel.md)** | Splits sequence dimension across GPUs (Ulysses-SP, Ring-Attention, or hybrid) for high-resolution images and videos |
-| **[CFG-Parallel](cfg_parallel.md)** | Runs CFG positive/negative branches on separate GPUs for ~1.8x speedup on guided generation |
-| **[VAE Patch Parallelism](vae_patch_parallel.md)** | Distributes VAE decode spatially across GPUs to reduce peak VAE memory |
-| **[HSDP](hsdp.md)** | Shards full model weights via PyTorch FSDP2 to enable large-model inference on memory-constrained GPUs |
-| **[Expert Parallelism](expert_parallel.md)** | Shards MoE expert blocks across GPUs for MoE models (e.g. HunyuanImage3.0) |
+| Method                                             | Description                                                                                                         |
+|----------------------------------------------------|---------------------------------------------------------------------------------------------------------------------|
+| **[Tensor Parallelism](tensor_parallel.md)**       | Shards DiT weights across GPUs to reduce per-GPU memory                                                             |
+| **[Sequence Parallelism](sequence_parallel.md)**   | Splits sequence dimension across GPUs (Ulysses-SP, Ring-Attention, or hybrid) for high-resolution images and videos |
+| **[CFG-Parallel](cfg_parallel.md)**                | Runs CFG positive/negative branches on separate GPUs for ~1.8x speedup on guided generation                         |
+| **[Pipeline Parallelism](pipeline_parallel.md)**   | Splits the denoising transformer block-wise across sequential GPU stages to reduce per-GPU model memory             |
+| **[VAE Patch Parallelism](vae_patch_parallel.md)** | Distributes VAE decode spatially across GPUs to reduce peak VAE memory                                              |
+| **[HSDP](hsdp.md)**                                | Shards full model weights via PyTorch FSDP2 to enable large-model inference on memory-constrained GPUs              |
+| **[Expert Parallelism](expert_parallel.md)**       | Shards MoE expert blocks across GPUs for MoE models (e.g. HunyuanImage3.0)                                          |
 
 See [Supported Models](../../diffusion_features.md#supported-models) for per-model compatibility.

docs/user_guide/diffusion/parallelism/pipeline_parallel.md

@@ -0,0 +1,205 @@
+# Pipeline Parallelism Guide
+
+## Table of Content
+
+- [Overview](#overview)
+- [Quick Start](#quick-start)
+- [Example Script](#example-script)
+- [Configuration Parameters](#configuration-parameters)
+- [Best Practices](#best-practices)
+- [Troubleshooting](#troubleshooting)
+- [Summary](#summary)
+
+---
+
+## Overview
+
+Pipeline Parallelism splits the denoising transformer block-wise into sequential stages across GPUs. Each rank owns only
+part of the transformer, which reduces per-GPU model memory and enables larger diffusion models to run across multiple
+devices.
+
+It can also be combined with other distributed methods such as CFG-Parallel, Tensor Parallelism, and Sequence
+Parallelism.
+
+See supported models list in [Supported Models](../../diffusion_features.md#supported-models).
+
+---
+
+## Quick Start
+
+### Basic Usage
+
+Simplest working example:
+
+```python
+from vllm_omni import Omni
+from vllm_omni.diffusion.data import DiffusionParallelConfig
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(
+    model="Wan-AI/Wan2.2-TI2V-5B-Diffusers",
+    parallel_config=DiffusionParallelConfig(
+        pipeline_parallel_size=2,
+    ),
+)
+
+outputs = omni.generate(
+    {"prompt": "A cinematic drone shot over snowy mountains"},
+    OmniDiffusionSamplingParams(
+        num_inference_steps=40,
+        num_frames=81,
+        height=704,
+        width=1280,
+    ),
+)
+```
+
+---
+
+## Example Script
+
+### Offline Inference
+
+Use python scripts under:
+
+- `examples/offline_inference/text_to_video/text_to_video.py`
+- `examples/offline_inference/image_to_video/image_to_video.py`
+
+Text-to-video example:
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+--model=Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+--width=1280 \
+--height=704 \
+--guidance-scale=5.0 \
+--prompt="Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage" \
+--output=t2v_5B_pp2.mp4 \
+--pipeline-parallel-size=2
+```
+
+Pipeline Parallelism can also be combined with CFG-Parallel:
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+--model=Wan-AI/Wan2.2-TI2V-5B-Diffusers \
+--width=1280 \
+--height=704 \
+--guidance-scale=5.0 \
+--prompt="Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage" \
+--output=t2v_5B_pp2_cfg2.mp4 \
+--pipeline-parallel-size=2 \
+--cfg-parallel-size=2
+```
+
+### Online Serving
+
+Enable Pipeline Parallelism in online serving:
+
+```bash
+# Default PP configuration
+vllm serve Wan-AI/Wan2.2-TI2V-5B-Diffusers --omni --port 8091 --pipeline-parallel-size 2
+
+# PP + CFG-Parallel
+vllm serve Wan-AI/Wan2.2-TI2V-5B-Diffusers --omni --port 8091 \
+  --pipeline-parallel-size 2 \
+  --cfg-parallel-size 2
+```
+
+---
+
+## Configuration Parameters
+
+In `DiffusionParallelConfig`
+
+| Parameter                | Type | Default | Description                                                                                                      |
+|--------------------------|------|---------|------------------------------------------------------------------------------------------------------------------|
+| `pipeline_parallel_size` | int  | 1       | Number of pipeline-parallel stages. Set to a value greater than 1 to split the denoising transformer across GPUs |
+
+> [!NOTE]
+> Total GPU count is the product of all enabled distributed dimensions, for example
+`pipeline_parallel_size * cfg_parallel_size * tensor_parallel_size * ulysses_degree * ring_degree`.
+
+### Manual Layer Partitioning
+
+By default, transformer layers are distributed evenly across PP ranks. You can override this with the
+`VLLM_PP_LAYER_PARTITION` environment variable to assign a specific number of layers to each rank:
+
+```bash
+# Example: 40 layers across 4 PP ranks, assigning 8 / 12 / 12 / 8 layers
+export VLLM_PP_LAYER_PARTITION=8,12,12,8
+```
+
+The value must be a comma-separated list of integers whose length equals `pipeline_parallel_size` and whose sum equals
+the total number of transformer layers. This is useful when you want to balance memory or compute asymmetrically across
+ranks.
+
+
+---
+
+## Best Practices
+
+### When to Use
+
+**Good for:**
+
+- Large diffusion transformers that do not fit comfortably on one GPU
+- Multi-GPU setups where reducing per-GPU model memory is more important than minimizing communication
+- Combining with CFG-Parallel or other distributed methods on supported models
+
+**Not for:**
+
+- Single GPU setups
+- Models that do not support Pipeline Parallelism
+  (check [supported models](../../diffusion_features.md#supported-models))
+- Very small models where inter-stage communication overhead may outweigh the benefit
+
+### Expected Behavior
+
+Pipeline Parallelism primarily reduces per-GPU memory usage by splitting the transformer into stage-local blocks.
+Depending on the model, topology, and resolution, it may also help execution fit into available hardware, but it is not
+primarily a latency optimization.
+
+---
+
+## Troubleshooting
+
+### Common Issue 1: No benefit from Pipeline Parallelism
+
+**Symptoms**: PP is enabled, but latency does not improve or becomes slightly worse.
+
+**Solutions**:
+
+1. **Check your goal:**
+
+```python
+# PP is mainly for memory scaling, not guaranteed latency speedup
+parallel_config = DiffusionParallelConfig(pipeline_parallel_size=2)
+```
+
+2. **Check model support:**
+    - Verify your model in [supported models](../../diffusion_features.md#supported-models)
+    - PP is currently validated only on selected pipelines
+
+3. **Combine with other methods when appropriate:**
+    - PP can be combined with CFG-Parallel, Tensor Parallelism, or Sequence Parallelism on supported models
+
+### Common Issue 2: PP pipeline fails at import
+
+**Symptoms**: Importing a custom pipeline raises a `TypeError` about `CFGParallelMixin` or mixin order.
+
+**Solutions**:
+
+1. Inherit both mixins.
+2. Put `PipelineParallelMixin` before `CFGParallelMixin` in the class base list.
+3. Use `class YourPipeline(nn.Module, PipelineParallelMixin, CFGParallelMixin): ...` as the reference pattern.
+
+---
+
+## Summary
+
+1. ✅ **Enable Pipeline Parallelism** - Set `pipeline_parallel_size > 1` in `DiffusionParallelConfig`
+2. ✅ **Use Supported Models** - Verify your model supports PP in
+   [supported models](../../diffusion_features.md#supported-models)
+3. ✅ **Combine When Needed** - PP can be combined with CFG-Parallel and other distributed methods on supported pipelines
+4. ✅ **Scale for Memory** - Use PP primarily to reduce per-GPU model memory and fit larger transformers

docs/user_guide/diffusion/step_execution.md

@@ -60,6 +60,9 @@ batching only becomes relevant when `--max-num-seqs > 1`.
 - `cache_backend` is not supported together with step execution.
 - Unsupported pipelines fail early during model loading.
 - Request-mode extras such as KV transfer are not wired into step mode yet.
+- LoRA is supported in step mode, but each batch must use a single adapter:
+  requests with different `lora_request` or `lora_scale` are scheduled into
+  separate batches.
 
 ## When To Use It
 

docs/user_guide/diffusion_features.md

@@ -35,14 +35,15 @@ Cache methods trade minimal quality for significant speedup. Quality loss is typ
 
 Parallelism methods distribute computation across GPUs without quality loss (mathematically equivalent to single-GPU).
 
-| Method | Description | Best For |
-|--------|-------------|----------|
-| **[Ulysses-SP](diffusion/parallelism/sequence_parallel.md)** | Sequence parallelism via all-to-all communication | High-resolution images (>1536px) or long videos with 2-8 GPUs |
-| **[Ring-Attention](diffusion/parallelism/sequence_parallel.md)** | Sequence parallelism via ring-based communication | Videos, very long sequences, memory-constrained, with 2-8 GPUs |
-| **[CFG-Parallel](diffusion/parallelism/cfg_parallel.md)** | Splits CFG positive/negative branches across devices | Image editing with CFG guidance (true_cfg_scale > 1) on 2 GPUs |
-| **[Tensor Parallelism](diffusion/parallelism/tensor_parallel.md)** | Shards model weights across devices | Large models that don't fit in single GPU, with 2+ GPUs |
-| **[HSDP](diffusion/parallelism/hsdp.md)** | Weight sharding via FSDP2, redistributed on-demand at runtime | Very large models (14B+) on limited VRAM, combinable with SP |
-| **[Expert Parallelism](diffusion/parallelism/expert_parallel.md)** | Shards MoE expert MLP blocks across devices | MoE diffusion models (e.g., HunyuanImage3.0) |
+| Method                                                                 | Description                                                              | Best For                                                          |
+|------------------------------------------------------------------------|--------------------------------------------------------------------------|-------------------------------------------------------------------|
+| **[Ulysses-SP](diffusion/parallelism/sequence_parallel.md)**           | Sequence parallelism via all-to-all communication                        | High-resolution images (>1536px) or long videos with 2-8 GPUs     |
+| **[Ring-Attention](diffusion/parallelism/sequence_parallel.md)**       | Sequence parallelism via ring-based communication                        | Videos, very long sequences, memory-constrained, with 2-8 GPUs    |
+| **[CFG-Parallel](diffusion/parallelism/cfg_parallel.md)**              | Splits CFG positive/negative branches across devices                     | Image editing with CFG guidance (true_cfg_scale > 1) on 2 GPUs    |
+| **[Tensor Parallelism](diffusion/parallelism/tensor_parallel.md)**     | Shards model weights across devices                                      | Large models that don't fit in single GPU, with 2+ GPUs           |
+| **[Pipeline Parallelism](diffusion/parallelism/pipeline_parallel.md)** | Splits the denoising transformer block-wise across sequential GPU stages | Large diffusion transformers that need lower per-GPU model memory |
+| **[HSDP](diffusion/parallelism/hsdp.md)**                              | Weight sharding via FSDP2, redistributed on-demand at runtime            | Very large models (14B+) on limited VRAM, combinable with SP      |
+| **[Expert Parallelism](diffusion/parallelism/expert_parallel.md)**     | Shards MoE expert MLP blocks across devices                              | MoE diffusion models (e.g., HunyuanImage3.0)                      |
 
 #### Startup Optimization
 
@@ -105,32 +106,33 @@ The following tables show which models support each feature:
 
 ### ImageGen
 
-| Model | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
-|-------|:----------:|:-----------:|:---------------------:|:--------------:|:-----------------:|:------:|:------------------------:|:--------------------:|:--------------:|:----------------:|
-| **Bagel** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅  | ❌ | ❌ | ❌ |
-| **FLUX.1-dev** | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-| **FLUX.1-schnell** | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-| **FLUX.2-klein** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-| **FLUX.1-Kontext-dev** | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-| **FLUX.2-dev** | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-| **GLM-Image** | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-| **HunyuanImage3** | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-| **LongCat-Image** | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-| **LongCat-Image-Edit** | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-| **MagiHuman** | ❌ | ❌ | ❌ | ❓ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-| **MammothModa2(T2I)** | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-| **Nextstep_1(T2I)** | ❓ | ❓ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-| **OmniGen2** | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-| **Ovis-Image** | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-| **Qwen-Image** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ✅ | ✅ |
-| **Qwen-Image-2512** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ✅ | ✅ |
-| **Qwen-Image-Edit** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ |
-| **Qwen-Image-Edit-2509** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ✅ | ❌ | ❌ |
-| **Qwen-Image-Layered** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ |
-| **SenseNova-U1** | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-| **Stable-Diffusion3.5** | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ (decode) | ❌ | ❌ |
-| **Z-Image** | ✅ | ✅ | ✅ | ❓ | ✅ (TP=2 only) | ✅ | ❌ | ✅ (decode) | ✅ | ❌ |
-| **ERNIE-Image** | ❌ | ✅ | ✅ | ❓ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+| Model                    | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | 🔀Pipeline-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
+|--------------------------|:---------:|:----------:|:---------------------:|:--------------:|:-----------------:|:-------------------:|:------:|:-------------------------:|:--------------------:|:--------------:|:----------------:|
+| **Bagel**                |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **FLUX.1-dev**           |     ✅     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ✅        |        ❌         |
+| **FLUX.1-schnell**       |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ✅        |        ❌         |
+| **FLUX.2-klein**         |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ✅        |        ❌         |
+| **FLUX.1-Kontext-dev**   |     ❌     |     ❌      |           ❌           |       ❌        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **FLUX.2-dev**           |     ✅     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **GLM-Image**            |     ❌     |     ❌      |           ❌           |       ✅        |         ✅         |          ❌          |   ✅    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **Hidream-I1-Full**        |     ❌     |     ❌      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **HunyuanImage3**        |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ✅        |        ❌         |
+| **LongCat-Image**        |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **LongCat-Image-Edit**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **MagiHuman**            |     ❌     |     ❌      |           ❌           |       ❓        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **MammothModa2(T2I)**    |     ❌     |     ❌      |           ❌           |       ❌        |         ❌         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **Nextstep_1(T2I)**      |     ❓     |     ❓      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **OmniGen2**             |     ❌     |     ✅      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **Ovis-Image**           |     ❌     |     ✅      |           ❌           |       ✅        |         ❌         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **Qwen-Image**           |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ✅        |        ✅         |
+| **Qwen-Image-2512**      |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ✅        |        ✅         |
+| **Qwen-Image-Edit**      |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
+| **Qwen-Image-Edit-2509** |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |        ✅ (decode)         |          ✅           |       ❌        |        ❌         |
+| **Qwen-Image-Layered**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
+| **SenseNova-U1**         |     ❌     |     ❌      |           ❌           |       ❌        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **Stable-Diffusion3.5**  |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
+| **Z-Image**              |     ✅     |     ✅      |           ✅           |       ❓        |   ✅ (TP=2 only)   |          ❌          |   ✅    |             ❌             |      ✅ (decode)      |       ✅        |        ❌         |
+| **ERNIE-Image**          |     ❌     |     ✅      |           ✅           |       ❓        |         ✅         |          ❌          |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
 
 > Notes:
 > 1. Nextstep_1(T2I) does not support cache acceleration methods such as TeaCache or Cache-DiT.
@@ -138,15 +140,15 @@ The following tables show which models support each feature:
 
 ### VideoGen
 
-| Model | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
-|-------|:----------:|:-----------:|:---------------------:|:--------------:|:-----------------:|:------:|:------------------------:|:--------------------:|:--------------:|:----------------:|
-| **Wan2.2** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (encode/decode) | ❌ | ❌ |
-| **Wan2.1-VACE** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ❌ | ❌ |
-| **LTX-2** | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-| **LTX-2.3** | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
-| **Helios** | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-| **HunyuanVideo-1.5 T2V I2V** | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ (decode) | ✅ | ❌ |
-| **DreamID-Omni** | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+| Model                        | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | Pipeline-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
+|------------------------------|:---------:|:----------:|:---------------------:|:--------------:|:-----------------:|:-----------------:|:------:|:-------------------------:|:--------------------:|:--------------:|:----------------:|
+| **Wan2.2**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ✅         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ❌        |        ❌         |
+| **Wan2.1-VACE**              |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
+| **LTX-2**                    |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ❌        |         ✅         |         ❌         |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
+| **Helios**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **HunyuanVideo-1.5 T2V I2V** |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ✅        |        ❌         |
+| **DreamID-Omni**             |     ❌     |     ❌      |           ❌           |       ✅        |         ❌         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
 
 **Frame Interpolation Support**
 
@@ -155,9 +157,9 @@ The following tables show which models support each feature:
 
 ### AudioGen
 
-| Model | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
-|-------|:----------:|:-----------:|:---------------------:|:--------------:|:-----------------:|:------:|:------------------------:|:--------------------:|:--------------:|:----------------:|
-| **Stable-Audio-Open** | ✅ | ❌ | ❓ | ❓ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ |
+| Model                 | ⚡TeaCache | ⚡Cache-DiT | 🔀SP (Ulysses & Ring) | 🔀CFG-Parallel | 🔀Tensor-Parallel | 🔀Pipeline-Parallel | 🔀HSDP | 💾CPU Offload (Layerwise) | 💾VAE-Patch-Parallel | 💾Quantization | 🔄Step Execution |
+|-----------------------|:---------:|:----------:|:---------------------:|:--------------:|:-----------------:|:-------------------:|:------:|:-------------------------:|:--------------------:|:--------------:|:----------------:|
+| **Stable-Audio-Open** |     ✅     |     ❌      |           ❓           |       ❓        |         ❌         |          ❌          |   ✅    |             ✅             |          ❌           |       ✅        |        ❌         |
 
 
 ## Feature Compatibility
@@ -183,7 +185,7 @@ The following tables show which models support each feature:
 | **💾VAE Patch Parallel** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | | | | |
 | **💾FP8 Quant** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❓ | ❓ | ✅ | ✅ | ✅ | | | |
 | **🔧LoRA Inference** | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | ❓ | | |
-| **🔄Step Execution** | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❓ | ❓ | ✅ | ❓ | ✅ | ✅ | ❌ | |
+| **🔄Step Execution** | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❓ | ❓ | ✅ | ❓ | ✅ | ✅ | ✅ | |
 
 !!! info
 
@@ -192,7 +194,7 @@ The following tables show which models support each feature:
     3. CPU Offloading (Layerwise) and CPU Offloading (Module-wise) are not compatible.
     4. CPU Offloading (Layerwise) supports single-card for now.
     5. Using FP8-Quant as an example of qunatization methods.
-    6. Step Execution is not compatible with cache backends (TeaCache, Cache-DiT) or LoRA.
+    6. Step Execution is not compatible with cache backends (TeaCache, Cache-DiT). LoRA is supported, but each scheduled batch must use a single adapter (requests with different `lora_request` or `lora_scale` are kept in separate batches).
 
 
 ## Multi-Thread Weight Loading
@@ -257,7 +259,7 @@ Measured on NVIDIA H800:
 
 **Parallelism Methods:**
 
-- **[Parallelism Overview](diffusion/parallelism/overview.md)** - Tensor Parallelism, Sequence Parallelism, CFG Parallelism, HSDP, and Expert Parallelism
+- **[Parallelism Overview](diffusion/parallelism/overview.md)** - Tensor Parallelism, Sequence Parallelism, CFG Parallelism, Pipeline Parallelism, HSDP, and Expert Parallelism
 
 **Memory Optimization:**
 

docs/user_guide/examples/offline_inference/text_to_speech.md

@@ -17,9 +17,9 @@ list of supported architectures across all modalities, see
 | Model | HuggingFace repo | Stages | Voice cloning | Streaming | Special modes | Sample rate |
 |---|---|---|---|---|---|---|
 | VoxCPM2 | `openbmb/VoxCPM2` | single (native AR) | ✓ | — | continuation (`--ref-audio` + `--ref-text`) | 48 kHz |
-| VoxCPM | local model dir | split | ✓ | ✓ (`voxcpm_async_chunk.yaml`) | — | 24 kHz |
 | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | 2 (talker + code2wav) | ✓ | ✓ (`async_chunk: true` default) | — | 22.05 kHz |
 | Fish Speech S2 Pro | `fishaudio/s2-pro` | dual-AR | ✓ | ✓ (`--streaming`) | — | 44.1 kHz |
+| GLM-TTS | `zai-org/GLM-TTS` | 2 (AR + DiT) | ✓ (required) | ✓ | — | 24 kHz |
 | OmniVoice | `k2-fsa/OmniVoice` | 2 (gen + dec) | ✓ | — | voice design (`--instruct`), language (`--lang`) | 24 kHz |
 | Qwen3-TTS | `Qwen/Qwen3-TTS-12Hz-1.7B-{CustomVoice,VoiceDesign,Base}` | 2 (talker + code2wav) | ✓ (Base) | ✓ | 3 task variants (`--query-type`) | 24 kHz |
 | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | varies | ✓ | ✓ | voice presets (`--voice`) | 24 kHz |
@@ -72,61 +72,6 @@ python examples/offline_inference/text_to_speech/voxcpm2/end2end.py \
 
 ---
 
-## VoxCPM
-
-Split-stage TTS. The hub example covers single TTS, single voice cloning, and streaming. Use `benchmarks/voxcpm/` for warmup, batch JSONL prompts, profiler injection, and offline TTFP / RTF measurement.
-
-### Prerequisites
-```bash
-pip install voxcpm soundfile
-# or use a local source tree:
-export VLLM_OMNI_VOXCPM_CODE_PATH=/path/to/VoxCPM/src
-```
-
-If the native VoxCPM `config.json` does not contain HF metadata such as `model_type`, prepare a persistent HF-compatible config directory and point the stage configs to it via `VLLM_OMNI_VOXCPM_HF_CONFIG_PATH`:
-
-```bash
-export VOXCPM_MODEL=/path/to/voxcpm-model
-export VLLM_OMNI_VOXCPM_HF_CONFIG_PATH=/tmp/voxcpm_hf_config
-mkdir -p "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"
-cp "$VOXCPM_MODEL/config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/config.json"
-cp "$VOXCPM_MODEL/generation_config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/generation_config.json" 2>/dev/null || true
-python3 -c 'import json, os; p=os.path.join(os.environ["VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"], "config.json"); cfg=json.load(open(p, "r", encoding="utf-8")); cfg["model_type"]="voxcpm"; cfg.setdefault("architectures", ["VoxCPMForConditionalGeneration"]); json.dump(cfg, open(p, "w", encoding="utf-8"), indent=2, ensure_ascii=False)'
-```
-
-### Quick start
-```bash
-python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This is a split-stage VoxCPM synthesis example running on vLLM Omni."
-```
-
-### Voice cloning
-```bash
-python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This sentence is synthesized with a cloned voice." \
-    --ref-audio /path/to/reference.wav \
-    --ref-text  "The exact transcript spoken in reference.wav."
-```
-
-### Streaming
-Pass the async-chunk stage config:
-```bash
-python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
-    --stage-configs-path vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml \
-    --text "This is a split-stage VoxCPM streaming example running on vLLM Omni."
-```
-
-### Notes
-- `voxcpm.yaml` is the default non-streaming stage config; `voxcpm_async_chunk.yaml` enables streaming.
-- Streaming is currently single-request oriented.
-- `--ref-text` must be the real transcript of `--ref-audio`; mismatched text degrades quality.
-- For online serving, see [`examples/online_serving/voxcpm`](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/voxcpm/README.md). For benchmark reporting, see [`benchmarks/voxcpm`](https://github.com/vllm-project/vllm-omni/tree/main/benchmarks/voxcpm/README.md).
-
----
-
 ## CosyVoice3
 
 2-stage TTS pipeline (`talker` + `code2wav`) at 22.05 kHz.
@@ -214,6 +159,35 @@ Streaming requires `async_chunk: true` in the stage config.
 
 ---
 
+## GLM-TTS
+
+2-stage TTS pipeline (AR + DiT flow-matching) at 24 kHz. Every request requires reference audio and its transcript for zero-shot voice cloning.
+
+### Quick start
+```bash
+python examples/offline_inference/text_to_speech/glm_tts/end2end.py \
+    --model zai-org/GLM-TTS \
+    --text "你好，这是语音合成测试。" \
+    --ref-audio /path/to/reference.wav \
+    --ref-text "这是参考音频的文本内容。" \
+    --output-dir ./output
+```
+
+### Architecture
+```
+Text → [Stage 0: AR] → Speech Tokens → [Stage 1: DiT + HiFT] → Audio (24 kHz)
+        (Llama-based)    (32k vocab)      (Flow Matching)
+```
+
+### Notes
+- `--ref-audio` and `--ref-text` are **required** together; GLM-TTS does not support text-only synthesis.
+- Reference audio should be 3-10 seconds.
+- First run may be slow due to lazy loading of WhisperVQ tokenizer and CampPlus ONNX speaker embedder.
+- Default sampling: temperature=1.0, top_k=25, top_p=0.8 (RAS method).
+- The `--model` path should point to the repository root (not `llm/` subdirectory).
+
+---
+
 ## OmniVoice
 
 Zero-shot multilingual TTS supporting 600+ languages, with three modes (auto / clone / design).
@@ -369,6 +343,10 @@ Available voice presets are listed on the HF model card (`mistralai/Voxtral-4B-T
     ``````py
     --8<-- "examples/offline_inference/text_to_speech/fish_speech/end2end.py"
     ``````
+??? abstract "glm_tts/end2end.py"
+    ``````py
+    --8<-- "examples/offline_inference/text_to_speech/glm_tts/end2end.py"
+    ``````
 ??? abstract "omnivoice/end2end.py"
     ``````py
     --8<-- "examples/offline_inference/text_to_speech/omnivoice/end2end.py"
@@ -381,10 +359,6 @@ Available voice presets are listed on the HF model card (`mistralai/Voxtral-4B-T
     ``````py
     --8<-- "examples/offline_inference/text_to_speech/qwen3_tts/end2end.py"
     ``````
-??? abstract "voxcpm/end2end.py"
-    ``````py
-    --8<-- "examples/offline_inference/text_to_speech/voxcpm/end2end.py"
-    ``````
 ??? abstract "voxcpm2/end2end.py"
     ``````py
     --8<-- "examples/offline_inference/text_to_speech/voxcpm2/end2end.py"

docs/user_guide/examples/online_serving/glm_image.md

@@ -101,7 +101,7 @@ guides.
 
 When using `/v1/chat/completions`, pass these inside `extra_body` in the curl
 JSON, or via the `extra_body` keyword argument in the OpenAI Python SDK (see the
-[Diffusion Chat API guide](../../../serving/diffusion_chat_api.md)).
+[Chat Completions API guide](../../../serving/chat_completions_api.md)).
 When using the dedicated [`/v1/images/generations`](../../../serving/image_generation_api.md)
 or [`/v1/images/edits`](../../../serving/image_edit_api.md) endpoints, pass
 the supported generation controls as top-level fields directly. For image

docs/user_guide/examples/online_serving/image_to_image.md

@@ -208,7 +208,7 @@ Wrap generation parameters inside `extra_body` in the request JSON:
     ```
 
     For details on how generation parameters are handled across different clients, see the
-    [Diffusion Chat API guide](../../../../serving/diffusion_chat_api.md).
+    [Chat Completions API guide](../../../../serving/chat_completions_api.md).
 
 ### Layered Image Generation (Qwen-Image-Layered)
 
@@ -361,7 +361,7 @@ Provide multiple images in `content` (order matters):
 
 When using `/v1/chat/completions`, pass these inside `extra_body` in the curl
 JSON, or via the `extra_body` keyword argument in the OpenAI Python SDK (see the
-[Diffusion Chat API guide](../../../../serving/diffusion_chat_api.md)).
+[Chat Completions API guide](../../../../serving/chat_completions_api.md)).
 When using the dedicated [`/v1/images/edits`](../../../../serving/image_edit_api.md)
 endpoint, pass the supported generation controls as top-level form fields
 directly. For image dimensions and count, use `size` and `n` rather than

docs/user_guide/examples/online_serving/text_to_image.md

@@ -216,7 +216,7 @@ Wrap generation parameters inside `extra_body` in the request JSON:
     ```
 
     For details on how generation parameters are handled across different clients, see the
-    [Diffusion Chat API guide](../../../../serving/diffusion_chat_api.md).
+    [Chat Completions API guide](../../../../serving/chat_completions_api.md).
 
 ### Multimodal Input (Text + Structured Content)
 
@@ -237,7 +237,7 @@ Wrap generation parameters inside `extra_body` in the request JSON:
 
 When using `/v1/chat/completions`, pass these inside `extra_body` in the curl
 JSON, or via the `extra_body` keyword argument in the OpenAI Python SDK (see the
-[Diffusion Chat API guide](../../../../serving/diffusion_chat_api.md)).
+[Chat Completions API guide](../../../../serving/chat_completions_api.md)).
 When using the dedicated [`/v1/images/generations`](../../../../serving/image_generation_api.md)
 endpoint, pass the supported generation controls as top-level JSON fields
 directly. For image dimensions and count, use `size` and `n` rather than

docs/user_guide/examples/online_serving/text_to_speech.md

@@ -17,14 +17,15 @@ For the full list of supported architectures across all modalities, see
 
 | Model | HuggingFace repo | Voice cloning | Streaming | Voice presets / upload | Gradio demo |
 |---|---|---|---|---|---|
+| CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✓ (`ref_audio`+`ref_text`) | ✓ (PCM stream) | — | — |
 | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✓ (`ref_audio`+`ref_text`) | ✓ (PCM stream) | — | ✓ |
+| GLM-TTS | `zai-org/GLM-TTS` | ✓ (`ref_audio`+`ref_text`, required) | ✓ (PCM stream) | — | ✓ |
 | OmniVoice | `k2-fsa/OmniVoice` | (offline only) | — | — | — |
 | Qwen3-TTS | `Qwen/Qwen3-TTS-12Hz-1.7B-{CustomVoice,VoiceDesign,Base}` | ✓ (Base) | ✓ (PCM + WebSocket) | ✓ (presets + `/v1/audio/voices` upload) | ✓ (standard + FastRTC) |
-| VoxCPM | local model dir | ✓ | ✓ (PCM stream) | — | — |
 | VoxCPM2 | `openbmb/VoxCPM2` | ✓ | ✓ (AudioWorklet via gradio) | — | ✓ |
 | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✓ (gated upstream) | ✓ | ✓ (presets) | ✓ |
 
-CosyVoice3 is intentionally absent: no online example exists for it yet. See its [offline section](https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/text_to_speech/README.md#cosyvoice3) instead.
+For offline inference of any of these models, see the [offline TTS hub](https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/text_to_speech/README.md).
 
 ## Common Quick Start
 
@@ -96,6 +97,54 @@ For full request-shape documentation (all parameters, response formats, error co
 
 ---
 
+## CosyVoice3
+
+2-stage TTS (`talker` + flow-matching `code2wav`) at 24 kHz. Voice cloning only — every request needs `ref_audio` + `ref_text`; there are no built-in voice presets.
+
+### Prerequisites
+```bash
+huggingface-cli download FunAudioLLM/Fun-CosyVoice3-0.5B-2512
+```
+
+If your downloaded checkpoint lacks `config.json`, add one with `{"model_type": "cosyvoice3", "architectures": ["CosyVoice3Model"]}` (the loader reads `model_type` to select the class).
+
+### Launch
+```bash
+vllm serve FunAudioLLM/Fun-CosyVoice3-0.5B-2512 --omni --port 8091 --trust-remote-code
+# or:
+./cosyvoice3/run_server.sh
+```
+
+Streaming is on by default via `async_chunk: true` in `vllm_omni/deploy/cosyvoice3.yaml`. Pass `--no-async-chunk` (or `NO_ASYNC_CHUNK=1 ./cosyvoice3/run_server.sh`) for the legacy synchronous path.
+
+### CLI client
+The client defaults to the official upstream zero-shot prompt, so it runs without extra flags:
+```bash
+cd examples/online_serving/text_to_speech/cosyvoice3
+python speech_client.py --text "收到好友从远方寄来的生日礼物。"
+```
+
+Pass your own reference clip and transcript for a different voice:
+```bash
+python speech_client.py --text "Hello, this is a cloned voice." \
+    --ref-audio /path/to/reference.wav \
+    --ref-text "Transcript of the reference audio."
+```
+
+Stream PCM instead of WAV:
+```bash
+python speech_client.py --text "Hello world" --stream --output output.pcm
+```
+
+The client supports `--api-base`, `--model`, `--text`, `--ref-audio`, `--ref-text`, `--response-format`, `--stream`, `--output`.
+
+### Notes
+- Stage 0 (`talker`) emits speech tokens; stage 1 (`code2wav`) runs flow matching + HiFiGAN to synthesize waveform.
+- Deploy config auto-loads from `vllm_omni/deploy/cosyvoice3.yaml` based on HF `model_type`. Pass `--deploy-config <path>` to override.
+- For offline inference and the end-to-end script, see the [offline CosyVoice3 section](https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/text_to_speech/README.md#cosyvoice3).
+
+---
+
 ## Fish Speech S2 Pro
 
 4B dual-AR TTS at 44.1 kHz. Server uses the DAC codec.
@@ -144,6 +193,45 @@ python fish_speech/gradio_demo.py --api-base http://localhost:8091  # if server
 
 ---
 
+## GLM-TTS
+
+2-stage TTS (AR + DiT flow-matching) at 24 kHz. Every request requires `ref_audio` + `ref_text`.
+
+### Launch
+```bash
+vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091
+# or:
+bash examples/online_serving/text_to_speech/glm_tts/run_server.sh /path/to/GLM-TTS
+```
+
+### Sending requests
+```bash
+# Voice cloning (required)
+python examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py \
+    --text "你好，这是语音克隆测试。" \
+    --ref-audio file:///path/to/ref.wav \
+    --ref-text "这是参考音频的文本内容。"
+
+# Custom format
+python examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py \
+    --text "Hello, this is a voice cloning test." \
+    --ref-audio file:///path/to/ref.wav \
+    --ref-text "Transcript of the reference audio." \
+    --response-format mp3 -o output.mp3
+```
+
+### Gradio demo
+```bash
+bash examples/online_serving/text_to_speech/glm_tts/run_gradio_demo.sh
+```
+
+### Notes
+- Output: 24 kHz mono WAV via HiFT vocoder.
+- `ref_audio` + `ref_text` are **required** together on every request. Reference audio should be 3-10 seconds.
+- Voice cloning feature extraction (WhisperVQ, CampPlus, mel) runs on the model side — no external dependency on the serving layer.
+
+---
+
 ## OmniVoice
 
 Zero-shot multilingual TTS (600+ languages). Online serving currently exposes **auto voice** only; voice cloning and voice design are available offline.
@@ -196,10 +284,10 @@ vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --omni --port 8091
 ./qwen3_tts/run_server.sh Base
 ```
 
-### Choosing an executor backend (uniproc vs mp)
-Stage configs ship with the chunked-streaming default. To use the uniproc executor (lower IPC overhead for the Base cloning task), pass `--stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts_uniproc.yaml`. See [#2603](https://github.com/vllm-project/vllm-omni/issues/2603) and [#2604](https://github.com/vllm-project/vllm-omni/pull/2604) for the full investigation.
+### Executor backend
+Single-GPU serves now default to the uniproc executor (lower IPC overhead, the Base cloning use case from [#2603](https://github.com/vllm-project/vllm-omni/issues/2603) / [#2604](https://github.com/vllm-project/vllm-omni/pull/2604)). `vllm_omni/deploy/qwen3_tts.yaml` is the only Qwen3-TTS deploy config; pass `--deploy-config <path>` to override.
 
-To opt out of chunked streaming, pass `--no-async-chunk` instead — the pipeline auto-dispatches to the end-to-end codec processor.
+To opt out of chunked streaming, pass `--no-async-chunk` — the pipeline auto-dispatches to the end-to-end codec processor.
 
 ### Sending requests
 ```bash
@@ -281,71 +369,11 @@ python qwen3_tts/gradio_fastrtc_demo.py --api-base http://localhost:8000
 `qwen3_tts/batch_speech_client.py` issues many concurrent requests for throughput measurement.
 
 ### Notes
-- Base voice cloning has uniproc-vs-mp tradeoffs depending on per-request reference audio cost; see the executor-backend section above.
+- Base voice cloning has per-request reference-audio cost; the uniproc default keeps IPC overhead off the critical path. See the executor-backend section above for background.
 - `vllm_omni/deploy/qwen3_tts.yaml` is the default deploy config (loaded by HF `model_type`); per-stage runtime overrides are available via `--stage-N-<field> <value>`.
 
 ---
 
-## VoxCPM
-
-Split-stage TTS at 24 kHz.
-
-### Prerequisites
-```bash
-pip install voxcpm
-# or use a local source tree:
-export VLLM_OMNI_VOXCPM_CODE_PATH=/path/to/VoxCPM/src
-```
-
-If the native VoxCPM `config.json` lacks HF `model_type`, set up an HF-compatible config dir:
-```bash
-export VOXCPM_MODEL=/path/to/voxcpm-model
-export VLLM_OMNI_VOXCPM_HF_CONFIG_PATH=/tmp/voxcpm_hf_config
-mkdir -p "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"
-cp "$VOXCPM_MODEL/config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/config.json"
-cp "$VOXCPM_MODEL/generation_config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/generation_config.json" 2>/dev/null || true
-python3 -c 'import json, os; p=os.path.join(os.environ["VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"], "config.json"); cfg=json.load(open(p, "r", encoding="utf-8")); cfg["model_type"]="voxcpm"; cfg.setdefault("architectures", ["VoxCPMForConditionalGeneration"]); json.dump(cfg, open(p, "w", encoding="utf-8"), indent=2, ensure_ascii=False)'
-```
-
-### Launch
-```bash
-export VOXCPM_MODEL=/path/to/voxcpm-model
-./voxcpm/run_server.sh                # async-chunk streaming (default)
-./voxcpm/run_server.sh sync           # non-streaming
-```
-Or directly:
-```bash
-vllm serve "$VOXCPM_MODEL" \
-    --stage-configs-path vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml \
-    --trust-remote-code --enforce-eager --omni --port 8091
-```
-
-### Sending requests
-```bash
-# Basic TTS
-python voxcpm/openai_speech_client.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This is a VoxCPM online text-to-speech example."
-
-# Voice cloning
-python voxcpm/openai_speech_client.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This sentence is synthesized with a cloned voice." \
-    --ref-audio /path/to/reference.wav \
-    --ref-text "The exact transcript spoken in reference.wav."
-
-# Streaming PCM
-python voxcpm/openai_speech_client.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This is a streaming VoxCPM request." \
-    --stream --output voxcpm_stream.pcm
-```
-
-### Notes
-- `voxcpm.yaml` for one-shot decode; `voxcpm_async_chunk.yaml` for single-request streaming. Do not use the async-chunk config for concurrent requests or `/v1/audio/speech/batch`.
-- Generic TTS fields not supported by VoxCPM: `voice`, `instructions`, `language`, `speaker_embedding`, `x_vector_only_mode`.
-- For benchmark measurement, see [`benchmarks/voxcpm`](https://github.com/vllm-project/vllm-omni/tree/main/benchmarks/voxcpm/README.md).
-
 ---
 
 ## VoxCPM2
@@ -407,6 +435,14 @@ The demo handles voice-preset selection and reference-audio upload. `voxtral_tts
 
 ## Example materials
 
+??? abstract "cosyvoice3/run_server.sh"
+
+    --8<-- "examples/online_serving/text_to_speech/cosyvoice3/run_server.sh"
+
+??? abstract "cosyvoice3/speech_client.py"
+
+    --8<-- "examples/online_serving/text_to_speech/cosyvoice3/speech_client.py"
+
 ??? abstract "fish_speech/gradio_demo.py"
     ``````py
     --8<-- "examples/online_serving/text_to_speech/fish_speech/gradio_demo.py"
@@ -423,6 +459,22 @@ The demo handles voice-preset selection and reference-audio upload. `voxtral_tts
     ``````py
     --8<-- "examples/online_serving/text_to_speech/fish_speech/speech_client.py"
     ``````
+??? abstract "glm_tts/gradio_demo.py"
+    ``````py
+    --8<-- "examples/online_serving/text_to_speech/glm_tts/gradio_demo.py"
+    ``````
+??? abstract "glm_tts/openai_speech_client.py"
+    ``````py
+    --8<-- "examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py"
+    ``````
+??? abstract "glm_tts/run_gradio_demo.sh"
+    ``````sh
+    --8<-- "examples/online_serving/text_to_speech/glm_tts/run_gradio_demo.sh"
+    ``````
+??? abstract "glm_tts/run_server.sh"
+    ``````sh
+    --8<-- "examples/online_serving/text_to_speech/glm_tts/run_server.sh"
+    ``````
 ??? abstract "omnivoice/run_server.sh"
     ``````sh
     --8<-- "examples/online_serving/text_to_speech/omnivoice/run_server.sh"
@@ -463,14 +515,6 @@ The demo handles voice-preset selection and reference-audio upload. `voxtral_tts
     ``````py
     --8<-- "examples/online_serving/text_to_speech/qwen3_tts/tts_common.py"
     ``````
-??? abstract "voxcpm/openai_speech_client.py"
-    ``````py
-    --8<-- "examples/online_serving/text_to_speech/voxcpm/openai_speech_client.py"
-    ``````
-??? abstract "voxcpm/run_server.sh"
-    ``````sh
-    --8<-- "examples/online_serving/text_to_speech/voxcpm/run_server.sh"
-    ``````
 ??? abstract "voxcpm2/gradio_demo.py"
     ``````py
     --8<-- "examples/online_serving/text_to_speech/voxcpm2/gradio_demo.py"

docs/user_guide/quantization/autoround.md

@@ -32,7 +32,9 @@ guide. AutoRound is Intel-supported.
 |-------|------------|-------|--------|---------|
 | FLUX.1-dev | `vllm-project-org/FLUX.1-dev-AutoRound-w4a16` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
 | Qwen-Image | Not listed | Diffusion transformer | W4A16 | Not validated |
-| Wan2.2 | Not listed | Diffusion transformer | W4A16 | Not validated |
+| Wan2.2-I2V | `Intel/Wan2.2-I2V-A14B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
+| Wan2.2-T2V | `Intel/Wan2.2-T2V-A14B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
+| Wan2.2-TI2V | `Intel/Wan2.2-TI2V-5B-Diffusers-int4-AutoRound` | Diffusion transformer | W4A16 | GPTQ-Marlin or Intel-supported AutoRound backend |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 

docs/user_guide/quantization/modelopt.md

@@ -6,34 +6,46 @@ ModelOpt quantization loads checkpoints produced by NVIDIA ModelOpt. The
 quantized weights and scale tensors are generated before serving, so inference
 does not run online calibration or convert a BF16 checkpoint at startup.
 
-vLLM-Omni currently validates the ModelOpt FP8 checkpoint path for diffusion
-transformers. The loader auto-detects supported ModelOpt FP8 checkpoint configs
-and keeps non-transformer components, such as the tokenizer, scheduler, text
+vLLM-Omni validates ModelOpt FP8, ModelOpt NVFP4, and ModelOpt mixed
+FP8/NVFP4 checkpoint loading for diffusion transformer stages. The loader
+auto-detects supported ModelOpt checkpoint configs and keeps non-transformer
+components, such as the tokenizer, scheduler, text encoder, vision/audio
 encoder, and VAE, on the base checkpoint unless a model-specific guide says
 otherwise.
 
 !!! note
-    `--force-cutlass-fp8` is an explicit runtime override for diffusion
-    checkpoints that already carry a supported ModelOpt FP8 config. It does not
-    quantize BF16 checkpoints and it does not apply to online `--quantization
-    fp8`. The flag only takes effect for ModelOpt FP8 diffusion stages on CUDA
-    SM89+ devices; other platforms and non-ModelOpt FP8 paths fall back to the
-    normal vLLM kernel selection.
+    ModelOpt checkpoints are pre-quantized checkpoints. Do not pass
+    `--quantization fp8` for these checkpoints. The checkpoint
+    `quantization_config` selects the ModelOpt path.
+
+!!! note
+    `--force-cutlass-fp8`, `--linear-backend cutlass`, and
+    `--moe-backend cutlass` are runtime backend selections for checkpoints that
+    already carry supported ModelOpt quantized weights and scales. They do not
+    quantize BF16 checkpoints at startup.
 
 ## Supported ModelOpt Checkpoint Formats
 
 vLLM-Omni treats ModelOpt checkpoints as pre-quantized checkpoints. The
 checkpoint config must identify ModelOpt as the quantization method or producer,
-and the quantization algorithm must be one of the validated FP8 algorithms.
+and the quantization algorithm must be one of the validated algorithms below.
 
 | Checkpoint field | Supported value |
 |------------------|-----------------|
-| `method` / `quant_method` | `modelopt` |
+| `method` / `quant_method` | `modelopt`, `modelopt_fp4`, `modelopt_mixed` |
 | `producer.name` | `modelopt` |
-| `quant_algo` | `FP8`, `FP8_PER_CHANNEL_PER_TOKEN` |
+| `quant_algo` | `FP8`, `FP8_PER_CHANNEL_PER_TOKEN`, `NVFP4`, `MIXED_PRECISION` |
 
-Other ModelOpt algorithms, such as NVFP4, are not enabled by this diffusion
-FP8 path until they have separate model and quality validation.
+| `quant_algo` | Runtime method | Typical use |
+|--------------|----------------|-------------|
+| `FP8`, `FP8_PER_CHANNEL_PER_TOKEN` | `modelopt` | FP8 diffusion transformer checkpoints |
+| `NVFP4` | `modelopt_fp4` | NVFP4 diffusion transformer checkpoints |
+| `MIXED_PRECISION` | `modelopt_mixed` | Mixed FP8/NVFP4 checkpoints with a ModelOpt per-layer policy |
+
+For multi-component diffusion or omni models, only the checkpoint component
+that contains ModelOpt quantized weights should use the ModelOpt quantization
+method. Encoders, decoders, tokenizers, schedulers, and other BF16 components
+stay unquantized unless the model-specific recipe validates otherwise.
 
 ## Hardware Support
 
@@ -47,7 +59,10 @@ FP8 path until they have separate model and quality validation.
 | Ascend NPU | ❌ |
 
 Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this guide.
-The optional CUTLASS FP8 runtime override requires CUDA SM89+.
+The optional CUTLASS FP8 runtime override requires CUDA SM89+. ModelOpt NVFP4
+and mixed FP8/NVFP4 diffusion checkpoints are currently validated on Blackwell
+CUDA systems in the recipes below; other CUDA generations require separate
+backend and quality validation.
 
 ## Model Type Support
 
@@ -56,12 +71,19 @@ The optional CUTLASS FP8 runtime override requires CUDA SM89+.
 | Model | HF checkpoint | Scope | Status |
 |-------|---------------|-------|--------|
 | Qwen-Image 2512 | `feizhai123/qwen-image-2512-modelopt-fp8-dynamic-all` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
+| Qwen-Image 2512 | `feizhai123/qwen-image-2512-modelopt-mixed-fp8-sensitive-nvfp4-heavy` | Diffusion transformer | Validated for ModelOpt mixed FP8/NVFP4 checkpoints |
 | Z-Image | `feizhai123/z-image-modelopt-fp8-conservative` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
 | FLUX.2-dev | `feizhai123/flux2-dev-modelopt-fp8` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
 | FLUX.2-klein 4B | `feizhai123/flux2-klein-4b-modelopt-fp8` | Diffusion transformer | Validated for ModelOpt FP8 checkpoints |
 | HunyuanImage-3.0 | `feizhai123/hunyuan-image3-modelopt-fp8` | MoE diffusion transformer | Validated for ModelOpt FP8 checkpoints |
+| HunyuanImage-3.0 | `feizhai123/hunyuan-image3-modelopt-mixed-experts-nvfp4-dense-fp8` | MoE diffusion transformer | Validated for ModelOpt mixed FP8/NVFP4 checkpoints |
 | Wan2.2 | Not available | Diffusion transformer | Not validated |
 
+For full serving commands and benchmark context, see
+[`recipes/Qwen/Qwen-Image.md`](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen-Image.md)
+and
+[`recipes/Tencent/HunyuanImage-3.0-Instruct.md`](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Tencent/HunyuanImage-3.0-Instruct.md).
+
 ### Multi-Stage Omni/TTS Model
 
 | Model | Scope | Status |
@@ -80,15 +102,28 @@ ModelOpt targets yet.
 
 ## Configuration
 
-For pre-quantized ModelOpt FP8 checkpoints, no `--quantization fp8` flag is
-needed. The checkpoint config selects the ModelOpt path.
+For pre-quantized ModelOpt checkpoints, no `--quantization fp8` flag is needed.
+The checkpoint config selects the ModelOpt path.
 
 Online serving:
 
 ```bash
-vllm serve <modelopt-fp8-checkpoint> \
+vllm serve <modelopt-checkpoint> \
   --omni \
   --tensor-parallel-size <N> \
+  --linear-backend cutlass \
+  --force-cutlass-fp8
+```
+
+For mixed FP8/NVFP4 MoE checkpoints, also select the validated MoE backend:
+
+```bash
+vllm serve <modelopt-mixed-moe-checkpoint> \
+  --omni \
+  --tensor-parallel-size <N> \
+  --enable-expert-parallel \
+  --linear-backend cutlass \
+  --moe-backend cutlass \
   --force-cutlass-fp8
 ```
 
@@ -96,14 +131,14 @@ Offline inference:
 
 ```bash
 python examples/offline_inference/text_to_image/text_to_image.py \
-  --model <modelopt-fp8-checkpoint> \
+  --model <modelopt-checkpoint> \
   --tensor-parallel-size <N> \
   --prompt "a red ceramic teapot on a wooden table" \
   --height 1024 \
   --width 1024 \
   --num-inference-steps 20 \
   --seed 42 \
-  --output outputs/modelopt_fp8.png
+  --output outputs/modelopt.png
 ```
 
 Python API:
@@ -112,7 +147,7 @@ Python API:
 from vllm_omni import Omni
 
 omni = Omni(
-    model="<modelopt-fp8-checkpoint>",
+    model="<modelopt-checkpoint>",
     tensor_parallel_size=2,
     force_cutlass_fp8=True,
 )
@@ -123,16 +158,34 @@ omni = Omni(
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
 | `force_cutlass_fp8` / `--force-cutlass-fp8` | bool | `False` | Force CUTLASS FP8 linear kernels for supported ModelOpt FP8 diffusion stages on CUDA SM89+ |
+| `--linear-backend cutlass` | str | auto | Select the validated CUTLASS linear backend for supported ModelOpt NVFP4 or mixed FP8/NVFP4 diffusion stages |
+| `--moe-backend cutlass` | str | auto | Select the validated CUTLASS MoE backend for supported ModelOpt mixed MoE checkpoints |
 
 ## Validation and Notes
 
-1. Compare the ModelOpt FP8 checkpoint against the BF16 baseline with the same
+1. Compare the ModelOpt checkpoint against the BF16 baseline with the same
    prompt, resolution, seed, and inference steps.
 2. Use `tests/diffusion/quantization/test_quantization_quality.py` with
    `VLLM_OMNI_QUALITY_CONFIGS` to validate local baseline and quantized model
    paths.
-3. Report LPIPS, PSNR, MAE, throughput, latency, and peak memory when adding a
-   new validated ModelOpt diffusion checkpoint.
-4. Keep `--quantization fp8` for online FP8 from BF16 checkpoints; use this
-   ModelOpt path only when the checkpoint already contains ModelOpt FP8 weights
-   and scales.
+3. For HunyuanImage-3.0 quantized DiT checkpoints, the opt-in accuracy check is:
+
+   ```bash
+   CUDA_VISIBLE_DEVICES=2,3 \
+   HUNYUAN_IMAGE3_RUN_QUANT_ACCURACY=1 \
+   HUNYUAN_IMAGE3_QUANT_DEVICES=0,1 \
+   HUNYUAN_IMAGE3_QUANT_TP=2 \
+   HUNYUAN_IMAGE3_BF16_MODEL=/path/to/hunyuan-image3-bf16 \
+   HUNYUAN_IMAGE3_FP8_MODEL=/path/to/hunyuan-image3-modelopt-fp8 \
+   HUNYUAN_IMAGE3_NVFP4_MODEL=/path/to/hunyuan-image3-modelopt-mixed-experts-nvfp4-dense-fp8 \
+   PYTHONPATH=/path/to/vllm-omni:${PYTHONPATH:-} \
+   python -m pytest -s -v \
+     tests/e2e/accuracy/test_hunyuan_image3.py \
+     -k quantized_dit_matches_bf16_accuracy
+   ```
+
+4. Report CLIP score deltas, SSIM, PSNR, throughput, latency, and peak memory
+   when adding a new validated ModelOpt diffusion checkpoint.
+5. Keep `--quantization fp8` for online FP8 from BF16 checkpoints; use this
+   ModelOpt path only when the checkpoint already contains ModelOpt quantized
+   weights and scales.

docs/user_guide/quantization/mxfp4.md

@@ -0,0 +1,479 @@
+# W4A4 MXFP4 Quantization
+
+## Overview
+
+W4A4 MXFP4 (Microscaling FP4) quantizes both weights and activations to FP4
+(`float4_e2m1fn_x2`, packed 2 values per byte) using the OCP MX format: groups
+of 32 K-dimension elements share a single `float8_e8m0fnu` exponent scale.
+
+vLLM-Omni provides two quantization methods with different scale structures:
+
+| Method | Scale structure | Mode | Use case |
+|--------|----------------|------|----------|
+| `mxfp4` | Single-scale (per-32 fine only) | Online only | Quick accuracy baseline; no checkpoint prep needed |
+| `mxfp4_dualscale` | Dual-scale (fine per-32 + coarse per-512 + per-channel `mul_scale`) | Online + Offline | Production; better accuracy; offline recommended |
+
+!!! tip "Recommended: `mxfp4_dualscale` offline"
+    For production deployments, use the `mxfp4_dualscale` offline mode with a
+    pre-quantized checkpoint produced by msModelSlim. Offline checkpoints load
+    calibrated `mul_scale` tensors from disk, providing measurably better accuracy
+    than any online method. The one-time preprocessing cost amortises across all
+    subsequent inference runs.
+
+    Use `mxfp4` online only for quick experimentation where preprocessing time
+    is not acceptable and accuracy loss is tolerable.
+
+!!! warning "Online single-scale ≠ Offline dual-scale"
+    `mxfp4_dualscale` offline mode uses `NPUMxfp4DualScaleLinearMethod`:
+    fine scale (per-32 K), coarse scale (per-512 K), and per-input-channel
+    `mul_scale` from calibration — all loaded from the checkpoint.
+    `mxfp4_dualscale` online mode uses `NPUMxfp4DualScaleOnlineLinearMethod`:
+    dual-level scales computed on the fly from BF16 weights; no calibration
+    `mul_scale` is available. Loading an offline checkpoint with the online
+    method (or vice versa) will produce incorrect results or shape errors.
+
+## Hardware Support
+
+| Device | Support |
+|--------|---------|
+| NVIDIA Blackwell GPU (SM 100+) | ⭕ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ⭕ |
+| NVIDIA Ampere GPU (SM 80+) | ⭕ |
+| AMD ROCm | ⭕ |
+| Intel XPU | ⭕ |
+| Ascend NPU (Atlas 950 A5) | ✅ |
+
+Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this guide.
+
+## Model Type Support
+
+### Diffusion Model (Wan2.2)
+
+| Model | Online | Offline | Notes |
+|-------|--------|---------|-------|
+| Wan2.2-T2V-A14B | `mxfp4` / `mxfp4_dualscale` | `mxfp4_dualscale` | MoE cascade (`transformer` + `transformer_2`); both transformers quantized with the same config |
+| Wan2.2-I2V-A14B | `mxfp4` / `mxfp4_dualscale` | `mxfp4_dualscale` | MoE cascade; same scheme as T2V-A14B |
+| Wan2.2-TI2V-5B | ❌ | ❌ | Parameter count too small; W4A4 causes unacceptable accuracy loss |
+
+The choice between `mxfp4` and `mxfp4_dualscale` in **online mode** is about
+quantization quality, not model compatibility — both work on cascade (A14B) and
+single-transformer models alike, the same as `mxfp8` online:
+
+- `mxfp4`: single-scale, lower overhead, simpler compute, online only
+- `mxfp4_dualscale`: dual-scale + optional BF16 fallback, better accuracy, online **and** offline
+
+**Offline** checkpoints for A14B are always in `mxfp4_dualscale` format (produced
+by the merge script); there is no offline `mxfp4` single-scale format.
+
+!!! note "Per-layer BF16 fallback in offline cascade models"
+    The A14B offline checkpoint uses `quant_method: mxfp4_dualscale`. Most
+    linear layers are stored as W4A4 MXFP4 DualScale; precision-sensitive layers
+    retain their original BF16 weights and are listed in `ignored_layers` inside
+    each transformer's `config.json`. The two transformers may have different
+    `ignored_layers` sets — the pipeline reads each transformer's own `config.json`
+    and rebuilds the config locally when they differ, so routing is always
+    per-transformer-accurate.
+
+!!! warning "TI2V-5B not supported"
+    Wan2.2-TI2V-5B is excluded from W4A4 quantization. Its smaller parameter
+    count makes it significantly more sensitive to 4-bit quantization noise,
+    resulting in unacceptable accuracy loss. Use [MXFP8](mxfp8.md) for TI2V-5B.
+
+## Configuration
+
+### `mxfp4` — Single-Scale Online Mode
+
+Online mode requires no pre-processing. vLLM-Omni quantizes BF16 weights to
+MXFP4 at load time using `npu_dynamic_mx_quant`. A single block scale
+(`float8_e8m0fnu`, one per 32 K elements) is computed on the fly; no
+calibration `mul_scale` is available. Applies equally to single-transformer
+and cascade (A14B) models — both transformers in a cascade receive the same
+quantization config automatically.
+
+```python
+from vllm_omni import Omni
+from vllm_omni.inputs.data import OmniDiffusionSamplingParams
+
+omni = Omni(model="<your-model>", quantization="mxfp4")
+outputs = omni.generate(
+    "A cat sitting on a windowsill",
+    OmniDiffusionSamplingParams(num_inference_steps=50),
+)
+```
+
+```bash
+# Single-transformer or cascade model — same command
+python text_to_video.py --model <your-model> --quantization mxfp4
+
+# Online serving
+vllm serve <your-model> --omni --quantization mxfp4
+```
+
+### `mxfp4_dualscale` — DualScale Online Mode
+
+Online DualScale mode computes both fine and coarse scales on the fly from BF16
+weights using `npu_dynamic_dual_level_mx_quant`. Applies equally to
+single-transformer and cascade (A14B) models. Compared to `mxfp4` online,
+DualScale provides better quantization accuracy at higher compute cost.
+
+The default configuration keeps the leading 5 transformer blocks in BF16
+(`num_bf16_fallback_layers=5`). Accuracy evaluation on Wan2.2-A14B shows this
+is sufficient to meet quality requirements and is the recommended setting.
+
+```python
+omni = Omni(model="<your-model>", quantization="mxfp4_dualscale")
+```
+
+```bash
+python text_to_video.py --model <your-model> --quantization mxfp4_dualscale
+```
+
+If accuracy debugging identifies additional precision-sensitive layers, they
+can be pinned to BF16 via the Python API:
+
+```python
+omni = Omni(
+    model="<your-model>",
+    quantization_config={
+        "method": "mxfp4_dualscale",
+        "ignored_layers": ["blocks.10.attn1.to_q"],   # explicit per-layer override
+    },
+)
+```
+
+BF16 fallback routing in online mode applies two rules in priority order:
+
+1. **`ignored_layers`** (explicit per-layer override): any layer whose prefix
+   matches is kept in BF16 regardless of block index.
+2. **`num_bf16_fallback_layers`** (coarse leading-block rule): the first N
+   transformer blocks (`blocks.0` … `blocks.N-1`) fall back to BF16. Defaults
+   to `5` (recommended). Layers outside `blocks.N.*`
+   (e.g. `condition_embedder`) are always quantized.
+
+### `mxfp4_dualscale` — DualScale Offline Mode (Recommended)
+
+Offline mode loads a pre-quantized DualScale checkpoint from msModelSlim. A
+preprocessing step converts the raw quantized output to the diffusers format
+expected by vLLM-Omni and injects the quantization config into each
+`transformer/config.json` so that vLLM-Omni auto-detects the offline path
+without a `--quantization` flag.
+
+BF16 fallback layers may be interleaved anywhere in the transformer — they are
+not restricted to leading blocks. The merge script detects them from
+`quant_model_description.json` and writes their prefixes into `ignored_layers`
+inside `config.json`. At runtime, each layer's prefix is matched against
+`ignored_layers` to decide BF16 vs. MXFP4 DualScale.
+
+#### Checkpoint tensor layout
+
+Each quantized linear layer stores four tensors:
+
+| Tensor | Shape | dtype | Description |
+|--------|-------|-------|-------------|
+| `weight` | `(N, K)` | float8_e4m3fn | FP4 packed (2 values per byte) |
+| `weight_scale` | `(N, K//32)` | uint8 | Fine block scale (`float8_e8m0fnu` bit pattern) |
+| `weight_dual_scale` | `(N, K//512, 1)` | float32 | Coarse block scale |
+| `mul_scale` | `(K,)` | float32 | Per-input-channel smooth pre-scale (from calibration) |
+
+BF16 fallback layers have no quantization tensors; only the original `weight`
+(and optional `bias`) are present, loaded directly from the base checkpoint.
+
+#### Step 1 — Quantize with msModelSlim
+
+```bash
+msmodelslim quant \
+  --model_path /path/to/Wan2.2-T2V-A14B-Diffusers \
+  --save_path  /path/to/wan2_2_t2v_quantized_raw \
+  --device npu \
+  --model_type Wan2_2 \
+  --config_path /path/to/wan2_2_w4a4_mxfp4_dualscale.yaml \
+  --trust_remote_code True
+```
+
+After this step, `--save_path` contains raw quantized safetensors files,
+scale files, and a metadata JSON (`quant_model_description*.json`).
+
+For cascade MoE models (T2V-A14B, I2V-A14B), msModelSlim outputs two
+subdirectories: `high_noise_model/` (transformer) and `low_noise_model/`
+(transformer_2).
+
+#### Step 2 — Preprocess with merge_mxfp4_dualscale_checkpoint.py
+
+The script (`vllm_omni/quantization/tools/merge_mxfp4_dualscale_checkpoint.py`):
+
+1. Copies the original diffusers model to `--output-path` (VAE, text encoder,
+   scheduler, etc. are preserved).
+2. Remaps tensor names from msModelSlim convention to diffusers convention and
+   strips `.linear.` / `.div.` wrappers added by the quantization tool.
+3. Overlays MXFP4 tensors (weight, fine/coarse scales, `mul_scale`) onto the
+   BF16 base checkpoint. Non-quantized layers keep their original BF16 weights.
+4. Detects all linear layers that remain in BF16 and writes their prefixes into
+   `ignored_layers` in `config.json`.
+5. Injects `quantization_config` so vLLM-Omni auto-detects offline MXFP4
+   DualScale.
+
+For cascade MoE models, steps 2–5 run separately for each transformer.
+
+```bash
+python vllm_omni/quantization/tools/merge_mxfp4_dualscale_checkpoint.py \
+  --model-type     Wan2.2-T2V-A14B \
+  --original-model /path/to/Wan2.2-T2V-A14B-Diffusers \
+  --quant-path     /path/to/wan2_2_t2v_quantized_raw \
+  --output-path    /path/to/Wan2.2-T2V-A14B-MXFP4-DualScale
+```
+
+| Argument | Description |
+|----------|-------------|
+| `--model-type` | Model variant: `Wan2.2-T2V-A14B` or `Wan2.2-I2V-A14B` |
+| `--original-model` | Root directory of the original BF16 diffusers model |
+| `--quant-path` | Root directory of the msModelSlim quantized output |
+| `--output-path` | Output directory for the merged model (created by the script) |
+
+The script outputs a complete diffusers model directory at `--output-path`,
+with each transformer subfolder containing:
+
+- `diffusion_pytorch_model.safetensors` — MXFP4 weights + scale tensors, with BF16 fallback layers from the base checkpoint
+- `config.json` — original transformer config with `quantization_config` injected
+- `quant_model_description.json` — quantization metadata (reference only)
+
+The `quantization_config` injected into `config.json` for each transformer:
+
+```json
+{
+  "quant_method": "mxfp4_dualscale",
+  "is_checkpoint_serialized": true,
+  "ignored_layers": [
+    "blocks.0.attn1.to_qkv",
+    "blocks.0.attn1.to_out",
+    "proj_out"
+  ]
+}
+```
+
+`ignored_layers` lists every linear layer that retains its original BF16 weight,
+using vllm-omni model parameter names (QKV-fused, FFN underscored, `to_out`
+unindexed). The exact entries are determined by the quantization tool (msModelSlim)
+and may differ between `transformer` and `transformer_2` in a cascade model.
+
+#### Step 3 — Serve
+
+```bash
+python text_to_video.py --model /path/to/Wan2.2-T2V-A14B-MXFP4-DualScale
+
+# Online serving
+vllm serve /path/to/Wan2.2-T2V-A14B-MXFP4-DualScale --omni
+```
+
+```python
+omni = Omni(model="/path/to/Wan2.2-T2V-A14B-MXFP4-DualScale")
+```
+
+!!! note
+    No `--quantization` flag is needed for offline mode. The preprocessing
+    script injects `quantization_config` into each `transformer/config.json`,
+    which vLLM-Omni reads automatically to activate the correct offline path.
+
+## Parameters
+
+### `mxfp4` (single-scale, online only)
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `method` | str | — | `"mxfp4"` |
+| `ignored_layers` | list[str] | `[]` | Layer prefixes to keep in BF16 |
+
+### `mxfp4_dualscale` (dual-scale, online + offline)
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `method` | str | — | `"mxfp4_dualscale"` |
+| `is_checkpoint_serialized` | bool | `False` | `True` for offline DualScale checkpoints; auto-set from `config.json` when using the preprocessing script |
+| `ignored_layers` | list[str] | `[]` | Layer prefixes to keep in BF16. **Works in both modes**: offline — populated by the merge script for interleaved sensitive layers; online — user-supplied for explicit per-layer precision override |
+| `num_bf16_fallback_layers` | int | `5` | **Online mode only**: leading N transformer blocks (`blocks.0` … `blocks.N-1`) kept in BF16. Applied after `ignored_layers`; ignored in offline mode. Default of `5` is the evaluated recommended value for Wan2.2-A14B |
+
+#### BF16 fallback priority (online mode)
+
+```
+for each linear layer:
+    if prefix in ignored_layers               → BF16  (explicit override, highest priority)
+    elif block_idx < num_bf16_fallback_layers → BF16  (coarse leading-block rule)
+    else                                      → MXFP4 DualScale online
+```
+
+Layers outside `blocks.N.*` (e.g. `condition_embedder.*`) are always quantized
+unless they appear in `ignored_layers`.
+
+## Validation and Notes
+
+1. **Online single-scale (`mxfp4`)** quantizes BF16 weights at load time using
+   `npu_dynamic_mx_quant` (single-scale). No calibration `mul_scale` is
+   available — all output partitions receive an identity pre-scale. No offline
+   checkpoint format exists for this method.
+
+2. **Online dual-scale (`mxfp4_dualscale`, `is_checkpoint_serialized=False`)**
+   quantizes BF16 weights using `npu_dynamic_dual_level_mx_quant` (fine + coarse
+   scales computed on the fly). No calibration `mul_scale`; leading blocks or
+   explicit `ignored_layers` stay in BF16 for accuracy.
+
+3. **Offline dual-scale (`mxfp4_dualscale`, `is_checkpoint_serialized=True`)** —
+   **recommended for production** — loads four tensors per quantized layer: FP4
+   weight, fine scale (`uint8` reinterpreted as `float8_e8m0fnu`), coarse scale
+   (`float32`), and per-input-channel `mul_scale` (`float32`). BF16 fallback
+   layers have no quantization tensors and are routed via `ignored_layers`.
+
+4. **Scale dtype**: fine scales are stored as `uint8` in safetensors (same bit
+   layout as `float8_e8m0fnu`) and reinterpreted at load time without a lossy
+   float32 round-trip.
+
+5. **Cascade model config propagation**: in a cascade model (transformer +
+   transformer_2), vLLM-Omni reads each transformer's own `config.json` and
+   rebuilds the quant config locally when `ignored_layers` differs between
+   transformers, ensuring per-layer routing is accurate for each. The first
+   transformer's config is propagated to `od_config` so the second transformer
+   can reuse it as a starting point.
+
+6. **Self-attention QKV fusion**: Q, K, V projection weights are fused into a
+   single `QKVParallelLinear` layer at runtime. `ignored_layers` entries use the
+   fused name (`attn1.to_qkv`), written automatically by the merge script.
+
+7. W4A4 carries higher quantization noise than W8A8 (16 vs 256 levels). The
+   DualScale offline method mitigates this with calibrated `mul_scale` smooth
+   quantization. Use `ignored_layers` and `num_bf16_fallback_layers` to trade
+   off compression vs. accuracy for precision-sensitive layers.
+
+## Adapting MXFP4 for a New Model
+
+This section is aimed at developers who want to add MXFP4 support to a model
+other than Wan2.2. The three integration points are: (1) discovering the correct
+runtime layer names, (2) wiring `ignored_layers` into the model, and (3) writing
+a merge script for offline checkpoints.
+
+### Step 1 — Discover runtime layer names
+
+`ignored_layers` entries must match the **runtime parameter names** used inside
+vllm-omni, which may differ from the names stored in the diffusers checkpoint.
+The canonical source of truth is the model's own `named_parameters()`.
+
+```python
+from vllm_omni import Omni
+
+# Load the model without quantization to inspect parameter names.
+omni = Omni(model="/path/to/your-model")  # no --quantization flag
+for name, _ in omni.pipeline.transformer.named_parameters():
+    if "weight" in name and "scale" not in name:
+        print(name)
+```
+
+Compare the printed names against the diffusers checkpoint keys
+(`safetensors.safe_open` or `torch.load`) to identify any renames your model
+applies. Common patterns that differ in Wan2.2 (and may appear in other
+models):
+
+| Diffusers checkpoint name | vllm-omni runtime name | Reason |
+|---------------------------|------------------------|--------|
+| `attn1.to_q`, `attn1.to_k`, `attn1.to_v` | `attn1.to_qkv` | Self-attention Q/K/V fused into `QKVParallelLinear` |
+| `ffn.net.0.proj` | `ffn.net_0.proj` | Dots in sub-module names replaced with underscores |
+| `ffn.net.2` | `ffn.net_2` | Same underscore rule |
+| `to_out.0` | `to_out` | Sequential index stripped |
+
+If your model has different fusion patterns, inspect `packed_modules_mapping`
+on the model class — this dict records how checkpoint keys are mapped to
+fused runtime parameters.
+
+!!! warning "Partial QKV fallback is not allowed"
+    If your model fuses Q, K, V into a single layer, `ignored_layers` must
+    include **all three or none**. A partial fallback (e.g. `to_q` in BF16 but
+    `to_k`, `to_v` quantized) cannot be expressed at runtime because they share
+    one `QKVParallelLinear`. The merge script enforces this and raises an error
+    if only some of the trio appear as non-quantized.
+
+### Step 2 — Add ignored_layers to the model
+
+#### Online mode
+
+Pass `ignored_layers` directly in the quantization config using the **runtime
+names** discovered in Step 1. No code changes to the model are required.
+
+```python
+omni = Omni(
+    model="/path/to/your-model",
+    quantization={
+        "method": "mxfp4_dualscale",
+        "ignored_layers": [
+            "blocks.0.attn1.to_qkv",   # runtime name, not diffusers name
+            "blocks.0.attn1.to_out",
+            "blocks.0.ffn.net_0.proj",
+        ],
+    },
+)
+```
+
+```bash
+# CLI does not support list-typed ignored_layers directly.
+# Use the Python API or set ignored_layers in config.json (offline).
+python your_script.py --model /path/to/your-model --quantization mxfp4_dualscale
+```
+
+The `num_bf16_fallback_layers` coarse rule is an alternative to listing layers
+individually: set it to N to keep all linear layers in blocks 0 … N-1 in BF16.
+The right value depends on the model's sensitivity; evaluate on a validation
+set and pick the smallest N that meets your accuracy target.
+
+#### Offline mode
+
+For offline checkpoints, `ignored_layers` is written into each transformer's
+`config.json` by the merge script (see Step 3). No manual editing is needed if
+the merge script is correct. The injected block:
+
+```json
+{
+  "quant_method": "mxfp4_dualscale",
+  "is_checkpoint_serialized": true,
+  "ignored_layers": [
+    "blocks.0.attn1.to_qkv",
+    "blocks.0.attn1.to_out"
+  ]
+}
+```
+
+To add a layer manually (e.g. to pin an additional layer to BF16 without
+re-running the merge script), edit `config.json` inside the transformer
+subfolder. Use runtime names, not diffusers checkpoint names.
+
+### Step 3 — Write a merge script for offline mode
+
+The merge script for a new model mirrors
+`vllm_omni/quantization/tools/merge_mxfp4_dualscale_checkpoint.py`. The four
+things it must do:
+
+1. **Remap tensor names** from the quantization tool convention to diffusers
+   convention (strip wrappers like `.linear.`, `.div.`; fix any prefix
+   differences).
+
+2. **Collect ignored_layers**: after loading, enumerate all `*.weight` keys that
+   have no corresponding `*.weight_scale` (i.e. layers the tool left in BF16).
+   Convert diffusers names to vllm-omni runtime names (fuse QKV, rename FFN
+   sub-modules, etc.). Write the result to `config.json`.
+
+3. **Inject `quantization_config`** into `config.json`:
+   ```python
+   config["quantization_config"] = {
+       "quant_method":              "mxfp4_dualscale",
+       "is_checkpoint_serialized":  True,
+       "ignored_layers":            ignored_layers,   # runtime names
+   }
+   ```
+
+4. **Save** the merged safetensors and the updated `config.json`.
+
+The key helper to implement is the diffusers-to-runtime name translator
+(equivalent to `_diffusers_to_vllm_ignored` in the Wan2.2 merge script).
+For each non-quantized diffusers weight key, apply your model's specific
+renaming rules and collect the results.
+
+!!! tip "Validate before serving"
+    After producing the offline checkpoint, load it without a `--quantization`
+    flag and verify that vLLM-Omni auto-detects the correct method. Check that
+    the layer count reported in the startup log matches expectations: quantized
+    layer count + `ignored_layers` count should equal total linear layer count.
+    Any mismatch indicates a name-mapping bug in the merge script.

docs/user_guide/quantization/online.md

@@ -9,20 +9,24 @@ separate quantized checkpoint.
 This mode is different from pre-quantized checkpoint formats such as GGUF,
 AutoRound, msModelSlim, or serialized Int8 checkpoints. Those formats are
 prepared before serving and are documented in their method-specific guides.
+For MXFP8 and MXFP4, use this page for load-time quantization from BF16
+checkpoints, and use the method-specific pages for offline checkpoints produced
+by msModelSlim and the merge tools.
 
 ## Hardware Support
 
-| Device | FP8 W8A8 | Int8 W8A8 |
-|--------|----------|-----------|
-| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ |
-| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ |
-| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ |
-| AMD ROCm | ⭕ | ⭕ |
-| Intel XPU | ⭕ | ⭕ |
-| Ascend NPU | ❌ | ✅ |
+| Device | FP8 W8A8 | Int8 W8A8 | MXFP8 W8A8 | MXFP4 W4A4 |
+|--------|----------|-----------|------------|------------|
+| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ⭕ | ⭕ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ⭕ | ⭕ |
+| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ⭕ | ⭕ |
+| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ |
+| Intel XPU | ⭕ | ⭕ | ⭕ | ⭕ |
+| Ascend NPU | ❌ | ✅ | ✅ | ✅ |
 
 Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
-guide. FP8 on Ampere may use a weight-only path where available.
+guide. FP8 on Ampere may use a weight-only path where available. MXFP8 and
+MXFP4 are documented for the Ascend NPU path.
 
 ## Model Type Support
 
@@ -32,6 +36,8 @@ guide. FP8 on Ampere may use a weight-only path where available.
 |--------|-------|----------------|--------|
 | FP8 W8A8 | [FP8](fp8.md) | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image family and other DiT models |
 | Int8 W8A8 | [Int8](int8.md) | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image and Z-Image |
+| MXFP8 W8A8 | [MXFP8](mxfp8.md) | Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, Wan2.2-TI2V-5B | Ascend NPU only |
+| MXFP4 W4A4 | [MXFP4](mxfp4.md) | Wan2.2-T2V-A14B, Wan2.2-I2V-A14B | Ascend NPU only; TI2V-5B is not supported |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 
@@ -53,6 +59,9 @@ from vllm_omni import Omni
 
 omni_fp8 = Omni(model="<your-model>", quantization="fp8")
 omni_int8 = Omni(model="<your-model>", quantization="int8")
+omni_mxfp8 = Omni(model="<your-model>", quantization="mxfp8")
+omni_mxfp4 = Omni(model="<your-model>", quantization="mxfp4")
+omni_mxfp4_dualscale = Omni(model="<your-model>", quantization="mxfp4_dualscale")
 ```
 
 CLI:
@@ -60,6 +69,9 @@ CLI:
 ```bash
 vllm serve <your-model> --omni --quantization fp8
 vllm serve <your-model> --omni --quantization int8
+vllm serve <your-model> --omni --quantization mxfp8
+vllm serve <your-model> --omni --quantization mxfp4
+vllm serve <your-model> --omni --quantization mxfp4_dualscale
 ```
 
 Per-component routing:
@@ -77,10 +89,11 @@ config = build_quant_config({
 
 | Parameter | Methods | Description |
 |-----------|---------|-------------|
-| `method` | FP8, Int8 | Quantization method, usually `"fp8"` or `"int8"` |
-| `ignored_layers` | FP8, Int8 | Layer name patterns to keep in BF16/FP16 |
+| `method` | FP8, Int8, MXFP8, MXFP4 | Quantization method: `"fp8"`, `"int8"`, `"mxfp8"`, `"mxfp4"`, or `"mxfp4_dualscale"` |
+| `ignored_layers` | FP8, Int8, MXFP8, MXFP4 | Layer name patterns to keep in BF16/FP16 |
 | `activation_scheme` | FP8, Int8 | The runtime value `"dynamic"` selects online activation scaling |
 | `weight_block_size` | FP8 | Optional block-wise FP8 weight quantization size |
+| `num_bf16_fallback_layers` | MXFP4 DualScale | Leading transformer blocks to keep in BF16 for online `mxfp4_dualscale`; defaults to `5` |
 
 ## Validation and Notes
 
@@ -91,3 +104,5 @@ config = build_quant_config({
    model as supported.
 4. If a model already ships quantized weights, use the matching pre-quantized
    method guide instead of online quantization.
+5. For Ascend MXFP4 deployments, prefer offline `mxfp4_dualscale` checkpoints
+   when production quality is more important than avoiding preprocessing.

docs/user_guide/quantization/overview.md

@@ -9,20 +9,20 @@ type has a different quantization scope.
 
 | Mode | Guide | Description | Methods |
 |------|-------|-------------|---------|
-| Online quantization | [Online Quantization](online.md) | vLLM-Omni computes quantized weights and scales while loading the model. | FP8 W8A8, Int8 W8A8, MXFP8 W8A8 |
+| Online quantization | [Online Quantization](online.md) | vLLM-Omni computes quantized weights and scales while loading the model. | FP8 W8A8, Int8 W8A8, MXFP8 W8A8, MXFP4 W4A4 |
 | Runtime attention quantization | [Quantized KV Cache](quantized_kvcache.md) | vLLM-Omni dynamically quantizes eligible diffusion Flash Attention tensors during inference. | FP8 FA |
-| Pre-quantized checkpoints | Method-specific guides | The checkpoint or an offline quantizer provides quantized weights and scales before serving. | ModelOpt, GGUF, AutoRound, msModelSlim, serialized Int8, offline MXFP8 |
+| Pre-quantized checkpoints | Method-specific guides | The checkpoint or an offline quantizer provides quantized weights and scales before serving. | ModelOpt, GGUF, AutoRound, msModelSlim, serialized Int8, offline MXFP8, offline MXFP4 DualScale |
 
 ## Hardware Support
 
-| Device | FP8 W8A8 | Int8 W8A8 | ModelOpt | MXFP8 W8A8 | GGUF | AutoRound | msModelSlim |
-|--------|----------|-----------|----------|------------|------|-----------|-------------|
-| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ⭕ | ✅ | ✅ | ❌ |
-| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ⭕ | ✅ | ✅ | ❌ |
-| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
-| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ❌ |
-| Intel XPU | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
-| Ascend NPU | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ |
+| Device | FP8 W8A8 | Int8 W8A8 | ModelOpt | MXFP8 W8A8 | MXFP4 W4A4 | GGUF | AutoRound | msModelSlim |
+|--------|----------|-----------|----------|------------|------------|------|-----------|-------------|
+| NVIDIA Blackwell GPU (SM 100+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
+| NVIDIA Ada/Hopper GPU (SM 89+) | ✅ | ✅ | ✅ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
+| NVIDIA Ampere GPU (SM 80+) | ✅ | ✅ | ⭕ | ⭕ | ⭕ | ✅ | ✅ | ❌ |
+| AMD ROCm | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ❌ |
+| Intel XPU | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ⭕ | ✅ | ❌ |
+| Ascend NPU | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ |
 
 Legend: `✅` supported, `❌` unsupported, `⭕` not verified in this
 guide. FP8 on Ampere may use a weight-only path where available.
@@ -42,6 +42,7 @@ otherwise.
 | Int8 W8A8 | [Int8](int8.md) | Online or serialized W8A8 | Qwen-Image; Wan2.2 is not validated | Validated for Qwen-Image and Z-Image |
 | ModelOpt | [ModelOpt](modelopt.md) | Pre-quantized FP8 checkpoints | Qwen-Image, Z-Image, FLUX.2, HunyuanImage-3.0 | Validated for ModelOpt FP8 diffusion checkpoints |
 | MXFP8 W8A8 | [MXFP8](mxfp8.md) | Online W8A8 or offline pre-quantized | Wan2.2-T2V-A14B, I2V-A14B, TI2V-5B | Ascend NPU only; validated for Wan2.2 |
+| MXFP4 W4A4 | [MXFP4](mxfp4.md) | `mxfp4`: online single-scale only; `mxfp4_dualscale`: online or offline dual-scale (offline recommended) | Wan2.2-T2V-A14B, I2V-A14B | Ascend NPU only; validated for Wan2.2 A14B cascade models; TI2V-5B not supported; offline `mxfp4_dualscale` uses calibrated `mul_scale` for best accuracy |
 | GGUF | [GGUF](gguf.md) | Pre-quantized transformer weights | Qwen-Image | Validated where a model-specific GGUF adapter exists |
 | AutoRound | [AutoRound](autoround.md) | Pre-quantized W4A16 checkpoints | FLUX.1-dev; Qwen-Image/Wan2.2 not validated | Checkpoint-driven |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Pre-quantized Ascend checkpoints | Wan2.2 recipe; HunyuanImage-3.0 inference target | Ascend/NPU path |
@@ -58,6 +59,7 @@ in BF16 unless the model guide explicitly adds support.
 | ModelOpt | [ModelOpt](modelopt.md) | Thinker or language-model checkpoint config | Qwen3-Omni thinker | ModelOpt checkpoint path |
 | Int8 | [Int8](int8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | MXFP8 | [MXFP8](mxfp8.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
+| MXFP4 | [MXFP4](mxfp4.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | GGUF | [GGUF](gguf.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
 | AutoRound | [AutoRound](autoround.md) | Thinker or language-model checkpoint config | Qwen2.5-Omni, Qwen3-Omni | Supported through AutoRound checkpoints |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Not currently validated for omni/TTS stages | Qwen3-Omni, Qwen3-TTS | Not validated |
@@ -73,6 +75,7 @@ attached to the intended stage rather than applied globally.
 | Int8 | [Int8](int8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Requires model-specific validation |
 | ModelOpt | [ModelOpt](modelopt.md) | Checkpoint-defined diffusion stage | BAGEL, GLM-Image | Requires model-specific validation |
 | MXFP8 | [MXFP8](mxfp8.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Not validated |
+| MXFP4 | [MXFP4](mxfp4.md) | Stage-specific DiT or transformer module | BAGEL, GLM-Image | Not validated |
 | GGUF | [GGUF](gguf.md) | Stage-specific transformer weights | BAGEL, GLM-Image | No validated adapter listed |
 | AutoRound | [AutoRound](autoround.md) | Checkpoint-defined stage | BAGEL, GLM-Image | No validated checkpoint listed |
 | msModelSlim | [msModelSlim](msmodelslim.md) | Ascend-generated stage weights | GLM-Image | Requires model-specific adaptation |
@@ -100,7 +103,7 @@ config = build_quant_config({
 
 | Component | Default quantized? | Notes |
 |-----------|--------------------|-------|
-| Diffusion transformer | Yes | Primary target for FP8, Int8, ModelOpt, GGUF, AutoRound, and msModelSlim |
+| Diffusion transformer | Yes | Primary target for FP8, Int8, ModelOpt, MXFP8, MXFP4, GGUF, AutoRound, and msModelSlim |
 | Text encoder | No | Keep BF16 unless a method-specific guide documents support |
 | VAE | No | Keep BF16; storage-only paths are method-specific |
 | Scheduler/tokenizer | No | Loaded from the base model repository |
@@ -138,3 +141,98 @@ build_quant_config({"method": "gguf", "gguf_model": "/path/to/model.gguf"})
 build_quant_config({"transformer": {"method": "fp8"}, "vae": None})
 build_quant_config(None)
 ```
+
+## Output Similarity Comparison Tool
+
+Use `vllm_omni.quantization.tools.compare_diffusion_trajectory_similarity`
+to compare a reference diffusion run with a quantized candidate run using the
+same prompt, seed, resolution, scheduler settings, and inference steps. The
+tool compares final decoded images or video frames, and also reports generation
+latency and worker-reported peak memory when available.
+
+This is useful when validating whether online quantization, an offline
+pre-quantized checkpoint, or a new `ignored_layers` choice keeps generation
+quality close to the BF16 reference.
+
+### Online Quantization Example
+
+```bash
+python -m vllm_omni.quantization.tools.compare_diffusion_trajectory_similarity \
+  --task t2i \
+  --model Qwen/Qwen-Image \
+  --candidate-quantization fp8 \
+  --ignored-layers img_mlp \
+  --prompt "a cup of coffee on the table" \
+  --height 512 --width 512 \
+  --num-inference-steps 20 \
+  --seed 142 \
+  --output-json /tmp/qwen_image_fp8_similarity/result.json \
+  --save-output-dir /tmp/qwen_image_fp8_similarity/images \
+  --enforce-eager
+```
+
+### Offline Checkpoint Example
+
+Use `--candidate-model` when the candidate is already quantized or lives at a
+different model path:
+
+```bash
+python -m vllm_omni.quantization.tools.compare_diffusion_trajectory_similarity \
+  --task t2i \
+  --reference-model Qwen/Qwen-Image \
+  --candidate-model /path/to/qwen-image-fp8-checkpoint \
+  --prompt "a cup of coffee on the table" \
+  --height 512 --width 512 \
+  --num-inference-steps 20 \
+  --seed 142 \
+  --output-json /tmp/qwen_image_fp8_checkpoint_similarity/result.json
+```
+
+If the checkpoint does not include a loadable quantization config, pass one
+explicitly:
+
+```bash
+--candidate-quantization-config-json '{"method":"fp8"}'
+```
+
+### Output Metrics
+
+The output JSON includes `output_metrics`, `reference_generation`, and
+`candidate_generation`.
+
+| Metric | Direction | Meaning |
+|--------|-----------|---------|
+| `cosine_similarity` | Higher is better | Vector direction similarity between output pixels or frames. Useful as a broad sanity check. |
+| `mae` | Lower is better | Mean absolute pixel or frame error. For decoded outputs, values are in uint8 pixel units. |
+| `mse` / `rmse` | Lower is better | Squared error and its square root. These penalize localized large differences more than `mae`. |
+| `max_abs` | Lower is better | Worst single-element absolute error. Treat it as an outlier/debug signal, not as a release gate. |
+| `l2` / `relative_l2` | Lower is better | Absolute and reference-normalized L2 distance. `relative_l2` is easier to compare across resolutions. |
+| `psnr_db` | Higher is better | Pixel-space signal-to-noise ratio in dB for uint8 images or frames. |
+| `avg_generation_time_s` | Lower is better | Average wall-clock generation time across measured runs. |
+| `max_peak_memory_mb` | Lower is better | Maximum worker-reported peak device memory across measured runs, when the worker reports it. |
+
+Recommended starting thresholds for same-seed diffusion comparisons:
+
+| Metric | Smoke threshold | Stricter target | Notes |
+|--------|-----------------|-----------------|-------|
+| `psnr_db` | `>= 20.0` | `>= 25.0` | Good for quick image or frame regression checks. |
+| `mae` | `<= 12.0` | `<= 6.0` | Interpreted in decoded uint8 pixel units. |
+| `cosine_similarity` | `>= 0.98` | `>= 0.995` | Less sensitive to global scale than L2-style metrics. |
+| `relative_l2` | `<= 0.20` | `<= 0.08` | Useful when comparing across prompts or resolutions. |
+
+These thresholds are heuristics. Tune them by model family, task, resolution,
+quantization method, and deployment tolerance. For release gating, pair the
+numeric report with visual inspection of saved reference and candidate outputs.
+
+The tool intentionally reports separate quality, latency, and memory metrics
+instead of a single consolidated similarity score. A single score can hide
+important tradeoffs, for example a candidate with good PSNR but a meaningful
+memory regression, or a candidate with low average error but localized visual
+artifacts. If you need a project-specific pass/fail gate, define it as an
+explicit policy over the individual metrics.
+
+Pixel-level metrics do not measure semantic consistency. For higher-cost
+evaluation, you can complement this report with a vision-language judge that
+describes the reference and candidate outputs and compares those descriptions.
+Keep that semantic check separate from this lightweight tool so users can
+choose whether the additional model cost and latency are appropriate.

docs/user_guide/quantization/quantized_kvcache.md

@@ -8,15 +8,20 @@ vLLM-Omni supports online FP8 quantization for eligible diffusion Flash
 Attention (FA) to reduce FA latency while keeping model weights in their
 original dtype.
 
-This feature is configured through `kv_cache_dtype`, matching the option name
-used by vLLM's language-model KV-cache quantization. In vLLM-Omni diffusion
-pipelines, however, it is a runtime FA path: Q/K/V tensors are dynamically
-quantized before the attention operator. It does not quantize model weights and
-is separate from [FP8 W8A8](fp8.md), [Int8 W8A8](int8.md), or pre-quantized
-checkpoint formats.
+This feature is configured through `diffusion_kv_cache_dtype` on
+`OmniDiffusionConfig` (CLI: `--diffusion-kv-cache-dtype`). It is intentionally
+**not** the same as vLLM's `--kv-cache-dtype`, which controls autoregressive
+language-model KV cache storage and defaults to `"auto"`. Diffusion FA
+quantization uses the dedicated diffusion flags so omni serve does not inherit
+that default.
 
-If `kv_cache_dtype` is not set, behavior is unchanged and attention runs in the
-native dtype.
+In vLLM-Omni diffusion pipelines, this is a runtime FA path: Q/K/V tensors are
+dynamically quantized before the attention operator. It does not quantize model
+weights and is separate from [FP8 W8A8](fp8.md), [Int8 W8A8](int8.md), or
+pre-quantized checkpoint formats.
+
+If `diffusion_kv_cache_dtype` is not set, behavior is unchanged and attention
+runs in the native dtype.
 
 ## Hardware Support
 
@@ -30,8 +35,8 @@ native dtype.
 Legend: `✅` supported, `❌` unsupported.
 
 FP8 FA is currently implemented only for the NPU Flash Attention backend. Other
-backends do not support `kv_cache_dtype="fp8"` for diffusion attention and fall
-back to native dtype execution.
+backends do not support `diffusion_kv_cache_dtype="fp8"` for diffusion attention
+and fall back to native dtype execution.
 
 ## Model Type Support
 
@@ -40,7 +45,7 @@ back to native dtype execution.
 | Model | Scope | Status | Notes |
 |-------|-------|--------|-------|
 | Wan2.2 | Eligible DiT full-attention FA on Ascend NPU | Tested | Compare quality and latency against a BF16 baseline before production use |
-| Other diffusion models | Eligible DiT full-attention FA on Ascend NPU | Not tested | You can try `kv_cache_dtype="fp8"`; tune `kv_cache_skip_steps` and `kv_cache_skip_layers` when higher precision is needed |
+| Other diffusion models | Eligible DiT full-attention FA on Ascend NPU | Not tested | You can try `diffusion_kv_cache_dtype="fp8"`; tune `diffusion_kv_cache_skip_steps` and `diffusion_kv_cache_skip_layers` when higher precision is needed |
 
 ### Multi-Stage Omni/TTS Model (Qwen3-Omni, Qwen3-TTS)
 
@@ -50,8 +55,8 @@ guide documents support.
 ### Multi-Stage Diffusion Model (BAGEL, GLM-Image)
 
 Not tested. If the diffusion stage uses the same NPU Flash Attention backend,
-`kv_cache_dtype` may apply in theory; validate quality and latency for each
-stage and model.
+`diffusion_kv_cache_dtype` may apply in theory; validate quality and latency for
+each stage and model.
 
 ## Configuration
 
@@ -67,15 +72,15 @@ python examples/offline_inference/image_to_video/image_to_video.py \
     --num-inference-steps 4 \
     --ulysses-degree 4 \
     --vae-patch-parallel-size 4 \
-    --kv-cache-dtype fp8 \
-    --kv-cache-skip-steps "0,1" \
-    --kv-cache-skip-layers "0-2"
+    --diffusion-kv-cache-dtype fp8 \
+    --diffusion-kv-cache-skip-steps "0,1" \
+    --diffusion-kv-cache-skip-layers "0-2"
 ```
 
 Online serving:
 
 ```bash
-vllm serve <your-model> --omni --kv-cache-dtype fp8
+vllm serve <your-model> --omni --diffusion-kv-cache-dtype fp8
 ```
 
 Stage config:
@@ -86,18 +91,23 @@ stage_args:
     stage_type: diffusion
     engine_args:
       model_stage: dit
-      kv_cache_dtype: "fp8"
-      kv_cache_skip_steps: "0,1"
-      kv_cache_skip_layers: "0-2"
+      diffusion_kv_cache_dtype: "fp8"
+      diffusion_kv_cache_skip_steps: "0,1"
+      diffusion_kv_cache_skip_layers: "0-2"
 ```
 
+Legacy YAML keys `kv_cache_dtype`, `kv_cache_skip_steps`, and
+`kv_cache_skip_layers` are still accepted when constructing
+`OmniDiffusionConfig` (for example via `from_kwargs`); prefer the `diffusion_*`
+names for new configs.
+
 ## Parameters
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `kv_cache_dtype` | str \| None | `None` | Set to `"fp8"` to enable dynamic FP8 FA on supported attention backends |
-| `kv_cache_skip_steps` | str \| None | `None` | Denoising step selector to keep in native dtype, for example `"0,1,4-6"` |
-| `kv_cache_skip_layers` | str \| None | `None` | Transformer layer selector to keep in native dtype, for example `"0-2,10"` |
+| `diffusion_kv_cache_dtype` | str \| None | `None` | Set to `"fp8"` to enable dynamic FP8 FA on supported attention backends |
+| `diffusion_kv_cache_skip_steps` | str \| None | `None` | Denoising step selector to keep in native dtype, for example `"0,1,4-6"` |
+| `diffusion_kv_cache_skip_layers` | str \| None | `None` | Transformer layer selector to keep in native dtype, for example `"0-2,10"` |
 
 Selectors use comma-separated integers and inclusive ranges. Listed steps or
 layers skip FP8 FA; all other eligible full-attention forwards use the FP8 path.
@@ -106,9 +116,9 @@ layers skip FP8 FA; all other eligible full-attention forwards use the FP8 path.
 
 1. Compare generated images or videos against a BF16 baseline with the same
    seed, prompt, resolution, frame count, and denoising steps.
-2. Use `kv_cache_skip_steps` for denoising steps where quality is more
+2. Use `diffusion_kv_cache_skip_steps` for denoising steps where quality is more
    sensitive.
-3. Use `kv_cache_skip_layers` for transformer layers that show visible quality
+3. Use `diffusion_kv_cache_skip_layers` for transformer layers that show visible quality
    regressions.
 4. Report both latency and quality results when enabling this option for a new
    model. For image or video models, include visual comparison and quantitative

examples/offline_inference/hunyuan_image3/README.md

@@ -112,6 +112,7 @@ python end2end.py --modality text2img \
                   --additional-config '{"torchair_graph_config":{"enabled":true}}'
 ```
 
+
 ## Key Arguments
 
 | Argument | Description |
@@ -123,16 +124,15 @@ python end2end.py --modality text2img \
 | `--steps` | Number of diffusion inference steps for image generation. |
 | `--guidance-scale` | Classifier-free guidance scale for image generation. |
 | `--height`, `--width` | Output image size for `text2img`. |
-| `--bot-task` | Prompt behavior. `auto` selects the default from `--modality`; `think` adds `<think>`; `recaption` adds `<recaption>`; `vanilla` uses the text-to-image pretrain template. |
+| `--bot-task` | Override prompt mode. `none`, `think`, `recaption`, `think_recaption`, or `vanilla`. |
 | `--sys-type` | Override the system prompt type, for example `en_unified` or `en_vanilla`. |
 | `--vae-use-tiling` | Enable VAE tiling for memory reduction. |
 
 ## Notes
 
-- `hunyuan_image3_ar.yaml` is a 4-card AR-only text/comprehension deploy. It sets `engine_output_type: text`, `final_output_type: text`, and text sampling defaults.
-- `hunyuan_image3_dit.yaml` is a single-stage DiT deploy with `stage_id: 0`; it does not require stage 1 or a running AR stage.
+- `hunyuan_image3_ar.yaml` is a 4-card AR-only text/comprehension deploy.
+- `hunyuan_image3_dit.yaml` is a single-stage DiT deploy with `stage_id: 0`.
 - The old HunyuanImage3 YAMLs under `model_executor/stage_configs/` and `platforms/*/stage_configs/` have been folded into the deploy YAMLs.
-- This PR does not keep the HunyuanImage3 AR-to-DiT KV reuse wiring. The deploy YAMLs describe the topology and platform settings only.
 
 ## Prompt Format
 
@@ -148,22 +148,8 @@ Assistant: {trigger_tag?}
 
 - `<img>`: Placeholder for each input image (single token; expanded by the multimodal pipeline).
 - Trigger tags: `<think>` for CoT and `<recaption>` for recaptioning, placed after `Assistant: `.
-- System prompt: Auto-selected based on task.
-- `t2i_vanilla` is the only task that uses the bare pretrain template without chat structure.
-- The example composes the internal prompt task from `--modality` and `--bot-task`
-  before calling `prompt_utils`; for example, `img2text + think` becomes
-  `i2t_think` for prompt and stop-token lookup.
+- System prompt: Auto-selected from `task` and `bot_task`.
+- `bot_task='vanilla'` with `task='t2i'` uses the bare pretrain template.
 
 The shared `vllm_omni.diffusion.models.hunyuan_image3.prompt_utils.build_prompt_tokens()`
 helper handles segment-by-segment tokenization and matches HF `apply_chat_template`.
-
-## FAQ
-
-- **OOM errors**: Decrease `gpu_memory_utilization` in the deploy YAML, use a smaller `max_num_batched_tokens`, or enable VAE tiling with `--vae-use-tiling`.
-- **Custom image sizes**: Use `--height` and `--width` flags (multiples of 16 recommended).
-
-| Stage | VRAM (approx) |
-| :--- | :--- |
-| Stage 0 (AR) | ~15 GiB + KV Cache |
-| Stage 1 (DiT) | ~30 GiB |
-| Total (8-GPU) | ~45 GiB + KV Cache |

examples/offline_inference/hunyuan_image3/end2end.py

@@ -1,16 +1,5 @@
 """
 HunyuanImage-3.0-Instruct unified end-to-end inference script.
-
-Supports all modalities through a single entry point:
-  - text2img:  Text → AR → DiT → Image
-  - img2img:   Text+Image → AR → DiT → Edited Image (IT2I)
-  - img2text:  Image+Text → AR → Text description (I2T)
-  - text2text: Text → AR → Text (comprehension, no image)
-
-Usage:
-    python end2end.py --modality text2img --prompts "A cute cat"
-    python end2end.py --modality img2img --image-path input.png --prompts "Make it snowy"
-    python end2end.py --modality img2text --image-path input.png --prompts "Describe this image"
 """
 
 import argparse
@@ -19,18 +8,25 @@ import os
 from pathlib import Path
 
 from vllm_omni.diffusion.models.hunyuan_image3.prompt_utils import (
-    _TASK_PRESETS,
+    MAX_IMAGES_PER_REQUEST,
     build_prompt_tokens,
     resolve_stop_token_ids,
+    resolve_sys_type,
 )
 from vllm_omni.entrypoints.omni import Omni
 from vllm_omni.inputs.data import OmniPromptType
 
-# Default deploy configs are absolute so this example works from any cwd.
 _REPO_ROOT = Path(__file__).resolve().parents[3]
 _DEFAULT_DEPLOY_CONFIG = str(_REPO_ROOT / "vllm_omni" / "deploy" / "hunyuan_image3.yaml")
 _DEFAULT_AR_DEPLOY_CONFIG = str(_REPO_ROOT / "vllm_omni" / "deploy" / "hunyuan_image3_ar.yaml")
 
+_MODALITY_TASK_MAP: dict[str, tuple[str, str | None]] = {
+    "text2img": ("t2i", "think"),
+    "img2img": ("it2i", "think"),
+    "img2text": ("i2t", None),
+    "text2text": ("t2t", None),
+}
+
 _MODALITY_DEFAULT_DEPLOY_CONFIG = {
     "text2img": _DEFAULT_DEPLOY_CONFIG,
     "img2img": _DEFAULT_DEPLOY_CONFIG,
@@ -45,78 +41,60 @@ _MODALITY_MODE = {
     "text2text": "text-to-text",
 }
 
-_MODALITY_TASK_MAP = {
-    "text2img": "t2i",
-    "img2img": "it2i",
-    "img2text": "i2t",
-    "text2text": "t2t",
-}
-
 
 def parse_args():
     parser = argparse.ArgumentParser(description="HunyuanImage-3.0-Instruct end-to-end inference.")
-    parser.add_argument(
-        "--model",
-        default="tencent/HunyuanImage-3.0-Instruct",
-        help="Model name or local path.",
-    )
+    parser.add_argument("--model", default="tencent/HunyuanImage-3.0-Instruct", help="Model name or local path.")
     parser.add_argument(
         "--modality",
         default="text2img",
-        choices=["text2img", "img2img", "img2text", "text2text"],
-        help="Modality mode to control stage execution.",
+        choices=list(_MODALITY_TASK_MAP),
     )
     parser.add_argument("--prompts", nargs="+", default=None, help="Input text prompts.")
     parser.add_argument(
         "--image-path",
         type=str,
         default=None,
-        help="Path to input image (for img2img/img2text).",
+        help="Input image path(s) for img2img/img2text. Comma-separated for multi-image (up to 3).",
     )
-    parser.add_argument(
-        "--output",
-        type=str,
-        default=".",
-        help="Output directory to save results.",
-    )
-
-    # Generation parameters
+    parser.add_argument("--output", type=str, default=".", help="Output directory to save results.")
     parser.add_argument("--steps", type=int, default=50, help="Number of inference steps.")
     parser.add_argument("--guidance-scale", type=float, default=5.0, help="Classifier-free guidance scale.")
     parser.add_argument("--seed", type=int, default=42, help="Random seed.")
     parser.add_argument("--height", type=int, default=1024, help="Output image height.")
     parser.add_argument("--width", type=int, default=1024, help="Output image width.")
-    parser.add_argument(
-        "--vae-use-tiling",
-        action="store_true",
-        help="Enable VAE tiling for memory optimization.",
-    )
-
-    # Prompt configuration
+    parser.add_argument("--vae-use-tiling", action="store_true", help="Enable VAE tiling.")
     parser.add_argument(
         "--bot-task",
         type=str,
-        default="auto",
-        choices=["auto", "think", "recaption", "think_recaption", "vanilla"],
-        help=(
-            "Prompt behavior. 'auto' selects the default for the modality; "
-            "'think' adds <think>; 'recaption' adds <recaption>; "
-            "'vanilla' uses the t2i pretrain template."
-        ),
-    )
-    parser.add_argument(
-        "--sys-type",
-        type=str,
         default=None,
-        help="Override system prompt type (e.g. en_unified, en_vanilla).",
+        choices=["none", "think", "recaption", "think_recaption", "vanilla"],
+        help="Override prompt mode. Default: auto from --modality.",
     )
-
-    # Omni init args
+    parser.add_argument("--sys-type", type=str, default=None, help="Override system prompt type.")
     parser.add_argument("--deploy-config", type=str, default=None, help="Custom deploy YAML path.")
     parser.add_argument("--stage-configs-path", type=str, default=None, help="Custom legacy stage config YAML path.")
     parser.add_argument("--log-stats", action="store_true", default=False)
     parser.add_argument("--init-timeout", type=int, default=300, help="Initialization timeout in seconds.")
     parser.add_argument("--enforce-eager", action="store_true", help="Disable torch.compile.")
+    parser.add_argument(
+        "--diffusion-kv-cache-dtype",
+        type=str,
+        default=None,
+        help="Diffusion attention KV cache dtype, for example 'fp8'. Separate from vLLM --kv-cache-dtype.",
+    )
+    parser.add_argument(
+        "--diffusion-kv-cache-skip-steps",
+        type=str,
+        default=None,
+        help="Denoising step selector to keep diffusion KV cache in native dtype, for example '0,1,4-6'.",
+    )
+    parser.add_argument(
+        "--diffusion-kv-cache-skip-layers",
+        type=str,
+        default=None,
+        help="Transformer layer selector to keep diffusion KV cache in native dtype, for example '0-2,10'.",
+    )
     parser.add_argument(
         "--additional-config",
         type=str,
@@ -157,22 +135,13 @@ def main():
     os.makedirs(args.output, exist_ok=True)
     additional_config = parse_additional_config(args.additional_config)
 
-    # Determine task for prompt formatting from modality + bot behavior.
-    task = _MODALITY_TASK_MAP[args.modality]
-    assert task is not None
-    bot_task = args.bot_task
-    if bot_task != "auto":
-        task = task + "_" + bot_task
-    if task not in _TASK_PRESETS:
-        valid_bot_tasks = {
-            "text2img": ["think", "recaption", "vanilla"],
-            "img2img": ["think", "recaption", "think_recaption"],
-            "img2text": ["auto"],
-            "text2text": ["auto"],
-        }[args.modality]
-        raise ValueError(
-            f"--bot-task {bot_task!r} is not supported for {args.modality}. Choose from: {valid_bot_tasks}"
-        )
+    task, default_bot_task = _MODALITY_TASK_MAP[args.modality]
+    if args.bot_task is None:
+        bot_task: str | None = default_bot_task
+    elif args.bot_task == "none":
+        bot_task = None
+    else:
+        bot_task = args.bot_task
 
     if args.deploy_config is not None and args.stage_configs_path is not None:
         raise ValueError("--deploy-config and --stage-configs-path are mutually exclusive.")
@@ -182,13 +151,16 @@ def main():
     if deploy_config is None and stage_configs_path is None:
         deploy_config = _MODALITY_DEFAULT_DEPLOY_CONFIG[args.modality]
 
-    # Build Omni
     omni_kwargs = {
         "model": args.model,
         "vae_use_tiling": args.vae_use_tiling,
         "log_stats": args.log_stats,
         "init_timeout": args.init_timeout,
         "enforce_eager": args.enforce_eager,
+        "mode": _MODALITY_MODE[args.modality],
+        "diffusion_kv_cache_dtype": args.diffusion_kv_cache_dtype,
+        "diffusion_kv_cache_skip_steps": args.diffusion_kv_cache_skip_steps,
+        "diffusion_kv_cache_skip_layers": args.diffusion_kv_cache_skip_layers,
     }
 
     if additional_config is not None:
@@ -197,71 +169,67 @@ def main():
         omni_kwargs["deploy_config"] = deploy_config
     else:
         omni_kwargs["stage_configs_path"] = stage_configs_path
-    omni_kwargs["mode"] = _MODALITY_MODE[args.modality]
 
     omni = Omni(**omni_kwargs)
 
-    # Prepare prompts
     prompts = args.prompts or ["A cute cat"]
-    if not prompts:
-        print("[Info] No prompts provided, using default.")
-        prompts = ["A cute cat"]
-
-    # Load image if needed
-    input_image = None
+    input_images: list = []
     if args.modality in ("img2img", "img2text"):
-        if not args.image_path or not os.path.exists(args.image_path):
+        if not args.image_path:
             raise ValueError(f"--image-path required for {args.modality}, got: {args.image_path}")
         from PIL import Image
 
-        input_image = Image.open(args.image_path).convert("RGB")
+        image_paths = [p.strip() for p in args.image_path.split(",") if p.strip()]
+        if len(image_paths) > MAX_IMAGES_PER_REQUEST:
+            raise ValueError(
+                f"--image-path accepts at most {MAX_IMAGES_PER_REQUEST} images for "
+                f"HunyuanImage-3.0 IT2I, got {len(image_paths)}: {args.image_path}"
+            )
+        for image_path in image_paths:
+            if not os.path.exists(image_path):
+                raise ValueError(f"Image path does not exist: {image_path}")
+            input_images.append(Image.open(image_path).convert("RGB"))
+        if not input_images:
+            raise ValueError(f"--image-path produced no usable paths: {args.image_path!r}")
 
-    # Load tokenizer for segment-wise prompt tokenization (matches HF
-    # apply_chat_template byte-for-byte; see build_prompt_tokens docstring).
     from transformers import AutoTokenizer
 
     tokenizer = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
+    mm_image_payload = (input_images[0] if len(input_images) == 1 else input_images) if input_images else None
 
-    # Format prompts
     formatted_prompts: list[OmniPromptType] = []
-    for p in prompts:
-        result = build_prompt_tokens(p, tokenizer, task=task, sys_type=args.sys_type)
+    for prompt in prompts:
+        build_kwargs: dict = {"task": task, "bot_task": bot_task, "sys_type": args.sys_type}
+        if input_images:
+            build_kwargs["num_images"] = len(input_images)
+        result = build_prompt_tokens(prompt, tokenizer, **build_kwargs)
         token_ids = result.token_ids
-        effective_sys_type = result.system_prompt_type
+        effective_sys_type = args.sys_type or resolve_sys_type(bot_task)
 
-        # `prompt_token_ids` drives the AR stage (matches HF byte-for-byte).
-        # `prompt` and `use_system_prompt` are forwarded by ar2diffusion to
-        # the DiT stage so the diffusion pipeline can rebuild the same
-        # system prefix when constructing its model inputs.
         prompt_dict: dict = {
             "prompt_token_ids": token_ids,
-            "prompt": p,
+            "prompt": prompt,
             "use_system_prompt": effective_sys_type,
         }
-
         if args.modality == "text2img":
             prompt_dict["modalities"] = ["image"]
         elif args.modality == "img2img":
             prompt_dict["modalities"] = ["image"]
-            prompt_dict["multi_modal_data"] = {"image": input_image}
-            prompt_dict["height"] = input_image.height
-            prompt_dict["width"] = input_image.width
+            prompt_dict["multi_modal_data"] = {"image": mm_image_payload}
+            prompt_dict["height"] = input_images[0].height
+            prompt_dict["width"] = input_images[0].width
         elif args.modality == "img2text":
             prompt_dict["modalities"] = ["text"]
-            prompt_dict["multi_modal_data"] = {"image": input_image}
-        elif args.modality == "text2text":
+            prompt_dict["multi_modal_data"] = {"image": mm_image_payload}
+        else:
             prompt_dict["modalities"] = ["text"]
-
         formatted_prompts.append(prompt_dict)
 
-    # Build sampling params from defaults
     params_list = list(omni.default_sampling_params_list)
 
-    # Override diffusion params if applicable
     from vllm_omni.inputs.data import OmniDiffusionSamplingParams
 
     ar_stop_token_ids = resolve_stop_token_ids(task=task, bot_task=bot_task, tokenizer=tokenizer)
-    assert ar_stop_token_ids is not None
     for sp in params_list:
         if isinstance(sp, OmniDiffusionSamplingParams):
             sp.num_inference_steps = args.steps
@@ -269,13 +237,12 @@ def main():
             sp.guidance_scale_provided = True
             if args.seed is not None:
                 sp.seed = args.seed
-            if args.modality in ("text2img",):
+            if args.modality == "text2img":
                 sp.height = args.height
                 sp.width = args.width
         elif hasattr(sp, "stop_token_ids"):
             sp.stop_token_ids = ar_stop_token_ids
 
-    # Print configuration
     print(f"\n{'=' * 60}")
     print("HunyuanImage-3.0 Generation Configuration:")
     print(f"  Model: {args.model}")
@@ -291,6 +258,9 @@ def main():
         print(f"  Inference steps: {args.steps}")
         print(f"  Guidance scale: {args.guidance_scale}")
         print(f"  Seed: {args.seed}")
+        print(f"  diffusion_kv_cache_dtype: {args.diffusion_kv_cache_dtype}")
+        print(f"  diffusion_kv_cache_skip_steps: {args.diffusion_kv_cache_skip_steps}")
+        print(f"  diffusion_kv_cache_skip_layers: {args.diffusion_kv_cache_skip_layers}")
     if args.modality == "text2img":
         print(f"  Output size: {args.width}x{args.height}")
     if args.image_path:
@@ -300,13 +270,9 @@ def main():
     print(f"  Prompts: {prompts}")
     print(f"{'=' * 60}\n")
 
-    # Generate
     omni_outputs = list(omni.generate(prompts=formatted_prompts, sampling_params_list=params_list))
-
-    # Process outputs
     img_idx = 0
     for req_output in omni_outputs:
-        # Text output (AR stage or text-only)
         ro = getattr(req_output, "request_output", None)
         txt = ""
         if ro and getattr(ro, "outputs", None):
@@ -320,11 +286,9 @@ def main():
         if txt:
             print(f"[Output] Text:\n{txt}")
 
-        # Image output (DiT stage)
         images = getattr(req_output, "images", None)
         if not images and ro and hasattr(ro, "images"):
             images = ro.images
-
         if images:
             for j, img in enumerate(images):
                 save_path = os.path.join(args.output, f"output_{img_idx}_{j}.png")

examples/offline_inference/image_to_video/image_to_video.py

@@ -103,22 +103,22 @@ def parse_args() -> argparse.Namespace:
         help="Sampling solver for Wan2.2 pipelines. Use 'euler' for Lightning/Distill setups.",
     )
     parser.add_argument(
-        "--kv-cache-dtype",
+        "--diffusion-kv-cache-dtype",
         type=str,
         default=None,
-        help="Config-level KV cache dtype (e.g. float8_e4m3fn).",
+        help="Diffusion attention KV cache dtype (e.g. float8_e4m3fn). Separate from vLLM --kv-cache-dtype.",
     )
     parser.add_argument(
-        "--kv-cache-skip-steps",
+        "--diffusion-kv-cache-skip-steps",
         type=str,
         default=None,
-        help="Config-level KV-cache quantization skip-step selector, e.g. '0-9,20,25-30'.",
+        help="Diffusion KV-cache quantization skip-step selector, e.g. '0-9,20,25-30'.",
     )
     parser.add_argument(
-        "--kv-cache-skip-layers",
+        "--diffusion-kv-cache-skip-layers",
         type=str,
         default=None,
-        help="Config-level KV-cache quantization skip-layer selector, e.g. '0,1,4-8'.",
+        help="Diffusion KV-cache quantization skip-layer selector, e.g. '0,1,4-8'.",
     )
     parser.add_argument("--output", type=str, default="i2v_output.mp4", help="Path to save the video (mp4).")
     parser.add_argument("--fps", type=int, default=None, help="Frames per second for the output video.")
@@ -142,6 +142,42 @@ def parse_args() -> argparse.Namespace:
         action="store_true",
         help="Enable layerwise (blockwise) offloading on DiT modules.",
     )
+    parser.add_argument(
+        "--enforce-eager",
+        action="store_true",
+        help="Disable torch.compile and force eager execution.",
+    )
+    parser.add_argument(
+        "--audio-sample-rate",
+        type=int,
+        default=24000,
+        help="Sample rate for audio output when saved (default: 24000).",
+    )
+    parser.add_argument(
+        "--cache-backend",
+        type=str,
+        default=None,
+        choices=["cache_dit", "tea_cache"],
+        help=(
+            "Cache backend to use for acceleration. "
+            "Options: 'cache_dit' (DBCache + SCM + TaylorSeer), 'tea_cache' (Timestep Embedding Aware Cache). "
+            "Default: None (no cache acceleration)."
+        ),
+    )
+    parser.add_argument(
+        "--enable-diffusion-pipeline-profiler",
+        action="store_true",
+        help="Enable diffusion pipeline profiler to display stage durations.",
+    )
+    parser.add_argument(
+        "--quantization",
+        type=str,
+        default=None,
+        choices=["fp8", "mxfp8", "mxfp4", "mxfp4_dualscale", "int8", "gguf"],
+        help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 (NPU). mxfp4: W4A4 MXFP4 (NPU). mxfp4_dualscale: W4A4 MXFP4 dual-scale + BF16 fallback mixed (NPU). fp8: online FP8 (GPU).",
+    )
+
+    # Distributed and parallel execution
     parser.add_argument(
         "--ulysses-degree",
         type=int,
@@ -173,28 +209,6 @@ def parse_args() -> argparse.Namespace:
         default=1,
         help="Number of GPUs used for VAE patch/tile parallelism (decode).",
     )
-    parser.add_argument(
-        "--enforce-eager",
-        action="store_true",
-        help="Disable torch.compile and force eager execution.",
-    )
-    parser.add_argument(
-        "--audio-sample-rate",
-        type=int,
-        default=24000,
-        help="Sample rate for audio output when saved (default: 24000).",
-    )
-    parser.add_argument(
-        "--cache-backend",
-        type=str,
-        default=None,
-        choices=["cache_dit", "tea_cache"],
-        help=(
-            "Cache backend to use for acceleration. "
-            "Options: 'cache_dit' (DBCache + SCM + TaylorSeer), 'tea_cache' (Timestep Embedding Aware Cache). "
-            "Default: None (no cache acceleration)."
-        ),
-    )
     parser.add_argument(
         "--use-hsdp",
         action="store_true",
@@ -219,16 +233,10 @@ def parse_args() -> argparse.Namespace:
         ),
     )
     parser.add_argument(
-        "--quantization",
-        type=str,
-        default=None,
-        choices=["fp8", "mxfp8", "int8", "gguf"],
-        help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 online quant (NPU). fp8: online FP8 (GPU).",
-    )
-    parser.add_argument(
-        "--enable-diffusion-pipeline-profiler",
-        action="store_true",
-        help="Enable diffusion pipeline profiler to display stage durations.",
+        "--pipeline-parallel-size",
+        type=int,
+        default=1,
+        help="Number of pipeline parallel stages.",
     )
     parser.add_argument(
         "--profiler-config",
@@ -326,6 +334,7 @@ def main():
         use_hsdp=args.use_hsdp,
         hsdp_shard_size=args.hsdp_shard_size,
         hsdp_replicate_size=args.hsdp_replicate_size,
+        pipeline_parallel_size=args.pipeline_parallel_size,
     )
     omni_kwargs = dict(
         model=args.model,
@@ -334,9 +343,9 @@ def main():
         vae_use_tiling=args.vae_use_tiling,
         boundary_ratio=args.boundary_ratio,
         flow_shift=args.flow_shift,
-        kv_cache_dtype=args.kv_cache_dtype,
-        kv_cache_skip_steps=args.kv_cache_skip_steps,
-        kv_cache_skip_layers=args.kv_cache_skip_layers,
+        diffusion_kv_cache_dtype=args.diffusion_kv_cache_dtype,
+        diffusion_kv_cache_skip_steps=args.diffusion_kv_cache_skip_steps,
+        diffusion_kv_cache_skip_layers=args.diffusion_kv_cache_skip_layers,
         enable_cpu_offload=args.enable_cpu_offload,
         parallel_config=parallel_config,
         enforce_eager=args.enforce_eager,
@@ -361,12 +370,13 @@ def main():
     print(f"  Inference steps: {args.num_inference_steps}")
     print(f"  Frames: {args.num_frames}")
     print(f"  Solver: {args.sample_solver}")
-    print(f"  kv_cache_dtype(config): {args.kv_cache_dtype}")
-    print(f"  kv_cache_skip_steps(config): {args.kv_cache_skip_steps}")
-    print(f"  kv_cache_skip_layers(config): {args.kv_cache_skip_layers}")
+    print(f"  diffusion_kv_cache_dtype(config): {args.diffusion_kv_cache_dtype}")
+    print(f"  diffusion_kv_cache_skip_steps(config): {args.diffusion_kv_cache_skip_steps}")
+    print(f"  diffusion_kv_cache_skip_layers(config): {args.diffusion_kv_cache_skip_layers}")
     print(
         f"  Parallel configuration: cfg_parallel_size={args.cfg_parallel_size},"
-        f" tensor_parallel_size={args.tensor_parallel_size}, vae_patch_parallel_size={args.vae_patch_parallel_size}"
+        f" tensor_parallel_size={args.tensor_parallel_size}, vae_patch_parallel_size={args.vae_patch_parallel_size},"
+        f" pipeline_parallel_size={args.pipeline_parallel_size}"
     )
     print(f"  Video size: {args.width}x{args.height}")
     print(f"{'=' * 60}\n")

examples/offline_inference/text_to_image/README.md

@@ -35,6 +35,7 @@ This folder provides several entrypoints for experimenting with text-to-image di
 | `black-forest-labs/FLUX.2-klein-9B` | 1024 x 1024 | 37.1 | 32.3 |
 | `black-forest-labs/FLUX.2-dev` | 1024 x 1024 | 65.7 | >80 (CPU offload required) |
 | `HunyuanImage-3.0` | 1024 x 1024 | 80.0 (TP≥3)  | 160 |
+| `HiDream-I1-Full` | 1024 x 1024 | 63.7  | 57.7 |
 
 !!! info
 *Peak VRAM:  based on basic single-card usage, batch size =1, without any acceleration/optimization features. FLUX.2-dev requires `--enable-cpu-offload` on a single 80 GiB GPU.
@@ -94,6 +95,7 @@ python text_to_image.py \
 | `--lora-scale` | float | `1.0` | Scale factor for LoRA weights |
 | `--use-system-prompt` | str | `None` | System prompt preset: `en_unified`, `en_vanilla`, `en_recaption`, `en_think_recaption`, `dynamic`, `None`, or custom text. Recommended: `en_unified`. Only for HunyuanImage-3.0.|
 | `--system-prompt` | str | `None` | Custom system prompt text. Only used when `--use-system-prompt` is set to `custom`. Only for HunyuanImage-3.0.|
+| `--auxiliary-text-encoder` | str | `None` | Supplementary auxiliary text encoder parameters model name or path (especially for Hidream-l1-full). |
 
 **NextStep-1.1 specific arguments:**
 
@@ -177,6 +179,24 @@ python examples/offline_inference/text_to_image/text_to_image.py \
   --output flux2-dev.png
 ```
 
+
+### HiDream-I1-Full Models
+
+The `--auxiliary-text-encoder` parameter is required when running HiDream‑I1‑Full:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model HiDream-ai/HiDream-I1-Full \
+  --prompt "The setting sun of late autumn dyes the riverside with a warm orange hue" \
+  --seed 42 \
+  --guidance-scale 5.0 \
+  --tensor-parallel-size 1 \
+  --num-images-per-prompt 1 \
+  --num-inference-steps 50 \
+  --auxiliary-text-encoder meta-llama/Meta-Llama-3.1-8B-Instruct \
+  --output /output.png
+```
+
 ### Batch Requests (Multiple Prompts)
 
 You can pass multiple prompts in a single `generate` call.

examples/offline_inference/text_to_image/text_to_image.py

@@ -300,6 +300,12 @@ def parse_args() -> argparse.Namespace:
         default=None,
         help=("Custom system prompt. Used when --use-system-prompt is custom. "),
     )
+    parser.add_argument(
+        "--auxiliary-text-encoder",
+        type=str,
+        default=None,
+        help="Supplementary auxiliary text encoder parameters model name or path (especially for Hidream-l1-full).",
+    )
     current_omni_platform.pre_register_and_update(parser)
     from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
 
@@ -392,6 +398,7 @@ def main():
         "profiler_config": args.profiler_config,
         "init_timeout": args.init_timeout,
         "stage_init_timeout": args.stage_init_timeout,
+        "auxiliary_text_encoder": args.auxiliary_text_encoder,
         **lora_args,
         **quant_kwargs,
     }

examples/offline_inference/text_to_speech/README.md

@@ -14,13 +14,13 @@ list of supported architectures across all modalities, see
 
 | Model | HuggingFace repo | Stages | Voice cloning | Streaming | Special modes | Sample rate |
 |---|---|---|---|---|---|---|
-| CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | 2 (talker + code2wav) | ✓ | ✓ | — | 22.05 kHz |
+| CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | 2 (talker + code2wav) | ✓ | ✓ | — | 24 kHz |
 | Fish Speech S2 Pro | `fishaudio/s2-pro` | dual-AR | ✓ | ✓ | — | 44.1 kHz |
+| GLM-TTS | `zai-org/GLM-TTS` | 2 (AR + DiT) | ✓ (required) | ✓ | — | 24 kHz |
 | Ming-flash-omni-TTS | `Jonathan1909/Ming-flash-omni-2.0` | single (talker only) | — (caption-controlled) | — | style / IP / basic captions | 44.1 kHz |
 | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | single (AR + codec) | ✓ (required) | ✓ | voice_clone, continuation | 48 kHz |
 | OmniVoice | `k2-fsa/OmniVoice` | 2 (gen + dec) | ✓ | — | voice design, language hint | 24 kHz |
 | Qwen3-TTS | `Qwen/Qwen3-TTS-12Hz-1.7B-{CustomVoice,VoiceDesign,Base}` | 2 (talker + code2wav) | ✓ (Base) | ✓ | 3 task variants | 24 kHz |
-| VoxCPM | `openbmb/VoxCPM-0.5B` | split | ✓ | ✓ | — | 24 kHz |
 | VoxCPM2 | `openbmb/VoxCPM2` | single (native AR) | ✓ | ✓ (online) | continuation | 48 kHz |
 | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | varies | ✓ | ✓ | voice presets | 24 kHz |
 
@@ -41,7 +41,7 @@ python examples/offline_inference/text_to_speech/<model>/end2end.py \
 
 ## CosyVoice3
 
-2-stage TTS pipeline (`talker` + `code2wav`) at 22.05 kHz.
+2-stage TTS pipeline (`talker` + `code2wav`) at 24 kHz.
 
 ### Prerequisites
 ```bash
@@ -73,13 +73,15 @@ python examples/offline_inference/text_to_speech/cosyvoice3/end2end.py \
 ```
 
 ### Voice cloning
-Pass a reference audio. Note that CosyVoice3's `--prompt-text` is a system-style prompt for the GPT stage, not a reference transcript:
+If `--ref-audio` is omitted, the script downloads the upstream
+[`zero_shot_prompt.wav`](https://github.com/FunAudioLLM/CosyVoice/blob/main/asset/zero_shot_prompt.wav) from the CosyVoice repo into the current directory.
+To use your own clip, pass `--ref-audio /path/to/reference.wav`, and modify `--prompt-text` correspondingly.
 ```bash
 python examples/offline_inference/text_to_speech/cosyvoice3/end2end.py \
     --model pretrained_models/Fun-CosyVoice3-0.5B \
     --tokenizer pretrained_models/Fun-CosyVoice3-0.5B/CosyVoice-BlankEN \
-    --ref-audio prompt.wav \
-    --prompt-text "You are a helpful assistant.<|endofprompt|>Testing my voices. Why should I not?"
+    --ref-audio /path/to/reference.wav \
+    --prompt-text "You are a helpful assistant.<|endofprompt|>Trascript in your ref audio clip"
 ```
 
 ### Streaming
@@ -91,6 +93,35 @@ Streaming is enabled by default via `async_chunk: true` in `vllm_omni/deploy/cos
 
 ---
 
+## GLM-TTS
+
+2-stage TTS pipeline (AR + DiT flow-matching) at 24 kHz. Every request requires reference audio and its transcript for zero-shot voice cloning.
+
+### Quick start
+```bash
+python examples/offline_inference/text_to_speech/glm_tts/end2end.py \
+    --model zai-org/GLM-TTS \
+    --text "你好，这是语音合成测试。" \
+    --ref-audio /path/to/reference.wav \
+    --ref-text "这是参考音频的文本内容。" \
+    --output-dir ./output
+```
+
+### Architecture
+```
+Text → [Stage 0: AR] → Speech Tokens → [Stage 1: DiT + HiFT] → Audio (24 kHz)
+        (Llama-based)    (32k vocab)      (Flow Matching)
+```
+
+### Notes
+- `--ref-audio` and `--ref-text` are **required** together; GLM-TTS does not support text-only synthesis.
+- Reference audio should be 3-10 seconds.
+- First run may be slow due to lazy loading of WhisperVQ tokenizer and CampPlus ONNX speaker embedder.
+- Default sampling: temperature=1.0, top_k=25, top_p=0.8 (RAS method).
+- The `--model` path should point to the repository root (not `llm/` subdirectory).
+
+---
+
 ## Fish Speech S2 Pro
 
 4B dual-AR text-to-speech model from FishAudio with the DAC codec at 44.1 kHz.
@@ -306,61 +337,6 @@ python examples/offline_inference/text_to_speech/qwen3_tts/end2end.py \
 
 ---
 
-## VoxCPM
-
-Split-stage TTS. The hub example covers single TTS, single voice cloning, and streaming. Use `benchmarks/voxcpm/` for warmup, batch JSONL prompts, profiler injection, and offline TTFP / RTF measurement.
-
-### Prerequisites
-```bash
-pip install voxcpm soundfile
-# or use a local source tree:
-export VLLM_OMNI_VOXCPM_CODE_PATH=/path/to/VoxCPM/src
-```
-
-If the native VoxCPM `config.json` does not contain HF metadata such as `model_type`, prepare a persistent HF-compatible config directory and point the stage configs to it via `VLLM_OMNI_VOXCPM_HF_CONFIG_PATH`:
-
-```bash
-export VOXCPM_MODEL=/path/to/voxcpm-model
-export VLLM_OMNI_VOXCPM_HF_CONFIG_PATH=/tmp/voxcpm_hf_config
-mkdir -p "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"
-cp "$VOXCPM_MODEL/config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/config.json"
-cp "$VOXCPM_MODEL/generation_config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/generation_config.json" 2>/dev/null || true
-python3 -c 'import json, os; p=os.path.join(os.environ["VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"], "config.json"); cfg=json.load(open(p, "r", encoding="utf-8")); cfg["model_type"]="voxcpm"; cfg.setdefault("architectures", ["VoxCPMForConditionalGeneration"]); json.dump(cfg, open(p, "w", encoding="utf-8"), indent=2, ensure_ascii=False)'
-```
-
-### Quick start
-```bash
-python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This is a split-stage VoxCPM synthesis example running on vLLM Omni."
-```
-
-### Voice cloning
-```bash
-python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This sentence is synthesized with a cloned voice." \
-    --ref-audio /path/to/reference.wav \
-    --ref-text  "The exact transcript spoken in reference.wav."
-```
-
-### Streaming
-Pass the async-chunk stage config:
-```bash
-python examples/offline_inference/text_to_speech/voxcpm/end2end.py \
-    --model "$VOXCPM_MODEL" \
-    --stage-configs-path vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml \
-    --text "This is a split-stage VoxCPM streaming example running on vLLM Omni."
-```
-
-### Notes
-- `voxcpm.yaml` is the default non-streaming stage config; `voxcpm_async_chunk.yaml` enables streaming.
-- Streaming is currently single-request oriented.
-- `--ref-text` must be the real transcript of `--ref-audio`; mismatched text degrades quality.
-- For online serving, see the [VoxCPM section in the online hub](../../online_serving/text_to_speech/README.md#voxcpm). For benchmark reporting, see [`benchmarks/voxcpm`](../../../benchmarks/voxcpm/README.md).
-
----
-
 ## VoxCPM2
 
 Single-stage native AR TTS at 48 kHz. Pipeline: `feat_encoder → MiniCPM4 → FSQ → residual_lm → LocDiT → AudioVAE`.

examples/offline_inference/text_to_speech/cosyvoice3/end2end.py

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

examples/offline_inference/text_to_speech/glm_tts/end2end.py

@@ -0,0 +1,194 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""GLM-TTS End-to-End Offline Inference Example.
+
+GLM-TTS is a two-stage TTS system:
+  - Stage 0 (AR): Llama-based model generates speech tokens from text
+  - Stage 1 (DiT): Flow matching model converts speech tokens to audio
+
+Usage:
+    # Sync two-stage (default)
+    python examples/offline_inference/text_to_speech/glm_tts/end2end.py \
+        --model /path/to/GLM-TTS \
+        --text "你好，这是一个语音合成测试。" \
+        --ref-audio /path/to/reference.wav \
+        --ref-text "参考音频的转录文本。" \
+        --output-dir ./output
+
+    # Async chunk mode (streaming DiT)
+    python examples/offline_inference/text_to_speech/glm_tts/end2end.py \
+        --model /path/to/GLM-TTS --async-chunk \
+        --text "你好，这是一个语音合成测试。" \
+        --ref-audio /path/to/reference.wav \
+        --ref-text "参考音频的转录文本。" \
+        --output-dir ./output
+"""
+
+import base64
+import io
+import logging
+import os
+import tempfile
+import time
+from typing import Any
+from urllib.request import urlopen
+
+import numpy as np
+import soundfile as sf
+import torch
+import yaml
+
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+from vllm_omni import Omni
+from vllm_omni.model_executor.models.glm_tts.glm_tts import build_glm_tts_prefill_metadata
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_DEPLOY_CONFIG = os.path.join(
+    os.path.dirname(__file__),
+    "..",
+    "..",
+    "..",
+    "..",
+    "vllm_omni",
+    "deploy",
+    "glm_tts.yaml",
+)
+SAMPLE_RATE = 24000
+
+
+def _load_ref_audio(ref_audio: str) -> tuple[torch.Tensor, int]:
+    """Load reference audio from file path, URL, or data URI."""
+    if ref_audio.startswith(("http://", "https://")):
+        with urlopen(ref_audio, timeout=60) as response:
+            audio_obj: Any = io.BytesIO(response.read())
+    elif ref_audio.startswith("data:"):
+        _, _, encoded = ref_audio.partition(",")
+        audio_obj = io.BytesIO(base64.b64decode(encoded))
+    else:
+        audio_obj = ref_audio
+    wav_np, sr = sf.read(audio_obj, dtype="float32")
+    if wav_np.ndim > 1:
+        wav_np = wav_np.mean(axis=1)
+    return torch.from_numpy(wav_np), int(sr)
+
+
+def _concat_audio(audio_val: Any) -> np.ndarray:
+    """Concatenate audio tensors from multimodal output."""
+    if isinstance(audio_val, list):
+        tensors = [torch.as_tensor(t).float().reshape(-1) for t in audio_val if t is not None]
+        if not tensors:
+            return np.zeros((0,), dtype=np.float32)
+        return torch.cat(tensors, dim=-1).cpu().numpy().astype(np.float32, copy=False)
+    if isinstance(audio_val, torch.Tensor):
+        return audio_val.float().cpu().numpy().reshape(-1)
+    return np.asarray(audio_val, dtype=np.float32).reshape(-1)
+
+
+def _extract_sample_rate(audio_mm: dict) -> int:
+    """Extract sample rate from multimodal output dict."""
+    sr_raw = audio_mm.get("sr", SAMPLE_RATE)
+    if isinstance(sr_raw, list):
+        sr_raw = sr_raw[-1] if sr_raw else SAMPLE_RATE
+    if hasattr(sr_raw, "item"):
+        return int(sr_raw.item())
+    return int(sr_raw)
+
+
+def _modify_deploy_config(base_path: str, async_chunk: bool) -> str:
+    """Build deploy config with explicit sync/async mode and eager execution.
+
+    Mirrors the logic in ``tests/e2e/offline_inference/test_glm_tts.py``
+    (``_get_deploy_config``) so that example runs match CI behavior.
+    """
+    with open(base_path) as f:
+        cfg = yaml.safe_load(f)
+    cfg["async_chunk"] = async_chunk
+    for stage in cfg.get("stages", []):
+        stage["enforce_eager"] = True
+        if stage.get("stage_id") == 0:
+            stage["async_scheduling"] = bool(async_chunk)
+    tmp = tempfile.NamedTemporaryFile(mode="w", suffix=".yaml", delete=False, prefix="glm_tts_")
+    yaml.dump(cfg, tmp)
+    tmp.close()
+    return tmp.name
+
+
+def main(args):
+    """Run offline GLM-TTS inference."""
+    os.makedirs(args.output_dir, exist_ok=True)
+    base_deploy_config = args.deploy_config or DEFAULT_DEPLOY_CONFIG
+    deploy_config_path = _modify_deploy_config(base_deploy_config, args.async_chunk)
+
+    ref_audio_wav, ref_audio_sr = _load_ref_audio(args.ref_audio)
+    if not args.ref_text:
+        raise ValueError("GLM-TTS requires --ref-audio and --ref-text.")
+
+    inputs = [
+        {
+            "prompt": args.text,
+            "multi_modal_data": {
+                "audio": (ref_audio_wav.float().cpu().numpy(), ref_audio_sr),
+            },
+            "modalities": ["audio"],
+            "mm_processor_kwargs": {"prompt_text": args.ref_text},
+            "additional_information": build_glm_tts_prefill_metadata(
+                args.model,
+                args.text,
+                args.ref_text,
+            ),
+        }
+    ]
+
+    omni = Omni(
+        model=args.model,
+        stage_configs_path=deploy_config_path,
+        log_stats=args.log_stats,
+        stage_init_timeout=args.stage_init_timeout,
+    )
+
+    t_start = time.perf_counter()
+    outputs = omni.generate(inputs)
+    elapsed = (time.perf_counter() - t_start) * 1000
+
+    assert outputs, "No outputs returned"
+    audio_mm = outputs[0].multimodal_output
+    assert "audio" in audio_mm, "No audio output found"
+
+    audio = _concat_audio(audio_mm["audio"])
+    sr = _extract_sample_rate(audio_mm)
+    out_path = os.path.join(args.output_dir, "output.wav")
+    sf.write(out_path, audio, samplerate=sr, format="WAV")
+
+    logger.info("Saved %s (%.2fs @ %dHz)", out_path, len(audio) / sr, sr)
+    logger.info("Total inference: %.1f ms", elapsed)
+
+
+def parse_args():
+    parser = FlexibleArgumentParser(description="GLM-TTS Text-to-Speech Example")
+    parser.add_argument("--model", type=str, required=True, help="Model path")
+    parser.add_argument("--text", type=str, default="你好，这是一个语音合成测试。")
+    parser.add_argument("--output-dir", type=str, default="./output")
+    parser.add_argument("--ref-audio", type=str, required=True, help="Reference WAV path/URL")
+    parser.add_argument("--ref-text", type=str, required=True, help="Transcript of ref audio")
+    parser.add_argument("--deploy-config", type=str, default=None)
+    parser.add_argument(
+        "--async-chunk",
+        action="store_true",
+        default=False,
+        help="Enable async_chunk mode (streaming DiT). Default: sync two-stage.",
+    )
+    parser.add_argument("--log-stats", action="store_true")
+    parser.add_argument("--stage-init-timeout", type=int, default=600)
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
+    )
+    main(parse_args())

examples/offline_inference/text_to_speech/voxcpm/end2end.py

@@ -1,208 +0,0 @@
-"""Minimal offline VoxCPM example for vLLM Omni."""
-
-from __future__ import annotations
-
-import asyncio
-import time
-from pathlib import Path
-from typing import Any
-
-import soundfile as sf
-import torch
-from vllm.utils.argparse_utils import FlexibleArgumentParser
-
-from tests.helpers.stage_config import get_deploy_config_path
-from vllm_omni import AsyncOmni, Omni
-from vllm_omni.engine.arg_utils import nullify_stage_engine_defaults
-
-DEFAULT_SYNC_STAGE_CONFIG = get_deploy_config_path("voxcpm.yaml")
-
-
-def _build_prompt(args) -> dict[str, Any]:
-    additional_information: dict[str, list[Any]] = {
-        "text": [args.text],
-        "cfg_value": [args.cfg_value],
-        "inference_timesteps": [args.inference_timesteps],
-        "min_len": [args.min_len],
-        "max_new_tokens": [args.max_new_tokens],
-    }
-    if args.streaming_prefix_len is not None:
-        additional_information["streaming_prefix_len"] = [args.streaming_prefix_len]
-    if args.ref_audio is not None:
-        additional_information["ref_audio"] = [args.ref_audio]
-    if args.ref_text is not None:
-        additional_information["ref_text"] = [args.ref_text]
-    return {
-        "prompt_token_ids": [1],
-        "additional_information": additional_information,
-    }
-
-
-def _extract_audio_tensor(mm: dict[str, Any]) -> torch.Tensor:
-    audio = mm.get("audio", mm.get("model_outputs"))
-    if audio is None:
-        raise ValueError("No audio output found in multimodal output.")
-    if isinstance(audio, list):
-        parts = [torch.as_tensor(item).float().cpu().reshape(-1) for item in audio]
-        audio = torch.cat(parts, dim=-1) if parts else torch.zeros(0)
-    if not isinstance(audio, torch.Tensor):
-        audio = torch.as_tensor(audio)
-    return audio.float().cpu().reshape(-1)
-
-
-def _extract_sample_rate(mm: dict[str, Any]) -> int:
-    sr_raw = mm.get("sr", 24000)
-    if isinstance(sr_raw, list) and sr_raw:
-        sr_raw = sr_raw[-1]
-    if hasattr(sr_raw, "item"):
-        return int(sr_raw.item())
-    return int(sr_raw)
-
-
-def _is_streaming_stage_config(stage_config_path: str) -> bool:
-    return "async_chunk" in Path(stage_config_path).stem
-
-
-def _save_audio(audio: torch.Tensor, sample_rate: int, output_dir: Path, request_id: str) -> Path:
-    output_dir.mkdir(parents=True, exist_ok=True)
-    output_path = output_dir / f"output_{request_id}.wav"
-    sf.write(
-        output_path,
-        audio.float().cpu().clamp(-1.0, 1.0).numpy(),
-        sample_rate,
-        format="WAV",
-        subtype="PCM_16",
-    )
-    return output_path
-
-
-async def _run_streaming(args) -> Path:
-    prompt = _build_prompt(args)
-    output_dir = Path(args.output_dir) if args.output_dir is not None else Path("output_audio_streaming")
-    request_id = "streaming_example"
-    sample_rate = 24000
-    buffered_samples = 0
-    chunks: list[torch.Tensor] = []
-    started = time.perf_counter()
-    omni = AsyncOmni(
-        model=args.model,
-        stage_configs_path=args.stage_configs_path,
-        log_stats=args.log_stats,
-        stage_init_timeout=args.stage_init_timeout,
-    )
-    try:
-        async for stage_output in omni.generate(prompt, request_id=request_id):
-            mm = getattr(stage_output, "multimodal_output", None)
-            if not isinstance(mm, dict):
-                request_output = getattr(stage_output, "request_output", None)
-                if request_output is None:
-                    continue
-                mm = getattr(request_output, "multimodal_output", None)
-                if not isinstance(mm, dict) and getattr(request_output, "outputs", None):
-                    mm = getattr(request_output.outputs[0], "multimodal_output", None)
-            if not isinstance(mm, dict):
-                continue
-            audio = _extract_audio_tensor(mm)
-            if audio.numel() == 0:
-                continue
-            sample_rate = _extract_sample_rate(mm)
-            if audio.numel() > buffered_samples:
-                delta = audio[buffered_samples:]
-                buffered_samples = int(audio.numel())
-            else:
-                delta = audio
-                buffered_samples += int(delta.numel())
-            if delta.numel() > 0:
-                chunks.append(delta)
-        if not chunks:
-            raise RuntimeError("No streaming audio chunks received from VoxCPM.")
-        output_audio = torch.cat(chunks, dim=0)
-        output_path = _save_audio(output_audio, sample_rate, output_dir, request_id)
-        print(f"Saved streaming audio to: {output_path} ({time.perf_counter() - started:.2f}s)")
-        return output_path
-    finally:
-        omni.shutdown()
-
-
-def _run_sync(args) -> Path:
-    prompt = _build_prompt(args)
-    output_dir = Path(args.output_dir) if args.output_dir is not None else Path("output_audio")
-    request_id = "sync_example"
-    started = time.perf_counter()
-    last_mm: dict[str, Any] | None = None
-    omni = Omni(
-        model=args.model,
-        stage_configs_path=args.stage_configs_path,
-        log_stats=args.log_stats,
-        stage_init_timeout=args.stage_init_timeout,
-    )
-    for stage_outputs in omni.generate(prompt):
-        request_output = getattr(stage_outputs, "request_output", None)
-        if request_output is None:
-            continue
-        outputs = getattr(request_output, "outputs", None)
-        if outputs:
-            for output in outputs:
-                mm = getattr(output, "multimodal_output", None)
-                if isinstance(mm, dict):
-                    last_mm = mm
-        mm = getattr(request_output, "multimodal_output", None)
-        if isinstance(mm, dict):
-            last_mm = mm
-    if last_mm is None:
-        raise RuntimeError("No audio output received from VoxCPM.")
-    output_path = _save_audio(
-        _extract_audio_tensor(last_mm),
-        _extract_sample_rate(last_mm),
-        output_dir,
-        request_id,
-    )
-    print(f"Saved audio to: {output_path} ({time.perf_counter() - started:.2f}s)")
-    return output_path
-
-
-def parse_args():
-    parser = FlexibleArgumentParser(description="Minimal offline VoxCPM example for vLLM Omni.")
-    parser.add_argument("--model", type=str, required=True, help="Local VoxCPM model directory.")
-    parser.add_argument(
-        "--stage-configs-path",
-        type=str,
-        default=DEFAULT_SYNC_STAGE_CONFIG,
-        help=("Stage config path. Use voxcpm.yaml for non-streaming or voxcpm_async_chunk.yaml for streaming."),
-    )
-    parser.add_argument("--text", type=str, required=True, help="Input text for synthesis.")
-    parser.add_argument("--ref-audio", type=str, default=None, help="Reference audio path for voice cloning.")
-    parser.add_argument("--ref-text", type=str, default=None, help="Transcript of the reference audio.")
-    parser.add_argument("--output-dir", type=str, default=None, help="Output directory for generated wav files.")
-    parser.add_argument("--cfg-value", type=float, default=2.0, help="Guidance value passed to VoxCPM.")
-    parser.add_argument("--inference-timesteps", type=int, default=10, help="Number of diffusion timesteps.")
-    parser.add_argument("--min-len", type=int, default=2, help="Minimum latent length.")
-    parser.add_argument("--max-new-tokens", type=int, default=4096, help="Maximum latent length.")
-    parser.add_argument(
-        "--streaming-prefix-len",
-        type=int,
-        default=3,
-        help="Streaming prefix length used by voxcpm_async_chunk.yaml.",
-    )
-    parser.add_argument("--stage-init-timeout", type=int, default=600, help="Stage initialization timeout in seconds.")
-    parser.add_argument("--log-stats", action="store_true", help="Enable vLLM Omni stats logging.")
-    nullify_stage_engine_defaults(parser)
-    args = parser.parse_args()
-    if (args.ref_audio is None) != (args.ref_text is None):
-        raise ValueError("Voice cloning requires --ref-audio and --ref-text together.")
-    return args
-
-
-def main(args) -> None:
-    route = "streaming" if _is_streaming_stage_config(args.stage_configs_path) else "sync"
-    print(f"Model: {args.model}")
-    print(f"Stage config: {args.stage_configs_path}")
-    print(f"Route: {route}")
-    if route == "streaming":
-        asyncio.run(_run_streaming(args))
-    else:
-        _run_sync(args)
-
-
-if __name__ == "__main__":
-    main(parse_args())

examples/offline_inference/text_to_video/text_to_video.py

@@ -142,6 +142,32 @@ def parse_args() -> argparse.Namespace:
         action="store_true",
         help="Enable layerwise (blockwise) offloading on DiT modules.",
     )
+    parser.add_argument(
+        "--audio-sample-rate",
+        type=int,
+        default=24000,
+        help="Sample rate for audio output when saved (default: 24000).",
+    )
+    parser.add_argument(
+        "--enable-diffusion-pipeline-profiler",
+        action="store_true",
+        help="Enable diffusion pipeline profiler to display stage durations.",
+    )
+    parser.add_argument(
+        "--profiler-config",
+        type=parse_profiler_config,
+        default=None,
+        help='JSON profiler config for torch/cuda profiling, e.g. \'{"profiler":"torch","torch_profiler_dir":"./perf"}\'.',
+    )
+    parser.add_argument(
+        "--quantization",
+        type=str,
+        default=None,
+        choices=["fp8", "mxfp8", "mxfp4", "mxfp4_dualscale", "int8", "gguf"],
+        help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 (NPU). mxfp4: W4A4 MXFP4 (NPU). mxfp4_dualscale: W4A4 MXFP4 dual-scale + BF16 fallback mixed (NPU). fp8: online FP8 (GPU).",
+    )
+
+    # Distributed and parallel execution
     parser.add_argument(
         "--ulysses-mode",
         type=str,
@@ -174,41 +200,23 @@ def parse_args() -> argparse.Namespace:
         default=1,
         help="Number of GPUs used for tensor parallelism (TP) inside the DiT.",
     )
-    parser.add_argument(
-        "--audio-sample-rate",
-        type=int,
-        default=24000,
-        help="Sample rate for audio output when saved (default: 24000).",
-    )
     parser.add_argument(
         "--vae-patch-parallel-size",
         type=int,
         default=1,
         help="Number of GPUs used for VAE patch/tile parallelism (decode).",
     )
+    parser.add_argument(
+        "--pipeline-parallel-size",
+        type=int,
+        default=1,
+        help="Number of pipeline parallel stages.",
+    )
     parser.add_argument(
         "--enable-expert-parallel",
         action="store_true",
         help="Enable expert parallelism for MoE layers.",
     )
-    parser.add_argument(
-        "--enable-diffusion-pipeline-profiler",
-        action="store_true",
-        help="Enable diffusion pipeline profiler to display stage durations.",
-    )
-    parser.add_argument(
-        "--profiler-config",
-        type=parse_profiler_config,
-        default=None,
-        help='JSON profiler config for torch/cuda profiling, e.g. \'{"profiler":"torch","torch_profiler_dir":"./perf"}\'.',
-    )
-    parser.add_argument(
-        "--quantization",
-        type=str,
-        default=None,
-        choices=["fp8", "mxfp8", "int8", "gguf"],
-        help="Quantization method for the transformer. mxfp8: W8A8 MXFP8 online quant (NPU). fp8: online FP8 (GPU).",
-    )
     parser.add_argument(
         "--use-hsdp",
         action="store_true",
@@ -229,6 +237,27 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
+def _extract_peak_memory_mb(result: Any) -> float:
+    """Pull worker-reported peak VRAM (MiB) generation result.
+
+    Mirrors vllm_omni/entrypoints/openai/serving_video.py:_extract_peak_memory_mb.
+    """
+    if isinstance(result, list):
+        result = result[0] if result else None
+    if result is None:
+        return 0.0
+    val = getattr(result, "peak_memory_mb", 0.0)
+    if not val:
+        inner = getattr(result, "request_output", None)
+        if isinstance(inner, list):
+            inner = inner[0] if inner else None
+        val = getattr(inner, "peak_memory_mb", 0.0)
+    try:
+        return float(val or 0.0)
+    except (TypeError, ValueError):
+        return 0.0
+
+
 def main():
     args = parse_args()
     model_class_name = args.model_class_name
@@ -262,6 +291,7 @@ def main():
         cfg_parallel_size=args.cfg_parallel_size,
         tensor_parallel_size=args.tensor_parallel_size,
         vae_patch_parallel_size=args.vae_patch_parallel_size,
+        pipeline_parallel_size=args.pipeline_parallel_size,
         enable_expert_parallel=args.enable_expert_parallel,
     )
 
@@ -307,7 +337,8 @@ def main():
     print(
         f"  Parallel configuration: ulysses_degree={args.ulysses_degree}, ring_degree={args.ring_degree},"
         f" cfg_parallel_size={args.cfg_parallel_size}, tensor_parallel_size={args.tensor_parallel_size},"
-        f" vae_patch_parallel_size={args.vae_patch_parallel_size}, enable_expert_parallel={args.enable_expert_parallel}"
+        f" vae_patch_parallel_size={args.vae_patch_parallel_size}, pipeline_parallel_size={args.pipeline_parallel_size},"
+        f" enable_expert_parallel={args.enable_expert_parallel}"
     )
     print(f"  Video size: {args.width}x{args.height}")
     print(f"{'=' * 60}\n")
@@ -338,6 +369,10 @@ def main():
     # Print profiling results
     print(f"Total generation time: {generation_time:.4f} seconds ({generation_time * 1000:.2f} ms)")
 
+    peak_mb = _extract_peak_memory_mb(frames)
+    if peak_mb:
+        print(f"Worker peak GPU memory (reserved): {peak_mb:.2f} MiB ({peak_mb / 1024:.2f} GiB)")
+
     audio = None
     if isinstance(frames, list):
         frames = frames[0] if frames else None

examples/offline_inference/x_to_video_audio/x_to_video_audio.md

@@ -101,3 +101,4 @@ Key arguments:
 - `--video-negative-prompt`: negative prompt for video generation.
 - `--audio-negative-prompt`: negative prompt for audio generation.
 - `--enable-cpu-offload`: enable CPU offload (defaults False).
+- `--cache-backend`: enable `cache_dit` for acceleration.

examples/offline_inference/x_to_video_audio/x_to_video_audio.py

@@ -63,6 +63,7 @@ def parse_args() -> argparse.Namespace:
         action="store_true",
         help="Enable layerwise (blockwise) offloading on DiT modules.",
     )
+    parser.add_argument("--cache-backend", type=str, default=None, choices=["cache_dit"], help="Cache backend.")
     parser.add_argument(
         "--use-hsdp",
         action="store_true",
@@ -146,12 +147,29 @@ def main() -> None:
         hsdp_replicate_size=args.hsdp_replicate_size,
     )
 
+    cache_config = None
+    if args.cache_backend == "cache_dit":
+        cache_config = {
+            "Fn_compute_blocks": 1,
+            "Bn_compute_blocks": 0,
+            "max_warmup_steps": 4,
+            "max_cached_steps": 20,
+            "residual_diff_threshold": 0.24,
+            "max_continuous_cached_steps": 3,
+            "enable_taylorseer": False,
+            "taylorseer_order": 1,
+            "scm_steps_mask_policy": None,
+            "scm_steps_policy": "dynamic",
+        }
+
     omni = Omni(
         model=args.model,
         parallel_config=parallel_config,
         model_type=args.model_type,
         enable_cpu_offload=args.enable_cpu_offload,
         enable_layerwise_offload=args.enable_layerwise_offload,
+        cache_backend=args.cache_backend,
+        cache_config=cache_config,
     )
     start = time.perf_counter()
     outputs = omni.generate(prompt, sampling_params)

examples/online_serving/text_to_speech/README.md

@@ -15,11 +15,11 @@ For the full list of supported architectures across all modalities, see
 | Model | HuggingFace repo | Voice cloning | Streaming | Voice presets / upload | Gradio demo |
 |---|---|---|---|---|---|
 | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✓ (`ref_audio`+`ref_text`) | ✓ (PCM stream) | — | ✓ |
+| GLM-TTS | `zai-org/GLM-TTS` | ✓ (`ref_audio`+`ref_text`, required) | ✓ (PCM stream) | — | ✓ |
 | Ming-flash-omni-TTS | `Jonathan1909/Ming-flash-omni-2.0` | — (caption-controlled) | — | caption fields (`instructions`) | — |
 | MOSS-TTS-Nano | `OpenMOSS-Team/MOSS-TTS-Nano` | ✓ (`ref_audio` required) | ✓ (PCM stream) | — | ✓ |
 | OmniVoice | `k2-fsa/OmniVoice` | (offline only) | — | — | — |
 | Qwen3-TTS | `Qwen/Qwen3-TTS-12Hz-1.7B-{CustomVoice,VoiceDesign,Base}` | ✓ (Base) | ✓ (PCM + WebSocket) | ✓ (presets + `/v1/audio/voices` upload) | ✓ (standard + FastRTC) |
-| VoxCPM | `openbmb/VoxCPM-0.5B` | ✓ | ✓ (PCM stream) | — | — |
 | VoxCPM2 | `openbmb/VoxCPM2` | ✓ | ✓ (AudioWorklet via gradio) | — | ✓ |
 | Voxtral TTS | `mistralai/Voxtral-4B-TTS-2603` | ✓ (gated upstream) | ✓ | ✓ (presets) | ✓ |
 
@@ -95,6 +95,45 @@ For full request-shape documentation (all parameters, response formats, error co
 
 ---
 
+## GLM-TTS
+
+2-stage TTS (AR + DiT flow-matching) at 24 kHz. Every request requires `ref_audio` + `ref_text`.
+
+### Launch
+```bash
+vllm serve zai-org/GLM-TTS --omni --trust-remote-code --port 8091
+# or:
+bash examples/online_serving/text_to_speech/glm_tts/run_server.sh /path/to/GLM-TTS
+```
+
+### Sending requests
+```bash
+# Voice cloning (required)
+python examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py \
+    --text "你好，这是语音克隆测试。" \
+    --ref-audio file:///path/to/ref.wav \
+    --ref-text "这是参考音频的文本内容。"
+
+# Custom format
+python examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py \
+    --text "Hello, this is a voice cloning test." \
+    --ref-audio file:///path/to/ref.wav \
+    --ref-text "Transcript of the reference audio." \
+    --response-format mp3 -o output.mp3
+```
+
+### Gradio demo
+```bash
+bash examples/online_serving/text_to_speech/glm_tts/run_gradio_demo.sh
+```
+
+### Notes
+- Output: 24 kHz mono WAV via HiFT vocoder.
+- `ref_audio` + `ref_text` are **required** together on every request. Reference audio should be 3-10 seconds.
+- Voice cloning feature extraction (WhisperVQ, CampPlus, mel) runs on the model side — no external dependency on the serving layer.
+
+---
+
 ## Fish Speech S2 Pro
 
 4B dual-AR TTS at 44.1 kHz. Server uses the DAC codec.
@@ -294,10 +333,10 @@ vllm serve Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --omni --port 8091
 ./qwen3_tts/run_server.sh Base
 ```
 
-### Choosing an executor backend (uniproc vs mp)
-Stage configs ship with the chunked-streaming default. To use the uniproc executor (lower IPC overhead for the Base cloning task), pass `--stage-configs-path vllm_omni/model_executor/stage_configs/qwen3_tts_uniproc.yaml`. See [#2603](https://github.com/vllm-project/vllm-omni/issues/2603) and [#2604](https://github.com/vllm-project/vllm-omni/pull/2604) for the full investigation.
+### Executor backend
+Single-GPU serves now default to the uniproc executor (lower IPC overhead, the Base cloning use case from [#2603](https://github.com/vllm-project/vllm-omni/issues/2603) / [#2604](https://github.com/vllm-project/vllm-omni/pull/2604)). `vllm_omni/deploy/qwen3_tts.yaml` is the only Qwen3-TTS deploy config; pass `--deploy-config <path>` to override.
 
-To opt out of chunked streaming, pass `--no-async-chunk` instead — the pipeline auto-dispatches to the end-to-end codec processor.
+To opt out of chunked streaming, pass `--no-async-chunk` — the pipeline auto-dispatches to the end-to-end codec processor.
 
 ### Sending requests
 ```bash
@@ -375,73 +414,11 @@ python qwen3_tts/streaming_speech_client.py --text "..." --simulate-stt --stt-de
 `qwen3_tts/batch_speech_client.py` issues many concurrent requests for throughput measurement.
 
 ### Notes
-- Base voice cloning has uniproc-vs-mp tradeoffs depending on per-request reference audio cost; see the executor-backend section above.
+- Base voice cloning has per-request reference-audio cost; the uniproc default keeps IPC overhead off the critical path. See the executor-backend section above for background.
 - `vllm_omni/deploy/qwen3_tts.yaml` is the default deploy config (loaded by HF `model_type`); per-stage runtime overrides are available via `--stage-N-<field> <value>`.
 
 ---
 
-## VoxCPM
-
-Split-stage TTS at 24 kHz.
-
-### Prerequisites
-```bash
-pip install voxcpm
-# or use a local source tree:
-export VLLM_OMNI_VOXCPM_CODE_PATH=/path/to/VoxCPM/src
-```
-
-If the native VoxCPM `config.json` lacks HF `model_type`, set up an HF-compatible config dir:
-```bash
-export VOXCPM_MODEL=/path/to/voxcpm-model
-export VLLM_OMNI_VOXCPM_HF_CONFIG_PATH=/tmp/voxcpm_hf_config
-mkdir -p "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"
-cp "$VOXCPM_MODEL/config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/config.json"
-cp "$VOXCPM_MODEL/generation_config.json" "$VLLM_OMNI_VOXCPM_HF_CONFIG_PATH/generation_config.json" 2>/dev/null || true
-python3 -c 'import json, os; p=os.path.join(os.environ["VLLM_OMNI_VOXCPM_HF_CONFIG_PATH"], "config.json"); cfg=json.load(open(p, "r", encoding="utf-8")); cfg["model_type"]="voxcpm"; cfg.setdefault("architectures", ["VoxCPMForConditionalGeneration"]); json.dump(cfg, open(p, "w", encoding="utf-8"), indent=2, ensure_ascii=False)'
-```
-
-### Launch
-```bash
-export VOXCPM_MODEL=/path/to/voxcpm-model
-./voxcpm/run_server.sh                # async-chunk streaming (default)
-./voxcpm/run_server.sh sync           # non-streaming
-```
-Or directly:
-```bash
-vllm serve "$VOXCPM_MODEL" \
-    --stage-configs-path vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml \
-    --trust-remote-code --enforce-eager --omni --port 8091
-```
-
-### Sending requests
-```bash
-# Basic TTS
-python voxcpm/openai_speech_client.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This is a VoxCPM online text-to-speech example."
-
-# Voice cloning
-python voxcpm/openai_speech_client.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This sentence is synthesized with a cloned voice." \
-    --ref-audio /path/to/reference.wav \
-    --ref-text "The exact transcript spoken in reference.wav."
-
-# Streaming PCM
-python voxcpm/openai_speech_client.py \
-    --model "$VOXCPM_MODEL" \
-    --text "This is a streaming VoxCPM request." \
-    --stream --output voxcpm_stream.pcm
-```
-
-### Notes
-- `voxcpm.yaml` for one-shot decode; `voxcpm_async_chunk.yaml` for single-request streaming. Do not use the async-chunk config for concurrent requests or `/v1/audio/speech/batch`.
-- Generic TTS fields not supported by VoxCPM: `voice`, `instructions`, `language`, `speaker_embedding`, `x_vector_only_mode`.
-- For benchmark measurement, see [`benchmarks/voxcpm`](../../../benchmarks/voxcpm/README.md).
-
----
-
 ## VoxCPM2
 
 Single-stage native AR TTS at 48 kHz.

examples/online_serving/text_to_speech/cosyvoice3/run_server.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# Launch vLLM-Omni server for CosyVoice3 TTS
+#
+# Usage:
+#   ./run_server.sh
+#   CUDA_VISIBLE_DEVICES=0 ./run_server.sh
+#
+# Streaming (async-chunk) is on by default via vllm_omni/deploy/cosyvoice3.yaml.
+# Set NO_ASYNC_CHUNK=1 to use the legacy synchronous path.
+
+set -e
+
+MODEL="${MODEL:-FunAudioLLM/Fun-CosyVoice3-0.5B-2512}"
+PORT="${PORT:-8091}"
+
+EXTRA_ARGS=()
+if [[ -n "${NO_ASYNC_CHUNK:-}" ]]; then
+    EXTRA_ARGS+=(--no-async-chunk)
+fi
+
+echo "Starting CosyVoice3 server with model: $MODEL"
+
+vllm serve "$MODEL" \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --trust-remote-code \
+    --omni \
+    "${EXTRA_ARGS[@]}"

examples/online_serving/text_to_speech/cosyvoice3/speech_client.py

@@ -0,0 +1,142 @@
+"""Client for CosyVoice3 TTS via /v1/audio/speech endpoint.
+
+CosyVoice3 has no built-in voice presets: every request is voice cloning
+driven by ``ref_audio`` + ``ref_text``. The defaults below point at the
+official upstream zero-shot prompt so the script runs out of the box.
+
+Examples:
+    # Voice cloning with the default upstream prompt
+    python speech_client.py --text "收到好友从远方寄来的生日礼物。"
+
+    # Custom reference clip + transcript
+    python speech_client.py --text "Hello, this is a cloned voice." \
+        --ref-audio /path/to/reference.wav \
+        --ref-text "Transcript of the reference audio."
+
+    # Streaming PCM output
+    python speech_client.py --text "Hello world" --stream --output output.pcm
+"""
+
+import argparse
+import base64
+import os
+
+import httpx
+
+DEFAULT_API_BASE = "http://localhost:8091"
+DEFAULT_API_KEY = "EMPTY"
+DEFAULT_MODEL = "FunAudioLLM/Fun-CosyVoice3-0.5B-2512"
+
+# Official CosyVoice zero-shot prompt and its transcript.
+DEFAULT_REF_AUDIO = "https://raw.githubusercontent.com/FunAudioLLM/CosyVoice/main/asset/zero_shot_prompt.wav"
+DEFAULT_REF_TEXT = "希望你以后能够做的比我还好呦。"
+
+
+def encode_audio_to_base64(audio_path: str) -> str:
+    """Encode a local audio file to a base64 data URL."""
+    if not os.path.exists(audio_path):
+        raise FileNotFoundError(f"Audio file not found: {audio_path}")
+    ext = audio_path.lower().rsplit(".", 1)[-1]
+    mime_map = {"wav": "audio/wav", "mp3": "audio/mpeg", "flac": "audio/flac", "ogg": "audio/ogg"}
+    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}"
+
+
+def run_tts(args) -> None:
+    """Generate speech via the /v1/audio/speech API."""
+    payload = {
+        "model": args.model,
+        "input": args.text,
+        "response_format": args.response_format,
+    }
+
+    if args.ref_audio.startswith(("http://", "https://")):
+        payload["ref_audio"] = args.ref_audio
+    else:
+        payload["ref_audio"] = encode_audio_to_base64(args.ref_audio)
+    payload["ref_text"] = args.ref_text
+
+    if args.stream:
+        payload["stream"] = True
+        payload["response_format"] = "pcm"
+
+    print(f"Model: {args.model}")
+    print(f"Text: {args.text}")
+    print(f"Voice cloning: ref_audio={args.ref_audio}, ref_text={args.ref_text}")
+    print("Generating audio...")
+
+    api_url = f"{args.api_base}/v1/audio/speech"
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": f"Bearer {args.api_key}",
+    }
+
+    if args.stream:
+        output_path = args.output or "output.pcm"
+        with httpx.Client(timeout=300.0) as client:
+            with client.stream("POST", api_url, json=payload, headers=headers) as resp:
+                if resp.status_code != 200:
+                    print(f"Error: {resp.status_code}")
+                    print(resp.read().decode())
+                    return
+                total_bytes = 0
+                with open(output_path, "wb") as f:
+                    for chunk in resp.iter_bytes():
+                        f.write(chunk)
+                        total_bytes += len(chunk)
+                print(f"Streamed {total_bytes} bytes to: {output_path}")
+    else:
+        with httpx.Client(timeout=300.0) as client:
+            response = client.post(api_url, json=payload, headers=headers)
+
+        if response.status_code != 200:
+            print(f"Error: {response.status_code}")
+            print(response.text)
+            return
+
+        try:
+            text = response.content.decode("utf-8")
+            if text.startswith('{"error"'):
+                print(f"Error: {text}")
+                return
+        except UnicodeDecodeError:
+            pass
+
+        output_path = args.output or "output.wav"
+        with open(output_path, "wb") as f:
+            f.write(response.content)
+        print(f"Audio saved to: {output_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="CosyVoice3 TTS client")
+    parser.add_argument("--api-base", default=DEFAULT_API_BASE, help="API base URL")
+    parser.add_argument("--api-key", default=DEFAULT_API_KEY, help="API key")
+    parser.add_argument("--model", "-m", default=DEFAULT_MODEL, help="Model name")
+    parser.add_argument("--text", required=True, help="Text to synthesize")
+    parser.add_argument(
+        "--ref-audio",
+        default=DEFAULT_REF_AUDIO,
+        help="Reference audio for voice cloning (path or URL)",
+    )
+    parser.add_argument(
+        "--ref-text",
+        default=DEFAULT_REF_TEXT,
+        help="Transcript of the reference audio",
+    )
+    parser.add_argument("--stream", action="store_true", help="Enable streaming (PCM output)")
+    parser.add_argument(
+        "--response-format",
+        default="wav",
+        choices=["wav", "mp3", "flac", "pcm", "aac", "opus"],
+        help="Audio format (default: wav)",
+    )
+    parser.add_argument("--output", "-o", default=None, help="Output file path")
+    args = parser.parse_args()
+    run_tts(args)
+
+
+if __name__ == "__main__":
+    main()

examples/online_serving/text_to_speech/glm_tts/gradio_demo.py

@@ -0,0 +1,291 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Gradio demo for GLM-TTS online serving via /v1/audio/speech API.
+
+Supports:
+  - Voice-cloned text-to-speech synthesis from reference audio (upload or URL)
+  - Streaming (progressive PCM chunks) and non-streaming modes
+
+Usage:
+    # Start the server first (see run_server.sh), then:
+    python gradio_demo.py --api-base http://localhost:8091
+
+    # Or use run_gradio_demo.sh to start both server and demo together.
+"""
+
+import argparse
+import base64
+import io
+
+try:
+    import gradio as gr
+except ImportError:
+    raise ImportError("gradio is required to run this demo. Install it with: pip install 'vllm-omni[demo]'") from None
+import httpx
+import numpy as np
+import soundfile as sf
+
+PCM_SAMPLE_RATE = 24000
+DEFAULT_API_BASE = "http://localhost:8091"
+
+
+def encode_audio_to_base64(audio_data: tuple) -> str:
+    """Encode Gradio audio input (sample_rate, numpy_array) to base64 data URL."""
+    sample_rate, audio_np = audio_data
+    if audio_np.dtype != np.int16:
+        if audio_np.dtype in (np.float32, np.float64):
+            audio_np = np.clip(audio_np, -1.0, 1.0)
+            audio_np = (audio_np * 32767).astype(np.int16)
+        else:
+            audio_np = audio_np.astype(np.int16)
+    buf = io.BytesIO()
+    sf.write(buf, audio_np, sample_rate, format="WAV")
+    wav_b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
+    return f"data:audio/wav;base64,{wav_b64}"
+
+
+def build_payload(
+    text: str,
+    ref_audio: tuple | None,
+    ref_audio_url: str,
+    ref_text: str,
+    response_format: str = "wav",
+    stream: bool = False,
+    max_new_tokens: int = 4096,
+) -> dict:
+    """Build the /v1/audio/speech request payload."""
+    if not text or not text.strip():
+        raise gr.Error("Please enter text to synthesize.")
+
+    payload: dict = {
+        "input": text.strip(),
+        "voice": "default",
+        "response_format": "pcm" if stream else response_format,
+        "stream": stream,
+        "max_new_tokens": max_new_tokens,
+    }
+
+    # Voice cloning: ref_audio takes priority over URL.
+    ref_url_stripped = ref_audio_url.strip() if ref_audio_url else ""
+    if ref_audio is not None:
+        payload["ref_audio"] = encode_audio_to_base64(ref_audio)
+    elif ref_url_stripped:
+        payload["ref_audio"] = ref_url_stripped
+    else:
+        raise gr.Error("GLM-TTS requires a reference audio file or URL.")
+
+    if not ref_text or not ref_text.strip():
+        raise gr.Error("GLM-TTS requires a transcript of the reference audio.")
+    payload["ref_text"] = ref_text.strip()
+
+    return payload
+
+
+def generate_speech(api_base: str, text: str, ref_audio, ref_audio_url, ref_text, response_format, max_new_tokens):
+    """Non-streaming: call /v1/audio/speech and return full audio."""
+    payload = build_payload(
+        text, ref_audio, ref_audio_url, ref_text, response_format, stream=False, max_new_tokens=max_new_tokens
+    )
+
+    try:
+        with httpx.Client(timeout=300.0) as client:
+            resp = client.post(
+                f"{api_base}/v1/audio/speech",
+                json=payload,
+                headers={"Content-Type": "application/json", "Authorization": "Bearer EMPTY"},
+            )
+    except httpx.TimeoutException:
+        raise gr.Error("Request timed out. The server may be busy.")
+    except httpx.ConnectError:
+        raise gr.Error(f"Cannot connect to server at {api_base}. Is the server running?")
+
+    if resp.status_code != 200:
+        raise gr.Error(f"Server error ({resp.status_code}): {resp.text}")
+
+    content_type = resp.headers.get("content-type", "")
+    if "application/json" in content_type:
+        try:
+            raise gr.Error(f"Server error: {resp.json()}")
+        except ValueError:
+            pass
+
+    try:
+        if response_format == "pcm":
+            audio_np = np.frombuffer(resp.content, dtype=np.int16).astype(np.float32) / 32767.0
+            return (PCM_SAMPLE_RATE, audio_np)
+        audio_np, sample_rate = sf.read(io.BytesIO(resp.content))
+        if audio_np.ndim > 1:
+            audio_np = audio_np[:, 0]
+        return (sample_rate, audio_np.astype(np.float32))
+    except Exception as e:
+        raise gr.Error(f"Failed to decode audio: {e}")
+
+
+def generate_speech_stream(
+    api_base, text, ref_audio, ref_audio_url, ref_text, response_format, max_new_tokens, chunk_seconds=0.5
+):
+    """Streaming: yield PCM chunks for progressive playback."""
+    payload = build_payload(
+        text, ref_audio, ref_audio_url, ref_text, response_format, stream=True, max_new_tokens=max_new_tokens
+    )
+
+    min_chunk_samples = int(PCM_SAMPLE_RATE * chunk_seconds)
+    prebuffer_chunks = 2
+    pending: list[np.ndarray] = []
+    pending_len = 0
+    chunks_yielded = 0
+    prebuffer: list[np.ndarray] = []
+
+    try:
+        leftover = b""
+        with httpx.Client(timeout=300.0) as client:
+            with client.stream(
+                "POST",
+                f"{api_base}/v1/audio/speech",
+                json=payload,
+                headers={"Content-Type": "application/json", "Authorization": "Bearer EMPTY"},
+            ) as resp:
+                if resp.status_code != 200:
+                    resp.read()
+                    raise gr.Error(f"Server error ({resp.status_code}): {resp.text}")
+                for chunk in resp.iter_bytes():
+                    if not chunk:
+                        continue
+                    raw = leftover + chunk
+                    usable = len(raw) - (len(raw) % 2)
+                    leftover = raw[usable:]
+                    if usable == 0:
+                        continue
+                    float_samples = np.frombuffer(raw[:usable], dtype=np.int16).copy().astype(np.float32) / 32767.0
+                    pending.append(float_samples)
+                    pending_len += len(float_samples)
+                    if pending_len >= min_chunk_samples:
+                        audio_chunk = np.concatenate(pending)
+                        pending.clear()
+                        pending_len = 0
+                        if chunks_yielded < prebuffer_chunks:
+                            prebuffer.append(audio_chunk)
+                            chunks_yielded += 1
+                            if chunks_yielded == prebuffer_chunks:
+                                yield (PCM_SAMPLE_RATE, np.concatenate(prebuffer))
+                                prebuffer.clear()
+                        else:
+                            yield (PCM_SAMPLE_RATE, audio_chunk)
+        remaining = prebuffer + pending
+        if remaining:
+            yield (PCM_SAMPLE_RATE, np.concatenate(remaining))
+    except httpx.TimeoutException:
+        raise gr.Error("Request timed out. The server may be busy.")
+    except httpx.ConnectError:
+        raise gr.Error(f"Cannot connect to server at {api_base}. Is the server running?")
+
+
+def on_stream_change(stream: bool):
+    """Lock format to PCM when streaming."""
+    if stream:
+        return gr.update(value="pcm", interactive=False)
+    return gr.update(value="wav", interactive=True)
+
+
+def build_interface(api_base: str, stream_chunk_seconds: float = 0.5):
+    """Build the Gradio interface."""
+    with gr.Blocks(title="GLM-TTS Demo") as demo:
+        gr.Markdown("# GLM-TTS - Voice Cloning")
+        gr.Markdown(f"**Server:** `{api_base}` | **Model:** GLM-TTS | **Output:** 24kHz")
+
+        with gr.Row():
+            with gr.Column(scale=3):
+                text_input = gr.Textbox(
+                    label="Text to Synthesize",
+                    placeholder="Enter text here (supports Chinese and English)...",
+                    lines=4,
+                )
+
+                with gr.Accordion("Reference Voice (required)", open=True):
+                    gr.Markdown("Upload or link a short reference audio (3-10s) and provide its transcript.")
+                    ref_audio = gr.Audio(
+                        label="Reference Audio",
+                        type="numpy",
+                        sources=["upload", "microphone"],
+                    )
+                    ref_audio_url = gr.Textbox(
+                        label="Reference Audio URL (alternative to upload)",
+                        placeholder="https://example.com/reference.wav",
+                        lines=1,
+                    )
+                    ref_text = gr.Textbox(
+                        label="Reference Audio Transcript (required for cloning)",
+                        placeholder="Exact transcript of the reference audio...",
+                        lines=2,
+                    )
+
+                with gr.Row():
+                    response_format = gr.Dropdown(
+                        choices=["wav", "mp3", "flac", "pcm"],
+                        value="wav",
+                        label="Audio Format",
+                        scale=1,
+                    )
+                    stream_checkbox = gr.Checkbox(
+                        label="Stream output",
+                        value=False,
+                        info="Progressive PCM streaming",
+                        scale=1,
+                    )
+                    max_new_tokens = gr.Slider(
+                        minimum=256,
+                        maximum=8192,
+                        value=4096,
+                        step=256,
+                        label="Max New Tokens",
+                        scale=1,
+                    )
+
+                generate_btn = gr.Button("Generate Speech", variant="primary", size="lg")
+
+            with gr.Column(scale=2):
+                audio_output = gr.Audio(
+                    label="Generated Audio",
+                    interactive=False,
+                    streaming=True,
+                    autoplay=True,
+                )
+                gr.Markdown(
+                    "### About\n"
+                    "- **GLM-TTS** by Zhipu AI: Llama-based AR + Flow Matching DiT\n"
+                    "- **Voice cloning**: upload 3-10s reference + transcript\n"
+                    "- **Streaming**: real-time PCM output\n"
+                    "- **24kHz** output via Vocos vocoder"
+                )
+
+        stream_checkbox.change(fn=on_stream_change, inputs=[stream_checkbox], outputs=[response_format])
+
+        all_inputs = [text_input, ref_audio, ref_audio_url, ref_text, response_format, max_new_tokens]
+
+        def dispatch(stream_enabled, *args):
+            if stream_enabled:
+                yield from generate_speech_stream(api_base, *args, chunk_seconds=stream_chunk_seconds)
+            else:
+                yield generate_speech(api_base, *args)
+
+        generate_btn.click(fn=dispatch, inputs=[stream_checkbox] + all_inputs, outputs=[audio_output])
+        demo.queue()
+    return demo
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Gradio demo for GLM-TTS")
+    parser.add_argument("--api-base", default=DEFAULT_API_BASE, help=f"API base URL (default: {DEFAULT_API_BASE})")
+    parser.add_argument("--host", default="0.0.0.0", help="Gradio host (default: 0.0.0.0)")
+    parser.add_argument("--port", type=int, default=7860, help="Gradio port (default: 7860)")
+    parser.add_argument("--share", action="store_true", help="Share publicly")
+    parser.add_argument("--stream-chunk-seconds", type=float, default=0.5, help="Seconds per streaming chunk")
+    args = parser.parse_args()
+
+    print(f"Connecting to vLLM server at: {args.api_base}")
+    demo = build_interface(args.api_base, stream_chunk_seconds=args.stream_chunk_seconds)
+    demo.launch(server_name=args.host, server_port=args.port, share=args.share)
+
+
+if __name__ == "__main__":
+    main()

examples/online_serving/text_to_speech/glm_tts/openai_speech_client.py

@@ -0,0 +1,173 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""OpenAI-compatible client for GLM-TTS via /v1/audio/speech endpoint.
+
+GLM-TTS is a two-stage TTS system (AR + DiT) that generates audio from text
+conditioned on reference speech. Each request requires ref_audio + ref_text.
+
+Usage:
+    # Voice cloning
+    python openai_speech_client.py --text "你好" --ref-audio file:///path/to/ref.wav --ref-text "参考文本"
+
+    # Streaming response, for async_chunk server mode
+    python openai_speech_client.py --text "你好" --stream --ref-audio file:///path/to/ref.wav --ref-text "参考文本"
+
+    # Specify output format
+    python openai_speech_client.py --text "你好" --ref-audio file:///path/to/ref.wav \
+        --ref-text "参考文本" --response-format mp3 -o output.mp3
+"""
+
+import argparse
+
+import httpx
+
+# Default server configuration
+DEFAULT_API_BASE = "http://localhost:8091"
+DEFAULT_API_KEY = "EMPTY"
+
+
+def run_tts_generation(args) -> None:
+    """Run TTS generation via OpenAI-compatible /v1/audio/speech API."""
+    if not args.ref_audio or not args.ref_text:
+        raise ValueError("GLM-TTS requires --ref-audio and --ref-text for voice cloning.")
+
+    payload = {
+        "model": args.model,
+        "voice": "default",
+        "input": args.text,
+        "response_format": args.response_format,
+        "stream": bool(args.stream),
+        "ref_audio": args.ref_audio,
+        "ref_text": args.ref_text,
+    }
+    if args.max_new_tokens:
+        payload["max_new_tokens"] = args.max_new_tokens
+
+    print(f"Model: {args.model}")
+    print(f"Text: {args.text}")
+    print(f"Voice cloning: ref_audio={args.ref_audio}, ref_text={args.ref_text}")
+    print(f"Stream: {args.stream}")
+    print("Generating audio...")
+
+    api_url = f"{args.api_base}/v1/audio/speech"
+    headers = {
+        "Content-Type": "application/json",
+        "Authorization": f"Bearer {args.api_key}",
+    }
+
+    if args.stream:
+        output_path = args.output or "tts_output.pcm"
+        with httpx.Client(timeout=300.0) as client, open(output_path, "wb") as f:
+            with client.stream("POST", api_url, json=payload, headers=headers) as response:
+                if response.status_code != 200:
+                    print(f"Error: {response.status_code}")
+                    response.read()
+                    print(response.text)
+                    return
+                for chunk in response.iter_bytes():
+                    f.write(chunk)
+        print(f"Streaming audio saved to: {output_path}")
+    else:
+        with httpx.Client(timeout=300.0) as client:
+            response = client.post(api_url, json=payload, headers=headers)
+        if response.status_code != 200:
+            print(f"Error: {response.status_code}")
+            print(response.text)
+            return
+        try:
+            text = response.content.decode("utf-8")
+            if text.startswith('{"error"'):
+                print(f"Error: {text}")
+                return
+        except UnicodeDecodeError:
+            pass
+        output_path = args.output or f"tts_output.{args.response_format}"
+        with open(output_path, "wb") as f:
+            f.write(response.content)
+        print(f"Audio saved to: {output_path}")
+
+
+def parse_args():
+    """Parse command line arguments."""
+    parser = argparse.ArgumentParser(
+        description="OpenAI-compatible client for GLM-TTS via /v1/audio/speech",
+    )
+
+    # Server configuration
+    parser.add_argument(
+        "--api-base",
+        type=str,
+        default=DEFAULT_API_BASE,
+        help=f"API base URL (default: {DEFAULT_API_BASE})",
+    )
+    parser.add_argument(
+        "--api-key",
+        type=str,
+        default=DEFAULT_API_KEY,
+        help="API key (default: EMPTY)",
+    )
+    parser.add_argument(
+        "--model",
+        "-m",
+        type=str,
+        default="glm-tts",
+        help="Model name/path",
+    )
+
+    # Input text
+    parser.add_argument(
+        "--text",
+        type=str,
+        required=True,
+        help="Text to synthesize",
+    )
+
+    # Generation parameters
+    parser.add_argument(
+        "--max-new-tokens",
+        type=int,
+        default=None,
+        help="Maximum new tokens to generate (default: model default)",
+    )
+
+    # Output
+    parser.add_argument(
+        "--response-format",
+        type=str,
+        default="wav",
+        choices=["wav", "mp3", "flac", "pcm", "aac", "opus"],
+        help="Audio output format (default: wav)",
+    )
+    parser.add_argument(
+        "--stream",
+        action="store_true",
+        help="Request a streaming audio response (use with async_chunk server mode).",
+    )
+    parser.add_argument(
+        "--output",
+        "-o",
+        type=str,
+        default=None,
+        help="Output audio file path (default: tts_output.<format>)",
+    )
+
+    # Voice cloning parameters
+    parser.add_argument(
+        "--ref-audio",
+        type=str,
+        default=None,
+        help="Reference audio URL, file:// URI, or base64 data URL for voice cloning",
+    )
+    parser.add_argument(
+        "--ref-text",
+        type=str,
+        default=None,
+        help="Transcript of the reference audio (required with --ref-audio)",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    run_tts_generation(args)

examples/online_serving/text_to_speech/glm_tts/run_gradio_demo.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Launch GLM-TTS server + Gradio demo together.
+#
+# Usage:
+#   ./run_gradio_demo.sh
+#   CUDA_VISIBLE_DEVICES=0 PORT=8091 GRADIO_PORT=7860 ./run_gradio_demo.sh
+
+set -e
+
+MODEL="${MODEL:-zai-org/GLM-TTS}"
+PORT="${PORT:-8091}"
+GRADIO_PORT="${GRADIO_PORT:-7860}"
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="$(cd "$SCRIPT_DIR/../../../.." && pwd)"
+
+echo "Starting GLM-TTS server (port $PORT)..."
+FLASHINFER_DISABLE_VERSION_CHECK=1 \
+vllm-omni serve "$MODEL" \
+    --deploy-config "$REPO_ROOT/vllm_omni/deploy/glm_tts.yaml" \
+    --host 0.0.0.0 \
+    --port "$PORT" \
+    --gpu-memory-utilization 0.9 \
+    --trust-remote-code \
+    --enforce-eager \
+    --omni &
+SERVER_PID=$!
+
+cleanup() {
+    echo "Stopping server (PID $SERVER_PID)..."
+    kill $SERVER_PID 2>/dev/null
+    wait $SERVER_PID 2>/dev/null
+}
+trap cleanup EXIT
+
+# Wait for server to be ready.
+echo "Waiting for server to start..."
+for i in $(seq 1 120); do
+    if curl -s "http://localhost:$PORT/health" > /dev/null 2>&1; then
+        echo "Server ready."
+        break
+    fi
+    sleep 2
+done
+
+echo "Starting Gradio demo (port $GRADIO_PORT)..."
+python "$SCRIPT_DIR/gradio_demo.py" \
+    --api-base "http://localhost:$PORT" \
+    --port "$GRADIO_PORT"

examples/online_serving/text_to_speech/glm_tts/run_server.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+# Launch vLLM-Omni server for GLM-TTS models
+#
+# Usage:
+#   ./run_server.sh                           # Default model path, async_chunk mode
+#   ./run_server.sh /path/to/GLM-TTS          # Custom model path, async_chunk mode
+#   ./run_server.sh /path/to/GLM-TTS sync     # Sync two-stage mode
+#
+# NOTE: The model path should point to the repo ROOT (not llm/ subdirectory).
+# model_subdir/tokenizer_subdir in the pipeline config resolve subdirectories.
+
+set -e
+
+MODEL="${1:-zai-org/GLM-TTS}"
+MODE="${2:-async}"
+
+EXTRA_ARGS=()
+case "$MODE" in
+    async|async_chunk)
+        ;;
+    sync|no_async_chunk)
+        EXTRA_ARGS+=("--no-async-chunk")
+        ;;
+    *)
+        echo "Unknown mode: $MODE (expected async or sync)" >&2
+        exit 1
+        ;;
+esac
+
+echo "Starting GLM-TTS server with model: $MODEL (mode: $MODE)"
+
+vllm-omni serve "$MODEL" \
+    --deploy-config vllm_omni/deploy/glm_tts.yaml \
+    --host 0.0.0.0 \
+    --port 8091 \
+    --trust-remote-code \
+    --omni \
+    "${EXTRA_ARGS[@]}"

examples/online_serving/text_to_speech/omnivoice/speech_client.py

@@ -6,6 +6,9 @@ Examples:
 
     # Specify language
     python speech_client.py --text "Bonjour, comment allez-vous?" --language French
+
+    # Use a specific uploaded/supported voice
+    python speech_client.py --text "Hello" --voice my_uploaded_voice
 """
 
 import argparse
@@ -21,15 +24,18 @@ def run_tts(args) -> None:
     payload = {
         "model": args.model,
         "input": args.text,
-        "voice": "default",
         "response_format": args.response_format,
     }
 
+    if args.voice:
+        payload["voice"] = args.voice
     if args.language:
         payload["language"] = args.language
 
     print(f"Model: {args.model}")
     print(f"Text: {args.text}")
+    if args.voice:
+        print(f"Voice: {args.voice}")
     if args.language:
         print(f"Language: {args.language}")
     print("Generating audio...")
@@ -68,6 +74,11 @@ def main():
     parser.add_argument("--api-key", default=DEFAULT_API_KEY, help="API key")
     parser.add_argument("--model", "-m", default="k2-fsa/OmniVoice", help="Model name")
     parser.add_argument("--text", required=True, help="Text to synthesize")
+    parser.add_argument(
+        "--voice",
+        default=None,
+        help="Voice name (omit for auto voice; must match a supported or uploaded speaker if set)",
+    )
     parser.add_argument("--language", default=None, help="Language hint (e.g., English, Chinese, French)")
     parser.add_argument(
         "--response-format",

examples/online_serving/text_to_speech/qwen3_tts/streaming_speech_client.py

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

examples/online_serving/text_to_speech/voxcpm/openai_speech_client.py

@@ -1,155 +0,0 @@
-"""OpenAI-compatible client for VoxCPM via /v1/audio/speech.
-
-Examples:
-    # Basic text-to-speech
-    python openai_speech_client.py --text "Hello from VoxCPM"
-
-    # Voice cloning
-    python openai_speech_client.py \
-        --text "This sentence uses the cloned voice." \
-        --ref-audio /path/to/reference.wav \
-        --ref-text "The exact transcript spoken in the reference audio."
-
-    # Streaming PCM output
-    python openai_speech_client.py \
-        --text "This is a streaming VoxCPM request." \
-        --stream \
-        --output output.pcm
-"""
-
-import argparse
-import base64
-import os
-
-import httpx
-
-DEFAULT_API_BASE = "http://localhost:8091"
-DEFAULT_API_KEY = "EMPTY"
-DEFAULT_MODEL = "OpenBMB/VoxCPM1.5"
-
-
-def encode_audio_to_base64(audio_path: str) -> str:
-    """Encode a local audio file to base64 data URL."""
-    if not os.path.exists(audio_path):
-        raise FileNotFoundError(f"Audio file not found: {audio_path}")
-
-    ext = audio_path.lower().rsplit(".", 1)[-1]
-    mime_map = {
-        "wav": "audio/wav",
-        "mp3": "audio/mpeg",
-        "flac": "audio/flac",
-        "ogg": "audio/ogg",
-    }
-    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}"
-
-
-def build_payload(args) -> dict[str, object]:
-    payload: dict[str, object] = {
-        "model": args.model,
-        "input": args.text,
-        "response_format": "pcm" if args.stream else args.response_format,
-    }
-
-    if args.ref_audio:
-        if args.ref_audio.startswith(("http://", "https://", "data:")):
-            payload["ref_audio"] = args.ref_audio
-        else:
-            payload["ref_audio"] = encode_audio_to_base64(args.ref_audio)
-    if args.ref_text:
-        payload["ref_text"] = args.ref_text
-    if args.max_new_tokens is not None:
-        payload["max_new_tokens"] = args.max_new_tokens
-    if args.stream:
-        payload["stream"] = True
-
-    return payload
-
-
-def run_tts(args) -> None:
-    payload = build_payload(args)
-    api_url = f"{args.api_base}/v1/audio/speech"
-    headers = {
-        "Content-Type": "application/json",
-        "Authorization": f"Bearer {args.api_key}",
-    }
-
-    print(f"Model: {args.model}")
-    print(f"Text: {args.text}")
-    if args.ref_audio:
-        print("Mode: voice cloning")
-        print(f"Reference audio: {args.ref_audio}")
-    else:
-        print("Mode: text-to-speech")
-
-    if args.stream:
-        output_path = args.output or "voxcpm_output.pcm"
-        with httpx.Client(timeout=300.0) as client:
-            with client.stream("POST", api_url, json=payload, headers=headers) as response:
-                if response.status_code != 200:
-                    print(f"Error: {response.status_code}")
-                    print(response.read().decode("utf-8", errors="ignore"))
-                    return
-
-                total_bytes = 0
-                with open(output_path, "wb") as f:
-                    for chunk in response.iter_bytes():
-                        if not chunk:
-                            continue
-                        f.write(chunk)
-                        total_bytes += len(chunk)
-        print(f"Streamed {total_bytes} bytes to: {output_path}")
-        return
-
-    with httpx.Client(timeout=300.0) as client:
-        response = client.post(api_url, json=payload, headers=headers)
-
-    if response.status_code != 200:
-        print(f"Error: {response.status_code}")
-        print(response.text)
-        return
-
-    try:
-        text = response.content.decode("utf-8")
-        if text.startswith('{"error"'):
-            print(f"Error: {text}")
-            return
-    except UnicodeDecodeError:
-        pass
-
-    output_path = args.output or "voxcpm_output.wav"
-    with open(output_path, "wb") as f:
-        f.write(response.content)
-    print(f"Audio saved to: {output_path}")
-
-
-def main():
-    parser = argparse.ArgumentParser(description="VoxCPM OpenAI-compatible speech client")
-    parser.add_argument("--api-base", default=DEFAULT_API_BASE, help="API base URL")
-    parser.add_argument("--api-key", default=DEFAULT_API_KEY, help="API key")
-    parser.add_argument("--model", "-m", default=DEFAULT_MODEL, help="Model name or path")
-    parser.add_argument("--text", required=True, help="Text to synthesize")
-    parser.add_argument("--ref-audio", default=None, help="Reference audio path, URL, or data URL")
-    parser.add_argument(
-        "--ref-text",
-        default=None,
-        help="The exact transcript spoken in the reference audio",
-    )
-    parser.add_argument("--stream", action="store_true", help="Enable streaming PCM output")
-    parser.add_argument(
-        "--response-format",
-        default="wav",
-        choices=["wav", "pcm", "flac", "mp3", "aac", "opus"],
-        help="Audio format for non-streaming mode (default: wav)",
-    )
-    parser.add_argument("--max-new-tokens", type=int, default=None, help="Maximum tokens to generate")
-    parser.add_argument("--output", "-o", default=None, help="Output file path")
-    args = parser.parse_args()
-    run_tts(args)
-
-
-if __name__ == "__main__":
-    main()

examples/online_serving/text_to_speech/voxcpm/run_server.sh

@@ -1,38 +0,0 @@
-#!/bin/bash
-# Launch vLLM-Omni server for VoxCPM online speech serving.
-#
-# Usage:
-#   ./run_server.sh                 # default: async_chunk stage config
-#   ./run_server.sh async           # async_chunk stage config
-#   ./run_server.sh sync            # no-async-chunk stage config
-#   VOXCPM_MODEL=/path/to/model ./run_server.sh
-
-set -e
-
-MODE="${1:-async}"
-MODEL="${VOXCPM_MODEL:-OpenBMB/VoxCPM1.5}"
-
-case "$MODE" in
-    async)
-        STAGE_CONFIG="vllm_omni/model_executor/stage_configs/voxcpm_async_chunk.yaml"
-        ;;
-    sync)
-        STAGE_CONFIG="vllm_omni/deploy/voxcpm.yaml"
-        ;;
-    *)
-        echo "Unknown mode: $MODE"
-        echo "Supported: async, sync"
-        exit 1
-        ;;
-esac
-
-echo "Starting VoxCPM server with model: $MODEL"
-echo "Stage config: $STAGE_CONFIG"
-
-vllm serve "$MODEL" \
-    --stage-configs-path "$STAGE_CONFIG" \
-    --host 0.0.0.0 \
-    --port 8091 \
-    --trust-remote-code \
-    --enforce-eager \
-    --omni

mkdocs.yml

@@ -120,7 +120,9 @@ plugins:
           inventories:
           - https://docs.python.org/3/objects.inv
           - https://typing-extensions.readthedocs.io/en/latest/objects.inv
-          - https://docs.aiohttp.org/en/stable/objects.inv
+          # Disabled because Read the Docs strict builds fail if this
+          # external inventory intermittently resets the connection.
+          # - https://docs.aiohttp.org/en/stable/objects.inv
           - https://pillow.readthedocs.io/en/stable/objects.inv
           - https://numpy.org/doc/stable/objects.inv
           # Temporarily disabled due to decompression errors

pyproject.toml

@@ -63,7 +63,7 @@ dev = [
     "funasr>=1.0.0",
     "lpips>=0.1.4", # for quantization fp8
     "orjson>=3.11.7", # to make v1/embeddings API fast
-    "voxcpm>=2.0.2", # for voxcpm model Test
+    "voxcpm>=2.0.2", # for VoxCPM2 path-locating in tests
     "torchdiffeq>=0.2.5" # for covo_audio Test
 ]
 
@@ -71,6 +71,11 @@ demo = [
     "gradio>=6.7.0",
 ]
 
+# Local-test-only extras
+local = [
+    "FlagEmbedding",
+]
+
 docs = [
     "mkdocs>=1.5.0",
     "mkdocs-api-autonav",
@@ -184,6 +189,7 @@ markers = [
     "core_model: L1&L2 tests (run in each PR)",
     "advanced_model: L3 level tests (run on each merge)",
     "full_model: L4 level tests (run nightly)",
+    "local_model: tests (run locally)",
     # function module markers
     "diffusion: Diffusion model tests",
     "omni: Omni model tests",
@@ -235,6 +241,8 @@ extend-ignore-identifiers-re = [
     ".*MoBA",
     ".*temperal_downsample",
     ".*_nd",
+    ".*nd$",
+    ".*EmbedND$",
     ".*nothink.*",
     ".*NOTHINK.*",
     ".*nin.*",
@@ -252,3 +260,4 @@ fullset = "fullset"
 Vai = "Vai"
 tockens = "tockens"
 CANN = "CANN"
+ND = "ND"

recipes/Baidu/ERNIE-Image.md

@@ -1,4 +1,6 @@
-# ERNIE-Image for text-to-image generation
+# ERNIE-Image
+
+> Text-to-image online serving (ERNIE-Image 8B)
 
 ## Summary
 

recipes/HiDream/HiDream-I1-Image.md

@@ -0,0 +1,106 @@
+# HiDream-I1-Full
+
+> Text-to-image serving and benchmark
+
+## Summary
+
+- Vendor: HiDream.ai
+- Model: `HiDream-ai/HiDream-I1-Full`
+- Task: Text-to-image generation
+- Mode: Online serving with the OpenAI-compatible API
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe when you want a known-good starting point for serving
+`HiDream-ai/HiDream-I1-Full` on a single 80 GB A800, validate the normal online-serving path.
+
+## References
+
+- Model: <https://huggingface.co/HiDream-ai/HiDream-I1-Full>
+- Upstream or canonical docs:
+  [`docs/user_guide/examples/online_serving/text_to_image.md`](../../docs/user_guide/examples/online_serving/text_to_image.md)
+- Related example under `examples/`:
+  [`examples/online_serving/text_to_image/README.md`](../../examples/online_serving/text_to_image/README.md)
+
+
+## Hardware Support
+
+This recipe currently documents one CUDA GPU serving configuration.
+
+## GPU
+
+### 1x A800 80GB
+
+#### Environment
+
+- OS: Linux
+- Python: 3.10+
+- Driver / runtime: NVIDIA CUDA environment with an A100 80 GB GPU
+- vLLM version: Match the repository requirements for your checkout
+- vLLM-Omni version or commit: Use the commit you are deploying from
+
+#### Command
+
+Start the baseline server:
+
+```bash
+export MODEL_NAME_OR_PATH=HiDream-ai/HiDream-I1-Full
+vllm serve ${MODEL_NAME_OR_PATH} \
+   --omni \
+   --port 8092 \
+   --auxiliary-text-encoder meta-llama/Llama-3.1-8B-Instruct \
+   --tensor-parallel-size 1 \
+   --vae_use_slicing \
+   --vae_use_tiling
+```
+
+
+You can also use the example launcher and pass the extra flags through:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+  --model HiDream-ai/HiDream-I1-Full \
+  --prompt "The setting sun of late autumn dyes the riverside with a warm orange hue" \
+  --seed 42 \
+  --guidance-scale 5.0 \
+  --tensor-parallel-size 1 \
+  --num-images-per-prompt 1 \
+  --num-inference-steps 50 \
+  --auxiliary-text-encoder meta-llama/Llama-3.1-8B-Instruct \
+  --output /workspace/mnt/cmss-wmd/result/output.png
+
+```
+
+#### Verification
+
+Run the existing client example after the server is ready:
+
+```bash
+python examples/online_serving/text_to_image/openai_chat_client.py \
+  --server http://localhost:8092 \
+  --prompt "The setting sun of late autumn dyes the riverside with a warm orange hue" \
+  --output /tmp/hidream_i1_recipe.png
+```
+
+For a direct API smoke test:
+
+```bash
+curl -s http://localhost:8092/v1/chat/completions   -H "Content-Type: application/json"   -d '{
+    "messages": [
+      {"role": "user", "content": "The setting sun of late autumn dyes the riverside with a warm orange hue"}
+    ],
+    "extra_body": {
+      "num_inference_steps": 50,
+      "seed": 42,
+      "height": 1024,
+      "width": 1024
+    }
+  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2- | base64 -d > output_online.png
+```
+
+
+
+#### Notes
+
+- Key flags: `--auxiliary-text-encoder` designates the path for the auxiliary text model meta-llama/Llama-3.1-8B-Instruct. For HiDream-I1-Full, unspecified use defaults meta-llama/Llama-3.1-8B-Instruct (downloaded from official Hugging Face), and custom paths are supported.

recipes/LTX/LTX-2.3.md

@@ -1,4 +1,4 @@
-# LTX-2.3 Text-to-Video with Audio on 1x GPU (96GB VRAM)
+# LTX-2.3
 
 > 22B parameter text-to-video + audio generation model served via vLLM-Omni
 

recipes/LTX/LTX-2.md

@@ -1,6 +1,6 @@
-# LTX-2 T2V and I2V
+# LTX-2
 
-> LTX-2 for text-to-video & image-to-video on 1x H200 141GB
+> Text-to-video and image-to-video serving
 
 ## Summary
 

recipes/Qwen/Qwen-Image-Edit.md

@@ -0,0 +1,171 @@
+# Qwen-Image-Edit
+
+> Image editing serving
+
+## Summary
+
+- Vendor: Qwen
+- Model: `Qwen/Qwen-Image-Edit`
+- Task: Text-guided single-image editing
+- Mode: Online serving with the OpenAI-compatible multimodal endpoint
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe as a baseline for serving `Qwen/Qwen-Image-Edit` for single-image editing on H200 141GB. Two configurations are validated below, sharing one software environment:
+
+| Configuration | Parallelism | Notes |
+| --- | --- | --- |
+| `1x H200` | none | Single-card baseline; ~60 GiB peak VRAM on a 141 GB card |
+| `2x H200` | TP=2 | Lower single-request latency (≈ 1.60× speedup) and lower per-GPU VRAM |
+
+## References
+
+- Model: <https://huggingface.co/Qwen/Qwen-Image-Edit>
+- Example guide: [`examples/online_serving/image_to_image/README.md`](../../examples/online_serving/image_to_image/README.md)
+- User guide: [`docs/user_guide/examples/online_serving/image_to_image.md`](../../docs/user_guide/examples/online_serving/image_to_image.md)
+
+## Hardware Support
+
+### GPU
+
+The two hardware configurations below share the same software environment.
+
+#### Environment
+
+- OS: Linux
+- Python: 3.12+
+- Driver / runtime: CUDA-capable runtime matching the repository build
+- vLLM: use versions from your current checkout, >=0.20.0
+- vLLM-Omni: use versions from your current checkout
+
+#### 1x H200 141GB
+
+##### Command
+
+```bash
+CUDA_VISIBLE_DEVICES=0 vllm serve Qwen/Qwen-Image-Edit --omni \
+  --port 8092
+```
+
+The bundled launcher is equivalent:
+
+```bash
+bash examples/online_serving/image_to_image/run_server.sh
+```
+
+The server is ready when the log shows `INFO:     Application startup complete.` (preceded by `Started server process [PID]` and `Waiting for application startup.`).
+
+##### Verification
+
+Run a single-image edit against the public example asset:
+
+```bash
+wget https://vllm-public-assets.s3.us-west-2.amazonaws.com/omni-assets/qwen-bear.png
+
+PROMPT="Let this mascot dance under the moon, surrounded by floating stars and poetic bubbles such as 'Be Kind'"
+IMG_B64=$(base64 -w0 qwen-bear.png)
+
+curl -s http://localhost:8092/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d "$(jq -n --arg prompt "$PROMPT" --arg img "$IMG_B64" '{
+    messages: [{
+      role: "user",
+      content: [
+        {type: "text", text: $prompt},
+        {type: "image_url", image_url: {url: ("data:image/png;base64," + $img)}}
+      ]
+    }],
+    extra_body: {
+      height: 1024,
+      width: 1024,
+      num_inference_steps: 50,
+      guidance_scale: 1.0,
+      seed: 42
+    }
+  }')" \
+  | jq -r '.choices[0].message.content[0].image_url.url' \
+  | cut -d',' -f2- | base64 -d > qwen_image_edit_1card.png
+```
+
+The decoded PNG is a 1024×1024 RGB image of the bear mascot dancing under stylized moonlight, surrounded by floating stars and a "Be Kind" text bubble.
+
+You can also run the offline CLI without standing up a server:
+
+```bash
+python examples/offline_inference/image_to_image/image_edit.py \
+  --image qwen-bear.png \
+  --prompt "Let this mascot dance under the moon, ..." \
+  --output qwen_image_edit_offline.png \
+  --num-inference-steps 50 \
+  --cfg-scale 4.0
+```
+
+Validated reference numbers for 1024×1024, 50 denoising steps, batch size 1, `guidance_scale=1.0`, seed 42 (`Cold` = first request after startup, `Warm` = identical subsequent request):
+
+| Cold latency | Warm latency | Peak GPU memory |
+| ---:         | ---:         | ---:            |
+| 17.80 s      | 17.78 s      | 61,091 MiB (~59.66 GiB) |
+
+##### Notes
+
+- Generation parameters live under `extra_body` for `/v1/chat/completions`, or as top-level form fields when using `/v1/images/edits`. See [`examples/online_serving/image_to_image/README.md`](../../examples/online_serving/image_to_image/README.md) for the full request / response schema.
+- Memory: a 1024×1024 / 50-step edit uses ~60 GiB on a single H200, well within the 141 GB headroom.
+- Multi-image variants of Qwen-Image-Edit (e.g. `Qwen-Image-Edit-2509`) are intentionally out of scope for this recipe and are not validated here.
+
+#### 2x H200 141GB (TP=2)
+
+##### Command
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 vllm serve Qwen/Qwen-Image-Edit --omni \
+  --port 8092 \
+  --tensor-parallel-size 2
+```
+
+In addition to the `Application startup complete.` line, the TP=2 server's `non-default args` log line echoes `'tensor_parallel_size': 2`, and a separate `Worker N: Initialized device and distributed environment` line is emitted for each rank — one per visible GPU.
+
+##### Verification
+
+Run the same `curl` command as in the **1x H200** Verification section above against the TP=2 server (same prompt, same seed, same image). Save the result to `qwen_image_edit_2card_tp2.png` so the two outputs sit side by side for comparison. During TP=2 inference, `nvidia-smi` shows sustained `~99 %` utilization on both GPUs, which is the expected signal that TP wiring is active.
+
+Same workload as the 1x table above (1024×1024, 50 steps, `guidance_scale=1.0`, seed 42):
+
+| Configuration | Cold latency | Warm latency | Peak per-GPU memory |
+| ---           | ---:         | ---:         | ---:                |
+| 1x H200       | 17.80 s      | 17.78 s      | 61,091 MiB (~59.66 GiB) |
+| 2x H200 TP=2  | 11.16 s      | 11.09 s      | 49,377 MiB (~48.22 GiB) |
+
+TP=2 reduces single-request warm latency from `17.78 s` to `11.09 s` (≈ **1.60×**) and drops per-GPU peak VRAM by ~11.4 GiB.
+
+##### Notes
+
+- TP=2 shards model weights across both cards, which is what produces both the latency reduction and the per-GPU VRAM drop above.
+- CFG-Parallel (`--cfg-parallel-size 2`) is supported on the Qwen-Image family but only when `cfg_scale > 1`. The verification command above uses `guidance_scale=1.0` to match the public example image, so it does not exercise CFG-parallel; a separate run with `cfg_scale > 1` is needed if you want to benchmark that path.
+- For deeper acceleration knobs (Cache-DiT, sequence parallel, HSDP) see [`docs/user_guide/diffusion/parallelism/overview.md`](../../docs/user_guide/diffusion/parallelism/overview.md); this recipe intentionally documents only the validated TP=2 baseline.
+
+#### Profiling
+
+The numbers below are from a warm request at 1024×1024, 50 steps, and `guidance_scale=1.0`, measured with `--enable-diffusion-pipeline-profiler` and a 250 ms `nvidia-smi --query-gpu=memory.used` sampler. For TP=2, latency uses rank max, and VRAM is per GPU.
+
+Per-stage latency:
+
+| Stage | TP=1 | TP=2 | Notes |
+| --- | ---: | ---: | --- |
+| `text_encoder.forward` | 0.132 s | 0.135 s | ~1.0× |
+| `vae.encode` | 0.040 s | 0.041 s | ~1.0× |
+| `diffuse` | 17.105 s | 10.382 s | **1.65×** speedup |
+| `vae.decode` | 0.064 s | 0.066 s | ~1.0× |
+| `forward` | 17.372 s | 10.652 s | 1.63× speedup |
+
+Per-GPU VRAM:
+
+| Snapshot | TP=1 | TP=2 | Notes |
+| --- | ---: | ---: | --- |
+| Static VRAM | 58,777 MiB | 47,065 MiB | after model load, before request |
+| Peak VRAM | 61,091 MiB | 49,377 MiB | during inference |
+| Request-time increment | 2,314 MiB | 2,312 MiB | peak minus static |
+
+The latency gain mainly comes from the DiT denoising stage (`diffuse` in the table above). The other measured stages, such as text encoder and VAE encode/decode, stay essentially unchanged between TP=1 and TP=2, which is consistent with them not benefiting meaningfully from TP in this pipeline. The `forward` row's 1.63× speedup tracks `diffuse`, since `diffuse` accounts for about 98% of total pipeline time.
+
+For memory, TP=2 saves about 11.4 GiB per GPU, mostly from static model-state memory, likely dominated by tensor-parallel sharding of the DiT weights. The request-time increment is almost unchanged, so TP=2 does not materially reduce transient inference-time memory for this single-image, batch-1 workload.

recipes/Qwen/Qwen-Image.md

@@ -1,4 +1,6 @@
-# Qwen-Image for text-to-image serving with optional continuous batching on 1x A100 80GB
+# Qwen-Image
+
+> Text-to-image serving with optional step-wise continuous batching
 
 ## Summary
 
@@ -102,7 +104,7 @@ To replay the batching benefit with matched warmup:
 
 ```bash
 python benchmarks/diffusion/diffusion_benchmark_serving.py \
-  --backend vllm-omni \
+  --endpoint /v1/chat/completions \
   --dataset vbench \
   --task t2i \
   --model Qwen/Qwen-Image \
@@ -128,3 +130,108 @@ and compare the output JSON or terminal metrics.
 - Key flags: `--step-execution` enables the step-wise runtime; `--max-num-seqs` controls how many compatible requests may stay active together.
 - Keep `--max-num-seqs 1` when you want the more conservative path, when traffic is mostly single-request, or when you are debugging correctness before measuring throughput.
 - Current batching is still shape-sensitive: different step progress can co-batch, but different resolutions do not yet co-batch.
+
+### 2x B200 ModelOpt mixed FP8/NVFP4
+
+#### Environment
+
+- OS: Linux
+- Python: 3.12
+- Driver / runtime: NVIDIA CUDA environment with two B200 GPUs
+- vLLM version: Match the vLLM-Omni checkout used for deployment
+- vLLM-Omni version or commit: Use the commit that contains ModelOpt mixed
+  FP8/NVFP4 checkpoint loading for diffusion models
+
+#### Command
+
+This configuration uses a pre-quantized ModelOpt mixed checkpoint for
+Qwen-Image-2512. The validated checkpoint keeps boundary, modulation,
+normalization, and output layers in BF16, uses FP8 for sensitive attention
+projections, and uses NVFP4 for heavier MLP / attention-output linear layers.
+
+Validated checkpoint:
+
+```text
+feizhai123/qwen-image-2512-modelopt-mixed-fp8-sensitive-nvfp4-heavy
+```
+
+Start the server with the CUTLASS quantized linear backend:
+
+```bash
+CUDA_VISIBLE_DEVICES=2,3 \
+vllm serve feizhai123/qwen-image-2512-modelopt-mixed-fp8-sensitive-nvfp4-heavy \
+  --omni \
+  --host 0.0.0.0 \
+  --port 8091 \
+  --trust-remote-code \
+  --tensor-parallel-size 2 \
+  --linear-backend cutlass \
+  --force-cutlass-fp8
+```
+
+#### Verification
+
+Run a direct image generation request after the server is ready:
+
+```bash
+curl http://localhost:8091/v1/images/generations \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "feizhai123/qwen-image-2512-modelopt-mixed-fp8-sensitive-nvfp4-heavy",
+    "prompt": "A cinematic Warhammer 40,000 Ultramarine warrior in blue power armor, armor stained with alien ichor, swinging a chainsword at a Tyranid creature, gothic battlefield, dramatic lighting, high detail, no text, no watermark",
+    "negative_prompt": "blurry, low quality, text, watermark, logo, malformed armor, bad hands",
+    "size": "1024x1024",
+    "num_inference_steps": 20,
+    "true_cfg_scale": 4.0,
+    "guidance_scale": 4.0,
+    "seed": 48
+  }'
+```
+
+Check that:
+
+- The server responds on `http://localhost:8091/health`.
+- The response contains one generated image.
+- Logs show CUTLASS selection for ModelOpt FP8 / mixed-precision linear
+  layers instead of falling back to the default automatic backend choice.
+
+#### Benchmark
+
+The numbers below are loaded-once request-level measurements, not an HTTP
+concurrency benchmark. Each case loads the model once, runs warmup requests,
+then measures 100 sequential requests.
+
+```text
+Hardware: 2x B200
+Tensor parallel size: 2
+Requests: 100
+Concurrency: 1
+Resolution: 1024x1024
+Denoising steps: 20
+```
+
+| Config | Mean | P90 | P99 | Peak VRAM |
+| --- | ---: | ---: | ---: | ---: |
+| BF16 | 2.696s | 2.700s | 2.707s | 99,000 MiB |
+| ModelOpt FP8 CUTLASS | 2.663s | 2.671s | 2.676s | 85,760 MiB |
+| Mixed FP8/NVFP4 CUTLASS | 2.584s | 2.590s | 2.594s | 84,412 MiB |
+
+The same 2-GPU setup was also used to sweep NVFP4 linear backends. Direct
+CUTLASS was the fastest validated backend for this checkpoint.
+
+| Backend | Mean | P90 | P99 | Peak VRAM |
+| --- | ---: | ---: | ---: | ---: |
+| auto | 2.927s | 2.948s | 3.535s | 85,340 MiB |
+| flashinfer-cudnn | 3.523s | 3.528s | 3.766s | 84,500 MiB |
+| flashinfer-trtllm | 2.711s | 2.727s | 3.010s | 84,996 MiB |
+| flashinfer-cutlass | 2.935s | 2.950s | 3.011s | 85,340 MiB |
+| cutlass | 2.578s | 2.584s | 2.592s | 84,412 MiB |
+
+#### Notes
+
+- Use `--linear-backend cutlass` to select the validated quantized linear
+  backend from the CLI.
+- `--force-cutlass-fp8` keeps ModelOpt FP8 layers on the CUTLASS FP8 path.
+- The benchmark table intentionally reports Qwen-Image-2512 as
+  concurrency-1 request-level data; do not compare it directly with an online
+  HTTP concurrency benchmark.

recipes/Qwen/Qwen3-Omni.md

@@ -1,5 +1,7 @@
 # Qwen3-Omni
 
+> Online serving for multimodal chat
+
 ## Summary
 
 - Vendor: Qwen